U.U.D.M.

Project Report 2021:38

Pseudo-differential operators and

their properties
David Torstadius

Examensarbete i matematik, 15 hp
Handledare: Anders Israelsson
Examinator: Martin Herschend
Juli 2021

Department of Mathematics
Uppsala University
 Pseudo-differential operators and their properties
David Torstadius
June 2021

Abstract
Pseudo-differential operators are a kind of generalization of differential operators with ap-
plications in PDE:s. The purpose of this thesis is to introduce pseudo-differential operators
and prove some results about their properties. One of the main results is that under certain
conditions, pseudo-differential operators have a so called approximate inverse. At the end of
the thesis we use this result along with some other theorems to prove some results about the
solutions to elliptic PDE:s.

1
 Acknowledgements
I express my deepest gratitude to my supervisor, Anders Israelsson, for introducing me to
the theory pseudo-differential operators and all the discussions and corrections that helped me
during the writing of this thesis.

2
Contents
1 Introduction 4

2 Notation and terminology 5

3 Prerequisites 6

4 Littlewood-Paley partition of unity 11

5 The Schwartz space and symbols 14

5.1 Schwartz space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
5.2 Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.3 Elliptic symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

6 Fourier analysis 25

7 Tempered distributions 26

8 Pseudo-differential operators 28
8.1 Definition of Pseudo-differential operators . . . . . . . . . . . . . . . . . . . . . . . . 28
8.2 Composition of pseudo-differential operators . . . . . . . . . . . . . . . . . . . . . . . 30
8.3 Elliptic pseudo-differential operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
8.4 Pseudo-differential operators on Lp (Rn ) . . . . . . . . . . . . . . . . . . . . . . . . . 49

9 Sobolev spaces 56

10 Applications to elliptic pseudo-differential equations 60

3
1 Introduction
A partial differential operator is a sum of partial derivatives multiplied by some function. For
∂2 ∂
example, x ∂x 2 + ∂y is a partial differential operator. Pseudo-differential operators are a kind of

generalization of partial differential operators, and they have applications in PDE:s. In this thesis,
pseudo-differential operators are introduced and some important results about them are proven.
A particular kind of pseudo-differential operators called elliptic will be introduced. An interesting
property about these operators is that we can find an approximate inverse to them. This is used
to prove some results about PDE:s of the form P (x, D)u = f where P (x, D) is an elliptic partial
differential operator.
The majority of this work is based on Chapters 1-12 and 15 and the exercises in An introduction
to pseudo-differential operators by M.W. Wong [6]. In that book, some theorems about pseudo-
differential operators are proven but the author leaves some lemmas and results in the theorems
as exercises. The purpose of this text is to prove these theorems but to fill in the details and the
propositions that were excluded from [6]. However some proofs in [6] do not exclude any details so
these are omitted from this thesis. There are also some theorems that require more advanced theory
that are excluded in this thesis. The exercises in [6] also give ideas to apply the theorems to prove
more results. In particular, the last chapter in this thesis is based on exercises 12.8, 12.9, 15.1 and
15.2 in [6].
The reader is assumed to have knowledge in multivariate calculus and Fourier analysis. But
some well-known theorems from Fourier analysis are included for completeness. There are some
terminology, like measurable functions and measure zero, that are in this text, but most of the
material can be understood without it.
This thesis is structured as follows: In Chapter 2, basic notation and terminology is stated and in
Chapter 3, we state some theorems without proof and prove some basic results that we will use but
does not require any new theory. Chapter 4 is devoted to the construction of some smooth functions
and the Littlewood-Paley partition of unity. In Chapter 5, we introduce some important function
spaces such as the Schwartz space. Chapter 6 and 7 is a brief overview of definitions and theorems
about Fourier analysis and tempered distributions respectively. Pseudo-differential operators, the
main topic of this thesis, are finally introduced in Chapter 8 and in each subsection we prove an
important theorem about pseudo-differential operators. In Chapter 9, we introduce Sobolev spaces
and in Chapter 10 we apply the theory from previous chapters to PDE:s.

4
2 Notation and terminology
In this section, we state some basic notation and terminology so that there is no ambiguity in the
text.
• The natural numbers N include 0.
• C ∞ (Rn ) denotes the set of smooth functions Rn → C and C0∞ (Rn ) denotes the set of smooth
functions with compact support.

• If we have a function f : Rn × Rm → C with values denoted f (x, y), for x ∈ Rn and y ∈ Rm ,

then we say that f has compact support in x if the set

S = {x ∈ Rn | ∃y ∈ Rm : (x, y) ∈ supp(f )}

is compact in Rn . Then we say that the x-support of f is S.

• For a function f : Rn × Rm → C, f (x, ·) is the function Rm → C obtained by keeping x ∈ Rn
fixed and y ∈ Rm as a variable.
• For f ∈ C 1 (Rn ), and 1 ≤ k ≤ n, let ∂k be the k:th partial derivative, i.e

∂f
∂k f = .
∂xk

• Occasionally the operator Dk = −i∂k will be used instead to simplify notation.

• To simplify higher order derivatives and multivariate polynomials, multi-indices are used. A
multi-index is an element α ∈ Nn . For any x ∈ Rn , define

xα = xα1 α2 αn
1 x2 · · · xn ,

where we use the convention 00 = 1 if αk = 0 and xk = 0. For derivatives, define

∂ α = ∂1α1 ∂2α2 · · · ∂nαn

Dα = D1α1 D2α2 · · · Dnαn ,

where ∂k0 is treated as the identity operator.

• The absolute value of a multi-index α is defined as the sum of its components, i.e

|α| = α1 + · · · + αn .

We occasionally refer to |α| as the length or size of α.

• We say that α ≤ β if αk ≤ βk for all 1 ≤ k ≤ n.

• Define the factorial α! as
α! = α1 ! · · · αn !
and the binomial coefficient      
α α1 αn
= ··· .
β β1 βn

5
3 Prerequisites
In this section we state some definitions and theorems and prove some results that will be used in
later chapters. We begin by stating the definition of Lp (Rn ) and some properties.
Definition 3.1. For any 1 ≤ p < ∞, define Lp (Rn ) as the set of measurable functions f : Rn → C
such that Z
|f (x)|p dx < ∞.
Rn

The norm on Lp (Rn ), denoted k · kp is defined as

Z  p1
kf kp = |f (x)|p dx .
Rn

One can prove that Lp (Rn ) is a vector space. However the Lp -norm as defined above does not
satisfy the axioms of a norm because there are nonzero functions whose Lp -norm is zero, those that
are nonzero on a set of measure zero. To solve this problem, one can think of two functions f and g
as being equal if f = g almost everywhere, which is equivalent to kf − gkp = 0. More precisely, let
V be the space of functions f ∈ Lp (Rn ) such that kf kp = 0. Then V is a subspace of Lp (Rn ) so we
can take the quotient Lp (Rn )/V . Then f + V = g + V if and only if f = g almost everywhere, and
we can define a norm on Lp (Rn )/V by kf + V k = kf kp which now satisfies the axioms of a norm.
One can also prove that Lp (Rn ) is complete. Thus it is a Banach space, if we identify it by
L (Rn )/V as explained above. This fact will be important in the forthcoming chapters.
p

The proof of this theorem can be found in [4, Theorem 3.4, p. 69-70] by W. Rudin.
Theorem 3.2. C0∞ (Rn ) is dense in Lp (Rn ).
Hölder’s inequality and Minkowski’s integral inequality are two inequalities involving Lp -norms
that will be used in multiple proofs in this text.
1 1
Theorem 3.3 (Hölder’s inequality). Assume p, q > 1 and p + q = 1. Then for any f ∈ Lp (Rn )
and g ∈ Lq (Rn ),
kf gk1 ≤ kf kp kgkq .
Theorem 3.4 (Minkowski’s integral inequality). Assume p ≥ 1 and f : Rn ×Rn → C is a measurable
function. Then
Z Z p 1/p Z Z 1/p
f (x, y) dy dx ≤ |f (x, y)|p dx dy.
Rn Rn Rn Rn

In normed vector spaces, there are a certain kind of linear operators called bounded.

Definition 3.5. Let X and Y be normed vector spaces. A linear operator T : X → Y is said to be
bounded if there is a constant C ≥ 0 such that

kT xk ≤ Ckxk, ∀x ∈ X.

Remark. Note that this use of the term bounded will only used when the function in question is a
linear operator. For general functions such as f : Rn → C, we say that f is bounded if there is a
constant M ≥ 0 such that
|f (x)| ≤ M, ∀x ∈ Rn
The following theorem is available in [2, Theorem 2.7-11, p. 100] by E. Kreyszig. It will be used
in Chapter 8.4 and 9.

6
Theorem 3.6. Let X and Y be Banach spaces and let Z ⊆ X be a dense subspace of X (with norm
inherited from X). If T : Z → Y is a bounded linear operator, then T can be extended to a bounded
linear operator T̃ : X → Y such that T̃ x = T x for all x ∈ Z.
This theorem follows from a theorem which can be found in [3, Theorem 7.17, p. 152], and allows
us to interchange limit and derivative.
Theorem 3.7. Suppose (fk ) is a sequence of functions in C ∞ (Rn ). If (∂ α fk ) is a uniformly
convergent sequence for all multi-indices α, then the limit f = limk→∞ fk is in C ∞ (Rn ) and ∂ α f =
limk→∞ ∂ α fk for all multi-indices α.
The proof of the following theorem can be found in [6, Theorem 2.1, p. 5], which allows us to
interchange derivatives and integrals.
Theorem 3.8. Let f : Rn × Rm → C be a measurable function. Assume that f (·, y) ∈ C ∞ (Rn ) for
all y ∈ Rm Also assume that for all multi-indices α,
Z
sup |∂xα f (x, y)| dy < ∞.
x∈Rn Rn

f (x, y) dy is in C ∞ (Rn ) and for all multi-indices α

R
Then the function of x defined by Rn
Z Z
∂α f (x, y) dy = ∂xα f (x, y) dy.
Rn Rn

The proof of this version Taylor’s formula can be found in [6, Theorem 7.3, p. 50].
Theorem 3.9 (Taylor’s formula). Suppose f ∈ C ∞ (Rn ) and k is a positive integer. Then
X ∂ α f (x) X hα Z 1
α
f (x + h) = h +k (1 − θ)k−1 (∂ α f )(x + θh) dθ,
α! α! 0
|α|<k |α|=k

for any x, h ∈ Rn .
The following is a generalization of Leibniz’s formula for multivariate functions.
Theorem 3.10 (Leibniz’s formula). Suppose f, g ∈ C ∞ (Rn ). Then for all multi-indices α,
X α 
α
∂ (f g) = ∂ β f ∂ α−β g.
β
β≤α

Now we prove some equalities and inequalities that will be useful. Proposition 3.11 will be used
frequently in the text and will not be referenced when used.
Proposition 3.11. For all multi-indices α and x ∈ Rn .

|xα | ≤ |x||α| .

Proof. Since |xi | ≤ |x| for all i = 1, · · · , n, we have

|xα | = |xα αn
1 · · · xn | = |x1 |
1 α1
· · · |xn |αn ≤ |x|α1 · · · |x|αn = |x|α1 +···+αn = |x||α| .

Proposition 3.12. Let k ∈ R. Then there are positive constants C1 and C2 such that

C1 (1 + |ξ|2 )k/2 ≤ (1 + |ξ|)k ≤ C2 (1 + |ξ|2 )k/2 , ∀ξ ∈ Rn .

7
Proof. For any x ≥ 0,
1 + x2 ≤ 1 + 2x + x2 = (1 + x)2 (1)
and
(1 + x)2 ≤ (1 + x)2 + (1 − x)2 = 2(1 + x2 ). (2)
Replacing x with |ξ| in (1) and (2) implies that

1 + |ξ|2 ≤ (1 + |ξ|)2 ≤ 2(1 + |ξ|2 ). (3)

If k ≥ 0, then if we raise (3) to the power of k2 and let C1 = 1 and C2 = 2k/2 we see that the
proposition holds. If k < 0, then the inequalities in (3) change direction when raising to the power
of k2 , so then we take C1 = 2k/2 and C2 = 1 instead.
The following proposition highlights a common method for proofs involving multi-indices which
is to use induction on the length of the multi-index.
Proposition 3.13. For any two multi-indices α and β,
(
β! α

α−β
β α β x , β≤α
∂ x = (4)
0, otherwise.

