Lecture Notes for MA5NO

Cohomology, Connections, Curvature and Characteristic Classes

David Mond
October 27, 2006
1 Introduction
Lets begin with a little vector analysis (also known as Physics).
Consider a point source of uid, such as a burst water-main, on a perfectly uniform plane.
The water spreads out uniformly from the source, with a uniform depth, and so we can measure
the fuid owing out from the source in units of area per second. The ow is regulated so that
2K units of area ow out each second. Once a steady state has been reached, then if the
source of the ow is taken as origin of coordinates, at the point with coordinates (x, y) the
velocity vector v(x, y) of the ow is
K
(x, y)
|(x, y)|
2
.
This is easy to see by symmetry the ow is radial, and, assuming that a steady state has
been reached, the amount of water crossing the circle of radius R centred at 0 is independent
of R. This quantity is equal to the circumference 2R times the norm of the velocity vector,
and thus this norm is inversely proportional to R; to give 2K, the constant of proportionality
must be K.
We can use this ow to derive a formula for the number of times a closed curve C in
R
2
(0, 0) winds around (0, 0). Suppose rst that C is simple, i.e. does not cross itself.
Then (0, 0) is either outside C, or inside. In the rst case, the net amount of water crossing
the curve from inside to outside per second is 2K; in the second case, it is 0 (see the diagram).
Inner
Region
1
(t), ( (t))
2
(t)
2
1

, -
(
(t))

Unit tangent and unit normal Amount out = amount in

Amount in = 0
Amount out = 2 C
1
The amount crossing C per second can also be calculated by an integral, namely the
integral, along C, of the component of the ow normal to C, with respect to arc length (see
Figure 2).
v n
.
Fluid flow across the part of curve curve between A and B
per unit time is approximately the area of the parallelogram
ABBA = base x height .
n
v B
A
A
B
= |AB|
Figure 2
If : [0, ] R
2
is a parametrisation of C by arc-length, then (

1
(t),

2
(t)) is a unit tangent,
and so (

2
(t),

1
(t)) is a unit vector normal to C. This vector is obtained from the tangent
vector by rotating it by /2 in a clockwise direction. If the area inside C is on the left, with
respect to the sense in which the curve is parametrised, then this unit vector points outwards
from this region. Thus, providing C is parametrised by arc length, the total ow out of the
region is given by
_

0
v((t)) n(t) dt. (1)
Again provided C is parametrised by arc length, this is equal to
K
_

0
(x, y) (

2
(t),

1
(t))
|(x, y)|
2
dt, (2)
which can be re-written as
K
_

0
(
y
x
2
+y
2
,
x
x
2
+y
2
)

(t)dt. (3)
Note that even if we remove the requirement that C be parametrised by arc-length, (3) (but
not (1)) still gives the right answer. Intuitively, if we go round C at twice the speed then

,
and hence the integrand, is multiplied by 2, but the domain of integration has half the length
it had before, so the integral is unchanged.
Let us now x the value of K to be 1/2. Then we have an operator on the set of simple
closed curves in R
2
(0, 0), which gives the answer 0, 1 or 1. If we now allow the curve
C to cross itself, then the range of possible answers becomes all of Z; one can see this by de-
composing an arbitrary closed curve C into a sequence of simple closed curves, and observing
that the integral is additive over disjoint domains. Whereas previously the integral gave 1 if
the origin was inside the curve C and 0 if it is outside, now it is better simply to say that the
integral measures the number of times C winds around the origin in an anticlockwise direction.
2
Exercise 1.1 Show that the integral really is unchanged under a reparametrisation
Exercise 1.2 Generalise this construction to 3 dimensions. That is, use a similar physical
argument to devise an integral formula which, given a parametrisation of a closed surface in
3-space, gives the value 1 if the origin is in the region enclosed by the surface, and 0 if it is not.
Exercise 1.3 How can we prove that these formulae really do what we claim?
2 The rst C: Cohomology
I assume you have met the denitions of smooth manifold, dierential form, the exterior
derivative of a dierential form, and the integral of a (compactly supported) dierential form
on an oriented manifold.
Given a smooth manifold M we denote by
k
(M) the space of smooth (C

) k-forms, and
by

(M) the complex

0
0
(M)
d

1
(M)
d

2
(M)
d

d

n
(M) 0,
where n = dim M and d is the exterior derivative. Complex here means that d
2
= 0, and so
imd :
k1
(M)
k
(M) kerd :
k
(M)
k+1
(M);
the quotient of kernel by image is the kth de Rham cohomology group, H
k
DR
(M). The subscript
DR indicates de Rham, and when G. de Rham rst studied this construction he thought he
had a new invariant of smooth manifolds. This was not quite so: it turns out that de Rham
cohomology coincides with cohomology dened by other means (e.g. singular cohomology)
(with coecients in R), so we will drop the subscript.
Example 2.1 The dierential form
ydx
x
2
+y
2
+
xdy
x
2
+y
2
on R
2
(0, 0) is known as d. This is because in any region of the plane in which the polar
coordinate (see the diagram below) is smooth and well-dened, then this form is indeed
equal to the exterior derivative of , as you should check (perhaps by writing = arctan(y/x)
some such expression is valid in a neighbourhood of each point in R
2
(0, 0), and any
two determinations of dier by a multiple of 2 (i.e. a constant), and thus have the same
exterior derivative.)
3

x
y
Figure 3
The name d is deliberately ironical (one might say, deliberately confusing), since the polar
coordinate is not a smooth and single-valued function on all of R
2
(0, 0). However, since
d is at least locally the exterior derivative of a function, it is closed, given that d
2
= 0. On
the other hand despite its name it is not exact: there can be no smooth function dened on
all of R
2
(0, 0) of which d is the exterior derivative. This follows, by Stokess Theorem
(2.9 below), from
Exercise 2.2 The integral in formula (3) of the Introduction is just
_
C
d.
If d were the exterior derivative of a function dened on all of R
2
(0, 0), then by Stokess
theorem its integral over any simple closed curve in R
2
(0, 0) would be zero, which, by the
discussion in the Introduction, we know not to be the case for d. The fact that d is closed
but not exact means that its class in H
1
(R
2
(0, 0)) is not 0. We shall see later that the
class of d generates (is a basis of) H
1
(R
2
(0, 0)).
We shall also see (in 2.20 below) that every closed k-form on a smooth manifold is, in
some neighbourhood U
x
of every point x, the exterior derivative of some (k 1)-form
x
,
so that our form d is not special in this regard. The point is whether or not these local
(k 1)-forms piece together to form a global (k 1)-form such that = d.
Example 2.1 continued In fact, if we break up C into pieces C
i
, each one contained in some
region in which can be represented as a smooth, single- valued function
i
, then
_
C
d =

i
_
C
i
d =

i
_
C
i
d
i
.
By Stokess theorem (which in the case of curves is just the fundamental theorem of calculus),
if P
i1
and P
i
are the end-points of C
i
, so that C
i
= P
i
P
i1
, then
_
C
i
d
i
=
i
(P
i
)
i
(P
i1
).
Note that each function
i
really is well dened and single-valued (unlike ), and so we really
can apply Stokess theorem after we have broken up the curve into bits. The choice of denition
4
of
i
makes no dierence, of course, since any two distinct versions dier by a constant. In the
curve shown here, we might dene the functions

1
: R
2
Ray 1 (/4, 7/4)

2
: R
2
Ray 2 (3/4, 5/4)

3
: R
2
Ray 3 (3/2, /2).
P
1
P
2
C
2
C
3
C
1
P = P
0 3
Ray 1
Ray 2
Ray 3
Figure 4
Observe that C
i
is contained in the domain of
i
for i = 1, 2, 3. We have
_
C
d =
_
C
1
d
1
+
_
C
2
d
2
+
_
C
3
d
3
= (
1
(P
1
)
1
(P
0
)) + (
2
(P
2
)
2
(P
1
)) + (
3
(P
3
)
3
(P
2
))
= (3/2 0) + ( (/2)) + (0 ()) = 4.
Exercise 2.3 Calculate H
0
(R) and H
1
(R). This is easy to do directly from the denition.
De Rham cohomology is functorial: if f : M N is a smooth map, then pull-back of forms
gives a map f

:
k
(N)
k
(M) for each k, which commutes with the exterior derivative d;
it follows that f

passes to the quotient to give a map H

k
(N) H
k
(M), also denoted f

, or
H
k
(f) is you are punctilious. We have (g f)

= f

, and (id
M
)

is the identity, so H
k
is
a contravariant functor.
The wedge product of forms,
j
(M)
k
(M)
j+k
(M), also passes to the quotient
(because d( ) = d + (1)
k
d) to give a product H
j
(M) H
k
(M) H
j+k
(M),
and this gives the direct sum
k
H
k
(M) a ring structure, which is also functorial.
5
Exercise 2.4 Show that the wedge product of forms passes to the quotient to dene a product
H
j
(M) H
k
(M) H
j+k
(M)
But what is a dierential form? I assume you know the denition, but is this an adequate
peg on which to hang a concept? Here are some more examples.
Example 2.5 1. Suppose that M is an n-dimensional oriented manifold equipped with a
Riemannian metric (an inner product on each tangent space T
p
M, varying smoothly with p).
Then we have the notion of orthonormal basis for T
p
M. An n-form which takes the value 1
on any (and therefore every) positively oriented orthonormal basis is called a volume form.
Indeed, if U M is a region with compact closure then the integral
_
U
vol
M
is the (oriented)
volume of M.
Exercise 2.6 If M R
2
or R
3
, we have a pre-existing notion of volume (called length
if M is 1-dimensional and area if M is 2-dimensional). Give a heuristic argument that if
we give M the Riemann metric it inherits naturally from R
n
, then
_
U
vol
M
agrees with this
pre-existing notion.
Exercise 2.7 Suppose that M R
n+1
is a smooth oriented hypersurface, with positively
oriented unit normal vector eld u(x) = (u
1
(x), . . . , u
n+1
(x)) (i.e. for each x M, a basis for
T
x
M is positive if this basis, preceded by n(x), is a positive basis for R
n+1
). Show that

i
(1)
i1
u
i
(x)dx
1


dx
i
dx
n+1
is the volume form. Hint: A positive orthonormal basis v
1
, . . . , v
n
of T
x
M gives rise to a
positive orthonormal basis u(x), v
1
, . . . , v
n
for R
n+1
.
In particular, write down an explicit volume form on the n-sphere S
n
.
Exercise 2.8 Show that if M is a compact oriented manifold (without boundary) and vol
M
its volume form, then vol
M
,= 0 in H
n
(M).
2. If M
n
R
n+1
is an oriented hypersurface, there is a Gauss map : M S
n
, (x) = the
positively oriented unit normal to M at x. Gauss used this to dene a the curvature (x) of
M at x:
(x) = lim
U{x}
volume((U))
volume(U)
.
Make a drawing to see that this is reasonable! With a bit of extra eort this can be given a
sign: +1 if locally preserves orientation, 1 if it reverses it. If is not a local dieomorphism
at x, then (x) = 0. Can you nd a heuristic argument for this? It is closely related to the
proof of Sards theorem, that the set of critical values of a smooth map has measure zero.
In fact, if vol
M
and vol
S
n are volume forms, then because the space of alternating n-tensors
on an n-dimensional vector-space is 1-dimensional,

(vol
S
n) must at each point on M be a
scalar multiple of vol
M
; and of course by continuity we nd that the scalar in question is
precisely the Gauss curvature of M.
6
3. Let X be a vector eld on R
3
, and imagine that it is the velocity eld of a uid ow. For
each point x R
3
, each pair of vectors v
1
, v
2
, and each positive real , imagine a parallelogram
P

based at x and spanned by v

1
and v
2
. Note that the area of P

, divided by
2
, is
independent of . Let
(v
1
, v
2
) = lim
0
uid ow through P

per unit time

2
.
We can give this a sign: ow through P

is positive if its projection to the normal direction

agrees with v
1
v
2
, and negative if it disagrees. Then is a smooth 2-form on R
3
.
When is closed?
4. The vector eld X also denes a 1-form : (v) = v X (where the dot means ordinary
scalar product). What is d?
What are forms for? One should, of course, ask this kind of question about every new
object one meets in mathematics. Asking it and trying to answer helps to free up intellectual
energy to devote to the topic in question, even when one does not nd a clear answer at rst.
Example 2.9 Stokess Theorem says that for an oriented manifold M
n
with boundary M,
and for any compactly supported form
n1
(M),
_
M
d =
_
M
.
This is also valid on a manifold with corners, such as a k-simplex, and leads to an interesting
pairing between de Rham cohomology and singular homology, which I now describe.
A singular k-simplex in M is a smooth map to M from the standard k-simplex
k
, dened
by

k
= (x
1
, . . . , x
k+1
) R
k+1
+
:

i
x
i
= 1.
If s is a singular k-simplex and
k
(M), we can integrate over s that is, the integral
_

k
s

() is dened.
A singular k-chain is a formal sum

i
n
i
s
i
, where the n
i
are integers and s
i
is a singular
k-simplex. Formal means you dont have to worry about what it is, only about what it does.
And what is does, is that you can integrate a k-form over it: if c is the k-chain

i
n
i
s
i
, then
we dene _
c
=

i
n
i
_
s
i
.
The collection of all singular k-chains forms a Z-module, C
k
(M). If s is a singular k-simplex,
one can think of the singular chain 1 s (or s) simply as s with the opposite orientation.
Certainly, that is how it behaves in integration: by denition of the integral over a k-chain,
_
s
=
_
s
w.
7
There are inclusions i
j
:
k1

k
for j = 1, . . . , k + 1: i
j
(x
1
, . . . , x
k
) = (x
1
, . . . , 0, . . . , x
k
)
with the zero in the j-th position. Using these we dene a boundary operator C
k
(M) C
k1
(M),
by setting
(s) =

j
(1)
j1
(s i
j
)
on a singular k-simplex s and extending linearly to formal sums of k-simplices. The coecient
(1)
j1
guarantees that the k 1-simplex s i
j
appears with the correct orientation (i.e. as
part of the boundary of s).
1
x
x
2
x
2
1
x
x
3

1
i
2

2
Figure 5: The dieomorphism i
2
gives the second edge of
2
the opposite orientation to its
boundary orientation
Abusing our notation slightly and denoting by
k
the singular (k 1)-chain in
k
equal to

(
1)
j1
i
j
, it follows that for any k 1-form dened on the standard k-simplex
k
,
_

k
=
_

k
(4)
From this and Stokes theorem we get the following simplicial version of Stokess Theorem:
Theorem 2.10 If c =

i
m
i

i
is a singular k-chain in the smooth manifold M, and

k1
(M), then
_
c
d =
_
c
.
2
Exercise 2.11 Prove this.
We nd that
2
= 0 (Exercise). Note that this doesnt mean that the double boundary is
empty, merely that in its expression as formal sum, all the coecients are 0. We dene the
k-th homology group H
k
(M) to be
ker : C
k
(M) C
k1
(M)
im : C
k+1
(M) C
k
(M)
.
1
1
Every continuous singular k-simplex in a smooth manifold can be well-enough approximated by a smooth
singular k-simplex, so this group coincides with the group dened in terms of continuous singular k-simplices
8
Every compact smooth manifold M can be nitely triangulated; that is, can be subdivided
into a nite number of simplices. Any such triangulation gives rise to a singular chain

C
k
(M), and in C
k
(N) if M is a smooth submanifold of N. If M is a manifold without
boundary, then (

) = 0.
Exercise: Regarding the torus
as a square with opposite sides
identified, find a triangulation
of the 2-torus, in which each
2-simplex is isometric to a
right-angled isosceles plane
triangle.
arcs are their edges.
vertices of the 2-simplices, and the
Triangulation of the 2-sphere with
eight 2-simplices. The black dots are
Figure 6
Exercise 2.12 Check that in each of these two cases, the boundary of the singular 1-chain

is equal to 0.
Two k-chains which dier by a boundary are said to be homologous. Stokess theorem tells
us that for any
k
(M),
_
c
=
_
c
d, and therefore
1. if d = 0 then
_
c
= 0. and
2. if c = 0 then
_
c
(d) = 0,
As a consequence, if is closed and c
1
and c
2
are homologous then
_
c
1
=
_
c
2
, and if c = 0
and
1
and
2
are cohomologous (i.e. dier by d for some ) then
_
c

1
=
_
c

2
. Thus,
integration descends to a pairing
H
k
(M) H
k
(M) R, ([c], [])
_
c
.
The de Rham Theorem (which we will not prove directly) says that by means of this pairing
H
k
(M) = Hom
Z
(H
k
(M), R)
and
H
k
(M)
Z
R = Hom
R
(H
k
(M), R).
Note that if our singular chains have coecients in R instead of in Z, these two formulae
simplify slightly.
9
Remark 2.13 If
1
and
2
are any two triangulations of a manifold M
n
, then the two n-
chains

1
and

2
are homologous. Thus they represent the same homology class in H
n
(M)
(and in H
n
(N) if M N), the fundamental class of M.
Exercise 2.14 The integral of a holomorphic function along a closed curve in C can be viewed,
in real terms (i.e. interpreting C as R
2
) as the integral of a pair of 1-forms along the curve.
Use Stokess Theorem, and the Cauchy-Riemann equations, to show that
(i) if C is a simple closed curve enclosing a domain U C in which the function f is holomor-
phic, then
_
C
f(z)dz = 0;
(ii) if C
1
and C
2
are simple closed curves which together make up the boundary of a region
within which f is holomorphic, then
_
C
1
f(z)dz =
_
C
2
f(z)dz.
What is the most general statement of this type that you can make, using Stokess Theorem
and the Cauchy-Riemann equations?
The Poincare Lemma
We now veer from the impressionistic to the technical, and prove a result which turns out to
explain, in some sense, why de Rham cohomology is the same as other standard cohomology
theories, such as singular cohomology.
Theorem 2.15 (The Poincare Lemma)
H
k
(R
n
) =
_
R if k = 0
0 otherwise
Proof We use a lemma:
Lemma 2.16 Let : R
n
R R
n
be projection, and let s : R
n
R
n
R be the inclusion
s(x) = (x, 0). Then for each k, s

: H
k
(R
n
R) H
k
(R
n
) and

: H
k
(R
n
) H
k
(R
n
R)
are mutually inverse isomorphisms.
Proof Since s = id
R
n
, it follows from functoriality that s

is the identity on
H
k
(R
n
). The other equality,

= id

R
n
R
is not obvious. To prove it, we construct
a chain homotopy between the maps (of complexes)

(R
n
R)

(R
n
R)
and the identity map on the same complex. That is, we construct a family of maps K :

k
(R
n
R)
k1
(R
n
R) such that
1

= (dK Kd).
For in that case, if
k
(M) is closed then (1

)() = d(K()) and so is zero in

cohomology, and thus 1

is the zero map on H

k
(R
n
R).
To construct K, note that every form on R
n
R is a linear combination (over R) of forms
of the following two kinds:
(1)

()f(x, t)
10
(2)

()f(x, t) dt
where is a form on R
n
. We dene K by
(1) K() = 0 if is of type (1), and
(2) if =

()f(x, t) dt is a form of type (2), then K() = ()

_
t
0
f(x, u)du
and extending by R-linearity.
If is a q-form of type (1), we have (1

)() =

()f(x, t)

()f(x, 0), while

(dK Kd)() = Kd = K
_
d

()f(x, t) + (1)
q

()
_

i
f
x
i
dx
i
+
f
t
dt
__
= K
_

(d)f(x, t) + (1)
q

( dx
i
)
f
x
i
+ (1)
q

()
f
t
dt
_
= (1)
q
pi

()
_
t
0
f
u
du
= (1)
q

()f(x, t) f(x, 0).

