Dr Jens Wirth

Department of Mathematics
Imperial College London

Analytical Methods for

Partial Differential Equations
(M4P41/MSP11)

London, May 6, 2010

Contents

1 Introduction 4

2 Basic concepts 6
2.1 Motivation: Why Fourier analysis is useful for us . . . . . . . . . . . . . . . . . 6
2.2 Basic properties of the Fourier transform . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Tempered distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Convolution revisited and fundamental solutions to PDE . . . . . . . . . . . . . 16
2.5 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Pseudo-differential operators 21
3.1 Integral representations of differential operators . . . . . . . . . . . . . . . . . . 21
3.2 Hörmander symbols and their calculus . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Mapping properties of pseudo-differential operators . . . . . . . . . . . . . . . . 28
3.4 Ellipticity and parametrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1 Introduction
1.1. (Linear) Partial differential differential equations of order m are equations for an unknown
function u : Rn → C involving its partial derivatives up to order m,
X
Au(x) := aα (x)∂ α u(x) = f (x). (1.1)
|α|≤m

Here we use the multi-index notation α = (α1 , . . . , αn ) ∈ Nn0 with the conventions |α| =
α1 + · · · + αn and ∂ α = ∂xα11 · · · ∂xαnn . We can view A as operator acting on differentiable1
functions A : C m (Rn ) → C(Rn ), provided that the coefficients aα : Rn → C are continuous.
We may ask several questions related to problem (1.1). As a motivation for the later following
considerations we will discuss a few of them.
1. Solvability. We fix a right hand side f : Rn → C. Can we find a function u : Rn → C such
that Au = f . Is that possible for all right hand sides f (from a certain class of functions,
e.g., f ∈ C ∞ (Rn ))?
2. Admissible problems. Assume we can solve the problem for arbitrary smooth right hand
sides. What kind of additional condition can we impose to make the solution unique?
This is a natural question to ask. If we consider the special case of n = 1, i.e. of ordinary
differential equations of order m, one usually specifies m initial conditions to obtain a
unique solutions (or m ‘boundary’ conditions at different points).
In higher dimensions such an additional condition might be of the form Bu|Σ = g for
a certain submanifold Σ ⊂ Rn , g : Σ → C and B a second partial differential operator
(usually of order < m).
3. Regularity. If f has a certain regularity, can we predict/calculate the regularity of solu-
tions u?
4. Singularities. Assume that regularity of f does not transfer to regularity of u. How
do singularities of the initial/boundary data g : Σ → Rn influence the regularity of the
solution u?
Of course we can replace Rn by an open subset of Rn and ask the same questions.

1.2. We will consider some concrete and simple examples. We consider on an open subset
Ω ⊂ Rn the Laplacian
n
X
∆= ∂x2j . (1.2)
j=1

If we know that the boundary ∂Ω is smooth, the following boundary value problem2 makes
1
We follow the standard conventions and denote for an arbitrary open set U ⊂ Rn by C(U ) the set of all
continuous functions on U , while C k (U ) denotes all k times continuously differentiable functions. C ∞ (Rn ) =
k n
T
k C (R ) denotes smooth functions.
2
In physics this problem models the potential of an electric field induced by the density of charge f and if
g ≡ 0 inside a metal surface of shape ∂Ω. If g 6≡ 0 we fix the potential at the boundary.

4
sense. If f ∈ C ∞ (Ω) and g ∈ C ∞ (∂Ω), then there exists a unique u ∈ C ∞ (Ω) such that

∆u = f, u ∂Ω
= g. (1.3)

Smoothness of g is not essential here, we could do with much less and still get a smooth solution
inside.

1.3. We consider on R1+n = Rt × Rnx the initial value problem3

∂t u − ∆u = f, u(0, ·) = g. (1.4)

Then for any bounded smooth f and bounded g (even with discontinuities) we find a unique
smooth solution in R1+n
+ = (0, ∞) × Rn . However, there exists in general no solution on the
1+n
full space R .

1.4. We consider on R1+n = Rt × Rnx the initial value problem4

∂t2 u − ∆u = f, u(0, ·) = g, ∂t u(0, ·) = h. (1.5)

Then for arbitrary smooth f , g and h we find a unique smooth solution on R1+n . However, if
g or h have singularities (points with a lack of smoothness), these singularities will extend to
singularities of u even away from the initial line t = 0.

1.5. We will not prove these statements here. Some of them will appear with a proof later on
or can be obtained as results of exercises. We will not address all problems mentioned under
point 1.1, some of them turn out to be rather hard in a general framework. We will mainly
focus on the development of tools to understand the existence of solutions and to investigate
regularity properties of solutions later on. This will be done on a rigorous basis and in the
framework of distribution theory and pseudo-differential operators.

3
This equation models propagation of heat. It is not surprising that we can only solve it in a forward direction.
4
This equation models propagation of waves, e.g. electromagnetic waves in vacuum or acoustic waves in air.
Solutions are supposed to propagate initial disturbances with the speed of sound or light and therefore also
singularities, which might be seen as very sharp disturbances, propagate.

5
2 Basic concepts

2.1 Motivation: Why Fourier analysis is useful for us

2.1.1. We start with a simple observation. If we want to find solutions to the problem
X
Au(x) = aα ∂ α u(x) = 0 (2.1.1)
|α|≤m

with constant coefficients aα ∈ C, we may ask for particular solutions of the form
uξ (x) = eix·ξ (2.1.2)
for ξ ∈ Rn and with x · ξ = j xj ξj the usual Euclidean inner product. Since ∂xj uξ = iξj uξ the
P
differential equation Au = 0 rewrites as algebraic equation (for this particular kind of solutions)
X
Auξ = a(iξ)uξ = 0, a(iξ) = aα (iξ)α . (2.1.3)
|α|≤m

We observe two things. First, such particular solutions exist if and only if iξ is a zero of the
polynomial a(iξ). And second, if we want to analyse Au = f and can write f as combination
of exponentials of the form (2.1.2), the problem is reduced to division by the polynomial a(iξ).
Unfortunately, exponentials of the form (2.1.2) do not belong to most of the interesting function
spaces1 .
2.1.2. One particular instance when the ideas from 2.1.1 can be applied in direct form are
periodic problems. Thus if we want to solve Au = f for the operator A from 2.1.1 with given
smooth and periodic f (2π-periodic in all variables), then we can write f as Fourier series
X Z
−n/2
f (x) = b ix·ξ
f (ξ)e , f (ξ) = (2π)
b f (x)e−ix·ξ dx ∈ C (2.1.4)
ξ∈Zn [−π,π]n

and ask whether corresponding periodic solutions u exists. Since at least on a formal level
!
X X X
A b(ξ)eix·ξ =
u u(ξ)eix·ξ =
a(iξ)b fb(ξ)eix·ξ , (2.1.5)
ξ∈Zn ξ∈Zn ξ∈Zn

this results in dividing Fourier coefficients fb(ξ) by a(iξ),

fb(ξ)
u
b(ξ) = , (2.1.6)
a(iξ)
and ensuring that we are indeed allowed to differentiate the series of u term by term. Most of
this goes well, if we assume that the set of real zeros of a(iξ) is bounded. We leave it as an
exercise for the interested reader to work out the details, especially,
1
They clearly belong to L∞ (Rn ), however most of the interesting theory is done in Lp (Rn ) with 1 < p < ∞.
In particular p = 2 is of interest as we will see later on. We have to spend a bit of work to use this form of
algebraifying PDE.

6
 2.2 Basic properties of the Fourier transform

1. for which f we can find periodic solutions,

2. whether they are smooth,
3. and how many different solutions can be found (i.e. what dimension is the set of all
solutions).

2.1.3. We are not interested in periodic solutions, so we can not use Fourier series. However,
we may ask for solutions represented by integrals of the form
Z
−
f (x) = eix·ξ fb(ξ)d ξ, (2.1.7)
Rn

−
where we use the convention d ξ = (2π)−n dξ for technical reasons (to become apparent later).
We will spend the main part of this first chapter to make the following formal calculation
rigorous: Assume u and f can be expressed as integrals of this form. Then formally
Z Z Z
ix·ξ − ix·ξ − ! −
Au = A e u b(ξ)d ξ = e a(iξ)b u(ξ)d ξ = eix·ξ fb(ξ)d ξ. (2.1.8)
Rn Rn Rn

If such representations are unique, this means

a(iξ)b
u(ξ) = fb(ξ) (2.1.9)

b(ξ) = fb(ξ)/a(iξ) (as long as a(iξ) 6= 0).

and we obtain u

2.2 Basic properties of the Fourier transform

2.2.1 Definition. Let f ∈ L1 (Rn ). We define its Fourier transform Ff as
Z
Ff (ξ) = fb(ξ) = e−ix·ξ f (x)dx. (2.2.1)
Rn

2.2.2 Lemma. The Fourier transform is a bounded linear operator F : L1 (Rn ) → L∞ (Rn )
satisfying
Z Z
−ix·ξ
kFf k∞ = ess sup e f (x)dx ≤ |f (x)|dx = kf k1 . (2.2.2)
ξ∈Rn Rn Rn

We will abbreviate this as kFk1→∞ ≤ 1.

We can do even a bit more, by Lebesgue’s dominated convergence theorem2 it follows that
the Fourier transform of an integrable function is in fact continuous. Also this is not the best
possible statement; it turns out to be quite difficult to characterise the image of L1 (Rn ) under
the Fourier transform F.

2.2.3 Theorem (Riemann-Lebesgue Lemma). Let f ∈ L1 (Rn ). Then its Fourier transform fb
is a uniformly continuous function on Rn vanishing at infinity, i.e. fb(ξ) → 0 as ξ → ∞.
2
Lebesgue’s dominated convergence theorem. Let {fk }∞ k=1 be a sequence of measurable functions on Ω
such that fk → f pointwise almost everywhere on Ω as k → ∞. R Suppose there is
R an integrable function
g ∈ L1 (Ω) such that |fk | ≤ g for all k. Then f is integrable and Ω f dx = limk→∞ Ω fk dx.

7
2 Basic concepts

Proof. We consider first the particular case of f (x) = 1Q (x) for Q = [−1, 1]n ⊂ Rn . Then for
any ξ not on the co-ordinate axes we can calculate the Fourier transform
n Z 1 n n
eiξj − e−iξj
Z
−ix·ξ
Y
−ixj ξj
Y
n
Y sin ξj
fb(ξ) = e dx = e dxj = =2 , (2.2.3)
Q j=1 −1 j=1
iξj j=1
ξj

which is clearly a smooth (even real-analytic) function decaying at infinity.

Next, we observe that for all scalings and translations of Q the Fourier transform is also
smooth and the linear span of characteristic functions of all such cubes is dense3 in L1 (Rn ).
Boundedness of Fourier transform implies that its image is contained in the closure of smooth
functions decaying at infinity in the uniform topology, which is the space of continuous functions
vanishing at infinity.
As last step we observe that continuous functions vanishing at infinity are continuous on the
one-point compactification of Rn and therefore uniformly continuous.

2.2.4. Following L. Schwartz we single out a special class of functions which are particularly
useful when dealing with the Fourier transform. The Schwartz space is defined as

S(Rn ) = {f ∈ C ∞ (Rn ) | ∀α, β ∈ Nn0 : xα ∂ β f (x) ∈ L1 (Rn )}. (2.2.4)

It becomes a Fréchet space4 if we endow it with the set of seminorms

pα,β (f ) = kxα ∂ β f (x)k1 . (2.2.5)

2.2.5 Examples. (a) All compactly supported smooth functions belong to Schwartz space.
To see that such functions exist, we are going to construct one. Let h : R → R be defined as
(
e−1/x , x > 0,
h(x) = (2.2.6)
0, x ≤ 0.