Proof. Fix α and use induction on |β|. The base case β = 0 clearly holds. For the induction step,
assume β > 0 and that (4) holds for all multi-indices of size less than β. β > 0 implies that βi ≥ 1
for some i ∈ {1, . . . , n}. Let δ be the multi-index with a 1 at position i and 0 at all other positions.
Then δ ≤ β, so we can let γ = β − δ, and then we see that β = γ + δ and γ < β. The induction
assumption now holds for γ. So now we want to show that (4) holds for β. There are three cases:
Case 1: Suppose that γ  α and β  α. Then ∂ γ xα =0 by the induction assumption, so

∂ β xα = ∂ γ+δ xα = ∂i ∂ γ xα = ∂i 0 = 0,

which shows that (4) holds for β.

Case 2: Suppose that γ ≤ α and β  α. The latter implies that βk > αk for some k. But β = γ + δ,
so βj = γj ≤ αj for all j 6= i. Therefore, the only possible option is that k = i. Since βi = γi + 1,
we see that
γi ≤ αi < βi = γi + 1 =⇒ γi = αi . (5)
By the induction assumption,  
γ α α α−γ
∂ x = γ! x . (6)
γ
αi −γi = 0 by (5) so xα−γ is independent of xi . Since ∂ β = ∂i ∂ γ , we then have by (6) that ∂ β xα = 0,
so (4) holds for β.
Case 3: Suppose that γ ≤ α and β ≤ α. Since βi = γi + 1, we then see that αi − γi > 0. Therefore,

∂i (xα−γ ) = ∂i (xα1 −γ1 · · · xαi −γi · · · xαn −γn ) = (αi − γi )xα1 −γ1 · · · xαi −γi −1 · · · xαn −γn
= (αi − γi )xα1 −β1 · · · xαi −βi · · · xαn −βn = (αi − γi )xα−β .

Then by the induction assumption

   
α α
∂ β xα = ∂i ∂ γ xα = γ! ∂i (xα−γ ) = γ! (αi − γi )xα−β . (7)
γ γ
Notice that
   
αi αi ! αi ! αi ! αi
γi ! (αi − γi ) = (αi − γi ) = = = βi ! .
γi (αi − γi )! (αi − γi − 1)! (αi − βi )! βi

8
We then obtain
    n    Y n    
α αi Y αj αi αj α
γ! (αi − γi ) = γi ! (αi − γi ) γj ! = βi ! βj ! = β! . (8)
γ γi j=1
γ j β i j=1
β j β
j6=i j6=i

By combining (7) with (8), we see that

 
α α−β
∂ β xα = β! x ,
β
which completes the induction step.
The following lemma will be used to prove Proposition 3.15, which is a very useful inequality.
Lemma 3.14. Let k ∈ N be fixed. Suppose that G(γ, δ) is a function depending on multi-indices
γ and δ which satisfy |γ| = k and |δ| = 1. Suppose that H(α, i) is a function depending on multi-
indices α which satisfy |α| = k + 1 and i ∈ {1, . . . , n}. Also assume that H(γ + δ, iδ ) = G(γ, δ) for
all |γ| = k and |δ| = 1 where iδ is the index of the only nonzero component of δ. Then
X X X X
G(γ, δ) = H(α, i). (9)
|γ|=k |δ|=1 |α|=k+1 αi ≥1
P
where ” αi ≥1 ” means that we sum over all i ∈ {1, . . . , n} that satisfy αi ≥ 1.
Proof. The main idea is to prove that the number of terms on both sides are the same by constructing
a bijection. Let
S = {(γ, δ) | γ and δ are multi-indices such that |γ| = k and |δ| = 1}
T ={(α, i) | α is a multi-index such that |α| = k + 1 and i ∈ {1, . . . , n}
is any index such that αi ≥ 1}.
Then (9) is equivalent to proving that
X X
G(γ, δ) = H(α, i).
(γ,δ)∈S (α,i)∈T

Let F : S → T be the mapping (γ, δ) 7→ (γ + δ, iδ ). If (γ (1) , δ (1) ), (γ (2) , δ (2) ) ∈ S and F (γ (1) , δ (1) ) =
F (γ (2) , δ (2) ), then
γ (1) + δ (1) = γ (2) + δ (2) (10)
iδ(1) = iδ(2) . (11)
(11) implies δ (1) = δ (2) by the definition of iδ so then (10) implies γ (1) = γ (2) . So F is injective.
If (α, j) ∈ T , then αj ≥ 1. We then let δ be the multi-index with a 1 at position j and zeroes
everywhere else and put
γ = (α1 , . . . , αj − 1, . . . , αn ).
Then (γ, δ) ∈ S and
(α, j) = (γ + δ, iδ ) = F (γ, δ),
so F is surjective. So there is a bijection between the elements of S and the elements of T . Notice
that by assumption H(F (γ, δ)) = G(γ, δ). By combining this with the fact that F is a bijection
between S and T , we get
X X X X
G(γ, δ) = H(F (γ, δ)) = H(F (γ, δ)) = H(α, i),
(γ,δ)∈S (γ,δ)∈S F (γ,δ)∈F (S) (α,i)∈T

which completes the proof.

9
Proposition 3.15. For any k ∈ N,
X
|z|2k ≤ nk (z α )2 , z ∈ Rn .
|α|=k

Proof. We use induction on k. For k = 0, we have |z|2k = 1 and nk |α|=k (z α )2 = n0 (z 0 )2 = 1 so

P
the base case holds. Now suppose that k > 0 and that the inequality holds for k − 1. Then we have
X
|z|2k = |z|2 |z|2(k−1) ≤ |z|2 nk−1 (z γ )2 . (12)
|γ|=k−1

Notice that there are exactly n different multi-indices δ with |δ| = 1. Also, z δ = zi where i is the
index of the only nonzero component of δ. Hence,
n
X X
|z|2 = zi2 = (z δ )2 . (13)
i=1 |δ|=1

So by (12) and (13),

X X X X
|z|2k ≤ nk−1 (z δ )2 (z γ )2 = nk−1 (z γ )2 (z δ )2
|δ|=1 |γ|=k−1 |γ|=k−1 |δ|=1
X X (14)
k−1 γ+δ 2
=n (z ) .
|γ|=k−1 |δ|=1

By Lemma 3.14 we can rewrite the sum as

X X X X
(z γ+δ )2 = (z α )2 . (15)
|γ|=k−1 |δ|=1 |α|=k αi ≥1

The terms on the right hand side do not depend on i. So for each multi-index α in the sum, the
number of terms of the form (z α )2 is just the number of indices i ∈ {1, . . . , n} that satisfy αi ≥ 1,
which is less than or equal to n. So by (14) and (15),
X X X X
|z|2k ≤ nk−1 (z α )2 ≤ nk−1 n(z α )2 = nk (z α )2 ,
|α|=k αi ≥1 |α|=k |α|=k

which completes the induction step.

We will encounter integrals on Rn where the integrand only depends |x|, and then one can do
the change of variable r = |x| to transform into a single-variable integral:
∞ ∞
2π n/2
Z Z Z Z
f (|x|)dx = rn−1 f (r) dr dω = rn−1 f (r) dr, (16)
Rn Sn−1 0 Γ(n/2) 0

where Sn−1 ⊆ Rn is the n − 1-dimensional unit sphere. By this change of variable and elementary
calculus we also obtain the following result which will be used a lot throughout the text.
Proposition 3.16. The integral Z
1
dx
Rn (1 + |x|)p
is convergent if and only if p > n.

10
4 Littlewood-Paley partition of unity
An important technique in the theory of pseudo-differential operators is to decompose the operators
into small pieces. In this chapter we construct the main building blocks of these decompositions.
We first construct a smooth function on Rn with compact support in Proposition 4.2, which will
be used in multiple proofs throughout this text.
Lemma 4.1. The function f : R → R defined by
(  
exp x21−1 , |x| < 1
f (x) =
0, |x| ≥ 1

is smooth.
Proof. f is clearly smooth on the set defined by |x| = 6 1. So the only assertion left to prove is that
all partial derivatives exist at x = 1 and x = −1 but since f is even it suffices to do this for x = 1.
But first we prove the following by induction: For all k ∈ N, there exists a polynomial pk such that
 
pk (x) 1
f (k) (x) = 2 exp , |x| < 1. (17)
(x − 1)2k x2 − 1
This is clearly true for k = 0 and for the induction step, assume (17) holds for some k = m ∈ N.
Then by using the chain, product and division rule,
  
d pm (x) 1
f (m+1) (x) = exp
dx (x2 − 1)2m x2 − 1
(x2 − 1)2m p0m (x) − 4mx(x2 − 1)2m−1 pm (x)
 
1
= exp
(x2 − 1)4m x2 − 1
 
pm (x) 2x 1
− 2 exp
(x − 1)2m (x2 − 1)2 x2 − 1
(x2 − 1)2 p0m (x) − (2x + 4mx(x2 − 1))pm (x)
 
1
= exp , |x| < 1,
(x2 − 1)2m+2 x2 − 1
so (17) holds for k = m + 1 if we take

pm+1 (x) = (x2 − 1)2 p0m (x) − (2x + 4mx(x2 − 1))pm (x),

which completes the induction proof.

1
By using (17) and the change of variable u = 1−x2 , we see that

u2k
 
k pk (x) 1
lim f (x) = lim− 2 exp = pk (1) lim = 0, k ∈ N.
x→1− x→1 (x − 1)2k x2 − 1 u→∞ eu

Since f k (x) = 0 for x > 1, we also see that

lim f k (x) = 0, k ∈ N. (18)

x→1

Now we will again use induction to prove that for all k ∈ N, f (k) (1) exists and is equal to 0
(where f (0) = f ), which will complete the proof of this lemma. The base case k = 0 holds by the
definition of f . For the induction step, assume that for a given k ∈ N, f (k) (1) exists and is equal to
0. Then
f k (1 + h) − f k (1) f k (1 + h)
f (k+1) (1) = lim = lim = lim f k+1 (1 + h) = 0,
h→0 h h→0 h h→0

11
where the second equality follows by the induction assumption, the third follows from (18) and
L’Hôpital’s rule and the fourth also follows from (18). Thus f (k+1) (1) exists and is equal to 0, which
completes the induction step.
Proposition 4.2. There exists a function ψ ∈ C0∞ (Rn ) such that

0 ≤ ψ(x) ≤ 1, x ∈ Rn
ψ(x) = 1, |x| ≤ 1
ψ(x) = 0, |x| ≥ 2.

Proof. Let f be the function from Lemma 4.1 and define ϕ(t) = Cf (2t + 3) where C is the positive
constant defined by Z ∞
1
= f (2t + 3) dt.
C −∞

Then ϕ ∈ C0∞ (R) and

Z ∞
ϕ(t) dt = 1.
−∞

Define χ by
Z x
χ(x) = ϕ(t) dt x ∈ R.
−∞

Since ϕ is nonnegative, Z ∞
0 ≤ χ(x) ≤ ϕ(t) dt = 1
−∞

and because ϕ has its support in [−2, −1],

χ(x) = 0, x ≤ −2
Z ∞
χ(x) = ϕ(t) dt = 1, x ≥ −1.
−∞

Also, for all a ≤ −2, Z x

χ(x) = ϕ(x) dt, x ≥ a.
a

So by the fundamental theorem of calculus χ is differentiable everywhere and χ0 (x) = ϕ(x). This
shows that χ is smooth since ϕ is. Next define ψ : Rn → R by

ψ(x) = χ(−|x|), x ∈ Rn .

From the properties of χ we see that

0 ≤ ψ(x) ≤ 1, x ∈ Rn
ψ(x) = 1, |x| ≤ 1
ψ(x) = 0, |x| ≥ 2.

Note that ψ is smooth in the set |x| < 1 because it is constant there. It is also smooth in Rn \ {0}
since the the mapping x 7→ |x| is smooth for x 6= 0 and ψ is the composition of that mapping with
χ. Thus ψ ∈ C0∞ (Rn ), which completes the proof.

