You are on page 1of 101

Global Analysis 2: Pseudodifferential Opera-

tors
Lecture course at the University of Bonn in the WS 2020/2021

Batu Güneysu

1. Introduction
One could say that the aim of this course is to provide a machinery (namely: pseudodif-
ferential operators) that allows to study the solvability and regularity of solutions of linear
partial differential equations of the form P f = h, where P is a linear partial differential
operator with smooth coefficients acting on sections of a vector bundle over a manifold,
f is the unknown section and h is a given section. We are also going to explain some
applications of this theory to differential topology.
Since the definition of pseudodifferential operators will turn out be directly motivated by
the Fourier transform, special topological function spaces such as Schwartz functions will
play a fundamental role in the course (ultimately, we would like pseudodifferential operators
to be continuous linear maps between such spaces). These function spaces are typically
not normed in a useful way, but come equipped with useful families of (semi-)norms. In
addition, we will also need the topological dual spaces of these functions (’distributions’),
which will then carry caonically given families of (semi-)norms. The abstract theory behind
such spaces is the one of locally convex spaces, which will cover the first part of the course.

We will then apply this abstract theory to particular function spaces and study the Fourier
transform and Sobolev embedding theorems.
The next part of the course will cover the local theory of pseudodifferential operators,
and finally we will present some global applications of this theory on manifolds, such as a
complete proof of the Hodge theorem (which relates geometry with topology).
Let us start with the most basic definition of a differential operator, namely that one acting
on vector-valued functions on an open subset of Rm : let m, `1 , `2 ∈ N, k ∈ N0 .
Definition 1.1. Given an open subset U ⊂ Rm , a linear map
P : C ∞ (U, C`1 ) −→ C ∞ (U, C`2 )
m
Pam (linear smooth partial) differential operator of order ≤ k, if for all α ∈ N0 with
is called
|α| := j=1 αj ≤ k there exist (necessarily uniquely determined) smooth maps
Pα : U −→ Mat`1 ×`2 (C)
1
2 B. GÜNEYSU

such that for all f ∈ C ∞ (U, C`1 ) one has P f = Pα Dα f , where


P
|α|≤k

Dα := i−α ∂xα11 · · · ∂xαmm .


There is nothing particularly remarkable about √ this definition, exceptα maybe that the
convention including the imaginary unit i = −1 in the definition of D will turn out to
be useful in context of the Fourier transform later on.
Our next aim is to generalize this definition by replacing the spaces C ∞ (U, C` ) with smooth
sections of a vector bundle over a manifold.
To this end, let M be a manifold with dim(M ) = m and let E → M , F → M be complex
vector bundles1 with `1 = rank(E), `2 = rank(F ).
Definition 1.2. A linear map
P : ΓC ∞ (M, E) −→ ΓC ∞ (M, F )
is called a differential operator of order of order ≤ k, if it can be restricted to open subsets
of M 2, and if for all manifold charts
ψ : M ⊃ U −→ U 0 ⊂ Rm
which admit vector bundle trivializations
φE : E|U −→ U × C`1 , φF : F |U −→ U × C`2 ,
the linear map defined by the diagram
P |U
ΓC ∞ (U, E|U ) / ΓC ∞ (U, F |U )
O

f 0 7→φ−1 0
E ◦f ◦ψ f 7→φF ◦f ◦ψ −1


C ∞ (U 0 , C`1 ) / C ∞ (U 0 , C`2 )
is a differential operator of order ≤ k in the sense of the previous definition.
The following simple lemma shows that the latter definition is consistent with the former
one in case the manifold is an open subset of Rm and the vector bundles are trivial:
Lemma 1.3. Assume U, U 0 ⊂ Rm are open, ψ : U → U 0 is a diffeomorphism and that
φ1 : U −→ GL(C`1 ), φ2 : U −→ GL(C`2 )
are smooth. Then for every differential operator
P : C ∞ (U, C`1 ) −→ C ∞ (U, C`2 )
1By default we understand these concepts in the category of finite dimensional smooth manifolds without
boundary. Note that if M × Cl → M is a trivial vector bundle, then sections of this bundle can be
canonically identified with functions M → Cl
2that is, if for all V ⊂ M open, there exists a (necessary uniquely determined) linear map P |
V :
ΓC ∞ (V, E|V ) → ΓC ∞ (V, F |V ) such that (P f )|V = PV f |V for all sections f ∈ ΓC ∞ (M, E)
Global Analysis 2 3

or order ≤ k in the sense of Definition 1.1, the linear operator


C ∞ (U 0 , C`1 ) −→ C ∞ (U 0 , C`2 ), f 0 7−→ φ2 P (φ−1 0
1 f ◦ ψ) ◦ ψ
−1

is again a differential operator of order ≤ k in the sense of Definition 1.1.


Proof : Exercise. 
It is easy to check that the composition of a differential operator of order ≤ k with one
of order ≤ l is a differential operator is a differential operator of order k + l; the sum of a
differential operator of order ≤ k with one of order ≤ k is one of order k; scalars from C
are differential operators of order 0.
Example 1.4. 1. Given j ∈ {0, . . . , m}, the de Rham differential
dj : ΩjC ∞ (M ) −→ Ωj+1
C ∞ (M )
is a differential operator of order ≤ 1. Note that by definitions 0-forms are understood to
be functions.
2. Let g be a Riemannian metric on M , that is, a smooth section of the vector bundle
T ∗ M ⊗ T ∗ M → M , such that for all x ∈ M the map
g(x) : Tx M × Tx M −→ R
is a scalar product. Then for every j ∈ {0, . . . , m − 1}, one canonically gets the de Rham
co-differential
dgj : Ωj+1 j
C ∞ (M ) −→ ΩC ∞ (M ),
which is a differential operator of order ≤ 1. It follows that the Hodge-Laplacian on j-forms
∆j,g := dgj dj + dj−1 gj−1
g
: ΩjC ∞ (M ) −→ ΩjC ∞ (M )
is a differential operator of order ≤ 2. For j = 0 one gets the Laplace-Beltrami operator3
∆g := ∆0,g = dg0 d0 : C ∞ (M ) −→ C ∞ (M ),
and this operator is locally in any chart
ψ = (x1 , . . . , xm ) : M ⊃ U −→ U 0 ⊂ Rm
given by
m  
X 1 √ ab
∆g = − √ ∂ xa gg ∂xb ,
a,b=1
g
where
√ p
gab := g(∂xa , ∂xb ), (g ab ) := (gab )−1 , g := det(g ab ).
3. In particular, if M = Rm and g is the Euclidean metric gab = δab , then we get the usual
Laplace operator
X m
∆=− ∂x2a : C ∞ (Rm ) −→ C ∞ (Rm )
a=1

3the minus here is a typical sign-convention in differential geometry, while it is left away in the physics
literature.
4 B. GÜNEYSU

(in cartesian coordinates).


Do not worry, if you are not familiar with these definitions: we are going to revisit them
in greater detail later on.
Let us take a preview of some of the main results of this course:
• Assume (M, g) is a Riemannian manifold, let z ∈ C, and let h be a smooth j-form
on M . Then every j-form f on M which solves (∆j,g + z)f = h in the weakest
possible sense (namely in the sense of distributions) is automatically smooth, and
is a posteriori a classical solution.
• There are nonsmooth solutions of
∂x1 ∂x2 f = 0 in M = R2
For example, pick f˜ : R → R nonsmooth and set f (x1 , x2 ) := f˜(x1 ). So what is the
difference between this and the previous example? Answer: ∆gj + z is an elliptic
operator, while ∂x1 ∂x2 is not elliptic.
• Let M be a m-dimensional manifold and consider its de Rham complex
d d dm−1
0 −→ Ω0 (M ) −→
0
Ω1 −→
1
. . . −−−→ Ωm (M )
j
with the de Rham cohomology groups HdR (M ) := ker(dj )/im(dj−1 ) (note: dj dj−1 =
0). Pick any Riemannian metric g on M and consider
H j (M, g) := ker(∆j,g ).
It follows that
H j (M, g) = ker(dj ) ∩ ker(δj−1,g ) ⊂ ker(dj )
so that each α ∈ H j (M, g) represents a de Rham cohomology class [α] ∈ HdR
j
(M ).
We are going to prove the Hodge-Theorem: If M is compact, then H j (M, g) is
finite dimensional and the map
H j (M, g) −→ HdR
j
(M ), α 7−→ [α]
j
is an isomorphism of (complex) linear spaces. In particular, HdR (M ) is finite dimen-
sional. This results allows to calculate differential topological data from geometric
j
data. Since by de Rham’s Theorem HdR (M ) is even canonically isomorphic to the
(singular) cohomology of the topological space M , in this way one even gets a
geometry ←→ differential topology ←→ topology
type result.
• Elliptic regularity is a local result (in the sense that it holds on any possibly non-
compact Riemannian manifold), while the Hodge-Theorem is a global result (in the
sense that ker(∆j,g ) is in general not finite dimensional on a noncompact M ).
• If time admits, we will present another global application of pseudodifferential
operators: namely, a sketch of proof of the Gauss-Bonnet-Chern theorem for even
dimensional compact manifolds, which allows to calculate the Euler characteristic
(a topological datum) geometrically.
Global Analysis 2 5

Let us close the introduction with a calculation that will ultimately motivate the definition
of a pseudodifferential operators on open subsets of Rm later on:
Example 1.5. Assume U ⊂ Rm is open,
X
P = Pα Dα : C ∞ (U, C`1 ) −→ C ∞ (U, C`2 )
|α|≤k

is a differential operator of order ≤ k and that f ∈ C ∞ (U, C`1 ) is compactly supported.


Then by the so called Fourier inversion formula one finds
Z

f (x) = ei(x,ζ) fˆ(ζ) ,
Rm (2π)m
with Z
da
fˆ(ζ) := ei(ζ,a) f (a)
Rm (2π)m
the Fourier transform of f . Differentiation under the integral yields
Z

(1) P f (x) = ei(x,ζ) σP (x, ζ)fˆ(ζ) ,
Rm (2π)m
with X
σP (x, ζ) := Pα (x)ζ α
|α|≤k

the full symbol of P . Ultimately, a pseudodifferential operator will be precisely an operator


of the form (1), where σP (a polynomial in ζ in the above situation of a differential operator)
will be replaced by a more general smooth function
σ : U × Rm −→ Mat`1 ×`2 (C)
subject to some decay properties that determinse its ’order’. It should then be at least
morally clear4, that if the inverse of a differential opertar exists in some sense, then it is
also pseudodifferential operator (namely, one can replace σ by ’1/σ’).

2. Basics of Locally Convex Spaces


In this section I have mainly followed the presentation from the manuscript by Matthias
Lesch [9]. Classical references are [10, 8, 14, 11]. We understand all our vector spaces over
C.

Definition 2.1. A topological vector space X is a vector space which is also a topological
space such that the multiplication with scalars C × X → X and the addition X × X → X
is continuous.
4This is a very rough statement!
6 B. GÜNEYSU

Remark 2.2. In the above situation, the maps


X −→ X, x 7−→ x + v, with v ∈ X fixed,
as well as
X −→ X, x 7−→ ax, with a ∈ C \ {0} fixed,
are easily checked to be homeomorphism. It follows that if U is a neighbourhood (nbh) of
0, then x + U is a nbh of x. Ultimately, the topology of X is uniquely determined by any
basis of neighbourhoods of 0 ∈ X.
Definition 2.3. Given a linear space X, a map p : X → [0, ∞) is called a seminorm on
X, if for all a ∈ C, x, y ∈ X one has the homogeneity p(ax) = |a|p(x) and the triangle
inequality p(x + y) ≤ p(x) + p(y).
Remark 2.4. It follows in the above situation that p(0) = 0, but p(x) = 0 need not imply
x = 0.
It is a bit tedious, but completely elementary, to check that if X is a linear space and
(pi )i∈I is a family of seminorms on X, then there exists a unique topology τ on X with the
following two properties
• (X, τ ) is a topological vector space,
• the family of sets A ⊂ X of the form
n
\
A= {pij < j }, n ∈ N, i1 , . . . , in ∈ I, 1 , . . . , n > 0,
j=1

forms a basis of neighbourhoods of 0 ∈ X.


This is in fact the coarsest topology on X such that pi is continuous for all i ∈ I.
Definition 2.5. A topological vector space X is called a locally convex space (LCS), if its
topology is induced by a family of seminorms in the above sense. Any such family is called
a defining family of seminorms.
Remark 2.6. The name ’locally convex’ stems from the fact that a topological vector
space X is locally convex, if and only if 0 ∈ X has a basis of neighbourhoods given by
convex sets. We are not going to use this fact so we ommit its proof.
The Hausdorffness of a LCS can be checked easily:
Lemma 2.7. Let X be a LCS. Then X is Hausdorff, if any only if there exists a defining
family (pi )i∈I of seminorms with the following property: for all x ∈ X \ {0} there exists
i ∈ I with pi (x) 6= 0.
Proof : ⇒: For all x 6= 0 there exists a nbh U of 0 such that x ∈
/ U . There exist n ∈ N,
i1 , . . . , in ∈ I, 1 , . . . , n > 0 with
n
\
U= {pij < j }.
j=1
Global Analysis 2 7

Since x ∈
/ U , there exists j ∈ {1, . . . , n} with pij (x) ≥ j > 0.
⇐: Pick any defining family (pi )i∈I . In this situation,
U := {pi < (1/2)pi (x)}
is a nbh of 0 which is disjoint from the nbh x + U of x. 
Remark 2.8. Assume X is a LCS which admits a countable defining family of seminorms.
Then one can easily construct a translation invariant pseudometric on X which induces
the original topology. Using the previous lemma one can furthermore easily check that
if X is a Hausdorff LCS which admits a countable defining family of seminorms, then X
carries a translation invariant metric which induces the original topology.
Any LCS admits a natural entourage (or: uniform) structure. In case you are not familiar
with these concepts, the following working definition is enough for our purposes:
Definition 2.9. 1. A net (xα )α∈A in a LCS X is called a Cauchy net, if there exists a
defining family (pi )i∈I of seminorms with the following property: for all i ∈ I, for all  > 0,
there exists α0 ∈ A such that for all α, β ≥ α0 one has pi (xα − xβ ) < .
2. A net (xα )α∈A in a LCS X is said to converge to x ∈ X, if there exists a defining family
(pi )i∈I of seminorms with the following property: for all i ∈ I, for all  > 0, there exists
α0 ∈ A such that for all α ≥ α0 one has p(x − xα ) < .
3. A LCS X is called complete, if every Cauchy net is convergent.
Lemma 2.10. Let p be a seminorm on a LCS X. Then the following statements are
equivalent:
i) p is continuous.

ii) p is continuous at 0.

iii) There exists a defining family (pi )i∈I of seminorms with the following property:
there exist r ∈ N, i1 , . . . ir ∈ I, C > 0, such that for all x ∈ X one has
(2) p(x) ≤ C max pij (x).
j=1,...,r

Proof : i) ⇒ ii): trivial.


ii) ⇒ i): Given  > 0 pick r ∈ N, i1 , . . . , ir ∈ I, 1 , . . . , r > 0 such that for all
r
\
(3) z ∈ U := {pij < j }
j=1

one has p(z) < . Fix an arbitrary x ∈ X. Then for all y ∈ x+U one has trivially y −x ∈ U
and by examining U one also finds x − y ∈ U , and so by the triangle inequality
p(x) ≤ p(x − y) + p(y) <  + p(y), p(y) ≤ p(y − x) + p(x) <  + p(x),
and finally |p(x) − p(y)| < . This shows that p is continuous in x.
8 B. GÜNEYSU

iii) ⇒ ii): If (xα )α∈A is a net which converges to 0, then by (2) we have
p(xα ) ≤ C max pij (xα ),
j=1,...,r

which converges to 0, as the pij ’s are continuous. Since 0 = p(0), this shows that p is
continuous at 0.
ii) ⇒ iii): Pick any defining family (pi )i∈I . Given  > 0 pick r ∈ N, i1 , . . . , ir ∈ I,
1 , . . . , r > 0 such that for all
r
\
z ∈ U := {pij < j }
j=1

one has p(z) < . Let x ∈ X be arbitrary and set Rx := maxj=1 pij (x) ∈ [0, ∞).
Case Rx = 0: Then we have x ∈ U and so p(x) < . Since  was arbitrary, this shows
p(x) = 0 and we are done.
Case Rx > 0: Then we have  
j
pij x < j /2 < j
2Rx

for all j ∈ {1, . . . , r}, and so 2Rjx x ∈ U and
 
2Rx j 2Rx
p(x) = p x < pij (x) := Cpij (x).
j 2Rx j

Definition 2.11. Let X be a vector space. Then two families of seminorms (pi )i∈I , (qj )j∈J
on X are called equivalent, if for all i ∈ I the seminorm pi is continuous with respect to
the topology induced by (qj )j∈J and if for all j ∈ J the seminorm qj is continuous with
respect to the topology induced by (pi )i∈J .
Remark 2.12. The equivalence i) ⇔ ii) in Lemma 2.10 easily shows that equivalent
families of seminorms induce the same topology. In particular, the family of all continuous
seminorms on a LCS X is a defining family of seminorms for X.
The natural way to turn LCS’s into a category is to pick continuous linear maps as mor-
phisms. The continuity of a linear map can be checked as follows in this context:
Proposition 2.13. Assume X and Y are LCS’s and that T : X → Y is a linear map.
Then the following statements are equivalent:
i) T is continuous at 0.
ii) T is continuous.
iii) There exist defining families of seminorms (pi )i∈I , (qj )j∈J for X and Y , respectively,
with the following property:
∀j ∈ J ∃r ∈ N ∃i1 , . . . , ir ∈ I ∃C > 0 ∀x ∈ X qj (T x) ≤ C max pij (x).
j=1,...,r

iv) For every continuous seminorm q on Y there exists a continuous seminorm q on


X with q ◦ T ≤ p.
Global Analysis 2 9

Proof : i) ⇒ ii): Let (xα )α∈A be a net which converges which converges to x ∈ X. Then
(x − xα )α∈A converges to 0 and so
(T xα )α∈A = (T x − T (x − xα ))α∈A converges to T x
ii) ⇒ iii): Pick any defining family (pi )i∈I on X, resp. (qj )j∈J in Y . For all j ∈ J the map
qj ◦ T is a continuous seminorm on X, and so the claim follows from the i) ⇒ iii) part of
Lemma 2.10.
iii) ⇒ i): If (xα )α∈A is a net which converges which converges 0 and if j ∈ J is arbitrary,
then by iii) we have that (qj (T xα ))α∈A converges to 0 and so (T xα )α∈A converges to 0.
iii) ⇒ iv): Lemma 2.10 and Remark 2.12.

For LCS’s X and Y we denote the linear space of continuous linear maps X → Y with
L (X, Y ). Specifically:
Definition 2.14. For a LCS X, the space of continuous linear functionals on X is defined5
to be X 0 := L (X, C).
In the above situation, X 0 is by default equipped with topology induced by the family of
seminorms T 7→ |T (x)|, where x runs through X. This is called the weak-*- topology on
X 0.
Proposition 2.13 immediately gives:
Corollary 2.15. Let T be a linear functional on a LCS X. Then the following assertions
are equivalent:
• T ∈ X 0.
• There exists a defining family of seminorms (pi )i∈I for X such that
∃r ∈ N ∃i1 , . . . , ir ∈ I ∃C > 0 ∀x ∈ X |T x| ≤ C max pij (x).
j=1,...,r

• There exists a continuous seminorm p on X such that |T | ≤ p.


One expects that X 0 should be a nontrivial space, in case the topology on X is not patho-
logical. This expectation is true, but its proof is somewhat obscure, for it relies on the
Hahn-Banach extension theorem (whose proof will not be given here; Zorn’s Lemma...):
Theorem 2.16 (Hahn-Banach extension theorem). Given a vector space X, a vector sub-
space M ⊂ X, a seminorm p on X, and a linear functional T0 on M with |T0 | ≤ p on M ,
there exists a linear functional T on X such that such T |M = T0 and |T | ≤ p.
Using this result we can prove:
Proposition 2.17. For every seminorm p on a vector space X and every x ∈ X with
p(x) 6= 0 there exists a linear functional T on X such that T (x) 6= 0 and |T | ≤ p.
5Here, C is considered as a LCS by equipping it with its Euclidean norm (keeping in mind that all
norms on C are equivalent.)
10 B. GÜNEYSU

Proof : Consider the linear functional T0 on span{x} given by T0 (ax) := ap(x). Then one
has |T0 | ≤ p on span{x}, so by the Hahn-Banach extension theorem there exists a linear
functional T on X with the asserted properties. 
Since a linear functional on a LCS is continuous, if and only if it is continuous on 0, we
immediately get:
Corollary 2.18. Assume there exists a nontrivial continuous seminorm on a LCS X.
Then one has X 0 6= {0}.
Lemma 2.7 in combination with Proposition 2.17 immediately give:
Corollary 2.19. Assume the LCS X is Hausdorff. Then for all x ∈ X \ {0} there exists
T ∈ X 0 with T (x) 6= 0.
Definition 2.20. Assume X and Y are LCS’s. Given T ∈ L (X, Y ), the transpose T 0 ∈
L (Y 0 , X 0 ) is defined by T 0 y 0 (x) := y 0 (T x).
Note that in the above situation T 0 is indeed continuous: we have |T 0 y 0 (x)| = |y 0 (T x)|, so
the claim follows immediately from the iii) ⇒ ii) part of Proposition 2.13 by examining
the definition of the underlying weak-*-topologies.
Let us come to come classes of LCS’s. Apart from normed spaces, the main examples of
LCS’s in our course will be Fréchet spaces (F) and limit Fréchet (LF) spaces:
Definition 2.21. A LCS X is called an F space, if X is complete, Hausdorff and admits
a countable defining family of seminorms.
Every Banach space is an F space, and any F space carries a translation invariant metric
which induces the original topology.
Assume that on a linear space X we are given an increasing sequence Xn ⊂ X, n ∈ N, of
linear subspaces which are all F spaces, such that
[
X= Xn ,
n∈N

and such that the inclusion map Xn → Xn+1 is continuous for all n. Call a seminorm p
on X admissible w.r.t. (Xn )n∈N , if p|Xn is continuous for all n. The family of all such
seminorms induces a LCS topology on X.
Definition 2.22. A LCS X is called an LF space, if there exists a sequence (Xn )n∈N as
above which induces the original topology on X. Such a sequence is called a defining
sequence (for the LF topology on X).
LF spaces need not be metrizable, however (as Dieudonne and Schwartz have shown in
1949) they are complete and Hausdorff:
Theorem 2.23. Let X be an LF space. Then:
a) X is Hausdorff and for any defining sequence (Xr )r∈N , the inclusion Xr → X is con-
tinuous.
Global Analysis 2 11

b) X is complete.
c) A sequence (xn )n∈N in X converges in the topology of X, if and only if there exists a
defining sequence (Xr )r∈N and r0 ∈ N such that
• xn ∈ Xr0 for all (large) n ∈ N,
• (xn )n∈N converges in the topology of Xr0 .
d) If Y is a LCS then for a linear map T : X → Y the following statements are equivalent6:
i) T is continuous.
ii) T is sequentially continuous.
iii) There exists a defining sequence (Xr )r∈N such that T |Xr : Xr → Y is continuous
for all r ∈ N.
Proof : The proof of part a) will be sketched in the exercises; the proof of b) and c) is
beyond the scope of this course (cf. Section 1.3 in [5]).
d) i) ⇒ ii): this is true for every map between topological spaces.
ii) ⇒ iii): Pick any defining sequence (Xr ). As Xr is metrizable, it suffices to show that
T |Xr is sequentially continuous. Any convergent sequence in Xr is by a) also convergent
T
in X, so the claim follows form observing that T |Xr is the composition Xr → X →− →Y.
iii) ⇒ i) Let q be a continuous seminorm on Y . Then q ◦ T is a seminorm on X and q ◦ T |Xr
is a continuous seminorm on Xr , so that p := q ◦ T is an admissible and thus continuous
seminorm on X with q ◦T ≤ p. This observation and Proposition 2.13 imply the continuity
of T . 
The following result is also due to Dieudonne and Schwartz (1949):
Theorem 2.24 (Open mapping theorem). Let X, Y be LF spaces and let T ∈ L (X, Y )
be surjective. Then T is open, that is, T maps open subsets to open subsets.
Proof : The proof in the LF case can be found in [11], p.78 (this result is again remarkable,
as X need not be metrizable). The proof in the F case is sufficiently close to the Banach
space case and will be an exercise. 
Corollary 2.25. Let X, Y be LF spaces and let T ∈ L (X, Y ) be bijective. Then T −1 is
automatically continuous.
A consequence of the latter corollary is the following variant of the closed graph theorem.
Note here that the product of two LCS’s with its canonically given structure of a topological
vector space is again a LCS (consider ’pi ⊗ qj ’), which is easily checked to be an F space,
if both factors are F spaces.
Theorem 2.26 (Closed graph theorem). Let X be an LF space and let Y be an F (!)
space, and let T : X → Y be linear. Then the following statements are equivalent:
i) T is continuous.
ii The graph G(T ) := {(x, y) : y = T x} ⊂ X × Y is a closed subset.
6The point of this result here is that X need not be metrizable.
12 B. GÜNEYSU

Proof : i) ⇒ ii): The graph of any continuous map from a topological space to a Hausdorff
space is closed: Given a net ((xα , yα ))α∈A ⊂ G(T ) which converges to (x, y) ∈ X × Y , we
have that (xα )α converges to x, (yα )α converges to y, yα = T xα . By continuity (T xα )α
converges to T x and by uniqueness of limits y = T x and so (x, y) ∈ G(T ).
ii) ⇒ i) For the general case see [14], p.173. If X and Y are F spaces, then G(T ) becomes
an F space (if X is only LF then G(T ) need not be LF), and with the projections
ΠX : X × Y −→ X, ΠY : X × Y −→ Y,
and the map ΠX |G(T ) is a linear continuous bijective map from an F space to an F space.
The previous corollary shows that (ΠX |G(T ) )−1 is continuous, too. It follows that
T = ΠY |G(T ) ◦ (ΠX |G(T ) )−1
is continuous.

Another central result is:
Theorem 2.27 (Uniform boundedness principle). Let X be an LF space and let Y be a
LCS. Suppose (Ti )i∈I is a family of continuous linear maps from X to Y with the following
property: for all x ∈ X and all continuous seminorms p0 on Y , there exists C > 0 such that
for all i ∈ I one has p0 (Ti x) ≤ C. Then for all continuous seminorms p on Y there exists a
continuous seminorm q on X such that for all i ∈ I and all x ∈ X one has p(Ti x) ≤ q(x).
Proof : The general case can be found on page 83 in [11]. We will give the proof for the
case that X is an F -space, which relies on the following (weak form of) Baire’s category
theorem (axiom of coice): assume a completely metrizable space is written as a countable
collection of closed subsets. Then at least one of these closed subsets has a nonemtpty
interior. This is applied as follows: Given any n ∈ N put
Xnp := {x ∈ X : for all i ∈ I one has p(Ti x) ≤ n}.
Since each Ti is continuous, Xnp is closed and the assumption on (Ti )i∈I implies
[
X= Xnp ,
n∈N

and as X is completely metrizable, we can conclude from Baire that there exist n0 ∈ N,
x0 ∈ X,  > 0, and a continuous seminorm q̃ on X with
x0 + {q̃ < } ⊂ Xnp0 .
Given an arbitrary x ∈ X we consider two cases:
Case q̃(x) = 0: Then for all λ > 0 we have q̃(λx) = 0 < , and so for all i ∈ I,
p(Ti λx) ≤ p(Ti (x0 + λx)) + p(Ti x0 ) ≤ 2n0 ,
and
p(Ti x) ≤ (2n0 )/λ, and finally p(Ti x) = 0.
Global Analysis 2 13

Case q̃(x) > 0: Then we have


 

q̃ x ≤ /2 < ,
2q̃(x)
and so with

x̃ := x,
2q̃(x)
for all i ∈ I we get
p (Ti x̃) ≤ p(Ti (x0 + x̃)) + p(Ti x0 ) ≤ 2n0 ,
and
2n0
p(Ti x) ≤ q̃(x).

Thus the proof is complete if we set
2n0
q := q̃.


The following result is a straightforward Corollary to the uniform boundedness principle.
Corollary 2.28. Let X,Y be LF spaces, and let
T : X × Y −→ C
be a bilinear and seperatly continuous map. Then T is already continuous.
Proof : In view of the bilinearity of T , it suffices to check the continuity of T at 0. Let
(xn , yn )n∈N ⊂ X × Y
be a sequence which converges to 0, and for every n ∈ N set
Tn : X −→ C, x 7−→ T (x, yn ),
which is a continuous map by assumption. For fixed x ∈ X we can pick C > 0 such that
|Tn (x)| ≤ C for all n ∈ N (note here that it is not true that convergent nets are bounded;
so we use Hausdorffness of X in an essential way). It follows from Lemma 2.10 that for all
continuous seminorms p on C one has p(Tn (x)) ≤ C, possibly after slightly modifying C
a little (we consider C as a normed space). By the principle of uniform boundedness, we
can thus pick a continuous seminorm q on X such that
|T (x, yn )| = |Tn (x)| ≤ q(x) for all x ∈ X, n ∈ N,
so
|T (xn , yn )| = |Tn (xn )| ≤ q(xn ) for all n ∈ N,
and limn q(xn ) = 0, showing limn T (xn , yn ) = 0. 
14 B. GÜNEYSU

3. Function spaces, distributions and the Fourier transform


We collect some functions spaces that are important examples of LCS’s and that will be
the central objects of the lecture course. We understand all our function spaces to be over
C.
In the sequel let Ω ⊂ Rm be an arbitrary open subset, and fix l ∈ N. Both Rm and Cl are
equipped with their standard Euclidean scalar products (·, ·) and the induced norms | · |.
Let e1 , . . . , el ∈ Cl denote the standard basis in Cl .
The Lebesgue measure on Ω is considered to be defined on the Borel sigma algebra of Ω,
rather than on the Lebesgue sigma algebra (although this makes no essential difference).
The Lebesgue integral of an appropriate function f : Ω → Cl is denoted with symbols such
as Z Z
f, f (x)dx,
Ω Ω
and ’a.e.’ will stand for ’Lebesgue almost everywhere’. The Lebesgue measure of a Borel
set A ⊂ Ω set will be denoted with
Z
|A| = 1 ∈ [0, ∞]
A
We record the following regularity property of the Lebesgue measure:
Remark 3.1. For every Borel set A ⊂ Ω with a finite Lebesgue measure and every  > 0
there exists an open subset U ⊂ Ω and a compact set K ⊂ Ω such that
K ⊂ A ⊂ U, |U \ K| < .
An index ’c’ in a function space such as Cc∞ (Ω) will stand for ’compactly supported’.
Open Euclidean balls will be denoted with
Br (x) = {y ∈ Rm : d(x, y) = |x − y| < r}, r > 0, x ∈ Rm .
The closed r-neighbourhood of a set A ⊂ Rm is defined by
(A)r := {x ∈ Rm : inf{d(x, a) : a ∈ A} ≤ r}, r > 0.
Example 3.2. Assume p ∈ [1, ∞]. Then the linear space Lp (Ω, Cl ) of all Lebesgue equiv-
alence classes of Borel maps f : Ω → Cl such that
(R
Ω Ω
|f |p , if p < ∞
kf kLp := < ∞,
inf{C ≥ 0 : |f | ≤ C a.e. in Ω}, if p = ∞

becomes a Banach space (thus an F space) with respect to k·kΩ Lp . The asserted completeness
is a classical result by Riesz and Fischer, which is in fact true on every measure space (and
which is proved in every course on Lebesgue integration theory). For p = 2 the space
Lp (Ω, Cl ) is in fact a Hilbert space w.r.t.
Z

hf1 , f1 i := (f1 , f2 ).