Then h(x) ∈ C ∞ (R) (which is easy to check and should be done as exercise) and, therefore,
the function
ψ(x) = h(x)h(1 − x) (2.2.7)
is an example of a compactly supported smooth function. Examples in higher dimensions can
be constructed as tensor products. We denote all compactly supported smooth functions as
C0∞ (Rn ) (or C0∞ (Ω) for functions supported within some open set Ω ⊆ Rn ).
2
(b) The Gaussian φ(x) = e−|x| belongs to Schwartz space S(Rn ).
(c) Schwartz space forms an algebra. Sums and products of Schwartz functions are Schwartz.

2.2.6 Lemma. Let f ∈ S(Rn ). Then ∂d

α αb \ α αb
x f (ξ) = (iξ) f (ξ) and (−ix) f (ξ) = ∂ξ f (ξ).

3
This fact follows from standard arguments of measure theory. The set of all axi-parallel cubes generates the
Borel algebra on Rn and therefore the Lebesgue measure is already determined by the volume measure of
these cubes. Etc.
4
Seminorms have all properties of a norm except that pα,β (f ) = 0 does not imply f = 0. Fréchet spaces are
defined in terms of a countable system of seminorms, where

fj → f ⇐⇒ ∀α, β : pα,β (f − fi ) → 0

and are complete with respect to them.

8
 2.2 Basic properties of the Fourier transform

Proof. For Schwartz functions it is clearly allowed to interchange differentiation and the Fourier
integral.5 Therefore,
Z Z
α −ix·ξ
αb
∂ξ f (ξ) = ∂ξ e f (x)dx = e−ix·ξ (−ix)α f (x)dx (2.2.8)
Rn Rn

and simply by integration by parts

Z Z Z
−ix·ξ α |α| α −ix·ξ
dα
∂x f (ξ) = e ∂x f (x)dx = (−1) (∂x e )f (x)dx = (iξ) α
e−ix·ξ f (x)dx. (2.2.9)
Rn Rn Rn

To simplify notation it is convenient to introduce the notation D = −i∂. Then Dαx corresponds
on the Fourier side to multiplication with ξ α . We will frequently use this convention later on.
One particular consequence of the previous Lemma is the boundedness of Fourier transform in
Schwartz space.

2.2.7 Lemma. The Fourier transform is a continuous map F : S(Rn ) → S(Rn ), i.e. if fj → f
in the sense that pα,β (fj − f ) → 0 for all α, β ∈ Nn0 then also pα,β (fbj − fb) → 0 for all α, β ∈ Nn0 .

Proof. We split the proof into two steps. First observe that with qα,β (f ) = k∂ α xβ f (x)k∞
Lemma 2.2.2 in combination with Lemma 2.2.6 implies

qα,β (fb) = k∂ α ξ β fb(ξ)k∞ = kx\

α ∂ β f (ξ)k
x
α β
∞ ≤ kx ∂ f (x)k1 = pα,β (f ). (2.2.10)

Hence fj → 0 implies pα,β (fj ) → 0 and therefore also qα,β (fbj ) → 0 for all α, β. It remains to
come back to the pα,β seminorms. This follows from
Z
α β 2 n α β dx
pα,β (f ) = kx ∂ f (x)k1 ≤ k(1 + |x| ) x ∂ f (x)k∞
Rn (1 + |x|2 )n
X
≤ Cn k cγ,δ ∂ γ xδ f (x)k∞
|γ|≤|β|,|δ|≤|α|+|β|+2n
X
≤ Cn |cγ,δ |qγ,δ (f )
|γ|≤|β|,|δ|≤|α|+|β|+2n

for certain constants cγ,δ ∈ R.

In fact the qα,β are equivalent seminorms on S(Rn ) and

S(Rn ) = {f ∈ C ∞ (Rn ) | ∀α, β ∈ Nn0 : ∂ α xβ f (x) ∈ L∞ (Rn )}. (2.2.11)

We are now in a position to formulate the first main result.

5
This is a consequence of Lebesgue’s dominated convergence theorem. For completeness we give the precise
result.
Lebesgue’s differentiation theorem. Let Ω ⊂ Rm be open and f : Rn × Ω → R satisfy (1) f (x, ·) ∈ C 1 (Ω)
for almost all x ∈ Rn , (2) f (·, y), ∂yj f (·, y) ∈ L1 (Rn ) for all y ∈ Ω and (3) for all Rcompact K ⊂ Ω there
exists a Rfunction ΦK ∈ RL1 (Rn ) such that supy∈K |∂yj f (·, y)| ≤ ΦK (y). Then y 7→ Rn f (x, y)dx ∈ C 1 (Ω)
and ∂yj Rn f (x, y)dx = Rn ∂yj f (x, y)dx.

9
2 Basic concepts

2.2.8 Theorem (Fourier inversion theorem, I). The Fourier transform F : f 7→ fb is an

isomorphism of S(Rn ) onto S(Rn ) whose inverse is given by
Z
− −
f (x) = eix·ξ fb(ξ)d ξ, d ξ = (2π)−n dξ. (2.2.12)
Rn

Proof. Step 1. Fourier transform of Gaussians. In a first step we show that

Z
2 n 2
eix·ξ e−|ξ| /2 d
−
ξ = (2π)− 2 e−|x| /2 . (2.2.13)
Rn

Note, that the integral on the left splits into a product of one-dimensional integrals and it is
therefore enough to consider the corresponding one-dimensional statement. This can be tackled
by complex methods, indeed Z ∞
t2 √
e− 2 dt = 2π (2.2.14)
−∞
implies by Cauchy integral formula
Z ∞ 2
itτ − t2 − τ2
2
Z ∞
(t−iτ )2
− 2 −τ 2 /2
Z ∞−iτ
t2 √ τ2
e e dt = e e dt = e e− 2 dt = 2π e− 2 . (2.2.15)
−∞ −∞ −∞−iτ

Step 2. Let f ∈ S(Rn ). Then fb ∈ S(Rn ) and therefore the iterated integral
ZZ
−
F (x) = ei(x−y)·ξ f (y)dyd ξ (2.2.16)

exists. We want to interchange the order of integration. For this purpose we take a scaled
2
version of the Gaussian φ(ξ) = e−|ξ| /2 to cut off all large frequencies. By the Lebesgue theorem
of dominated convergence this implies
ZZ
−
F (x) = lim ei(x−y)·ξ φ(ξ)f (y)dyd ξ
→0
Z Z 
i(x−y)·ξ −
= lim e φ(ξ)d ξ f (y)dy
→0
 
x−y
Z
−n −n
= lim (2π) 2  φ f (y)dy
→0 

where we used that the Fourier integral (2.2.13) preserves the Gaussian. Substituting y = x+z
yields, again by the aid of Lebesgue’s theorem of dominated convergence,
Z
−n
= lim (2π) 2 φ(z)f (x + z)dz = f (x).
→0

2.2.9 Remark. Instead of a Gaussian we could use an arbitrary function φ ∈ S(Rn ) with
φ(0) = 1. But then we have to make use of the inversion theorem or a similar statement to
R
conclude that φ(ξ) = ψ(ξ)
b for some ψ with ψ(x)dx = 1.
2.2.10. We will collect some ideas used in the previous proof. First, we define for two functions
f, g ∈ L1 (Rn ) their convolution
Z
f ∗ g(x) = f (x − y)g(y)dy. (2.2.17)
Rn

10
 2.2 Basic properties of the Fourier transform

This is well-defined since by Fubini theorem

Z Z
kf ∗ gkL1 (Rn ) = f (x − y)g(y)dy dx
ZZ
≤ |f (x − y)| |g(y)|dydx = kf kL1 (Rn ) kgkL1 (Rn ) . (2.2.18)

In particular we observe that f ∗ g ∈ L1 (Rn ) and we can apply the Fourier transform to it. As
in the middle of the previous proof it follows that

Lemma. Let f, g ∈ L1 (Rn ). Then f[

∗ g(ξ) = fbgb.

Proof. It is easily calculated that

Z Z Z
−ix·ξ
f[ ∗ g(ξ) = e f (x − y)g(y)dydx = e−i(x−y)·ξ f (x − y)e−iy·ξ g(y)dydx

and the desired statement follows by Fubini theorem.

Convolutions are noteworthy because of several further properties. First, they are commutative,
f ∗ g = g ∗ f , and associative6 such that L1 (Rn ) becomes a commutative algebra if endowed
with the ∗ product.
Furthermore, if f, g ∈ S(Rn ) it is easy to prove that

∂ α (f ∗ g) = ∂ α f ∗ g = f ∗ ∂ α g. (2.2.19)

2.2.11. We will use convolutions to prove that we can indeed approximate Lp -functions for
p ∈ [1, ∞) by Schwartz functions with respect to the Lp -norm. The precise statement is as
follows

2.2.12 Lemma. Let f ∈ Lp (Rn ) for p ∈ [1, ∞). Then there exists a sequence fj ∈ S(Rn ) with
fj → f in Lp (Rn ), i.e. kf − fj kp → 0 as j → ∞.

Proof. Step 1. The case p = 1. Let φ ∈ C0∞ (Rn ) satisfy φ(x)dx = 1. We define the rescaled
R

functions
Z Z
−n
φ (x) =  φ(x/), φ (x)dx = φ(x)dx = 1, kφ k1 = kφk1 (2.2.20)

and consider for f ∈ L1 (Rn ) the function f = φ ∗ f . Evidently kf k1 ≤ kφk1 kf k1 and
kf k∞ ≤ kφ k∞ kf k1 ' −n kf k1 . By (2.2.19) it follows that ∂ α f = (∂ α φ ) ∗ f and therefore
f ∈ C ∞ (Rn ). We show that these functions approximate f in L1 -norm,
Z Z
kf − f k1 = f (x) − φ (y)f (x − y)dy dx
ZZ
≤ |φ (y)(f (x) − f (x − y))|dydx
Z Z
= |φ(y)| |f (x) − f (x − y)|dxdy → 0, →0

6
at least for functions; this is wrong for distributions as we will see later on

11
2 Basic concepts

where the inner integral tends to zero uniformly on the support of φ due to the continuity in
mean 7 property of the Lebesgue integral,
Z
lim sup |f (x) − f (x − y)|dx = 0.
→0 y∈supp φ

Hence, for any n there exists an f with kf − f k1 ≤ 1/(2n). To obtain Schwartz functions we
use a second trick and cut off large values of x. Let χ ∈ C0∞ (Rn ). Then
Z
kf − χ(δ·)f k1 = |f (x)(1 − χ(δx))|dx → 0 (2.2.21)

by Lebesgue theorem on dominated convergence. Hence we find δ such that kf − χ(δx)f k1 ≤
1/(2n) and therefore by triangle inequality |f − χ(δ·)f k1 ≤ 1/n. This can be done for all n
and we found a suitable sequence.
Step 2. We only sketch the major differences to the case p = 1. We define f in the same way
and conclude similarly that f ∈ Lp (Rn ) ∩ C ∞ (Rn ). Then
Z Z p
p
kf − f kp = φ(y)(f (x) − f (x − y))dy dx
Z Z
p
≤ kφkq |f (x) − f (x − y)|p dydx
Z supp φ
= kφkpq kf − Ty f kpp dy → 0, →0
supp φ

based on Hölder inequality 8 and the (Lp -) continuity in mean.