12
 We will use Proposition 4.2 to construct a partition of unity, which is a sequence of functions
with compact support that sums to 1. In Chapter 8.2, this will be used to split up a smooth function
into smooth functions with compact support. The sequence (ψk )∞ k=1 in the following proposition is
called a Littlewood-Paley partition of unity.
Proposition 4.3 (Partition of unity). Let ψ0 ∈ C0∞ (Rn ) be a function such that
0 ≤ ψ0 (ξ) ≤ 1, x ∈ Rn
ψ0 (ξ) = 1, |x| ≤ 1
ψ0 (ξ) = 0, |x| ≥ 2.
Define ψk ∈ C0∞ (Rn ) for integers k ≥ 1 by
ψk (ξ) = ψ0 (2−k ξ) − ψ0 (21−k ξ).
Then the following properties hold:
P∞
i) k=0 ψk (ξ) = 1, ∀ξ ∈ Rn .
ii) For k ≥ 1, the support of ψk is in {ξ ∈ Rn | 2k−1 ≤ |ξ| ≤ 2k+1 }.
iii) 0 ≤ ψk (ξ) ≤ 1 ∀k ∈ N.
Proof. Observe that if |ξ| ≤ 2k−1 or |ξ| ≥ 2k+1 , then ψ0 (2−k ξ) and ψ0 (21−k ξ) are both 0 or both 1.
Hence ii) holds. To prove i), note that for any ξ ∈ Rn , one can pick m > 0 such that |ξ| ≤ 2m−1 .
Then by ii), ψk (ξ) = 0 for all k ≥ m. Thus,
∞
X m
X m
X
ψk (ξ) = ψ0 (ξ) + ψk (ξ) = ψ0 (ξ) + ψ0 (2−k ξ) − ψ0 (21−k ξ)
k=0 k=1 k=1
= ψ0 (ξ) + ψ0 (2−m ξ) − ψ0 (ξ) = ψ0 (2−m ξ) = 1,
since 2−m |ξ| ≤ 2−m 2m−1 = 21 . So i) holds.
Property iii) follows directly from the definition of ψk and the properties of ψ0 since
ψk (ξ) = 1 − ψ0 (21−k ξ), |ξ| ≤ 2k
ψk (ξ) = ψ0 (2−k ξ), |ξ| ≥ 2k .

We also estimate the size of the derivatives of ψk :

Proposition 4.4. Let (ψk )∞ k=0 be the Littlewood-Paley partition of unity from Proposition 4.3. Then
for all multi-indices α, there is a constant Cα > 0 such that for all k ∈ N,
|∂ α ψk (ξ)| ≤ Cα 2−k|α| , ξ ∈ Rn .
Proof. For each multi-index α, let
Aα = sup |∂ α ψ0 (ξ)|.
ξ∈Rn

This is well defined since ψ0 and all its derivatives have compact support. For any multi-index α
and integer k > 0, we obtain by applying the chain rule |α| times that

|∂ α ψk (ξ)| = ∂ α (ψ0 (2−k ξ)) − ∂ α (ψ0 (21−k ξ)) = 2−k|α| ∂ α ψ0 (2−k ξ) − 2(1−k)|α| ∂ α ψ0 (21−k ξ)
 
≤ 2−k|α| |∂ α ψ0 (2−k ξ)| + 2|α| |∂ α ψ0 (21−k ξ)| ≤ 2−k|α| (Aα + Aα 2|α| ) = Cα 2−k|α| ,

where Cα = (Aα + Aα 2|α| ). By the definition of Aα , this holds for k = 0 as well, which completes
the proof.

13
5 The Schwartz space and symbols
In this section we will introduce some important function spaces that are necessary for defining
pseudo-differential operators.

5.1 Schwartz space

Definition 5.1. The Schwartz space, denoted by S , is the set of functions ϕ ∈ C ∞ (Rn ) such that
for all multi-indices α and β, the function xα ∂ β ϕ(x) is bounded on Rn . In other words,

sup |xα ∂ β ϕ(x)| < ∞.

x∈Rn

A function in S is called a Schwartz function.

In S , there is a notion of convergence. This definition will be used in Chapter 7 to define
tempered distributions.
Definition 5.2. Let (ϕi ) be a sequence in S . (ϕi ) is said to converge to 0 in S if for all multi-
indices α and β, the sequence of functions (xα ∂ β ϕi ) converges uniformly to 0 as i → ∞. In other
words,
sup |xα ∂ β ϕi (x)| → 0
x∈Rn

as i → ∞.
The following properties of S will be used frequently throughout this text and will not be explic-
itly referenced. They imply that S is closed under addition, scalar multiplication, multiplication
by another Schwartz function, multiplication by polynomials and differentiation.

Proposition 5.3. The Schwartz space satisfies the following properties:

i) S is a vector space.
ii) If ϕ, ψ ∈ S , then ϕψ ∈ S .
iii) If ϕ ∈ S , then xα ϕ ∈ S for any multi-index α.

iv) ϕ ∈ S , then ∂ α ϕ ∈ S for any multi-index α.

v) C0∞ (Rn ) ⊂ S
Proof. i) This follows directly from the definition of S and the fact that the set of bounded
functions is closed under addition and scalar multiplication.

ii) By using Leibniz’s formula, we see that for any multi-indices α and β,
X β 
xα ∂ β (ϕψ)(x) = xα ∂ γ ϕ(x)∂ β−γ ψ(x).
γ
γ≤β

Since ϕ, ψ ∈ S , the functions xα ∂ γ ϕ(x) and ∂ β−γ ψ(x) are bounded for all γ ≤ β. Since
addition and multiplication of bounded functions yield a bounded function, the above equality
shows that ϕψ ∈ S .

14
iii) By using Leibniz’s formula and Proposition 3.13, we see that for any multi-indices β and γ,
X γ 
xβ ∂ γ (xα ϕ(x)) = xβ ∂ δ (xα )∂ γ−δ ϕ(x)
δ
δ≤γ
X γ α
= δ! xα+β−δ ∂ γ−δ ϕ(x).
δ δ
δ≤γ
δ≤α

In the above sum, every term xα+β−δ ∂ γ−δ ϕ(x) is a bounded function by assumption so the
whole sum is a bounded function, hence xα ϕ ∈ S .

iv) For all multi-indices β and γ,

xβ ∂ γ (∂ α ϕ(x)) = xβ ∂ γ+α ϕ(x).

The right hand side is bounded since ϕ ∈ S . So ∂ α ϕ ∈ S by the definition of S .

v) This follows directly from the definition of S and the fact that continuous functions with
compact support are bounded.

The set of sequences in S that converge to 0 is closed under addition, multiplication and differ-
entiation in an analogous way to Proposition 5.3:

Proposition 5.4. The set of sequences in S that converge to 0 satisfies the following properties:
i) If (ϕi ) and (ψi ) converges to 0 in S , then (ϕi + ψi ) converges to 0 in S .
ii) If (ϕi ) converges to 0 in S and λ ∈ C, then (λϕi ) converges to 0 in S .
iii) If (ϕi ) and (ψi ) converges to 0 in S , then (ϕi ψi ) converges to 0 in S .

iv) If (ϕi ) converges to 0 in S , then (xα ϕi ) converges to 0 in S for any multi-index α.

v) If (ϕi ) converges to 0 in S , then (∂ α ϕi ) converges to 0 in S for any multi-index α.
Proof. This is essentially proved in the same way as the corresponding properties in Proposition 5.3.
The main difference is that in Proposition 5.3 we used the fact that the set of bounded functions
Rn → C is closed under addition, multiplication and multiplication by scalar. In this case we use
this closeness property to the set of sequences of bounded functions that converge uniformly to 0.
For example, to prove iii ) we use the calculation from Proposition 5.3.ii ) to get that for all
multi-indices α and β,
X β 
xα ∂ β (ϕi ψi )(x) = xα ∂ γ ϕi (x)∂ β−γ ψi (x).
γ
γ≤β

By assumption, each term xα ∂ γ ϕi (x) and ∂ β−γ ψi (x) are bounded functions that converge uniformly
to 0, so the whole sum converge uniformly to 0 as i → ∞. Thus (ϕi ψi ) converges to 0 in S . The
other properties are proven by modifying the proof of the corresponding properties from Proposition
5.3 in the same way.

There is an alternative definition of S and convergence in S that will be useful.

15
Proposition 5.5. ϕ ∈ S if and only if for all multi-indices β and k ∈ N, the function (1 + |x|)k ∂ β ϕ
is bounded, i.e
sup (1 + |x|)k |∂ β ϕ(x)| < ∞.
x∈Rn

The equivalence also holds if N is replaced by R in the above statement.

Proof. =⇒ : Suppose ϕ ∈ S . Let k ∈ N and the multi-index β be arbitrary. By Proposition 3.12,
there is a constant C > 0 such that

(1 + |x|)k |∂ β ϕ(x)| ≤ C(1 + |x|2 )k/2 |∂ β ϕ(x)| ≤ C(1 + |x|2 )k |∂ β ϕ(x)|. (19)

The expression (1 + |x|2 )k is a polynomial in x and therefore a sum of powers of x. By Proposition

5.3, this implies that (1 + |x|2 )k ∂ β ϕ ∈ S so it is a bounded function. Then (1 + |x|)k ∂ β ϕ is bounded
by (19). If m ∈ R, then we pick k ∈ N such that k ≥ m. Because (1 + |x|)m ≤ (1 + |x|)k , the
statement holds for all m ∈ R as well.
⇐= : Suppose that for all multi-indices β and k ∈ N, the function (1 + |x|)k ∂ β ϕ is bounded.
Then for all multi-indices α and β,

|xα ∂ β ϕ(x)| ≤ |x||α| |∂ β ϕ(x)| ≤ (1 + |x|)|α| |∂ β ϕ(x)|. (20)

The last expression in (20) is bounded by assumption, so the inequality shows that xα ∂ β ϕ is bounded.
This proves that ϕ ∈ S .
Proposition 5.6. A sequence (ϕi ) in S converges to 0 in S if and only if for all multi-indices β
and k ∈ N, the function sequence ((1 + |x|)k ∂ β ϕi ) converges uniformly to 0, i.e

sup (1 + |x|)k |∂ β ϕi (x)| → 0

x∈Rn

as i → ∞.
Proof. The proof is similar to the one of Proposition 5.5. We just use Definition 5.2 and the
inequalities (19) and (20) to prove the uniform convergence to 0.
From Proposition 5.5 it follows that for all k ∈ N, a Schwartz function is bounded by an expression
of the form C(1+|x|)−k . Hence S ⊂ Lp (Rn ) and by Theorem 3.2 and Proposition 5.3.v), we directly
get the following proposition:
Proposition 5.7. For any p ≥ 1, S is dense in Lp (Rn ).
The following is a multivariate version of integration by parts which will be used frequently in
this text.
Proposition 5.8 (Integration by Parts). Let g ∈ S and let f ∈ C ∞ (Rn ) be such that all its deriva-
tives are bounded by a polynomial, or more precisely, for every multi-index α there are constants
Cα , mα > 0 such that
|∂ α f (x)| ≤ Cα (1 + |x|)mα , x ∈ Rn .
Then for all multi-indices α,
Z Z
∂ α f (x)g(x) dx = (−1)|α| f (x)∂ α g(x) dx.
Rn Rn

Proof. The proof is by induction on |α|. The base case α = 0 clearly holds. Now assume |α| > 0
and that the equality holds for all multi-indices α0 of length less than |α| and all functions f 0 and
g 0 that satisfy the assumptions in the proposition. Write α = γ + δ where |γ| = |α| − 1 and |δ| = 1.
Let i be the index of the nonzero component of δ.

16
 Note that since g ∈ S there is by Proposition 5.5 a constant Dα > 0 such that

|g(x)| ≤ Dα (1 + |x|)mγ −1 .

Thus
|∂ γ f (x)g(x)| ≤ Cγ Dγ (1 + |x|)mγ (1 + |x|)mγ −1 ≤ Cγ Dγ (1 + |x|)−1 → 0
as |x| → ∞. So by single-variable integration by parts,
Z ∞ Z ∞
γ γ ∞
∂i ∂ f (x)g(x) dxi = [∂ f (x)g(x)]xi =−∞ − ∂ γ f (x)∂i g(x) dxi
−∞ −∞
Z ∞ (21)
=− ∂ γ f (x)∂i g(x) dxi .
−∞

Therefore by (21), the induction assumption and the fact that ∂i g ∈ S ,

Z Z
α
∂ f (x)g(x) dx = ∂i ∂ γ f (x)g(x) dx
Rn Rn
Z ∞ Z ∞
= ··· ∂i ∂ γ f (x)g(x) dxi dx1 · · · dxi−1 dxi+1 · · · dxn
−∞ −∞
Z ∞ Z ∞
=− ··· ∂ γ f (x)∂i g(x) dxi dx1 · · · dxi−1 dxi+1 · · · dxn
Z−∞ −∞
Z
=− ∂ f (x)∂i g(x) dx = −(−1)|γ|
γ
f (x)∂ γ ∂i g(x) dx
R n R n
Z
|α| α
= (−1) f (x)∂ g(x) dx,
Rn

which completes the induction step.