Global Analysis 2 15

Example 3.3. For k ∈ N≥0 the linear space


Cbk (Ω, Cl )
n m
X o
k l α m
:= f ∈ C (Ω, C ) : ∂ f is bounded for all α ∈ N0 with |α| := αi ≤ k
i=1
becomes a Banach space with respect to the norm
X
kf kΩ
∞,k := sup |∂ α f (x)|.
x∈Ω
|α|≤k

Indeed, given a Cauchy sequence (fn )n∈N , α ∈ Nm 0 with |α| ≤ k, it follows immediately
that for all x ∈ Ω the sequence (∂ fn (x))n is Cauchy in Cl and thus has a limit which
α

we denote by Fα (x), where we set f (x) := F(0,...,0) (x). Then using the triangle inequality
(and an /2 argument) one easily checks that the function x 7→ f (x) lies in Cbk (Ω, Cl ) wit
Fα = ∂ α f and
lim kfn − f kΩ∞,k = 0.
n
Another important space is the Banach space C0k (Ω, Cl ), which is defined as the closure of
Cck (Ω, Cl ) with respect to the norm k·kΩ
∞,k .

Example 3.4. The space C ∞ (Ω, Cl ) becomes a locally convex space with respect to the
family of seminorms
(4) pΩ α
α,K (f ) := sup |∂ f (x)|, α ∈ Nm
0 , K ⊂ Ω compact.
x∈K

This locally convex space (that is, the pair given by C ∞ (Ω, Cl ) and the above topology)
plays a distinguished role and is usually
denoted by E (Ω, Cl ).
This is in fact an F space: Every sequence (Kr )r∈N of compact subsets of Ω with
[
(5) Kr = Ω, K1 ⊂ K̊2 ⊂ K2 ⊂ K̊3 ⊂ · · ·
r∈N

induces the countable family of (semi)norms pΩ m


α,Kr , α ∈ N0 , r ∈ N, which is equivalent
to (4). The Hausdorff property of (pΩ α,Kr )α,r is obvious in view of Lemma 2.7, and the
completeness can be seen much as in the previous example: If (fn )n∈N is a Cauchy sequence
with respect to (pΩ m
α,Kr )α,r , then certainly for all x ∈ Ω and all α ∈ N0 the sequence
(∂ α fn (x))n is a Cauchy sequence in Cl thus has a limit which we denote with Fα (x), where
f (x) := F(0,...,0) (x). In fact, x 7→ f (x) is in C ∞ (Ω, Cl ) with Fα = ∂ α f (x) and one has
lim pΩ
α,Kr (fn − f ) = 0 for all α ∈ Nm
0 , r ∈ N.
n
As above, this follows from the subadditivity

pΩ Ω Ω l
α,Kr (ψ + φ) ≤ pα,Kr (ψ) + pα,Kr (φ) for all ψ, φ ∈ C (Ω, C )

and an /2 argument.


16 B. GÜNEYSU

In a complete analogy to the smooth case one can consider C k -functions:


Example 3.5. Given k ∈ Nk≥0 the space C k (Ω, Cl ) becomes a locally convex space with
respect to the family of seminorms pΩ m
α,K , where α ∈ N0 satisfies |α| ≤ k and K ⊂ Ω is
compact. As above one sees that this is in fact an F space.
Example 3.6. Given a compact subset K ⊂ Ω set
DK (Ω, Cl ) := f ∈ C ∞ (Ω, Cl ) : supp(f ) ⊂ K .


Then DK (Ω, Cl ) is a F space with respect to the above defined seminorms (pΩ
α,K )α∈Nm
0
, and
if K ⊂ K 0 are compact subsets of Ω, then the natural linear injective map
DK (Ω, Cl ) ,→ DK 0 (Ω, Cl )
is continuous. Note that
[
Cc∞ (Ω, Cl ) = DK (Ω, Cl ).
K ⊂ Ω compact

We equip Cc∞ (Ω, Cl ) with the family of seminorms p on Cc∞ (Ω, Cl ) such that p|DK (Ω,Cl ) is
continuous for all compact K ⊂ Ω. Then Cc∞ (Ω, Cl ) together with this family of seminorms
is denoted by D(Ω, Cl ) and called the space of test functions on Ω. The space D(Ω, Cl )
becomes an LF space: in fact, for every sequence (Kr )r∈N of compact subsets of Ω with
[
(6) Kr = Ω, K1 ⊂ K̊2 ⊂ K2 ⊂ K̊3 ⊂ · · · ,
r∈N

the topology on D(Ω, Cl ) is induced by the sequence of F spaces (DKr (Ω, Cl ))r∈N .
A sequence (fn )n∈N in D(Ω, Cl ) converges to f ∈ D(Ω, Cl ), if and only if there exists a
compact set K ⊂ Ω such that
• supp(fn ) ⊂ K for all n ∈ N,
• ∂ α fn converges to ∂ α f uniformly as n → ∞ for all α ∈ N.
Again there is a C k -variant:
Example 3.7. Given k ∈ N≥0 and a compact subset K ⊂ Ω set
k
(Ω, Cl ) := f ∈ C k (Ω, Cl ) : supp(f ) ⊂ K .

CK
Then CKk
(Ω, Cl ) is a F space with respect to the seminorms (pΩ
α,K )α∈Nm
0 ,|α|≤k
, and if K ⊂ K 0
are compact subsets of Ω, then the natural linear injective
k
CK (Ω, Cl ) ,→ CK
k l
0 (Ω, C )

is continuous. Note that as above


[
Cck (Ω, Cl ) = k
CK (Ω, Cl ).
K ⊂ Ω compact

One equips Cck (Ω, Cl )


with the family of seminorms p on Cck (Ω, Cl ) such that p|CKk (Ω,Cl ) is
continuous for all compact K ⊂ Ω. Then Cck (Ω, Cl ) together with this family of seminorms
becomes an LF space, which checked with the same arguments as in the smooth case.
Global Analysis 2 17

Example 3.8. The space of Schwartz test functions or tempered test functions S (Rm , Cl )
is the linear space defined by
n o
S (Rm , Cl ) := f ∈ C ∞ (Rm , Cl ) : sup |xα ∂ β f (x)| < ∞ for all α, beta ∈ Nm
0 .
x∈Rm

It is an F space with respect to the family of seminorms given by


pα,β (f ) := sup |xα ∂ β f (x)|, α, β ∈ Nm
0 .
x∈Rm

The Hausdorff property and the completeness will be proved as an exercise. The Schwartz
space is the natural space on which the Fourier transform is a priori defined (later!).
Example 3.9. Assume p ∈ [1, ∞]. Then the linear space Lploc (Ω, Cl ) of all Lebesgue
equivalence classes of Borel maps f : Ω → Cl with7
k1K f kΩ
Lp < ∞ for all compact K ⊂ Ω,
becomes a locally convex space with respect to the family of seminorms
k1K ·kΩ
Lp , K ⊂ Ω compact,
called the space of locally Lp -integrable ’functions’ on Ω. It is in fact a F space, which is
checked by picking a compct exhaustion on Ω as above.
Example 3.10. Assume p ∈ [1, ∞]. Define the linear space
Lpc (Ω, Cl ) := f ∈ Lp (Ω, Cl ) : f has a compactly supported Lebesgue representative


of compactly supported Lp -’functions’ on Ω. Given K ⊂ Ω compact define


LpK (Ω, Cl ) :=
f ∈ Lp (Ω, Cl ) : f has a Lebesgue representative which is supported in K .


Then LpK (Ω, Cl ) is a Banach space with respect to k1K ·kΩ p l


Lp and we turn Lc (Ω, C ) into a
locally convex space with the family of seminorms p on Lpc (Ω, Cl ) such that p|LpK (Ω,Cl ) is
continuous for all compact K ⊂ Ω. It is then easily checked (with arguments as above)
that Lpc (Ω, Cl ) is in fact an LF space.
Remark 3.11. In the sequel, whenever there is no danger of confusion, we will omit ’Ω’
m
in the notation for the various (semi-)norms, so k·kLp := k·kRLp etc. Furthermore, if l = 1
we are going to omit Cl in the notation for the function spaces, so Lp (Ω) := Lp (Ω, C) etc.
Our next aim is to determine nice dense subspaces of these spaces. The key tool in this
context is a construction that goes back to K. Friedrichs (1940!), usually referred to as
Friedrichs mollifiers. The essential idea is to convolute a ’bad function’ with a nice function
(smooth and compactly supported). To this end, we are going to need precise Lp -estimates
7that with 1A the indicator function of a subset A
18 B. GÜNEYSU

for convolutions.
Given Borel functions f : Rm → C, g : Rm → Cl with
Z
|f (y)g(x − y)|dy < ∞ for every or (a.e.) x ∈ Rm ,
Rm
we define their convolution to be the function
Z
m l
f ∗ g : R −→ C , f ∗ g(x) := f (y)g(x − y)dy.
Rm
Then from the transformation formula for Lebesgue integrals we have
Z
f ∗ g(x) = f (x − y)g(y)dy.
Rm

Proposition 3.12. Let p ∈ [1, ∞], let f ∈ L1 (Rm ), g ∈ Lp (Rm , Cl ). Then f ∗ g(x) is
well-defined for a.e. x ∈ Rm , and one has the estimate
kf ∗ gkLp ≤ kf kL1 kgkLp .
Proof : Case p = ∞: Obviously
Z
|f ∗ g(x)| ≤ |f (y)||g(x − y)|dy ≤ kf kL1 kgkL∞ .
Rm
Case p < ∞: We are going to prove that for all x ∈ Rm one has
Z Z p
|f (y)||g(x − y)|dy dx ≤ kf kpL1 kgkpLp ,
Rm Rm

which also shows a posteriori that f ∗ g is well-defined a.e. in Rm (if the integral of
nonnegative function taking a priori values in [0, ∞] is finite, then the function must be
finite a.e.).
Case p = 1: One has
Z Z
|f (y)||g(x − y)|dydx = kf kpL1 kgkpL1
Rm Rm
by Fubini-Tonelli.
Case 1 < p < ∞: Set q := p/(p − 1), so 1/q + 1/p = 1. Then by Hoelder’s inequality
Z p Z p
1/p 1/q
|f (y)||g(x − y)|dy = |f (x − y)| |g(y)||f (x − y)| dy
Rm Rm
Z  Z p/q
p
≤ |f (x − y)||g(y)| dy |f (y)|dy ,
Rm Rm
so that Z Z p
1+p/q
|f (y)||g(x − y)|dy dx ≤ kf kL1 kgkpLp ,
Rm Rm
completing the proof in view of 1 + p/q = p.

Global Analysis 2 19

More generally, one can prove Young’s inequality, which states that for all p, q, r ∈ [1, ∞]
with 1/p + 1/q = 1/r + 1 one has
kf ∗ gkLr ≤ kf kLq kgkLp ,
whenever the right hand side is finite. We will not need the latter generalization, which
also holds on general unimodular groups, if one replaces x − y with y −1 x and the Lebesgue
measure with the (up to a constant uniquely determined) Haar measure. If p, q, r ∈ (1, ∞),
then one can even prove an inequality of the form
kf ∗ gkLr ≤ cp,q kf kLq kgkLp ,
where cp,q < 1. This is a highly nontrivial result by Brascamp and Lieb from 1976.
Definition 3.13. Let 0 ≤ % ∈ Cc∞ (Rm ) be such that
• supp(%)
R ⊂ B1 (0).
• Rm % = 1.
For all λ > 0 define 0 ≤ %λ ∈ Cc∞ (Rm ) by the scaling %λ (x) := λ−m %(x/λ). Then for all
p ∈ [1, ∞), all f ∈ Lploc (Rm , Cl ), the net of functions
(%λ ∗ f )λ>0
is called the Friedrichs mollification of f (with respect8 to %).
Note that the definition of %λ is justified by the factRthat the support of this function is
controlled according to supp(%λ ) ⊂ Bλ (0), while still Rm %λ = 1.
Remark 3.14. In the above sitution, the following assertions hold true for all λ > 0:
i) One has %λ ∗ f ∈ C ∞ (Rm , C, ). This follows from
Z
%λ ∗ f (x) = %λ (x − y)f (y)dy
Rm
and differentiating under the integral (make yourself clear that the latter is possible!). In
|α|
fact, for all α ∈ Nm α α m l
0 one has ∂ (%λ ∗ f ) = (∂ %λ ) ∗ f , and if f ∈ C (R , C ) then the latter
is also equal to %λ ∗ (∂ α f ), which follows
 from integrating by parts.
ii) One has supp(%λ ∗ f ) ⊂ supp(f ) λ .
A first application of Friedrichs mollifiers is provided by the construction of nice cut-off
functions:
Proposition 3.15. Let K ⊂ Ω be compact. Then there exists a function χ ∈ Cc∞ (Ω) with
0 ≤ χ ≤ 1 and χ = 1 in an open neighbourhood of K.
Proof : Pick open and relatively compact subsets V1 , V2 ⊂ Ω (meaning that the closure of
Vj is a compact subset of Ω) which satisfy K ⊂ V1 ⊂ V1 ⊂ V2 . Then for all λ > 0 one has
%λ ∗ 1V2 ∈ C ∞ (Rm ), and clearly
Z Z
0 ≤ %λ ∗ 1V2 (x) = %λ (x − y)dy ≤ %λ (y)dy = 1 for all x ∈ Rm .
V2 Rm

8The particular choice of % with the above properties will play no role in the sequel.
20 B. GÜNEYSU

If λ is sufficiently small (depending on V2 ), the function %λ ∗ 1V2 is compactly supported


in Ω. Moreover, if λ is sufficiently small (depending on Vj and K) we have Bλ (x) ⊂ V2 for
all x ∈ V1 and so 1V2 |Bλ (x) = 1 for all x ∈ V1 . Thus for such x’s we have
Z Z
%λ ∗ 1V2 (x) = %λ (x − y)1V2 (y)dy = %λ (x − y)dy = 1.
Rm Bλ (x)

Thus we can pick some λK,Vj > 0 such that χ := (%λ ∗ 1V2 )|Ω does the job for all 0 < λ <
λK,Vj . 
All density results can be derived from the following two auxiliary results:
Proposition 3.16. For all f ∈ C(Rm , Cl ) one has %λ ∗ f → f as λ → 0+ in the topology
of C(Rm , Cl ).
Proof : Given an arbitrary compact subset K ⊂ we have to show that %λ ∗ f → f uniformly
on K. For all λ < 1 and all x ∈ Rm we have
Z

|%λ ∗ f (x) − f (x)| = %λ (y)(f (x − y) − f (x))dy ≤ sup |f (x − y) − f (y)| ,
Bλ (0) y∈Bλ (0)

where we have used Z


%λ (y)dy = 1.
Bλ(0)
Thus,
sup |%λ ∗ f (x) − f (x)| ≤ sup |f (x − y) − f (x)| for all x ∈ K, λ > 0,
x∈K x∈K,y∈Bλ (0)

which goes to zero as λ → 0+, as f is locally uniformly continuous. 


Proposition 3.17. For all p ∈ [1, ∞), f ∈ Lp (Rm , Cl ) one has %λ ∗ f → f as λ → 0+ in
the topology of Lp (Rm , Cl ).
Proof : We can assume l = 1. Note first that indeed %λ ∗ f ∈ Lp (Rm ) for all λ > 0, which
follows e.g. from Lemma 3.12.
We first show that Cc∞ (Rm ) is dense. To this end, as simple functions of the form rj=1 cj 1Aj
P
with cj ∈ C and Aj Borel with finite Lebesgue measure are dense, given a Borel set A ⊂ Rm
with finite Lebesgue measure and  > 0 we need to find h ∈ Cc∞ (Rm ) with
kh − 1A kpLp < .
By the regularity of the Lebesgue measure we can pick U ⊂ Rm open and K ⊂ Rm compact
with
K ⊂ A ⊂ U, |U \ K| < .
By Proposition 3.15 we can find h ∈ Cc∞ (U ) ⊂ Cc∞ (Rm ) with 0 ≤ h ≤ 1 in U and h = 1 in
K. Thus one has h = 1 = 1A in K, h = 0 = 1A in Rm \ U , and |h − 1A | ≤ 1 in U \ K, so
that Z
p
kh − 1A kLp = |h − 1A |p ≤ |U \ K| < .
U \K
Global Analysis 2 21

Now given f ∈ Lp (Rm ),  > 0, pick h ∈ Cc∞ (Rm ) ⊂ Cc (Rm ) with kf − hkpLp < /3. Then
we can write
(7) kf − %λ ∗ f kpLp ≤ kf − hkpLp + kh − %λ ∗ hkpLp + k%λ ∗ (h − f )kpLp
(8) ≤ kf − hkpLp + kh − %λ ∗ hkpLp + kh − f kpLp
(9) < /3 + /3 + /3,
where we have used the Young-inequality, and that kh − %λ ∗ hkpLp can be made < /3
for small λ by Proposition 3.16 (as this integral is over a compact and there uniform
convergence implies Lp -convergence. 
Remark 3.18. Proposition 3.17 fails for p = ∞; in fact, one can show that smooth
L∞ (Rm , Cl ) functions are not dense in L∞ (Rm , Cl ) (exercise).
The latter two results remain true on open subsets of Rm : To this end, given a function
f : Ω → Cl we denote by f : Rm → Cl its extension by 0 to Rm .
Proposition 3.19. For all h ∈ C(Ω, Cl ) one has (%λ ∗ h)|Ω → h as λ → 0+ in the topology
of C(Ω, Cl ).
Proof : Given a compact set K ⊂ Ω pick a continuous function ψ : Rm → Cl which coincides
with h in an open neighboorhood V ⊂ Ω of K. For example you can use Proposition 3.15
to do that, or simply use Urysohn’s lemma. Then for all sufficiently small λ > 0 (depending
on K and V ) we have Bλ (x) ⊂ V for all x ∈ K, thus
Z
%λ ∗ h(x) = %λ (x − y)h(y)dy = %λ ∗ ψ(x),
Bλ (x)

which tends to h uniformly in K as λ → 0+ by Proposition 3.16.



Proposition 3.20. For all p ∈ [1, ∞), h ∈ Lp (Ω, Cl ) one has (%λ ∗ h)|Ω → h as λ → 0+
in the topology of Lp (Ω, Cl ).
Proof : Apply Proposition 3.17 with f = h. 
We collect all possible convergence results for Friedrichs mollifiers in the following Theorem:
Theorem 3.21. i) For all p ∈ [1, ∞), f ∈ Lp (Ω, Cl ) one has (%λ ∗ f )|Ω → f as λ → 0+
in the topology of Lp (Ω, Cl ).
ii) For all p ∈ [1, ∞), f ∈ Lploc (Ω, Cl ) one has (%λ ∗ f )|Ω → f as λ → 0+ in the topology of
Lploc (Ω, Cl ).
iii) For all p ∈ [1, ∞), f ∈ Lpc (Ω, Cl ) one has (%λ ∗ f )|Ω → f as λ → 0+ in the topology of
Lpc (Ω, Cl ).
iv) For all k ∈ N≥0 , f ∈ Cck (Ω, Cl ) one has (%λ ∗ f )|Ω → f as λ → 0+ in the topology of
Cck (Ω, Cl ).
v) For all f ∈ D(Ω, Cl ) one has (%λ ∗ f )|Ω → f as λ → 0+ in the topology of D(Ω, Cl ).
vi) For all k ∈ N≥0 , f ∈ C k (Ω, Cl ) one has (%λ ∗ f )|Ω → f as λ → 0+ in the topology of
22 B. GÜNEYSU

C k (Ω, Cl ).
vii) For all f ∈ E (Ω, Cl ) one has (%λ ∗ f )|Ω → f as λ → 0+ in the topology of E (Ω, Cl ).
viii) For all f ∈ S (Rm , Cl ) one has (%λ ∗ f ) → f as λ → 0+ in the topology of S (Rm , Cl ).
Proof : Exercise. 

As a consequence we get that Cc∞ is dense in any of the above space:


Theorem 3.22. Cc∞ (Ω, Cl ) is dense in...
i) Lp (Ω, Cl ) for all p ∈ [1, ∞),
ii) Lploc (Ω, Cl ) for all p ∈ [1, ∞),
iii) Lpc (Ω, Cl ) for all p ∈ [1, ∞),
iv) Cck (Ω, Cl ) for all k ∈ N≥0 ,
v) C k (Ω, Cl ) for all k ∈ N≥0 ,
vi) E (Ω, Cl ),
vii) S (Rm , Cl ), if Ω = Rm .
Proof : i) Clearly Lpc (Ω, Cl ) is dense in Lp (Ω, Cl ) (approximate f ∈ Lp (Ω, Cl ) with 1Kn f
using dominated convergence, where (Kn ) is a compact exhaustion of Ω), but the elements
of Lpc (Ω, Cl ) can be approximated with Friedrichs mollifiers, which now are compactly
supported.
ii) By the same argument as in i) one finds that Lpc (Ω, Cl ) is dense in Lploc (Ω, Cl ), but the
latter functions can be approximated by Friedrichs mollifiers. Noting that the support of
the Friedrichs mollification of a compactly supported function gets smaller with λ getting
smaller, this also proves iii).
iii) This immediately from Friedrichs mollifiers.
iv) This follows immediately using Friedrichs mollifiers, using again that the support of
the Friedrichs mollification of a compactly supported function gets smaller with λ getting
smaller.
v) Note first that Cck (Ω, Cl ) is dense in C k (Ω, Cl ) (approximate f ∈ C k (Ω, Cl ) with fn (x) =
φ(x/n)f (x) where φ ∈ Cc∞ (Ω)). Then one can use Friedrichs mollifiers.
vi) This follows from applying v) order by order.
vii) Approximate f ∈ S (Ω, Cl ) with fn (x) = φ(x/n)f (x) where φ ∈ D(Rm ). 

The following result is usually referred to as the fundamental lemma of distribution theory:
Lemma 3.23. Assume f ∈ L1loc (Ω, Cl ) and that
Z
(f, Ψ) = 0 for all Ψ ∈ Cc∞ (Ω, Cl ).

Then one has f = 0 (a.e. in Ω).


Proof : We are going to prove
(10) φf (i) = 0 a.e. in Ω for all φ ∈ Cc∞ (Ω), j = 1, . . . , l.
Global Analysis 2 23

To see this, note that for all ψ ∈ Cc∞ (Rm ) we have φψ|Ω ∈ Cc∞ (Ω), and so by using the
assumption with Ψ = φψ|Ω ej we get the second equality in
Z Z
(j)

ψ · φf = f, φψ|Ω ej = 0,
Rm Ω
where for the first equality we have used the support properties and assumed that ψ is
real-valued. Applying this with ψ = %1/n (x − ·) gives
Z
(j)
%1/n ∗ (φf )(x) = %1/n (x − y)φf (j) (y)dy = 0 for all n ∈ N,
Rm

which implies φf (j) = 0 a.e. in Rm and so (10), as


φf (j) = lim %1/n ∗ (φf (j) ) in L1 (Rm , Cl ),
n

which completes the proof. 


Let us now come to spaces of distributions. In the sequel, we call that the space of
continuous linear functionals X 0 on a locally convex space X is always equipped with its
weak-∗-topology.
Definition 3.24. a) D 0 (Ω, Cl ) := D(Ω, Cl )0 is called the space of distributions on Ω.
b) E 0 (Ω, Cl ) := E (Ω, Cl )0 is called the space of distributions with compact support 9.
c) S 0 (Rm , C) := S (Rm , C)0 is called the space of tempered distributions or the space of
Schwartz distributions.
Lemma 3.25. The inclusion maps
D(Ω, Cl ) ,→ E (Ω, Cl ), D(Rm , Cl ) ,→ S (Rm , Cl ) ,→ E (Rm , Cl )
are continuous.
b) The restriction maps
E 0 (Ω, Cl ) −→ D 0 (Ω, Cl ), E 0 (Rm , Cl ) −→ S 0 (Rm , Cl ) −→ D 0 (Rm , Cl )
are continuous.
Proof : a) Convergence in D (support in a fixed compact set and local uniform conver-
gence of all derivatives thereon) implies convergence in E (local uniform convergence of
all derivatives) and in S (global uniform convergence of all derivatives multiplied with
polynomials). Likewise, convergence in S implies convergence in E .
b) Note first that D in dense in any of these spaces (cf. Theorem 3.22). Now the claim
follows from the following simple observation: If X and Y are LCS’s such that there is an
continuous linear map ι : X ,→ Y with a dense image, then the map
r : Y 0 −→ X 0 , r(T )(φ) := T (ι(φ))
is a continuous linear map. 
9This notion will become clear later.
24 B. GÜNEYSU

Our abstract continuity results for linear maps between locally convex spaces immediately
imply:
Lemma 3.26. a) A linear functional T : D(Ω, Cl ) → C is a distribution, if and only if
for every compact set K ⊂ Ω there exist N ∈ N≥0 , C > 0, such that
|T (φ)| ≤ C max pΩ
α,K (φ) for all φ ∈ DK (Ω, Cl ).
|α|≤N

b) A linear functional T : E (Ω, Cl ) → C is a distribution with compact support, if and only


if there exist a compact set K ⊂ Ω, N ∈ N≥0 , C > 0, such that
|T (φ)| ≤ C max pΩ
α,K (φ) for all φ ∈ E (Ω, Cl ).
|α|≤N

c) A linear functional T : S (Rm , Cl ) → C is a tempered distribution, if and only if there


exist α, β ∈ Nm , C > 0, such that
|T (φ)| ≤ Cpα,β (φ) for all φ ∈ S (Ω, Cl ).
We continue with some important examples:
Example 3.27. Every f ∈ L1loc (Ω, Cl ) induces the distribution Tf ∈ D 0 (Ω, Cl ) given by
Z
Tf (φ) := (f, φ), φ ∈ D(Ω, Cl ).

Indeed, for all K ⊂ Ω compact and all φ ∈ DK (Ω, Cl ) one has


Z 
(11) |Tf (φ)| ≤ |f | sup |φ| =: CK,f pΩ
(0,...,0),K (φ)
K K

The assignment f 7→ Tf induces a continuous antilinear injective map


L1loc (Ω, Cl ) ,→ D 0 (Ω, Cl ).
Indeed, the injectivity follows from the fundamental lemma of distribution theory, and the
continuity follows from
|Tf (φ)| ≤ CK,φ k1K f kΩ
L1 ,
which is implied by (11). Distributions that arise from locally integrable functions in the
above way are called regular distributions.
In a complete analogy, one gets continuous antilinear injective maps
Z
Lc (Ω, C ) ,→ E (Ω, C ), f 7→ Tf := (f, ·)
1 l 0 l

and Z
L (R , C ) ,→ S (R , C ),
p m l 0 m l
f 7→ Tf := (f, ·) for all p ∈ [1, ∞].
Rm

The prototype of a distribution which is not regular is given by the δ-distribution:


Global Analysis 2 25

Example 3.28. For every fixed a ∈ Ω put


δa : E (Ω) −→ C, φ 7−→ φ(a).
Then δa is a distribution with compact support, called the δ-distribution with mass in a.
Indeed, this follows from the inequality
|δa (φ)| = |φ(a)| ≤ pΩ
(0...,0),{a} (φ).

In particular (by restriction) one has δa ∈ D 0 (Ω) and δa ∈ S 0 (Rm ). Let us show that δa
is not a regular distribution: for if δa = Tf ∈ D 0 (Ω) for some f ∈ L1loc (Ω), then for all
φ ∈ D(Ω) with support in Ω \ {a} one has
Z Z
0 = δa (φ) = Tf (φ) = fφ = f φ,
Ω Ω\{a}

which implies f = 0 a.e. in Ω \ {a} by the fundamental lemma of distribution theory and
thus f = 0 a.e in Ω. But this shows δa (φ) = Tf (φ) = 0 for all φ ∈ D(Ω), which obviously
cannot be true.
Althoug the δ-distribution is not regular, it can be approximated by regular distributions:
Example 3.29. One has
lim T%λ (a−·) = δa in D 0 (Ω).
λ→0+

Indeed, by definition one has to show that


Z
lim T%λ (a−·) (φ) = lim %λ (a − y)φ(y)dy = φ(a) for all φ ∈ D(Ω),
λ→0+ λ→0+ Ω

which follows from Theorem 3.21 v).


Although x 7→ e−i(x,a) is not in L1 (Rm ), one can use distribution theory to give meaning
to Rm e−i(x,a) dx by considering b as a variable. More precisely:
R

Proposition 3.30. Fix a ∈ Rm . Then for the net of bounded functions


Z
m
R 3 y 7−→ e−i(x,a−y) dx ∈ C, R > 0,
BR (0)

the limit
lim TRB e−i(x,a−•)dx exists in S 0 (Rm ) (and thus in D 0 (Rm )).
R→∞ R (0)

In fact, the limit is proportional to δa .


Proof : Exercise. 
Distribution theory also allows to differentiate nondifferentiable functions in a consistent
’weak’ sense. To this end, we add:
26 B. GÜNEYSU

Lemma 3.31. For all k ∈ N≥0 and all differential operators


0
P : E (Ω, Cl ) −→ E (Ω, Cl ) of order ≤ k
there exists a unique differential operator
0
P † : E (Ω, Cl ) −→ E (Ω, Cl ) of order ≤ k,
the formal adjoint of P , such that
Z Z
0
(12) (P φ, ψ) = (φ, P † ψ) for all φ ∈ E (Ω, Cl ), ψ ∈ E (Ω, Cl ),
Ω Ω

at least one of which having a compact support.