Thus, up to sign, dK Kd and 1

agree on forms of type (1).

An equally straightforward calculation (Exercise) proves the result also for forms of type
(2). 2
The Poincare Lemma now follows by induction, from e.g. a calculation of H

(R). 2
The same method of proof shows
Corollary 2.17 For any manifold M, the projection : M R M and the zero-section
M M R induce mutually inverse isomorphisms on cohomology. 2
Corollary 2.18 (Homotopy invariance of de Rham cohomology) If f and g are smoothly
homotopic smooth maps from M to N, then f

and g

agree on cohomology.
Proof Given a smooth homotopy F : M R N, with f(x) = F(x, 0) and g(x) =
F(x, 1), it follows that f

= s

0
F

and g

= s

1
F

, where s
0
(x) = (x, 0) and s
1
(x) = (x, 1).
Both s

0
and s

1
are inverse isomorphisms to

, and thus coincide. Hence f

= g

. 2
Corollary 2.19 If f : M N is a smooth homotopy equivalence, then f

is an isomorphism
on cohomology. 2
Remark 2.20 By functoriality, de Rham cohomology is a dieomorphism- invariant. Every
point on an n-manifold M has arbitrarily small neighbourhoods U dieomorphic to R
n
. It
follows from the Poincare Lemma that inside such a neighbourhood, every closed form is exact,
or, in other words, the complex
0 R
0
(U)
1
(U)
n
(U) 0
11
is exact. In case you are familiar with sheaf theory, this means that the complex of sheaves of
germs of dierential forms on M,

M
, is a resolution of the sheaf R
M
:
0 R
M

0
M

1
M

n
M
0
is an exact sequence of sheaves.
The Mayer Vietoris Sequence
Every n-manifold can be put together from pieces dieomorphic to R
n
. Where the number
of these pieces is nite, the Mayer-Vietoris sequence in principle gives a way of calculating
H

(M). To obtain it, we begin by constructing a short exact sequence of complexes.

Suppose that U
1
and U
2
are open subsets of the n-manifold M. We have a commutative
diagram
U
1
i
1
j
1
U
1
U
2
U
1
U
2
i
2
j
2
U
2
from which we derive an evidently exact sequence
0
k
(U
1
U
2
)
(j

1
,j

2
)

k
(U
1
)
k
(U
2
)
i

1
i

2

k
(U
1
U
2
).
Lemma 2.21 i

1
i

2
is surjective.
Proof Choose a partition of unity subordinate to the open cover U
1
, U
2
of U
1
U
2
.
That is, choose
1
and
2
, smooth functions on U
1
U
2
, such that supp
i
U
i
for i = 1, 2,
and such that
1
+
2
is identically equal to 1 on U
1
U
2
. Given a form
k
(U
1
U
2
),
the form
1
can be smoothly extended to give a form on all of U
2
by declaring it equal to 0
on U
2
U
1
U
2
, and similarly
2
extends to a form on U
1
. Regarding
1
and
2
in this
way as forms on U
2
and U
1
, we have =
1
+
2
= i

1
(
2
) i

2
(
1
), and this proves
surjectivity. 2
So the exact sequence can be augmented by adding a 0 on the right. We thus have a
short exact sequence for each k; as is easily seen, these piece together to give a short exact
sequence of complexes
0

(U
1
U
2
)
(j

1
,j

2
)

(U
1
)

(U
2
)
i

1
i

(U
1
U
2
) 0
(that is, the morphisms shown commute with the exterior derivatives in the complexes).
Proposition 2.22 A short exact sequence 0 A

0 of complexes gives rise

to a long exact sequence of cohomology
H
k
(A

) H
k
(B

) H
k
((

)
d

H
k+1
(A

) .
12
Proof The morphisms A
k
B
k
and B
k
C
k
descend to morphisms of cohomology
H
k
(A

) H
k
(B

) and H
k
(B

) H
k
(C

) because they commute with the dierentials in

the three complexes. And exactness at H
k
(B) is easy to prove.
The only mystery is the denition of d

, which is constructed by a diagram chase. Begin

with an element of H
k
(C

), which you can represent by an element c

k
of of C
k
such that
d
C
(c
k
) = 0. By the surjectivity of the map B
k
C
k
, there is a b
k
B
k
mapping to c
k
. By
commutativity, the image in C
k+1
of d
B
(b
k
) is 0; by exactness of the sequence of complexes,
there exists a
k+1
A
k+1
mapping to d
B
(b
k
). One proves:
d
A
(a
k+1
) = 0, so a
k+1
denes a cohomology class in H
k+1
(A

), and
this class is independent of all of the choices made in its construction, so we have a
well-dened map d

: H
k
(C

) H
k+1
(A

).
d

is linear, because the choices made in the construction of a

k+1
can be made linearly.
The proof of exactness at the other spots in the sequence is then straightforward, and fun. 2
The long exact sequence arising from the short exact sequence of complexes
0

(U
1
U
2
)
(j

1
,j

2
)

(U
1
)

(U
2
)
i

1
i

(U
1
U
2
) 0
is called the Mayer-Vietoris sequence.
Exercise 2.23 Go through the construction of d

in the Mayer-Vietoris sequence.

Exercise 2.24 Prove the rank theorem: if 0 A B C 0 is an exact sequence of
nite dimensional vector spaces and linear maps, then dim B = dim A+dim C. Hint: choose
bases a
1
, . . . , a
n
for A and c
1
, . . . , c
m
for C. Use them to get a basis for B.
Exercise 2.25 Show that if 0 A
1
A
n
0 is an exact sequence of nite-dimensional
vector spaces and linear maps, then

i
(1)
i
dim A
i
= 0. Hint: break the long exact sequence
up into a collection of short exact sequences, and apply the rank theorem.
Exercise 2.26 Use the Mayer-Vietoris sequence to calculate the cohomology of the circle
S
1
(Hint: Cover S
1
with open sets U
1
and U
2
each dieomrphic to R). By following the
construction of d

in Mayer- Vietoris, give as precise a description as you can of a non-zero

element of H
1
(S
1
).
Exercise 2.27 Use the Mayer-Vietoris sequence, and the previous exercise, to calculate the
cohomology of the sphere S
2
, and, inductively, of the sphere S
n
.
Exercise 2.28 Show that the inclusion S
n
R
n
0 is a homotopy-equivalence and deduce
that it induces an isomorphism in cohomology.
Exercise 2.29 Calculate the cohomology of R
n
P
1
, P
2
, where P
1
and P
2
are any two points.
Generalise to k points.
13
Exercise 2.30 For future use we need the following statement, the Five Lemma: suppose
that in the following diagram of abelian groups and homomorphisms, the rows are exact and
all the vertical maps except possibly for
k
are ismorphisms.
A
k2
A
k1
A
k
A
k+1
A
k+2


k2

k1

k

k+1

k+2
B
k2
B
k1
B
k
B
k+1
B
k+2

Then
k
also is an isomorphism.
Two Applications
Theorem 2.31 On the sphere S
n
there is a nowhere-vanishing vector eld if and only if n is
odd.
Proof
Step 1: If v is a nowhere vanishing vector eld on S
n
then after dividing by its length, we can
assume |v(x)| = 1 for all x. The map
F(x, t) = cos(t)x + sin(t)v(x)
is then a homotopy between the identity map and the antipodal map a(x) = x.
Step 2: In H
n
(S
n
), the class of the volume form is not zero, by Stokess Theorem (Exercise).
But if n is even, a

(vol
S
n) = vol
S
n (Exercise). This means that a cannot be homotopic to
the identity, since homotopic maps induce the same morphism on cohomology. 2
Theorem 2.32 Brouwers Fixed Point Theorem: Any continuous map of the unit ball D
n+1

R
n+1
to itself has a xed point.
Proof First assume f : D
n+1
D
n+1
is smooth, and has no xed point. Dene a
smooth map r : D
n+1
S
n
by mapping each x to the point where the line-segment f(x) to
x, continued, meets the boundary, S
n
. Clearly r is the identity map on S
n
; in other words, if
i : S
n
D
n+1
is inclusion, we have r i = id
S
n. It follows that the composite
H
n
(S
n
)
r

H
n
(D
n+1
)
i

H
n
(S
n
)
is the identity on H
n
(S
n
). But H
n
(R
n+1
) = 0 and H
n
(S
n
) ,= 0 so this is impossible.
If we assume that the continuous map f : D
n+1
D
n+1
has no xed point, then by
compactness there exists > 0 such that for all x D
n+1
, |f(x)x| . We can approximate
f by a smooth map g : D
n+1
D
n+1
such that |f(x) g(x)| < for all x (how?); it follows
that g also has no xed point, a contradiction. 2
3 Compactly Supported Cohomology and Poincare Duality
Heuristic Introduction
If C
1
and C
2
are two oriented closed curves on the oriented 2-torus T
2
, we can assign them
an intersection index C
1
C
2
Z (after slightly shifting one to make them transverse to one
14
another, if they are not transverse to start with), and then counting their intersection points,
with sign, as follows: let c
i
be a positive basis for T
x
C
i
, i = 1, 2; then
(C
1
C
2
)
x
=
_
1 if c
1
, c
2
is a positive basis for T
x
T
2
1 if c
1
, c
2
is a negative basis for T
x
T
2
(This slight shifting needs some justication: we really mean a homotopy of the embedding
C
i
T
2
. Any two slight shiftings of the same curve C
i
are homotopic to one another, and
thus have the same intersection index with any curve they are both transverse to.)
The set of all oriented closed curves can be made into an abelian group ((T
2
) by taking
formal sums with integer coecients. The pairing on curves extends to a pairing on formal
sums of curves, in the obvious way:

i
n
i
C
i

j
m
j
D
j
=

i,j
n
i
m
j
C
i
D j.
One checks that it is skew-symmetric.
Like any bilinear pairing, this pairing gives rise to a duality map
((T
2
) Hom
Z
(((T
2
), Z),
sending C to C and

i
C
i
to

i
C
i
. This map is not injective, however. Exercise Find a
closed curve C T
2
such that for every closed curve C

T
2
, C C

= 0
In order to get an injective duality map, we have to kill elements of ((T
2
) whose intersection
index with every curve is 0. We can do this by imposing an equivalence relation on formal
sums of closed curves: for example, homotopy. However, we get a better result if we impose a
still weaker equivalence relation, that of homology. We may as well go straight to the point:
instead of ((T
2
) we consider the group H
1
(T
2
; Z).
Poincare observed that the duality map
H
1
(T
2
; Z) Hom
Z
(H
1
(T
2
; Z), Z)
induced by the intersection pairing is an isomorphism, and for that reason it, and its gener-
alisation to other compact manifolds and other dimensions (i.e. not just curves), is known as
Poincare duality.
If M is a compact n-dimensional manifold, there is a well-dened intersection pairing
H
k
(M; Z) H
nk
(M; Z) Z;
given homology classes [c
k
] and [c
nk
] it is possible to represent them by chains c
k
and c
nk
which are in general position with respect to one another, and then count intersection points,
with sign, much as we did for closed curves on the torus. This pairing induces a duality map
H
k
(M; Z) Hom
Z
(H
nk
(M; Z), Z);
but in general this is not an isomorphism. The problem is the existence of torsion elements
in homology. A torsion element is a non-zero homology class [c] such that for some integer
15
m ,= 0, m[c] = 0. Exercise Suppose that [c] H
k
(M) is a torsion element. Show that for any
[c
nk
] H
nk
(M; Z), we have [c
k
] [c
nk
] = 0.
A familiar example of torsion element can be found in H
1
(RP
2
; Z); if we think of RP
2
as the
quotient of the 2-sphere S
2
by the equivalence relation identifying antipodal points, then we
can represent a non-zero element of H
1
(RP
2
; Z) by a half-circle in S
2
joining a pair of antipodal
points (so that it becomes a closed curve in RP
2
). One can think of this curve as the central
circle of the M obius strip; it is well known that cutting the strip along its central circle does
not disconnect it, but that cutting it along a curve which winds twice around the strip does
disconnect it. This means that this twice-winding curve is the boundary of a 2-chain either
one of the two halves into which it disconnects the M obius strip will do.
.
Figure 7: The image of the shaded band in RP
2
is a Mobius strip; its central circle (the
image of the thick black curve) is non-zero in H
1
(RP
2
- in fact its a generator.
In order for the duality map to be an isomorphism, it turns out to be necessary to kill torsion,
by taking coecients in Q rather than in Z, or, equivalently, by tensoring the homology groups
H

(M; Z) with Q.
Theorem 3.1 Homological Poincare Duality: if M
n
is a compact oriented manifold then the
intersection pairing H
k
(M; Q) H
nk
(M; Q) Q gives rise to an isomorphism
H
k
(M; Q) Hom(H
nk
(M; Q), Q).
2
Poincare Duality in de Rham Cohomology
16
An apparently quite dierent duality arises from the wedge product of dierential forms: if
M
n
is compact and oriented, we get a pairing
H
k
(M) H
nk
(M) R
([
1
], [
2
])
_
M

1

2
.
Exercise Show that this pairing is well-dened.
It turns out that this is closely related to the intersection form in homology. It is this version
of Poincare Duality that we will study in detail. Later we will see how it corresponds to
homological Poincare duality.
Our proof of (cohomological) Poincare duality for compact manifolds will go by induction
on the number of contractible open sets necessary to cover the manifold, using the Mayer
Vietoris sequence. However, in the course of assembling a compact manifold from open sets,
one has a non-compact manifold until the nal step. In order for an inductive proof to be
possible, we therefore need a version of Poincare duality which holds on non-compact manifolds.
The key is to consider a special class of dierential forms, the class of compactly supported
dierential forms. A form
k
(M) is compactly supported if outside some compact set
X M it is identically zero. Even if M is not compact, one can integrate a compactly
supported form over it; thus there is a pairing

k
(M)
nk
c
(M) R
(
1
,
2
)
_
M

1

2
.
Here
nk
c
(M) denotes the vector space of all compactly supported n k-forms.
If
j
c
(M) then d
j+1
c
(M), and thus (

c
(M), d), is a subcomplex of the de Rham
complex (

(M), d). Its cohomology groups are the compactly supported cohomology groups of
M, and are denoted H
k
c
(M). Note that although (

c
(M), d) is a subcomplex of (

(M), d), it
is not in general the case that the compactly supported cohomology space H
k
c
(M) is a subspace
of H
k
(M).
Exercise Why not?
Exercise Compute H
0
c
(R) and H
1
c
(R)
Of course, if M is compact then every form is compactly supported, so the rings H

c
(M)
and H

(M) coincide.
Theorem 3.2 Poincare Duality in de Rham Cohomology. The integration pairing H
k
(M)
H
nk
c
(M) R induces an isomorphism H
k
(M) (H
nk
c
(M))

. 2
Warning: the pairing also induces a map H
nk
c
(M) (H
k
(M))

; however, this is not neces-

sarily an isomorphism.
17
Exercise (i) Show that if a bilinear pairing V
1
V
2
R of nite dimensional vector-spaces
induces an isomorphism V
1
V

2
then it also induces an isomorphism V
2
V

1
(ii) Conclude that if H
k
(M) is nite dimensional then the integration pairing induces an iso-
morphism H
nk
c
(M) (H
k
(M))

.
(iii) Give an example of vector spaces and a pairing V
1
V
2
R inducing one isomorphism
but not the other. Hint: take V
1
to be a suitable innite dimensional vector space and V
2
to
be V

1
.
(iv) Give an example of a manifold M such that H
nk
c
(M) H
k
(M)

is not an isomorphism.
Before beginning the proof of 3.2, we note an important consequence:
Corollary 3.3 If M is a connected oriented n-manifold, then H
n
c
(M) R. If M is also
compact, then H
n
(M) R.
Proof H
0
(M) = R. 2
Exercise 3.4 Show that if M is connected, oriented and n-dimensional but is not compact
then H
n
(M) = 0.
Compactly Supported Forms and Cohomology
Our proof of Theorem 3.2 will be by induction on the number of open sets in a cover, using the
Mayer-Vietoris sequence in the inductive step. Our rst step will be to prove it for M = R
n
.
This is another reason why it is useful to introduce compactly supported cohomology: if we
were trying to prove an assertion valid only for compact manifolds, we couldnt use induction
beginning with something non-compact like R
n
.
Exercise 3.5 Suppose that M is an oriented n-dimensional manifold without boundary and
let
n
c
(M) be a compactly supported n-form on M, such that
_
M
,= 0. Show that
[] ,= 0 in H
n
c
(M).
Since H
k
(R
n
) = R if k = 0 and is 0 otherwise, we have to prove that H
k
c
(R
n
) = R if k = n
and is 0 otherwise. Not surprisingly, we need a version of the Poincare Lemma for compactly
supported cohomology. This will be somewhat dierent from the previous version. For a start,
the dimension in which the compactly supported cohomology of R
n
is non-zero changes with n
(it is n) and so in place of the isomorphism H

