Global Analysis

Lecture notes

Andriy Haydys

May 24, 2024

This is a draft. If you spot a mistake, please let me know.

Contents

1 Introduction 2
1.1 The de Rham cohomology groups . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Some linear algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Riemannian metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Harmonic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Riemann surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 The Laplacian on Riemann surfaces . . . . . . . . . . . . . . . . . . . . . . . 12
1.7 Some consequences of Theorem 1.48 . . . . . . . . . . . . . . . . . . . . . . 13

2 Vector bundles, Sobolev spaces, and elliptic partial differential operators 17

2.1 Vector bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Sobolev spaces on Rn . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Sobolev spaces on manifolds . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Differential operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.1 Symbols of differential operators . . . . . . . . . . . . . . . . . . . . . 29
2.3.2 The formal adjoint of a linear differential operator and its symbol . . . 31
2.4 Pseudodifferential operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.1 Elliptic operators and their parametrices . . . . . . . . . . . . . . . . . 38
2.4.2 Elliptic estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5 Elliptic complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.5.1 The Dolbeault complex . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.6 Harmonic forms on complex projective spaces . . . . . . . . . . . . . . . . . . 47
2.7 The elliptic calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3 The Atiyah–Singer index theorem 52

3.1 Fredholm maps and their indices . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2 Characteristic Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.1 The curvature of a connection . . . . . . . . . . . . . . . . . . . . . . 54
3.2.2 Local representation of the curvature form . . . . . . . . . . . . . . . . 55
3.2.3 Chern classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2.4 Some computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.2.5 Other characteristic classes . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3 A prototype of the index theorem . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4 On the de Rham cohomolgy with compact supports . . . . . . . . . . . . . . . 66
3.5 Basics of K-theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.6 The symbol of an elliptic operator revisited . . . . . . . . . . . . . . . . . . . 67
3.7 An application: The Riemann-Roch formula . . . . . . . . . . . . . . . . . . . 68

1
Chapter 1

Introduction

1.1 The de Rham cohomology groups

Let M be a manifold of dimension n. Denote by Ωk (M ) the space of differential k-forms on
M . Recall that there exist a unique R-linear map d : Ωk (M ) → Ωk+1 (M ) with the following
properties:

(i) df is the differential of f if f ∈ C ∞ (M ) = Ω0 (M );

(ii) d (ω ∧ η) = dω ∧ η + (−1)q ω ∧ dη if ω ∈ Ωq (M );
(iii) d2 = 0.

The last property simply means that d (dω) = 0 for each ω ∈ Ωk (M ). This yields the de
Rham complex
d d d d d d d
0 → Ω0 (M ) −→ Ω1 (M ) −→ . . . −→ Ωk−1 (M ) −→ Ωk (M ) −→ Ωk+1 (M ) −→ . . . −→ Ωn (M ) → 0
(1.1)
Remark 1.2. The map d : Ωk (M ) → Ωk+1 (M ) depends on k, however this is suppressed in
the notations.
Property (iii) means that (1.1) is a complex, that is the kernel of d : Ωk → Ωk+1 contains
the image of d : Ωk−1 → Ωk and therefore we can define
k k+1


k Ker d : Ω (M ) → Ω (M )
HdR (M ) := (1.3)
Im (d : Ωk−1 (M ) → Ωk (M ))

This is called the k th de Rham cohomology group, which is in fact a vector space.
Assume that M is closed, that is M is compact and has no boundary. Then the number
k
bk (M ) := dim HdR (M )

is finite. It is called the k th Betti number of M and is an invariant of M .

Remark 1.4. While it is by no means obvious from the above description, Betti numbers are
topological invariants of M , that is bk (M ) = bk (N ) provided M and N are homeomorphic. In
paticular, Betti numbers do not depend on the smooth structure.
k
Coming back to the de Rham cohomlogy groups, each element in HdR (M ) is represented
by the equivalence class
[ω] = ω + dη | η ∈ Ωk−1 (M ) ,


2
 Global analysis

where ω ∈ Ωk (M ) is closed: dω = 0. Hence, we may ask the following question: What is the
best representative in [ω]?
Of course, at this point the above question is vague, since the notion of being "the best" is
undefined. One possibility to make this more precise is as follows. Just by its definition, the set
[ω] is an affine subspace of Ωk (M ). We could call an element in [ω] "the best" if it is the closest
one to the origin just as shown schematically on Figure 1.1.

[ω]

Ωk (M )

Figure 1.1: A choice of a representative in the de Rham cohomology class.

However, this raises our next question: How do we measure distance in Ωk (M )? A suitable
answer to this question requires a detour, which we do next.

1.2 Some linear algebra

Let V be a vector space of dimension n over R. Recall that for each basis e = (e1 , . . . , en ) of
V there exist a unique basis e∗ = (e∗1 , . . . , e∗n ) of the dual vector space V ∗ such that
(
∗ 1, if i = j,
ei (ej ) = δij =
0, if i ̸= j.

e∗ is called the dual basis to e.

Assume ⟨·, ·⟩ is a scalar product on V . If e = (e1 , . . . , en ) is an orthonormal basis of V , that

is
⟨ei , ej ⟩ = δij ,
then the dual basis e∗ is given explicitly by

e∗i := ⟨ei , ·⟩ ⇐⇒ e∗i (v) = ⟨ei , v⟩ for v ∈ V.

Then V ∗ has a unique scalar product such that e∗ = (e∗1 , . . . , e∗n ) is an orthonormal basis.
Explicitly, for ξ, η ∈ V ∗ define

ξ = P ξi e∗i
P
ξi := ξ (ei )
⇐⇒
ηi := η (ei ) η = ηj e∗j

Draft 3 May 24, 2024

 Global analysis

n
X n
X
=⇒ ⟨ξ, η⟩ = ξi ηi = ξ (ei ) η (ei ) .
i=1 i=1

To sum up, for any scalar product on V there exists a unique scalar product on V ∗ such that
the dual basis of an orthonormal basis is itself orthonormal.
More generally, any basis e of V yields a basis of Λk V ∗ . Explicitly,

Λk e∗ := e∗i1 ∧ . . . ∧ e∗ik | 1 ≤ i1 < i2 < . . . < ik ≤ n


(1.5)

is a basis of Λk V ∗ consisting of  
n n!
=
k (n − k)!k!
elements. Just like in the case of V ∗ = Λ1 V ∗ , we can define a scalar product on each Λk V ∗ by
declaring (1.5) to be an orthonormal basis.
Recall that any two bases e and f of V are related by a change-of-basis matrix A. This
means n
X
f =e·A ⇐⇒ fi = aij ej
j=1

Then e and f are said to be cooriented, if det A > 0. It is easy to check that

e∼f ≡ e and f are cooriented

yields an equivalence relation on the set of all bases of V . Moreover, there are exactly two
equivalence classes represented by e and ē = (−e1 , e2 , . . . , en ).
Definition 1.6. An orientation on V is a choice of an equivalence class of bases of V . Any basis
in the chosen class is said to be positively oriented and any basis, which does not belong to the
selected class is said to be negatively oriented.
Example 1.7. For Rn the class of the standard basis is called the standard orientation of Rn .
Example 1.8. Any ω ∈ Λn V ∗ , ω ̸= 0, determines an orientation of V by the rule: e is positively
oriented if and only if
ω (e1 , . . . , en ) > 0.
For example, if e∗ = (e∗1 , . . . , e∗n ) is the dual basis to the standard one, then

ω := e∗1 ∧ . . . ∧ e∗n (1.9)

determines the standard orientation of Rn .

Definition 1.10. Let V be an oriented Euclidean vector space of dimension n. An n-form ω is
said to be the Euclidean volume form, if

ω (e1 , . . . , en ) = 1 (1.11)

holds for any positively oriented orthonormal basis e of V .

For example, in the case V = Rn , which is equipped with the standard scalar product and
orientation, (1.9) is the Euclidean volume form.
Example 1.12. Show that any oriented Euclidean vector space admits a unique Euclidean
volume form. This is sometimes denoted by vol.

Draft 4 May 24, 2024

 Global analysis

Proposition 1.13. There is a unique linear map

∗ : Λk V ∗ → Λn−k V ∗ satisfying ξ ∧ ∗ η = ⟨ξ, η⟩ vol (1.14)

for all ξ , η ∈ Λk V ∗ .

Proof (Sketch). Let e be a positively oriented orthonormal basis of V . Set

η := e∗i1 ∧ . . . ∧ e∗ik ∈ Λk e∗

Then assuming that ∗ exists, for

X
ξ= ξj1 ...jk e∗j1 ∧ . . . ∧ e∗jk

we must have
ξ ∧ ∗ η = ⟨ξ, η⟩ vol = ξi1 ...ik · e∗1 ∧ . . . ∧ e∗n .
This yields
∗ η = ∗ e∗i1 ∧ . . . ∧ e∗ik = ε · e∗l1 ∧ . . . ∧ e∗ln−k ,


(1.15)
where ε ∈ {±1} and 1 ≤ l1 < . . . < ln−k ≤ n consists of those integers in the interval [1, n]
which are complementary to {i1 , . . . , ik }.
For example, if n = 6 and η = e2 ∧ e4 , then

∗ (e∗2 ∧ e∗4 ) = ε e∗1 ∧ e∗3 ∧ e∗5 ∧ e∗6 .

To determine ε, we compute

e∗2 ∧ e∗4 ∧ ε (e∗1 ∧ e∗3 ∧ e∗5 ∧ e∗6 ) = −ε vol ,

e∗2 ∧ e∗4 ∧ ∗ (e∗2 ∧ e∗4 ) = ∥ e∗2 ∧ e∗4 ∥2 · vol = vol,

which yields ε = −1 so that we finally obtain

∗ (e∗2 ∧ e∗4 ) = −e∗1 ∧ e∗3 ∧ e∗5 ∧ e∗6 .

In general, ε is the signature of the permutation (i1 , . . . , ik , j1 , . . . , jn−k ).

Thus, (1.15) defines ∗ on the elements of the basis Λk e∗ . This yields a unique linear map
∗ : Λk V ∗ → Λn−k V ∗ , which satisfies (1.14). □
The map ∗ defined in the above proposition is called the Hodge operator.
Remark 1.16. It follows from the proof of the above proposition that the Hodge operator satisfies

∗ ∗ ω = (−1)k(n−k) ω for all ω ∈ Λk V ∗ . (1.17)

1.3 Riemannian metrics

Definition 1.18. A Riemannian metric g on M is a smooth section of T ∗ M ⊗ T ∗ M such that
the following holds:

(i) g is symmetric, that is

g (v, w) = g (w, v) for all v, w ∈ Tm M and any m ∈ M ;

Draft 5 May 24, 2024

 Global analysis

(ii) g is positive-definite, that is

g (v, v) > 0 for all v ∈ Tm M, v ̸= 0, and any m ∈ M.

In other words, g is a family {gm | m ∈ M } of scalar products on each Tm M and gm

depends smoothly on m. In particular, each Tm M is an Euclidean vector space. Hence, each
Λk Tm∗ M is also an Euclidean vector space.
An orientation of a manifold M is (informally speaking) a choice of coherent orientations
of Tm M for each m ∈ M . More formally, we have the following.
Definition 1.19. A manifold M of dimension n is said to be orientable, if there exists ω ∈
Ωn (M ) such that ωm ̸= 0 for all m ∈ M .
By Example 1.8, for each m ∈ M the n-form on Tm M determines a class of positively
oriented bases of Tm M , that is an orientation. Notice that for any function f , which is positive
everywhere, the forms ω and f · ω determine the same orientation on each Tm M .
Albeit not all manifolds are orientable, orientability is a mild restriction. In particular, for
any connected non-orientable manifold M there exists a unique connected double covering
M2 → M , which is orientable. The reader may find more information on this for example
in [Hir94, Sect. 4.4].
Definition 1.20. An orientation of an n-manifold M is a class of n-forms JωK, where
• ω is a nowhere vanishing n-form on M .
• ω1 ∼ ω2 if and only if there exists an everywhere positive function f such that ω2 = f ·ω1 .
Notice that J·K above is not the de Rham cohomology class.
Just like in the preceding section, a differential n-form ω on M is said to be a Riemannian
volume form, if
ωm (e1 , . . . , en ) = 1 (1.21)
holds for any m ∈ M and any oriented orthonormal basis (e1 , . . . , en ) of Tm M . Property (1.21)
determines a Riemannian volume form uniquely. This volume form is denoted by vol.
Thus, by the preceding subsection, a Riemannian metric and orientation on M induce for
each k ≤ n the Hodge operator ∗ : Ωk (M ) → Ωn−k (M ) such that

ω ∧ ∗ η = ⟨ω, η⟩ vol and ∗ ∗ ω = (−1)k(n−k) ω

holds for all ω, η ∈ Ωk (M ).

1.4 Harmonic forms

Let us come back to the original question about "the best" representatives of the de Rham
cohomology classes. Thus, we define the L2 -scalar product on each Ωk (M ) by setting
Z Z
⟨ω, η⟩L2 := ⟨ωm , ηm ⟩ volm = ω ∧ ∗ η.
M M

b ∈ [ω] "the best", if ω

With this at hand, we could call an element ω b minimizes the distance
to the origin, that is if
inf ∥ ω + dη ∥2L2 = ∥ ω
b ∥2L2 . (1.22)
η∈Ωk−1 (M )

Draft 6 May 24, 2024

 Global analysis

Then, if (1.22) holds, for any η ∈ Ωk−1 (M ) we must have

d
0= b + t dη ∥2L2
∥ω
dt t=0
d (1.23)
b ∥2L2 + 2t⟨b
ω , dη⟩L2 + t2 ∥ dη ∥2L2


= ∥ω
dt t=0
= 2⟨bω , dη⟩L2

Proposition 1.24. Denote d∗ := (−1)nk+1 ∗ d ∗ : Ωk+1 (M ) → Ωk (M ). If M is closed, then

the following holds:

(i) d∗ is the formal adjoint of d, that is

⟨dω, η⟩L2 = ⟨ω, d∗ η⟩L2 for all ω ∈ Ωk (M ) and η ∈ Ωk+1 (M ) . (1.25)

(ii) (1.23) is equivalent to

d∗ ω
b = 0. (1.26)

Proof. Notice first that (1.14) and (1.17) imply the equality

ω ∧ ζ = (−1)k(n−k) ⟨ω, ∗ ζ⟩vol for all ω ∈ Ωk (M ) and ζ ∈ Ωn−k (M ) .

Using this, we obtain

Z
∗ kn+1
⟨ω, d η⟩L2 = (−1) ⟨ω, ∗ d ∗ η⟩ vol
MZ
kn+1+k(n−k)
= (−1) ω∧d∗η
Z M
1−k2
= (−1) (−1)k (d (ω ∧ ∗ η) − dω ∧ ∗ η)
M

Here the last equality follows from the Leibnitz rule:

d (ω ∧ ζ) = dω ∧ ζ + (−1)k ω ∧ dζ

provided ω ∈ Ωk (M ) and ζ ∈ Ωl (M ). Hence, applying Stokes’ theorem and recalling that

∂M = ∅, we obtain
Z
k−k2 +2 2
∗
⟨ω, d η⟩L2 = (−1) dω ∧ ∗ η = (−1)k−k ⟨dω, η⟩L2 .
M

By noticing that k 2 is even/odd if and only if k is even/odd, we arrive finally at (1.25).

To prove (ii), notice that

d∗ ω
b=0 =⇒ 0 = ⟨d∗ ω
b , η⟩L2 = ⟨b
ω , dη⟩L2 .

Conversely, setting η = d∗ ω
b in (1.23), we obtain

ω , dd∗ ω
0 = ⟨b b ⟩L2 = ⟨d∗ ω b ⟩L2 = ∥ d∗ ω ∥2L2
b , d∗ ω =⇒ d∗ ω
b = 0.

Draft 7 May 24, 2024

 Global analysis

Notice that (1.26) is nothing else but the Euler-Lagrange equation for the functional

f : ω + dη | η ∈ Ωk−1 (M ) −→ R, f (ω + dη) = ∥ ω + dη ∥2L2


(1.27)

Definition 1.28. The map

∆ = dd∗ + d∗ d : Ωk (M ) → Ωk (M )

is called the Laplace operator (or, simply, the Laplacian). A k-form ω such that ∆ω = 0 is
called harmonic.

Proposition 1.29. On a closed manifold a k-form ω is harmonic if and only if

dω = 0 and d∗ ω = 0. (1.30)

Proof. If (1.30) holds, then ω is clearly harmonic. To show the converse, consider

0 = ⟨∆ω, ω⟩L2 = ⟨dd∗ ω, ω⟩L2 + ⟨d∗ dω, ω⟩ = ∥ d∗ ω ∥2L2 + ∥ dω ∥2L2 .

Here the last equality follows by Proposition 1.24, (i). Since both summands are non-negative,
we obtain (1.30). □

Theorem 1.31. If a minimizer ω

b of (1.27) exists, then ω
b is harmonic. Moreover, if ω
b exists, then
it is unique.

Proof. Assume ω b = ω + dη for some η ∈ Ωk−1 (M ), we have

b exists. Since ω

db
ω = dω + d (dη) = 0 + 0.

Combining this with (1.26), we obtain by Proposition 1.29, that ω

b is harmonic.
Furthermore, let ω
b = ω + dη and ω b = ω + dζ be two harmonic representatives of [ω]. Then
b
 
∗ ∗b ∗
0=d ω b−d ω b=d ω b−ω b = d∗ d (η − ζ) .
b

Denoting temporarily ξ := η − ζ, we obtain

0 = ⟨d∗ dξ, ξ⟩L2 = ⟨dξ, dξ⟩ = ∥ dξ ∥2L2 =⇒ dξ = 0 =⇒ dη = dζ =⇒ ω

b=ω
b.
b

This proves the uniquesness. □

Our aim is to prove the following.

Theorem 1.32. (Hodge) Each de Rham cohomology class is represented by a unique harmonic
form.

Notice that since we have already proved the uniqueness, it is the existence, which remains
to be proved. It turns out that this is somewhat harder and requires certain technology, which
we will consider first.
Notice that for any oriented Riemannian manifold, the Laplacian on Ω0 (M ) = C ∞ (M ) is
given by
∆f = d∗ df = − ∗ d ∗ df. (1.33)

Draft 8 May 24, 2024

 Global analysis

Example 1.34. Consider the case M = R3 equipped with the standard Euclidean matric. If
(x, y, z) are coordinates on R3 , then
 
∂f ∂f ∂f ∂f ∂f ∂f
∗ df = ∗ dx + dy + dz = dy ∧ dz − dx ∧ dz + dx ∧ dy.
∂x ∂y ∂z ∂x ∂y ∂z

Hence,

∂ 2f ∂ 2f ∂ 2f ∂ 2f ∂ 2f ∂ 2f
   
d ∗ df = + + dx ∧ dy ∧ dz =⇒ ∆f = − + + .
∂x2 ∂y 2 ∂z 2 ∂x2 ∂y 2 ∂z 2

Sometimes this is called "the non-negative Laplacian", since its eigenvalues are non-negative.
More generally, if the Riemannian metric on Rn is given by a positive-definite matrix (gij ),
where gij = gij (x), then the corresponding Laplacian on functions is given explicitly by

1 p
ij

∆f = − p ∂i |g| g ∂j f , (1.35)
|g|

where |g| = | det (gij ) | and (g ij ) = (gij )−1 . This is sometimes called the Laplace-Beltrami
operator.

Let M be a closed oriented Riemannian manifold.

Proposition 1.36 (Green’s identity). For any ω, η ∈ Ωk (M ) we have

⟨∆ω, η⟩L2 = ⟨ω, ∆η⟩L2 = ⟨dω, dη⟩L2 + ⟨d∗ ω, d∗ η⟩L2 .

Proof. By Proposition 1.24 (i), we have

⟨dd∗ ω + d∗ dω, η⟩ = ⟨d∗ ω, d∗ η⟩ + ⟨dω, dη⟩.

By the same token,

⟨ω, dd∗ η + d∗ dη⟩ = ⟨d∗ ω, d∗ η⟩ + ⟨dω, dη⟩,
which yields the claim of this proposition. □

Corollary 1.37. On a closed connected manifold any harmonic function is constant.

Proof. If f ∈ C ∞ (M ) is harmonic, by Green’s identity (with ω = η = f ) we obtain

0 = ⟨∆f, f ⟩ =∥ df ∥2L2 =⇒ df = 0 =⇒ f is constant.

The Hodge theorem follows from the following more general result, which is also attributed
to Hodge.

Theorem 1.38. (Hodge) Let M be a closed oriented Riemannian manifold and η ∈ Ωk (M ).

The equation ∆ω = η has a solution if and only if

⟨η, ω0 ⟩L2 = 0 (1.39)

for any harmonic k-form ω0 .

Draft 9 May 24, 2024

 Global analysis

It is easy to see that (1.39) is a necessary condition. Indeed, if there is ω ∈ Ωk (M ) such

that ∆ω = η, then Green’s identity yields

⟨η, ω0 ⟩L2 = ⟨∆ω, ω0 ⟩L2 = ⟨ω, ∆ω0 ⟩L2 = 0

for any harmonic k-form ω0 .

The proof that (1.39) is a sufficient condition will be given later. Taking this as granted for
now, from Corollary 1.37 we obtain the following.
Corollary 1.40. Let M be a closed oriented Riemannian manifold. The equation

∆f = h, f, h ∈ C ∞ (M )

has a solution if and only if Z

h · vol = 0.
M

Let me also show that Theorem 1.38 implies that any de Rham cohomolgy class is represented
by a harmonic form. To this end, pick any closed k-form ω. Then by Proposition 1.29, ω + dη
is harmonic if and only if

d (ω + dη) = 0
⇐⇒ d∗ (ω + dη) = 0 ⇐⇒ d∗ dη = −d∗ ω. (1.41)
d∗ (ω + dη) = 0

Furthermore, consider instead the equation

∆η = −d∗ ω. (1.42)

If η0 is any harmonic (k − 1)-form, then

⟨d∗ ω, η0 ⟩L2 = ⟨ω, dη0 ⟩L2 = 0,

since η0 is closed. Hence, Theorem 1.38 guarantees that (1.42) has a solution η.
Furthermore, notice that (1.42) is equivalent to d∗ (dη + ω) + dd∗ η = 0. However, the two
summands appearing on the left hand side are L2 -orthogonal:

d∗ (dη + ω), dd∗ η L2

= dη + ω, d2 d∗ η L2
= 0.

Hence, if η is a solution of (1.42), then in fact η solves (1.41) so that ω + dη is harmonic indeed.
Remark 1.43. The argument above shows in fact that Im d and Im d∗ are L2 -orthogonal subspaces
provided M is closed.
With these preliminary considerations at hand we proceed by showing that Theorem 1.38
has useful applications, in particular in the theory of Riemann surfaces.