Proof : Uniqueness
Pfollows from the fundamental lemma of distribution theory. For the
α
existence, if P = |α|≤k Pα ∂ , one may set
X
P † ψ := (−1)|α| ∂ α (Pα† ψ),
|α|≤k

where (Pα† )ij (x) = (Pα )ji (x) denotes the adjoint matrix for x ∈ Ω. By the Leibnitz rule,
P † indeed is a partial differential operator of order ≤ k. The equality (12) then follows
from integrating by parts. 
In particular, the formal adjoint of ∂ α is (−1)|α| ∂ α . Using the formal adjoint, we can give:
Definition 3.32. Let
0
P : E (Ω, Cl ) −→ E (Ω, Cl )
be a differential operator.
a) Given T ∈ D 0 (Ω, Cl ) we set
0
P T (φ) := T (P † φ) for all φ ∈ D(Ω, Cl ).
b) Given T ∈ E 0 (Ω, Cl ) we set
0
P T (φ) := T (P † φ) for all φ ∈ E (Ω, Cl ).
c) Given T ∈ S 0 (Rm , Cl ) and if P = |α|≤k Pα ∂ α is such that (Pαij ) ∈ S (Rm ) for all α
P

with |α| ≤ k and all i = 1, . . . , l, j = 1, . . . , l0 , we again set


0
P T (φ) := T (P † φ) for all φ ∈ S (Rm , Cl ).
Note that these definitions are consistent with the antilinear injective maps
(13) E (Ω, Cl ) ,→ D 0 (Ω, Cl )
(14) D(Ω, Cl ) ,→ E 0 (Ω, Cl ),
(15) S (Rm , Cl ) ,→ S 0 (Rm , Cl ),
Global Analysis 2 27

In the sense that in each case one has P Tf = TP f . This follows from integrating by parts.
Moreover, the linear maps
0
P : D(Ω, Cl ) −→ D(Ω, Cl ),
0
P : E (Ω, Cl ) −→ E (Ω, Cl ),
0
P : S (Rm , Cl ) −→ S (Rm , Cl )
are easily checked to be continuous (where in the Schwartz case we assume Schwartz
coefficients!), which implies that so are the induced maps
0
P : D 0 (Ω, Cl ) −→ D 0 (Ω, Cl ),
0
P : E 0 (Ω, Cl ) −→ E 0 (Ω, Cl ),
0
P : S 0 (Rm , Cl ) −→ S 0 (Rm , Cl ),
since these maps are just the transposed maps in the sense of LCS’s of continuous maps.
Example 3.33. 1. The derivative of the distribution TH ∈ D 0 (R) which is induced by
the locally integrable function H := 1[0,∞) ∈ L1loc (Rm ), the so called Heaviside function, is
given by ∂TH = δ0 , where ∂ = d/dx. Indeed, for all φ ∈ D(R) one has
Z
∂TH (φ) = −TH (∂φ) = − φ0 = φ(0).
[0,∞)

2. The derivative of the distribution Tf ∈ D 0 (−1, 1) which is induced by the locally


integrable function f (x) := log(x) for x > 0, f (x) := 0 for x ≤ 0 is given (exercise) by
Z 1
φ(x) − φ(0)
∂Tf (φ) = dx for all φ ∈ D(−1, 1).
0 x
3. The partial derivatives of δa ∈ E 0 (Ω) are given by the compactly supported distribution
φ 7→ (−1)|α| ∂ α φ(a), for we have
∂ α δa (φ) = δa ((−1)|α| ∂ α φ) = (−1)|α| ∂ α φ(a) for all φ ∈ E (Ω).
The first two examples shows that the derivatives of regular distributions that are induced
by nondifferentiable functions need not be regular.
Let us now come to the Fourier transform:
Definition 3.34. Given f ∈ S (Rm , Cl ) its Fourier transform fb : Rm → Cl is the function
defined by Z
fb(ζ) := (2π)−m/2 e−i(x,ζ) f (x)dx.
Rm

The Fourier transform of f is well-defined since S (Rm , Cl ) ⊂ L1 (Rm , Cl ), and in fact


a smooth function (differentiate under the integral). Moreover, this operation switches
multiplication by monomials to differentiation and vice versa. More precisely:
28 B. GÜNEYSU

Proposition 3.35. Let f ∈ S (Rm , Cl ), ζ ∈ Rm .


a) One has Dα fb(ζ) = xd α f (ζ), where D α := (−i)|α| ∂ α , and xα (ζ) := ζ α .

b) One has ζ α fb(ζ) = D


d α f (ζ).

c) The linear map


F : S (Rm , Cl ) −→ S (Rm , Cl ), f 7−→ fb
is well-defined and continuous.
d) For all λ ∈ R \ {0}, v ∈ Rm , f ∈ S (Rm , Cl ) one has the following scaling/translation
behaviour:
−m b ·
 
[
f (λ·)(ζ) = λ f (ζ), f \(· − v)(ζ) = e−i(v,ζ) fb(ζ).
λ
Proof : a) Differentiation under the integral gives
Z Z
−m/2 α −i(x,ζ) −m/2
(2π) ∂ζ e f (x)dx = (2π) (−i)|α| xα e−i(x,ζ) f (x)dx,
Rm Rm

which is the claimed formula.


b) One has
Z Z
−m/2 α −i(x,ζ) |α| −m/2
(2π) ζ e f (x)dx = (−1) (2π) e−i(x,ζ) ∂xα f (x)dx,
Rm Rm

where we have integrated by parts.


c) By parts a), b) we can estimate as follows

α βb α\ β
|ζ D f (ζ)| = D (x f )(ζ)

Z
≤ |Dα (xβ f )(x)|dx
m
ZR
= |Dα (xβ f )(x)|(1 + |x|2 )(m+1)/2 (1 + |x|2 )−(m+1)/2 dx
Rm
≤ C sup |Dα (xβ f )(x)|(1 + |x|2 )(m+1)/2 ,
x∈Rm

where Z
C := (1 + |x|2 )−(m+1)/2 dx < ∞.
Rm
Using the Leibniz formula to calculate Dα (xβ f ), we can pick multi-indices α0 , β 0 and C 0 > 0
such that
C sup |Dα (xβ f )(x)|(1 + |x|2 )(m+1)/2 ≤ C 0 pα0 ,β 0 (f ).
x∈Rm

d) Transformation formula for integrals. 


The following corollary will later on show that one can solve PDE’s with constant co-
efficients in the Schwartz space, if the polynomial underlying the differential operator is
invertible.
Global Analysis 2 29

Corollary 3.36. Assume


X
P (X1 , . . . , Xm ) = Pα X α ∈ Mat`×` (C)[X 1 , . . . , X m ]
|α|≤m

is a polynomial with constant matrix coefficients, and let P (D1 , . . . , Dm ) = α Pα D√α be the
P
associated differential operator with constant matrix coefficients, where Dj = (1/ −1)∂j .
Then one has
F P (D1 , . . . , Dm )f = P F f for all f ∈ S (Rm , Cl ) .
Our next aim will be to prove that the Fourier transform is continuously invertible. To
this end, the Fourier transform of the Gaussian function will be useful (for it is a fixed
point of F ):
Example 3.37. For the Schwartz function exp(−| · |2 ) ∈ S (Rm ) one has
2
−|·|2 /2 = e−|·| /2 .
e\
There are many proofs of this formula (complex analysis, ...). We use ODE theory: firstly,
in view of m
−|x|2 /2 2
Y
e = e−|xj | /2
j=1
and Fubini, it is enough to consider the case m = 1. Set
Z
−1/2 2
h(t) := e\
−|·| 2
/2(t) = (2π) e−ixt e−x /2 dx.
R
Then one has Z
−1/2 2
(d/dt)h(t) = i(2π) e−ixt (−x)e−x /2 dx
R
Z Z
−1/2 −ixt −x2 /2 −1/2 2
= i(2π) e (d/dx)e dx = −i(2π) (−it)e−ixt e−x /2 dx
R R
= −th(t).
Thus Z
−t2 /2 −1/2 2 /2 2 /2 2 /2
h(t) = h(0)e = (2π) e−x dx · e−t = e−t .
R

The following lemma is a small variation of the Friedrichs construction:


Lemma 3.38. Assume f0 ∈ L1 (Rm ), f ∈ L∞ (Rm , Cl ) are such that
R
• R m f0 = 1
• f is continuous at 0.
Then one has Z
lim λ−m f0 (x/λ)f (x)dx = f (0).
λ→0+ Rm

Proof : The proof is similar to the proof of Proposition 3.16 and can be left to the reader.

30 B. GÜNEYSU

Now we can prove the following result:


Theorem 3.39. F is continuously invertible and its inverse is given by
F −1 : S (Rm , Cl ) −→ S (Rm , Cl ), F −1 f (x) = fˇ(x) = F f (−x),
the so called inverse Fourier transform.
Proof : Let G be defined by
G : S (Rm , Cl ) −→ S (Rm , Cl ), Gf (x) := F f (−x).
The fact that G is continuous is checked similarly to the continuity of F . Let f ∈
S (Rm , Cl ).
To see that GF f = f it suffices to prove
GF f (0) = f (0).
Indeed, together with the translation behaviour from Proposition 3.35 d), the latter equality
allows the following calculation:
Z
−m/2
f (x) = f (0 + x) = GF f (· + x)(0) = (2π) F f (· + x)(ζ)ei(0,ζ) dζ
Rm
Z
−m/2 i(x,ζ)
= (2π) e F f (ζ)dζ = GF f (x).
Rm
2
To show GF f (0) = f (0), let h := e−|·| /2 , f0 := (2π)−m/2 h. Using the previous lemma,
F h = h, and the scaling behaviour from Proposition 3.35 d), we can calculate
(2π)m/2 f (0)
Z
= lim λ−m h(x/λ)f (x)dx
λ→0+ Rm
Z
= lim λ−m F h(ζ/λ)f (ζ)dζ
λ→0+ Rm
Z
= lim F h(λ·)(ζ)f (ζ)dζ
λ→0+ Rm
Z
= lim h(λζ)F f (ζ)dζ
λ→0+ Rm
Z
= F f (ζ)dζ,
Rm
where we have used Fubini for the fourth equality and dominated convergence for the last
equality. Using the magic trick 1 = ei(0,ζ) , the latter expression is
Z
= ei(0,ζ) F f (ζ)dζ.
Rm
Alltogether we have shown
Z
−m/2
f (0) = (2π) ei(0,ζ) F f (ζ)dζ = GF f (0).
Rm
Global Analysis 2 31

The proof of F Gf (x) = f (x) is very similar, showing G = F −1 . 


Remark 3.40. The invertibility of F gives a simple recipe for solving PDE’s with constant
coefficients: assume that in the situation of Corollary 3.36 we want to find a solution f of
P (D1 , . . . , Dm )f = g, where g is a Schwartz function. Then we have
F P (D1 , . . . , Dm )f = P F f = F g,
so if P (ζ1 , . . . , ζm ) is an invertible matrix for all ζ ∈ Rm and if P −1 F g is Schwartz
again, then f = F −1 P −1 F g is the unique solution in the space of Schwartz functions
of P (D1 , . . . , Dm )f = g.
We extend the Fourier transform by duality:
Definition 3.41. The continuous linear map
F : S 0 (Rm , Cl ) −→ S 0 (Rm , Cl ), F T (f ) := T (F −1 f )
is called the distributional Fourier transform.
The distributional Fourier transform is continuously invertible, too. This follows immedi-
ately from Theorem 3.39. The reason for taking F −1 in the above definition comes from
the anti linearity of the map f 7→ Tf and makes sure that F Tf = F f if f ∈ S 0 (Rm , Cl ).
One can replace S with S 0 everywhere in Remark 3.40, obtaining a unique Schwartz dis-
tribution f which solves P (D1 , . . . , Dm )f = g, given g is a Schwartz distribution. However,
in both cases, the assumption that P (ζ1 , . . . , ζm ) is invertible for all ζ ∈ Rm is restrictive
in applications. Even if this is the case, the assumption that P −1 F g is Schwartz is often
not satisfied. However, if we restrict ourselves to compactly supported distributions g’s, it
turns out that one can find a (not uniquely determined) solution f in D 0 assuming only
that P (ζ1 , . . . , ζm ) is invertible for some ζ, for one has the famous:
Theorem 3.42 (Malgrange-Ehrenpreis Theorem, 1955). Assume
X
P (X1 , . . . , Xm ) = Pα X α ∈ Mat`×` (C)[X 1 , . . . , X m ]
|α|≤m

is a polynomial with constant matrix coefficients such that P (ζ1 , . . . , ζm ) is invertible for
some ζ ∈ Rm (note that for the scalar case l = 1 this assumption just means that P is not
the trivial polynomial). Then for every g ∈ E 0 (Rm , Cl ) there exists an f ∈ D 0 (Rm , Cl ) with
P (∂1 , . . . , ∂m )f = g.
If g ∈ D(Rm , Cl ), then then f can be chosen to be in E (Rm , Cl ).
Proof : The original proofs by Malgrange and Ehrenpreis were rather involved and relied
on the Hahn-Banach theorem. I will sketch a remarkably simple constructive proof for the
scalar case which is due to Peter Wagner [15] from 2009: first of all, it suffices to consider
the case g = δ0 (in which case any solution is called a fundamental solution). Indeed,
32 B. GÜNEYSU

define the convolution f1 ∗ f2 ∈ D 0 (Rm ) of f1 ∈ D 0 (Rm ), f2 ∈ E 0 (Rm ) as follows: first of


all, if f2 ∈ D(Rm ), then f1 ∗ f2 ∈ E (Rm ) is defined by
f1 ∗ f2 (x) := f1 (f2 (x − •)).
In the general case, f1 ∗ f2 be defined to be the uniquely determined T ∈ D 0 (Rm ) which
satisfies T (φ) = f1 ∗ (f2 ∗ φ) for all φ ∈ D(Rm ) (neither the smoothness of f1 ∗ f2 in the
former case, nor the existence/uniqueness part in the latter case are trivial; cf. [4]). Then
one has
∂ α (f1 ∗ f2 ) = (∂ α f1 ) ∗ f2 , δ0 ∗ f2 = f2 .
Thus if we can find f˜ ∈ D 0 (Rm ) with
P (∂1 , . . . , ∂m )f˜ = δ0 ,
then f := f˜ ∗ g solves
P (∂1 , . . . , ∂m )f = g,
and f is smooth, if g is so.
Now assume P has degree k and denote with
X
P (k) (X1 , . . . , Xm ) := Pα X α
|α|=k

its k-principal part. Pick η ∈ Rm with P (k) (η) 6= 0. Given any λ ∈ R, using P (k) (η) 6= 0 it
is straightforward to check that

|{x : P ( −1x + λη) = 0}| = 0,
so that √
P ( −1x + λη)
x 7−→ P̃λ (x) := √
P ( −1x + λη)
is a well-defined element of L∞ (Rm ) ,→ S 0 (Rm ). Consider also the smooth function
x 7−→ fλ (x) := eλ(η,x)
as a differential operator of order 0. Then one calculates
k−1
X
−1
(16) P (∂1 , . . . , ∂m )fλ F P̃λ = P (k) (2η)λk δ 0 + λi Ti
i=0

for some T0 , . . . , Tk−1 ∈ D (R ). It follows that if one picks λ0 , . . . , λk ∈ R pairwise


0 m

disjoint, and sets


Yk
aj := (λj − λi )−1 ,
i=0,i6=j
then
k
1 X
f := (k) aj fλj F −1 P̃λj
P (2η) j=0
Global Analysis 2 33

satisfies
P (∂1 , . . . , ∂m )f = δ0 .
Indeed, this follows immediately from combining (16) with the following algebraic fact:
given disjoint complex numbers λe0 , . . . , λek , the unique solution (a˜0 , . . . , a˜k ) ∈ Ck+1 of the
linear system (
k
X i 0 if i = 0, . . . , k − 1
a˜j λej =
j=0
1 if i = k
is given by
k
Y
a˜j = (λej − λei )−1 , j = 0, . . . , k.
i=0,i6=j

Remark 3.43. 1. The fact that all the machinery from the proof has been established by
Schwartz should you give a feeling of how well-developed distribution theory is (and why
it deserved a fields medal in 1950).
2. It is not obvious at all whether the above proof for the existence of a fundamental
solution (replacing δ0 with δ0 ⊗ 1 extends to the vector-valued case: is the right definition
of P̃λ (x) given by √ √
P̃λ (x) := P ( −1x + λη)† P ( −1x + λη)−1
or √ √
P̃λ (x) := P ( −1x + λη)−1 P ( −1x + λη)† ,
or do we even need that the coefficients of P are normal to make the proof work?
3. To give an idea of how research sometimes arises from examining proofs (assume for
simplicity l = 1): once one has a fundamental solution, the proof shows that ’arbitrary’
right-hand sided can be dealt with convolution. Thus, to obtain regularity results for
solutions, one could examine the following question: Given subspace V ⊂ E 0 (Rm ), W ⊂
D 0 (Rm ), is it then true that f˜ ∗ g ∈ W for all f˜ ∈ D 0 (Rm ), g ∈ V ? If so, given any g ∈ V
one finds a solution f ∈ W of
P (∂1 , . . . , ∂m )f = g.
Since f := f˜ ∗ g solves the above equation, if f˜ ∈ D 0 (Rm ) is a fundamental solution.
4. It would interesting to see if one can actually calculate fundamental solutions using the
construction from the proof in some particular cases, such as
• the heat equation in Rm , where
Xm
P (∂0 , . . . , ∂m ) = ∂0 − ∂j2 = ∂0 + ∆Rm ,
j=1

the Laplace equation in Rm , where


m
X
P (∂1 , . . . , ∂m ) = − ∂j2 = ∆Rm ,
j=1
34 B. GÜNEYSU

or the wave equation in Rm , where


m
X
P (∂0 , . . . , ∂m ) = ∂02 − ∂j2 = ∂02 + ∆Rm .
j=1

Next, we want to consider the Fourier transform in L2 . To this end we recall the following
facts from functional analysis:
Remark 3.44. 1. A continuous linear operator A : X → Y between Hilbert spaces is
called unitary, if A is invertible with A−1 = A∗ . Unitary maps preserve scalar products
and thus norms.
2. If X is a normed space, Y is a Banach space, D ⊂ X is a dense linear subspace,
and T0 : D → Y is a continuous linear map, then there exists a unique continuous linear
extension T : X → Y of T0 . One has the following relation for the operator norms:
kT0 kD→Y := sup{kT0 f k : kf k ≤ 1, f ∈ D} = kT kX→Y := sup{kT f k : kf k ≤ 1, f ∈ X}.
Theorem 3.45 (Plancherel’s Theorem). The linear map
(17) S (Rm , Cl ) −→ L2 (Rm , Cl ), f 7−→ F f,
where the Schwartz space is equipped with the L2 -topology, is continuous. Thus this map
has a unique continuous linear extension L2 (Rm , Cl ) → L2 (Rm , Cl ), which will be denoted
again by F . This extension is unitary.
Proof : For all f1 , f2 ∈ S (Rm , Cl ) one has
Z
hF f1 , F f2 i = (F f1 , F f2 )
Rm
Z Z
= (2π)−m/2 e+i(x,ζ) (f1 (x), F f2 (ζ))dxdζ
m m
ZR ZR
= (2π)−m/2 e+i(x,ζ) (f1 (x), F f2 (ζ))dxdζ
m m
ZR R

= (f1 (x), F −1 F f2 (x))dx


Rm
= hf1 , f2 i ,
showing that (17) is continuous in the asserted way. Let us denote for a moment with
FL2 : L2 (Rm , Cl ) −→ L2 (Rm , Cl )
its continuous linear extension. As above, one shows that the map
G : S (Rm , Cl ) −→ L2 (Rm , Cl ), f 7−→ F −1 f,
satisfies
hGf1 , Gf2 i = hf1 , f2 i ,
Global Analysis 2 35

for all f1 , f2 ∈ S (Rm , Cl ), and so is continuous, too (again with the Schwartz space is
equipped with the L2 -topology). Denoting with
GL2 : L2 (Rm , Cl ) −→ L2 (Rm , Cl )
its continuous linear extension, it follows now easily from a density argument and GL2 =
FL−12 when considered as acting in S (R , C ) that GL2 = FL2 .
m l −1
Finally, for all f, g ∈
2 m l
L (R , C ) one has
hf, FL2 gi = lim hfn , F gn i
n
= lim F fn , F −1 F gn 2 = hGL2 f, gi ,

−1
n

for some sequences fn , gn in S (Rm , Cl ) with fn → f and gn → g in L2 (Rm , Cl ). Thus we


have shown FL−1 ∗
2 = GL2 = FL2 , which completes the proof. 
Remark 3.46. Plancherel’s theorem has important applications in quantum mechanics
(which allows to switch from the ’position representation’ to the ’momentum representa-
tion’. It also allows to prove a refined version of Heisenberg’s uncertainty principle (cf. Satz
11.14 in [16], which states that momentum and position cannot be localized simultaniously
on sets of finite Lebesgue measure.
While the Fourier transform in L2 is ’abstractly’ defined (by density), it turns out that
the Fourier transform on L1 ∩ L2 can be evaluated explicitly. This results in the Riemann-
Lebesgue Lemma, which in addition states that the Fourier transform of an L1 -function is
continuous and vanishes at ∞.
Theorem 3.47 (Riemann-Lebesgue Lemma). The linear map
(18) L1 (Rm , Cl ) ∩ L2 (Rm , Cl ) −→ C0 (Rm , Cl ), f 7−→ F f,
where the LHS is equipped with the L1 -topology, is well-defined and continuous. Its unique
continuous linear extension
L1 (Rm , Cl ) −→ C0 (Rm , Cl )
is denoted with F again, and this map is injective with
kF kL1 →C0 ≤ (2π)−m/2 ≤ 1.
One has
Z
−m/2
(19) F f (ζ) = (2π) e−i(x,ζ) f (x)dx for all f ∈ L1 (Rm , Cl ), ζ ∈ Rm .
Rm

Proof : Let us first check that for all f ∈ L1 (Rm , Cl ) ∩ L2 (Rm , Cl ) one has (19) for a.e.
ζ ∈ Rm , where the Fourier transform F f has already been defined above, as f ∈ L2 (Rm , Cl ).
Indeed, pick a sequence fn in S (Rm , Cl ) with fn → f in L1 (Rm , Cl ) and in L2 (Rm , Cl ).
Then from the L1 -convergence one gets
Z Z
−m/2 −i(x,ζ) −m/2
lim (2π) e fn (x)dx = (2π) e−i(x,ζ) f (x)dx for all ζ ∈ Rm ,
n Rm Rm
36 B. GÜNEYSU

while the L2 -convergence and the continuity of F in L2 (Rm , Cl ) implies F fn → F f in


L2 (Rm , Cl ). Thus, after possibly picking a subsequence one gets
Z
−m/2
F f (ζ) = lim F fn (ζ) = lim(2π) e−i(x,ζ) fn (x)dx for a.e. ζ ∈ Rm ,
n n Rm

ultimately proving (19) for a.e. ζ ∈ ζ ∈ Rm .


Formula (19) shows that for all f ∈ L1 (Rm , Cl ) ∩ L2 (Rm , Cl ), the RHS of (19) defines a
continuous representative of F f , which is absolutely bounded by (2π)−m/2 kf kL1 . Thus we
have a continuous linear map
(20) L1 (Rm , Cl ) ∩ L2 (Rm , Cl ) −→ Cb (Rm , Cl ), f 7−→ F f,
where the LHS is equipped with the L1 (Rm , Cl )-topology. It remains to show that F f
vanishes at ∞ for all f ∈ L1 (Rm , Cl ) ∩ L2 (Rm , Cl ). Pick a sequence fn in S (Rm , Cl ) with
fn → f in L1 (Rm , Cl ). Then
F fn ∈ S (Rm , Cl ) ⊂ C0 (Rm , Cl ) for all n
and by the continuity of (20) one has F fn → F f in Cb (Rm , Cl ). It follows that F f ∈
C0 (Rm , Cl ), as C0 (Rm , Cl ) is a closed subspace of Cb (Rm , Cl ).
In order to show that
(21) F : L1 (Rm , Cl ) −→ C0 (Rm , Cl )
is injective, assume h, g ∈ L1 (Rm , Cl ) are given with F h = F g. As for all f ∈ L1 (Rm , Cl )
the formula (19) remains true (pick a sequence
(fn ) ⊂ S (Rm , Cl ) ⊂ L1 (Rm , Cl ) ∩ L2 (Rm , Cl )
with fn → f in L1 (Rm , Cl ) and use the continuity of (21)), it follows using Fubini that for
all φ ∈ S (Rm , Cl ) one has
Z Z Z
−1
(h, F φ) = (2π)−m/2 ei(x,ζ) (h(ζ), φ(x))dxdζ
Rm m m
Z ZR R Z
= (F h, φ) = (F g, φ) = (g, F −1 φ).
Rm Rm Rm

Thus we have shown F Th = F Tg in S 0 (Rm , Cl ) and so Th = Tg in in S 0 (Rm , Cl ), as the


Fourier transform is bijective on Schwartz distributions. Finally, we arrive at h = g a.e. in
Rm by the fundamental lemma of distribution theory. 
Thus, putting all mapping properties together, the Fourier transform induces consistently
defined continuously invertible maps
S (Rm , Cl ) −→ S (Rm , Cl ),
S 0 (Rm , Cl ) −→ S 0 (Rm , Cl ),
L2 (Rm , Cl ) −→ L2 (Rm , Cl ),
Global Analysis 2 37

which are furthermore consistent with the continuous injective (and in fact not surjective)
map
L1 (Rm , Cl ) −→ C0 (Rm , Cl ),
and we have agreed on denoting all these maps by F . If f ∈ L1 (Rm , Cl ), then one has the
explicit formula Z
F f (ζ) = (2π)−m/2 e−i(x,ζ) f (x)dx, ζ ∈ Rm ,
Rm
and in all other cases F f is defined implictly (either by duality in the case of Schwartz
distributions, or by density in the L2 -case).
Remark 3.48. We have seen that F is bounded from L1 to L∞ and from L2 to L2 . Using
an interpolation result for bounded operators from Lp to Lq (Riesz-Thorin) one can now
easily show that F also defines a continuous linear map
F : Lp (Rm , Cl ) −→ Lq (Rm , Cl )
for all p ∈ [1, 2] and q ∈ [1, ∞] with 1/p + 1/q = 1, with operator norm
kF kLp →Lq ≤ (2π)m(1/2−1/p) .
The later bound is called the Hausdorff-Young inequality. We are not going to use this
result in the sequel.
We continue with the Fourier convolution theorem. To this end, it will be convenient to
define a slightly modified convolution by
Z
m/2
f ∗˜g(x) := 1/(2π) f (x − y)g(y)dy.

Theorem 3.49 (Fourier convolution theorem). Let f, g ∈ L2 (Rm ). Then the following
statements are equivalent:
• f ∗ g ∈ L2 (Rm ),
• F f · F g ∈ L2 (Rm ),
• F −1 f · F −1 g ∈ L2 (Rm ).
In this case one has the formulae
(22) F f · F g = F (f ∗˜g).
and
(23) F −1 f · F −1 g = F −1 (f ∗˜g).
Proof : We only prove (22), as the proof of (22) is very similar. For all f, g in S (Rm ) one
has
Z Z Z
−m/2 −i(ζ,x) −m/2 −m/2
F (f ∗g)(ζ) = (2π)
˜ e f ∗g(x)dx = (2π)
˜ (2π) e−i(ζ,x) f (x − y)dxg(y)dy
Z Z
−m/2 −m/2
= (2π) (2π) e−i(ζ,z+y) f (z)dzg(y)dy = F f (ζ)F g(ζ).
38 B. GÜNEYSU

Assume now f, g ∈ L2 (Rm ) and pick sequences fn , gn in S (Rm ) with fn → f and gn → g


in L2 (Rm ). This implies F fn · F gn → F f · F g in L1 (Rm ): to see this, write
F f n · F gn − F f · F g = F f n · F gn + F f · F gn − F f · F gn − F f · F g
and use Cauchy-Schwarz. On the other hand, one has fn ∗˜gn → f ∗˜g in L∞ (Rm ), which is
seen likewise, writing
fn ∗˜gn − f ∗˜g = fn ∗˜gn + f ∗˜gn − f ∗˜gn − f ∗˜g
and using Cauchy-Schwarz. As Lp -convergence implies convergence in the sense of Schwartz
distributions and as F is continuous, we have shown
F (f ∗˜g) = lim F (fn ∗˜gn ) = lim F fn · F gn = F f · F g in S 0 (Rm ),
n n

finishing the proof the proof of (22), as F leave L2 invariant. 


Remark 3.50. Note that in the above situation one also has the formula F (f g) = F f ∗˜F g,
because of
F f ∗˜F g = F (F −1 F f · F −1 F g) = F (f g),
and likewise
F −1 (f g) = F − f ∗˜F −1 g.
Let us note/calculate some Fourier transform of Schwartz distributions:
Example 3.51. 1. The Fourier transform of δa ∈ S 0 (Rm ) is
F δa = Tx7→(2π)−m/2 e−i(a,x) .
Indeed,
Z Z
−1 −m/2 i(x,ζ) −m/2
F δa (φ) = δa (F φ) = (2π) e φ(x)dx|ζ=a = (2π) e+i(a,x) φ(x)dx.

2. The Fourier transform of the Heaviside function H := 1[0,∞) ∈ L∞ (R) ,→ S 0 (R) is given
by
F TH = (2π)−1/2 lim Tζ7→(+iζ)−1 .
→0+

This follows from approximating TH by TH in S 0 (R), where H (x) := 1[0,∞) e−x .
Let us finally come to the definition of the support of a distribution:
Definition 3.52. 1. Given T ∈ D 0 (Ω, Cl ), U ⊂ Ω open, we set
T |U := T |D(U,Cl ) ∈ D 0 (U, Cl )
and say that T is smooth in U , if T |U = Tψ for some smooth ψ : U → Cl .
2. For T ∈ D 0 (Ω, Cl ) one defines its support by
[
supp(T ) := Ω \ Ω0 ,
Ω0 ⊂Ω, T |Ω0 =0
Global Analysis 2 39

and its singular support by


[
sing − supp(T ) := Ω \ Ω0 .
Ω0 ⊂Ω, T |Ω0 is smooth

Note that in view of E 0 ,→ S 0 ,→ D 0 , the above notions are defined in all spaces of
distributions. Furthermore, the assignment U 7→ D 0 (U, Cl ), together with the restriction
morphisms
D 0 (V, C) −→ D 0 (U, Cl ), T 7−→ T |U ,
where U ⊂ V ⊂ Ω are open, is a sheaf over Ω: all sheaf axioms are more or less trivial
to check, except locality, which requires the existence of smooth partitions of unity (cf.
Theorem 3.58). This will be shown as an exercise.

