Differential Forms

DIFFERENTIAL FORMS
AND
THEIR INTEGRALS
(preliminary, incomplete version)
Sandro Buoncristiano, Francesco Mercuri and Alcibiades Rigas

Contents
Chapter 0. A review of basic vector calculus 5

1. Differentiable functions 5
2. Integration 13
3. Vector fields, distributions and the local Frobenius Theorem 17
4. Exercises 24
Chapter 1. The de Rham cohomology for open sets of Rn 29

1. Exterior forms 29
2. Differential forms and the de Rham cohomology 35
3. Algebraic aspects of cohomology 39
4. Basic properties of the de Rham cohomology 45
5. An application: the Jordan-Alexander duality Theorem 50
6. Appendix: partitions of unity and smooth approximations 52
7. Exercises 54
Chapter 2. Integration and the singular homology of open sets of Rn 59

1. Integration on singular chains and Stokes Theorem 59
2. Singular homology 62
3. The de Rham Theorem for open sets of Rn 66
4. Tensor product of vector spaces and the Künnet’s Theorem 68
5. Integration of 1-forms and some applications 70
6. Appendix: baricentric subdivision and the proof of Theorem 2.13 74
7. Exercises 75
Chapter 3. A short course on differentiable manifolds 81

1. Submanifolds of Euclidean spaces 81
2. Differentiable manifolds 88
3. Special atlas and partitions of unity 93
4. Immersions, embeddings and the Theorems of Whitney 95
5. Integration of vector fields an distributions 100
6. Lie groups 102
7. The adjoint representation 107
8. Covering groups 109
9. Group actions and homogeneous spaces 111
3
4
10. Exercises 113
Chapter 4. The theory of de Rham for differentiable manifolds 119

1. Cohomology and homology for differentiable manifolds 119
2. The Theorems of de Rham and Künnet for differentiable manifolds 123
3. The Jordan separation Theorem revisited 126
4. Orientable manifolds, integration of n-forms and the Theorem of Stokes 128
5. Invariant cohomology and the cohomology of compact Lie groups 132
6. Exercises 136
Chapter 5. Harmonic forms and the Theorem of Hodge 139

1. Some basic facts in Riemannian geometry 139
2. The Laplace-Beltrami Operator 150
3. The Theorem of Hodge 153
4. The Weitzenbök formula 155
5. Some applications 156
6. Exercises 158
Bibliography 161
CHAPTER 0
A review of basic vector calculus
In this chapter we will review some basic facts of vector calculus which will be used extensively along
these notes. We will assume the reader familiar with the differential and integral calculus for real valued
functions of one real variable, as well as with the basic topology of Euclidean spaces: open and closed
sets, continuity, compactness, Cauchy sequences etc. The material in sections 1 and 2 are quite standard in
Calculus courses, while the one in section 3 is probably less “popular.”
1. Differentiable functions
We will consider the Euclidean space Rn with its canonical inner product and associated norm.
For a point x ∈ Rn and r ∈ R, r > 0, we denote by B n (p, r) := {x ∈ Rn : kx − pk < r} the ball of radius
r centred at p. When p = 0 we simply write B n (r) for B n (0, r).
We will denote by L(Rn , Rm ) the space of linear maps of Rn into Rm . There is a natural identification
of L(Rn , Rm ) with Rnm , associating to a linear transformation L, the entries (in a fixed order) of the matrix
representing L in the canonical bases. This identification induces a scalar product in L(Rn , Rm )
hA, Bi = trace At B,
where At is the transpose of A. Often it is more convenient to consider the operator norm, defined by
kLk = sup{kLxk : x ∈ Rn , kxk = 1}.
The two norms are equivalent (see Exercise 4.1), so for the basic topological concepts like convergence,
continuity etc., it does not matter which one we use. In what follows we will consider the operator norm,
unless otherwise stated. Observe that, for the operator norm, we have the inequality kL ◦ T k ≤ kLkkT k.
Let U ⊆ Rn be an open set and f : U −→ Rm a function.
1.1. Definition. f is differentiable at x ∈ U if there exist a linear map df (x) : Rn −→ Rm such that1
kf (x + h) − f (x) − df (x)(h)k
lim = 0.
khk→0 khk
The map df (x) is called the differential of f at x.
f is differentiable in U if it is differentiable at every point of U .
1.2. Remark. For a function f : U ⊆ R −→ R the derivative of f at x ∈ U, f 0 (x), is defined as

df
f 0 (x) := |t=x := lim [f (x + t) − f (x)]t−1 ,
dt t→0
1Observe that, if khk is sufficiently small, x + h ∈ U .
5
1. DIFFERENTIABLE FUNCTIONS 6
if the limit exists. The differential of f at x is the linear map
df (x) : R −→ R, df (x)h = f 0 (x)h.
The following facts are easy to prove and and we leave the proofs to the reader (Exercise 4.2).
1.3. Proposition.
• If f, g are differentiable at x and a ∈ R, then f + g and af are differentiable and d(f + g)(x) =
df (x) + dg(x), d[af ](x) = a[df (x)].
• If f is differentiable at x, f is continuous at x.
• If f is differentiable at x, the differential is unique.
• (The chain rule) If f : U ⊆ Rn −→ Rm , g : Σ ⊆ Rm −→ Rp are differentiable at x and f (x)
respectively, then g ◦ f is differentiable at x and d[g ◦ f ](x) = dg(f (x)) ◦ df (x).
1.4. Example. Let L : Rn −→ Rm be a linear map. Then L is differentiable and dL(x) = L, ∀ x ∈ Rn ,

as follows directly from the definition.
1.5. Example. If B : Rn × Rm −→ Rp is a bi-linear map, B is differentiable and dB(x, y)(z, w) =

B(x, w) + B(z, y). In particular, if f, g : U −→ R are differentiable functions, the map F : U −→ R2 , F (x) =
(f (x), g(x)) is differentiable with dF (x) = (df (x), dg(x)). Since the product R × R −→ R is bilinear, by the
chain rule the function f g is differentiable and the product rule d(f g)(x) = f (x)dg(x) + g(x)df (x) holds.
In the theory of real valued functions of one real variable, an elementary but useful result is the Mean
Value Theorem.
1.6. Theorem. [Mean Value Theorem] If f : [a, a + h] ⊆ R −→ R is a differentiable function, then there
exists t0 ∈ [0, 1] such that
f (a + h) − f (a) = f 0 (a + t0 h)h.
The Theorem extends, with essentially the same proof, to the case of a differentiable function f : U ⊆
Rn −→ R. For functions with values in Rm , m > 1, such a Theorem does not hold any longer (see Exercise
4.4) but, at least, we have an inequality. The result will still be called the Mean Value Theorem.
1.7. Theorem. [Mean Value Theorem] Let f : U ⊆ Rn −→ Rm be a continuous function. Suppose that
the segment with endpoints a, a + h is contained in U and f is differentiable at the points of the segment.
Then
kf (a + h) − f (a)k ≤ khk sup{kdf (a + th)k : t ∈ [0, 1]}.
Proof. Consider the function φ : [0, 1] −→ Rm , φ(t) = f (a + th). φ is differentiable, by the chain rule,
φ(0) = f (a), φ(1) = f (a + h) and dφ(t)(1) = df (a + th)(h). Let M = sup{kdφ(t)k : t ∈ [0, 1]}. It is then
sufficient to prove that kφ(1) − φ(0)k ≤ M . For this purpose we will show that, given ² > 0, kφ(1) − φ(0)k ≤
M + ². Consider the set
A = {t ∈ [0, 1] : kφ(s) − φ(0)k ≤ (M + ²)s, ∀ s ∈ [0, t]}.

CHAPTER 0. A REVIEW OF BASIC VECTOR CALCULUS 7
It is easy to see that A = [0, a] for some a ∈ (0, 1]. We wish to prove that a = 1. Suppose a < 1. Then there
exists a positive δ such that a + δ < 1 and for k ∈ [0, δ)
φ(a + k) − φ(a) = dφ(a)k + r(k) with kr(k)k ≤ ²k
(by the definition of differentiability at a). Then kφ(a + k) − φ(a)k ≤ (M + ²)k. But a ∈ A, hence
kφ(a) − φ(0)k ≤ (M + ²)a. Therefore kφ(a + k) − φ(0)k ≤ (M + ²)(a + k). In particular a + k ∈ A, a
contradiction. ¤
1.8. Definition. Let f : U −→ Rm be differentiable at x ∈ U and X ∈ Rn . The directional derivative

of f at x in the X direction is defined as
∂f
(x) := df (x)(X).
∂X
1.9. Remark. For reasons that will be clear later on will use often the notation Xx (f ) for df (x)(X).
∂f
If {e1 , . . . , en } is the canonical basis of Rn ,
(x) is the ith partial derivative at x and will be denoted,
∂ei
∂f Pn
as usual, by (x). If f is differentiable at x and h = i=1 αi ei , then
∂xi
n
X n
X ∂f
df (x)h = αi df (x)(ei ) = αi (x).
i=1 i=1
∂xi
In particular, if f (x) = (f1 (x), . . . , fm (x)), where fi : U −→ R are the coordinate functions of f ,
∂fj
then the Jacobian matrix [ (x)] is the matrix that represents df (x) in the canonical bases. This is the
∂xi
multidimensional analogue of Remark 1.2.
Let γ : (a, b) ⊆ R −→ Rn be a differentiable map. We will also say that γ is a differentiable curve. For
such a function, the tangent vector at t ∈ (a, b) (or, sometimes, at γ(t)) is the vector
d
γ̇(t) := dγ(t)(1) = |s=t γ(s).
ds
It is easy to see that if γ : (−², ²) −→ Rn is a differentiable curve with γ̇(0) = X,
∂f d
(x) = d(f ◦ γ)(0)(1) := |t=0 f (γ(t)).
∂X dt
1.10. Remark. The latter fact gives a geometric interpretation of the differential of f : the image, via
df , of the vector tangent to a given curve γ is the vector tangent to the image curve, f ◦ γ.
In particular the right hand side of the formula above does not depend on γ as long the curve passes
trough x and its tangent vector, at x, is X. This observation allow us to define the directional derivative,
hence partial derivatives, even for a class of not necessarily differentiable functions (in the sense of Definition
1.1). If f : U −→ Rm is a function and X ∈ Rn , we define the directional derivative of f at x ∈ U , in the
direction of X as
∂f d
(x) := |t=0 f (x + tX),
∂X dt
if it exists. The partial derivatives may exist even if the function is not differentiable (see Exercise 4.7).
However we have
1.11. Proposition. Let f : U ⊆ Rn −→ Rm be a function. If the partial derivatives of f exist and are
continuous2, f is differentiable.
Proof. We will prove the Proposition for n = 2 to avoid notational complications. We want to show
∂f ∂f
that the linear map L(x, y)(h, k) = h+ k is the differential of f at (x, y) ∈ R2 . Hence we have to show
∂x ∂y
that, given ² > 0,
∂f ∂f
kf (x + h, y + k) − f (x, y) − h− kk ≤ ²k(h, k)k,
∂x ∂y
if k(h, k)k is sufficiently small. Adding and subtracting f (x, y + k) and using Exercise 4.5, we have that the
quantity on the left of the inequality sign is less or equal to
∂f ∂f ∂f ∂f
khk sup{k (x + th, y + k) − (x, y)k : t ∈ [0, 1]} + kkk sup{k (x, y + tk) − (x, y)k : t ∈ [0, 1]}.
∂x ∂x ∂y ∂x
The conclusion follows from the continuity of the partial derivatives. ¤
1.12. Remark. Let f : U ⊆ Rn −→ Rm be a function. Let πi : Rm −→ R, πi (x1 , . . . , xm ) = xi be the

th
i projection. Then fi (x) = πi ◦ f (x) are the coordinates of f (x). It is easy to see that f is differentiable
at x ∈ U if and only if the coordinate functions are differentiable at x and, in this case,
X
df (x)(X) = (df1 (x)(X), . . . , dfm (x)(X)) = [dfi (x)X]ei .
Partial derivatives take care of the “opposite” situation. Given a splitting of Rn = E1 ⊕ E2 as a direct sum of
complementary subspaces and a point (x0 , y0 ) ∈ E1 ⊕E2 , we can consider the inclusion ij : Ej −→ Rn , i1 (x) =
(x, y0 ), i2 (y) = (x0 , y) and the functions f (i) = f ◦ ij : Ej ∩ U −→ Rm . If f is differentiable at (x0 , y0 ), f (1)
(resp. f (2) ) is differentiable at x0 (resp. y0 ) and df (x0 , y0 )(X, Y ) = df (1) (x0 )(X) + df (2) (y0 )(Y ). So we
can define the partial differentials relative to the given splitting, dj f = df ◦ ij . The existence of the partial
differentials does not implies the existence of the differential of f . However, as in Proposition 1.11, if the
partial differentials exist and are continuous, then f is differentiable. Obviously the same arguments work for
a decomposition of Rn into the direct sum of k complementary subspaces. Partial derivatives are, essentially,
partial differentials relative to the canonical splitting of Rn as the direct sum of the coordinate lines.
If f : U −→ Rm is a differentiable function, the differential can be seen as a map df : U −→

L(Rn , Rm ), x Ã df (x).
1.13. Definition. We will say that f is twice differentiable at x ∈ U , if the function df : U −→

L(R , Rm ) is differentiable at x. In this case, the differential of df at x will be called the second differential
n
of f at x and will be denoted by d2 f (x).

Inductively, we define, if it exists, the k th differential of f at x, dk f (x), as the differential, at x, of
dk−1 f U −→ L(Rn , L(Rn , . . .)
We will say that f is of class C k in U if dk (x) exists, ∀ x ∈ U , and is continuous, as a function of x.
We will say that f is of class C ∞ , if it is of class C k , ∀ k. If f is C ∞ we will also say that f is smooth.
1.14. Remark. It is easy to produce examples of C k functions that are not C k+1 (see Exercise 4.6).
One of the important features of the class of smooth functions is that it is closed under differentiation, i.e.
f is smooth if and only if df is smooth.
2As maps ∂f : U −→ Rm .
∂xi
If f is twice differentiable at x, then d2 f (x) ∈ L(Rn , L(Rn , Rm )), and so it can be seen as the bilinear
map d2 f (x)(X, Y ) = d(df )(x)(X)(Y ). In a similar way, dk f (x) can be viewed as a k-multilinear map.
1.15. Theorem. [Schwarz’s Theorem] If d2 f (x) exists, it is a symmetric bilinear form.
We will sketch a proof in the case that f is C 2 in Exercise 4.22
1.16. Remark. If f is k times differentiable at x, dk f (x) is a k-multilinear symmetric map. Moreover

we can define higher order partial derivatives. Schwarz Theorem 1.15 and a simple induction imply that if
f is of class C k , the result of successive partial derivatives, up to order k, does not depend on the order of
derivations.
1.17. Example. If L : Rn −→ Rm is a linear map dL(x) = L, ∀ x ∈ Rn . In particular the differential

dL : Rn −→ L(Rn , Rm ) is the constant map. Hence dk L = 0, if k ≥ 2, and L is C ∞ . Similarly, if
B : Rn × Rm −→ Rp is a bilinear map, dB(x, y)(z, w) = B(x, w) + B(z, y). In particular dB is a linear map,
hence C ∞ , and so is B.
1.18. Example. Let M (n, R) be the space of n × n matrices with real coefficients. The product map
m : M (n, R) × M (n, R) −→ M (n, R), m(A, B) = AB,
is a bi-linear map, hence smooth. Also the map
g : M (n, R) × M (n, R) −→ L(M (n, R), M (n, R)), g(A, B)H = AHB
is bilinear, hence smooth.

Since the determinant, det : M (n, R) −→ R, is a continuous function, the set of invertible matrixes,
GL(n, R), is an open subset of M (n, R). Consider the inversion map
ι : GL(n, R) −→ M (n, R), ι(A) = A−1 .
Claim The map ι is smooth.
Proof. Differentiating the identity (A + tB)(A + tB)−1 = 11 we get

d
|t=0 (A + tB)−1 = −A−1 BA−1 .
dt
So, if ι is differentiable at A ∈ GL(n, R), dι(A)(B) = −A−1 BA−1 . It is easy to check that, in fact, the
formula defines a linear map which is the differential of ι at A. In particular ι is continuous. The differential
of ι is then given by the composition
∆ ι×ι −g
GL(n, R) −→ GL(n, R) × GL(n, R) −→ GL(n, R) × GL(n, R) −→ L(M (n, R), M (n, R)),
where ∆ : GL(n, R) −→ GL(n, R) × GL(n, R) is the diagonal map, ∆(A) = (A, A), and g is as above. Hence
dι is continuous and ι is of class C 1 . At this point a simple induction proves the Claim. ¤
As we have seen, the differential of a function f at a point x, provides the best linear approximation
of f − f (x) in a neighborhood of x. The Taylor formula provides the best polynomial approximation, for
functions with more differentiability.
1.19. Theorem. [Infinitesimal version of Taylor Theorem] Let f : U −→ Rm be a function s times

differentiable in an open neighborhood of a ∈ U and such that ds+1 (a) exists. Then
s+1
X 1 k
f (a + h) = f (a) + d f (a)(h, . . . , h) + r(h) with limh→0 r(h)khk−(s+1) = 0.
k!
k=1
Proof. The proof is a simple consequence of the lemma below, which is of interest on its own
1.20. Lemma. Let r : B = B n (R) −→ Rm be a function s times differentiable in B and s + 1 times

differentiable in 0. Assume dj r(0) = 0, 0 ≤ j ≤ s + 1. Then limx→0 r(x)kxk−(s+1) = 0.
Proof. We proceed by induction on s. If s = 0 the conclusion follow from the definition of differentia-
bility. Suppose the conclusion true for s. By the mean value Theorem we have
kr(x)k ≤ M kxk, M = sup{kdr(tx)k| : t ∈ [0, 1]}.
Applying the inductive hypothesis to dr, given ² > 0 there exist δ > 0 such that, if kyk < δ, kdr(y)k < ²kyks .
Hence if kxkδ, M ≤ ²kxks and kr(x)k ≤ ²kxks+1 . ¤
In the linear context, i.e. vector spaces and linear maps, we study properties that are invariant for linear
isomorphisms, i.e. changes of bases. The analogue in the differential context are properties that are invariant
for (local) diffeomorphisms , i.e. change of variables (or coordinates).
1.21. Definition. A map φ : U ⊆ Rn −→ V ⊆ Rm between open sets is a C k diffeomorphism if there

exists a C k map ψ : V −→ U , such that ψ ◦ φ = 11U , φ ◦ ψ = 11V . φ is a local diffeomorphism if ∀ x ∈ U , there
exists an open neighborhood Ũ ⊆ U of x, such that φ|Ũ is a diffeomorphism onto an open neighborhood Ṽ
of φ(x) in V . A local diffeomorphism will also be called a change of variables (or coordinates).
1.22. Remark. If φ is a diffeomorphism, then dφ(x) is an isomorphism, by the chain rule. Hence n = m.
The following fact will be useful.
1.23. Lemma. If φ : U ⊆ Rn −→ V ⊆ Rm is a C k map, k ≥ 1, between open sets, and φ admits a

differentiable inverse, then the inverse is of class C k .
Proof. From the chain rule d[φ−1 ](φ(x)) ◦ dφ(x) = 11. In particular d[φ−1 ] is given by the composition
φ−1 dφ ι
V −→ U −→ GL(n, R) −→ GL(n, R),
where ι is the matrix inversion map, that, by Example 1.18, is smooth. Hence d[φ−1 ] is continuous and φ−1
is of class C 1 . In general the argument gives that, if φ−1 is C s , s < k, d[φ−1 ] is also of class C s so φ−1 is of
class C s+1 and this concludes the proof.
¤
Let E, F be real, finite dimensional vector spaces and L : E −→ F be a linear map. Then, in suitable
bases, F has a very simple expression. In fact we can chose a basis {e1 , . . . , en } of E, such that {ek+1 , . . . , en }
is a basis of the kernel of F . Then {f1 = F (e1 ), . . . , fk = F (ek )} is a basis of the image of F that we can
complete with vectors {fk+1 , . . . , fm } to have a basis for F. Then, in terms of coordinates in these bases,
F (x1 , . . . , xn ) = (x1 , . . . , xk , 0, . . . , 0).
Since a differentiable function is (locally) approximated by a linear one, we can expect something similar
to hold, locally, for differentiable maps, up to change of coordinates. In fact this is the case.
1.24. Theorem. [Rank Theorem] Let f : U ⊆ Rn −→ Rm be a function of class C k . Let p ∈ U be such

that rank df (x) = k in an open neighborhood of p. Then there exist open neighborhoods U of p and V of
f (p), and diffeomorphisms φ : U −→ Ũ , ψ : V −→ Ṽ such that:
ψ ◦ f ◦ φ−1 (x1 , . . . , xn ) = (x1 , . . . xk , 0, . . . , 0), f or (x1 , . . . , xn ) ∈ Ũ .
This Theorem will follow from the next three.
1.25. Theorem. [Inverse function Theorem] Let f : U ⊆ Rn −→ Rn be a C k map, k ≥ 1, such that

df (p) : Rn −→ Rn is an isomorphism. Then the exist an open neighborhood U 0 of p such that f |U 0 is a C k
diffeomorphism onto an open neighborhood of f (p).
Proof. Without loss of generality we may assume p = 0 = f (p). Moreover, by composing f with
df (0)−1 , we may assume df (p) = 11. Consider the function g(x) = f (x) − x. Then g(0) = 0, dg(0) = 0. Let
r be a positive real number such that if x ∈ B n (r), df (x) is invertible and kdg(x)k < 21 .
Claim 1. If y ∈ B n ( 2r ) then there exists a unique x ∈ B n (r) such that f (x) = y.
Proof. Define x0 = 0, x1 = y and, inductively, xn+1 = y + g(xn ). By the mean value Theorem we
have:
1
kxn+1 − xn k = kg(xn ) − g(xn−1 )k ≤
kxn − xn−1 k
2
1 1 r
kxn+1 k = kg(xn ) + yk ≤ kg(xn )k + kyk < kxn k + kyk ≤ kxn k + − ²,
2 2 2
for some ² > 0, independent of n. Hence:
(1) kxn+1 − xn k ≤ 2−n kyk (from the first equation),
(2) kxn k < r − ² (from the second equation and induction).
By condition (1), {xn } is a Cauchy sequence, hence it converges to a point x, and by condition (2), x ∈ B n (r).
Let us show that x is unique. Suppose f (z) = f (x) = y, z ∈ B n (r). Then kx − zk = kg(z) − g(x)k ≤
1
2 kz − xk which implies z = x. ¤
So we have a well defined surjective map f −1 : B n ( 2r ) −→ U 0 = B n (r) ∩ f −1 (B n ( 2r )). Observe that U 0

is the intersection of two open sets, hence it is open.
Claim 2. f −1 : B n ( 2r ) −→ U 0 is C k .
Proof. We start by observing that kf (x1 ) − f (x2 )k ≥ kx1 − x2 k − kg(x1 ) − g(x2 )k ≥ 21 kx1 − x2 k, hence
f −1 is continuous. In order to show that f −1 is differentiable we observe that, since f is differentiable,
h(x, x1 )
f (x) − f (x1 ) = df (x1 )(x − x1 ) + h(x, x1 ) with lim = 0.
x→x1 kx − x1 k
Applying A := df (x1 )−1 to the equality above we have
A(y − y1 ) + Ah1 (y, y1 ) = f −1 (y) − f −1 (y1 ),
where y = f (x), y1 = f (x1 ), h1 (y, y1 ) = −h(f −1 (y), f −1 (y1 )). Then

h1 (y, y1 ) h(x, x1 ) kx − x1 k
=− .
ky − y1 k kx − x1 k ky − y1 k
kx − x1 k h1 (y, y1 )
Since ≤ 2, lim = 0 and d[f −1 ](y1 ) = [df (x1 )]−1 . Hence f −1 is differentiable , the
ky − y1 k y→y1 ky − y1 k
Claim follows from Lemma 1.23. ¤
1.26. Theorem. [Local form of immersions] Let f : U ⊆ Rn −→ Rn+p be a C k map, k ≥ 1, such that
0 ∈ U, f (0) = 0. If df (0) : Rn −→ Rn+p = Rn × Rp is injective, there exists an open neighborhood U 0 of 0
and a C k diffeomorphism φ between neighborhoods of 0 ∈ Rn+p such that if x ∈ U 0 ,
φ ◦ f (x) = (x, 0).
Proof. Up to an isomorphism of Rn+p which sends df (0)ei to ei , we can assume that df (0)v = (v, 0).
Consider the function
F : U × Rp −→ Rn+p , F (x, y) = f (x) + y.
Observe that F (x, 0) = f (x) and dF (0) = 11. By the Inverse Function Theorem there is a diffeomorphism φ
between neighborhoods of 0 ∈ Rn+p such that φ ◦ F = 11. Then
φ ◦ f (x) = φ ◦ F (x, 0) = (x, 0).
1.27. Theorem. [Local form of subimmersions] Let f : U ⊆ Rn+p −→ Rn be a C k map, k ≥ 1, such that
0 ∈ U, f (0) = 0. If df (0) : Rn+p −→ Rn is surjective, then there exists a C k diffeomorphism ψ : U 0 −→ V ,
between open neighborhoods of 0 ∈ Rn+p , such that if (x, y) ∈ U ,
f ◦ ψ(x, y) = x.
Proof. Up to an isomorphism of Rn+p = Rn × Rp , we can assume that ker df (0) = {0} × Rp and
df (0)(v, 0) = v. Consider the function
F : U −→ Rn × Rp , F (x, y) = (f (x, y), y).
Then dF (0) = 11 and f = π ◦ F , where π : Rn+p −→ Rn is the canonical projection. By the Inverse Function
Theorem there exists a local inverse ψ of F . Then
f ◦ ψ(x, y) = π ◦ F ◦ ψ(x, y) = π(x, y) = x.
At this point we leave to the reader the task of proving Theorem 1.24.
2. Integration
We will recall now the basic fact of Riemann integration theory. We will take a limited approach which
is enough for our purposes.
We will start with the case of functions of one real variable. Let [a, b] ⊆ R be a closed interval. A partition
of the interval is a set P = {t0 , . . . , tk } ⊆ R such that a = t0 < · · · < tk = b. We will set |P | = sup{ti − ti−1 }.
Given a function f : [a, b] −→ Rm and a partition P of [a, b], we define
k−1
X
Σ(f, P ) = (ti+1 − ti )f (ti ).
i=0
2.1. Definition. A vector X ∈ Rm is said to be an integral of f on [a, b] if, given ² > 0, there exists
δ > 0 such that
kΣ(f, P ) − Xk < ² for all partitions P with |P | < δ.
If an integral exists we will say that f is integrable on [a, b] and we use the notation
Z b Z b
X= f (t)dt or, when clear from the contex, simply X = f.
a a
It is easy to see that if a function is integrable, the integral is unique.
2.2. Remark. Let f : [a, b] −→ Rm be an integrable function. Then we can compute the integral as
limit of the sequence Σ(f, Pn ) where Pn is a sequence of partitions such that limn→∞ |Pn | = 0.
The proof of the following Proposition is simple and left to the reader (see Exercise 4.16).
2.3. Proposition. Let f, g : [a, b] −→ Rm be integrable functions, k ∈ R and T : Rm −→ Rp be a linear

map. Then f + g, kf and T ◦ f are integrable and
Rb Rb Rb Rb Rb
(1) a f + g = a f + a g, a
kf = k a f,
Rb Rb
(2) a T ◦ f = T ( a f )
Rb
(3) k a f k ≤ (b − a)kf k0 .
Z x
m
(4) The function F : [a, b] −→ R , defined by F (x) = f (t)dt is well defined and continuous.
a
We will denote by B := B([a, b], Rm ) the set of bounded functions of [a, b] into Rm . B([a, b], Rm ) is a real
vector space, with the obvious operations, and
kf k0 = sup{f (t) : t ∈ [a, b]}
is a norm in B, called the sup norm or the norm of uniform convergence.
2.4. Definition. A sequence of functions fn : [a, b] −→ Rm in B is uniformly convergent to f ∈ B if

limn−→∞ kfn − f k0 = 0.
2.5. Remark. Proposition 2.3 tell us that the set of bounded integrable functions I is a linear subspace
of B, the integral maps I into Rm linearly (items (1)) and continuously (item (3)). The next Proposition
tell us that I is closed in B.
2. INTEGRATION 14
2.6. Proposition. Let fn : [a, b] −→ Rm be a sequence of bounded integrable functions. If the sequence
converges uniformly to a function f , the f is integrable and
Z b Z b
f = limn→∞ fn .
a a
Z b
Proof. Set In = fn . By Proposition 2.3 (3), kIm − Ik k ≤ (b − a)kfm − fk k0 . Hence {In } is a Cauchy
a
m
sequence in R and therefore converges to a vector I ∈ Rm . We claim that I is the integral of f . Fix ² > 0.
Then there exist n such that kfm − f k0 < ²/3(b − a), kI − Im k < ²/3 if m > n. Also there exist δ > 0
such that kΣ(fm , P ) − Im k < ²/3 if |P | < δ. Observe that kΣ(f, P ) − Σ(fm , P )k ≤ (b − a)kf − fm k0 . So,
if |P | < δ, kI − Σ(f, P )k ≤ kI − Im k + kIm − Σ(Fm , P )k + kΣ(fm , P ) − Σ(f, P )k < ² and this prove the
claim. ¤
We will describe classes of integrable functions.
2.7. Definition. A function f : [a, b] −→ Rm is a step function if there exists a partition P of [a, b], P =
{t0 , . . . , tk } and vectors {X0 , . . . , Xk−1 } such that f (t) = Xi , t ∈ (ti , ti+1 ).
2.8. Lemma. Let f : [a, b] −→ Rm be a step function relative to a partition P . Then f is integrable and
Z b X
f= (ti+1 − ti )Xi .
a
Proof. We can suppose f (ti ) = Xi (see Exercise 4.19). Observe that Σ(f, P 0 ) = Σ(f, P ) if P 0 is
obtained from P adding new points. Therefore Σ(f, P ) = Σ(f, P ∪ Q) for all partitions Q and the conclusion
follows. ¤
Since every continuous function is uniform limit of step functions (see Exercise 4.20), combining the last
two Proposition we have
2.9. Proposition. If f : [a, b] −→ Rm is continuous, then it is integrable.
We will recall now the basic relation between differentiation and integration.
2.10. Lemma. Let f : [a, b] −→ Rm be a continuous function and x ∈ [a, b]. Then
Z b Z x Z b
f= f+ f.
a a x
Proof. Let P be a partition such that x ∈ P . then P = P 0 ∪ P 00 where P 0 is a partition of [a, x] and
P 00 a partition of [x, b]. Since all three integrals exist, by Lemma 4.20, we can compute the integrals as limit
of Σ(f, Pn ) where Pn is a partition as above and limn→∞ |Pn | = 0 (see Remark 2.2). Then the conclusion
follows easily. ¤
2.11. Remark. The Lemma still holds for functions that are just integrable. We just have to prove that
an integrable function is integrable on any subinterval (see Exercise 4.18).
2.12. Theorem. [Fundamental Theorem of Calculus] Let f : [a, b] −→ Rm be a continuous function.

Define Z x
F : (a, b) −→ Rm , F (x) = f.
a
Then F is differentiable and F 0 (x) = f (x).
Proof. F is continuous by Proposition 2.3 (item (4)). Fix x ∈ (a, b) and let h > 0 be such that
x + h ∈ (a, b). Then
Z x+h
F (x + h) − F (x) 1
| − f (x)| = | f − f (x)| ≤ sup{|f (t) − f (x)| : t ∈ [x, x + h)}.
h h x
Since f is continuous the expression on the right hand side goes to 0, when h goes to 0. The same argument
works for h < 0 and the Claim follows. ¤
We will define now the integral of functions of several real variables. We will consider the case of two
variables and the reader should not have any difficulty to extend these considerations for n variables.
Let f : [a, b] × [c, b] −→ Rm be a function and let t, s be the first and second coordinate respectively. For
t ∈ [a, b] fixed, we set ft (s) = f (t, s). Suppose the function ft integrable, ∀ t ∈ [a, b]. Then we define the
iterated integral (if it exists), as
Z bZ d Z b Z d Z b Z d
f (t, s) ds dt := [ ft (s) ds] dt := dt f (t, s)ds.
a c a c a c
The elementary properties of the iterated integrals follows from the corresponding ones for the integrals
of functions of one real variable. For example
Z bZ d
| f (t, s)dsdt| ≤ (b − a)(d − c)kf k0 , where kf k0 = sup{kf (t, s)k : (t, s) ∈ [a.b] × [c, d]}.
a c
2.13. Example. Let P = {t0 , . . . , tk } be a partition of [a, b], Q = {s0 , . . . , sh } a partition of [c, d] and let
Xij ∈ Rn be fixed vectors. Let g : [a, b] × [c, d] −→ Rn be a function such that g(t, s) = Xij , t ∈ (ti , ti+1 ), s ∈
(sj , sj+1 ). For s ∈ [c, d] the function gs (t) = g(t, s) is a step function and
Z b k−1
X
gs dt = (tj+1 − tj )Xij if s ∈ (si , si+1 ).
a 0
Z b
Therefore h(s) = gs dt is also a step function, therefore integrable and
a
Z d Z b X
ds g(t, s)dt = (si+1 − si )(tj+1 − tj )Xij .
c a ij
Observe, in particular, that the iterated integral does not depend on the order of integration.
2.14. Proposition. If f is continuous, the iterated integrals exist and

Z b Z d Z d Z b
f (t, s) ds dt = f (t, s) dt ds.
a c c a
Proof. We will start with a general fact

Claim Let U ⊆ Rn and let f : U × [a, b] −→ Rm be a continuous function. Then the function
Z b
m
F : U −→ R , F (x) = f (x, t)
a
is a continuous function.
2. INTEGRATION 16
Proof. Fix x0 ∈ U and ² > 0. The set V = {(x, t) ∈ X × [a, b] : |f (x, t) − f (x0 , t)| < ²(b − a)−1 is
an open neighborhood of x0 × [a, b]. Since [a, b] is compact, there exists a neighborhood W of x0 such that
W × [a, b] ⊆ V . In particular, for all x ∈ W, |f (x, t) − f (x0 , t)| < ²(b − a)−1 , ∀t ∈ [a, b]. Hence, if x ∈ W
Z b
|F (x) − F (x0 )| ≤ |f (x, t) − f (x0 , t)| ≤ (b − a) sup{|f (x, t) − f (x0 , t)|} < ².
a
¤
The Claim implies, in particular, that a continuous function admits iterated integrals. We will prove
now the commutativity relation. More precisely, given ² > 0, we will show that, if P, Q are partitions as in
Example 2.13, there exists δ > 0 such that, if |P |, |Q| < δ,
Z b Z d X
| dt f (t, s)ds − (si+1 − si )(tj+1 − tj )f (tj , si )| < ².
a c
The conclusion will follows, since the other integral is, by symmetry, approximated by a sum of the same type.
By uniform continuity of f , it follows that there exists δ > 0 such that |f (t, s) − f (t0 , s0 )| < ²/(b − a)(d − c)
if |s − s0 |, |t − t0 | are smaller than δ. Consider the function g as in Example 2.13, with g(t, s) = f (tj , si ), t ∈
[tj , tj+1 ), s ∈ [si , si+1 ). Then kf − gk < ²/(b − a)(d − c). Therefore
Z b Z d X Z b Z d Z b Z d
| dt f (t, s)ds − (si+1 − si )(tj+1 − tj )f (tj , si )| = | dt f (t, s)ds − dt g(t, s)ds| =
a c a c a c
Z b Z d
=| dt [f (t, s) − g(t, s)]ds ≤ (b − a)(d − c)kf − gk < ².
a c
¤
We will define now the integral of a function f : C = [a, b] × [c, d] −→ Rn . Let P, Q be partitions of the
two intervals and set, in analogy with the 1-dimensional case,
X
Σ(f, P, Q) = (ti+1 − ti )(sj+1 − sj )f (ti , sj ).
2.15. Definition. We will say that X = lim|P |,|Q|→0 Σ(f, P, Q) if, given ² > 0 there exist δ > 0 such
that kX − Σ(f, P, Q)k < ² if |P |, |Q| < δ. If such a limit exists we will say that f is integrable and define the
duple integral of f on C as Z
f (t, s)dtds = X.
C
Rp
2.16. Lemma. If f is integrable over C and one of the simple integral, let say a
f (t, s)dt, exists, ∀s ∈
[c, d], then the other simple integral exists and the iterated integrals are equal to the duple integral.
Proof. The claim follows from the general relation between duple limits and iterated limits:
if lim Σ(f, P, Q) exists and lim Σ(f, P, Q) exists ∀Q, lim [ lim Σ(f, P, Q)] = lim Σ(f, P, Q).
|P |,|Q|→0 |P |→0 |Q|→0 |P |→0 |P |,|Q|→0
2.17. Theorem. [Baby Fubini] If f : C −→ Rn is continuous, then the duple integral exists and is equal
to the iterated integrals.
Proof. This is a corollary of the proof of Theorem 2.14. ¤

In particular we can define the integral of a continuous function with compact support.
2.18. Definition. Let f : R2 −→ Rn be a continuous function with compact support and C a rectangle
containing the support of F . We define
Z Z
f (t, s)dtds = f (t, s)dtds.
R2 C
2.19. Remark. It is easy to see that the definition does not depends on the choice of the rectangle C.
Beside Fubini’s Theorem, that allows us to reduce the calculus of multiple integrals to the case of simple
integrals, the other basic fact on integration that we will need is the formula of change of variables.
2.20. Theorem. Let U, V ⊆ R2 be open sets and let F : U −→ V be a diffeomorphism. If f : V −→ Rn

is an integrable function with compact support,
Z Z
f= |J[F ]|f ◦ F,
V U
where J[F ] := det[dF ] is the Jacobian determinant.
Proof. ¤
We invite the reader to extend the concepts and results above for the case of integration of function of
several variables.
3. Vector fields, distributions and the local Frobenius Theorem
3.1. Definition. Let U be an open set of Rn . A (tangent) vector field on U is a smooth map X : U −→
Rn . We will denote by H(U ) the space of vector fields on U .
3.2. Remark. Let X be a vector field. We want to think of X(x) as a vector based at x. This is the
reason why we use different names for the same thing3. We can make this point more precise as follows.
• The tangent space of U at x ∈ U is the vector space
Tx U = {(x, v) : v ∈ Rn }
with the obvious operations on the second component.

• The tangent bundle of U is
T U = ∪x∈U Tx U = U × Rn .
A vector field on U should be defined as a smooth map X̃ : U −→ T U of the form X̃(x) = (x, X(x)), X :
U −→ Rn . Of course, in our context, we are just complicating notations, but this point of view, that seems
silly now, will prove to be useful when these concepts are extended to the case of differentiable manifolds.
We will review now some facts about solutions of differential equations.
3B. Russel used to say that “Mathematics is the art of calling different things with the same name and the same thing
with different names”.

3. VECTOR FIELDS, DISTRIBUTIONS AND THE LOCAL FROBENIUS THEOREM 18
3.3. Definition. Let X ∈ H(U ), x ∈ U . An integral curve of X with initial condition x is a smooth
map γx : (a, b) ⊆ R −→ U such that:
dγx (t)(1) := γ̇x (t) = X(γx (t)), 0 ∈ (a, b) and γx (0) = x.
When it is clear from the context, or irrelevant, we will ignore the subscript relative to the initial
condition.
The basic result about integral curves is the following
3.4. Theorem. If X ∈ H(U ), x ∈ U , there exists an integral curve with initial condition x ∈ U . This
curve is unique in the sense that two such curves, with the same initial condition, coincide in the intersection
of the domains. In particular there is a maximal interval of definition, (α(x), β(x)) ⊆ R. Moreover the curve
is smooth and depends smoothly on the initial condition.
3.5. Remark. Smooth dependence on the initial condition means that, for fixed x, the exists a neigh-
borhood U of x and ² > 0 such that the map
Γ : U × (−², ²) −→ U, Γ(y, t) = γy (t),
is a smooth map.
3.6. Remark. Integral curves exist even if the vector field X is merely continuous. For unicity we need
the field to be locally lipschitzian. If X is of class C k , the curves are of class C k+1 .
3.7. Definition. The vector field is complete if its integral curves are defined on all of R.
3.8. Proposition. If X is complete, the map
γt : U −→ U, γt (x) = γx (t),
is well defined and smooth. Moreover

(1) γ0 = 11,
(2) γt+s = γt ◦ γs .
Proof. The first property is obvious, by definition. As regards the second, we observe that, for fixed
s, the curves γγx (s) (t) and γx (t + s) are integral curves of X with the same initial condition. The conclusion
follows from the unicity of integral curves. ¤
In particular γt is a diffeomorphism of U with inverse γ−t , and the map t ∈ R Ã γt is a homomorphism

of the additive group R into the group of diffeomorphisms of U .
3.9. Remark. If X is not complete, the considerations above hold locally. We leave to the reader the
task of making this claim precise.
3.10. Definition. A point x ∈ U is a singularity (or a singular point) of X ∈ H(U ), if X(x) = 0.
The behavior of X near a singularity can be quite complicated. On the contrary, the behavior near a
non singular point is quite simple.
3.11. Theorem. Let X ∈ H(U ), x ∈ U and X(x) 6= 0. Then there is a diffeomorphism φ : U ⊆ U −→ V

∂
of a neighborhood U of 0 ∈ Rn onto a neighborhood V of x, such that dφ(y)( ) = X(φ(y)).
∂x1
∂
Proof. We can assume x = 0, X(0) = (0). For p = (0, x2 , . . . , xn ) ∈ U , consider the integral curve
∂x1
of X with initial condition p, γp (t). Then the map φ(p, t) = γp (t) is well defined and smooth if |t| < ², with ²
sufficiently small and p is in a sufficiently small neighborhood U 0 of 0 ∈ Rn−1 = {(x1 , . . . xn )} ∈ Rn : x1 = 0}.
∂
It is clear that dφ(p, t)( ∂x 1
) = X(φ(p, t)) (see Remark 1.10). Also dφ(0, 0) = 11, hence, by Theorem 1.25, φ
is a diffeomorphism of a (possible smaller) neighborhood U ⊆ U 0 × (−², ²) of 0, onto its image. ¤
We can ask for a natural generalization of Theorem 3.11: given linearly independent vector fields
∂
X1 , . . . , Xk ∈ H(U ), do there exist local coordinates (x1 , . . . , xn ) in Rn such that Xi = ? In order
∂xi
to answer this question we will take a slight different approach to vector fields. First a few definitions.
3.12. Definition. An algebra over the reals is a real vector space E together with a bilinear map, the
product, b : E ⊕ E −→ E. The algebra is said to be associative if b(x, b(y, z)) = b(b(x, y), z) and commutative
if b(x, y) = b(y, x) ∀ x, y, z ∈ E.
When clear from the context we will write xy for b(x, y).
Examples of such a structure are
• The real or complex numbers with the usual multiplication. They are associative and commutative
algebras.
• The spaces M (n, K) of n × n matrices with entries in K = R or C, with the usual product of matrices.
They are associative but non commutative algebras (if n > 1!).
• The space F(U ) of smooth real valued functions defined in U ⊆ Rn .
3.13. Definition. An algebra homomorphism h : E −→ E0 between the algebras E and E0 is a linear

map such that the image of the product of two elements in E is the product of the images (in E0 ).
3.14. Definition. An ideal I of an algebra E is a vector subspace of E such that if x ∈ I, y ∈ E, then

b(x, y), b(y, x) ∈ I
It is not difficult to see that if I is an ideal of E, the quotient vector space E/I has a natural product
(and hence a structure of algebra) such that the quotient map is an algebra homomorphism. Moreover, given
an algebra homomorphism h : E −→ E0 , the kernel of h, ker h, is an ideal and, in fact, every ideal I is the
kernel of an algebra homomorphism, the projection π : E −→ E/I.
Let F(U ) be the algebra of smooth real valued functions defined in U .
3.15. Definition. A derivation of F(U ) (resp. a derivation at x ∈ U ) is a R-linear map Y : F(U ) −→

F(U ) (resp. Y (x) : F(U ) −→ R), such that:
Y (f g) = Y (f )g + f Y (g) (resp. Y (x)(f g) = Y (x)(f )g(p) + f (p)Y (x)(g)) ∀ f, g ∈ F(U ).
Both the set of derivations and the set of derivations at x have a natural structure of real vector space.
We will denote by Der(U ) and Derx (U ) these spaces. Observe that Der(U ) is infinite dimensional (if n > 0!)
while, as we will see soon, Derx (U ) is n-dimensional.
3.16. Example. Let X ∈ Rn , x ∈ U . Then the directional derivative of f ∈ F (U ) at x ∈ U , in the

direction X is a derivation at x. As we shall soon see, all derivations at x are directional derivatives.
3.17. Example. If X ∈ H(U ), we define a derivation Der(U ), still denoted by X, X(f )(x) := X(x)(f ),
where X(x)(f ) is the directional derivative at x as in the example above. It is easily seen that X(f )(x) ∈ F(U )
so X is, in fact, a derivation in Der(U ).
Some simple but basic facts are the following:
3.18. Lemma. Let f ∈ F(U ) and Xx ∈ Derx (U ).

• If f vanishes on an open neighborhood V ⊆ U , then Xx (f ) = 0. In particular, if two functions
f, g ∈ F(U ) coincide in a neighborhood of x, then Xx f = Xx g.
• If f is constant in a neighborhood of x, then Xx f = 0.
• If f is (locally) a product of functions vanishing at x, then Xx f = 0.
Proof. Let φ ∈ F (U ) be a function which vanishes in a neighborhood V1 of x and is identically 1

outside V (see Exercise 4.13 for the existence of such a function). Then f = φf and
Xx (f ) = (Xx φ)f (x) + φ(x)Xx f = 0.
The second claim follows from 1·1 = 1 and the definition of a derivation. The third one is also immediate. ¤
Let x ∈ Rn . Consider the set
F̃x := {(f, V ) : V is a neighborhood of x, f ∈ F(V )}.
3.19. Definition. The algebra of germs of smooth functions at x, Fx , is the quotient of F̃x by the
equivalence relation (f, U ) ∼ (g, V ) ⇐⇒ f = g in a neighborhood of x (contained in U ∩ V ). The operations
are the usual sum and product of functions (which are defined in the intersections of the domains).
3.20. Remark. The advantage of working with germs instead that with functions is that we do not have
to worry about the domain of definition of the functions involved. Anyway, when clear from the context we
will make no difference between a function and its germ.
We will denote by Dx the space of derivations of Fx at x (with the obvious definition). Lemma 3.18
implies, in particular, that an element of Derx (U ) induces a derivation of Fx . We shall see next that all
derivations in Dx are of this type.
3.21. Theorem. Given x ∈ Rn and a derivation Xx ∈ Dx , there exist a unique vector v ∈ Rn such that
∼ Tx Rn ∼
Xx = v(x). In particular Dx = = Derx (U ).
Proof. Let f ∈ Fx . In a suitable neighborhood of x consider the Taylor formula

n
X ∂f
f (x1 , . . . , xn ) = f (x) + (x)(xi − xi (x)) + Φ(x1 . . . . , xn ),
1
∂xi
where Φ(x1 . . . . , xn ) is a sum of products of functions vanishing at x (see Theorem 1.19 and Exercise 4.26).
Applying Xx to both sides and using Lemma 3.18 we have:

n
X ∂f
Xx (f ) = Xx (xi ) (x).
1
∂xi
Therefore:
n
X ∂
X= X(xi ) (x),
1
∂xi
∂
and the map that associates to ei the derivation (x) extends to an isomorphism of Rn (or, better Tx U )
∂xi
onto Dx . ¤
In what follows we will identify Tx U with Dx and H(U ) with Der(U ).

The composition of two derivations is not a derivation, in general. However the commutator of two
derivations is a derivation ( Exercise 4.28). This fact suggests the following
3.22. Definition. Let X, Y ∈ Der(U ). The Lie product (or bracket) of X and Y is the commutator
[X, Y ] := X ◦ Y − Y ◦ X.
The following properties are easy to prove and we leave the details to the reader (Exercise 4.29).
3.23. Proposition. The Lie product [ ·, · ] : H(U ) × H(U ) −→ H(U ) is a R-bilinear map. Moreover
(1) [X, Y ] = −[Y, X],
(2) [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0 (Jacoby identity).
3.24. Remark. An algebra with a product which satisfies the properties above is called a Lie algebra.
∂ ∂
3.25. Example. By Theorem 1.15 [ , ] = 0.
∂xi ∂xj
We go back to the original question: given vector fields X1 , . . . , Xk ∈ H(U ), linearly independent at
∂
each point, there exist local coordinates (x1 , . . . , xn ) in Rn such that Xi = ?
∂xi
There is a natural necessary condition for a positive answer, the condition being [Xi , Xj ] = 0. It turns
out that the condition is also sufficient, at least locally. We will take a slightly more general approach.
3.26. Definition. Let U ⊆ Rn be an open set. A k-dimensional distribution D on U is a law that

associates to a point x ∈ U a k-dimensional subspace Dx ⊆ Rn . Moreover:
• The distribution D is smooth if there exist, locally, k smooth vector fields X1 , . . . , Xk such that
Dx = span{X1 (x), . . . , Xk (x)}.
• A smooth distribution D is involutive (or integrable) if for all vector fields X, Y ∈ H(U ) such that
X(x), Y (x) ∈ Dx , ∀ x ∈ U , then [X, Y ](x) ∈ Dx .
3.27. Theorem. [Frobenius Theorem, local version] Let D be a k-dimensional involutive smooth distribu-
∂ ∂
tion on U ⊆ Rn . Then there exist (local) coordinates x1 , . . . , xn such that Dx = span{ (x), . . . , (x)}.
∂x1 ∂xk
3.28. Remark. The word “local” means that the claim of the Theorem holds in a sufficiently small open
neighborhood of a fixed point, that we can assume to be 0 ∈ Rn .
Proof. We will proceed by induction on k. If k = 1, the Theorem follows directly from Theorem 3.11.
So we assume that the Theorem is true for (k − 1)-dimensional involutive distributions. Let us suppose that
D is a k-dimensional distribution spanned, locally, by smooth vector fields X1 , . . . , Xk . By Theorem 3.11 we
∂
can assume that there are coordinates y1 , . . . , yn such that X1 = . Consider the set
∂y1
D = {X ∈ H(U ) : X(x) ∈ Dx , X(y1 ) = 0}.
Claim 1. D is a smooth (k − 1)-dimensional involutive distribution spanned by the vector fields
Yi = Xi − Xi (y1 )X1 , i = 2, . . . , k.
Proof. It is easy to see that the vector fields X1 , Y2 , . . . , Yk are linearly independent at every point.
Moreover Yi ∈ D and X1 6∈ D, since X1 (y1 ) = 1. So D is a smooth (k − 1)-dimensional distribution.
Let us show that D is involutive. If Y, Z ∈ D, [Y, Z] ∈ D since D is involutive. Moreover [Y, Z](y1 ) =
Y (Z(y1 )) − Z(Y (y1 )) = 0, hence [Y, Z] ∈ D. ¤
Observe that D is tangent to the slides Rn−1

c := {(y1 , . . . , yn ) ∈ Rn : y1 = c}, since the first coordinate
of Yi is Yi (y1 ) = 0. By the inductive hypothesis the are (local) coordinates z2 , . . . zn , in Rn−1 0 such that
∂
D(0,z) = span{ , i = 2, . . . , k}. Consider the coordinates x1 = y1 , xi = zi , i = 2, . . . , n. The proof of the
∂zi
Theorem follow from the following
∂
Claim 2. D is spanned by , i = 1, . . . , k.
∂xi
∂ ∂
Proof. We want to show that Y1 := X1 = , Y2 , . . . , Yk are linear combinations of , i = 1, . . . , k.
∂x1 ∂xi
For this is sufficient to show that Yi (xj ) = 0 for i ≤ k, j > k (this is obviously true for i = 1).
Since the distribution is involutive, there are real valued smooth functions girs such that [Yi , Yr ] =
Pk
s=1 girs Ys . Now
k
X
Y1 (Yi )(xj ) = [Y1 , Yi ](xj ) = girs Ys (xj ).
1
Hence the functions Yi (xj ) are solutions of the system of differential equations
X k
∂
Yi (xj ) = girs Ys (xj ).
∂x1 1
This is a linear homogeneous system of ordinary differential equations, along the x1 curves, hence it admits
the zero functions as solutions. Now the initial condition, for x1 = 0, is Yi (xj )(0, x2 , . . . , xn ) which vanishes
(for j > k) since there xj = zj . Hence, by unicity of the solutions of the initial value problem, the solutions
vanish identically. ¤
3.29. Remark. The Frobenius Theorem is really a result on existence and unicity of solutions of first
order partial differential equations. We will sketch the proof of a simple fact that will explain this claim.
Let a, b : R2 −→ R be smooth functions and consider the problem of finding a function f : R2 −→ R
such that
∂f ∂f
= a, = b.
∂x ∂y
The Theorem of Schwarz (Theorem 1.15) gives an obvious necessary condition for the existence of such
∂b ∂a
a function, that is = . We will use the Theorem of Frobenius to shows that, at least locally, such
∂x ∂y
condition is also sufficient4. Consider the vector fields
∂ ∂ ∂ ∂
X= +a Y = +b .
∂x ∂z ∂y ∂z
∂b ∂a ∂
A simple calculation gives [X, Y ] = ( − ) . Hence the distribution spanned by X, Y is involutive if
∂x ∂y ∂z
and only if [X, Y ] = 0. In this case, by the Frobenius Theorem, there is a local diffeomorphism Φ : R3 −→ R3
∂ ∂
such that dΦ( ) = X, dΦ( ) = Y . The “surface” Φ(x, y, c) has the distribution spanned by X and Y
∂x ∂y
as “tangent space” and, since the normal vector is not horizontal, it project (locally) on the plane e⊥ 3,
injectively. Hence it is the graph of a function f that is, as it is easily seen, a solution of our problem.
The differential equation above is the simplest case of a class of differential equation, called total differ-
ential equations, for which necessary and sufficient conditions for existence and unicity of solutions may be
given in terms of the Theorem of Frobenius.
There is an interpretation of the Lie product of vector fields worth mentioning.
3.30. Proposition. Let X, Y ∈ H(U ) and φt be the (local) flow of X. Then, for x ∈ U ,
[X, Y ](x) = lim t−1 [dφ−t (φt (x))Y (φt (x)) − Y (x)].
t→0
∂ X ∂
Proof. By Theorem 3.11 we can assume X = . Let Y = yi . By linearity we can assume
∂x1 ∂xi
∂
Y = yi . Observe that the flow of X is just translations, i.e. φt (x1 , . . . , xn ) = (x1 + t, . . . , xn ). Then the
∂xi
∂Y ∂yi ∂ ∂yi ∂
right hand side is just (x) = (x) (x). On the other hand, the left hand side is also (x) (x)
∂x1 ∂x1 ∂yi ∂x1 ∂yi
(see Exercise 4.30). ¤
An important fact about Lie product is that it “behaves well with respect to smooth maps”. First a
definition to make the statement precise. Let F : U ⊆ Rn −→ V ⊆ Rm be a smooth map between open
sets.
3.31. Definition. We say that X̃ ∈ H(V ) is F -related to X ∈ H(U ) if dF (x)(X) = X̃(F (x)), ∀ x ∈ U .
3.32. Proposition. If X̃, Ỹ ∈ H(V ) are F -related to X, Y ∈ H(U ), then [X̃, Ỹ ] is F -related to [X, Y ].
Proof. Let f ∈ F(V ). We must show that, fixed x ∈ U, dF ([X, Y ](x))(f ) = [X̃, Ỹ ](F (x))(f ).
dF ([X, Y ](x))(f ) = [X, Y ](x)(f ◦F ) = X(x)(Y (f ◦F ))−Y (x)(X(f ◦F )) = X(x)(Ỹ (f )◦F )−Y (x)(X̃(f )◦F ) =
= dF (X(x))(Ỹ (f )) − dF (Y (x))(X̃(f )) = X̃(F (x))(Ỹ (f )) − Ỹ (F (x))(X̃(f )) = [X̃, Ỹ ](x)(f ).
4A different proof will be given in Chapter 1.