In a final step we again cut off large values of x to get Schwartz functions.
2.2.13 Theorem (Fourier inversion theorem, II). The Fourier transform F : f 7→ fb extends
−
to an unitary isomorphism (also denoted by F) acting L2 (Rn , dx) → L2 (Rn , d ξ), i.e., for all
2 n
f, g ∈ L (R ) the Plancherel identity
Z Z
−
(f, g)L2 (Rn ,dx) = f (x)g(x)dx = fb(ξ)b
g (ξ)d ξ = (fb, gb)L2 (Rn ,d− ξ) (2.2.22)
Rn Rn

holds true and F −1 = F ∗ , kFf kL2 (Rn ,d− ξ) = kf kL2 (Rn ,dx) .
7
Theorem (Continuity in mean). Let f ∈ Lp (Rn ) with p ∈ [1, ∞) and denote Ty f (x) = f (x − y). Then
limy→0 kf − Ty f kp = 0 uniform on compact sets in y.
This theorem can be obtained from the somewhat simpler statement of:
Theorem (Lebesue’s differentiation theorem). Let f ∈ L1 (Rn ). Then with B(x, r) = {y ∈ Rn | |x − y| ≤ r}
the limit Z
1
lim |f (y) − f (x)|dy = 0
r→0 |B(x, r)| B(x,r)

exists for almost every x ∈ Rn .

8
Theorem (Hölder inequality). Let f ∈ Lp (Rn ) and g ∈ Lq (Rn ) with 1/p + 1/q = 1. Then f g ∈ L1 (Rn ) with
norm estimated by kf gk1 ≤ kf kp kgkq .
For a proof of this fact note first that for 1 < p, q < ∞, 1/p + 1/q = 1 and a, b > 0 the inequality
ab ≤ ap /p + bq /q holds true. This is sometimes called Young’s inequality. Then consider two functions f, g
with kf kp = 1 and kgkq = 1. Then it follows that
1 1 1 1
kf gk1 ≤ kf kpp + kgkqq = + = 1 = kf kp kgkq .
p q p q
The general case is either trivial (p = 1, q = ∞ and vice versa) or can be reduced by scaling to the special
case treated here.

12
 2.3 Tempered distributions

Proof. Since S(Rn ) is sequentially dense9 in L2 (Rn ) it is sufficient to prove (2.2.22) for f, g ∈
S(Rn ) and use an approximation argument. Fubini theorem together with the Fourier inversion
formula on S(Rn ) implies for f, g ∈ S(Rn )
Z ZZ Z Z Z
ixξ b − −ixξ − −
f (x)g(x)dx = e f (ξ)d ξg(x)dx = f (ξ) e
b g(x)dxd ξ = fb(ξ)b g (ξ)d ξ. (2.2.23)

Assume now that f ∈ L2 (Rn ) is limit of a sequence fj ∈ S(Rn ). Then by the first part
−
kfbj − fbk kL2 = kfj − fk kL2 and fbj is a Cauchy sequence in the Hilbert space L2 (Rn , d ξ). There-
2 n −
fore it must converge to a limit v = limj→∞ fj ∈ L (R , d ξ). Furthermore, kvkL2 (Rn ,d− ξ) =
b
limj→∞ kfbj kL2 (Rn ,d− ξ) = limj→∞ kfj kL2 (Rn ,dx) = kf kL2 (Rn ,dx) . It is now easy to see that the limit
v does not depend on the choice of the sequence fj approximating f and we denote this limit
as Ff .
For the Plancherel identity we proceed similarly with two sequences (or use Hilbert space theory
to see that it is equivalent to the statement for norms).

2.3 Tempered distributions

2.3.1. To restrict all considerations to Schwartz functions would be too restrictive for a real
applicable theory. A first possible choice would be to use L2 (Rn ) or smaller spaces like the
(standard) Sobolev spaces
−
H s (Rn ) = {f ∈ L2 (Rn ) | hξis fb(ξ) ∈ L2 (Rn , d ξ)}, s≥0 (2.3.1)
with hξi = (1 + |ξ|2 )1/2 . These spaces fit well to differential operators (e.g. (1.1) with bounded
smooth coefficients aα (x) ∈ B ∞ (Rn ) = {f ∈ C ∞ (Rn ) | ∀β : ∂ β f ∈ L∞ (Rn )} maps H s (Rn ) →
H s−m (Rn ) as bounded operator for all s ≥ m) and can be described and handled in terms of
Fourier transform.
The idea of tempered distribution theory is to generalise as far as possible and extend Fourier
transform to a broad class of objects containing Lp -spaces, measures and more abstract entities.
2.3.2 Definition. The space of tempered distributions S 0 (Rn ) is the space of all continuous
linear functionals on S(Rn ), i.e. each u ∈ S 0 (Rn ) is a map u : S(Rn ) → C sucht that
1. u is linear, u(αf + βg) = αu(f ) + βu(g) for all α, β ∈ C and f, g ∈ S(Rn );
2. u is continuous, u(fj ) → u(f ) in C whenever fj → f in S(Rn ).
To simplify notation (and avoid too many levels of parentheses) we will denote u(f ) by hu, f i.
2.3.3 Examples. We will give the main examples here. Note, that the assumption of continuity
has to be checked in each case and that this breaks down to prove an estimate of the form
X
|hu, φi| ≤ cα,β pα,β (φ) (2.3.2)
α,β

for all φ ∈ S(Rn ) and a certain finite sum of seminorms.

(a) Functions f ∈ Lp (Rn ) for 1 ≤ p ≤ ∞ can be understood as distributions via
Z
hf, φi = f (x)φ(x)dx. (2.3.3)
Rn
9
By this we mean that for each f ∈ L2 (Rn ) there exists a sequence fj ∈ S(Rn ) auch that fj → f in L2 (Rn ).
This is a special case of the previous statement.

13
2 Basic concepts

To be precise we show three things. First, we recall Hölder inequality 10 for Lp -spaces. Namely,
for f ∈ Lp (Rn ) and g ∈ Lq (Rn ) with 1/p + 1/q = 1 the product f g belongs to L1 (Rn ). Using
this inequality we easily see that S(Rn ) is continuously embedded into Lq (Rn ). Then continuity
follows from |hf, φi| ≤ kf kp kφkq .
It holds even more, Lp (Rn ) is embedded into S 0 (Rn ). Different functions give rise to different
distributions as the following Lemma implies:

Lemma. Assume that f ∈ Lp (Rn ) satisfies hf, φi = 0 for all φ ∈ S(Rn ). Then f (x) = 0 a.e.

(b) If we want to remain in the framework of functions, the weakest possibility is as follows:
A function f is called locally integrable, f ∈ L1loc (Rn ) if for all compact sets K
Z
|f (x)|dx < ∞. (2.3.4)
K

We call a locally integrable function temperate (or polynomially bounded) if there exists a
polynomial p(x) > 0 such that f (x)/p(x) ∈ L1 (Rn ). All temperate locally integrable functions
give rise to a temperate distribution via (2.3.3). Such distributions are called regular.
(c) (Signed) Measures can be understood as temperate distributions via
Z
hµ, φi = φ(x)dµ(x), (2.3.5)

provided they are polynomially bounded. By the latter we mean that there exist constants C
and N such that for all balls Br = {x ∈ Rn | |x| ≤ r} of radius r around the origin and the
absolute value of the measure11
|µ|(Br ) ≤ C(1 + r)N (2.3.6)
holds true. Again, these measures can be embedded into S 0 (Rn ), different measures give rise
to different distributions.
(d) As simple concrete example we give the Dirac distribution

hδ, φi = φ(0), (2.3.7)

which corresponds to the Dirac measure δ = #(A ∩ {0}).

(e) A more sophisticated example on R are principle value distributions. We give one example,
  Z − Z ∞ 
1 φ(x) φ(x)
v.p. , φ = lim dx + dx . (2.3.8)
x →0 −∞ x  x
The abbreviation v.p. stands for valeur principale and goes back to J. Hadamard. Similar
constructions are possible for higher order singularities.

2.3.4. The space S 0 (Rn ) carries a natural notion of convergence12 . We say that uj → u in
S 0 (Rn ) if for all φ ∈ S(Rn ) the sequence huj , φi → hu, φi converges in C.
10
see page 12
11
If measures are positive, |µ| = µ. For signed real/complex measures the absolute value is defined in terms of
its Hahn/polar decomposition. We don’t want to go into details here.
12
It can be shown by a bit more of functional analysis that this weak convergence is indeed the strong conver-
gence on the dual space, i.e. it is uniform on bounded subsets of S(Rn ). Reason for this is that S(Rn ) is a
Montel space, which means that all (in all seminorms) bounded subsets of S(Rn ) are compact.

14
 2.3 Tempered distributions

2.3.5. Before defining the Fourier transform of tempered distributions we want to explain the
general philosophy behind. Let A : S(Rn ) → S(Rn ) be a continuous linear operator. Then we
can define an operator on S 0 (Rn ) by transposition,

hAt u, φi = hu, Aφi. (2.3.9)

It is evident from At u = u ◦ A that At u is again a temperate distribution. Furthermore, if

uj → u in S 0 (Rn ) also At uj → At u and therefore At : S 0 (Rn ) → S 0 (Rn ) is continuous.
2.3.6. Recall first that by Fubini theorem for two functions φ, ψ ∈ S(Rn ) the identity
hFφ, ψi = hφ, Fψi holds. Therefore, F t extends F from S(Rn ) to a continuous linear op-
erator on S 0 (Rn ). We denote this operator again as F and define thus the Fourier transform
on tempered distributions as
hFu, φi = hu, Fφi. (2.3.10)
We were fast in denoting this operator as F. The notation is justified by the following state-
ments.
Lemma. The Fourier transform F : L1 (Rn ) → L∞ (Rn ) coincides with the distributional
Fourier transform.

Proof. Fubini theorem gives hFf, φi = hf, Fφi for all f ∈ L1 (Rn ) and φ ∈ S 0 (Rn ).
−
Lemma. The unitary operator F : L2 (Rn , dx) → L2 (Rn , d ξ) defined in Theorem 2.2.13 coin-
cides with the distributional Fourier transform.

Proof. Convergence in L2 (Rn ) implies convergence in S 0 (Rn ) and therefore the identity of both
operators follows just from continuity of the distributional Fourier transform.
2.3.7 Example. F[δ] = 1 as can be seen from
Z
hFδ, φi = hδ, Fφi = φ(0)
b = φ(x)dx = h1, φi. (2.3.11)
Rn

2.3.8 Theorem (Fourier inversion theorem, III). The Fourier transform F : S 0 (Rn ) → S 0 (Rn )
is an isomorphism. The inverse is given by F −1 := (F −1 )t .
2.3.9. For two functions φ, ψ ∈ S(Rn ) the integration by parts formula gives

h∂ α φ, ψi = (−1)|α| hφ, ∂ α ψi. (2.3.12)

Because derivatives are continuous linear operators on S(Rn ) we can use this formula to extend
the notion of derivatives to tempered distributions and define ∂ α u for u ∈ S 0 (Rn ) by

h∂ α u, φi = (−1)|α| hu, ∂ α φi. (2.3.13)

This definition coincides with the classical derivatives on all polynomially bounded functions
from C |α| (Rn ). (Why?)
2.3.10 Examples. (a) The Heaviside function H(x) = 1[0,∞) is not differentiable on R, but it
can be understood as temperate distribution. Therefore, we can calculate its derivative. This
yields
Z ∞
0 0
hH , φi = −hH, φ i = − φ0 (x)dx = φ(0) = hδ, φi (2.3.14)
0

15
2 Basic concepts

for all φ ∈ S(R). Hence, H 0 = δ in the sense of distributions.