(R
n
) H

(R
n
R), we look for an ismorphism
H

c
(R
n
) H
+1
c
(R
n
R). In fact at no extra cost we can prove this for an arbitrary oriented
manifold M.
We construct a map H
k1
c
(M) H
k
c
(M R) as follows: let e be a compactly supported
1-form on R such that
_
R
e = 1. By the exercise above, [e] ,= 0 in H
1
c
(R) (and in fact
it generates H
1
c
(R)). Let : M R M and : M R R be projections; we dene
e

:
k1
c
(M)
k
c
(M R) by
e

() =

()

(e)
18
(which we will write simply as e

() = e.) This map descends to a map on cohomology

(Exercise), which we also denote e

, H
k1
c
(M) H
k
c
(M R). We will show that it is an
isomorphism.
Exercise Show, using iterated integration, that if
_
M
,= 0 then
_
MR
e

() ,= 0.
In fact we will not need to use the result of this exercise in our proof.
We construct the cohomological inverse to e

by integration along the bre. For each k

dene

:
k
c
(M R)
k1
c
(M)
by :
(1) if =

()f(x, t) we set

() = 0,
(2) if =

()f(x, t) dt, then

() =
_

f(x, t)dt.
Every form is a sum of forms of these two types; we extend

linearly. Note that in

0
c
(MR)
there are only forms of type (1), and so here

= 0.
Exercise 3.6 Show that d

d (so that

induces a morphism H
k
c
(MR) H
k1
c
(M)).
It is easy to see that

= 1 on H
k1
c
(M). To show that e

= 1 on H
k
c
(M R), as
in the proof of the previous version of the Poincare Lemma we construct a homotopy operator,
K :

c
(M R)
1
c
(M).
It is dened as follows:
(1) If =

()f(x, t) then K() = 0

(2) If =

()f(x, t) dt then
K() =

()
__
t

f(x, u)du A(t)

_

f(x, u)du
_
where A(t) =
_
t

e.
Lemma 3.7 1 e

= (1)
k1
(dK Kd) on
k
c
(M R).
Proof Exercise (see Bott and Tu pages 38-39). 2
Proposition 3.8 The maps

: H
k
c
(M R) H
k1
c
(M) and e

: H
k1
c
(M) H
k
c
(M R)
are mutually inverse isomorphisms. 2
Corollary 3.9
H
k
c
(R
n
) =
_
R if k = n
0 otherwise
2
19
Exercise (i) Find a generator for H
n
c
(R
n
).
(ii) Show that if U R
n
is any open set, it is possible to choose a generator for H
n
c
(R
n
)
with supp() U.
Exercise Prove Poincare Duality (3.2) for R
n
.
Good Covers
An open cover U

A
of the n-manifold M is good if for every choice
0
, . . . ,
k
A,
U

0
U

k
is dieomorphic to R
n
. Such covers are also sometimes called acyclic covers.
They are important in all cohomology theories; one can think of them as a way of assembling
M out of cohomologically trivial pieces, so that in computing the cohomology of M, only the
combinatorics of the cover (i.e. which open sets intersect which) intervenes.
Lemma 3.10 If V

B
is any open cover of M, there is a good cover U

A
such that
each U

is contained in some V

.
The cover U

A
is a renement of the cover V

B
if every U

is contained in some V

;
so the lemma says that every cover has a good renement. This property of open covers is
described by saying that they are conal in the set of all open covers of M, partially ordered
by renement.
Proof of 3.10 If M = R
n
, any cover all of whose members is convex is good since the inter-
section of an arbitrary family of convex open sets is a convex open set, and hence dieomorphic
to R
n
; such covers are evidently conal in the set of all covers. Convexity is a metric property
it involves the notion of straight line and to use the same idea on a manifold, we endow
it with a Riemannian metric, which allows us to speak of geodesics, the Riemannian equivalent
of straight lines. A set U in a Riemannian manifold is geodesically convex if for every two
points in U there is a unique geodesic in M of minimal length joining them, and moreover this
geodesic is entirely contained in U.
The intersection of any collection of geodesically convex sets is also geodesically convex. So
to obtain a good renement of an open cover V

B
, it is enough to choose, for every point
in M, some geodesically convex neighbourhood, small enough to be contained in one of the V

.
A theorem of Riemannian geometry assures us that this can be done: if x M is any point
and V is any open neighbourhood of x, there is a geodesically convex open neighbourhood U
of x contained in V . 2
Note that any sub-cover of a good cover is also good. Thus, every compact manifold has
a nite good cover.
From 3.10 we easily deduce (by way of practice in the technique):
Theorem 3.11 If the manifold M has a nite good cover, then its de Rham cohomology spaces
are all nite dimensional.
Proof Induction, using Mayer Vietoris: let U
1
, . . . , U
m
be a nite good cover of M,
and for each k = 1, . . . , m let M
k
= U
1
U
k
. Clearly the de Rham cohomology of any
manifold dioemorphic to R
n
is nite dimensional; this is the start of our induction.
20
Suppose that the de Rham cohomology of M
k
is nite dimensional; we now compare it
with the de Rham cohomology of M
k+1
= M
k
U
k+1
using the Mayer-Vietoris long exact
sequence. This made up of 3-term segments
H
q
(M
k
U
k+1
)
d

H
q+1
(M
k+1
)
(j

1
,j

2
)
H
q+1
(M
k
) H
q+1
(U
k+1
).
From this we obtain the short exact sequence
0 ker d

H
q+1
(M
k+1
) im(j

1
+j

2
) 0.
In a short exact sequence, nite dimnesionality of any two of the three non-trivial spaces
implies nite dimensionality of the third.
Besides M
k+1
, the short exact sequence here involves the three spaces M
k
, U
k+1
and M
k

U
k+1
. Do we know enough about the cohomology of all three to conclude that the middle
term is nite dimensional? In order to reach this conclusion, we must be careful to choose
the right inductive hypothesis: it is not simply that the de Rham cohomology of M
k
is nite
dimensional, but that every manifold having a good cover consisting of no more than k open
sets has nite dimensional de Rham cohomology, since this implies that the cohomology of M
k
and of M
k
U
k+1
is nite-dimensional. This is evidently true when k = 1; in view of the short
exact sequence we derived from Mayer Vietoris, if true for k it is true for k + 1. So the proof
is complete. 2
Now we proceed with the proof of Poincare duality, for manifolds having a nite good cover.
This involves comparing two Mayer-Vietoris sequences, one for ordinary de Rham cohomology
and one for compactly supported cohomology.
First we develop the Mayer-Vietoris sequence for compactly supported cohomology. Recall
the commutative diagram
U
1
i
1
j
1
U
1
U
2
U
1
U
2
i
2
j
2
U
2
Inclusion of open sets makes possible a new morphism: push-forward. If i : U V is inclusion,
dene i

:
k
(U)
k
(V ) (note: this is covariant, not contravariant) by extending by zero.
That is, if
k
c
(U) then outside some compact K U, is identically zero. Thus we can
dene a form i

() on the bigger open set V by setting i

() to be 0 at every point in V U.
V
U

21
Figure 8
It is now easy to see, as in the case of non-compactly supported forms, that the following
sequence is exact:
0
k
c
(U
1
U
2
)
(i
1
,i
2
)

k
c
(U
1
)
k
c
(U
2
)
j
1
+j
2

k
c
(U
1
U
2
) 0.
The last arrow needs a little clarication: if
k
c
(U
1
U
2
), choose
1

0
c
(U
1
) and

2

0
c
(U
2
) such that (if we extend each of them by 0 to functions on all of U
1
U
2
))

1
+
2
= 1 everywhere on U
1
U
2
. Then
i

k
c
(U
i
) for i = 1, 2 and
= j
1
(
1
) +j
2
(
2
)
As before, the maps in this short exact sequence commute with the exterior derivatives,
and thus we get a short exact sequence of complexes, and hence a long exact sequence of
cohomology, the Mayer-Vietoris sequence for compactly supported cohomology.
0 H
0
c
(U
1
U
2
) H
0
c
(U
1
) H
0
c
(U
2
) H
0
c
(U
1
U
2
)
d

H
1
c
(U
1
U
2
) H
1
c
(U
1
) H
1
c
(U
2
) H
1
c
(U
1
U
2
)
d

H
n
c
(U
1
U
2
) H
n
c
(U
1
) H
n
c
(U
2
) H
n
c
(U
1
U
2
) 0
Exercise Use Mayer Vietoris for compactly supported cohomology to compute the compactly
supported cohomology of the circle S
1
, and, inductively, of the sphere S
n
.
Now suppose that M is a manifold with a nite good cover U
1
, . . . , U
r
, and write M
k
= U
1

U
k
. By an earlier exercise, we know that the Poincare Duality morphismH
k
(M
1
) H
nk
c
(M
1
)

is an isomorphism; as inductive hypothesis we assume that it is an isomorphism for all mani-

folds having a good cover consisting of no more than k open sets (and in particular for M
k
and
M
k
U
k+1
). If we dualise the Mayer-Vietoris sequence for compactly supported cohomology,
it remains exact.
(Exercise: (i) Prove that if
A
k1

k1
A
k

k
A
k+1

is an exact sequence of vector spaces and linear maps then so is
A
k1

k1
A
k

k
A
k+1

(ii) Give an example of a short exact sequence of Abelian groups where the corresponding
statement (in which A

is replaced by Hom
Z
(A, Z)) fails. (Hint: try practically any s.e.s.))
The dualised Mayer-Vietoris for compactly supported cohomology plays to just the right
rhythm for us to compare it with Mayer-Vietoris for ordinary cohomlogy via the Poincare
duality morphism, and we get a large (and hard to typeset) diagram:
H
q
(M
k+1
) H
q
(M
k
) H
q
(U
k+1
) H
q
(M
k
U
k+1
)
d

H
q+1
(M
k+1
)

H
nq
c
(M
k+1
)

H
nq
c
(M
k
)

H
nq
c
(U
k+1
)

H
nq
c
(M
k
U
k+1
)

(d)

H
nq1
c
(M
k+1
)

22
in which the vertical maps are Poincare duality morphisms. The induction step uses the
5-lemma (cf. Exercise at the foot of page 8) to deduce the statement for M
k+1
from the state-
ments for M
k
, U
k+1
and M
k
U
k+1
. After all, the vertical map H
q
(M
k+1
) H
nq
c
(M
k+1
)

is anked by two maps on each side, each one of which is an isomorphism, by the induction
hypothesis. However, despite its naturality, the diagram is not obviously commutative, and
we need to check that in fact it is. Well, in fact its not: its only sign commutative. That
is, any two compositions of arrows starting at the same point and nishing at the same point
agree up to multiplication by 1. We establish this below. Fortunately, the proof of the
5-Lemma survives this slight weakening of its hypotheses, and so the induction is complete.
This completes the proof of 3.2, except for
Lemma 3.12 The above diagram is sign-commutative.
Proof The diagram shows three squares. The left-most one, in more detail, is
H
q
(M
k+1
)
j

1
+j

2
H
q
(M
k
) H
q
(U
k+1
)

H
nq
c
(M
k+1
)

(j
1
)

+(j
2
)

H
nq
c
(M
k
)

H
nq
c
(U
k+1
)

Checking that this is commutative is easy, provided you dont lose your grip of what the
bottom arrow is. In what follows we denote both vertical maps by PD.
Let [] H
q
(M
k+1
); then PD(j

1
+j

2
)() is a linear map from H
nq
(M
k
) H
nq
(U
k+1
)
to R; it takes ([], []) to
(
_
M
k
j

1
() ,
_
U
k+1
j

2
() ).
Meanwhile, (j
1
)

+ (j
2
)

PD([]) is also a linear map from H

nq
(M
k
) H
nq
(U
k+1
) to R;
it takes ([], []) to
(
_
M
k+1
j
1
(),
_
M
k+1
j
2
().
It is straightforward to see that the two pairs of integrals are equal.
Commutativity of the middle square is equally straightforward, and you should check it
yourself. The only diculty comes with two square involving d

and (d

,
H
q
(M
k
U
k+1
)
d

H
q+1
(M
k+1
)

H
nq
c
(M
k
U
k+1
)

(d)

H
nq1
c
(M
k+1
)

Recall the denition of d

: choose functions
1
,
2
such that supp(
1
) M
k
, supp(
2
) U
k+1
and
1
+
2
= 1 on M
k+1
. Then given [] H
q
(M
k
U
k+1
), d

[] is the cohomology class of

any form d

q+1
(M
k
U
k+1
) such that
d

=
_
d(
2
) on M
k
d(
1
) on U
k+1
.
Note that supp(d

) M
k
U
k+1
.
23
Similarly, for [] H
nq1
c
(M
k+1
), d

([]) is the cohomology class of any form d

nq
c
(M
k
U
k+1
) such that
i
1
(d

) = d(
1
) on M
k
i
2
(d

) = d(
2
) on U
k+1
Thus, for [] H
q
(M
k
U
k+1
) and [] H
nq1
c
(M
k+1
),
PD d

([])([]) =
_
M
k+1
d

=
_
M
k
U
k+1
d

(as supp(d

) M
k
U
k+1
)
=
_
M
k
U
k+1
d(
1
) =
_
M
k
U
k+1
(d
1
)
as is closed.
Meanwhile,
(d

PD([])([]) =
_
M
k
U
k+1
d

=
_
M
k
U
k+1
d(
1
) =
_
M
k
U
k+1
d
1
,
as is closed. It follows that up to sign, the two integrals coincide. 2
Our proof of Cohomological Poincare Duality, Theorem 3.2, is now complete, at least, that
is, for manifolds having a nite good cover. In fact its true even without the hypothesis on
the existence of a nite good cover - a careful proof can be found in Madsen and Tornehave,
Chapter 13. Their proof contains just one extra step, the following theorem on induction on
open sets:
Theorem 3.13 Let M
n
be a smooth n-manifold with an open cover U

A
. Suppose that
there is a collection ( of open sets of M such that
(1) (.
(2) Any open set V dieomorphic to R
n
and contained in some open set U

of the cover belongs

to (.
(3) If V
1
, V
2
and V
1
V
2
belong to ( then so does V
1
V
2
.
(4) If V
1
, V
2
, . . . is a sequence of pairwise disjoint open sets all belonging to ( then their union

i
V
i
belongs to (.
Then M (. 2
Exercise Assuming this theorem, prove 3.2 for an arbitrary smooth oriented manifold without
boundary.
Recall that the Euler Characteristic of a topological space X, (X), is dened to be

k
(1)
k
dim H
k
(X; R).
24
Exercise 3.14 Suppose that M is a compact oriented manifold without boundary, of dimen-
sion n.
(i) Prove that if n is odd then (M) = 0;
(ii) Prove that if n is of the form 4k + 2 then (M) is even. (Hint: the intersection form on
H
2k+1
(M) must be symplectic, i.e. non-degenerate and skew-symmetric.)
3.1 The Poincare dual of a submanifold
Suppose that M is an oriented n-dimensional manifold. If Y is an oriented (nk)-dimensional
submanifold (without boundary) integration over Y denes a linear map
nk
c
(M) R send-
ing to
_
Y
:
_
Y
:
nk
c
(M) R,
_
Y
.
(We ought to be writing the integral as
_
Y
i

(), where i : Y M is inclusion, but will usually

omit the i

in this context.) This map passes to the quotient in the usual way, to dene a map
_
Y
: H
nk
c
(M) R;
for if = d then _
Y
=
_
Y
=
_

= 0.
The map
_
Y
: H
nk
c
(M) R is clearly linear. In other words, it belongs to H
nk
c
(M)

.
Poincare duality 3.2 tells us that the map
H
k
(M) H
nk
c
(M)

sending [
1
] to the linear map
_
M

1
, []
_
M

1

is an isomorphism. Unwinding the denitions, this means that there exists a closed k-form

Y

k
(M) such that for all closed compactly supported forms
nk
c
(M),
_
Y
=
_
M

Y
.
Also, although this form
Y
is not unique, its cohomology class is unique; in other words any
two such forms
Y
and

Y
dier by an exact form d. The cohomology class of
Y
is called the
Poincare dual of the submanifold Y . By abuse of notation the form
Y
itself is also sometimes
called the Poincare dual of Y . It is also useful sometimes to denote the cohomology class dual
to the submanifold Y by PD(Y ).
Exercise 3.15 Suppose that Y
0
and Y
1
are k-dimensional oriented submanifolds of M, and
that Y
0
is homotopic to Y
1
in the sense that there is an oriented manifold Y and a map
F : Y [0, 1] M such that F(Y 0) = Y, F(Y 1) = Y
1
, and F
0
: Y Y
0
and
F
1
: Y Y
1
are orientation-preserving dieomorphisms.
(i) Show that for every closed form
k
c
(M),
_
Y
0
=
_
Y
1
.
(ii) Deduce that PD(Y
0
) = PD(Y
1
).
25
How can we construct such a form
Y
? In some cases it is clear. Suppose that Y = y
0
is
a single point. Here an orientation is just a sign, +1 or 1. Then
_
Y
: H
0
c
(M) R. If M is
not compact then H
0
c
(M) = 0 = H
n
(M), so PD(Y ) = 0. On the other hand, if M is compact
then H
0
c
(M) R (we assume M connected). The closed 0-forms are just constant functions.
For any function f
0
c
(M) we have
_
Y
f = f(y
0
), where the sign is the orientation of Y . In
order that
_
Y
f =
_
M

Y
f for all constant functions f, as required for
Y
to be the Poincare
dual of Y , we must thus have
f(y
0
) =
_
M

Y
f = f(y
0
)
_
M

Y
(recall that f is constant). Hence
Y
must be an n-form whose integral over M is either 1 (if
the orientation of Y is +1) or 1, if the orientation of Y is 1.
Exercise 3.16 What is the Poincare dual of M itself ?
The relation between cohomological Poincare duality and the intersection of submanifolds is
neatly expressed by the following theorem.
Theorem 3.17 If X
k
and Y
nk
are compact oriented submanifolds of the compact oriented
manifold M
n
, then
_
M
PD(X) PD(Y ) = (X Y )
M
.
This is really a special case of the following naturality property of Poincare duality.
Theorem 3.18 Suppose that X
k
and M
n
are smooth oriented manifolds and that f : X M
is a smooth map. Let Y

be an oriented submanifold of M with f transverse to Y , and give
f
1
(Y ) the transverse preimage orientation. Then
f