Remark 3.53. 1. For all T ∈ D 0 (Ω, Cl ) one has sing − supp(T ) ⊂ supp(T ), which are
both closed subsets of Ω.
2. For all T ∈ D 0 (Ω, Cl ) one has T |Ω\supp(T ) = 0, and T |Ω\sing−supp(T ) is smooth (use sheaf
property).
3. For all T ∈ D 0 (Ω, Cl ) one has T ∈ E 0 (Ω, Cl ), if and only if supp(T ) is a compact set
(exercise).
4. For all f ∈ L1loc (Ω, Cl ) and all p ∈ Ω one has p ∈ supp(T ), if and only if
Z
|f | > 0 for all  > 0 with B (p) ⊂ Ω.
B (p)

This result follows from the fact that the Lebesgue measure is a Borel measure with full
support, and it implies that for all f ∈ C(Ω, Cl ) one has
supp(f ) = {f 6= 0} = supp(Tf ),
that is, in the continuous case the distributional support is equal to the support in the
sense of functions on topological spaces.
Example 3.54. 1. One has sing − supp(δa ) = supp(δa ) = {a}.
2. For f := 1(0,∞) log ∈ L1loc (R) one has
sing − supp(Tf ) = {0}, supp(Tf ) = [0, ∞) = supp(f ).
3. Although in the case of continuous functions and even in the previous example we had
supp(Tf ) = supp(f ), this equality fails in general, e.g. for f = 1Q on Ω = R. In fact, in
the context of analysis, the distributional support of a function turns out to be the more
useful concept.
We now turn to Sobolev spaces, the localized versions of which will be the natural mapping
spaces for pseudodifferential operators:
Definition 3.55. For all k ∈ N≥0 define H k (Ω, Cl ) to be the space of all f ∈ D 0 (Ω, Cl )
such that ∂ α f ∈ L2 (Ω, Cl ) for all |α| ≤ k. This space is called the Sobolev space of order
40 B. GÜNEYSU

k and is equipped with the scalar product


X
hf, giΩ
H k := h∂ α f, ∂ α f iΩ .
|α|≤k

Note that the assumption ∂ α f ∈ L2 (Ω, Cl ) is a short way of expressing that ∂ α f is a


regular distribution which stems from an L2 (Ω, Cl ) function; in other words, there exists a
(necessarily uniquely determined) g ∈ L2 (Ω, Cl ) such that for all φ ∈ D(Ω, Cl ) one has
Z Z
|α|
(φ, g) = (−1) (∂ α φ, f ),
Ω Ω
α
and then ∂ f is the regular distribution given by g. In particular, using this with α =
(0, . . . , 0), we find that f is a regular distribution which stems from an L2 (Ω, Cl ) function.
Lemma 3.56. H k (Ω, Cl ) is a Hilbert space for all k ∈ N≥0 .
Proof : Assume fn is a Cauchy in H k (Ω, Cl ). Then Dα fn is Cauchy in L2 (Ω, Cl ) for all
|α| ≤ k, and we denote its limit by F (α) ∈ L2 (Ω, Cl ). Then one can easily prove that
f := F (0,...,0) is the limit of fn in H k (Ω, Cl ), and in fact F (α) = Dα f . 
One has the following important result:
Theorem 3.57 (Meyers-Serrin, 1964). H k (Ω, Cl ) ∩ E (Ω, Cl ) is dense in H k (Ω, Cl ) for all
k ∈ N≥0 .
The proof relies on the existence of smooth partitions of unity:
Theorem 3.58. Assume (Ui )i∈I is a cover of Ω with open subsets. Then there exists a
smooth partition (ψi )i∈I subordinate to (Ui )i∈I , that is,
• ψi ∈ E (Ω), 0 ≤ ψi ≤ 1, supp(ψi ) ⊂ Ui for all i ∈ I,
• P(supp(ψi ))i∈I is a locally finite family,
• i∈I ψi = 1.
Proof : Since Ω is metrizable, it is a paracompact Hausdorff space. Thus we can find locally
finite open covers (Vi ), (Wi ) of Ω such that Vi ⊂ Ui , Wi ⊂ Vi (’shrinking lemma’). By
the proof of (a small generalization of) Proposition 3.15 one can pick φi ∈ E (Ω) with
0 ≤ φi ≤ 1 with φ1 = 1 on Wi and φi = 0 on Ω \ Vi . Then
X
ψi := φi / φj
j∈I

has the desired properties. 


Proof of Theorem 3.57: I will give a somewhat simplified version of the original proof by
Meyers-Serrin:
Let us first show that for all open relatively compact U ⊂ Ω one has
%λ ∗ f |Ω − f U k → 0

(24) H
Global Analysis 2 41

as λ → 0+. To this end, observe that ∂ α %λ ∗ f |Ω = %λ ∗ ∂ α f for all sufficiently small λ > 0,
which can be seen as follows: Pick λ0 > 0 such that for all x ∈ U one has Bλ (x) ⊂ Ω. Then
for fixed x ∈ U the function y 7→ %λ (x − y) is a smooth compactly supported function in
Ω, and so
Z
α
∂ %λ ∗ f (x) = ∂xα %λ (x − y)f (y)dy

Z
= (−1)|α| ∂yα %λ (x − y)f (y)dy

Z
= %λ (x − y)∂ α f (y)dy.

Thus, (24) follows from
%λ ∗ (∂ α f )|Ω − ∂ α f U2 → 0.

L
Now pick countable (Ui ) and (ψi ) as in Theorem 3.58. Then fi := ψi f ∈ H k (Ω, Cl ) with
supp(fi ) ⊂ Ui (note that Ui is relatively compact). For fixed  > 0, by the above, we can
pick λi > 0 such that
%λ ∗ fi |Ω − fi Ω k <  .

i H 2i+1
Now we set
X∞
v := %λi ∗ fi |Ω ,
i=0
l
which is a smooth function Ω → C , as the sum is locally finite. For an arbitrary open
relatively compact U ⊂ Ω we get
∞ ∞
U
X Ω X 
kv − f kH k ≤ %λi ∗ fi |Ω − fi H k <

i+1
< ,
i=0 i=0
2
thus letting U run through an open relatively compact exhaustion of Ω, we end up with
kv − f kΩ
H k < ,

which a posteriori also shows v ∈ H k (Ω, Cl ). 


Remark 3.59. 1. The paper by Meyers-Serrin has the title H = W . The probably shortest
title of any math paper stems from the fact that before this result some authors defined
H k (Ω, Cl ) to be the closure of all smooth functions Ω → Cl with all derivatives of order
≤ k in L2 (Ω, Cl ), while some authors called W k (Ω, Cl ) what we defined to be H k (Ω, Cl ).
In that sense one has H = W . When published, the result clarified some confusion in the
PDE literature.
2. D(Ω, Cl ) is not dense in H k (Ω, Cl ) (however it is, if Ω = Rm ; exercise). This justifies to
give the closure of D(Ω, Cl ) in H k (Ω, Cl ) a new symbol: typically something like H0k (Ω, Cl ).
3. Sobolev spaces have obvious Lp variants that are denoted by H k,p (Ω) and H0k (Ω, Cl ).
The above results hold for all p < ∞, too.
4. There are generalizations of Meyers-Serrin, where one replaces the ∂ α with an arbitrary
42 B. GÜNEYSU

finite family of differential operators (under a subtle additional condition on the underlying
highest order operator). This is a result by Güneysu/Guidetti/Pallara from 2017 [3].
Example 3.60. If I = (a, b) is an interval, where −∞ ≤ a < b ≤ −∞, then for all
k ∈ N≥1 the space H k (I, Cl ) is given by all f ∈ C k−1 (I, Cl ), such that ∂ j f ∈ L2 (I, Cl ) for
j = 0, . . . , k − 1 and such that ∂T∂ k−1 f is absolutely continuous and in L2 (I, Cl ). Thus,
e.g., H 1 (I, Cl ) does not contain any noncontinuous functions. One can show10 that in fact
for all f ∈ H k (I, Cl ) the functions ∂ j f have limits as x → a and x → b for j = 0, ..., k − 1.
There is no simple way known to me to define H k (Ω, Cl ) for general Ω’s, if k ∈ R. However,
if Ω = Rm , we can observe that for k ∈ N≥0 and f ∈ D 0 (Rm ) one has f ∈ H k (Rm , Cl ), if
and only f ∈ S 0 (Rm , Cl ) and (1 + |ζ|2 )k/2 F f ∈ L2 (Rm , Cl ), and that the scalar product
m
hf, giRH k is equivalent to
Z
hf, giH k := (F f (ζ), F g(ζ))(1 + |ζ|2 )k dζ.
Rm
The latter data make sense for k ∈ R, and justify:
Definition 3.61. For all s ∈ R one defines H s (Rm , Cl ) to be the space of all f ∈
S 0 (Rm , Cl ) such that Z
|F f (ζ)|2 (1 + |ζ|2 )s dζ < ∞.
Rm
One defines a scalar product on H s (Rm , Cl ) by setting
Z
hf, giH s = (F f (ζ), F g(ζ))(1 + |ζ|2 )s dζ
Rm
s m l
and calls H (R , C ) the Sobolev space of order s.
Note that it is somewhat hidden in the definition that one assumes F f to be a regular
distribution. Note also that
H 0 (Rm , Cl ) = L2 (Rm , Cl )
and
H t (Rm , Cl ) ⊂ H s (Rm , Cl ) if s < t, continuously.
Let us see how the definition works:
Example 3.62. For all x ∈ Rm we have seen previously that
F δx (ζ) = (2π)−m/2 e−i(x,ζ) ,
a bounded function of ζ ∈ Rm . It follows that δx ∈ H s (Rm ), if and only if s < −m/2, as
precisely for such s one has
Z
|F δx (ζ)|2 (1 + |ζ|2 )s dζ < ∞.
Rm
10On
can also prove with an interpolation inequality that for all f ∈ C k−1 (I, Cl )∩L2 (I, Cl ) with ∂T∂ k−1 f
absolutely continuous and in L2 (I, Cl ), one automatically has ∂ j f ∈ L2 (I, Cl ) for j = 0, . . . , k − 1.
Global Analysis 2 43

It is straighforward to show that for all s < −m/2 the map


Rm −→ H s (Rm , Cl ), x 7−→ δx
is continuous.
Some basic properties of the H s -Sobolev spaces are collected in the proposition below. For
its proof we will need:
Lemma 3.63 (Peetre’s inequality). Assume X is a seminormed space and s ∈ R. Then
for all x, y ∈ X one has
(1 + kxk)s ≤ (1 + kx − yk)|s| (1 + kyk)s ,
in particular,
(1 + kxk2 )s ≤ Cs (1 + kx − yk2 )|s| (1 + kyk2 )s ,
where Cs > 0 is a universal constant that only depends on s.
Proof : Both inequalities have only to be proved for s ≥ 0 (the s < 0 then follows from
changing the roles of x and y). If we can show that the first inequality holds for all s ≥ 0,
the second one follows from
(1 + kxk2 )s ≤ (1 + 2 kxk + kxk2 )s = (1 + kxk)2s
≤ (1 + kx − yk)2s (1 + kyk)2s ≤ 2s 2s (1 + kx − yk2 )s (1 + kyk2 )s ,
where we have used
(a + b)2 ≤ 2(a2 + b2 ) for all a, b ≥ 0.
Finally, the first inequality follows for s ≥ 0 simply from applying
(1 + ku − vk)s ≤ (1 + kuk)s (1 + kvk)s
which follows from
1 + ku − vk ≤ 1 + kuk + kvk ≤ (1 + kuk)(1 + kvk),
with u = x − y, v = −y. 
Proposition 3.64. Let s ∈ R.
a) For all t ∈ R the map
Λs,t : H s (Rm , Cl ) −→ H s−t (Rm , Cl ), f 7−→ F −1 (1 + |ζ|2 )t/2 F f
is an isomorphism of normed spaces.
b) H s (Rm , Cl ) is a Hilbert space.
c) S (Rm , Cl ) is dense in H s (Rm , Cl ).
d) For all α ∈ Nm ≥0 the operator

Dα : H s (Rm , Cl ) −→ H s−|α| (Rm , Cl )


is well-defined and continuous.
e) For all ψ ∈ S (Rm ) the operator
H s (Rm , Cl ) −→ H s (Rm , Cl ), f 7−→ ψf
44 B. GÜNEYSU

is well-defined and continuous.


f ) The scalar product h·, ·i on S (Rm , Cl ) extends continuously to an antidual pairing be-
tween H s (Rm , Cl ) and H −s (Rm , Cl ). The pairing incuces a canonical antiisomorphism of
Banach spaces H s (Rm , Cl )0 ∼
= H −s (Rm , Cl ).

Proof : a) Clear.
b) Follows from a).
c) The map

Λ0,−s : L2 (Rm , Cl ) −→ H s (Rm , Cl )

is an isometric isomorphism. As S (Rm , Cl ) is dense in L2 (Rm , Cl ), it follows that

Λ0,−s (S (Rm , Cl )) ⊂ H s (Rm , Cl )

is a dense subspace. But we have

Λ0,−s (S (Rm , Cl )) ⊂ S (Rm , Cl ).

d) For all ψ ∈ S (Rm , Cl ), ζ ∈ Rm , one has

|F Dα ψ(ζ)| = |ζ α F ψ(ζ)| ≤ |(1 + |ζ|2 )|α|/2 ||F ψ(ζ)|,

so that
Z
kD α
ψk2H s−|α| = |F Dα ψ(ζ)|2 (1 + |ζ|2 )s−|α| dζ ≤ kψk2H s .

Now the claim follows from the density of S in H s .


e) Assume g ∈ S (Rm , Cl ). Then for all t > 0 we have

Z 2
2 2
|F (ψ · g)(x)| ≤ |F ψ ∗ F g(x)| ≤ |F ψ(y)||F g(x − y)|dy
Z 2
2 t/2 2 −t/2
= |F ψ(y)|(1 + |y| ) |F g(x − y)|(1 + |y| ) dy
Z
2
≤ kψkH t |F g(x − y)|2 (1 + |y|2 )−t dy,
Global Analysis 2 45

where we have used Cauchy-Schwarz. So


Z Z
2 2
kψgkHs ≤ kψkH t |F g(x − y)|2 (1 + |y|2 )−t (1 + |x|2 )s dydx
Z Z
2
≤ kψkH t |F g(x − y)|2 (1 + |y|2 )−t (1 + |x|2 )s dxdy
Z Z
2
= kψkH t |F g(z)|2 (1 + |z + y|2 )s dz(1 + |y|2 )−t dy
Z Z
2
≤ Cs kψkH t |F g(z)|2 (1 + |z|2 )s dz(1 + |y|2 )−t+|s| dy
Z
2 2
= Cs kψkH t kgkH s (1 + |y|2 )−t+|s| dy

≤ Cs,t kψk2H t kgk2H s ,


where we have used Peetre’s inequality, and where
Z
Cs,t := Cs (1 + |y|2 )−t+|s| dy < ∞,

if t > 0 is chosen large enough. As S is dense in H s , the proof is complete.


f) This follows from the fact that for Schwartz functions f, g one has
|hf, gi| = |hF f, F gi|
Z
F f (ζ)(1 + |ζ|2 )s/2 , F g(ζ)(1 + |ζ|2 )−s/2 dζ ≤ kf kH s kgkH −s ,

=

in view of Cauchy-Schwarz. That the pairing induces an isomorphism H s (Rm , Cl )0 ∼


=
H −s (Rm , Cl ) is left as an exercise. 
As a consequence of the fourth and fifth statement of the previous proposition, we imme-
diately get:
α
P
Corollary 3.65. Every partial differential operator P = α Pα D of order ≤ k with
Pαij ∈ S (R ) for all i = 1, . . . , l, j = 1, . . . , l maps
m 0

0
P : H s (Rm , Cl ) −→ H s−k (Rm , Cl )
continuously.
If f ∈ H s (Rm , Cl ), one might expect that f should canonically induce an element of
H s (Rm−1 , Cl ) by some sort of restriction. However, f need not be a function and if so, it
need not have a canonically given representative, so that it doesn’t make sense to consider
f (0, ·). There is a functional analytic way out of this dilemma, the price being that one
looses 1/2 of the Sobolev order:
Theorem 3.66 (Trace Theorem). For all s > 1/2 the map
S (Rm , Cl ) −→ S (Rm−1 , Cl ), f 7−→ f (0, ·)
46 B. GÜNEYSU

extends to a continuous linear map


H s (Rm , Cl ) −→ H s−1/2 (Rm−1 , Cl ).
Proof : Let f ∈ S (Rm , Cl ) and write y = (y1 , y 0 ) ∈ R × Rm−1 . By the invertibility of the
Fourier transform we have
Z
0 0 0 0
g(x ) := f (0, x ) = C ei(x ,ζ ) ei(0,ζ1 ) F f (ζ1 , ζ 0 )dζ
m
Z Z R
0 0
=C ei(x ,ζ ) F f (ζ1 , ζ 0 )dζ1 dζ 0 = CF −1 h(x0 ),
Rm−1 R

where Z
h(y) := F f (ζ1 , y)dζ1 .
R
Thus Z Z
0 0
F g(ζ ) = C F f (ζ1 , ζ )dζ1 = C F f (ζ)dζ1 .
R R
Using Cauchy-Schwarz we estimate as follows,
Z 2
0 2 2 s/2 2 −s/2
|F g(ζ )| ≤ C |F f (ζ)| (1 + |ζ| ) (1 + |ζ| ) dζ1
R
Z Z
2
≤ C |F f (ζ)| (1 + |ζ| ) dζ1 (1 + |ζ|2 )−s dζ1
2 s
R R
Z Z
0 2 −s+1/2
= C(1 + |ζ | ) (1 + t ) dt · |F f (ζ)|2 (1 + |ζ|2 )s dζ1
2 −s

ZR R

= C 0 (1 + |ζ 0 |2 )−s+1/2 |F f (ζ)|2 (1 + |ζ|2 )s dζ1 .


R

Thus we have
Z
0 2
|F g(ζ )| (1 + |ζ | ) 0 2 s−1/2
≤C 0
|F f (ζ)|2 (1 + |ζ|2 )s dζ1
R

and by integrating over Rm−1 we end up with


kf (0, ·)k2H s−1/2 (Rm−1 ,Cl ) = kgk2H s−1/2 (Rm−1 ,Cl ) ≤ C 0 kf k2H s (Rm ,Cl ) ,
finishing the proof. 
Theorem 3.67 (Sobolev embedding theorem). Let k ∈ N≥0 . Then for all s ∈ R with
s > k + m/2 the inclusion
S (Rm , Cl ) ⊂ C0k (Rm , Cl )
extends to a continuous linear injective map
H s (Rm , Cl ) ,→ C0k (Rm , Cl ).
Global Analysis 2 47

Note that this result implies that if s > k + m/2 > 0 every
f ∈ H s (Rm , Cl ) ⊂ L2 (Rm , Cl )
has a (of course uniquely determined) Lebesgue representative f˜ ∈ C0k (R, Cl ) with

˜
f ≤ C kf kH s ,
∞,k

where C > 0 does not depend on f . Indeed, pick a sequence fn of Schwartz function with
fn → f in H s (Rm , Cl ). Then by the continuity of the above linear injective map fn is
Cauchy in C0k (Rm , Cl ) and can take f˜ to be the limit.
Proof of Theorem 3.67: We are going to prove the existence of a constant C = C(k, m, s) >
0 such that for all f ∈ S (Rm , Cl ) one has
(25) kf k∞,k ≤ C kf kH s .
As S is dense in H s , this will prove the claim. To see the latter inequality, let |α| ≤ k.
We have
|F Dα f (ζ)| = |ζ α F f (ζ)| = |(1 + |ζ|2 )(s−k)/2 ζ α F f (ζ)(1 + |ζ|2 )−(s−k)/2 |
≤ C(1 + |ζ|2 )−(s−k)/2 (1 + |ζ|2 )s/2 |F f (ζ)|.
Thus, by the Riemann-Lebesgue Lemma, the formula
F −1 Dα f (x) = F Dα f (−x),
and Cauchy-Schwarz, one gets
kDα f k∞ = F F −1 Dα f ∞ ≤ F −1 Dα f L1

= kF Dα f kL1 ≤ C 0 kf kH s ,
where
Z 1/2
0 2 −s+k
C =C· (1 + |ζ| ) dζ < ∞,
Rm
as we have assumed −s + k < −m/2. 
Theorem 3.68 (Rellich’s compactness lemma). Let s ∈ R and let fn be a bounded sequence
in H s (Rm , Cl ) with support in a fixed compact set K ⊂ Rm . Then there is a subsequence
fnl of fn which for all t < s converges in H t (Rm , Cl ).
Proof : We start by noting that for all f ∈ H s (Rm , Cl ) and all g ∈ D(Rm ) the function
F (gf ) is smooth, and that one has ∂ α F (gf ) = (∂ α F g)˜∗F f (exercise). Pick now g ∈ D(Rm )
with g = 1 in K. As we have gfn = fn , it follows from the above observation (with α = 0)
that
Z
0
F fn (ζ) = F g˜∗F fn (ζ) = C F g(ζ − x)(1 + |x|2 )−s/2 F fn (x)(1 + |x|2 )s/2 dx.
48 B. GÜNEYSU

With Cauchy-Schwarz this leads to the bound


Z 1/2
2 2 −s
|F fn (ζ)| ≤ kfn kH s κ(ζ) ≤ Cκ(ζ) := C |F g(ζ − x)| (1 + |x| ) dx ,

where κ is a continuous function. Likewise, one sees by replacing F fn with ∂j F fn in the


above argument and using ∂j F fn = (∂j F g)˜∗F fn ,
|∂j F fn (ζ)| ≤ Cκj (ζ),
where κj is a continuous function. The first inequality implies that F fn is a locally uni-
formly bounded family of continuous functions. As gradient estimates lead to Lipschitz
estimates, the second inequality implies that F fn is a locally equicontinuous family. It
follows from Arzela-Ascoli that there exists a subsequence F fnj , which converges locally
uniformly. We now want to prove that fnj is Cauchy in H t (Rm , Cl ) for all t < s. To this
end, fix  > 0 and let r > 0 be arbitrary (to be chosen later). One has
kfnl − fnk k2H t
Z
= |F fnl (ζ) − F fnk (ζ)|2 (1 + |ζ|2 )s (1 + |ζ|2 )t−s dζ
Rm \Br (0)
Z
+ |F fnl (ζ) − F fnk (ζ)|2 (1 + |ζ|2 )t dζ.
Br (0)

Using Cauchy-Schwarz and the boundedness of fn in H s (Rm , Cl ), one has


Z
|F fnl (ζ) − F fnk (ζ)|2 (1 + |ζ|2 )s (1 + |ζ|2 )t−s dζ ≤ C(1 + r2 )t−s ,
Rm \Br (0)

which is < /2 for some r > 0. Fixing the latter r, the second integral is, with
Z
0
C := (1 + |ζ|2 )t dζ
Br (0)

one has Z
|F fnl (ζ) − F fnk (ζ)|2 (1 + |ζ|2 )t dζ ≤ C 0 sup |F fnl − F fnk |2 ,
Br (0) Br (0)

which is < /2 for large k, l, as F fnj is uniformly convergent in the closure of Br (0). This
completes the proof. 
As we have noted, there is no useful way to define H s on an arbitrary Ω. However, localized
versions of these spaces turn out to be useful:
Definition 3.69. Let s ∈ R.
a) The space Hcs (Ω, Cl ) is given by all T ∈ H s (Rm , Cl ) such that T is compactly supported
in Ω.
b) The space Hlocs
(Ω, Cl ) is given by all T ∈ D 0 (Ω, Cl ) such that φT ∈ Hcs (Ω, Cl ) for all
φ ∈ D(Ω).
Global Analysis 2 49

s
The space HK (Ω, Cl ) of all T ∈ Hcs (Ω, Cl ) with support in the compact K ⊂ Ω is a Banach
space with respect to k·kHs , and Hcs (Ω, Cl ) becomes an LF space with respect to the family
of all seminorms p on Hcs (Ω, Cl ) such that p|HKs (Ω,Cl ) is continuous for all compact K ⊂ Ω.
s
The space Hloc (Ω, Cl ) becomes an F space with respect to the family of seminorms
kφT kH s , φ ∈ D(Ω).
We have continuous inclusions
Hcs (Ω, Cl ) ⊂ Hct (Ω, Cl ), if t > s,
s
Hloc (Ω, Cl ) ⊂ t
Hloc (Ω, Cl ) ⊂ D 0 (Ω, Cl ), if t > s.

Moreover, the isomorphism of linear spaces


(26) E 0 (Ω, Cl ) −→ {T ∈ D 0 (Rm , Cl ) : supp(T ) ⊂ Ω is compact},

(27) T 7−→ φ 7−→ T (φ|Ω ) .
induces continuous linear injective maps
Hcs (Ω, Cl ) ,→ E 0 (Ω, Cl ),
Hcs (Ω, Cl ) ,→ Hloc
s
(Ω, Cl ).

Proposition 3.64.3 implies that D(Ω, Cl ) is dense in Hloc


s
(Ω, Cl ). Moreover, every differential
operator
0
P : E (Ω, Cl ) −→ E (Ω, Cl )
or order ≤ k induces continuous linear maps
0
P : Hcs (Ω, Cl ) −→ Hcs−k (Ω, Cl ),
s s−k 0
P : Hloc (Ω, Cl ) −→ Hloc (Ω, Cl ),

which follows from Corollary 3.65. The Sobolev embedding theorem can be localized as
follows: For all k ∈ N≥0 , s ∈ R with s > k + m/2 the inclusion
D(Ω, Cl ) ⊂ C k (Ω, Cl )
extends to a continuous linear injective map
s
Hloc (Ω, Cl ) ,→ C k (Ω, Cl ).
Finally, the scalar product h·, ·iΩ on D(Ω, Cl ) extends continuously to an anti-dual pairing
between Hlocs
(Ω, Cl ) and Hc−s (Ω, Cl ). The pairing induces an anti-isomorphism of locally
convex spaces
s
Hloc (Ω, Cl )0 ∼
= Hc−s (Ω, Cl ).
Let us present an important application of this theory: the existence of smooth integral
kernels. To this end we record that by the Riesz duality theorem, given a complex Hilbert
50 B. GÜNEYSU

space X and a continuous linear functional T in X there exists a unique ψT ∈ X such that
for all φ ∈ X one has T (φ) = hψT , φi. We also use the fact that the Hilbertian pairing
X × X −→ C, (φ1 , φ2 ) 7−→ hφ1 , φ2 i
is smooth, and that a map from Ω to a Banach space is weakly C k , if and only of it is
strongly C k−1 , where k ∈ N≥1 . For m = 1 this was shown in an exercise, and since being
C k means the existence and continuity of all partial derivatives of order ≤ k, the general
case follows in fact from the m = 1 case.
Theorem 3.70. Assume we are given a linear operator
T : L2 (Ω, Cl ) −→ L2 (Ω, Cl )
for which there exists continuous linear operators
T1 , T2 : L2 (Ω, Cl ) −→ L2 (Ω, Cl ) with T = T1 T2 and T2 = T2∗ ,
and numbers s ∈ R, k ∈ N≥1 with
s > k + m/2, Tj (L2 (Ω, Cl )) ⊂ Hloc
s
(Ω, Cl ) for both j = 1, 2.
Then one has T (L2 (Ω, Cl )) ⊂ C k (Ω, Cl ) and there exists a unique map
T (·, ·) ∈ C k−1 (Ω × Ω, Matl×l (C))
with
Z Z
2
|T (x, y)| dy < ∞, T f (x) = T (x, y)f (x)dy for all x ∈ Ω, f ∈ L2 (Ω, Cl ).
Ω Ω

Proof : I will give a complete proof for the case l = 1. Even more, I will assume that the
function spaces are defined over R, so that I don’t have to care about complex conjugation
in the scalar products etc. The general complex and matrix-valued case is left as an
exercise.
Note first that T is continuous as the Tj ’s are. By the local Sobolev embedding theorem
we have
Tj (L2 (Ω, R)) ⊂ C k (Ω, R),
thus for all f ∈ L2 (Ω, R), one has T f = T1 g, where g = T2 f , and so T f ∈ C k (Ω, R).
Let us now come to T (·, ·): clearly, if T (·, ·) exists, it unique by the fundamental lemma
of distribution theory. For the existence, note that for all open relatively compact sets V
with V ⊂ Ω one has
Tj (L2 (Ω, R)) ⊂ Cbk (V, R).
In fact, the induced maps
L2 (Ω, R) −→ Cbk (V, R), f 7−→ Tj f
are continuous: Indeed, if fn → f in L2 (Ω, R) and Tj fn converges in Cbk (V, R) to some
ψ ∈ Cbk (V, R), then T fn → Tj f in L2 (V, R), thus w.l.o.g Tj fn (x) → T f (x) for a.e. x ∈ V
and so ψ = Tj f . This implies the asserted continuity by the closed graph theorem. The
Global Analysis 2 51

latter continuity in turn implies that for all V as above there is a constant C = C(V ) > 0
such that for all x ∈ V , f ∈ L2 (Ω, R) one has
|Tj f (x)| ≤ C kf kΩ
L2 .

Thus, by the Riesz duality theorem for Hilbert spaces for all x ∈ Ω there exists Tj,x ∈
L2 (Ω, R) such that for all f ∈ L2 (Ω, R) one has
Z
(28) Tj f (x) = Tj,x (y)f (y)dy = hTj,x , f iΩ .

Define now
T (·, ·) : Ω × Ω −→ R
by
T (x, y) := hT1,x , T2,y iΩ .
By (28), the map z 7→ Tj,z is weakly C k , thus strongly C k−1 , so that by the smoothness
of the Hilbertian pairing, the map (x, y) 7→ T (x, y) is C k−1 and so is its adjoint. On the
other hand, in view of (28) one has
T (x, y) = T2 [T1,x ](y),
which is L2 in y. Finally,
Z
T f (x) = T (x, y)f (y)dy for all x ∈ Ω, f ∈ L2 (Ω, R)

follows from a simple calculation using Fubini: one has
Z Z Z
T (x, y)f (y)dy = T1,x (z)T2,y (z)dzf (y)dy,
Ω Ω Ω
while
T f (x) = T1 T2 f (x) = hT1,x , T2 f iΩ
= hT2 T1,x , f iΩ
Z
= T2 T1,x (y)f (y)dy

Z Z
= T2,y (z)T1,x (z)dzf (y)dy.
Ω Ω


4. Pseudodifferential operators on open subsets of Rm


In this section I will mainly follow the presentation given by E. van den Ban and M. Crainic
in their lecture notes [1]. Classical references are [12, 13]. Whenever there is no danger of
confusion, I will from now on denote the distribution Tf which corresponds to a function
f with f again.
52 B. GÜNEYSU

We begin by an intuitive calculation which will justify the ultimate definition of pseudo-
differential operators: assume we are given a differential operator
0
X
P = Pα Dα : D(Ω, Cl ) −→ E (Ω, Cl )
|α|≤k

of order ≤ k. Then for all f ∈ D(Ω, Cl ) ⊂ S (Rm , Cl ), x ∈ Ω, the Fourier inversion


theorem implies Z
1
P f (x) = ei(x,ζ) p(x, ζ)F f (ζ)dζ,
(2π)m Rm
where
X
(29) p(x, ζ) = Pα (x)ζ α ∈ Matl×l0 (C)
|α|≤k

is the so called (full) symbol of P .