17
5.2 Symbols
Definition 5.9. For any m ∈ R, the space S m is defined to be the set of functions σ ∈ C ∞ (Rn ×Rn )
such that for all multi-indices α and β, there exists a constant Cα,β such that

|∂xα ∂ξβ σ(x, ξ)| ≤ Cα,β (1 + |ξ|)m−|β| , x, ξ ∈ Rn . (22)

S m is called a symbol.
S
A function in m∈R

Remark. If a symbol σ is independent of either the x-variable or ξ-variable, we will simply write
σ(ξ) = σ(x, ξ) or σ(x) = σ(x, ξ) respectively and consider its domain to be Rn instead of Rn × Rn .
We list some properties of symbols. They will be used frequently throughout this text and
sometimes even without reference.
Proposition 5.10. Symbols satisfy the following properties:
i) S m is a vector space.

ii) If σ ∈ S m1 and τ ∈ S m2 , then στ ∈ S m1 +m2 .

iii) If σ ∈ S m , then for all multi-indices α and β, ∂xα ∂ξβ σ ∈ S m−|β| .

iv) If m1 ≤ m2 , then S m1 ⊆ S m2 .

v) If σ ∈ S m , ϕ ∈ S , and τ is defined by τ (x, ξ) = σ(x, ξ)ϕ(ξ), then τ ∈ k∈R S k .

T

vi) If ϕ ∈ S and τ is defined by τ (x, ξ) = ϕ(ξ), then τ ∈ m∈R S m .

T

vii) If σ ∈ m∈R S m , then for any fixed x, the function ϕx defined by ϕx (ξ) = σ(x, ξ) is in S .
T

viii) If σ ∈ S m has compact support in the ξ-variable, then σ ∈ k∈R S k

T

ix) If σ ∈ S m , ϕ ∈ S , and τ is defined by τ (x, ξ) = σ(x, ξ)ϕ(x), then τ ∈ S m .

Proof. i) Clearly 0 ∈ S m . If σ, τ ∈ S m , then for all multi-indices α and β there are constants Aα,β
and Bα,β such that

|∂xα ∂ξβ σ(x, ξ)| ≤ Aα,β (1 + |ξ|)m−|β|

|∂xα ∂ξβ τ (x, ξ)| ≤ Bα,β (1 + |ξ|)m−|β| .

By the triangle inequality,

|∂xα ∂ξβ (σ + τ )(x, ξ)| ≤ |∂xα ∂ξβ σ(x, ξ)| + |∂xα ∂ξβ τ (x, ξ)|
≤ (Aα,β + Bα,β )(1 + |ξ|)m−|β| .

So σ + τ ∈ S m . If λ ∈ C, then

|∂xα ∂ξβ (λσ)(x, ξ)| = |λ||∂xα ∂ξβ σ(x, ξ)| ≤ |λ|Aα,β (1 + |ξ|)m−|β| .

So λσ ∈ S m . Thus S m is a vector space.

18
ii) By using the triangle inequality and Leibniz’s formula twice, we get that for all multi-indices α
and β,
X β 
|∂xα ∂ξβ (στ )(x, ξ)| = ∂xα ( ∂ δ σ(x, ξ)∂ξβ−δ τ (x, ξ))
δ ξ
δ≤β

X X β α
= ∂ γ ∂ δ σ(x, ξ)∂xα−γ ∂ξβ−δ τ (x, ξ)
δ γ x ξ
δ≤β γ≤α
X X β α
≤ |∂xγ ∂ξδ σ(x, ξ)||∂xα−γ ∂ξβ−δ τ (x, ξ)|.
δ γ
δ≤β γ≤α

By assumption, there are constants Aγ,δ and Bα,β,γ,δ such that

|∂xγ ∂ξδ σ(x, ξ)| ≤ Aγ,δ (1 + |ξ|)m1 −|δ|

|∂xα−γ ∂ξβ−δ τ (x, ξ)| ≤ Bα,β,γ,δ (1 + |ξ|)m2 −|β|+|δ| .

Thus,
X X β α
|∂xα ∂ξβ (στ )(x, ξ)| ≤ Aγ,δ Bα,β,γ,δ (1 + |ξ|)m1 +m2 −|β|
δ γ
δ≤β γ≤α

≤ Cα,β (1 + |ξ|)m1 +m2 −|β|

where
X X β α
Cα,β = Aγ,δ Bα,β,γ,δ .
δ γ
δ≤β γ≤α

Hence στ ∈ S m1 +m2 .
iii) We fix α and β. By the definition of S m ,

|∂xγ ∂ξδ (∂xα ∂ξβ σ)(x, ξ)| = |∂xα+γ ∂ξβ+δ σ(x, ξ)| ≤ Cα+γ,β+δ (1 + |ξ|)m−|β|−|δ| .

Since α and β are fixed, Cα+γ,β+δ depends only on γ and δ, so ∂xα ∂ξβ σ ∈ S m−|β| by definition.
iv) This follows directly from the definition of S m and the fact that m1 ≤ m2 implies

(1 + |ξ|)m1 −|β| ≤ (1 + |ξ|)m2 −|β| .

v) Let k ∈ R be arbitrary. By Leibniz’s formula and the triangle inequality, we get for all multi-
indices α and β,
X β 
α β
|∂x ∂ξ τ (x, ξ)| ≤ |∂xα ∂ξγ σ(x, ξ)||∂ β−γ ϕ(ξ)|.
γ
γ≤β

Since σ ∈ S m , there is a constant Aα,γ such that

|∂xα ∂ξγ σ(x, ξ)| ≤ Aα,γ (1 + |ξ|)m−|γ| .

Because ϕ ∈ S , there is by Proposition 5.5 a constant Bβ,γ such that

(1 + |ξ|)m+|β|−|γ|−k |∂ β−γ ϕ(ξ)| ≤ Bβ,γ .

19
 Thus,
X β 
|∂xα ∂ξβ τ (x, ξ)| ≤ Aα,γ (1 + |ξ|)m−|γ| (1 + |ξ|)k+|γ|−|β|−m Bβ,γ
γ
γ≤β

= Cα,β (1 + |ξ|)k−|β|

where
X β 
Cα,β = Aα,γ Bβ,γ .
γ
γ≤β

So τ ∈ S k . Because k was arbitrary, this implies that τ ∈ Sk.

T
k∈R

vi) This is a special case of v) for σ = 1.

vii) Since σ ∈ m∈R S m , there exist for every m ∈ R and multi-index α a constant Cm,α such that
T

|∂ξα σ(x, ξ)| ≤ Cm,α (1 + |ξ|)m−|α| ≤ Cm,α (1 + |ξ|)m .

Now for all k ∈ N and multi-indices α, we get by taking m = −k in the above inequality that

(1 + |ξ|)k |∂ α ϕx (ξ)| = (1 + |ξ|)k |∂ξα σ(x, ξ)| ≤ (1 + |ξ|)k C−k,α (1 + |ξ|)−k

= C−k,α .

By Proposition 5.5, this implies that ϕx ∈ S .

viii) Fix k ∈ R and let W be the support of σ. Then the support of every derivative of σ is contained
in W . Since W is compact in the ξ-variable, there exists a constant M such that

(1 + |ξ|)m−k ≤ M, ∀(x, ξ) ∈ W.

Because σ ∈ S m , we then get for all multi-indices α and β,

|∂xα ∂ξβ σ(x, ξ)| ≤ Cα,β (1 + |ξ|)m−|β| = Cα,β (1 + |ξ|)m−k (1 + |ξ|)k−|β| ≤ Cα,β M (1 + |ξ|)k−|β| .

So σ ∈ S k . Because k ∈ R was arbitrary, this completes the proof.

ix) By the definition of S , there exists for each multi-index γ a constant Dγ such that

|∂xγ ϕ(x)| ≤ Dγ .

So by Leibniz’s formula and (22),

X α  X α 
α β γ α−γ β
|∂x ∂ξ τ (x, ξ)| ≤ |∂x ϕ(x)||∂x ∂ξ σ(x, ξ)| ≤ Dγ Cα−γ,β (1 + |ξ|)m−|β|
γ γ
γ≤α γ≤α
X α 
0
≤ Dγ Cα−γ,β (1 + |ξ|)m−|β| = Cα,β (1 + |ξ|)m−|β|
γ
γ≤α

where
X α
0
Cα,β = Dγ Cα−γ,β .
γ
γ≤α

It follows that τ ∈ S m .

20
 The following example of symbols will be used in Chapter 8 for the theory of partial differential
operators.
Proposition 5.11. Let X
P (x, ξ) = aα (x)ξ α , x, ξ ∈ Rn ,
|α|≤m

where m ∈ N and all the coefficients aα ∈ C ∞ (Rn ) and the derivatives ∂ β aα are bounded for all
multi-indices β. Then P ∈ S m .
Proof. For multi-indices |α| ≤ m and β, let Cα,β = supx∈Rn ∂ β aα (x). By Proposition 3.13, we get
for all multi-indices β and γ,
X X
|∂xβ ∂ξγ P (x, ξ)| ≤ |∂xβ ∂ξγ (aα (x)ξ α )| = |∂ β aα (x)||∂ γ ξ α |
|α|≤m |α|≤m
  X α 
X α
≤ γ! Cα,β |ξ α−γ
|≤ γ! Cα,β |ξ||α|−|γ|
γ γ
|α|≤m |α|≤m
α≥γ α≥γ
X α 
0
≤ γ! Cα,β |ξ|m−|γ| ≤ Cβ,γ (1 + |ξ|)m−|γ|
γ
|α|≤m
α≥γ

where  
0
X α
Cβ,γ = γ! Cα,β .
γ
|α|≤m
α≥γ

It follows that P ∈ S m .
An important concept in the theory of pseudo-differential operators are asymptotic expansions.
In some cases, like in the proof of Theorem 8.8, we will construct symbols in terms of asymptotic
expansions.

Definition 5.12. Suppose σ ∈ S m and we have a sequence of symbols (σi )∞ i=0 such that σi ∈ S
mi
,
∞
where (mi )i=0 is a strictly decreasing sequence of real numbers, m0 = m and mi → −∞ as i → ∞ .
If
Xk
σ− σi ∈ S mk+1
i=0
P∞
for every k ∈ N, then we say that i=0 σi is an asymptotic expansion of σ and write
∞
X
σ∼ σi .
i=0

The proof of the following proposition can be found in [6, Theorem 6.10, p. 36-39].
Proposition 5.13. Suppose (σi )∞ i=0 is a sequence of symbols such that σi ∈ S
mi
, where (mi )∞
i=0 is
a strictly decreasing sequence of real numbers and mi → −∞ as i → ∞. Then there exists a symbol
σ ∈ S m0 such that
X∞
σ∼ σi .
i=0

21
5.3 Elliptic symbols
A special type of symbols called elliptic symbols are of particular importance in the theory of pseudo-
differential operators. Symbols are bounded from above by an expression of the form C(1 + |ξ|)m .
Elliptic symbols are additionally bounded from below by D(1 + |ξ|)m for some positive constant D.
Definition 5.14. A symbol σ is called m-elliptic if σ ∈ S m and there exist constants C > 0 and
R ≥ 0 such that
|σ(x, ξ)| ≥ C(1 + |ξ|)m , |ξ| ≥ R.
If the real number m is not important in the context, σ is just called elliptic.
Note that unlike general symbols, the set of elliptic symbols is not closed under addition. For
example, if σ is m-elliptic then −σ is as also m-elliptic but the sum σ − σ = 0 is not. But it is closed
under multiplication:
Proposition 5.15. If σ ∈ S m1 is m1 -elliptic and τ ∈ S m1 is m2 -elliptic, then στ is m1 +m2 -elliptic.
Proof. Pick C1 , C2 > 0 and R1 , R2 > 0 such that

|σ(x, ξ)| ≥ C1 (1 + |ξ|)m1 , |ξ| ≥ R1

m2
|τ (x, ξ)| ≥ C2 (1 + |ξ|) , |ξ| ≥ R2 .

Then for all |ξ| ≥ max(R1 , R2 ),

|σ(x, ξ)τ (x, ξ)| ≥ C1 C2 (1 + |ξ|)m1 +m2 .

This together with Proposition 5.10.ii) shows that στ is m1 + m2 -elliptic.