(PD
M
(Y )) = PD
X
(f
1
(Y )).
(Here we use subscripts to distinguish between Poincare duality on M and on X.)
Proof of 3.17 from 3.18 Assume rst that X and Y are transverse, and denote the inclusion
of X in M by i. Give i
1
(Y ) = X Y its transverse preimage orientation. Note that
(X Y )
M
=

xi
1
(Y )
sign(x).
By what we observed before concerning the Poincare dual of a point, we have
PD
X
(i
1
(Y )) =

xXY
sign(x)
0
, (5)
where
0
is a k-form on X such that
_
X

0
= 1. By 3.18,
i

(PD
M
(Y )) = PD
X
(i
1
(Y )). (6)
26
By denition of PD, we have
_
M
PD
M
(X) PD
M
(Y ) =
_
X
i

(PD
M
(Y ))
and by (6) this is equal to
_
X
PD
X
(i
1
(Y ))
which, by (5), is equal to
_
X

xXY
sign(x)
0
.
The right hand side evaluates to

xXY
sign(x), i.e. to (X Y )
M
.
If X and Y are not transverse, we can nevertheless deform the embedding of X in M in
a homotopy, so that it becomes transverse to Y . Denote the deformed X by X

. By Exercise
3.15, PD
M
(X

) = PD
M
(X). Hence
_
M
PD
M
(X) PD
M
(Y ) =
_
M
PD
M
(X

) PD
M
(Y ),
and this is equal to (X

Y )
M
by what we have proved for the transverse case. Finally,
(X Y )
M
= (X

Y )
M
; indeed, we dene (X Y )
M
by perturbing X so that it becomes
transverse to Y and then counting intersection points with their signs. 2
We will not give a complete proof of 3.18, but instead consider some special cases and give
an overview of the proof, which can be found e.g. in Bott and Tu, pages 65-67. Consider rst
the very simple case of a cylinder M = R S
1
and let Y = S
1
0. Locally we can take
coordinates t, on M, although of course is not well-dened globally. Since dierent branches
of dier by a constant, their exterior derivatives coincide, and dene a (global) 1-form d.
If = fd
1
(M), then d = 0 if and only if f/t = 0, i.e. if f = f() is independent of
t. For such a form there is no mystery in nding a form
Y
such that
_
Y
=
_
M

Y
:
simply take any smooth compactly supported function c(t) such that
_
R
c(t)dt = 1, and let

Y
= c(t)dt. Then
_
M
o
Y
=
_
M
f()c(t)dt d =
_
R
c(t)dt
_
Y
f()d =
_
Y
.
The second equality here is obtained simply by expressing the integral over M as an iterated
integral.
2
If = f(, t) d+g(, t)dt is a more general closed form, the situation is scarcely more dicult:
if i : S
1
M is the inclusion (, 0) and p : M S
1
is projection, then we know by 2.17
2
This may seem more familiar if we parametrise M by the obvious map [0, 2] R M, which is a
dieomorphism o a set of measure 0. Then
Z
M
c(t)dt f()d =
Z
R[0,2]
c(t)f()ddt =
Z
2
0
Z
R
c(t)f()dt d.
27
that p

: H
1
(Y ) H
1
(M) and i

: H
1
(Y ) H
1
(M) are mutually inverse. That is, and
p

dier by an exact form d. Now p

is a 1-form of the type we considered in the

previous paragraph. It follows that with the same
Y
as before, we have
_
M

Y
=
_
M

Y
p

=
_
Y
p

=
_
Y

(note that on Y , is equal to p

).
Now suppose Y
0
is any oriented manifold and M = R
k
Y
0
, and let Y = 0 Y
0
M.
Esssentially the same argument as above, using Fubinis Theorem (evaluation of a multiple
integral by iterated integration) shows that we can take, as
Y
, the pull-back to M of a k form
c = c(t)dt
1
dt
k
on R
k
such that
_
R
k c = 1. For
_
M
c =
_
M
c p

=
_
R
k
c(t)dt
1
dt
k
_
Y
0
p

=
_
Y
0
.
The proof of 3.18 consists of two steps. Both involve vector bundles, and it you are not familiar
with them it may be best to postpone reading the remainder of this subsection until you have
gained some familiarity with them (Chapter 15 of Madsen and Tornehave, which you will read
later in the course, is concerned with vector bundles).
The rst step is to generalise the previous observation to the situation where E is the
total space of an oriented vector bundle of rank k and Y is its zero section (in the previous
paragraph, M = R
k
Y is the total space of a trivial vector bundle). Since now E is no longer
globally a product, we have to work a little to recreate in this new situation the form
Y
we
used in the previous paragraphs. In fact it is not hard, by piecing together local constructions,
to nd a form
Y

k
(E) such that the integral of
Y
over each bre of the vector bundle
E Y is equal to 1. It is a little harder to translate to this new context the property of being
independent of the Y -variables which we used to reduce the integral of
Y
over M to
an iterated integral. To do this, we make use of the idea of integration in the bre direction
which appeared briey in the proof of the Poincare Lemma. This is a well-dened morphism
I :

(E)
k
(Y ) dened simply by integrating out the bre variables.
Denition 3.19 The Thom class of the oriented vector bundle E Y is (the cohomology
class of ) a k-form
E

k
cv
(E) such that I(
E
) = 1
M
, where 1
M
is the function on M with
contant value 1.
Here the subscript cv means compactly supported in the vertical (i.e. bre) direction.
Lemma 3.20 (i) The Thom class exists that is, there always is a form
Y
with the property
described in Denition 3.19.
(ii) If E is an oriented vector bundle over the oriented manifold Y , and we identify Y with
the zero section of E, then the Poincare dual
Y
of Y is equal to
E
. 2
I omit the proof, though it amounts to little more than using a partition of unity to piece
together local constructions like those in the special case dealt with above where M was a
28
trivial vector bundle.
The second step uses the fact that an oriented submanifold Y of the oriented manifold M
has a neighbourhood V
0
in M which is dieomorphic to the total space of its normal bundle
(Y, M) in M. The standard proof of the tubular neighbourhood theorem (see e.g. Guillemin
and Pollack, or my Manifolds lecture notes, page 34) can easily be adapted to show this.
The tubular neighbourhood theorem is proved in the simplest case where Y R
N
= M by
considering the map
F : (Y, R
N
) R
N
dened by F(y, v) = y + v, and showing that there is a neighbourhood U of the zero section
of (Y, R
N
) on which F is a dieomorphism onto a neighbourhood V of Y in R
N
. It is easy to
see that U contains a neighbourhood U
0
of the zero section which is dieomorphic to the total
space (Y, R
N
); it follows that V also contains a neighbourhoood of Y also dieomorphic to
(Y, R
N
).
We obtain the Poincare dual
Y
of Y by identifying this neighbourhood V
0
of Y in M with
(Y, M) and pushing forward the Thom class of (Y, M) to all of M by extending it by zero
(as described on page 21). It follows easily from Lemma 3.20 that the cohomology class of this
extension of the Thom class of (Y, M) is the Poincare dual of Y .
Now that we have an eective construction of
Y
, we can easily prove 3.18.
Proposition 3.21 (i) If f : X M is transverse to the submanifold Y of M then the normal
bundle of f
1
(Y ) in X is isomorphic to the pull-back by f of the normal bundle of Y in M:
(f
1
(Y ), X) f

(Y, M).
(ii) If E M is a vector bundle and f : X M is a smooth map, then the Thom class of
the vector bundle f

(E) over X is equal to the pull-back by f of the Thom class of E:

(E))
= f

(
E
).
2
The proofs of both parts of the proposition are straightforward. The proof of 3.18 follows.
3.2 The degree of a smooth map
An immediate consequence of Poincare Duality is that for any oriented n-manifold M without
boundary, H
n
c
(M) R, with the isomorphism given by integration over M.
Let M and N be smooth connected boundaryless oriented n-manifolds, and let f : M N
be a proper map (that is, the preimage in M of every compact set in N is compact). Then
f induces a pull-back map on compactly supported cohomology, f

: H
n
c
(N) H
n
c
(M). The
degree of f, deg(f) is dened by the commutative diagram
H
n
c
(N)
f

H
n
c
(M)
_
N

_
M

R
deg(f)
R
29
Proposition 3.22 If y N is a regular value of f then
deg(f) =

xf
1
(y)
sign
x
(f),
where sign
x
(f) is +1 if f preserves orientation at x and 1 if it reverses it.
Proof We need the stack of records lemma:
Lemma 3.23 If y N is a regular value of the proper map f : M
n
N
n
, there is a neigh-
bourhood V of y in N such that f
1
(y) =
r
i
U
i
, such that U
i
U
j
= for i ,= j and f
|
: U
i
V
is a dieomorphism for all i.
Proof See e.g. Madesen and Tornehave, Lemma 11.8 page 100, or Guillemin and
Pollack page ??. 2
Exercise Prove 3.22 from the stack of records lemma. 2
Corollary 3.24 deg(f) is an integer.
If M is compact, every map f : M N is of course proper.
Exercise Show that if f : M N is not surjective then deg(f) = 0.
Exercise Suppose that W is an oriented n + 1-manifold, N is an oriented boundaryless n-
manifold, and F : W N is a proper map. Let f : W N be the restriction of F. Show
that deg(f) = 0.
Given disjoint oriented simple closed curves C
1
, C
2
R
3
, dene the linking number (C
1
, C
2
)
as follows: (C
1
, C
2
) is the degree of the smooth map f : C
1
C
2
S
2
sending (x
1
, x
2
) to
(x
1
x
2
)/|x
1
x
2
|. Here C
1
C
2
is given the product orientation.
Exercise Show that if we deform C
1
to C

1
in a family C
1,t
of closed curves such that C
1,t
and
C
2
are always disjoint, then (C
1
, C
2
) = (C

1
, C
2
).
Exercise Show that if C
1
and C
2
can be untangled from one another i.e. if it is possible
to deform them to new curves C

1
, C

2
lying on opposite sides of some hyperplane H R
3
(with
the two curves disjoint from one another at all times in the deformation), then (C
1
, C
2
) = 0.
Exercise Find the linking numbers of the pairs of curves shown:
30
C
D
E
F
Figure 9
Hint: count preimages (under f : C
1
C
2
S
2
) of the unit vector in S
2
pointing out of
the plane of the paper towards you.
Mayer-Vietoris for Singular Homology
3
If X is any topological space and U
1
, U
2
are open subsets, there is a Mayer-Vietoris long exact
sequence
H
k
(U
1
U
2
)
(i
1
,i
2
)
H
k
(U
1
) H
k
(U
2
)
j
1
+j
2
H
k
(U
1
U
2
)

H
k1
(U
1
U
2
) .
Here the morphisms i

, j

are induced by the inclusions in the obvious way: if U

i
V is any
continuous map and s :
k
U is a singular k-simplex, i

(s) = i s is a singular k simplex

in V .
This long exact sequence can be constructed by almost the same procedure used for the
Mayer Vietoris sequence of cohomology. As before, there is an exact sequence of complexes
0 C

(U
1
U
2
)
(i
1
,i
2
)
C

(U
1
) C

(U
2
)
j
1
j
2
C

(U
1
U
2
);
however this time the the arrow C
k
(U
1
) C
k
(U
2
) C
k
(U
1
U
2
) is plainly not surjective for
k > 0 (see for example the diagram below). However, this lack of surjectivity can easily be
remedied by subdivision: any singular k-chain in U
1
U
2
can be subdivided into a singular
chain of the form j
1
(c
1
) +j
2
(c
2
), where c
1
C
k
(U
1
) and c
2
C
k
(U
2
). This is schematically
shown in the diagram on the right, where the 2-simplex c is subdivided into c
1
U
1
and
c
2
U
2
.
U1
U2
U1
U2
c
c1
c2
Figure 10
3
This and anything else with a star is here for entertainment only, and is not examinable
31
Building on this idea, it is possible to prove the following Subdivision Lemma:
Lemma 3.25 Suppose that U
1
and U
2
are open sets in the toplogical space X. If c C
k
(U
1

U
2
) is a k-cycle then there is a k-chain c

C
k
(U
1
U
2
) such that
(i) c c

(C
k+1
(U
1
U
2
))
(ii) every one of the singular simplices s making up c

lies either in U
1
or in U
2
; thus
c

j
1
(C
k
(U
1
)) +j
2
(C
k
(U
2
)).
2
Exercise Using the Subdivision Lemma to make up for the failure of surjectivity of j
1
+j
2
:
C
k
(U
1
) C
k
(U
2
) C
k
(U
1
U
2
), prove Mayer-Vietoris for singular homology
Exercise* Can you nd an inductive proof of the de Rham theorem using Mayer-Vietoris for
homology and cohomology?
The K unneth Formula for de Rham Cohomology
The K unneth Formula gives us a way of computing the cohomolgy of a product M N in
terms of the cohomology of the two factors M and N:
Theorem 3.26 H

(M N) H

(M)
R
H

(N)
This means the following:
(1) For each q,
H
q
(M N)

i+j=q
H
i
(M) H
j
(N)
(2) Something about the ring structure, which as yet I do not attempt to state.
In fact, where nite-dimensional vector-spaces are concerned, the mere existence of an iso-
morpphism is a pretty feeble statement; it is equivalent to their having the same dimension.
Much more interesting is a statement describing the isomorphism in concrete terms. Here it is:
For each q, the projections
M
: M N M and
N
: M N N induce linear maps

M
:
q
(M)
q
(M N) and

N
:
q
(N)
q
(M N) , and hence a bilinear map

M
:
p
(M)
q
(N)
p+q
(M N)
given by
(
p
,
q
)

M
(
p
)

N
(
q
).
This passes to the quotient to give a bilinear map H
p
(M) H
q
(N) H
p+q
(M N). By
denition of tensor product, there is a unique linear map H
p
(M) H
q
(N) H
p+q
(M N)
making the diagram
H
p
(M) H
q
(N)

H
p
(M) H
q
(N) H
p+q
(M N)
32
(in which the vertical map is the canonical projection (u, v) u v) commutative. The
K unneth theorem asserts that for each q the sum

i+j=q
H
i
(M) H
j
(N) H
q
(M N)
of all these maps is an isomorphism.
Example 3.27 The torus T
2
is dieomorphic to S
1
S
1
; so H
1
(T
2
) H
0
(S
1
) H
1
(S
1
)
H
1
(S
1
) H
0
(S
1
), and thus is 2-dimensional, and H
2
(T
2
) = H
1
(S
1
) H
1
(S
1
).
I leave you to decide what the natural (and correct) statement about multiplication is.
To prove the K unneth Theorem, we use Mayer Vietoris and induction on the number of
open sets in a good cover of M; (we prove it only when either M or N has a nite good cover).
Let U
1
, . . . , U
N
be a good cover of M, and write U
1
U
k
= M
k
.
Step 1 Its true if M itself is dieomorphic to R
n
for some n, for then MN R
n
N, and

N
: H
q
(N) H
q
(R
n
N) is an isomorphism, so that

N
: H
0
(R
n
)H
q
(N) H
q
(R
n

N) is also an isomorphism.
Step 2: the induction step Assume the theorem is true whenever M is any manifold
having a good cover consisting of no more than k open sets. We use Mayer Vietoris for the
pair M
k
, U
k+1
of open sets.
One of the goood things about exact sequences of vector spaces (as opposed to exact
sequences of abelian groups, or of modules over a more general ring) is that their exactness is
extremely robust. Not only does dualising leave them exact, as we saw in the proof of Poincare
Duality, but so does tensoring with another vector space: if A

,
A
k1

k1
A
k

k
A
k+1

is an exact sequence of vector spaces and linear maps, and V is any other vector space, then
A
k1
V

k1
1
V
A
k
V

k
1
V
A
k+1
V ,
which we denote by A

V , is exact too. Here

k
1
V
is the linear map sending a
k
v to

k
(a
k
) v.
Exercise (i) Prove this.
(ii) What happens when you tensor the short exact sequence
0 Z
2
Z Z
2
0
of abelian groups, with the abelian group Z
2
? Here tensor product is over Z, of course - were
looking at abelian grops (i.e. Z-modules) instead of vector spaces (i.e. k-modules, where k is
a eld).
33
Remark 3.28 It may be that at this point in your studies you havent really used tensor
products before. Tensor product is one of those notions which it is best simply to use without
worrying too much, at rst, about what it means. You have to remember only that if V and
W are vector spaces over the eld k then V
k
W is the vector space generated by elements
v w, where v V, w W, subject only to the rules
(v
1
+v
2
) w = v
1
w +v
2
w
v (w
1
+w
2
) = v w
1
+v w
2
(v w) = (v) w = v (w).
It is being able to pass scalars across from one factor in the tensor product to the other (the
third rule) that makes this the tensor product over k rather than over anything else. If, for
example, k is a subeld of K, then any K-vector space is also a k-vector space; and if V and
W are two such, then
V
K
W and V
k
W
are dierent spaces. Which one is bigger?
Denote by MV

k
the Mayer-Vietoris long exact sequence of the pair M
k
, U
k+1
of open sets.
Tensoring with H
j
(N) we get the exact sequence MV

k
H
j
(N).
The direct sum of any number of exact sequences is also exact. However one needs to be
a little careful to interpret this statement correctly. For example, if
0 A
1
A
2
A
3
A
4
0
and
0 B
1
B
2
B
3
0
are both exact, then so are
0 A
1
B
1
A
2
B
2
A
3
B
3
A
4
0
and
0 A
1
A
2
B
1
A
3
B
2
A
4
B
3
0.
(Check it!). We want to sum the exact sequences MV

k
H
j
(N) over j, but in such a way that
each spot in the resulting exact sequence, the sum i +j of the indices on the tensor products
H
i
(something) H
j
(N) are all the same. To make clear how to do this, imagine extending
each sequence MV

k
H
j
(N) by an innite sequence of zeros on each end. Place them in
vertical array, with MV

k
H
0
(N) at the top and MV

k
H
n
(N) (where n = dim N) at the
bottom, and then slide each row three spots to the right with respect to the one above it. This
is indicated schematically in the following diagram:
H
0
H
0
(H
0
H
0
) H
0
H
0
H
0
H
1
H
0
(H
1
H
1
) H
0
H
1
H
0
H
2
H
0
0 0 0 H
0
H
1
(H
0
H
0
) H
1
H
0
H
1
H
1
H
1
0 0 0 0 0 0 H
0
H
2
34
Now consider the grand total exact sequence you get by summing over columns. This has
the form

q

i=0
H
i
(M
k+1
) H
qi
(N)

i=0
(H
i
(M
k
) H
i
(U
k+1
)) H
qi
(N)

i=0
H
i
(M
k
U
k+1
) H
qi
(N)
Each of the sums of spaces here is precisely what is called for in the K unneth Formula, and
maps, via