(b) The function ln |x| is locally integrable on R and clearly polynomially bounded. It’s deriva-
tive is calculated via
Z − Z ∞ 
0 0 0
h∂x ln |x|, φi = −hln |x|, φ (x)i = − lim ln |x|φ (x)dx + ln |x|φ (x)dx
→0 −∞ 
Z − Z ∞ 
φ(x) φ(x)
= lim dx + φ(−) ln  + dx − φ() ln 
→0 −∞ x  x
 
1
= v.p. , φ , (2.3.15)
x
where we used that lim→0 (φ() − φ(−)) ln  = 0 due to differentiability of φ. We have to use
the same  in both integrals to cancel the boundary terms arising from integration by parts
(otherwise we would get the same result but complicated formulæ).

2.4 Convolution revisited and fundamental solutions to PDE

2.4.1. We have transferred Fourier transform and partial derivatives to distributions. There is
one operation left out, we have not defined products of distributions. In fact, this is in general
impossible. The best we can do is to consider products of distributions u ∈ S 0 (Rn ) with certain
smooth functions f ∈ C ∞ (Rn ) ∩ S 0 (Rn ).
Definition. We say f ∈ C ∞ (Rn ) belongs to the multiplier algebra OM (Rn ) if for all multi-
indices α ∈ Nn0 we can constants Cα and Nα such that

|∂ α f (x)| ≤ Cα hxiNα . (2.4.1)

For each such function f the multiplication operator f · : φ 7→ f φ is a continuous linear map
on S(Rn ) and we can therefore define

hf u, φi = hu, f φi (2.4.2)

for all tempered distributions u ∈ S 0 (Rn ).

Clearly, OM (Rn ) is embedded into S 0 (Rn ). We denote by Oc0 (Rn ) the Fourier transforms of ele-
ments from OM (Rn ). Elements of Oc0 (Rn ) are sometimes called rapidly decaying distributions.
2.4.2. Let u ∈ S 0 (Rn ) and w ∈ Oc0 (Rn ). Then we define the convolution of u and w by the
identity
F[w ∗ u] = F[w]F[u]. (2.4.3)
This definition is rather impractical and also does not contain the L1 ∗ L1 case. However, at
least for some particular cases we can reformulate it in a more comprehensive way.
(a) If both u and w belong to Schwartz class, the convolution coincides with the usual one.
(b) Assume that u and w are regular and that w has compact support. Then a convolution
may be defined by
ZZ ZZ
0
hw ∗ u, φi = w(x − y)u(y)dyφ(x)dx = w(z)u(y)φ(z + y)dydz. (2.4.4)

It can be shown that it satisfies (2.4.3) and coincides therefore with *.

16
 2.4 Convolution revisited and fundamental solutions to PDE

(c) Assume that w has compact support13 . Then w ∈ Oc0 (Rn ) and the convolution
hw ∗ u, φi = hw ⊗ u, φ(x + y)ix,y (2.4.5)
defined in a similar way coincides with our definition. This is the ‘usual’ definition of a convo-
lution.
It is usual to think of ∗ as being commutative.
2.4.3 Example. We give only one example here, which shows that we have to be careful when
dealing with several convolutions. Let δ 0 be the derivative of the Dirac distribution. Then for
any u ∈ S 0 (R)
F[δ 0 ∗ u] = F[δ 0 ]F[u] = iξ F[u] = F[u0 ] (2.4.6)
such that δ 0 ∗ u = u0 . Hence,
1 ∗ (δ 0 ∗ H) = 1 ∗ δ = 1, (1 ∗ δ 0 ) ∗ H = 0 ∗ H = 0 (2.4.7)
and, therefore, there is no associativity for convolutions.
2.4.4 Lemma. Let φ ∈ S(Rn ) and u ∈ S 0 (Rn ). Then φ ∗ u ∈ C ∞ (Rn ).

Proof. For each x ∈ Rn we can write

D E
φ ∗ u(x) = F −1 [φ(ξ)F[u]]
b = (2π)−n F[u], eix·ξ φ(ξ)
b .
ξ

The statement follows since x 7→ eix·ξ φ(ξ)

b ∈ S(Rn ) is continuous, similarly x 7→ ξ α eix·ξ φ(ξ)
b ∈
n
S(R ) for all multi-indices α.
2.4.5. Let P (D) be a partial differential operator with constant coefficients,
X
P (D) = aα Dα , aα ∈ C. (2.4.8)
|α|≤m

A S 0 -fundamental solution to P (D) is a distribution E ∈ Oc0 (Rn ) such that P (D)E = δ0 .

Such fundamental solutions are particularly useful. For any given f ∈ S 0 (Rn ) a solution to
P (D)u = f is given by the convolution u = E ∗ f .
If we apply the Fourier transform to the defining equation it means that
p(ξ)E
b = 1, (2.4.9)
hence we want to divide by the polynomial p(ξ). If p(ξ) has no real zeros, 1/p(ξ) is smooth and
belongs (obviously) to OM (Rn ). Hence, E = F −1 p(ξ)
1
gives such a fundamental solution and in
this case it is also uniquely determined.
2.4.6 Example. We consider the operator 1 − ∆. Its symbol is 1 + |ξ|2 , the corresponding
fundamental solution is
eix·ξ
Z Z
−
hE, φi = 2
φ(x)dxd ξ. (2.4.10)
Rn Rn 1 + |ξ|
The distribution E can be calculated explicitely, it is regular and given by
(
eix·ξ − O(|x|2−n ), |x| → 0,
Z
n n−2
2 |x|
− 2
E(x) = d ξ = (2π) K n−2 (|x|) = (2.4.11)
Rn 1 + ξ
2 2 O(e−c|x| ), |x| → ∞,

with the Bessel function of third kind (or MacDonald function) Kν (x).
13
Precisely, w ∈ E 0 (Rn ). See Nr. 2.4.7 below.

17
2 Basic concepts

2.4.7. If the polynomial p(ξ) has real zeros, equation (2.4.9) does not have solutions in OM (Rn ).
Solutions to P (D)u = f are not unique; any ξ0 with p(ξ0 ) = 0 gives the solution exp(ix · ξ0 ) of
the corresponding homogeneous problem P (D)u = 0. Furthermore, we need conditions on the
right hand side f for solutions to exist; fb(ξ) has to be ‘regular enough’ at zeros of p(ξ).
A way out of this is to consider the distribution spaces D0 (Rn ), the dual of C0∞ (Rn ), and E 0 (Rn ),
the dual to C ∞ (Rn ). The elements of E 0 (Rn ) can be identified with elements of D0 (Rn ) with
compact support, where
x 6∈ supp u ⇐⇒ ∃U open, x ∈ U : ∀φ ∈ C0∞ (U ) : hu, φi = 0. (2.4.12)
It follows E 0 (Rn ) ⊂ S 0 (Rn ) ⊂ D0 (Rn ) and via (2.4.5) the convolution can be defined for all
w ∈ E 0 (Rn ) and u ∈ D0 (Rn ). We will denote a solution to P (D)E = δ with E ∈ D0 (Rn ) as (E 0 -)
fundamental solution and only quote the following (fundamental) result:
Theorem (Malgrange-Ehrenpreis). For any partial differential operator P (D) with constant
coefficients there exists a fundamental solution E ∈ D0 (Rn ). Therefore, for any f ∈ E 0 (Rn ) the
distribution u = E ∗ f ∈ D0 (Rn ) satisfies P (D)u = f .
2.4.8 Example. In R the operator d
dx
has the fundamental solution H(x) (since H 0 = δ).
2.4.9 Example. We consider the operator ∆. Then the symbol −|ξ|2 vanishes in ξ = 0 and
we have to be careful about calculations. Assume we want to solve ∆u = f with f ∈ S(Rn )
and n ≥ 3. Then |ξ|−2 ∈ L1loc (Rn ) and a solution is given by
1
b = −fb 2 ,
u i.e. u=f ∗E (2.4.13)
|ξ|
with
Γ(n/2 − 1) 2−n
Z
1 −
E(x) = eix·ξ
2
dξ = |x| ∈ L1loc (Rn ). (2.4.14)
Rn |ξ| 2π n/2
The formula does not extend to all f ∈ S 0 (Rn ) (since E(x) 6∈ Oc0 (Rn )), however, it makes
perfect sense as soon as f is a compactly supported distribution, f ∈ E 0 (Rn ). Hence, E(x) is a
fundamental solution in the above sense.
For n = 1 or n = 2 the same reasoning can not be applied. We leave n = 1 as an exercise and
mention that E(x) = −(2π)−1 ln |x| for n = 2. For n = 3 the above given formula simplifies to
E(x) = (4π)−1 |x|−1 . Solutions to ∆u = f are not unique, ∆w = 0 for any affine linear function
w(x) = a · x + b, a ∈ Cn , b ∈ C.
2.4.10 Example. Consider the operator 1+∆. Then the same reasoning applied to (1+∆)u =
f for f ∈ S(Rn ) yields
(1 − |ξ|2 )b
u = fb. (2.4.15)
But (1 − |ξ|2 )−1 6∈ L1loc (Rn ). So we have to explain how we understand it as distribution (in
such a way that it gives 1 if multiplied by (1 − |ξ|2 )); we use principle value integrals
Z Z
2 −1 φ(ξ)
hv.p.(1 − |ξ| ) , φi := lim ( + ) 2
dξ. (2.4.16)
→0 |ξ|<1− |ξ|>1+ 1 − |ξ|

Applying the inverse Fourier transform to v.p.(1−|ξ|2 )−1 gives a fundamental solution to 1+∆.
Again this is not unique; e.g., any regular distribution given as Fourier integral over the sphere
Sn−1 = {ξ ∈ Rn : |ξ| = 1},
Z
−
u(x) = eix·ξ $(ξ)d ξ, $(ξ) ∈ C ∞ (Sn−1 ), (2.4.17)
Sn−1

satisfies the homogeneous problem ∆u + u = 0.

18
 2.5 Sobolev spaces

2.5 Sobolev spaces

2.5.1. We want to combine Lp -spaces with the ideas of distribution theory and define therefore
certain spaces of regular distributions. For p ∈ [1, ∞] and k ∈ N0 we denote

W k,p (Rn ) = {f ∈ S 0 (Rn ) | ∀α, |α| ≤ k : ∂ α f ∈ Lp (Rn )} (2.5.1)

as Sobolev space of order k over Lp (Rn ). Sobolev spaces are Banach spaces endowed with the
norm14 X  p1
α p
kf kk,p = k∂ f kp , kf kk,∞ = max k∂ α f k∞ . (2.5.2)
|α|≤k
|α|≤k

2.5.2. For the particular case p = 2 these Sobolev spaces are Hilbert spaces with inner product
XZ
(f, g)W k,2 = ∂ α f (x)∂ α g(x)dx. (2.5.3)
|α|≤k

By Plancherel’s identity (2.2.22) this can be rewritten as

XZ Z
2α b − −
(f, g)W k,2 = g (ξ)d ξ ' hξi2k fb(ξ)b
ξ f (ξ)b g (ξ)d ξ (2.5.4)
|α|≤k

and therefore, up to equivalence of norms, the Sobolev space W k,2 (Rn ) coincides with the
previously defined Sobolev space H k (Rn ).
2.5.3. We will consider an example of a point singularity. Let d ∈ R and χ(x) ∈ C0∞ (Rn ) be
supported around x = 0. Then we can consider the ’singularity’ f (x) = |x|−d χ(x) in x = 0 and
ask to which Sobolev spaces it belongs.
It follows that f (x) ∈ Lp (Rn ) if and only if for some small 
Z Z 
−dp
|x| dx = |S | n−1
|x|−dp+n−1 dx < ∞, (2.5.5)
B 0

which is equivalent to dp < n. Similarly we see that f ∈ W k,p (Rn ) if and only if (d + k)p < n.