The following proposition will be used later in the proof of Theorem 5.16.
Proposition 5.16. Suppose that σ ∈ S m is m-elliptic and let C > 0 and R > 0 be such that

|σ(x, ξ)| ≥ C(1 + |ξ|)m , |ξ| ≥ R. (23)

By Proposition 4.2, take ψ ∈ C ∞ (Rn ) such that

0 ≤ ψ(ξ) ≤ 1, ξ ∈ Rn
ψ(ξ) = 0, |ξ| ≤ R
ψ(ξ) = 1, |ξ| ≥ 2R.

Then the function τ defined by (

ψ(ξ)
σ(x,ξ) , |ξ| ≥ R
τ (x, ξ) =
0, |ξ| < R
is a −m-elliptic symbol.
Proof. First we need to prove τ ∈ C ∞ (Rn × Rn ). Since σ ∈ C ∞ (Rn × Rn ) and is nonzero for
|ξ| ≥ R by (23), it follows that τ is smooth on the open set of (x, ξ) ∈ Rn satisfying |ξ| > R and
obviously it is smooth on |ξ| < R. Now assume that (x, ξ) ∈ Rn and |ξ| = R. Since σ is continuous,
ψ(ξ 0 )
there is an open neighborhood U around (x, ξ) where σ is nonzero. Then τ (x0 , ξ 0 ) = σ(x 0 ,ξ 0 ) for all
0 0
(x , ξ ) ∈ U by the definition of τ and the properties of ψ. Thus τ is smooth on U . It follows that
τ ∈ C ∞ (Rn × Rn ).
To prove τ ∈ S m , we need an expression for the partial derivatives of τ . Clearly ∂xα ∂ξβ τ (x, ξ) =
0 for |ξ| ≤ R and all multi-indices α and β from the definition of τ and the continuity of its

22
derivatives. For |ξ| > R we first prove that for all multi-indices α and β, there exists a symbol
υα,β ∈ S (|α|+|β|)m−|β| such that

υα,β (x, ξ)
∂xα ∂ξβ τ (x, ξ) = , |ξ| > R. (24)
(σ(x, ξ))|α|+|β|+1

This is done by induction on |α| + |β|. First define υ0,0 (x, ξ) = ψ(ξ). Since 1 − ψ ∈ S , it follows
that 1 − υ0,0 ∈ S 0 , so υ0,0 ∈ S 0 . So the base case α = β = 0 holds.
0 0 0
Now suppose that |α| + |β| > 0 and that there exist υα0 ,β 0 ∈ S (|α |+|β |)m−|β | that satisfy (24)
for all α0 ,β 0 with |α0 | + |β 0 | < |α| + |β|. There are two cases:
Case 1: If |α| > 0, then we can write α = γ + δ where |γ| = |α| − 1, |δ| = 1 as has been done in
previous proofs. Then by the induction assumption and the division rule,
 υ 
γ,β
∂xα ∂ξβ τ = ∂xδ (∂xγ ∂ξβ τ ) = ∂xδ
σ |γ|+|β|+1
|γ|+|β|+1 δ
σ ∂x υγ,β − (|γ| + |β| + 1)υγ,β σ |γ|+|β| ∂xδ σ
= (25)
σ 2(|γ|+|β|+1)
σ∂xδ υγ,β − (|γ| + |β| + 1)υγ,β ∂xδ σ υα,β
= |γ|+|β|+2
= |α|+|β|+1
σ σ
where
υα,β = σ∂xδ υγ,β − (|γ| + |β| + 1)υγ,β ∂xδ σ. (26)
Notice that by Proposition 5.10.iii), ∂xδ σ ∈ S m and ∂xδ υγ,β ∈ S (|γ|+|β|)m−|β| . By combining this with
(26) and Proposition 5.10.i) and ii), we see that

υα,β ∈ S (|γ|+|β|)m−|β|+m = S (|α|+|β|)m−|β| .

Case 2: If |β| > 0, then we can write β = γ + δ as in Case 1 and do the same calculation (25) but
with ∂ξδ instead of ∂xδ . We then obtain
υα,β
∂xα ∂ξβ τ = |α|+|β|+1
,
σ
where
υα,β = σ∂ξδ υα,γ − (|α| + |γ| + 1)υα,γ ∂ξδ σ. (27)
By using Proposition 5.10 as before, ∂ξδ σ ∈ S m−1 and ∂ξδ υα,γ ∈ S (|α|+|γ|)m−|γ|−1 . Combining this
with (27) shows that
υα,β ∈ S (|α|+|γ|)m−|γ|+m−1 = S (|α|+|β|)m−|β| ,
which completes the induction proof.
Now to prove that τ ∈ S m . Since υα,β ∈ S (|α|+|β|)m−|β| for any multi-indices α and β we can
take Dα,β > 0 such that
|υα,β (x, ξ)| ≤ Dα,β (1 + |ξ|)(|α|+|β|)m−|β| . (28)
Then by (23),(24) and (28),

|υα,β (x, ξ)| Dα,β (1 + |ξ|)(|α|+|β|)m−|β|

|∂xα ∂ξβ τ (x, ξ)| = ≤
|σ(x, ξ)||α|+|β|+1 C |α|+|β|+1 (1 + |ξ|)(|α|+|β|+1)m
Dα,β
= (1 + |ξ|)−m−|β| , |ξ| > R,
C
so τ ∈ S −m .

23
 Finally, to prove that τ is −m-elliptic, take E > 0 such that

|σ(x, ξ)| ≤ E(1 + |ξ|)m .

Then
ψ(ξ) 1 1
|τ (x, ξ)| = = ≥ (1 + |ξ|)−m , |ξ| ≥ 2R.
|σ(x, ξ)| |σ(x, ξ)| E

24
6 Fourier analysis
In this chapter, some basic definitions and theorems from Fourier analysis are stated without proof.
Most of these are well known theorems and to prove them is not the purpose of this text. The
interested reader can find the details elsewhere, e.g. [6, Section 3-4] which this chapter is based on.
Definition 6.1 (Fourier transform and inverse Fourier transform). For any f ∈ L1 (Rn ), define the
Fourier transform fˆ by Z
n
fˆ(ξ) = (2π)− 2 e−iξ·x f (x) dx
Rn

and the inverse Fourier transform fˇ by

Z
n
fˇ(ξ) = (2π)− 2 eiξ·x f (x) dx.
Rn

The Riemann-Lebesque lemma says that the Fourier transform is continuous and vanishes at
infinity:
Theorem 6.2 (Riemann-Lebesque lemma). If f ∈ L1 (Rn ), then fˆ ∈ C(Rn ) and

lim fˆ(ξ) = 0.
ξ→∞

The above statements hold for fˇ as well.

It is not necessarily the case that the Fourier transform of f ∈ L1 (Rn ) is in L1 (Rn ). However,
the Fourier transform map S into S .
Proposition 6.3. If f ∈ S , then fˆ ∈ S and fˇ ∈ S .
One can also prove that the Fourier transform and inverse Fourier transform restricted to S
is a bounded linear operator in L2 -norm. So by Theorem 3.6, they can be extended to bounded
linear operators on L2 (Rn ). For these extensions, we have the following two theorems. The first is
Plancherel’s formula which says that the Fourier transform is an isometry in L2 -norm.
Theorem 6.4 (Plancherel’s formula). If f ∈ L2 (Rn ), then

kf k2 = kfˆk2 = kfˇk2 .

The Fourier inversion theorem says that on L2 (Rn ), the inverse Fourier transform that we defined
actually is an inverse.
ˇ
Theorem 6.5 (Fourier inversion theorem). If f ∈ L2 (Rn ), then fˆ = f and fˆˇ = f almost everywhere.
The following formulas follow from the definition of the Fourier transform and integration by
parts.
Proposition 6.6. Suppose ϕ ∈ S . Then for all multi-indices α, the following equalities hold:
i) (∂ α ϕ)∧ (ξ) = (iξ)α ϕ̂(ξ).
ii) ∂ α ϕ̂(ξ) = ((−ix)α ϕ)∧ (ξ).
iii) (∂ α ϕ)∨ (ξ) = (−iξ)α ϕ̌(ξ).
iv) ∂ α ϕ̌(ξ) = ((ix)α ϕ)∨ (ξ).

25
7 Tempered distributions
We give a short introduction to tempered distributions. The main purpose is to define Sobolev
spaces in Chapter 9.
Definition 7.1. A tempered distribution is a linear functional U : S → C such that for every
sequence (ϕi ) converging to 0 in S , (U ϕi ) converges to 0 in C. The space of tempered distributions
is denoted by S 0 .
Definition 7.2. A tempered function f : Rn → C is a measurable function such that for some k ∈ N
|f (x)|
Z
k
dx < ∞. (29)
Rn (1 + |x|)

Proposition 7.3. All functions in Lp (Rn ) are tempered functions.

Proof. The case p = 1 is trivial (just take k = 0 in (29)). So assume p > 1 and f ∈ Lp (Rn ). Take
q > 1 such that p1 + 1q = 1. Then by Hölder’s inequality,
1
|f (x)|
Z Z
1 q

n+1 dx < kf kp dx < ∞,

Rn (1 + |x|) q RN (1 + |x|)n+1

n+1
since the integral on the right hand side is finite. So if we take k ∈ N such that k ≥ q , then the
above inequality shows that (29) holds, thus f is a tempered function.
Tempered distributions can be constructed from tempered functions in the following way:
Proposition 7.4. Let f be a tempered function and define Uf : S → C by
Z
Uf (ϕ) = f (x)ϕ(x) dx.
Rn

Then Uf is a tempered distribution.

Proof. First we prove that the integral is finite so that Uf is well defined. Pick k ∈ N such that (29)
holds for f . Then by Proposition 5.5,
Z Z
f (x)ϕ(x) dx ≤ |f (x)||ϕ(x)| dx
Rn Rn
(30)
|f (x)|
Z
≤ sup ((1 + |x|)k |ϕ(x)|) k
dx < ∞,
x∈Rn Rn (1 + |x|)

so Uf is well defined. Uf is clearly linear, since the integral is linear. Finally, if (ϕi ) is a sequence
converging to 0 in S , then by (30) and Proposition 5.6,

|f (x)|
Z
k
|Uf (ϕi )| ≤ sup ((1 + |x|) |ϕi (x)|) dx → 0
x∈R n
R n (1 + |x|)k

as i → ∞. Thus Uf is a tempered distribution.

The function defining a tempered distribution is unique, meaning that if Uf = Ug , then f = g
almost everywhere. A proof of this fact when f and g are continuous can be found in [6, Lemma
6.6, p. 34] but the general proof is omitted.
Proposition 7.5. Let f and g be tempered functions such that Uf = Ug . Then f = g almost
everywhere.

26
 Proposition 7.5 together with Proposition 7.3 implies that there is a bijection between Lp (Rn )
and a subset of S 0 defined by f 7→ Uf . This is also a linear isomorphism. Therefore we can view
Lp (Rn ) and S as subsets of S 0 .
The Fourier transform can be defined on the set S 0 as follows:
Definition 7.6. For every U ∈ S 0 , the Fourier transform Û is defined as the tempered distribution

Û (ϕ) = U (ϕ̂), ϕ ∈ S.

The inverse Fourier transform Ǔ is defined as the tempered distribution

Ǔ (ϕ) = U (ϕ̌), ϕ ∈ S.
ˇ
It follows directly from the definition and the Fourier inversion theorem that Û = U .