N
, to H
q
(M
k+1
N), to H
q
(M
k
N)H
q
(U
k+1
N) or to H
q
(M
k
U
k+1
N),
respectively. Thus, from the grand total exact sequence we have a sequence of K unneth maps
to the Mayer Vietoris sequence for the pair M
k
N, U
k+1
N of open sets in M N. By
induction, we can assume that those mapping to H
q
(M
k
N), to H
q
(U
k+1
N) and to
H
q
(M
k
U
k+1
) N are all isomorphisms. It will follow from the 5-Lemma that the same is
true for the map
q

i=0
H
i
(M
k+1
) H
qi
(N) H
q
(M
k+1
N),
provided we can show that the diagram is commutative. This I leave to you as an Exercise.2
Exercise Devise a form of notation which makes it possible to describe the argument involving
sliding the exact sequences 3 spots to the right without drawing such large diagrams. (I am
serious).
Exercise Formulate and prove the correct statement of the K unneth formula for the multi-
plicative operation (i.e. wedge product) on cohomology.
Exercise Can you use Induction on Open Sets (3.13) to prove the K unneth Theorem without
the hypothesis that M or N has a nite good cover?
The Leray-Hirsch Theorem
The map : E M is smooth manifolds is a locally trivial bre bundle with bre F if every
point x M has a neighbourhood U such that there is a commutative diagram

1
(U)

U
U F

U
U
in which
U
is a dieomorphism and
U
: U F U is simply projection. Commutativity
of the diagram means that
U
restricts to give a map (also a dieomorphism, of course) from

1
(y) to y F (which we identify with F) for each y U.
The dieomorphism
U
is a local trivialisation of : E M over U.
35
Example 3.29 1. If E = M F then the projection : E M is a locally trivial bre
bundle. Ineed, it is globally trivial.
2. If M is a smooth n-dimensional manifold, its tangent bundle is a locally trivial bre
bundle with bre R
n
. Of course, it has additional structure: it is possible to choose the
dieomorphisms
U
so that their restriction to the bre, T
y
M =
1
(y) F = R
n
is a
linear isomorphism for every y U.
3. If : E M is a locally trivial bre bundle then is evidently a submersion. Not every
submersion is a locally trivial bre bundle, but we have
Theorem 3.30 The Ehresmann Fibration Theorem: If : E M is a proper submer-
sion, then it is a locally trivial bre bundle.
4. Exercise The map S
1
S
1
S
1
sending (z
1
, z
2
) to z
1
z
2
is a locally trivial bre bundle.
What is its bre? What about S
1
S
1
S
1
S
1
dened by (z
1
, z
2
, z
3
) z
1
z
2
z
3
?
5. Exercise Find an example of a submersion : E M which is not a locally trivial
bre bundle. Hint: cut a hole in the source of a locally trivial bre bundle.
Suppose that f : E M is any smooth map (not necessarily a locally trivial bre bundle).
The morphism f

: H

(M) H

(E) of cohomology rings makes makes H

(E) into a module

over the ring H

(M): if [] H
q
(M) and [] H
p
(E) then we can wedge f

() with and
get a new cohomology class on E. This extends linearly in an obvious way, and gives a pairing
H

(M) H

(E) H

(E).
The structure of H

(E) as a module over H

(M) can reect the dierential topology of

the map f in an interesting way.
The most uninteresting modules are the free modules: theirs is the most transparent of
structures.
Denition 3.31 Let R be a ring and M an R-module: M is a free R-module if there exists
some collection m
{}
of elements of M such that every m M can be written uniquely as
a linear combination m =

, with all except nitely many of the r

equal to 0. In this
case m
{}
is called a free basis, or free R-basis to be more precise, for M.
Every vector space is a free k-module, where k is the eld of scalars; this freeness reects the
fact that as a ring, k is pretty uninteresting.
On the other hand, an abelian group with torsion elements is not a free Z-module. Nor, in
fact, is Q, even though it has no torsion (Exercise, if you enjoy this kind of thing).
If E = M N then H

(E) is a free module over H

(M): for suppose that c

1
, . . . , c
r

H

(N) form an R-basis. Let e

1
, . . . , e
r
be the cohomology classes on MN obtained by pulling
back the c
i
, e
i
=

N
(c
i
) for i = 1, . . . , r. Then I claim that the e
i
form a free H

(M)-basis for
H

(M N).
In fact this is simply another way of stating the K unneth Theorem, and I leave it to you
to prove it - it amounts to no more than unravelling denitions.
Now let : E M be a locally trivial bre bundle, and suppose that there are cohomology
classes e
1
, . . . , e
m
H

(E) whose restriction to each bre generates the cohomology of the bre
36
(as vector space over R) (note that this is certainly the case if E = M N; the classes e
i
described two paragraphs above have this property). Then we can dene a map
: H

(M)
r
i=1
R e
i
H

(E)
by
(

j
) (

i
e
i
)

i,j

j
e
i
.
Here
r
i=1
R e
i
means the vector space of formal linear combinations

i
e
i
with real coe-
cients (i.e.
1
e
1
+ +
r
e
r
=
1
e
1
+ +
r
e
r
if and only if
i
=
i
for each i). It should
be distinguished from the subspace of H

(E) generated over R by the cohomology classes

e
1
, . . . , e
r
, which we denote Re
1
, . . . , e
r
. In this space, two disctinct linear combinations of
the e
i
may be equal.
There is an obvious map
r

i=1
R e
i
Re
1
, . . . , e
r

sending a formal linear combination

1
e
1
+ +
r
e
r
to the informal linear combination

1
e
1
+ +
r
e
r
H

(E).
Theorem 3.32 Let : E M be a locally trivial bre bundle with bre F. If there are coho-
mology classes e
1
, . . . , e
r
on E whose restriction to each bre is an R-basis for its cohomology,
then H

(E) is a free module over H

(M) with free basis e

1
, . . . , e
r
. Thus,
H

(E) H

(M) Re
1
, ..., e
r
H

(M) H

(F).
Proof The hypothesis is preserved under restriction of the base: that is, if U
M and we write E
U
for
1
(U), and denote by i the inclusion E
U
E, then the classes
i

(e
1
), . . . , i

(e
r
) in H

(E
U
) have the same property as the e
i
: their restrictions to each bre
of : E
U
U form a basis for its cohomology. If U M is an open set over which E is
trivial, then the K unneth Theorem tells us that the map
H

(U) Ri

(e
1
), . . . , i

(e
r
) H

(E
U
)
is an isomorphism.
Exercise Complete the proof, using Mayer Vietoris. 2
4 Morse Theory
We begin with a proof of the Ehresmann Fibration Theorem. Recall that we are assuming
f : E M is a proper submersion; for the moment, we assume also that E and M are mani-
folds without boundary, and that M is connected.
Step 1 At each point e E, choose a complement H
e
in T
e
E to ker d
e
f, in such a way that
the H
e
vary smoothly with e. This can be done, for example, by giving E a Riemann metric
37
and setting H
e
= (ker d
e
f)

. As f is a submersion, the restriction of d

e
f gives an isomorphism
H
e
T
f(e)
M. The H in H
e
stands for horizontal, of course; in pictures of bre bundles
one usually shows the bre as arrayed vertically above the base M, which is drawn horizontal.
The choice of this eld of horizontal subspaces H
e
is called an Ehresmann connection.
Step 2 Suppose that p and q are two points in M, and are joined by a simple smooth curve C
parametrised by : [0, a] M. We use an Ehresmann connection to dene a dieomorphism
E
p
E
q
as follows: rst,

denes a vector eld on C, which we denote by /t. We lift this

to a smooth vector eld X on f
1
(C) by setting
X(e) = unique v in H
e
such that d
e
f(v) = /t.
The theory of ODEs assures us that for every point e f
1
(C), there is an integral curve
of X passing through e. Note that if
e
is an integral curve of X, then for all t [0, a] and
e f
1
((t)),
d
e
f(

e
(t)) =

(t).
Because f is proper, for each t [0, a] there exists
t
> 0 such that for all e f
1
((t)) there is
an integral curve
e
through e satisfying
e
(t) = e and dened on the interval [t
t
, t+
t
][0, a].
f (C)
-1
E
q
E
p
p
q
C
f
M
E

e

(e)
p,q
e
Figure 11
By compactness, a nite number of the intervals [t
t
, t +
t
] cover [a, b], and it follows that
any integral curve through a point e f
1
(p) can be extended (by integral curves) to a curve
38
dened on all of [0, a]. Let e f
1
(p) and let
e
: [0, a] f
1
(C) be the (unique) integral
curves of X satisfying
e
(0) = e. Then we dene a map
p,q
: E
p
E
q
by e
e
(a). General
theory of ODEs tells us that
p,q
is smooth varying e f
1
(p) corresponds to varying the
intial values of the solution of an ODE, and the solution of an ODE depends smoothly on the
initial conditions. Moreover,
p,q
is a dieomorphism: its inverse is
q,p
.
Step 3 Given p M, choose a coordinate chart centred on x and by means of the chart
identify some neighbourhood U of x with the open unit ball in R
n
. Each point q U is then
joined to p by a unique radial segment. These segments will play the role of the curve C in
the previous step. The maps
p,q
t together to give a dieomorphism
U
: E
p
U f
1
(U)
sending (e, q) to
p,q
(e), and clearly the diagram
f
1
(U)

U
U E
p
f
U
U
is commutative.
The inverse of
U
is the dieomorphism
U
required by the denition of local trivialisa-
tion.
If M is connected then it is path-connected (Exercise) and it follows from Step 2 that all
of the bres E
p
are dieomorphic to one another. If we call any one of them F, then it follows
from what we have just done that f : E M is a locally trivial bre bundle with bre F. 2
Remark 4.1 If we allow E to have boundary, then we must insist that not only f but also
its restriction to E be proper submersions. Assuming this is the case, we take care to choose
our Ehresmann connection so that at each point e E, H
e
T
e
E. Once this is done,
the integral curve
e
through any point e E remains in E at all times, and thus the
dieomorphism
p,q
: E
p
E
q
maps (E
p
) = E
p
E to (E
q
) = E
q
E.
Example 4.2 The argument used in the proof of the Ehresmann bration theorem can easily
be used to dene the monodromy associated to a loop in the target of a proper submersion
f : E B. That is, if p B and is a loop based at p (i.e. a smooth map : [0, 1] B
with (0) = (1) = p), then our construction gives us a dieomorphism

p,p
: E
p
E
p
,
the geometric monodromy associated to . To emphasize its dependence on the choice of ,
we denote it by

. Of course,

depends on the choice of Ehresmann connection, but it can

quite easily be shown that any two dierent Ehresmann connections give rise to homotopic,
indeed isotopic, dieomorphisms. Thus, for example, the induced morphism of cohomology

: H
k
(E
p
) H
k
(E
p
)
(the cohomologoical monodromy) is independent of the choice of connection. Indeed, it is also
easy to show that the homotopy class of

depends only on the homotopy class of (as a

loop based at p).
39
Exercise (i) Let E be the M obius strip, which you can think of as the rectangle [R, R][1, 1]
subject to the equivalence relation
(R, t) (R, t).
There is an obvious map from E to its central circle B=[R, R] 0, in which each bre is
the interval [1, 1]. Let : [0, 1] B be a parametrisation of the central circle. Find the ge-
ometric monodromy dieomorphism determined by a sensible choice of Ehresmann connection.
Exercise If p : E B is a locally trivial bre bundle, and if f : S B is any smooth map,
then there is a locally trivial bre bundle f

(E B) over S with total space S

B
E :=
(s, e) S E : f(s) = p(e) and projection (s, e) s.
1. Prove local triviality of f

(E B).
2. Given a horizontal distribution H
e
on E (as in the proof of the Ehresmann bration
theorem), construct a pull-back horizontal distribution on S
B
E.
Although locally trivial bre bundles are extremely important, maps from manifolds to R are
rarely submersions. In particular, if M is compact then any smooth map f : M R must
have a global maximum and a global minimum, and these are of necessity critical points.
Morse Theory is concerned with what happens for generic smooth maps f : M R where
M is compact; that is, with how the presence of critical points of the simplest sort alters the
description of the map.
Denition 4.3 Suppose that x M is a critical point of the smooth map f : M R, and
let : U V R
n
be a chart arround x. The Hessian matrix of f at x, with respect to , is
the matrix of second order partial derivatives
[
2
(f
1
)/x
i
x
j
].
evaluated at (x).
Lemma 4.4 The following properties of the Hessian matrix of the function f at a critical
point are independent of the choice of chart: its rank, the number of negative eigenvalues, the
number of positive eigenvalues.
Proof Exercise. Note that as the Hessian is a symmetric matrix, all of its eigenvalues
are real. 2
Denition 4.5 (1) The critical point x of f : M R is non-degenerate, or a Morse critical
point, if the Hessian at x is a non-singular matrix. In this case the index of the critical point
is the number of negative eigenvalues it has.
(2) The function f : M R is a Morse function if all of its critical points are non-degenerate,
and if for no two distinct critical points p
1
, p
2
are the critical values f(p
1
), f(p
2
) equal.
40
We remark that the condition that x be a non-degenerate critical point is a transversality
condition on f
1
: it is equivalent to the map d(f
1
) : V M
1,n
(R) (sending x to
the 1 n real matrix [d
x
(f
1
]) being transverse to 0. With a bit of eort, this condition
can be rephrased without reference to charts, as a property of f, and it is a general fact
(the Thom Transversality Theorem) that most maps will satisfy any given transversality
condition. The rephrasing is as follows: any function f : M R denes a section of the
cotangent bundle T

M, the vector bundle whose bre over x M is T

x
M := (T
x
M)

, simply
sending x to d
x
f T

x
M. A critical point of f is a point x such that d
x
f lies in the zero
section M 0 T

M; then
Lemma 4.6 The critical point x of f : M R is non-degenerate if and only if df : M T

M
is transverse to the zero-section at x.
Proof The association M T

M is a functor on the category of smooth manifolds and

dieomorphisms, in the sense that a dieomorphism : M N induces a dieomorphism
T

() : T

M T

N by (x, ) ((x), (d
x
)
1
). In fact, of course, this is how one proves
in the rst place that T

M is a manifold (and a vector bundle over M). The dieomorphism

T

() maps zero section to zero-section; it follows that to prove the lemma, it is enough to
use a chart : U R
n
, where U is a chart on M around x. For the diagram
T

U
T

()
T

R
n
df d(f
1
)
U

R
n
commutes, and thus df M0 if and only if d(f
1
) R
n
0. To lighten notation, write
f
1
simply as h : R
n
R. The map dh : R
n
T

(R
n
) = R
n
R
n
maps x to (x,

h/x
i
)
(derivatives evaluated at x). It is transverse to R
n
0 if and only the projection to the second
copy of R
n
is a submersion i.e. if and only if the matrix [
2
h/x
i
x
j
] is non-singular. This
proves the lemma. 2
Morse functions are in fact open and dense (i.e. form an open and dense set) in the space
of all smooth functions M R, equipped with a sensible topology (e.g. the Whitney C

topology). We do not show that here, but we will show the following simpler result:
Theorem 4.7 Suppose that M R
N
is a smooth submanifold, and let f : M R be any
smooth map. Then for almost all a R
N
, the function f
a
: M R dened by f
a
(x) =
f(x) +a x has only non-degenerate critical points.
Proof Consider the map F : M R
N
R dened by F(x, a) = f(x) + a x. I claim
that its dierential with respect to x M, i.e. the map d
M
F : M R
N
T

M sending (x, a)
to (x, d
x
f
a
), is transverse to the zero section of T

M. The point is that d

x
f
a
( x) = df
x
( x)+a x;
this can better be written as d
x
f
a
= d
x
f +a. Now every linear map T
x
M R is equal to a
for suitable a R
N
indeed, for suitable a T
x
M if so desired. The map
_
x R
N
T

x
M
(x, a) d
x
f +a
41
is just a translate of the linear epimorphism
_
R
N
T

x
M
a a
and the derivative of this last map is the map itself, and thus an epimorphism; it follows that the
derivative of the preceding map is also an epimorphism. And this implies that d
M
F M0,
as claimed.
Now we use a well-known elementary lemma due to Rene Thom:
Lemma 4.8 Suppose that X, Y and Z are smooth manifolds and that W is a submanifold of
Z. If G : X Y Z is transverse to W, then for almost all y Y , the map G
y
: X Z
dened by G
y
(x) = G(x, y), is transverse to W.
Proof One checks (Exercise) that
f
y
W if and only if y is a regular value of : F
1
(W) Y
and then applies Sards Theorem. 2
We now use the lemma, taking X = M, Y = R
N
, Z = T

M and W = the zero section of

T

M. The map G, of course, is d

M
F : M R
N
T

M.
Applying the lemma, we deduce that for almost all a R
n
, the map (d
M
F)
a
: M T

M
is transverse to the zero section. But (d
M
F)
a
is just the derivative of f
a
; thus, we have shown
that for almost all a R
N
, df
a
: M T