Lemma 4.1. In the above situation, for all α, β ∈ Nm
≥0 , K ⊂ Ω, there exists a constant
C = C(K, α, β) > 0 such that
(30) |∂xα ∂ζβ p(x, ζ)| ≤ C(1 + |ζ|)k−|β| for all x ∈ K, ζ ∈ Rm ..
Proof : From formula (29). 
Thus, searching for a natural generalization of the theory of differential operators one is
naturally lead to:
Definition 4.2. For k ∈ R the space S k (Ω, Matl×l0 (C)) of symbols of order k ∈ R is defined
to be the space of all smooth maps
p : Ω × Rm −→ Matl×l0 (C)
which satisfy the bounds (30).
In case l = l0 = 1, we will simply write S k (Ω) and likewise in analogous situations in the
sequel.
The space S k (Ω, Matl×l0 (C)) becomes an F space with respect to the seminorms
(31) sup |∂xα ∂ζβ p(x, ζ)|(1 + |ζ|)|β|−k ,
(x,ζ)∈K×Rn

where α, β runs through Nm


≥0 and K runs through all compact subsets of Ω. We have
continuous inclusions
S k1 (Ω, Matl×l0 (C)) ⊂ S k2 (Ω, Matl×l0 (C)), if k1 ≤ k2 .
One sets
[
S ∞ (Ω, Matl×l0 (C)) := S k (Ω, Matl×l0 (C)),
k∈R
\
S −∞ (Ω, Matl×l0 (C)) := S k (Ω, Matl×l0 (C)).
k∈R
Global Analysis 2 53

Letting k run through R in (31), the space S −∞ (Ω, Matl×l0 (C)) becomes a locally convex
space again.
The following examples are illustrative:
Example 4.3. 1. We have seen that all matrix-valued polynomials of the form
X
p(x, ζ) = P α (x)ζα
|α|≤k

are in S k (Ω, Matl×l0 (C)).


2. The function
p(x, ζ) := (1 + |ζ|2 )k/2

is in S k (Ω), for all k ∈ R.


3. The function
 m
X −1
2
p(x, ζ) := 1 + |ζ| + xi xj ζi ζj
i,j=1

is in S −2 (Ω).
4. The pointwise multiplication
S a (Ω, . . . ) × S b (Ω, . . . ) −→ S a+b (Ω, . . . ), (v, w) 7−→ vw
is well-defined for all a, b ∈ R.
For the following definition, note that every map
A ∈ E (Ω × Ω, Matl×l0 (C))
determines a continuous linear operator
Z
l0
QA : D(Ω, C ) −→ E (Ω, C ),
l
QA f (x) = A(x, y)f (y)dy.

Definition 4.4. An operator of the form QA as above is called a smoothing operator. The
0
linear space of smoothing operators is denoted by Ψ−∞ (Ω; Cl , Cl ).
Remark 4.5. 1. The map
0
E (Ω × Ω, Matl×l0 (C)) −→ L (D(Ω, Cl ), E (Ω, Cl )), A 7−→ QA
is a linear injective map by the fundamental lemma of distribution theory.
2. Nontrivial smoothing operators are never local, that is, for all QA as above
supp(QA f ) ⊂ supp(f ) for all f ∈ D(Ω, Cl ) ⇒ A = 0.
We can now give the definition of pseudodifferential operators:
54 B. GÜNEYSU

Definition 4.6. 1. Let k ∈ R ∪ {−∞}. For every p ∈ S k (Ω, Matl×l0 (C)) the continuous
linear operator
0
Op(p) : D(Ω, Cl ) −→ E (Ω, Cl )
is defined by Z
Op(p)f (x) = ei(x,y) p(x, ζ)F f (ζ)dζ.
Rm
2. Let k ∈ R. A linear operator
0
P : D(Ω, Cl ) −→ E (Ω, Cl )
is called a pseudodifferential operator of order k, if there exists p ∈ S k (Ω, Matl×l0 (C)) and
0
Q ∈ Ψ−∞ (Ω; Cl , Cl ) such that P = Op(p) + Q. The linear space of such operators is
denoted by
0 0
Ψk (Ω; Cl , Cl ) ⊂ L (D(Ω, Cl ), E (Ω, Cl )).
At this point, we only remark that requiring P = Op(p) + Q rather than just P = Op(p) is
needed to prove coordinate invariance later on, while smoothing operators play of course
no role in the local regularity theory (so the price we pay in order to obtain for coordinate
invariance is little).
Remark 4.7. 1. Note that every differential operator of order ≤ k is a pseudodifferential
operator of order k.
2. In view of Remark 4.5.2, pseudodifferential operators need not be local, while of course
differential operators are local. In fact these are essentially all local operators, as Peetre’s
theorem states that a (not necessarily continuous) local linear map
0
D(Ω, Cl ) −→ E (Ω, Cl )
is automatically a differential operator, when restricted to D(U, Cl ), where U ⊂ Ω is open
and relatively compact. However the order of this operator can go to ∞ as U → Ω. This
result implies, e.g, that local differential operators on compact manifolds are automatically
differential operators!
The map p 7→ Op(p) seems to be not injective for general Ω’s (although I am not aware of
a counterexample). For Ω = Rm we can use the Fourier transform to obtain:
Proposition 4.8. The map
0
S k (Rm , Matl×l0 (C)) −→ L (D(Rm , Cl ), E (Rm , Cl )), p 7−→ Op(p)
is injective.
Proof : Assume Op(p)f = 0 for all f ∈ D(Rm , Cl ). Fix x ∈ Rm . Then we have
Z
ei(x,ζ) p(x, ζ)g(ζ)dζ = 0
Rm

for all g ∈ F (D(R , C )). As D(Rm , Cl ) is dense in S (Rm , Cl ) and F is an isomorphism


m l

of locally convex spaces from S (Rm , Cl ) to S (Rm , Cl ), it follows that F (D(Rm , Cl )) is


Global Analysis 2 55

dense11 in S (Rm , Cl ). Thus, given φ ∈ D(Rm , Cl ) arbitrary, we can pick a sequence φn in


F (D(Rm , Cl )) with φn → φ in S (Rm , Cl ). Then we have
Z Z
i(x,ζ)
e p(x, ζ)φ(ζ)dζ = lim ei(x,ζ) p(x, ζ)φn (ζ)dζ = 0,
Rm n Rm

thus by the fundamental lemma of distribution theory we have ei(x,ζ) p(x, ζ) = 0 for all
x, ζ ∈ Rm , thus p = 0. 
0
Given f : Ω → Cl , g : Ω → Cl let
f ⊗ g : Ω −→ Matl×l0 (C)
be the function defined by (f ⊗ g)ij (x, y) := f (i) (x)g (j) (y). Then Schwartz’s kernel theorem
states that that in fact for every continuous linear operator
0
P : D(Ω, Cl ) −→ D 0 (Ω, Cl )
there exists a unique distribution KP ∈ D 0 (Ω×Ω, Matl×l0 (C)), its so called Schwartz kernel,
0
such that for all f ∈ D(Ω, Cl ), g ∈ D(Ω, Cl ) one has
P f (g) = KP (f ⊗ g).
The proof of this result is very technical and can be found in Hoermander’s book. We will
only need this statement for pseudodifferential operators for which we will provide a sketch
of proof:
0
Theorem 4.9. a) Assume k ∈ R and P ∈ Ψk (Ω; Cl , Cl ). Then there exists a unique
0
distribution KP ∈ D 0 (Ω × Ω, Matl×l0 (C)) such that for all f ∈ D(Ω, Cl ), g ∈ D(Ω, Cl ) one
has
P f (g) = KP (f ⊗ g).
0
The distribution KP is called the Schwartz kernel of P . In case P = QA ∈ Ψ−∞ (Ω; Cl , Cl )
one has KQA = A, where the continuous embedding
L1loc (Ω × Ω, Matl×l0 (C)) ,→ D 0 (Ω × Ω, Matl×l0 (C))
is given by Z
 
A 7−→ φ 7−→ tr(A(x, y)∗ φ(x, y))dxdy .
Ω×Ω
b) Let
0
P : D(Ω, Cl ) −→ E (Ω, Cl )
be a continuous linear operator. Then the following statements are equivalent:
0
i) one has P (E 0 (Ω, Cl )) ⊂ E (Ω, Cl ) and the induced map
0
P : E 0 (Ω, Cl ) −→ E (Ω, Cl )
11This is the place where we use Ω = Rm : F (D(Ω, Cl )) is not dense in S (Rm , Cl ) and of course not
even a subspace of D(Rm , Cl ).
56 B. GÜNEYSU

is continuous.
0
ii) For all s, t ∈ N one has P (Hc−s (Ω, Cl )) ⊂ Hloc
t
(Ω, Cl ) and the induced map
0
P : Hc−s (Ω, Cl ) −→ Hloc
t
(Ω, Cl )
is continuous.
0
iii) One has P ∈ Ψ−∞ (Ω; Cl , Cl ).
Proof : a) The uniqueness is seen by showing the density of the span of functions of the
form f ⊗ g in D(Ω × Ω, Matl×l0 (C)) (exercise). For the existence, assume first P = QA is
smoothing. Then one has
Z Z Z
QA f (g) = (P f (x), g(x))dx = (A(x, y)f (y), g(x))dydx
ZΩ Ω Ω

tr A(x, y)∗ (f ⊗ g)(x, y) dxdy.



=
Ω×Ω

It remains to consider the case P = Op(p), where p ∈ S k (Ω, Matl×l0 (Ω). We get the linear
map
(k)
F2 : S k (Ω, Matl×l0 (C)) −→ D 0 (Ω × Rm , Matl×l0 (C))
defined by Z
(k)
F2 q(φ) := tr (q(x, ζ)∗ F2 φ(x, ζ)) dζdx,
Ω×Rm
where F2 denotes the Fourier transform acting on the second variable. Then with the
diffeomorphism
L : Ω × Rm −→ Ω × Rm , L(x, y) := (x, y − x),
the distribution
0 (k)
KP := ((L−1 )∗ ) F2 p |Ω×Ω ∈ D 0 (Ω × Ω, Matl×l0 (C))
does the job, where
(L−1 )∗ : D(Ω × Ω, Matl×l0 (C)) −→ D(Ω × Ω, Matl×l0 (C))
denotes the pullback with respect to L−1 , and
0
((L−1 )∗ ) : D 0 (Ω × Ω, Matl×l0 (C)) −→ D 0 (Ω × Ω, Matl×l0 (C))
its dual.
b) We are going to show iii) ⇔ ii) ⇔ i).
iii) ⇒ ii): Note that for r ∈ N one has
r
Hloc (Ω, Cl ) = {f ∈ L2loc (Ω, Cl ) : ∂ α f ∈ L2loc (Ω, Cl ) for all |α| ≤ r}
r
and that the F topology on Hloc (Ω, Cl ) is equivalently given by the family of seminorms
X
k1K ∂ α f kΩ , where K ⊂ Ω is compact.
|α|≤k

Likewise, one has


Hcr (Ω, Cl ) = {f ∈ L2c (Ω, Cl ) : ∂ α f ∈ L2c (Ω, Cl ) for all |α| ≤ r}
Global Analysis 2 57

and that the LF topology on Hcr (Ω, Cl ) is equivalently given by the family of seminorms
p on Hcr (Ω, Cl ), such that p|HKr (Ω,Cl ) is continuous for all compact K ⊂ Ω. Now let P̃ ∈
0
Ψ−∞ (Ω; Cl , Cl ). Then using that ∂ α P̃ is smoothing with
K∂ α P̃ (x, y) = ∂xα KP̃ (x, y)
and Cauchy-Schwartz one can easily show that ∂ α P , where |α| ≤ s, induces a continuous
linear map
∂ α P̃ : L2c (Ω, Cl ) −→ L2loc (Ω, Cl ),
so that P̃ induces a continuous linear map
P̃ : L2c (Ω, Cl ) −→ Hloc
s
(Ω, Cl ).
By duality,
P̃ : Hc−s (Ω, Cl ) −→ L2loc (Ω, Cl )
continuously. Applying this with P̃ = ∂ β P , where |α| ≤ s, shows that
P : Hc−s (Ω, Cl ) −→ Hloc
t
(Ω, Cl )
continuously.
ii) ⇒ iii): Let k ∈ N be arbitrary. The map
Ω −→ Hc−t (Ω, Cl ), x 7−→ φ 7−→ δx ⊗ ej (φ) := φ(j) (x)


is C k , if we pick t > m/2 + k, and one has


XZ
f= f (j) (x)δx ⊗ ej dx for all f ∈ D(Ω, Cl ),
j Ω

where the integral converges in the strong sense (exercise), and thus commutes with con-
tinuous linear operators. By assumption, the map
Hc−t (Ω, Cl ) × Hc−t (Ω, Cl ) −→ Matl×l0 (C), (f, g) 7−→ hP f, giΩ
Hc−t ,H −t
loc

is the composition of a pairing with a continuous linear map, thus smooth. It follows that
the map
A : Ω × Ω −→ Matl×l0 (C), A(x, y)ij := hP δy ⊗ ei , δx ⊗ ej iΩ
Hc−t ,H −t
loc

k
is C for all k, so smooth. Finally, we can calculate
Z XZ

(P f, g) = hP f, giHc−t ,H −t = hP f, δx ⊗ ej iΩ (j)
Hc−t ,H −t g (x)dx
loc loc
Ω j Ω
XZ
= f (i) (y)A(x, y)ij dyg (j) (x)dx
ij Ω
Z Z
= (A(x, y)f (y), g(x))dydx.
Ω Ω
58 B. GÜNEYSU

Thus, we have shown P = QA .


ii) ⇒ i): The essential observation is that by the local Sobolev embedding theorem one
has
E (Ω, Cl ) ∼
\
r
= Hloc (Ω, Cl )
r∈Z
as locally convex spaces, so that by duality
E 0 (Ω, Cl ) ∼
[
= Hcr (Ω, Cl )
r∈Z
0
as locally convex spaces (exercise). Thus, if fn → f in E 0 (Ω, Cl ), then fn → f in Hcs (Ω, Cl )
for some s0 ∈ Z and so fn → f in Hc−s (Ω, Cl ) for some s ∈ N, and by assumption P fn → P f
t0
t
in Hloc (Ω, Cl ) for all t ∈ N. This implies P fn → P f in Hloc (Ω, Cl ) for all t0 ∈ Z and so
P fn → P f in E (Ω, Cl ).
i) ⇒ ii) If s, t ∈ N and fn → f in Hc−s (Ω, Cl ), then fn → f in E 0 (Ω, Cl ) and P fn → P f in
E (Ω, Cl ) and so P fn → P f in Hloc
t
(Ω, Cl ). 
As an immediate consequence of part b) we get:
Corollary 4.10. There exists a canonical isomorphism of linear spaces
Ψ−∞ (Ω; Cl , Cl ) ∼
0 0
= L (E 0 (Ω, Cl ), E (Ω, Cl )).
Symbols in S −∞ correspond to smoothing operators and vice versa, showing that we could
have allowed k = −∞ in Definition 4.6.2.:
Proposition 4.11. a) For every p ∈ S −∞ (Ω, Matl×l0 (C)) there exists a unique
Ap ∈ E (Ω × Ω, Matl×l0 (C))
0
with Op(p) = QAp , in particular, Op(p) ∈ Ψ−∞ (Ω; Cl , Cl ). If Op(p)|D(Ω\K;Cl ) = 0 for some
compact K ⊂ Ω, then one has supp(A) ⊂ Ω × K.
b) If
A ∈ E (Ω × Ω, Matl×l0 (C)))
satisfies supp(A) ⊂ Ω × K for some compact K ⊂ Ω, then there exists a
p ∈ S −∞ (Ω, Matl×l0 (C))
such that QA = Op(p).
Proof : Exercise. The key observation is that elements of S −∞ are Schwartz functions with
respect to their second variable, and that the Fourier transform on the second variable
(−∞)
F2 : S −∞ (Ω, Matl×l0 (C)) −→ S −∞ (Ω, Matl×l0 (C))
is an isomorphism of locally convex spaces. 
Next, we show that pseudodifferential operators can be localized:
0
Proposition 4.12. Given ψ1 , ψ2 ∈ D(Ω), k ∈ R ∪ {−∞}, P ∈ Ψk (Ω; Cl , Cl ), one has
0 0 0
ψ1 P ∈ Ψk (Ω; Cl , Cl ), P ψ2 ∈ Ψk (Ω; Cl , Cl ), ψ1 P ψ2 ∈ Ψk (Ω; Cl , Cl ).
Global Analysis 2 59

Proof : Clearly it suffices to show


0 0
ψ1 P ∈ Ψk (Ω; Cl , Cl ), P ψ2 ∈ Ψk (Ω; Cl , Cl ).
Case k = −∞: For P = QA one has
ψ1 QA ψ2 = QÃ ,
where Ã(x, y) = ψ1 (x)A(x, y)ψ2 (y).
Case k ∈ R: in view of the k = −∞ case, it remains to show that
0
ψ1 Op(p), Op(p)ψ2 ∈ Ψd (Ω; Cl , Cl ) for all p ∈ S k (Ω, Matl×l0 (C)).
CLearly, ψ1 Op(p) = Op(p̃), where
p̃ ∈ S k (Ω, Matl×l0 (C))
is given by ψ1 (x)p(x, ζ). It remains to prove
0
Op(p)ψ2 ∈ Ψd (Ω; Cl , Cl ).
For all f ∈ D(Ω, Cl ), x ∈ Ω we calculate using the convolution theorem,
Z
Op(p)(ψ2 f )(x) = ei(x,ζ) p(x, ζ)F (ψ2 f )(ζ)dζ
Rm
Z
= ei(x,ζ) p(x, ζ)F ψ2 (ζ)F f (ζ)dζ
Rm
Z Z
1
= ei(x,ζ) p(x, ζ)F ψ2 (ζ − y)F f (y)dydζ
(2π)m Rm Rm
Z Z
1
= ei(x,ζ) ei(x,y) p(x, ζ + y)F ψ2 (ζ)F f (y)dζdy
(2π)m Rm Rm
Z
1
= ei(x,ζ) q(x, ζ)F f (y)dζ,
(2π)m Rm
where Z
1
q(x, ζ) := m
ei(x,ζ) F f (y)p(x, ζ + y)dy.
(2π) Rm
By differentiating under integral one can now straighforwardly check that
q ∈ S k (Ω, Matl×l0 (C)).

Next, we show that the Schwartz kernel of a pseudodifferential operator can only be singular
on the diagonal:
0
Proposition 4.13. Assume k ∈ R ∪ {−∞} and P ∈ Ψk (Ω; Cl , Cl ). Then one has
KP |(Ω×Ω)\diag(Ω) ∈ E ((Ω × Ω) \ diag(Ω), Matl×l0 (C)),
in other words,
sing − supp(KP ) ⊂ diag(Ω).
60 B. GÜNEYSU

Proof : As the Schwartz kernels of smoothing operators are smooth, we can assume P =
Op(p) for some p ∈ S k (Ω, Matl×l0 (C)). Consider again the maps
(k)
F2 : S k (Ω, Matl×l0 (C)) −→ D 0 (Ω × Rm , Matl×l0 (C))

and
L : Ω × Rm −→ Ω × Rm .
from the proof of Theorem 4.9. As we have
0 (k)
KP = ((L−1 )∗ ) F2 p |Ω×Ω ,

and
diag(Ω) = L−1 (Ω × {0})
(k)
it suffices to show that F2 p is smooth away from Ω × {0}, in other words, it suffices to
(k)
show that for all ψ ∈ D(Rm ) with 0 ∈ / supp(ψ) one has that ψ(0) F2 p is smooth, where
for all n ∈ N the function ψ(n) ∈ D(Ω × Rm , Matl×l0 (C)) is given by

ψ(n) (x, ζ)ij := δij |ζ|−2n ψ(ζ),

noting that divisian by |ζ|2n causes no problem, because ψ is supported away from 0. To
this end, one checks as usual that for all n ∈ N one has
(k) (k−2n)
ψ(0) F2 p = ψ(n) F2 (∆n2 p)

where ∆n2 is the n-th power of the Laplace-operator ∆ = m 2


P
j=1 ∂j acting on the second
variable. Let s ∈ N be arbitrary. We are going to prove in a moment that for n large
(k−2n)
enough such that k − 2n < −m − s the map F2 maps into C s (Ω × Rm , Matl×l0 (C)),
which will complete the proof. In order to prove the latter mapping property, we assume
for notational convenience l = l0 = 1. Then for all q ∈ S k−2n (Ω), φ ∈ D(Ω × Rm ), one has
Z
(k−2n)
F2 q(φ) = q(x, ζ)F2 φ(x, ζ)dζdx
Ω×Rm
Z Z Z
0
= q(x, ζ)e−i(ζ ,ζ) dζφ(x, ζ 0 )dxdζ 0
m m
ZΩ R R
= r(x, ζ 0 )φ(x, ζ 0 )dxdζ 0 ,
Ω×Rm

where
Z
0 0
r(x, ζ ) := q(x, ζ)ei(ζ ,ζ) dζ.
Rm
Global Analysis 2 61

Let N ⊂ Ω be a compact set with supp(φ) ⊂ N × N . Using the Leibnitz rule and (30) we
get for all ζ, ζ 0 ∈ Rm , x ∈ N , |α| + |β| ≤ s the estimates
0
|∂xα ∂ζβ0 q(x, ζ)ei(ζ ,ζ) |
0
≤ Cα,β,m |∂xα q(x, ζ)| + Cα,β,m |q(x, ζ)||∂ζβ0 ei(ζ ,ζ) |
0
≤ Cα,β,m,q,N (1 + |ζ|)k−2n + Cα,β,m,q,N (1 + |ζ|)k−2n |∂ζβ0 ei(ζ ,ζ) |
≤ Cα,β,m,q,N (1 + |ζ|)k−2n + Cα,β,m,q,N (1 + |ζ|)k−2n+s .
This shows that one may differentiate under the integral to prove that r is a C s map. 
Let prj : Ω × Ω → Ω, j = 1, 2, denote the projection map onto the j-th slot.
0
Definition 4.14. Assume k ∈ R∪{−∞} and P ∈ Ψk (Ω; Cl , Cl ). Then P is called properly
supported, if the induced projection maps
prj |supp(KP ) : supp(KP ) −→ Ω, j = 1, 2
are proper, that is, pr−1
j (N ) ⊂ supp(KP ) is compact for all compact N ⊂ Ω.

Remark 4.15. 1. P is above is properly supported, if and only if for all compact N ⊂ Ω
one has that
(N × Ω) ∩ supp(KP ), (Ω × N ) ∩ supp(KP )
are compact. Note here a subset N ⊂ supp(KP ) is compact (in the subspace topology), if
and only if it is compact as a subset of Ω.
2. Every differential operator is properly supported: the Schwartz kernel of P = α Pα ∂ α
P
is easily seen to be (in a physicists’ notation)
KP (x, y) = (P δ)(x − y),
whose support is equal to diag(Ω).
3. If P is a pseudodifferential operator and φ1 , φ2 ∈ D(Ω), then φ1 P φ2 is properly sup-
ported.
Properly supported pseudodifferential operators will turn out to have many technical ad-
vantages. It is thus important to know that every pseudodifferential operator can be
approximated by a properly supported one, whose kernel is supported arbitrarily closely
to diag(Ω):
0
Theorem 4.16. Assume k ∈ R ∪ {−∞}, P ∈ Ψk (Ω; Cl , Cl ) and that U ⊂ Ω × Ω is an
open neighbourhood of diag(Ω). Then there is a q ∈ S k (Ω, Matl×l0 (C)) such that
• Op(q) is properly supported,
• supp(KOp(q) ) ⊂ U ,
0
• P − Op(q) ∈ Ψ−∞ (Ω; Cl , Cl ).
The proof relies on:
62 B. GÜNEYSU

Lemma 4.17. Assume U ⊂ Ω × Ω is an open neighbourhood of diag(Ω). Then there exists


an open relatively compact locally finite covering (Vi )i∈I of Ω and a smooth partition of
unity (φi )i∈I subordinate to (Vi )i∈I such that for all i, j ∈ I with
supp(φi ) ∩ supp(φj ) 6= ∅
one has
supp(φi ) × supp(φj ) ⊂ U.
Proof : Start from any open relatively compact cover W 00 of Ω. Pick a locally finite refine-
ment W 0 of W 00 such that W × W ⊂ U for all W ∈ W 0 . Then for all x ∈ Ω there exists
an open nbh Wx of x such that Wx ⊂ W for all W ∈ W 0 with x ∈ W and Wx ∩ W = ∅ for
all W ∈ W 0 with x ∈/ W . Then any locally finite refinement V = (Vi )i∈I of the open cover
(Wx )x∈X and any partition of unity (φi )i∈I subordinate to V do the job. 
Proof of Theorem 4.16: We can and we will assume that the smoothing part of P is zero.
By the Lemma we can pick a family of functions (φi )i∈I ⊂ D(Ω) such that (supp(φi ))i∈I is
a locally finite collection of sets, and
X
φi = 1,
i∈I
and for all i, j ∈ I with
supp(φi ) ∩ supp(φj ) 6= ∅
one has
supp(φi ) × supp(φj ) ⊂ U.
Since the smoothing part of P is zero, With Pij := φi P φj , one finds that for all i, j there
exists qij ∈ S k (Ω, Matl×l0 (C)) such that Pij = Op(qij ). One has
Kij := KPij = (φi ⊗ φj )KP ,
where φi ⊗ φj acts as a scalar. It follows that
supp(Kij ) ⊂ supp(φi ) × supp(φj ).
If
supp(φi ) ∩ supp(φj ) = ∅,
then Kij is smooth away from diag(Ω) and supported away from diag(Ω), so smooth on
Ω × Ω. If
supp(φi ) ∩ supp(φj ) 6= ∅,
then
supp(Kij ) ⊂ U.
Set
J := {(i, j) ∈ I × I : supp(φi ) ∩ supp(φj ) 6= ∅}.
Then one has
pr1 (supp(qij )) ⊂ supp(φi ),
so that
pr1 (supp(qij )), (i, j) ∈ J
Global Analysis 2 63

is a locally finite collection of sets. Thus


X
q := qij ∈ S k (Ω, Matl×l0 (C)),
(i,j)∈J

while X
K := Kij ∈ E (Ω × Ω, Matl×l0 (C)).
(i,j)∈(I×I)\J

Then one easily finds P = Op(q) + QK , and


X
KOp(q) = Kij ,
(i,j)∈J

in particular,
supp(KOp(q) ) ⊂ U.
Finally, it remains to show that Op(q) is properly supported: given A ⊂ Ω compact, the
set
JA := {(i, j) ∈ J : supp(φi ) ∩ A 6= ∅}
is finite. Then
[
(A × Ω) ∩ supp(KOp(q) ) ⊂ supp(φi ) × supp(φj ),
(i,j)∈JA

so that (A × Ω) ∩ supp(KOp(q) ) is compact. The compactness of (Ω × A) ∩ supp(KOp(q) ) is


seen analogously. 
Definition 4.18. Given a countable set N , a function N → R, n 7→ kn , with kn → −∞
as n → ∞12, k ∈ R, a symbol p ∈ S k (Ω, Matl×l0 (C)), and for each n ∈ N a symbol
pn ∈ S kn (Ω, Matl×l0 (C)). Then one writes
X
p∼ pn ,
n∈N
0
if for all k ∈ R there exists a finite set F0 ⊂ N such that for all finite subsets F ⊂ N with
F0 ⊂ F one has
0
X
p− pn ∈ S k (Ω, Matl×l0 (C)).
n∈F
P
One then calls n∈N pn an asymptotic expansion of p.

Remark 4.19. If p, q ∈ S k (Ω, Matl×l0 (C)) and


X X
p∼ pn , q ∼ pn ,
n∈N n∈N

then one has p − q ∈ S −∞ (Ω, Matl×l0 (C)). Thus any asymptotic expansion of p determines
p modulo S −∞ (Ω, Matl×l0 (C)).
12that is, for all r ∈ R there exists a finite set F ⊂ N such that for all n ∈ N \ F one has kn < r.
64 B. GÜNEYSU

Proposition 4.20. Let k ∈ R, p ∈ S k (Ω, Matl×l0 (C)), φ ∈ D(Ω) and define q ∈ S k (Ω, Matl×l0 (C))
by
Z
q(x, ζ) := ei(x,y) p(x, y + ζ)F φ(y)dy.
Rm
Then one has
X 1
q∼ Dxα φ∂yα p.
α∈Nm
α!
≥0

Proof : Exercise. The only idea one needs is to use the multidimensional Taylor formula
with remainder to rewrite p(x, y + ζ). 

Given k ∈ R ∪ {−∞} and a symbol q ∈ S k (Ω, Matl×l0 (C)) and Ω0 ⊂ Ω open we define
qΩ0 ∈ S k (Ω0 , Matl×l0 (C)) by qΩ0 := q|Ω0 ×Rm .
Corollary 4.21. Let k ∈ R, p ∈ S k (Ω, Matl×l0 (C)), let Ω0 ⊂ Ω open with Ω0 ⊂ Ω compact,
φ ∈ D(Ω) with φ = 1 in an open neighbourhood of Ω0 . Then there is a q ∈ S k (Ω, Matl×l0 (C))
with
Op(q) = Op(p)φ
and
qΩ0 − pΩ0 ∈ S −∞ (Ω0 , Matl×l0 (C)).
Proof : We can assume k ∈ R. Define q as in Proposition 4.20. Then Op(q) = Op(p)φ by
the proof of Proposition 4.12, and Proposition 4.20 implies
X 1
qΩ0 ∼ (Dxα φ∂yα p)Ω0 ,
α∈Nm
α!
≥0

0
but in view of Dα φ = 0 for all α ∈ Nm
≥0 \ {0} and φp = p in Ω one trivially
X 1
pΩ 0 ∼ (Dxα φ∂yα p)Ω0 ,
α∈Nm
α!
≥0

thus
qΩ0 − pΩ0 ∈ S −∞ (Ω0 , Matl×l0 (C))
by Remark 4.19. 