4. EXERCISES 24
4. Exercises
4.1. For L, T ∈ L(Rn , Rm ) consider the norms kLk2 = traceLt L, kLk = sup{kL(x)k : kxk = 1}. Prove
that kLk ≤ kLk2 ≤ nkLk and kL ◦ T k ≤ kLkkT k.
4.2. Prove Proposition 1.3.
4.3. Let f, g : U −→ Rn be differentiable functions. Define F : U −→ R, F (x) = hf (x), g(x)i. Prove

that F is differentiable and compute dF (x) (see Example 1.5).
4.4. Consider the function f : R −→ R2 , f (t) = (cos t, sin t). Compute df (t) and show that there is no
t0 ∈ [0, 2π] such that f (2π) − f (0) = df (t0 )(1)2π (So the mean value Theorem, in the form 1.6, is not true
if the dimension of the target space is greater than 1).
4.5. Let f : U −→ Rm be a differentiable function. Use Theorem 1.7 to prove

(1) if df (x) = 0, ∀ x ∈ U , then f is locally constant. In particular, if U is connected, f is constant,
(2) if T ∈ L(Rn , Rm ). Then
kf (a + h) − f (a) − T (h)k ≤ khk sup{kdf (a + th) − T k : t ∈ [0, 1]}.
4.6. Prove that the function (

tk if t > 0
fk (t) =
0 if t ≤ 0
k−1 k
is of class C but is not of class C .
4.7. Consider the function f : R2 −→ R

x2 y
f (x, y) = , (x, y) 6= (0, 0), f (0, 0) = 0.
x2 + y 2
Prove that the partial derivatives at (0, 0) exist, but f is not differentiable at (0, 0).
4.8. Let f : Rn −→ Rm be a function.

(1) Prove that if f (tx) = tf (x), ∀ t ∈ R, and f is differentiable at 0, then f is linear.
(2) Prove that if f (tx) = |t|f (x), ∀ t ∈ R, and f is differentiable at 0, then f vanishes identically.
(3) Prove that if f (tx) = t2 f (x), ∀ t ∈ R, and f is twice differentiable at 0, then f is bilinear.
(4) Prove that if f (tx) = tk f (x), ∀ t ∈ R and f is of class C k , then
1
di f (x)(h1 , . . . , hi ) = dk f (0)(x, . . . , x, h1 , . . . , hi ).
(k − i)!
4.9. Let k · k : Rn −→ R be a norm.
(1) Prove that the function f (x) = kxk is not differentiable at 0.
(2) Prove that the function f 2 (x) = kxk2 is differentiable at 0 and df 2 (0) = 0.
(3) Prove that the norm is induced by a scalar product if and only if f 2 is twice differentiable at 0.
4.10. Prove that, if {xn } ⊆ Rn is a Cauchy sequence admitting a subsequence converging to x ∈ Rn ,

then the whole sequence converges to x.
P
4.11. Let {xn } ⊆ Rn be a sequence. Define convergence for the series xn and prove the Cauchy
convergence criterion for series.
4.12. Let M (n, R) the space of n × n matrices with real entries. Consider the natural identification with
n2
R and define exp : M (n, R) −→ M (n, R), by:
∞
X 1 k
exp(A) = A .
k!
k=0
(1) Prove that exp is well defined (i.e. the series

Ã converges).
!
P k k
(2) Prove that, if AB = BA, (A + B)k = i=0 Ai B k−i . Conclude that, if AB = BA, exp(A +
i
B) = exp(A) exp(B).
(3) Prove that exp(P AP −1 ) = P exp(A)P −1 .
(4) Let A be an upper triangular matrix. Compute the diagonal entries of exp(A).
(5) Show that det(exp(A)) = etrace(A) , ∀A ∈ M (n, R). Conclude that exp(A) is invertible ∀A ∈
M (n, R). Hint: put A in upper diagonal form.
d
(6) Show that exp is differentiable and compute d exp(A)(B). Hint: compute dt |t=0 exp(A + tB).
(7) Show that d exp(0) = 11. In particular exp maps diffeomorphically a neighborhood of 0 onto a
neighborhood of exp(0) = 11. The (local) inverse is the logarithm.
4.13. Consider the function

( 1
e− t if t > 0
f (t) =
0 if t ≤ 0
(1) Prove that f is smooth.
(2) Let 0 < δ1 < δ2 . Prove that the function
f (kxk2 − δ12 )
φ(x) =
f (kxk2 − δ12 ) + f (δ22 − kxk2 )
is a well defined smooth function with values in [0, 1], that vanishes for kxk ≤ δ1 and is identically
1 for kxk ≥ δ2 .
4.14. Consider the map φ : B n (1) −→ Rn , φ(x) = x(1 − kxk2 )−1 . Prove that φ is a diffeomorphism.
4.15. Use the local form of subimmersions (Theorem 1.27), to prove the following
Theorem [Implicit function Theorem] Let U ⊆ Rn ×Rm be an open set. f : U −→ Rm a smooth function
such that, for z0 = (x0 , y0 ) ∈ U, f (z0 ) = 0, and d2 f (z0 ) : Rm −→ Rm is an isomorphism. Then there exists
a neighborhood V ⊆ Rn of x0 and a unique smooth function g : V −→ Rm such that f (x, g(x)) = 0 ∀ x ∈ V .
Moreover
dg((x) = −[d2 f (x, g(x))]−1 ◦ d1 f (x, g(x))
(di f is defined in Remark 1.12).
4.16. Prove Proposition 2.3

4. EXERCISES 26
4.17. Let T : [a, b] ⊆ R −→ L(Rn , Rm ) be an integrable function and Y ∈ Rm . Prove that the function
f (t) = T (t)Y is integrable and
Z b Z b
f =[ T ]Y.
a a
4.18. Prove that, if f : [a, b] −→ Rm is integrable and x ∈ [a, b] then f |[a,x] and f |[x,b] are integrable and
Z b Z x Z b
f= f+ f.
a a x
m
4.19. Let f : [a, b] −→ R be such that f (t) = 0 for t outside a finite set. Prove that f is integrable and
Z b
f = 0. Conclude that if two functions f, g : [a, b] −→ Rm differ only on a finite set, then one is integrable
a
if and only if the other one is integrable and, in this case, the two integral coincide.
4.20. Prove that any continuous function f : [a, b] ⊆ R −→ Rm is uniform limit of step functions.
4.21. A curve γ : [a, b] −→ Rm is said to be rectificable if there exists l = l(γ) ∈ R (called the length of
γ) such that for all ² > 0 there exists δ > 0 such that if P = {t0 , . . . , tk } is a partition with |P | < δ, we have
k−1
X
|l − kγ(ti+1 ) − γ(ti )k | < ².
0
Z b Z b
Prove that if γ is of class C 1 , γ is rectificable and l(f ) = γ̇(t)dt := dγ(t)(1)dt.
a a
4.22. Use Fubini’s Theorem (Theorem 2.17) to prove Theorem 1.15. Hint: it is nor restrictive to assume
∂f ∂f
f : R2 −→ R (why?). If − > 0 at z0 = (x0 , y0 ) so it is in a small rectangle C = [a, b] × [c, d]
∂x∂y ∂y∂x
containing z0 . Show that the integral over C of the difference is zero.
4.23. Let X : U ⊆ Rn −→ Rn be a smooth vector field. Prove that if supp(X) := {x ∈ U : X(x) 6= 0} is

compact, then X is complete.
4.24. Let X : Rn −→ Rn be a smooth vector field. Prove that, if there is a constant M with kX(x)k ≤
M, ∀ x ∈ Rn , then X is complete (hint: show that an integral curve γ : [0, a) −→ Rn has finite length, if
a < ∞, so its image has compact closure).
4.25. Give an example of a non complete vector field in R.
4.26. Let f ∈ F(U ), 0 ∈ U ⊆ Rn , f (0) = 0. Prove that there exist functions gi ∈ F(Ui ) where U 0 ⊆ U
Pn ∂f
is an open neighborhood of 0, such that f (xi , . . . xn ) = i=1 xi gi (xi , . . . xn ) and gi (0) = (0). Hint: write
Z 1 ∂xi
df (tx)
f (x) = dt.
0 dt
4.27. Consider F0 , the algebra of germs of smooth functions at 0 ∈ Rn and I0 = {[f ] ∈ F0 : f (0) = 0}.
(1) Prove that I0 is the unique maximal (non trivial) ideal of F0 .
(2) Let I02 be the ideal generated by products of two elements in I0 . Prove that I0 /I02 is a n-dimensional
real vector space spanned by the (equivalence classes of) the germs of the coordinate functions.
Conclude that I0 /I02 is canonically isomorphic to [Rn ]∗ .
4.28. Prove that if X, Y ∈ Der(U ) then [X, Y ] := X ◦ Y − Y ◦ X ∈ Der(U ).
4.29. Prove Proposition 3.23.
4.30. Let
X ∂ X ∂
X= ak (x) , Y = bk (x)
∂xk ∂xk
k k
be smooth vector fields in Rn .
∂
(1) Compute [X, Y ] in the basis .
∂xk
(2) Let X1 , . . . , Xp be linear independent vectors in Rn . Show that there exist smooth vector fields
X̃1 , . . . , X̃p in Rn such that, for a fixed x ∈ U, X̃i (x) = Xi and [X̃i , X̃j ] = 0 in Rn .
CHAPTER 1
The de Rham cohomology for open sets of Rn
1. Exterior forms
Let E be a finite dimensional real vector space and E∗ its dual. We will identify, as usual, E with the
double dual (E∗ )∗ := E∗∗ .
1.1. Definition. A tensor of type (p, q) in E is a multilinear1 map:
t : E∗ × · · · × E∗ × E × · · · × E −→ R
| {z } | {z }
p times q times
We will denote by E(p,q) the space of these tensors. This is a real vector space with the operations of sum
of multilinear maps (summing the values) and product by a scalar (multiplying the values by the scalar).
1.2. Examples.
• E(0,1) = E∗ , E(1,0) = E∗∗ = E.
• A scalar product in E is an element of E(0,2) .
• It is convenient to define E(0,0) := R.
We will be interested mainly in tensors of type (0, q). To simplify the notations we will set Eq := E(0,q) .
Beside adding tensors, we can multiply them.
1.3. Definition. Given ω ∈ Ep , τ ∈ Eq , we define the tensor product ω ⊗ τ ∈ Ep+q as
ω ⊗ τ (x1 , . . . , xp+q ) := ω(x1 , . . . , xp )τ (xp+1 , . . . xp+q ).
1.4. Remark. It is easy to see that the tensor product is associative and distributive (Exercise 7.1) and
therefore, suitable extended, define an associative algebra structure in E∗ := ⊕Ep . Whit this structure E∗ is
called the tensor algebra.
1.5. Proposition. Let {ω1 , . . . , ωn } be a basis of E1 = E∗ . Then the set {ωi1 ⊗ · · · ⊗ ωiq : i1 , . . . , iq ∈
{1, . . . , n}} is a basis of Eq .
Proof. Let {e1 , · · · , en } be the dual basis, i.e., ωi (ej ) = δij . Then:
X
ai1 ···iq ωi1 ⊗ · · · ⊗ ωiq (ej1 , . . . , ejq ) = aj1 ···jq .
1i.e. linear in each variable.
29
1. EXTERIOR FORMS 30
It follows, by a standard argument, that the the elements of the set in question are linearly independent.
P
Now, given ω ∈ Eq we define ai1 ···iq = ω(ei1 , . . . , eiq ). It is easy to check that ω = ai1 ···iq ωi1 ⊗ · · · ⊗ ωiq ,
and this concludes the proof.
¤
We will be interested in special elements of Eq . Let Σ(p) be the group of permutations of {1, . . . , p} ⊆ N.
If π ∈ Σ(p), we will denote by |π| the sign of π, i.e. |π| = 1 if π is the product of an even number of
transpositions and |π| = −1 otherwise.
1.6. Definition. Let ω ∈ Ep . We will say that

• ω is a symmetric form if ω(x1 , . . . , xp ) = ω(xπ(1) , . . . , xπ(p) ), ∀ π ∈ Σ(p).
2
• ω is an exterior form if ω(x1 , . . . , xp ) = |π|ω(xπ(1) , . . . , xπ(p) ), ∀ π ∈ Σ(p).
We will denote by Σp (E) the space of symmetric tensors in Ep and with Λp (E) the space of exterior
p-forms. These are subspaces of Ep . Clearly Λ0 (E) = R = Σ0 (E), Λ1 (E) = E1 = E∗ = Σ1 (E).
We will be mostly interested in exterior forms and we will describe now the basic example.
1.7. Example. Let {e1 , . . . , en } be a fixed basis of E and {φ1 , . . . , φn } be the dual basis. Let us fix
indexes 1 ≤ i1 < · · · < ip ≤ n and define:
ω(i1 ,...,ip ) (x1 , . . . , xp ) := det(φij (xk )).
In other words we consider the matrix whose k th column is given by the coordinates of xk in the fixed
basis, and compute the determinant of the sub matrix obtained considering only the lines (i1 , . . . , ip ) of the
original matrix. The ω(i1 ,...,ip ) ’s are exterior p-forms since the determinant is multilinear in the columns
and, permuting the columns the sign changes according to the parity of the permutation. As we shall see
(Proposition 1.20 and Remark 1.18), these forms are a basis of Λp (E).
1.8. Remark. By Example 1.7 p-forms are, essentially, determinants of p×p matrices and, therefore, “p-
dimensional (oriented) volume elements”. So they appear as the natural integrands of the multiple (oriented)
integrals. This statement will be made precise in the next chapter.
The tensor product of exterior forms is not, in general, an exterior form. But we can “alternate” the
tensor product in order to obtain an exterior form. Define the linear operator
1 X
A : Ep −→ Ep , A(τ )(x1 , . . . , xp ) = |π|τ (xπ(1) , . . . , xπ(p) ).
p!
π∈Σ(p)
1.9. Proposition.
(1) If τ ∈ Ep , A(τ ) ∈ Λp (E).
(2) If τ ∈ Λp (E), A(τ ) = τ .
In particular A2 = A.
2The terms alternating tensor or skew symmetric tensor are also used in the literature.
CHAPTER 1. THE DE RHAM COHOMOLOGY FOR OPEN SETS OF RN 31
Proof. If p = 1 there is nothing to prove, so we assume p > 1. For i, j ∈ {1, . . . , p}, we will denote by
(ij) the element of Σ(p) that interchanges i and j and leaves the other integers fixed. If π ∈ Σ(p), we set
π 0 = π ◦ (ij). Then |π 0 | = −|π| and
1 X
A(τ )(x1 , . . . , xj , . . . , xi , . . . , xp ) = |π|τ (xπ(1) , . . . , xπ(j) , . . . , xπ(i) , . . . , xπ(p) ) =
p! π
1 X
|π|τ (xπ0 (1) , . . . , xπ0 (i) , . . . , xπ0 (j) , . . . , xπ0 (p) ) =
p! π
1 X
−|π 0 |τ (xπ0 (1) , . . . , xπ0 (i) , . . . , xπ0 (j) , . . . , xπ0 (p) ) = −A(τ )(x1 , . . . , xi , . . . , xj , . . . , xp )
p! 0
π
It is easy to see that the equation above implies that A(τ ) ∈ Λp (E) (see Exercise 7.2). Moreover, if τ ∈ Λp (E),
1 X 1 X 2
A(τ )(x1 , . . . , xp ) = |π|τ (xπ(1) , . . . xπ(p) ) = |π| τ (x1 , . . . xp ) = τ (x1 , . . . , xp )
p! π p! π
and this proves the second claim.

¤
Observe that, in general, A(φ ⊗ ψ) 6= A(φ) ⊗ A(ψ). However we have
1.10. Lemma. If φ1 , . . . , φp ∈ E∗ , then:

1 X
A(φ1 ⊗ · · · ⊗ φp ) = |σ|φσ(1) ⊗ · · · ⊗ φσ(p) .
p!
σ∈Σ(p)
Proof.
1 X
A(φ1 ⊗ · · · ⊗ φp )(x1 , . . . , xp ) = |σ|φ1 ⊗ · · · ⊗ φp (xσ(1) , . . . , xσ(p) ) =
p!
σ∈Σ(p)
1 X 1 X
|σ|φ1 (xσ(1) ) · · · φp (xσ(p) ) = |σ|φσ(1) (x1 ) · · · φσ(p) (xp ).
p! p!
σ∈Σ(p) σ∈Σ(p)
Using the operator A we can define product of exterior forms.
1.11. Definition. The exterior (or wedge) product is defined as the map
(p + q)!
∧ : Λp (E) × Λq (E) −→ Λp+q (E), ∧(ω, τ ) := ω ∧ τ = A(ω ⊗ τ ).
p!q!
(p + q)!
(The reason for the coefficient will be discuss in Remark 1.19.)
p!q!
It is easy to prove that the exterior product is distributive (see Exercise 7.3). In particular, suitable
extended, it defines an algebra structure on Λ∗ (E) := ⊕Λp (E). Λ∗ (E) is called the exterior algebra.
It is also true that the exterior product is associative, but this fact is a little bit tricky. The proof
involves a characterization of the kernel of A. The problem is that A is not an algebra homomorphism, hence
we can not conclude, directly, that ker A is an ideal. We will prove that, in fact, ker A is an ideal.
Consider the ideal I ⊆ E∗ generated by φ ⊗ φ, φ ∈ E∗ . This is the vector subspace of E∗ generated by

elements of the form τ ⊗ φ ⊗ φ, ψ ⊗ ψ ⊗ η, φ, ψ ∈ E∗ , τ, η ∈ E∗ or, alternatively, it is the intersection of all
ideals containing the elements of the form φ ⊗ φ, φ ∈ E∗ .
1.12. Theorem. ker A = I.
Proof. It is easily seen that I ⊆ ker A. We will prove that ker A ⊆ I. Consider the quotient algebra
E∗ /I. Denote by · the product in this quotient and by π : E∗ −→ E∗ /I the projection map, which is an
algebra homomorphism. First observe that, if φ, ψ ∈ E∗ :
0 = π((φ + ψ) ⊗ (φ + ψ)) = π(φ ⊗ φ + φ ⊗ ψ + ψ ⊗ φ + ψ ⊗ ψ) = π(φ ⊗ ψ) + π(ψ ⊗ φ),
i.e. π(φ ⊗ ψ) = −π(ψ ⊗ φ). Therefore, for φ1 , . . . , φp ∈ E∗ and σ ∈ Σ(p), we have

¡ ¢
π(φσ(1) , ⊗ . . . , ⊗φσ(p) ) = π φσ(1) ) · · · π(φσ(p) = |σ|π(φ1 ) · · · π(φp ) = |σ|π(φ1 ⊗ · · · ⊗ φp ).
Hence
1 X 1 X
π(A(φ1 ⊗· · ·⊗φp )) = π( |σ|π(φσ(1) ⊗ · · · ⊗ φσ(p) )) = |σ|2 π(φ1 ⊗ · · · ⊗ φp ) = π(φ1 ⊗ · · · ⊗ φp ).
p! p!
σ∈Σ(p) σ∈Σ(p)
So any element in ker A is in I := ker π. ¤
1.13. Corollary. Let ω ∈ Ep , τ ∈ Eq . If A(ω) = 0, A(ω ⊗ τ ) = 0 = A(τ ⊗ ω).
Proof. This follows from the fact that ker A is an ideal. ¤
At this point we can prove the announced result
1.14. Proposition. The wedge product is associative.
Proof. First we observe that:
A(A(ω ⊗ η) ⊗ θ)) = A(ω ⊗ η ⊗ θ) = A(ω ⊗ A(η ⊗ θ)).
In fact, by 1.9, A(A(η ⊗ θ) − η ⊗ θ) = 0 and, by 1.13, we have that:
0 = A(ω ⊗ [A(η ⊗ θ) − η ⊗ θ]) = A(ω ⊗ A(η ⊗ η) − ω ⊗ η ⊗ θ) = A(ω ⊗ A(η ⊗ θ)) − A(ω ⊗ η ⊗ θ),
which proves the second equality. The first one is proved in a similar way.
Therefore, if ω ∈ Λk (E), η ∈ Λl (E), θ ∈ Λm (E), we have:
(k + l + m)! (k + l + m)! (k + l)!
(ω ∧ η) ∧ θ = A((ω ∧ η) ⊗ θ) = A(ω ⊗ η ⊗ θ),
(k + l)!m! (k + l)!m! k!l!
and the associativity follows from the associativity of the tensor product. ¤
1.15. Example. Let φ1 , φ2 ∈ E∗ = E1 , x1 , x2 ∈ E. Then:

1
φ1 ∧ φ2 (x1 , x2 ) = 2 (φ1 (x1 )φ2 (x2 ) − φ1 (x2 )φ2 (x1 )) = det[φi (xj )].
2
More generally, an induction on p gives:
1.16. Proposition. Let φi ∈ E∗ , xj ∈ E i, j = 1, . . . , p. Then:
φ1 ∧ · · · ∧ φp (x1 , . . . , xp ) = det[φi (xj )].
In particular if σ ∈ Σ(p), φ1 ∧ · · · ∧ φp = |σ|φσ(1) ∧ · · · ∧ φσ(p) .
1.17. Remark. Observe that, by 1.14, the form φ1 ∧ · · · ∧ φp is well defined.
1.18. Remark. In the Example 1.7 the form ω(i1 ,...,ip ) is just φi1 ∧ · · · ∧ φip .
(p + q)!
1.19. Remark. The coefficient in 1.11 is convenient in order to avoid unpleasant coefficients in
p!q!
1.16 and also for a geometric reason: let E be an inner product space, {e1 , . . . , en } an orthonormal basis and
{φ1 , . . . , φn } the dual basis (so φi (ej ) = hei , ej i = δij ). Given vectors x1 , . . . , xn ∈ E, φ1 ∧· · ·∧φn (x1 , . . . , xn )
is the “volume” of the parallelepiped of edges the x0i s. The coefficient above is such that the “unit cube”,
i.e. the parallelepiped spanned by the ei ’s, has volume 1 (see Definition 1.28).
1.20. Proposition. Let {φ1 , . . . , φn } be a basis for E∗ . Then
{φi1 ∧ · · · ∧ φip : 1 ≤ i1 < · · · < ip ≤ n}

µ ¶
n
is a basis of Λp (E). In particular Λp (E) has dimension and Λp (E) = {0}, if p > n.
p
Proof. Let {e1 , . . . , en } be the dual basis. First observe that φ1 ∧ · · · ∧ φn (e1 , . . . , en ) = det[φi (ej )] = 1.
Also observe that φi ∧ φj = −φj ∧ φi and, in particular, if we interchange two elements in the product
φi1 ∧ · · · ∧ φip the form changes sign. We will prove now that the forms {φi1 ∧ · · · ∧ φip : i1 < · · · < ip } are
linearly independent. Suppose
X
ai1 ···ip φi1 ∧ · · · ∧ φip = 0.
i1 <···<ip
We want to show that ai1 ···ip = 0. We will do it for a1···p , the other cases being analogous. Observe that
X
ai1 ···ip φi1 ∧ · · · ∧ φip ∧ φp+1 ∧ · · · ∧ φn (e1 , . . . , en ) = a1···p = 0,
i1 <···<ip
since the terms with {i1 , . . . , ip } 6= {1, . . . , p} vanish (they contain two equal indexes), and the conclusion
follows. We leave to the reader the task of showing that they span Λp (E) (Exercise 7.4).
¤
1.21. Corollary. The algebra Λ∗ (E) is graded commutative3 , i.e. if ω ∈ Λp (E), τ ∈ Λq (E)
ω ∧ τ = (−1)pq τ ∧ ω.
In particular the square of a form of odd degree is zero.
Proof. As we have seen this is true for products of decomposable elements (i.e. elements of the form
φi1 ∧ · · · ∧ φip ). The general case follows from the fact that such forms span the exterior algebra. ¤
1.22. Remark. There is a restriction, in Proposition 1.20, on the set of indexes with respect to Propo-
sition 1.5 and this is due to the graded commutativity of the exterior algebra.
3An algebra E, with product b : E ⊕ E −→ E is a graded algebra if there is a sequence of vector subspaces E such that
i
E = ⊕Ei and b(Ei ⊕Ej ) ⊆ Ei+j . Such an algebra is said to be graded commutative if for ω ∈ Ep , τ ∈ Eq , b(ω, τ ) = (−1)pq b(τ, ω).
Let L : E −→ F be a linear map. Recall that the transpose of L is the map
L∗ : F∗ (= F1 ) −→ E∗ (= E1 ), L∗ (φ)(x) := φ(Lx).
This map extends to a linear map
Ep (L) : Fp −→ Ep , Ep (L)(ω)(x1 , . . . , xp ) = ω(L(x1 ), . . . , L(xp )).
It is simple to see that if ω ∈ Λp (F) then Ep (L)(ω) ∈ Λp (E). So we get, by restriction, a linear map
Λp (L) := Ep (L)|Λp (F) : Λp (F) −→ Λp (E),
and, by additivity, a linear map Λ∗ (L) : Λ∗ (F) −→ Λ∗ (E).

When clear from the context we will write L∗p , or just L∗ , for Λp (L) and Λ∗ (L).
1.23. Proposition. L∗ (ω ∧ τ ) = L∗ (ω) ∧ L∗ (τ ). This means that L induces a graded algebra homomor-
phism L∗ : Λ∗ (F) −→ Λ∗ (E). Moreover we have the following properties, called the funtorial properties4
(1) (11E )∗ = 11Λ∗ (E) .
(2) If L : E −→ F and T : F −→ G are linear maps, then (T ◦ L)∗ = L∗ ◦ T ∗ .
Proof. To prove the first assertion, we just observe that, if φi ∈ E∗ , xj ∈ E, i, j = 1, . . . , p, we have:
L∗p (φ1 ∧ · · · ∧ φp )(x1 , . . . , xp ) = det[φi (Lxj )] = det[L∗ (φi )(xj )] = L∗ (φ1 ) ∧ · · · ∧ L∗ (φp )(x1 , . . . , xp ).
Since Λp (E) is spanned by elements of the form φ1 ∧ · · · ∧ φp , by Proposition 1.20, the conclusion follows by
linearity. The functorial properties are obvious. ¤
1.24. Remark. We will meet often, along these notes, “functorial properties”. These properties are
usually trivial to prove, but important. For example, in the context of Proposition 1.23, they imply that, if
L is an isomorphism, then L∗ is also an isomorphism (see Exercise 7.15).
Let E be a finite dimensional real vector space with an inner product h·, ·i : E × E −→ R.
1.25. Definition. The isomorphisms
[ : E −→ E∗ , [(x)(y) = hx, yi, ] : E∗ −→ E, ] := [−1 ,
are called the musical isomorphisms.
We define an inner product in E∗ by requiring [ to be an isometry. We can also define an inner product
in Λp (E) extending, by bi-linearity, the formula
hφ1 ∧ · · · ∧ φp , ψ1 ∧ · · · ∧ ψp i = det(hφi , ψj i).

Observe that, if {ωi } is an orthonormal basis for E∗ , the basis {ωi1 ∧ · · · ∧ ωip : i1 < · · · < ip } is orthonormal.
We recall that two bases of a n-dimensional real vector space E are equioriented if the matrix that gives
the change of bases has positive determinant. This relation is an equivalence relation and the set of bases of
E is divided into two equivalence classes.
4In the language of category theory this means that the low that associate to a finite dimensional real vector space E
the graded algebra Λ∗ (E) and to a linear maps L : E −→ F the map L∗ is a contravariant functor from the category of finite
dimensional real vector spaces and linear maps, to the category of algebras and their homomorphisms.
1.26. Definition. An orientation on E is the choice of one of two equivalence classes of equioriented
bases. E is oriented if such a choice has been made and the bases in the chosen class will be called positive.
1.27. Remark. Naturally an orientation in E induces an orientation on E∗ , by declaring positive the

bases that are dual of positive bases of E.
1.28. Definition. Let E be a n-dimensional oriented inner product space and let {ω1 , . . . , ωn } be a
positive orthonormal basis of E∗ . The volume form of E is the n-form v = ω1 ∧ · · · ∧ ωn .
1.29. Lemma. The volume form is well defined, i.e. it does not depend on the choice of the basis.
P
Proof. Let {ωi }, {φj } be bases of E∗ and A = (aij ) such that φk = akj ωj . Then
X
φ1 ∧ · · · ∧ φn = |σ|a1σ(1) · · · anσ(n) ω1 ∧ · · · ∧ ωn = det(A)ω1 ∧ · · · ∧ ωn .
σ∈Σ(n)
If the bases are orthonormal and positive, then A ∈ SO(n). In particular det(A) = 1. ¤
1.30. Definition. Let E be a n-dimensional oriented inner product space. The Hodge (star) operator
is the operator
∗p : Λp (E) −→ Λ(n−p) (E), ∗p (η)(x1 , . . . , x(n−p) ) := hη ∧ [(x1 ) ∧ · · · ∧ [(x(n−p) ), vi,
where v is the volume form. When clear from the context, we will write simply ∗ instead of ∗p .
1.31. Remark. Let {ωi } be a positive orthonormal basis for E∗ . Then the Hodge operator may be
defined by extending, linearly, the map
∗(ωi1 ∧ · · · ∧ ωip ) = ωj1 ∧ · · · ∧ ωjn−p ,
where {i1 , . . . , ip , j1 , . . . jn−p } is an even permutation of {1, . . . , n}.
The following properties are easily established
1.32. Proposition. ∗ is a linear isometry and ∗n−p ◦ ∗p = (−1)p(n−p) 11Λp (E) .
2. Differential forms and the de Rham cohomology
2.1. Definition. A differential p-form on an open set U ⊆ Rn is a smooth map ω : U −→ Λp (Rn ) ∼

=
( )
n
R . When clear from the context we will just say that ω is a differential form or simply a form.
p
2.2. Remark. According to Remark 3.2 of Chapter 0, we can complicate the definition in order to have
one that make sense in the context of smooth manifold. Consider the bundle of exterior p-forms
Λp (U ) := ∪x∈U Λp (Tx U )
that can be identified with U × Λp (Rn ). Then a differential p-form is a smooth map ω̃ : U −→ Λp (U ) such
that ω̃(x) ∈ Λp (Tx U ), i.e, ω̃(x) = (x, ω(x)), ω(x) ∈ Λp (Rn ).
2. DIFFERENTIAL FORMS AND THE DE RHAM COHOMOLOGY 36
We will denote by Ωp (U ) the set of differential p-forms on U . Ωp (U ) has an obvious structure of real
vector space. Moreover we can multiply a differential form by a function and this operation is associative
and distributive, in the appropriate sense, i.e. Ωp (U ) is a module over F(U ).
A differential form ω ∈ Ωp (U ) induces a F(U )-multilinear map, denoted by the same symbol,
ω : H(U ) × · · · × H(U ) −→ F(U ), ω(X1 , . . . , Xp )(x) = ω(x)(X1 (x), . . . , Xp (x)).
Conversely, we have
2.3. Theorem. [Tensoriality Criterion] A R-multilinear map
ω : H(U ) × · · · × H(U ) −→ F(U ),
is induced by a differential form if and only if it is F(U )-multilinear.
Proof. Clearly, if ω is induced by a form, it is F(U )-multilinear. Suppose that ω is F(U )-multilinear.
P
Let x ∈ U, Xi ∈ Tx U . Extend the Xi ’s to vector fields X̃i ∈ H(U ), X̃i (y) =j aij (y)ej , and define:
ω(x)(X1 , . . . , Xp ) := ω(X̃1 , . . . , X̃p )(x).
In order to show that the above equality defines a form it is sufficient to show that it does not depend on
the extensions. In fact, by F(U )-multilinearity,
n
X
ω(X̃1 , . . . , X̃p )(x) = a1i1 (x) · · · apip (x)ω(ei1 , . . . , eip ).
i1 ,...,ip =1
2.4. Example. Since Λ0 (Rn ) = R, Ω0 (U ) = F(U ).
The basic example of a differential form is the following. Let f ∈ F(U ). Then the differential of f is
the the 1-form
(df )(x)(X) := X(x)(f ), X ∈ Der(U ).
In particular, we can consider the coordinate functions xi : Rn −→ R. At each point x ∈ U , the
differentials at x, dxi (x) 5 are a basis of Λ1 (Rn ). Therefore {dxi1 (x) ∧ · · · ∧ dxip (x) : 1 ≤ ii < · · · < ip ≤ n}
is a basis of Λp (Rn ). So we have
2.5. Proposition. Let ω ∈ Ωp (U ). Then ω can be written in a unique way as:

X
ω= ωi1 ,...,ip dxi1 ∧ · · · ∧ dxip ,
i1 <···<ip
where ωi1 ,...,ip ∈ F(U ).

n
X ∂f
2.6. Example. If f ∈ F(U ), df = dxi .
1
∂xi
2.7. Remark. As a real vector

µ ¶space, Ωp (U ) is infinite dimensional (if n > 0 !), but as a F(U )-module,
n
it is a free module of dimension .
p
5Since x is linear, dx = x , and dx is the form that associates to a vector its ith coordinate in the canonical basis.
i i i i
Let U ⊆ Rn , V ⊆ Rm be open sets and F : U −→ V a smooth function, F (x) = (F1 (x), . . . , Fm (x)).
Then dF (x) : Rn −→ Rm is a linear map and we have an induced map F ∗ : Λp (Rm ) −→ Λp (Rn ). This map
induces a linear map:
F ∗ : Ωp (V ) −→ Ωp (U ), F ∗ (ω)(X1 , . . . , Xp )(x) := ω(dF (x)(X1 ), . . . , dF (x)(Xp )).

If x1 , . . . , xn , y1 , . . . ym are the canonical coordinates in Rn , Rm respectively, we have
n
X ∂Fi
(1) F ∗ (dyi ) = dxj ,
i=1
∂xj
X
and therefore, if ω = ωi1 ,...,ip dyi1 ∧ · · · ∧ dyip ,
i1 ,...,ip
X
F ∗ (ω)(x) = ωi1 ,...,ip (F (x))F ∗ (dyi1 ) ∧ . . . ∧ F ∗ (dyi1 ).
i1 ,...,ip
We have the functorial properties:

• 11∗U = 11Ωp (U ) ,
• If F1 : U1 −→ U2 e F2 : U2 −→ U3 are smooth maps, (F2 ◦ F1 )∗ = F1∗ ◦ F2∗ .
In particular, if F is a diffeomorphism, F ∗ is an isomorphism.
2.8. Example. Let U ⊆ Rn and j : U −→ U × Rm , j(x1 . . . , xn ) = (x1 . . . , xn , 0 . . . , 0), be the in-

clusion. If ω = f (x1 , . . . , xn+m )dxi1 ∧ · · · ∧ dxip , i1 < · · · < ip , j ∗ ω = 0, if ip > n, and j ∗ ω =
f (x1 , . . . xn , 0, . . . , 0)dxi1 ∧ · · · ∧ dxip is ip ≤ n.
Differentiating a function can be viewed as a R-linear map:
d : Ω0 (U ) = F(U ) −→ Ω1 (U ).
Now we extend extend now this operation to higher dimensional forms.
2.9. Theorem. There exists a unique family of R linear operators dp : Ωp (U ) −→ Ωp+1 (U ), p =

0, . . . , n, such that:
(1) d0 = d (the usual differential).
(2) dp+1 ◦ dp = 0.
(3) If ω ∈ Ωp (U ), τ ∈ Ωq (U ), dp+q ω ∧ τ = dp ω ∧ τ + (−1)p ω ∧ dq τ.
Moreover, if F : U −→ V is a smooth map and ω ∈ Ωp (V ), then dp F ∗ ω = F ∗ dp ω.
When clear from the context we will write simply d for dp .
Proof. Let us suppose that such a family exists. If ω = f (x) dxi1 ∧ · · · ∧ dxip , we have:
dω = (df ) ∧ dxi1 ∧ · · · ∧ dxip + f d(dxi1 ∧ · · · ∧ dxip ).

Xn
∂f
Now, from (1), df = dxi , and, from (2) and (3)
i=1
∂xi
X
d(dxi1 ∧ · · · ∧ dxip ) = ±dxi1 ∧ · · · ∧ ddxij ∧ · · · ∧ dxip = 0.
j
2. DIFFERENTIAL FORMS AND THE DE RHAM COHOMOLOGY 38
X
Therefore, if ω = ωi1 ...ip dxi1 ∧ · · · ∧ dxip ,
i1 <···<ip
X X ∂ωi1 ...ip
dω = dxk ∧ dxi1 ∧ · · · ∧ dxip .
∂xk
k i1 <···<ip
This shows that if such a family exists, it is unique. Conversely, if we define dp by the formula above we
obtain a family of operators that, as it is easily seen, has the desired properties.
The last claim follows from
X ∂Fi
F ∗ (dyi ) = dxj = d(yi ◦ F ) = d(F ∗ (yi ))
j
∂xj
and the fact that F ∗ is an algebra homomorphism. ¤
The operator d is called the de Rham differential or exterior differential or simply the differential.
A simple but useful consequence of the properties above is the following
2.10. Corollary. d is a local operator, i.e. if ω ≡ τ in an open set U , then dω = dτ in U .
Proof. The proof is essentially the same as the proof of the first claim in Lemma 3.18 of Chapter 0. ¤
We can also give an alternative definition of the exterior differential that does not depend on coordinates.
2.11. Proposition. Let ω ∈ Ωp (U ), X0 , . . . , Xp ∈ H(U ). Then

p
X X
dω(X0 , . . . , Xp ) = (−1)i Xi · ω(X0 , . . . , X̂i , . . . Xp ) + (−i)i+j ω([Xi , Xj ], X0 , . . . , X̂i , . . . , X̂j , . . . , Xp ).
i=0 i<j
Proof. We sketch the proof leaving the details to the reader (Exercise 7.25). First observe that the right
hand side of the equality above is F(U )-multilinear, and so, by the tensoriality criterium, it is a differential
form. In particular, to compute dω(X0 , . . . , Xp ) at a given point x0 ∈ U , we can take arbitrarily extensions
of the Xi (x0 ). So will be enough to prove the equality for the case the Xi ’s are coordinate vector fields.
In this case [Xi , Xj ] = 0 so the second term on the right hand side vanishes while the firs term is just the
expression of dω given in Theorem 2.9. ¤
We have a sequence of vector spaces and R-linear maps:
d0 d1 dn−1
0 −→ Ω0 (U ) −→ Ω1 (U ) −→ · · · −→ Ωn (U ) −→ 0
which is a cochain complex, i.e. dp+1 ◦ dp = 0, or, equivalently, Im dp−1 ⊆ ker dp (see next section for the
definition and basic properties of cochain complexes). This sequence is called the de Rham complex of U .
We define
• Z p (U ) := ker dp , the space of p-cocycles or closed p-forms.
• B p (U ) := Im dp−1 , the space p-coboundaries or exact p-forms.
• H p (U ) := Z p (U )/B p (U ), the p-dimensional (de Rham) cohomology of U .
2.12. Remark. Let ω, τ be closed forms in U . Then d(ω ∧ τ ) = 0, i.e. ω ∧ τ is closed. Moreover if
τ = dβ, ω ∧ τ = ±d(ω ∧ β), i.e. ω ∧ τ is exact. In particular the wedge product induces a well defined
bilinear map ∪ : H p (U ) ⊕ H q (U ) −→ H p+q (U ), [ω] ∪ [τ ] = [ω ∧ τ ]. This product, suitably extended, define
an algebra structure on H ∗ (U ) := ⊕H p (U ). With this structure H ∗ (U ) is called the cohomology algebra of
U 6.
Let U ⊆ Rn , V ⊆ Rm be open sets and F : U −→ V a smooth function. As we already observed, F

induces a map F ∗ : Ωp (V ) −→ Ωp (U ). Since, by Theorem 2.9, F ∗ ◦ d = d ◦ F ∗ , F ∗ maps closed forms to
closed forms and exact forms to exact forms. Therefore it induces a R-linear map, that we still denote by
F ∗,
F ∗ : H p (V ) −→ H p (U ).
It is also simple to see that F ∗ induces an algebra homomorphism F ∗ : H ∗ (V ) −→ H ∗ (U ).
The basic functorial properties are easily verified
• 11∗U = 11H p (U ) ,
• If F1 : U1 −→ U2 and F2 : U2 −→ U3 are smooth maps, then (F2 ◦ F1 )∗ = F1∗ ◦ F2∗ .
In particular, if F is a diffeomorphism, F ∗ is an isomorphism. So the de Rham cohomology is a
(differential) topological invariant of U .
3. Algebraic aspects of cohomology
The construction of the de Rham cohomology fits into a general algebraic setting called homological
algebra. In this section we will discuss some elementary facts that will be used in these notes. For simplicity
we will restrict to the case of real vector spaces (not necessarily finite dimensional) although most of the
matter could be extended to the case of modules over commutative rings (see Remarks 3.10 and 3.22 ).
The objects we study are sequences of (real) vector spaces and linear maps of the type
E := {(Ep , dp ) : dp : Ep −→ Ep+1 }.
When we introduce “objects” it is a good strategy to introduce “morphisms” between such objects, i.e.
maps that preserves the structure of the objects.
3.1. Definition. A morphism φ : E −→ F, between two sequences is a sequence of linear maps φp :

E −→ Fp such that the diagrams
p
dp
· · · −→ Ep −→ Ep+1 −→ · · ·
p p+1
↓φ ↓φ
p dp
· · · −→ F −→ Fp+1 −→ · · ·
commute, i.e. dp ◦φp = φp+1 ◦dp (we are using the same symbols dp for the linear maps in the two sequences).
The morphism is an isomorphism if all φp are vector spaces isomorphisms.
We have some special sequences.

6This product is usually called the cup product. The use of this terminology, instead of the more natural wedge product, is
due to the fact that the cup product can be defined for different cohomology theories, where the wedge product is not defined.
3. ALGEBRAIC ASPECTS OF COHOMOLOGY 40
3.2. Definition. A sequence E = {Ep , dp } is exact at Ep if Im dp−1 = ker dp . The sequence is an exact
sequence if it is exact at each Ep .
3.3. Examples.
φ
(1) A sequence of the type {0} −→ E −→ F is exact at E if and only if φ is injective.
φ
(2) A sequence of the type E −→ F −→ {0} is exact at F if and only if φ is surjective.
φ
(3) A sequence of the type {0} −→ E −→ F −→ {0} is exact if and only if φ is an isomorphism.
3.4. Definition. A sequence of the type:
{0} −→ E −→ F −→ G −→ {0}
is called a short sequence.
3.5. Remark. Short (exact) sequences are important since they are the “building blocks” of (long exact)
sequences. Let
φi−1 φi
· · · −→ Ei−1 −→ Ei −→ Ei+1 −→ · · ·
be a sequence. Consider the short sequence
φ̃i−1 φ̃i
{0} −→ Ei−1 / ker φi−1 −→ Ei −→ Im(φi ) −→ {0}
where φ̃i−1 , φ̃i are the induced maps. Since Im(φ̃i−1 ) = Im(φi−1 ), ker(φ̃i ) = ker(φi ), the (long) sequence is
exact at Ei if and only if the short sequence is exact.
3.6. Proposition. A short exact sequence

φ ψ
{0} −→ E −→ F −→ G −→ {0}
is isomorphic to the sequence

i π
{0} −→ E −→ E ⊕ G −→ G −→ {0},
where i(v) = (v, 0) and π(v, w) = w.
Proof. Let G̃ be a complement7 of Im φ = ker ψ, i.e F = φ(E) ⊕ G̃. The map ψ |G̃ : G̃ −→ G is an
isomorphism. Therefore the map k : F −→ E ⊕ G, k(v + w) = (φ−1 (v), ψ(w)) (v ∈ φ(E), w ∈ G̃) is the
required isomorphism. ¤
The following result appears often in the applications
3.7. Lemma. [The five Lemma] Consider the diagram:
f1 f2 f3 f4
E1 −→ E2 −→ E3 −→ E4 −→ E5
↓ φ1 ↓ φ2 ↓ φ3 ↓ φ4 ↓ φ5
g1 g2 g3 g4
F1 −→ F2 −→ F3 −→ F4 −→ F5
If the squares commute, the lines are exact and the φi ’s are isomorphisms for i = 1, 2, 4, 5 then φ3 is an
isomorphism.
7Recall that a complement of a subspace is obtained by starting from a basis {e } of the subspace and completing it to a
α
basis of the ambient space with elements {fβ } and then considering the subspace spanned by the {fβ }.
Proof. Suppose φ3 (e3 ) = 0. Then φ4 (f3 (e3 )) = g3 (φ3 (e3 )) = 0. Therefore f3 (e3 ) = 0 and, by the
exactness of the first line, e3 = f2 (e2 ). Now g2 (φ2 (e(2)) = φ3 (e3 ) = 0, and therefore φ2 (e2 ) = g1 (µ1 ), for
some µ1 ∈ F1 , by the exactness of the second line. Since φ1 is surjective, there exists e1 ∈ E1 such that
φ1 (e1 ) = µ1 . Finally
0 = f2 (f1 (e1 )) = f2 (φ−1
2 g1 φ1 (e1 )) = f2 (e2 ) = e3
and therefore φ3 is injective. We will show now that φ3 is surjective. Let µ3 ∈ F3 , µ4 = g3 (µ3 ) and
e4 ∈ φ−1
4 (µ4 ). Now φ5 (f4 (e4 )) = g4 (µ4 ) = 0 and therefore f4 (e4 ) = 0, since φ5 is injective. In particular
there exists e3 ∈ E3 such that f3 (e3 ) = e4 . Let µ3 = φ3 (e3 ) and ω = µ3 − µ3 . Now g3 (ω) = 0 and
therefore ω = g2 (µ2 ). Let e2 = φ−1
2 (µ2 ). We have φ3 (f2 (e2 )) = g2 (φ2 (e2 )) = ω = φ(e3 ) − µ3 and therefore
µ3 = φ3 (e3 − f2 (e2 )) ∈ Im φ3 .
¤
3.8. Remark. We observe that in the proof of Theorem 3.7 we use only that φ2 , φ4 are isomorphisms,
φ1 is surjective and φ5 is injective. However the lemma is used, generally, as it is stated.
A more general and very important class of sequences is the class of cochain complexes.
3.9. Definition. A sequence E = {Ep , dp } is semiexact or a cochain complex if Im dp−1 ⊆ ker dp , ∀p.
Equivalently, it is a cochain complex if dp ◦ dp−1 = 0.
If E is a cochain complex we define:

• Z p (E) := ker dp , the group of p-dimensional cocycles,
• B p (E) := Im dp−1 , the group of p-dimensional coboundaries,
• H p (E) := Z p (E)/B p (E), the p-dimensional cohomology group.
3.10. Remark. Naturally Z p (E), B p (E), H p (E) are vector spaces. The use of the term “group” is due
to the fact that they can be defined in the more general context of complexes of Abelian groups, or modules
over a commutative ring.
The cohomology gives a measure of how much the complex is not an exact sequence.
dp
3.11. Example. The de Rham complex · · · −→ Ωp (U ) −→ Ω(p+1) (U ) −→ · · · is a cochains complex
whose cohomology is the de Rham cohomology H p (U ).
Consider now a morphism between two cochain complexes, φ : E −→ F. The commutativity condition
implies that cocycles are sent to cocycles and coboundaries to coboudaries. In particular φ induces linear
maps
φ∗,p : H p (E) −→ H p (F).
When clear from the context we will write simply φ∗ or φp .

The following “functorial” properties are easily verified:
• 11∗ = 11,
• (φ ◦ ψ)∗ = φ∗ ◦ ψ ∗ .
In particular if φ is an isomorphism, φ∗ is also an isomorphism.

It is convenient to consider also sequences with “decreasing indexes”, i.e. a sequence of the type
E := {(Ep , ∂p ) : ∂p : Ep −→ Ep−1 }.
If such a sequence is semiexact, we will call it a chain complex. For such a chain complex we define:
• Zp (E) := ker ∂p , the group of p-dimensional cycles.
• Bp (E) := Im ∂p+1 , the group of p-dimensional boundaries.
• Hp (E) := Zp (E)/Bp (E), the p-dimensional homology group.
As in the case of cochains, a morphism φ : E −→ F , between two chain complexes, sends cycles to cycles
and boundaries to boundaries, so it induces a sequence of maps φ∗,p : Hp (E) −→ Hp (F) and the functorial
properties are easily verified. When clear from the context we will write simply φ∗ or φp .
3.12. Remark. Naturally chain and cochain complexes are, essentially, the same objects. For example,
changing the index p in −p we pass from a chain complex to a cochain complex. But a more interesting
approach is duality and we will discuss this now.
Let E := {(Ep , ∂p ) : ∂p : Ep −→ Ep−1 } be a chain complex. We define the dual complex E ∗ = {(Ep , dp )}
where Ep := (Ep )∗ is the dual space, and dp = (∂p )∗ is the transpose of ∂p . It is simple to show that
dp ◦ dp−1 = 0 so E ∗ is, in fact, a cochain complex. We will denote with Hp (resp. H p ) the homology of E
(resp. the cohomology of E ∗ ). Consider the bi-linear map
b : Ep × Ep −→ R, b(φ, c) = φ(c).
It is easily seen that this map induces a bi-linear map
b̃ : H p × Hp −→ R, b̃([φ], [c]) = φ(c),
and therefore a linear map

K : H p −→ [Hp ]∗ , K([φ])([c]) = φ(c).
3.13. Theorem. [Universal coefficients Theorem] The map K is an isomorphism.
Proof. We start observing that we have two short exact sequences
∂p
(2) {0} −→ Zp −→ Ep −→ Bp−1 −→ {0}, {0} −→ Bp−1 −→ Zp−1 −→ Hp−1 −→ {0}
where the maps are the obvious ones. By Proposition 3.6, we have the decompositions
(3) Ep ∼
= Zp ⊕ Bp−1 , Zp−1 ∼
= Bp−1 ⊕ Hp−1
Claim 1.: K is surjective. Let [φ] ∈ [Hp ]∗ . Consider the map φ ◦ π : Zp −→ R, where π : Zp −→ Hp is
the quotient map. Using the first decomposition in (3), we can extend this map to a map φ̃ : Ep −→ R with
φ̃ = 0 on Bp−1 . Let e ∈ Ep . Then dφ̃(e) = φ̃(∂(e)) = 0, hence φ̃ is a cocycle and K([φ̃]) = [φ].
Claim 2.: K is injective. Let ψ ∈ Z p be such that ψ(c) = 0 ∀ c ∈ Zp . The map φ = ψ◦∂ −1 : Bp−1 −→ R
is well defined since, by the first sequence in (2), the difference of two elements in ∂ −1 (Bp−1 ) is a cycle.
Using the decompositions in (3), we can extend φ to a map φ̃ : Ep−1 −→ R. Now, ∀e ∈ Ep , we have:
dφ̃(e) = φ̃(∂e) = ψ ◦ ∂ −1 (∂e) = ψ(e).
Hence [ψ] = [dφ̃] = 0. ¤
A useful consequence is the following
3.14. Corollary. If a sequence is exact, the dual sequence is also exact.
Proof. An exact sequence is a chain complex with vanishing homology. Therefore the dual sequence
8
is a cochain complex with vanishing cohomology, by Theorem 3.13, hence an exact sequence ¤
We will study now when two morphism between cochain (resp. chain) complexes induces the same map
in cohomology (resp. homology).
3.15. Definition. An algebraic homotopy between two morphisms φ, ψ : E −→ F of cochain (resp.

chain) complexes is a family of maps K p : Ep −→ Fp−1 (resp. Kp : Ep −→ Fp+1 ), such that:
φ−ψ =d◦K +K ◦d (resp. φ − ψ = ∂ ◦ K + K ◦ ∂).
If there exists such an algebraic homotopy, we will say the the two morphisms are (algebraically) homotopic.
From the very definition of induced morphisms we have:
3.16. Proposition. Two algebraically homotopic maps induce the same morphism in cohomology (resp.
in homology).
Consider now a short exact sequence of cochain complexes:
φ ψ
{0} −→ E −→ F −→ G −→ {0}.
In particular φi is injective and ψi is surjective. In general, at the cohomology level, φ∗ is not injective and
ψ ∗ is not surjective. In any case, we still have a good relation between the cohomology groups of the three
complexes.
3.17. Theorem. [Algebraic Mayer-Vietoris Theorem] In the situation above there exists a family of
linear maps ∆∗,p : H p (G) −→ H p+1 (E) such that the sequence:
φ∗ ψ∗ ∆∗,p
· · · −→ H p (E) −→ H p (F) −→ H p (G) −→ H p+1 (E) −→ · · ·
is a (long) exact sequence. When clear from the context we will write dimply ∆p or ∆∗ .
8We could also give a direct proof, and the reader is invited to do so (Exercise 7.17).
Proof. We have the commutative diagram
0 0 0
² dp
² dp+1
²
··· / Ep / Ep+1 / Ep+2 / ···
φp φp+1 φp+2
² dp
² dp+1
²
··· / Fp / Fp+1 / Fp+2 / ···
ψp ψp+1 ψp+2
² dp
² dp+1
²
··· / Gp / Gp+1 / Gp+2 / ···
² ² ²
0 0 0
where the columns are exact and the rows are the cochain complexes under consideration. The idea is to
construct a map from Gp to Ep+1 . A natural choice would be (φp+1 )−1 ◦ dp ◦ ψp−1 . The point is that this map
is not well defined. Let us see how we can overcome this problem. Consider a cocycle c ∈ Gp . Since ψp is
surjective, there exists b ∈ Fi such that c = ψp (b). The element dp (b) ∈ Fp+1 is in ker ψp+1 since the diagrams
commute and c is a cocycle. Since ker ψp+1 = Im φp+1 we have dp (b) = φp+1 (a) for some a ∈ Ep+1 and this
a is unique since φp+1 is injective. Observe that dp+1 (a) = 0, since φp+2 (dp+1 (a)) = dp+1 (φp+1 (a)) = dp+1 ◦
dp (b) = 0 and φi+2 is injective. Therefore a is a cocycle. We define: ∆∗ : H p (G) −→ H p+1 (E), ∆∗ ([c]) = [a].
We have to show that [a] is well defined. The first choice we made was b ∈ Fp . If b0 is an other choice, i.e.
ψ p (b0 ) = ψ p (b), then b − b0 ∈ ker ψp = Im φp . Therefore b0 − b = φp (a0 ), for some a0 ∈ Ep , and b0 = b + φp (a0 ).
So, changing b by b + φp (a0 ), we change a by a + dp (a0 ) and this does not change [a]. Next we shall show
that [a] does not depend on the choice of c ∈ [c]. Consider c + dp (c0 ). Since c0 = ψp−1 (b̃), for some b̃ ∈ Fp−1 ,
we have c + dp−1 (c0 ) = c + dp−1 (ψp−1 (b̃)) = c + ψp (dp−1 (b̃)) = ψp (b + dp−1 (b̃)). Therefore b is substituted by
b + dp−1 (b̃), and this does not change dp (b) and, therefore, [a].
It is easy to see that ∆∗ is linear. We leave to the reader the task of proving the exactness of the
sequence (Exercise 7.22). ¤
3.18. Remark. The map ∆∗ is well defined in cohomology but not at the cocycles level.
3.19. Definition. The sequence in Theorem 3.17 is called the (algebraic) Mayer-Vietoris sequence. The
maps ∆∗ , are the Mayer-Vietoris coboundaries.
3.20. Remark. Naturally we have a similar sequence in homology, associated to a short exact sequence
of chain complexes. The similar maps ∆∗,p or simply ∆∗ , are called the Mayer-Vietoris boundaries. We
leave the details to the reader.
An important aspect of the Mayer-Vietoris (co)boundaries is that they are “natural” in the sense of the
following Proposition, whose proof we leave to the reader (Exercise 7.22).
3.21. Proposition. A morphism between short exact sequences of (co)chain complexes induces a mor-
phism between the associated Mayer-Vietoris exact sequences, i.e. the Mayer-Vietoris (co)boundaries com-
mutes with the induced maps.
3.22. Remark. As suggested in Remark 3.10, instead of chain and cochain complexes of vector spaces
we could consider chain and cochain complexes of Abelian groups (or modules over a commutative ring).
Almost all we have done in this section extends to the case of complexes of Abelian groups. The “almost”
refers to two exceptions:
• Proposition 3.6 does not hold in this more general setting. For example the sequence of abelian groups
·2
{0} −→ Z −→ Z −→ Z2 −→ {0}, ·2(a) := 2a,
is a short exact sequence, but it is not isomorphic to the sequence
{0} −→ Z −→ Z ⊕ Z2 −→ Z2 −→ {0}.
A short exact sequence of Abelian groups that verify Proposition 3.6 is called a split short exact sequence.
A sufficient condition for splitting is given by the following simple fact
3.23. Proposition. A short exact sequence of Abelian groups

φ ψ
{0} −→ A −→ B −→ C −→ {0}
splits if and only if there is a map r : C −→ B such that ψ ◦ r = 11C . This always happens if C is free 9.
• We can consider “duality” in the context Abelian groups. If G is such a group, G∗ := Hom(G, Z) is
the group of homomorphisms from G to Z. Therefore we can define the dual of a chain complex of Abelian
groups. However Theorem 3.13 does not hold in this context. In fact, one of the points in the proof was
that the sequence of vector spaces
{0} −→ Bp−1 −→ Zp−1 −→ Hp−1 −→ {0}
splits. As observed above, this is not the case, in general, for short exact sequences of Abelian groups.
However, if Hp−1 is a free Abelian group, then the sequence splits, by Proposition 3.23, and the Theorem
holds true. In the general case there is still a relation between the homology of a chain complex of Abelian
groups and the cohomology of the dual complex, still known as the Universal Coefficients Theorem.
4. Basic properties of the de Rham cohomology
The natural problem that the de Rham cohomology attacks is the problem of (indefinite) integration,
i.e. the problem of solving the equation dω = β, for a given β ∈ Ωp+1 (U ). A necessary condition for the
existence of a solution ω is dβ = 0. In general the problem has two aspects:
• The local problem: given x ∈ U, β ∈ Ωp+1 (U ) do there exist a neighborhood V ⊆ U of x and a solution
ω ∈ Ωp (V ) of the equation dω = β|V ? In this case, as we shall see, the condition dβ = 0 is also sufficient.
• The global problem: given β ∈ Ωp+1 (U ), does there exist a solution ω ∈ Ωp (U ) of the equation dω = β?
In this case, the condition dβ = 0 is no longer sufficient, in general, and the answer will depend on the
particular β and/or the topology of U .
We will start computing the de Rham cohomology in some simple cases.
9A free Abelian group G is an Abelian group that admits a basis, i.e. a subset B ⊆ G such that for any Abelian group H
and map φ : B −→ H, there exists a homomorphism φ̃ : G −→ H, extending φ.