M is transverse to M 0 in T

M, and thus that

f
a
has only non-degenerate critical points. 2
We will refer to a function with only non-degenerate critical points as locally Morse.
Exercise Suppose that in the previous theorem M is compact. Show that
(i) If f
a
has only non-degenerate critical points then there are only nitely many of them.
(ii) The set a R
N
: f
a
is locally Morse is open and dense in R
N
; indeed, if f
a
is locally
Morse and b is close enough to a then f
b
has the same number of critical points as f
a
.
(iii) If f
a
is locally Morse, there exist b arbitrarily close to a in R
N
such that f
b
is (globally)
Morse (i.e. such that f
b
is locally Morse and has all its critical values distinct).
(iv) Is the set of such b open and dense in R
N
?
One way of viewing this result is that any smooth function f : M R can be perturbed
an arbitrarily small amount and become a Morse function. This is in fact the key step in
showing that Morse functions are dense.
Exercise If a R
N
is a unit vector, the function h
a
: M R dened simply by x a x is
called a height function. Show that for almost all unit vectors the height function h
a
is locally
Morse.
Example 4.9 1. If S
n
R
n+1
then every height function is Morse, having just two critical
points.
42
2. If T
2
R
3
is the standard picture of a 2-torus (i.e. a doughnut lying on a table) then
the height function h
e
3
is not locally Morse, but h
e
1
and h
e
2
are Morse.
c1 c2 c3 c4
Figure 12: Over each interval (c
i
, c
i+1
) the height function on the torus is a locally
trivial bre bundle. As t passes through a critical value c
i
, the bre f
1
(t) changes.
3. If M R
N
is a smooth submanifold and p R
N
, the function
_
f
p
: M R
x |p x|
2
is a called a distance squared function. Madsen and Tornehave show in Theorem 12.4
on page 114 that for almost all p R
N
, f
p
is a locally Morse function. In fact a great
deal of information about the dierential geometry of the embedding M R
N
can be
obtained from a study of the critical points of the distance squared functions (Morse and
non-Morse).
Suppose that f : M R is a Morse function. Morse theory is concerned with how the bre
f
1
(t) and the sub-level set the manifold with boundary f
1
((, t]) change as t
passes through a critical value. A rst step is to describe them in the neighbourhood of the
critical point.
Theorem 4.10 The Morse Lemma: Suppose that the function f : M R has a non-degenerate
critical point of index k at the point p M. Then there is a chart on M around p, with
(p) = 0, such that f
1
takes the form
f
1
(x
1
, . . . , x
n
) = f(p) x
2
1
x
2
k
+x
2
k+1
+ +x
2
n
.
43
Proof A detailed proof can be found in e.g. Madsen and Tornehave, pages 117-118.
However, it is complicated and in my opinion not very illuminating.
Singularity Theory provides a much more interesting approach to this kind of question. It
falls into two parts, one of which is easy:
Step 1 A suitable chart can be found in which f
1
has the required form modulo a re-
mainder term of order 3.
This is just the classication of quadratic forms: by means of any chart mapping x to 0,
we can identify some neighbourhood of x in M with R
n
, so forget M and think of f as being
dened on some neighbourhood of 0 in R
n
. Let H be the Hessian matrix of f at 0.
If L : R
n
R
n
is a linear isomorphism, then the Hessian matrix of f L at 0 is just
L
t
HL
(we are assuming that f has a critical point at 0). The theory of real quadratic forms assures
us that we can choose appropriate L so that L
t
HL is a diagonal matrix with each diagonal
entry equal to 1, to zero or to 1. As f has a non-degenerate critical point, there can be no
zeros. Now re-order the coordinates to get all the 1s at the start. This proves Step 1.
Step 2 I claim that any function with a non-degenerate critical point is 2-determined: that
is, if I add to it any function g with vanishing rst and second order partials at 0, then there
exists a dieomorphism , dened on some neighbourhood of 0 in R
n
and mapping 0 to 0,
such that (f +g) = f. The way to prove this, due to John Mather, is to show that in fact
there is a family of dieomorphisms,
t
, depending smoothly on t, such that (f +tg)
t
= f,
with
0
= id
R
n
. To construct such a family of dieomorphisms, we integrate a certain vector
eld X constructed on a neighbourhood of (0, 0) in R
n
R.
Think of the functions f
t
= f + tg, as t varies, as together making a function F : R
n

R R, F(x, t) = f(x) +tg(x). We have

F/t = g;
thus, if we can nd X
1
, . . . , X
n
(functions of x and t) such that
g = X
1
F/x
1
+ +X
n
F/x
n
and we dene a vector eld X by
X = /t X
1
/x
1
X
n
/x
n
then X F = 0. This means that F is constant along the integral curves of X. Let

x
: R R
n
R be the integral curve of X satisfying
x
(0) = (x, 0). Because the component
of X in the t direction has constant length 1,
x
(t) has the form

x
(t) = (
x
(t), t)
for some smooth curve
x
. By what we have just said,
F(
x
(t)) = F(
x
(0)) = F(x, 0) = f(x).
44
Moreover
F(
x
(t)) = F(
x
(t), t) = f(
x
(t)) +tg(
x
(t)) = (f +tg)(
x
(t)).
As in the proof of the Ehresmann bration theorem, the map

t
: x
x
(t)
is a dieomorphism; the previous two equations show that
(f +tg)
t
= f
as required.
Arrows denote vector field
X; curves denote integral
curves
x
R
n
x 0
R
n
0
0
taxis
x 1
Figure 13
In order to guarantee that
t
(0) = 0, we make the additional requirement on X that it be
tangent to the t-axis; that is, that X
i
(0, t) = 0 for i = 1, . . . , n.
We have yet to describe the construction of the vector eld X. I will only give a sketch. It
begins with the observation that invertibility of the Hessian matrix H(f)(0) is equivalent to
the solvability of the system of linear equations in unknown functions a
i,j
(x)
x
1
= a
1,1
f
x
1
+ +a
1,n
f
xn

x
n
= a
n,1
f
x
1
+ +a
1,n
f
xn
(7)
in some neighbourhood of 0 in R
n
(by Cramers rule). For by Taylors Theorem, modulo terms
of order 2 we have
_
_
_
f
x
1

f
xn
_
_
_ = H(f)(0)
_
_
x
1

x
n
_
_
45
where H(f)(0) is the Hessian matrix of f at 0; since H(f)(0) is invertible, we get
(H(f)(0))
1
_
_
_
f
x
1

f
xn
_
_
_
=
_
_
x
1

x
n
_
_
(again modulo terms of order 2). That is, we can take the a
i,j
in the system of equations (7)
to be the entries of (H(f)(0)
1
. A result from elementary commutative algebra (Nakayamas
lemma) shows that this is good enough: if we can solve the equation (1) to rst order (i.e.
ignoring terms of order 2 and higher), then we can solve it precisely.
Techniques from elementary commutative algebra (in particular Nakayamas Lemma), plus
a patching argument (for (3)) show that this implies that
1. any function g, all of whose rst and second order partials vanish at 0, can be written
as a linear combination
g = a
1
f
x
1
+ +a
n
f
x
n
,
where the a
i
are functions of x, vanishing at 0, and dened on some neighbourhood of 0;
2. the same is true if f is replaced by f +tg;
3. the same function g can be written as a linear combination
g = A
1
F
x
1
+ +A
n
F
x
n
,
where now the A
i
are functions of x and t, dened on some neighbourhood of 0[0, 1]
in R
n
R and vanishing when x = 0.
Note that (3) is, more or less, the construction of the vector eld X (just take X
i
= A
i
for
i = 1, . . . , n).
Step 3 By Step 1, we have brought f to the form
f(p) x
2
1
x
2
k
+x
2
k+1
+ +x
2
n
+h
where h has order 3 at 0. Now apply Step 2, taking g = h. 2
Remark 4.11 The same method of proof gives a criterion for a function f to be k-determined
in some neighbourhood of a critical point: if every equation
x
i
1
1
x
in
n
= a
1
f
x
1
+ +a
n
f
x
n
in which i
1
+ +i
n
= k1 can be solved for the unknown functions a
i
in some neighbourhood
of 0, then for every function g of order k +1 at 0, f +g can be transformed to f by a suitable
change of coordinates.
The denition of k-determinacy, and this criterion for it, are expressed more succinctly in
the language of germs:
46
(i) two functions or maps f
1
, f
2
dened in some neighbourhoods of 0 R
n
have the same
germ at 0 if there is a neighbourhood of 0 on which both are dened and on which they
coincide. This is clearly an equivalence relation, and a germ is an equivalence class.
(ii) The set of germs at 0 R
n
of smooth functions is a ring, under the obvious operations of
pointwise addition and multiplication; it is denoted c
n
.
(iii) c
n
has a unique maximal ideal, m
n
, consisting of all germs whose value at 0 is 0. It
is generated by the germs of the coordinate functions x
1
, . . . , x
n
. The j-th power m
j
n
of m
n
consists of all germs of functions f such that f and all partials of order less than j vanish at
0.
(iv) The ideal in c
n
generated by the rst order partial derivatives of f is denoted J
f
.
(v) The set of germs at 0 of dieomorphisms of R
n
mapping 0 to 0 is a group under composition,
and is denoted (in this context) by . It acts on c
n
by composition on the right
4
: (f, )
f . If f
1
and f
2
are in the same orbit, we say they are right equivalent.
Then by denition, f c
n
is k-determined if, for all g m
k+1
n
, f +g is right-equivalent to
f. The criterion quoted above becomes
Theorem 4.12 (John Mather, 1968) If m
k1
n
J
f
then f is k-determined. 2
This theorem is proved by exactly the same argument we used for the Morse Lemma, which
of course is just a special case.
The point of having a normal form for a non-degenerate critical point is that it allows
us to describe the change in the topology of the level set f
1
(a) and the sub-level set M
(,a]
as a passes through a critical value. First, we note
Proposition 4.13 Suppose that M is a compact manifold without boundary and that f :
M R is a smooth function. If the interval [a
1
, a
2
] contains no critical value of f then the
manifolds with boundary M
(,a
1
]
and M
(,a
2
]
are dieomorphic.
Proof The hypothesis implies that there is an open interval (b
1
, b
2
) containing [a
1
, a
2
]
and containing no critical value of f. For any set X R, denote f
1
(X) by M
X
. By the
Ehresmann bration theorem, f
|
: M
(b
1
,b
2
)
(b
1
, b
2
) is a locally trivial bre bundle. Indeed,
it is trivial: in the argument we gave in the proof of the bration theorem, we can take U to
be all of (b
1
, b
2
). Thus M
[b
1
,a
1
]
is dieomorphic to M
[b
1
,a
2
]
(simply stretch the interval). This
dieomorphism gives rise to a dieomorphism M
(,a
1
]
= M
(,b
1
]
M
[b
1
,a
1
]
to M
(,a
2
]
=
M
(,b
1
]
M
[b
1
,a
2
]
. We have to be a little careful, though: in order to piece together the
identity dieomorphism on M
(,b
1
]
and the stretching dieomorphism M
[b
1
,a
1
]
M
[b
1
,a
2
]
,
we have to start the stretching very slowly in the vicinity of a
1
. In other words, we replace
the linear stretch sending b
1
+t in [b
1
, a
1
] to b
1
+t in [b
1
, a
2
] (where = (a
2
b
1
)/(a
1
b
1
))
by a slow-starting stretch b
1
+ t b
1
+ (t)t, where is a smooth function equal to 1 on a
neighbourhood of 0 and equal to in a neighbourhood of a
1
b
1
. 2
The value of 4.10 becomes clear with the following result:
4
Strictly speaking, in order to comply with the denition of group action, this should be (f, ) f
1
;
but were only interested in the orbits, so we ignore this dierence
47
Lemma 4.14 Let f : M R be a Morse function, with M compact, and suppose that a is a
critical value of f, with the (unique) critical point p f
1
(a) having index k. Then if > 0
is so small that a is the only critical point in [a , a +], there is a closed neighbourhood

U
of p such that
1.

U is homeomorphic to D
k
D
nk
;
2.

U M
(,a]
is contained in the level set M
a
= f
1
(a) and is homeomorphic to S
k1

D
nk
,
3. M
(,a+]
is homeomorphic to M
(,a]


U. 2
This lemma is often stated as
M
(,a+]
is homeomorphic to the space obtained from M
(,a]
by gluing D
k
D
nk
to
M
a
along (D
k
) D
nk
.
Example 4.15 1. If k = 0 (i.e. if p is a local minimum of f) then U M
(,a]
=
(since S
k1
= ): M
(,a+]
is dieomorphic to the disjoint union of M
(,a]
and an
n-ball. Compare the picture of the torus on page 32.
2. Let f : R
3
R be f(x, y, z) = x
2
+ y
2
z
2
. Here the index is 1. The following gure
shows R
3
(,1]
, R
3
(,0]
and R
3
(,1]
.
Figure 14
It is clear that R
3
(,1]
is homeomorphic to R
3
(,1]
with a cylinder D
1
D
2
glued in
along its top and bottom i.e., along (D
1
) D
2
:
homeomorphic
to
+ =
Figure 15
48
3. Let f : R
3
R be f(x, y, z) = x
2
y
2
+z
2
. Here the index is 2. The following gure
shows R
3
(,1]
, R
3
(,0]
and R
3
(,1]
(in each case the complement of the solid shown).
Figure 16
It is clear that R
3
(,1]
is dieomorphic to R
3
(,1]
with a cylinder D
2
D
1
glued in,
but this time along (D
2
) D
1
:
+ =
homeomorphic
to
Figure 17
4. Exercise In the picture of the Morse function on the torus on page 32,
(i) Find the index of each of the critical points by choosing suitable local coordinates;
(ii) For each critical point c
i
, make a drawing showing that T
2
(,c
i
+]
is homeomorphic
to T
2
(,c
i
]
with D
k
i
D
2k
i
glued in along (D
k
i
) D
2k
i
, where k
i
is the index of
the critical point lying over c
i
.
Up to now our information about how M
(,a]
changes as a passes through a critical point,
has been in terms of homeomorphisms. This can be improved to statement in terms of dieo-
morphisms, at a slight cost in terms of precision:
Lemma 4.16 Let f : M R be a Morse function, with M compact, and suppose that a is a
critical value of f, with the (unique) critical point p f
1
(a) having index k. Then if > 0
is so small that a is the only critical point in [a , a +], there is an open neighbourhood U
of p such that
1. U is dieomorphic to a contractible open set in R
n
;
2. U M
(,a)
is dieomorphic to S
k1
B
nk+1
(where B
j
is the open unit ball in R
j
);
3. M
(,a+)
is dieomorphic to M
(,a)
U. 2
49
Note the dierence between (ii) here and in 4.14: here we consider the M
(,a)
, which are
open subsets of M, in place of the manifolds with boundary M
(,a]
; the intersection of U
and M
(,a)
is an open set, and thus of dimension n; it is dieomorphic to S
k1
B
nk+1
,
which one should think of as a the product of a thickened sphere S
k1
(1, 1) with an open
ball B
nk
. It is worth trying to understand how the pictures in Example 4.15 change in this
version of the lemma: the intersection

U M
(,a]
, which lay in the level set f
1
(a) in 4.14,
is now thickened to an open set in M
(,a]
.
I will not prove either 4.14 or 4.16; a careful proof of the latter can be found in Appendix
C of Madsen and Tornehave. It will not gure in the exam!
We go on to consider the consequences of 4.16 for cohomology. The rst concerns the
Euler characteristic. Recall that if M is a manifold then the Euler characteristic (M) is, by
denition, equal to

q
(1)
q
dim H
q
(M).
Exercise (Subadditivity of the Euler characteristic) Suppose that U
1
and U
2
are open subsets
of the manifold M. Show, using Mayer Vietoris, that
(U
1
U
2
) = (U
1
) +(U
2
) (U
1
U
2
).
Proposition 4.17 In the situation of 4.16, suppose that M
(,a)
has nite dimensional
cohomology. Then so does M
(,a+)
, and
(M
(,a+)
) = (M
(,a)
) + (1)
k
.
Proof The open set U of 4.16 is contractible, so H
0
(U) = R and H
q
(U) = 0 for q > 0.
Hence (U) = 1. Now apply subadditivity of the Euler characteristic:
(M
(,a+)
) = (M
(,a)
) +(U) (M
(,a)
U).
Since M
(,a+)
U is dieomorphic to S
k1
B
nk+1
it is homotopy-equivalent to S
k1
and
its Euler characteristic is 1 + (1)
k1
. The proposition follows. 2
Corollary 4.18 Suppose that f : M R is a Morse function, with M compact. Let c
k
be
the number of critical points of f with index k. Then
(M) =

k
(1)
k
c
k
.
Proof Exercise (use 4.17, and work your way through the critical points of f, ordered
by the size of their critical values). 2
Exercise (The Morse alternative) Suppose that M is compact, the function f : M R has
a single critical value in the interval [a , a +], and the corresponding critical point is non-
degenerate and has index k. Show that for q ,= k1, k the inclusion j : M
(,a)
M
(,a+)
induces an isomorphism on qth cohomology groups, and if k 2 then either
50
1.
j

: H
k1
(M
(,a+)
) H
k1
(M
(,a+)
)
is an isomorphism and
j

: H
k
(M
(,a+)
) H
k
(M
(,a+)
)
has 1-dimensional kernel, or
2.
j

: H
k1
(M
(,a+)
) H
k1
(M
(,a+)
)
has 1-dimensional cokernel and
j

: H
k
(M
(,a+)
) H
k
(M
(,a+)
)
is an isomorphism.
What happens if k = 0 or 1?
Exercise 4.19 Recall that SO(n) is the group of orientation-preserving linear isometries of
R
n
.
1. Show that A Gl(n, R) is in SO(n) if and only if A
t
A = I
n
, where I
n
is the identity
matrix, and det A = 1.
2. Let M
n
(R) be the space of all n n real matrices, and let Sym
n
(R) be the subspace
consisting of symmetric matrices. Show that the map M
n
(R) Sym
n
(R) sending A to
A
t
A has I
n
as a regular value. (By (1), this shows that SO(n) is a manifold of dimension
n(n 1)/2).
3. Use (2) to show that the Lie algebra T
In
SO(n) is equal to the space of all skew-symmetric
n n real matrices, and go on nd an expression for T
A
SO(n).
4. Consider the trace function tr : M
n
(R) R, tr(A) = the sum of the diagonal elements
of A. Show that I
n
is a critical point of tr, and go on to show that every diagonal matrix
in SO(n) is a critical point of tr.
5. Are the diagonal matrices non-degenerate critical points of tr?
51
The Poincare-Hopf Theorem
From Corollary 4.18 one obtains an easy proof of the Poincare- Hopf theorem on the sum of
the indices of the isolated zeros of a vector eld. We now sketch this, leaving out many details
due to lack of time.
Let X be a vector eld on the manifold M, and let p M be an isolated zero of X. We
dene the index of X at p, denoted
p
(X), as follows:
choose a chart : U V R
n
around p, and dene a vector eld

(X) (the push-forward

of X by ) on V by

(X)(y) = d
x
(X(x))
where x =
1
(y). Evidently

(X) has an isolated zero at (p). For convenience let us now

assume (p) = 0. Choose a closed ball

B

with centre 0, so small it contains no other zero of

(X), and denote its boundary by S

. Then

p
(X) = deg
_

(X)
|

(X)|
: S

S
n1
_
.
Lemma 4.20 This is well-dened: it does not depend on the choice of , nor on the choice
of chart .
Proof The rst assertion is easy: for any , there is an orientation-preserving dieo-
morphism g