1.5 Riemann surfaces

Definition 1.44. A Riemann surface Σ is a complex manifold of dimension one. In particular, Σ
admits an atlas {(Uα , ψα ) | α ∈ A}, where ψα : Uα −→ C, such that each coordinate transformation
map
ψα ◦ ψβ−1 : C −→ C
is holomorphic on the domain of its definition.

Draft 10 May 24, 2024

 Global analysis

Writing zα = z = x + iy = ψα , we obtain a holomorphic coordinate on Σ (defined on Uα ).

Notice that
x = Re z and y = Im z
are real coordinates on Uα . In particular, for any p ∈ Uα we have a basis
∂ α := (∂x , ∂y ) p

of Tp Σ. Define a linear map

Ip : Tp Σ −→ Tp Σ by Ip ∂x = ∂y and Ip ∂y = −∂x . (1.45)
Using the fact that ψα ◦ ψβ−1 is holomorphic, it is easy to check that Ip does not depend on
the choice of a holomorphic coordinate. It follows also from (1.45) that Ip2 = −id.
For any smooth (not necessarily holomorphic) function f : Σ −→ C define ∂f, ∂f ¯ ∈
Ω (Σ; C) by
1

∂f (v) := 12 (df (v) − i df (Iv))

for v ∈ T Σ.
¯ (v) := 1 (df (v) + i df (Iv))
∂f 2

In particular, we have df = ∂f + ∂f ¯ and f is holomorphic if and only if ∂f

¯ = 0. If
fα := f ◦ ψα−1 : C −→ C is the local representation of f , then

¯ α = ∂fα dz̄
∂f and ∂fα =
∂fα
dz
∂ z̄ ∂z
are local representations of ∂f ¯ α = ∂f
¯ and ∂f respectively, that is ψ ∗ ∂f ¯ and ψ ∗ ∂fα = ∂f .
α α
Furthermore, since
dz = dx + i dy, dx = 21 (dz + dz̄) ,
⇐⇒
dz̄ = dx − i dy, dy = 2i1 (dz − dz̄) ,
any ω ∈ Ω1 (Σ; C) can be written uniquely as ω = a dz + b dz̄ for some functions a and b.
Denote by Ω1,0 (Σ) the space of all those complex valued 1-forms, whose local representation
is a dz for some a ∈ C ∞ (Σ; C). More invariantly, we have
Ω1,0 (Σ) = ω ∈ Ω1 (Σ; C) | ω (I·) = i ω (·) .


Similarly, denote by Ω0,1 (Σ) the space of all those 1-forms, whose local representation is
b dz̄ for some complex valued function b, or equivalently,
Ω0,1 (Σ) := ω ∈ Ω1 (Σ) | ω (I·) = −i ω (·) .


Thus, we obtain the decomposition

Ω1 (Σ; C) = Ω1,0 (Σ) ⊕ Ω0,1 (Σ) .
Then ∂f and ∂f ¯ are nothing else but the components of df lying in Ω1,0 (Σ) and Ω0,1 (Σ)
respectively.
Thus, in the case of a Riemann surface, the complexified de Rham complex has the following
form:
Ω1,0 (Σ)
∂ ∂¯

0 Ω0 (Σ; C) ⊕ Ω2 (Σ, C) 0 (1.46)

∂¯ ∂

Ω0,1 (Σ)

Draft 11 May 24, 2024

 Global analysis

Here ∂¯ : Ω1,0 (Σ) → Ω2 (Σ; C) is just the restriction of d to Ω1,0 (Σ) and ∂ : Ω0,1 (Σ) →
Ω (Σ, C) is the restriction of d to Ω0,1 (Σ). Locally, we have
2

 
1,0 ∂a ∂a ∂a
ω ∈ Ω (Σ) =⇒ ω = a dz =⇒ dω = da ∧ dz = dz + dz̄ ∧ dz = − dz ∧ dz̄,
∂z ∂ z̄ ∂ z̄
 
∂b ∂b ∂b
ω ∈ Ω0,1 (Σ) =⇒ ω = b dz̄ =⇒ dω = db ∧ dz̄ = dz + dz̄ ∧ dz̄ = dz ∧ dz̄.
∂z ∂ z̄ ∂z
Notice that we have
d2 = 0 ⇐⇒ ∂ ∂¯ + ∂∂
¯ = 0.
The splitting (1.46) of the de Rham complex yields the Dolbeault cohomology groups:

¯
H 1,0 (Σ) := ker ∂,

¯ ¯

H 0 (Σ) := ker ∂, H 1,1 (Σ) := Ω2 Im ∂,

¯

H 0,1 (Σ) := Ω0,1 Im ∂.

Clearly, H 0 (Σ) is the space of holomorphic functions on Σ. In particular, if Σ is compact,

connected, and boundaryless, then H 0 (Σ) ∼= C.
H 1,0 (Σ) is the space of holomorphic differentials, that is 1-forms ω such that locally ω =
a dz and a is a holomorphic function. The geometric meaning of H 0,1 (Σ) and H 1,1 (Σ) is
somewhat less straightforward. An interested reader may wish to consult for example [Don11]
for further details.

1.6 The Laplacian on Riemann surfaces

A peculiar feature of Riemann surfaces is that the Laplace-Beltrami operator can be defined
without a reference to a Riemannian metric. Indeed, set
¯ = −2i ∂ ∂¯ : Ω0 (Σ; C) −→ Ω2 (Σ; C) .
∆ := 2i ∂∂

If z = x + iy is a local holomorphic coordinate as above, then we have

¯ = ∂f dz̄ = 1 (∂x f + ∂y f i) dz̄

∂f =⇒
∂ z̄ 2
¯ = 1 ∂ (∂x f + ∂y f i) dz ∧ dz̄
∂ ∂f
2 ∂z
1
= (∂x − i ∂y ) (∂x f + ∂y f i) dz ∧ dz̄
4
1 2 2


= ∂xx f + ∂yy f dz ∧ dz̄.
4
Furthermore, since

dz ∧ dz̄ = (dx + i dy) ∧ (dx − i dy) = −2i dx ∧ dy,

we obtain
¯ = − ∂ 2 f + ∂ 2 f dx ∧ dy.


∆f = −2i ∂ ∂f xx yy

Draft 12 May 24, 2024

 Global analysis

Remark 1.47. To relate this operator to the Laplacian in the sense of (1.33), let g be a Hermitian
metric on Σ, that is g (I·, I·) = g (·, ·).
If z = x + yi is a local holomorphic coordinate, then
   −1 
λ 0 ij

λ 0
(gij ) = =⇒ g =
0 λ 0 λ−1
for some positive function λ = λ (x, y).
Substituting this into (1.35), we obtain
1 1 2
∂x λ · λ−1 ∂x f + ∂y λ · λ−1 ∂y f 2





∆f = − = − ∂xx f + ∂yy f .
λ λ
Furthermore, the Riemannian volume form on Σ in terms of local coordinates (x, y) is
vol = λ dx ∧ dy.
Hence, ∗ vol = 1 =⇒ ∗ (dx ∧ dy) = λ−1 . This yields
1 2 2


¯ .


∆f = − ∂xx f + ∂yy f = ∗ −2i ∂ ∂f
λ
Hence, up to the application of the isomorphism ∗ : Ω2 (Σ) → Ω0 (Σ), the Laplacian coincides
with −2i ∂ ∂¯ indeed.
In the current setting, Corollary 1.40 yields the following.
Theorem 1.48. Let Σ be a compact connected Riemann ¯ = η,
surface. The equation −2i ∂ ∂f
where η ∈ Ω (Σ; C), has a solution if and only if Σ η = 0.
2
R
□

1.7 Some consequences of Theorem 1.48

In this section we assume that Σ is a compact connected Riemann surface throughout.
We have a natural skew-symmetric pairing
Z
Ω (Σ) × Ω (Σ) −→ C,
1,0 0,1
(ω, η) 7−→ ω ∧ η.
Σ

This yields a bilinear map

Z
B: H 1,0
(Σ) × H 0,1
(Σ) −→ C, B (ω, [η]) = ω ∧ η. (1.49)
Σ

Lemma 1.50. B is well defined.

Proof. Notice first that for any α, β ∈ Ω1,0 (Σ) we have α ∧ β = 0. Indeed, locally α = a · dz
and β = b · dz so that α ∧ β = ab dz ∧ dz = 0. Using this observation, we obtain
Z Z Z
¯ = ¯



ω ∧ η + ∂f ω ∧ η + ω ∧ ∂f + ∂f
Σ ZΣ ZΣ
= ω ∧ η + ω ∧ df
ZΣ ZΣ Z
= ω ∧ η − d (f · ω) + f dω
ZΣ Σ Σ

= ω ∧ η − 0 + 0.
Σ

Draft 13 May 24, 2024

 Global analysis

In the last equality the second summand vanishes by the Stokes’ thm; The last summand
vanishes, since ω is closed: Locally ω = a (z) dz so that
 
∂a ∂a
dω = da ∧ dz = dz + dz̄ ∧ dz = 0 + 0 = 0,
∂z ∂ z̄
since a is a holomorphic. This finishes the proof of this lemma. □
Notice that we have a natural map

i : H 1,0 (Σ) −→ H 1 (Σ; C) , ω 7−→ [ω] .

The class [ω] is well-defined, since ω is closed as it has been shown in the proof Lemma 1.50.
Furthermore, the conjugation Ω1 (Σ; C) −→ Ω1 (Σ, C) , η 7−→ η̄, induces a map

σ : H 1,0 (Σ) −→ H 0,1 (Σ) ,

which is antilinear, that is σ (iη) = −iσ (η).

Theorem 1.51.
(i) σ is an isomorphism;
(ii) For a complex vector space V denote V ∗ = {ψ : V → C | ψ is antilinear}. Then B : H 1,0 (Σ) →
H 0,1 (Σ)∗ is an isomorphism;
(iii) The map

H 1,0 (Σ) ⊕ H 0,1 (Σ) −→ H 1 (Σ; C) , (ω, [η]) 7−→ i (ω) + i (σ −1 ([η])) (1.52)

is an isomorphism.

Proof. Notice that σ : H 1,0 (Σ) −→ H 0,1 (Σ) is well-defined. We want to show that for each
class [η] ∈ H 0,1 (Σ) there exists a unique ω ∈ H 1,0 (Σ) such that σ (ω) = [η]. Indeed, any
representative of [η] ∈ H 0,1 (Σ) can be written as
¯
η ′ = η + ∂f

for some f ∈ C ∞ (Σ; C). We want to find a representative η ′ such that

∂η ′ = 0 ⇐⇒ ¯ = −∂η.
∂ ∂f (1.53)

The equation on the right hand side of (1.53) has a solution, since
Z Z
∂η = dη = 0.
Σ Σ

In fact, f is defined uniquely by (1.53) up to the addition of a constant. Hence, for each class
[η] ∈ H 0,1 (Σ) there exists a unique representative η ′ such that ∂η ′ = 0.
Define ω := η̄ ′ . Then
¯ = ∂¯η̄ ′ = ∂η ′ = 0,
∂ω
that is ω lies in H 1,0 (Σ) and is a unique preimage of [η].
To prove (ii), we only need to show that (1.49) is non-degenerate, that is

∀ ω ∈ H 1,0 (Σ) ∃ [η] ∈ H 0,1 (Σ) such that B (ω, [η]) ̸= 0.

Draft 14 May 24, 2024

 Global analysis

Indeed, for ω ∈ H 1,0 (Σ) set η := σ (ω) = [ω̄]. Then

Z
B (ω, [ω̄]) = ω ∧ ω̄ ̸= 0,
Σ

since a dz ∧ ā dz̄ = |a|2 dz ∧ dz̄.

To prove the injectivity of (1.52), notice first that i is injective, since
Z
B (i (ω) , σ (i (ω))) = ω ∧ ω̄ ̸= 0
Σ

provided ω ̸= 0.
Furthermore, denote ζ := i (σ −1 ([η])) and assume that ω + ζ̄ = df = ∂f + ∂f¯ for some
1,0 ¯ ∈
function f , that is (ω, [η]) is in the kernel of (1.52). Since ω − ∂f ∈ Ω (Σ) and ζ̄ − ∂f
Ω0,1 (Σ), we obtain
ω = ∂f and ζ = ∂ f¯.
¯ = ∂ω
Hence, ∂∂f ¯ = 0 and therefore f is constant, which yields ω = 0 and

ζ = i(σ −1 ([η])) = 0 ⇐⇒ [η] = 0,

since i is injective and σ is an isomorphism. Hence, (ω, [η]) = 0.

To see that (1.52) is surjective, pick any [ζ] ∈ HdR 1
(Σ; C). I claim that there exists a
′
representative ζ = ζ + df such that

¯ ′=0
∂ζ ⇐⇒ ¯ + ∂df
0 = ∂ζ ¯ = ∂ζ
¯ − ∂ ∂f.
¯ (1.54)

The existence of a function f satisfying (1.54) follows from Theorem 1.48 just like in the proof
of (i).
Furthermore, write
1 ′ 1 ′
ζ ′ = ω + η, where ω = (ζ − iζ ′ (I·)) and η= (ζ + iζ ′ (I·)) .
2 2
Locally, we have

ζ ′ = a dz + b dz̄ =⇒ ω = a dz and η = b dz̄.

¯ ′ = 0, we obtain
Since ∂ζ ∂a
= 0, that is ω is a holomorphic differential.
∂ z̄
Furthermore,

¯ ′=0 ∂b
∂ζ and dζ ′ = 0 =⇒ ∂ζ ′ = 0 =⇒ =0 =⇒ ∂η = 0
∂z
=⇒ ∂¯η̄ = 0.

Hence, η̄ is also a holomorphic differential. Combining this with (i), we obtain (iii). □

Corollary 1.55. If b1 (Σ) = 0, then Σ admits a meromorphic function with a single simple pole.

Proof. Pick any point p ∈ Σ and a local holomorphic coordinate z centered at p. Let χ be a
bump function at p, that is χ ≡ 1 on a neighbourhood V of p and χ vanishes identically outside
of a slightly larger neighbourhood W ⊃ V .

Draft 15 May 24, 2024

 Global analysis

We wish to show that there exists a meromorphic function f on Σ with a unique simple pole
at p. Assume for a moment, however, that such f does exist and consider
a
h := f − χ ,
z
where a is the residue of f at p. Then h is smooth everywhere, albeit may fail to be holomorphic
on W \V . In any case, we have

∂h ¯ − a ∂χ
¯ = ∂f ¯ = − a ∂χ.
¯ (1.56)
z z
¯ is a smooth (0, 1)-form supported in W \V .
Notice that η := − az ∂χ
Conversely, if h is a smooth solution of (1.56), we can define f by
a
f := h + χ ,
z
which is meromorphic and has a unique simple pole at p.
Thus, our task reduced to showing that (1.56) has a solution. By the definition of H 0,1 (Σ),
this is the case if and only if the class of η in H 0,1 (Σ) vanishes. However, by (ii), we have

dimR H 1,0 (Σ) + dimR H 0,1 (Σ) = 2b1 (Σ) = 0 =⇒ H 0,1 (Σ) = {0}

so that [η] vanishes trivially. This finishes the proof. □

Theorem 1.57. Let Σ be a closed Riemann surface. Then Σ is homeomorphic to CP 1 ∼

= S 2 if
and only if Σ is biholomorphic to CP 1 .

Idea of the proof. By the classification theorem for closed surfaces, Σ is homeomorphic to S 2
if and only if b1 (Σ) = 0. By Corollary 1.55, there exists a meromorphic function f on Σ with a
unique simple pole. We can view f as a holomorphic map f : Σ −→ CP 1 . A simple topological
argument yields that f must be in fact bijective (this uses crucially that the pole of f is unique
and simple). The reader can find the details of this topological argument in [Don, Sec 4.1]. □
Another reformulation of the above theorem is that the sphere S 2 admits a unique structure
of a complex manifold. This is in contrast with the torus T2 (or, in fact, any closed orientable
surface with b1 ≥ 2), which admits continuous families of inequivalent complex structures.
Developing these ideas somewhat further one can obtain also a classification of all elliptic
curves, that is Riemann surfaces homeomorphic to the torus. Or, one can show that any closed
Riemann surface can be embedded into some projective space. However, this goes somewhat
beyond the purposes of this course.

Draft 16 May 24, 2024

Chapter 2

Vector bundles, Sobolev spaces, and

elliptic partial differential operators

2.1 Vector bundles

Basic definitions. Roughly speaking, a vector bundle is just a family of vector spaces parametrized
by points of a manifold (or, more generally, of a topological space).
More formally, the notion of a vector bundle is defined as follows.

Definition 2.1. Choose a non-negative integer k. A real smooth vector bundle of rank k is a
triple (π, E, M ) such that the following holds:

(i) E and M are smooth manifolds, π : E → M is a smooth submersion (the differential is

surjective at each point);
(ii) For each m ∈ M the fiber Em := π −1 (m) has the structure of a vector space and Em ∼
=
Rk ;
(iii) For each m ∈ M there is a neighborhood U ∋ m and a smooth map ψU such that the
following diagram

ψU
π −1 (U ) U × Rk
π
pr1
U

commutes. Moreover, ψU is a fiberwise linear isomorphism.

The following terminology is commonly used: E is the total space, M is the base, π is the
projection, and ψU is the local trivialization (over U ).
It is worth pointing out that one can equally well talk about complex and quaternionic vector
bundles. This requires only cosmetic changes, which are left to the reader. The preference for
real vector bundles in this section is given for the sake of definitness mainly. I shall feel free to
use complex vector bundles below without further explanations.
Examples of vector bundles known already to the reader include the product bundle M × Rk
and the tangent bundle T M of any smooth manifold M .

17
 Global analysis

Example 2.2 (Tautological bundle of a projective space). Consider the set

E := [x], y ∈ RPn × Rn+1 y = 0 or [y] = [x] .


I leave to the reader to check that E is a smooth manifold and that the natural map π([x], y) = [x]
is a smooth submersion. To show that E is locally trivial, consider a standard chart (or, rather,
a parametrization) of RPn :
h0 : Rn → U0 := [x0 : . . . : xn ] | x0 ̸= 0 ,

h0 (z1 , . . . , zn ) = [1 : z1 : . . . : zn ].
Over U0 we have the isomorphism
U0 × R → π −1 (U0 ),


[1 : z1 : . . . : zn ], λ 7→ [λ : λz1 : . . . : λzn ].
The inverse of this map yields a trivialization ψ0 of E over U0 . Trivializations over other charts
Uj = {xj ̸= 0} can be constructed in a similar way. Thus, E is a rank k = 1 real vector bundle.
Let E and F be two vector bundles over a common base M . A homomorphism between E
and F is a smooth map φ : E → F such that the diagram
ϕ
E F
πE
πF
M

commutes and φ is a fiberwise linear map.

Two bundles E and F are said to be isomorphic, if there is a homomorphism φ, which is
fiberwise an isomorphism.
A bundle E is said to be trivial, if E is isomorphic to the product bundle.

Operations on vector bundles. Let E and F be two vector bundles over a common base M .
Then we can construct new bundles E ∗ , Λp E, E ⊕ F, E ⊗ F, and Hom(E, F ) as follows:
(∗) (E ∗ )m = (Em )∗ ;
(Λ) (Λp E)m = Λp (Em );
(⊕) (E ⊕ F )m := Em ⊕ Fm ;
(⊗) (E ⊗ F )m := Em ⊗ Fm ;
(Hom) Hom(E, F )m := Hom(Em , Fm ).

If f : M ′ → M is a smooth map, we can define the pull-back of E → M via

(f ∗ E)m′ := Ef (m′ ) .
For example, if M ′ is an open subset of M and ι is the inclusion, then E|M ′ := ι∗ E is just the
restriction of E to M ′ .
The reader should check that the families of vector spaces defined above satisfy the properties
required by Definition 2.1.
Exercise 2.3. Prove that E ∗ ⊗ F is isomorphic to Hom(E, F ).
Exercise 2.4. Prove that the tangent bundle of the 2-sphere is non-trivial. (Hint: Apply the
hairy ball theorem).

Draft 18 May 24, 2024

 Global analysis

Sections. Speaking informally, a section is an assignment of a vector s(m) ∈ Em to each

point m ∈ M such that s(m) depends smoothly on m. More formally, we have the following.

Definition 2.5. A smooth map s : M → E is called a section, if π ◦ s = idM .

Sections of the tangent bundle T M are called vector fields. Sections of Λp T ∗ M are called
differential p-forms.

Exercise 2.6. Let E → M be a vector bundle of rank k and U ⊂ M be an open subset. Prove
that E is trivial over U if and only if there are k sections e = (e1 , . . . , ek ), ej ∈ Γ(U ; E),
such that e(m) is a basis of Em for each m ∈ U . More precisely, given e show that ψU can be
constructed according to the formula

ψU−1 : U × Rk −→ E|U , (m, x) 7→ e(m) · x.

In fact this establishes a one-to-one correspondence between k-tuples of pointwise linearly

independent sections and local trivializations of E.

We denote by Γ(E) = Γ(M ; E) the space of all smooth sections of E. Clearly, Γ(E) is a
vector space, where the addition and multiplication with a scalar are defined pointwise. In fact,
Γ(E) is a C ∞ (M )-module.
Given a local trivialization e over U (cf. Exercise 2.6) and a section s, we can write
k
X
s(m) = σj (m)ej (m)
j=1

for some functions σj : U → R. Thus, locally any section of a vector bundle can be thought of
as a map σ : U → Rk .
It is important to notice that σ depends on the choice of a local trivialization. Indeed, if e′ is
another local trivialization of E over U ′ , then there is a map

g : U ∩ U ′ −→ GLn (R) such that e = e′ · g. (2.7)

If σ ′ : U ′ → Rk is a local representation of s with respect to e′ , we have

s = e′ σ ′ = eg −1 σ ′ = eσ =⇒ σ ′ = gσ.

Covariant derivatives. The reader surely knows from the basic analysis course that the notion
of the derivative is very useful. It is natural to ask whether there is a way to differentiate sections
of bundles too.
To answer this question, recall the definition of the derivative of a function f : M → R.
Namely, choose a smooth curve γ : (−ε, ε) → M and denote m := γ(0), v := γ̇(0) ∈ Tm M .
Then
f (γ(t)) − f (m)
df (v) = lim . (2.8)
t→0 t
Trying to replace f by a section s of a vector bundle, we immediately run into a problem,
namely the difference s(γ(t)) − s(m) is ill-defined in general since these two vectors may lie
in different vector spaces.
Hence, instead of trying to mimic (2.8) we will define the derivatives of sections axiomatically,
namely asking that the most basic property of the derivative—the Leibnitz rule—holds.

Draft 19 May 24, 2024

 Global analysis

Definition 2.9. Let E → M be a vector bundle. A covariant derivative is an R-linear map

∇ : Γ(E) → Γ(T ∗ M ⊗ E) such that

∇(f s) = df ⊗ s + f ∇s (2.10)

holds for all f ∈ C ∞ (M ) and all s ∈ Γ(E).