Now we can prove the following structure theorem:


Theorem 4.22. Let k ∈ R. Then the map p 7→ Op(p) induces an isomorphism
0 0
(32) S k (Ω, Matl×l0 (C))/S −∞ (Ω, Matl×l0 (C)) −→ Ψk (Ω; Cl , Cl )/Ψ−∞ (Ω; Cl , Cl )
of linear spaces.
Global Analysis 2 65

Proof : Clearly the map p 7→ Op(p) induces a surjective linear map


0 0
S k (Ω, Matl×l0 (C)) −→ Ψk (Ω; Cl , Cl )/Ψ−∞ (Ω; Cl , Cl ).
Thus it remains to prove that the kernel of the latter map is S −∞ (Ω, Matl×l0 (C)), which
follows if we can show that given an arbitrary p ∈ S k (Ω, Matl×l0 (C)) with Op(p) ∈
0
Ψ−∞ (Ω; Cl , Cl ) one has p ∈ S −∞ (Ω, Matl×l0 (C)). Pick
A ∈ E (Ω × Ω, Matl×l0 (C))
with Op(p) = QA . We will assume for the moment that there exists a compact set N ⊂ Ω
with
(33) supp(p) ⊂ N × Rm , supp(A) ⊂ N × Rm .
Assuming (33), we are going to prove that for an arbitrary Ω0 ⊂ Ω open with Ω0 ⊂ Ω
compact, one has pΩ0 ∈ S −∞ (Ω0 , Matl×l0 (C)). To this end, note that by (33) we have
m m
pR ∈ S k (Rm , Matl×l0 (C)), where pR denotes the extension by zero of p to Rm × Rm . In
0
particular, Op(pR ) ∈ Ψk (Rm ; Cl , Cl ) coincides with Op(p) on D(Ω, Cl ). Pick φ ∈ D(Ω)
m

with φ = 1 in an open neighbourhood of Ω0 . Then one has


m 0
(34) Op(pR )φ = Op(p)φ = QA(1⊗φ) ∈ Ψ−∞ (Rm ; Cl , Cl ),
which, as A(1 ⊗ φ) is compactly supported in Ω × Ω, implies by Proposition 4.11 b) the
existence of some
(35) r ∈ S −∞ (Rm , Matl×l0 (C))
with
(36) QA(1⊗φ) = Op(r).
Applying Corollary 4.21 shows
m
(37) Op(pR )φ = Op(q),
where
q ∈ S k (Rm , Matl×l0 (C))
and
m
qΩ0 − pRΩ0 ∈ S −∞ (Rm , Matl×l0 (C)).
It follows from (34), (36), (37) that Op(r) = Op(q) and so p = q, as the quantization map
is injective on Rm . Thus
pΩ0 − rΩ0 = pΩ0 − qΩ0 ∈ S −∞ (Ω0 , Matl×l0 (C)),
which completes the proof in view of (35) under (33).
0
It remains to remove (33): so assume p ∈ S k (Ω, Matl×l0 (C)) with Op(p) ∈ Ψ−∞ (Ω; Cl , Cl )
and pick
A ∈ E (Ω × Ω, Matl×l0 (C))
66 B. GÜNEYSU

with Op(p) = QA . Given Ω0 ⊂ Ω open with Ω0 ⊂ Ω compact, pick χ ∈ D(Ω) with χ = 1 in


an open neighbourhood of Ω0 . Then χp and (χ ⊗ 1)A satisfy (33) and Op(χp) = Q(χ⊗1)A
is smoothing, too. Thus by the previous case we have
pΩ0 = (χp)Ω0 ∈ S −∞ (Ω0 , Matl×l0 (C)).
This completes the proof. 
Definition 4.23. Let k ∈ R.
1. The full symbol map of order k is the isomorphism of linear spaces
0 0
(38) σ : Ψk (Ω; Cl , Cl )/Ψ−∞ (Ω; Cl , Cl ) −→ S k (Ω, Matl×l0 (C))/S −∞ (Ω, Matl×l0 (C))
given by the inverse of the map (32).
2. The principal symbol map of order k is the isomorphism of linear spaces
0 0
(39) σk : Ψk (Ω; Cl , Cl )/Ψk−1 (Ω; Cl , Cl ) −→ S k (Ω, Matl×l0 (C))/S k−1 (Ω, Matl×l0 (C))
which is induced by the map (38).
The following ’completeness result’ will be the key tool in the construction of parametri-
ces for elliptic pseudodifferential operators, for it will replace the typical Neumann-series
argument from functional analysis:
Theorem 4.24 (Asymptotic completeness lemma). Let (kj )j∈N≥0 ⊂ R be a sequence with
kj & −∞ as j → ∞, and assume for every j we are given a symbol pj ∈ S kj (Ω, Matl×l0 (C)).
Then there exists a symbol p ∈ S k0 (Ω, Matl×l0 (C)) with the following properties:
• one has

[
pr1 (supp(p)) ⊂ pr1 (supp(pj ))
j=0

• one has

X
p∼ pj .
j=0

Moreover, by Remark 4.19, such a p is uniquely determined modulo S −∞ (Ω, Matl×l0 (C)).
Lemma 4.25. Let (kj )j∈N≥0 ⊂ R be a sequence and assume for every j we are given a
symbol pj ∈ S kj (Ω, Matl×l0 (C)). Assume furthermore that χ ∈ E (Rm ) with χ = 0 on B1 (0)
and χ = 1 on Rm \ B2 (0) and for all t > 0 define
χt ∈ E (Ω × Rm ), χt (x, ζ) := χ(tζ).
Then for all j ∈ N≥0 , α, β ∈ Nm ≥0 , N ⊂ Ω compact, there exists Cj,α,β,N > 0 with

α β
(40) ∂x ∂ζ (χt pj )(x, ζ) ≤ Cj,α,β,N (1 + |ζ|)kj −|β| for all x ∈ N , ζ ∈ Rm , 0 < t ≤ 1

Global Analysis 2 67

Proof : Fix x ∈ N , 0 < t ≤ 1.


For all ζ ∈ Rm with |ζ| ≤ t−1 one has χt (x, ζ) = 0 and the asserted estimate holds trivially.
For all ζ ∈ Rm with |ζ| ≥ 2t−1 one has χt (x, ζ) = 1 and the asserted estimate follows from
pj ∈ S −kj (Ω, Matl×l0 (C).
For all ζ ∈ Rm with t−1 < |ζ| < 2t−1 we infer from the Leibnitz formula and
pj ∈ S −kj (Ω, Matl×l0 (C)
the following inequalities,

α β
∂x ∂ζ (χt pj )(x, ζ)

X γ
α β−γ
≤ Cβ ∂ζ χt (x, ζ)∂x ∂ζ pj (x, ζ)

γ≤β
X
|γ| γ α β−γ
= Cβ t ∂ζ χ(tζ)∂x ∂ζ pj (x, ζ)

γ≤β
 X
γ
≤ Cβ max ∂ζ χ ∞ t|γ| Cj,α,γ,N (1 + |ζ|)kj −|β|+|γ|
γ≤β
γ≤β
 X
γ
≤ Cβ max ∂ζ χ ∞
t|γ| Cj,α,γ,N (1 + |ζ|)kj −|β|+|γ| .
γ≤β
γ≤β

Because of t ≤ 1, |ζ| ≤ 2t−1 one has


t|γ| = 3|γ| (3t−1 )−γ ≤ 3|γ| (1 + 2t−1 )−γ ≤ 3|γ| (1 + |ζ|)−γ ,
so that the above can be further estimated by
≤ Cj,α,β,χ,N (1 + |ζ|)kj −|β| ,
completing the proof.

Lemma 4.26. Let (kj )j∈N≥0 ⊂ R be a sequence with kj & −∞ as j → ∞, and assume
for every j we are given a symbol pj ∈ S kj (Ω, Matl×l0 (C)). Assume furthermore that
χ ∈ E (Rm ) with χ = 0 on B1 (0) and χ = 1 on Rm \ B2 (0). Then there exists a sequence
(tj ) ⊂ (0, ∞) such that
• tj → 0 as j → ∞,
• for all l ∈ N≥0 , α, β ∈ Nm
≥0 the series
X
(1 + |ζ|)|β|−kl ∂xα ∂ζβ (χtj pj )
j≥l

converges absolutely and uniformly in any subset of Ω × Rm of the form N × Rm ,


where N ⊂ Ω is compact.
68 B. GÜNEYSU

Proof : We can assume that the kj0 s are negative. Pick a sequence of open subsets (Ωj ) of
Ω with compact closure such that
[
Ωj = Ω, Ωj ⊂ Ωj+1 .
j∈N

Pick Cj,α,β,Ωj as in the above lemma and tj > 0 with

max(Cj,α,β,Ωj , 1)(1 + t−1


j )
kj −k0
< 2−j ,
and so
max(Cj,α,β,Ωj , 1)(1 + t−1
j )
kj −kl
< 2−j for all l ≤ j;
then we have tj → 0 as j → ∞. Let x ∈ Ωj , ζ ∈ R.
If |ζ| ≥ t−1
j , then we have

α β
∂x ∂ζ (χtj pj )(x, ζ)

≤ max(Cj,α,β,Ωj , 1)(1 + |ζ|)kj −|β|


≤ max(Cj,α,β,Ωj , 1)(1 + |ζ|)kj −kl (1 + |ζ|)kl −|β|
≤ max(Cj,α,β,Ωj , 1)(1 + t−1
j )
kj −kl
(1 + |ζ|)kl −|β|
≤ 2−j (1 + |ζ|)kl −|β| .
If |ζ| < t−1
j , then we have χtj (x, ζ) = 0 so the above estimate remains trivially valid.
If x ∈ N , ζ ∈ Rm , we can pick j0 with N ⊂ Ωj0 . Thus for all j ≥ j0 one has

α β
∂x ∂ζ (χtj pj )(x, ζ) ≤ 2−j (1 + |ζ|)kl −|β| ,

which completes the proof. 


Proof of Theorem 4.24: Pick χ, (tj ) as in the previous lemma. Because of
supp(χtj ) ∩ (Ω × B1/tj (0)) = ∅
and 1/tj → ∞ as j → ∞, the sum

X
p(x, ζ) := χtj (x, ζ)pj (x, ζ)
j

is locally finite and thus defines a smooth function


p : Ω × Rm → Matl×l0 (C).
Because of the previous lemma, for all l ∈ N the series
X
rl := χtj pj
j≥l
Global Analysis 2 69

converges absolutely with respect to all seminorms on S kl (Ω, Matl×l0 (C)) so that by the
completeness13 of the latter space one gets
(41) rl ∈ S kl (Ω, Matl×l0 (C)),
in particular,
p = r0 ∈ S k0 (Ω, Matl×l0 (C)) = S k (Ω, Matl×l0 (C)).
It remains to show

X
(42) p∼ pj
j=0

To this end, note that


l−1
X l−1
X
p− pj = (χtj − 1)pj + rl .
j=0 j=0

As l−1 m
P
j=0 (χtj −1)pj is a smooth function on Ω×R with compact support in the ζ-variable,
this function is an element of
S −∞ (Ω, Matl×l0 (C)) ⊂ S kl (Ω, Matl×l0 (C)),
which together with (41) implies
l−1
X
p− pj ∈ S kl (Ω, Matl×l0 (C)).
j=0

This shows (42), as kl → −∞ as l → ∞.


Finally, the asserted support property follows from examining the construction of p as a
sum of the form φj pj , where φj is constant in the x-slot. 

We now prove that the formal adjoint of a pseudodifferential operator is again in that class:
0
Theorem 4.27. Let k ∈ R ∪ {−∞}. Then for every P ∈ Ψk (Ω; Cl , Cl ) there exists a
0
unique P † ∈ Ψk (Ω; Cl , Cl ) such that
Z Z
0
(43) (P f, g) = (f, P † g) for all f ∈ D(Ω, Cl ), g ∈ D(Ω, Cl ).

Moreover, if σ(P ) is represented by a and σ(P † ) is represented by b, then


X 1
b∼ Dxα ∂ζα a† .
α∈Nm
α!
≥0

13A P
sequence j aj in a complete locally convex space X converges, if for all continuous seminorms ν
P
on X one has j ν(aj ) < ∞. Completeness is essential here.
70 B. GÜNEYSU

Note that (43) defines a linear operator


0
P † : D(Ω, Cl ) −→ E (Ω, Cl )
which is clearly uniquely determined by its defining property. The point of Theorem 4.27 is
to show that P † is a pseudodifferential operator of order k and to determine its full symbol.
We prepare the proof with the following auxiliary result, which contains the whole algebra
behind the argument:
Lemma 4.28. Define
e(Dx ,∂ζ ) := F −1 ei(x,ζ) F : S 0 (Rm × Rm , Matl×l0 (C)) −→ S 0 (Rm × Rm , Matl×l0 (C))
and let k ∈ R ∪ {−∞}. Then the following statements hold true:
i) Upon restriction, T induces a map
T : Sck (Ω, Matl×l0 (C)) −→ S k (Ω, Matl×l0 (C)),
where Sck is understood to be the space of symbols having a compact support in the x-
variable.
ii) If p ∈ Sck (Ω, Matl×l0 (C)), then one has Op(p)† = Op(q), where q ∈ S k (Ω, Matl×l0 (C)) is
given by q = T p† (T applied to the ajoint of the matrix p), and one has
X 1
q∼ Dxα ∂ζα p† .
α∈Nm
α!
≥0

Proof : Exercise. 
Given the lemma, the general case reduces to a partition of unity argument (which is,
however, a little subtle):
Proof of Theorem 4.27: Since smoothing operators obviously have formal adjoints (Q†A is
induced by A(y, x)† ) and since every pseudodifferential operator is equal to a properly
supported quantization up to a smoothing error, we can assume that P is properly sup-
ported and that there exists p ∈ S k (Ω, Matl×l0 (C)) with P = Op(p). If p was compactly
supported we could apply the previous lemma directly. To do so, we pick a partition of
unity (ψj )j∈N ⊂ D(Ω) and for each j set
Pj := ψj P, Kj := KPj .
Then Pj is properly supported, too and Pj = Op(pj ), where now
pj := (ψj ⊗ 1)p ∈ Sck (Ω, Matl×l0 (C)).
Now we can use the lemma to pick qj ∈ S k (Ω, Matl0 ×l (C)) with Op(pj )† = Op(qj ) and
X 1
qj ∼ Dxα ∂ζα p†j .
α∈N
α!
As we have (qj corresponds to the adjoint of Pj )
pr1 (supp(qj )) ⊂ pr2 (supp(Kj )),
Global Analysis 2 71

the family (pr1 (supp(qj )))j is locally finite, so that we can define
X
q := qj ∈ S k (Ω, Matl0 ×l (C))
j
P
Using p = j pj it is now easily checked that
X 1
q∼ Dxα ∂ζα p† ,
α∈Nm
α!
≥0

and
P † = Op(p)† = Op(q),
finishing the proof. 
l0
It follows that for every P ∈ Ψk (Ω; Cl , C ) we can define the continuous linear map
0
P : E 0 (Ω, Cl ) −→ D 0 (Ω, Cl )
by
P f (g) = f (P † g).
0
Lemma 4.29. a) Assume k ∈ {−∞} ∪ R and P ∈ Ψk (Ω; Cl , Cl ). The P is properly
supported, if and only if P and P † map
0 0
P : D(Ω, Cl ) −→ D(Ω, Cl ), P † : D(Ω, Cl ) −→ D(Ω, Cl ).
continuously. In particular, P is properly supported, if and only if P † is properly supported.
0 00 0
b) Assume k, r ∈ R ∪ {−∞} and that P ∈ Ψk (Ω; Cl , Cl ), Q ∈ Ψr (Ω; Cl , Cl ) are properly
supported. Then P Q is properly supported.

Proof : Exercise. 
l0
By density, it follows that every properly supported P ∈ Ψk (Ω; Cl , C ) also maps
0
P : E (Ω, Cl ) −→ E (Ω, Cl ).
In particular, we can define the continuous linear maps
0 0
P : E 0 (Ω, Cl ) −→ E 0 (Ω, Cl ), P : D 0 (Ω, Cl ) −→ D 0 (Ω, Cl ),
both by
P f (g) = f (P † g).
Next we want to establish mapping properties that will play a crucial role in the proof
of local elliptic regularity. To this end, we recall that Schur’s test, which states that if
(Xj , µj ), j = 1, 2, are measure spaces and
k : X1 × X2 −→ Matl×l0 (C))
is Borel measurable with
Z Z
C1 := sup |k(x, y)|dµ2 (y) < ∞, C2 := sup |k(x, y)|dµ1 (x) < ∞,
x∈X1 X2 y∈X2 X1
72 B. GÜNEYSU

then Z
Kf (x) := k(x, y)f (y)dµ2 (y)
X2
is a bounded linear operator
p
K : L2 (X2 , µ2 ) −→ L2 (X1 , µ1 ), with kKk ≤ C1 C 2 .
Applying this with Xj = Rm , dµj (x) = dx, and k satisfying
|k(x, y)| ≤ C(1 + |x − y|)−p for some p > m,
we get
Z
(44) kKk ≤ C (1 + |x|)−p dx.
Rm
0
Theorem 4.30. Assume k, s ∈ R and P ∈ Ψk (Ω; Cl , Cl ).
a) P maps
s−k 0
P : Hcs (Ω, Cl ) −→ Hloc (Ω, Cl )
continuously.
b) If P is properly supported, then P maps
s s−k 0
P : Hloc/c (Ω, Cl ) −→ Hloc/c (Ω, Cl )
continuously.
Proof : WLOG l = l0 = 1. Also, using a partition of unity, resp., cut-off function arguments,
all asserted results can be deduced [13] from the following result: assume k ∈ R, p ∈
Sck (Rm ). Then
Op(p) : H s (Rm ) −→ H s−k (Rm )
continuously. To prove the latter mapping property, we note that we have the unitary
maps
A : H s (Rm ) −→ L2 (Rm , (1 + | · |2 )s ), A := F,
B : L2 (Rm , (1 + | · |2 )s ) −→ L2 (Rm ), f 7−→ (1 + | · |2 )s/2 f,
C : H s−k (Rm ) −→ L2 (Rm , (1 + | · |2 )s−k ), C := F,
D : L2 (Rm , (1 + | · |2 )s−k ) −→ H s−k (Rm ), f 7−→ (1 + | · |2 )(s−k)/2 f.
By a simple calculation using Fubini (here we use that p has a compact support in its
x-slot), one finds that the operator
^ := COp(p)A−1
Op(p)
has the integral kernel Z
0
k (τ, ζ) := ei(x,ζ−τ ) p(x, ζ)dx.
Rm
Thus it remains to show that the integral kernel
k(τ, ζ) := (1 + |τ |)s−k k 0 (τ, ζ)(1 + |ζ|)−s
Global Analysis 2 73

defined a bounded operator on L2 (Rm ). To this end, for any multiindex α and any ζ, τ ∈
Rm , have
Z
α 0
i(x,ζ−τ ) α

|τ k (τ, ζ)| = e
∂x p(x, ζ)dx ≤ |K|Cα,K,p (1 + |ζ|)k .
K
Thus, for all N ∈ N≥0 ,
|k 0 (τ, ζ)| ≤ C(1 + |ζ − τ |)−N (1 + |ζ|k ).
This implies
|k(τ, ζ)| ≤ C(1 + |τ |)s−k (1 + |ζ − τ |)−N (1 + |ζ|)k−s .
By Peetre’s inequality, this implies
|k(τ, ζ)| ≤ C(1 + |ζ − τ |)−|k−s|−N .
Since N was arbitrary, the asserted L2 -boundedness now follows from (44) from taking N
large enough that |k − s| + N > m. 

Theorem 4.27 has a simple but nevertheless important consequence: pseudofifferential


operators are psuedo-local:
0
Corollary 4.31. Let k ∈ R ∪ {−∞}. Then every P ∈ Ψk (Ω; Cl , Cl ) is pseudo-local, that
is, one has
0
sing − supp(P u) ⊂ sing − supp(u) for all u ∈ E 0 (Ω, Cl ).
Proof : The only essential observation is that P † has a Schwartz kernel KP † which is smooth
away from the diagonal. The rest is routine: let
x ∈ Ω \ sing − supp(u),
let Ω1 be an open nbh of x and let Ω2 be an open nbh of sing − supp(u) such that Ω1 ∩Ω2 =
∅. It now suffices to show that for all real-valued φ1 ∈ D(Ω1 ) and all real-valued φ2 ∈ D(Ω2 )
with φ2 = 1 in an open nbh of sing − supp(u) one has that φ1 (P φ2 u) is smooth. To this
end, define a function
A : Ω × Ω −→ Matl0 ×l (C),
A(x, y) := φ1 (x)φ2 (y)KP † |(Ω×Ω)\diag(Ω) (x, y),
which is indeed a well-defined smooth function, as φ1 ⊗ φ2 is supported away from (Ω ×
0
Ω) \ diag(Ω), the region where KP † is smooth. Now we have, for every f ∈ D(U, Cl ),
(45) [φ1 (P φ2 u)](f ) = u(φ2 P † φ1 f ) = u(QA f ) = [QA† u](f ),
so that φ1 (P φ2 u) = QA† u, the latter being a smooth function. 

Our next aim is to prove that the composition of two properly supported pseudodifferential
operators is a pseudodifferential operator. The main technical tool behind this result is:
74 B. GÜNEYSU

Proposition 4.32. Let k, r ∈ R ∪ {−∞}, l, l0 , l00 ∈ N, and


p ∈ S k (Ω, Matl0 ×l00 (C)), q ∈ Scr (Ω, Matl×l0 (C)).
Then one has Op(p)Op(q) = Op(r), where (cf. Lemma 4.28)
r ∈ S r+k (Ω, Matl×l00 (C))
is given by
r(x, ζ) = e(Dy ,∂ζ ) p(x, ζ)q(y, η)|y=x,η=ζ .
In particular, one has pr1 (supp(r)) ⊂ pr1 (supp(p)) and (cf. exercise sheet 11)
X 1
r∼ ∂ζα pDxα q.
α∈Nm
α!
≥0

Proof : Step 1: Let us first assume


p ∈ D(Ω × Rm , Matl0 ×l00 (C)), q ∈ D(Ω × Rm , Matl×l0 (C)),
a case which in fact shows the whole algebra behind the result. Considering p and q as
smooth compactly supported functions on R2m in the obvious way, for all f ∈ D(Ω, Cl )
one can calculate, using Fubini,
Z Z
(46) Op(p)Op(q)f (x) = ei(ζ,x−y) p(x, ζ)Op(q)f (y)dydζ
Z Z Z
(47) = ei(ζ,x−y) ei(y,ζ) p(x, ζ)q(y, η)F f (η)dηdydζ
Z
(48) = ei(η,x) r(x, η)F f (η)dη,

where
Z Z
(49) r(x, η) := ei(ζ−η,x−y) p(x, ζ)q(y, η)F f (η)dζdy.

For fixed (x, η) ∈ R2m define R ∈ D(Rm × Rm , Matl×l00 (C)) by


Rx,η (y, ζ) = p(x, ζ)q(y, η).
Using that for the bounded smooth function u(y, ζ) := e−i(ζ,y) on R2m one has F u(y, ζ) =
ei(ζ,y) (using the convolution theorem), using the convolution theorem, and using the defi-
nition of the operator e(Dy ,∂ζ ) , we get
F e(Dy ,∂ζ ) p(x, ζ)q(y, η)|y=x,η=ζ = F e(Dy ,∂ζ ) Rx,η (x, η) = ei(x,η) F Rx,η (x, η)
= F u · F Rx,η (x, η) = F (u ∗ Rx,η )(x, η),
and so by applying F −1 to this equation,
e(Dy ,∂ζ ) p(x, ζ)q(y, η)|y=x,η=ζ = e−i(x,ζ) ∗ Rx,η (x, η) = r(x, η),
where the latter formula follows immediately from (49). This proves the claim in this case.
Global Analysis 2 75

Step 2: If
p ∈ Sck (Ω, Matl0 ×l00 (C)), q ∈ Scr (Ω, Matl×l0 (C)),
we use that the map
ρ : Sck (Ω, Matl0 ×l00 (C)) × Scr (Ω, Matl×l0 (C)) −→ Sck+r (Ω, Matl×l00 (C))
given by
ρ(p0 , q 0 )(x, ζ) := e(Dy ,Dζ ) p0 (x, ζ)q 0 (y, η)|y=x,η=ζ
is bilinear and continuous, and that for fixed f ∈ D(Ω, Cl ), the map
Sck+r (Ω, Matl×l00 (C)) −→ E (Ω, Cl ), r0 7−→ Op(r0 )f
is linear and continuous. Then the asserted formula for r follows from the density of
D(Ω × Rm , Matl0 ×l00 (C)) in Sck (Ω, Matl×l00 (C)) in the Sck+1 (!) topology, the density of the
density of D(Ω × Rm , Matl×l0 (C)) in Scr (Ω, Matl×l0 (C)) in the Scr+1 topology, and Step 1.
Step 3: In the general case, with a partition of unity (ψj ) associatedP to a countable relatively
compact open cover of Ω we can write p as a locally finite sum j ψj p with
ψj p ∈ Sck (Ω, Matl0 ×l00 (C)) for all j.
By step 2, for all j we have Op(ψj p)Op(q) = Op(rj ), where
rj (x, ζ) = e(Dy ,∂ζ ) (ψj p)(x, ζ)q(y, η)|y=x,η=ζ .
The latter formula
P implies that (pr1 (supp(rj )))j is a locally finite collection of subsets of
Ω, so that r := ψj r defines an element
r ∈ Sck+r (Ω, Matl×l00 (C))
which satisfies
r(x, ζ) = e(Dy ,∂ζ ) p(x, ζ)q(y, η)|y=x,η=ζ .

We continue with auxiliary results:
0
Lemma 4.33. For all k ∈ R ∪ {−∞} and all properly supported P ∈ Ψk (Ω; Cl , Cl ) there
exists a symbol
p ∈ S k (Ω, Matl×l0 (C))
with P = Op(p).
Proof : Exercise. One first assumes k = −∞ and uses Proposition 4.11 together with a
partition of unity argument. The general case then follows easily from Proposition 4.16. 
0 00 0
Theorem 4.34. Assume k, r ∈ R∪{−∞} and that P ∈ Ψk (Ω; Cl , Cl ), Q ∈ Ψr (Ω; Cl , Cl )
00
are properly supported. Then one has P Q ∈ Ψk+r (Ω; Cl , Cl ). If σ(P ) is represented by a,
σ(Q) is represented by b and σ(P Q) by c, then one has
X 1
c∼ ∂ζα aDxα b,
α∈N
α!
76 B. GÜNEYSU

in particular,
σ k+r (P Q) = σ k (P )σ r (Q).
Proof : By Lemma 4.33, we can pick
p ∈ S k (Ω, Matl0 ×l00 (C)), q ∈ S r (Ω, Matl×l0 (C))
with P = Op(p) and Q = Op(q). Pick also a partition of unity (ψj ) associated to an
open P
relatively compact cover of Ω and set pj := ψj p so that with Pj := Op(pj ) one has
P = j Pj . As P is properly supported, the sets

Aj := pr2 pr1−1 (supp(ψj )) ∩ supp(KP ) ⊂ Ω




are compact, so that we can pick χj ∈ D(Ω) with χj = 1 in Aj . Since


KPj χj = (ψj ⊗ χj )KP = (ψj ⊗ 1)KP

it follows that Pj χj = Pj and so Pj Q = Pj Qj , where Qj = Op(qj ) with


qj := χj q ∈ Scr (Ω, Matl×l0 (C)).
It follows from the above proposition that Pj Q = Op(rj ), where

rj ∈ S r+k (Ω, Matl×l00 (C))


is given by
rj (x, ζ) = eDy ,∂ζ pj (x, ζ)qj (y, η)|y=x,η=ζ
and
X 1
rj ∼ ∂ζα pj Dxα qj .
α∈Nm
α!
≥0

It follows that
pr1 (supp(rj )) ⊂ supp(ψj ),
so that
X
r := rj ∈ S r+k (Ω, Matl×l00 (C))
j

is well-defined. Now one has


X X
Op(r) = Pj Qj = Pj Q = P Q.
j j

Moreover, qj = q in an open nbh of supp(ψj ), so that


∂ζα pj Dxα qj = ∂ζα pj Dxα q,
Global Analysis 2 77

and so
X X 1
r∼ ∂ζα pj Dxα qj
α∈Nm j
α!
≥0
X X 1
= ∂ζα pj Dxα q
α∈Nm j
α!
≥0
X 1
= ∂ζα pDxα q,
α∈Nm
α!
≥0

completing the proof. 


Definition 4.35. Assume k ∈ R ∪ {−∞}.
a) A symbol
p ∈ S k (Ω, Matl×l0 (Ω))
is called elliptic, if
(50)
pq − 1 ∈ S −1 (Ω, Matl0 ×l0 (C)), qp − 1 ∈ S −1 (Ω, Matl×l (C)) for some q ∈ S −k (Ω, Matl0 ×l (C)) .
b) An equivalence class
[p]k ∈ S k (Ω, Matl×l0 (C))/S k−1 (Ω, Matl×l0 (C))
is called elliptic, if p is elliptic.
0
c) P ∈ Ψk (Ω; Cl , Cl ) is called elliptic, if σ k (P ) is elliptic.
As the pointwise multiplication
S a (Ω, . . . ) × S b (Ω, . . . ) −→ S a+b (Ω, . . . ), (v, w) 7−→ vw
is well-defined, it follows that part b) of the above definition is well-defined. Indeed, if
p, p0 ∈ S k (Ω, Matl×l0 (C))
are given with
p0 − p ∈ S k−1 (Ω, Matl×l0 (C))
and (50), then one has
p0 q − 1 = (p0 − p)q + pq − 1 ∈ S −1 (Ω, Matl0 ×l0 (C)),
and likewise
qp0 − 1 = q(p0 − p) + qp − 1 ∈ S −1 (Ω, Matl0 ×l0 (C)),
proving the claim.
As the multiplication factors
   
S a (Ω, . . . )/S a−1 (Ω, . . . ) × S b (Ω, . . . )/S b−1 (Ω, . . . ) → S a+b (Ω, . . . )/S a+b−1 (Ω, . . . ),
([v]a , [w]b ) 7→ [vw]a+b ,
78 B. GÜNEYSU

we can state that


p ∈ S k (Ω, Matl×l0 (C))
is elliptic, if and only if there exists
q ∈ S −k (Ω, Matl0 ×l (C))
with
[p]k [q]−k = [1]0 , [q]−k [p]k = [1]0 .
The above notion of ellipticity is equivalent to the standard notion of (local) ellipticity of
partial differential operators:
Lemma 4.36. Assume k ∈ N and that
0
X
P = Pα Dα : D(Ω, Cl ) −→ E (Ω, Cl )
α
is a differential operator of order ≤ k. Then P is elliptic, if and only if
X
(51) Pα (x)ζ α ∈ GLl×l0 (C) for all x ∈ Ω, ζ ∈ Rm \ {0},
|α|=k

in particular, one necessarily has l = l0 .