27
8 Pseudo-differential operators
8.1 Definition of Pseudo-differential operators
Now we finally come to the definition of pseudo-differential operators.
Definition 8.1 (Pseudo-differential operator). For any symbol σ the pseudo-differential operator
Tσ : S → S is defined to be the linear operator such that for any ϕ ∈ S , Tσ ϕ is the function
Z
−n
(Tσ ϕ)(x) = (2π) 2 eix·ξ σ(x, ξ)ϕ̂(ξ) dξ, x ∈ Rn .
Rn
To see that the above definition makes sense, we need to T check that Tσ ϕ ∈ S for every ϕ ∈ S .
First notice that the function τ (x, ξ) = σ(x, ξ)ϕ̂(ξ) is in m∈R S m by Proposition 5.10.v) and
therefore τ (x, ·) ∈ S for each fixed x Proposition 5.10.vii). To prove Tσ ϕ ∈ C ∞ (Rn ), notice that
for all multi-indices β, Leibniz’s formula implies,
Z X β  Z
|∂xβ (eix·ξ τ (x, ξ))| dξ = |ξ γ eix·ξ ∂xβ−γ τ (x, ξ)| dξ
Rn γ R n
γ≤β
X β  Z (31)
|γ| β−γ
≤ |ξ| |∂x τ (x, ξ)| dξ.
γ Rn
γ≤β

Since τ ∈ m∈R S , it follows by Proposition 5.10.iii) that ∂xβ−γ τ ∈

m
S m ⊆ S −n−|γ|−1 .
T T
m∈R
Therefore there is a constant Cβ,γ such that
|∂xβ−γ τ (x, ξ)| ≤ Cβ,γ (1 + |ξ|)−n−|γ|−1 .
(1 + |ξ|)−n−1 ∈ L1 (Rn ) so by (31)
Z X β  Z
|∂xβ (eix·ξ τ (x, ξ))| dξ ≤ Cβ,γ (1 + |ξ|)−n−1 dξ < ∞.
Rn γ Rn
γ≤β

The right hand side in the above expression is a constant independent of x ∈ Rn , so by Theorem
3.8, Tσ ϕ ∈ C ∞ (Rn ) and we can interchange the order of differentiation and integration.
Then for all multi-indices α and β, Leibniz’s formula and Proposition 5.8 (which can be applied
below because ∂xβ−γ τ (x, ·) ∈ S ) implies
Z
n
|xα ∂ β (Tσ (ϕ))(x)| = (2π)− 2 xα ∂ β eix·ξ τ (x, ξ) dξ
Rn
Z
n
= (2π)− 2 xα ∂xβ (eix·ξ τ (x, ξ)) dξ
Rn
Z X β 
−n α
= (2π) 2 x (iξ)γ eix·ξ ∂xβ−γ τ (x, ξ)) dξ
Rn γ≤β γ

n
X β  Z
≤ (2π)− 2 ξ γ (ix)α eix·ξ ∂xβ−γ τ (x, ξ)) dξ (32)
γ Rn
γ≤β

n
X β  Z
= (2π)− 2 ξ γ ∂ξα eix·ξ ∂xβ−γ τ (x, ξ)) dξ
γ Rn
γ≤β

n
X β  Z
= (2π)− 2 eix·ξ ∂ξα (ξ γ ∂xβ−γ τ (x, ξ)) dξ
γ Rn
γ≤β
X β  Z
−n
≤ (2π) 2 |∂ξα (ξ γ ∂xβ−γ τ (x, ξ))| dξ.
γ Rn
γ≤β

28
Since ξ γ ∈ S |γ| and τ ∈ m∈R S m , it follows by Proposition 5.10 that ∂ξα (ξ γ ∂xβ−γ τ (x, ξ)) ∈
T
m
⊆ S −n−1 . So there is a constant Cα,β,γ such that
T
m∈R S

|∂ξα (ξ γ ∂xβ−γ τ (x, ξ))| ≤ Cα,β,γ (1 + |ξ|)−n−1 .

Then by (32),
X β  Z
−n
α β
|x ∂ (Tσ (ϕ))(x)| ≤ (2π) 2 Cα,β,γ (1 + |ξ|)−n−1 dξ < ∞.
γ Rn
γ≤β

Thus xα ∂ β (Tσ (ϕ)) is bounded for all multi-indices α and β. This together with the proof that
Tσ ϕ ∈ C ∞ (Rn ) shows that Tσ ϕ ∈ S , and therefore Tσ maps S into S . That Tσ is a linear map is
obvious from the definition. It is also obvious from the definition that Tσ is linear in σ, i.e

Tλ1 σ+λ2 τ = λ1 Tσ + λ2 Tτ (33)

for arbitrary symbols σ, τ and scalars λ1 , λ2 ∈ C.

The symbol defining a pseudo-differential operator is unique, meaning that if two pseudo-
differential operators are equal, then their corresponding symbols are equal.
Proposition 8.2. Suppose σ and τ are symbols such that Tσ = Tτ . Then σ = τ .
Proof. If Tσ = Tτ , then
Z Z
eix·ξ σ(x, ξ)ϕ̂(ξ) dξ = eix·ξ τ (x, ξ)ϕ̂(ξ) dξ
Rn Rn

for all x ∈ Rn and ϕ ∈ S . Replacing ϕ by ϕ̌ ∈ S in the above equality, it follows that

Z Z
ix·ξ
e σ(x, ξ)ϕ(ξ) dξ = eix·ξ τ (x, ξ)ϕ(ξ) dξ
Rn Rn

for all x ∈ Rn and ϕ ∈ S . For fixed x ∈ Rn , we see that the above expressions are just the
tempered distributions corresponding to eix·ξ σ(x, ξ) and eix·ξ τ (x, ξ) (which are tempered functions
in the ξ-variable since they are bounded by C(1 + |ξ|)m by the definition of a symbol), so it follows
by Proposition 7.5 that σ = τ .
Normal partial differential operators are pseudo-differential operators under certain conditions as
the following proposition shows. This is actually what motivates the definition of pseudo-differential
operators, since pseudo-differential operators can be thought of as a generalization of partial differ-
ential operators.
Proposition 8.3. Let P (x, D) = |α|≤m aα (x)Dα be a partial differential operator where all the
P

coefficients aα ∈ C ∞ (Rn ) and the derivatives ∂ β aα are bounded

Pfor all multi-indices β. Then P (x, D)
is also a pseudo-differential operator, with symbol P (x, ξ) = |α|≤m aα (x)ξ α .
Proof. We already proved in Proposition 5.11 that P (x, ξ) is a symbol. By the Fourier inversion
theorem and Proposition 6.6. ii),
Z Z
n n
X
(2π)− 2 eix·ξ P (x, ξ)ϕ̂(ξ) dξ = (2π)− 2 aα (x) eix·ξ ξ α ϕ̂(ξ) dξ
Rn |α|≤m Rn
Z
n
X X
= (2π)− 2 aα (x) eix·ξ D
[ α ϕ(ξ) dξ = aα (x)Dα ϕ(x) = P (x, D)ϕ(x),
|α|≤m Rn |α|≤m

which proves that P (x, D) is a pseudo-differential operator with symbol P (x, ξ).

29
8.2 Composition of pseudo-differential operators
It follows from the definition of a pseudo-differential operator that the sum of two pseudo-differential
operators is also a pseudo-differential operator, since Tσ + Tτ = Tσ+τ . Another question is if the
composition Tσ Tτ is a pseudo-differential operator. The composition is well defined since a pseudo-
differential operator maps S into itself, but the question is if there exist a symbol λ such that
Tλ = Tσ Tτ ? The answer is yes, but calculating the symbol λ is not something that can be done
directly from the definition in most cases. It is also not necessarily the case that λ = στ as one
might guess by comparing with the case of addition.
To prove that the set of pseudo-differential operators is closed under composition we first partition
σ into symbols with compact support by taking the partition of unity (ψk ) from Proposition 4.3 and
defining
σk (x, ξ) = ψk (ξ)σ(x, ξ), k ∈ N. (34)
P∞
Then σ = k=0 σk . It turns out that treating each σk separately and then T summing up will be easier.
Notice that since ψk ∈ S , we have by Proposition 5.10 that σk ∈ m∈R S m so σk is a Schwartz
function in the ξ-variable, so one can take the Fourier transform or inverse Fourier transform in the
ξ-variable, as is done in the following lemma.
Lemma 8.4. Suppose σ ∈ S m . Let
Z
−n
Kk (x, z) = (2π) 2 eiz·ξ σk (x, ξ) dξ, k ∈ N.
Rn

Then for all N ∈ N and multi-indices α and β, there exists a constant CN,α,β > 0 such that
Z
|z|N |∂xα ∂zβ Kk (x, z)| dz ≤ CN,α,β 2(m+|β|−N )k (35)
Rn

for all k ∈ N.
Proof. We will start by estimating (35) but with its integrand squared because this will allow us to
use Plancherel’s formula to simplify the integral. First notice that by Proposition 3.15,
X
|z|2N |∂xα ∂zβ Kk (x, z)|2 ≤ nN |z γ ∂xα ∂zβ Kk (x, z)|2 . (36)
|γ|=N

We estimate the integral of each term in the right hand side of (36). To this end, let

W0 = {ξ ∈ Rn | |ξ| ≤ 2}
Wk = {ξ ∈ Rn | 2k−1 ≤ |ξ| ≤ 2k+1 }, k > 0.

ThenTthe support of σk in the ξ-variable is in Wk by the definition of σk and Proposition 4.3. Since
σk ∈ m∈R S m , all its partial derivatives are bounded so for each multi-index α,
Z Z
α iz·ξ
|∂x (e σk (x, ξ))| dξ ≤ |∂xα σk (x, ξ)| dξ
Rn Rn
Z Z (37)
= |∂xα σk (x, ξ)| dξ ≤ Cα dξ < ∞,
Wk Wk

where Cα = sup(x,ξ)∈R2n |∂xα σk (x, ξ))| and the compactness of Wk gives a finite integral. (37) shows
that the conditions in Theorem 3.8 holds so that we can take the x-derivative inside the integral:
Z
n
∂xα Kk (x, z) = (2π)− 2 eiz·ξ ∂xα σk (x, ξ) dξ.
Rn

30
This shows that for fixed x, ∂xα Kk (x, ·) is the inverse Fourier transform of ∂xα σk (x, ·). So by Propo-
sition 6.6 iv),
∂zβ ∂xα Kk (x, z) = i|β| (ξ β ∂xα σk )∨ (x, z),
where the inverse Fourier transform is taken in the ξ-variable. Now by Proposition 6.6 iii),

z γ ∂zβ ∂xα Kk (x, z) = i|γ|+|β| (∂ξγ (ξ β ∂xα σk ))∨ (x, z).

So by Plancherel’s formula,
Z Z
|z γ ∂xα ∂ξβ Kk (x, z)|2 dz = |∂ξγ (ξ β ∂xα σk (x, ξ))|2 dξ. (38)
Rn Rn

By Leibniz’s formula, the definition of σk and Proposition 4.4,

|∂ξγ (ξ β ∂xα σk (x, ξ))| = |∂ξγ (ξ β ∂xα σ(x, ξ)ψk (ξ))|

X γ 
≤ |∂ξδ (ξ β ∂xα σ(x, ξ))||∂ξγ−δ ψk (ξ)|
δ (39)
δ≤γ
X γ 
≤ |∂ξδ (ξ β ∂xα σ(x, ξ))|Cγ,δ 2−k(|γ|−|δ|) ,
δ
δ≤γ

where Cγ,δ is the constant from Proposition 4.4. Since ξ β ∈ S |β| and ∂xα σ ∈ S m , we have that
ξ β ∂xα σ ∈ S m+|β| . So by the definition of a symbol, there exists a constant Cα,β,δ such that

|∂ξδ (ξ β ∂xα σ(x, ξ))| ≤ Cα,β,δ (1 + |ξ|)m+|β|−|δ| .

Hence for all (x, ξ) ∈ Rn × Wk ,

X γ 
|∂ξγ (ξ β ∂xα σk (x, ξ))| ≤ Cα,β,δ (1 + |ξ|)m+|β|−|δ| Cγ,δ 2−k(|γ|−|δ|)
δ
δ≤γ
X γ 
≤ Cα,β,δ (1 + 2k+1 )m+|β|−|δ| Cγ,δ 2−k(|γ|−|δ|)
δ
δ≤γ
X γ 
≤ Cα,β,δ (2k+2 )m+|β|−|δ| Cγ,δ 2−k(|γ|−|δ|)
δ
δ≤γ
X γ 
= Cα,β,δ Cγ,δ 2(k+2)(m+|β|−|δ|)−k(|γ|−|δ|)
δ
δ≤γ
X γ 
= Cα,β,δ Cγ,δ 22(m+|β|−|δ|) 2k(m+|β|−|γ|)
δ
δ≤γ
0
= Cα,β,γ 2k(m+|β|−|γ|) ,

where
X γ 
0
Cα,β,γ = Cα,β,δ Cγ,δ 22(m+|β|−|δ|) .
δ
δ≤γ

This implies that

Z Z
|∂ξγ (ξ β ∂xα σk (x, ξ))|2 dξ = |∂ξγ (ξ β ∂xα σk (x, ξ))|2 dξ
Rn Wk
Z (40)
0 2 2k(m+|β|−|γ|)
≤ (Cα,β,γ ) 2 dξ.
Wk

31
The volume of every ball with radius r in n dimensions can be written as Cn rn where Cn > 0 is
some constant independent of r. Since Wk is contained in a ball with radius 2k+1 , we get that
Z
dξ ≤ Cn 2(k+1)n .
Wk

So by (40) Z
|∂ξγ (ξ β ∂xα σk (x, ξ))|2 dξ ≤ Cα,β,γ
00
2k(n+2(m+|β|−|γ|)) , (41)
Rn
where
00 0
Cα,β,γ = (Cα,β,γ )2 Cn 2n .
Now by (36), (38) and (41),
Z X
00
|z|2N |∂xα ∂zβ Kk (x, z)|2 dz ≤ nN Cα,β,γ 2k(n+2(m+|β|−|γ|)) ≤ DN,α,β 2k(n+2(m+|β|−N )) , (42)
n
R |γ|=N

where X
00
DN,α,β = nN Cα,β,γ .
|γ|=N

We apply the Cauchy-Schwartz inequality for integrals and use (42) to obtain

Z Z !1/2 Z !1/2
N
|z| |∂xα ∂zβ Kk (x, z)| dz ≤ |z| 2N
|∂xα ∂zβ Kk (x, z)|2 dz dz
|z|≤2−k |z|≤2−k |z|≤2−k (43)
1/2
≤ DN,α,β 2k(n/2+m+|β|−N ) Cn1/2 2−kn/2 = EN,α,β 2(m+|β|−N )k .