4. BASIC PROPERTIES OF THE DE RHAM COHOMOLOGY 46
4.1. Example. For U = R0 we have:

(
p 0 R if p = 0
H (R ) '
{0} if p > 0
` Q
4.2. Example. Let U = α Uα be the union of disjoint open sets Uα . Then Ωp (U ) = α Ωp (Uα ) (direct
product) and the differential preserves the decomposition, i.e. if ω = {ωα }, dω = {dωα }. It follows that
Y
H p (U ) ∼
= H p (Uα ).
α
4.3. Example. Let us analyze the 0-dimensional cohomology. In this case, the only exact 0-form is the
zero form so H 0 (U ) is the space of closed 0-forms, i.e. functions in F(U ) with zero differential. Such a
function is locally constant, in particular it is constant on the connected components of U . It follows that
H 0 (U ) is the direct product of copies of R, as many as the connected components of U .
Let us give a further look at the 0-dimensional cohomology. Let U ⊆ Rn , V ⊆ Rm be open connected
sets, and F : U −→ V a smooth map. As we observe in 4.3, the zero dimensional cohomology of U is the
space of constant functions, and the same for V . Given a 0-form f ∈ Ω0 (V ) = F(V ), F ∗ (f ) = f ◦ F and
therefore F ∗ : H 0 (V ) −→ H 0 (U ) is an isomorphism. Modulo the identification of the zero dimensional
cohomology groups with R, we have F ∗ = 11 : R −→ R.
We want to look now at the induced maps in higher dimensional cohomology groups. The question is
the following: when do two smooth maps Fi : U −→ V, i = 0, 1 induce the same morphism in cohomology?
We will give a sufficient condition in terms of homotopy.
4.4. Definition. Let U ⊆ Rn , V ⊆ Rm be open sets and Fi : U −→ V, i = 0, 1 be smooth functions.

• A homotopy between the two functions is a smooth map10
H : U × [0, 1] ⊆ Rn+1 −→ V,
such that H(x, i) = Fi (x), i = 0, 1.

• We will say that the two functions are homotopic if there exist a homotopy between them. In this
case we write F0 ∼ F1 .
• We will say that U and V are homotopy equivalent if there exist functions F : U −→ V, G : V −→ U ,
such that G ◦ F ∼ 11U , F ◦ G ∼ 11V . F (resp. G) is called a homotopy inverse of G (resp. of F )11.
• We will say that U is contractible if U is homotopy equivalent to R0 .
4.5. Example. A subset U ⊆ Rn is star shaped if there exists p ∈ U such that, for all q ∈ U , the
segment joining p and q is contained in U . For example convex sets are star shaped. Star shaped subsets
are contractible since the map H(q, t) := tp + (1 − t)q is a homotopy between 11U and the constant map
F (q) = p. It follows that 11 and F are homotopy inverses.
4.6. Remark. Given a homotopy H : U × [0, 1] −→ V , there is a smooth function H : U × R −→ V ,

such that H(x, i) = Fi (x), i = 0, 1. In fact, if λ : R −→ [0, 1] is a smooth function such that λ(t) = 0 if t ≤
0, λ(t) = 1 if t ≥ 1, just take H(x, t) = H(x, λ(t)).
10A map f : V ⊆ RN −→ RM , defined in a non necessarily open subset V ⊆ RN is smooth, if for all p ∈ V, f extends to
a smooth map defined in an open neighborhood of p.
11Observe that a homotopy inverse is not, in general, unique.
A homotopy between two functions may be viewed as a curve in the space of smooth maps joining the
two functions. Also it may be viewed as a “smooth deformation” of one function to the other.
4.7. Theorem. [Homotopy invariance for cohomology] If Fi : U −→ V, i = 0, 1 are two homotopic

smooth functions, then F0∗ = F1∗ : H p (V ) −→ H p (U ), for all p.
Proof. By Remark 4.6 we can suppose that there is a homotopy H : U × R −→ V . Let ji : U −→

U × R, i = 0, 1, ji (x) = (x, i), be the canonical inclusions. We claim that it is sufficient to prove that
j0∗ = j1∗ . In fact, if so, we have:
F0∗ = (H ◦ j0 )∗ = j0∗ ◦ H ∗ = j1∗ ◦ H ∗ = (H ◦ j1 )∗ = F1∗ .
To prove that j0∗ = j1∗ we will construct an algebraic homotopy between j0∗ and j1∗ (at the cochain level,
see Definition 3.15 and Proposition 3.16), i.e. an R-linear map H̃ : Ωp (U × R) −→ Ωp−1 (U ) such that
H̃dω + dH̃ω = j1∗ ω − j0∗ ω.
Let us construct such a map. If ω ∈ Ωp (U × R), ω = dt ∧ α + β, with

X X
α= αi1 ,...,ip−1 (x, t)dxi1 ∧ · · · ∧ dxip−1 , β = βj1 ,...,jp (x, t)dxj1 ∧ · · · ∧ dxjp .
i1 <...<ip−1 j1 <···<jp
We define µZ ¶
X 1
H̃(ω) = αi1 ,...,ip−1 (x, t)dt dxi1 ∧ · · · ∧ dxip−1 .
i1 <...<ip−1 0
Then
X ∂αi1 ...ip−1
dω = −dt ∧ dα + dβ = −dt ∧ dxj ∧ dxi1 ∧ · · · ∧ dxip−1 +
j,i1 <···<ip
∂xj
X ∂βj1 ,...,jp
+dt ∧ dxj1 ∧ · · · ∧ dxjp + γ
j1 <···<jp
∂t
where γ does not contain terms with dt. Therefore

X µ Z 1 ∂βj1 ,...,jp ¶
H̃dω = dt dxj1 ∧ · · · ∧ dxjp −
j1 <···<jp 0 ∂t
X µ Z 1 ∂αi1 ...ip−1 ¶
dt dxj ∧ dxi1 ∧ · · · ∧ dxip−1 ,
j,i <···<i 0 ∂xj
1 p
X µZ 1 ¶
∂αi1 ...ip−1
dH̃ω = dt dxj ∧ dxi1 ∧ · · · ∧ dxip−1 ,
j,i1 <···<ip 0 ∂xj
and (see Example 2.8)
X µZ 1 ¶
∂βj1 ,...,jp
H̃dω + dH̃ω = dt dxj1 ∧ · · · ∧ dxjp =
j1 <···<jp 0 ∂t
X
= [βj1 ,...,jp (x, 1) − βj1 ,...,jp (x, 0)]dxj1 ∧ · · · ∧ dxjp = j1∗ ω − j0∗ ω.
j1 <···<jp
From 4.7, and the funtorial properties, we have

4. BASIC PROPERTIES OF THE DE RHAM COHOMOLOGY 48
4.8. Corollary. If U ⊆ Rn , V ⊆ Rm are homotopically equivalent open sets, then they have isomorphic
cohomology.
In particular we have the so called Poincaré Lemma
4.9. Corollary. [Poincaré Lemma] If U is a star shaped open set in Rn , H p (U ) = {0} if p ≥ 1.
4.10. Remark. Theorem 4.7 allows to define the map induced in cohomology by a continuous map.
In fact, as we shall see in the Appendix, a continuous map F : U −→ V is homotopic, via a continuous
homotopy H : U × [0, 1] −→ V , to a smooth map F̃ : U −→ V and if there is a continuous homotopy
between two smooth maps, there is a smooth one. So F ∗ := F̃ ∗ is well defined and invariant by continuous
homotopies.
A basic method to compute the cohomology of an open set U ⊆ Rn is to write U as union of two,
possibly simpler open sets U1 , U2 , and look for relations between the cohomology of U, Ui and V := U1 ∩ U2 .
4.11. Lemma. Consider the sequence :

(j ∗ ,j ∗ ) (k∗ −k∗ )
{0} −→ Ωp (U ) −→
1 2
Ωp (U1 ) ⊕ Ωp (U2 ) −→
1 2
Ωp (V ) −→ {0},
where ji : Ui −→ U and ki : V −→ Ui are the inclusions. Then the sequence is a short exact sequence of
cochain complexes.
Proof. Observe that ji∗ ω = ω|Ui and, if (ω1 , ω2 ) ∈ Ωp (U1 ) ⊕ Ωp (U2 ), (k1∗ − k2∗ )(ω1 , ω2 ) = ω1 |V − ω2 |V
(see Example 2.8). So the exactness of the sequence is obvious, except for the surjectivity of (k1∗ − k2∗ ). To
prove that (k1∗ − k2∗ ) is surjective we consider a partition of unity dominated by the covering {U1 , U2 }, i.e.
smooth functions φi : U −→ [0, 1], i = 1, 2 such that:
φ1 (x) + φ2 (x) = 1 ∀x ∈ U, supp(φi ) := {x ∈ U : φi (x) > 0} ⊆ Ui
(see Theorem 6.2 for a proof of the existence of partitions of unity).

Given ω ∈ Ωp (V ), we define:
(
φj (x)ω(x) if x ∈ V
ωi (x) =
0 if x ∈ Ui \ V
where i 6= j. Then ωi is well defined since φj vanishes outside Uj , j 6= i. Moreover,
(k1∗ − k2∗ )(ω1 , −ω2 ) = ω1 |V + ω2 |V = φ1 ω + φ2 ω = ω.
Therefore (k1∗ − k2∗ ) is surjective. ¤
At this point Theorem 3.17 gives:
4.12. Theorem. [Mayer Vietoris sequence for de Rham cohomology] There exists a sequence of linear
maps ∆∗ : H p (V ) −→ H p+1 (U ), such that the sequence below is exact:
(j ∗ ,j ∗ ) (k∗ −k∗ ) ∆∗
· · · −→ H p (U ) −→
1 2
H p (U1 ) ⊕ H p (U2 ) −→
1 2
H p (V ) −→ H i+1 (U ) −→ · · ·
4.13. Definition. The sequence above is called the Mayer-Vietoris sequence for the de Rham cohomology
and the maps ∆∗ are called the Mayer-Vietories coboundaries.
4.14. Remark. The Mayer-Vietoris coboundaries can be decribed explicitly. If [ω] ∈ H p (V ), ∆∗ [ω] is
the class of the form (
−d(φ2 ω)(x) if x ∈ U1
τ (x) =
d(φ1 ω)(x) if x ∈ U2
∗
Since d commutes with induced maps, so does ∆ . We invite the reader to check the details.
4.15. Example. Let us apply the Mayer-Vietoris sequence to compute the cohomology of Σn := Rn \{x =
(x1 , . . . , xn ) ∈ Rn : |xi | ≤ 1}.
Consider the open sets:
U1 = {(x1 , . . . , xn ) ∈ Σn : xn > −1/2}, U2 = {(x1 , . . . , xn ) ∈ Σn : xn < 1/2}.
The following facts are easy to prove:

• Σn = U1 ∪ U2 .
• Ui is contractible, i = 1, 2. In fact the projection (x1 , . . . , xn ) Ã (x1 , . . . , xn−1 , 2) is a homotopy
equivalence between U1 and the hyperplane xn = 2. Similarly for U2 .
• U1 ∩ U2 is homotopy equivalent to Σ(n−1) (the projection of U1 ∩ U2 into the hyperplane xn = 0 is
a homotopy equivalence).
If n = 1, Σ1 is the disjoint union of two contractible sets, hence by Corollary 4.8 and Example 4.3
(
p ∼ R ⊕ R if p = 0
H (Σ1 ) =
{0} if p > 0
For the case n ≥ 2 we will prove that
(
p R if p = 0, n − 1
H (Σn ) =
{0} if p 6= 0, n − 1
= H 0 (Ui ) ∼
We proceed by induction. Let n = 2. Since Σ2 and the Ui ’s are connected, H 0 (Σ2 ) ∼ = R.
Consider the Mayer-Vietoris sequence:
{0} −→ H 0 (Σ2 ) −→ H 0 (U1 ) ⊕ H 0 (U2 ) −→ H 0 (Σ1 ) −→ H 1 (Σ2 ) −→ H 1 (U1 ) ⊕ H 1 (U2 ) −→

−→ · · · −→ H p−1 (Σ1 ) −→ H p (Σ2 ) −→ H p (U1 ) ⊕ H p (U2 ) −→ · · · .
The first row reduces to:
{0} −→ R −→ R ⊕ R −→ R ⊕ R −→ H 1 (Σ2 ) −→ {0}.
The first arrow is injective hence the kernel of the second one, as well as its image, are 1-dimensional.
Hence the the third one is surjective with 1-dimensional kernel and H 1 (Σ2 ) ∼
= R.
From the second row we get H p (Σ2 ) = {0} if p > 1. Hence the formula holds true.
Suppose now n ≥ 3 and that the formula holds true for n − 1. Consider again the Mayer-Vietoris
sequence:
H p−1 (Σn ) −→ H p−1 (U1 ) ⊕ H p−1 (U2 ) −→ H p−1 (Σn−1 ) −→ H p (Σn ) −→ H p (U1 ) ⊕ H p (U2 ) −→
5. AN APPLICATION: THE JORDAN-ALEXANDER DUALITY THEOREM 50
If p > 1 we have H p (Σn ) ∼

= H p−1 (Σn−1 ), and, for p = 1 we get
{0} −→ R −→ R ⊕ R −→ R −→ H 1 (Σn ) −→ {0}.
Hence H 1 (Σn ) = {0} and the formula holds true for n.
4.16. Remark. Observe that the inclusion Σn −→ Rn \ {0} is a homotopy equivalence (Exercise 7.28).
5. An application: the Jordan-Alexander duality Theorem
It is convenient, as we shall see, in order to avoid special arguments for the 0-dimensional case and to
have more clean statements, to introduce reduced cohomology. Define
Ω−1 (U ) := R d(−1) : Ω−1 (U ) −→ Ω0 (U ), d(−1) (a) := a ∈ Ω0 (U ).
Then the sequence

d(−1) d
{0} −→ Ω−1 (U ) −→ Ω0 (U ) −→ Ω1 (U ) −→ · · ·
is a cochain compex called the augmented de Rham complex.
5.1. Definition. The reduced de Rham cohomology of U, H̃ p (U ), is the cohomology of the augmented
de Rham complex.
5.2. Remark. It is clear that H̃ −1 (U ) = {0}, H 0 (U ) ∼

= H̃ 0 (U ) ⊕ R and H̃ p (U ) = H p (U ), if p > 0. In
particular H̃ p (U ) = {0}, ∀ p ≥ 0, if U is contractible.
The basic properties, such as homotopy invariance and the Mayer-Vietoris exact sequence, continue to
hold true for the reduced cohomology and we will leave the proofs to the reader (see Exercise 7.24).
We will discuss now a nice application of the Mayer-Vietoris argument, the so called Jordan-Alexander
duality principle, that has, as a simple consequence, the celebrated Jordan closed curve Theorem. We will
follow closely [4].
Let Fi , i = 1, 2 be closed subsets of Rn . Suppose that there exists a homeomorphism φ : F1 −→ F2 .
It is natural to ask if there exists some relation between the complementary sets Rn \ Fi . The illusion
that they are homeomorphic or, at least, homotopy equivalent is soon frustrated. For example consider
F1 = {x ∈ R2 : kxk = 1} ∪ {x ∈ R2 : kxk = 2} and F2 = {x ∈ R2 : kxk = 1} ∪ {x ∈ R2 : kx − (3, 0)k = 1}.
The complement of F1 is homotopy equivalent to the disjoint union of a point and two circles, while the
complement of F2 is homotopy equivalent to the disjoint union of two points and the wedge 12 of two circles.
It is easily seen that these spaces are not homotopy equivalent.
5.3. Remark. (For the reader familiar with the concept of foundamental group,) The fact that the
complements of two homeomorphic closed set are not homotopy equivalent is important in several contexts,
for example in Knot Theory. Recall that a knot in R3 is a function γ : S 1 −→ R3 which is a homeomorphism
onto its image. Two knots are equivalent if there exists an isotopy, i.e. a homotopy through homeomorphisms,
which takes one into the other. One of the most important invariants for equivalence classes of knots is the
fundamental group of the complement of the image. Now, the images of two knots are homeomorphic and
12Recall that the wedge of two topological spaces is the space obtained from the disjoint union identifying a fixed point in
the first space with one in the second one.

if the complements were homotopy equivalent, they would have isomorphic fundamental group and so the
invariant would be trivial.
There is, however, an interesting relation between the complements of homeomorphic closed sets.
5.4. Theorem. [Jordan Alexander duality Theorem] Let Fi , i = 1, 2, be closed sets in Rn and φ : F1 −→
F2 an homeomorphism. Then:
H̃ k (Rn \ F1 ) ∼
= H̃ k (Rn \ F2 ).
Proof. We will consider Rn as the subspace of vectors in Rn+k with the last k coordinates zero. The
proof of the Theorem will be an easy consequence of the following two Lemmas.
5.5. Lemma. Let F ( Rn be a closed subset. Then H̃ i+1 (Rn+1 \ F ) ∼

= H̃ i (Rn \ F ), i ≥ −1.
Proof. Consider the subsets of Rn+1 :

• Z+ := Rn+1 \ F × {t ∈ R : t ≤ 0}.
• Z− := Rn+1 \ F × {t ∈ R : t ≥ 0}.
• Z := Z+ ∪ Z− = Rn+1 \ F.
• Z+ ∩ Z− ∼ Rn \ F.
The orthogonal projection of Z+ onto the hyperplane xn+1 = 1 is a homotopy equivalence. Hence the
reduced cohomology of Z+ vanishes in all dimensions. The same is true for Z− and the Lemma follows from
the Mayer-Vietoris sequence for the reduced cohomology:
H̃ i (Z+ ) ⊕ H̃ i (Z− ) = {0} −→ H̃ i (Z+ ∩ Z− ) −→ H̃ i+1 (Z) −→ H̃ i+1 (Z+ ) ⊕ H̃ i+1 (Z− ) = {0}.
5.6. Corollary. If F ⊆ Rn is a closed set, then H̃ i+k (Rn+k \ F ) ∼

= H̃ i (Rn \ F ), ∀ i ≥ −k.
5.7. Lemma. Let Fi ⊆ Rn , i = 1, 2 be closed subsets and φ : F1 −→ F2 an homeomorphism. Then

R2n \ F1 × {0} is homeomorphic to R2n \ {0} × F2 .
Proof. Let ψ = φ−1 . The homeomorphisms φ, ψ extend, by Tietze’s Theorem13, to continuous maps
Φ, Ψ : Rn −→ Rn . Define:
• L : R2n −→ R2n , L(x, y) = (x, y − Φ(x)).
• R : R2n −→ R2n , R(x, y) = (x − Ψ(y), y).
The maps L, R are homeomorphisms. In fact L−1 (x, y) = (x, y + Φ(x)), R−1 (x, y) = (x + Ψ(y), y). Consider
Γ := {(x, y) ∈ R2n : x ∈ F1 , y = φ(x)} = {(x, y) ∈ R2n : y ∈ F2 , x = ψ(y)}. We have L(F1 × {0}) = Γ =
R({0} × F2 ) and therefore a homeomorphism:
L R−1
R2n \ F1 × {0} −→ R2n \ Γ −→ R2n \ {0} × F2 .
¤
13Tietze’s Theorem states that a continuous real valued function defined in a closed subset of Rn extends to a continuous
function defined in the all of Rn (this fact is true, more generally, for normal topological spaces).
6. APPENDIX: PARTITIONS OF UNITY AND SMOOTH APPROXIMATIONS 52
The proof of the Theorem is, at this point, immediate:
H̃ i (Rn \ F1 ) ∼
= H̃ i+n (R2n \ F1 ) ∼
= H̃ i+n (R2n \ F2 ) ∼
= H̃ i (Rn \ F2 ).
As an immediate consequence of the Jordan-Alexander duality we have have the celebrated Jordan curve
Theorem.
5.8. Theorem. [Jordan curve Theorem] Let γ : S 1 −→ R2 be a homeomorphism onto its image14. Then
R2 \ γ(S 1 ) has exactly two connected components.
Proof. Consider the unit circle S 1 ⊆ R2 . It is clear that the complement of S 1 in R2 has exactly two
connected components and therefore H̃ 0 (R2 \ S 1 ) ∼
= R. By the duality Theorem 5.4, H̃ 0 (R2 \ γ(S 1 )) ∼
=R
and therefore the complement of γ(S 1 ) in R2 has also exactly two connected components. ¤
5.9. Remark. It is clear that the argument in the proof of Theorem 5.8 may be extended to the case
of a closed hypersurface M n ⊆ Rn+1 (see Chapter 3 for definitions) any time we have a “model”, i.e. a
close hypersurface homeomorphic to M n and information on the complement of the model. For example
this happens in the case of closed oriented surfaces in R3 or in the case of closed hypersurfaces of Rn+1 ,
homeomorphic to a sphere. A different approach will be discussed in Chapter 4 (Theorem 3.1).
6. Appendix: partitions of unity and smooth approximations
In order not to interrupt the flow of the arguments, we left, in the previous sections, a couple of “gaps”,
namely the proof of existence of partitions of unity (in the proof of Theorem 4.12) and the approximation
of continuous maps by smooth ones (see Remark 4.10). In this Appendix we will fill up these gaps.
Partitions of unity is a basic tool that allows you to glue together locally defined objects (such as
functions, forms etc.) in order to obtain a globally defined object. We start with the basic definition.
6.1. Definition. Let U ⊂ Rn be an open set and let {Vα } be an open covering of U . A partition of
unity dominated by the covering {Vα } is a family of smooth functions λi : Rn −→ [0, 1] such that:
(1) For all i there exist α such that supp(λi ) := {x ∈ Rn : λi (x) 6= 0} ⊆ Vα .
(2) For all x ∈ U there exist a neighborhood Ux of x such that Ux ∩ supp(λi ) = ∅ for all but finitely
many of the λi ’s.
P
(3) For x ∈ U, i λi (x) = 1 (observe that, by (2), the sum is finite).
Our aim is to prove the following result:
6.2. Theorem. Let U ⊂ Rn be an open set and let {Vα } be an open covering of U . Then there exist a
partition of unity dominated by {Vα }.
Proof. We will use the following notations:
B(p, r) = {x ∈ Rn : kx − pk < r}, D(p, r) = {x ∈ Rn : kx − pk ≤ r} = B(p, r) .
14Such a map is usually called a Jordan curve.

We recall (Exercise 4.13 of Chapter 0) that given δ1 , δ2 ∈ R, 0 < δ1 < δ2 ,and p ∈ Rn , there exists a smooth
function φ : Rn −→ [0, 1] such that φ(x) = 0 in B(p, δ1 ) and φ(x) = 1 in Rn \ B(p, δ2 ).
Claim 1. Let K ⊆ Rn be a compact set and V ⊆ Rn an open set with K ⊆ V . Then there exist a
smooth function ψ : Rn −→ [0, 1] such that ψ(x) = 1, if x ∈ K and ψ(x) = 0 if x 6∈ V .
Proof. For any p ∈ K consider δ(p) such that D(p, 2δ(p)) ⊆ V . Then there is a finite number of points,
S
p1 , . . . , pr ∈ K, such that K ⊆ D(pi , δ(pi )). For each i we have a function φi : Rn −→ [0, 1] such that
φi (x) = 0, x ∈ D(pi , δ(pi )) and φ(y) = 1, y 6∈ D(pi , 2δ(pi )). Then the function
ψ(x) = 1 − φ1 (x) · · · φr (x)
has the required properties. ¤
Claim 2. There exist a continuous proper function15 φ : U −→ [0, ∞).
Proof. Since Rn is homeomorphic to the open ball B(0, 1) (Exercise 4.14, Chapter 1) and the composi-
tion of a proper continuous function with a homeomorphism is still proper, we can assume that U ⊆ B(0, 1)
. For x ∈ U , define d(x) to be the distance of x to the boundary of U . Then d : U −→ R is a positive contin-
uous function. Consider φ : U −→ [0, ∞), φ(x) = d(x)−1 . Then φ is continuous and for all n ∈ N, φ−1 [0, n]
is a closed bounded set in U , hence compact. So φ is proper. ¤
We will now prove the Theorem. Consider a proper function φ : U −→ [0, ∞) and set
1 3
An = φ−1 [n, n + 1], Wn = φ−1 (n − , n + ).
2 2
Then An is compact and therefore may be covered with a finite number of balls Bk,n such that each disk
Dk.n := Bk,n is contained in some Vα ∩ Wn . For each such disk we have a smooth function φk,n : U −→ [0, 1]
vanishing outside Vα ∩ Wn and identically 1 in Dk,n . It is clear from the construction that the An ’s cover
U and so, for all x ∈ U , there is at least one of the φn.k ’s not vanishing at x. Also Wn ∩ Wn+2 = ∅ so the
P
supports of the φn,k are a locally finite covering and k,n φk,n (x) < ∞, ∀x ∈ U . So the family of functions
φn,k
λn,k = P
i,j φi,j
is a well defined partition of unity dominated by the covering Vα . ¤
6.3. Remark. Observe that the partition of unity we constructed is a countable set of smooth functions.
We shall prove now that a continuous function may be approximate by a smooth function, homotopic
to it. The proof is a good example of how to use partition of unity.
6.4. Theorem. Let U ⊆ Rn be an open set and let F : U −→ W ⊆ Rm be a continuous function which
is smooth on a closed subset N ⊆ U . Then, given a real valued positive continuous function δ : U −→ R
there exists a smooth function G : U −→ W such that kF (x) − G(x)k < δ(x), ∀x ∈ U and F (x) = G(x) if
x ∈ N . Moreover G ∼ F .
15A function is proper if the inverse image of a compact set is compact.

7. EXERCISES 54
Proof. We recall that F smooth on N means that for all x ∈ N there exists a neighborhood Vx of x
and a smooth extension hx of F |[Vx ∩ N ]. For x ∈ U we consider a neighborhood Vx of x and a function
hx : Vx −→ R with the following conditions:
(1) F (Vx ) is contained in a subset of an open ball contained in W .
δ(x)
(2) If x ∈ N, hx is a smooth extension of F |[Vx ∩ N ] and khx (y) − F (x)k < .
2
(3) If x 6∈ N, Vx ∩ N = ∅ and hx (y) = F (x), ∀y ∈ Vx .
δ(x)
(4) ∀y ∈ Vx , kF (y) − F (x)k < < δ(y).
2
Consider a smooth partition of unity, λi , dominated by the covering Vx . Then ∀ i there exists x = x(i)
with supp(λi ) ⊆ Vx(i) . For every i fix such a x(i) and set
X
G(z) = λi (z)hx(i) (z).
i
Then G is a smooth function since in a neighborhood of a point G is a finite sum of smooth functions.
Let z ∈ N and λi1 , . . . λik be the functions of the partition which do non vanish at z. Then hx(ij ) is an
extension of F , hence equal, in z, to F (z). Hence G(z) = F (z) and G is an extension of F |N .
Let y ∈ M \ N . If λi (y) 6= 0, y ∈ supp(λi ) ⊆ Vx(i) . Hence kF (y) − hx(i) k < δ(x(i))/2. Hence
X X X δ(x(i))
kF (y) − G(y)k = k λi (y)F (y) − λi hx(i) (y)k ≤ λi (y)kF (y) − hx(i) (y)k < < δ(y).
i i i
2
Finally H(x, t) = tF (x) + (1 − t)G(x) is the required homotopy.

¤
6.5. Corollary. If two smooth maps F, G : U −→ W are homotopic via a continuous homotopy, then
they are homotopic via a smooth one.
7. Exercises
7.1. Prove that the tensor product of tensors is associative and distributive.
7.2. Prove that ω ∈ Ep is an exterior form if and only if
ω(x1 , . . . , xi , . . . , xj , . . . xp ) = −ω(x1 , . . . , xj , . . . , xi , . . . xp ).
7.3. Prove that the exterior product is distributive with respect to the sum.
7.4. Complete the proof of Proposition 1.20.
7.5. Prove that φ1 , . . . , φp ∈ E∗ are linearly independent if and only if φ1 ∧ · · · ∧ φp 6= 0.
7.6. Prove that two sets of linearly independent elements of E∗ , {φ1 , . . . , φp } and {ψ1 , . . . , ψp } span the
same subspace of E∗ , if and only if φ1 ∧ · · · ∧ φp = d ψ1 ∧ · · · ∧ ψp , d ∈ R. In this case, d is the determinant
of the matrix that gives the change of basis for the subspace.
Pn
7.7. Let ω ∈ Λ∗ (E), ω = 0 ωi , ωi ∈ Λi (E). Prove that ω is invertible in Λ∗ (E)16 if and only if ω0 6= 0.
16i.e. there exists ω −1 ∈ Λ∗ (E) such that ω ∧ ω −1 = 1.

7.8. Let E be a n-dimensional vector space. Let π : E∗ × · · · × E∗ −→ Λp (E) be the p-linear extension
of (φ1 , . . . , φp ) −→ φ1 ∧ · · · ∧ φp . Prove that the following universal property of the exterior algebra holds:
• (UP∧) If K is a vector space and b : E∗ × · · · × E∗ −→ K is an alternated p-linear map, then there
exists a unique linear map l : Λp (E) −→ K such that l ◦ π = b.
7.9. Prove that the universal property (UP∧) characterizes Λp (E) i.e., given a vector space L and a
p-linear map π̃ : E∗ × · · · × E∗ −→ L such that (π̃, L) verifies UP∧, then L ∼
= Λp (E).
7.10. Prove that Λp (E∗ ) ∼

= [Λp (E)]∗ .
7.11. Let v ∈ Λn (E) \ {0}. Define a map:
bv : Λp (E) × Λ(n−p) (E) −→ R, bv (ω, τ )v := ω ∧ τ.
Prove that bv is non degenerate and hence defines an isomorphism b̃v : Λp (E) −→ [Λ(n−p) (E)]∗ .
P
7.12. Let φ1 , . . . , φr ∈ E∗ be linearly independent. Let ψ1 , . . . , ψr ∈ E∗ be such that i φi ∧ ψi = 0.
P
Prove that ψi = j aij φj with aij = aji .
7.13. A form ω ∈ Λp (E) is decomposable if ω = φ1 ∧ · · · ∧ φp , φi ∈ E∗ . By Proposition 1.20, any p-form

is a sum of decomposable forms.
(1) Show that, if dim(E) = n, any (n − 1)-form is decomposable.
(2) Show that, if dim(E) = 4 and {φ1 , . . . , φ4 } is a basis of E∗ , then φ1 ∧φ2 +φ3 ∧φ4 is not decomposable.
7.14. Let E be a n-dimensional vector space. A vector space G(E), with an associative product denoted
by ∧, is called a Grassman algebra for E if
(1) G(E) contains a subspace isomorphic to R ⊕ E and is generated, as an algebra, by this subspace,
(2) 1 ∧ x = x, x ∧ x = 0, ∀x ∈ E,
(3) dim(G(E)) = 2n .
Prove that G(E) is isomorphic, as an algebra, to Λ∗ (E∗ ).
7.15. Prove, using the functorial properties, that if L : E −→ F is an isomorphism, L∗ : Λ∗ (F) −→ Λ∗ (E)
is an isomorphism (see Remark 1.24).
7.16. Let φ ∈ E∗ \ {0} and ω ∈ Λp (E). Show that, if φ ∧ ω = 0, then there exists τ ∈ Λp−1 such that
ω = φ ∧ τ . Conclude that the sequence:
φ∧ φ∧
· · · −→ Λp−1 (E) −→ Λp (E) −→ Λp+1 (E) −→ · · ·
is exact. Hint: choose a basis containing φ.
7.17. Prove directly, i.e. without using Theorem 3.13, Proposition 3.14.
7.18. Let L be a finite dimensional real Lie algebra, i.e. a finite dimensional real vector space with a
bi-linear map [ , ] : L × L −→ L, (X, Y ) −→ [X, Y ] such that, ∀ X, Y, Z ∈ L we have:
(1) [X, Y ] = −[Y, X],
(2) [[X, Y ]Z] + [[Y, Z], X] + [[Z, X], Y ] = 0 (Jacobi identity).
7. EXERCISES 56
Define a map dp : Λp (L) −→ Λp+1 (L),

X
dp (ω)(X1 , . . . , Xp+1 ) = (−1)i+j ω([Xi , Xj ], X1 . . . , X̂i , . . . , X̂j , . . . , Xp+1 ).
i<j
Show, at least for p = 1, that dp+1 ◦ dp = 0.

In particular the sequence above is a cochain complex and its cohomology is called the cohomology of
the Lie algebra L.
7.19. Let L be a Lie algebra. ω ∈ Λp (L) is said to be Ad-invariant if, ∀ Y, X1 , . . . , Xp ∈ L, we have

X
(−1)i−1 ω([Y, Xi ], X1 , . . . , X̂i , . . . , Xp ) = 0.
i
(The terminology will be clear in the section on Lie groups).

(1) Show that if ω ∈ Λp (L) is Ad-invariant, then dp ω = 0.
(2) Show that span {[X, Y ] : X, Y ∈ L} = L if and only if the only Ad-invariant 1-form is zero.
(3) Show that if the only Ad-invariant 1-form is the zero form, the only Ad-invariant 2-form is zero.
Remark: Under suitable hypothesis the cohomology of the Lie algebra is isomorphic to the space of Ad-
invariant forms.
7.20. Let E = {0} −→ En −→ · · · −→ E0 −→ {0} be a chain complex. Assume that the Ei ’s are finite
dimensional and let Hi be the homology groups of the complex. Prove that
n
X n
X
χ(E) := (−1)i dim(Ei ) = (−1)i dim(Hi ).
0 0
χ(E) is called the Euler characteristic of the complex.
7.21. Let E, F be chain complexes as in Exercise 7.20, and let φ : E −→ F be a morphism. Prove that:
X X
λ(φ) := (−1)i trace(φi ) = (−1)i trace(φ∗,i ).
λ(φ) is called the Leftchetz number of φ (this number is of great importance in fixed point theory).
7.22. Show that the (algebraic) Mayer-Vietoris sequence (Theorem 3.17) is exact and the (co)boundaries
are natural (Proposition 3.21).
7.23. Prove that the Mayer-Vietoris cobundary, for the de Rham cohomology, are given by
7.24. Show that the Mayer-Vietoris sequence for the reduced cohomology (see Definition 5.2) is exact.
7.25. Give details of the proof of Proposition 2.11.
7.26. Use Example 4.15 and Remark 4.16 to prove the Theorem of invariance of dimension:
Theorem: If h : Rn −→ Rm is a homeomorphism, then n = m.
7.27. Redo the computations in Example 4.15, using reduced cohomology.
7.28. Prove the Claim in Remark 4.16

7.29. Let U ⊆ Rn be an open set and v = dx1 ∧ · · · ∧ dxn be the volume form. We will identify vectors
fields and 1-forms via the “musical isomorphisms” [ : H(U ) −→ Ω1 (U ) and its inverse ] : Ω1 (U ) −→ H(U ).
Also ∗ will denote the Hodge operator. We define the classical differential operators of calculus:
X ∂f ∂
• The gradient ∇ : F(U ) −→ H(U ), ∇f := ]df = .
µX ∂xi¶∂xi
∂ X ∂Xi
• The divergence div : H(U ) −→ F (U ), div Xi = .
∂xi ∂xi
• The (geometers) Laplacian ∆ : F(U ) −→ F (U ), ∆f = −div ∇f .
• The rotational rot : Ω1 (U ) −→ Ωn−2 (U ) rot ω = ∗dω.
.
Prove that:
n
X ∂2f
(1) ∆f = −d ∗ (df ) = − .
∂x2i
1
(2) ∆(f g) = g∆f + f ∆g − 2h∇f, ∇gi.
(3) ω is closed if and only if rot ω = 0.
(4) rot ∇f = 0.
X ∂
(5) If n = 3 compute rot Xi and show that div rot ω = 0.
∂xi
7.30. Let U ⊆ Rn be an open set. Show that H n (U ) = {0} if and only if ∀ f ∈ F(U ) there exists a
vector field X ∈ H(U ) such that div X = f .
Remark: It can be shown that the Laplacian ∆ : F(U ) −→ F(U ) is surjective (this is a non trivial
fact). In particular the equation div X = f has a solution ∀ f ∈ F(U ). Hence H n (U ) = {0}.
√
7.31. Identify R2 with the complex line C, (x, y) −→ x + iy, i = −1. If U ⊆ R2 is an open set and
f : U −→ C, we will write f (z) := f (x, y) = u(x, y) + iv(x, y), u, v ∈ F(U ). f is said to be holomorphic if it
is C 1 and
∂u ∂v ∂v ∂u
= and =− (Cauchy-Riemann equations).
∂x ∂y ∂x ∂y
It can be shown that a holomorphic function is smooth, and, even more than that, complex analytic, i.e. it
is locally the sum of its (complex) Taylor series.
(1) Show that the Cauchy-Riemann equations just say that the differential df (z) : R2 −→ R2 is C-linear
√
(i.e. commutes with multiplication by i = −1).
(2) Define complex 1-forms:
dz := dx + idy, f dz := (u + iv)dz := (udx − vdy) + i(udy + vdx).
and the complex derivative f 0 (z) by the identity f 0 (z)dz = df . Prove that f is holomorphic if and
∂u ∂u
only if the real and imaginary parts of f dz are closed. In this case f 0 (z) = −i .
∂x ∂y
(3) Prove that if f = u + iv is holomorphic, then u, v : U −→ R are harmonic functions (i.e. ∆u =
∆v = 0).
(4) Show that, if U is star shaped, given a harmonic function u : U −→ R, there exists a harmonic
function v : U −→ R such that f (x, y) = u(x, y) + iv(x, y) is holomorphic. The function v is unique,
up to an additive constant (if U is connected), and is called the harmonic conjugate of u.
7. EXERCISES 58
7.32. Let E be a real vector space and J : E −→ E a linear map such that J 2 = −11. Prove that the
dimension of E is even an J induces a structure of complex vector space on E.
7.33. Let E be a real vector space with an inner product, dim E = 2n and let J : E −→ E be a linear
isometry such that J 2 = −11.
(1) Prove that there exist orthonormal vectors {e1 , . . . , en } ⊆ E such that the set {ei , J(ei ), i = 1, . . . n}
is an orthonormal basis for E.
(2) Prove that ω(x, y) := hx, J(y)i is an exterior form.
P
(3) Let φi = [ei , ψi = [J(ei ). Prove that ω = − φi ∧ ψi .
(4) Prove that ω n = (−1)n n! ∗ 1
CHAPTER 2
Integration and the singular homology of open sets of Rn
In Remark 1.8 of Chapter 1, we observed that p-forms are “p-dimensional (oriented) volume elements”
and hence the natural integrands for the (oriented) multiple integrals. In this Chapter we will make this
statement precise, we will introduce the singular homology of open sets in Rn and see how integration gives
a duality between singular homology and the de Rham cohomology.
1. Integration on singular chains and Stokes Theorem
1.1. Definition. Let U ⊆ Rn be an open set and ω = f (x)dx1 ∧ · · · ∧ dxn ∈ Ωn (U ). Let D ⊆ U be the
closure of an open bounded set. We define
Z Z
ω= f (x1 , . . . , xn )dx1 · · · dxn ,
D D
where the integral on the right hand side is the usual Riemann integral.
1.2. Remark. The integral defined above is “oriented” in the sense that if ωσ = f (x)dxσ(1) ∧ · · · ∧
dxσ(n) , σ ∈ Σ(n), then
Z Z
ω = |σ| ωσ .
D D
In particular the integral depends on the ordering of the coordinates, i.e., it depends on the choice of an
orientation in Rn , while the usual Riemann integral of a function does not depend on such a choice (see also
Exercise 6.5).
In order to define the integral of a p-form, we first define the “domain of integration”.
1.3. Definition.
• A p-simplex in Rn is the convex hull1 of (p + 1) points {v0 , . . . , vp } ⊂ Rn in general position2. The
points vi are called the vertices of the simplex. Any subset of q + 1 (distinct) vertices determines
a q-simplex called a face of the original one.
• Let {e1 , . . . , ep } be the canonical basis of Rp and e0 = 0. The standard p-simplex, ∆p ⊂ Rp is the
simplex with vertices {e0 , e1 , . . . , ep }.
• A differentiable singular p-simplex in U is a smooth map σ : ∆p −→ U (i.e. σ extends to a smooth
map of an open neighborhood of ∆p ) . If it is clear from the context we shall omit the term
differentiable.
1We recall that the convex hull of a subset of Rn is the smallest convex set that contains the given set. More precisely, it
is the intersection of all convex sets that contain the given set.
2The points {v , . . . , v } are in general position if they are not contained in any affine subspace of dimension less than p.
0 p
This is equivalent to the fact that the vectors {vi − v0 : i = 1, . . . p} are linearly independent.
59
1. INTEGRATION ON SINGULAR CHAINS AND STOKES THEOREM 60
1.4. Remark. Given a p-simplex with vertices {vo , . . . , vp }, a point in the simplex can be written in
Pp Pp
a unique way in the form v = i=0 λi vi with λi ∈ [0, 1] ⊂ R and i=0 λi = 1. The numbers λi are the
barycentric coordinates of v.
1.5. Example. An important example of a singular simplex is the following: Let {v0 , . . . , vp } be points
of Rn , not necessarily in general position. Define L(v0 , . . . , vp ) to be the singular simplex of Rn that maps
Pp
the point of ∆p with barycentric coordinates {λ0 , . . . , λp } to the point i=0 λi vi ∈ Rn . This simplex will
be called the linear simplex with vertices {v0 , . . . , vp }.
1.6. Definition. Let ω ∈ Ωp (U ) be a differential p-form and σ : ∆p → U a singular p-simplex. We

define Z Z
ω := σ ∗ ω,
σ ∆p
where the integral on the right hand side is in the sense of Definition 1.1.
1.7. Example. If f ∈ F(U ) is a smooth function, i.e. a 0-form, and p ∈ U a fixed point, i.e. a 0-simplex,
then the integral of the form on the simplex is just f (p).
P
1.8. Example. If ω = ωi dxi ∈ Ω1 (U ) is a 1-form and σ : ∆1 −→ U a smooth 1-simplex, then
n
X
σ ∗ ω = ω̃(t)dt, with ω̃(t) = σ ∗ ω(t)(1) = ω(σ(t))(dσ(t)(1)) = ω(σ(t))(σ̇(t)) = ωi (σ(t))σ̇i (t),
i=1
where σi (t) = hσ(t), ei i is the ith coordinate of σ. Hence

Z Z 1 "Xn
#
ω= ωi (σ(t))σ̇i (t) dt.
σ 0 i=1
The fundamental result in the elementary integration theory is Stokes Theorem. It relates the integral
of a p-form on a domain to the integral of a primitive on the boundary. For p = 1 Stokes Theorem is just
the foundamental Theorem of calculus
Z b Z
df (t)dt = f = f (b) − f (a) (see Example 1.7).
a ∂[a,b]
We will define now the ingredients necessary to state this Theorem in higher dimensions. We start by
introducing more general domains of integration for a p-form.
1.9. Definition. A singular p-chain is a (formal) finite linear combination of singular p-simplices, with
real coefficients. The set Cp (U ) of all such p-chains is a real vector space, with the obvious operations.
P
If ω ∈ Ωp (U ), c ∈ Cp (U ), c = ai σi , we define the integral of ω on c by:
Z X Z
I(c, ω) := ω := ai ω.
c σi
Next we have to define the boundary of a p chain. Intuitively, the boundary of a singular simplex will
be the restriction of the simplex to the boundary of the standard p-simplex ∆p (which is a chain and not a
simplex). More precisely
CHAPTER 2. INTEGRATION AND THE SINGULAR HOMOLOGY OF OPEN SETS OF RN 61
1.10. Definition. The boundary operator ∂p : Cp (U ) −→ Cp−1 (U ) is defined as the linear extension of
p
X
∂p σ := (−1)i σ ◦ Fi ,
0
p−1
where σ is a singular p-simplex and Fi : ∆ −→ ∆p is the linear simplex Fi = L(e0 , . . . , eî , . . . , ep ).
1.11. Remark. The signs in the definition above guarantee that the (p − 1) faces of ∆p are taken with
the induced orientations.
1.12. Example. For a linear simplex, we have the formula:

p
X
∂p L(v0 , . . . , vp ) = (−1)i L(v0 , . . . , vî , . . . , vp ).
i=0
In our context we have the following version of the classical Stokes Theorem:
1.13. Theorem. [Stokes Theorem] If c ∈ Cp+1 (U ), ω ∈ Ωp (U ), then

Z Z
I(∂c, ω) := ω = dω := I(c, dω).
∂c c
Proof. By linearity, it is sufficient to prove the Theorem when c is a singular simplex σ : ∆p+1 −→ U .
In this case Z Z Z
dω = σ ∗ dω = dσ ∗ ω
σ ∆p+1 ∆p+1
(see Theorem 2.9 of Chapter 1 for the last equality). Also
Z Z
ω= σ ∗ ω,
∂σ ∂∆p+1
p+1
Pp+1
where ∂∆ is the linear chain i=0 (−1) L(e0 , . . . , eî , . . . ep+1 ) ∈ Cp (∆p+1 ).
i
P ˆ i · · · ∧ dxp+1 . Again by linearity, it is sufficient to prove
Now η := σ ∗ ω = i fi (x1 , . . . , xp+1 )dx1 ∧ · · · dx
the Theorem for each monomial. Since we can permute coordinates, up to sign, it is not restrictive to assume
η = f (x1 , . . . , xp+1 )dx1 ∧ · · · ∧ dxp .
Then:
∂f
dη = (−1)p dx1 ∧ · · · ∧ dxp+1 .
∂xp+1
Hence, by Fubini’s Theorem
Z Z Z "Z 1−Ppi xi #
p ∂f p ∂f
dη = (−1) dx1 · · · dxp+1 = (−1) dxp+1 dx1 · · · dxp =
∆p+1 ∆p+1 ∂xp+1 ∆p 0 ∂xp+1
Z " p
X
#
p
= (−1) f (x1 , . . . , xp , 1 − xi ) − f (x1 , . . . , xp , 0) dx1 · · · dxp ,
∆p i=1
where ∆ is the standard simplex {e0 , . . . ep } ⊆ Rp ⊆ Rp+1 .
p
Now ∂∆p+1 = L(e1 , . . . ep+1 ) + (−1)p+1 L(e0 , . . . , ep ) + γ where γ is a chain of linear simplices that are
faces of ∆p+1 containing both e0 and ep+1 . Since on each of such faces at least one of the first p coordinates
vanishes, η = 0 on γ. Hence:
Z Z Z
p+1
η= η + (−1) η=
∂∆p+1 L(e1 ,...ep+1 ) L(e0 ,...ep )
2. SINGULAR HOMOLOGY 62
Z p
X Z Z
p p+1
= (−1) f (x1 , . . . , xp , 1 − xi )dx1 · · · dxp + (−1) f (x1 , . . . , xp , 0)dx1 · · · dxp = dη.
∆p i=1 ∆p ∆p+1
2. Singular homology
We will now look a little more deeply at the boundary operator.
2.1. Lemma. ∂(p−1) ◦ ∂p = 0.
Proof. Let σ be a singular simplex. From Example 1.12 we have

X
∂p (σ) = (−1)i σ ◦ L(e0 , . . . , eî , . . . , ep ).
i
Therefore:
p
X X
∂(p−1) ∂p (σ) = (−1)i (−1)j σ ◦ L(e0 , . . . , eˆj , . . . , eî , . . . , ep )+
i=0 j<i
p
X X
+ (−1)i (−1)(j−1) σ ◦ L(e0 , . . . , eî , . . . , eˆj , . . . , ep ).
i=0 j>i
Note that the term σ ◦ L(e0 , . . . , eî , . . . , eˆj , . . . , ep ), i, j fixed, appears twice in the above sum with opposite
signs, and therefore ∂(p−1) ∂p (σ) = 0.
¤
In particular the sequence:

∂(p+1) ∂p ∂(p−1)
· · · −→ C(p+1) (U ) −→ Cp (U ) −→ C(p−1) (U ) −→ · · · ,
is a chain complex and we define:
• Zp (U ) := ker ∂p the group of p-dimensional cycles.

• Bp (U ) := Im ∂(p+1) the group of p-dimensional boundaries.
• Hp (U ) := Zp (U )/Bp (U ) the pth dimensional (singular smooth) homology group.
From Stokes Theorem 1.13 we get:
2.2. Theorem. If a ∈ Zp (U ), I(a, dω) = 0. If σ ∈ Z p (U ), I(∂b, σ) = 0. Therefore the operator

I : Cp (U ) × Ωp (U ) → R induces a R-bilinear operator:
I˜ : Hp (U ) × H p (U ) −→ R, ˜
I([c], [ω]) := I(c, ω).
˜
2.3. Remark. The classical Theorem of de Rham states that the linear map induced by I,
Z
dRU : H p (U ) −→ [Hp (U )]∗ , dRU ([ω])([c]) = ω,
c
is an isomorphism, called de de Rham isomorphism. We will prove this Theorem in the next section.
Let F : U ⊆ Rn → V ⊆ Rm be a smooth map. Then F induces a linear map F∗ : Cp (U ) −→ Cp (V ),

obtained by extending by linearity the map which sends a singular simplex σ : ∆p −→ U to the singular
simplex F ◦ σ : ∆p → V . It is easy to check that F∗ commutes with the boundary operator and hence it is
a morphism between chain complexes. Therefore it induces a morphism in homology, that we will denote
with the same symbol,
F∗ : Hp (U ) −→ Hp (V ).
The following functorial properties are easily established3

• (11U )∗ = 11Hp (U ) ,
• (G ◦ F ) = G∗ ◦ F∗ .
An important feature of the de Rham map is that it is natural with respect to smooth maps.
2.4. Proposition. Let F : U −→ V be a smooth map. Then
[F∗ ]∗ (dRU (ω)) = dRV (F ∗ ω).
Proof. Let σ ∈ Cp (U ), ω ∈ Ωp (V ). Then

Z Z
ω= F ∗ω
F ◦σ σ
(essentially by definition), and the conclusion follows. ¤
Now we will look at some examples that are the analogs, for homology, of Examples 4.1 4.2, and 4.3 of
Chapter 1.
2.5. Example. Let U = R0 . Then there is a unique singular p-simplex, the constant one. His boundary
is the alternated sum of (p + 1) elements, all equal to the (unique) (p − 1)-simplex. Therefore the boundary
operator is null if p is odd and it is the identity if p is even. The complex of singular chains is given by:
0 11 0 0
−→ C(2p+1) (U ) = R −→ C2p (U ) = R −→ C(2p−1) (U ) = R −→ · · · −→ C0 (U ) = R −→ {0}.
Therefore:
(
0 R if p = 0
Hp (R ) '
{0} if p > 0
2.6. Remark. It might appear more natural and, in fact, some times more convenient, to define chains
and homology using singular cubes, i.e., smooth maps of the unit cube [0, 1]p ⊆ Rp into U . Since a p-cube
has always an even number of (p − 1)-faces, this construction gives, for U = R0 , a chain complex with
p-dimensional chain group R and null boundary operators. So the homology would be isomorphic to R in
all dimensions, which is not what we would like to have. However if we take the quotient of the complex
of singular cubes by a suitable subcomplex, we obtain a new complex whose homology is the same as the
homology of the complex of singular simplices.
3This means that the homology is a covariant functor from the category of open sets of Rn and smooth maps into the
category of (graded) vector spaces and linear maps.