Example 2.11. Let M ⊂ RN be an embedded submanifold. Then the tangent bundle T M is

naturally a subbundle of the product bundle RN := M × RN . In particular, any section s of
T M can be regarded as a map M → RN . With this at hand we can define a connection on T M
as follows
∇s := pr(ds),
where pr is the orthogonal projection onto T M . A straightforward computation shows that this
satisfies the Leibniz rule, i.e., ∇ is a connection indeed.

Theorem 2.12. For any vector bundle E → M the space of all connections A(E) is an affine
space modelled on Ω1 (End E) = Γ T ∗ M ⊗ End(E) .

To be somewhat more concrete, the above theorem consists of the following statements:

(a) A(E) is non-empty.

ˆ the difference ∇−∇
(b) For any two connections ∇ and ∇ ˆ is a 1-form with values in End(E);
(c) For any ∇ ∈ A(E) and any a ∈ Ω1 (End E) the following

(∇ + a)s := ∇s + as

is a connection.

For the proof of Theorem 2.12 we need the following elementary lemma, whose proof is
left as an exercise.

Lemma 2.13. Let A : Γ(E) → Ωp (F ) be an R-linear map, which is also C ∞ (M )-linear, i.e.,

A(f s) = f A(s) ∀f ∈ C ∞ (M ) and ∀s ∈ Γ(E).

Then there exists a ∈ Ωp Hom(E, F ) such that A(s) = a · s. □

Proof of Theorem 2.12. Notice first that A(E) is convex, i.e., for any ∇, ∇ ˆ ∈ A(E) and any
ˆ
t ∈ [0, 1] the combination t∇ + (1 − t)∇ is also a connection.
If ψU is a local trivialization of E over U , then we can define a connection ∇U on E|U by
declaring
∇U s := ψU−1 d(ψU (s)).
Using a partition of unity and the convexity property, a collection of these local covariant
derivatives can be sewed into a global covariant derivative just like in the proof of the existence
of Riemannian metrics on manifolds, cf. [BT03, Thm. 3.3.7]. This proves (a).
By (2.10), the difference ∇ − ∇ˆ is C ∞ (M )–linear. Hence, (b) follows by Lemma 2.13.

The remaining step, namely (c), is straightforward. This finishes the proof of this theorem.
□

Draft 20 May 24, 2024

 Global analysis

While Theorem 2.12 answers the question of the existence of connections, the reader may
wish to have a more direct way to put his hands on a connection. One way to do this is as
follows.
Let e be a local trivialization. Since e is a pointwise base we can write

∇e = e · A, (2.14)

where A = A(∇, e) is a k × k-matrix, whose entries are 1-forms defined on U . A is called the
connection matrix of ∇ with respect to e.
If σ is a local representation of a section s, then


∇s = ∇(eσ) = ∇(e)σ + e ⊗ dσ = e Aσ + dσ .

Hence, it is common to say that locally

∇ = d + A,

which means that dσ + Aσ is a local representation of ∇s. In particular, ∇ is uniquely

determined by its connection matrix.

2.2 Sobolev spaces

2.2.1 Sobolev spaces on Rn
Recall that the space of square-integrable functions is defined by
n Z o
L (R ) := u : R → C | u is measurable &
2 n n
|u (x) |2 dx < ∞ .
Rn

This is a (complex) Hilbert space with respect to the Hermitian scalar product
Z
⟨u, v⟩L2 := u (x) v̄ (x) dx.
Rn

Remark 2.15. Strictly speaking, we should also identify those functions, which differ only on a
subset of measure zero so that L2 (Rn ) consists of classes of functions. However, this will not
be an issue for us and we shall treat square-integrable functions as honest functions.

The Fourier transform

For f : Rn → C decaying sufficiently fast at ∞ the Fourier transform fb: Rn → C is defined by
Z
1
fb(ξ) := f (x) e−i⟨x,ξ⟩ dx.
(2π)n Rn

The Plancherel theorem states that

2 1 2
∥fb∥L2 = n ∥f ∥L2 , (2.16)
(2π)

that is (an extension of) the map f 7→ fb is essentially an isometry of L2 (Rn ).

Draft 21 May 24, 2024

 Global analysis

Let α = (α1 , . . . , αn ) ∈ Zn≥0 be a multi-index. Denote

 |α|
α 1 ∂ |α| f
D f := ,
i ∂xα1 1 · · · ∂xαnn
where |α| = α1 + . . . + αn . The basic property of the Fourier transform is

D\
α f (ξ) = ξ α fb (ξ) , (2.17)
where ξ α = ξ1α1 . . . ξnαn .
Another basic property of the Fourier transform is the following.
Lemma 2.18 (Riemann-Lebesgue). If u ∈ L1 (Rn ), then ub ∈ C 0 (Rn ) and u
b decays at ∞. □
Denote by f ∨ (ξ) = (2π)n fb(−ξ) = f (x) ei⟨x,ξ⟩ dx.
R

1
Theorem 2.19 (The Fourier inversion theorem). If f ∈L (Rn ) and fb
∈ L
1 n
 (R ), then f agrees
∨
almost everywhere with a continuous function f0 and fb = f0 = fc∨ . □

Sobolev spaces
Let u ∈ L2 (Rn ). For α ∈ Zn≥0 , we say that v ∈ L1loc (Rn ) is the αth -weak derivative if

∂ |α| ψ
Z Z
|α|
u = (−1) v·ψ
Rn ∂xα1 1 . . . ∂xαnn Rn

holds for all test functions ψ contained in

C0∞ (Rn ) := {ψ ∈ C ∞ (Rn ) | supp ψ is compact} .
The αth weak derivative does not need to exist in general, however if it does exist we simply
write
∂ |α| u
v=
∂xα1 1 . . . ∂xαnn
keeping in mind that the above equality holds in the weak sense.
With this understood, given any integer k ≥ 0 we can define
∂ |α| u
 
k n 2 n 2 n
H (R ) = u ∈ L (R ) | exists and belongs to L (R ) ∀α s.t. |α| ≤ k .
∂xα1 1 . . . ∂xαnn
Somewhat less formally, this is expressed as follows:
X
u ∈ H k (Rn ) iff ∥u∥2H k := ∥Dα u∥2L2 < ∞. (2.20)
|α|≤k

k
b (ξ) ∈ L2 (Rn ) .
Properties (2.16) and (2.17) imply that u ∈ H k (Rn ) if and only if (1 + |ξ|2 ) 2 u
Hence, we can equally well define H k (Rn ) by
 Z 
k n 2 2 k 2


H (R ) := u | ∥u∥H k := 1 + |ξ| |b u (ξ) | dξ < ∞ .
Rn

Moreover, the norm appearing in the above definition is equivalent to (2.20). The advantage of
the above definition is that this makes sense for any k ∈ R, not just non-negative integers.
The following are basic results about Sobolev spaces.

Draft 22 May 24, 2024

 Global analysis

Proposition 2.21. Assume u ∈ H k (Rn ), where κ := k − n2 > l ∈ N0 . Then u ∈ C l (Rn ) after

changing u on a subset of measure zero if necessary. Moreover, the natural inclusion

H k (Rn ) −→ C l (Rn )

is bounded, that is ∥u∥C l ≤ C∥u∥H k for some positive constant C independent of u.

Proof. Using (2.17) we obtain
Z Z Z
α

|α|
|D u| (ξ) dξ = |ξ ||b
d α u (ξ) |dξ ≤ 1 + |ξ|2 2 |b u (ξ) |dξ
Z

l
≤ 1 + |ξ|2 2 |b
u (ξ) |dξ,

provided |α| ≤ l.
Furthermore, by the Cauchy-Schwartz inequality we have
Z Z

l
2 2

k
l−k
1 + |ξ| |b u (ξ) |dξ = 1 + |ξ|2 2 |b
u (ξ) | · 1 + |ξ|2 2 dξ
Z  21 Z  21
2 k 2 2 l−k



≤ 1 + |ξ| |b
u (ξ) | dξ 1 + |ξ| dξ

The last integral converges since |ξ|2(l−k) · |ξ|n−1 = |ξ|β , where β = 2 (l − k) + n − 1 <
−n + n − 1 = −1. Hence, we obtain
Z
|D
d α u| (ξ) dξ ≤ C∥u∥ k .
H

Hence, Dα u ∈ C 0 by a combination of the Riemann-Lebesgue lemma and the Fourier inversion

theorem.
Moreover,
∥Dα u∥C 0 ≤ ∥D d α u∥
L1 ≤ C∥u∥H k ,

which finishes the proof. □

Proposition 2.22 (Rellich). Suppose uj ∈ H k (Rn ) is a sequence such that there exists a
compact subset K ⊂ Rn containing supp uj for all j. If ∥uj ∥H k is bounded, then for any
s < k there is a subsequence ujl , which converges in H k (Rn ). □

2.2.2 Sobolev spaces on manifolds

In the case when the base is a closed oriented Riemannian manifold M rather than Rn , the
definition of the L2 -spaces generalizes in a straightforward manner. Namely,
( Z 1 )
2
L2 (M ) := u : M → R | ∥u∥L2 := |u|2 vol <∞ .
M

Let E be an Euclidean vector bundle over a manifold M . That means that each fiber Em is
equipped with a scalar product ⟨·, ·⟩m and this scalar product depends smoothly on m. More
formally, just like in Definition 1.18, an Euclidean structure on E is a smooth section ⟨·, ·⟩ of
E ∗ ⊗ E ∗ such that
⟨v, w⟩ = ⟨w, v⟩ and ⟨v, v⟩ > 0

Draft 23 May 24, 2024

 Global analysis

holds for all v, w ∈ Em and all m ∈ M ; In addition, in the last inequality we assume v ̸= 0.
In any case, the definition of the L2 -spaces for sections of Euclidean bundles generalizes in
a straightforward manner. Namely, if M is a compact oriented Riemannian manifold, then
( Z 1 )
2
L2 (E) := s | ∥s∥L2 := |s|2 vol <∞ .
M

Furthermore, pick a connection ∇ ∈ A (E). It is convinient to assume that ∇ is Euclidean, that

is
d⟨s1 , s2 ⟩ = ⟨∇s1 , s2 ⟩ + ⟨s1 , ∇s2 ⟩
holds for all smooth sections s1 and s2 of E.
Remark 2.23. Any Euclidean vector bundle admits an Euclidean connection. Moreover, the
space of all Euclidean connections on E is an affine space modeled on

Ω1 (M ; o (E)) := Γ (T ∗ M ⊗ o (E)) , where o (E) := {A ∈ End (E) | A∗ = −A} .

These facts can be established by a straightforward modification of the proof of Theorem 2.12.
We can define
1
n
2 2

21 o
H (E) := s | ∥s∥H 1 := ∥s∥L2 + ∥∇s∥L2 < ∞ .

Remark 2.24. If s is smooth, then ∇s is a 1-form with values in E. Then

n
X
|∇s|2m := |∇ei s|2 ,
i=1

where (e1 , . . . , en ) is an orthonormal basis of Tm M . Thus, somewhat more precisely by ∥∇s∥L2

we mean Z  21
2
∥∇s∥L2 = |∇s|m volm .
M

Just like in the case of R , H (E) can be understood in at least two following ways. First,
n 1

we can define H 1 (E) as the completion of Γ (E) with respect to the ∥ · ∥H 1 -norm. Secondly,
for s ∈ L2 (E) we can first define the weak covariant derivative ∇s as a functional acting on
Ω1 (E) and ask ∇s to lie in L2 (T ∗ M ⊗ E).
To describe some details concerning the second approach, notice that akin to the de Rham
complex for any connection ∇ we have the sequence
∇=d d d
0 → Ω0 (E) = Γ(E) −−−−∇
→ Ω1 (E) −−∇→ . . . −−∇→ Ωn (E) → 0, (2.25)

where for ω ⊗ s ∈ Ωk (E) the map d∇ is defined by

d∇ (ω ⊗ s) = (dω) ⊗ s + (−1)k ω ∧ ∇s.

Notice, however, that (2.25) is not a complex in general, that is d∇ ◦ d∇ does not necessarily
vanish.
In any case, just like for the de Rham differential we can also define

d∗∇ : Ωk+1 (E) → Ωk (E) by d∗∇ := (−1)nk+1 ∗ d∇ ∗,

Draft 24 May 24, 2024

 Global analysis

cf. Proposition 1.24. Here the Hodge operator acts as follows: ∗(ω ⊗ s) = (∗ω) ⊗ s. Assuming
∇ is Euclidean, d∗∇ is the formal adjoint of d∇ , that is

⟨d∇ α, β⟩L2 = ⟨α, d∗∇ β⟩L2

holds for all α ∈ Ωk (E) and all β ∈ Ωk+1 (E).

With these preliminaries at hand, if s ∈ L2 (E) the value of the weak derivative ∇s on
ψ ∈ Ω1 (E), which has a compact support, is declared to be

⟨∇s, ψ⟩ := ⟨s, d∗∇ ψ⟩L2 .

Then we can define H 1 (E) as a subspace of L2 (E) consisting of those s, whose weak derivative
belongs to L2 , that is if there exists w ∈ L2 (T ∗ M ⊗ E) such that

⟨w, ψ⟩L2 = ⟨s, d∗∇ ψ⟩L2

holds for any ψ ∈ Γ (E). Of course, in this case we must have w = ∇s.
Our definition of H 1 (E) depends a priory on the choice of the connection ∇. It turns out
that this dependence is not really essential as the following result shows.

Proposition 2.26. Pick any two Euclidean connections ∇ and ∇′ on E. If M is compact, then
the corresponding norms ∥ · ∥H 1 and ∥ · ∥′H 1 are equivalent.

Proof. By Theorem 2.12, we have ∇′ = ∇ + a, where a ∈ Ω1 (End (E)) (in fact a takes values
in o (E), but this is immaterial here). Denote

A := sup |am | = sup sup |am (v) | < ∞.

m∈M m∈M v∈Tm M
|v|=1

Then for any v ∈ Tm M of unit norm and any smooth section s we have

∇′v s = ∇v s + a (v) · s =⇒ |∇′v s| ≤ |∇v s| + A|s|.

Hence, for any orthonormal basis (e1 , . . . , en ) of Tm M we obtain

n n n
!
X X X
|∇′ s|2m = |∇′ei s|2 ≤ (|∇ei s| + A|s|)2 ≤ 2 |∇ei s|2 + nA2 |s|2
i=1 i=1 i=1
(2.27)
= 2|∇s|2m 2
+ 2n A |s| . 2

Here to obtain the second inequality we used the mean inequality

 12
a2 + b 2

a+b
≤ ,
2 2
where a and b are positive real numbers.
With the help of (2.27) we obtain

∥∇′ s∥L2 ≤ C1 ∥∇s∥L2 + C2 ∥s∥L2

for some positive constants C1 and C2 independent of s. This yields the claim of this proposition.
□

Draft 25 May 24, 2024

 Global analysis

Our next aim is to define the Sobolev spaces H k (E) with k ≥ 1 being an integer. To this
end notice first the following.

Lemma 2.28.
(i) If ∇ is a connection on a vector bundle E, then there exists a unique connection ∇∗ on
E ∗ such that
d⟨σ, s⟩ = ⟨∇∗ σ, s⟩ + ⟨σ, ∇s⟩ (2.29)
holds for all σ ∈ Γ (E ∗ ) and all s ∈ Γ (E). Here ⟨·, ·⟩ denotes the natural pointwise
pairing E ∗ ⊗ E → R.
(ii) If ∇E and ∇F are connections on vector bundles E and F respectively, then there exists
a unique connection ∇ on E ⊗ F such that

∇ (s ⊗ t) = ∇E s ⊗ t + s ⊗ ∇F t (2.30)

holds for all s ∈ Γ (E) and all t ∈ Γ (F ).

The proof of this lemma can be obtained by a straightforward verification that ∇∗ and ∇
defined by (2.29) and (2.30) respectively satisfies the Leibnitz rule. I leave this as an exercise
to the reader.
With this understood, we can proceed as follows. Pick an Euclidean connection ∇E on
E
TM E TM
and an Euclidean connection ∇ on T M . This yields a connection ∇ = ∇ ∇ , ∇ on
T ∗ M ⊗ E. Hence, for any smooth section s of E we can define the second derivative by

∇2 s = ∇ ∇E s ∈ Γ (T ∗ M ⊗ T ∗ M ⊗ E) .

In a less regular case, for example when s ∈ H 1 (E) we can still define the weak second
derivative as a functional on Γ (T ∗ M ⊗ T ∗ M ⊗ E) just like we defined the weak first derivative.
Then we may set

H 2 (E) = s ∈ H 1 (E) | ∇2 s ∈ L2 (E) ,



= s | ∥s∥pH 2 := ∥s∥2L2 + ∥∇s∥2L2 + ∥∇2 s∥2L2 < ∞ ,



where the second equality should be treated with care just like in the case of functions.
Furthermore, we can define H k (E) for any integer k ≥ 2 by induction. The details are left
to the reader.
Alternatively, we can also define H k (E) as the closure of Γ (E) with respect to ∥ · ∥H k ,
where
Xk
2
∥s∥H k := ∥∇j s∥2L2 . (2.31)
j=0

Yet another way to define Sobolev spaces, which works for any real k, is as follows. Since
M is compact, we can pick a finite covering (Uα , ψα ), where ψα is a trivialization of E Uα just
like in Definition 2.1. Moreover, we can assume that each Uα is a coordinate chart. Let {ρα } be
a partition of unity subordinate to {Uα }, that is

• supp ρα ⊂ Uα ;

• ρα ≥ 0 everywhere;
P
• α ρα (m) = 1 for each m ∈ M .

Draft 26 May 24, 2024

 Global analysis

If s is a section of E, over each Uα we can identify s with some sα : Uα → Rl , where l is the

rank of E, as follows:
ψα ◦ s = (id, sα ) .
Hence, we can think of ρα · sα as a map defined on Rn with compact support. Finally, we set
X
∥s∥2H k := ∥ρα · sα ∥2H k .
α

This norm depends on the choice of {(Uα , ψα , ρα )}, however turns out to be equivalent to (2.31)
if k is an integer. Hence, we can define H k (E) in the usual way, for example as the completion
of Γ (E) with respect to the above norm.
With this understood, we have the sequence of inclusions

L2 (M ; E) = H 0 (M ; E) ⊃ H 1 (M ; E) ⊃ H 2 (M ; E) ⊃ . . .

Relations between all these spaces is given by the following theorem, which is of fundamental
importance in the theory of PDEs.
Theorem 2.32. Let M be a compact manifold.
(i) The natural inclusion
j : H k (M ; E) ⊂ H l (M ; E)
is a compact operator provided k > l. This means that if k > l any sequence bounded in
H k has a subsequence, which converges in H l .
(ii) We have a natural continuous embedding

H k (M ; E) ⊂ C l (M ; E)

provided k − n
2
> l. In particular, if s ∈ H k (M ; E) for all k ≥ 0, then s ∈ C ∞ (M ; E).

2.3 Differential operators

Let E be a smooth complex vector bundle of rank a over M. This means that each fiber Em
has the structure of a complex vector space and its dimension equals a; Also, Property (iii) of
Definition 2.1 should be read as follows: For each m ∈ M there exists a neighbourhood U ∋ m
and a smooth map ψU show that the following diagram
ψU
π −1 (U ) U × Ca
π
pr2

U
commutes. Moreover, ψU is a fiberwise complex linear isomorphism.
Given a local trivialization ψU as above, any section s ∈ Γ (E) can be identified with a
smooth map σ : U → Ca in the following sense:

ψU ◦ s = (idU , σ) .

In other words, we have a well-defined C ∞ (M ; C)-linear isomorphism

ΨU : Γ E U −→ C ∞ (U ; Ca ) .

Draft 27 May 24, 2024

 Global analysis

Let F be another complex vector bundle over M of rank b. Then F admits a local trivialization
over some neighbourhood U ′ of M . Replacing U and U ′ by U ∩ U ′ if necessary, we can assume
U ′ = U . Notice also that by shrinking U further if necessary, we can assume that U is a chart.
Denote by (x1 , . . . , xn ) local coordinates on U . Let ψUF be the trivialization of F over U so that
sections of F over U can be thought of as maps U −→ Cb .
Definition 2.33. A C-linear map L : Γ (E) → Γ (F ) is called a differential operator of order
at most l, if for each choice of local trivializations of E and F as above there is a differential
operator L̃ : C (U ; C ) → C U ; C of order at most l such that the following diagram
∞ ∞
a b

L

Γ E U
Γ F U

ΨU ΨF
U

L̃
C ∞ (U ; Ca ) C ∞ U ; Cb

commutes. Here L̃ is said to be a differential operator of order at most l, if L̃ admits a

representation
X ∂ |α| σ
L̃σ = aα (x) α1 , (2.34)
∂x1 . . . ∂xαnn
|α|≤l
∞
where aα ∈ C (U ; Ma×b (C)) and Ma×b (C) denotes the space of all a × b-matrices with
complex entries.
Remark 2.35. We can equally well define real differential operators acting on sections of real
vector bundles. This requires cosmetic changes only. The choice to focus on complex differential
operators will be somewhat clearer below.
Example 2.36. Let M be any manifold of dimension n. Recall that any chart (U, x1 , . . . xn )
yields a trivialization of T ∗ M . The corresponding map ΨU : Ω1 (U ) → C ∞ (U ; Rn ) is given by
 
n ω 1
ωi (x) dxi 7−→  ...  .
X
ω=
 
i=1 ωn
Consider the differential as the map d : Ω0 (M ) = Γ (R) → Ω1 (M ). Since R is globally trivial,
there is no need to choose an extra local trivialization of R. Then, relative to the above choice
of the local trivialization of T ∗ M , the local representation of d is given by
   
 ∂f  1 0
∂x1  0  ∂f  0  ∂f
f 7−→  ...  =  ..  · + . . . +  ..  · .
     
∂f
 .  ∂x 1  .  ∂x n
∂xn 0 1
In particular, d is a first order (real) differential operator.
By the same token, d : Ωk (M ) → Ωk+1 (M ) is a first order real differential operator for
each k.
Let Diff l (E, F ) denote the vector space of all differential operators of order at most l.
Proposition 2.37. Any L ∈ Diff l (E, F ) extends as a bounded map L : H k (E) → H k−l (F ).
The proof of this proposition follows immediately from the property D
d α f (ξ) = ξ α fb (ξ)

and the definition of H k -norm in terms of the Fourier transform.