Proof : Note first that  
X
σ k (P ) =  Pα ζ α  ,
|α|=k
k
where with the usual abuse of notation understand Pα ζ α to be the function (x, ζ) 7→
Pα (x)ζ α .
Assume now (51) and pick a smooth function χ ∈ E (Ω × Rm ) such that χ = 1 in a nbh of
Ω × {0} and such that χ is compactly supported in its ζ-slot. Define
 −1
X
q : Ω × Rm −→ Matl×l (C), q(x, ζ) := (1 − χ(x, ζ))  Pα (x)ζ α  .
|α|=k
−k
Then one has q ∈ S (Ω, Matl×l (C)) and
X X
q(x, ζ) Pα (x)ζ α − 1 = −χ(x, ζ), Pα (x)ζ α − 1 = −χ(x, ζ).
|α|=k |α|=k

The right hand sides define smoothing symbols, showing that P is elliptic.
Assume now P is elliptic. Pick
q ∈ S −k (Ω, Matl×l (C)), r ∈ S −1 (Ω, Matl×l (C))
with qp = 1 + r. Fix x ∈ Ω. Then one has
q(x, ζ)p(x, ζ) − 1 = Ox (1/(1 + |ζ|))
as |ζ| → ∞. Likewise, we see that
p(x, ζ)q(x, ζ) − 1 = Ox (1/(1 + |ζ|)).
Global Analysis 2 79

This shows that there exists R > 0 with p(x, ζ) invertible for all ζ ∈ Rm with |ζ| > R.
Since p(x, ζ) is a polynomial in its ζ-slot, ultimately p(x, ζ) is invertible for all ζ ∈ Rm \{0},
by a simple scaling argument.

0
Proposition 4.37. Assume k ∈ {−∞} ∪ R and that P ∈ Ψk (Ω; Cl , Cl ) is properly sup-
0
ported and elliptic. Then there exists a properly supported Q ∈ Ψ−k (Ω; Cl , Cl ) such that
0 0
QP − 1 ∈ Ψ−1 (Ω; Cl , Cl ), P Q − 1 ∈ Ψ−1 (Ω; Cl , Cl ).
Proof : Pick
q ∈ S −k (Ω, Matl×l (C))
such that
σ k (P )[q]−k = [1]0 , [q]−k σ k (P ) = [1]0 .
0
Pick Q ∈ Ψ−k (Ω; Cl , Cl ) with σ −k (Q) = [q]−k . We can even make sure that Q is properly
supported by Theorem 4.16. It follows that
σ 0 (QP ) = σ 0 (Q)σ 0 (P ) = [1]0 , σ 0 (P Q) = [1]0 .
In view of σ 0 (1) = [1]0 , it follows from applying (σ 0 )−1 to the last formulae that
0 0
QP − 1 ∈ Ψ−1 (Ω; Cl , Cl ), P Q − 1 ∈ Ψ−1 (Ω; Cl , Cl ).

Definition 4.38. Let (kj ) be a sequence of real numbers with kj & −∞ as j → ∞.
0
Assume that for every j we are given Qj ∈ Ψj (Ω; Cl , Cl ) and that we are given Q ∈
0
Ψk (Ω; Cl , Cl ) where k := maxj kj ∈ R. Then one writes
X
Q∼ Qj ,
j

if for all k 0 ∈ R there exists N ∈ N such that for all n ≥ N one has
n
0 0
X
Q− Qj ∈ Ψk (Ω; Cl , Cl ).
j=0

We recall:
0
Lemma 4.39. Assume s, k ∈ R and that P ∈ Ψk (Ω; Cl , Cl ). Then
s−k
P : Hcs (Ω, Cl ) −→ Hloc (Ω, Cl ).
In particular,
0 0
\
Ψ−k (Ω; Cl , Cl ) ⊂ Ψ−∞ (Ω; Cl , Cl ).
k∈Z

The following the the asymptotic completeness lemma at the level of operators:
80 B. GÜNEYSU

Theorem 4.40. Let (kj ) be a sequence of real numbers with kj & −∞ as j → ∞


and set k := maxj kj . Assume that for every j we are given a properly supported Qj ∈
0 0
Ψj (Ω; Cl , Cl ). Then there exists Q ∈ Ψk (Ω; Cl , Cl ), such that
X
Q∼ Qj .
j
0
The operator Q is uniquely determined modulo Ψ−∞ (Ω; Cl , Cl ).
Proof : Uniqueness: If Q0 is another operator with the above properties, then by Lemma
4.39,
0 0
\
Q − Q0 ∈ Ψk (Ω; Cl , Cl ) ⊂ Ψ−∞ (Ω; Cl , Cl ).
k∈R
Existence: This follows from the asymptotic summation lemma for symbols: we can pick
qj ∈ S kj (Ω, Matl×l (C))
with Op(qj ) = Qj by Lemma 4.33 and
q ∈ S k (Ω, Matl×l (C))
P
with q ∼ j qj . Then Q := Op(q) does the job. 
Finally we can prove:
0
Theorem 4.41. Assume k ∈ {−∞} ∪ R and that P ∈ Ψk (Ω; Cl , Cl ) is properly supported
0
and elliptic. Then there exists Q ∈ Ψ−k (Ω; Cl , Cl ) such that
0 0
QP − 1 ∈ Ψ−∞ (Ω; Cl , Cl ), P Q − 1 ∈ Ψ−∞ (Ω; Cl , Cl ).
0
Moreover, Q is uniquely determined modulo Ψ−∞ (Ω; Cl , Cl ).
Remark 4.42. An operator Q as above is called a parametrix for P . As the proof below
0
shows, every properly supported Q ∈ Ψ−k (Ω; Cl , Cl ) with
QP − 1 ∈ Ψ−∞ (Ω; Cl , Cl )
automatically satisfies
0 0
P Q − 1 ∈ Ψ−∞ (Ω; Cl , Cl ).
Proof of Theorem 4.41: Existence: We are going to bootstrap Proposition 4.37 using the
0
last Theorem: by Proposition 4.37 we can pick a properly supported Q0 ∈ Ψ−k (Ω; Cl , Cl )
such that
0 0
Q0 P − 1 ∈ Ψ−1 (Ω; Cl , Cl ), P Q0 − 1 ∈ Ψ−1 (Ω; Cl , Cl ).
Then
R := 1 − Q0 P ∈ Ψ−1 (Ω; Cl , Cl )
and Theorem 4.34 implies
Rj ∈ Ψ−j (Ω; Cl , Cl ) for all j ∈ N,
Global Analysis 2 81

and these are properly supported. Then Theorem 4.40 implies the existence of A ∈
Ψ0 (Ω; Cl , Cl ) such that

X
A∼ Rj .
j=0

In particular one has


\
A(1 − R) − 1 ∈ Ψ−n (Ω; Cl , Cl ) ⊂ Ψ−∞ (Ω; Cl , Cl ).
n∈N

Defining the properly supported


0
Q := AQ0 ∈ Ψ−k (Ω; Cl , Cl ),
one now finds
QP − 1 = AQ0 P − 1 = A(1 − R) − 1 ∈ Ψ−∞ (Ω; Cl , Cl ).
On the other hand, -
B := QP − 1 ∈ Ψ−∞ (Ω; Cl , Cl )
is properly supported.
0
Analogously, one finds a properly supported P 0 ∈ Ψ−k (Ω; Cl , Cl ) such that
P 0 Q − 1 ∈ Ψ−∞ (Ω; Cl , Cl ).
Then
C := P 0 Q − 1 ∈ Ψ−∞ (Ω; Cl , Cl )
is properly supported. We have
P 0 QP 0 = P (1 + B) = P + P B, P 0 QP 0 = (1 + C)P 0 = P 0 + CP 0 ,
so that (by Theorem 4.9),
0
S := P − P 0 = P B + CP 0 ∈ Ψ−∞ (Ω; Cl , Cl )
and
0 0
D := P Q − 1 = SQ + P 0 Q − 1 ∈ Ψ−∞ (Ω; Cl , Cl ).
This proves the existence of a parametrix.
Uniqueness: assume
0
Q̃ ∈ Ψ−k (Ω; Cl , Cl )
is another parametrix for P , so that in particular
0 0
E := Q̃P − 1 ∈ Ψ−∞ (Ω; Cl , Cl )
Then we have
0
Q̃ − Q = Q̃P Q − Q̃D − Q = EQ − Q̃D ∈ Ψ−∞ (Ω; Cl , Cl ),
completing the proof. 
82 B. GÜNEYSU
0
Corollary 4.43 (Local Elliptic regularity). Assume k ∈ R and that P ∈ Ψk (Ω; Cl , Cl ) is
properly supported and elliptic. Then for all s ∈ R one has the implication ( strong local
elliptic regularity)
0
s
P f ∈ Hloc s+k
(Ω, Cl ) ⇒ f ∈ Hloc (Ω, Cl ) for all f ∈ D 0 (Ω, Cl ).
In particular, one has the implication ( weak local elliptic regularity)
0
P f ∈ E (Ω, Cl ) ⇒ f ∈ E (Ω, Cl ) for all f ∈ D 0 (Ω, Cl ).
0
Proof : Pick a parametrix Q ∈ Ψ−k (Ω; Cl , Cl ) for P . Then QP − 1 is smoothing, so that
\
QP f − f =: h ∈ E (Ω, Cl ) = r
Hloc (Ω, Cl ).
r∈R
s+k
But the mapping properties of pseudodifferential operators show QP f ∈ Hloc (Ω, Cl ), so
s+k
f = QP f − h ∈ Hloc (Ω, Cl ). 
For P = ∆ and f ∈ L1loc (Ω, Cl ) this result goes back to Hermann Weyl (1940) and is
typically called Weyl’s lemma.

5. Function spaces and distributions on manifolds


Assume in the sequel that M is a smooth m-dimensional (without boundary), so for exam-
ple an open subset Ω ⊂ Rm . We remark that M admits smooth partitions of unity, that
is, given an open cover (Ui )i∈I we can pick (φi )i∈I ⊂ C ∞P
(M ) such that supp(φi ) ⊂ Ui and
0 ≤ φi ≤ 1 for all i and (supp(φi ))i is locally finite and i φi = 1.
We also fix a smooth complex vector bundle E → M with rank l.
We denote for all open U ⊂ M the space of smooth sections in E → M which are
defined on U by ΓC ∞ (U, E), and its subspace of smooth compactly supported sections by
ΓCc∞ (U, M ). Then a local frame for E → M over U is by definition a collection of local
sections b1 , . . . , bl ∈ ΓC ∞ (U, M ) such that b1 (x), . . . , bl (x) ∈ Ex is a basis for all x ∈ U .

By a local control datum for E → M , we understand a pair (U, x), {b1 , . . . , bl } given by
a chart
x = (x1 , . . . , xm ) : U −→ V ⊂ Rm
and a local frame b1 , . . . , bl ∈ ΓC ∞ (U, M ) for E → M .
We want turn ΓC ∞ (M, E) into a F space. To this end, assume we are given a family A of
local control data for E → M which also form an atlas  for M (in short, an atlas of local
control data for E → M ). For all (U, x), {b1 , . . . , bl } ∈ A , all K ⊂ U compact, and all
n ∈ N≥0 we define a seminorm pU,x,K,{b1 ,...,bl } on ΓC ∞ (M, E) by setting
|α| j
∂ ψ
pU,x,{b1 ,...bl },K,n (ψ) := max ,
K,|α|≤n,j=1,...,l ∂xα

where we have written ψ = lj=1 ψ j bj for (uniquely determined) ψ 1 , . . . , ψ l ∈ C ∞ (U ). In


P

this way we obtain a locally convex topology on ΓC ∞ (M, E), and it is rather easy to show
Global Analysis 2 83

that this topology does not depend on the choice of an atlas A of local control data for
E → M . It is easy to check that this topology turns ΓC ∞ (M, E) into an F space, which is
denoted by ΓE (M, E).
Let us now come to the LC topology on ΓCc∞ (M, E). Denote by ΓDK (M, E) the space
of smooth sections with support in K ⊂ M compact, and equip ΓCc∞ (M, E) with all
seminorms p on ΓCc∞ (M, E) such that the restriction of p to ΓDK (M, E) is continuous for
all compact K.
It is easily checked that this produces an LF space (by picking an atlas of local control data
and the induced seminorms, and on each local control datum with domain U a compact
exhaustion of U ). We denote this LF space which we denote by ΓD (M, E).

Definition 5.1. We call the LC space ΓD 0 (M, E) := ΓD (M, E)0 the space of distributions
on E → M , and the LC space ΓE 0 (M, E) := ΓE (M, E)0 the space of compactly supported
distributions on E → M .
Lemma 5.2. The embedding ΓD (M, E) ,→ ΓE (M, E) is continuous and has a dense image.
In particular, there is a canonical continuous embedding ΓE 0 (M, E) ,→ ΓD 0 (M, E).
Proof : The continuity of ΓD (M, E) ,→ ΓE (M, E) is checked easily using the pU,x,{b1 ,...,bl },K,r
’s.To see the density, pick a compact exhaustion (Kn ) of M and for each n some χn ∈ D(M )
with χn = 1 in Kn and supp(χn ) ⊂ Kn+1 . Then given ψ ∈ ΓE (M, E) one has χn ψ → ψ
with respect to any pU,x,{b1 ,...,bl },K,r , r ∈ N≥0 , which follows from the Leibnitz rule. 
Let us produce some subspaces of these spaces: Assume p ∈ [1, ∞]. We call two Borel
measurable sections f1 , f2 of E → M equivalent, if for some (and then automatically any!)
Borel measure µ on M which in each chart has a smooth strictly positive density with
respect to the Lebesgue measure (in short: a smooth meausre), one has f1 = f2 µ-a.e.
Then we get the F space ΓLploc (M, E) of equivalence classes f of Borel sections in E → M
such that for all local control data ((U, φ), {b1 , . . . bl }) one has
l
X
1 l
(f , . . . , f ) ∈ Lploc (φ(U ), Cl ), if f = f j bj in U .
j=1

In addition we get the LF space ΓLpc (M, E) of all f ∈ ΓLploc (M, E) such that f has a
compactly supported representative. In both cases, the LC structure is defined by let-
ting ((U, φ), {b1 , . . . , bl }) run through an atlas, and using the transformation formula for
Lebesgue integrals one can check that theses LCS’s do not depend on the atlas.
For f1 , f2 ∈ ΓL1loc (M, E) we have f1 = f2 , if and only if for some/all pairs (µ, h) such that
µ is a smooth measure on M and h a metric on E → M , and all ψ ∈ ΓD (M, E) one has
Z Z
(f1 , ψ)h dµ = (f2 , ψ)h dµ,

and there are canonical continuous embeddings


ΓD (M, E) ,→ ΓLpc (M, E) ,→ ΓLploc (M, E), ΓE (M, E) ,→ ΓLploc (M, E).
84 B. GÜNEYSU

There is no way to produce a canonical embedding of ΓL1loc (M, E) into ΓD 0 (M, E). On the
other hand, any (µ, h) produces the continuous embedding
 Z 
µ,h µ,h
T : ΓL1loc (M, E) ,→ ΓD 0 (M, E), f 7−→ ψ 7−→ Tf (ψ) := (f, ψ)h dµ ,
M
and likewise
T µ,h : ΓL1c (M, E) ,→ ΓE 0 (M, E).
Definition 5.3. 1. Given T ∈ ΓD 0 (M, E), U ⊂ Ω open, we set
T |U := T |ΓD (U,E) ∈ ΓD 0 (U, E)
and say that T is smooth in U , if there exists/for all (µ, h) as above, there exists ψ ∈
ΓE (U, E) such that T |U = Tψµ,h .
2. The support of T is defined by
[
supp(T ) := M \ U.
U ⊂M, T |U =0

The assignment U 7→ ΓD 0 /E 0 (U, E), together with the restriction morphisms


ΓD 0 /E 0 (V, E) −→ ΓD 0 /E 0 (U, E), T 7−→ T |U ,
where U ⊂ V ⊂ Ω are open, is a linear sheaf over M .

Remark 5.4. 1. For all T ∈ ΓD 0 (M, E) the set supp(T ) is closed.


2. For all T ∈ D 0 (M, E) one has T ∈ ΓE 0 (M, E), if and only if supp(T ) is a compact set.
3. For all f ∈ ΓL1loc (M, E) and all x ∈ M one has x ∈ supp(T ), if and only if there
exists/for all (µ, h) one has
Z
|f |h dµ > 0 for all open neighbourhoods U of x.
U
In particular, for all continuous sections f ∈ ΓC (M, E) one has
supp(f ) = {f 6= 0} = supp(Tf ),
that is, in the continuous case the distributional support is equal to the support in the
sense of functions on topological spaces.
We can define local Sobolev spaces as follows:
Definition 5.5. Let s ∈ R. Then T ∈ ΓD 0 (M, E) is by definition in ΓHloc s (M, E), if for all

local control data ((U, φ), {b1 , . . . , bl }) the linear form defined by the diagram below is in
s
Hloc (φ(U ), Cl ),
T |U
(52) ΓD (U, E) /
O 9 C
Pl
(ψ 1 ,...,ψ l )7→ j=1 (ψ
j ◦φ)b
j

D(φ(U ), Cl )
Global Analysis 2 85

s (M, E) ∩ ΓE 0 (M, E).


One sets ΓHcs (M, E) := ΓHloc

The space ΓHcs (M, E) canonically becomes an LF space, and ΓHloc s (M, E) becomes an F

s l
space, by taking the induced topology from Hloc/c (φ(U ), C ) and letting ((U, φ), {b1 , . . . bl })
run through an atlas. One can check that these LCS’s do not depend on the atlas. Let
us sketch a proof of the fact that at least the underlying linear spaces do not depend
on the atlas. To this end, it suffices to show that for every open Ω, Ω0 ⊂ Rm every
smooth diffeomorphism φ : Ω → Ω0 and every smooth map A : Ω → Gll×l (C) the induced
isomorphism of linear spaces

∗ : D (Ω, C ) −→ D (Ω , C )
0 0 0
φA l l

restricts to an isomorphism of linear spaces


0
φA s l s l
∗ : Hloc (Ω, C ) −→ Hloc (Ω , C ),

as then we also get an isomorphisn of linear spaces


0
φA s l s l
∗ : Hc (Ω, C ) −→ Hc (Ω , C ).

Note first that given any elliptic properly supported P ∈ Ψs (Ω, Cl , Cl ) allows an equivalent
characterization of Hlocs
(Ω, Cl ): namely, for all f ∈ D 0 (Ω, Cl ) one has f ∈ Hloc s
(Ω, Cl ), if and
only if P f ∈ Lloc (Ω, C ). Indeed, if f ∈ Hloc (Ω, C ), then P f ∈ Hloc (Ω, C ) = L2loc (Ω, Cl ).
2 l s l 0 l
0
Conversely, if P f ∈ Hloc (Ω, Cl ) = L2loc (Ω, Cl ), then we pick a parametrix Q for P . The
we have QP f ∈ Hloc (Ω, Cl ) and h := QP f − f is smooth and so in Hloc
s s
(Ω, Cl ), thus so
is f = QP F − h. Given this characterization, we can proceed as follows: pick some P as
0
above. Then φA s l l
∗ (P ) ∈ Ψ (Ω , C , C ) (this ’invariance of pseudodifferential operators under
coordinate transformations’ [1] is hard to prove; we have seen this in the first exercise
for differential operators), and then it is easily seen that φA ∗ (P ) is elliptic and properly
supported, too. Since by the transformation formula for integrals we get the isomorphism
of linear spaces
0
φA 2 l 2 l
∗ : Lloc (Ω, C ) −→ Lloc (Ω , C ),

∗ (P )φ∗ (f ) = φ∗ (P f ) for all f ∈ D (Ω, C ), we finally have


0
since φA A A l

s 0
f ∈ Hloc (Ω, Cl ) ⇔ P f ∈: L2loc (Ω, Cl ) ⇔ φA 2 l
∗ (P f ) ∈ Lloc (Ω , C )
0 0
⇔ φA A 2 l A s l
∗ (P )φ∗ (f ) ∈ Lloc (Ω , C ) ⇔ φ∗ (f ) ∈ Hloc (Ω , C ),

s
where we have used the characteization of Hloc (Ω0 , Cl ) in terms of φA
∗ (P ). This completes
the proof.
One has
ΓHc/loc
0 (M, E) = ΓL2c/loc (M, E),
86 B. GÜNEYSU

and for all t ≥ s there are canonical coninuous embeddings


(53) ΓHc/loc
s (M, E) ,→ ΓHc/loc
s (M, E),
(54) ΓHloc
s (M, E) ,→ ΓD 0 (M, E),

(55) ΓHcs (M, E) ,→ ΓE 0 (M, E),


(56) ΓHcs (M, E) ,→ ΓHloc
s (M, E).

In addition by the local Sobolev embedding theorem we canonically have an isomorphism


of LCS’s
s (M, E) ∼
\
ΓHloc = ΓE (M, E),
s∈R
and every (µ, h) induces an isomorphism of LCS’s
ΓH s (M, E)0 ∼
=µ,h Γ −s (M, E), loc Hc

so that
ΓHcs (M, E) ∼
[
=µ,h ΓE 0 (M, E),
s∈R

6. Pseudodifferential operators on manifolds


Fix now a second smooth complex vector bundle F → M of rank l0 .
Definition 6.1. Let k ∈ R ∪ {−∞}. A linear map
P : ΓD (M, E) −→ ΓE (M, F )
is called a pseudodifferential operator of order k (properly supported, elliptic), if for all local
control data (U, φ, {b1 , . . . , bl }) for E → M and all local frames c1 , . . . , cl0 ∈ ΓE (U, F ) the
induced map

P |U
(57) ΓD (U, E) / ΓE (U, F )
O

Pl Pl
(ψ 1 ,...,ψ l )7→ j=1 (ψ
j ◦φ)b
j j=1 ψ j bj 7→(ψ 1 ◦φ−1 ,...,ψ l ◦φ−1 )


D(φ(U ), Cl ) / E (φ(U ), Cl )
0

0
is in Ψk (φ(U ); Cl , Cl ) (properly supported, elliptic).
The linear space of such operators is denoted with Ψk (M ; E, F ). Note that every pseudo-
differential operator is automatically continuous from ΓD to ΓE .
The above definition is consistent with the ones we gave for M = Ω and E = Ω × Cl × Ω
(that being a pseudo is invariant under coordinate transformations is actually not so easy
to check!).
Global Analysis 2 87

We have the following result which is central for geometry, and which follows straightfor-
wardly from a partition of unity argument. Note for its formulation that every properly
supported P ∈ Ψk (M ; E, F ) maps
P : ΓD (M, E) −→ ΓD (M, F )
continuously and thus by density induces a continous linear map
P : ΓE (M, E) −→ ΓE (M, F ),
and that the composition of properly supported pseudifferential operators is again a prop-
erly supported differential operator. This follows immediately from the Euclidean results.
Theorem 6.2. Let k ∈ R ∪ {−∞} and assume P ∈ Ψk (M ; E, F ) is elliptic and properly
supported. Then there exists Q ∈ Ψ−k (M ; F, E) properly supported, such that
QP − 1 ∈ Ψ−∞ (M ; E, E), P Q − 1 ∈ Ψ−∞ (Ω; F, E).
Any such Q is uniquely determined modulo Ψ−∞ (M ; F, E), and called a parametrix for P .
From now on, let (M, g) be a Riemannian manifold and let (E, hE ) → M and (F, hF ) → M
be metric vector bundles. The Riemannian volume measure is denoted with with µg . It is a
smooth measure. For example, M could be a Riemannian submanifold of Euclidean space,
or a Riemannian spin manifold, or the Hyperbolic space. The bundle E could be ∧k T ∗ M ,
so that its sections are k-forms Ωk (M ) on M , or the spinor bundle, if M is spin and so on.
We get the global Lp -spaces of sections ΓLp (M, g; E, hE ) given by all f ∈ ΓLploc (M, E) such
that Z
|f |phE dµg < ∞, if p < ∞,
M
and analogously
inf{C ≥ 0 : |f |hE ≤ C µg -a.e.} < ∞, if p = ∞.
In particular ΓL2 (M, g; E, hE ) is a Hilbert space according to
Z
hf1 , f2 ig,hE := hE (f1 , f2 )dµg .
M

Given P ∈ Ψk (M ; E, F ) there is a unique P g,hE ,hF ∈ Ψk (M ; F, E) such that


Z Z
hE ψ1 , P g,hE ,hF ψ2 dµg

hF (P ψ1 , ψ2 ) dµg =
M M
for all ψ1 ∈ ΓD (M, E), ψ2 ∈ ΓD (M, F ). The above embeddings are denoted by
Tg,hE := Tµg ,hE : ΓL1loc (M, E) ,→ ΓD (M, E).
Next, P induces a continuous linear map
Pg,hE ,hF : ΓE 0 (M, E) −→ ΓD 0 (M, F )
given by
Pg,hE ,hF f1 (f2 ) := f1 (P g,hE ,hF f2 ) for all f1 ∈ ΓE 0 (M, F ), f2 ∈ ΓD (M, F ).
88 B. GÜNEYSU

If k > −∞ then by restriction one gets a continuous linear map


(58) Pg,hE ,hF : ΓHcs (M, E) −→ ΓH s−k (M, F ).
loc

Likewise, if P is properly supported, then we canonically get continuous linear maps


Pg,hE ,hF : ΓD 0 /E 0 (M, E) −→ ΓD 0 /E 0 (M, F )
which by restriction induce continuous linear maps
(59) Pg,hE ,hF : ΓHloc/c
s (M, E) −→ ΓH s−k (M, F ).
loc/c

It suffices to prove all these facts locally (cf. Theorem 4.30).


Corollary 6.3 (Local elliptic regularity). Assume k, s ∈ R, f ∈ ΓD 0 (M, E), and that
P ∈ Ψk (M ; E, F ) is properly supported and elliptic. Then one has
 
Pg,hE ,hF f ∈ ΓHloc
s (M, F ) for some triple (g, hE , hF ),
if and only if one has
 
Pg,hE ,hF f ∈ ΓHloc
s (M, F ) for all such triples,

and in this case one has f ∈ ΓH s+k (M, E). In particular, one has
loc
 
Pg,hE ,hF f ∈ ΓE (M, F ) for some triple (g, hE , hF ),
if and only if one has
 
Pg,hE ,hF f ∈ ΓE (M, F ) for all such triples,

and in this case one has f ∈ ΓE (M, E).


Proof : The asserted independence of the assumption
 
Pg,hE ,hF f ∈ ΓHloc (M, F )
s

0 0 0
on the metrics follows from calculating P g,hE ,hF in terms of P g ,hE ,hF explicitly, which is
not so complicated (calculate the smooth density of one measure with respect to the other,
and calculate each fiber metric in terms of the other using a smooth endomorphism on
the bundle). Now one can proceed with the proof of elliptic regularoty precisely as in the
Euclidean case. 

7. Basic facts on linear operators in Hilbert spaces


We collect some standard facts for unbounded operators on Hilbert spaces. A classical
reference is [6].
Let H1 , H2 be infinite dimensional seperable complex Hilbert spaces. The underlying scalar
products will simply be denoted with hf, gi, and the induced norms with kf k := hf, f i1/2 .
Global Analysis 2 89

A continuous linear map A : H1 → H2 is called unitary, if A is invertible with A−1 = A∗ .