2. SINGULAR HOMOLOGY 64
`
2.7. Example. Let U = Uα be the disjoint union of the open sets Uα . Since ∆p is connected, the
α
L
image of a singular simplex is contained in some Uα . Therefore Cp (U ) = α Cp (Uα ) (direct sum) and the
boundary maps preserve the decomposition, i.e. if c = {cα }, ∂c = {∂cα }. It follows that
M
Hp (U ) ∼= Hp (Uα ).
α
2.8. Remark. We observe explicitly that we are dealing with finite linear combinations of simplices,
hence we have a direct sum instead of a direct product, as in the case of cohomology. Furthermore, this is
in agreement with the de Rham Theorem 2.3, since the dual of the direct sum of vector spaces is the direct
product of the duals.
2.9. Example. Let us analyze the 0-dimensional homology. Let us suppose first that U is connected.
A 0-simplex is a constant map, i.e. a point in U . Such a simplex is a cycle, by definition. On the other
hand, given two points in U they can be joined by a smooth curve, i.e. a 1-simplex. The boundary of such
simplex is the difference of the two points, so the two points are in the same homology class. It follows
that H0 (U ) ∼
= R. Also, as in the case of cohomology, if U ⊆ Rn , V ⊆ Rm are connected open sets and
F : U −→ V is a smooth map, the induced map F∗ : H0 (U ) −→ H0 (V ) is an isomorphism.
If U is not connected, let us say with connected components Uα , it follows from Example 2.7 that
M
H0 (U ) ∼
= R.
α
Next we will prove the homotopy invariance for homology:
2.10. Theorem. Let F, G : U → V be homotopic smooth maps. Then F∗ = G∗ .
Proof. Let H : U × [0, 1] → V be a homotopy between F and G. We will construct an algebraic

homotopy between the induced maps, i.e. a map H̃p : Cp (U ) −→ C(p+1) (V ) such that
∂ ◦ H̃(σ) = G∗ (σ) − F∗ (σ) − H̃ ◦ ∂σ.

The Theorem then follows since if c ∈ Zp (U ), G∗ (c) − F∗ (c) ∈ Bp (V ), i.e. [G∗ (c)] = [F∗ (c)] in Hp (V ).
Consider the product ∆p × [0, 1] ⊂ Rp+1 . If σ is a singular p-simplex of U , we consider the map
H ◦ (σ × 11) : ∆p × [0, 1] −→ V . The problem is that ∆p × [0, 1] is not a simplex4. The strategy will be to
subdivide ∆p × [0, 1] into simplices and to take a suitable alternated sums of the restrictions of H ◦ (σ × 11)
to such simplices.
Consider vi = (ei , 0), wi = (ei , 1), and the linear (p+1)-simplices L(v0 , . . . , vi , wi , . . . wp ). If σ : ∆p −→ U
is a singular p-simplex, we define
p
X
H̃(σ) = (−1)i H ◦ (σ × 11) ◦ L(v0 , . . . , vi , wi , . . . , wp ) ∈ Cp+1 (V ).
i=0
Extending the formula by linearity we get a morphism H̃ : Cp (U ) −→ Cp+1 (V ).

We observe that, geometrically, the left hand side of the first equation is the restriction of σ × 11 to the
boundary of the prism ∆p × [0, 1] while the right hand side is, with appropriate signs, the restriction of σ × 11
4This is a case in which would be more convenient to work with singular cubes instead that simplices since the product of
two cubes is a cube (Remark 2.6).

to the bases of the prism, ∆ × {0, 1}, essentially G∗ (σ) − F∗ (σ), plus the restriction of σ × 11 to the “lateral
faces” ∂∆p × [0, 1]. We will make this precise.
Using 1.12 and the functorial properties, we get:
XX
∂ H̃(σ) = (−1)i (−1)j H ◦ (σ × 11) ◦ L(v0 , . . . , vˆj , . . . , vi , wi , . . . wp )+
i j≤i
XX
+ (−1)i (−1)j+1 H ◦ (σ × 11) ◦ L(v0 , . . . , vi , wi , . . . , ŵj , . . . , wp ).
i j≥i
For i = j the terms on the right hand side cancel except for
H ◦ (σ × 11) ◦ L(vˆ0 , w0 , . . . , wp ) = G ◦ σ and − H ◦ (σ × 11) ◦ L(v0 , . . . , vp , wˆp ) = −F ◦ σ.
The rest of the sum is the opposite of
XX
(−1)i−1 (−1)j H ◦ (σ × 11) ◦ L(v0 , . . . , vˆj , . . . , vi , wi , . . . wp )+
i j<i
XX
+ (−1)i (−1)j H ◦ (σ × 11) ◦ L(v0 , . . . , vi , wi , . . . , ŵj , . . . , wp ) = H̃∂(σ).
i j>i
H̃ is then the required homotopy. ¤
From Theorem 2.10 and the funtorial properties we have
2.11. Corollary. If F : U −→ V is a homotopy equivalence, then F∗ : Hp (U ) → Hp (V ) is an

isomorphism. In particular, a contractible space has the same homology as R0 .
2.12. Remark. As in the case of cohomology, the homotopy invariance allows us to define the map
induced in homology by a continuous map (see Remark 4.10 in Chapter 1).
We also have a Mayer-Vietoris exact sequence for homology. Let Ui ⊆ Rn , i = 1, 2 be open sets and
define U = U1 ∪ U2 , V = U1 ∩ U2 . Consider the sequence of chain complexes
((j1 )∗ ,(j2 )∗ ) ((k1 )∗ −(k2 )∗ )

{0} −→ Cp (V ) −→ Cp (U1 ) ⊕ Cp (U2 ) −→ Cp (U ) −→ {0},
where ji : V → Ui , ki : Ui → U are the inclusions and the boundary maps are the obvious ones.
We would like to proceed like in the case of cohomology. The problem we have here is that the sequence
above is not exact. More precisely, ((k1 )∗ − (k2 )∗ ) is not surjective, since a chain in U might not be the sum
of chains in Ui . To overcome this problem, we consider the chain complex Cp (U1 + U2 ) ⊆ Cp (U ) spanned by
the singular simplices of U1 and U2 . Substituting Cp (U ) with this complex, we have a short exact sequence
of chain complexes. The point that makes this idea work is the following result
2.13. Theorem. [Small simplicies Theorem] The inclusion Cp (U1 + U2 ) −→ Cp (U ) induces an isomor-
phism in homology.
The proof requires some new constructions and we will give it in the Appendix in order not to interrupt
the flow of our discussion.
Using Theorem 2.13 and Theorem 3.17 of Chapter 1, we deduce, as for cohomology
3. THE DE RHAM THEOREM FOR OPEN SETS OF RN 66
2.14. Theorem. There are linear maps ∆∗,p : Hp (U ) −→ H(p−1) (V ) such that the sequence
((j1 )∗ ,(j2 )∗ ) ((k1 )∗ −(k2 )∗ ) ∆∗,p
· · · −→ Hp (V ) −→ Hp (U1 ) ⊕ Hp (U2 ) −→ Hp (U ) −→ H(p−1) (V ) −→ · · ·
is a (long) exact sequence. Again, we often write ∆∗ or ∆p for ∆∗,p .
2.15. Definition. The exact sequence above is called the Mayer-Vietoris sequence for singular homology
and the maps ∆∗ , the Mayer-Vietoris boundary operators.
3. The de Rham Theorem for open sets of Rn
Let U ⊆ Rn be an open set. As we have seen, integration induces a linear map:

Z
p ∗
dR : H (U ) −→ [Hp (U )] , dR([ω])([c]) = ω.
c
We have already announced that this map is an isomorphism and the aim of this section is to prove this
fact. We will start with a Lemma, known as the Mayer Vietoris argument, useful in several situations.
3.1. Lemma. [Mayer Vietoris argument]5 Let U ⊆ Rn be an open set and P a statement about the open
subsets V ⊆ U . Suppose that:
(1) P is true for open convex sets,
(2) If P is true for disjoint sets, then it is true for their union,
(3) If P is true for two sets and for their intersection, then it is true for their union.
Then P is true for U .
Proof. First we observe that P is true for the union of n convex sets. In fact, for n = 2 this follows
from (3) observing that the intersection of two convex sets is convex. Suppose that P is true for the union
of (n − 1) convex sets. Let V1 , . . . , Vn be convex sets and V = V1 ∪ . . . ∪ V(n−1) . Then P is true for Vn and,
by the inductive hypothesis, for V . But it is also true for V ∩ Vn since
V ∩ Vn = (V1 ∩ Vn ) ∪ . . . ∪ (V(n−1) ∩ Vn )
is the union of (n − 1) convex sets. From (3), P is true for the union of all the Vi ’s.
Let φ : U −→ [0, ∞) be a proper function (see Claim 2. in the proof of Theorem 6.2, Chapter 1). Define:
An = φ−1 ([n, n + 1]).
Since φ is proper, An is compact and we can cover it with a finite number of open convex sets, Uk,n , contained
¡ ¢
in φ−1 (n − 21 , n + 32 ) . Let Un = ∪k Uk,n . Now P is true for Un , since it is a finite union of convex sets.
Let us consider Ueven = ∪n U2n and Uodd = ∪n U2n+1 . Then, by (2), P is true for Ueven and Uodd since
each one is a disjoint union of sets for which P is true. Finally Ueven ∩ Uodd = ∪n,k,h Uk,2n ∩ Uh,2n+1 and
therefore it is a disjoint union of sets that are finite unions of convex sets. Therefore, by (3), P is true for
U = Ueven ∪ Uodd . ¤
We can now prove the de Rham Theorem.
3.2. Theorem. The map dR : H p (U ) −→ [Hp (U )]∗ is an isomorphism.
5The lemma is also called the onion lemma and the reason for this will be clear from the proof (see [2]).
Proof. Since we will work with several open sets, it is convenient to denote with dRV the de Rham
map relative to the open set V ⊆ U ⊆ Rn . We are going to use Lemma 3.1. Let us consider the statement
P(V ) := dRV : H p (V ) −→ [Hp (V )]∗ is an isomorphism.
Clearly the statement is true for convex sets. In fact they are contractible and we have to check the
statement in dimension 0, which is trivial. Also, if it is true for a family of disjoint open sets, it is also true
for their union (recall that the dual of the direct sum is the direct product).
Let us suppose that P is true for the open sets V, W and for V ∩ W . Consider the diagram:
··· / H p (V ∩ W ) / H p+1 (V ∪ W ) / H p+1 (V ) ⊕ H p+1 (W ) / ···
dRV ∩W dRV ∪W dRV ⊕dRW

² ² ²
··· / [Hp (V ∩ W ))]∗ / [Hp+1 (V ∪ W )]∗ / [Hp+1 (W )]∗ ⊕ [Hp+1 (W )]∗ / ···
where the upper row is the Mayer-Vietoris sequence for cohomology and the lower row is the dual of the
Mayer-Vietoris sequence in homology. The latter is exact by Proposition 3.14 of Chapter 1. Since integration
commutes with induced maps (Proposition 2.4), the diagram above are commutative. Since dRV ∩W and
dRV ⊕ dRW are isomorphisms by hypothesis, it follows from the five Lemma (Lemma 3.7 of Chapter 1)
that dRV ∪W is an isomorphism. So P verifies the hypothesis of Lemma 3.1 and hence dR = dRU is an
isomorphism. ¤
3.3. Remark. Starting with the singular complex C(U ) = {Cp (U ), ∂p }, we can consider the dual complex
C (U ) = {Cp (U )∗ , ∂p∗ } (see Remark 3.12 of Chapter 1). The cohomology of C ∗ (U ) is called the singular
∗
cohomology of U and it is isomorphic, by Theorem 3.13 of Chapter 1, to the dual of the singular homology
of U . So the de Rham Theorem states that the singular cohomology and the de Rham cohomology are
isomorphic. The de Rham cohomology H ∗ (U ) = ⊕p≥0 H p (U ) has a natural product, induced by the exterior
product of forms, which is distributive, associative and graded commutative, (see Remark 2.12 of Chapter
1). In the singular cohomology it is possible to introduce, by geometric arguments, a product, also called
the cup product, which is distributive, associative and graded commutative. The de Rham Theorem actually
says that dR, extended by linearity, is an isomorphism of algebras.
3.4. Remark. Singular homology is usually defined by starting with continuous simplices i.e., continuous
maps σ : ∆p −→ U 6. The singular (continuous) chain complex C 0 (U ) = {Cp0 (U ), ∂p } is defined in the obvious
way, i.e. the spaces Cp0 (U ) are the vector spaces with basis the singular continuous simplices and the boundary
operator is defined just as in the smooth case. The basic properties, such as homotopy invariance and the
Mayer-Vietoris exact sequence, are also proved just as in the smooth case. The inclusion C(U ) −→ C 0 (U ) is a
morphism of chain complexes, so it induces a map between the homology groups. Using the same arguments
as in the proof of the de Rham Theorem, it is easy to prove that the inclusion induces an isomorphism in
homology.
6Here U can be any topological space.

4. TENSOR PRODUCT OF VECTOR SPACES AND THE KÜNNET’S THEOREM 68
4. Tensor product of vector spaces and the Künnet’s Theorem
A natural question to ask is the following: given open sets U1 ⊆ Rn and U2 ⊆ Rm find a relation between
the cohomology groups of U1 , U2 and U1 × U2 ⊆ Rn × Rm .
To answer this question we need some preliminary algebraic facts. To start with we need a slightly
different approach to tensors.
4.1. Definition. Let E, F be two real vector spaces (not necessarily finite dimensional). Consider the
vector space freely generated by {(x, y) : x ∈ E, y ∈ F} and the subspace generated by the elements
• (x1 + x2 , y) − (x1 , y) − (x2 , y), (x, y1 + y2 ) − (x, y1 ) − (x, y2 ), xi ∈ E, yi ∈ F.
• r(x, y) − (rx, y), r(x, y) − (x, ry), x ∈ E, y ∈ F, r ∈ R.
The quotient space is called the tensor product of E and F and will be denoted by E ⊗ F. The class of (x, y)
in E ⊗ F will be denoted by x ⊗ y.
In other words we may think of E ⊗ F as the space of finite (formal) linear combinations of elements of
the type x ⊗ y with the “calculus rules”
• (x1 + x2 ) ⊗ y = x1 ⊗ y + x2 ⊗ y, x ⊗ (y1 + y2 ) = x ⊗ y1 + x ⊗ y2 ,
• r(x ⊗ y) = rx ⊗ y = x ⊗ ry.
The following facts are easily verified
4.2. Proposition.
(1) E ⊗ F ∼
= F ⊗ E, E ⊗ R ∼= E.
∼
(2) (E ⊗ F) ⊗ P = E ⊗ (F ⊗ P).
(3) E ⊗ (F ⊕ P) ∼
= E ⊗ F ⊕ E ⊗ P.
(4) If {ei }, {fj } are bases for E, F respectively, then {ei ⊗ fj } is a basis for E ⊗ F. In particular, if E, F
are finite dimensional, dim(E ⊗ F) = dim(E) dim(F).
(5) If E is finite dimensional, E∗ ⊗ E∗ ∼
= E2 .
Let π : E × F −→ E ⊗ F be the bi-linear extension of π(x, y) = x ⊗ y.
4.3. Proposition. The following universal property of the tensor product holds
• (UP⊗) If K is a vector space and b : E × F −→ K, is a bilinear map, then there exists a unique linear
map l : E ⊗ F −→ K such that l ◦ π = b.
Proof. Set l(x ⊗ y) = b(x, y). By the “calculus rules”, l extend to a linear map of E ⊗ F into K such
that l ◦ π = b. If l0 : E ⊗ F −→ K is a linear map with l0 ◦ π = b, then l0 (x ⊗ y) = b(x, y) = l(x ⊗ y). Since
the elements of the type x ⊗ y span E ⊗ F, we have l = l0 . ¤
The general philosophy is that objects defined by universal properties are unique.
4.4. Proposition. If H is a vector space and π̃ : E × F −→ H is a bi-linear map such that UP⊗ is
verified for (π̃, H), then H ∼
= E ⊗ F.
Proof. From the universal property for π : E × F −→ E ⊗ F follows that there is a unique linear map
l : E ⊗ F −→ H such that l ◦ π = π̃. From the universal property of π̃ : E × F −→ H follows that there
is a unique map l0 : H −→ E ⊗ F such that l0 ◦ π̃ = π. Now, l ◦ l0 : H −→ H is such that π̃ ◦ (l ◦ l0 ) = π̃.

But also π̃ ◦ 11 = π̃. Hence, by uniqueness, (l ◦ l0 ) = 11. Analogously l0 ◦ l = 11, hence l and l0 are inverse
isomorphisms. ¤
The important feature of the tensor product is that it allows us to transform a bi-linear problem into a
linear problem, which is, generally, easier to solve.
Let Ei , Fi , i = 1, 2 be vector spaces and let Li : Ei −→ Fi be linear map. We define
L1 ⊗ L2 : E1 ⊗ E2 −→ F1 ⊗ F2 , L1 ⊗ L2 (v ⊗ w) := L1 (v) ⊗ L2 (w).
We will need also the following result , whose proof we will leave to the reader (Exercise 7.23)
φ ψ
4.5. Proposition. Let · · · −→ E1 −→ E2 −→ E3 −→ · · · be an exact sequence and let F be a vector
φ⊗11 ψ⊗11
space. Then the sequence · · · −→ E1 ⊗ F −→ E2 ⊗ F −→ E3 ⊗ F −→ · · · is exact.
The result we have promised, called the Künnet Theorem, or also the Künnet formula, is the following
4.6. Theorem. [Künnet’s Theorem] Let U1 ⊆ Rn and U2 ⊆ Rm be open set. Then
H k (U1 × U2 ) ∼
= ⊕p+q=k H p (U1 ) ⊗ H q (U2 ).
Proof. Let us denote by πi : U1 × U2 −→ Ui the projection maps. Let W ⊆ U1 be an open set and
consider the map
κW : Ωp (W ) ⊗ Ωq (U2 ) −→ Ωk (W × U2 ), k = p + q, κW (ω ⊗ τ ) = π1∗ ω ∧ π2∗ τ.
Since πi∗ commutes with d, κW induces a morphism in cohomology. Summing up these morphism, for
p + q = k, we get a map, still denoted by κW ,
κW : ⊕p+q=k H p (W ) ⊗ H q (U2 ) −→ H k (W ⊗ U2 ).
We want to prove that κU1 is an isomorphism. For this we will use the Mayer Vietoris argument (Lemma
3.1) as in the proof of the de Rham Theorem. Let us consider the statement
P(W ) = κW : ⊕p+q=k H p (W ) ⊗ H q (U2 ) −→ H k (W ⊗ U2 ) is an isomorphism.
We have to show that the conditions of the Lemma 3.1 are verified. Clearly P(W ) is true if W is convex.
Also if Wα are disjoint open sets such that P(Wα ) is true, the same holds for W = ∪α Wα . It remains to
show that if V, W ⊆ U1 are open sets such that P(V ), P(W ) and P(V ∩ W ) are true, them P(V ∪ W ) is
true. Consider the Mayer-Vietoris sequence
∆∗
· · · −→ H p (V ∪ W ) −→ H p (V ) ⊕ H p (W ) −→ H p (V ∩ W ) −→ H p+1 (V ∪ W ) −→ · · ·
Tensoring with H q (U2 ) and summing for p + q = k we obtain the diagram
⊕∆∗ ⊗11
· · · −→ ⊕p+q=k H p (V ∩ W ) ⊗ H q (U2 ) −→ ⊕p+q=k H p+1 (V ∪ W ) ⊗ H q (U2 ) −→ · · ·
↓ κV ∩W ↓ κU ∪V
k ∆∗ k+1
· · · −→ H ((V ∩ W ) × U2 ) −→ H ((V ∪ W ) × U2 ) −→ · · ·
The upper line is exact by Corollary 4.5, the lower one is exact being the Mayer-Vietoris sequence of
V × U2 , W × U2 ⊆ (V ∪ W ) × U2 . Moreover κV ∩W and κV ⊕ κW are isomorphisms, by hypothesis. So we
5. INTEGRATION OF 1-FORMS AND SOME APPLICATIONS 70
can use the five Lemma to conclude that κV ∪W is an isomorphism, once we show that the squares commute.
This is obvious for all squares but the one that appear in the picture above. For this square we have
• κV ∪W ◦ (∆∗ ⊗ 11)(ω ⊗ φ) = π1∗ ∆∗ ω ∧ π2∗ φ,
• ∆∗ ◦ κV ∩W (ω ⊗ φ) = ∆∗ π1∗ ω ∧ π2∗ φ.
The conclusion follows from the fact that ∆∗ commutes with induced maps (Remark 4.14 of Chapter 1). ¤
5. Integration of 1-forms and some applications
Let U ⊆ Rn be an open set. A smooth curve γ : [a, b] −→ U can be seen as the smooth 1-simplex
γ̃ = γ ◦ L(a, b) where L(a, b) = (1 − t)a + tb. If ω ∈ Ω1 (U ) is a 1-form, we define
Z Z Z 1 hX i Z b hX i
ω := ω= ωi (γ̃(t))γ̃˙ i (t) dt = ωi (γ(t))γ̇i (t) dt,
γ γ̃ 0 a
where the second integral is the integral of ω on the 1-simplex γ̃ and the last equality comes from the formula
of change of variable in 1-dimensional integrals (see also Example 1.8). For the rest of this section, when
clear from the context, we will make no difference between the curve γ and the 1-simplex γ̃.
Let γ : [a, b] ⊆ R −→ U be a piecewise smooth curve, i.e. a continuous curve such that there exists a
partition t0 = a < t1 < · · · < tk = b of [a, b] such that γi := γ|[ti , ti+1 ] is smooth. Then γ can be viewed either
P
as the (smooth) 1-chain γ = γi or as a continuous 1-simplex. Clearly, in both cases, ∂γ = γ(b) − γ(a).
Let γ : [a, b] ⊆ R −→ U be a continuous closed curve, i.e. γ(a) = γ(b). Consider the map π : [a, b] −→
S := {x ∈ R2 : kxk = 1}, π((1 − t)a + tb)) = (cos 2πt, sin 2πt). Since γ is closed, γ = γ ◦ π −1 is a well
1
defined continuous map of S 1 into U . Conversely, any such map defines a continuous closed curve. From this
point of view, continuous closed curves and continuous maps of the circle into U look like the same thing.
However, there are some differences:
• If γ is a smooth curve γ will be just piecewise smooth. It will be smooth if and only if the derivatives
of all orders of γ at a, coincide with the derivatives of the corresponding order of γ at b.
• Any curve γ : [a, b] −→ U is homotopic to a constant (see Exercise 7.3). This is not the case for maps
of S 1 into U . The following result,whose proof is quite obvious, relates the two situations:
5.1. Lemma. Let γ i : S 1 −→ U, i = 0, 1 be continuous maps and γi be the corresponding closed curves.
Then γ 0 ∼ γ 1 if and only if there is a homotopy H : [a, b] × [0, 1] −→ U between γ0 and γ1 such that
H(a, s) = H(b, s) ∀s ∈ [0, 1].
5.2. Remark. A homotopy like the one in Lemma 5.1 is called a free homotopy and the maps γ i are
said to be freely homotopic. The word “free” is to distinguish this concept from the one of based homotopy,
frequently used in homotopy theory, for example in the definition of the fundamental group.
When clear from the context we will make no distinction between γ and γ.
Let γ : [a, b] −→ U be a closed piecewise smooth curve. Then, if we think of γ as a smooth 1-chain,
∂γ = 0 and therefore it determines an element [γ] ∈ H1 (U ).
5.3. Lemma. If γ0 and γ1 are freely homotopic piecewise smooth closed curves, then [γ0 ] = [γ1 ] in H1 (U ).
Proof. Let H : [a, b]×[0, 1] −→ U be a free homotopy between the two curves. Subdividing [a, b]×[0, 1]
into triangles and using linear simplices as in the proof of homotopy invariance for singular homology (see
Theorem 2.10), we get a chain H̃ with ∂ H̃ = γ1 − γ0 . ¤
An other important variant of the concept of homotopy of curves is the following:
5.4. Definition. Let γi : [a, b] −→ U, i = 0, 1 be curves such that γ0 (a) = γ1 (a), γ0 (b) = γ1 (b).
An endpoints fixing homotopy between the two curves is a homotopy H : [a, b] × [0, 1] −→ U such that
H(a, s) = γ0 (a), H(b, s) = γ0 (b), ∀ s ∈ [0, 1].
If such homotopy exists, we will say that the curves are homotopic relative to the endpoints.
The following (well known) facts follow easily from the Theorem of de Rham (we invite the reader to
give a more elementary proof, see Exercise 7.11).
5.5. Proposition. Let ω ∈ Ω1 (U ) be a closed 1-form.

(1) If γi , i = 0, 1 are freely homotopic piecewise smooth closed curves (resp. curves homotopic relative
to the endpoints) then: Z Z
ω= ω.
γ0 γ1
(2) ω is exact if and only if for all closed curves γ
Z
ω = 0.
γ
5.6. Definition. A connected open set U ⊆ Rn is simply connected if every closed curve is freely
homotopic to a constant curve 7.
From Proposition 5.5 we have:
5.7. Corollary. If U is simply connected, then H 1 (U ) = {0}.
5.8. Remark. A natural question is whether H 1 (U ) = {0} implies that U is simply connected. The
answer to this question is affirmative for n = 2 (see Exercise 7.25) and negative if n ≥ 3. For example there
are, in R3 , (complicated) closed sets, homeomorphic to the 3-dimensional closed disk, whose complements
are not simply connected (for example the so called “horned sphere”). The complement of such a disk has,
by the Jordan-Alexander duality (see Theorem 5.4 of Chapter 1), the same cohomology as the complement
of the standard 3-dimensional disk, hence vanishing first cohomology group (see Example 4.15 of Chapter
1). We do not know of any simpler example in dimension 3. For n ≥ 4 there are simpler examples that we
will discuss in Chapter 4.
We will focus now on closed curves in U = R2 \ {0}. In U there is an important 1-form, the angle form
−y x
ω= dx + 2 dy.
x2 + y 2 x + y2
7The concept of simply connectedness is usually defined in terms of the vanishing of the fundamental group. In this group,
two freely homotopic closed curves are in the same conjugacy class (and conversely), but they may not be the same element of
the group. However, the vanishing of the fundamental group is equivalent to the fact that every two closed curves are freely
homotopic.
5. INTEGRATION OF 1-FORMS AND SOME APPLICATIONS 72
It is easily seen that dω = 0, in fact, locally, ω = d arctan(y/x). But ω is not exact since, if γ(t) =
(cos 2πt, sin 2πt),
Z Z 1
ω= 2π[sin2 (2πt) + cos2 (2πt)]dt = 2π 6= 0.
γ 0
In particular, dR([ω])([γ]) = 2π. Since H 1 (U ) ∼

= R, by Examples 4.15 of Chapter 1, [ω] spans H 1 (U )
and [γ] spans H1 (U ) ∼
= R.
5.9. Definition. Let γ : [0, 1] −→ U be a piecewise smooth curve. An angular function for γ is a
piecewise smooth function θ : [0, 1] −→ R such that θ(t) is one of the determinations, in radians, of the
(oriented) angle between e1 and γ(t).
5.10. Lemma. Any piecewise smooth curve γ : [0, 1] −→ U admits angular functions and two angular
functions for γ differ by an entire multiple of 2π.
Proof. Let θ0 ∈ [0, 2π) be the angle between e1 and γ(0), and ω the angle form. Define
Z
θ(t) = ω + θ0 .
γ|[0,t]
Since, locally, ω = d arctan(y/x), θ is an angular function for γ. Finally we observe that two angular
functions, at a given time, are determinations of the same angle, so they differ, at that time, by an entire
multiple of 2π. This multiple does not depend on the time since the difference of the two angular functions
is, divided by 2π, an integers valued continuous function defined on a connected set, hence constant. ¤
5.11. Remark. The advantage of having angular functions is that we can write γ in polar coordinates
γ(t) = kγ(t)k eiθ(t) = kγ(t)k (cos θ(t), sin θ(t)).
Let γ : [0, 1] −→ U be a closed curve and θ an angular function. Since γ(0) = γ(1), θ(1) − θ(0) is an
entire multiple of 2π.
5.12. Definition. The winding number of γ is the integer

θ(1) − θ(0)
w(γ) = .
2π
5.13. Remark. Since two angular functions differ by a multiple of 2π, the winding number does not
depend on the particular angular function. Moreover
Z
1
w(γ) = ω.
2π γ
5.14. Example. Consider the curve ξn (t) = (cos 2πnt, sin 2πnt), t ∈ [0, 1], n a given integer. Then
θ(t) = 2πnt is an angular function and w(ξn ) = n.
The main fact about winding numbers is the following
5.15. Theorem. [Homotopy classification] Two piecewise smooth closed curves γi : [0, 1] −→ U, i = 0, 1,
are freely homotopic if and only if they have the same winding number.
Proof. If the two curves are freely homotopic, by Proposition 5.5 and Remark 5.13, they have the same
winding number. Let γ be a piecewise smooth closed curve with angular function θ and winding number
w(γ) = n ∈ Z. Let ξn be as in Example 5.14. Define
H : [0, 1] × [0, 1] −→ U, H(t, s) = [skγ(t)k + (1 − s)](cos(sθ(t) + (1 − s)2πnt), sin(sθ(t) + (1 − s)2πnt)).
Then H(t, 0) = ξn (t), H(t, 1) = γ(t) and the condition w(γ) = n implies H(0, s) = H(1, s), ∀ s ∈ [0, 1].
Hence H is a free homotopy between ξn and γ. This concludes the proof since the relation of being freely
homotopic is an equivalence relation. ¤
5.16. Remark. Any continuous curve in U admits continuous angular functions. Once we have angular
functions, we can define the winding number for a continuous closed curve. Theorem 5.15 holds true in this
more general situation (see Exercise 7.15).
We will see now some applications of the homotopy invariance of the winding number.
Let D2 (r) := {x ∈ R2 : kxk ≤ r} be the disk of radius r and S 1 (r) := {x ∈ R2 : kxk = r} be its
boundary. Consider a smooth function8 f : D2 (r) −→ R2 . A basic question is to find solutions of the
equation f (x) = 0. In the case of a function f : [−r, r] ⊆ R −→ R, the celebrated Theorem of Bolzano
states that if f (r)f (−r) < 0 the equation has a solution. We will prove a similar result for our case, similar
in the sense that we shall give a condition on f , at the boundary of the disk, that is be sufficient (but not
necessary, in general) for the existence of solutions of our equation.
5.17. Definition. Let f : D2 (r) −→ R2 be a smooth function. Suppose f (x) 6= 0 if kxk = r. The
degree of f , dg(f ), is defined as the winding number of the closed curve:
γf : [0, 1] −→ U := R2 \ {0}, γf (t) = f (r(cos 2πt, sin 2πt))
5.18. Example. Consider the complex plane C ∼

= R2 with complex variable z = x + iy, and the map
g(z) = z n . Then γg (t) = r(cos 2πnt, sin 2πnt). Hence dg(g) = n.
The announced result is the following:
5.19. Theorem. If dg(f ) 6= 0 then the equation f (x) = 0 has a solution.
Proof. Suppose dg(f ) 6= 0, f (x) 6= 0 ∀ x ∈ D2 (r). Consider the map
H : [0, 1] × [0, 1] −→ R2 \ {0}, H(t, s) = f (sr(cos 2πt, sin 2πt)).
Since f (x) 6= 0, for kxk ≤ r, H is a free homotopy, in R2 \{0}, between γf and the constant curve α(t) = f (0).
Therefore, by Theorem 5.15, dg(f ) := w(γf ) = w(α) = 0, a contradiction. ¤
In order to compute degrees, the following fact is often useful
5.20. Lemma. [Lemma of Poincaré-Bohl] Let γi : [0, 1] −→ R2 \ {0}, i = 0, 1 be two closed curves. If
kγ0 (t) − γ1 (t)k < kγ0 (t)k ∀ t ∈ [0, 1], the two curves are freely homotopic.
8By Remark 5.16 we only need continuity of the function.

6. APPENDIX: BARICENTRIC SUBDIVISION AND THE PROOF OF THEOREM 2.13 74
Proof. Consider the map:
H : [0, 1] × [0, 1] −→ R2 , H(t, s) = sγ1 (t) + (1 − s)γ0 (t).
The condition kγ0 (t) − γ1 (t)k < kγ0 (t)k implies that the segment joining γ0 (t) and γ1 (t) does not contain
the origin. Hence H([0, 1] × [0, 1]) ⊆ R2 \ {0} and H is a free homotopy between the two curves. ¤
As an application of Theorem 5.19, we prove now the Fundamental Theorem of Algebra:
5.21. Theorem. Let f (z) = z n + a1 z n−1 + · · · + an−1 z + an be a polynomial in the complex variable z.
If n ≥ 1, f has a complex root.
Pn
Proof. Let r > 1 + 1 |ai |. If f (z) = 0, for some z ∈ S 1 (r), there is nothing to prove. Suppose
f (z) 6= 0 for kzk = r and consider the function g(z) = z n . For kzk = r we have:
n
X
kf (z) − g(z)k ≤ |ai |kzkn−i < rn = kg(z)k.
1
Hence, by Lemma 5.20, f and g have the same degree and dg(g) = n 6= 0, by Example 5.18. Hence, by
Theorem 5.19, the polynomial f has a root in D2 (r). ¤
5.22. Remark. We can take a slightly different approach to the winding number. Let γ : S 1 −→ R2 \{0}
be a closed smooth curve. Then we can extend γ to a map
Γ : R2 \ {0} −→ R2 \ {0}, Γ(tx) = tγ(x), x ∈ S 1 .
Hence we have an induced map
Γ∗ : H1 (R2 \ {0}) ∼
= R −→ H1 (R2 \ {0}) ∼
=R
which is multiplication by a real number, which is, as it is easily seen, the winding number of γ. In this
context the winding number is also called the degree of γ and is denoted by dg(γ). This point of view is
useful in extending the concept to higher dimensions (see Exercise 7.21).
6. Appendix: baricentric subdivision and the proof of Theorem 2.13
Let U = U1 ∪ U2 , U, Ui open sets in Rn . We want to show that the inclusion i : Cp (U1 + U2 ) −→ Cp (U )

induces an isomorphism in homology. The idea of the proof goes as follow: consider a singular simplex
σ : ∆p −→ U and the covering of ∆p given by σ −1 Ui , i = 1, 2. We “subdivide” ∆p into sub simplices of very
small diameter, smaller than the Lebesgue number of the covering, so that σ sends each one of the small
simplecis into U1 or U2 . With this operation we pass from a chain c ∈ Cp (U ) to a chain c̃ ∈ CP (U1 + U2 ).
Finally we must show that if c is a cycle, c̃ is also a cycle that represent, in Hp (U ), the same class of c and
if c̃ is the boundary of a chain in Cp+1 (U ) it is also the boundary of a chain in Cp+1 (U1 + U2 ). To formalize
this idea we start introducing the concept of baricentric subdivision.
Let Γ = [v0 , . . . , vp ] be an (affine) p-simplex in Rn .
7. Exercises
7.1. Let ω = dx1 ∧ · · · ∧ dxp ∈ Ωp (Rn ) and ∆p be the standard p-simplex. Show that
Z
1
ω= (= volume of ∆p ).
∆ p p!
7.2. Let U, V ⊆ Rn be connected open sets and F : U −→ V be a diffeomorphism. Let D ⊆ U be the
closure of a bounded open set and f : V −→ R a smooth function. The change of variables Theorem for
multiple integrals states that:
Z Z
f (y1 , . . . , yn )dy1 · · · dyn = f (F (x1 , . . . , xn ))| det(dF )|dx1 · · · dxn .
F (D) D
Let ω ∈ Ωn (V ). Prove that:

Z Z
ω=± F ∗ ω,
F (D) D
with the sign + (resp. −) if F preserves (resp. reverse) the orientation, i.e. det(dF ) > 0 (resp. det(dF ) < 0).
7.3. Let U ⊆ Rm , V ⊆ Rm be open sets and F : U −→ V a continuous map. Prove that if U (resp. V )
is contractible, then F is homotopic to a constant map.
7.4. Let Dn+1 = {x ∈ Rn+1 : kxk ≤ 1}, S n = {x ∈ Rn+1 : kxk = 1} = ∂Dn+1 and V ⊆ Rm . Show that
a continuous map F : S n −→ V is continuously homotopic to a constant map if and only if it extends to a
continuous map F̃ : Dn+1 −→ V .
7.5. Prove that an open set U ⊆ Rn is connected if and only if H0 (U ) ∼

= R (see Example 2.9).
7.6. Let U ⊆ Rn , V ⊆ Rm be open sets and F : U −→ V a smooth map. Prove that if U is connected,
F∗ : H0 (U ) −→ H0 (V ) is injective. Study the case in which U is not connected (see Example 2.9).
7.7. For an open set U ⊆ Rn define the reduced homology, H̃p (U ), as the homology of the augmented
chain complex
· · · −→ Cp (U ) −→ Cp−1 (U ) −→ · · · −→ C0 (U ) −→ R −→ {0},
where the last map sends any singular 0-simplex to 1 ∈ R and is extended by linearity (the other maps are
the usual boundaries). Find the relation between Hp (U ) and H̃p (U ) and prove the homotopy invariance and
the exactness of Mayer-Vietoris sequence for reduced homology.
7.8. Compute the homology of Σn ⊆ Rn using the Mayer Vietoris sequence for reduced homology (see
Example 4.15 of Chapter 1 for the definition of Σn ).
7.9. Let U ⊆ Rn be an open set and p ∈ U . Assume known that Hn (U ) = {0} (see Remark in Exercise
7.30 of Chapter 1). Find the relation between Hk (U \ {p}) and of Hk (U ).
7.10. Prove the claim made in Remak 3.4 that the homology of the complex of continuous singular
simplices is isomorphic to the homology of the complex of the smooth singular simplices.
7.11. Give a proof, without using the Theorem of de Rham, of Proposition 5.5.
7. EXERCISES 76
7.12. Let U ⊆ Rm be an open set and let α, β : [0, 1] −→ V be continuous curves with α(1) = β(0).
Define α ∗ β as
(
1
α(2t) se 0 ≤ t ≤ 2
α ∗ β(t) = 1
β(2t − 1) se 2 ≤t≤1
−1
and α (t) = α(1 − t). Assuming that the products below are well defined. Prove that
(1) If α1 ∼ β1 , α2 ∼ β2 , then α1 ∗ α2 ∼ β1 ∗ β2 .
(2) (α ∗ β) ∗ γ ∼ α ∗ (β ∗ γ).
(3) α ∗ ²p ∼ α ∼ ²p ∗ α.
(4) α ∗ α−1 ∼ ²p ∼ α−1 ∗ α.
where the homotopies are relative to the endpoints and ²p is the constant path ²p (t) = p.
Hint: consider the homotopies
(1) If Hi are homotopies between αi e βi ,
(
1
H1 (2t, s) se 0 ≤ t ≤ 2
H(t, s) = 1
H2 (2t − 1, s) se 2 ≤t≤1
(2) 
 4t
 α( s+1 )
 se 0 ≤ 4t ≤ s + 1
H(t, s) = β(4t − s − 1) se s + 1 ≤ 4t ≤ s + 2


 γ( 4t−s−2 se s + 2 ≤ 4t ≤ 4
2−s
(3)
(
²p (t) se 0 ≤ 2t ≤ 1 − s
H(t, s) =
α( 2t−1+s
1+s ) se 1 − s ≤ 2t ≤ 2
(4) 
 1−s
 α(2t)
 se 0 ≤ t ≤ 2
1−s
H(t, s) = α(1 − s) se 2 ≤t ≤ 1+s


2
 α−1 (2t − 1) 1+s
se 2 ≤ t ≤1
Give a “geometric” interpretation of the homotopies above.
7.13. Let U ⊆ Rn be an open set and p ∈ U . Consider the set Ω(U, p) = {γ : [0, 1] −→ U :
γ is continuous and γ(0) = γ(1) = p}. Prove that ∗ induces a group structure on the quotient set
π1 (U, p) := Ω(U, p) modulo the equivalence relation α ∼ β if and only if α, β are homotopic relative to
the endpoints.
Remark. With this structure, π1 (U ) is called the fundamental group of U with respect to p.
7.14. Prove that an open set U ⊆ Rn is simply connected if and only if any two curves γi : [0, 1] −→
U, i = 0, 1 with with the same endpoints are homotopic relative to the endpoints.
7.15. Prove that any continuous curve γ : [a, b] −→ R2 \ {0} admits angular functions (hint: use polar
coordinates to prove the claim when the image of γ is contained in a half plane. Then...). Extend Theorem
5.15, Definition 5.17 and Theorem 5.19 to the case of continuous functions.
7.16. Let γ : S 1 −→ R2 \ {0} be an odd closed curve, i.e. γ(−t) = −γ(t), t ∈ S 1 . Prove that w(γ) is
odd.
7.17. Prove the following Theorem of Borsuk: if f, g : S 2 −→ R are odd continuous functions, there
exists p ∈ S 2 such that f (p) = 0 = g(p) (hint: use the projection of the closed upper hemisphere onto the
unit disk to define a function of the disk in R2 ).
7.18. Let f, g : S 2 −→ R be continuous functions. Prove that there exists p ∈ S 2 such that f (p) =
f (−p), g(p) = g(−p).
7.19. Prove that there are no injective continuous functions F : S 2 −→ R2 .
7.20. Let ω = a(x, y)dx + b(x, y)dy be a smooth closed 1-form in R2 \ {0}. Suppose that, for 0 <
x2 + y 2 ≤ K, the functions a, b are bounded. Prove that ω is exact (hint: use homotopy invariance to show
R
that for all closed curves γ : S 1 −→ R2 \ {0}, γ ω = 0).
7.21. Let F : S n −→ S n be a smooth function and F̃ : Rn+1 \ {0} −→ Rn+1 \ {0}, F̃ (tx) = tF (x).
Then we have an induced linear map F̃∗ : Hn (Rn+1 \ {0}) ∼
= R −→ Hn (Rn+1 \ {0}) ∼= R. This map is
9
multiplication by a real number dg(F ), called the degree of F . It is known that dg(F ) ∈ Z . Let Dn+1
be the unit disk and G : Dn+1 −→ Rn+1 a smooth function not vanishing on the unit sphere S n = ∂Dn+1 .
G(x)
Then the degree of G, dg(G), is defined as the degree of the map G̃(x) = kG(x)k . Prove that, if dg(G) 6= 0,
then the equation G(x) = 0 has a solution.
7.22. Prove that there are no smooth maps F : Dn+1 −→ S n = ∂Dn+1 such that F (x) = x ∀x ∈ S n .
Use this fact to prove the celebrated Brouwer fixed point Theorem: any continuous map G : Dn+1 −→ Dn+1
has a fixed point, i.e. a point x ∈ Dn+1 such that G(x) = x (hint for the Brouwer fixed point Theorem:
suppose G(x) 6= x ∀ x ∈ Dn+1 . For x Dn+1 consider the ray starting at G(x) containing x and define F (x)
to be the intersection of this half line with S n . Then ...).
7.23. Let L : E1 −→ E2 be a linear map and let F be a given vector space. Prove that ker(L ⊗ 11) =
ker(L) ⊗ F and Im(L ⊗ 11) = Im(L) ⊗ F. Prove Proposition 4.5.
7.24. Compute the cohomology of Rn \ {0} × Rm \ {0}.
7.25. Let γ : [0, 1] −→ R2 be a Jordan curve. Then, by Theorem 5.8 of Chapter 1, R2 \ γ([0, 1]) has
two connected components. It is easily seen that one component is bounded and the other is unbounded.
Assume the following (non trivial) Theorem
Theorem. The bounded component of R2 \ γ([0, 1]) is homeomorphic to a disk.
Let U ⊆ R2 be an open set such that H 1 (U ) = {0}. Prove that any smooth Jordan curve γ : S 1 −→ U
is homotopic, in U , to a constant curve (hint: by the Theorem above, γ(S 1 ) is the boundary of a disk in R2 .
If the disk is contained in U , the curve is contractible by Exercise 7.4. If not, use the angle form to get a
contradiction).
Remark: This fact implies that U is simply connected (see Remark 5.8), and, by the Riemann mapping
theorem, U is diffeomorphic to R2 .
9It follows, from homotopy invariance, that homotopic maps have the same degree. A basic fact in homotopy theory is
the Theorem of Hopf: if two maps from S n to S n have the same degree, then they are homotopic.
7. EXERCISES 78
7.26. Let U ⊆ R2 be an open set and X : U −→ R2 a smooth vector field. Let D² ⊆ U be a disk of
radius ², with center p ∈ U , and assume that X(q) 6= 0, ∀q ∈ D² \ {p}. The point p is called an (isolated)
singularity of X. The index of X at p, i(X, p), is defined as the degree of X|D² , i.e. the winding number of
the curve X(p + ² cos 2πt, p + ² sin 2πt), t ∈ [0, 1].
(1) Let γ : [0, 1] −→ U be a piecewise smooth, positively oriented closed Jordan curve bounding a disk
in U containing p in its interior. Prove that i(X, p) is the winding number of X ◦ γ.
(2) If X(x, y) = (f (x, y), g(x, y)), prove that
Z
1
i(x, p) = θ,
2π γ
where γ is as in the preceding item and

−gdx f dy
θ= 2 2
+ 2 = X ∗ ω,
f +g f + g2
where ω is the angle form.
(3) Prove that if X(p) 6= 0, then i(X, p) = 0.
(4) Let X : R2 −→ R2 be a linear isomorphism. Prove that i(x, 0) = 1 if det X > 0 and i(x, 0) = −1 if
det X < 0.
(5) Assume that X(p) = 0 and dX(p) is invertible. In this case we say that p is a simple singularity
of X, positive, if det dX(p) > 0, negative otherwise. Prove that a simple singularity is isolated
and i(X, p) = ±1, depending on whether p is a positive or negative simple singularity (hint:
by Taylor’s formula X(q) = dX(0)(q) + R(q)kqk, with limq→0 R(q) = 0. Prove that H(q, s) =
dX(0)(q) + (1 − t)R(q)kqk 6= 0, if kqk is sufficiently small. Hence...).
(6) Prove the following formula, called the Kronecker formula.
Let D ⊆ R2 be a closed disk, with center q and radius r, and X : D −→ R2 be a vector field with
only simple singularities, none of which is in ∂D. Then
Z
1
θ = P − N,
2π γ
where γ(t) = p + r(cos 2πt, sin 2πt), P is the number of the positive singularities and N the number
of the negative ones.
Remark: The condition i(X, p) = 0 does not imply X(p) 6= 0 (find an example!). However, if i(X, p) = 0,
given ² > 0, we can find a vector field X̃ which coincides with X outside a disk of radius ² and center p and
has no zeros in that disk.
7.27. Let f : U ⊆ C = R2 −→ C be a holomorphic function (see Exercise 7.31 of Chapter 1), f = u + iv.
(1) Prove the following Theorem:
Theorem: [Cauchy] If U is simply connected and γ : S 1 −→ U is a closed piecewise smooth curve
then Z Z Z
f (z)dz := (udx − vdy) + i (udy + vdx) = 0.
γ γ γ
(2) Suppose that f 0 (z) 6= 0 for z in a disk D ⊆ U and f (z) 6= 0 for z ∈ ∂U . Prove that the number of
zeros in D is given by Z
1 df
2πi ∂D f
(hint: prove that the singularities of the vector field X(x, y) = (u(x, y), (v(x, y)) are all simple and
positive. Then....).
7.28. Use Exercise 7.21 to define the index of a vector field X : U ⊆ Rn −→ Rn at a point p ∈ U and
try to extend, as much as you can, the facts claimed in Exercise 7.26 for this situation.
CHAPTER 3
A short course on differentiable manifolds
The concept of a differentiable manifold is a natural extension of the concept of a regular surface in R3
or, more generally, of a submanifold of RN . In this Chapter we will recall, briefly, some basic facts about
submanifolds of RN , introduce (abstract) differentiable manifolds, as a natural extension of the concept of
submanifold, and discuss the basic facts of the theory of differentiable manifolds.
We will use the words smooth and differentiable as synonymous of C ∞ .
1. Submanifolds of Euclidean spaces
1.1. Definition. An n-dimensional submanifold of RN is a subset M n ⊆ RN , such that ∀p ∈ M there

exist an open set U ⊆ RN , an open set A ⊆ Rn and a smooth map φ : U −→ RN , p ∈ φ(U ), such that:
(1) φ is a homeomorphism of U onto A ∩ M , i.e. φ has a continuous inverse φ−1 : A ∩ M −→ U ,
(2) ∀ x ∈ U, dφ(x) : Rn −→ RN is injective.
• The integer n is called the dimension of M n .
• The map φ, or better the pair (U, φ), is called a (local regular) parametrization or a (local) chart and, if
p ∈ φ(U ), (x1 , . . . , xn ) = φ−1 (p) are called the local coordinates of p in the chart (U, φ).
• A set of regular parameterizations whose images cover M will be called an atlas.
1.2. Remark. It may appear, at first, that the first condition in the definition is, at least locally, a
consequence of the second one. In fact it follows from Theorem 1.26 of Chapter 0 that there is a neigh-
borhood U 0 of x and a local diffeomorphism Φ of a neighborhood of φ(x) such that Φ ◦ φ(x1 , . . . , xn ) =
(x1 , . . . , xn , 0, . . . , 0) ∀x ∈ U 0 . The point is that the image of U 0 may not be the intersection of M with an
open set of RN . For example the set M = {(x, y) ∈ R2 : xy = 0} may be covered with images of maps with
injective differentials but it is not a submanifold of R2 , since the intersection of an open set containing (0, 0)
with M is not homeomorphic to an open interval. We will came back to this question later on, when we will
talk about immersions and embeddings of manifolds (see Remark 4.2).
1.3. Definition. If φi : Ui ⊆ Rn −→ M, i = 1, 2 are regular parameterizations of M , the map φ−1

2 ◦ φ1 ,
defined in φ−1
1 (φ1 (U1 ) ∩ φ2 (U2 )), is called a change of coordinates.
It follows easily from Theorem 1.26 of Chapter 0 that, for a local regular parametrization φ, φ−1 is,
locally, the restriction of a smooth map defined on an open set of RN . Therefore
1.4. Lemma. If φi : Ui ⊆ Rn −→ M, i = 1, 2 are regular parameterizations of M , the change of

coordinates φ−1
2 ◦ φ1 is a diffeomorphism from its domain onto its image.
81
1. SUBMANIFOLDS OF EUCLIDEAN SPACES 82
1.5. Definition. Let M be a submanifold of RN , p ∈ M and φ : U −→ M be a regular parametrization

with image in a neighborhood of p. Set x = φ−1 (p). The n-dimensional vector space
Tp M := (dφ)(x)(Rn ) ⊆ RN ,
is called the tangent space of M at p.
1.6. Remark. It is easy to see that the definition does not depend on the choice of the local chart.
Observe that, in general, p 6∈ Tp M . The “geometric tangent space” is the affine subspace p + dφ(x)(Rn ).
1.7. Definition. Let M be a submanifold of RN and F : M −→ RK a function. We will say that F is

smooth if, ∀ p ∈ M , there exists a regular parametrization φ : U ⊂ Rn −→ M of a neighborhood of p in M ,
such that the function φ ◦ F : U −→ RK is smooth1. Moreover
• A map F : M ⊆ Rn −→ N ⊆ RK between submanifolds is smooth if it is smooth as a map into RK .
• A smooth map F : M −→ N is a diffeomorphism if it has a smooth inverse.
1.8. Remark. It follows from Lemma 1.4 that the definition of smooth maps does not depend on the
parametrization.
1.9. Remark. We can give an “intrinsic” definition of smooth maps between submanifolds that will be
useful later on. In fact, as it is easily seen, the definition above is equivalent to the following one
1.10. Definition. F is smooth if and only if, given regular parameterizations φ : U −→ M, ψ : Σ −→ N

of neighborhoods of p and f (p), the map ψ −1 ◦ F ◦ φ is smooth, where defined.
1.11. Remark. If U ⊆ M is an open set, p ∈ U , we can consider the algebra of smooth real valued
functions defined on U , F(U ), the algebra of germs at p ∈ U, Fp , and their derivations, Der(U ), Dp (see
Chapter 0, Definitions 3.15 and 3.19). At this point we can look at the tangent space in three ways
• As in the definition: Tp M = dφ(x)(Rn ) ⊆ RN , p = φ(x). A basis for this space is given by the vectors
∂φ
(dφ)(x)(ei ) = (x) := φxi (x), i = 1, . . . , n.
∂xi
We will call those vectors the coordinate tangent vectors.
• Tp M = {γ̇(0) : γ : (−², ²) −→ M ⊆ RN is a smooth curve with γ(0) = p}.
• As the space Dp of derivations of the algebra of germs at p. Given a regular parametrization φ : U ⊂
R −→ RN of a neighborhood of p and a smooth function f : U −→ R, we define the derivations
n
∂ ∂f ◦ φ −1
(p)(f ) = (φ (p)), i = 1, . . . n.
∂xi ∂xi
These derivations are a basis of Dp .
We leave to the reader the task of proving the equivalence of these definitions (Exercise 10.2).
We shall define the differential of a smooth map.
1This is equivalent, again by Theorem 1.26, to the fact that ∀ p ∈ M there exist a smooth extension of F to an open
neighborhood of p ∈ RN .
CHAPTER 3. A SHORT COURSE ON DIFFERENTIABLE MANIFOLDS 83
1.12. Definition. Let M, N be submanifolds of RN and RM respectively and F : M −→ N a smooth

map. We define the differential of F at p
dF (p) : Tp M −→ TF (p) N, dF (p)(dφ(φ−1 (p))(v) = dψ(ψ −1 (F (p))d(ψ −1 ◦ F ◦ φ)(φ−1 (p))(v)
where φ : U −→ M, ψ : Σ −→ N are regular parameterizations of neighborhoods of p and F (p) respectively

and v ∈ Rn . It is easily seen that the definition does not depend on the choice of the charts.
1.13. Remark. Since we have alternative definitions of Tp M (see Remark 1.11) it is convenient to have
alternative definitions of dF (p).
• If we look at Tp M as the space of tangent vectors to smooth curves γ : (−², ²) −→ M, γ(0) = p,
d
dF (p)(γ̇(0)) = |t=0 F ◦ γ(t).
dt
• If we look at Tp M as derivations, we have
dF (p)(Xp )(g) = Xp (F ◦ g).
We leave to the reader the task of proving the equivalences.
We will discuss now a few examples.
1.14. Example. If M is a submanifold of RN , any open subset U of M is a submanifold of RN (of the

same dimension) and Tp U = Tp M .
1.15. Example. Let U ⊂ Rn be an open set and f : U → R a smooth function. Then
φ : U −→ Rn+1 , φ(x) = (x1 , . . . , xn , f (x)), x = (x1 , . . . , xn ) ∈ U,
is a regular parametrization of the graph of the function f , Γ(f ) = {(x, f (x)) : x ∈ U } ⊆ Rn+1 . The vectors
∂f
{(ei , (x)) ∈ Rn × R}, where {ei } is the canonical bases of Rn , span the tangent space at (x, f (x)).
∂xi
1.16. Example. If Mini ⊆ RNi , i = 1, 2 are two submanifolds, then M1n1 × M2n1 ⊆ RN1 +N2 is a
submanifold. Local charts for M1n1 × M2n1 are given by products of local charts for the Mi ’s. The tangent
space at (x, y) is isomorphic to Tx M1 × Ty M2 .
1.17. Example. Consider S n = {x ∈ Rn+1 : kxk = 1}. We shall to show that S n is a submanifold.
Consider a point x ∈ S n . Up to an orthogonal transformation we can suppose x = en+1 , the last vector of
the canonical bases. Then the neighborhood U = S n ∩ {x ∈ Rn+1 : xn+1 > 0} is parameterized by
X 1
φ : B(1) = {y ∈ Rn : kyk < 1} −→ U, φ(y1 , . . . , yn ) = (y1 , . . . yn , (1 − yi2 ) 2 ).
For any other point x ∈ S n we can define a parametrization φx = Ax ◦ φ where Ax is an orthogonal

operator such that Ax (en+1 ) = x. Observe that, in order to cover S n , we need at least 2(n + 1) of such
parameterizations.
A more economic way of parameterize S n is via the stereographic projections. The stereographic projec-
tion from the point en+1 , πen+1 : S n \ {en+1 } −→ Rn = {x ∈ Rn+1 : hx, en+1 i = 0} is the map that associate
to x ∈ S n \ {en+1 } the point of intersection of the straight line tx + (1 − t)en+1 with Rn . The condition
htx + (1 − t)en+1 , en+1 i = 0 gives
πen+1 (x) = [1 − hx, en+1 i]−1 x − hx, en+1 i[1 − hx, en+1 i]−1 en+1 ,
whose inverse is given by
πe−1
n+1
(y) = 2[1 + kyk2 ]−1 y + [kyk2 − 1][1 + kyk2 ]−1 en+1 .
Clearly πe−1
n+1
: Rn −→ S n \ {en+1 } is a regular parametrization. Taking the inverse of the stereographic
projection from −en+1 we get an other regular parametrization, and the two parametrization are an atlas
for S n .
Let x ∈ S n and γ : (−², ²) −→ S n be a smooth curve with γ(0) = x. Differentiating the identity
hγ(t), γ(t)i = 1 at t = 0, we get hx, γ̇(0)i = 0. It follows that
Tx S n = x⊥ := {v ∈ Rn+1 : hx, vi = 0}.
1.18. Example. Let M ⊆ RN be an n-dimensional submanifold. Consider the tangent bundle
T M := {(p, X) ∈ M × RN : X ∈ Tp M } ⊆ RN × RN .
This is a 2n-dimensional submanifold of R2N . Il fact, if φ : U ⊆ Rn −→ M is a regular parametrization,

n
X
Φ : U × Rn −→ T M, Φ(x1 , . . . , xn , t1 , . . . , tn ) = (φ(x), ti φxi ) ∈ RN × RN ,
1
is a regular parametrization for T M . If {(Ui , φi )} is an atlas for M , the corresponding parameterizations