Draft 28 May 24, 2024

 Global analysis

2.3.1 Symbols of differential operators

Consider the following second order differential operator
n n
X ∂ 2u X ∂u
Lu = aij (x) + bi (x) + c (x) u, (2.38)
i,j=1
∂xi ∂xj i=1
∂xi

where u is a function of n variables (x1 , . . . , xn ). It is well-known from the theory of linear

PDEs that in many cases the most essential properties of L depend on the highest order terms
only, that is on
n
(2)
X ∂ 2u
L u= aij (x) .
i,j=1
∂xi ∂xj

It is convenient to represent this operator by the expression

n
X
σ (L) := − aij (x) ξi ξj ,
i,j=1

where ξ = (ξ1 , . . . , ξn ) can be understood as a formal variable and the negative sign is a
convention. This leads to the notion of the symbol of a differential operator.
It is important to understand how the symbol changes if we change the variables, since on a
manifold there are no preferred coordinates in general. Thus, let

y1 = y1 (x1 , . . . , xn )
...
yn = yn (x1 , . . . , xn )

be new coordinates on Rn . If u (x) = v (y (x)), then we have

n n n
∂u X ∂v ∂yp ∂ 2u X ∂ 2 v ∂yp ∂yq X ∂v ∂ 2 yp
= =⇒ = + .
∂xi p=1
∂yp ∂xi ∂xi ∂xj p,q=1 ∂yp ∂yq ∂xi ∂xj p=1 ∂yp ∂xi ∂xj

Substituting this into (2.38) we obtain

n
X ∂yp ∂yq ∂ 2 v
Lu = aij + . . . =: L̃v
i,j,p,q=1
∂xi ∂xj ∂yp ∂yq

where ". . ." denotes the lower order terms. Therefore, the symbol of L̃ is given by
n n
!

X X ∂yp ∂yq
σ L̃ (η) = − aij ηp ηq .
p,q=1 i,j=1
∂xi ∂xj

Hence, if we set
n n
X ∂yp X ∂yq
ξi = ηp =⇒ ξj = ηq ,
p=1
∂xi q=1
∂xj

then

σ(L̃) (η) = σ(L) (ξ) ⇐⇒ σ(L̃) (η) = σ L (Jη)

Draft 29 May 24, 2024

 Global analysis

holds identically for all η ∈ Rn , where

 ∂y1 ∂y2 ∂yn

∂x ∂x1
... ∂x1
 ∂y11 ∂y2 ∂yn  t
...

 ∂x ∂x2 ∂x2 
 ∂y
J =  2 = .
... ... ... ...
 ∂x
∂y1 ∂y2 ∂yn
∂xn ∂xn
... ∂xn

This implies the following. Think of (x, ξ) as coordinates on T ∗ Rn ∼ = Rn × Rn . If we change

the coordinates x on Rn for y as above, we obtain new coordinates (y, η) on T ∗ Rn . Moreover,
these coordinates are related by
 t
∂y
y = y (x) and ξ= · η.
∂x



Hence σ L̃ (η) = σ L̃ (y, η) is just the expression for σ (L) in these new coordinates. Put
differently, σ (L) (ξ) = σ (L) (x, ξ) is well-defined as a function on T ∗ Rn .
Notice that σ (L) (ξ) is homogeneous of degree 2 in ξ, that is

σ (L) (λξ) = λ2 σ (L) (ξ)

for all λ ∈ R.
In general, we can proceed as follows. For vector bundles E and F as above, consider
π E −→ T ∗ M and π ∗ F −→ T ∗ M , where π : T ∗ M −→ M is the cotangent bundle. For any
∗

l ∈ Z set

Smbl (E; F ) := σ ∈ Γ (Hom (π ∗ E, π ∗ F )) | σ (m, λξ) = λl σ (m, ξ) ,


(2.39)

where λ > 0 and (m, ξ) ∈ T ∗ M .

For any linear differential operator L : Γ (E) −→ Γ (F ) of order at most l as above, its
symbol σ (L) is an element of Smbl (E; F ). This can be defined via the local representations
of L just like we did above in the case of the second order operators acting on functions. To be
more precise, if L̃ is a local representation of L just like in (2.34), define
X
σ L̃ (x, ξ) := il aα (x) ξ α ∈ Ma×b (C) .

|α|=l

This yields a map

π −1 (U ) = U × Rn −→ Ma×b (C) ,
where U
⊂ M is as in Definition 2.33. A computation similar to the one we did above shows
that σ L̃ is a local representation of a well-defined section σ (L) of Smbl (E; F ).
Alternatively, somewhat more invariantly, one can also define the symbol as follows. Given
(m, ξ) ∈ T ∗ M and e ∈ Em , pick f ∈ C ∞ (M ) and s ∈ Γ (E) such that f (m) = 0, dfm = ξ,
and s (m) = e. Define  l 
i l
σ (L) (m, ξ) e := L fs ∈ Fm
l! m
It is then easy to check that this is equivalent to our definition of the symbol in terms of local
representations of L.
Denote by Diff l (E; F ) the vector space of all linear differential operators L : Γ (E) →
Γ (F ) of order at most l. Thus, we obtain a linear map σ : Diff l (E; F ) → Smbl (E; F ), which
has the following properties.

Draft 30 May 24, 2024

 Global analysis

Proposition 2.40.

(i) For any l ∈ Z, l ≥ 1, the sequence

0 −→ Diff l−1 (E, F ) −→ Diff l (E, F ) −→ Smbl (E; F )

is exact.
(ii) σ (L1 ◦ L2 ) = σ (L1 ) ◦ σ (L2 ). □

Example 2.41. Recall that in local coordinates the Laplacian (acting on functions) is given by

1 p 
∆f = − p ∂i |g|g ij ∂j f .
|g|

Hence, for the symbol we have

1 p 
σ (∆) (ξ) = + p ξi |g|g ξj = g ij ξi ξj = |ξ|2 .
ij
|g|

Example 2.42. Consider d : Ωp (M ) −→ Ωp+1 (M ), which is clearly the first order linear
differential operator. Pick a point m ∈ M and a function f ∈ C ∞ (M ) such that f (m) = 0.
We have
d (f ω) m = df ∧ ω m + 0 · dω m = df ∧ ω m .
Hence, σ (d) (m, ξ) (α) = iξ ∧ α.

2.3.2 The formal adjoint of a linear differential operator and its symbol
An important role in what follows is played by the adjoint of a differential operator. We first
prove an auxiliary result, which will be useful below, define the (formal) adjoint operator, and
prove its existence afterwards.
Let E be Hermitian vector bundle. The Hermitian structure is defined just like the Euclidean
one, namely as a family of Hermitian scalar products (·, ·)m on each fiber depending smoothly
on m.

∗
Lemma 2.43. For any non-negative k ∈ R, we have H k (M ; E) ∼ = H −k (M ; E). In
particular, the natural semilinear map

s, t 7−→ (s, t)L2 , s, t ∈ C ∞ (M ; E)

extends to the semilinear map

H k (M ; E) × H −k (M ; E) −→ C

such that | (s, t) | ≤ C∥s∥H k ∥t∥H −k .

Idea of the proof. The proof boils down to showing that for any u, v ∈ C0∞ (Rn ; C) we have
Z
u (x) v̄ (x) dx ≤ C∥u∥H k ∥v∥H −k .
Rn

Draft 31 May 24, 2024

 Global analysis

To see that this inequality holds indeed, notice that the Plancherel theorem yields
Z Z
n
u (x) v̄ (x) dx = (2π) u
b (ξ) vb (ξ)dξ
Rn Rn
Z
n
k
− k
≤ (2π) u (ξ) | 1 + |ξ|2 2 |b
|b v (ξ) | 1 + |ξ|2 2 dξ
Rn
Z  21 Z  21
n 2 2 k 2 2 −k



≤ (2π) u (ξ) | 1 + |ξ|
|b dξ v (ξ) | 1 + |ξ|
|b
Rn Rn
= (2π)n ∥u∥H k · ∥v∥H −k ,

where the last inequality follows from the Cauchy–Schwartz inequality. □

Let F be another Hermitian bundle. Slightly abusing notations, we denote the Hermitian
structure of F still by (·, ·).

Definition 2.44. Let L : Γ (E) → Γ (F ) be a C-linear map. A C-linear map L∗ : Γ (F ) → Γ (E)

is called the formal adjoint of L, if
Z
∗
(Ls, t)L2 = (s, L t)L2 = (s, L∗ t)m volm
M

holds for all s ∈ Γ (E) and all t ∈ Γ (F ).

Proposition 2.45. If L ∈ Diff l (E; F ), then L∗ exists and is unique. Moreover, L∗ ∈ Diff l (F ; E)
and σ (L∗ ) = σ (L)∗ , where σ (L)∗ denotes the pointwise Hermitian-dual map.

Proof. The proof consists of the following steps.

Step 1. The formal adjoint operator is unique.
Indeed, if L∗1 and L∗2 are two formal adjoint operators, then for all s, t as in the definition we
have
0 = (s, (L∗1 − L∗2 ) t)L2 =⇒ L∗1 = L∗2 .
Step 2. L∗ exists.
For each t ∈ C ∞ (F ) consider the functional

s 7−→ ψt (s) = (Ls, t)L2 .

Since
|ψt (s) | ≤ ∥Ls∥H −k ∥t∥H k ≤ C∥t∥H k ∥s∥H l−k , (2.46)
ψt can be viewed as a bounded functional on H l−k . Hence, by the Riesz representation theorem
there exists a unique u = u (t) ∈ H k−l such that ψt (s) = (s, u (t)).
Since ψt depends semilinearly on t, the map t 7−→ u (t) is C-linear. Moreover,

|ψt (s) |
∥ψt ∥H l−k = sup ≤ C ∥u (t) ∥H k−l . (2.47)
s̸=0 ∥s∥H l−k

In fact, ∥ψt ∥ = C ∥u (t) ∥H k−l , since1 , roughly speaking, (2.47) follows from the Cauchy–
Schwartz inequality, which is optimal.
1
A proper justification of this equality requires an extra argument, which I omit here.

Draft 32 May 24, 2024

 Global analysis

We set L∗ t := u (t). Then by (2.46) we obtain

∥L∗ t∥H k−l = ∥u (t) ∥H k−l ≤ C∥t∥H k
for any k ∈ R. In particular, L∗ yields a bounded map H k → H k−l . If t ∈ C ∞ , then L∗ t does
not depend on k and belongs to H k−l for any k ∈ R. Hence, L∗ : C ∞ (F ) → C ∞ (E).
Step 3. Let L̃ : C ∞ (Rn ; Ca ) → C ∞ Rn ; Cb be a linear differential operator of order l. Define

the weighted Hermitian L2 -scalar product on C0∞ (Rn ; Ca ) by

Z
(s, σ)L2 , ρ := (s (x) , σ (x)) ρ (x) dx,
Rn

where ρ is a positive smooth function. Then L̃∗ defined by

L̃s, t L2 , ρ = s, L̃∗ t L2 , ρ ∀s ∈ C0∞ (Rn ; Ca ) and ∀t ∈ C0∞ Rn ; Cb

∗
is also a linear differential operator of order l. Moreover, σ L̃∗ = σ L̃ .

Write L̃ = |α|≤l aα (x) Dα and consider the term corresponding to α = (1, 0, . . . , 0). We
P
have Z Z
α tr
α
(aα D s, t)L2 , ρ = (aα D s) t̄ ρ dx = −i (∂1 s)tr atr
α t̄ ρ dx
RnZ Rn

∂1 str atr tr tr




= −i α t̄ ρ − s ∂1 aα t̄ ρ dx
RZn
Z
tr tr
str ∂1 atr



=0+i s aα ∂1 t ρ dx + i α ρ t̄ dx
Rn Rn
Z Z
str a∗α Dα t ρ dx − str bα t ρ dx,


=
Rn Rn
tr
where bα = − ρi ∂1 (a∗α ρ) and a∗α = (ā) denotes the Hermitian dual matrix. In other words,
for α = (1, 0, . . . , 0) we have (aα Dα )∗ = a∗α Dα − bα . Iterating this argument, we obtain
(aα Dα )∗ = a∗α Dα + Sα , where Sα ∈ Diff |α|−1 . This yields
X
L̃∗ = a∗α (x) Dα + S,
|α|=l

∗
where S ∈ Diff l−1 . In particular, we have σ L̃∗ = σ L̃ .

Step 4. If L ∈ Diff l (E; F ), then L∗ ∈ Diff l (F ; E) and σ (L∗ ) = σ (L)∗ .

Let ψ be a local trivialization of E over an open neighbourhood U . Recall that such a
trivialization is given by a collection (e1 , . . . , ea ) of pointwise linearly independent sections over
U . By applying the Gram-Schmidt orthogonalization process, we can assume that (e1 (m) , . . . , ea (m))
is an orthonormal basis of Em for each m.
Assume U is a chart with local coordinates (x1 , . . . , xn ) and (f1 , . . . , fb ) P
is a trivialization
of F such that (f1 (m) , . . . , fb (m)) is orthonormal at each m ∈ U . Let L̃ = |α|≤l aα (x) Dα
be the local representation of L over U . Write also vol = ρ (x) dx1 ∧ . . . ∧ dxn = ρ (x) dx,
where ρ is a positive function on U . Then for any s ∈ Γ (E) and t ∈ Γ (F ) such that supp s and
supp t are contained in U , we have
(Ls, t)L2 = L̃σ, τ L2 , ρ = σ, L̃∗ τ L2 , ρ ,

where σ and τ are local representations of s and t respectively. Moreover, L̃∗ ∈ Diff l . □

Draft 33 May 24, 2024

 Global analysis

2.4 Pseudodifferential operators

Before giving the formal definition of a pseudodifferential operator, let us consider the following
model case. Assume we want to solve the Poisson equation

∆u (x) = f (x) , x ∈ Rn , (2.48)

where f is a given function on Rn . Assume f ∈ L2 and we are looking for a solution u ∈ H 2

of (2.48). Applying the Fourier transform to both sides, we obtain

1 b
ξ12 + . . . + ξn2 u


b (ξ) = fb(ξ) =⇒ u
b (ξ) = f (ξ)
|ξ|2
ei⟨x,ξ⟩ b
Z
=⇒ u (x) = f (ξ) dξ.
Rn |ξ|
2

In particular, the map, which assigns to f ∈ L2 a solution u ∈ H 2 of (2.48) is not a differential

operator. This leads to the concept of a pseudodifferential operator, which will be very useful
below.
Let U ⊂ Rn be an open set and let l be an integer.

Definition 2.49. The class S̃ l (U ) consists of functions p = p (x, ξ) on U × Rn satisfying the

following: For any compact K ⊂ U we have

|Dxβ Dξα p (x, ξ) | ≤ Cα,β,K (1 + |ξ|)l−|α| , x ∈ K, ξ ∈ Rn , (2.50)

where α and β are multiindices.

Definition 2.51. The class S l (U ) consists of those p ∈ S̃ l (U ) for which the limit σl (p) (x, ξ) =
limλ→∞ p(x,λξ)
λl
exists for ξ ̸= 0 and

p (x, ξ) − ψ (ξ) σl (p) (x, ξ) ∈ S̃ l−1 (U ) , (2.52)

where ψ is a cut-off function such that ψ ≡ 0 near the origin and ψ ≡ 1 outside of the unit ball.
α r
P
Example 2.53. Let p (x, ξ) = |α|≤l pα (x) ξ . If each pα is bounded in any C (U ), then
p ∈ S l (U ).

Definition 2.54 (Local pseudodifferential operator). For any p ∈ S̃ l (U ) and any u ∈ C0∞ (U )
set Z
Lp u (x) := b (ξ) ei⟨x,ξ⟩ dξ.
p (x, ξ) u
Rn

Example 2.55. If p (x, ξ) = |ξ|2 = ξ12 + . . . + ξn2 , then ξj2 u

b (ξ) = D d2
j u so that Lu (x) =
P ∂2u P α
− ∂x2 = ∆u. More generally for any p (x, ξ) = |α|≤l pα (x) ξ the map Lp is a linear
j
differential operator of order l.

Example 2.56. Assume K ∈ C ∞ (U × U ; C) and supp K (x, ·) is compact. Consider the map
Z
Lu (x) = K (x, y) u (y) dy,
U

Draft 34 May 24, 2024

 Global analysis

where u ∈ C0∞ (U ). We have

Z Z Z 
i⟨y,ξ⟩
K (x, y) u (y) dy = K (x, y) e u
b (ξ) dξ dy
U U Rn
Z Z 
i⟨x,ξ⟩ i⟨y−x,ξ⟩
= e e K (x, y) dy ub (ξ) dξ.
Rn U
R
Setting p (x, ξ) := U ei⟨y−x,ξ⟩ K (x, y) dy, we see that L = LUp , that is integral operators are
also pseudodifferential operators at least formally. In fact,
Z
p (x, ξ) = e −i⟨x,ξ⟩
ei⟨y,ξ⟩ K (x, y) dy = e−i⟨x,ξ⟩ K (x, ·)∨ (ξ) .
U

Hence ei⟨x,ξ⟩ p (x, ξ) is essentially the Fourier transform of a compactly supported function and
therefore is rapidly decaying, that is (1 + |ξ|)N |p (x, ξ) | → 0 as ξ → ∞ for any N > 0, see
Step 1 of the proof of Theorem 2.57 for details.
Theorem 2.57. Assume p belongs to the class
n o
l
S̃0 (U ) := p ∈ S̃ (U ) | ∃ cmpt K ⊂ U with supp p (·, ξ) ⊂ K ∀ξ ∈ R .
l n

Then Lp maps C0∞ (Rn ) into C ∞ (Rn ) and its closure yields a bounded map Lp : H k (Rn ) →
H k−l (Rn ).
Proof. The proof consists of the following steps.
Step 1. Lp (C0∞ (Rn )) ⊂ C ∞ (Rn ).
If u ∈ C0∞ (Rn ), then
Z
−n
α
ξ u
b (ξ) = D
d α u (ξ) = (2π) Dα u (x) e−i⟨x,ξ⟩ dx
Rn

u (ξ) | ≤ Cα , since supp u is compact. Hence, for any N ∈ N we have

yields that |ξ α ||b

u (ξ) | ≤ CN (1 + |ξ|)−N ,
|b (2.58)

which implies in turn

b (ξ) | ≤ CN,β (1 + |ξ|)l (1 + |ξ|)−N .

|Dxβ p (x, ξ) u
R

Furthermore, assume N ≫ 1. Then R the integral Dxβ p(x, ξ) û(ξ)ei⟨x,ξ⟩ dξ converges
absolutely for any β. This implies that p(x, ξ)û(ξ)ei⟨x,ξ⟩ dξ is smooth in x.
Step 2. For any ξ, η ∈ Rn and any k ≥ 0 we have the inequality

k2 k
k2
k2
1 + |ξ|2 ≤ 2 2 1 + |ξ − η|2 1 + |η|2 .

By the Cauchy-Schwartz inequality we have

1 + |ζ + η|2 ≤ 1 + (|ζ| + |η|)2 ≤ 1 + 2 |ζ|2 + |η|2 ≤ 2 1 + |ζ|2 1 + |η|2 .

Substituting ζ = ξ − η we obtain 1 + |ξ|2 ≤ 2 (1 + |ξ − η|2 ) (1 + |η|2 ), which implies the claim

of this step.

Draft 35 May 24, 2024

 Global analysis

Step 3. We prove the inequality ∥Lp u∥H k ≤ C∥u∥H k+l .

First notice that we have
Z Z
−n −i⟨x,ξ⟩
L
dp u (ξ) = (2π) e b (η) ei⟨x,η⟩ dηdx
p (x, η) u
Rn
Z Rn
−n
e−i⟨x,ξ−η⟩ p (x, η) dx u


= (2π) b (η) dη
Z
= pb (ξ − η, η) u
b (η) dη.
Rn

Just like (2.58), we obtain

−N
2l
p (ζ, η) | ≤ C 1 + |ζ|2
|b 1 + |η|2 .

Furthermore, using Step 2 we have

Z

−N
l
|L
d p u (ξ) | ≤ C 1 + |ξ − η|2 1 + |η|2 2 |b u (η) |dη
−N
(1 + |ξ − η|2 )
Z

l+k
2 2
≤C k 1 + |η| |b
u (η) |dη
(1 + |η|2 ) 2
Z

k
2 −2

−N + k2
l+k
≤ C 1 + |ξ| 1 + |ξ − η|2 1 + |η|2 2 |b u (η) |dη,

which yields in turn

Z
2

k2
−N + k2
l+k
|L
d p u (ξ) | 1 + |ξ| ≤C 1 + |ξ − η|2 1 + |η|2 2
|b
u (η) |dη.

Applying Young’s inequality ∥f ∗ g∥L2 ≤ ∥f ∥L1 ∥g∥L2 , for N ≫ 1 we obtain

−N + k2
∥Lp u∥H k ≤ C∥ 1 + |ξ|2 ∥L1 ∥u∥H k+l .

□
Let L : C0∞ (M ; C) → C ∞ (M ; C) be a linear map.
Definition 2.59. L is said to be a pseudodifferential operator of order l if for any chart U ⊂ M
and any open subset U ′ ⊂ U such that Ū ′ ⊂ U there exists some

p ∈ S0l (U ) = S̃0l (U ) ∩ S l (U )

with the following property: For any u ∈ C0∞ (U ′ ) we have Lu = Lp u.

Remark 2.60. Pseudodifferential operators are non-local in general, that is supp Lu ̸⊂ supp u
in general. This lack of locality for pseudodifferential operators explains the appearence of U ′
in the above definition. In particular, the function p may depend on U ′ .
Just like in the case of the differential operators we can define the l-symbol of a pseudodifferential
operator L by
p (x, λξ)
σl (p) (x, ξ) = lim , ξ ̸= 0
λ→∞ λl
assuming that this limit exists.
Just like in the case of differential operators we have the following basic result.

Draft 36 May 24, 2024

 Global analysis

Proposition 2.61. Let U be an open bounded subset of Rn equipped with coordinates (x1 , . . . , xn )
and let p ∈ S0l (U ). Let y = f (x) be new coordinates on U . Let L̃ be the representation of
L = Lp in these new coordinates, i.e.,

L̃v := L v ◦ f −1 ◦ f.

Then there is a function q ∈ S0l (f (U )) such that L̃ = Lq . Moreover,

 t !
∂y
σl (q) η = σl (p) η .
∂x

□
The above proposition yields that to any L ∈ PDiff l (M ) we can associate a well-defined
(principal) symbol
σl (L) ∈ Smbl (C; C) ⊂ C ∞ (T ∗ M ; C) .
Notice, that if l is negative, the homogeneity property of the symbol implies that σl (L) may be
singular along the zero section. Thus, strictly speaking, σl is defined away from the zero section
only.
A simple consequence of Theorem 2.57 is the following.

Proposition 2.62. If M is compact, then the closure of any L ∈ PDiff l (M ) yields a linear
bounded operator
L : H k (M ) −→ H k−l (M ) .

□
Let E and F be complex vector bundles over M .