2.5.4. Let hDis f := F −1 [hξis fb(ξ)] for hξi = (1 + |ξ|2 )1/2 . Then hDis : H s (Rn ) → L2 (Rn ) as
can be easily seen from Plancherel identity and the definition of H s spaces in (2.3.1)
Z
s 2 −
khDi f k2 = hξi2s |fb(ξ)|2 d ξ = kf kH s (Rn ) (2.5.6)
Rn
14
That these are indeed norms follows easily from the fact that the corresponding Lp -norms,
Z 1/p
p
kf kp = |f (x)| dx , kf k∞ = esssupx |f (x)|,

are norms. For this we recall in particular the Minkowski inequality kf + gkp ≤ kf kp + kgkp , which can
be proved directly from Hölder inequality. Assume 1 < p < ∞ (the remaining cases are trivial) and apply
Hölder inequality with q = p/(p − 1). Then
Z Z (p−1)/p  Z 1/p Z 1/p 
kf + gkpp ≤ |f + g|p−1 (|f | + |g|)dx ≤ |f + g|p dx |f |p dx + |g|p dx

and the inequality follows after dividing both sides by kf + gkp−1

p .
The fact that Sobolev spaces are complete follows directly from the completeness of the Lebesgue spaces.

19
2 Basic concepts

and gives an isometric isomorphism between these spaces.

We want to extend this to spaces over Lp (Rn ). For this it is convenient to use15

Theorem (Hörmander–Mikhlin multiplier theorem). Assume that m(ξ) ∈ L∞ (Rn ) satisfies

n
|∂ α m(ξ)| ≤ C|ξ|−|α| , |α| ≤ + 1.
2

Then the operator Tm : f 7→ F −1 [m(ξ)fb(ξ)] is a bounded linear operator mapping Lp (Rn , dx)
into itself.

It implies that hDik : W k,p (Rn ) → Lp (Rn ) is a bounded and boundedly invertible linear operator
for all 1 < p < ∞ and therefore

kf kW k,p (Rn ) ' khDik f kLp (Rn , 1 < p < ∞. (2.5.7)

This can be used to define the W s,p (Rn ) spaces for real numbers s ∈ R. We leave it as an
exercise to work out the details.

2.5.5. We will conclude this chapter with embedding relations between Sobolev spaces. As in
Example 2.4.6 one can prove that

hDis f = g ⇐⇒ f = Gs ∗ g (2.5.8)

with some Gs ∈ Oc0 (Rn ), Gs = F −1 [hξi−s ]. It turns out that Gs ∈ C ∞ (Rn \ {0}), Gs (x) =
O(exp(−c|x|)), x → ∞ and Gs (x) = O(|x|s−n ) as x → 0. Hence, Gs ∈ Lp (Rn ) if s > n(1−1/p).
Young’s inequality16 implies

2.5.6 Theorem (Sobolev embedding theorem). Assume s > n(1/p − 1/r). Then the Sobolev
space of order s over Lp is contained in Lr , W s,p (Rn ) ⊆ Lr (Rn ).

2.5.7 Remark. The statement of this theorem can be improved to s = n(1/p − 1/q) provided
r < ∞. The proof of this is slightly more involved and omitted here.

15
A proof can be found in any good text book on harmonic analysis, e.g. Stein& Weiss, Introduction to Fourier
analysis in Euclidean spaces.
16
Lemma (Young’s inequality). Let f ∈ Lp (Rn ), g ∈ Lq (Rn ) then f ∗ g ∈ Lr (Rn ) with

||f ∗ g||r ≤ ||f ||p ||g||q

for 1/r = 1/p + 1/q − 1.

The proof for the particular cases p = 1 and arbitrary q and p = ∞ and q = 1 are obvious. The remaining
cases can be treated via Riesz-Thorin interpolation. We omit the details.

20
3 Pseudo-differential operators

3.1 Integral representations of differential operators

3.1.1. So far we understand differential operators with constant coefficients on Rn ; using Fourier
transform we can rewrite them as multiplication with a certain polynomial—its symbol. To be
precise, let X
P (D) = aα Dα , (3.1.1)
|α|≤m

where D = −i∂ and the aα ∈ C are constants. Then P (D)u ∈ S 0 (Rn ) can be represented for
all u ∈ S 0 (Rn ) as
P (D)u = F −1 [p(ξ)b
u], (3.1.2)
where the polynomial X
p(ξ) = aα ξ α (3.1.3)
|α|≤m

denotes the symbol of the differential operator. At least for u ∈ S(Rn ) we can write (3.1.2) as
iterated integral ZZ
−
P (D)u(x) = ei(x−y)·ξ p(ξ)u(y)dyd ξ. (3.1.4)

If we know that p(ξ) has no zeros, we can solve P (D)u = f for f ∈ S(Rn ) similarly by an
iterated integral ZZ
−1 1 −
u = P (D) f = ei(x−y)·ξ f (y)dyd ξ. (3.1.5)
p(ξ)
This formula can be extended by continuity to a broader class of f (and even to distributions
as u = E ∗ f with the fundamental solution E = F −1 [1/p(ξ)] ∈ Oc0 (Rn )).
3.1.2. As a next step we want to consider differential operators with variable coefficients, i.e.
X
P (x, D) = aα (x)Dα (3.1.6)
|α|≤m

with certain functions aα (x). Such operators are continuous on S(Rn ) if (and only if) the
coefficients satisfy aα (x) ∈ OM (Rn ) and using the arguments of the previous chapter, §-2.3.5,
we can extend such operators by transposition to S 0 (Rn ). Note for this that
X
P t (x, D)φ(x) = (−1)|α| Dα (aα (x)φ(x)), (3.1.7)
|α|≤m

which can be brought to the form 3.1.6 by product rule of differentiation. Then P (x, D)u for
u ∈ S 0 (Rn ) is defined via hP (x, D)u, φi = hu, P t (x, D)φi for all φ ∈ S(Rn ).
Again we want to write the action of P (x, D) as iterated integral; the Fourier inversion formula
on S(Rn ) tells us that
ZZ
−
P (x, D)u(x) = ei(x−y)·ξ p(x, ξ)u(y)dyd ξ (3.1.8)

21
3 Pseudo-differential operators

for all u ∈ S(Rn ) and with X

p(x, ξ) = aα (x)ξ α (3.1.9)
|α|≤m

again denoted as symbol of the differential operator. We anticipate that under certain condi-
tions on this symbol the operator P (x, D) might be invertible and that the inverse also has a
representation of such a form. Main aim of this chapter is to show that this is indeed the case.
3.1.3. Following the scheme of the motivational Section 2.1 we include again a short comment
on periodic problems. Assume that all coefficients aα (x) are 2π-periodic and that we are looking
for 2π-periodic solutions of P (x, D)u = f for some 2π-periodic right hand side f . In this case
it might be a good idea to write P (x, D) in terms of a series (applying the discrete Fourier
inversion formula)
XZ
−n/2
P (x, D)u = (2π) ei(x−y)·ξ p(x, ξ)u(y)dy (3.1.10)
ξ∈Zn

and again anticipate that under certain conditions on the symbol p(x, ξ) the inverse operator
exists and solutions can be written as
XZ
−1 −n/2
u(x) = P (x, D) f (x) = (2π) ei(x−y)·ξ q(x, ξ)f (y)dy (3.1.11)
ξ∈Zn

for a certain ‘symbol’ q(x, ξ). We will not pursue this any further, however, most of the following
theory can be developed parallel for this case. This can be seen as an (enlightening) exercise.

3.2 Hörmander symbols and their calculus

3.2.1. In order to define continuous operators P (x, D) on S(Rn ) via integrals of the form (3.1.8),
we have to make assumptions on their symbol p(x, ξ). There are several possibilities to do this,
the following one goes back to L. Hörmander.1
Definition. The symbol p(x, ξ) belongs to the symbol class S m (Rn ) if it is a smooth function,
p ∈ C ∞ (Rn × Rn ), and satisfies the symbol estimates

sup |∂xβ ∂ξα p(x, ξ)| ≤ Cα,β (1 + |ξ|)m−|α| (3.2.1)

x∈Rn

for all multi-indices α, β ∈ Nn0 . The number m ∈ R is called order of the symbol p.
3.2.2. Some properties of these symbol classes can be checked easily. First, we see that S m (Rn )
is a vector space over C. It can be given the structure of a Fréchet space if we use the optimal
constants Cα,β as seminorms (it is in fact easy to check that S m (Rn ) is complete in these norms).
Furthermore,
0
m ≤ m0 =⇒ S m (Rn ) ⊂ S m (Rn ) (3.2.2)
p ∈ S m1 (Rn ), q ∈ S m2 (Rn ) =⇒ pq ∈ S m1 +m2 (Rn ), (3.2.3)
p ∈ S m (Rn ) =⇒ ∂ β ∂ α p ∈ S m−|α| (Rn ), (3.2.4)
1 m
More generally, these are symbols of type 1,0 (written S1,0 (Rn )). For symbols of type ρ, δ one assumes an
estimate by
(1 + |ξ|)m−ρ|α|+δ|β| .
For 1 ≥ ρ > δ ≥ 0 most of the following considerations transfer.

22
 3.2 Hörmander symbols and their calculus

and multiplication and forming derivatives are continuous operations between these spaces. It
is also convenient to introduce
[ \
S ∞ (Rn ) = S m (Rn ), S −∞ (Rn ) = S m (Rn ). (3.2.5)
m∈R m∈R

For a given symbol a we define a corresponding operator.

3.2.3 Theorem. Let p ∈ S m (Rn ) be a symbol of order m. Then for any f ∈ S(Rn ) the integral
ZZ
−
op[p]f (x) = ei(x−y)·ξ p(x, ξ)f (y)dyd ξ (3.2.6)

is well-defined and defines a continuous operator op[p] : S(Rn ) → S(Rn ). We denote op[p] as
pseudo-differential operator with symbol p and define Ψm (Rn ) = op S m (Rn ).

Proof. Note first that the integral is well defined, since fb ∈ S(Rn ) and thus for all fixed x the
iterated integral exists in Lebesgue sense. It remains to estimate op[p]f . For this we introduce
the operator
1 1 + |x|2 ix·ξ
L(x, Dξ ) = (I − ∆ξ ), L(x, Dξ )eix·ξ = e = eix·ξ , (3.2.7)
1 + |x|2 1 + |x|2
and apply integration by parts N times in the integral
Z
−
op[p]f (x) = eix·ξ (Lt (x, Dξ ))N (p(x, ξ)fb(ξ))d ξ. (3.2.8)

Note that the symbol estimates imply that

k(Lt (x, Dξ ))N (p(x, ξ)fb(ξ))| ≤ (1 + |x|)−2N × some Schwartz function, (3.2.9)

such that
| op[p]f (x)| ≤ CN (1 + |x|)−2N , ∀N. (3.2.10)
The estimate for derivatives is analogous.