{(Ui × Rn , Φi )} form an atlas for T M . We also observe that the tangent bundle comes with a natural
projection π : T M −→ M, π(p, X) = p, which is a smooth map. T M is a prototype of what is called a
(smooth) vector bundle. Although locally T M is diffeomorphic to a product, it is not so globally, in general,
as we will see along those notes.
1.19. Example. A similar example is the normal bundle of a submanifold M ⊆ RN . It is defined as
νM = {(p, ξ) ∈ M × RN : ξ ∈ [Tp M ]⊥ }.
We want to produce regular parameterizations for νM . Let φ : U −→ M, φ(x) = p, be a parameterizations

of a neighborhood of p ∈ M .
Claim. There is a neighborhood U 0 ⊆ U of x and smooth functions ξ1 , . . . , ξN −n : U 0 −→ RN , such that
the ξi (y)’ span [Tφ(y) M ]⊥ , ∀ y ∈ U 0 .
Proof. The maps φxi are smooth maps that span Tφ(y) M, ∀ y ∈ U . Using the orthonormalization
process we can find smooth maps Xi : U −→ Rn such that hXi (y), Xj (y)i = δij and they span Tφ(y) M, ∀ y ∈
P
U . Fix now a basis {ξi } of [Tφ(x) M ]⊥ and define ξi (y) = ξi − j hξi , Xj (y)iXj (y). Clearly ξi (y) ∈ [Tφ(y) M ]⊥ .
Since ξi (x) = ξi , they are linearly independent at x, hence in a neighborhood U 0 ⊆ U of x. ¤
At this point we define local parameterizations (with the notations of the Claim) by
X
Φ : U 0 × RN −n −→ νM, Φ(y, t1 , . . . , tN −n ) = (φ(y), tj ξj (y)).
Also the normal bundle cames with a natural projection π : νM −→ M which is a smooth map.
1.20. Example. We will discuss now an example that will be useful later on. Consider the space
M (p, n, R) of p × n matrices with real entries, and its natural identification with Rpn . Consider the subspace
M (p, n; k) of the matrices of rank k. We want to show that M (p, n; k) is a submanifold of M (p, n, R). Let
" #
A0 B0
E0 =
C0 D 0
be in M (p, n; k). We will suppose initially that A0 ∈ GL(k, R), i.e. det A0 6= 0. In particular there is ² > 0
such that, if kA − A0 k < ², A ∈ Gl(k, R). Set
" #
A B
U ={ ∈ M (p, n, R) : kA − A0 k < ²}.
C D
U is an open neighborhood of E0 in M
" (p, n, R).#
A B
Claim. If A ∈ GL(k), then E = ∈ M (p, n; k) if and only if D = CA−1 B.
C D
Proof. Consider " #" # " #
11 0 A B A B
=
X 11 C D XA + C XB + D
Since the first matrix is invertible,
" #
A B
rankE = rank .
XA + C XB + D
Take now X = −CA−1 . The matrix at the right hand side become
" #
A B
.
0 D − CA−1 B
Hence E has rank k if and only if CA−1 B − D = 0. ¤
k(p+n−k)
We exhibit now a (local) regular
" parametrization
# of a neighborhood of E0 . "Identify R
# with the
A B A B
space of matrices of the type , and consider the open subset U = { : kA − A0 k < ²}.
C 0 C 0
Consider the map Ã" #! " #
A B A B
φ : U −→ M (p, n, R), φ = .
C 0 C CA−1 B
It is easily seen that φ is a homeomorphism onto its image. Moreover
Ã" #! Ã" #! " #
A B X Y X Y
dφ = .
C 0 Z 0 Z ∗
Hence dφ is injective φ is a regular parametrization with image U ∩ M (p, n; k).
In the general case we can proceed in the following way. Let E ∈ M (p, n; k). Then there exists a
submatrix A ∈ GL(k). By suitably permutations of rows and columns, which do not affect the rank, we
can construct an isomorphism L of M (p, n, R) in itself that takes E onto a matrix L(E) ∈ M (p, n; k) such
that the upper left block is in GL(k). Then, by the argument above, we have a regular parametrization of
a neighborhood of L(E). Composing this parametrization with L−1 we obtain a regular parametrization of
a neighborhood of E.
Hence M (p, n; k) is a submanifold of M (p, n, R) and its dimension is k(p + n − k).
A useful tool for constructing examples of submanifolds is given by the following
1.21. Proposition. Let U ⊆ RN be an open set and F : U → Rp a smooth map. Let c ∈ Rp be a regular
value of F , i.e. dF (x) is surjective ∀ x ∈ F −1 (c). Then M := F −1 (c) is a (N − p)-dimensional submanifold
of RN and Tp M = ker dF (p).
Proof. Let p ∈ M . We can suppose, without loss of generality, p = 0, F (0) = 0. By Theorem 1.27 there
is a neighborhood of U of 0 and a local diffeomorphism Ψ : U −→ Ψ(U ) ⊆ RN such that F ◦ Ψ(xi , . . . , xN ) =
(x1 , . . . , xN −p ). Then φ(x1 , . . . , xN −p ) = Φ−1 (x1 , . . . , xN −p , 0, . . . 0) is a regular parametrization of U ∩
f −1 (c). We also claim that Tp M = ker df (p). In fact, if γ(−², ²) −→ M is a smooth curve with γ(0) =
d
p, f (γ(t)) = c. Therefore df (p)(γ̇(0)) = dt |t=0 f (γ(t)) = 0. Hence Tp M ⊆ ker df (p)2. Since both subspaces
are (N − p)-dimensional, they agree. ¤
1.22. Example. Consider again the sphere S n = {x ∈ Rn+1 : kxk = 1}. Then S n = F −1 (1) where
F (x) = kxk2 . Now 1 is a regular value of F . In fact, since dF (x)(v) = 2hx, vi, dF (x) : Rn+1 −→ R is non
zero, if x 6= 0, hence surjective. Finally Tx S n = ker dF (x) = x⊥ := {v ∈ Rn+1 : hx, vi = 0}.
2
1.23. Example. Let M (n, R) be the space of n × n matrices with real entries, identified with Rn .
Consider the group of linear isometries of Rn , O(n) = {A ∈ M (n, R) : AAt = 11} (At being the transpose of
A). We want to show that O(n) is a submanifold of M (n, R). It is natural to consider the function
F : M (n, R) −→ M (n, R), F (A) = AAt .
Then O(n) = F −1 (11). The first idea would be to show that 11 is a regular value of F . This is not the case
since, if it were, O(n) would be 0-dimensional. More formally, a simple computation gives
dF (A)(B) = BAt + AB t .
Observe that BAt + AB t is a symmetric matrix, so the image of dF (A) is not the whole M (n, R). But this
observation suggests a way out. First observe that F (A) belongs to the subspace Sym(n, R) ⊆ M (n, R) of
symmetric matrices. So we may consider F as a map with values in Sym(n, R). As such, 11 ∈ Sym(n, R)
1
is a regular value, since, given A ∈ O(n) and C ∈ Sym(n, R), dF (A)(B) = C, for B = 2 CA, hence
dF (A) : M (n, R) −→ Sym(n, R) is surjective. Since Sym(n, R) has dimension n(n + 1)/2, O(n) has
dimension n(n − 1)/2.
Finally since ker dF (11) is the space Skw(n, R) of antisymmetric matrices, T11 O(n) = Skw(n, R).
1.24. Example. Consider the set V2,n = {(x, y) ∈ Rn ⊕ Rn : kxk = 1 = kyk, hx, yi = 0}. This set is
called the Stiefel manifold of 2-frame (orthonormal) in Rn . We will show that V2,n is a submanifold of R2n
of dimension 2n − 3. Observe that V2,2 = O(2), which is two (disjoint) copies of S 1 . Consider the function
F : Rn ⊕ Rn −→ R3 , F (x, y) = (kxk2 , kyk2 , hx, yi).
Clearly V2,n = F −1 (1, 1, 0). We shall show that (1, 1, 0) is a regular value of F . A simple computation gives
dF (x, y)(X, Y ) = (2hx, Xi, 2hy, Y i, hX, yi + hx, Y i).

2We are using the second definition of T M in Remark 1.11
p
Now, given (x, y) ∈ V2,n , (a, b, c) ∈ R3 , we have dF (x, y)(X, Y ) = (a, b, c) for X = ax/2 + cy/2, Y =
by/2 + cx/2. Then (1, 1, 0) is a regular value for F and V2,n is a submanifold of dimension 2n - 3.
More in general, we can consider Vk,n = {(x1 , . . . , xk ) ∈ Rkn : hxi , xj i = δij . Vk,n is a submanifold of
Rkn .
1.25. Definition. Let M be a submanifold of RN . A smooth (tangent) vector field on M is a smooth

map X : M −→ RN such that X(p) ∈ Tp M, ∀ p ∈ M . We will denote by H(M ) the real vector space of
such vector fields on M .
1.26. Remark. If W ⊆ M is an open set (in M ), then W is a submanifold of Rn and we can talk about
vector fields on W and of the restrictions of vector fields on M to W .
1.27. Remark. We can think of a tangent vector field as a smooth map X : M −→ T M , such that
π ◦ X = 11M (see Example 1.18). This point of view will be useful in the theory of (abstract) differentiable
manifolds.
1.28. Example. Let M be a submanifold and W ⊆ M an open subset, image of a regular parametrization
φ : U ⊆ Rn −→ W . Then φxi , i = 1, . . . , n, are tangent vector fields on W . Since they are, at each point
Pn
p ∈ W , a basis for Tp M (= Tp W ), any other vector field in H(W ) may be written as X = 1 Xi φxi , where
the Xi : W −→ R are smooth functions. If we think of tangent vectors as derivations we can write
n
X ∂
X= Xi .
1
∂xi
∂
The vector fields are called coordinate vector fields.
∂xi
As we have observed previously, the tangent bundle to a n-dimensional submanifold M n ⊆ RN , may
not be, globally, diffeomorphic to the product M n × Rn .
1.29. Definition. A submanifold M n ⊆ RN is parallelizable if there exists a diffeomorphism t : M n ×

Rn −→ T M n such that t|{p} × Rn is a (linear) isomorphism onto Tp M n . Such a diffeomorphism is called a
trivialization of T M n .
1.30. Proposition. Let M n ⊆ RN be a submanifold. Them M n is parallelizable if and only if there

exist vector fields X1 , . . . , Xn ∈ H(M n ) such that ∀ p ∈ M n , X1 (p), . . . , Xn (p) are linearly independent.
Proof. Let t : M n × Rn −→ T M n be a trivialization. Then Xi (x) = t(x, ei ) are smooth vector fields
linearly independent at each point. Conversely, given n vector fields, X1 , . . . , Xn , linearly independent at
P
each point, we define t(x, t1 , . . . , tn ) = ti Xi (x). t is clearly smooth, a bijection and t|{p} × Rn is a linear
isomorphism onto Tp M . We have just to show that t−1 is smooth. Let φ : U ⊆ Rn −→ M be a local chart
P
and Φ : U × Rn −→ T M be the associated local chart for T M as in Example 1.18. Then φxi = aij Xj for
some smooth functions aij : U −→ R. Therefore
X X X X
t−1 Φ(x, t1 , . . . , tn ) = t−1 (φ(x), ti φxi ) = t−1 (φ(x), ti aij Xj ) = (x, ti ai1 , . . . ti ain ).
i ij i i
−1
It follows that t is smooth and this conclude the proof. ¤
2. DIFFERENTIABLE MANIFOLDS 88
1.31. Example. If M can be cover by only one regular parametrization, in particular if it is an open
set of Rn , then M is parallelizable.
1.32. Example. Consider the unit sphere S n ⊆ Rn+1 . As we have seen in Example 1.22, Tx S n = x⊥ =
{v ∈ Rn+1 : hx, vi = 0}. Hence a vector field on S n can be viewed as a smooth map X : S n −→ Rn+1 such
that hX(x), xi = 0. If n = 2k − 1 a nowhere zero vector field is given by
X(x1 , x2 , . . . , x2k−1 , x2k ) = (−x2 , x1 , . . . , −x2k , x2k−1 ).
We will see later that, if n = 2k, every vector field has must have a zero. Hence, by Proposition 1.30,
even dimensional spheres are not parallelizable. It turns out that the only parallelizable spheres of positive
dimension are S 1 , S 3 , S 7 . This is a highly non trivial result, connected with the existence of a division
algebra structure3 on Rn+1 (see also the comments in Remark 2.13).
1.33. Example. Consider the submanifold O(n) ⊆ Rn (Example 1.23). We have seen that T11 O(n) ∼
2
=
Skw(n, R). For A ∈ O(n), consider the map LA : O(n) −→ O(n), LA (B) = AB. Then LA is a smooth
map, being the restriction of a linear map, and, in fact, it is a diffeomorphism, with [LA ]−1 = LA−1 . Let
{E1 , . . . , Ek }, k = n(n − 1)/2, be a basis of T11 O(n). The maps Ẽi : O(n) −→ T O(n), Ẽi (A) = dLA (Ei )
are smooth vector fields on O(n) 4 and, since dLA (11) : T11 O(n) −→ TA O(n) is an isomorphism, they are
linearly independent at each point. Therefore O(n) is parallelizable 5.
2. Differentiable manifolds
The careful reader has probably noted that all we have done up to now depends, essentially, only on
the fact that the change of coordinates are smooth functions. This observation leads to the concept of an
abstract differentiable manifold, that we will introduce in this section.
In what follows we will always assume that the topological spaces involved are Hausdorff and second
countable. We recall that a topological space is second countable if there is a countable bases6 for the
topology. Such spaces, as it easily seen, have the Lindeloff property: any open covering contains a countable
subcovering. The reader more familiar with metric spaces may think of separable metric spaces (which are
Hausdorff and second countable). This assumption is not restrictive (see Remark 3.3).
We start with a few definitions.
2.1. Definition. Let M be a topological space. A topological atlas, or simply an atlas, for M is a
collection of open sets Uα ⊆ Rn and continuous maps φα : Uα −→ M such that:
(1) φα (Uα ) is open in M and φα is a homeomorphism onto its image,
S
(2) α φα (Uα ) = M .
3An algebra A is a division algebra if the maps L (x) = ax and R (x) = xa are surjective, if a 6= 0. If the algebra has a
a a
multiplicative unit, this is equivalent to the fact that any non zero element is invertible.
4Observe that, since multiplication of matrices is bilinear, hence smooth, L is smooth also with respect to A.
A
5This is a particular case of the fact that a Lie group is parallelizable (see section 5).
6If X is a topological space, a bases for the topology of X is a family of open sets such that every open set of X is union
of elements of the bases.

The maps φα : Uα −→ M are called charts or local coordinates and the maps φ−1
β ◦ φα (where defined)
are the change of coordinates.
A topological space which admits an atlas is called a topological manifold.
2.2. Definition. An atlas is a differentiable (or a smooth) atlas if the change of coordinates φ−1
β ◦ φα
are smooth functions (where defined).
Two smooth atlases on M are said to be equivalent if their union is a smooth atlas.
An equivalence class of smooth atlases is called a differentiable or smooth structure.
2.3. Remark. It is clear that each equivalence class of differentiable atlases contains a maximal element,
the union of all atlases in the class. We can think that a differential structure is such a maximal atlas.
At this point we can give the desired definition.
2.4. Definition. A differentiable or smooth manifold is a topological manifold together with a differen-
tiable structure.
2.5. Remark. A word on the topological conditions in the definition of smooth manifold. The Hausdorff
condition is a natural one to avoid pathologies. It is clear that the existence of an atlas on M implies that
M is a locally Hausdorff spaces, i.e. ∀ p ∈ M there exists an open neighborhood U of p which is an
Hausdorff space (for example the image of a chart). However there are spaces that admit atlases which are
not Hausdorff. A classical example is the real line with two origins, i.e. the quotient of R × {−1, 1} modulo
the equivalence relation (t, α) ∼ (s, β) ⇐⇒ t = s 6= 0 or (t, α) = (s, β). In fact M = R × {−1, 1}/ ∼, with
the quotient topology7, can be covered by the charts φ± : R −→ M, φ± (t) = [t, ±1]8. But [0, 1] and [0, −1]
can not be separated by disjoint open sets.
The second countability condition will allow us to construct partitions of unity, a basic tool in order to
glue together locally defined objects to obtain a globally defined one.
Clearly submanifolds of RN are example of smooth manifolds. We will carry on the program of extending
to this context what we have done for submanifolds.
For differentiable manifolds we can copy Definition 1.10 in order to define smooth maps.
2.6. Definition. Let M, N be a differentiable manifolds. A map F : M −→ N is smooth if for every

pair of chart φα : Uα −→ M and ψβ : Σβ −→ N, ψβ−1 ◦ F ◦ φα is smooth, where defined.
F is a diffeomorphism if it is smooth and has a smooth inverse.
2.7. Remark. Since change of coordinates are smooth, it is sufficient to check differentiability for given
local charts (see also Remark 1.8).
7Let X be a topolgical space and ∼ an equivalence relation.Consider the quotient set X/ ∼ and the quotient map
π : X −→ X/ ∼. The quotient topology on X/ ∼ is the topology such that U ⊆ X/ ∼ is open if and only if π −1 (U ) is open in
X. This topology is characterized by the property that given a topological space Y and a map f : X/ ∼−→ Y , f is continuous
if and only if f ◦ π : X −→ Y is continuous.
8[t, ²] is the class of (t, ²) in M .
2.8. Remark. The natural equivalence relation in the theory of smooth manifolds is the relation of being
diffeomorphic, i.e. we consider two smooth manifolds M, N to be the same if there exists a diffeomorphism
F : M −→ N . This relation is not to be confused with equivalence of atlases. We make this point
clear. Suppose that M is a topological manifold with two smooth atlases, A, B. Then the two atlases are
equivalent if and only if 11 : (M, A) −→ (M, B) is a diffeomorphism. This is more restrictive that requiring
the existence of a diffeomorphism F : (M, A) −→ (M, B). For example consider in R the two differentiable
structures containing φ(t) = t and ψ(t) = t3 respectively. Those structure are not the same since the changes
√
of coordinates ψ −1 φ(t) = 3 t is not differentiable at 0 ∈ R. However the two structure are diffeomorphic,
√
since the map F (t) = 3 t is a diffeomorphism. We invite the reader to check the details.
The considerations above lead us to the following questions
(1) Given a topological manifold M there exists a smooth structure on M ?
(2) Given a topological manifold M there are differentiable structures on M that are not diffeomorphic?
Both questions have a definitive answer in low dimension.
2.9. Theorem. A n-dimensional topological manifold admits a unique (up to diffeomorphism) differen-
tiable structure if n ≤ 3.
But the situation in higher dimensions is quite complicated and it is still object of active research. To
give a very approximative idea we will state some known facts and an important open question.
• There are very many n-dimensional topological manifolds, n ≥ 4, which do not admit differentiable
structures.
• Rn ha a unique differentiable structure (up to diffeomorphisms), if n 6= 4. R4 has infinitely many non
diffeomorphic differentiable structure, in fact an uncountable number.
• The number Γn of (non diffeomorphic) differentiable structures on S n is finite, if n 6= 4. Γn = 1 if
n = 5, 6, Γ7 = 28. It is still an open question if Γ4 = 1 (differentiable Poincarè conjecture).
2.10. Example. The real projective space RP n is defined as the quotient of Rn+1 \ {0} modulo the
equivalence relation x ∼ λx, λ ∈ R \ {0}. The topology is the quotient topology. It can be seen also as
the quotient of S n = {x ∈ Rn+1 : kxk = 1} modulo the equivalence relation x ∼ ±x. In fact the inclusion
S n −→ Rn+1 \ {0} induces a homeomorphism of the second model onto the first one. RP n is compact,
being the image, by the quotient map, of S n , which is compact. It is easily seen that RP n is Hausdorff and
second countable. We shall define a smooth atlas on RP n . We will denote by [x0 , . . . , xn ] the equivalence
class determined by (x0 , . . . , xn ) ∈ Rn+1 \ {0}.
Let Ai ⊆ RP n = {[x0 , . . . , xn ] ∈ RP n : xi 6= 0}. Observe that Ai is well defined sice if the ith coordinate
of a (non zero) vector is non zero, it is non zero for all equivalent vectors. Moreover Ai is open in RP n since
the inverse image via the quotient map π : Rn+1 \ {0} −→ RP n is open in Rn+1 \ {0}. Define the maps
φi : Rn −→ Ai , φi (y1 , . . . , yn ) = [y1 , . . . , yi−1 , 1, yi , . . . , yn ], i = 0, . . . , n.
Then φi is a bijection and its inverse is given by
φ−1 ([x0 , . . . , xn ]) = x−1

i (x0 , . . . , xi−1 , xi+1 , . . . , xn ).
The changes of coordinates are given by
φ−1 −1
j ◦ φi (y1 , . . . , yn ) = yj (y1 , . . . , yi−1 , 1, yi , . . . , yj−1 , yj+1 , . . . , yn ) (i < j),
hence they are smooth. Observe that RP n = ∪i Ai , so {(Rn , φi ), i = 0, . . . n} is a smooth atlas and determines
a smooth structure on RP n .
2.11. Remark. The numbers (x0 , . . . , xn ) are called the homogeneous coordinates of [x].
If we think at RP n as the quotient of S n , as above, we can use the parameterizations of S n as in Example

1.17, to produce an atlas for RP n . In fact we can define maps
X 1
ψi : B n (1) −→ RP n , ψi (y1 , . . . , yn ) = [y1 , . . . , yi−1 , (1 − yj2 ) 2 , yi , . . . , yn ].
j6=i
the inverse maps are

xi
ψi−1 ([x0 , . . . , xn ]) =
(x1 , . . . , xi−1 , xi+1 , . . . , xn ).
|xi |
The change of coordinates are the same as the ones in Example 1.17, hence smooth.
2.12. Example. We can consider the complex projective space, CP n , as the quotient of Cn+1 \ {0} by
the equivalence relation z ∼ λz, λ ∈ C \ {0}. This is equivalent to consider the unit sphere S 2n+1 = {z =
Pn
(z0 , . . . , zn ) ∈ Cn+1 : 0 |zi |2 = 1} and the equivalence relation z ∼ λz, λ ∈ S 1 = {w ∈ C : |w| = 1}.
In particular CP n is compact, as a quotient of a compact space, and, as it is easily seen, Hausdorff and
second countable. Proceeding, formally, as in the Example above, we can construct smooth atlases on CP n .
Thus CP n has, at least, one differentiable structure and with this structure it is a differentiable manifold of
dimension 2n.
2.13. Remark. The the construction of real and complex projective spaces relays on the fact that we
have an associative division algebra structure9 on R and in C ∼
= R2 . The existence of such a structure is
a stringent constrain on the dimension n. In fact it can be proved that such a product exist if and only if
n = 1, 2, 4, 8 (a non trivial result).
For n = 1 (resp. n = 2), a product is given by the usual product of real (resp. complex) numbers.
For n = 4 such a structure can be defined in the following way: let us denote by {1, i, j, k} the canonical
bases of R4 . Define 1 to be the neutral element for the product and
ij = k = −ji, jk = i = −kj, ki = j = −ik.
We can extend this product, by bi-linearity, to a product in R4 . With this product R4 is a division algebra,
associative but not commutative, called the algebra of quaternions, that we will denote by H. In analogy to
the case of complex numbers we can define conjugation: if q = x0 + x1 i + x2 j + x3 k, q = x0 − x1 i − x2 j − x3 k.
Then qq = kqk2 . For q ∈ H \ {0} we have q −1 = qkqk−2 .
For n = 8 the situation is a bit more complicated. We can define the algebra of octonions or the Caley
algebra as O = H ⊕ H, as a vector space, equipped with the product
(a, b) ∗ (c, d) = (ac − db, da − bc).

9An algebra A is a division algebra if for a ∈ A \ {0}, the maps L (x) = ax and R (x) = xa are surjective. If A has a
a a
(multiplicative) neutral element, this is equivalent to require that any nonzero element is invertible.
With this product O becomes a non commutative, non associative division algebra.
We can define the quaternionic projective space HP n as the quotient of Hn+1 \ {0} by the equivalence
relation x ∼ λx, λ ∈ H \ {0}. Proceeding formally as above we can construct a smooth structure on HP n
that, with this structure, is a differentiable manifold of dimension 4n.
We can try to define the Caley projective space OP n as the quotient of On+1 \ {0} by the relation
z ∼ λz, λ ∈ O \ {0}. It turns out that, for the lack of associativity, this relation is not an equivalence
relation. We can still define the Caley projective line (n = 1) as the 1-point compactification O, which is S 8 ,
in analogy to the cases of the real, complex and quaternionic projective lines. Also it is possible to define a
16-dimensional manifold which deserves the name of Caley projective plane, and it is denoted by OP 2 , but
not higher dimensional Caley projective spaces.
Once we have the concept of real valued differentiable functions, we have the concept of algebra of germs
of smooth functions at p ∈ M, Fp and we can use it to define the tangent space.
2.14. Definition. The tangent space of M at p is the space of derivations of Fp .
Also, given a smooth map between differentiable manifolds, F : M −→ N , we can define the differential
of F at p ∈ M as the linear map:
dF (p)(X)(f ) = X(f ◦ F ), ∀ X ∈ Tp M, f ∈ FF (p) .
Also we can define the tangent vector to a curve γ : (a, b) ⊆ R −→ M , at t ∈ (a, b), as the tangent vector
γ̇(t) that, on a function f ∈ F(M ), takes the value:
d
γ̇(t)(f ) = |s=t f (γ(s)).
ds
Naturally, if M is a submanifold of RN , γ̇(t) is the usual tangent vector to γ.
2.15. Remark. Starting with the observation above we will suggest, in Exercise 10.4, a more geometric
definition of Tp M .
2.16. Definition. The tangent bundle of a smooth manifold M is the the set
T M = {(p, V ) : p ∈ M, V ∈ Tp M }.
So the tangent bundle is, roughly speaking, the disjoint union of all tangent spaces. We want to put a
topology and a differentiable structure on T M . Consider a maximal smooth atlas (Uα , φα ) for M . Define
Φα : Uα × Rn −→ T M, Φα (x, X) = (φα (x), dφα (x)(X)).
It is easily seen that Φα is a bijection onto {(p, V ) ∈ T M : p ∈ φα (Uα )} ⊆ T M . We define a topology on

T M declaring open the sets that are union of images of open sets in Uα × Rn . With this topology the Φ0α s
are homeomorphisms onto their images (which are open). Therefore we have a topological atlas. The change
of coordinates are given by
Φ−1 −1 −1
β ◦ Φα (x, X) = (φβ ◦ φα (x), d[φβ ◦ φα ](x)(X)).
Hence the change of coordinates are smooth and the atlas define a differentiable structure on T M .
Also we define the projection map π : T M −→ M, π(p, V ) = p. It is obvious that π is smooth (with
the smooth structure on T M defined above).
3. Special atlas and partitions of unity
In this section we will prove the existence of partition of unity for smooth manifolds. For the case
M = U ⊆ Rn (see Appendix of Chapter 1) the essential fact in proving the existence of partitions of unity
was that we could write U as the union of an incrising sequence of compact sets. We will extend this fact to
the case of smooth manifolds and proceed as in the case of open sets of Rn . We start recalling a few concept
and facts in General Topology.
3.1. Definition. Lat X be a topological space and let U = {Uα , α ∈ A} be a covering of X.

(1) The covering U is locally finite if ∀ x ∈ X there exists a neighborhood U of x such that U ∩ Uα = ∅
for all but a finite number of the α0 s.
(2) A covering V = {Vi , i ∈ I} is a refinement of U if for each i ∈ I there exists α ∈ A such that
Vi ⊂ Uα .
(3) An Hausdorff space is paracompact if every open covering has a locally finite refinement.
3.2. Theorem. An Hausdorff space which is second countable and locally compact10 is paracompact.
Proof. We start by proving the following

Claim. There exists a sequence of compact sets Ki such that ∪i Ki = X and Ki ⊆ Int(Ki+1 ).11
Proof. Let {Ui , i ∈ N} be a countable bases for the open sets. Since X is locally compact, every point
x ∈ X has a relatively compact neighborhood U . Then there is Ui such that x ∈ Ui ⊆ U . Hence X is
covered by a sequence of relatively compact open sets and we can suppose, to start with, that Ui is compact.
Define K1 = U1 . We define Kk inductively. Given Ki compact, let k be the smaller integer such that
Ki ⊆ U1 ∪ · · · ∪ Uk . Define Ki+1 = U1 ∪ · · · ∪ Uk ∪ Ui+1 . Then the sequence Ki has the required property. ¤
We conclude now the proof of the Theorem. Let U = {Uα , α ∈ K} be an open covering of X. Cover
the compact set Ki+1 \ Int(Ki ) by open sets V1 , . . . , Vk such that Vj ⊆ Uα , for some α ∈ K and Vj ⊆
Ki+2 \ Int(Ki−1 ). Let Pi be the collection of such sets. Then {Pi } is a locally finite refinement of U. ¤
3.3. Remark. We should mention also a classical result due to Uryson.
3.4. Theorem. [Uryson’s Theorem] An Hausdorff, paracompact and second countable topological space
X is metrizable, i.e. there is a metric d on X such that the identity map 11X : (X, d) −→ X is a homeomor-
phisms.
In particular differentiable manifolds are metrizable spaces.
We will need special atlases.
3.5. Definition. A countable smooth atlas {(Rn , φi )} on a smooth manifold M is a special atlas if
(1) V = {φi (B n (3))} is a locally finite open covering of M ,
(2) W = {φi (B n (1))} is a (necessarily locally finite) covering of M .
10A topological space is locally compact if every point has an open neighborhood with compact closure.
11The interior of a subset Y ⊆ X, Int(Y ) is the union of all open sets contained in Y.
3. SPECIAL ATLAS AND PARTITIONS OF UNITY 94
For such an atlas we will use, consistently, the notations
Vi = φi (B n (3)), Ui = φ(B(2)), Wi = φi (B n (1)).
If U is a covering such that V refines U we will say that the special atlas is dominated by U.
Also, since the domain of the charts is fixed, we will denote a special atlas simply by {φi }.
3.6. Proposition. Let M be an n-dimensional smooth manifold and let U = {Uα } be an open covering
of M . Then there exists a special atlas such that V = {Vi } refines U.
Proof. We start with the following

Claim 1. Given p ∈ M there is a chart φ : Rn −→ M with φ(0) = p.
Proof. Let ψ : U −→ M be a chart such that ψ(x) = p. Let r ∈ R be positive and such that
B(x, r) ⊆ U . Consider a diffeomorphism F : Rn −→ B(x, r), F (0) = x (see Exercise 4.14 in Chapter 0).
Then φ = φ ◦ F : Rn −→ M is a chart with the desired property. ¤
Then we can proceed as in the proof of Theorem 3.2 and construct a special atlas such that V refines
U. We really need a little extra attention in order to guarantee that W is a covering of M . For this we need
the following fact
Claim 2. Given a chart φ : Rn −→ M and an opens set W̃ ⊇ W , the exist ² > 0 and a chart
φ̃ : Rn −→ M such that φ̃|B(1 + ²) = φ|B(1 + ²) and φ̃(B(3)) ⊆ W̃ .
We leave to the reader the tasks of proving the Claim and to conclude the proof of the Proposition. ¤
3.7. Definition. Let M be an n-dimensional smooth manifold and U = {Uα } an open covering of M .
A partition of unity dominated by U is a countable family of smooth functions λi : M −→ [0, 1] such that
(1) supp(λi ) := {x ∈ M : λi (x) 6= 0} ⊆ Uα for some α = α(i),
(2) ∀ x ∈ M there exists an open neighborhood U 3 x such that supp(λi ) ∩ U = ∅ for all but a finite
number of i’s.
P
(3) For all x ∈ M, i λi (x) = 1 (∀ x the sum is finite by (2)).
Our aim is to prove the following result
3.8. Theorem. Given an open covering U of M there exists a partition of unity dominated by U.
Proof. Let {φi } be a special atlas such that V refines U and λ : Rn −→ R a smooth function such that
λ(x) = 1 if kxk ≤ 1, λ(x) = 0 if kxk ≥ 2, λ(x) ∈ (0, 1) if kxk ∈ (1, 2)
(see Exercise 4.13, Chapter 1, for the existence of such a function). Define λ̃i : M −→ R,
(
λ ◦ φ−1
i (x) if x ∈ Vi
λ̃i (x) =
0 if x ∈ M \ Ui
P
Clearly λ̃i is a well defined smooth function. Set λ̃(x) = i λ̃i (x) (observe that there are only finitely
many non zero summands at every point). Then λ̃(x) is a well defined positive smooth function, since the
Wi ’s cover M , and {λi = λ̃i /λ̃} is a partition of unity dominated by U . ¤
The following approximation Theorem is proved just like in the case of open sets of Rn and we will leave
the details to the reader.
3.9. Theorem. Let M be a smooth manifold, F : M −→ Rm a continuous function and δ : M −→ R a

continuous positive function. Suppose that F is smooth on a closed set N ⊆ M . Then there exists a smooth
function G : M −→ R such that kF (x) − G(x)k < δ(x), ∀x ∈ U and F (x) = G(x) if x ∈ N . Moreover G can
be chosen homotopic to F .
3.10. Remark. Theorem 3.9 holds also if the target space is a differentiable manifold P , modulo in-
terpreting appropriately “approximation”. For example if P is a submanifold of some RN , we can consider
the induced metric space structure and define approximation in this context. In the next section we will see
that, in fact, any differentiable manifold is a submanifolds of some RN .
4. Immersions, embeddings and the Theorems of Whitney
In this sections we will prove some results, essentially due to Whitney, on the existence of immersions
and embeddings of smooth manifolds into Euclidean spaces. As a consequence we will see that any smooth
manifold is diffeomorphic to a submanifold of RN , for some N sufficiently large. We will follow closely [?].
We will start with some definitions.
4.1. Definition. Let M, N be smooth manifolds and F : M −→ N be a smooth map. We will say that
• F is an immersion if dF (x) : Tx M −→ TF (x) N is injective, ∀ x ∈ M ,
• F is a 1-1 immersion if it is an injective immersion,
• F is an embedding if F is an immersion and a homeomorphism onto its image (with the relative
topology).
4.2. Remark. A few observations and examples.

• By Theorem 1.26, an immersion is locally injective. In fact a little bit more: for any x ∈ M there
exists a neighborhood U of x, such that F |U : U −→ N is an embedding. But an immersion may
not be injective . For example the map F : R −→ R2 , F (t) = (cos t, sin t) in an immersion which
is not injective (see also Remark 1.2).
• An injective immersion may not be an embedding. However, if M is compact, an injective immersion
is an embedding (see Exercise 10.14).
• A diffeomorphism is a surjective embedding.
4.3. Definition. A submanifold of a smooth manifold N is the image of an embedding F : M −→ N .
4.4. Remark. Sometimes submanifolds are defined, roughly speaking, as images of injective immersion.
Such immersions are important because they appear naturally in many contexts, for example in the case
of integration of vector fields an distributions (see next section). To be coherent with our definition of
submanifolds of RN , we will stick with Definition 4.3.
A fact that will be useful in few occasions is the following
4.5. Lemma. Let F : M −→ N be an immersion and let N ⊆ M be a closed subset. Then, if F |N is a

1-1 immersion, there exists an open neighborhood U of N such that F |U is a 1-1 immersion.
4. IMMERSIONS, EMBEDDINGS AND THE THEOREMS OF WHITNEY 96
Proof. ¤
The first result we will prove in this section is that any function F : U ⊆ Rn −→ Rp can be “approximate”
by an immersion, if p ≥ 2n. We need a few preliminaries.
We will denote by C(a) ⊆ Rn a cube of side a and define the volume of C(a), v(C(x, a)) := an .
4.6. Definition. A set A ⊆ Rn has measure zero if, given ² > 0, there exists a countable family of
P n
cubes, C n (ai ), that cover A, with total volume ai < ². If this is the case, we will write m(A) = 0.
4.7. Remark. A few observations are in order.

(1) If B ⊂ A ⊂ Rn and m(A) = 0 then m(B) = 0.
(2) if m(A) = 0 then A does not contain open sets. In particular Rn \ A is dense.
(3) if Ai is a countable family of subsets with m(Ai ) = 0, then m(∪i Ai ) = 0.
(4) if A is a proper affine subspace of Rn , then m(A) = 0.
We leave to the reader the task of proving the claims above (Exercise 10.10).
4.8. Lemma. Let F : U ⊆ Rn −→ Rn be a smooth map and A ⊆ U with m(A) = 0. Then m(F (A)) = 0.
Proof. Let C = C(a) be a cube contained in U . Let b = sup{kdF (y)k : y ∈ C}. By the mean value
Theorem kF (x) − F (y)k ≤ bkx − yk if x, y ∈ C. In particular F (C) is contained in a cube of side ba.
P n
Let ² > 0 be given. Since m(A ∩ C) has measure 0, we can cover it with cubes C(ai ) with ai ≤ b−n ².
Then F (A ∩ C) can be covered with cubes of total volume less then ², hence F (A ∩ C) has measure zero.
Now F (A) is countable union of sets of the type F (A ∩ C) and, therefore, has measure zero. ¤
4.9. Corollary. If F : U ⊆ Rn −→ Rm be a smooth map, m > n, m(F (U )) = 0.
Proof. Consider G : U × Rm−n −→ Rm , G(x, y) = F (x). By Lemma 4.8 and Remark 4.7 (4),
0 = m(G(U × {0}) = m(F (U )). ¤
4.10. Definition. Let M be a smooth manifold and A ⊆ M be a subset contained in the image of a
local chart φ : U ⊆ Rn −→ M . We say that m(A) = 0 if m(φ−1 A) = 0.
A subset A ⊆ M has measure zero if for every chart φ : U −→ M, m(A ∩ φ(U )) = 0.
4.11. Remark. By Lemma 4.8, this definition does not depends on the choice of the local chart. Observe
also that we have defined sets of measure zero but not the “measure of a set”, and we can not really define
measure of a set, in our context, since any reasonable definition of “measure of a set” is hardly invariant for
diffeomorphisms. To define measures we need some extra structure.
4.12. Remark. The claims in Remark 4.7 continue to be true in this more general context (substituting
affine subspace in claim (4) by submanifold of lower dimension). Also Corollary 4.9 extends immediately to
the case of smooth functions between smooth manifolds. More precisely
4.13. Corollary. Let F : M n −→ N m be a smooth map with n < m. Then m(F (M )) = 0.
Observe that the condition that a manifold is a second countable topological space plays is essential.
We will prove now a “local version” of the Whitney immersion Theorem.

4.14. Lemma. Let U ⊆ Rn be an open set and F : U −→ Rp be a smooth map, p ≥ 2n. Then, given
² > 0, there exist A ∈ L(Rn , Rp ) with kAk < ² and such that the map G(x) = F (x) + Ax is an immersion.
Proof. We want to determine A such that G = F +A is an immersion. Consider the manifold M (p, n; k)
of Example 1.20 and the map
Fk : M (p, n; k) × U −→ M (p, n, R), Fk (Q, x) = Q − dF (x).
Claim: If k < n, the image of Fk has measure zero.
Proof. We will show that, if k < n, the dimension of M (p, n; k)×U is smaller than pn = dim(M (p, n, R))
and the conclusion will follow from Corollary 4.13. In fact, as we have seen in Example 1.20, dim(M (p, n; k)) =
k(p + n − k). Hence the domain of Fk has dimension k(p + n − k) + n. The latter expression is an increasing
function of k as long as k < n (take derivative with respect to k), hence it is, at most, 2n − p + pn − 1, as
long as k < n. But we are assuming p ≥ 2n, hence 2n − p + pn − 1 < pn = dim(M (p, n, R)). ¤
We conclude now the proof of the Theorem. By the Claim there exists A ∈ M (p, n, R) with kAk < ²
and not in the image of Fk , k < n. If dG(x) = A + dF (x) has rank k < n, then A = Fk (dG(x), x), a
contradiction. Hence G is an immersion. ¤
We want to prove a “global” version of Lemma 4.14.
4.15. Theorem. [Whitney immersion Theorem] Let M n be a smooth manifold and let F : M n −→ Rp ,
be a smooth function, p ≥ 2n. Given a positive continuous function δ : M −→ R, there exists an immersion
G : M −→ Rp such that kF (x) − G(x)k < δ(x). Moreover, if dF has rank n on a closed subset N ⊆ M , we
can choose G such that G|N = F |N .
Proof. The idea of the proof is to construct a sequence of functions, Fk : M −→ Rp , by subsequent

local modifications of F , using Lemma 4.14, in such a way that G = limk→∞ Fk is a well defined smooth
function with the required properties. Let go to work.
If rank dF = n in N , the same happens in an open neighborhood A of N . We consider the open
covering of M given by A and M \ N . Let {φi } be a special atlas dominated by the covering (see Definition
3.5 and Proposition 3.6, also for notations). Without loss of generality we can suppose that i ∈ Z and
Vi := φi (B(3)) ⊆ A if and only if i ≤ 0.
Let F0 = F . We will define, inductively, a sequence of functions Fk : M −→ Rp whose limit, as k → ∞,
will be a well defined function G with the desired properties.
Let ²j be the minimum of δ in Uj = φj (B(2)). Suppose we have defined Fk−1 such that
(1) dFk−1 (x) has rank n on Nk−1 = ∪j<k Wj
(2) kFk−1 (x) − Fk−2 (x)k < 21−k ²k−1 .
Let λ : Rn −→ [0, 1] be a smooth function which is identically 1 on B n (1), identically zero outside B n (2).
Claim. There exists A ∈ L(Rn , Rp ) such that the function
FA : B n (3) −→ Rp , FA (x) = Fk−1 ◦ φk (x) + λ(x)Ax
has the following properties

(1) FA is an immersion in K = φ−1 n
k (Nk−1 ) ∩ B (2),
4. IMMERSIONS, EMBEDDINGS AND THE THEOREMS OF WHITNEY 98
(2) kAxk ≤ 2−k ²k , ∀ x ∈ B n (3),

(3) FA is an immersion on B n (1).
Proof. Consider the function
Φ : K × M (p, n, R) −→ M (p, n, R), Φ(x, A) = dFA (x).
Observe that Φ(K × {0}) ⊆ M (p, n; n). Since M (p, n; n) is open in M (p, n, R), dFA (x) ∈ M (p, n; n), ∀x ∈
K, if kAk is sufficiently small. So there exist η > 0 such that, if kAk < η, FA verifies the first two
conditions. Lemma 4.14 guarantee that there exist A ∈ M (p, n, R) such that kAk < η and FA satisfies the
last condition. ¤
Fix A ∈ L(Rn , Rp ) that verifies the condition of the claim and define
(
FA ◦ φ−1
k (p) if p ∈ Vk
Fk (p) =
Fk−1 (p) if p ∈ M \ Uk
Observe that Fk is a well defined smooth map since the two functions above are smooth and coincide on the
intersection of the domains. Moreover
• dFk has rank n on Nk−1 (by the first condition on A).
• kFk (p) − Fk−1 (p)k < 2−k δ(p) ∀ p ∈ M (by the second condition on A).
• dFk has rank n on Wk (by the third condition on A), hence on Nk .
Define G(p) = limk→∞ Fk (p). Observe that, due to the fact that the covering {Vj } is locally finite, there
exist a neighborhood Up of p such that G|Up = Fk |Up for k sufficiently large. So G is a well defined smooth
function and has the required properties. ¤
A natural question to ask is whether we can approximate an immersion by a 1-1 immersion. For example,
if we have a regular curve in R2 with a self intersections point, say p, we can not, in general, approximate
it by a regular curve without self intersections. But if we look at the picture as being in R3 we can “lift” a
branch of the curve near p to obtain a new curve in R3 , which is an approximation of the given one, without
self intersections. This is, essentially, the content of our next result.
4.16. Theorem. [Whitney injective immersions Theorem] Let M n be a differentiable manifold, F :

M −→ Rp an immersion and δ : M −→ R a positive continuous function. If p > 2n, there is a 1-1
immersion G : M −→ Rp such that kF (p) − G(p)k < δ(p), ∀p ∈ M . Moreover, if N ⊆ M is a closed set
such that F |N is 1-1, we can chose G such that G|N = F |N .
Proof. Since F |N is 1-1, there exist, by Lemma 4.5, an open neighborhood A of N such that F |A is
1-1. Since F is a local embedding there exist an open covering {Vα } of M , that refines {A, M \ N }, such that
F |Vα is an embedding. Let {φi } be a special atlas that refines this covering. We will use the same notations
as in the proof of Theorem 4.15. Define λk : M −→ R as the obvious extension of λ ◦ φ−1
k : Vk −→ R.
Let F0 = F . We define, inductively, the sequence of functions
Fk (p) = Fk−1 (p) + λk (p)bk ,

where bk ∈ Rp has to be chosen. To explain this choice we start considering the set Pk = {(p, q) ∈ M × M :
λk (p) 6= λk (q)}. Then Pk is an open subset of M × M hence a 2n-dimensional manifold. Consider the map
Fk−1 (p) − Fk−1 (q)
Φk : Pk −→ Rp , Φk (p, q) = − .
λk (p) − λk (q)
By Lemma 4.13, the image of Φk has measure zero in Rp (since p > 2n).
Therefore we can choose bk in such a way that
• Fk is an immersion,
• Fk is a δ/2k approximation of Fk−1 ,
• bk is not in the image of Φk .
It follows that Fk (p) = Fk (q) if and only if λk (p) = λk (q) and Fk−1 (p) = Fk−1 (q). By a backward
induction, the same hold for any h, 0 < h < k.
Define G(p) = limk→∞ Fk (p). As in the proof of the preceding Theorem, G is a well defined immersion
and G|N = F |N . We want to show that G is 1-1. Suppose G(p) = G(q), p 6= q. Then, for some k sufficiently
large, Fk (p) = Fk (q). Then F (p) = F (q) and p and q can not be in the same Ui . Since φk (p) = φk (q), p, q
do not belong to Uk , for k > 0. Therefore p, q ∈ U , a contradiction. ¤
The next natural question is if we can approximate an immersion by an embedding. In order to answer
this question we need some preliminary ingredients.
4.17. Definition. Let F : M −→ Rp be a continuous map. The limit set of F , L(F ), is the set of points
y ∈ Rp such that y = limk→∞ F (xk ) where {xk } ⊆ M is a sequence without limit points.
We will prove now the basic properties of the limit set that we will need in the proof of the Whitney
embedding Theorem.
4.18. Proposition. Let F : M −→ Rp be a continuous map. Then

(1) F (M ) is closed in Rp if and only if L(F ) ⊆ F (M ).
(2) F is a homeomorphism onto its image if and only if it is injective and L(F ) ∩ F (M ) = ∅.
(3) If L(F ) = ∅ and G : M −→ Rp is a continuous function such that kF (x)−F (x)k ≤ K ∀ x ∈ M, K
a constant, then L(G) = ∅.
Proof. Observe that L(F ) ⊆ F (M ). Hence, if F (M ) is closed, L(F ) ⊆ F (M ). Conversely, suppose

L(F ) ⊆ F (M ) and let y ∈ F (M ). Then y = limk→∞ F (xk ). If the sequence has no limit points, y ∈ L(F ) ⊂
F (M ). Suppose there is a subsequence xkj converging to x ∈ M . Then y = F (x) ∈ F (M ). Therefore F (M )
is closed.
Suppose now that F is a bijection onto its image. Let {xn } be a sequence in M with y = limn→∞ f (xn ).
If y ∈ F (M ) and F −1 is continuous, the sequence {xn } converges to F −1 (y) hence y 6∈ L(F ). Conversely
if y ∈ F (M ) and I and L(F ) ∩ F (M ) = ∅, then any subsequence of {xn } converges and the limit points
coincide, since F is injective. This imply that F −1 is continuous.
Let y ∈ L(G). Then there exist a sequence {xn } in M , without limit points, such that G(xn ) converges
to y. Consider {F (xn )}. This is a bounded sequence in Rp , hence has a convergent subsequence and the
limit is in L(F ), a contradiction. Hence L(G) = ∅. ¤
5. INTEGRATION OF VECTOR FIELDS AN DISTRIBUTIONS 100
4.19. Theorem. [Whitney closed embedding Theorem] Let M be a n-dimensional differentiable mani-
fold. Then there exists an embedding G : M −→ R2n+1 such that G(M ) is closed.
Proof. We start with the following general fact

Claim. There exists a smooth function F : M −→ R with L(F ) = ∅.
Proof. Let {φi }, i ∈ N, be a special atlas and {λi } a partition of unity dominated by this atlas. Define
P
F (x) = i iλi (x). We claim that L(F ) = ∅. Set Ki = ∪j≤i Wj . If {xn } is a sequence without limit points,
∀i there exist xn(i) 6∈ Ki , since Ki is compact. In particular F (xni ) is unbounded and F (xn ) does not
converges. ¤
Consider now the function F : M −→ R ⊆ R2n+1 , F as in the Claim. By Theorem 4.16 there exist a
1-1 immersion G : M −→ R2n+1 such that kG(x) − F (x)k < 1. Then L(G) = ∅, by item (3) of Proposition
4.18. By item (2) of the same Proposition, G is an embedding, and by item (1) G(M ) is closed. ¤
4.20. Remark. We have proved that every n-dimensional manifold admit an immersion in R2n and
an embedding in R2n+1 . Those results are not sharp. In fact it is possible to prove that there are always
immersions of an n-dimensional manifold into R2n−1 and embeddings into R2n . These results, known as the
strong Whitney Theorems, are sharp. For example the real projective space RP n does not immerse into
R2n−2 and does not embed into R2n−1 , if n = 2k . However a question arise naturally: given n the exist an
integer p(n) such that every n-dimensional manifold can be immersed into Rp(n) and there are n-dimensional
manifolds that can not be immersed in Rp(n)−1 ? As remarked above, p(n) = 2n − 1 if n is “very even”, i.e.
n = 2k , so there is a suspect that p(n) depends on “haw even” is n. Let be a bit more precise. Let α(n) be
the number of 1’s in the in the binary expression of n, i.e. if we write n = 2i1 + . . . + 2il , 0 ≤ i1 < . . . < il ,
then α(n) = l. Observe that α(2k ) = 1. Using algebraic topological methods (characteristic classes) it can be
seen that p(n) ≥ 2n − α(n) and it was conjectured, for some times, that p(n) = 2n − α(n). In 1985 R. Cohen
proved that a compact n-dimensional manifold can be immersed in R2n−α(n) , giving so a positive answer to
the conjecture, at least in the compact case. For the embedding problem it can be shown that a compact
orientable12 n-dimensional manifold embeds into R2n−1 (observe the even dimensional real projective spaces
are not orientable).
5. Integration of vector fields an distributions
5.1. Definition. Let M be a smooth manifold and U ⊆ M be an open set. A (tangent) vector field on
U is a smooth map X : U −→ T M such that π ◦ X = 11U . The space of vector fields on U will be denoted
by H(U ).
5.2. Remark. A tangent field is essentially a function that associate to p ∈ U a derivation of the
algebra Fp . It is easy to see that such a vector field define a derivation of the algebra of smooth real valued
function defined in U i.e. an element of Der(U ) and, conversely, an element of Der(U ) defines a vector
field. So we can identify H(U ) with Der(U ). Therefore we can define the Lie product of vector fields, as the
commutator of derivations, Therefore define a Lie algebra structure on H(U ), and prove, using local charts,
12We will define orientability later on.

that if F : M −→ N is a smooth map, the Lie product of F -related vector fields is F -related to the Lie
product of the vector fields (see Proposition ?? of Chapter 0).
5.3. Definition. Let X ∈ H(M ) and p ∈ M . An integral curve of X with initial condition p is a smooth
curve γ : (a, b) ⊆ R −→ M such that 0 ∈ (a, b), γ(0) = p and γ̇(t) = X(γ(t)).
Using local charts and Theorem ?? of Chapter 0, we have
5.4. Theorem. For all p ∈ M there exists an integral curve γ of X with initial condition p and such a
curve is unique in the sense that two integral curves with the same initial condition agree on the intersection
of their domains. In particular there is a unique maximal integral curve with initial condition p and this
curve depends smoothly on p.
Then we can define complete vector fields as these vector fields whose integral curves are defined in the
whole of R and prove that vector fields with compact support are complete, just like in the case of vector
fields on open sets of Rn . For complete vector fields we have the flow map
Γ : M × R −→ M, Γ(p, t) = γp (t),
where γp is the integral curve of X with initial condition p.
5.5. Remark. It is now a good time to spend a word on the behavior of the integral curves of a vector
field. Let X ∈ H(M ) and assume, for simplicity, that X is complete. If X(p) = 0 then, by unicity, the
integral curve of X with initial condition p is the constant curve γ(t) = p. Suppose now X(p) 6= 0 and let
γ be the integral curve of X with initial condition p. Then, again by unicity, γ̇(t) 6= 0, ∀ t ∈ R. Therefore
γ is an immersion. Then we have two possibilities
• γ(t) 6= γ(s), s 6= t, and γ is a 1-1 immersion of R into M (it may not be an embedding),
• γ(t) = γ(s) for some s > t. It will be not restrictive to assume t = 0, γ(r) 6= γ(0), ∀r ∈ (0, s). In
this case, again by unicity, γ̇(s) = γ̇(0) and γ is a periodic curve if period s. In particular γ is a
closed curve and induces an embedding γ̃ : S 1 −→ M .
Next we consider next the case of distributions.
5.6. Definition. Let M be a n-dimensional manifold. A smooth k-dimensional distribution D on M is

the choice of a k-dimensional subspace Dp ⊆ Tp M, ∀ p ∈ M , that depends smoothly on p. This means that,
∀ p ∈ M , there are smooth vector fields {X1 , . . . , Xk }, defined in a neighborhood U of p, such that Dq is
spanned by {X1 (q), . . . , Xk (q)}, ∀ q ∈ U .
A distribution is a subset of T M , hence we ca talk about vector field (with values) in the distribution.
5.7. Definition.
(1) The distribution D is involutive if the Lie product of two vector fields in the distribution is a vector
field in the distribution.
(2) An integral manifold (N, φ) for the distribution D is a k-dimensional smooth manifold N and a 1-1
immersion φ : N −→ M such that dφ(x)(Tx N ) = Dφ(x) .
6. LIE GROUPS 102
(3) A maximal integral manifold for D is an integral manifold (N, φ) such that N is connected and
φ(N ) is not properly contained in the image of any connected integral manifold of D.
As we have seen in Chapter 0, Theorem ??, a necessary and sufficient condition for the existence of
integral manifolds for D is D to be involutive. The delicate question is the existence of maximal integral
manifolds. This is the content of the next result.
5.8. Theorem. [Frobenius Theorem, global version] Let D be an involutive k-dimensional distribution
on a manifold M and p0 ∈ M . Then the exist a unique maximal integral manifold of D containing p0 .
Proof. We start by defining the candidate to be the integral manifold we are looking for. Consider
N = {p ∈ M : there exist a piecewise smooth curve γ : [0, 1] −→ M with γ(0) = p0 , γ(1) = p, γ̇(t) ∈ Dγ(t) }.
From Theorem ?? in Chapter 0 and the fact that M is second countable it follows that the exists a
countable atlas
φi : U = {(x1 , . . . , xn ) ∈ Rn : |xi | < 1} −→ M,
such that the images of the slices xk+j = constant are integral manifolds of D.
We want to define a topology on N . We do this by defining a system of open neighborhoods for p ∈ N .
Without loss of generality we can suppose p = φi (0). Observe that φ(U ∩ Rk ) ⊆ N and we declare open
neighborhoods of p in N , the images of open sets in U ∩ Rk containing 0. In particular φi |U ∩Rk is a smooth
atlas for N . This topology is Hausdorff since two points in distinct slices have disjoint open neighborhoods.
We observe also that a piecewise smooth curve γ, joining p0 with p and with tangent vector in the distribution,
is contained in N and is a continuous curve in this topology. Therefore N is connected. The delicate point
in the proof is the following
Claim The topology on N defined above is second countable.
Proof. ¤
6. Lie groups
A very important class of differentiable manifolds is the class of Lie group that we will study in this
section. These are object with two structure: a structure of differentiable manifold and a structure of group.
We require that the two structures are compatible, in the sense that the algebraic operations are smooth
maps. More precisely
6.1. Definition. A Lie group is a smooth manifold G with a preferred element e ∈ G and smooth maps:
m : G × G −→ G, r : G −→ G,
such that the axioms of groups are verified, i.e.