Definition 2.63. A linear map L : C0∞ (M ; E) → C ∞ (M ; F ) = Γ (F ) is called a pseudodifferential

operator if for any chart U such that E and F admit trivializations over U and any U ′ ⊂ U such
that Ū ′ ⊂ U there exists a matrix p = (pij ) with pij ∈ S0l (U ) so that the diagram

L
C0∞ (U ′ ; E) C ∞ (U ; F )
ΨE ΨF
Lp
C0∞ (U ′ ; Ca ) C ∞ U ′ ; Cb

commutes. The matrix p is called the local symbol of L (this depends on the choices made). If
L ∈ PDiff l (E; F ), then we have a well-defined principal symbol σl (L) ∈ Smbl (E; F ).

Theorem 2.64. For any smooth manifold M the sequence

l σ
0 −→ Kl−1 (E; F ) −→ PDiff l (E; F ) −−−
→ Smbl (E; F ) −→ 0 (2.65)

is exact. Here Kl−1 := Ker σl . Morever, if M is compact, then

L ∈ Kl−1 =⇒ L : H k (E) −→ H k−(l−1) (F ) is bounded. (2.66)

Sketch of proof. We first prove (2.66). Thus, assume L ∈ Kl−1 , that is σl (L) = 0. Then locally
we have L = Lp and
σl (p) = 0 =⇒ p ∈ S̃0l−1

Draft 37 May 24, 2024

 Global analysis

by (2.52). Then (2.66) follows from Theorem 2.57.

We also need to show that σl : PDiff l (E; F ) → Smbl (E; F ) is surjective. Thus given
σ ∈ Smbl (E; F ), pick a chart (U, (x1 , . . . , xn )) such that both E and F admit trivializations
over U . Hence, given these choices σ can be written as a matrix p = (pij (x, ξ)) so that the local
pseudodifferential operator Lp can be defined. Then the global map L : Γ (E) → Γ (F ) can be
obtained with the help of a partition of unity. The reader may find the details in [Wel08, Thm.
IV.3.16]. □
Just like in the case of linear differential operators we have the following result.

Theorem 2.67. Let E, F, and G be Hermitian vector bundles over a compact manifold M .
Then the following holds:

(a) If L ∈ PDiff l (E; F ) and S ∈ PDiff s (F ; G), then S ◦L ∈ PDiff l+s (E; G) and σl+s (S ◦ L) =
σs (S) ◦ σl (L).

(b) If L ∈ PDiff (E; F ), then the formal adjoint L∗ exists and belongs to PDiff l (F ; E).
Moreover, σl (L∗ ) = σl (L)∗ .

The proof, which is omitted here, can be found in [Wel08, Thm. IV.3.17].

2.4.1 Elliptic operators and their parametrices

Definition 2.68. An operator L ∈ PDiff l (E; F ) is said to be elliptic, if for all (m, ξ) ∈ T ∗ M
such that ξ ̸= 0 the symbol σ (L) (m, ξ) : Em → Fm is an isomorphism.

Example 2.69. Consider the Laplacian ∆ acting on functions. By Example 2.41, σ (∆) (m, ξ) =
|ξ|2 . Hence, ∆ is elliptic.
More generally, consider the Laplacian ∆ = dd∗ +d∗ d acting on p-forms. By Example 2.42,
we know that σ (d) (ξ) α = i ξ ∧ α. For ξ ∈ T ∗ M denote by ξ # ∈ Tm M the metric dual of ξ,
that is ξ # is defined by requiring that the equality

gm ξ # , v = ξ (v)

holds for any v ∈ Tm M . A computation yields that the adjoint of σ (d) is given by

σ (d)∗ (ξ) β = −i ιξ# β ,

where ıξ# : Λp+1 Tm∗ M → Λp Tm∗ M is the contraction:

ıξ# ω = ω ξ # , · , . . . , · .

This yields in turn:

σ (∆) ω = (σ (d) ◦ σ (d∗ ) + σ (d∗ ) ◦ σ (d)) ω = ξ ∧ ιξ# ω + ιξ# (ξ ∧ ω)

= ξ ∧ ιξ# ω + |ξ|2 ω − ξ ∧ iξ# ω = |ξ|2 ω.

Thus, σ (∆) is still given by the multiplication with |ξ|2 and, hence, ∆ is elliptic.

This example largely explains our interest to elliptic operators. It turns out that this class
contains many other interesting operators and has particularly good properties, which we consider
next.

Draft 38 May 24, 2024

 Global analysis

Definition 2.70.
(i) If L : C ∞ (E) → C ∞ (F ) is a linear map, then we say that L ∈ OPl (E; F ) if the closure
of L yields a bounded map H k (E) → H k−l (F ) for each k ∈ R.
(ii) For L : C ∞ (E) → C ∞ (F ), a linear map S : C ∞ (F ) → C ∞ (E) is called a parametrix,
if
L ◦ S − idF ∈ OP−1 (F ) and S ◦ L − idE ∈ OP−1 (E) .

For example, any pseudodifferential operator of order l on a compact manifold belongs to

OPl by Proposition 2.62.
A parametrix should be understood as an approximate inverse in some sense. It should not
be surprising that explicit examples are not easy to construct. However, one general way to
obtain a class of examples is as follows.
Suppose Ω ⊂ Rn is a bounded domain with smooth boundary ∂Ω. Consider the classical
Dirichlet boundary value problem
(
∆u = f in Ω,
(2.71)
u ∂Ω = 0,
P ∂2
where ∆ = − ∂x 2 is the standard Laplacian acting on functions. I will use the well-known
j
fact that the solution operator G, which assigns to f a unique solution u = uf of (2.71) belongs
to OP−2 (this is proved below in fact). Taking this as granted for now, consider a slightly more
general boundary value problem
(
(∆ + K) u = 0 in Ω,
u ∂Ω = 0,

where K ∈ Diff 1 . Then for L := ∆+K we have G ◦ L−id = G ◦ ∆+G ◦ K −id = G ◦ K ∈ OP−1
and also L ◦ G − id = K ◦ G ∈ OP−1 . Thus, G is a parametrix for ∆ + K.
One of the main theorems about elliptic operators is the following.

Theorem 2.72. For any elliptic L ∈ PDiff l (E; F ) there exists a parametrix S ∈ PDiff −l (F ; E).

Proof. For any elliptic L the symbol σ (L) is invertible at each (m, ξ) , ξ ̸= 0. Then by
Theorem 2.64 there exists some S ∈ PDiff −l such that σ−l (S) = σl (L)−1 . We have

σ0 (L ◦ S − idF ) = σl (L) ◦ σ−l (S) − σ0 (idF ) = 0,

so that L ◦ S − idF ∈ OP−1 .

A similar argument yields that S ◦ L − idE belongs to OP−1 too. □

2.4.2 Elliptic estimate

Let B1 and B2 be two Banach spaces. Recall that an operator K : B1 → B2 is said to be
compact, if the image of the unit ball in B1 is relatively compact in B2 . This means that for
any sequence uj in B1 such that ∥uj ∥B1 ≤ 1 there exists a subsequence ujk such that Kujk
converges in B2 .

Proposition 2.73. Let M be a closed manifold, L ∈ Diff l (E; F ) an elliptic operator, and k a
real number. Then the following holds:

Draft 39 May 24, 2024

 Global analysis

(i) HL := Ker L : H k → H k−l is finite-dimensional and consists of smooth sections only;
In particular, HL does not depend on k.
(ii) Denote V := HL⊥ ∩ H k+l (E), where ⊥ means the orthogonal complement in L2 (E).
Then for all k ≥ 0 there exists a constant Ck > 0 such that

∥Lv∥H k ≥ Ck ∥v∥H k+l (2.74)

holds for all v ∈ V provided k ≥ 0.

Proof. The proof consists of the following steps.

Step 1. K ∈ OP−1 (E; F ) =⇒ K : H k (E) → H k (F ) is compact provided M is compact.
Indeed, K : H k (E) → H k+1 (F ) is bounded and the embedding j : H k+1 (F ) → H k (F )
is compact, since M is compact. Then K : H k (E) → H k (F ) is compact as the composition
of bounded and compact linear maps.
Step 2. We prove (i).
For the proof we need the following standard fact of functional analysis: The unit sphere
in a Banach space B is compact if and only if B is finite-dimensional. With this understood,
we obtain that for a compact operator R : B → B, where B is a Banach space, any eigenspace
Bλ := {u | Ru = λu} is finite dimensional provided λ ̸= 0.
Assume Lu = 0. Applying a parametrix of L to both sides of this equation, we obtain
u + Ru = 0, where R = S ◦ L − id : H k → H k is compact. Hence, the space

HLk := u ∈ H k | Lu = 0 = u ∈ H k | Ru = −u
 

is finite-dimensional. Moreover, since R ∈ OP−1 ,

u ∈ Hk =⇒ u ∈ H k+1 =⇒ u ∈ H k+2 =⇒ . . .

so that u ∈ ∩H s = C ∞ . In particular, each u ∈ HLk is smooth so that HLk in fact does not
depend on k.
Step 3. We prove (2.74).
Assume (2.74) fails, that is there exists a sequence vj ∈ V such that

∥vj ∥H k+l = 1 and Lvj −→ 0 in H k .

Applying the parametrix S we obtain that vj + Rvj → 0 in H k+l . Since R is compact, vj has a
subsequence, which converges to some v in H k+l . Then we must have

∥v∥H k+l = 1, Lv = 0, and v ∈ (Ker L)⊥L2 .

This contradiction finishes the proof. □

Corollary 2.75 (Elliptic estimate). Let M be a closed manifold and L ∈ Diff l (E; F ) be an
elliptic operator. Then for any k ≥ 0 there exists a constant Ck > 0 such that the estimate

∥u∥H k+l ≤ Ck (∥Lu∥H k + ∥u∥L2 )

holds for any u ∈ H k+l .

Draft 40 May 24, 2024

 Global analysis

Proof. Since HL ⊂ H k+l , for any u ∈ H k+l we can write u = v + w, where v ∈ HL

and (v, w)L2 = 0. Notice that since v is smooth, w = u − v ∈ H k+l , that is we have the
decomposition
⊥
H k+l = HL ⊕ HL L2 ∩ H k+l .
Since HL is finite dimensional, the restriction of ∥ · ∥H k+l to HL is equivalent to the restriction
of ∥ · ∥L2 . Combining this with (2.74) we obtain

∥u∥H k+l ≤ ∥v∥H k+l + ∥w∥H k+l ≤ Ck′ ∥v∥L2 + Ck′′ ∥Lw∥H k
= Ck′ ∥v∥L2 + Ck′′ ∥Lu∥H k ≤ Ck (∥Lu∥H k + ∥u∥L2 ) .

This finishes the proof. □

Theorem 2.76. Let M be a closed manifold and L ∈ Diff l (E; F ) be an elliptic operator. Then
for any k ∈ R the operator L : H k+l (E) → H k (F ) is Fredholm, i.e.,
1. dim Ker L < ∞;
2. The image of L is a closed subspace of H k (F );
3. The cokernel Coker L := H k (F ) / Im L is finite dimensional.

Proof. We shall give the proof only in the case k ≥ 0.

We know already from Proposition 2.73, (i) that Ker L is finite dimensional. The rest of the
proof consists of the following steps.
Step 1. Im L is closed.
Clearly Im L = L (V ), where V = HL⊥ ∩ H k+l (E). Assume w lies in the closure of L (V ),
that is there exist a sequence vj ∈ V such that Lvj → w. Then by (2.74) we have

∥vi − vj ∥H k+l ≤ C∥Lvi − Lvj ∥H k ≤ C (∥Lvi − w∥H k + ∥Lvj − w∥H k ) .

This yields that (vj ) is a Cauchy sequence, hence converges to some v ∈ V . Then we must have

Lv = lim Lvj = w.
j→∞

Step 2. Let L∗ be the formal adjoint of L. Then w ∈ Im (L)⊥L2 ∩ H k (F ) if and only if

w ∈ Ker L∗ .
If w ∈ Im (L)⊥L2 ∩ H k (F ), then for any u ∈ H k+l (E) we have

0 = (Lu, w)L2 = (u, L∗ w)L2 =⇒ L∗ w = 0.

Conversely, if L∗ w = 0, then (Lu, w)L2 = 0 for any u ∈ H k+l (E).

Step 3. We prove this theorem.
It only remains to show that dim Coker L < ∞. To see this, identify Coker L with (Im L)⊥L2 ∩
H k (F ) which equals Ker L∗ by the preceding step. It remains to notice that L∗ is elliptic too,
so that dim Ker L∗ < ∞ by Proposition 2.73, (i). □

The proof of the above theorem actually implies the following.

Draft 41 May 24, 2024

 Global analysis

Corollary 2.77. Let M be a compact manifold and L ∈ Diff l (E; F ) be an elliptic operator.
Given w ∈ H k (F ) the equation
Lu = w (2.78)
has a solution u ∈ H k+l (E) if and only if w ⊥ Ker L∗ . Moreover, if w ∈ C ∞ (M ; F ), then any
solution u of (2.78) is smooth.
Proof. We only need to show that w ∈ C ∞ ⇒ u ∈ C ∞ . However, this follows immediately
from ∩ H k (E) = C ∞ (E). □
Corollary 2.79 (Fredholm’s alternative). Let M be a compact manifold and L ∈ Diff l (E; F )
an elliptic operator. Then one and only one of the following statements holds:
(a) The inhomogeneous equation Lu = w has a solution for any w ∈ Γ (F ).
(b) The homogeneous equation L∗ v = 0 has a non-trivial solution. □
Corollary 2.77 implies in particular Theorem 1.38. To see this, it just suffices to recall that
the Laplacian is formally self-adjoint, that is ∆∗ = ∆, which is the content of Proposition 1.36.
In the case of the Laplacian, it is also possible to give a more detailed version of Corollary 2.77
as follows.
Theorem 2.80 (Hodge). Let
M be a closed oriented Riemannian manifold. Denote Hk :=
Ker ∆ : Ωk (M ) → Ωk (M ) . We have the following decomposition

Ωk (M ) = Hk ⊕ Im d ⊕ Im d∗ ,

where all three spaces are orthogonal with respect to the L2 -scalar product.
Proof. First we show that Hk is orthogonal to Im d in the L2 -sense. Indeed, if ω ∈ Hk and
η ∈ Ωk−1 (M ), then
⟨ω, dη⟩L2 = ⟨d∗ ω, η⟩L2 = 0,
since d∗ ω = 0 by Proposition 1.29.
The claims that Hk ⊥ Im d∗ and Im d ⊥ Im d∗ can be obtained by using similar arguments.
Furthermore, to prove the rest of the claim it suffices to show that Im ∆ = Im (d) ⊕ Im d∗ .
Clearly, we have Im ∆ ⊂ Im d ⊕ Im d∗ by the definition of ∆. The converse inclusion follows
from Corollary 2.77 and the self-adjointness of ∆. □

2.5 Elliptic complexes

The de Rham complex can be generalized as follows. Let E0 , E1 , . . . , EN be Hermitian vector
bundles over a smooth manifold M . Assume we have the sequence
L
0 1 L LN −1
Γ (E0 ) −→ Γ (E1 ) −→ Γ (E2 ) −→ . . . −−−→ Γ (EN ) , (2.81)

where each Lj ∈ Diff l (Ej ; Ej+1 ) and l does not depend on j. Assume also that (2.81) is a
complex, i.e., Lj+1 ◦ Lj = 0.
Definition 2.82. (2.81) is said to be an elliptic complex, if the associated sequence of symbols
σl (L0 ) σl (L1 ) σl (LN −1 )
0 −→ π ∗ E0 −−−−→ π ∗ E1 −−−−→ ... −−−−−−→ π ∗ EN −→ 0

is exact away from the zero section of T ∗ M .

Draft 42 May 24, 2024

 Global analysis

Just like in the case of the de Rham complex, we can define the corresponding cohomology
groups by
Ker (Lj : Γ (Ej ) −→ Γ (Ej+1 ))
Hj (E) := ,
Im (Lj−1 : Γ (Ej−1 ) −→ Γ (Ej ))
where H0 (E) = Ker L0 and HN (E) = Γ (EN ) / Im LN−1 .
Example 2.83. Any elliptic operator L : Γ (E) → Γ (F ) is a (very short) elliptic complex.
Example 2.84. The de Rham complex of any smooth manifold.
New examples of elliptic complexes will be given below.
Associated to (2.81), is the (generalized) Laplacian

∆j := L∗j Lj + Lj−1 L∗j−1 : Γ (Ej ) −→ Γ (Ej ) ,

which is clearly self-adjoint.

Exercise 2.85. Show that ∆j is an elliptic operator provided (2.81) is an elliptic complex.
Any element of the space

H (Ej ) := Ker ∆j ⊂ Γ (Ej )

is called harmonic.
Theorem 2.86. If M is closed connected oriented Riemannian manifold, then the following
holds:
(i) s ∈ H (Ej ) ⇐⇒ Lj s = 0 and L∗j−1 s = 0.
(ii) There is an orthogonal decomposition

Γ (Ej ) = H (Ej ) ⊕ Im Lj−1 ⊕ Im L∗j .

(iii) dim H (Ej ) < ∞ and there is a canonical isomorphism

H (Ej ) −→ Hj (E) , s 7−→ [s] .

The proof of this theorem can be obtained by a straightforward modification of the proofs
of the corressponding statements for the de Rham complex. I leave the details to the reader.

2.5.1 The Dolbeault complex

In this section M denotes a complex manifold of complex dimension n. This means that M is
a real manifold (of dimension 2n), each point admits a chart of the form ψ : U → Cn and each
change-of-coordinates map ψ ◦ ϕ−1 : Cn → Cn is holomorphic.
Just like in the case of Riemann surfaces, writing ψ = (z1 , . . . , zn ) we obtain real coordinates
(x1 , y1 , . . . , xn , yn ), where

xj = Re zj and yj = Im zj .

Then we can define the complex structure I : T M −→ T M by

∂ ∂ ∂ ∂
I = and I =− .
∂xj ∂yj ∂yj ∂xj

Draft 43 May 24, 2024

 Global analysis

Using the fact that the change-of-coordinates maps are holomorphic, it can be shown that I is
well-defined. Since I 2 = −id by the very definition, we obtain that each tangent space Tm M is
equipped with the structure of a complex vector space via
(a + bi) · v = a v + b Iv, v ∈ Tm M.
Then the complexification of the cotangent space splits:
Tm∗ M ⊗ C = (Tm∗ M )1,0 ⊕ (Tm∗ M )0,1 ,
(Tm∗ M )1,0 := {ψ : Tm M −→ C | ψ (Iv) = iψ (v)} ,
(Tm∗ M )0,1 := {ψ : Tm M −→ C | ψ (Iv) = −iψ (v)} .

In other words, (Tm∗ M )1,0 consists of C-linear functionals, whereas (Tm∗ M )0,1 consists of C-
antilinear ones.
Since (Tm∗ M )1,0 and (Tm∗ M )0,1 are isomorphic via ψ 7→ ψ̄, we obtain
1
dimC (Tm∗ M )1,0 = dimC (Tm∗ M )0,1 = dimC (Tm∗ M ⊗ C) = n.
2
Remark 2.87. By the way of reminder, for any real vector space V its complexification V ⊗ C is
a complex vector space with the obvious complex structure, namely v ⊗ z 7→ v ⊗ iz. Suppose V
is itself a complex vector space regarded as a real vector space equipped with an endomorphism
I such that I 2 = −id. Then its complexification V ⊗R C inherits a complex structure from
V , namely I ⊗ id. The default complex structure on V ⊗ C is v ⊗ z 7→ v ⊗ iz. Then on
V 1,0 = {v ⊗ z − Iv ⊗ iz | v ∈ V, z ∈ C} these two complex structures coincide, while
on V 0,1 = {v ⊗ z − Iv ⊗ iz | v ∈ V, z ∈ C} they have opposite signs. The conjugation
v⊗z 7→ v⊗z̄ is a complex anti-linear isomorphism with respect to the default complex structure
but a complex linear isomorphism between (V 1,0 , I ⊗ id) and (V 0,1 , I ⊗ id). In particular,
dimC V 1,0 = dimC V 0,1 = dimR V , although V 1,0 and V 0,1 are not naturally isomorphic.
If (z1 , . . . , zn ) are local holomorphic coordinates in a neighbourhood of m, denote
dzj = dxj + i dyj ∈ (Tm∗ M )1,0 and dz̄j = dxj − i dyj ∈ (Tm∗ M )0,1 ,
where j = 1, . . . , n. It is easy to see that dz := (dz1 , . . . dzn ) consists of linearly independent
1-forms. Hence, dz is a basis of (Tm∗ M )1,0 at each point m where the coordinates (z1 , . . . , zn )
are defined. Likely, dz̄ = (dz̄1 , . . . , dz̄n ) is a basis of (Tm∗ M )0,1 .
Furthermore, using the isomorphism
Λk (U ⊕ V ) = ⊕ Λp U ⊗ Λq V
p+q=k
p,q≥0

we obtain the decomposition

Λk (T ∗ M ⊗ C) = ⊕ Λp (T ∗ M )1,0 ⊗ Λq (T ∗ M )0,1 =: ⊕ Λp,q .
p+q=k p+q=k

Denote
Ωp,q (M ) := Γ (Λp,q ) .
Thus, ω ∈ Ωp,q (M ) if and only if ω is a smooth complex-valued differential form of degree
p + q and locally ω can be written in the following form
X
ω= ωi1 ...ip ;j1 ...jq dzi1 ∧ . . . ∧ dzip ∧ dz̄j1 ∧ . . . ∧ dz̄jq .
i1 <...<ip
j1 <...<jq

Draft 44 May 24, 2024

 Global analysis

For any ω ∈ Ω0,0 (M ) ∼ ¯ where

= C ∞ (M ; C) we have dω = ∂ω + ∂ω,
n n
X ∂ω ¯ =
X ∂ω
∂ω = dzj ∈ Ω1,0 (M ) and ∂ω dz̄j ∈ Ω0,1 (M ) .
j=1
∂z j j=1
∂ z̄j

More generally, if ω ∈ Ωp,q (M ), then we still have dω = ∂ω + ∂ω, ¯ where

X
∂ω = ∂ωi1 ...ip ;j1 ...jq ∧ dzi1 ∧ . . . ∧ dzip ∧ dz̄j1 ∧ . . . ∧ dz̄jq ,
i1 <...<ip
j1 <...<jq
X
¯ =
∂ω ¯ i ...ip ;j ...jq ∧ dzi ∧ . . . ∧ dzip ∧ dz̄j ∧ . . . ∧ dz̄jq .
∂ω 1 1 1 1
i1 <...<ip
j1 <...<jq

In particular, ∂ : Ωp,q → Ωp+1,q and ∂¯ : Ωp,q → Ωp,q+1 .