By the the Cauchy-Schwartz inequality again, (42) (but with n + N in place of N ) and the same
kind of change of variable as in (16),
Z Z
N α β
|z| |∂x ∂z Kk (x, z)| dz = |z|N +n |∂xα ∂zβ Kk (x, z)||z|−n dz
|z|≥2−k |z|≥2−k
Z !1/2 Z !1/2
2(n+N ) −2N
≤ |z| |∂xα ∂zβ Kk (x, z)|2 dz |z| dz
|z|≥2−k |z|≥2−k
Z Z ∞ 1/2
1/2
≤ DN +n,α,β 2k(n/2+m+|β|−N −n) r −2n n−1
r dr dω
Sn−1 2−k (44)
Z 1/2   1/2
−n ∞
1/2 r
= DN +n,α,β 2k(n/2+m+|β|−N −n) dω
Sn−1 n 2−k
Z 1/2
1/2 1 k(n/2+m+|β|−N −n) kn/2
= DN +n,α,β dω 2 2
Sn−1 n1/2
= FN,α,β , 2k(m+|β|−N ) ,

where Sn−1 ⊆ Rn is the unit sphere. By (43) and (44),

Z
|z|N |∂xα ∂zβ Kk (x, z)| dz ≤ (EN,α,β + FN,α,β )2k(m+|β|−N ) ,
Rn

which completes the proof.

32
 Now we prove the main theorem of this section, that the set of pseudo-differential operators is
closed under composition.
Theorem 8.5. Suppose σ ∈ S m1 and τ ∈ S m2 . Then the composition Tσ Tτ is a pseudo-differential
operator and its corresponding symbol λ is in S m1 +m2 and has the asymptotic expansion
X (−i)|µ|
λ∼ (∂ξµ σ)(∂xµ τ ),
µ
µ!

or more precisely, for all integers N > 0

X (−i)|µ| µ
λ− (∂ξ σ)(∂xµ τ ) ∈ S m1 +m2 −N .
µ!
|µ|<N

Proof. We again use (34) to partition σ into a sequence (σk ) of functions with compact support in
the ξ-variable. By definition,
Z
−n
(Tσk Tτ ϕ)(x) = (2π) 2 eix·ξ σk (x, ξ)Td
τ ϕ(ξ) dξ
Rn
Z Z
= (2π)−n eix·ξ σk (x, ξ) e−iξ·y (Tτ ϕ)(y) dy dξ
Rn Rn
Z Z
−n
= (2π) ei(x−y)·ξ σk (x, ξ) dξ(Tτ ϕ)(y) dy
Rn Rn
Z
−n
= (2π) 2 Kk (x, x − y)(Tτ ϕ)(y) dy,
Rn
where Kk is the same function as in Proposition 8.4. Then
Z
−n
(Tσk Tτ ϕ)(x) = (2π) 2 Kk (x, x − y)(Tτ ϕ)(y) dy
Rn
Z Z
= (2π)−n Kk (x, x − y) eiy·η τ (y, η)ϕ̂(η) dη dy
Rn Rn
Z Z
−n
= (2π) eiy·η Kk (x, x − y)τ (y, η) dy ϕ̂(η) dη
Rn Rn
Z Z
−n
= (2π) e ix·η
e−i(x−y)·η Kk (x, x − y)τ (y, η) dy ϕ̂(η) dη
n n
ZR ZR
−n
= (2π) e ix·η
e−iz·η Kk (x, z)τ (x − z, η) dz ϕ̂(η) dη,
Rn Rn

where the last equality follows from the change of variable z = x − y. Now define
Z
n
λk (x, η) = (2π)− 2 e−iz·η Kk (x, z)τ (x − z, η) dz. (45)
Rn

λk is well defined since Kk (x, ·) is the inverse Fourier transform of σk (x, ·) ∈ S and hence Kk (x, ·) ∈
S so the above integral is convergent. Then
Z
n
(Tσk Tτ ϕ)(x) = (2π)− 2 eix·η λk (x, η)ϕ̂(η) dη. (46)
Rn

Now define
∞
X
λ(x, η) = λk (x, η). (47)
k=0

33
 We will prove that λ ∈ S m1 +m2 and it is the symbol of Tσ Tτ and has the asymptotic expansion
given in the theorem. It also needs to be checked that λ is well defined, i.e the sum given in (47) is
convergent. By applying Theorem 3.9 in the x-variable, we obtain
X (−z)µ
τ (x − z, η) = ∂xµ τ (x, η) + RN (x, z, η) (48)
µ!
|µ|<N

for each positive integer N where RN is defined by

X (−z)µ Z 1
RN (x, z, η) = N (1 − θ)N −1 ∂xµ τ (x − θz, η) dθ. (49)
µ! 0
|µ|=N

Then by (45) and (48),

Z
n
λk (x, η) = (2π)− 2 e−iz·η Kk (x, z)τ (x − z, η) dz
Rn
X 1 Z
−n
= (2π) 2 µ
∂ τ (x, η) e−iz·η Kk (x, z)(−z)µ dz (50)
µ! x R n
|µ|<N
Z
n
+ (2π)− 2 e−iz·η Kk (x, z)RN (x, z, η) dz.
Rn
n
For fixed x ∈ R , Kk (x, ·) is the inverse Fourier transform of σk (x, ·). So Proposition 6.6 iii) implies
that Kk (x, z)(−z)µ is the inverse Fourier transform of (−i)µ ∂ µ σk (x, ·). Since
Z
−n
(2π) 2 e−iz·η Kk (x, z)(−z)µ dz
Rn

is the Fourier transform of Kk (x, z)(−z)µ in the z-variable, it follows by the Fourier inversion theorem
that Z
−n
(2π) 2 e−iz·η Kk (x, z)(−z)µ dz = (−i)µ ∂ηµ σk (x, η). (51)
Rn
So by (50) and (51),
X (−i)µ
λk (x, η) = ∂ µ τ (x, η)∂ηµ σk (x, η) + TN,k (x, η), (52)
µ! x
|µ|<N

where Z
−n
TN,k (x, η) = (2π) 2 e−iz·η Kk (x, z)RN (x, z, η) dz. (53)
Rn
Before continuing the main proof, we will prove some estimates for RN and TN,k .
Lemma 8.6. For arbitrary multi-indices α,β and γ and integers N > 0, there is a constant CN,α,β,γ
such that X
|∂zγ ∂xα ∂ηβ RN (x, z, η)| ≤ CN,α,β,γ (1 + |η|)m2 −|β| |z|N −|δ| . (54)
δ≤γ
|δ|≤N

Proof. By the definition of RN ,

X (−z)µ Z 1
∂xα ∂ηβ RN (x, z, η)
=N (1 − θ)N −1 ∂xα ∂ηβ (∂xµ τ (x − θz, η)) dθ
µ! 0
|µ|=N
(55)
X (−z)µ Z 1
=N (1 − θ)N −1 ∂xµ+α ∂ηβ τ (x − θz, η) dθ.
µ! 0
|µ|=N

34
By Leibniz’s formula and Proposition 3.13,

|∂zγ ∂xα ∂ηβ RN (x, z, η)|

X X γ  ∂ δ (−z)µ Z 1
= N z
(1 − θ)N −1 ∂zγ−δ (∂xµ+α ∂ηβ τ (x − θz, η)) dθ
δ µ! 0
|µ|=N δ≤γ
X X γ µ |(−z)µ−δ | Z 1
≤N δ! |1 − θ|N −1 ∂xµ+α+γ−δ ∂ηβ τ (x − θz, η)||θ||γ|−|δ| dθ (56)
δ δ µ! 0
|µ|=N δ≤γ
δ≤µ
X X γ µ |z||µ|−|δ| Z 1
≤N δ! |∂xµ+α+γ−δ ∂ηβ τ (x − θz, η)| dθ.
δ δ µ! 0
|µ|=N δ≤γ
δ≤µ

Since τ ∈ S m2 , there are constants Cα,β,γ,δ,µ such that

X X γ µ |z|N −|δ| Z 1

|∂zγ ∂xα ∂ηβ RN (x, z, η)| ≤ N δ! Cα,β,γ,δ,µ (1 + |η|)m2 −|β| dθ
δ δ µ! 0
|µ|=N δ≤γ
δ≤µ
X X γ µ |z|N −|δ|
≤N δ! Cα,β,γ,δ,µ (1 + |η|)m2 −|β|
δ δ µ! (57)
|µ|=N δ≤γ
δ≤µ
X X γ µ |z|N −|δ|
m2 −|β|
≤ (1 + |η|) N δ! Cα,β,γ,δ,µ .
δ δ µ!
|µ|=N δ≤γ
|δ|≤N

The last equality follows from the fact that |µ| = N and δ ≤ µ implies |δ| ≤ N and we define
Cα,β,γ,δ,µ = 0 if δ  µ. Now we change the order of summation and get
X X γ µ δ!
|∂zγ ∂xα ∂ηβ RN (x, z, η)| ≤ (1 + |η|)m2 −|β| N |z|N −|δ| Cα,β,γ,δ,µ
δ δ µ!
δ≤γ |µ|=N
|δ|≤N
X (58)
≤ (1 + |η|)m2 −|β| N 0
CN,α,β,γ,δ |z|N −|δ| ,
δ≤γ
|δ|≤N

where
X γ µ δ!
0
CN,α,β,γ,δ = Cα,β,γ,δ,µ .
δ δ µ!
|µ|=N

Now let
00 0
CN,α,β,γ = sup CN,α,β,γ,δ .
δ≤γ
|δ|≤N

Then by (58),
X
00
|∂zγ ∂xα ∂ηβ RN (x, z, η)| ≤ CN,α,β,γ (1 + |η|)m2 −|β| N |z|N −|δ| ,
δ≤γ
|δ|≤N

which proves the lemma.

35
Lemma 8.7. For all multi-indices α and β and integers M ≥ |β| and N > 0, there exists a constant
Cα,β,M,N > 0 such that for all k ∈ N,

|∂xα ∂ηβ TN,k (x, η)| ≤ Cα,β,M,N 2(m1 +2M −N )k (1 + |η|)m2 −M .