• m(g, m(h, k)) = m(m(g, h), k),
• m(e, g) = g = m(g, e),
• m(g, r(g)) = e = m(r(g), g).
As usual in group theory we will use the notation gh for m(g, h) and g −1 for r(g). In the case of Abelian
groups, i.e. m(g, h) = m(h, g) ∀ g, h ∈ G, we will also use the notation m(g, h) := g + h, r(g) = −g.
6.2. Definition. Let G be a Lie group. We define the left translation by g ∈ G as the map Lg : G −→
G, Lg (x) = gx. Similarly the right translation by g is the map Rg (x) = xg.
6.3. Lemma. Lg and Rg are diffeomorphisms of G.
Proof. Lg is the composition of m, which is smooth, with the smooth map (cg , 11) : G −→ G × G, cg
being the constant map cg (h) = g. Hence Lg is smooth. Also (Lg )−1 = Lg−1 so Lg has a smooth inverse.
Analogously for Rg . ¤
6.4. Definition. A vector field X ∈ H(G) is said to be left invariant if ∀ g, h ∈ G, X(gh) = dLg (X(h)).
Analogously, X is right invariant if ∀ g, h ∈ G, X(gh) = dRh (X(g)).
We will denote by L(G) the space of left invariant vector fields. Obviously L(G) is a (real) vector subspace
of H(G).
Since a left invariant vector field X is Lg -related to itself, we have
6.5. Lemma. If X, Y ∈ L(G), [X, Y ] ∈ L(G). In particular LG is a Lie subalgebra of H(G).
6.6. Lemma. L(G) is canonically isomorphic to Te G, as a vector space.
Proof. Consider the evaluation map ev : L(G) −→ Te G, ev(X) = X(e). Clearly ev is linear and
−1
has a linear inverse ev (Xe )(g) = dLg (e)(Xe ). We should only check that dLg (e)(Xe ) is a smooth vector
field. For this we consider a smooth function f : G −→ R and a smooth curve γ : (−², ²) −→ G such that
γ(0) = e, γ̇(0) = Xe . Then
d
dLg (e)(Xe )(f ) = Xe (f ◦ Lg ) = |t=0 f (gγ(t)),
dt
which is smooth in g. Therefore dLg (e)(Xe ) is a derivation of F(G), hence a smooth vector field. ¤
6.7. Definition. Te G, endowed with the (Lie) product induced by ev, is a Lie algebra, called the Lie
b
algebra of G, and will be denoted by G.
6.8. Remark. The Lie algebra of a Lie group is a basic invariant. In fact it is known that there exist
a bijection between isomorphism classes of (finite dimensional real) Lie algebras and simply connected Lie
groups, up to smooth isomorphisms. We will comment a bit more on this fact later on.
We will study now the integral curves of a left invariant vector field .
6.9. Proposition. Let X ∈ L(G) be a left invariant vector field and γ : (a, b) −→ G be an integral curve
of X with initial condition e ∈ G.
(1) The curve γg = Lg ◦ γ is an integral curve of X with initial condition g ∈ G.
(2) If t, s, t + s ∈ (a, b), γ(t + s) = γ(t)γ(s).
(3) X is complete, i.e. the integral curves are defined on all R.
6. LIE GROUPS 104
d
Proof. Since X is left invariant, Lg ◦ γ = dLg (γ̇(t)) = X, hence Lg ◦ γ is an integral curve of X.
dt
The second Claim follows by observing that, for a fixed s, γ(t + s) and γ(t)γ(s) are integral curves of X
with the same initial condition. We will prove now that X is complete.
Let t0 = sup{t ∈ R : γ is defined in [0, t)}, and suppose, by absurd, that t0 < ∞. Let ² be a small
positive constant and consider the curve
(
γ(t), if t ≤ t0 − ²
γ(t) =
γ(t0 − ²)γ(t − t0 + ²) if t0 − ² ≤ t < 2t0 − ²
Then γ is an integral curve of X defined in [0, 2t0 − ²) ) [0, t0 ), a contradiction. ¤
6.10. Example. Consider GL(n, R) = {A ∈ M (n, R) : det(A) 6= 0}. Then GL(n, R) is an open subset of
∼ Rn2 , hence a differentiable manifold. The product and inversion are smooth maps (see Example
M (n, R) =
??, Chapter 1). Hence GL(n, R) is a n2 -dimensional Lie group. Here some simple observations.
(1) If A ∈ GL(n, R), LA (B) = AB. In particular LA is linear and dLA (C) = LA , ∀C ∈ GL(n, R).
(2) If A ∈ T11 GL(n, R) ∼
= M (n, R), the left invariant field (uniquely) determined by A is Ã(C) = CA.
(3) The integral curve of the left invariant extension Ã of A, with initial condition 11, is γ Ã (t) =
P∞
exp(tA) := k=0 [k!]−1 tk Ak . Therefore the integral curve with initial condition C is C exp(tA).
(4) The flow of Ã is φÃ
t (B) = B exp(tA).
\R) ∼
We can then compute the Lie algebra structure on GL(n, = T11 GL(n, R) ∼
= M (n, R) using Proposition
∼ M (n, R),
?? of Chapter 1. If A, B ∈ T11 GL(n, R) =
d
[A, B] = [Ã, B̃](11) = |t=0 exp(tA)B exp(−tA) =
dt
= [A exp(tA)B exp(−tA) − exp(tA)BA exp(−tA)]|t=0 = AB − BA.
An observation on the topology of GL(n, R). Since det : GL(n, R) −→ R \ {0} is a continuous surjective
map, GL(n, R) is not connected. Consider GL+ (n, R) = {A ∈ GL(n, R) : det A > 0} and GL− (n, R) =
{A ∈ GL(n, R) : det A < 0}. These are disjoint open submanifolds of GL(n, R), and we claim that they
are connected. Let Ri be the diagonal matrix with entries aj = 1, j 6= i, ai = −1. Multiplication by Ri is
a diffeomorphism of GL+ (n, R) onto GL− (n, R). So it is sufficient to show that GL+ (n, R) is connected.
Let A ∈ GL+ (n, R). We want to show that A belongs to the connected component of 11 in GL+ (n, R). By
basic linear algebra we know that if A ∈ GL(n, R) there is an invertible matrix P such that P −1 AP is upper
triangular. Conjugation by P sends GL+ (n, R) into itself (and fix 11), so we can suppose that A is upper
triangular. Multiplying by t ∈ [0, 1] the non diagonal elements we get a path joining a diagonal matrix to A,
and, along this path, the determinant is constant. So we can suppose that A is diagonal. By convex linear
combination we can find a path in GL+ (n, R) between A and a diagonal matrix with entries ±1. The matrix
Ri Rj is in the connected component of 11, since the rotation of an angle θ ∈ [0, π] in the plane {ei , ej } is a
curve in GL+ (n, R) between 11 and Ri Rj . So multiplication by Ri Rj sends the connected component of 11
into itself. Now we multiply A by the Ri ’s, for i corresponding to the negative eigenvalues. Since there are
an even number of such Ri ’s, this map sends the connected component of 11 into itself and A to 11. Hence A
belongs to the connected component of 11 and this component coincides with GL+ (n, R).
Once we introduce “objects” (Lie groups), it is good strategy to introduce “sub objects” and morphisms,
i.e. maps that preserve the structure.
6.11. Definition. Let G, H be Lie groups.

• A Lie groups homomorphism φ : G −→ H is a smooth map which is a group homomorphism. If φ
has an inverse, which is a Lie groups homomorphism, we will say that φ is a Lie groups isomorphism.
• A Lie subgroup of H is a pair (G, φ) where G is a Lie group and φ : G −→ H is a group homomor-
phism which is a 1-1 immersion.
We will prove some results on uniqueness of subgroups that verify certain condition. So we will make
clear what we mean that two subgroups are “equal”.
6.12. Definition. Let H be a Lie group and (Gi , φi ), i = 1, 2 be two subgroup. We will say that two
subgroups are equivalent if there exists a Lie group isomorphism ψ : G1 −→ G2 such that φ2 ◦ ψ = φ1 .
b −→ H
6.13. Proposition. If φ : G −→ H is a Lie group homomorphism, then dφ(e) : G b is a Lie
algebras homomorphism.
b and X̃ be the left invariant field in H with X̃(e) = dφ(e)X(e) (we are denoting by
Proof. Let X ∈ G
e the identity element both in G and in H). Since φ is a homomorphism, Lφ(g) ◦ φ = φ ◦ Lg . Hence
X̃(φ(g)) = dLφ(g) X̃(e) = dLφ(g) dφ(e)(X(e)) =
= d(Lφ(g) ◦ φ)(e)(X(e)) = d(φ ◦ Lg )(e)(X(e)) = dφ(g)(X(g)).

b
Hence X and X̃ are φ-related. In particular, if X, Y ∈ G,
^
[X, Y ](e) = [X̃, Ỹ ](e) = dφ(e)([X, Y ]).
b ⊆ G,
A Lie subgroup φ : H −→ G, determines a Lie subalgebra dφ(e)(H) b by Proposition 6.13. We will
prove that, conversely, a Lie subalgebra determines a Lie subgroup. We will start with a preliminary result.
6.14. Lemma. Let G be a connected Lie group and U an open neighborhood of e ∈ G. Then G = ∪∞ n
n=1 U ,
where U n is the set of products on n elements in U .
Proof. Consider U −1 := {g −1 : g ∈ U }. Then U −1 is also an open neighborhood of e and so is

V := U ∩ U −1 . We will show that G = ∪∞ n ∞ n ∞ n
i=1 V ⊂ ∪n=1 U . Set H = ∪i=1 V . Then
• H is an abstract subgroup (if g, h ∈ H, gh−1 ∈ H).
• H is open (if g ∈ H, gV ⊆ H).
Now G is covered by the cosets {gH}, g ∈ G. Since H is open these cosets are open. In particular G\H is
an open set, since is union of the cosets of H, different from H. Since G is connected and H 6= ∅, G = H. ¤
6.15. Theorem. Let G be a Lie group and let H e ⊂ G

b be a subalgebra. Then there exists a unique
b = H.
connected subgroup (H, φ) of G such that dφ(e)(H) e
6. LIE GROUPS 106
Proof. Consider the distribution

e
D(g) = dLg (e)(H).
e is a subalgebra, it follows easily that D is involutive. Let (H, φ) a maximal integral manifold of
Since H
D containing e. Set e = φ−1 (e).
Claim 1: φ(H) is an abstract subgroup of G.
Proof. Let g ∈ φ(H). Consider the map Lg−1 ◦ φ. This map is an injective immersion whose image
contains e. Moreover
e = D(g −1 ),
dLg−1 ◦ φ(e)(Te P ) = dLg−1 (H)
so it is also an integral manifold of D. By maximality Lg−1 ◦φ(H) ⊆ φ(H). Hence, given g, h ∈ φ(H), g −1 h ∈
φ(H) and the Claim is proved. ¤
Claim 2: H is a Lie group.
Proof. We can define an abstract group structure on H by requiring that φ : H −→ φ(H) to be

an abstract groups isomorphism. Since H is a smooth manifold, we have only to check that the map
β : H × H −→ H, β(g, h) := g −1 h, is smooth. This follows from Lemma ??. ¤
To conclude the proof of the Theorem we need to prove unicity (in the sense of Definition 6.12). Suppose
that (K, ψ) is an other such subgroup. Then (K, ψ) must be an other integral manifold of D, with e ∈ ψ(K).
Since (H, φ) is maximal, ψ(K) ⊆ φ(K). From Lemma ?? oh Chapter 2, it follows that we have a well defined
smooth map κ : K −→ H such that φ ◦ κ = φ. It follows that κ is an injective Lie groups homomorphism.
We want to show that κ is onto. Now dκ(e) is an isomorphism an so, by the inverse function Theorem it
maps a neighborhood of e ∈ H diffeomorphically onto a neighborhood of e ∈ H. Now, by Lemma 6.14, both
K and H are generated by such neighborhoods, a fact that clearly implies that κ is onto. ¤
6.16. Corollary. Let G be a Lie group. Then there is a 1-1 correspondence between connected Lie
b
subgroups of G and Lie subalgebras of G.
A particularly interesting case is the following
6.17. Definition. A 1-parameter subgroup of a Lie group G is a 1-dimensional Lie subgroup.
b and X̃ be the left invariant extension of X. By Proposition 6.9 X̃ is complete hence the
Let X ∈ G
integral curve of X̃ thought e ∈ G is a smooth map γX : R −→ G which, again by Proposition 6.9, it is
an abstract group homomorphism. Moreover, if X 6= 0, X̃ is everywhere nonzero and γ is an immersion.
Now the integral curves of a nonzero complete vector field are either injective or periodic (see Remark ??
of Chapter 2). In the first case (R, γX ) is a Lie subgroup. In the second case, γX induces an embedding
γ X : S 1 −→ G. Let p be the period of γX . We can define a Lie group structure on S 1 , identifying it with
R/pZ and, with this structure, (S 1 , γ X ) is a Lie subgroup of G.
6.18. Definition. Let G be a Lie group. We define the exponential map :
exp : Te −→ G, exp(X) = γX (1),
where γX is the integral curve of the left invariant extension of X with γX (0) = e.
6.19. Lemma. exp(tX) = γX (t).
Proof. Fix t an define σ(s) = γX (st). Then σ is an integral curve of tX (which is left invariant) with
σ(0) = e. Hence exp(tX) = σ(1) = γ(t). ¤
6.20. Lemma. The exponential map is smooth and d exp(0) : T0 Te G ∼

= T0 G −→ T0 G is the identity map.
Proof. d exp(0)(X) is the tangent vector, at e ∈ G, of the curve exp(tX) = γX (t). Hence the conclusion.
¤
In particular exp is a diffeomorphism of a neighborhood of 0 ∈ Te G onto a neighborhood of e ∈ G.

Composition of this chart with left translations give an atlas for G.
A useful property of the exponential map is that it is natural with respect to Lie groups homomorphisms
6.21. Proposition. Let φ : G −→ H be a Lie groups homomorphism. Then φ ◦ exp = exp ◦dφ(e).
b Then γ(t) = φ(exp(tX)) is a smooth curve in H with γ̇(0) = dφ(e)(X). Since φ is

Proof. Let X ∈ G.
a homomorphism, γ is also a 1-parameter subgroup, hence φ(exp(tX)) = exp(tdφ(e)(X)). ¤
Proposition 6.21 allows us to prove that, under natural conditions, an abstract subgroup of a Lie group
is, in fact, a Lie subgroup.
6.22. Theorem. Let G be a Lie group and H ⊆ G be an (abstract) subgroup which is a closed subset.
Then there exists a unique manifold structure in H such that the inclusion i : H −→ G is a Lie subgroup.
Proof. ¤
A useful consequence is the following
6.23. Proposition. Let ψ : G −→ H be a Lie groups homomorphism. Then ker ψ is a Lie subgroup of
G with Lie algebra ker dψ(e).
Proof. Set K = ker ψ. Then K is a closed subgroup, hence a Lie subgroup, by Theorem 6.22. Now
X∈Kb if and only if exp(tX) ∈ K, by Proposition 6.21, and this occurs if and only if X ∈ ker dψ(e). ¤
7. The adjoint representation
Let E be a (finite dimensional) vector space. We will denote by End(E) the space of linear maps of E
into itself and by Aut(E) the subspace of the invertible ones. Once fixed a bases of E we will identify such
spaces with M (n, R) and GL(n, R) respectively.
7.1. Definition. Let G be a Lie group. A representation of G in E is a Lie groups homomorphism

ρ : G −→ Aut(E).
We will introduce now a very important representation. Consider a Lie group G and the map κg : G −→
G, κg (h) = ghg −1 . Clearly κg (e) = e ∀ g ∈ G.
b
7.2. Definition. The adjoint representation of G is the representation Ad : G −→ Aut(G), Ad(g) =
dκg (e).
7. THE ADJOINT REPRESENTATION 108
We also define the map

b −→ End(G)
ad : G b = T Aut(G),
b ad(X) = dAd(e).
11
By Proposition 6.21 we have the commutative diagrams
Ad κg
G −→ Aut(G) G −→ G
exp ↑ ↑ exp exp ↑ ↑ exp
ad Ad(g)
b
G b
−→ End(G) Gb −→ b
G
In particular
exp(tAd(g)X) = g exp(tX)g −1 .
In the case G = GL(n, R) we have
d d
Ad(X)(Y ) = |t=0 X exp(tY )X −1 = |t=0 exp(tXY X −1 ) = XY X −1 ,
dt dt
d d
ad(X)(Y ) = |t=0 Ad(11 + tX)(Y ) = |t=0 (11 + tX)Y (11 + tX)−1 = XY − Y X = [X, Y ].
dt dt
The last formula is a general fact.
b ad(X)(Y ) = [X, Y ].
7.3. Proposition. If G is a Lie group, X, Y ∈ G,
Proof. First observe that

d
ad(X)(Y ) = dAd(e)(X)(Y ) = |t=0 Ad(exp(tX))(Y ) =
dt
d d
= |t=0 Ad(exp(tX))(Y ) = |t=0 d(κexp(tX) )(Y ).
dt dt
We extend X, Y to left invariant fields X̃, Ỹ and denote by φt is the flux of X̃. Then
d d
ad(X)(Y ) = |t=0 dRexp(−tX) ◦ dLexp(tX) (Y ) = |t=0 dRexp(−tX) Ỹ (exp(tX)) =
dt dt
d
= |t=0 dφ−t (Ỹ (φt (e)) = [X, Y ].
dt
¤
b be a Lie algebra, We define
We will study now some special subgroups. Let G be a group and let G
• the center of G, Z(G) = {g ∈ G : gh = hg, ∀ h ∈ G},
b Z(G)
• the center of G, b = {X ∈ G b : [X, Y ] = 0, ∀ Y ∈ G}.
b
b is an Abelian Lie algebra if
G is Abelian if and only if Z(G) = G. By analogy we will say that G
b b
Z(G) = G.
If G is a Lie group, Z(G) is a closed subgroup, hence, by Theorem 6.22 a Lie subgroup.
[
b = Z(G).
7.4. Proposition. If G is a connected Lie group, ker Ad = Z(G) and Z(G)
Proof. Let g ∈ Z(G). Then exp(tX) = g exp(tX)g −1 = exp(tAd(g)(X)). If t is sufficiently small,

b where exp is a diffeomorphism onto its image. Hence
tX, tAd(g)(X) belong to a neighborhood of 0 ∈ G
tX = Ad(g)(X), i,e g ∈ ker Ad. Conversely, suppose g ∈ KerAd. Then exp(X) = exp(Ad(g)(X)) =
g exp(H)g −1 . In particular g commutes with elements in the image of exp. But the image of exp contain an
open neighborhood U of e. Since G is connected, any element h ∈ G is product of elements in U , by Lemma

b
6.14, and g ∈ Z(G). The last Claim follows from Proposition 6.23 applied to Ad : G −→ Aut(G). ¤
7.5. Corollary. A connected Lie group is Abelian if and only is its Lie algebra is Abelian.
b are commuting tangent vectors (i.e. [X, Y ] = 0),

7.6. Corollary. If G is a Lie group and X, Y ∈ G
then exp(X + Y ) = exp(X) exp(Y ).
b = span{X, Y }. Then H
Proof. Consider H b is an Abelian subalgebra of G
b and the associated subgroup
H is Abelian. Then the map
α : R −→ H, α(t) = exp(tX) exp(tY )
is a smooth homomorphism. Therefore α(t) = exp(tα̇(0)) = exp(t(X + Y )). ¤
7.7. Proposition. Let H be a connected Lie subgroup of a connected Lie group G. Then H is normal
b is an ideal of G.
if and only if H b
b is an ideal of G.
Proof. Assume that H b We want to prove that H is normal. Let Y ∈ H,
b X ∈ G,
b g=
exp(X). Then
∞
X
g exp(Y )g −1 = exp(Ad(exp(X))Y ) = exp(exp(ad(X))Y ) = exp( [k!]−1 ad(X)k (Y ).
k=0
b is an ideal, ad(X) (Y ) ∈ H,
Since H k b as follows from Proposition 7.3. Hence ghg −1 ∈ H, for g, h in small
neighborhoods of e in G and H respectively. But G, H are connected, hence ghg −1 ∈ H, ∀ g ∈ G, h ∈ H,
by Lemma 6.14, and H is normal.
b X ∈ G,
Conversely, suppose H normal in G. Let Y ∈ H, b t, s ∈ R, g = exp(tX). Then
g exp(sY )g −1 = exp(Ad(g)(sY )) = exp(s[exp((ad(tX)(Y ))]) ∈ H.

b f orall t ∈ R. Now
Hence exp((ad(tX)(Y )) ∈ H,
∞
X
exp((ad(tX)(Y )) = exp(t ad(X)(Y )) = Y + t ad(X)(Y ) + [k!]−1 tk ad(X)k (Y )
k=2
b with tangent vector, at zero, ad(X)(Y ) = [X, Y ]. Hence [X, H] ∈ H

is a smooth curve in H b and H
b is an
ideal. ¤
8. Covering groups
In this subsection we will study coverings of a Lie group. We refer to the Appendix for the basic facts
on covering spaces theory.
We will start with a simple fact.
8.1. Proposition. Let φ : G −→ H be a surjective Lie group homomorphism. The following conditions
are equivalent
(1) ker φ is discrete,
(2) dφ(e) : Gb −→ H b is an isomorphism,
(3) φ is a covering map.
8. COVERING GROUPS 110
Proof. Since φ is smooth and surjective, dim(G) ≥ dim(H) by Theorem ?? of Chapter 2. Let K :=
b = ker dφ(e) and dφ(e) is an isomorphism. Conversely, suppose that dφ(e) is
ker φ. If K is discrete, {0} = K
an isomorphism. Then φ is a diffeomorphism of a neighborhood U of e ∈ G onto an open neighborhood V
of e ∈ H. In particular U ∩ K = {e}. Let g ∈ K. Then gU ∩ K = {g}. In fact, if h ∈ gU ∩ K, g −1 h ∈ U ∩ K
and therefore g = h. Therefore K is discrete. Suppose now that φ is a covering map. Then K = φ−1 (e) is
discrete, being the fibre over e. So we are left to prove that condition (1) (or (2)) implies that φ is a covering
map. Let U, V be as above. Then φ−1 (V ) = ∪g∈K gU . Now the (distinct) gU ’s are disjoint, by the argument
used above, and φ|gU : gU −→ V is a diffeomorphism, hence V is evenly covered. Let h ∈ H and consider
the open neighborhood hV . Then φ−1 (hV ) = ∪g∈φ−1 (h) gU . Then these open sets are disjoint, φ restricted
to each of them is a diffeomorphism onto hV , and the latter is evenly covered. So φ is a covering map. ¤
8.2. Corollary. Let G be a simply connected Abelian Lie group. Then G is isomorphic to the additive
group Rn .
b ∼
Proof. We consider the (additive) vector group G b −→ G is
= Rn . Since G is Abelian, exp : G
a group homomorphism, by Corollary 7.6. We will show that it is surjective. Let g ∈ G. Since G is
connected, g is product of elements in a small neighborhood of e, which is in the image of exp. Hence
g = exp(X1 ) · · · exp(Xk ) = exp(X1 + · · · + Xk ). Since d exp(0) = 11, exp is a covering map by Proposition
8.1. Since G is simply connected, exp is a diffeomorphism and, being a homomorphism, it is a Lie groups
isomorphism. ¤
We will classify now all Abelian Lie groups. We need the following
8.3. Lemma. Let K ⊆ Rn be a discrete non trivial subgroup. Then there exist linearly independent
vectors, v1 , . . . , vk , that generates K.
Proof. We proceed by induction on n. Let n = 1 and let g ∈ K \ {e} be an element minimal norm.
Suppose h ∈ K ∩ (mg, (m + 1)g then h − mg ∈ K and |h − mg| < |g|, contradicting the minimality of |g|.
Hence h ∈ gZ and K is generated by g. Suppose the Lemma holds for discrete subgroups of Rn−1 . Take
again an element g ∈ K \ {e} of minimal norm. Let π : Rn −→ g ⊥ be the orthogonal projection. ¤
8.4. Theorem. An Abelian Lie group G is isomorphic to the direct product of the k-torus S 1 ×· · ·×S 1 , (k
factors), and Rn−k .
b −→ G is a homomorphism. Also exp is surjective,
Proof. First observe that, by Corollary 7.6 exp : G
∼ G/K.
by the same argument used in Corollary 8.2. Hence K = ker exp is a discrete subgroup and G = b By
Lemma 8.3 K is generated by k independent vectors. Consider a linear isomorphism L : G b −→ R which
n
sends these vectors in the first k vectors of the canonical basis. Then L induces a Lie group isomorphism
e : H/K
L b −→ Rn /Γ where Γ is the subgroup spanned by the first k vectors of the canonical basis. The
conclusion follows from the simple fact that Rn /Γ is Lie isomorphic to S 1 × · · · × S 1 × Rn−k . ¤
8.5. Theorem. Let G be a Lie group and let π : G −→ G be a covering map. Then G has a (unique)
manifold structure such that π is a Lie groups homomorphism.
Proof. (Sketch) Consider the commutative diagram

¤
In particular, since manifolds are locally semi simply connected, there is a (unique) simply connected
e that covers G. For simply connected groups we have the following important result
Lie group G,
b −→ H
8.6. Theorem. Let G, H be Lie groups with G simply connected. If ψ : G b is a Lie algebras
homomorphism, then there exists a unique Lie groups homomorphism φ : G −→ H such that ψ = dφ(e). In
particular simply connected Lie groups with isomorphic Lie algebras are Lie isomorphic.
Proof. ¤
8.7. Remark. A celebrated result of Ado, that we will not prove here, states that a (finite dimensional)
\R), for some n. In particular there is a Lie group H with Lie algebra
b is a subalgebra of GL(n,
Lie algebra H
b by ??, and , in particular, a (unique) simply connected one, by Theorem 8.6. Hence
H,
8.8. Theorem. There is a 1-1 correspondence between finite dimensional Lie algebras and simply con-
nected Lie groups.
9. Group actions and homogeneous spaces
Let G be a Lie group, H a subgroup and π : G −→ G/H be the quotient map. If H is not normal, the
quotient set G/H does not have a natural group structure. However has a natural topology, the quotient
topology. In this topology the open sets are the ones whose inverse image is open.
9.1. Theorem. If (H, φ) is a Lie subgroup of G, and φ is an embedding with closed image13 , then the
quotient space G/φ(H) has a unique smooth structure such that
(1) π is smooth,
(2) there are local smooth sections, i.e., for every point in the quotient there exists an open neighborhood
U and a smooth map s : U −→ G such that π ◦ s = 11U .
Proof. Since φ is an embedding we will identify H with its image φ(H). ¤
The following corollary will be useful to prove connectness of some Lie groups
9.2. Corollary. If H, G/H are connected, then G is connected.
Proof. ¤
9.3. Definition. Manifolds of the type G/H where G is a Lie group and H is a closed subgroup, are
called homogeneous manifolds.
Let M be a smooth manifold and G a Lie group.
9.4. Definition. A smooth left action of G on M is a smooth map µ : G × M −→ M such that

(1) µ(e, x) = x ∀ x ∈ M ,
(2) µ(g, µ(h, x)) = µ(gh, x) ∀ x ∈ M, g, h ∈ G.
13It is known that φ is an embedding if and only if it has closed image. We will not prove this fact in those notes.
9. GROUP ACTIONS AND HOMOGENEOUS SPACES 112
When the action is fixed we will write gx for µ(g, x).

We can define the “left translations” µg : M −→ M, µg (x) = gx. These are clearly smooth maps an,
in fact, diffeomorphisms, since (µg )−1 = µg−1 , as follows easily from the definition. So we can think of an
action as a homomorphism of the group into the group of diffeomorphism of the manifold (with suitable
conditions of differentiability).
9.5. Definition. Given an action of G on M we define

(1) O(x) := {gx : g ∈ G}, the orbit of x,
(2) H(x) := {g ∈ G : gx = x}, the isotropy (sub)group of x.
Moreover we will say that the action is transitive is O(x) = M ,
9.6. Remark. If the action is transitive then O(y) = M, ∀ y ∈ M . In fact, given z ∈ M, z = hx, y = kx
and z = hk −1 y ∈ O(y).
9.7. Example. Identify S 1 = SO(2) with the group of rotations of R3 around the z-axis. Then S 1 acts
on S 2 in a natural way. If p ∈ S 2 is not on the z-axes, O(p) is the “parallel” trough p. If p is the north or
south pole, then O(p) = p. In the first case the isotropy subgroup is the identity, in the second one is the
all S 1 .
This action induces an action on RP 2 . If we look at RP 2 as the upper hemisphere S + , modulo the
identification x ∼ ±x, x ∈ ∂S + , we have that S 1 still fixes the north pole, whose isotropy is the all S 1 , the
other orbits are copies of S 1 , with trivial isotropy, except for the orbit of x ∈ ∂S + that is a RP 1 , which is
diffeomorphic to S 1 , but the isotropy is {±1} ⊆ S 1 .
9.8. Theorem. Let M be a differentiable manifold and G a Lie group that acts transitively on M . Fix
p ∈ M an let H = H(p) be the isotropy subgroup. Then the map
h : G/H −→ M, h(gH) = gp
is well defined and a diffeomorphism.
Proof. ¤
9.9. Remark. Let G be a Lie group and H a closed subgroup. Then there is a natural transitive action
G × G/H −→ G/H, g(hH) = (gh)H.
In particular a manifold M is (diffeomorphic to) a homogeneous manifold if and only if there is a Lie group
G that acts transitively on M . We want to stress the fact that G has to be a Lie group and not only a
topological group. In fact, for any connected manifold M , the group of diffeomorphism of M , which is a
topological group with a suitable topology, acts transitively on M (see Exercise ?? of Chapter 2). However
this group is not a Lie group.
Naturally there may be several groups that acts transitively on the same differentiable manifold.
9.10. Example. Let S n ⊆ Rn+1 be the unit sphere. There is a natural action of O(n + 1) on S n , namely
µ(A, x) = A(x). Let e1 = (1, 0, . . . , 0) ∈ S n . This action is clearly transitive. The isotropy group is the
H(e) = {[aij ] ∈ O(n + 1) : a1i = ai1 = δ1i which is clearly isomorphic to O(n). Then S n is diffeomorphic to
O(n + 1)/O(n). The action above induces a transitive action of O(n + 1) on RP n = S n /{±1}. The isotropy
group is H([e1 ]) = {[aij ] ∈ O(n + 1) : a1i = ai1 = ±δ1i .
The subgroup SO(n + 1) ⊆ O(n + 1) also acts transitively on S n and RP n . We leave to the reader the
task of determine the isotropy subgroups.
The following result is useful to prove connectness of some Lie groups.
9.11. Proposition. let G be a Lie group and H ⊆ G a closed subgroup. If H and F/H are connected,
then G is connected.
Proof. ¤
9.12. Example. As an example we will use the result above to prove to prove that SO(n) is connected.
We proceed by induction. SO(2) = S 1 hence it is connected. Suppose SO(n − 1) connected and consider
the homogeneous manifold S n−1 = SO(n)/SO(n − 1) (see Example 9.10). Since S n−1 and SO(n − 1) are
connected, the latter by the inductive hypothesis, it follows that SO(n) is connected.
10. Exercises
10.1. Prove that a smooth (sub)manifold is connected if and only if it is path-connected.
10.2. Prove that the definitions of tangent space given in Remark 1.11 are equivalent.
10.3. Let M ⊆ RN be a submanifold. For p ∈ M let {Xi }, {ξj } be bases for Tp M and [Tp M ]⊥ respec-
tively. Show that {(Xi , 0), (0, Xj )} and {(Xi , 0), (0, ξj )} are bases for T(p,0) T M and T(p,0) νM respectively.
10.4. Let M be a smooth manifold and p ∈ M . Consider the set T] p M = {γ : I −→ M : γ is smooth and γ(0) =
p}, where I ⊆ R is an interval containing 0. In T]

p M define the relation
d[φ−1 ◦ γ] d[φ−1 ◦ σ]
γ∼σ ⇐⇒ (0) = (0),
dt dt
for a chart φ : Ω −→ M with p ∈ φ(Ω).
(1) Prove that ∼ does not depend on the chart and it is an equivalence relation.
(2) Define a vector space structure on T] ] ∼
p M / ∼ such that Tp M / ∼ = Tp M .
10.5. Prove that RP 1 is diffeomorphic to S 1 and CP 1 is diffeomorphic to S 2 . What about the quater-
nionic and Cayley projective lines?
10.6. Prove that if X is a smooth vector field in RP n and π : S n −→ RP n is the projection, then there
exist a smooth vector field X̃ on S n such that dπ(X̃) = X. Under which condition a vector field on S n is
induced in this way by a vector field on RP n ?
10.7. Prove that S 3 and S 7 are parallelizable.
10.8. Prove that if RP n is parallelizable, then S n is parallelizible.
10.9. Prove that RP 3 and RP 7 are parallelizable.
10.10. Prove the claims in Remark 4.7.

10. EXERCISES 114
10.11. Let M, N be smooth manifolds and let f : M −→ N be a smooth map. Suppose that f is a
bijection and df (x) is injective, ∀x ∈ M . Prove that f is a diffeomorphism (hint: use Remark 4.12 to prove
that df (x) is an isomorphism, i.e. M, N have the same dimension).
10.12. Prove that if F : M −→ RN is an embedding, F (M ) is a submanifold of RN in the sense of

Definition 1.1.
10.13. Let M be a compact manifold and N a connected non compact manifold of the same dimension.
Prove that there are no immersions of M into N .
10.14. Prove that if M is compact and F : M −→ N is a 1-1 immersion, then F is an embedding.
10.15. Consider the map V : S n−1 ⊆ Rn −→ M (n, R), V (x) = xt x, where x = [x1 , . . . , xn ] is viewed as
an 1 × n matrix and xt is the transpose of x.
(1) Prove that V is an immersion (called the Veronese immersion).
(2) Prove that V induces an embedding Ṽ : RP n−1 −→ M (n, R) (called the Veronese embedding).
n(n + 1)
(3) Prove that the image of V is contained in a sphere of an [ − 1]-dimensional affine subspace
2
of M (n, R).
10.16. Consider the sphere S 2n−1 ⊆ Cn . Define V : S 2n−1 −→ M (n, C), V (z) = z ∗ z, where z =
[z1 , . . . , zn ] and z ∗ is the transpose conjugate of z.
(1) Prove that V induces an embedding Ṽ : CP n−1 −→ M (n, C), the Veronese embedding.
(2) Use the same idea to define an embedding of the quaternionic projective space into M (n, H), the
space of n × n matrices with quaternionic entries.
10.17. Give an example of a 1-1 immersion φ : R −→ R2 such that there exists a smooth curve σ :
(−², ²) −→ R2 with image contained in φ(R), but σ̇(0) not tangent to γ at σ(0).
10.18. Let π : M −→ N be a surjective local diffeomorphism. We will say that X̃ ∈ H(M ) project onto
X ∈ H(N ), if dπ(x)(X̃) = X(π(x)).
(1) Prove that, if π is surjective, ∀ X ∈ H(N ) there is a unique X̃ ∈ H(M ) that project onto X.
(2) Prove that if X̃ projects onto X, the map π sends integral curves of X̃ onto integral curves of X.
10.19. Consider the torus T 2 = S 1 ×S 1 ⊆ R4 and the map π : R2 −→ R4 , π(t, s) = (cos t, sin t, cos s, sin s).
(1) Prove that π is an immersion.
(2) Prove that π, as a map onto T 2 , is a local diffeomorphism.
(3) Let X = (x0 , y0 ) ∈ R2 and extend X to a constant vector field in R2 . Prove that X projects, via
π, onto a vector field on the torus.
(4) Prove that, if y0 /x0 is rational, the image of the integral curve γ(t) = tX is an (embedded) closed
curve.
(5) Prove that, if y0 /x0 is irrational, the image of the integral curve γ(t) = tX is an 1-1 immersed
curve with dense image (in T 2 ). Conclude that such an integral curve is not embedded.
10.20. Prove that, for all x, y ∈ Rn there exists a diffeorphism φ : Rn −→ Rn such that φ(x) = y and
φ(z) = z for kzk sufficiently large (hint: consider the (constant) vector field X = y − x. If γt is the flow of
X, γ1 (x) = y. Take now a vector field that coincide with X in a ball containing x, y and vanish outside a
bigger ball).
10.21. Let M be a connected smooth manifold, p, q ∈ M . Prove that there is a diffeomorfism φ : M −→

M such that φ(p) = q (hint: use the result above for the case in which p and q belong to the image of a
given chart and use this fact to show that the set of points q 0 such that there exist a diffeomorphism taking
a fixed p to q 0 is open and closed in M ).
10.22. Let G be a Lie Group. Prove that the connected component of the identity, Ge is a normal open
and closed subgroup. in particular, if G is compact, Ge is compact
10.23. Determine, up to isomorphisms, all 1-dimensional Lie groups and 2-dimensional simply connected
Lie groups.
10.24. Let K be the field of real or complex numbers.Consider the canonical scalar product in Kn , hx, yi =
P
i xi yi . Prove that a K-linear map A : Kn −→ Kn preserves the scalar product if and only is AA∗ = 11,
where A∗ is the conjugate transpose (with respect to h·, ·i) of A.
10.25. Consider the groups G = {A ∈ L(K n , Kn ) : AA∗ = 11}. Define a Lie group structure on G and
determine Gb (as a Lie algebra). These groups are called the orthogonal an the unitary group, respectively,
and are denoted by O(n), U (n).
10.26. Prove that O(n) and U (n) are compact, U (n) is connected and O(n) has two connected com-
ponents. The connected components of O(n) that contains 11 is called the special orthogonal group and is
denoted by SO(n).
10.27. Consider SU (n) = {A ∈ U (n) : det(A) = 1}.

(1) Prove that SU (n) is a Lie subgroup of U (n).
(2) Prove that the map m : S 1 × SU (n) −→ U (n), m(z, A) = zA is a surjective Lie group homomor-
phism.
(3) Prove that ker m is a cyclic group of order n.
10.28. Let Z2 = {±1} be the multiplicative group with two elements.

(1) Prove that O(2n + 1) is isomorphic, as Lie group, to SO(2n + 1) × Z2 .
(2) Prove that O(2n) is diffeomorphic to SO(2n) × Z2 .
(3) Define a Lie group structure on SO(2n) × Z2 in such a way that it is isomorphic to SO(2n).
10.29. Prove that SL(n, R) := {A ∈ M (n, R) : det(A) = 1} is a connected Lie subgroup of GL(n, R) and
\R) (hint: compute d det(A)(A)). This group is called the special linear group.
compute SL(n,
10.30. Prove directly that A ∈ Skw(n, R) if and only if exp(A) ∈ O(n) and det(A) = 1 if and only if
trace(A) = 0.
10. EXERCISES 116
10.31. Let A ∈ M (n, R) be a positive symmetric matrix. Prove that exp−1 (A) 6= ∅. Use this fact to
prove that there exists a positive symmetric matrix B such that B 2 = A. The matrix B is also denoted by
√
A.
10.32. Let Sym(n, R)+ be the set of positive definite symmetric matrices (observe that it is an open set
of the vector space of symmetric matrices, hence a differentiable manifold).
(1) Prove the polarization formula: Given A ∈ GL(n, R), there exist P ∈ Sym(n, R)+ , R ∈ O(n) such
√
that A = P R (hint set P = AAt and R = P −1 A).
(2) Prove that m : Sym(n, R)+ × O(n) −→ GL(n, R), m(P, T ) = P T is a diffeomorphism.
10.33. Let D be the group of real upper triangular matrices with positive diagonal entries.
(1) Prove that, given A ∈ GL(n, R), there exists a unique pair of matrices (P, R) ∈ D × O(n) such that
A = P R (hint: this is equivalent to the Gram-Schmidt orthonormalization process).
(2) Prove that that m : D × O(n) −→ GL(n, R), m(P, T ) = P T is a diffeomorphism. In particular
GL(n, R) is diffeomorphic to Rn(n+1)/2 × O(n).
10.34. Let K = R, C. Consider K∗ = K \ {0}. Identify K ∗ with the multiple of 11 ∈ GL(n, K). Show that
K∗ is a normal subgroup and define P GL(n, K) = GL(n, K)/K∗ . Prove that GL(n, K)/K∗ is a Lie group.
This group is called the projective linear group.
10.35. Consider S 3 as the space of unit quaternions (see Remark 2.13 of Chapter 2 for the definition of
quaternions).
(1) Prove that S 3 is a (non commutative) Lie group (with the induced product).
(2) Compute Sc3 and show that it is isomorphic, as a Lie algebra, to R3 with the usual vector product.
(3) Prove that {±1} ⊆ S 3 is a normal subgroup.
(4) Define a Lie group structure on RP 3 such that the quotient map q : S 3 −→ RP 3 is a Lie group
homomorphism.
10.36. Consider S 3 as above. If q ∈ S 3 , define Lq : H ∼

= R4 −→ H, Lq (x) = qxq.
(1) Prove that Lq ∈ O(4) and Lq (1) = 1.
(2) Define L̃q = Lq |1⊥ (1⊥ = {p ∈ H : p = xi + yj + zk} ∼
= R3 ). Prove that the map π : S 3 −→
SO(3), π(q) = L̃p is a Lie group homomorphism.
(3) Prove that SO(3) is isomorphic, as a Lie group to RP 3 .
10.37. Consider the map Lp,q : H −→ H, Lp,q (x) = pxq.

(1) Prove that, if p, q ∈ S 3 , Lp,q ∈ SO(4).
(2) Prove that the map L : S 3 ×S 3 −→ SO(4), L(p, q) = Lp,q is a surjective Lie group homomorphism.
(3) Compute ker L and show that L induces a diffeomorphism of SO(4) onto S 3 × RP 3 .
(4) Determine a Lie group structure on S 3 × RP 3 that makes the diffeomorphism above a Lie group
isomorphism.
10.38. Exhibit the Stieffel and Grasmann manifold as homogeneous manifolds (see Chapter 2 for defini-
tion). Prove that they are compact and connected.
10.39. Define transitive actions of P GL(n + 1, K) on KP n . Determine the isotropy subgroups.

" #
−2 0
10.40. Prove that the matrix is not not in the image of exp : M (2, R) −→ GL(2, R).
0 −1
Remark: Just for the records. If G is a connected compact Lie group, exp : G b −→ G is surjective.
CHAPTER 4
The theory of de Rham for differentiable manifolds
In this Chapter we will extend the concepts and results of the first two Chapters to the case of differ-
entiable manifolds. Since we have done most of the hard work in the first two Chapters, this extension will
very “smooth.”
1. Cohomology and homology for differentiable manifolds
1.1. Definition. Let M be a differentiable manifold. A differential p-form is a law that associate to
each x ∈ M an exterior p-form ω(x) ∈ Λp (Tx M ) that depends differentiably on x. This means that, given
smooth vector fields X1 , . . . , Xp ∈ H(M ), the function y Ã ω(y)(X1 (y), . . . , Xp (y)) is smooth.
We will denote by Ωp (M ) the set of differential p-forms on M .
Ωp (M ) has an obvious structure of real vector space. Moreover we can multiply a differential form by
a smooth function and this operation is associative and distributive, in the appropriate sense, i.e. Ωp (M )
is a module over F(M ). Finally, the wedge product induces a wedge product of differential forms. So
Ω∗ (M ) := ⊕Ωp (M ) is an associative, graded commutative algebra.
A differential form ω ∈ Ωp (M ) induces a F(M )-multilinear map
ω̃ : H(M ) × · · · × H(M ) −→ F(M ), ω̃(X1 , . . . , Xp )(x) = ω(x)(X1 (x), . . . , Xp (x)).
As for the case of open sets of Rn we have the simple but important converse
1.2. Theorem. [Tensoriality criterion] A R-multilinear map
ω̃ : H(M ) × · · · × H(M ) −→ F(M ),
is induced by a differential form if and only if it is F(M )-multilinear.
1.3. Example. Since Λ0 (Rn ) = R, Ω0 (M ) = F(M ).
1.4. Example. If f ∈ F(M ), df is the differential 1-form df (X) = X(f ).
Let F : M −→ N be a smooth map. Then, for all x ∈ M we have a linear map dF (x) : Tx M −→ TF (x) N
and an induced linear map
F ∗ : Ωp (N ) −→ Ωp (M ), F ∗ (ω)(Xi , . . . , Xp ) = ω(dF (X1 ), . . . , dF (Xp )).
It is easily seen that such map commutes with multiplications by functions in F(N ), in the sense that
F ∗ (f ω)(x) = (f ◦ F )F ∗ (ω). Also, F ∗ (ω ∧ τ ) = F ∗ (ω) ∧ F ∗ (τ ), i.e. F ∗ : Ω∗ (N ) −→ Ω∗ (M ) is an algebras
homomorphism.
Finally we have the functorial properties
119
1. COHOMOLOGY AND HOMOLOGY FOR DIFFERENTIABLE MANIFOLDS 120
• 11∗M = 11Ωp (M ) ,
• If F1 : M1 −→ M2 and F2 : M2 −→ M3 are smooth maps, (F2 ◦ F1 )∗ = F1∗ ◦ F2∗ .
In particular, if F is a diffeomorphism, F ∗ is an isomorphism.
Also we have thAs for the case of open sets of Rn we have the exterior differential of a differential form.
1.5. Theorem. There exists a unique sequence of R-linear operators dp : Ωp (M ) −→ Ωp+1 (M ) p =

0, . . . , n, such that:
(1) d0 = d (the usual differential, see Example 1.4).
(2) dp+1 ◦ dp = 0.
(3) If ω ∈ Ωp (U ), τ ∈ Ωq (U ), dp+q ω ∧ τ = dp ω ∧ τ + (−1)p ω ∧ dq τ.
Moreover, if F : M −→ N is a smooth map and ω ∈ Ωp (N ), dp F ∗ ω = F ∗ dp ω.
When clear from the context we will write simply d for dp .
Proof. We will sketch a couple of possible approach to the proof, leaving the details to the reader. We
can use the hint given by Proposition 2.11 in Chapter 1 and define, for ω ∈ Ωp (M ),
p
X X
dω(X0 , . . . , Xp ) = (−1)i Xi · ω(X0 , . . . , X̂i , . . . Xp ) + (−1)i+j ω([Xi , Xj ], X0 , . . . , X̂i , . . . , X̂j , . . . , Xp ).
i=0 i<j
Alternatively, we can proceed as follows: if ω has support in the image of a chart φ : U ⊆ Rn −→ M ,

we can define dω = (φ−1 )∗ dφ∗ ω (dφ∗ ω is the exterior differential of a form defined in U ⊆ Rn ). Again two
ways to prove that this definition does not depend on the choice of the chart. Either we can prove that this
definition coincides with the one given above, or we can use the unicity part of Theorem 2.9 of Chapter 1.
Finally observe that any form is a (locally finite) sum of forms supported in images of charts. ¤
1.6. Remark. Just like in the case of open sets in Rn , d is a local operator (Lemma 2.10 of Chapter 1).
So we have a sequence of vector spaces and R-linear maps:
d0 d1 dn−1
0 −→ Ω0 (M ) −→ Ω1 (M ) −→ · · · −→ Ωn (M ) −→ 0
which is a cochain complex, called the de Rham complex of M . We define, as in Chapter 1,
• Z p (M ) := ker dp , the space of p-cocycles or closed p-forms.
• B p (M ) := Im dp−1 , the space p-coboundaries or exact p-forms.
• H p (M ) := Z p (M )/B p (M ), the p-dimensional (de Rham) cohomology of M .
Let F : M −→ N be a smooth function and Fp∗ : Ωp (N ) −→ Ωp (M ) the induced map. It follows from
the last claim in Theorem 1.5, that F ∗ sends closed forms to closed forms and exact forms to exact forms.
Hence it induces a linear map, denoted by the same symbol,
F ∗ : H p (N ) −→ H P (M ), F ∗ ([α]) = [F ∗ α],
where [τ ] denotes the class of the form τ in the quotient space.

The basic functorial properties
• 11∗M = 11H p (M ) ,
• If F1 : M1 −→ M2 and F2 : M2 −→ M3 are smooth maps, then (F2 ◦ F1 )∗ = F1∗ ◦ F2∗ ,
CHAPTER 4. THE THEORY OF DE RHAM FOR DIFFERENTIABLE MANIFOLDS 121
are obvious. In particular, if F is a diffeomorphism, F ∗ is an isomorphism. So the de Rham cohomology is

a (differential) topological invariant of M .
The basic properties of the de Rham cohomology proved in Chapter 1 hold true with essentially the
same proofs.
1.7. Proposition.
• If M is a point, H 0 (M ) ∼
= R, H p (M ) = {0}, p > 0.
• M is connected if and only if H 0 (M ) ∼
= R.
` Q
• If M = Mα (disjoint union) then H p (M ) = H p (Mα (direct product).
• If M, N are connected and F : M −→ N is a smooth map, F ∗ : H 0 (N ) −→ H 0 (M ) is an
isomorphism.
1.8. Theorem. [Homotopy invariance for cohomology] Let F, G : M −→ N be a smooth maps between
smooth manifolds. If F is homotopic to G, then F ∗ = G∗ : H p (N ) −→ H p (M ), ∀ p. In particular, if
∗ p p
F : M −→ N is a homotopy equivalence, F : H (N ) −→ H (M ) is an isomorphism.
Proof. It is enough to show, just like in the case of open sets of Rn , that the inclusions ji : M −→
M × R, ji (x) = (x, i), i = 0, 1, induce the same homomorphism in cohomology. For this is enough to
produce an algebraic homotopy H̃ : Ωp (M × R) −→ Ωp−1 (M ) between the maps induced by the inclusions.
We fix an atlas {Ui , φi } and a partition of unity λi dominated by the atlas. Given ω ∈ Ωp (M × R) set
P
ωi = λi ω. Then ωi is supported in φi (Ui ) × R and ω = ωi . For fixed i, we have defined an algebraic
P
homotopy H̃i : Ω (Ui × R) −→ Ω (Ui ). We define H̃(ωi ) = (φ−1
p p−1 ∗ ∗
i ) H̃i (φi × 11) ωi and H̃(ω) = H̃(ωi ). It
is easy to see that H̃ is a well defined algebraic homotopy between the maps induced by the inclusions. ¤
1.9. Theorem. [Mayer Vietoris sequence for de Rham cohomology] Let M be a differentiable manifold,
U1 , U2 open sets such that M = U1 ∪ U2 . Let V = U1 ∩ U2 and ki : V −→ Ui , ji : Ui −→ M be the
inclusions. Then there exists a sequence of linear maps ∆∗ : H p (V ) −→ H p+1 (M ), such that the sequence
below is exact:
(j ∗ ,j ∗ ) (k∗ −k∗ ) ∆∗
· · · −→ H p (M ) −→
1 2
H p (U1 ) ⊕ H p (U2 ) −→
1 2
H p (V ) −→ H i+1 (M ) −→ · · ·
Proof. The result is a purely algebraic consequence of the exactness of the sequence
(j ∗ ,j ∗ ) (k∗ −k∗ )
{0} −→ Ωp (U ) −→
1 2
Ωp (U1 ) ⊕ Ωp (U2 ) −→
1 2
Ωp (V ) −→ {0}.
The exactness of this sequence is proved exactly as in the case of open sets of Rn . ¤
x
1.10. Example. Let S n be the unit sphere. The map r : Rn+1 \ {0} −→ S n , r(x) = is a homotopy
kxk
equivalence. Hence (see Example 4.15 and Exercise 7.28 in Chapter 1)
(
R if p = 0, n
H p (S n ) =
{0} if p =
6 0, n
1.11. Example. Consider the complex projective space CP n . CP 1 = S 2 and its cohomology was
computed in the example above. We will prove, by induction, that
1. COHOMOLOGY AND HOMOLOGY FOR DIFFERENTIABLE MANIFOLDS 122
(
p n R if p = 2k, 0 ≤ k ≤ n
H (CP ) =
{0} otherwise
n
We can look at CP in two different ways
• as the quotient of the (closed) unit disk D ⊆ Cn modulo the equivalence relation x ∼ y ⇐⇒ x = y or
kxk = kyk = 1 and x = λy, λ ∈ C,
• as the quotient of the unit sphere S 2n+1 ⊆ Cn+1 modulo the equivalence relation x ∼ y ⇐⇒ x = λy, λ ∈ C.
Look at the first model and define U1 = {[x] ∈ CP n : kxk < 1}, U2 = CP n \ {0}. Then we have
• U1 is diffeomorphic to the open ball, hence contractible.
x
• The map r : D \ {0} −→ S 2n−1 , r(x) = , is a homotopy equivalence which induces an homotopy
kxk
equivalence between U2 and CP n−1 (second model)
• U1 ∩ U2 = {x ∈ D : 0 < kxk < 1}, hence homotopy equivalent to S 2n−1 .
For this decomposition the Mayer Vietoris sequence is
· · · H p−1 (S 2n−1 ) −→ H p (CP n ) −→ H p (U1 ) ⊕ H p (CP n−1 ) −→ H p (S 2n−1 ) −→ H p+1 (CP n ) −→ · · · .
Hence, if 0 < p ≤ 2n − 2, H p (CP n ) −→ H p (U1 ) ⊕ H p (CP n−1 ) = H p (CP n−1 ) is an isomorphism. For
p = 2n − 1 we have
{0} = H 2n−2 (S 2n−1 ) −→ H 2n−1 (CP n ) −→ {0},
hence H 2n−1 (CP n ) = {0}. For p = 2n we have
H 2n−1 (U1 ) ⊕ H 2n−1 (CP n−1 ) = {0} −→ H 2n−1 (S 2n−1 ) −→ H 2n (CP n ) −→ H 2n (U1 ) ⊕ H 2n (CP n−1 ) = {0},
hence H 2n (CP n ) ∼
= R. The conclusion follows by the inductive hypothesis (observe that H 0 (CP n ) ∼
= R, by
connectedness).
1.12. Example. Essentially the same argument used in the example above allow us to compute the
cohomology of the quaternionic projective space HP n . We obtain
(
R if p = 4k, 0 ≤ k ≤ n
H p (HP n ) =
{0} otherwise
For the singular homology we also have very little new. If M is a differentiable manifold, a smooth
p-simplex is a smooth map σ : ∆p −→ M , ∆p being the standard p-simplex, and the space of p-chains of
M , Cp (M ), is the real vector space with basis the singular simplices. We can define the boundary operator
∂p : Cp (M ) −→ Cp−1 (M ) just like in Chapter 2 and obtain the singular chain complex of M
∂p+1 ∂p
· · · −→ Cp+1 (M ) −→ Cp (M ) −→ Cp−1 (M ) −→ · · ·
whose homology, Hp (M ), is the singular homology of M .

olso, if F : M −→ N is a smooth manifold, we have an induced morphism F∗ : Cp (M ) −→ Cp (N )
(composing a singular simplex with F ) and this morphism induces a linear map (denoted by the same
symbol) F∗ , F∗ : HP (M ) −→ Hp (N ). This construction verifies the functorial properties
• 11M = 11Hp (M ) ,
• (F ◦ G)∗ = F∗ ◦ G∗ .
The basic properties of homology also hold true, with the same proofs as in the case of open sets of Rn .
1.13. Proposition.
• If M is a point, H0 (M ) ∼
= R, Hp (M ) = {0}, p > 0.
• M is connected if and only if H0 (M ) ∼
= R.
`
• If M = Mα (disjoint union) then Hp (M ) = ⊕Hp (Mα (direct sum).
• If M, N are connected and F : M −→ N is a smooth map, F∗ : H0 (N ) −→ H0 (M ) is an isomor-
phism.
1.14. Theorem. [Homotopy invariance for homology] Let F, G : M −→ N be a smooth maps between
smooth manifolds. If F is homotopic to G, then F∗ = G∗ : Hp (N ) −→ Hp (M ), ∀ p. In particular, if
F : M −→ N is a homotopy equivalence, F∗ : Hp (N ) −→ Hp (M ) is an isomorphism.
1.15. Theorem. [Mayer Vietoris sequence for singular homology] Let M be a differentiable manifold,
U1 , U2 open sets such that M = U1 ∪ U2 . Let V = U1 ∩ U2 and ki : V −→ Ui , ji : Ui −→ M be the
inclusions. Then there exists a sequence of linear maps ∆∗ : Hp (M ) −→ Hp−1 (M ), such that the sequence
below is exact
[(k1 )∗ ,(k2 )∗ ] [(j1 )∗ −(j2 )∗ ] ∆∗
· · · −→ Hp (V ) −→ Hp (U1 ) ⊕ Hp (U2 ) −→ Hp (M ) −→ Hp−1 (V ) −→ · · · .
2. The Theorems of de Rham and Künnet for differentiable manifolds
As in the case of open sets of Rn we can define, for a differentiable manifold M , integration of p-forms
over p-chains of M . This construction induces, via the Theorem of Stokes, a de Rham map
dRM : H p (M ) −→ [Hp (M )]∗ ,
which is natural with respect to smooth maps.