¯ we have
Notice that since d = ∂ + ∂,
d2 = 0 ⇐⇒ ∂ 2 = 0, ∂¯2 = 0, ∂ ∂¯ = −∂∂.
¯
Therefore, for any fixed p such that 0 ≤ p ≤ n we obtain the following complex
∂¯ ∂¯ ∂¯ ∂¯
0 −→ Ωp,0 (M ) −→ Ωp,1 (M ) −→ Ωp,2 (M ) −→ . . . −→ Ωp,n (M ) −→ 0,
which is called the Dolbeault complex. The corresponding cohomology groups
¯ : Ωp,q −→ Ωp,q+1


Ker ∂
H p,q (M ) :=
Im ∂¯ : Ωp,q−1 −→ Ωp,q

are called the Dolbeault cohomolgy groups of M . Notice, however, that unlike the de Rham
cohomology groups, the Dolbeault cohomology groups depend on the complex structure of M .
In particular, they are not topological invariants of M .
Proposition 2.88. For any complex manifold M the Dolbeault complex is elliptic.
¯
Proof. Just like in Example 2.42, for the symbol of the ∂-operator we have
σ ∂¯ (ξ) ω = i ξ ∧ ω,
0,1

where ξ 0,1 = 12 (ξ + iξ (I·)) is the (0, 1)-part of ξ and ω ∈ Λp,q . We have to show the following
ξ 0,1 ∧ ω = 0 ⇐⇒ ω = ξ 0,1 ∧ η
for some η ∈ Λp,q−1 , where ξ ̸= 0 ⇔ ξ 0,1 ̸= 0.
Assume ξ is based at a point m. Since ξ 0,1 ̸= 0, we can find a complex basis (ψ1 , . . . , ψn )
of (Tm∗ M )1,0 such that ψ1 = ξ 0,1 . Then ω ∈ Λp,q m can be uniquely written in the form
X
ω= ωi1 <...<ip ;j1 <...<jq ψi1 ∧ . . . ∧ ψip ∧ ψ̄j1 ∧ . . . ∧ ψ̄jq .
i1 <...<ip
j1 <...<jq

Then ξ 0,1 ∧ ω = 0 ⇔ ψ̄1 ∧ ω = 0 yields ωi1 ,...ip ;j1 ,...jq = 0 provided j1 ̸= 1. Hence,
X
ω= ωi1 ...ip ;1 j2 ...jq ψi1 ∧ . . . ∧ ψip ∧ ψ̄1 ∧ ψ̄j2 ∧ . . . ∧ ψ̄jq
i1 <...<ip
j1 =1<j2 <...<jq
X
= (−1)p ψ̄1 ∧ ωi1 ...ip ;1 j2 ...jq ψi1 ∧ . . . ∧ ψip ∧ ψ̄j2 ∧ . . . ∧ ψ̄jq
i1 <...<ip
j2 <...<jq

= ξ 0,1 ∧ η.
This finishes the proof of this proposition. □

Draft 45 May 24, 2024

 Global analysis

Corollary 2.89. If M is a compact complex manifold, then hp,q (M ) := dimC H p,q (M ) < ∞.

□
p,q
The integers h (M ) are called the Hodge numbers of M .
Let ∂¯∗ be the formal adjoint of ∂.
¯ The Laplacian corresponding to the Dolbeault complex is
then
□ := ∂¯∂¯∗ + ∂¯∗ ∂,
¯

which is called the Hodge-Laplacian. By Theorem 2.86 we have an isomorphism

H p,q (M ) ∼
= Ker (□ : Ωp,q (M ) −→ Ωp,q (M )) .

Definition 2.90. Let g be a Riemannian metric on a complex manifold M such that g (Iv, Iw) =
g (v, w) for all v, w (such metrics always exist). Define a 2-form ω on M by

ω (v, w) := g (Iv, w) .

Then g is said to be Kähler, if ω is closed.

Kähler metrics should be understood as those metrics, which fit with the complex structure
particularly well (this may not be obvious from the definition and I shall not try to justify this
claim here).

Lemma 2.91. Any complex submanifold of a Kähler manifold is itself Kähler.

Proof. Let ι : N → M be a complex submanifold. Then for the inherited Riemannian metric
on N , we have ωN = ι∗ ωM . Hence, ωN is closed. □

While we shall not use the above lemma directly, it immediately yields a vast number of
examples of Kähler manifolds. Indeed, it is shown in the following section that the induced
metric on CP n = S 2n+1 /U(1) is Kähler (this is known as the Fubini–Study metric). Then the
common zero locus of a collection of homogenious polynomials with complex coefficients is a
Kähler manifold provided it is a smooth manifold.

Theorem 2.92. If g is Kähler, then ∆ maps Ωp,q (M ) into Ωp,q (M ) and ∆ = 2 □.

In particular, if M is compact and Kähler, for any k ≥ 0 we have

Hk (M ) ∼
= ⊕ H p,q (M ) and H p,q (M ) ∼
= H q,p (M ) (2.93)
p+q=k

for all p and q. □

Corollary 2.94. If M is a compact Kähler manifold, then all odd Betti numbers of M are even.

Proof. By (2.93) we have

b2k+1 (M ) = h0,2k+1 + h1,2k + . . . + h2k,1 + h2k+1,0

= 2 h0,2k+1 + h1,2k + . . . + hk,k+1 ,

where the last equality follows from hp,q (M ) = hq,p (M ). □

Corollary 2.95. There exist compact complex manifolds, which can not be embedded into any
complex projective manifold.

Draft 46 May 24, 2024

 Global analysis

Sketch of the proof. Consider the following action of Z on C2 \ {0} :

a · (z1 , z2 ) = e−a z1 , e−a z2 .


(2.96)
Think of C2 \ {0} as R × S 3 via the diffeomorphism
R × S 3 −→ C2 \ {0}, (t, σ) 7−→ et · σ
Then the above action becomes a · (t, σ) = (t − a, σ) . Hence, the quotient M = C2 \ {0}/Z is
diffeomorphic to S 1 × S 3 . In particular, M is compact and b1 (M ) = b1 (S 1 × S 3 ) = 1.
Furthermore, notice that M is a complex manifold, since action (2.93) preserves the standard
complex structure of C2 \ {0}. However, M cannot be embedded as a complex submanifold
into CP N , since otherwise M would be Kähler, which would contradict Corollary 2.94. □
Remark 2.97. It should be noticed, that any holomorphic function on a compact complex
manifold is constant. This fact follows easily from the maximum modulus principle. Hence, a
compact complex manifold can never be embedded (as a complex submanifold) into CN . Thus,
CP N can be seen as the "best" replacement for CN for complex manifolds. Then Corollary 2.95
claims essentially that the Whitney embedding theorem is false in the context of complex
manifolds.
Also, Corollary 2.95 should be contrasted with the fact that any compact Riemann surface
can be embedded into some complex projective space.

2.6 Harmonic forms on complex projective spaces

Think of Cn+1 ∼= R2n+2 as a flat Kähler manifold. If (z0 , . . . , zn ) are standard holomorphic
coordinates on Cn+1 , the Kähler structure is given by
n
X n
X
g♭ = dx2j + dyj2 , ω♭ = dxj ∧ dyj ,
j=0 j=0

where zj = xj + yj i.
To endow CPn with a Riemannian metric, think of CPn as a quotient of the sphere, namely
CP = S 2n+1 /U(1). Here we think of S 2n+1 as a subspace of Cn+1 . For p ∈ S 2n+1 , the scalar
n

product on Tp S 2n+1 ∼= p⊥ is simply the restriction of the standard scalar product on Cn+1 .
Denoting by
K ∈ X(S 2n+1 ), K(p) = ip,
the infinitesimal action of U(1), we obtain that T[p] CPn can be identified with the orthogonal
complement of K(p) in Tp S 2n+1 , i.e.,
T[p] CPn ∼
= span{p, ip}⊥ ,
where the latter orthogonal complement is meant in Cn+1 . Since span{p, ip} is a complex
subspace of Cn+1 , T[p] CPn inherits a scalar product g, a complex structure I, and a symplectic
form ω. All these three tensors are independent of the choice of a representative in [p]. This
follows easily from the U(1)-invariance of the Kähler structure on Cn+1 .
The reader should check that the inherited complex structure coincides with the standard
complex structure of CPn . Moreover, I claim that the 2-form ω is closed and, thus, the name
“symplectic form” is justified. Indeed, if ι : S 2n+1 → Cn+1 denotes the embedding and π : S 2n+1 →
CPn the projection, ω is uniquely characterized by the following property: π ∗ ω = ι∗ ω♭ . Hence,
∗ ∗ ∗ ∗



π dω = d π ω = d ι ω♭ = ι dω♭ = 0. This yields that dω vanishes itself as claimed above.
Thus, the Fubini–Study metric on CPn is Kähler.

Draft 47 May 24, 2024

 Global analysis

1 n
Exercise 2.98. Show that for any Kähler manifold M 2n we have n!
ω = volM . Deduce from
here that ω k ∈ Ωk,k (M ) ⊂ Ω2k (M ) is harmonic.
Proposition 2.99. There are no non-trivial harmonic forms of odd degrees on CPn . Any non-
trivial harmonic 2k-form on CPn is of the form Cω k , where C ̸= 0 is a constant. In particular,
the Hodge numbers of CPn are given by
(
1 if p = q ≤ n,
hp,q (CPn ) =
0 otherwise.

Proof. It is well-known that the Betti numbers of CPn are given by

(
1 if k is even and k ≤ 2n,
bk (CPn ) =
0 otherwise.

Combining this with Theorem 2.92, we obtain that ω k is a unique (up to multiplication by
constants) harmonic 2k-form on CPn , where ω is the Fubini-Study form. □

2.7 The elliptic calculus

In this section I describe the main ideas discussed in this chapter from a more abstract point of
view. This section is not essential for understanding the subsequent material; it is meant more
to demonstrate logical connections between various notions that have appeared in this chapter.
The reader may safely skip this section by the first reading and only return once the material
preceding this section is well digested.
If M is a closed oriented Riemannian manifold and E → M is a vector bundle, we have
seen that Sobolev spaces form a sequence

H 0 (E) = L2 (E) ⊃ H 1 (E) ⊃ H 2 (E) ⊃ . . .

In fact, H k (E) is well-defined for all k ∈ Z (or even k ∈ R, however, this will be of minor
importance below). These have the following properties:
(1) H k (E), ∥ · ∥k is a Banach (Hilbert) space for each k ∈ Z;

(2) For each k ≥ l we have an inclusion H k (E) ⊂ H l (E), which is a bounded map, i.e.,

u ∈ Hk =⇒ u ∈ Hl and ∥u∥l ≤ Ck,l ∥u∥k .

Moreover, if k > l, then the inclusion H k (E) ,→ H l (E) is compact.

H (E) = C ∞ (E) is dense in each H k (E).
T k
(3)
k∈Z


One says that H k (E), ∥ · ∥k is a scale of Banach spaces (or a Banach scale).
Leaving Banach scales aside for a moment, recall that Smbl (E; F ) denotes the space of
symbols of degree l, where F is a vector bundle over M , see (2.39). To simplify the exposition
somewhat, it will be convenient to assume that F = E, albeit this is certainly non-essential as
the reader will be able to see on his own. Thus,
M
Smb = Smb(E) := Smbl (E)
l∈N0

Draft 48 May 24, 2024

 Global analysis

is a graded algebra with respect to the obvious multiplication.

Likewise, we have the filtered algebra of differential operators
[
Diff = Diff(E) := Diff l (E).
l∈N0

The qualifier ‘filtered’ means that we have an increasing sequence Diff 0 = C ∞ End(E) ⊂

Diff 1 ⊂ Diff 2 ⊂ . . . and if L1 ∈ Diff l1 and L2 ∈ Diff l2 , then L1 ◦ L2 ∈ Diff l1 +l2 . These come
equipped with “the principal symbol” homomorphism


σ : Diff → Smb such that σ Diff l ⊂ Smbl
for all l ∈ N0 . In other words, σ has degree zero as a homomorphism of filtered algebras.
Moreover, the sequence
σ
0 → Diff l−1 → Diff l −−−−→ Smbl (2.100)
is exact for any l ∈ N0 and each L ∈ Diff l yields a bounded map L : H k → H k−l , i.e., L has
degree −l viewed as a map of the Banach scale.
With this understood, the main objective of this chapter is a construction of enhancements
of algebras Diff and Smb. In the case of Smb, there are a few possibilities to proceed, which
correspond to the choices of various classes of functions such as S l , S̃ l or S0l . Ignoring non-
essential details, the outcome is a filtered algebra
[
S= Sl , where . . . ⊂ Sl−1 ⊂ Sl ⊂ Sl+1 ⊂ . . .
l∈Z

Denoting S[l] := Sl /Sl−1 , we have the natural projection πl : Sl → S[l] . Also, one can construct
a filtered algebra [
Ψ= Ψl
l∈Z
with the following properties:
(a) Diff l ⊂ Ψl provided l ≥ 0;
(b) There is a degree zero homomorphism σl : Ψl → S[l] . Moreover, for l ≥ 0 we have
Smbl ⊂ S[l] and the restriction of σl to Diff l yields the principal symbol of a differential
operator;
(c) For any k ∈ Z there is a homomorphism Ψl → B H k ; H k−l , where B denotes the space

of linear bounded

maps. Moreover, this extends the natural homomorphism Diff l →
k k−l
B H ;H and yields a homomorphism of filtered algebras: Ψ → B(H);
(d) The sequence
σ
0 → Ψl−1 −→ Ψl −−−−→ S[l] → 0 (2.101)
is exact for each l ∈ Z (cf. (2.100)).
Notice that at this point I deviate slightly from the exposition of Section 2.4, namely the
kernel of σ in that setting is Kl−1 rather than Ψl−1 . This is a minor point which allows one to
prove the main result omitting certain technical details. By choosing an appropriate class of
symbols, one can construct an algebra of pseudo-differential operators Ψ so that (2.101) holds.
Below I show that the main results about elliptic operators (elliptic estimates and Fredholm-
ness) are formal consequences of Properties (a)–(d). We begin with the following simple
observation.

Draft 49 May 24, 2024

 Global analysis

Proposition 2.102. Any R ∈ Ψ−1 viewed as a map H k → H k+1 ,→ H k is compact.

The proof of this proposition follows immediately from the observation that the composition
of a bounded and compact operators is a compact operator. Thus, the above proposition is a
simple observation indeed; I stated it formally for convenience only.

Definition 2.103. An operator L ∈ Ψl is said to be elliptic, if its principal symbol is invertible

in S.

Clearly, if σ(L) ∈ Sl and σ(L)−1 exists, then it must be of degree −l, because 1 = σ(id) =
σ(L ◦ L−1 ) = σ(L)σ(L)−1 and 1 ∈ S0 .

Proposition 2.104. Any elliptic operator L admits a parametrix, i.e., there exists an operator
P ∈ Ψ such that
L ◦ P = id + R1 and P ◦ L = id + R2 ,
where R1 , R2 ∈ Ψ−1 .

Proof. Since L is elliptic, there exists σ(L)−1 and σ(L)−1 ∈ S−l . By the exactness of (2.101),
we can find some P ∈ Ψ−l with σ(P ) = σ(L)−1 . Hence, σ(P L−id) = σ(P )σ(L)−σ(id) = 0.
Using the exactness of (2.101) again, we conclude that R2 := P L − id belongs to Ψ−1 . The
proof that R1 := LP − id ∈ Ψ−1 requires cosmetic changes only. □

The following statements follow rather easily from the existence of a parametrix. Notice,
that I do not try to prove the most general statements here, rather just to illustrate the ideas.

Corollary 2.105 (Elliptic estimate). Let L ∈ Ψl be elliptic. There exists a positive constant C
such that for any u ∈ H k we have


∥u∥k ≤ C ∥Lu∥k−l + ∥u∥k−1 . (2.106)

Proof. Indeed, since u = P Lu − R2 u, using (c) we immediately obtain (2.106). □

Corollary 2.107 (Fredholmness). Any elliptic operator L ∈ Ψl viewed as a map H k → H k−l

is Fredholm. Moreover, if u ∈ Ker L and u ∈ H k , then u ∈ C ∞ .

Proof. Pick k ∈ Z an assume that un is any sequence in Ker L : H k → H k−l such that

∥un ∥k = 1. If P is a parametrix, we have un = −R2 un for all n. Using the compactness of

R2 , we obtain that there is a subsequence u
nj , which converges in H k . In other words, the unit
sphere in the space Ker L : H k → H k−l equipped with the norm ∥ · ∥k
is compact. Thus,
by a well-known result from functional analysis, dim Ker L : H k → H k−l < ∞. Moreover,
utilizing the equality u = −R2 u, we obtain

u ∈ Hk =⇒ u ∈ H k+1 =⇒ u ∈ H k+2 . . . =⇒ u ∈ C ∞.

In particular, Ker L does not depend on the choice of k.

It remains to prove that Im L is closed and dim Coker L < ∞. To this end, we have the
following.
Exercise 2.108. If H is a Hilbert space and R : H → H is any compact operator, then id + R
is Fredholm. In particular, the image of id + R is closed.

Draft 50 May 24, 2024

 Global analysis

With this at hand, observe that

Im id + R1 = Im LP : H k−l → H k−l ⊂ Im L : H k → H k−l .

Let V be a complement of Im id + R1
in H k−l . Thus,
dim V < ∞ and there exists a subspace
V1 ⊂ V such that Im L : H k → H k−l = Im id + R1 ⊕ V1 . In particular, Im L is closed and
dim Coker L ≤ dim Coker(id + R1 ) < ∞. □
It is worthwhile to note that Corollary 2.107 is a very satisfactory statement about the
solvability of the equation Lu = v. Indeed, choosing a basis f1 , . . . , fN of the cokernel of
L, Corollary 2.107 simply says that Lu = v is solvable, if and only if

f1 (v) = 0, ... fN (v) = 0.

R example, in the case of the Laplacian (acting on functions), we have N = 1 and f (v) =
For
M
v and Im ∆ is necessarily contained in Ker f so that the above corollary is clearly optimal.
In general, checking a finite number of linear conditions on v is typically not too hard.
The power of Corollary 2.107 is perhaps best seen by contrasting with the case where the
base manifold may be non-closed. In this case, the image of an elliptic operator may well fail
to be closed and even if one obtains a finite number of linear necessary conditions, there is still
no guarantee that the equation Lu = v is solvable. Even if it is solvable for some v0 , this may
fail to be the case for some v arbitrarily close to v0 . . .
Having said this though, for suitable classes of non-closed manifolds and/or elliptic operators
one can still prove the Fredholm property. However, this goes beyond the goals of these notes.

Draft 51 May 24, 2024

Chapter 3

The Atiyah–Singer index theorem

3.1 Fredholm maps and their indices

Let H1 and H2 be separable Hilbert spaces. Recall that a linear bounded map T : H1 → H2 is
said to be Fredholm, if dim Ker T < ∞, Im T ⊂ H2 is a closed subspace, and dim Coker T =
dim H2 / Im T < ∞. Denote by F (H1 ; H2 ) the set of all Fredholm maps. Also, let B (H1 ; H2 )
denote the space of all linear bounded maps H1 → H2 equipped with the operator norm

∥T h∥H2
∥T ∥ := sup .
h∈H1 \{0} ∥h∥H1

Theorem 3.1. F (H1 ; H2 ) is an open subset of B (H1 ; H2 ). The map

ind : F (H1 ; H2 ) −→ Z, T 7−→ ind T := dim Ker T − dim Coker T

is constant on each connected component of F (H1 ; H2 ).

Proof. The proof consists of the following steps.

Step 1. F (H1 ; H2 ) is open in B (H1 ; H2 ).
Pick T0 ∈ F (H1 ; H2 ) and denote V := (Ker T0 )⊥ ⊂ H1 and W = (Im T0 )⊥ ∼
= Coker T0 .
For any T ∈ B (H1 ; H2 ) consider the map

T̃ : V ⊕ W −→ H2 , T̃ (v, w) = T v + w.

Since T̃0 is an isomorphism, T̃ is also an isomorphism provided T lies in a sufficiently small

neighbourhood U ⊂ B (H1 , H2 ) of T0 . In particular, for T ∈ U we have

(Ker T ) ∩ V = {0} =⇒ dim Ker T < ∞.

Moreover, T (V ) = T̃ (V ) ⊂ H2 is a closed subspace of finite codimension. Hence, Im T is

also closed and of finite codimension. Thus, T is Fredholm.
Step 2. ind is constant on U , where U is as in the preceding step.
Keeping the notations of the preceding step, we have the (non-orthogonal) decomposition

H1 = Ker T ⊕ Z ⊕ V,

52
 Global analysis

where Z := (Ker T ⊕ V )⊥ . Then T : Z ⊕ V → T (Z) ⊕ T (V ) = Im T is an isomorphism.

Moreover,

Ker T0 ∼
= H1 /V ∼
= Ker T ⊕ Z and Coker T0 ∼
= H2 / Im T0 = H2 /T0 (V ) . (3.2)

Furthermore, since for any T ∈ U the homomorphism T̃ is in fact an isomorphism and T (V ) =

T̃ (V ), we obtain
H2 /T (V ) = H2 /T̃ (V ) ∼
= T̃ (W ) ∼
= W.
The left hand side can be interpreted as a complement of T (V ) in H2 . This equals the complement
of T (V ) in Im T plus the complement of Im T in H2 . Hence,

Coker T0 = W ∼
= H2 /T (V ) = T (Z) ⊕ Coker T ∼
= Z ⊕ Coker T.

Combining this with (3.2), we obtain ind T = ind T0 . □

Remark 3.3. One can show that ind T1 equals ind T2 if and only if T1 and T2 lie in the same
connected component of F (H1 ; H2 ).

Proposition 3.4. If T is Fredholm and K is compact, then T + K is also Fredholm and

ind (T + K) = ind (T ).

Proof. The proof consists of the following steps.

Step 1. A bounded map T : H1 → H2 is Fredholm if and only if there exist bounded maps
S1 , S2 : H2 → H1 such that

S1 ◦ T = idH1 + R1 and T ◦ S2 = idH2 + R2 ,

where both R1 and R2 are compact.

The proof of this step is left as an exercise (see, however, the proof of Theorem 2.76).
Step 2. We prove the statement of this proposition.
Let T be a Fredholm map. Keeping the notations of the preceding step, we have

S1 (T + K) = S1 T + S1 K = id + (R1 + S1 K) ,

where R1 + S1 K is compact. Also, a similar argument yields that (T + K) S2 − id is also

compact. Hence, T + K is Fredholm. Moreover, for each λ ∈ [0, 1] the map T + λK is
also Fredholm and therefore ind (T + λK) is a locally constant function of λ. Hence, in fact
ind (T + λK) is constant in λ so that ind (T + K) = ind T . □

Corollary 3.5. Let L ∈ Diff l (E; F ) be an elliptic differential operator. If M is closed, then

ind L : H k+l (E) −→ H k (F )

depends only on the principal symbol of L.