Unitary maps preserve scalar products and thus norms.
We also recall that for a closed subspace M of H1 one has
H1 = M ⊕ M ⊥ ,
and that for every continuous linear functional T : H1 → K there exists a unique fT ∈ H1
such that T (ψ) = hfT , ψi for all ψ ∈ H1 .
The assignment T 7→ fT induces an antilinear isomorphism of LCS’s between H10 and H1
called the Riesz-Fischer duality theorem.
Let S be linear operator from H1 to H2 which is defined on some dense linear subspace
Dom(S) ⊂ H1 , that is, S is a linear map S : Dom(S) → H2 .
Then S is called bounded, if its operator norm is finite,
kSk := sup{kSψk : ψ ∈ Dom(S), kψk ≤ 1} < ∞,
that is, if S : Dom(S) → H2 is continuous when Dom(S) ⊂ H1 is equipped with subspace
topology. Otherwise S is called unbounded. Given another densely defined operator T from
H1 to H2 one calls T an extension of S and writes S ⊂ T , if Dom(S) ⊂ Dom(T ) and
T = S on Dom(S). Any bounded operator can always be uniquely extended to a bounded
operator which is defined everywhere (and this extension has the same operator norm)
The operator S is called closed, if its graph {(f, g) : f ∈ Dom(S), g = Sf } ⊂ H1 × H2 is a
closed subset (bounded everyhwere defined linear operators are automatically closed), and
S is called closable, if it has a closed extension.
Lemma 7.1. If S is closable, then S has a smallest closed extension S which is given by
Dom(S) = {f ∈ H1 : there ex. sequ. (fn ) ⊂ Dom(S) s.t. fn → f and s.t. (Sfn ) conv.},
and then Sf = limn Sfn .
The adjoint S ∗ of S is the linear operator from H2 to H1 defined by
Dom(S ∗ ) := {f ∈ H2 : there ex. f ∗ ∈ H1 w. hf ∗ , hi = hf, Shi for all h ∈ Dom(S)}
and then S ∗ f := f ∗ .
Adjoints need not be densely defined (they can even be trivial!), but:
Lemma 7.2. The operator S ∗ is always closed. One has
Ran(S)⊥ = Ker(S ∗ ).
If H2 = H1 = H , then the resolvent set ρ(S) is defined by all λ ∈ C s.t. S − λ is an
invertible linear map Dom(S) → H , such that the resolvent (S − λ)−1 , which is a linear
map H → Dom(S), is bounded as a linear map H → H .
Lemma 7.3. If H2 = H1 = H and if S is closed, then ρ(S) is an open set and if S − λ
is invertible for some λ ∈ C, then λ is automatically in ρ(S), that is, (S − λ)−1 is bounded
as a map from H → H . If we equip Dom(S) with the graph norm k·k + kS(·)k, then
(S − λ)−1 is also bounded as linear map Dom(S) → H .
90 B. GÜNEYSU

If H2 = H1 = H , then a point λ ∈ C is called an eigenvalue of S, if Ker(S −λ) 6= {0}, and


then the dimension of this space is called the multiplicity of λ. Vectors in Ker(S − λ) \ {0}
are called eigenvectors corresponding to the eigenvalue λ; the spectrum of S is defined
by σ(S) := C \ ρ(S), a set which contains all eigenvalues of S and which by the above
considerations is closed, if S is so.
S is called symmetric if S ⊂ S ∗ , and self-adjoint, if S = S ∗ (both imply of course H1 = H2 ).
If S is everywhere defined, then S is symmetric if and only if S is self-adjoint. Note however
the following result (which somewhat exludes this situation in typical apllications):
Theorem 7.4 (Hellinger/Toeplitz 1910). An everywhere defined symmetric operator is
automatically bounded.
Unlike self-adjointness, symmetry is easy to check: it just means
hSψ1 , ψ2 i = hψ1 , Sψ2 i for all ψ1 , ψ2 ∈ Dom(S),
which in particular implies
hSψ, ψi ∈ R for all ψ ∈ H1 .
If S is symmetric, one writes S ≥ a for some a ∈ R and calls S semibounded from below
by a, if
hSψ, ψi ≥ a kψk2 for all ψ ∈ H1 .
If S is self-adjoint, then this is equivalent to σ(S) ⊂ [a, ∞).
Lemma 7.5. a) Symmetric operators have real eigenvalues, and self-adjoint operators have
a real spectrum.
b) A symmetric operator is self-adjoint, if and only if its spectrum is real (!).
c) The closure of a symmetric operator S is self-adjoint, if Ker((S ± i)∗ ) = {0}.
d) If the closure of a symmetric operator is self-adjoint, then its closure is its unique self-
adjoint extension and the operator is called essentially self-adjoint.
c’) If S is symmetric with S ≥ a, then S is essentially self-adjoint if Ker((S + b)∗ ) = {0}
for some b > a.
Example 7.6. Let Ω be an arbitrary open
Pm subset of Rm and consider the operator with

Dom(∆ ) = Cc (Ω) given by ∆ = − j=1 ∂j in L2 (Ω). Integrating by parts (Green’s
Ω Ω 2

formula) one finds that Z




∆ ψ1 , ψ2 = (∇ψ1 , ∇ψ2 )dx,

so that ∆ is symmetric (thus closable) and ≥ 0; ∆Ω is unbounded and not closed. One

has
Dom((∆Ω )∗ ) = {ψ ∈ L2 (Ω) : ∆Ω ψ ∈ L2 (Ω)}, (∆Ω )∗ ψ = ∆Ω ψ,
in particular, ∆Ω is not self-adjoint. However, ∆Ω has self-adjoint extensions. For example
the Dirichlet realization,
Dom(S Ω ) := {ψ ∈ H01 (Ω) : ∆Ω ψ ∈ L2 (Ω)}, S Ω ψ := ∆Ω ψ
Global Analysis 2 91

is a self-adjoint extension. In general, ∆Ω has more than one self-adjoint extension and so
is not essentially self-adjoint. For example, if Ω has finite measure and Rm \ Ω 6= ∅, then
the Dirichlet realization differs from the so called Neumann realization. If Ω = Rm , then
m
∆R is essentially self-adjoint: the closure of this operator is equal to the closure of the
same operator with domain of definition S (Rm ), and the latter symmetric operator is via
Fourier transform unitarily equivalent to the maximally defined multiplication operator Q
induced by | · |2 , which is given as follows:
Dom(Q) := {ψ ∈ L2 (Rm ) : | · |2 ψ ∈ L2 (Rm )}, Qψ(x) := |x|2 ψ(x).
Since a maximally defined multiplication operator induced by a function h on an L2 -space
is self-adjoint, if and only if h is almost everywhere real-valued, it follows that Q is self-
adjoint.
If S is self-adjoint, then S has a functional calculus (its spectral calculus), that is, one can
define f (S) in a ’consistent’ way for every Borel function f : R → C, and then f (S) is
self-adjoint if f is real-valued, and bounded (resp. ≥ a for some a ∈ R) if f is bounded
(resp. ≥ a) on the spectrum of S.
Example 7.7. Assume S is self-adjoint.
1. Given ψ ∈ H , ψ(t) := e−itS ψ is the unique solution of the abstract Schrödinger operator
ψ̇(t) = −iSψ(t), t ∈ R, ψ(0) = ψ ∈ Dom(S).
2. If S ≥ c for a constant c ∈ R, that is, hSf, f i (which is always a real number as S is
symmetric) is ≥ c kf k2 for all f ∈ Dom(S), then one can also define ψ(t) := e−tS ψ which
solves the abstract heat equation ψ̇(t) = −Sψ(t), t > 0, ψ(0) = ψ ∈ H (in the Schrödinger
case the initial vector has to be in the domain of S, in the heat case the exponential has a
smoothing effect, as S ≥ c!).
A bounded linear map K : H1 → H2 is called compact, if for all ONB (en )n∈N of H1 and
(fn )n∈N of H2 one has hKen , fn i → 0. This is equivalent to requiring that for every bounded
sequence ψn in H1 the sequence Kψn in H2 has a convergent subsequence. Compositions
of compact linear maps with bounded linear maps are again compact.
Example 7.8. Given a Riemann manifold (M, g) and complex metric vector bundles
(E, hE ) → M , (F, hF ) → M , we can canonically construct the complex metric vector
bundle
(E  F, hE  hF ) −→ M × M
whose fiber at (x, y) ∈ M × M is given by Hom(Ey , Fx ), and then every ’integral kernel’
k ∈ ΓL2 (M × M, g ⊕ g; E  F, hE  hF )
determines the continuous linear operator
Z
K : ΓL2 (M, g; E, hE ) −→ ΓL2 (M, g; F, hF ), Kf (x) = k(x, y)f (y)dµg (y).
M
92 B. GÜNEYSU

One can easily prove using Fubini that


X
| hKen , fn ig,hF |2 < ∞,
n

showing that K is compact (in fact, the above square summaability is called Hilbert-Schmidt
property).
Theorem 7.9. Assume H1 = H2 = H , that S is closed and that (S − λ)−1 is compact
for some/all (resolvent identity!) λ ∈ ρ(S). Then:
a) S has at most countably many eigenvalues all having a finite multiplicity, and the spec-
trum of S contains no other points (one says that S has a purely discrete spectrum).
b) If (λj )j∈N denotes the eigenvalues, then either ]N < ∞ or |λj | → ∞ as j → ∞.
c) If S is self-adjoint (or more generally: normal), then there exists an orthonormal basis
of H given by eigenvectors of S,
m
Example 7.10. 1. The unique self-adjoint extension of ∆R does not have a purely discrete
spectrum (thus the spectrum contains some ’continuous part’). In fact, from the above
unitary equivalence to a continuous multiplication operator one finds that this extension
has no eigenvalues at all.
2. If Ω is bounded (= relatively compact), then ’often’ (!) self-adjoint extensions ∆Ω have
a compact resolvent and thus a purely discrete spectrum. For example, this is the case for
the Dirichlet realization S Ω (cf. Example 7.6). A simple proof which is in the spirit of our
methods: the resolvent (S Ω + 1)−1 : L2 (Ω) → L2 (Ω) factors as
(S Ω + 1)−1 = AB,
where
B : L2 (Ω) −→ H01 (Ω)
denotes the resolvent with different target space (a bounded operator), and where the
embedding
A : H01 (Ω) −→ L2 (Ω)
is compact (which follows ultimately from the Rellich compactness theorem, using that Ω
is bounded). The fact that the resolvent maps into H01 (Ω) follows from the fact that the
resolvent always maps into the domain of definition.
The resolvent of a general self-adjoint extension need not map into H01 (Ω), and in fact
even if Ω is bounded there exist self-adjoint extensions that do not have a purely discrete
spectrum (boundary regularity plays an additional role in this context for certain self-
adjoint realizations such as the Neumann one).
Theorem 7.11. Assume S is closed and invertible modulo compact operators, that is,
there exist continuous linear maps A1 , A2 : H2 → H1 , and compact linear maps Kj : Hj →
Hj , such that
A1 S = 1 + K1 , SA2 = 1 + K2 .
Then S is Fredholm, that is, Ker(S) is finite dimensional, Ran(S) is closed and its codi-
mension in H2 is finite.
Global Analysis 2 93

Fredholm operators have a well-defined Fredholm index


ind(S) := dimKer(S) − codimRan(S).
Such quantities play a crucial role in the context of the Atiyah-Singer index theorem, which
relates geometry and topology (and surely is in the top ten of important mathematical
results of the 20th century).

8. Functional analytic properties of pseudodifferential operators on Riemann-


ian manifolds
Assume now (M, g) is a Riemannian manifold, and that (E, hE ) → M , (F, hF ) → M
are Hermitian vector bundles. If k ∈ R and P ∈ Ψk (M ; E, F ) is properly supported (for
example, if M is compact or if P is a differential operator, this assumption is automatically
satisfied), then P maps
P : ΓD (M, E) −→ ΓD (M, F )
and thus induces a densely defined linear operator
P̃g,hE ,hF from ΓL2 (M, g; E, hE ) to ΓL2 (M, g; F, hF )
with domain of definition Dom(P ) = ΓD (M, E).
Lemma 8.1. P̃g,hE ,hF is closable.
max
Proof : A closed extension is given by Pg,hE ,hF
with domain of definition
max
Dom(Pg,hE ,hF
) = {f ∈ ΓL2 (M, g; E, hE ) : Pg,hE ,hF f ∈ ΓL2 (M, g; F, hF )}.

In particular for the closure
P g,hE ,hF := P̃g,hE ,hF
we have
max
Dom(P g,hE ,hF ) ⊂ Dom(Pg,h E ,hF
).
If P is elliptic, then by local elliptic regularity one has
max
Dom(Pg,h E ,hF
) ⊂ ΓHloc
k (M, E).

Theorem 8.2. Assume M is compact, that k ∈ (0, ∞) and that P ∈ Ψk (M ; E, F ) is


elliptic.
a) One has
max
Dom(Pg,h E ,hF
) = ΓHloc/c
k (M, E),
max
and P g,hE ,hF = Pg,h E ,hF
.
b) P g,hE ,hF is invertible modulo compact operators and so is Fredholm.
c) Assume (E, hE ) = (F, hF ). Then for all14 λ ∈ ρ(P g,hE ,hF ) the resolvent (P g,hE ,hF − λ)−1
14In general ρ(P
g,hE ,hF ) could be empty; in case P g,hE ,hF is self-adjoint, however, we have σ(P g,hE ,hF ) ⊂
R and so C \ R ⊂ ρ(P g,hE ,hF ).
94 B. GÜNEYSU

is compact, in particular, P g,hE ,hF has a purely discrete spectrum. All eigenvectors of
P g,hE ,hF are smooth.
d) Assume (E, hE ) = (F, hF ) and P = P g,hE ,hF , that is, P is formally self-adjoint w.r.t.
(g, hE , hF ). Then P g,hE ,hF is self-adjoint (equivalently: P̃g,hE ,hF is essentially self-adjoint),
in particular, σ(P g,hE ,hF ) ⊂ R and ρ(P g,hE ,hF ) is not empty).
Proof : By the compactness of M one has
ΓE (M, E) = ΓD (M, E), ΓHloc
k (M, E) = ΓH k (M, E),
c

ΓL2 (M, g; E, hE ) = ΓL2c (M, E) = ΓL2loc (M, E)


as locally convex spaces, and similarly for (F, hF ). I will always indicate in which argument
compactness is used.
max
a) We already know Dom(Pg,h E ,hF
) ⊂ ΓHloc
k (M, E) as P is elliptic. If f ∈ ΓH k (M, E) ⊂
loc
ΓL2loc (M, E) (note k ≥ 0), then
Pg,hE ,hF f ∈ ΓL2loc (M, F )
max
by the mapping properties of pseudos, and so f ∈ Dom(Pg,hE ,hF
) since
ΓL2loc (M, E) = ΓL2 (M, g; E, hE ), ΓL2loc (M, F ) = ΓL2 (M, g; F, hF )
by the compactness of M .
k (M, E). Since locally E is
max max
To show that Pg,h E ,hF
= P g,hE ,hF , let f ∈ Dom(Pg,h E ,hF
) ⊂ ΓHloc
k
dense in Hloc , using a partition of unity which is subordinate to a countable atlas of control
data, we can pick a sequence
(fn ) ⊂ ΓE (M, E)
such that fn → f in ΓHloc k (M, E), so that

fn → f in ΓL2loc (M, E) (note: k ≥ 0)


and
Pg,hE ,hF fn → Pg,hE ,hF f in ΓL2loc (M, F ).
Since M is compact, (fn ) ⊂ ΓD (M, E), and the first convergence also implies convergence
in ΓL2 (M, g; E, hE ), and the second convergence implies convergence in ΓL2 (M, g; F, hF ),
max
and so f ∈ P g,hE ,hF , showing that Pg,hE ,hF
= P g,hE ,hF .
−k
b) Pick a parametrix Q ∈ Ψ (M ; F, E) for P , so that
QP = 1 + R : ΓD (M, E) −→ ΓD (M, E), P Q = 1 + S : ΓD (M, F ) −→ ΓD (M, F ),
where R, S have a smooth integral kernel. It follows that
Qmax max max
g,hE ,hF Pg,hE ,hF = 1 + Rg,hE ,hF ,
max
Pg,hE ,hF
Qmax max
g,hE ,hF = 1 + Sg,hE ,hF .

Since
Qg,hE ,hF : ΓL2c (M, F ) −→ ΓHck (M, E) ,→ ΓL2c (M, E)
is continuous (note: k ≥ 0) and since by the compactness of M we have
ΓL2c (M, F/E) = ΓL2 (M, g; F/E, hF /hE ),
Global Analysis 2 95

it follows that Qmax max max


g,hE ,hF is bounded. Moreover, Rg,hE ,hF /Sg,hE ,hF is an integral operator with
a smooth kernel, and so by compactness of M in fact a Hilbert-Schmidt operator and so
max
compact. It follows that Pg,h E ,hF
= P g,hE ,hF is invertible modulo compact operators.
c) The assumption f ∈ Dom(P g,hE ,hF ) with (P g,hE ,hF − λ)f = 0 implies (Pg,hE ,hF − λ)f = 0
as a distribution, so f is smooth by local elliptic regularity.
Using Lemma 7.3 with part a) of this theorem, the compactness of M implies
(P g,hE ,hF − λ)−1 : ΓL2 (M, g; E, hE ) −→ ΓHloc/c
k (M, F )
continuously. Using the Rellich compactness theorem (M compact!) implies that the
embedding
ΓHloc/c
k (M, F ) ,→ ΓL2 (M, g; F, hF )
is compact, so
(P g,hE ,hF − λ)−1 : ΓL2 (M, g; E, hE ) −→ ΓL2 (M, g; F, hF )
is compact, being a product of a bounded and a compact operator.
d) Clearly P̃g,hE ,hF is symmetric, and so it remains to show

Ker(P̃g,hE ,hF
± i) = {0}.
∗ max
The operator P̃g,hE ,hF
is easily seen to be equal to Pg,h E ,hF
. Thus we have to show that
f ∈ ΓL2 (M, g; E, hE ), Pg,hE ,hF f ∈ ΓL2 (M, g; F, hF ), (Pg,hE ,hF ± i)f = 0 as a distribution
implies f = 0. But by local elliptic regularity (Pg,hE ,hF ± i)f = 0 as a distribution implies
that f is smooth, and so (P̃g,hE ,hF ± i)f = 0 in the usual sense. Thus, if f 6= 0 then f
would be an eigenvector of P̃g,hE ,hF corresponding to a complex eigenvalue, contradicting
the symmetry of P̃g,hE ,hF . Note here that compactness of M is used through to conclude
that f is compactly supported!


9. The Hodge-Theorem for compact manifolds


Let (M, g) be a Riemannian manifold of dimension m := dim M . Given j ∈ {0, . . . , m} set
Ωj× (M ) := Γ× (M, ∧j (T ∗ M )C ),
where × = D, E , E 0 , D 0 and so on (spaces of j-forms that do not depend on g). Each
(complexified) vector bundle ∧j (T ∗ M )C → M canonically inherits a metric g∧j from the
Riemannian metric g on M , and thus canonically becomes a metric vector bundle. We
then set
ΩjL2 (M, g) := ΓL2 (M, g; ∧j (T ∗ M )C , g∧j ).
Let
dj := d|Ωj (M ) : ΩjD (M ) −→ Ωj+1D (M )
D
be the exterior derivative acting on smooth j-forms, a first order differential operator.
Then the Laplace-Beltrami operator acting on j-forms is defined by
∆j,g := dgj dj + dj−1 dgj−1 : ΩjD (M ) −→ ΩjD (M ),
96 B. GÜNEYSU

where P g denotes the formal adjoint with respect to (g, g∧j , g∧k ). This is a second order
elliptic differential operator which is formally self-adjoint: ∆gj,g = ∆j,g . While the formal
self-adjointness is obvious, the ellipticity follows from a somewhat lenghty calculation which
shows that in each chart U the second order symbol of ∆j,g is given for x ∈ U , ζ ∈ Rm , by
!
X
g kl (x)ζk ζl idCr : Cr −→ Cr ,
k,l

where r is the rank of ∧j (T ∗ M )C (one says that ∆j,g has a scalar symbol ).
Let us denote the induced densely defined operator from ΩjL2 (M, g) to ΩjL2 (M, g) again with
∆j,g (so Dom(∆j,g ) = ΩjD (M )). This is operator is obviously symmetric and nonnegative.
Moreover, Theorem 8.2 shows that ∆j,g is essentially self-adjoint, if M is compact. In other
words, is M is compact then ∆j,g is the unique self-adjoint extension of ∆j,g . 15
Denote with dj,g the operator dj acting from ΩjL2 (M, g) to Ωj+1
L2 (M, g) with domain of
definition ΩD (M, g) and with δj,g the operator dj acting from Ωj+1
j g j
L2 (M, g) to ΩL2 (M, g)
with domain of definition Ωj+1
D (M, g).

Define
b j (M, δg ) := Ran δj,g ,


b j (M, dg ) := Ran dj−1,g ,


b j (M, g) := Ker ∆j,g , Ωjhar b j (M, g) ∩ Ωj (M ).

Ω har (M, g) := Ω har E

Set also
Ωj (M, δg ) := Ran(δj,g ), Ωj (M, d) := Ran (dj−1 ) .
Theorem 9.1 (Hodge). Assume M is compact.
a) One has
Ωb j (M, g) = Ωj (M, g),
har har
and this space is finite dimensional and called the space of harmonic j-forms on (M, g).
b) One has
(60) ΩjL2 (M, g) = Ω
b j (M, g) ⊕ Ω
har
b j (M, δg ) ⊕ Ω
b j (M, dg ).

c) One has
ΩjE (M ) = Ωjhar (M, g) ⊕ Ωj (M, δg ) ⊕ Ωj (M, d).
15in
fact, one can show that ∆j,g always has self-adjoit extensions even in the noncompact case, and by
a theorem of Chernoff, ∆j,g is essentially self-adjoint if (M, g) is geodesically complete, which means that
the distance function
%g (x, y) := inf{`g (γ) : γ : [0, 1] → M is a smooth curve from x to y}
R1
on M is complete, where `g (γ) := 0 |γ̇(s)|g ds denotes the lenght of γ. Think of M = Rm with its
Euclidean metric (but not an open subset thereof!)
Global Analysis 2 97

d) One has
Ker (dj ) = Ωjhar (M, g) ⊕ Ωj (M, d).
e) Let Hj (M ) denote the j-th singular cohomology group of the topological space (!) under-
lying M . Then one has Hj (M ) ∼ j
= Ωhar (M, g) naturally as groups, in particular, the Betti
numbers bj (M ) := rankHj (M ) are finite.
Proof : a) This follows from Theorem 8.2 c).
b) The spaces on the RHS of (60) are easily seen to pairwise orthogonal, because of d2 =
0 = (δg )2 = 0. It remains to show that every α ∈ ΩjL2 (M, g) can be written as a sum of
this form. As ∆j,g is Fredholm by Theorem 8.2 b) and so has a closed range, and as this
operator is self-adjoint, we have

(61) ΩjL2 (M, g) =Ran(∆j,g )⊥ ⊕ Ran(∆j,g ) = Ker(∆j,g ) ⊕ Ran(∆j,g )
(62) = Ker(∆j,g ) ⊕ Ran(∆j,g ),
and so because of
∆j,g = δj,g dj,g + dj−1,g δj−1,g
we have
α = α1 + ∆j α2 = α1 + δj,g dj,g α2 + dj−1,g δj−1,g α2 ,

where α1 ∈ Ker ∆j,g . This proves b).
c) Let α ∈ ΩjE (M ) and decompose α = α1 + ∆j,g α2 according to (61). Then
∆j,g α2 = α − α1
is smooth and so α2 is smooth by local elliptic regularity, and the proof is completed upon
writing
α = α1 + ∆j,g α2 = α1 + ∆j,g α2 = α1 + δj,g dj α2 + dj−1 δj−1,g α2 .
d) This follows by combining
Ωjhar (M, g) ⊕ Ωj (M, d) ⊂ Ker (dj ) , Ωj (M, δg ) ∩ Ker (dj ) = {0}
with c).
e) By de Rham’s Theorem and part d) we have
 
Hj (M ) ∼
= Ker (dj ) /Ωj (M, d) ∼
= Ωhar (M, g) ⊕ Ωj (M, d) /Ωj (M, d) ∼
j j
= Ωhar (M, g).


10. Some remarks on the Atiyah-Singer index theorem


Let (M, g) be a Riemannian manifold of dimension m.
Definition 10.1. A Dirac bundle over (M, g) is given by a tuple (E, hE , c, ∇), which is
also depicted as (E, hE , c, ∇) → (M, g), such that
• (E, hE ) → M is a Hermitian vector bundle
98 B. GÜNEYSU

• c is a Clifford multiplication on (E, hE ) → M , in the sense that c is a homomorphism


of real (!) vector bundles
c : T ∗ M −→ End(E),
such that for all real-valued α ∈ Ω1E (M ) one has
(63) c(α) = −c(α)hE , c(α)hE c(α) = |α|2g ;
in other words, (E, hE , c) → (M, g) is a Clifford module
• ∇ is a Clifford connection on the Clifford module (E, hE , c) → (M, g), that is, ∇
is a metric covariant derivative on (E, hE ) → M such that for all α as above, all
vector fields A on M , and all ψ ∈ ΓE (M, E), one has
(M,g)
∇A (c(α)ψ) = c(∇A α)ψ + c(α)∇A ψ,
where ∇(M,g) denotes the (dual) Levi-Cevita connection.
Every such structure canonically induces a Dirac operator:
Definition 10.2. If (E, hE , c, ∇) → (M, g) is a Dirac bundle, then the associated Dirac
operator is the first order differential operator
D(g, hE , c, ∇) : ΓE (M, E) −→ ΓE (M, E)
given by
m
X
D(g, hE , c, ∇)ψ(x) := c(e∗j )(x)∇ej ψ(x), ψ ∈ ΓE (M, E),
j=1

where e1 , . . . , em is an orthonormal frame on (T M, g) → M which is defined on an open


neighborhood U of x ∈ M , and where e∗1 , . . . , e∗m ∈ Ω1E (U ) is the corresponding dual
orthonormal frame.
It is easy to check that D(g, hE , c, ∇) is formally self-adjoint and elliptic.
Typical applications in differential topology and global analysis require an additional Z2 -
grading of (or a “super structure” on) the underlying Dirac bundle:
Definition 10.3. A Z2 -grading of a Dirac bundle (E, hE , c, ∇) → (M, g) is an endomor-
phism of vector bundles ω : E → E which satisfies the following assumptions:
• ω hE = ω, ω 2 = idE
• for all real-valued α ∈ Ω1E (M ), ψ ∈ ΓE (M, E), one has
c(α)ωψ + ωc(α)ψ = 0
• for all vector fields A on M , ψ ∈ ΓE (M, E), one has
∇A (ωψ) − ω∇A ψ = 0.
In other words, ω is a fiberwise unitary map onto its image, which anti-commutes with the
Clifford multiplication and commutes with the covariant derivative. A tuple of the form
(E, hE , c, ∇, ω) → (M, g) as above will be called a Z2 -graded Dirac bundle.
Global Analysis 2 99

If (E, hE , c, ∇, ω) → (M, g) is a Z2 -graded Dirac bundle, then we get the hermitian vector
bundles (E± , hE ) → M which are given by E± := Ker(ω ± 1), so that E = E+ ⊕ E− → M
in an hE -orthogonal sense. One can easily show that D(g, hE , c, ∇) maps ΓE (M, E+/− ) to
ΓE (M, E−/+ ) and the induced differential operators are denoted by
Dω (g, hE , c, ∇) : ΓE (M, E+ ) −→ ΓE (M, E− ).

Remark 10.4. If M is compact, our results entail that the closure D(g, hE , c, ∇) of
D(g, hE , c, ∇) acting in ΓL2 (M, g; E, hE ) is self-adjoint and Fredholm. However, its in-
dex is (by self-adjointness) = 0. This is why one considers the index of the Fredholm
operator Dω (g, hE , c, ∇) from ΓL2 (M, g; E+ , hE ) to ΓE (M, g; E− , hE ) instead.

Now one canonically constructs a (non-homogeneous) closed differential form Â(M, g) ∈


Ω∗ (M ) whose cohomology class A(M ) ∈ H ∗ (M ) does not depend on g and is called the
A-hat-genus.
To each Z2 -graded Dirac bundle (E, hE , c, ∇, ω) → (M, g) one can canonically construct
a (non-homogeneous) closed differential form Chrel (g, hE , c, ∇, ω, g) ∈ Ω∗ (M ) whose coho-
mology class Chrel (E) ∈ H ∗ (M ) does not depend on all data is called the relative Chern
cgaracter of E.
Theorem 10.5 (Atiyah-Singer index theorem, 19). If M is compact, oriented, and even-
dimensional, then for every Z2 -graded Dirac bundle (E, hE , c, ∇, ω) → (M, g) one has
Z

ind Dω (g, hE , c, ∇) = Â(M, g) ∧ Chrel (g, hE , c, ∇, ω, g),
M

where the integral of a non-homogeneuous differential form is understood to be the integral


of its top-degree part.

Remark 10.6. 1. The form Â(M, g) ∧ Chrel (g, hE , c, ∇, ω, g) is closed and by Stokes’
Theorem, its integral only depends on its cohomology class Â(M )Chrel (E), so the RHS of
the Atiyah-Singer index theorem can be written as
Z
Â(M )Chrel (E),
M

to indicate that this number only depends on the topology of M and E and not on any
geometry! So Atiyah-Singer compares a topologically defined number (the RHS) with a
geometrically defined number (the LHS).
2. Vice versa, sometimes, the LHS is easily seen to be given by a topologically defined
number, and one likes to see that it can be calculated geometrically.
Example 10.7. 1. Given (M, g), the choice E = ∧(T ∗ M )C → M with its natural metric,
∇ the induced Levi-Civita connection, and
c(α)β := α ∧ β − ιαg β,
100 B. GÜNEYSU

with ιA denoting the operation of contraction with the vector field A, define a Dirac bundle.
In this case, the induced Dirac operator turns (after a short calculation) to be
d + dg : Ω∗E (M ) −→ Ω∗E (M ),
so that
∗ ∗
(d + dg )2 =: ∆g = ⊕mj=0 ∆j,g : ΩE (M ) −→ ΩE (M ).
If M is even-dimensional, then the above Dirac bundle carries the natural Z2 -grading
ω(βk ) := (−1)k βk for βk ∈ ΩkE (M ), which splits the exterior bundle into even and odd
degree forms. In particular, if M is compact and even-dimensional, then by the Hodge-
Theorem one easily finds
ind((d + dg )ω ) = χ(M ),
the Euler characteristic of M (a topological invariant!). If M is in addition oriented, Atiyah-
Singer applies, and the integrand of its RHS can be evaluated (by a lengthy calculation)
to be e(M, g) ∈ Ω∗ (M ), the Euler form, and we obtain
Z
χ(M ) = e(M, g),
M
the Gauss-Bonnet-Chern theorem. If m = 2, that is, if (M, g) is a compact Riemannian
surface, then e(M, g) is the 2-form Kg dµg , where Kg : M → R is the scalar curvature and
dµg is read as a volume-form.
2. If (M, g) is even-dimensional and oriented, then there is a second natural Z2 -grading ω̃,
which is usually called the chiral grading, and which is given by
√ m +k(k−1)
ω̃(βk ) := −1 2 ∗ βk , βk ∈ ΩkE (M ),
where ∗ is the Hodge star operator. If in addition M is compact and m/4 ∈ N, then the
Atiyah-Singer index theorem produces Hirzebruch’s signature theorem in this situation.
3. If (M, g) is an even-dimensional spin-manifold (the existence of a spin structure is a
topological obstruction), then one canonically gets a Z-graded Dirac bundle over (M, g).
In this case, one has Chrel (E) = 1, so
Â(M )Chrel (E) = Â(M ).
R
History: Why is M
Â(M ) in integer??

References
[1] Crainic, M. & Van den Ban, E.: Analysis on manifolds. Lecture notes, available online.
[2] Cycon, H.L. & Froese, R.G. & Kirsch, W. & Simon, B.: Schrödinger Operators: With Applications
to Quantum Mechanics and Global Geometry. (book).
[3] Güneysu, B. & Guidetti, D. & Pallara, D.: L1 -elliptic regularity and H = W on the whole Lp -scale
on arbitrary manifolds. Ann. Acad. Sci. Fenn. Math. 42 (2017), no. 1, 497–521.
[4] Hörmander: The analysis of linear partial differential operators, vol. I.
[5] Infusino, M.: Topological Vector Spaces II. (lecture notes). Available from http://www.math.uni-
konstanz.de/ infusino/TVS-II-WS19-20/Notes-TVS-II-WS2019-20.pdf
[6] Kato, T.: Perturbation theory for linear operators. (book).
Global Analysis 2 101

[7] Kobayashi, S. & Nomizu, K..: Foundations of differential geometry. Vol. I, Vol. II, Reprint of the
1963 original. Wiley Classics Library. A Wiley-Interscience Publication. John Wiley & Sons, Inc.,
New York, 1996.
[8] Köthe: Topological vector spaces 1, 2 (books).
[9] Lesch, M.: Global Analysis 2, Lecture notes from 2011.
[10] Rudin: Functional analysis (book).
[11] Schaefer & Wolff: Topological vector spaces (second edition) (book).
[12] Shubin: Pseudodifferential operators (book).
[13] Treves: Fourier integral operators, part 1 (book).
[14] Treves: Topological vector spaces, distributions, and kernels (book).
[15] Wagner, P. A New Constructive Proof of the Malgrange-Ehrenpreis Theorem, The American Mathe-
matical Monthly Vol. 116, No. 5 (May, 2009), pp. 457–462.
[16] Weidmann, J.: Lineare Opertoren in Hilberträumen, Teil 1. (book).

You might also like