The de Rham Theorem, proved in Chapter 2 for open sets of Rn , states that, in the latter case, dR is an
isomorphism. In this section we will extend this result, and the Künnet formula, to the case of differentiable
manifolds.
We can try to reproduce the proofs given for the case of open sets of Rn . The careful reader understand
quickly the this reproduction is immediate once we have a result of the type of Lemma 3.1 of Chapter 2. For
the latter we encounter a couple of problems.
• We need a proper function.
• We need a concept of convex sets
The first problem is easily solved. In fact, by Whitney’s Theorem we can assume that M is a submanifold
of RN , which is a closed subset. Then the the restriction to M of the function φ(x) = kxk2 is a smooth
proper function.
For the second we need a family of subsets of M , {Cα , α ∈ A} with the following properties
• each Cα is contractible,
• the intersection of two sets in the family is still in the family,
• ∀ x ∈ M and neighborhood U ⊆ M of x, there exist α such that Cα ⊆ U .
It can be proved that such a family indeed exists, the geodetically convex sets for example, and we will
comment on this class in Chapter 5.
2. THE THEOREMS OF DE RHAM AND KÜNNET FOR DIFFERENTIABLE MANIFOLDS 124
In this section we will take a different approach, which is interesting in itself. In fact we will prove that
a smooth n-dimensional manifold M is homotopy equivalent to an open subset of RN , N sufficiently large,
and use directly the case of open sets of RN to obtain the desired generalizations.
Let M be a n-dimensional manifold. By Whitney’s Theorem we can assume, without loss of generality,
that M is a submanifold of RN and a closed subset. Consider the normal bundle (see Example 1.19 of
Chapter 3) and the endpoint map
νM = {(p, ξ) ∈ M × RN : ξ ∈ [Tp M ]⊥ }, E : νM −→ RN , E(p, ξ) = p + ξ.
Observe that E is a smooth map between manifolds of the same dimension. Let ² : M −→ R be a positive
continuous function. We set
ν² M = {(p, ξ) ∈ M × RN : ξ ∈ [Tp M ]⊥ , kξk < ²(p)}
2.1. Theorem. [Tubular neighborhood Theorem] There exists a positive continuous function ² : M −→
R such that E|ν² M is a diffeomorphism onto an open set T ub(M ) := T ub(M )² ⊆ RN .
Proof. We first compute dE(p, 0), p ∈ M . Let {Xi }, {ξj } be bases for Tp M and [Tp M ]⊥ respectively.
Then T(p,0) νM is spanned by {(Xi , 0), (0, ξj )} (see Exercise 10.3 of Chapter 3). Let γi : (−², ²) −→ M be
curves such that γ̇i (0) = Xi . Then
d d
dE(p, 0)(0, ξj ) = |t=0 tξj = ξj , dE(p, 0)(Xi , 0) |t=0 γi (t) = Xi .
dt dt
Therefore dE(p, 0) is an isomorphism and, by the inverse map Theorem, E maps a neighborhood Ũ of
(p, 0) diffeomorphically onto an open neighborhood of p ∈ RN . Ũ contains a neighborhood of the type
Ur(p) = {(q, ξ) ∈ νM : q ∈ U 0 , kξk < r(p)} where U 0 is a neighborhood of p ∈ M and r(p) > 0.
We will assume, for the moment, that M is compact. Then we can cover {(p, 0) : p ∈ M } ⊆ νM with a
finite number of neighborhoods Ur(pi ) . Let r < r(pi ). Then E|νs M is a local diffeomorphism for s < r, in
particular an open map. Then the Theorem will follows from the following fact
Claim 1. There exists a constant ² ∈ (0, r) such that E|ν² M is a diffeomorphism onto the open set
T ub(M ) = E(ν² M ) ⊆ RN .
Proof. It is sufficient to prove that there exists ² > 0 such that E|ν² M is injective. Suppose that
this not the case. Then, for all n there are distinct points (pn , ξn ), (qn , ηn ) with kξn k, kηn k < 1/n and
E(pn , ξn ) = E(qn , ηn ). By compactness, there are subsequences converging, for n → ∞, to points (p, 0) and
(q, 0) respectively and, by continuity E(p, 0) = E(q, 0). But this implies p = q. Then, for n sufficiently large,
(pn , ξn ), (qn , ηn ) are distinct points in the same Ur(pi ) , a contradiction since E|Ur(pi ) is injective. ¤
.
We consider now the case when M is not necessarily compact. We will start with some notations. Given
p ∈ M and ² > 0 we define the normal ball at p of radius ² as
Bp⊥ (²) = {p + ξ ∈ RN : ξ ∈ [Tp M ]⊥ . kξk < ²}.
Given K ⊆ M we will say that ²K is an admissible radius if Bp⊥ (²K ) ∩ Bq⊥ (²K ) = ∅ if p, q ∈ K, p 6= q. We
will also set TK = ∪p∈K Bp⊥ (²K ).
As in the proof of Theorem 3.2 of Chapter 3 we have a sequence of compact sets K1 ⊂ · · · ⊂ Ki ⊂ · · ·

with M = ∪Ki and Ki ⊆ Int(Ki+1 ). Then, as in the proof of Claim 1., we can find constant ²i > 0 such that
E|{(p, ξ) : p ∈ Int(Ki ), kξk < ²i } is a diffeomorphism onto an open set Ui ⊆ RN . We can suppose, without
loss of generality, ²i ≥ ²i+1 .
Claim 2. There is a sequence of constants αi ≥ αi+1 such that if p ∈ Ki , q ∈ Kj , p 6= q, Bp⊥ (αi ) ∩
Bq⊥ (αj ) = ∅.
Proof. Set α1 = ²2 , α2 = ²3 . Suppose to have defined α1 ≥ α2 ≥ · · · ≥ αs and the Claim holds for
i, j ≤ s. Define
0 < αs+1 < min{²s+2 , αs , d(Ks+1 \ intKs , ∪s−1
i=1 TKi αi )},
where d is the distance in RN .

We want to show that the Claim holds for i, j ≤ s + 1. For i, j ≤ s this is the inductive hypothesis. For
i = s, j = s + 1 the Claim follows from the fact that αs ≤ ²s+1 . Finally, suppose p ∈ Ks+1 \ IntKs , q ∈
Ki , i < s.
¤
2.2. Definition. The set T ub(M ) is called a tubular neighborhood of M .
2.3. Corollary. The inclusion i : M −→ T ub(M ) is a homotopy equivalence.
Proof. Define r : T ub(M ) −→ M, r(E(p, ξ)) = p. Then r ◦ i = 11M and i ◦ r ∼ 11T ub(M ) via the
homotopy H(p, ξ) = E(p, sξ). ¤
At this point we can prove the announced results
2.4. Theorem. [de Rham Theorem for smooth manifolds] Let M be a smooth manifold. The de Rham
map dRM : H p (M ) −→ [Hp (M )]∗ is an isomorphism.
Proof. Since integration commutes with induced maps, we have a commutative diagram
dRT ub(M )
H p (T ub(M )) −→ [Hp (T ub(M ))]∗
i∗ ↓ ↓ (i∗ )∗
dRM
H p (M ) −→ [Hp (M )]∗
Since the vertical maps, as well as the top horizontal one, are isomorphisms, the same is true for the
bottom horizontal one. ¤
Using essentially the same argument we have
2.5. Theorem. [Künnet Theorem for smooth manifolds] Let Mi , i = 1, 2 be smooth submanifolds of
Ni
R . Then
H p (M1 × M2 ) ∼
= ⊕k+l=p H k (M1 ) ⊗ H l (M2 ).
3. THE JORDAN SEPARATION THEOREM REVISITED 126
3. The Jordan separation Theorem revisited
As a further application of the existence of a tubular neighborhood, we prove now a version of the
Theorem of Jordan. The proof is interesting since it put light on the topological properties that make the
Theorem work.
3.1. Theorem. [Jordan Theorem for smooth hypersurfaces] Let M n ⊆ Rn+1 be a connected submanifold
which is a closed subset. Suppose that there exist a smooth function η : M −→ Rn+1 such that η(p) ∈
[Tp M ]⊥ , kη(p)k = 1. Then Rn+1 \ M has exactly two connected components.
Proof. We start observing that the map t : M × R −→ νM, t(p, t) = (p, tη) is a diffeomorphism. Then
νM \ M × {0} has exactly two connected components ν± = {(p, ξ) ∈ νM : hξ, ηi > 0 (respectively < 0)}.
Consider a tubular neighborhood E|ν² M : ν² M −→ T ub² (M ). It is easily seen that ν² M \ M × {0}
has also exactly two connected components which are the intersection of ν² M with ν± . Since E|ν² M is a
diffeomorphism onto its image and E|{(p,0)∈νM } is a diffeomorphism onto M, T ub² (M ) \ M has also exactly
two connected components, say T± . The Theorem will follows from the Claims below.
Claim 1. Any connected component of Rn \ M contains either T+ or T− . In particular Rn \ M has, at
most, two connected components.
Proof. Let C be a connected component of Rn \ M and x ∈ C. Let γ : [0, 1] −→ Rn a curve with

γ(0) = x, γ(1) ∈ M . Let t0 = sup{t ∈ [0, 1] : γ(s) 6∈ T ub² (M ) ∀ s ≤ t}. Then, for r sufficiently small
γ|[0, t0 + r] does not intersect M and γ(t0 + r) ∈ T± Let say γ(t0 + r) ∈ T+ . Then C ∩ T+ 6= ∅. Since T+
is connected, T+ ⊆ C. Finally, given three connected components of Rn \ M two of them, by the argument
above, contains T+ or T− , hence have non empty intersection and therefore they coincide. ¤
Claim 2. Rn \ M is not connected.
Proof. We will prove that there exists a smooth function f : Rn −→ R such that f assumes positive and
negative values and f (x) = 0 if and only if x ∈ M . This will prove the claim since a continuous real valued
function defined on a connected set and assuming positive and negative values must vanish somewhere.
On T ub² (M ) we have such a function, namely g̃ = π2 ◦ t−1 ◦ [E|ν² M ]−1 , where π2 : M × R −→ R is the
projection on the second factor. The problem is how to extend g̃ to a function defined on all Rn . We could be
tempted to use general results on existence of extensions of continuous functions but we can not guarantee
that such an extension does not vanish outside M . The strategy we will use is to modify the function g̃ near
the boundary of T ub² (M ) in such a way that the new function, call it g, still has the properties we want and
it is locally constant near ∂T ub² (M ).
Suppose we have constructed such a function g. Again we have problem extending it. In order to
construct such an extension we will proceed as follows: consider the 1-form ω = dg ∈ Ω1 (T ub² (M ). Since g
is locally constant near in a neighborhood U of ∂ T ub² (M ), ω = 0 in U and we can extend ω to a form defined
in all Rn setting ω = 0 outside T ub² (M ). This form is clearly closed and therefore, since H 1 (Rn ) = {0},
there exists a function f : Rn −→ R such that ω = df . Moreover this function is unique up to a constant and
therefore unique if we require f (p) = 0 for some p ∈ M . Observe that, with this choice, f agree with g on
T ub² M . In particular f assumes positive and negative values and, on T ub² (M ), does not vanish outside M .
We want to show that f does not vanish outside T ub² (M ). Let x ∈ Rn \ T ub² (M ) and let γ : [0, 1] −→ Rn
be a curve with γ(0) = x, γ(1) ∈ M . As in Claim 1., let t0 = sup{t ∈ [0, 1] : γ(s) 6∈ T ub² (M ) ∀s ≤ t}. Then,
for r sufficiently small γ|[0, t0 + r] does not intersect M and γ(t0 + s) ∈ U for s ≤ r. Hence
Z
f (x) = f (γ(t0 + r)) − ω = f (γ(t0 + r)) 6= 0,
γ|[0,t0 +r]
since ω = 0 along γ([0, t0 + r]).

We will prove now that a function g with the properties required above exists. We will identify ν² M
with T ub² (M ). For p ∈ M we consider the piecewise affine function λp : R −→ R which is the identity in
(−²(p)/2, ²(p)/2), it is identically 1 for t > 2²(p)/3 and identically -1 for t < −2²(p)/3. Then λ̃(p, t) := λp (t)
is a continuous function, smooth for (p, t) 6∈ V = {(q, s) : s ∈ (−2²(q)/3, −²(q)/2) ∪ (²(q)/2, 2²(q)/3)}.
Using Theorem 3.9 of Chapter 3, we can approximate λ̃ by a smooth function λ which coincide with λ̃ in
M × R \ V . We define g : M × R −→ R, g(p, ξ) = λ(p, g̃(p, ξ)). In this way we have a smooth function on
T ub² (M ), which assume positive and negative values, vanishes only on M and is constant un a neighborhood
of ∂T ub² (M ).
¤
3.2. Remark. The condition of the existence of the function η, equivalent to the fact that M is orientable
(see next section), is a consequence of the fact that M is closed in Rn+1 . We will prove this fact in the next
section. With respect to the situation discussed in Remark 5.9 of Chapter 1, we observe that we do not
need a “model” for our conclusion. However we need differentiability hypothesis, which were no needed in
Remark 5.9.
A natural question is if the Theorem holds for a connected closed hypersurface M of a connected
differentiable manifold N . We will assume that N is embedded as a closed subset of some RN . For this
situation we can define the normal bundle of M in N as
νM = {(p, ξ) ∈ T N : ξ ∈ [Tp M ]⊥ }.
Using a suitable endpoint map1 we can define tubular neighborhood also in this context, and try to
reproduce the proof given above. There are two essential points needed
• the fact that the tubular neighborhood are diffeomorphic to M × R,
• the fact that any closed 1-form in N is exact.
For the first point it is sufficient to assume that both M and N are orientable. For the second one we
need to assume that H 1 (N ) = {0}. With these two assumption is possible to reproduce the proof given
above and conclude that N \ M has exactly two components (if M, N are connected). The two conditions
are essential as the following two examples show.
• Consider the real projective plane RP 2 as the quotient of the closed disk {x ∈ R2 : kxk ≤ 1} modulo
the equivalence relation x ∼ y ⇔ x = y or x = ±y, kxk = 1. Consider the map γ : [0, 1] −→ RP 2 , γ(t) =
[cos πt, sin πt] (where [·] is the equivalence class). Then γ induces an embedding of the circle into RP 2 whose
complement is the open disk, hence a connected set. The problem here is that RP 2 is not orientable.
1The Riemannian exponential map that will be defined in the next chapter
4. ORIENTABLE MANIFOLDS, INTEGRATION OF N -FORMS AND THE THEOREM OF STOKES 128
• Consider the torus T 2 = S 1 × S 1 and the embedded circle Γ = S 1 × {(0, 1). Then T 2 \ Γ is connected.
The problem here is that H 1 (T 2 ) = R ⊕ R 6= {0}, by the Künnet formula.
Observe that in both examples, the complement of the submanifold is connected. This a general fact
since the simple existence of the tubular neighborhoods implies that the complement of the submanifold has,
at most, two connected components (see the proof of Claim 1).
4. Orientable manifolds, integration of n-forms and the Theorem of Stokes
We want to define the integral of a differential n-form on an n-dimensional manifold. We will start
recalling the change of coordinate formula for multiple integrals in Rn .
4.1. Theorem. Let F be a diffeomorphism of an open set A ⊆ Rn onto an open subset of Rn . We will
denote by J[F ] : A −→ R the Jacobian detrminant, J[F ] := det[dF ]. Then, if D ⊆ A is the closure of an
open bounded set and f : φ(D) −→ R is a continuous function,
Z Z
f= f ◦ F |J[F ]|.
F (D) D
Let now ω ∈ Ω (M ) be a differential form and let φ : U ⊆ Rn −→ M be a local chart. Suppose first
n
that ω has compact support contained in φ(U ). It is natural to (try to) define
Z Z Z
ω= φ∗ ω = f (x1 , . . . , xn ) dx1 · · · dxn , φ∗ ω = f dx1 ∧ · · · ∧ dxn .
M U U
We would like to prove to prove that the definition does not depend on the chart. If ψ : V −→ M is an
other chart containing the support of ω, we should have
Z Z
φ∗ ω = ψ∗ ω
U V
that is, using the formula of change of variables in the multiple integrals, J[ψ −1 ◦ φ] = |J[ψ −1 ◦ φ]|, and this
is not, in general, the case. However these considerations lead to the concept of orientable manifolds.
We recall that an orientation on a vector space is a choice of an equivalence class of equioriented bases
(see Definition 1.26 in Chapter 1).
4.2. Definition. Let M be a differentiable manifold. An orientation on M is the choice of an orientation

on each Tx M that depends differentiably on x. This means that for all x ∈ M there is a neighborhood U of
x and smooth vector fields X1 , . . . , Xn ∈ H(U ) such that {X1 (y), . . . , Xn (y)} is a positive bases, ∀ y ∈ U .
We will say that M is orientable if such an orientation exists, and that it is oriented if an orientation has
been fixed. If F : M −→ N is a diffeomorphism, we will say that f preserve the orientation or is orientation
preserving if dF sends positive bases onto positive bases. Otherwise we will say that F invert the orientation.
4.3. Remark. Rn comes with a canonical orientation, the one defined by the canonical basis. If F is a
diffeomorphism between open sets of Rn , it preserves the orientation if and only if J[F ] > 0.
4.4. Example. A parallelizable manifold, in particular a Lie group, is orientable.
4.5. Example. The sphere S n = {x ∈ Rn+1 : kxk = 1} is orientable. In fact Tp S n = {x ∈ Rn+1 :

hx, pi = 0}. Define {X1 , . . . Xn } ⊆ Tp S n to be positive if and only if {X1 , . . . , Xn , p} is a positive bases
of Rn+1 . More in general a submanifold M n ⊆ Rn+1 is orientable if and only if there exist a smooth map
N : M −→ Rn+1 such that N (p) ∈ [Tp M ]⊥ and N (p) 6= 0, ∀ p ∈ M .
The example above is a particular case of a much more general fact mentioned in Remark 3.2
4.6. Theorem. Let M n ⊆ Rn+1 be a submanifold which is a closed subset. Then M is orientable.
Proof. ¤
4.7. Theorem. Let M be a n-dimensional differentiable manifold. Then the following conditions are
equivalent
(1) M is orientable.
(2) There is a differential form ω ∈ Ωn (M ) such that ω(x) 6= 0, ∀ x ∈ M .
(3) There is a spacial atlas {ψi } such that J[ψi−1 ◦ ψj ] is positive, where defined.
Proof. By Whitney’s Theorem we may assume that M is a submanifold of RN . Let p ∈ M and let
{X1 , . . . , Xn } ba a local positive frame of vector fields in a neighborhood of p. Using the orthonormalization
process we may assume that the frame is orthonormal. If {ω1 , . . . , ωn } is the dual frame, we define ω =
ω1 ∧ · · · ∧ ωn . It is sufficient to prove that ω is well defined, i.e. does not depend on the chosen frame. In fact,
if {Y1 , . . . , Yn } is an other orthonormal positive frame and ω is the associated n-form, ω = det A ω where
A is the matrix of the change of bases. Since both bases are orthonormal and positive, A ∈ SO(n), hence
det A = 1. So (1) ⇒ (2).
We will show now that (2) ⇒ (3). Let ω be a nowhere zero n-form. Start with a special atlas {φi , Rn }.
Let F : Rn −→ Rn , F (x1 , x2 , . . . , xn ) = (−x1 , x2 , . . . xn ). So F reverses the orientation. Now φ∗i ω =
fi dx1 ∧ · · · ∧ dxn where fi : Rn −→ R is a nowhere vanishing function. Define ψi = φi if fi > 0 and
ψi = φi ◦ F if fi < 0. It is easily seen that the (special) atlas {ψi } has the required property.
Finally we will show that (3) ⇒ (1). Start with a special atlas {ψi } as constructed above. Consider the
local orientation on M defined by the local frame {Xji = dψi (ej )} where {ej } is the canonical basis of Rn .
In ψi (Rn ) ∩ ψj (Rn ) the frames are related by the matrix [d(ψi−1 ◦ ψj )] which has, by hypothesis, positive
determinant. So the two local orientations agree on the intersection. ¤
At this point we can define the integral of n-form on an oriented differentiable n-dimensional manifold.
Let M be such a manifold and ω ∈ Ωn (M ). To avoid convergence problem we will assume that ω has compact
support. Consider a special atlas {ψk } with J[ψi−1 ◦ ψj ] > 0. Let {λi } be a partition of unity dominated by
such a covering. Then, by the considerations above, the integral of λi ω over M is well defined.
4.8. Definition. With the notations above we define

Z XZ
ω= λi ω.
M i M
It is easily seen that the definition does not depends on the partition of unity.
At this point we would like to have, in this context, a suitable version of the Theorem of Stokes. For
this we need to define the concept of manifold with boundary or ∂-manifold. Consider the (closed) half space
and its boundary,
Hn = {(x1 , . . . xn ) ∈ Rn : xn ≥ 0}, ∂Hn = {(x1 , . . . xn ) ∈ IHn : xn = 0} ∼

= Rn−1 .
The following lemma will be useful

4. ORIENTABLE MANIFOLDS, INTEGRATION OF N -FORMS AND THE THEOREM OF STOKES 130
4.9. Lemma. Let F : Hn −→ Hn be a homeomorphism. If p ∈ ∂Hn , then F (p) ∈ ∂Hn .
Proof. Suppose F (p) 6∈ ∂Hn . Then Hn \ {F (p)} is homotopy equivalent to S n−1 while Hn \ {p} is
contractible. ¤
4.10. Definition. Let M be a topological (metric is enough) space.

(1) A ∂-chart is a map φ : U ⊆ Hn −→ M, U open in Hn , such that ψ(U ) is open and ψ is a
homeomorphis onto its image.
(2) A topological ∂-atlas is a collection of ∂-charts whose images cover M . If such an atlas exists, we
will say that M is a topological ∂-manifold.
Just as in the case of smooth manifolds we have the associated “differentiable concepts” for ∂-manifolds.
4.11. Definition. Let M be a topological ∂-manifold.

(1) A differentiable ∂-atlas is a topological ∂-atlas such that the change of coordinates are smooth.
(2) Two differentiable ∂-atlases are equivalent if their union is a differentiable ∂-atlas.
(3) A differentiable ∂-structure is an equivalence class of differentiable ∂-atlases (or a maximal one).
(4) A differentiable (or smooth) ∂-manifold is a topological ∂-manifold together with a differentiable
∂-structure.
Smooth maps between (smooth) ∂-manifolds and the tangent space of a ∂-manifold are defined as in
the case of empty boundary. Therefore we can talk, also in this context, of smooth curves and their tangent
vectors, vector field on ∂-manifolds, differential of a smooth map etc.. Let M be a ∂-manifold and p ∈ ∂M .
A smooth curve γ : [0, ²) −→ M with γ(0) = p is called an inner curve if γ̇(0) 6∈ Tp ∂M . In this case we will
say that γ̇(0) is an inner vector. In particular Tp M is spanned by Tp ∂M and an inner vector. It is easy to
see that if γ : [0, ²) −→ M is an inner curve, γ((0, η)) ⊆ M \ ∂M if η is sufficiently small.
4.12. Example. Let M be a ∂-manifold, p ∈ ∂M and let ψ : U −→ M be a chart with x = ψ −1 (p).

Then ψ(x + ten ), t ≥ 0 is an inner curve and dψ(x)(en ) = γ̇(0) is an inner vector.
4.13. Lemma. Let M be a smooth ∂-manifold. Define the boundary of M as
∂M = {p ∈ M : there exist a chart φ : U ⊆ Hn −→ M such that φ−1 (p) ∈ ∂Hn }.
Then ∂M is well defined, it is a differentiable manifold with ∂(∂M ) = ∅, and, if M is orientable, ∂M is

orientable.
Proof. The fact that ∂M is well defined follows from Lemma 4.9. Also, given a chart ψ : U ⊆ Hn −→
M, ψ|U ∩ ∂Hn : U ∩ ∂Hn −→ ∂M is a chart for ∂M . It is also clear that, given two charts for M such that
the change of coordinates is smooth, so happens for the induced charts for ∂M . Finally, if M is orientable
and p ∈ ∂M we can define a basis {X1 , . . . , Xn−1 } of Tp ∂M to be positive if {X1 , . . . , Xn−1 , X} is a positive
basis for Tp M , where X is an inner vector. We leave to the reader the task of proving the differentiability
condition in the definition of orientability. ¤
The concept of special atlas and partition of unity dominated by an open covering are the same as in
the case of manifolds without boundary. It is worth to observe that a partition of unity for M , restricted to
∂M is a partition of unity for ∂M . Also we can define integration of a n-form with compact support on a
∂-manifold, if it is oriented.
At this point we can state the Theorem of Stokes in our context.
4.14. Theorem. [Stokes Theorem] Let M be a ∂-manifold of dimension n and ω ∈ Ωn−1 (M ) a form
with compact support. Then
Z Z
dω = i∗ ω,
M ∂M
Z
where i : ∂M −→ M is the inclusion. In particular if ∂M = ∅, dω = 0.
M
Proof. We start proving the Theorem “locally”.

Claim 1. Let ω ∈ Ωn−1 (Rn ) be a form with compact support. Then
Z
dω = 0.
Rn
Proof. We can suppose, without loss of generality, that supp(ω) ⊆ Int(∆n ), where ∆n is the standard
n-simplex. Then, by Stokes Theorem for chains (see Theorem ?? of Chapter 2),
Z Z Z
dω = dω = ω=0
Rn ∆n ∂∆n
since ω = 0 on ∂∆n . ¤
Claim 2. Let ω ∈ Ωn (Hn ) be a form with compact support. Then

Z Z
dω = ω.
Hn ∂Hn
Proof. Again, without loss of generality, we can suppose that supp(ω) ⊆ Int(∆n ) ∪ Int(∆n−1 ). Then,
by Stokes Theorem for chains,
Z Z Z Z Z
dω = dω = ω= ω= ω,
Hn ∆n ∂∆n ∆n−1 ∂Hn
since ω = 0 on all faces of ∆n having en as a vertex. ¤
At this point we can “glue together” the local information using a partition of unity. Fix a special atlas
{U, φi } with U = Rn or Hn and let {λi } be a partition of unity dominated by φi (Ui ). Set ωi = λi ω. Since
integration is additive, we have to prove the Theorem just for ωi . From the Claims and the definition of
integral, we have
Z Z Z Z
dωi = dφ∗i ωi = φ∗i ωi = ωi .
M Ui ∂Ui ∂M
¤
5. INVARIANT COHOMOLOGY AND THE COHOMOLOGY OF COMPACT LIE GROUPS 132
5. Invariant cohomology and the cohomology of compact Lie groups
Let G be a Lie group, M a smooth manifold and µ : G × M −→ M a smooth action.
5.1. Definition. A differential form ω ∈ Ωp (M ) is µ-invariant, or simply invariant when clear from the
context, if µ∗g ω = ω, ∀ g ∈ G. We will denote by Ωµ (M ) the space of invariant differential p-forms.
Here two simple facts whose proof is left to the reader.

• The exterior product of invariant forms is invariant.
• If ω ∈ Ωpµ (M ), dω ∈ Ωµp+1 (M )
In particular the sequence
dp−1 dp
· · · −→ Ωµp−1 (M ) −→ Ωpµ (M ) −→ Ωp+1
µ (M ) −→ · · ·
is well defined and it is a subcomplex of the de Rham complex of M , the invariant de Rham complex.
5.2. Definition. The cohomology of the invariant de Rham complex is called the invariant cohomology
of M and will be denoted by Hµp (M ).
5.3. Example. A particular but important case is when M = G and the action is by left translations.
In this case an invariant form is called a left invariant form and the space of left invariant p-forms will be
denoted by ΩpL (G). We can also consider the action of G on itself by right translations, µ : G × G −→
G, µ(g, h) = hg −1 (the inverse is needed for the map satisfy the conditions to be an action). Differential
p-forms invariant for this action will be called right invariant forms and the space of such forms will be
denoted by ΩpR (G).
A left (resp. right) invariant p-form is uniquely determined by its value at the identity. More precisely
the evaluation map ev : ΩpL (G) −→ Λp (G),b ev(ω) = ω(e) is an isomorphism of algebras (with respect to
the exterior products). Since a left invariant form is constant on left invariant vector fields, the differential
of ω ∈ ΩpL (G) is given by
X c̃ , . . . , X
c̃ , . . . , X̃ )
dω(X0 , . . . , Xp ) = ω([X̃i , X̃j ], X̃0 , . . . , Xi j p
i<i
where X̃k is the left invariant extention of Xk (observe that, since dω is a form, dω(X0 , . . . , Xp ) does not
b the invariant de Rham complex
depend on the extensions of the Xk ’s). If we identify Ωp (G) with Λp (G),
L
becomes
p−1 p
b d−→ Λp (G)
· · · −→ Λp−1 (G)
d
b −→ b −→ · · ·
Λp+1 (G)
where
X
dp ω(X0 , . . . , X p ) = ci , . . . , X
ω([Xi , Xj ], X0 , . . . , X cj , . . . , Xp ).
i<i
b
The cohomology of the complex above is called the cohomology of G.
We want to compare the invariant cohomology with the usual de Rham cohomology. For this we need
to introduce integration of functions on a Lie group.
Let G be a n-dimensional Lie group. A left invariant n-form is uniquely determined up to a multiplicative
constant. Let ω 6= 0 be such a form, fixed once for all. For a smooth function f : G −→ R, with compact
support, we define the Haar integral of f , Z Z
f := f ω.
G G
The Haar integral is well defined since a Lie group is orientable and it is linear on the space of compactly
supported smooth real valued functions. Moreover, as it easily seen, the integral is left invariant, i.e.
Z Z
f= f ◦ Lg .
G G
We will be interested in the case of compact Lie groups. For such a group we can choose ω such that
Z
ω = 1.
G
5.4. Lemma. If G is a compact Lie group, the Haar integral is right invariant.
Proof. Consider the right action Rg (x) = xg −1 . Since the left and right actions commutes, we have
L∗h (Rg∗ ω) = Rg∗ (L∗h ω) = Rg∗ ω
since ω is left invariant. Then Rg∗ ω is left invariant, hence, by unicity of left invariant n-forms (up to constant),
Rg∗ ω = c(g)ω, where c : G −→ R \ {0}. Since Rgh
∗
= Rh∗ ◦ Rg∗ , c is a homomorphism into the multiplicative
group R \ {0}. But the only compact subgroups of this group are {1}, {±1}. Since G is connected, c(G) = 1.
Hence Z Z Z Z Z
f ◦ Rh = (f ◦ Rh )Rh∗ ω = Rh∗ (f ω) = fω = f.
G G G G G
¤
Let µ : G × M −→ M be a smooth action, G compact. Using integration we can “average” a differential

p-form to obtain an invariant one.For g ∈ G, X ∈ H(M ), we set X g := dµg (X). Consider the operator
Z
I : Ωp (M ) −→ Ωp (M ), I(ω)(X1 , . . . , Xp ) := ω(X1g , . . . , Xpg ).
G
5.5. Proposition.
(1) I(ω) ∈ Ωpµ (M ).
(2) If ω ∈ Ωpµ (M ), I(ω) = ω.
(3) dI(ω) = I(dω), i.e I is a morphism of the complex Ω∗ (M ) into the complex Ω∗µ (M ).
Proof. For the first claim we have

Z Z
µ∗g (I(ω))(X1 , . . . , Xp ) = I(ω)(X1g , . . . , Xpg ) = ω(X1gh , . . . Xpgh ) = ω(X1h , . . . , Xph ) = I(ω)(X1 , . . . , Xp ).
G G
The second claim is immediate and for the last one we have
p
X X
d(I(ω))(X0 , . . . , Xp ) = ci , . . . Xp )+ (−1)i+j I(ω)([Xi , Xj ], . . . , X
(−1)i Xi I(ω)(X0 , . . . , X ci , . . . , X
cj , . . . , Xp ) =
i=0 i<j
p
X Z X Z
= (−1)i Xi cg , . . . X g +
ω(X0g , . . . , X (−1)i+j cg , . . . , X
ω([Xi , Xj ]g , . . . , X cg , . . . , X g ) =
i p i j p
i=0 G i<j G
5. INVARIANT COHOMOLOGY AND THE COHOMOLOGY OF COMPACT LIE GROUPS 134
Z
= dω(X0g , . . . , Xpg ) = I(dω)(X0 , . . . , Xp ).
G
¤
The group G acts on H p (M ) in a natural way, g[ω] = [µ∗g ω]. We will denote by [H p (M )]µ the fixed
points set of this action, i.e. [H p (M )]µ = {[ω] ∈ H p (M ) : g[ω] = [ω] ∀ g ∈ G}.
5.6. Theorem. The inclusion J : Ωpµ (M ) −→ Ωp (M ) induces an isomorphism
J∗ : Hµp (M ) −→ [H p (M )]µ .
Proof. First notice that J is a chain map, i.e. d ◦ J = J ◦ d and hence induces a morphism J∗ :
Hµp (M ) −→ [H p (M )]. Since I ◦ J = 11, I∗ ◦ J∗ = 11 and J∗ is injective. We will show that ImJ∗ = [H p (M )]µ .
Let [α] ∈ [H p (M )]µ . Then [α] = [µ∗g α] and hence α − µ∗g α = dβ for some β depending on g. For any
cycle c ∈ Zp (M ) and any g ∈ G we have
Z Z Z
α− µ∗g α = dβ = β = 0,
c c ∂c
by Stokes Theorem. Hence

Z Z Z Z Z Z Z Z Z Z
I(α) = µ∗g α = µ∗g α = α= α 1 = α.
c c G G c G c c G c
Therefore, by the Theorem of de Rham, [I(α)] = [α]. Observe now that JI(ω) = I(ω). Hence J ∗ ([I(ω)]) =
[Iω] = [ω]. ¤
5.7. Lemma. Il G is connected, [H p (M )]µ = H p (M ).
Proof. Let g ∈ G and let γ : [0, 1] −→ G be a curve with γ(0) = g, γ(1) = e. Then H(x, t) := µγ(t) (x)
is a homotopy between µg and 11. Hence µ∗g = 11, in cohomology, and the action of G on H p (M ) is trivial. ¤
5.8. Corollary. If G is a compact connected Lie group, its cohomology is isomorphic to the cohomology
of its Lie algebra.
Proof. Let L denote the action by left translations. Since G is connected, [H p (G)]L = H p (G) and
HLp (G) b
is the cohomology of G. ¤
In particular, the cohomology of a compact connected Lie group can be computed by purely algebraic
means. In fact we can do even better. Consider the action
κ : (G × G) × G −→ G, κ((g1 , g2 ), h) = g1 hg2−1 .
If G is compact and connected, the cohomology of G is the cohomology of the complex of κ-invariant forms.
Such forms are left and right invariant and therefore are determined by their values at e ∈ G. We extend
the adjoint representation to act on forms
b −→ Λp (G),
Ad : G × Λp (G) b Ad(g)ω(X1 , . . . , Xp ) = ω(Ad(g −1 )X1 , . . . , Ad(g −1 )Xp ).
b is κ-invariant if and only if it is Ad-invariant. We want a characterization

It is easily seen that ω ∈ Λp (G)
of Ad-invariant forms. First a Lemma of general character.
5.9. Lemma. Let E be a vector space, G be a Lie group and θ : G −→ End(E) be a representation. Then
θ(g)v = v, ∀ g ∈ G if and only if dθ(e)(X)v = 0 ∀X ∈ G. b
Proof. Fix X ∈ G. b Then θ(exp tX) is a 1-parameter subgroup of End(E), hence θ(exp tX) = exp tA :=
P∞ k −1 k
k=0 t [k!] A for some A ∈ End(E). Therefore
d
dθ(e)(X)v =
|t=0 θ(exp tX)v = Av.
dt
d b Conversely, if Av = 0, exp(tA)v = v
If v is fixed by θ(g), ∀ g ∈ G, dθ(e)(X)v = v = 0, ∀X ∈ G.
dt
which implies θ(g)v = v, ∀g ∈ G. ¤
b is Ad-invariant if and only if

5.10. Theorem. ω ∈ Λp (G)
p
X
ω(X1 , . . . , Xi−1 , [Y, Xi ], Xi+1 , . . . , Xp ) = 0.
i=1
b Then
Proof. Let θ be the adjoint action on E = Λp (G).
d d
dθ(e)Y ω(X1 , . . . , Xp ) = |t=0 θ(exp Y )ω(X1 , . . . , Xp ) = |t=0 ω(Ad(exp(−tY )X1 , . . . , Ad(exp(−tY )Xp ) =
dt dt
Xp Xp
d
= ω(X1 , . . . , |t=0 Ad(exp(−tY )Xi , . . . , Xp ) = ω(X1 , . . . , ad(−Y )Xi , . . . , Xp ) =
i=1
dt i=1
p
X
= ω(X1 , . . . , [Xi , Y ], . . . , Xp ),
i=1
where the third equality follow from the fact that ω is multilinear. Then the Theorem follows from the
preceding Lemma.
¤
To compute the cohomology of G from the complex of Ad-invariant forms we need to compute the
differential of such a form.
b is Ad-invariant, then dω = 0.
5.11. Proposition. If ω ∈ Λp (G)
Proof. Set ²ii = 0, ²ij = (−1)j , i < j, ²ij = (−1)j+1 , i > j. Then
X
dω(X0 , . . . , Xp ) = 2−1 ci , . . . X
(−1)i ²ij ω([Xi , Xj ], X0 , . . . , X cj , . . . , Xp ) =
i6=j
X X
= 2−1 (−1)i ci , . . . X
²ij ω(([Xi , Xj ], X0 , . . . , X cj , . . . , Xp ) = 0,
i j
by Theorem 5.10. ¤
5.12. Corollary. The cohomology of a compact connected Lie group G is isomorphic to the space of
Ad-invariant forms.
5.13. Remark. If G is not connected, we can apply the Corollary to the connected component of the
identity, Ge (which is a compact connected subgroup, see Exercise 10.22 of Chapter 3). Since all connected
components are diffeomorphic, the cohomology of G is the direct product of as many copies of the cohomology
of Ge as many connected components there are.
6. EXERCISES 136
We will give now an application of Corollary 5.12. First a Definition.
b defined by
5.14. Definition. Let G be a Lie group. The derived algebra is the subalgebra DG of G
b : Z = [X, Y ], X, Y ∈ G}.
DG = {Z ∈ G b
5.15. Proposition. Let G be a compact, connected Lie group. Then

b
(1) H 1 (G) = {0} if and only if DG = G.
(2) If H 1 (G) = {0}, then H 2 (G) = {0}.
b there is a nonzero 1-form ω ∈ Λ1 (G)

Proof. If DG 6= G b vanishing on DG . By Proposition 5.11, ω is
Ad-invariant and defines a nonzero cohomology class. Conversely, if there is a nonzero Ad-invariant form,
b
this form vanishes on DG , hence DG 6= G.
b and let ω be an Ad-invariant 2-form. Then
Suppose now DG = G
0 = dω(X, Y, Z) = −ω([X, Y ], Z) − {ω([Z, X], Y ) + ω(X, [Z, Y ])} = −ω([X, Y ], Z),

b ω = 0 and H 2 (G) = {0}.
where the last equality follows from Proposition 5.11. Since DG = G, ¤
In the next Chapter we will give a “geometric” characterization of Ad-invariant forms (see Remark ??).
6. Exercises
6.1. Prove that if M is a compact manifold, dim(H p (M )) is finite.
6.2. Prove the Künnet formula for smooth manifolds (Theorem 2.5).
6.3. Prove that, if a manifold can be covered by two local charts such that the intersection of the domains
is connected, then it is orientable. Use this fact to give an alternative proof that S n , n > 1, is orientable.
6.4. The (open) Möebius band M is defined as the quotient of the square {(t, s) ∈ R2 : t ∈ [0, 1], s ∈ (0, 1)}
modulo the equivalence relation generated by (0, s) ∼ (1, 1 − s). Prove that M admits a smooth atlas
consisting of two charts, but it is not orientable. Find an embedding i : M −→ R3 .
6.5. Let U, V ⊆ Rn be connected open set and F : U −→ V be a diffeomorphism. Let ω ∈ Ωn (V ) be a

form with compact support. Prove that:
Z Z
ω=± F ∗ ω,
F (U ) U
with the sign + (resp. −) if F preserves (resp. inverts) the orientation. Extend the result to the case of
smooth oriented manifolds.
6.6. Let M be a compact, orientable n-dimensional manifold and p ∈ M . Let φ : Rn −→ M, φ(0) = p

be a local chart. Let f : Rn −→ R be a smooth non negative function with compact support and f (0) 6= 0.
Consider the form ω = f dx1 ∧ · · · ∧ dxn ∈ Ωn (Rn ). Prove that (φ−1 )∗ ω extends tho a smooth form in Ωn (M )
which represent a nonzero element of H n (M ).
Remark It can be shown that, if M is a compact connected smooth manifold, H n (M ) ∼
= R if M is
orientable and H n (M ) = {0} if M is not orientable. The first claim follows from Poincaré duality, a fact
that we will prove in the next Chapter (Theorem 3.9).
6.7. Prove that the antipodal map A : S n −→ S n , A(x) = −x is orientation preserving if and only if n
is odd. Prove that A∗ = (−1)n+1 : H n (S n ) ∼
= R −→ H n (S n ) ∼
= R.
6.8. Consider the vector field X : S 2k−1 −→ T S 2k−1 , X(x1 , . . . , x2k ) = (−x2 , x1 , . . . , −x2k , x2k−1 ) (see
Example 1.32). Prove that H(x, t) = (cos πtx + sin πtX(x)) is a homotopy between A and 11. Conclude that
A ∼ 11 if and only if n is odd, and, if n is even, every vector field on S n has a zero.
CHAPTER 5
Harmonic forms and the Theorem of Hodge
In this Chapter we will introduce the Laplace-Beltrami operator, a generalization at level of differentiable
p-forms of the classical Laplacian of functions. Hence we will have the concept of harmonic forms, i.e. forms
in the kernel of this operator and we will give a very rough idea of the proof of the classical Theorem of
Hodge which states that in each cohomology class of a compact Riemannian manifold, there is a unique
harmonic form. We will give also some applications of this Theorem to the topology of manifold with
“positive curvature”.
1. Some basic facts in Riemannian geometry
In this section we will describe the basic concepts and results in Riemannian geometry that we will use
in this Chapter. It is not intended to be an introduction to the subject.
1.1. The Levi-Civita connection of a submanifold. For vector fields X, Y ∈ H(RN ), we will denote
by ∇Y X the usual directional derivative, i.e. if γ : (−², ²) −→ RN is a smooth curve with γ(0) = p, γ̇(0) =
Y (p),
d
∇Y X(p) = |t=0 X(γ(t)).
dt
1.1. Remark. It is well known that ∇Y X(p) depends only on Y (p) and the values of X along any curve
γ : (−², ²) −→ RN with γ̇(0) = Y (p).
Let M ⊆ RN be an n-dimensional submanifold and X, Y ∈ H(M ). By Remark 1.1 we can compute

∇Y X(p), since we only have to know X along a curve γ with γ̇(0) = Y (p) and this curve can be taken in
M . This is not, in general, a tangent vector to M . But we can project it on Tp M .
1.2. Definition. We define the covariant derivative, at p, of X in the direction Y as
∇Y X(p) = Pp> (∇Y X(p)),

where Pp> : RN −→ Tp M is the orthogonal projection.
1.3. Remark. ∇Y X is smooth, as a function of p, hence ∇Y X ∈ H(M ). Hence we have a R-bilinear

map
∇ : H(M ) × H(M ) −→ H(M ), ∇(Y, X) := ∇Y X
.
1.4. Definition. The operator ∇ is called the Levi-Civita (or Riemannian) connection of M
1.5. Proposition. The Levi-Civita connection verify the following properties:

139
1. SOME BASIC FACTS IN RIEMANNIAN GEOMETRY 140
(1) ∇ is R-bilinear and F(M ) linear in the first variable,

(2) if f ∈ F(M ) then ∇Y f X = Y (f )X + f ∇Y X,
(3) ∇Y X − ∇X Y = [X, Y ],
(4) Y hX, Zi = h∇Y X, Zi + hX, ∇Y Zi.
Proof. The properties are obvious for ∇ and are preserved by Pp> : RN −→ Tp M . ¤
Those properties characterize the Levi-Civita connection in view of the following result, sometimes called
the Fundamental Theorem of Riemannian Geometry:
1.6. Theorem. There exist a unique operator ∇ : H(M ) × H(M ) −→ H(M ) with the properties of
Proposition 1.5
˜ : H(U ) × H(U ) −→ H(U ) be an operator satisfying the properties above. For X, Y, Z ∈
Proof. Let ∇
H(M ) we have:
˜ X Y, Zi + hY, ∇
XhY, Zi = h∇ ˜ X Zi,
˜ Y Z, Xi + hZ, ∇
Y hZ, Xi = h∇ ˜ Y Xi,
˜ Z X, Y i + hX, ∇
ZhX, Y i = h∇ ˜ Z Xi.
Adding the first two and subtracting the last one we have
˜ X Z, Y i = XhY, Zi + Y hX, Zi − ZhY, Xi + h[X, Y ], Zi − h[X, Z], Y i + h[Z, Y ], Xi.
2h∇
˜ = ∇.
On the other hand the same formula holds for the Levi-Civita connection ∇, hence ∇ ¤
1.7. Remark. The formula obtained in the proof of Theorem 1.6 is called the Kozul formula. An
important observation is that the right hand side of the Kozul formula depends only on the fact that we
have a scalar product on each tangent space and on the differential topological properties of M (Lie product
of vector fields).
1.8. Lemma. ∇Y is a local operator, i.e. if two vector fields X, Z coincide on an open set U ⊆ M then
∇Y X = ∇Y Z in U .
Proof. The statement is true for ∇ and preserved by P > , hence is true for ∇. ¤
1.9. Remark. Lemma 1.8 allows us to take covariant derivatives of locally defined vector fields.
∂
Let φ : Σ ⊆ Rn −→ U = φ(Σ) ⊆ M be a local chart and the associated coordinate vector fields.
∂xi
Define the functions Γkij : U −→ R by the condition:
X n
∂ ∂
∇ ∂ = Γkij .
∂xj ∂xk
k=1
∂xi
1.10. Definition. The functions Γkij are called the Christoffel symbols of the connection.
It is convenient to define the covariant derivative of vector fields along a curve. Let γ : (a, b) ⊆ R −→ M
be a smooth curve. We recall that a vector field along γ is a smooth map X : (a, b) −→ RN such that
X(t) ∈ Tγ(t) M, ∀ t ∈ (a, b). We will denote by H(γ) the space of such vector fields.
CHAPTER 5. HARMONIC FORMS AND THE THEOREM OF HODGE 141
∇
1.11. Theorem. There is a unique R-linear operator : H(γ) −→ H(γ) such that:
dt
∇ d ∇
(1) if X ∈ H(γ), f ∈ F((a, b)), f X = [ f ]X + f X,
dt dt dt
d ∇ ∇
(2) hX(t), Y (t)i = h X(t), Y (t)i + hX(t), Y (t)i, ∀X, Y ∈ H(γ),
dt dt dt
∇
(3) if X extends to a vector field in a neighborhood of γ(t), then X = ∇γ̇(t) X.
dt
∇ > d
Proof. It is clear the the operator X(t) = Pγ(t) ( X(t)) verifies the properties in question. Suppose
dt dt
∇˜
that is an other operator with the same properties. It is easily seen that it is a local operator in the
dt
sense that if two vector fields along γ coincide in an open neighborhood of t0 , so do their images. So we
can work in a coordinate neighborhood. Let {x1 , . . . , xn } be local coordinates in that neighborhood, and
γ(t) = {x1 (t), . . . , xn (t)} the local coordinate expression of γ. Then
n
X ˜ Xn
∂ ∇ d ∂ ∂
X(t) = Xi (t) and X= [ Xi ] + Xi ∇γ̇(t) .
i=1
∂xi dt i=1
dt ∂x i ∂x i
∇ ˜
∇ ∇
Since the same holds for X, we have X = X. ¤
dt dt dt
∇
1.12. Remark. We stress the difference between ∇γ̇ and . In fact a vector field along γ may not be,
dt
even locally, the restriction of a vector field in M . The extreme case is when γ is a constant curve and X is
∇
a non constant vector field along γ. Then γ̇ = 0 but, in general, X 6= 0.
dt
1.13. Remark. We can take also covariant derivatives of exterior forms. For a 1-form ω ∈ Ω1 (M ),
and X ∈ Tp M , we have the dual vector field Y = ] ω (] is the inverse of the “musical isomorphism”
[ : Tp M −→ [Tp M ]∗ ) and it is natural to define ∇X ω to be dual of ∇X Y . This lead to the formula:
[∇X ω](W ) = (]∇X Y )(W ) = h∇X Y, W i = XhY, W i − hY, ∇X W i = Xω(W ) − ω(∇X W ).
Next we observe that any ω ∈ Ωp (M ) is, locally, sum of decomposable forms, i.e. exterior products of
1-form. On such decomposable form we can extend ∇X to act as a derivation, i.e. ∇X (ω1 ∧ · · · ∧ ωp ) =
Pp
k=1 ω1 ∧ · · · ∧ ∇X ωk ∧ · · · ωp . A direct calculation gives the expression:
p
X
∇X ω(X1 , . . . , Xp ) = X(ω(X1 , . . . , Xp )) − ω(X1 , . . . , Xi−1 , ∇X Xi , Xi+1 , . . . , Xp ).
i=1
It is easy to check that ∇X ω, given by the formula above, is an alternating F(M ) multilinear map,
hence a differential form. So the formula above can be taken as definition of ∇X ω.
1.2. The curvature tensor.
1.14. Definition. The curvature of the Levi-Civita connection is the map
R(X, Y )Z := ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z, X, Y, Z ∈ H(M ).
1.15. Remark. Some authors use the opposite sign in the definition of R.
The basic property is that R is a tensor, i.e.:

1.16. Lemma. R is F(M )-trilinear.
Proof. R is clearly R-trilinear. The linearity with respect to F(M ) follows by a direct calculation
using the properties of the connection and of the Lie product. ¤
1.17. Remark. Lemma 1.16 implies that the value of R(X, Y )Z at a given point p ∈ M depends only
on the value of X, Y, Z at that point, i.e. to compute [R(X, Y )Z](p) we can use arbitrary extensions of
X(p), Y (p) and Z(p) (see Theorem 1.2 of Chapter 1). For example, if M is an open set of Rn we can
extends vectors to constant vector fields and therefore we have the expected fact that the curvature of an
open subset of Rn is zero.
The curvatures tensors have interesting symmetries:
1.18. Proposition. For X, Y, Z, W ∈ H(M ) we have:

(1) R(X, Y )Z = −R(Y, X)Z,
(2) hR(X, Y )Z, W i = −hR(X, Y )W, Zi,
(3) R(X, Y )Z + R(Y, Z)X + R(Z, X)Y = 0 (first Bianchi identity),
(4) hR(X, Y )Z, W i = hR(Z, W )X, Y i.
Proof. (1) follows directly from the definition. (2) is equivalent to hR(X, Y )Z, Zi = 0. The latter
follows from a direct calculations taking in account
1 1
h∇X ∇Y Z, Zi = Xh∇Y Z, Zi − h∇Y Z, ∇X Zi = XY hZ, Zi − h∇Y Z, ∇X Zi, h∇[X,Y ] Z, Zi = [X, Y ]hZ, Zi.
2 2
Next we prove (3).
R(X, Y )Z + R(Y, Z)X + R(Z, X)Y = ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z+

+∇Y ∇Z X − ∇Z ∇Y X − ∇[Y,Z] X + ∇Z ∇X Y − ∇X ∇Z Y − ∇[Z,X] Y =
= ∇Y [Z, X] + ∇Z [X, Y ] + ∇X [Y, Z] + ∇[X,Z] Y + ∇[X,Y ] Z + ∇[Z,Y ] X =
= [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0,
where the last equality is the Jacobi identity for the Lie product of vector fields (see ? Chapter ?).
We will sketch now the proof of (4), leaving the details to the reader. Consider the Bianchi identity
hR(X, Y )Z, W i + hR(Y, Z)X, W i + hR(Z, X)Y, W i = 0.
Permuting cyclically the vectors X, Y, Z and W , we obtain three more equations. Summing the four equations
we obtain
hR(X, Z)Y, W i + hR(Y, W )Z, Xi = 0
and the conclusion follows from (2). ¤
It will be useful, later on, to look at the curvature tensor in a slight different way. For p ∈ M we define
ρp : Λ2 (Tp M ∗ ) −→ Λ2 (Tp M ∗ ), hρ(X ∧ Y ), Z ∧ W i = hR(X, Y )W, Zi.
By Proposition 1.18, ρp is a well defined linear symmetric endomorphism of Λ2 (Tp M ∗ ).
1.19. Definition. The operator ρ is called the curvature operator of M .