Proof. If σl (L) is the principal symbol of L, pick any L0 ∈ Diff l (E; F ) such that σl (L0 ) =
σl (L). Then K := L − L0 is of order at most l − 1. Hence, its closure (still denoted by the same
letter) K : H k+l (E) −→ H k (F ) is compact as a composition of a bounded map H k+l (E) →
H k+1 (F ) and a compact one H k+1 (F ) → H k (F ). Hence, ind L = ind (L0 + K) = ind L0 .
□

Draft 53 May 24, 2024

 Global analysis

Exercise 3.6. Find a mistake in the following "proof" of the claim that any two elliptic operators
of the same order have equal indices. Thus, assume that L0 and L1 are two such operators. Then
ind ((1 − t) L0 + tL1 ) does not depend on t implying that ind L0 = ind L1 .

Exercise 3.7. Show that the index of a formally self-adjoint differential operator vanishes.

The above corollary raises naturally the question how one can actually compute the index of
a differential operator just in terms of its principal symbol. This task has been accomplished by
Atiyah and Singer in 1960s. However, even to formulate an answer we need to make a detour.

3.2 Characteristic Classes

3.2.1 The curvature of a connection
Recall that given a vector bundle E → M equipped with a connection ∇ we have the sequence
d d d d
Ω0 (E) −→
∇
Ω1 (E) −→
∇
Ω2 (E) −→
∇ ∇
. . . −→ Ωn (E) −→ 0,

see (2.25) for details. As it was already pointed out above, this sequence is not a complex in
general, that is d2∇ := d∇ ◦ d∇ does not necessarily vanish. Nevertheless, d2∇ has the following
interesting property.

Proposition 3.8. d2∇ : Ωk (E) −→ Ωk+2 (E) is C ∞ (M )-linear.

Proof. Pick any s ∈ Γ (E) and f ∈ C ∞ (M ). We have

d∇ d∇ (f s) = d∇ (∇ (f s)) = d∇ (df ⊗ s + f ∇s)

= d (df ) ⊗ s − df ∧ ∇s + df ∧ ∇s + f d∇ (∇s)
= f d∇ (∇s) .

This proves this proposition for k = 0. The rest of the proof is left as an exercise to the
reader. □

Remark 3.9. In our setting it is convenient to work over the field of complex numbers, although
we could have considered R as a basic field equally well. Hence, we assume that ∇ is compatible
with the complex structure of E, that is ∇ (is) = i∇s. In particular, in the above proposition
C ∞ (M ) means C ∞ (M ; C). I shall not always be very picky concerning such details below.

Lemma 3.10. Let A : Γ (E) → Ωk (M ; F ) be a C-linear map, which is also C ∞ (M ; C)-linear,

that is
A (f · s) = f · A (s) ∀f ∈ C ∞ (M ; C) and s ∈ Γ (E) .
Then there exists a ∈ Ωk (Hom (E; F )) such that A (s) = a · s.

The proof of this proposition is also left as an exercise.

Corollary 3.11. For any connection ∇ there exists a 2-form F∇ ∈ Ω2 (End (E)) such that for
any ω ∈ Ωk (E) we have d∇ (d∇ ω) = F∇ ∧ ω. □

Definition 3.12. The 2-form F∇ as in the above corollary is called the curvature form of ∇.

Draft 54 May 24, 2024

 Global analysis

To establish a geometric meaning of the curvature form, pick local coordinates (x1 , . . . xn )
on M . We could call
∇j s = ı ∂ (∇s)
∂xj

the j th partial covariant derivative of s ∈ Γ (E). In particular, ∇s = nj=1 dxj ⊗ ∇j s. Hence,

P

n n n
!
X X X
d∇ (d∇ s) = − dxj ∧ ∇ (∇j s) = − dxj ∧ dxi ⊗ ∇i (∇j s)
j=1 j=1 i=1
X
= dxi ∧ dxj (∇i (∇j s) − ∇j (∇i s)) .
i<j
P
In other words, by writing F∇ = i<j Fij dxi ∧ dxj where Fij ∈ End (E), we have Fij s =
∇i (∇j s) − ∇j (∇i s). Thus, the curvature measures the extend to which partial covariant
derivatives fail to commute.

3.2.2 Local representation of the curvature form

Recall that locally a connection can be identified with a 1-form. This means the following:
Given an open subset U such that E admits a trivialization ψ over U , any section s over U can
be identified with a map σ : U → Ca , where a = rk E. Then

∇=d+A ⇐⇒ ∇s ≡ dσ + Aσ,

where A ∈ Ω1 (U ; Ma (C)). Hence,

d∇ (∇s) ≡ d (dσ + A · σ) + A ∧ (dσ + A · σ)

= 0 + dA · σ − A ∧ dσ + A ∧ dσ + A ∧ A · σ
= (dA + A ∧ A) · σ.

Hence, locally F∇ can be identified with the 2-form dA + A ∧ A, which takes values in Ma (C).
Somewhat informally, one simply writes

F∇ = dA + A ∧ A (3.13)

keeping in mind that the right hand side depends on the trivialization chosen.
Exercise 3.14. For any A ∈ Ω1 (M ; Ma (C)) ∼= Ma (Ω1 (M )) define [A, A] ∈ Ω2 (M ; Ma (C))
by
[A, A] (v, w) = A (v) A (w) − A (w) A (v) ∈ Ma (C) .
Show that A ∧ A = 21 [A, A], where the factor 1
2
stems from the following definition of the
wedge-product for 1-forms:
1 1
ω∧η = (ω ⊗ η − η ⊗ ω) ⇐⇒ ω ∧ η (v, w) = (ω (v) η (ω) − ω (w) η (v)) .
2 2
In particular,
1
F∇ = dA + [A, A].
2
1
Warning: Sometimes the wedge-product is defined without the factor 2
in the literature. This
may lead to confusions.

Draft 55 May 24, 2024

 Global analysis

Exercise 3.15. Show that for anny A, B ∈ Ω1 (M ; Ma (C)) we have

[A, B] = +[B, A]. (3.16)

Exercise 3.17. If e′ is another trivialization of E such that e = e′ · g just like in Section 2.1,
then
F∇′ = g −1 F∇ g, (3.18)
where F∇′ denotes the local representation of the curvature of ∇ with respect to e′ .
Proposition 3.19 (Bianchi identity). If A ∈ Ω1 (U ; Ma (C)) is a local representation of ∇ as
above, then dF∇ + [A, F∇ ] = 0.
Sketch of proof. We have
   
1 1 1 1 1
d dA + [A, A] + A, dA + [A, A] = [dA, A] − [A, dA] + [A, dA] + [A, [A, A]]
2 2 2 2 2
1 1 1
= − [A, dA] − [A, dA] + [A, dA] + [A, [A, A]]
2 2 2
1
= [A, [A, A]] ,
2
where the second equality follows from [A, B] = − [B, A] if A ∈ Ω1 (U ; Ma (C)) and B ∈
Ω2 (U ; Ma (C)), cf. (3.16). The proof is finished by the following observation, whose proof is
left to the reader: The Jacobi identity

[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0,

which is valid for any X, Y, Z ∈ Ma (C), implies that [A, [A, A]] = 0. □
Remark 3.20. More invariant interpretation of the Bianchi identity is as follows. By Lemma 2.28,
∇ induces a connection on E ∗ . Hence, using Lemma 2.28 again we obtain a connection, which
is still denoted by the same letter, on E ∗ ⊗E ∼
= End (E). Then we also obtain the corresponding
k k+1
map d∇ : Ω (M ; End (E)) −→ Ω (M ; End (E)) and the above proposition may be restated
simply as d∇ F∇ = 0.
Proposition 3.21. If ∇ is Hermitian, then F∇ takes values in skew-Hermitian endomorphisms.
In other words, for any v, w ∈ T M we have

F∇ (v, w)∗ = −F∇ (v, w) .

Proof. Let e = (e1 , . . . , ea ) be a local trivialization of E such that e is a pointwise orthonormal

basis of the corresponding fiber. If A is the local representation of a Hermitian connection ∇
with respect to such a trivialization, then for any σ1 , σ2 : U −→ Ca we have

d⟨σ1 , σ2 ⟩ = ⟨(d + A) σ1 , σ2 ⟩ + ⟨σ1 , (d + A) σ2 ⟩,

since ∇ is Hermitian. Here ⟨·, ·⟩ denotes the standard Hermitian scalar product on Ca . The right
hand side of the above equality equals

d⟨σ1 , σ2 ⟩ + ⟨Aσ1 , σ2 ⟩ + ⟨σ1 , Aσ2 ⟩,

so that we must have ⟨Aσ1 , σ2 ⟩ = −⟨σ1 , Aσ2 ⟩. Hence, A ∈ Ω1 (U ; u (a)) and therefore F∇ =
dA + 21 [A, A] ∈ Ω2 (U ; u (a)), since the commutator of any pair of matrices from u (a) belongs
to u (a). □

Draft 56 May 24, 2024

 Global analysis

3.2.3 Chern classes

Let p : u (a) → C be a homogeneous polynomial of degree d. This P means the following:
Assume ξ1 , . . . , ξn is a basis of u (a) (in particular, n = a2 ). If ξ = nj=1 xj ξj , then p (ξ) =
P 
n
p j=1 j j is a polynomial of degree d in (x1 , . . . , xn ). Assume also that p is invariant, that
x ξ
is
p U ξU −1 = p (ξ)


for all ξ ∈ u (a) and U ∈ U (a) .
Example 3.22.
1) pd (ξ) := i tr ξ d is an invariant homogeneous polynomial of degree d.
2) p (ξ) := det ξ is an invariant homogeneous polynomial of degree a.
3) More generally, define polynomials c1 , . . . , ca on u (a) by the equality
 
i
det λ · 1 + ξ = λa + c1 (ξ) λa−1 + . . . + ca−1 (ξ) λ + ca (ξ) ,
2π
where λ ∈ C is a parameter and 1 is the identity matrix. For example, c1 (ξ) = 2πi tr ξ
ia
and ca (ξ) = (2π) a det ξ. The reader should check that cj is an invariant homogeneous

polynomial of degree j.
Moreover, notice that the equality
   
i i
det λ · 1 + ξ = det λ̄ · 1 + ξ
2π 2π

implies that cj (ξ) = cj (ξ), that is each cj takes values in R rather than C.
Definition 3.23. Let E be a Hermitian vector bundle of rank a. Pick an invariant homogeneous
polynomial p as above and a Hermitian connection ∇. In a local trivialization e = (e1 , . . . , ea )
such that e is pointwise an orthonormal basis, think of F∇ as a 2-form F∇loc = dA + 21 [A, A]
with values in u (a). Finally, set p (F∇ ) = p F∇loc ∈ Ω2d (M ; C).

P to the above definition, choose a basis2 (ξ1 , . . . , ξn ) of u(a) as above.

To provide some details
If ξ ∈ u(a), then ξ = xj ξj and p is just a polynomial in n = a variables xj :
X
p = p(x1 , . . . , xn ) = pj1 ...jn xj11 . . . xjnn .
j1 +···+jn =d

Furthermore, suppose that e is a local trivialization over an open set U as above.

We can think
of F∇loc as a 2-form with values in u(a), or, equivalently, as a matrix Fij , whose entries are
2-forms on U and Fji P = −Fij . With respect to a chosen basis (ξ1 , . . . , ξn ) we obtain a unique
representation F∇loc = Fj ξj , where Fj ∈ Ω2 (U ). Then
X
p(F∇ ) = pj1 ...jn F1j1 . . . Fnjn ,
j1 +···+jn =d

where the multiplication is understood as the wedge-product. Notice that all F1 , . . . , Fj commute
as forms of even degrees thus manifesting that p(F∇loc ) is well-defined as an element of Ω2d (U ).
Of course, one is free to choose the basis of u(a) in the form of skew-symmetric matrices ξij
with just one non-trivial entry above the diagonal (or a single non-trivial entry on the diagonal
if i = j). In this case the corresponding ‘variables’ are just Fij . This particular choice may be
convenient sometimes but is immaterial for our exposition.

Draft 57 May 24, 2024

 Global analysis

Lemma 3.24. The following holds:

(i) p (F∇ ) is well-defined.

(ii) p (F∇ ) is closed.
(iii) The de Rham cohomology class of p (F∇ ) depends neither on the choice of ∇ nor on the
Hermitian scalar product on E.

Proof. Claim (i) follows easily from (3.18) and the invariance of p. The rest of the proof consists
of the following steps.
Step 1. We prove (ii).
Let ξ, ξ1 , . . . , ξd be arbitrary elements of u (a). Slightly abusing notations, denote by p : u (a)×
. . . × u (a) → C the symmetrization of p, that is a multilinear symmetric map, whose restriction
to the diagonal {(ξ, . . . , ξ)} yields the original polynomial p.
Differentiating the equality

p etξ ξ1 e−tξ , . . . , etξ ξd e−tξ = p (ξ1 , . . . , ξd )

with respect to t and setting t = 0, we obtain that the equality

p ([ξ, ξ1 ] , ξ2 , . . . , ξd ) + p (ξ1 , [ξ, ξ2 ] , . . . , ξd ) + . . . + p (ξ1 , ξ2 , . . . , [ξ, ξd ]) = 0 (3.25)

holds for any ξ, ξ1 , . . . ξd ∈ u (a). This implies the following identity

p F∇loc , A , A, . . . , A + . . . + p A, A, . . . , F∇loc , A = 0.
 
 

Hence,

dp F∇loc = p dF∇loc , F∇loc , . . . , F∇loc + p F∇loc , dF∇loc , . . . , F∇loc + . . . + p F∇loc , . . . , dF∇loc

= p FAloc , A , F∇ , . . . , F∇ + p F∇loc , F∇loc , A , . . . , F∇loc + . . .

 
 

+ p F∇loc , F∇loc , . . . , F∇loc , A

 

= 0,

where the first equality follows from the Bianchi identity and the second one from (3.25).
Step 2. Let I = [0, 1] be the interval and ı0 , ı1 : M → I be the natural inclusions corresponding
to the endpoints of the interval. There exists a linear map Q : Ωk (M × I) → Ωk−1 (M ) such
that for any ω ∈ Ωk (M × I) we have ı∗1 ω − ı∗0 ω = dQω − Qdω. ref
Step 3. The cohomology class of p (F∇ ) does not depend on the choice of ∇.
Pick any two Hermitian connections ∇0 and ∇1 and think of ∇t := (1 − t) ∇0 + t∇1 as a
connection on pr1∗ E → M × I, where pr1 : M × I → M is the natural projection. Then we
have
p (F∇0 ) − p (F∇1 ) = ı∗1 p (F∇t ) − ı∗0 p (F∇t )
= dQp (F∇t ) − Qdp (F∇t )
= dQp (F∇t ) .
Here the second equality follows from Step 2 and the third one from Step 1.
Step 4. The cohomology class of p (F∇ ) does not depend on the Hermitian scalar product of E.

Draft 58 May 24, 2024

 Global analysis

The proof of this step is similar to the proof of the preceding one and follows in essense
from the fact that the space of all Hermitian scalar products is convex. I leave the details to the
reader. □
Let cj be as in Example 3.22. Then by Lemma 3.24, cj (F∇ ) is a closed real-valued form
and the de Rham cohomology class of cj (F∇ ) depends on E only.
2j
Definition 3.26. The class cj (E) := [cj (F∇ )] ∈ HdR (M ; R) is called the j th Chern class of E
and
c (E) := 1 + c1 (E) + c2 (E) + . . . + ca (E) ∈ H • (M ; R)
is called the total Chern class of E.

Theorem 3.27. The Chern classes satisfy the following properties:

(i) c0 (E) = 1 for any complex vector bundle E;

(ii) The Chern classes depend on the isomorphism class of E only;
(iii) cj (f ∗ E) = f ∗ cj (E) for all complex vector bundles E → M and all (smooth) maps
f : N → M;
(iv) c (E1 ⊕ E2 ) = c (E1 ) ∪ c (E2 );
(v) If E ∼ = E1 ⊕ Cb , then cj (E) = 0 for j > rk E − b. In particular, if E is trivial, then
cj (E) = 0 for all j > 0.

Proof. Property (i) is just the definition of c0 . The rest of the proof consists of the following
steps.
Step 1. We prove (ii).
Suppose ψ : E0 → E1 is an isomorphism. Given a Hermitian scalar product ⟨·, ·⟩1 and a
Hermitian connection ∇1 on E1 we can define a Hermitian scalar product ⟨·, ·⟩0 and a Hermitian
connection ∇0 on E0 as follows:

⟨t1 , t2 ⟩0 = ⟨ψ t1 , ψ t2 ⟩1 , t1 , t2 ∈ E0 .
∇0 s0 := ψ −1 (∇1 (ψs0 )) , s0 ∈ Γ (E0 ) .

If e1 = (e11 , . . . , e1a ) is a local trivialization of E1 such that e1 is pointwise orthonormal, then

e0 := (ψ −1 e11 , . . . , ψ −1 e1a ) is a local trivialization of E0 with the same property. Hence, if
A ∈ Ω1 (U ; u (a)) is a local representation of ∇1 , then the local representation of ∇0 with
respect to e0 is also A, since by (2.14) we have

∇1 e1 = e1 · A ∇0 e0 = ψ −1 ∇1 e1 = ψ −1 e1 · A = e0 · A.



=⇒

This yields that the curvature forms of ∇0 and ∇1 with respect to the above trivializations
coincide, so that we trivially have c (E0 ) = c (E1 ).
Step 2. We prove (iii).
Given a Hermitian scalar product ⟨·, ·⟩ on E and a local trivialization e = (e1 , . . . , ea ) of
E over U such that e is pointwise an orthonormal basis, we can construct a Hermitian scalar
product and a local trivialization of f ∗ E over f −1 (U ) by

⟨t1 , t2 ⟩n := ⟨t1 , t2 ⟩f (n) , t1 , t2 ∈ En , n ∈ N

f ∗e n
:= e f (n)
.

Draft 59 May 24, 2024

 Global analysis

Furthermore, if a Hermitian connection ∇ on E is represented by some A ∈ Ω1 (U ; u (a)), then

we can define a new connection f ∗ ∇ on f ∗ E by declaring

(f ∗ ∇) (f ∗ e) = (f ∗ e) · f ∗ A,

that is f ∗ A is a local representation of f ∗ ∇ with respect to f ∗ e. Then

 
1 ∗ 1
Ff ∗ ∇ = d (f A) + [f A, f A] = f dA + [A, A] = f ∗ F∇ ,
∗ ∗ ∗
2 2

which implies (iii).

Step 3. We prove (iv).

Let ∇j be a connection on Ej , j = 1, 2. Define a connection ∇ on E1 ⊕ E2 by

∇ (t1 , t2 ) = (∇1 t1 , ∇2 t2 ) .

If ej is a local trivialization of Ej , then e = (e1 , e2 ) is a local trivialization of E1 ⊕ E2 . With

respect to these trivializations we have
 
F∇loc1 0
F∇loc =  ,
0 F∇loc2

where the right hand side is a block-matrix with non-trivial blocks in the upper left and lower
right corners only.
Since for any block-matrix we have
!
A 0
det = det A · det B,
0 B

we obtain  
1 + 2πi F∇1 0
c (F∇ ) = det   = c (F∇1 ) ∧ c (F∇2 )
0 1+ i
F
2π ∇2

This implies (iv).

Step 4. We prove (v).

If E ∼
= Ca , then we can use the trivial connection ∇ = d to deduce that c (E) = 1. If
E∼
= E1 ⊕ Cb , by the preceding step we have

c (E) = c (E1 ) ∪ 1 = c (E1 ) ,

which yields (v). □

Draft 60 May 24, 2024

 Global analysis

3.2.4 Some computations

The Chern number of the 2-sphere
Recall that S 2 ∼ = CP1 is a complex manifold and, hence, T S 2 is a complex line bundle.
Explicitly, if w ∈ Tp S 2 , then Iw is a unique vector in Tp S 2 such that |Iw| = |w|, Iw ⊥ w, and
(w, Iw) is positively oriented provided w ̸= 0. It is therefore reasonable to ask for the Chern
class of T S 2 .
First recall that H k (S 2 ) ̸= 0 if and only if k ̸= 0, 2. Moreover, H 2 (S 2 ) ∼ = R and an
isomorphism is given by Z
[ω] 7→ ω.
S2

Therefore, the total Chern class of any complex line bundle on S 2 is determined by a single
number, 2πi S 2 F∇ , where ∇ is an arbitrary connection on T S 2 . This is called the Chern number
R

of S 2 . A well-known fact is that the Chern number is in fact an integer, but we shall see this
directly.
Define a connection on T S 2 as in Example 2.11. In other words, thinking of a vector field
v
on S 2 as a smooth map v : S 2 → R3 such that v(p) ⊥ p for all p ∈ S 2 , define ∇v := prT S 2 dv ,
where prT S 2 denotes the orthogonal projection onto T S 2 . Even more explicitly, for w ∈ X(S 2 )
we have
dv D dv E
∇w v = − , n n,
dw dw
where n is the unit normal field on S 2 : n(p) = p. The reader should check that this connection
preserves both the Riemannian metric and the complex structure on S 2 .
Furthermore, pick w ∈ Tp S 2 such that |w| = 1 and local coordinates (x1 , x2 ) on S 2 near p
such that
∂ ∂
(p) = w and (p) = Iw.
∂x1 ∂x2
∂v
Denoting ∂j v = ∂xj
and ∇j v := ∇ ∂ v, we obtain
∂xj

2 2
∇2 ∇1 v = ∂21 v − ⟨∂21 v, n⟩ n − ⟨∂1 v, ∂2 n⟩ n − ⟨∂1 v, n⟩ ∂2 n,
2 2
∇1 ∇2 v = ∂12 v − ⟨∂12 v, n⟩ n − ⟨∂2 v, ∂1 n⟩ n − ⟨∂2 v, n⟩ ∂1 n.
2 2
Using the equality ∂12 v = ∂21 v, we have



F∇ (∂1 , ∂2 )v = ∇1 ∇2 v − ∇2 ∇1 v = ⟨∂1 v, n⟩ ∂2 n − ⟨∂2 v, n⟩ ∂1 n.