A detailed analysis of the previous proof shows that the operator convention op is continuous, if
a sequence of symbols converges in S m (Rn ) the corresponding sequence of operators converges
(we will stumble upon this later on again). Using Lebesgue’s theorem of dominated convergence
we can do with less strict assumptions (and get a variant of strong continuity):
3.2.4 Theorem. Assume pk ∈ S m (Rn ) is a uniformly bounded sequence in S m (Rn ) (i.e. all
elements of the sequence satisfy the symbol estimates (3.2.1) with the same constants) which
converges point-wise together with all of its derivatives to a symbol p ∈ S m (Rn ) and its deriva-
tives. Then op[pk ]f → op[p]f in S(Rn ) as k → ∞ for any f ∈ S(Rn ).
3.2.5. We always emphasised that the integral representation of op[p], p ∈ S m (Rn ) is an iterated
integral. There is an alternative approach to define the operator by a convergent double integral
in the following way: We take a cut-off function χ ∈ C0∞ (Rn ) which equals 1 near the origin
and consider the sequence of symbols p (x, ξ) = χ(ξ)p(x, ξ). Then clearly p (x, ξ) → p(x, ξ)
as  → 0 point-wise and the family {p }0<<1 is uniformly bounded in S m (Rn ). Thus, by the
previous theorem
ZZ
−
op[p]f = lim op[p ]f = lim ei(x−y)·ξ χ(ξ)p(x, ξ)f (y)dyd ξ (3.2.11)
→0 →0

23
3 Pseudo-differential operators

and the latter integral exists as double integral. Note, that this is exactly the trick we used to
prove the Fourier inversion formula in Section 2.2.
Since the double integral exists, we can apply Fubini theorem now and represent op[p ] by an
integral kernel
Z Z
−
op[p ]f = K (x, y)f (y)dy, K (x, y) = ei(x−y)·ξ χ(ξ)p(x, ξ)d ξ. (3.2.12)

The following Lemma gives some insight in these kernels.

3.2.6 Lemma. Let p ∈ S m (Rn ). Then for arbitrary cut-off χ and x 6= y the limit
K(x, y) = lim K (x, y) (3.2.13)
→0

exists, is independent of the choice of χ and defines a smooth function in x 6= y decaying rapidly
as |x − y| → ∞ and satisfying
β
|∂x,y K(x, y)| ≤ CN,β |x − y|−N , ∀N > m + n + |β|. (3.2.14)
In particular, K(x, y) ∈ C(R2n ) if m < −n and smooth if p ∈ S −∞ (Rn ).

Proof. We represent K (x, y) = R (x, x − y) as integral

Z
R (x, z) = eiz·ξ χ(ξ)p(x, ξ)d− ξ (3.2.15)

and derive estimates uniform in  for it. For this we use N integrations by parts with the
operator
1
L(z, Dξ ) = − 2 ∆ξ , L(z, Dξ )eiz·ξ = eiz·ξ , (3.2.16)
|z|
which give Z
−
R (x, z) = eiz·ξ (Lt (z, Dξ ))N p (x, ξ)d ξ. (3.2.17)

Because χ(x, ξ) ∈ S 0 (Rn ) is uniformly bounded in  ∈ (0, 1), p ∈ S m (Rn ) is uniformly
bounded and the symbol estimates imply
|R (x, z)| ≤ CN |z|−2N (3.2.18)
uniform in  as soon as the bound CN |z|−2N (1+|ξ|)m−2N becomes integrable, i.e. m−2N < −n.
Furthermore, Lebesgue’s theorem on dominated convergence allows to take the limit  → 0.
The estimates for derivatives are similar.

Now we come to the main theorems of this part, the symbolic calculus. Operators are deter-
mined by their symbols, so it is not surprising that we can express composition of operators in
terms of their symbols, etc. We will state theorems first and then prove them in one goal. At
first we need a preparatory Lemma.
3.2.7 Lemma. Let {pk }∞ k=0 be a sequence of symbols pk ∈ S
m−k
(Rn ). Then there exists a
symbol p ∈ S m (Rn ) such that for any number N ∈ N
N
X −1
p− pk ∈ S m−N (Rn ). (3.2.19)
k=0
−∞ n
The symbol p is unique
Pmodulo S (R ) and called an asymptotic sum of the sequence pk . We
use the notation p ∼ k pk .

24
 3.2 Hörmander symbols and their calculus

Proof. Let χ ∈ C0∞ (Rn ), χ(ξ) = 1 near ξ = 0 and |χ(ξ)| ≤ 1. We will show that there exists a
sequence k → 0 such that the sequence (1 − χ(k ξ))pk (x, ξ) satisfies

∂xβ ∂ξα (1 − χ(k ξ))pk (x, ξ) ≤ 2−k (1 + |ξ|)m−k+1−|α| , |α| + |β| ≤ k. (3.2.20)

If we then form the series

∞
X
p(x, ξ) = (1 − χ(k ξ))pk (x, ξ) (3.2.21)
k=0

we see that it is locally finite for all (x, ξ) and it satisfies

 N
X −1  ∞
X
∂xβ ∂ξα p(x, ξ) − (1 − χ(k ξ))pk (x, ξ) ≤ ∂xβ ∂ξα (1 − χ(k ξ))pk (x, ξ)
k=0 k=N
X∞
≤ Cα,β (1 + |ξ|)m−N −|α| + 2−k (1 + |ξ|)m−k+1−|α|
k=max(N +1,|α|+|β|)
0
≤ Cα,β (1 + |ξ|)m−N −|α| ,

the first (finitely many) terms estimated with the symbol estimate from S m−N (Rn ), the remain-
der with (3.2.20).
It remains to show that a sequence with (3.2.20) exists. For this consider q (x, ξ) = 1 − χ(ξ).
This sequence is bounded in S 0 (Rn ) (obviously) and converges to zero in S 1 (Rn ). For |α| > 0
this follows from

(1 + |ξ|)−1+|α| |∂ξα (1 − χ(ξ))| ≤ Cα sup (1 + |ξ|)−1+|α| |α| ≤ Cα0 (1 + |ξ|)−1 , (3.2.22)

ξ∈supp ∇χ

while the case |α| = 0 can be done by a small modification of the argument. This implies
(3.2.20), all q pk tend to zero in S m−k+1 (Rn ).

3.2.8 Theorem. 1. (Composition formula) Let p ∈ S m1 (Rn ) and q ∈ S m2 (Rn ). Then

there exists a symbol p]q ∈ S m1 +m2 (Rn ) such that op[p]q]f = op[p] op[q]f for all f ∈
S(Rn ). Moreover,
X 1
p]q(x, ξ) ∼ (Dαξ p(x, ξ))(∂xα q(x, ξ)). (3.2.23)
α
α!

2. (Transposed operator) Let p ∈ S m (Rn ). Then the operator op[p]t defined via
hop[p]φ, ψi = hφ, op[p]t ψi for all φ, ψ ∈ S(Rn ) is a pseudo-differential operator, i.e. there
exists a symbol pt ∈ S m (Rn ) with op[pt ] = op[p]t . Moreover,

X (−1)|α|
pt (x, ξ) ∼ (Dαξ ∂xα p)(x, −ξ) (3.2.24)
α
α!

3. (Adjoint operator) Let p ∈ S m (Rn ). Then the operator op[p]∗ defined via
∗
(op[p]φ, ψ)L2 = (φ, op[p] ψ)L2 for all φ, ψ ∈ S(Rn ) is a pseudo-differential operator, i.e.
there exists a symbol p∗ ∈ S m (Rn ) with op[p∗ ] = op[p]∗ . Moreover,
X 1
p∗ (x, ξ) ∼ Dαξ ∂xα p(x, ξ) (3.2.25)
α
α!

25
3 Pseudo-differential operators

Idea of the proof (composition formula). Let p ∈ S m1 (Rn ) and q ∈ S m2 (Rn ) and assume both
are given as limits of approximating sequences p and q as before2 . Then
Z ZZ
ix·η − −
op[p]q]f (x) = lim e p (x, η) eiy(ξ−η) q (y, ξ)fb(ξ)d ξdyd η
→0
Z
−
= eix·ξ p]q(x, ξ)fb(ξ)d ξ

with ZZ
p]q(x, ξ) = lim e−i(x−y)·(ξ−η) p(x, η)q(y, ξ)χ(y, η)d
−
ηdy. (3.2.26)
→0

Integrals of this type are usual referred to as oscillatory integrals and play a fundamental rôle in
the theory of pseudo-differential operators. The following (meta) theorem concludes the proof
(and similar for all other statements).

We denote by Am1 ,m2 (Rn × Rn ) the set of all amplitudes a(x, y, ξ, η) ∈ C ∞ (Rn × Rn × Rn × Rn )
subject to the conditions
0 0 0
|∂yβ ∂ηα ∂xβ ∂ξα a(x, y, ξ, η)| ≤ Cα,β,α0 ,β 0 (1 + |ξ|)m1 −|α| (1 + |η|)m2 −|α | (3.2.27)

for all multi-indices α, β, α0 , β 0 ∈ Nn .

3.2.9 Theorem (Asymptotic expansions for oscillatory integrals). Let a ∈ Am1 ,m2 (Rn ). Then
ZZ
p(x, ξ) = lim e−i(x−y)·(ξ−η) a(x, y, ξ, η)χ(y, η)dyd
−
η (3.2.28)
→0

is well-defined, the limit independent of χ ∈ C0∞ (R2n ), χ(s) = 1 for |s| ≤ 1, and belongs to
S m1 +m2 (Rn ) with
X 1
p(x, ξ) ∼ ∂yα Dαη a(x, y, ξ, η) . (3.2.29)
α
α! x=y,ξ=η

Proof. We assume for the moment that a(x, y, ξ, η) is compactly supported in y and η. Then
p(x, ξ) is well-defined as double integral and the basic idea will be that p(x, ξ) depends essentially
on the neighbourhood of x = y and ξ = η. This suggests the use of Taylor’s theorem and we
consider the integral
ZZ X
1 α
pN (x, ξ) = (∂ a(x, y, ξ, ξ))(η − ξ)α e−i(x−y)·(ξ−η) dyd
−
η. (3.2.30)
α! η
|α|≤N

Integration by parts allows to reformulate it as

Z Z X
1 α
pN (x, ξ) = ( (∂η a(x, y, ξ, ξ))(−Dy )α e−i(x−y)·(ξ−η) dy)d
−
η
α!
|α|≤N
Z Z X
1 α α
= ( (D ∂ a(x, y, ξ, ξ))e−i(x−y)·(ξ−η) dy)d −
η
α! y η
|α|≤N
X 1
= (Dα ∂ α a(x, y, ξ, ξ))e−i(x−y)ξ
α! y η x=y
|α|≤N

2
But for technical reasons now in all variables, i.e., p (x, ξ) = p(x, ξ)χ(x, ξ) with a smooth cut-off function
χ ∈ C0∞ (R2n ) identically equal to 1 near the origin.