Associated to the curvature tensor there are several “measures of curvature”. We will describe three of
the most important.
• The sectional curvature, K. This is a function defined, in principle, on pairs of linearly independent
vectors tangent at the same point
hρ(X ∧ Y ), X ∧ Y i hR(X, Y )Y, Xi
K(X, Y ) := = .
kX ∧ Y k2 kXk2 kY k2 − hX, Y i2
It follows from Exercise 7.6 of Chapter 1, that K(X, Y ) depends only on the plane spanned by X and Y .
The sectional curvature K is then a real valued function defined on 2-dimensional tangent subspaces.
• The Ricci curvature, Ricc. For fixed X, Y ∈ Tp M consider the linear map LX,Y : Tp M −→
Tp M, LX,Y (Z) = R(Z, X)Y . The Ricci tensor is defined as
Q̃(X, Y ) = trace LX,Y .
By Proposition 1.18, Q̃ is a symmetric tensor. The Ricci curvature is defined as the quadratic form
Ricc(X) = Q̃(X, X).
A couple of observations will be useful. Fix an orthonormal basis {X1 , . . . , Xn }. Then the map
X
Q : Tp M −→ Tp M, Q(X) = R(X, Xi )Xi
Pn
is well defined and Q̃(X, Y ) = hQ(X), Y i. Also, if X = X1 , Ricci(X) = 2 K(X, Xi ).
The Ricci curvature is then a real valued funtion defined on the tangent bundle.
• The scalar curvature S. This is a function defined on the manifold M as the average of the Ricci
curvatures. If X1 , . . . Xn is an orthonormal basis, then
n
X
S = traceQ = Ricc(Xi ).
1
The scalar curvature is a real valued function defined on the manifold.
1.20. Remark. If n = 2 the sectional curvature coincides with the scalar curvature and is usually called
the Gaussian curvature of the surface.
1.21. Remark. Some authors define the Ricci and scalar curvatures dividing our expressions by n − 1
and n(n − 1) respectively.
1.3. Parallel tranlation and geodesics. Let γ : [a, b] ⊆ R −→ M be a smooth curve. We have
∇
introduced the operator: : H(γ) −→ H(γ) of covariant derivative along γ. When no confusion will arise,
dt
∇
we will set X = X 0 .
dt
1.22. Definition. Let γ : [a, b] ⊆ R −→ M be a smooth curve. A vector field X ∈ H(γ) is parallel along
∇
γ if X 0 := X = 0.
dt
1.23. Remark. If M is an open set of Rn , a vector field is parallel if and only if it is constant. But
in general we should not confuse parallel fields in the manifold with parallel field in the ambient Euclidean
space. For example, if M = S n , the unit sphere in Rn+1 , and γ(t) = sin(t)e1 + cos(t)e2 , the tangent vector
γ̇(t) is parallel along γ, in the sense of the definition, but not parallel in Rn+1 . However parallelism has
a “mechanical interpretation”. Suppose, for simplicity, we have a surface M ⊆ R3 , a curve γ in M and a

vector field X along γ. We can “roll” the surface on a plane along the curve obtaining a curve in the plain
and a vector field along this curve. Then the original vector field is parallel if and only if the the vector field
obtained by this process is parallel in the sense of the plain, i.e. is constant. The reader is invited to think
in the example above: we can “roll” the sphere S 2 on a plane in such a way that the curve γ correspond to
a straight line and the tangent vector field to the tangent to the line. Then the latter is parallel in the sense
of the plain. These intuitive arguments can be made precise.
1.24. Remark. It is convenient to have an expression of the equation of parallelism in terms of local
coordinates. Let γ(t) = (x1 (t), . . . , xn (t)) be the expression of the curve in local coordinates and let X(t) =
P ∂
Xi (t) ∈ H(γ). Then
∂xi
X ∂ X ∂ X X ∂
X 0 (t) = Ẋk + Xi ∇ X ∂ = [Ẋk + ẋi Xj Γkij ] .
∂xk i ( ẋj ) ∂xi i,j
∂xk
k k
j
∂xj
The equation X 0 = 0 is a system of ordinary differential equations. For such systems we have we have
(local) existence and unicity1 for the solutions with given initial conditions. We also have smooth dependence
on the initial conditions. In our case, since the system is linear, the solutions are defined on the entire interval
of definition of the coefficients.
1.25. Proposition. Let γ : [a, b] −→ M be a smooth curve and Xa ∈ Tγ(a) M , The there exists a unique
vector field X ∈ H(γ)) such that X 0 = 0 and X(a) = Xa .
Proof. Cover γ([a, b]) with a finite number of charts {Ui , φi }. Then γ −1 φi (Ui ) is an open covering of
[a, b]. Then there exist a subdivision t0 = a < t1 < · · · < tk = b such that γ([ti , ti+1 ]) ⊆ φi (Ui ). By Remark
1.24 there is a unique parallel vector field X1 ∈ H(γ|[t0 , t1 + ²)), ² > 0, with X(a) = Xa . By the same
reason there is a parallel vector field X2 ∈ H(γ|(t1 − ², t2 + ²)) with X2 (t1 ) = X1 (t1 ). The two vector fields
coincide in (t1 − ², t1 + ²) since in this interval they are parallel and coincide at t1 . So we have a well definite
smooth parallel vector field in [a, t2 + ²). Proceeding in the same way, we obtain, after a finite number of
steps, a vector field X ∈ H(γ) with the required properties. ¤
Proposition 1.25 allows us to define a map
Pγ : Tγ(a) M −→ Tγ(b) ,
which associate to a given vector Xa ∈ Tγ(a) M the value of the parallel translation of Xa at the value t = b
of the parameter. This map is called parallel translation along γ. Pγ is linear, since the equation X 0 = 0 is
linear, and it is an isometry since, if X, Y ∈ H(γ) are parallel vector fields,
d
hX(t), Y (t)i = hX 0 (t), Y (t)i + hX(t), Y 0 (t)i = 0.
dt
1.26. Remark. The word “connection” came from the fact that, using parallel translation, we can
“connect” tangent spaces at different points. We should point out, however, that this “connection” depends
strongly on the curve.
1Unicity has to be intended as the fact that two solutions with the same initial condition coincide in the intersection of
the domains of definition.

The next results link parallel translation with curvature and will be of use later on. First some prelimi-
naries. Let U ⊆ R2 and F : U −→ M a smooth map (sometimes called a surface). A vector field along F is a
smooth map X : U −→ T M such that X(t, s) ∈ TF (t,s) M . We will denote by H(F ) the vector space of vector
∇ ∇
fields along F . We have the operators , , the covariant derivative along the curves F (t, s0 ), F (t0 , s).
dt ds
1.27. Definition. A vector field X ∈ H(F ) is parallel along F if
∇ ∇
X=0= X.
dt ds
The existence of parallel vector fields along surfaces is not at all obvious since the equations involved
are not any more ordinary differential equations, but they are partial differential equations. The results we
are after are integrability conditions for the existence of solutions for such differential equations.
1.28. Lemma. If X ∈ H(F ),

∇∇ ∇∇ ∂F ∂F
X− X = R( , )X.
dt ds ds dt ∂t ∂s
Proof. Compute both sides in local coordinates. ¤
∂F ∂F
Suppose that X ∈ H(F ) is parallel. Then, by Lemma 1.28, R( , )X = 0. So the vanishing of some
∂t ∂s
curvatures is a necessary condition for the existence of parallel vector fields alon a surface. We will show
next that, in some sense, they are also sufficient.
Let F : R2 −→ M be a surface and let X0 be a vector in TF (0,0) M . The natural way to try to
construct a parallel vector field along F , with initial condition X0 is to parallel translate X0 along the curve
σ(t) = F (t, 0) obtaining a vector field X(t, 0), parallel along σ. Then parallel translate X(t, 0) along the
curve γt (s) = F (t, s) (t fixed). In such a way we obtain a vector field X ∈ H(F ) with X(0, 0) = X0 .
1.29. Theorem. The vector field X defined above is parallel along F if
∂F ∂F
R( , )X = 0.
∂t ∂s
Proof. We have seen that the condition is necessary. We will show that it is sufficient. By construction
∇ ∇
X = 0 so we have to show that X = 0. This is true, again by construction, along σ. Moreover
ds dt
∇∇ ∇∇ ∂F ∂F ∇∇
X= X − R( , )X = X = 0.
ds dt dt ds ∂t ∂s dt ds
∇ ∇
Therefore X is parallel along the curves γt . But, along these curves, X vanishes at s = 0, hence
dt dt
∇ ∇
X = 0 along the all curve γt . Therefore X = 0. ¤
dt dt
A concept very much related to the one of parallel translation is the one of geodesic.
1.30. Definition. A curve γ : (a, b) ⊆ R −→ M is a geodesic if γ̇ is parallel along γ.
1.31. Remark. If γ is a geodesic, γ̇ is parallel along γ and therefore has constant norm, since parallel
translation is an isometry. It follows that the geodesics came with a special parametrization, a parametrization
proportional to arc length, i.e. kγ̇(t)k is constant.
1.32. Remark. It is useful to have a local coordinate expression for the geodesic equation γ̇ 0 = 0. Let
γ(t) = (x1 (t), . . . xn (t)) be the expression of γ in some local coordinate system. Then, from Remark 1.24,
the equation γ̇ 0 = 0 takes the form
X
ẍk + Γkij ẋi ẋj = 0, k = 1, . . . , n.
i,j
The equation above is a system of second order ordinary differential equations, easily reduced to a system
of the first order by introducing the auxiliary variable y = ẋ (in this way we double the number of equations).
Then we have existence, unicity and smooth dependence on the initial conditions. However the system is
not linear and we do not have existence in the large. A simple example is the following.
1.33. Example. Consider M = R2 \ {0}. The geodesic with initial condition p = (−², 0), X(p) = (1, 0)
is γX (t) = (−² + t, 0) and is defined for t < ².
1.34. Example. Consider the unit sphere in Rn+1 . Then the curve γ(t) = sin(t)e1 +cos(t)e2 , is a geodesic
(see Remark 1.23).
1.35. Remark. In general the integration of the geodesic equations is a quite hard task even for very
simple surfaces.
Let γX : [0, a) −→ M be the geodesic with initial conditions γ(0) = p, γ̇(0) = X. Fix s ∈ R and
consider the curve σ(t) = γX (st). Then σ is a solution of the geodesics equation with initial condition sX,
and it is defined in the interval [0, a/s). This observation leads to the following
1.36. Lemma. The exists ² > 0 such that, if kXk < ², the geodesic γX is defined for t = 1.
Proof. For Z ∈ Tp M , with kZk = 1, define r(Z) = sup{t ∈ [0, ∞) : γZ is defined in [0, t)}. The r is a
well defined function on the unit sphere of Tp M . Moreover, by the smooth dependence from the initial data,
r is continuous, hence has a positive minimum ² = ²(p). ¤
1.37. Definition. Let p ∈ M and ² as in Lemma 1.36. We define
expp : {X ∈ Tp M : kXk < ²} −→ M, exp(X) = γX (1).
1.38. Remark. Observe that expp (tX) = γX (t) (when defined). This means, geometrically, that expp
sends the ray {tX} ⊆ Tp M , “linearly”, onto the geodesic γX .
Again, by the smooth dependence, expp is a smooth function.
1.39. Lemma. d expp (0) = 11 (modulo the identification T0 Tp M ∼

= Tp M ).
Proof. d expp (0)X is the tangent vector to exp(tX) at t = 0, i.e. X. ¤
In particular, by the inverse function Theorem, expp is a diffeomorphism of a small ball V = {Y ∈ Tp M :

kY k < δ} onto an open neighborhood expp (V ) ⊂ M . In other words, (V, expp |V ) is a chart for M .
1.40. Definition. The set expp (V ) is called a normal neighborhood of p and the local coordinates are
called normal coordinates.
2
1.41. Remark. The name exponential come from the fact that, in the case of O(n) ⊆ Rn (see Example
P∞
? of Chapter ?), exp11 (X) = 0 X k X ∈ Skw(n, R) = T11 O(n). This is a particular case of a more general
fact that we will discuss in subsection 1.5
The following result is often useful to simplify computations.
1.42. Proposition. Let {E1 , . . . , En } be an orthonormal basis for Tp M . Then there exist local orthonor-
mal vector fields X1 , . . . , Xn such that Xi (p) = Ei , and
∇Xi Xj (p) = 0.
Proof. Let U be a normal neighborhood of p. At q = expp (Y ) we define Xi (q) as the parallel translated
of Ei along the curve expp (tY ). The vector fields Xi are smooth since geodesics and parallel translation
depends smoothly on the initial conditions. Moreover, since parallel translation is an isometry, they are
orthonormal. Finally
∇
∇Y Xi (p) = Xi (p) = 0,
dt
since Xi is parallel along exp(tY ). ¤
1.4. Riemannian manifolds. The careful reader has probably realized that all we have done up to
now depend only on the existence of a scalar product in each tangent space. This suggest the following
1.43. Definition. A Riemannian manifold is a differentiable manifold together with a scalar product
h·, ·ip on each tangent space Tp M , that depends smoothly on p. This means that, given smooth vector fields
X, Y ∈ H(M ), the function that associate to p ∈ M the number hX(p), Y (p)ip is smooth. When clear from
the context we will omit the subindex p.
We will call this assignment of scalar products a Riemannian metric
In a Riemannian manifold, we can define the Levi-Civita connection using the Kozul formula (see Remark
1.7), prove the basic properties stated in Proposition 1.5 and show the uniqueness as in Theorem 1.6.
Once we have the Levi-Civita connection, we can define the curvature tensor, parallel translation,
geodesics etc.. The proofs of the respective properties go just as in the case of submanifolds.
A priori the “universe” of Riemannian manifold is larger than the one of submanifolds of Euclidean
space. It turns out that this is not the case. We will discuss shortly this fact now.
1.44. Definition. Let M, N be Riemannian manifolds and f : M −→ N a smooth map.

• We will say that f is an isometry if it is a diffeomorphism and
hdf (X), df (Y )i = hX, Y i ∀ X, Y ∈ H(M ),
(we use the same symbol for the Riemannian metrics of M and N ).
• We will say that f is a local isometry if ∀ p ∈ M there exist an open neighborhood U ⊆ M of p such
that f (U ) is open and f |U is an isometry onto F (U ).
• We will say that f is an isometric immersion (resp. an isometric embedding) if f is an immersion
(resp. an embedding) and the formula above holds.
It follows from the unicity of the Levi-Civita connection that a isometry sends covariant derivative of
vector fields into the covariant derivative of the images via the differential. In particular it preserve curvature
and sends geodesics to geodesics. For the case of isometric immersions the image of a geodesic is a geodesic
of the image which is not, in general, a geodesic of the ambient space.
1.45. Remark. It is worthwhile to observe, for the case of submanifolds, that the existence of an isometry
between two submanifolds does not means that they sit similarly in the ambient space2. The classical example
is the strip S = {(t, θ, 0) ∈ R3 : θ ∈ (0, π)} and the cylinder C = {(cos θ, sin θ, t) ∈ R3 : θ ∈ (0, π)}. The
map F (t, θ, 0) = (cos θ, sin θ, t) is an isometry, but the two submanifolds do not differ by an isometry of
R3 . The way that a submanifold sits inside the ambient manifold is described by the so called second
fundamental form which is the difference between the connection in the ambient space and the connection
of the submanifold. If the second fundamental form vanish, the submanifold is called totally geodesic and,
in this case, the geodesics of the submanifold are geodesic of the ambient manifold (and viceversa).
In principle the category of Riemannian manifolds, up to isometries, seems larger that the one of sub-
manifolds of Euclidean spaces. However, as in the case of differentiable manifolds, this is not the case, due
to the following result of J. Nash (that we will not prove here):
1.46. Theorem. [Nash] Let M be a Riemannian manifold. Then there exist an isometric embedding
F : M −→ RN , for N sufficiently large.
1.47. Remark. If F : M −→ N is an isometric embedding, F (M ) is a submanifold of N and F : M −→

F (M ) is an isometry.
1.5. Lie groups. Let G be a Lie group.
1.48. Definition. A Riemannian metric on G is left (resp. right) invariant if the left (resp. right)
translations are isometries. A Riemannian metric is bi-invariant if it is left and right invariant.
b = Te G. The we can define a Riemannian metric on G setting

Fix an inner product h·, ·ie on G
hX, Y ig = hdLg−1 X, dLg−1 Y ie .
This metric is, essentially by definition, a left invariant metric. Let now {E1 , . . . , En } be an orthonormal
b Then the left invariant translations are an orthonormal basis of vector fields. If {ω1 , . . . , ωn }, n =
basis for G.
dim G is the dual basis, ω = ω1 ∧ · · · ∧ ωn is a left invariant form, called the volume form (see ?). Naturally
the same considerations work for right invariant objects. However, in general does not exist a bi-invariant
metric. A particular interesting case when such a metric exists is the following.
1.49. Proposition. Let G be a compact Lie group. Then there exist a bi-invariant metric on G.
Proof. Start with a left invariant metric and “average” it under right translations, i.e. define a new
metric by Z
hhX, Y ii = hdRg X, dRg Xidω.
G
¤
2i.e. there exist an isometry of the ambient space that takes one onto the other.
For Lie groups with bi-invariant metric, the relevant geometric objects as connection, geodesics and
curvature may be described in algebraic terms. For the rest of this subsection, G will denote a Lie group
with a bi-invariant metric (that we will denote by h·, ·i).
1.50. Lemma. Let X, Y, Z ∈ L(G). Then
h[X, Y ], Zi = hX, [Y, Z]i.
Proof. Let γ(t) = exp(tY ). Since the metric is bi-invariant
hX, Zi = hdRγ(−t) ◦ dLγ(t) X, dRγ(−t) ◦ dLγ(t) Zi = hAd(γ(t))X, Ad(γ(t))Zi = constant.

Differentiating with respect to t, and computing for t = 0, we have
0 = h[dAd](e)(Y )X, Ad(e)Zi+hAd(e)X, [dAd](e)(Y )Zi = had(Y )X, Zi+hX, ad(Y )Zi = h[Y, X], Zi+hX, [Y, Z]i.
1.51. Remark. It is not difficult to show that a left invariant metric for which the conclusion of Lemma
1.50 holds is actually a bi-invariant metric.
1.52. Theorem. Let X, Y, Z ∈ L(G). Then

(1) 2∇X Y = [X, Y ],
(2) 4R(X, Y )Z = [Z, [X, Y ]].
Proof. First we prove that ∇X X = 0. In fact
2h∇X X, Y i = hX, [Y, X]i = −hX, [X, Y ]i = 0,
where the first equality is the Kozul formula (see Remark 1.7) and the last one is by Lemma 1.50. In particular
0 = ∇X+Y (X + Y ) = ∇X Y + ∇Y X.
On the other hand

∇X Y − ∇Y X = [X, Y ].
Adding the two equation we have (1). In order to prove (2) we compute
R(X, Y )Z = ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z
using (1). Thus we obtain

1 1 1
R(X, Y )Z = [X, [Y, Z]] − [Y, [X, Z]] − [[X, Y ], Z]
4 4 2
and the conclusion follows from the Jacobi identity. ¤
1.53. Corollary. The geodesics of G are left translated of 1-parameter subgroups. In particular they
are defined for all values of the parameter3.
Proof. Let X ∈ L(G). Then the integral curves of X, which are translated of the 1-parameter subgroup
exp(tX), are geodesics since, by Theorem 1.52, ∇X X = 0. ¤
3Riemannian manifolds for which the geodesics are defined for all values of the parameter are called complete.
2. THE LAPLACE-BELTRAMI OPERATOR 150
1.54. Corollary. Let X, Y ∈ L(G) be orthonormal. Then the sectional curvature of the plane spanned
by X, Y is given by K(X, Y ) = k[X, Y ]k2 . In particular
(1) K(X, Y ) ≥ 0 and K(X, Y ) = 0 if and only if [X, Y ] = 0.
b is trivial.
(2) The Ricci curvature is non negative and positive if and only if the center of G
b be a
Proof. The claims for the sectional curvature come directly from Theorem 1.52 (2). Let X ∈ G
b Then
unit vector and let {X1 = X, . . . , Xn } be an orthonormal basis for G.
n
X n
X
Ricc(X) = K(X, Xi ) = k[X, Xi ]k2 ≥ 0.
2 2
b
Moreover Ricc(X) = 0 if and only if X commutes with all the Xi ’s, i.e. X is in the center of G. ¤
2. The Laplace-Beltrami Operator
We want to define the (geometers) Laplacian of a function on a Riemannian manifold. We proceed as

in the case of Rn (see Exercise 7.29 of Chapter 1). We have the gradient operator, ∇f = ]df , and we will
define the divergence of a vector field.
Let x ∈ M and X ∈ H(M ). Consider the linear transformation LX : Tx M −→ Tx M given by LX (Y ) =
∇Y X and define
div (X) = trace LX .
In particular, if {X1 , . . . , Xn } is an orthonormal basis,

n
X
div X = h∇Xi X, Xi i,
i=1
2.1. Definition. TheLaplacian is the operator
∆ : Ω0 (M ) = F(M ) −→ Ω0 (M ) = F(M ), ∆f = −div df.
2.2. Remark. Let {X1 , . . . , Xn } be orthonormal vector fields. Then

n
X
∆f = − Xs Xs (f ) + ∇Xs Xs (f ).
1
Pn
In particular if the Xi ’s are like in Lemma 1.42, ∆f (p) = − 1 Xs Xs (f )(p). Also the product formula in
Exercise 7.29 of Chapter? extend to our case.
2.3. Lemma. Given f, g ∈ F(M ), ∆(f g) = g∆f + f ∆g − 2h∇f, ∇gi.
We want to extend the Laplacian to an operator on p-forms. Since we have the differential we need to
define the divergence of a p-form.
For 1-forms ω, the divergence is defined as the divergence of the dual vector field ]ω (which is a function,
that is a 0-form). For p-forms we give the following
2.4. Definition. Let ω ∈ Ωp (M ), X1 , . . . , Xp−1 ∈ Tp M . Consider the bi-linear map LX1 ,...,Xp−1 (X, Y ) =
[∇X ω](Y, X1 , . . . , Xp−1 ). The divergence of ω is defined as the (p − 1)-form
div ω(X1 , . . . , Xp−1 ) = trace LX1 ,...,Xp−1 .
In particular, if {X1 , . . . , Xn } is an orthonormal basis of Tx M , we have the formula:

n
X
(div ω)(Xi1 , . . . , Xip−1 ) = (∇Xj ω)(Xj , Xi1 , . . . , Xip−1 ).
j=1
It is useful to have an expression of the differential in terms of the connection.
2.5. Lemma. If ω ∈ Ωp (M ),
p
X
dω(X0 , . . . Xp ) = (−1)i ∇Xi ω(X0 , . . . , X̂i , . . . , Xp ).
0
Proof. Let p ∈ M . Choosing suitable normal coordinates, we can extend the Xi0 s to vector fields such
that, at p, ∇Xi Xj (p) = 0. In particular, [Xi , Xj ](p) = 0. Since we are working with differential forms, the
values, at p, do not depend on the extensions. Hence, at p, formula (2.5) coincide with the one that defines
d (see Definition ?? of Chapter 3). ¤
2.6. Definition.
• The codifferential is the operator:
δ : Ωp (M ) −→ Ωp−1 (M ) δω = −div ω.
• The Laplace-Beltrami operator, or simply the Laplacian, is the operator:
∆ : Ωp (M ) −→ Ωp (M ) ∆ω = d ◦ δω + δ ◦ dω.
• A form ω ∈ Ωp (M ) is coclosed if δω = 0 and harmonic if ∆ω = 0.
2.7. Remark. If ω ∈ Ω0 , δω = 0 and the definitions of the Laplacian in 2.6 and 2.1coincide.
Suppose now that M is oriented and let ∗ : Ωp (M ) −→ Ωn−p (M ) be the Hodge operator (see ?? of
Chapter 1).
2.8. Proposition. div ω = (−1)n(p−1) ∗ d ∗ ω ∀ω ∈ Ωp (M ).
Proof. Let {X1 , . . . , Xn } be a positive orthonormal basis of Tp M and extend the X’s to local vector
fields such that (∇Xi Xj )(p) = 0 (see Remark 1.42). Then, up to terms vanishing at p,
n−p+1
X
(∗ div ω)(X1 , . . . , Xn−p+1 ) = div ω(Xn−p+2 , . . . Xn ) = Xk ω(Xk , Xn−p+2 , . . . , Xn ),
k=1
n−p+1
X
(d ∗ ω)(X1 , . . . , Xn−p+1 ) = (−1)k+1 ∇Xk (∗ω)(X1 , . . . , X̂k , . . . , Xn−p+1 )
k=1
n−p+1
X
= (−1)k+1+n−p+1−k Xk · ω(Xk , Xn−p+2 , . . . , Xn ).
k=1
Then ∗ div ω = (−1)n−p d ∗ ω. Applying ∗ to both sides we get the conclusion. ¤

2. THE LAPLACE-BELTRAMI OPERATOR 152
2.9. Remark. In principle, to make sense to the statement of Proposition 2.8, we need M to be oriented.
However, we observe that ∗ appears twice and therefore the formula does not depend on the choice of the
orientation.
2.10. Corollary. A form ω ∈ Ωp (M ) is closed (resp. coclosed) if and only if ∗ω is coclosed (resp.
closed).
Proof. Suppose dω = 0. Then, up to sign, δ ∗ ω = ∗ ∗ δ ∗ ω = ∗dω = 0. The other statement is proved

in a similar way. ¤
4
We will suppose, from now on, that M is a compact, oriented Riemannian manifold. We will define
2 p
the L inner product in Ω (M ). We recall that the volume form of M is the n-form dM := ∗1, i.e. the
form that takes the value 1 on orthonormal positive basis. We define:
Z Z
(4) (ω1 , ω2 ) = hω1 (x), ω2 (x)idM = ω1 ∧ ∗ω2 .
M M
Now (·, ·) is a symmetric, positive definite bi-linear map, i.e. a scalar product in Ωp (M ).
2.11. Remark. We observe explicitly that Ωp (M ) is not complete, in relation to the induced metric
structure, if dim(M ) > 0. In particular, with this scalar product, Ωp (M ) is a pre-Hilbert space and not an
Hilbert space. This is the main difficulty in the proof of the Theorem of Hodge that we will discuss in the
next section.
2.12. Proposition. For ω ∈ Ωp−1 (M ), τ ∈ Ωp (M ) we have:
(dω, τ ) = (ω, δτ ).
In particular (∆ω1 , ω2 ) = (ω1 , ∆ω2 ), i.e ∆ is self adjoint with respect to the L2 inner product.
Proof. We observe that d ∗ τ ∈ Ω(n−p+1) (M ). Then

2
d ∗ τ = (−1)(p−1)(n−p+1) ∗ ∗d ∗ τ = (−1)(p−1)(n−p+1)+n(p−1)+1 ∗ δτ = (−1)(p−1) +1
∗ δτ.
Therefore
d(ω ∧ ∗τ ) = dω ∧ ∗τ + (−1)(p−1) ω ∧ d ∗ τ = dω ∧ ∗τ − ω ∧ ∗δτ.
and, from Stokes Theorem,
Z Z Z
0 = d(ω ∧ ∗τ ) = dω ∧ ∗τ − ω ∧ ∗δτ = (dω, τ ) − (ω, δτ ).
2.13. Corollary. If M is compact, a p-form ω is harmonic if and only if it is closed and coclosed.
Proof. It is clear tat if dω = 0 = δω, the ∆ω = 0. Conversely, if ∆ω = 0, we have
0 = (∆ω, ω) = (dδω, ω) + (δdω, ω) = (dω, dω) + (δω, δω) = kdωk2 + kδωk2 .
Hence o ω is closed and coclosed. ¤

4Orientability is not really needed for most of what we will do. We will comment on this point later on.
2.14. Remark. If M in not compact, still a closed and coclosed form is harmonic. However the converse
is not true. For example the 0-form f : R −→ R, f (x) = x is harmonic but not closed.
2.15. Corollary. [Hopf’s Lemma] If M is compact and f ∈ Ω0 (M ), then

Z
(∆f ) ∗ 1 = 0.
M
In particular, if ∆f ≥ 0, then ∆f = 0 and f is locally constant.
Proof. Z
(∆f ) ∗ 1 = (∆f, 1) = (f, ∆1) = 0.
M
Moreover, if ∆f ≥ 0, ∆f = 0. Hence f is harmonic, hence closed, and therefore locally constant. ¤
3. The Theorem of Hodge
Let M be a compact Riemannian manifold. We will denote by Hp (M ) the space of harmonic p-forms
on M . Since a harmonic form is closed, we have a map:
hp : Hp (M ) −→ H p (M ), hp (ω) = [ω].
The aim of this section is to give a very rough idea of the following result, known Hodge’s Theorem.
3.1. Theorem. [Hodge] The map hp is an isomorphism.
3.2. Remark. Clearly Hp (M ) is a vector space and hp is linear. But, in general, the wedge product
of harmonic forms is not harmonic, so does not induces a graded algebra structure in H∗ (M ) := ⊕Hp (M ).
Hence does not make sense to ask if h∗ := ⊕hp is an algebras morphism.
The geometric idea behind the proof 3.1 is quite simple.
3.3. Proposition. A closed form ω is harmonic if and only if it is a minimul of the fuction g(τ ) =
(τ, τ ) = kτ k2 restricted to the cohomology class [ω] = {ω + dβ : β ∈ Ω(p−1) (M )}.
Proof. Let ω ∈ Ωp (M ) be a closed form. If β ∈ Ω(p−1) (M ), we define:
gβ : R −→ R, gβ (t) = (ω + tdβ, ω + tdβ) = kω + tdβk2 .
Then:
dgβ
(0) = 2(ω, dβ) = 2(δω, β).
dt
dgβ dgβ
If ω is a minimum of g in [ω], (0) = 0, ∀β ∈ Ω(p−1) (M ) and, therefore, (0) = 2(δω, β) = 0 ∀β ∈
dt dt
(p−1)
Ω (M ). In particular δω = 0 and ω is coclosed, hence harmonic, being closed. Conversely, if ω is
coclosed, then it is a critical point of g restricted to [ω]. But [ω] is convex and the only critical points of g
are minimums. ¤
3.4. Corollary. The map hp is injective.
Proof. It follows from the general fact that in an inner product space a convex subset has, at most,
one point of minimal norm 5. ¤
5This fact follows essentially from the parallelogram low and is not true, in general, for normed spaces.
3. THE THEOREM OF HODGE 154
3.5. Corollary. Hp (M ) is finite dimensional.
Proof. It follows from the fact that H p (M ) is finite dimensional (see Exercise ?? of Chapter 3), an the
injectivity of hp . ¤
To show that hp is surjective we have to show that there exist a form of minimal norm in [ω]. This is
not clear since Ωp (M ) is not complete. The natural idea is to complete Ωp (M ) with respect to a suitable
scalar product, and look, in the completition Ωp (M ), for the minimum of the L2 norm in [ω]. This minimum
exists and is unique by general arguments. However it may not be in Ωp (M ) hence may not be a harmonic
p-form. The fact that the minimum is in Ωp (M ) is a highly non trivial result and came from the regularity
theory for elliptic partial differential equations.
We have an apparently different, but essentially equivalent approach to the Theorem of Hodge, in
particular to the regularity problem.
From the fact that Hp is finite dimensional (see Corollary 3.5), it follows that is closed in Ωp (M ) 6. Then
we have an orthogonal decomposition:
(5) Ωp (M ) = Hp (M ) ⊕ Hp (M )⊥ .
Moreover ∆(Ωp (M )) ⊆ Hp (M )⊥ , since ∆ is self adjoint.

It is well known, from basic linear algebra, that if A : Rn → Rn is a self adjoint linear operator, the
equation Ax = y has a solution if and only if y ∈ [ker A]⊥ . The regularity theorem is essentially equivalent
to the fact that the same result holds for ∆. More precisely:
3.6. Theorem. If β ∈ Ωp (M ), the equation ∆ω = β has a solution if and only if β ∈ Hp (M )⊥ .
We have an immediate but important consequence known as the Hodge’s decomposition Theorem:
3.7. Theorem.
Ωp (M ) = Hp (M ) ⊕ ∆(Ωp (M )) = Hp (M ) ⊕ δ(Ω(p+1) (M )) ⊕ d(Ω(p−1) (M )).
Proof. Theorem 3.6 implies ∆(Ωp (M )) = Hp (M )⊥ , hence the conclusion. ¤
3.8. Remark. Theorem 3.7 gives also a proof of the Theorem of Hodge. In fact, if ω is a closed form we
have an orthogonal decomposition ω = ωH +ωδ +ωd . Since ω is closed, (ω, δτ ) = (dω, τ ) = 0 ∀τ ∈ Ω(p+1) (M ).
Hence ω = ωH + dβ, i.e. [ω] = [ωH ].
As a first application of the Theorem of Hodge we will give now a fast proof of the Poincaré dualty
Theorem.
3.9. Theorem. [Poincaré duality] Let M be a compact oriented differentiable manifold. Then there exist
an isomorphism, called Poincaré duality, P p : H p (M ) −→ H (n−p) (M ).
Proof. Fix a Riemannian metric on N 7. If ω ∈ Ωp (M ) is an harmonic form, Corollary 2.10 implies

that ∗ω is harmonic. Hence P p = ∗ : Hp (M ) −→ H(n−p) (M ) is an isomorphism. ¤
6Since any finite dimensional normed space is complete, Hp (M ) is complete, hence closed in Ωp (M ).
7For example look at M as a submanifold of RN .
3.10. Corollary. Let M be a compact, connected, orientable smooth manifold of dimension n. Then
H (M ) ∼
n
= R.
Proof. Put a Riemannian metric on M . Then H n (M ) ∼

= H 0 (M ) ∼
= R (since M is connected). ¤
3.11. Remark. In the arguments related to the Theorem of Hodge orientability allows the use the ∗-
operator and integration of n-forms (or functions). But this is not essential since we can integrate a function
on a Riemannian manifold with respect to the Riemannian density (whatever this mens) and the ∗-operator
appears twice in the formula of the Laplacian, so a change of orientation does not affect the formulas.
However, in Theorem 3.9 the ∗-operator appears only once. It is a fact that this Theorem does not holds for
non orientable manifolds. For example it is possible to prove that a connected non orientable n-dimensional
manifold has vanishing n-dimensional cohomology. In particular, for compact manifolds, orientability is
detected by the topological invariant H n .
4. The Weitzenbök formula
We want to describe the Laplacian of a p-form in terms of geometric object like curvature and covariant
derivative. We start observing that, since we can take covariant differentiation of form, we can define an
action of the curvature tensor on forms.
4.1. Definition. Let X, Y ∈ H(M ), ω ∈ Ωp (M ). We define
R(X, Y )ω = ∇X ∇Y ω − ∇Y ∇X ω − ∇[X,Y ] ω.
We state some facts whose proofs are left to the reader:
4.2. Proposition.
(1) The map R : H(M ) × H(M ) × Ωp (M ) −→ Ωp (M ), R(X, Y, ω) = R(X, Y )ω is is F(M )-trilinear.
In particular it is induced by a R-trilinear map Rq : Tp M ×Tq M ×ΛP (Tq M ) −→ Λp (Tq M ), q ∈ M .
(2) If ω ∈ Ω1 (M ) and Z = ]ω, then R(X, Y )ω = [R(X, Y )Z ( [ : Tq M −→ [Tq M ]∗ = Λ1 (Tq M ) and
] = [−1 are the “musical isomorphisms”).
P
(3) If ω = φ1 ∧ · · · ∧ φp , φi ∈ Ω1 (M ), R(X, Y )ω = φ1 ∧ · · · ∧ R(X, Y )φk ∧ · · · ∧ φp .
(4) The map R(X, Y ) : Ωp (M ) −→ Ωp (M ) is antisymmetric.
We will be interested in a “curvature” that generalizes the Ricci curvature.
4.3. Definition. The generalized Ricci curvature is the map

p
n X
X
Qp : Ωp (M ) −→ Ωp (M ), Qp (ω)(Xi1 , . . . , Xip ) = R(Xs , Xik )ω(Xi1 , . . . , X̂ik , Xs , . . . Xip ),
s=1 k=1
where {X1 , . . . , Xp } is an orthonormal basis of Tp M .
4.4. Remark. Observe that Qp is a “trace”, in particular the definition does not depends on the choice
of the basis.
We will describe ∆ω, ω ∈ Ωp (M ), in terms of Levi-Civita connection and the generalized Ricci curvature.
5. SOME APPLICATIONS 156
4.5. Theorem. Let {X1 , . . . , Xn } be orthonormal vector fields defined in a neighborhood of q ∈ M and
ω ∈ Ωp (M ). Then
n
X
∆ω(q) = − ∇Xs ∇Xs ω(q) + Qp (ω)(q).
s=1
Proof. Since we are interest in the value of ∆ω at q ∈ M , we can assume that the Xi ’s are like in
Proposition 1.42. Using the expressions of d, δ in 2.4 and 2.5 we have, up to terms vanishing at q,
Pp
dδω(X1 , . . . , Xp ) = k=1 (−1)k+1 (∇Xk δω)(X1 , . . . , X̂k , . . . , Xp ) =
Pp Pp Pn
= k=1 (−1)k+1 Xk [δω(X1 , . . . , X̂k , . . . , Xp )] = k=1 (−1)k Xk [ s=1 ∇Xs ω(Xs , X1 , . . . , X̂k , . . . , Xp )] =
Pp Pn
= k=1 (−1)k s=1 ∇Xs ∇Xs (Xs , X1 , . . . , X̂k , . . . , Xp ).
Pn
δdω(X1 , . . . , Xp ) = − s=1 ∇Xs dω(Xs , X1 , . . . , Xp ) =
Pn Pp
= − s=1 Xs [∇Xs ω(X1 , . . . , Xp ) + k=1 (−1)k ∇Xk ω(Xs , X1 , . . . , X̂k , . . . , Xp )] =
Pn Pn Pp
= − s=1 ∇Xs ∇Xs ω(X1 , . . . , Xp ) − s=1 k=1 (−1)k ∇Xs ∇Xk ω(Xs , X1 , . . . , X̂k , . . . , Xp ).
The conclusion follows adding the two equations above.
¤
4.6. Corollary. [Weitzenbök’s formula] If ω ∈ Ωp (M ),

X n n
X
1
h∆ω(p), ω(p)i = ∆(kωk2 ) + k∇Xs ωk2 + hQp (ω), ωi, (∆ω, ω) = k∇Xs ωk2 + (Qp (ω), ω).
2
k=1 k=1
Proof. Consider vector fields {X1 , . . . , Xn } as above. Then

1 1
h∇Xs ∇Xs ω, ωi = Xs h∇Xs ω, ωi − k∇Xs ωk2 = Xs Xs hω, ωi − k∇Xs ωk2 = − ∆(kωk2 ) − k∇Xs ωk2 ,
2 2
where, in the last equality we use Remark 2.2. The first equation follows from Theorem 4.5. The second
equation is obtained from the first one by integration, remembering Hopf’s Lemma 2.15. ¤
5. Some applications
From Corollary 4.6 and the Theorem of Hodge it follows that, if M is a compact Riemannian manifold
with Qp positive definite, H p (M ) = {0}. Since the argument is an integral argument, it follows that the
same is true if Qp ≥ 0 and Qp > 0 for some point q ∈ M . We want similar results under “more usual”
curvature conditions.
We will start with the case p = 1. In this case, Q1 is, essentially, the Ricci curvature (see Exercise 6.6).
5.1. Theorem. Let M be a Riemannian manifold with Ricc ≥ 0 and Ricc(p) > 0, for some p ∈ M ,
then H 1 (M ) = {0}.
5.2. Remark. An observation is due for the reader familiar with the fundamental group. A classical
result in Riemannian geometry is the Bonnet-Myers Theorem that states that a compact manifold M with
positive Ricci curvature has finite fundamental group. From this fact it follows, by standard algebraic topol-
ogy, that H 1 (M ) = {0}. So, on one side the result above is weaker the the Bonnet-Myers Theorem but, on
the other hand, we need only that the Ricci curvature is non negative and positive at some point to conclude
that H 1 vanishes.
If we require that that the Ricci curvature is non negative, then the conclusion does not follows (in fact
it is false). However we still have some relevant informations. Given a harmonic form ω ∈ Ω1 (M ), it follows
from Corollary 4.6, that it is parallel, or, equivalently, the dual vector field X = ]ω is globally parallel, i.e
∇Z X = 0, ∀Z ∈ T M . The existence of parallel vector fields is a very strong restriction on the geometry of
the manifold as we have already seen in Theorem 1.29. In this situation we have
5.3. Corollary. Let M be a compact connected n-dimensional manifold with non negative Ricci cur-
vature. Then dim H 1 (M ) ≤ n.
Proof. A parallel vector field (or 1-form) is uniquely determined by its value at a point. Hence the
space of parallel 1-forms has dimension at most n. ¤
5.4. Remark. It is possible to show, in the hypothesis of the Corollary, that if dim H 1 (M ) = n, M is
isometric to the product of n circles. This is simple to see in the case that M is a Lie group since the existence
of n linearly independent parallel vector fields implies that the curvature tensor vanishes identically, hence
the group is Abelian , and therefore isomorphic the product of n circles.
We consider now a compact connected Lie group G with a bi-invariant metric. We have seen that the
cohomology of such group is isomorphic to the space of Ad-invariant forms. Such formas have the following
“geometric” interpretation
5.5. Proposition. A form ω ∈ Ωp (G) is Ad-invariant if and only if it is parallel, if and only if it is
harmonic.
Proof. The first equivalence follows from Theorem 1.52 and Theorem 5.10 of Chapter 4. For the second
we observe that clearly a parallel form is harmonic and conversely if ω is harmonic, there is an Ad-invariant,
hence parallel, form in [ω] which, by unicity is ω. ¤
b has trivial center if and only if

5.6. Theorem. Let G be a compact connected Lie group G. Then G
H (G) = {0}. In this case H 2 (G) = {0}, H 3 (G) 6= {0}.
1
b has trivial center, this metric has positive

Proof. Consider a bi-invariant Riemannian metric on G. If G
Ricci curvature, by Theorem 1.54, hence H 1 (G) = {0}, by Theorem 5.1. If H 1 (G) = {0}, DG = G, by
b has trivial center. In this case H 2 (G) = {0}, again by Proposition
Proposition 5.15 of Chapter 4, hence G
b×G
5.15 of Chapter 4. For the last assertion we consider the tri-linear map ω : G b×Gb −→ R, ω(X, Y, Z) :=
b by Lemma 1.50. Observe that ω =
h[X, Y ], Zi. This map is an Ad-invariant exterior form in Λ3 (G), 6 0, since
b is non Abelian, hence ω defines a non zero element in H 3 (G).
G ¤
b is not Abelian. In
5.7. Remark. The last assertion in Theorem 5.6 depends only on the fact that G
particular a compact connected two dimensional Lie group must be Abelian, hence isomorphic to S 1 × S 1 .
5.8. Corollary. S n admits a Lie group structure if and only if n = 0, 1, 3.
Proof. For n = 0, 1, 3, S n is the unit sphere of the real, complex and quaternionic lines respectively, so
it admits a Lie group structure. For n 6= 0, 1, 3, H i (S n ) = {0}, i = 1, 3, contraddincting Theorem 5.6. ¤
6. EXERCISES 158
5.9. Definition. A Kähler manifold is a Riemannian manifold M together with a smooth map J :
T M −→ T M such that
(1) π ◦ J = π and, ∀p ∈ M and Jp := J|Tp M : T pM −→ Tp M is a linear isometry,
(2) J 2 = −11,
(3) J is parallel in the sense that, if X, Y ∈ H(M ), J(∇X Y ) = ∇X (J(Y )).
A Kähler manifold has even dimension, say 2n, and ω(X, Y ) := hX, JY i is a parallel 2-form, called the
Kähler form or fundamental form. Moreover ω n 6= 0 (see Exercise 7.33 in Chapter 1). In particular we have
5.10. Proposition. If M is a compact Kähler manifold H 2k (M ) contains the non zero element [ω]k , k ≤
n.
5.11. Example. The main example of a compact Kähler manifold is the complex projective space CP n .
This space is the orbit space of the action
µ : S 1 × S 2n+1 −→ S 2n+1 , µ(θ, p) = eiθ p.
The orbit O(p) is the great circle {eiθ p}. We will denote by ξ(p) the unit tangent vector to the orbit
at p. The tangent space Tp S 2n+1 splits as the orthogonal sum of the real line Vp = {tξ(p) : t ∈ R} and
its orthogonal complement Hp in Tp S 2n+1 . The space Vp (resp. Hp ) is called the vertical (resp. horizontal)
space. Consider the quotient map and its differential,
π : S 2n+1 −→ CP n , π(p) = [p], dπ(p) : Tp S 2n+1 −→ T[p] CP n .
The kernel of dπ(p) is the vertical subspace Vp and dπ(p) maps Hp isomorphically onto T[p] CP n . Given
X ∈ T[p] CP n , the horizontal lift of X, at p, is the vector X̃(p) = [dπ(p)|Hp ]−1 (X). The fact that µθ :
S 2n+1 −→ S 2n+1 is an isometry allows us to define a Riemannian metric on CP n . Given X1 , X2 ∈ T[p] CP n
we define hX1 , X2 i := hX̃1 (p), X̃2 (p)i, where the second scalar product is the one in Tp S 2n+1 . This scalar
product is well defined, since if q = µθ (p), X̃i (q) = dµθ (p)(X̃i ) and dµθ (p) is an isometry.
We will define now a smooth map J : T CP n −→ T CP n that will define a structure of Kähler manifold
on CP n . Given X ∈ T [p]CP n we define JX := dπ(iX̃(p)). We have to check, first of all, that J is well
defined and this follows from the fact that dµθ (p)|Hp : Hp −→ Heiθ p is C-linear. Clearly J is smooth and
J 2 = −11. So we are left to prove that J is parallel. This is consequence of more general facts that we will
discuss in Exercise 6.7.
In particular we can compute the product in the algebra H ∗ (CP n ) = ⊕H k (CP n ). The Kähler form
ω generates H 2 (CP n ) and its powers, ω k generates H 2k (CP n ), k ≤ n. So H ∗ (CP n ) is isomorphic, as an
algebra, to the algebra of polynomials in one real variable t divided by the ideal generated by tn+1 .
6. Exercises
6.1. product metric and connections.
6.2. geodesics of O(n).
6.3. Prove that if X ∈ Tp M , and ω1 , ω2 are forms, ∇X (ω1 ∧ ω2 ) = ∇X ω1 ∧ ω2 + ω1 ∧ ∇X ω2 .
6.4. Prove that for ω, τ ∈ Ωp (M ), X ∈ Tq M, Xhω, τ i = h∇X ω, τ i + hω, ∇X τ i.

6.5. Prove Proposition 4.2
6.6. Prove that, if ω ∈ Ω1 (M ) and X = ]ω, then Q1 (ω) = [Q(X), where [ : Tp M −→ [Tp M ]∗ is the
“musical” isomorphism and Q is the Ricci tensor defined in Section 1.
6.7. Riemannian subimmersion sand CP n
6.8. Prove that H k (S 2 × S 4 ) ∼

= H k (CP 3 ) (as vector spaces). Use the Künnet formula to compute the
algebra structure of H ∗ (S 2 × S 4 ) and Example 5.11 to conclude that S 2 × S 4 is not homotopy equivalent to
CP 3 .
Bibliography
[1] Bott, R.; Tu, L. W. : Diferential Forms in Algebraic Topology, Graduate Texts in Mathematics, Springer-Verlag, New
York-Berlin, 1982.
[2] Bredon, G. E. : Topology and geometry, Graduate texts in Math, Springer Verlag, New York, 1993
[3] do Carmo, M. P. : Differential forms and applications, Universitext, Springer-Verlag, Berlin, 1994.
[4] Dold, A. : A simple proof of the Jordan-Alexander complement theorem. Amer. Math. Montly, 100, n. 9, 856-857.
[5] Lima, E. L. : Curso de Análise, Volume 2, Projeto Euclides, IMPA, Rio de Janeiro, Brazil, 1989.
[6] Lima, E. L. : Introduccion a la Cohomologia de de Rham, IMCA, PUC del Perú, Lima, Perú, 2001.
[7] Lima, E. L. : Àlgebra exterior, Coleção Matemática Universitária, IMPA, Rio de Janeiro, Brazil 2005.
[8] Lima, E. L. : Àlgebra linear, Coleção Matemática Universitária, IMPA, Rio de Janeiro, Brazil 2008.
[9] Milnor, J : Differential Topology, Mimiographed Notes, Princeton, 1966 (?) see also Stong notes on cobordism theory
[10] Singer, M. Thorpe, J. A. : Lecture Notes on Elementary Topology and Geometry, Undergraduate Texts in Mathematics,
Springer-Verlag, New York-Heidelberg, 1976.
[11] Spivak, M. : Calculus on Manifolds, Addison-Wesley Company, 1965.
161

Differential Forms - No Untitled

Uploaded by

Copyright:

Available Formats

Differential Forms - No Untitled

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Differential Forms - No Untitled

Uploaded by

Copyright:

Available Formats

Sandro Buoncristiano, Francesco Mercuri and Alcibiades Rigas

Chapter 0. A review of basic vector calculus 5

Chapter 1. The de Rham cohomology for open sets of Rn 29

Chapter 2. Integration and the singular homology of open sets of Rn 59

Chapter 3. A short course on differentiable manifolds 81

10. Exercises 113

Chapter 4. The theory of de Rham for differentiable manifolds 119

Chapter 5. Harmonic forms and the Theorem of Hodge 139

A review of basic vector calculus

kLk = sup{kLxk : x ∈ Rn , kxk = 1}.

1.2. Remark. For a function f : U ⊆ R −→ R the derivative of f at x ∈ U, f 0 (x), is defined as

1Observe that, if khk is sufficiently small, x + h ∈ U .

if the limit exists. The differential of f at x is the linear map

df (x) : R −→ R, df (x)h = f 0 (x)h.

1.4. Example. Let L : Rn −→ Rm be a linear map. Then L is differentiable and dL(x) = L, ∀ x ∈ Rn ,

1.5. Example. If B : Rn × Rm −→ Rp is a bi-linear map, B is differentiable and dB(x, y)(z, w) =

A = {t ∈ [0, 1] : kφ(s) − φ(0)k ≤ (M + ²)s, ∀ s ∈ [0, t]}.

φ(a + k) − φ(a) = dφ(a)k + r(k) with kr(k)k ≤ ²k

1.8. Definition. Let f : U −→ Rm be differentiable at x ∈ U and X ∈ Rn . The directional derivative

1.12. Remark. Let f : U ⊆ Rn −→ Rm be a function. Let πi : Rm −→ R, πi (x1 , . . . , xm ) = xi be the

If f : U −→ Rm is a differentiable function, the differential can be seen as a map df : U −→

1.13. Definition. We will say that f is twice differentiable at x ∈ U , if the function df : U −→

of f at x and will be denoted by d2 f (x).

1.15. Theorem. [Schwarz’s Theorem] If d2 f (x) exists, it is a symmetric bilinear form.

We will sketch a proof in the case that f is C 2 in Exercise 4.22

1.16. Remark. If f is k times differentiable at x, dk f (x) is a k-multilinear symmetric map. Moreover

1.17. Example. If L : Rn −→ Rm is a linear map dL(x) = L, ∀ x ∈ Rn . In particular the differential

m : M (n, R) × M (n, R) −→ M (n, R), m(A, B) = AB,

is a bi-linear map, hence smooth. Also the map

is bilinear, hence smooth.

ι : GL(n, R) −→ M (n, R), ι(A) = A−1 .

Claim The map ι is smooth.

Proof. Differentiating the identity (A + tB)(A + tB)−1 = 11 we get

1.19. Theorem. [Infinitesimal version of Taylor Theorem] Let f : U −→ Rm be a function s times

1.20. Lemma. Let r : B = B n (R) −→ Rm be a function s times differentiable in B and s + 1 times

kr(x)k ≤ M kxk, M = sup{kdr(tx)k| : t ∈ [0, 1]}.

1.21. Definition. A map φ : U ⊆ Rn −→ V ⊆ Rm between open sets is a C k diffeomorphism if there

The following fact will be useful.

1.23. Lemma. If φ : U ⊆ Rn −→ V ⊆ Rm is a C k map, k ≥ 1, between open sets, and φ admits a

F (x1 , . . . , xn ) = (x1 , . . . , xk , 0, . . . , 0).

1.24. Theorem. [Rank Theorem] Let f : U ⊆ Rn −→ Rm be a function of class C k . Let p ∈ U be such

ψ ◦ f ◦ φ−1 (x1 , . . . , xn ) = (x1 , . . . xk , 0, . . . , 0), f or (x1 , . . . , xn ) ∈ Ũ .

This Theorem will follow from the next three.

1.25. Theorem. [Inverse function Theorem] Let f : U ⊆ Rn −→ Rn be a C k map, k ≥ 1, such that

So we have a well defined surjective map f −1 : B n ( 2r ) −→ U 0 = B n (r) ∩ f −1 (B n ( 2r )). Observe that U 0

Applying A := df (x1 )−1 to the equality above we have

A(y − y1 ) + Ah1 (y, y1 ) = f −1 (y) − f −1 (y1 ),

where y = f (x), y1 = f (x1 ), h1 (y, y1 ) = −h(f −1 (y), f −1 (y1 )). Then

φ ◦ f (x) = (x, 0).

φ ◦ f (x) = φ ◦ F (x, 0) = (x, 0).

F : U −→ Rn × Rp , F (x, y) = (f (x, y), y).

f ◦ ψ(x, y) = π ◦ F ◦ ψ(x, y) = π(x, y) = x.

It is easy to see that if a function is integrable, the integral is unique.

2.3. Proposition. Let f, g : [a, b] −→ Rm be integrable functions, k ∈ R and T : Rm −→ Rp be a linear

kf k0 = sup{f (t) : t ∈ [a, b]}

is a norm in B, called the sup norm or the norm of uniform convergence.

2.4. Definition. A sequence of functions fn : [a, b] −→ Rm in B is uniformly convergent to f ∈ B if

We will describe classes of integrable functions.