Notice that I also ignored the terms that are parallel to n, since the left hand side of the above
equality is a vector field on S 2 and ∂j n is pointwise orthogonal to n, that is a vector field on S 2 .
The equality ⟨v, n⟩ ≡ 0 implies ⟨∂j v, n⟩ = −⟨v, ∂j n⟩. Hence, we obtain

F∇ (∂1 , ∂2 )v = ⟨v, ∂2 n⟩ ∂1 n − ⟨v, ∂1 n⟩ ∂2 n. (3.28)

Notice that the right hand side of the above expression does not contain derivatives of v; this is
expected since F∇ (∂1 , ∂2 ) is a section of End(T S 2 ).
Furthermore, recalling that ∇ preserves the Riemannian metric and the complex structure,
we conclude that in fact F∇ (∂1 , ∂2 ) must take values in skew-Hermitian endomorphisms of
T S 2 . Hence, F∇ (∂1 , ∂2 ) can be thought of as a function with values in iR, since up to the
multiplication with a real number the only skew-Hermitian endomorphism of T S 2 is I. Explicitly,
if v is a local vector field on S 2 such that |v| = 1 everywhere (on an open subset), then this

Draft 61 May 24, 2024

 Global analysis

purely imaginary valued function is given by ⟨F∇ (∂1 , ∂2 )v, Iv⟩ (up tp the factor i). Using (3.28),
we obtain
 
⟨v, ∂1 n⟩ ⟨v, ∂2 n⟩
⟨F∇ (∂1 , ∂2 )v, Iv⟩ = ⟨v, ∂2 n⟩ ⟨∂1 n, Iv⟩−⟨v, ∂1 n⟩ ⟨∂2 n, Iv⟩ = − det .
⟨Iv, ∂1 n⟩ ⟨Iv, ∂2 n⟩
Notice that the left hand side of the above formula is independent of the choice of v.
2
Think of n as the Gauss map of the sphere, i.e., n : S → S 2 . Its differential at p is a linear
∂ ∂


map dp n : Tp S 2 → Tn(p)S 2 = Tp S 2 that maps ∂x1 , ∂x2 to (∂1 n, ∂2 n). Of course, n is just the
identity map so that at p we have (∂1 n, ∂2 n) = (w, Iw). Since we are free to choose v so that
v(p) = w we see that for these choices the matrix appearing above is just the identity matrix.
Thus, taking into the account the ommitted factor i we obtain F∇ = −i volS 2 and therefore
Z Z
2 i 4π
c1 (T S ) = i −volS 2 = = 2.
S2 2π S 2 2π
Thus, the Chern number of S 2 equals 2.

The Chern class of the tautological line bundle

Just like in Example 2.2, consider the tautological bundle of the complex projective space1

O(−1) := [z], w ∈ CP1 × C2 w = 0 or [w] = [z] .



We could proceed just like in the example above, that is define a connection on O(−1) by
utilizing the obvious projection CP1 ×C2 → O(−1) and compute the curvature of this connection
explicitly.
To simplify computations somewhat, we shall take a different
route. Notice first, that O(−1)
is a Hermitian vector bundle and its sphere bundle S O(−1) = {ξ ∈ O(−1) | |ξ| = 1} is
diffeomorphic to S 3 . Indeed, thinking of S 3 as a subset of C2 ∼ = R4 a diffeomorphism may be
given explicitly as follows:

S 3 ∋ (z0 , z1 ) 7→ [z0 : z1 ], (z0 , z1 ) ∈ O(−1).

In particular, the standard U(1)-action on S 3 induces a transitive action on each fiber so that we
can think of CP1 ∼ = S 2 as the quotient S 3 /U(1). Explicitly, a diffeomorphism is induced by the
map
(z0 , z1 ) 7→ z0 z̄1 , 21 (|z0 |2 − |z1 |2 ) ∈ C × R ∼
= R3 .

The reader should check that the induced metric on S 3 /U(1) is that of the sphere of radius 1/2.
Let a ∈ Ω1 (S 3 ) be any U(1)-invariant 1-form such that a(K) = 1, where K ∈ X(S 3 ) the
infinitesimal U(1)-action, i.e., K(z) = iz. In real coordinates (x0 , x1 , x2 , x3 ) on R4 we have
K(x) = (−x1 , x0 , −x3 , x2 ) and we can choose

a = −x1 dx0 + x0 dx1 − x3 dx2 + x2 dx3 .

Let s be any trivialization of O(−1) over an open subset U . Without loss of generality, we
can assume that |s| = 1 everywhere on U and think of s as a map U → S 3 . Define

∇a s := i(s∗ a)s,

cf. (2.14). Clearly, this defines a unique connection on O(−1)|U .

1
The notation for the tautological bundle will become somewhat clearer towards the end of the computation.

Draft 62 May 24, 2024

 Global analysis

Exercise 3.29. Show that the U(1)-invariance implies that the connection ∇a does not depend
on the choice of s. Moreover, show that ∇a is Hermitian.
Thus, the above exercise implies that there is a unique Hermitian connection on O(−1) such
that for any local trivialization s such that |s| = 1 we have ∇a s = i(s∗ a)s. Hence,
F∇a s = d∇a (∇a s) = i d(s∗ a) s + is∗ a ∧ is∗ a s = i d(s∗ a) s + 0 = i s∗ da s.

Furthermore, observe that T[z] S 3 /U(1) can be identified with K(z)⊥ and the latter space is
spanned by
v1 (x) = (−x2 , x3 , x0 , −x1 ) and v2 = (−x3 , −x2 , x1 , x0 ).
In fact, (v1 , v2 ) is an orthonormal oriented basis of K(z)⊥ . If ω = dx0 ∧ dx1 + dx2 ∧ dx3 , we
have
ıK ω = −x1 dx1 − x0 dx0 − x3 dx3 − x2 dx2 = − 21 d(x20 + x21 + x22 + x23 ) = 0 on S 3 and
ω(v1 , v2 ) = x22 + x23 + x20 + x21 = 1.

Hence, ω is the pull-back of the volume form on S 3 /U(1) and therefore da = 2ω is twice the
pull-back of the volume form. It follows that Fa = 2i vol and therefore the Chern number of
O(−1) is
2i2
Z
i 2
Fa = Vol(S1/2 ) = −1.
2π S 2 2π
Exercise 3.30. Show that for any natural number n the Chern class of the tautological bundle
of CPn is the negative of the standard generator of H 2 (CPn ).

The Chern class of the tangent bundle of the complex projective space
Let L be a complex linear subspace in Cn+1 of dimension 1, i.e., a line through the origin.
Observe that any line near L can be represented as the graph of a unique linear map A : L → L⊥ .
In other words, T[L] CPn ∼
= Hom(L; L⊥ ) or, as complex vector bundles,
T CPn ∼ = Hom O(−1); O(−1)⊥ ,

where O(−1)⊥ denotes the quotient vector bundle:

O(−1)⊥ := [z], w | ⟨w, z⟩ = 0 .


In the above formula, ⟨·, ·⟩ denotes the standard Hermitian scalar product on Cn+1 .
Observe that by the definition of O(−1)⊥ we have O(−1) ⊕ O(−1)⊥ ∼ = Cn+1 . Hence,
Hom O(−1); O(−1) ⊕ O(−1)⊥ ∼ = Hom O(−1); Cn+1 ∼ = O(−1)∨ ⊗ Cn+1 ,

where O(−1)∨[z] := Hom O(−1)[z] ; C is the (complex) dual vector bundle. Using the isomorphism

Hom O(−1); O(−1) ∼ = C, we obtain

Hom O(−1); O(−1) ⊕ O(−1)⊥ ∼ = Hom O(−1); O(−1) ⊕ T CPn ∼ = C ⊕ T CPn .

Denoting by x ∈ H 2 (CPn ) the first Chern class of O(−1)∨ and comparing the right hand
sides of the above two formulae, we obtain
c(T CPn ) = c(T CPn ⊕ C) = c O(−1)∨ ⊕ . . . ⊕ O(−1)∨

= (1 + x)n+1 = 1 + (n + 1)x + . . .

Draft 63 May 24, 2024

 Global analysis

Exercise 3.31. Show that for any complex vector bundle E we have cj (E ∨ ) = −cj (E).
Combining the above exercise with our computation of the Chern class of O(−1), we
see that x is in fact the standard generator of H 2 (CPn ). In particular, for n = 1 we obtain
c(T CP1 ) = (1 + x)2 = 1 + 2x, i.e., c1 (T S 2 ) = 2x, which is in line with our previous
computation.

3.2.5 Other characteristic classes

Assume E can be written as the Whitney sum of line bundles: E = L1 ⊕ . . . ⊕ La . Then for the
total Chern class of E we have
c (E) = c (L1 ) · . . . · c (La ) = (1 + c1 (L1 )) · . . . · (1 + c1 (La )) ,
where · denotes the product in H • (M ) i.e., the cup-product. Denoting xj := c1 (Lj ), we obtain
a
Y
c (E) = (1 + xj ) = 1 + σ1 + σ2 + . . . + σa ,
j=1

where σk = σk (x1 , . . . , xa ) is the k th elementary symmetric function of x1 , . . . , xk . In other

words,
c1 (E) = σ1 (x1 , . . . , xa ) = x1 + . . . + xa ,
c2 (E) = σ2 (x1 , . . . , xa ) = x1 x2 + x1 x3 + . . . + xa−1 xa ,
...
ca (E) = σa (x1 , . . . , xa ) = x1 x2 . . . xa .
The following algebraic fact will be useful below.
Fact: For any symmetric polynomial p = p (x1 , . . . , xa ) there exists a unique polynomial q in
a variables such that
p (x1 , . . . , xa ) = q (σ1 (x) , . . . , σa (x)) .

This fact implies the following: Given any symmetric polynomial p (x1 , . . . , xa ) we can
construct a characteristic class p (E) by setting
p (E) := q (c1 , c2 , . . . , ca ) ∈ H • (M ; R) .
For example, choose
a
X
ch (x1 , . . . , xa ) = e xj
j=1
   
1 2 1 2
= 1 + x1 + x1 + . . . + . . . + 1 + xa + xa + . . .
2 2
1 2
x1 + . . . + x2a + . . .


= a + (x1 + . . . + xa ) +
2
1 2

= a + σ1 + σ − 2σ2 + . . .
2 1
Notice that here and below, we think of xj as nilpotent elements of the ring H • (M ; R) so that
all sums above are finite.
Thus we have
1 2

ch (E) = rk E + c1 (E) + c1 (E) − 2c2 (E) + . . .
2

Draft 64 May 24, 2024

 Global analysis

Remark 3.32. It is not true in general that any vector bundle can be written as the Whitney sum
of line bundles. Nevertheless, this is true in some sense, which I shall not try to describe here.
In particular, when manipulating with characteristic classes it is admissable to imagine that
bundles decompose as Whitney sums of line bundles. This is called "the splitting principle",
which is explained somewhat more concretely in the proof below.
Proposition 3.33. The Chern character satisfies the following:
(i) ch (E ⊕ F ) = ch (E) ⊕ ch (F );
(ii) ch (E ⊗ F ) = ch (E) · ch (F ).
Proof. Assume that both E and F split as Whitney sums of line bundles:

E = L1 ⊕ . . . ⊕ La and F = L1 ⊕ . . . ⊕ Lb .

Denote xj := c1 (Lj ) and yk := c1 (Lk ). The bundle E ⊕ F also splits as the Whitney sum of
line bundles so that we have
a
X b
X
xj
ch (E ⊕ F ) = e + eyj = ch (E) + ch (F )
j=1 k=1

for all bundles, which split as the sum of line bundles. This imples that we have the algebraic
identity
ch (x1 , . . . xa , y1 , . . . , yb ) = ch (x1 , . . . , xa ) + ch (y1 , . . . , yb ) (3.34)
where we think of ch simply as a polynomial in the corresponding number of variables. Notice
that (3.34) can be easily established purely algebraically without any reference to vector bundles.
Furthermore, if qa+b (σ1 , . . . , σa+b ) is the expression of ch (x1 , . . . , xa ; y1 , . . . , yb ) in terms
of elementary symmeric polynomials in a + b variables, then we must have

qa+b (σ1 (x, y) , . . . , σa+b (x, y)) = qa (σ1 (x) , . . . , σa (x)) + qb (σ1 (y) , . . . , σb (y)) .

This implies that (i) holds not only for those vector bundles, which split, but rather for all
bundles.
The proof of (ii) goes along similar lines and the details are left to the reader. □
Definition 3.35. The Todd class td (E) ∈ H • (M ; R) is the characteristic class corresponding
to the polynomial
a
Y xj 1 1 2


td (x1 , . . . , xa ) := = 1 + σ1 + σ 2 + σ1 + ...
j=1
1 − e−xj 2 12

3.3 A prototype of the index theorem

Denote

Ωev (M ) := Ω0 (M ) ⊕ Ω2 (M ) ⊕ . . . and Ωodd (M ) := Ω1 (M ) ⊕ Ω3 (M ) ⊕ . . .

Exercise 3.36. Show that

D := d + d∗ : Ωev (M ) −→ Ωodd (M ) (3.37)

is elliptic.

Draft 65 May 24, 2024

 Global analysis

If M is compact, which is assumed throughout in this section, by Proposition 1.29 we have

Ker D = H0 (M ) ⊕ H2 (M ) ⊕ . . . ∼ ev
= HdR (M ) .

Moreover, Theorem 2.80 implies

Coker D = H1 (M ) ⊕ H3 (M ) ⊕ . . . ∼ odd
= HdR (M ) .

Therefore, for the index of D we obtain

ev odd
ind D = dim HdR (M ) − dim HdR (M )
= b0 (M ) + b2 (M ) + . . . − b1 (M ) − b3 (M ) − . . .
= χ (M ) ,

where χ (M ) is the Euler characteristic of M . Thus, we summarize.

Theorem 3.38. The index of (3.37) equals χ (M ). □

This is an interesting theorem, since this is an example of a relation between solutions of

PDEs and a purely topological quantity. For example, if χ (M ) > 0, we deduce that M must
support at least one non-trivial harmonic form. Of course, the reader surely can give a more
precise relation between topological invariants of M and the number of linearly independent
harmonic forms on M , however notice the following:
• The Euler number is easier to compute than the individual Betti numbers.
• In more general situations the index is relatively easily computable, whereas more detailed
information about solutions of PDEs is typically hard to obtain.

3.4 On the de Rham cohomolgy with compact supports

For non-compact manifolds the de Rham cohomology groups are typically poorly behaved.
There are many ways to adapt the definition of the de Rham cohomology groups to this setting.
One possibility is to replace each Ωk (M ; R) by

Ωk0 (M ) := ω ∈ Ωk (M ) | supp ω is compact .



That is instead of the de Rham complex (1.1) we consider

d d d
0 → Ω00 (M ) −→ Ω10 (M ) −→ . . . −→ Ωn0 (M ) → 0.

Akin to (1.3), we define H0k (M ) to be the k th cohomolgy group of the above complex.
Of course, if M is compact, we trivially have H k (M ) = H0k (M ), however for non-compact
manifolds H k (M ) ̸= H0k (M ) in general. For example, one can show that H0n (Rn ) ∼ = R, see
[BT82, Cor. 4.7.1].
The only property which will be of concern for us is the following. If M n is oriented, the
integration Z
: Ωn0 (M ) −→ R

is well-defined and yields an isomorphism H0n (M ) ∼

= R [BT82, Cor. 5.8].

Draft 66 May 24, 2024

 Global analysis

3.5 Basics of K-theory

Let N be a semiring, that is N is endowed with addition and multiplication just like a ring,
however additive inverse of an element n ∈ N does not need to exist.

Example 3.39. The set of natural numbers N is a semiring.

Any semiring N can be canonically ’enlarged’ to a ring as follows. Denote

K (N ) = {n − m | n, m ∈ N } / ∼,

where the difference n − m is understood formally, and

n − m = n′ − m′ ⇐⇒ ∃ k ∈ N such that n + m′ + k = n′ + m + k.

Then K (N ) is a ring, since the additive inverse of n − m exists and equals m − n.

Example 3.40. K (N) ∼

= Z.

Let M be a compact manifold. Then the set of isomorphism classes of (complex) vector
bundles over M is a semiring with respect to ⊕ and ⊗ operations. Hence, we can construct a
ring K (M ), which consists of (classes of) formal diferences E − F . By setting

ch (E − F ) := ch (E) − ch (F )

we obtain by Proposition 3.33 that ch : K (M ) −→ H • (M ) is a well-defined homomorphism

of rings.
If M is non-compact, one has to modify the above construction slightly so that the resulting
ring behaves nicely. Thus, for a non-compact M roughly speaking the ring K (M ) consists of
classes of formal differences E − F such that E and F are isomorphic on the complement of a
compact subset A ⊂ M . By the proof of Theorem 3.27, there exist connections ∇E and ∇F on
E and F respectively such that F∇E = F∇F on M \A. Hence, cj (E) and cj (F ) are represented
by 2j-forms, which agree on M \A so that ch (E − F ) ∈ H0• (M ).
The upshot of this section is that it is legitimate to consider formal differences of vector
bundles and if E and F are isomorphic away from a compact subset of M , then ch (E · F )
takes values in the compactly supported cohomology ring of M .
Let me note in passing that the ring K (M ) is an important invariant of the underlying space
M . This invariant has been extensively studied, however this goes beyond the scope of these
lecture notes.

3.6 The symbol of an elliptic operator revisited

Let us consider first the following construction of vector bundles on the spheres. Denote

S±n := {(x0 , . . . xn ) ∈ S n | ± xn ≥ 0}

so that S+n ∩ S−n = S n−1 is the equator. Given a (smooth) map g : S n−1 → GLa (C), we can
construct a complex vector bundle E of rank a on S n as follows. Declare E to be

S+n × Ca ⊔ S−n × Ca \ ∼,



(3.41)

Draft 67 May 24, 2024

 Global analysis

where (x, v) ∼ (x, g(x)v) if and only if x ∈ S n−1 . This is usually called the clutching
construction, which yields all complex vector bundles on S n up to an isomorphism.
Furthermore, pick σ ∈ Smbl (E; F ) and a point m ∈ M . We can do the following
parameterized version of the clutching construction. Denote first by Tm∗ M + the one-point
= Rn so that Tm∗ M + ∼
compactification of Tm∗ M ∼ = S n . We can think of Tm∗ M + as the union of
the "hemispheres"
−
= ξ ∈ Tm∗ M | |ξ| ≤ 1 +
= ξ ∈ Tm∗ M | |ξ| ≥ 1 ∪ {∞}
 
Sm and Sm
−
+
so that Sm ∩ Sm = {|ξ| = 1} ⊂ Tm∗ M . Therefore, we can construct the bundle Σm by attaching
− +
Sm × Em to Sm × Fm by means of σ. To be more precise,
− +





Σm := Sm × Em ⊔ Sm × Fm / ∼, where (ξ, v) ∼ ξ, σ(m, ξ)v

provided |ξ| = 1. As m varies over M , we obtain a complex vector bundle Σ over S n M ,

where S n M is a fibered space arising as pointwise one-point compactifications of the fibers of
T ∗ M . Sometimes this vector bundle Σ is referred to as the symbol of L ∈ Diff l (E; F ) provided
σ = σl (L).
In any case, Σ − π ∗ F can be viewed as an element of K (S n M ). It will be more convenient
for us to view Σ − π ∗ F as an element of K (T ∗ M ). This makes sense, since Σ and π ∗ F are
isomorphic on the complement of the set {|ξ| ≤ 1}, which is compact. Therefore, the class
ch (Σ − F ) ∈ H0• (T ∗ M ) is well defined.
With these preparations at hand, we can state the index theorem.

Theorem 3.42 (Atiyah-Singer). Let M be a smooth closed oriented manifold of dimension n.

Let L ∈ Diff l (E; F ) be an elliptic differential operator. Let Σ be a vector bundle associated
with σl (L) as above. Then
Z
n
ind L = (−1) ch (Σ − π ∗ F ) ∪ π ∗ td (T M ⊗ C) . (3.43)
T ∗M

Remark 3.44.

wj ∈ H0• (M ; R), where wj ∈ H0j (M ; R), by definition we have

P
1) RFor any Rclass ω =
M
ω = M ωn .
2) A priori it is by no means clear that the right hand side of (3.43) is an integer (though one
can show that this is a rational number by purely topoligical means).
3) The right hand side of (3.43) can be expressed as an integral over M rather than T ∗ M .
Most frequently one finds the Atiyah-Singer theorem in the literature in the form with the
integration over M .

3.7 An application: The Riemann-Roch formula

Let E → Σ be a complex vector bundle over a Riemann surface Σ.

Definition 3.45. A holomorphic structure on E is an open covering U = {Uα } of Σ with the

following properties:

• E admits a trivialization eα over each Uα ;

Draft 68 May 24, 2024

 Global analysis

• On each Uαβ := Uα ∩Uβ the corresponding transition maps gαβ : Uαβ → GLa (C) defined
by eβ = eα · gαβ are holomorphic.

If E is a holomorphic vector bundle, that is a complex vector bundle equipped with a

holomorphic structure, we can define the Dolbeault operator ∂¯ : Ω0 (E) → Ω0,1 (E) by the
following rule: If s = eα · σα over Uα , then ∂s¯ := eα · ∂σ¯ α . In other words, we declare each
α
ej to be a local holomorphic section of E over Uα . The reader will have no difficulties to check
that ∂¯ is well-defined. Moreover, ∂¯ is elliptic just like the Dolbeault operator considered in
Section 2.5.1.
Clearly H 0 (E) := Ker ∂¯ is the space of global holomorphic sections. Denote also H 1 (E) :=
Coker ∂.¯ If Σ is closed, the Atiyah-Singer index theorem yields an expression for

ind ∂¯ = dim H 0 (E) − dim H 1 (E)

in terms of characteristic classes of E. Since Σ is one-dimensional, the only relevant class is

c1 (E). A computation shows that
Z
0 1
dim H (E) − dim H (E) = a (1 − g) + c1 (E) ,
Σ

where a = rk E and g is the genus of Σ. This is the celebrated Riemann-Roch formula. In

particular, if Z
c1 (E) > (g − 1) a,
Σ
then E admits a non-trivial holomorphic section.
For example, for E = T ∗ Σ ∼ 1,0
= (T ∗ Σ) we have
Z
c1 (T ∗ Σ) = 2g − 2
Σ

so that the Riemann-Roch formula yields

dim H 0 (T ∗ Σ) − dim H 1 (T ∗ Σ) = 1 − g + 2g − 2 = g − 1.

Hence, a compact Riemann surface admits a non-trivial holomorphic (1, 0) − form provided
g (Σ) > 1.

Draft 69 May 24, 2024

Bibliography

[BT03] D. Barden and C. Thomas, An introduction to differential manifolds, Imperial College Press, London,
2003. MR1992457 ↑20
[BT82] R. Bott and L. Tu, Differential forms in algebraic topology, Graduate Texts in Mathematics, vol. 82,
Springer-Verlag, New York-Berlin, 1982. MR658304 ↑66
[Don11] S. Donaldson, Riemann surfaces, Oxford Graduate Texts in Mathematics, vol. 22, Oxford University
Press, Oxford, 2011. MR2856237 ↑12
[Hir94] M. Hirsch, Differential topology, Graduate Texts in Mathematics, vol. 33, Springer-Verlag, New York,
1994. Corrected reprint of the 1976 original. MR1336822 ↑6
[Wel08] R. Wells, Differential analysis on complex manifolds, Third, Graduate Texts in Mathematics, vol. 65,
Springer, New York, 2008. With a new appendix by Oscar Garcia-Prada. MR2359489 ↑38

70