: S
n1
S

, which we may choose to depend smoothly on . Compose this with

the map used to dene
p
(X) to get a map S
n1
S
n1
. As the degree of g

is 1, this new
map has degree equal to the degree of the map S

S
n1
used to dene
p
(v). As varies,
we get a smooth homotopy, so the degree does not change.
The second assertion is more dicult; see Madsen and Tornehave, Lemma 11.18 page 107. 2
We say the zero p of X is non-degenerate if the derivative at (p) of

X (which we think
of as a smooth map V R
n
) is non-singular.
Lemma 4.21 p is a non-degenerate zero of X if and only if X is transverse to the zero-section
M 0 of TM at p. In this case

p
(X) = (X(M) M 0)
(p,0)
(Here (X(M) M0)
(p,0
is the (oriented) intersection number at p of the image of X in TM
and the zero section M. It is dened to be +1 if a positive basis for T
(p,0)
(X(M)), followed
by a positive basis for T
(p,0)
M, gives a positive basis for T
(p,0)
(TM), and 1 if they give a
negative basis. Although dened using the orientation of M, it is independent of the choice of
orientation, because the orientation of TM is independent of that of M, and when we reverse
the orientation of M then we also reverse the orientation of the zero section, so the changes
cancel each other out.)
Proof Madsen and Tornehave, in Lemma 11.20 page 109, show that if p is a non-
degenerate zero of X then

p
(

(X)) =
_
1 if d
(p)

(X) has positive determinant

1 if d
(p)

(X) has negative determinant

52
The result follows. 2
The key to the proof of Poincare-Hopf is the fact that a Morse function f : M R gives
rise to a vector eld on M with only non-degenerate zeros. The idea is to construct a gradient
vector eld. Recall that the gradient f(x) of a function f : R
n
R is dened by
f(x) = (
f
x
1
, . . . ,
f
x
n
).
Since on a manifold M there is in general no canonical system of coordinates it is not imme-
diately clear how to dene a gradient vector of a function on M. To make such a denition,
we single out a crucial property of the gradient of a function on R
n
: for any vector v,
< f(x), v >= d
x
f(v).
(the expression on the left is the inner product of f(x) and v). To transfer this denition to
a manifold M, we need a Riemann metric on M. Then with respect to this metric we dene
f(x) to be the unique vector in T
x
M such that for all v T
x
M,
< f(x), v >= d
x
f(v).
Another way of viewing this is to say that the Riemann metric gives rise to a bundle isomor-
phism
g : TM T

M, dened by (x, v) (x, < v, >

x
)
where <, >
x
is the scalar product in T
x
M, so that for v T
x
M < v, >: T
x
M R is a linear
map. Then
f = g
1
df.
Lemma 4.22 g is an orientation-preserving dieomorphism.
Proof Exercise You have to supply the (natural) orientations on TM and T

M. 2
Lemma 4.23 If f : M R has a non-degenerate critical point of index k at p, then
(df(M), M 0)
(p,0)
= (1)
k
.
Proof Exercise 2
Corollary 4.24 If M is compact and f : M R is a Morse function then

pM

p
(f) = (M).
Proof Exercise (using 3.20 and 3.21). 2
Lemma 4.25 If M is compact and X
1
and X
2
are any two vector elds with only non-
degenerate zeros, then

pM

p
(X
1
) =

pM

p
(X
2
).
53
Proof Both sums are equal to the self-intersection number of the diagonal in M M.
To understand this, recall that if Z is any compact oriented manifold of even dimension and
W
1
, W
2
are compact oriented submanifold with dim W
i
= (1/2)dim Z, then W
1
W
2
is dened
by moving one of them in a homotopy, say W
1
, until it becomes transverse to W
2
, and then
counting intersection points with their signs. The usual argument about homotopy-invariance
shows that the intersection number does not depend on the choice of perturbation of W
1
: any
two are homotopic to one another. This applies in particular if W
1
= W
2
. Thus, if is the
diagonal in MM, then is well-dened. It does not depend on the choice of orientation
of M (indeed, it doesnt really even require M to be orientable at all for M M is always
orientable).
To show that

pM

p
(X
1
) = , we show that there are neighbourhoods U of in
MM and V of M0 in TM, and a dieomorphism : U V such that (x, x) = (x, 0)
and such that the diagram
M M TM

U

V
proj
M
(in which the vertical arrows are inclusions and proj is projection to the rst factor) commutes.
This is easy: embed M in some R
N
(so that M M and TM are embedded in R
2N
), and
dene a map : M M TM by
(x, y) (x,
x
(y x)),
where for each x M,
x
: R
n
T
x
M is orthogonal projection.
We have
(1) For all x M, (x, x) = (x, 0), as required, and
(2) At each point (x, x) , d
(x,x)
: T
(x,x)
M M T
(x,0)
TM is an isomorphism (you
should check this). Thus, at each point of , is a local dieomorphism.
It follows that
(3) is a local dieomorphism at each point of some neighbourhood of .
As is injective on , it also follows that
(4) it is also injective on a neighbourhood of (this uses (3) also - see Guillemin and Pollack
Section 3 Exercise 10 page 19).
Thus,
(5) maps some neighbourhood U of in MM dieomomorphically to some neighbourhhod
V = (U) of M 0.
This dieomorphism turns a small perturbations of into (the image of) a vector eld.
That is, if is shifted to

, so that

U and

, then (provided : (

) M is a
dieomorphism), in an obvious way we can dene a vector eld X on M so that (

) is the
image of X. As : () M is a dieomorphism, then provided

is close enough to ,
: (

) M will indeed be a dieomorphism. By Lemma 4.21 X has only non-degenerate

54
zeros. In fact preserves intersection numbers (it respects orientations) and thus

(x,x)

)
(x,x)
=

xM
(X(M) M 0)
(x,0)
=

xM

x
(X).
The proof is complete. 2
Before completing the proof of Poincare Hopf, we need
Lemma 4.26 Suppose that the vector eld X has an isolated zero of index k at p M. If U
is a neighbourhood of p in which X has no other zero, then
(i) there exists a vector eld Y which coincides with X outside U and has only non-degenerate
zeros in U, and
(ii)

qU

q
(Y ) = k.
Proof (i) Choose a chart : U
1
V around p, with U
1
U, and let Z denote

(X).
We can think of Z simply as a map V R
n
. By Sards theorem, almost all v R
n
are
regular values of Z, and so it follows that for almost all v R
n
, the map Z + v dened
by (Z + v)(x) = Z(x) + v is transverse to 0 and thus the vector eld Z + v has only
non-degenerate zeros in V .
The basic idea of the proof now is to replace Z by Z + v, for suitable v, in some neigh-
bourhood of (p).
Choose radii
1
,
2
such that 0 <
1
<
2
and B((p),
2
) V . As Z has no zero in
V p, there exists > 0 such that for all x B((p),
2
) B((p),
1
), |Z(x)| > . Choose
a regular value v of Z with |v| < , and choose a non-negative smooth function on V such
that is identically zero outside B((p),
2
) and is identically 1 in B((p),
1
). Then if
Z

(x) = Z(x) (x)v, we have

(a) Z

= Z outside B((p),
2
)
(b) Z

has only non-degenerate zeros in V . (You should check this.)

To construct Y , we simply replace X inside U
1
by
1

(Z

).
(ii) Its clearly enough to show that

qB((p),
2
)

q
(Z

) = k. This follows from the second

exercise after corollary 2.13: inside B((p),
2
) we choose disjoint balls B
i
around the zeros of
Z

, and a ball B
0
around (p). Apply the exercise rst with W = B((p),
2
) B
0
to deduce
that

(p)
(Z) = deg
_
Z
|Z|
: B((p),
2
) S
n1
_
,
and then apply it with W = B((p),
2
)
i
B
i
to deduce that

qB((p),
2
)

q
(Z

) = deg
_
Z

|Z

|
: B((p),
2
) S
n1
_
.
55
W
B
B
B
1
2
B
3
B
B
4
5
6
S
n-1
Figure 18: The orientation of B
i
is reversed when we consider it as part of W
The right hand sides of these two equalities coincide, since Z = Z

on B((p),
2
). 2
Corollary 4.27 If X is any vector eld with only isolated zeros on the manifold M, there is
a vector eld Y on M having only non-degenerate zeros and such that

pM

p
(X) =

pM

p
(Y ).
2
The Poincare-Hopf Theorem now follows:
Theorem 4.28 Let M be a compact oriented manifold without boundary, and let X be a
vector eld on M with only isolated zeros. Then

pM

p
(X) = (M).
Proof Immediate from 4.24,4.25 and 4.27. 2
Note that 4.25 and 4.28 give another interpretation of the Euler characteristic, as the
self-intersection number of the diagonal in M M.
Exercise In the proof of Lemma 4.25 we (more or less) proved the following statement:
suppose that f : M N is a smooth map of smooth manifolds of the same dimension, that
at every point of some compact subset K of M, d
x
f is an isomorphism, and that f is 1-1 on
K. Then there is a neighbourhood U of K and a neighbourhood V of f(K) in N, such that
f
|
: U V is a dieomorphism. Use this result to prove that if M is a compact submanifold
of R
N
, then there is a dieomorphism from some neighbourhood of the zero section in the
normal bundle
5
NM to some neighbourhood V of M in R
N
. Using r :
1
: V M, where
: NM M is the bundle projection, we get a left inverse to the inclusion i : M V . If we
choose U sensibly, we can ensure that i : M V is a homotopy-equivalence, with homotopy
inverse r.
5
The normal bundle is by denition the set NM = {(x, v) M R
N
: v (TxM)

}. It is a smooth
manifold, and indeed a smooth vector bundle over M.
56
5 The complex projective space CP
n
CP
n
is the space of all complex lines through 0 in C
n+1
. Thus, it is equal, as a set, to
the quotient of C
n+1
0 by the equivalence relation which identies points (z
0
, . . . , z
n
)
and (z
0
, . . . , z
n
) (where is any non-zero complex number). We use square brackets to
denote points in CP
n
by the coordinates of any of their preimages in C
n=1
: thus [z
0
, . . . , z
n
] =
[z
0
, . . . , z
n
]. This description also enables us to endow it with the natural quotient topology:
U CP
n
is open if its preimage in C
n+1
is open (in the usual Euclidean topology). Since CP
n
is also the image under the quotient map C
n+1
0 CP
n
of the unit sphere S
2n+1
=
(z
0
, . . . , z
n
) :

i
[z
i
[
2
= 1, it is compact.
CP
n
is also a complex manifold of complex dimension n. We take
U
i
= [z
0
, . . . , z
n
] CP
n
: z
i
,= 0
and dene
i
: U
i
C
n
by

i
[z
1
, . . . , z
n
] = (z
0
/z
i
, . . . , z
n
/z
i
)
(leaving out the z
i
/z
i
, of course).
Exercise U
i
is open and
i
is a homeomorphism.
Note that CP
n
U
i
is dieomorphic to CP
n1
.
Each change of coordinates
i

1
j
is a rational map, and thus holomorphic. This shows
that CP
n
is a complex manifold. Any holomorphic map on C
n
is also a smooth map of the
underlying R
2n
, so that CP
n
is also a smooth 2n-dimensional manifold.
Exercise CP
1
is dieomorphic to the 2-sphere S
2
.
The Hopf bration
We have observed that the quotient map C
n+1
0 CP
n
restricts to a surjection S
2n+1
CP
n
.
This map is smooth, and a submersion (Exercise). As it is evidently proper, by the Ehres-
mann bration theorem it is a locally trivial bre bundle, the Hopf bration. Its bre over
CP
n
is the unit circle in the complex line . When n = 1, the Hopf bration is a map
S
3
CP
1
S
2
. H. Hopf showed that it is not homotopic to a constant map, and thus
represents a non-trivial element of
3
(S
2
). This was the one of the rst examples known of
non-vanishing of a higher homotopy group (i.e. where
k
(M) ,= 0 for some k > dim M).
Theorem 5.1
H
q
(CP
n
) =
_
R if q is even
0 if q is odd
Proof The proof is essentially an induction, based on the fact that the open set U
i

CP
n
is dieomorphic to C
n
= R
2n
and has complement in CP
n
dieomorphic to CP
n1
. To
make use of this decomposition, we need a general cohomological result:
57
Proposition 5.2 Suppose that N is a smooth compact manifold and M is a smooth compact
submanifold. Let U = N M, and let j : M N and i : U N denote the inclusions. Then
there is a long exact sequence
H
q1
(M)

H
q
c
(U)
i
H
q
(N)
j

H
q
(M) .
Proof Here i

is the push-forward dened on page 21. We prove rst

(1) j

:
q
(N)
q
(M) is an epimorphism.
(2) If
q
(M) is closed, then there exists
q
(N) such that j

() = and such that d

is identically zero on some open set in N containing M.
(3) If
q
(N) and the support of d does not meet M, and if j

() is exact, then there

exists
q1
(N) such that d is identically zero on some open neighbourhood of M in N.
We use a tubular neighbourhood of M in N. For the moment, this means just a neigh-
bourhood V of M in N together with a retraction r : V M such that r j = id
M
, and
such that j r is homotopic to the identity (so that r and j are homotopy-inverses to one
another). Choose a smooth function on N such that supp() V and 1 on a (smaller)
tubular neighbourhood V
1
of M.
To prove (1), simply observe that for
q
(M), the form r

() extends to a form
dened on all of N, and restricts to on M.
Exactly the same argument proves (2) also.
For (3), we can shrink V so that V supp(d) = . Thus,
| V
is closed, and hence represents
a cohomology class in H
q
(V ). As j : M V is a homotopy-equivalence, j

: H
q
(V ) H
q
(M)
is an isomorphism. It follows that (
| V
) is exact. Let
q1
(V ) satisfy d =
| V
. Extend
to a form dened on all of N by multiplying by the bump function described above. Then
on the neighbourhood V
1
of M, d 0.
Now we proceed with the proof of the proposition. Let
q
(N, M) denote the subset of

q
(N) consisting of forms such that j

() = 0. Then by (1) above, we have a short exact

sequence of complexes
0

(N, M)

(N)
j

(M) 0.
As usual, this gives a long exact sequence of cohomology,
H
q1
(M) H
q
(N, M) H
q
(N) H
q
(M) .
To complete the proof, we show that the chain map i

c
(U)

(N, M) gives rise to an

isomorphism on cohomology. We can then substitute H
q
c
(U) for H
q
(N, M) in the above long
exact sequence.
To prove that i

gives an isomorphism on cohomology, we use (2) and (3) above. First we

show that it is surjective. Let
q
(N, M) represent a cohomology class. Trivially, j

() is
58
exact on M, and moreover d = 0 on all of N. Thus by (3) above, there exists
q
(N)
such that d 0 on some neighbourhood of M. The form d is in i

(
q
c
(U)); it
represents the same cohomology class as in H
q
(N), but unfortunately not in H
q
(N, M)
since is not necessarily in
q1
(N, M). To remedy this, we have to show that we can indeed
modify to make it lie in
q1
(N, M). For this we use (2): we know that j

() is closed,
since dj

() = j

() = 0. Thus by (2) there exists

q1
(N) such that j

() = j

() and
such that d 0 on some neighbourhood of M. So we replace d by d(). For now

q1
(N, M), so that [] = [ d( )] in H
q
(N, M); moreover since d 0 on
a neighbourhood of M and d 0 on a neighbourhood of M, it follows that d( ) 0
on a neighbourhood of M, and thus lies in the image of i

:
q
c
(U)
q
(N, M). This shows
that i

is surjective on cohomology.
To see that it is injective, suppose that
q
c
(U) represents a cohomology class in H
q
c
(U),
and suppose that i

() is exact on N. That is,

i

() = d
for some
q1
(M, N). We have to replace in this equation by some
1

q1
c
(U), in
order to conclude that [] = 0 in H
q
c
(U). Now j

() = 0, and supp(d) M = ; it follows by

(3) above that there exists
q1
(N) such that d 0 on some neighbourhood of M
in N. This means that d
q1
c
(U); and of course d( d) = d = , so that [] = 0
in H
q
c
(U), as required. 2
Now we continue with the proof of 5.1. We apply the proposition with N = CP
n
, M =
CP
n1
. We assume the result for n = 1, since we know CP
1
S
2
. The long exact sequence
from the lemma gives us
H
q
c
(U
i
)
i
H
q
(CP
n
)
j

H
q
(CP
n1
) H
q+1
c
(U
i
) ;
as U
i
C
n
= R
2n
, we know H
q
c
(U
i
) = 0 if q ,= 2n and H
2n
c
(U
i
) = R. Thus if q < 2n 1,
j

: H
q
(CP
n
) H
q
(CP
n1
) is an isomorphism; moreover, H
q
(CP
n1
) = 0 if q > 2n 2, and
so the exact sequence gives us H
2n1
(CP
n
) = 0, H
2n
(CP
n
) = R, as stated. 2
Theorem 5.3 The cohomology algebra H

(CP
n
) is a truncated polynomial algebra
H

(CP
n
) = R[c]/(c
n+1
)
where c is any non-zero cohomology class in H
2
(CP
n
) and (c
n+1
) is the ideal generated by
c
n+1
.
Proof Its true for n = 1, so by induction assume it true for n 1. Let c
1
,= 0 in
H
2
(CP
n1
), and let c = j

(c
1
) in H
2
(CP
n
). By induction hypothesis, c
q
1
,= 0 in H
2q
(CP
n1
)
for q = 1, . . . , n 1.
The inclusion CP
n1
CP
n
induces a map
j