26
 3.2 Hörmander symbols and their calculus

the last line due to Fourier inversion formula. This is exactly the N -term approximation
from (3.2.29). Due to our assumptions pN ∈ S m1 +m2 (Rn ) and it remains to show p − pN ∈
S m1 +m2 −N −1 . Using Taylor’s theorem (with integral remainder), we observe that
ZZ
p(x, ξ) − pN (x, ξ) = rN (x, y, ξ, η)e−i(x−y)·(ξ−η) dyd
−
η (3.2.31)

with
X 1
rN (x, y, ξ, η) = a(x, y, ξ, η) − (∂ α a(x, y, ξ, ξ))(η − ξ)α
α! η
|α|≤N

N +1 1
X Z
= (1 − θ)N (∂ηα a)(x, y, ξ, ξ + θ(η − ξ))dθ(η − ξ)α .
α! 0
|α|=N +1

We use integration by parts in (3.2.31). First we understand the (η − ξ)α terms as derivatives
Dαy applied to the amplitude function. Then we use

e−i(x−y)·(ξ−η) = (1 − ∆y )` e−i(x−y)·(ξ−η) (1 + |η − ξ|2 )−` (3.2.32)

(and the corresponding (1 − ∆η )` version) such that

X N +1Z Z 1 Z
2 −`1
p(x, ξ) − pN (x, ξ) = (1 + |η − ξ| ) (1 − θ)N
e−i(x−y)·(ξ−η)
α! 0
|α|=N +1

× (1 + |x − y|2 )−`2 (1 − ∆η )`2 (1 − ∆y )`1 Dαy (∂ηα a)(x, y, ξ, ξ + θ(η − ξ))dydθd

−
η.

We are going to estimate this integral entirely in terms of the symbol-estimates of a independent
of our support assumption from the beginning. Chosing `2 > n/2 the y-integral exists as
Lebesgue integral and gives
Z 1 Z
|p(x, ξ) − pN (x, ξ)| ≤ CN (1 + |η − ξ|2 )−`1 (1 + |ξ|)m1 (1 + |ξ + θ(η − ξ)|)m2 −N −1 dηdθ.
0

Here we estimated roughly (1 − θ)N θ2`2 (1 + |ξ + θ(η − ξ)|)−2`2 ≤ 1. Choosing now `1 big enough
we estimate the remaining integrals. For this we use the following Peetre type inequality3 ,
 s
1 + |ξ|
≤ (1 + |ξ − η|)|s| , ξ, η ∈ Rn , s ∈ R (3.2.33)
1 + |η|
to get

(1 + |ξ + θ(η − ξ)|)m2 −N −1 ≤ (1 + |ξ|)m2 −N −1 (1 + θ|η − ξ|)|m2 −N −1|

≤ (1 + |ξ|)m2 −N −1 (1 + |η − ξ|)|m2 −N −1| .

Choosing 2`1 > n + |m2 − N − 1| implies convergence of the integral and gives the desired
estimate |p(x, ξ) − pN (x, ξ)| ≤ C(1 + |ξ|)m1 +m2 −N −1 . Derivatives are handled similarly and this
proves p − pN ∈ S m1 +m2 −N −1 .
For the general situation we have to use the cut-off function χ in (3.2.28) to make the inte-
gral convergent. For fixed  ∈ (0, 1] we can use the same arguments as above, and observe
3
In fact, this one directly follows from triangle inequality, | |ξ| − |η| | ≤ |ξ − η|. The original Peetre inequality
uses squares inside and reads (1 + |ξ|2 )s ≤ 2|s| (1 + |η|2 )s (1 + |ξ − η|2 )|s| .

27
3 Pseudo-differential operators

that all estimates depend entirely on the constants in symbol estimates for a (x, y, ξ, η) =
a(x, y, ξ, η)χ(y, η), which are in uniform in . This implies that the estimates are uniform in 
and Lebesgues dominated convergence theorem applied after all the integrations by parts have
been done proves the existence of the limit(s). It also proves uniqueness (how?). Furthermore,
the obtained asymptotic expansion is uniform in  ∈ (0, 1] and the statement is proven.

Proof of Theorem 3.2.8. We collect the remaining bits of the proof of Theorem 3.2.8. For the
composition we already know that p]q(x, ξ) is given by an oscillatory integral and the previous
theorem gives the corresponding asymptotic expansion.
For the transposed operator we observe that
ZZZ
−
hop[p]φ, ψi = ei(x−y)·ξ p(x, ξ)φ(y)dyd ξψ(x)dx
Z ZZ
−
= φ(y) ei(x−y)·ξ p(x, ξ)ψ(x)dxd ξdy

such that (substituting ξ by −ξ)

ZZ ZZ ZZ
t i(x−z)·ξ i(z−y)·η − − −
op[p] ψ(x) = e p(z, −ξ) e ψ(y)dyd ηd ξdz = ei(x−y)·η pt (x, η)ψ(y)dyd η

holds true with ZZ

t −
p (x, η) = ei(x−z)·(ξ−η) p(z, −ξ)dzd ξ. (3.2.34)

Again the previous theorem gives the desired expansion. For the adjoint operator one proceeds
similarly.

3.2.10. We want to draw some notable consequences from the calculus theorem.

Corollary. Let p ∈ S m1 (Rn ) and q ∈ S m2 (Rn ).

1. The Leibniz product p]q ∈ S m1 +m2 (Rn ) is given modulo symbols from S m1 +m2 −1 (Rn ) by
the ordinary (pointwise) product pq.
2. Let [p, q] = p]q − q]p such that op[[p, q]] = [op[p], op[q]] = op[p] ◦ op[q] − op[q] ◦ op[p] is
the commutator of the corresponding operators. Then [p, q] ∈ S m1 +m2 −1 (Rn ) and modulo
symbols from S m1 +m2 −2 (Rn ) the symbol [p, q] is given by the Poisson bracket
n
X
{p, q}(x, ξ) = −i (∂ξj p(x, ξ))(∂xj q(x, ξ)) − (∂ξj q(x, ξ))(∂xj p(x, ξ)). (3.2.35)
j=1

3.3 Mapping properties of pseudo-differential operators

3.3.1. In a first step we want to apply pseudo-differential operators to distributions. Since
pseudo-differential operators map Schwartz space into itself, we can extend them via transpo-
sition to tempered distributions.

Definition. Let u ∈ S 0 (Rn ) and p ∈ S m (Rn ). Then we define op[p]u via

hop[p]u, φi = hu, op[pt ]φi (3.3.1)

for all test functions φ ∈ S(Rn ).

28
 3.3 Mapping properties of pseudo-differential operators

Pseudo-differential operators act continuously S 0 (Rn ) → S 0 (Rn ) and the operator convention
op : S m (Rn ) → L(S 0 (Rn )) is continuous. However, in most applications this statement is too
weak and one asks for boundedness of operators between Sobolev spaces.

3.3.2 Theorem. Let p ∈ S 0 (Rn ). Then op[p] : L2 (Rn ) → L2 (Rn ) acts as continuous operator
and op : S 0 (Rn ) → L(L2 (Rn )) is continuous.

Proof. It is sufficient to prove that for all functions φ ∈ S(Rn ) the estimate k op[p]φk2 ≤ Ckφk2
where the constants depend only on finitely many symbol seminorms of p ∈ S 0 (Rn ).
In a first step we assume that p is compactly supported in x. Then we can apply a Fourier
transform in the x-variable, Z
pb(λ, ξ) = e−ix·λ p(x, ξ)dx, (3.3.2)

which gives a Schwartz function p(·, ξ) ∈ S(Rn ) uniformly in ξ. To see this uniformity we notice
that Z
Dβλ λα pb(λ, ξ) = e−ix·λ (−iλ)β Dαx p(x, ξ)dx (3.3.3)

p(λ, ξ)| ≤ CN (1 + |λ|)−N

can be estimated uniform in ξ due to p ∈ S 0 (Rn ). In particular, supξ |b
for all numbers N . Using
ZZ
op[p]φ(x) = b − ξd
eix·(ξ+λ) pb(λ, ξ)φ(ξ)d −
λ
Z
−
= Sφ(λ, x)d λ

with Sφ(λ, x) = eix·λ op[b

p(λ, ·)]φ. The latter can be estimated using Plancherel theorem

kSφ(λ, ·)k2 = k op[b

p(λ, ·)]φk2 = kb
p(λ, ·)φk
b 2 ≤ kb
p(λ, ·)k∞ kφk2 . (3.3.4)

Hence Z
k op[p]φk ≤ C (1 + |λ|)−N dλkφk2 ≤ C 0 kφk2 . (3.3.5)

The appearing constants depend on the symbol estimates for p and the size of the support in
x (but not its position). To go over to general symbols we use this fact in order to estimate
op[p]φ(x) near a particular x = x0 and consider only χ(x − x0 ) op[p]φ(x) for some fixed χ ∈
C0∞ (Rn ) supported in a ball of radius 1 (and equal to 1 near the origin). We prove

|φ(x)|2
Z Z
2
|χ(x − x0 ) op[p]φ(x)| dx ≤ CN (3.3.6)
(1 + |x − x0 |)N

for any x0 and arbitrary N . Integrating this in x0 and using Fubini implies
Z Z
kχk22 k op[p]φk22 = |χ(x − x0 ) op[p]φ(x)|2 dxdx0
Z Z
2
≤ CN |φ(x)| (1 + |x − x0 |)N dx0 dx = CN kφk22 , (3.3.7)

which is the desired statement. To prove (3.3.6) we split φ into two parts, φ = φ1 + φ2 , φ2
supported outside a ball of radius 2 around x0 and φ1 having compact support within a ball of

29
3 Pseudo-differential operators

radius 3 around x0 . Furthermore, we require |φj | ≤ |φ|. For φ1 estimate (3.3.6) follows from
the first part of the proof,
|φ(x)|2
Z Z Z
2 2
|χ(x − x0 ) op[p]φ1 (x)| dx ≤ C |φ1 (x)| dx ≤ CN . (3.3.8)
|x−x0 |≤3 (1 + |x − x0 |)N

For φ2 we use integration by parts,

ZZ
−
χ(x − x0 ) op[p]φ2 (x) = ei(x−y)·ξ χ(x − x0 )p(x, ξ)φ2 (y)dyd ξ
χ(x − x0 )φ2 (y) −
ZZ
= ei(x−y)·ξ ∆`ξ p(x, ξ) dyd ξ,
|x − y|2`

such that by Cauchy-Schwarz inequality

|φ(y)|2
Z Z
2 dy
|χ(x − x0 ) op[p]φ2 (x)| ≤ C dy
1 + |x − y|2` 1 + |x − y|2`
holds for all sufficiently large `. Integrating in x gives the desired estimate (3.3.6) and the proof
is complete.
3.3.3 Corollary. Let p ∈ S m (Rn ). Then op[p] : H s (Rn ) → H s−m (Rn ) as bounded linear
operator for all s ∈ R.

Proof. The isomorphism (1 − ∆)s/2 : H s (Rn ) → L2 (Rn ) is a pseudo-differential operator of

class Ψs . The previous theorem combined with the calculus of pseudo-differential operators
concludes the proof.