: H
q
(CP
n
) H
q
(CP
n1
),
59
which is an isomorphism for q 2n 2. Hence c
q
,= 0 in H
2
q(CP
n
) for q = 1, . . . , n 1, and
thus c
q
generates H
2q
(CP
n
) for q = 1, . . . , n 1. It remains only to show that c
n
,= 0. But by
Poincare duality, the wedge pairing
H
2n2
(CP
n
) H
2
(CP
n
) R
is non-degenerate; since H
2n2
(CP
n
) = R c
n1
and H
2
(CP
n
) = R c, and the pairing takes
(c
n1
, c) to
_
CP
n
c
n1
c =
_
CP
n
c
n
,
we must have c
n
,= 0. 2
Like any complex manifold, CP
n
acquires a canonical orientation coming from the complex
structure on its tangent spaces. This orientation is dened as follows. Suppose that V is a
nite dimensional complex vector space. Then V is also a real vector space. In what follows
it will be convenient to denote V , thought of as a real vector space, by rV , though of a course
as a set it is the same as V . If e
1
, . . . , e
n
is any basis for V , then e
1
, ie
1
, . . . , e
n
, ie
n
is a (real)
basis for rV . Any complex-linear map T : V V can also be thought of as a real linear map
rV rV , in which case it will be denoted rT.
Lemma 5.4 det(rT) = [ det(T)[
2
.
Proof Induction on n = dim
C
V . If n = 1, then T is just multiplication by a complex
number z = x +iy. With respect to a basis e, ie of rV , rT has matrix
_
x y
y x
_
and thus det(rT) = x
2
+y
2
= [z[
2
, as claimed.
Now assume the result true for all spaces of dimension n1, and suppose V has dimension
n. Choose a complex line V such that T() ( can be any line generated by an
eigenvector). Then T induces C-linear maps T
0
: and T
1
: V/ V/. By induction
hypothesis, we can assume det(rT
0
) = [ det(T
0
)[
2
and det(rT
1
) = [ det(T
1
)[
2
. Finally, we have
det(T) = det(T
0
) det(T
1
) and det(rT) = det(rT
0
) det(rT
1
),
and thus det(rT) = [ det(T)[
2
. 2
From the lemma it follows that if u
1
, . . . , u
n
and v
1
, . . . , v
n
are bases for V , then the two
bases u
1
, iu
1
, . . . , u
n
, iu
n
and v
1
, iv
1
, . . . , v
n
, iv
n
for rV are equivalent, i.e., the change of basis
matrix is positive. For this matrix is the same as the matrix of the map rT : rV rV arising
from the complex linear map T : V V sending a
i
to b
i
for i = 1, . . . , n. By the lemma,
det(rT) > 0.
Corollary 5.5 (1) If V is a complex vector space of dimension n then rV has a well-dened
orientation in which every basis v
1
, iv
1
, . . . , v
n
, iv
n
is positive.
(2) If T : V
1
V
2
is a complex-linear isomorphism then rT : rV
1
rV
2
preserves this orien-
tation.
60
Proof (1) is proved in the paragraph preceding the lemma. For (2), just observe that
tT sends a basis of rV
1
of the form v
1
, iv
1
, . . . , v
n
, iv
n
to a basis of rV
2
of the same form. 2
Now suppose that M is a complex manifold. That is, it is a 2n-dimensional manifold
with an atlas U

A
with the property that if we identify R
2n
with C
n
then each of the
crossover maps

1 is complex analytic. Write

simply as h. Like any smooth

map, h has a derivative d
x
h at each point; and in fact d
x
h is not only real linear but complex
linear. It follows that each T
p
M has a natural complex vector space structure.
It follows from the corollary that the derivative of h preserves the orientation of R
2n
coming
from C
n
; as all the crossover maps preserve this orientation, it thus denes an orientation on
M.
Despite this general fact, we consider the complex structure on CP
n
from a slightly dierent
point of view, which enables us also to dene a Hermitian metric on each tangent space T
p
CP
n
.
Let < , > denote the usual hermitian inner product on C
n+1
,
< (w
1
, . . . , w
n+1
), (z
1
, . . . , z
n+1
) >=

j
w
j
z
j
.
Then Re < , > is the usual real inner product on the underlying R
2n+2
(check this!). For
v C
n+1
let (Cv)

denote the orthogonal complement of Cv with respect to < , >. It is an

n-dimensional complex subspace of C
n+1
. Similarly, let (Rv)

be the orthogonal complement

of Rv with respect to the real inner product Re < , >. Evidently, (Cv)

(Rv)

.
Lemma 5.6 (1) Let p CP
n
and v
1
(p) S
2n+1
. Then there is an open neighbourhood
U of p in CP
n
and a smooth map s : U S
2n+1
such that s(p) = v and s = id
U
.
(2) Let v S
2n+1
and p = (v). The dierential d
v
induces an R-linear isomorphism
(Cv)

T
p
(CP
n
).
(3) T
p
CP
n
has a well-dened structure as complex vector space with a Hermitian inner product,
with respect to which the isomorphism of (2) is a C-linear isometry.
Proof Choose U +U
j
such that p U
j
. Dene s
j
: U
j
S
2n+1
by
s
j
([z
0
, . . . , z
n
]) =
_
n

k=0
[z
k
[
2
_
1/2
(z
0
, . . . , z
k
)
where homogeneous coordinates are chosen so that z
j
= 1. In other words, s
j
(q) is just
j
(q),
translated by 1 in the e
j
direction, and then divided by its norm, to pull it onto the unit
sphere.
(Exercise Show that (thinking of q CP
n
as a line through 0 in C
n+1
) s
j
(q) is the unique
point of q C
j
1 C
nj
, divided by its norm.)
Note that s
j
= id
U
. We then dene s by multiplying s
j
by the unique S
1
such that
s
j
(p) = v. Clearly s = id
U
. This proves (1). It follows by the chain rule that
d
v
: T
v
S
2n+1
T
p
CP
n
61
is surjective. Its kernel contains iv, for iv spans the tangent space at v to the unit circle
e
it
v : 0 t 2 though v in S
2n+1
, all of which is mapped to p by . As kerd
v
is
1-dimensional, it is spanned by iv, and thus d
v
is an isomorphism on any complement in
T
v
S
2n+1
to the linear span of iv. Now (Cv)

T
v
S
2n+1
(since T
v
S
2n+1
= (Rv)

), and
moreover (Cv)

= (Civ)

, so (Cv)

is a complement to iv in T
v
S
2n+1
. This proves (2).
We use the isomorphism d
v

|
: (Cv)

T
p
CP
n
to give T
p
CP
n
a complex structure in the
obvious way : for C and p T
p
CP
n
, p = d v
|
((d
v

|
)
1
( p)). All this uses is that d
v

|
is a bijection. However, because d
v
is a real linear isomorphism, this complex structure is
compatible with its pre-existing structure as real vector space (remember that R C).
By an analagous procedure we transfer to T
p
CP
n
the Hermitian metric from (Cv)

.
Only one thing remains to check: that the complex structure and Hermitian metric we
have just constructed on T
p
CP
n
do not depend on the choice of v in
1
(p). In fact, if v

is
any other point in S
2n+1
mapping to p, then there exists C of unit modulus such that
v

= v. The map on C
n+1
dened by multiplying by restricts to a map L : S
2n+1
S
2n+1
,
and we have a commutative diagram
T
v
S
2n+1
dvL
T
v
S
2n+1
d
v
d
v

T
p
CP
n
d
v
L = L restricts to a complex linear isomorphism (Cv)

(Cv

, and it follows that

the complex structure on T
p
CP
n
is independent of choice of v, and thus well dened;
L is a Hermitian isometry, and it follows that the Hermitian metric on T
p
CP
n
is well
dened.
2
Proposition 5.7 Let V be an n-dimensional complex vector space with Hermitian inner prod-
uct < , >. Then
(1) g(v
1
, v
2
) = Re < v
1
, v
2
> denes an inner product on rV .
(2) (v
1
, v
2
) = g(iv
1
, v
2
) = Im < v
1
, v
2
> denes an element of Alt
2
(rV ).
(3) If vol Alt
2n
(rV ) denotes the volume form with respect to the metric g and the orientation
dened in 5.5, then
n
= n!vol.
Proof (1) and (2) I leave as an Exercise. For (3), suppose that v
1
, . . . , v
n
is an
orthonormal basis for V with respect to < , >. Then v
1
, iv
1
, . . . , v
n
, iv
n
is an orthonormal
basis for rV with respect to g (check this!). Let
1
,
1
, . . . ,
n
,
n
denote the dual basis for
(rV )

= Alt
1
(rV ). We have (v
j
, iv
j
) = (iv
j
, v
j
) = 1, and vanishes on all other pairs of
bais vectors. Thus,
=
n

j=1

j

j
.
62
Moreover, vol =
1

1

n

n
(both sides are 1 on (v
1
, iv
1
, . . . , v
n
, iv
n
)). The equality

n
= n!vol can now easily be proved by induction on n. 2
Now recall that on each tangent space T
p
CP
n
we have a complex structure and a Hermitian
inner product. We dene
a 2-form on CP
n
by taking the alternating 2-tensor dened in 5.7 on each tangent
space T
p
CP
n
, and
a Riemannian metric (the Fubini-Study metric) on CP
n
by taking the inner product g
dened in 5.7 on each tangent space T
p
CP
n
.
Theorem 5.8 and g are, respectively, a smooth 2-form and a Riemannian metric. More-
over, is closed, and

n
= n!vol
CP
n
where vol
CP
n
is the volume form determined by g and the orientation coming from the complex
structure.
Proof Smoothness of both and g is clear: both come from the Hermitian metric
on the spaces T
p
CP
n
which is induced by the R-linear ismorphism d
v
: (Cv)

T
p
CP
n
,
and thus depends smoothly on p. To see that is closed, we show that in fact it is the
pull-back, via the smooth section s of that we dened in 5.6, of a closed 2-form on S
2n+1
.
The 2-form in question is the restriction to S
2n+1
of the canonical 2-form
C
n+1 dened by

C
n+1( x
1
, z
2
) = Im < z
1
, z
2
>.
Exercise Show that
C
n+1 =

n
j=0
dx
j
dy
j
on C
n+1
, where x
j
and y
j
are the real and
imaginary parts of the j-th complex coordinate function z
j
. (This form is obviously closed on
C
n+1
and therefore on S
2n+1
.)
Suppose that v
1
, v
2
(Cv)

. Via d
v
they map to p
1
, p
2
T
p
CP
n
. We have
( p
1
, p
2
) = Im < p
1
, p
2
>= Im < v
1
, v
2
>=
C
n+1( v
1
, v
2
)
This almost shows what we want; the only problem is that although p
k
= d
v
( v
k
) for k = 1, 2,
we dont know that v
k
= d
p
s( p
k
), which is apparently what we need to conclude that =
s

(
C
n+1). What we do know is that d
p
s( p
1
) = v
1
+
1
iv, d
p
s( p
2
) = v
2
+
2
iv for some real
scalars
1
, a
2
. But now since v
k
(Civ)

, we have
Im < v
1
+
1
iv, v
2
+
2
iv >= Im(< v
1
, v
2
> + <
1
iv,
2
iv >) = Im < v
1
, v
2
>
and we have won.
The fact that
n
= n!vol follows directly from the last part of 5.7. 2
We conclude this section by remarking that the closed 2-form we have dened on CP
n
generates its cohomology ring. This is now clear:
n
= n!vol and therefore
_
CP
n
n
,= 0, so
[
n
] ,= 0 in H
2n
(CP
n
). It follows that [] ,= 0 in H
2
(CP
n
).
63
Exercise Show that if n is even, there can be no orientation-reversing dieomorphismCP
n
CP
n
;
show also that if n is odd then the map induced by complex conjugation,
[z
0
, . . . , z
n
] [ z
0
, . . . , z
n
],
reverses orientation.
Exercise Show that any smooth map f : CP
m
CP
n
induces the zero map on cohomology if
m > n.(Hint: use the structure of the cohomology algebra described in 5.3).
Note (for the purposes of the next exercise) that any continuous map between manifolds
can be approximated by a smooth map homotopic to it, and that any two smooth approxi-
mations both homotopic to f are therefore homotopic to one another, and hence induce the
same morphism on cohomology. This morphism is therefore determined by f alone. In this
way we can associate to a merely continuous map between manifolds a well-dened morphism
on cohomology. The result of the last exercise still holds, of course, when f : CP
m
CP
n
is a
continuous map.
Exercise

Show that the Hopf map S

2n+1
CP
n
is not homotopic to a constant map. Here
is a sketch of how to do it (taken from Madsen and Tornehave): suppose that F : S
2n+1

[0, 1] CP
n
is a homotopy from a constant map F
0
to the Hopf map . Then it is possible
to extend to a continuous map g : D
2n+2
CP
n
, dened by g(x) = F(x/|x|, |x|). Now
dene h : D
2n+2
CP
n+1
by
h(z
0
, . . . , z
n
) = [z
0
, . . . , z
n
,
_
_
1
n

j=0
[z
j
[
2
_
_
1/2
].
Observe that h maps the interior B
2n+2
of D
2n+2
bijectively onto U
n+1
= [w
0
, . . . , w
n+1
]
CP
n+1
: w
n
,= 0 = CP
n+1
CP
n
. Also, h
| S
2n+1 is the composite of the Hopf map :
S
2n+1
CP
n
with the inclusion j : CP
n
CP
n+1
. Now nd f : CP
n+1
CP
n
such that
f j = id
CP
n
, and pass to de Rham cohomology to obtain a contradiction.
Exercise Show that H
q
(S
2
S
4
) H
q
(CP
3
) for every q, but that the cohomology algebras
are not isomorphic.
Exercise (i) Show that any element T Gl(n + 1, C) passes to the quotient to give a map

T : CP
n
CP
n
.
(ii) Which elements of Gl(n + 1, C) give rise in this way to isometries of CP
n
(with respect to
the Fubini-Study metric)?
(iii)

If T Gl(n + 1, C) and is the 2-form of 5.8, what is

T

()?
Exercise (i) Use the long exact sequence constructed in 5.2 to give another calculation of
H

(S
n
).
(ii) Ditto for the cohomology of the 2-torus.
64
Exercise Suppose that M
m
is a compact submanifold of S
n
, with 0 < m < n, and let
U = S
n
M. Construct isomorphisms
H
q
(U) H
nq1
(M)

for 1 q n 2. Show that H

n
(U) = 0 and nd short exact sequences
0 H
n1
(U) H
0
(M)

R 0
and
0 R H
0
(U) H
n1
(M)

0.
Exercise A symplectic manifold is a manifold M equipped with a closed 2-form such that
on each tangent space T
p
M, induces a non-degenerate pairing. Show that
1. A symplectic manifold must be even-dimensional (Hint: After choosing a basis for T
p
M,
the the pairing on T
p
M can be written in the form (v
1
, v
2
) = [v
1
]
t
A[v
2
], where A is
a skew-symmetric matrix and [v
1
], [v
2
] are the expressions of v
1
, v
2
in the chosen basis.
Show that det A ,= 0 to conclude that dimM must be even.)
2. Suppose M is a symplectic manifold of dimension 2n. Show that
n
must be a nowhere-
vanishing 2n-form. Deduce that if M is compact then [] ,= 0 in H
2
(M), and that
H

(M) contains a subalgebra isomorphic to H

(CP
n
).
Exercise On the manifold T

M there is a canonical 1-form , dened as follows: a point

x TM is a linear form on the vector space T
(x)
M, where : TM M is the bundle
projection. If v T
x
T

M, then d
x
(v) T
(x)
M, so we can evaluate x on it. This is how we
dene the 1-form :
(v) = x(d
x
(v)).
(i) Find coordinate expressions for and for d when M = R
n
.
(ii) Find an expression for using local coordinates on a manifold M.
(iii) Show that := d is a symplectic form on M.
(iv) Show that if f : M R is a smooth function and L = df(M) T

M then the restriction

of to L is zero.
6 Vector Bundles
There are two equivalent denitions:
(1) E

M is a smooth real vector bundle of rank k if each bre E
p
=
1
(p) is a real
vector space of dimension k, and if for all p M there is a neighbourhood U of p in M and a
dieomorphism
U
:
1
(U) U R
k
such that proj
U
= and such that the restriction
of to each bre is a linear isomorphism.
(2) E

M is a smooth real vector bundle of rank k if there is an open cover U

A
of
M and dieomorphisms

:
1
(U

) U

R
k
such that for all , A, the composite
(U

) R
k

1
a

1
(U

(U

) R
k
65
restricts to a linear isomorphism p R
k
R
k
for all p U
a
U

.
In the second denition, the vector space structure on the bre E
p
is the one pulled back from
p R
k
= R
k
by

. The second part of this denition is in order that the structure be

independent of the choice of local trivialisation

.
The denition for complex vector bundle is analogous.
Example 6.1 1. For any smooth n-manifold M, : TM M and its dual T

M are real
vector bundles of rank n.
2. If M
m
Y
n
and Y has a Riemannian metric then the normal bundle N(M, Y ) =
(x, v) M TY : v (T
x
M)

is a smooth vector bundle over M of rank n m,

with the obvious projection to X. It seems somewhat unsatisfactory to have to use a
Riemann metric on Y to dene N(M, Y ). Later we will see how to do it without.
3. There is a canonical complex line bundle (i.e. complex bundle of rank 1) over CP
n
,
denoted H
n
and dened as follows: its total space is
(, y) CP
n
C
n+1
: y
and the bundle projection simply maps (, x) to . That is, the bre over is the line .
Exercise Show that H
n
is a smooth complex line bundle.
A section of the bundle : E M is a map s : M E such that s = id
M
.
Example 6.2 1. If M R
N
and g is a smooth function vanishing on M and dened on
some neighbourhood V of M in R
N
, then restriction to M of the vector eld grad g
is a section of the normal bundle N(M, R
N
). If M R
N
is the preimage of a regular
value of a map g : V
k
(where V is some neighbourhood of M), then the k sections
grad g
i | M
of N(M, R
N
) are everywhere linearly independent, and can be used to dene
a global trivialisation N(M, R
N
) M R
k
of N(M, R
N
).
2. Suppose that G is a Lie group with group identity element e, and for each g G let

g
: G G denote left-multiplication by g,
g
(h) = gh. For any vector v T
e
G we use
the maps
g
to propagate v to give a vector eld
v
on G:

v
(g) = d
e

g
(v).
In this way we get n vector elds on G; together they dene a global trivialisation of
TG.
At this point these Lecture Notes end: the remainder of the course consists of
1. Morita, Geometry of Dierential Forms, Chapter 5.
2. Madsen and Tornehave, From Calculus to Cohomology, Chapters 15, 16, 17 and 18.
Read Morita rst: he gives a clearer and more motivated account, occasionally skipping tech-
nical details, which can be found in Madsen and Tornehave. If time runs short, make sure you
understand Morita, and worry less about Madsen and Tornehave. I strongly advise you to do
the relevant exercises in both books.
66