Without proof we mention the following result. Again the case m = 0 implies the general
statement.
3.3.4 Theorem. Let p ∈ S m (Rn ). Then for all p ∈ (1, ∞) and all s ∈ R the pseudo-differential
operator op[p] : W s,p (Rn ) → W s−m,p (Rn ) is continuous.
3.3.5. We conclude this section with an application of the calculus to prove estimates for
pseudo-differential operators. To make matters as simple as possible we consider only classical
operators, i.e., the class Sclm (Rn ) ⊆ S m (Rn ) of symbols a(x, ξ) with asymptotic expansions into
homogeneous components,
∞
X
a(x, ξ) ∼ χ(ξ)a(m−j) (x, ω)|ξ|m−j , (3.3.9)
j=0

with functions a(k) (x, ω) ∈ C ∞ (Rn × Sn−1 ), ω = ξ/|ξ| and a cut-off function χ ∈ C ∞ (Rn ),
χ(ξ) = 1 for |ξ| ≥ 1 and χ(ξ) = 0 for |ξ| ≤ 1/2. Particular examples are all differential
operators.
The advantage of classical pseudo-differential operators is that we have to care (up to S −∞ (Rn ))
only about the homogeneous components. Furthermore, they fit to the calculus expansions.
Leibniz products of classical symbols are classical again.
3.3.6 Theorem (Gårding inequality, 1D). Assume p ∈ Scl2 (R) is real-valued and p(x, ξ) ≥ c|ξ|2 ,
|ξ|  1. Then there exist constants C1 and C2 such that for all u ∈ H 2 (R)

Re (op[p]u, u)L2 ≥ C1 kuk2H 1 − C2 kuk2L2 . (3.3.10)

30
 3.4 Ellipticity and parametrices

Proof. We first note that it is enough to show that p + p∗ = q ∗ ]q mod S 0 for some q ∈ Scl1 (R).
Indeed looking at the corresponding operators P = op[p] and Q = op[q] we get P +P ∗ = Q∗ Q+R
with some R ∈ Ψ0 (Rn ). Hence

2Re (P u, u)L2 = ((P + P ∗ )u, u)L2 = kQuk22 + (Ru, u)L2 (3.3.11)

and therefore by |(Ru, u)L2 | ≤ C2 kuk22 we get

Re (P u, u)L2 ≥ kQuk22 − C2 kuk22 . (3.3.12)

Combined with an estimate of the form kQuk2 ≥ CkukH 1 − C 0 kuk2 the statement follows.
So it is enough to understand how to construct q. Furthermore it is enough to use the 2-
and 1-homogeneous components of p to construct the 1- and 0- homogeneous components of q.
Based on

p(x, ξ) ∼ p± 2 ±
2 (x)ξ + p1 (x)ξ mod S 0 , p±
i (x) ∈ R (3.3.13)
∗ ± ± ±
p (x, ξ) ∼ p2 (x)ξ + p1 (x)ξ − 2i∂x p2 (x)ξ mod S 0
2
(3.3.14)
q(x, ξ) ∼ q1± (x)ξ + q0± (x) mod S −1 (3.3.15)
q ∗ (x, ξ) ∼ q1± (x)ξ + q0± (x) − i∂x q1± (x) mod S −1 (3.3.16)

as ξ → ±∞ and

q ∗ ]q(x, ξ) = |q(x, ξ)|2 − iq(x, ξ)∂ξ ∂x q(x, ξ) − i(∂ξ q(x, ξ))∂x q(x, ξ) mod S 0
∼ |q1± (x)|2 ξ 2 + 2Re (q1± (x)q0± (x))ξ − 2iRe (∂x q1± (x)q1± (x))ξ mod S 0
∼ |q1± (x)|2 ξ 2 + 2Re (q1± (x)q0± (x))ξ − i∂x |q1± (x)|2 ξ mod S 0 (3.3.17)

a comparison of coefficients yields |q1± (x)|2 = 2p±

2 (x)
p and Re(q1± (x)q0± (x)) = p±
1 (x).
p Therefore,
we can simply take real-valued symbols q1 (x) = 2p2 (x) and q0 (x) = p1 (x)/ 2p±
± ± ± ±
2 (x) and
the desired operators are constructed.
The constructed operator Q is invertible modulo Ψ−1 , indeed choosing q̃(x, ξ) ∼ (p±
2 (x))
−1/2 −1
ξ
mod S −2 as inverse we get q̃]q = 1 mod S −1 . Therefore, we obtain operators Q̃, R̃ ∈ Ψ−1
with I = Q̃Q − R̃ such that kukH 1 = kQ̃Qu − R̃ukH 1 ≤ CkQuk2 + C 0 kuk2 , i.e., as desired
kQuk2 ≥ C −1 kukH 1 − C −1 C 0 kuk2 .

3.4 Ellipticity and parametrices

3.4.1. An operator A ∈ Ψ0 (Rn ) is invertible (with inverse in Ψ0 (Rn )) if we can find a Leibniz
inverse of the corresponding symbol, i.e. if A = op[a] we are looking for b ∈ S 0 (Rn ) with
a]b = 1. In general this might be too much to ask, however based on the symbolic calculus we
can try to construct b with a]b = 1 mod S −∞ (Rn ).
Why is this of interest? Consider the following partial differential equation
n
X
Au = aij (x)∂i ∂j u(x) = f (x) (3.4.1)
i,j=1

with f (x) ∈ H s (Rn ). Then A ∈ Ψ2 (Rn ) and assume there exists B ∈ Ψ−2 (Rn ) such that
B ◦ A = I + R with R ∈ Ψ−∞ (Rn ). Assume now that we have a solution u ∈ H −∞ (Rn ) =

31
3 Pseudo-differential operators

H s (Rn ) to the above problem, i.e., the equation is satisfied. Applying B on both sides of
S
s
the equation implies
BAu = u + Ru = Bf. (3.4.2)
Since R ∈ Ψ−∞ (Rn ), we know that Ru ∈ H ∞ (Rn ) = H s (Rn ) and from Bf ∈ H s+2 (Rn ) we
T
immediately conclude u ∈ H s+2 (Rn ). Hence, we obtain a regularity statement for solutions.

3.4.2 Definition. Let A ∈ Ψm (Rn ). An operator B ∈ Ψ∞ (Rn ) = s Ψs (Rn ) is called a left-

S
parametrix to A if B ◦ A − I ∈ Ψ−∞ (Rn ). Similarly, B is called right-parametrix if A ◦ B − I ∈
Ψ−∞ (Rn ). An operator which is left- and right-parametrix is simply called parametrix .

3.4.3 Lemma. 1. Assume A has a left-parametrix B and B 0 = B + R for some R ∈

Ψ (R ). Then B 0 is also a left-parametrix.
−∞ n

2. Assume A has a left-parametrix B1 and a right-parametrix B2 . Then B1 − B2 ∈ Ψ−∞ . In

particular, A has a parametrix and each left-parametrix is also right-parametrix.

Proof. The first part is evident. For the second one we use that B1 ◦A−I = R1 , A◦B2 −I = R2
for some smoothing operators R1 , R2 ∈ Ψ−∞ (Rn ). Therefore

B1 = B1 ◦ (A ◦ B2 − R2 ) = B1 ◦ A ◦ B2 − B1 ◦ R2 = B2 + R1 ◦ B2 − B1 ◦ R2 , (3.4.3)

where R1 ◦ B2 and B1 ◦ R2 are smoothing due to our calculus rules.

3.4.4. For the following we assume that A ∈ Ψm n m n

cl (R ) with symbol a ∈ Scl (R ) is classical.
Then the topmost homogeneous component a(m) ∈ C ∞ (Rn ×Sn−1 ) is called its principal symbol.
We call A elliptic if
|a(m) (x, ω)| ≥ c > 0 (3.4.4)
is non-vanishing.
−m
3.4.5 Theorem. Let A ∈ Ψm n n
cl (R ) be elliptic. Then there exists a left-parametrix B ∈ Ψcl (R )
to A.

Proof. It is sufficient to construct all homogeneous components b(−m−j) (x, ξ) =

|ξ|−m−j b(−m−j) (x, ω) in terms of the components a(m−k) (x, ξ) and this will be done using the
composition formula. Note for this that derivatives of homogeneous components are again
homogeneous. Hence
X 1
1∼ (Dαξ b(x, ξ))(∂xα a(x, ξ)) (3.4.5)
α
α!

implies

1 = b(−m) (x, ξ)a(m) (x, ξ) (3.4.6)

X
0 = b(−m−1) (x, ξ)a(m) (x, ξ) + b(−m) (x, ξ)a(m−1) (x, ξ) + (Dαξ b(m) (x, ξ))(∂xα a(m) (x, ξ)) (3.4.7)
|α|=1

0 = b(−m−2) (x, ξ)a(m) (x, ξ) + some terms containing only b(−m) and b(−m−1) (3.4.8)

etc. This implies that

1
b(−m) (x, ξ) = (3.4.9)
a(m) (x, ξ)

32
 3.4 Ellipticity and parametrices

(which is smooth on Rn × Sn−1 due to the ellipticity assumption) and the following equations
determine
1
b(−m−j) (x, ξ) = × some terms containing only b(−m) up to b(−m−j+1) (3.4.10)
a(m) (x, ξ)
This procedure constructs all homogeneous components and by constructing b as asymptotic
sum we obtain the required left-parametrix.
3.4.6. We leave it as an exercise to fill in the gaps in the previous proof. In particular it
can be shown that under the same assumption also a right-parametrix exists and therefore the
constructed operator B is indeed a parametrix.
3.4.7 Example. The second order partial differential operator
X n
A= aij (x)∂i ∂j , aij (x) = aji (x) ∈ R (3.4.11)
i,j=1

from the introduction to this section has a parametrix B ∈ Ψ−2 n

cl (R ) if we require that its
(homogeneous!) symbol
n
X
a(x, ξ) = − aij (x)ξi ξj (3.4.12)
i,j=1

satisfies the ellipticity assumption

|a(x, ξ)| ≥ C|ξ|2 (3.4.13)
for a suitable constant C, which is equivalent to assume that the matrix (aij (x))ij is either
positive or negative definite uniform in x.
3.4.8 Example. Assume (aij (x))ij is negative definite uniformly in x. Then the operator
n
X n
X
A= aij (x)∂i ∂j + bj (x)∂j + c(x) (3.4.14)
i,j=1 j=1

has a parametrix in Ψ−2 (Rn ) for any choice of lower order terms bj (x), c(x) ∈ B ∞ (Rn ).
3.4.9. The parametrix constructions can be localised. By this we mean that we want to solve
Au = f around a given point x0 and use for this a function χ ∈ C0∞ (Rn ) identically equal to 1.
near x0 . Hence, we require that χ(x)B ◦ A = χ(x)I modulo smoothing operators.
It can also be micro-localised. By this we mean that for given (x0 , ξ0 ) ∈ R2n we find a smooth
cut-off function ψx0 ,ξ0 (x, ξ) ∈ S 0 (Rn ) equal to one in a conic neighbourhood Cx0 ,ξ0 of (x0 , ξ0 ),
i.e.,
(x, ξ) ∈ Cx0 ,ξ0 =⇒ (x, λξ) ∈ Cx0 ,ξ0 , ∀λ > 0, (3.4.15)
such that op[ψx0 ,ξ0 ] ◦ B ◦ A = op[ψx0 ,ξ0 ] modulo smoothing operators. It turns out that such
micro-local parametrices exist at (x0 , ξ0 ) ∈ R2n (i.e. we find Cx0 ,ξ0 , a suitable ψx0 ,ξ0 and B), if
the principal symbol of A satisfies a(m) (x0 , ξ0 ) 6= 0.
This allows the detailed study of singularities of solutions to partial differential equations, e.g.
one can show that
W F (u) = R2n \ {(x0 , ξ0 ) | ∃ψx0 ,ξ0 : op[ψx0 ,ξ0 ]u ∈ C ∞ (Rn )} (3.4.16)
for u ∈ S 0 (Rn ) satisfies
W F (Au) ⊆ W F (u) ⊆ W F (Au) ∪ Char(A) (3.4.17)
with Char(A) = {(x0 , ξ0 )|a(m) (x0 , ξ0 ) = 0} for all A ∈ Ψm n
cl (R ).

33