Proof. Let the multi-indices α and β and integers M ≥ |β| and N > 0 be arbitrary. By using
Leibniz’s formula twice,
Z
∂xα ∂ηβ (e−iz·η Kk (x, z)RN (x, z, η)) dz
Rn
 
Z X β 
= ∂xα Kk (x, z) ∂ δ (e−iz·η )∂ηβ−δ RN (x, z, η) dz
Rn δ η
δ≤β

Z X 
β δ −iz·η α
= ∂ (e )∂x (Kk (x, z)∂ηβ−δ RN (x, z, η)) dz (59)
Rn δ η
δ≤β

Z X β  X α 
= ∂ηδ (e−iz·η ) ∂ γ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz
δ
Rn δ≤β γ x
γ≤α
X X αβ  Z

≤ ∂ηδ (e−iz·η )∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz .
γ δ R n
γ≤α δ≤β

Notice that
n
X n
X
(1 + |η|2 )(e−iz·η ) = (1 + ηj2 )e−iz·η = (1 − ∂z2j )e−iz·η = (1 − ∆z )e−iz·η ,
j=1 j=1

so
(1 + |η|2 )M +|δ|−|β| (e−iz·η ) = (1 − ∆z )M +|δ|−|β| e−iz·η (60)
by induction. Here the assumption that M ≥ |β| is used, since we need that M + |δ| − |β| is a
nonnegative integer for (60) to work. Note that (1 + |η|2 )M +|δ|−|β| is a 2(M + |δ| − |β|)-degree
differential operator, so we can write
X
(1 − ∆z )M +|δ|−|β| = a ∂z . (61)
||≤2(M +|δ|−|β|)

The coefficients a ∈ C are not important. Now for each γ ≤ α and δ ≤ β, (60) implies
Z
∂ηδ (e−iz·η )∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz
R n
Z
= (1 + |η|2 )−(M +|δ|−|β|) (1 + |η|2 )M +|δ|−|β| ∂ηδ (e−iz·η )∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz (62)
R n
Z
= (1 + |η|2 )−(M +|δ|−|β|) (1 − ∆z )M +|δ|−|β| ∂ηδ (e−iz·η )∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz .
Rn

Note that the expression

∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η)
is a Schwarz function in the z-variable since Kk (x, ·) is the inverse Fourier transform of σk (x, ·) ∈ S
and for fixed x, ξ ∈ Rn , all the derivatives of RN are bounded by a polynomial in z by Lemma 8.6.

36
Therefore we can use integration by parts from Proposition 5.8 and by combining that with (61),
we obtain
Z
(1 − ∆z )M +|δ|−|β| ∂ηδ (e−iz·η )∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz
Rn
Z
= (−iz)δ (e−iz·η )(1 − ∆z )M +|δ|−|β| (∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η)) dz
Rn
Z (63)
≤ |z||δ| |(1 − ∆z )M +|δ|−|β| (∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η))| dz
Rn
X Z
≤ |a | |z||δ| |∂z (∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η))| dz.
||≤2(M +|δ|−|β|) Rn

By Leibniz’s formula and Lemma 8.6, we see that for all || ≤ 2(M + |δ| − |β|),

|∂z (∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η))|

X 
≤ |∂zζ ∂xγ Kk (x, z)||∂z−ζ ∂ηβ−δ ∂xα−γ RN (x, z, η))|
ζ
ζ≤
  (64)
X    X
m2 −|β|+|γ|
ζ γ
|z|N −|ρ| 

≤  ζ |∂z ∂x Kk (x, z)|CN,α,β,γ,δ,,ζ (1 + |η|)

.
ζ≤ ρ≤−ζ
|ρ|≤N

We have by Proposition 8.4, (64) and the fact that || ≤ 2(M + |δ| − |β|) ≤ 2M (since δ ≤ β),
Z
|z||δ| |∂z (∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η))| dz
Rn
X  
 X Z 
m2 −|β|+|γ| N +|δ|−|ρ| ζ γ
≤ CN,α,β,γ,δ,,ζ (1 + |η|) |z| |∂z ∂x Kk (x, z)| dz
ζ Rn
ζ≤ ρ≤−ζ
|ρ|≤N
X  
 X 
m2 −|β|+|γ| (m1 +|ζ|−N −|δ|+|ρ|)k
≤ CN,α,β,γ,δ,,ζ (1 + |η|) DN,δ,ρ,ζ,γ 2
ζ (65)
ζ≤ ρ≤−ζ
|ρ|≤N
X  
 X 
≤ CN,α,β,γ,δ,,ζ (1 + |η|)m2 −|β|+|γ| DN,δ,ρ,ζ,γ 2(m1 +|ζ|−N −|δ|+||−|ζ|)k
ζ
ζ≤ ρ≤−ζ
|ρ|≤N

≤ EN,α,β,γ,δ, 2(m1 −N −|δ|+||)k (1 + |η|)m2 −|β|+|γ|

≤ EN,α,β,γ,δ, 2(m1 −N +2M )k (1 + |η|)m2 −|β|+|γ| ,

where
X   X


EN,α,β,γ,δ, = CN,α,β,γ,δ,,ζ DN,δ,ρ,ζ,γ .
ζ
ζ≤ ρ≤−ζ
|ρ|≤N

So by (63) and (65),

Z
(1 − ∆z )M +|δ|−|β| ∂ηδ (e−iz·η )∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz
Rn (66)
≤ FN,M,α,β,γ,δ 2(m1 −N +2M )k (1 + |η|)m2 −|β|+|δ| ,

37
where X
FN,M,α,β,γ,δ = |a |EN,M,α,β,γ,δ, .
||≤2(M +|δ|−|β|)

By Proposition 3.12, there is a constant GM,β,δ > 0 such that

(1 + |η|2 )−(M +|δ|−|β|) ≤ GM,β,δ (1 + |η|)−2M −2|δ|+2|β| . (67)

Then by (62), (66) and (67),

Z
∂ηδ (e−iz·η )∂xγ Kk (x, z)∂ηβ−δ ∂xα−γ RN (x, z, η) dz
Rn
≤ (1 + |η|2 )−(M +|δ|−|β|) FN,M,α,β,γ,δ 2(m1 −N +2M )k (1 + |η|)m2 −|β|+|δ|
(68)
≤ GM,β,δ FN,M,α,β,γ,δ 2(m1 −N +2M )k (1 + |η|)−2M −2|δ|+2|β|) (1 + |η|)m2 −|β|+|δ|
= GM,β,δ FN,M,α,β,γ,δ 2(m1 −N +2M )k (1 + |η|)m2 −2M +|β|−|δ|
≤ GM,β,δ FN,M,α,β,γ,δ 2(m1 −N +2M )k (1 + |η|)m2 −M ,

since |β| ≤ M . Now by (59) and (68),

Z
−n
(2π) 2 ∂xα ∂ηβ (e−iz·η Kk (x, z)RN (x, z, η)) dz
Rn (69)
(m1 −N +2M )k m2 −M
≤ CN,M,α,β 2 (1 + |η|) ,

where
X X αβ 
−n
CN,M,α,β = (2π) 2 GM,β,δ FN,M,α,β,γ,δ .
γ δ
γ≤α δ≤β

For any multi-indices α and β, pick an integer M such that M ≥ m2 and M ≥ |β|. Then the
integral in (69) is bounded by a constant, so by Theorem 3.8, TN,k ∈ C ∞ (Rn ) and integration and
differentiation can be interchanged, i.e
Z
−n
α β
∂x ∂η TN,k (x, η) = (2π) 2 ∂xα ∂ηβ (e−iz·η Kk (x, z)RN (x, z, η)) dz. (70)
Rn

By combining (69) with (70), we see that the lemma holds.

We now return to the proof of Theorem 8.5. For arbitrary multi-indices α and β, we can choose
an integer Mα,β such that m2 − Mα,β ≤ 0 and Mα,β ≥ |β|. We can then choose another positive
integer Nα,β such that m1 + 2Mα,β − Nα,β ≤ −1. Then Lemma 8.7 implies that
∞
X ∞
X
| ∂xα ∂ηβ TNα,β ,k (x, η)| ≤ |∂xα ∂ηβ TNα,β ,k (x, η)|
k=0 k=0
∞
X
≤ Cα,β,Mα,β ,Nα,β 2(m1 +2Mα,β −Nα,β )k (1 + |η|)m2 −Mα,β
k=0 (71)
X∞
≤ Cα,β,Mα,β ,Nα,β 2−k
k=0
≤ 2Cα,β,Mα,β ,Nα,β .

38
 P∞
(52) and the fact that σ = σk implies that
k=0
 
∞ ∞
X X X (−i)µ µ µ 
∂xα ∂ηβ TNα,β ,k = ∂xα ∂ηβ λk − ∂xα ∂ηβ ∂ τ ∂ σk
µ! x η
k=0 k=0 |µ|<Nα,β
∞ µ ∞
X X (−i) µ µ X
= ∂xα ∂ηβ λk − ∂xα ∂ηβ ∂ τ∂ σk (72)
µ! x η
k=0 |µ|<Nα,β k=0
∞
X X (−i)µ µ µ
= ∂xα ∂ηβ λk − ∂xα ∂ηβ ∂ τ ∂ σ.
µ! x η
k=0 |µ|<Nα,β

By the way we the partition of unity in Proposition 4.3 was constructed, around each point (x, η)
only a finite number of the σk are nonzero. It follows that the expression
∞
X X (−i)µ µ µ
∂xα ∂ηβ ∂ τ ∂ σk
µ! x η
k=0 |µ|<Nα,β

only contains a finite number nonzero terms around each point, so the interchange of Psum and
∞
derivatives in (72)Pis justified. By choosing α = β = 0 in (71) and (72), we see that k=0 k is
λ
∞ P∞
convergent, since k=0 TN0,0 ,k is convergent. Therefore λ = k=0 λk is well defined. By (71) and
the Weierstrass M-test,
∞
X
∂xα ∂ηβ TNα,β ,k
k=0

is uniformly convergent for all multi-indices α and β, so

∞
X
∂xα ∂ηβ λk
k=0

is also uniformly convergent by (72). Then by Theorem 3.7, λ ∈ C ∞ (Rn ) and differentiation and
integration can be interchanged, i.e
∞
X ∞
X
∂xα ∂ηβ λk = ∂xα ∂ηβ λk . (73)
k=0 k=0

Now we will prove that λ ∈ S m1 +m2 : For any multi-indices α and β and integer N > 0, choose
a positive integer Mα,β,N such that Mα,β,N ≥ |β| and m2 − Mα,β,N ≤ m1 + m2 − N − |β|. We then
choose another positive integer Kα,β,N such that m1 + 2Mα,β,N − Kα,β,N < −1 and Kα,β,N > N .
Then by (73), (72) and Lemma 8.7 there is a constant Cα,β,N such that
 
µ ∞
X (−i) X X (−i)µ µ µ
∂xα ∂ηβ λ − ∂xµ τ ∂ηµ σ  = ∂xα ∂ηβ λk − ∂xα ∂ηβ ∂ τ∂ σ
µ! µ! x η
|µ|<Kα,β,N k=0 |µ|<Kα,β,N
∞
X ∞
X
≤ |∂xα ∂ηβ TKα,β,N ,k | ≤ Cα,β,N 2(m1 +2Mα,β,N −Kα,β,N )k (1 + |η|)m2 −Mα,β,N
(74)
k=0 k=0
X∞ ∞
X
≤ Cα,β,N 2−k (1 + |η|)m1 +m2 −N −|β| = Cα,β,N (1 + |η|)m1 +m2 −N −|β| 2−k
k=0 k=0
m1 +m2 −N −|β|
= 2Cα,β,N (1 + |η|) ,

39
so
X (−i)µ µ µ
λ− ∂ τ ∂ σ ∈ S m1 +m2 −N . (75)
µ! x η
|µ|<Kα,β,N

By Proposition 5.10 ii)-iv), we see that

∂xµ τ ∂ηµ σ ∈ S m1 +m2 −|µ| ⊆ S m1 +m2

for all multi-indices µ. Therefore,

X (−i)µ µ µ
∂ τ ∂ σ ∈ S m1 +m2 . (76)
µ! x η
|µ|<Kα,β,N

So by (75) and (76), λ ∈ S m1 +m2 .

Next we will prove that λ has the given asymptotic expansion. Notice that by Proposition 5.10
ii)-iv) again,
∂xµ τ ∂ηµ σ ∈ S m1 +m2 −|µ| ⊆ S m1 +m2 −N
for all multi-indices µ that satisfy |µ| ≥ N . Hence,
X (−i)µ µ µ
∂ τ ∂ σ ∈ S m1 +m2 −N (77)
µ! x η
N ≤|µ|<Kα,β,N

(recall that Kα,β,N was chosen so that N < Kα,β,N ). Then by (75) and (77),
X (−i)µ
λ− ∂µτ ∂µσ
µ! x η
|µ|<N
(78)
X (−i)µ µ µ X (−i)µ µ µ
=λ− ∂ τ∂ σ + ∂ τ ∂ σ ∈ S m1 +m2 −N .
µ! x η µ! x η
|µ|<Kα,β,N N ≤|µ|<Kα,β,N

Since N was an arbitrary positive integer, this proves that λ has the asymptotic expansion given in
the theorem. Finally, we will prove that Tσ Tτ = Tλ . By the definition of λ, σk and (46),
Z
n
(Tλ ϕ)(x) = (2π)− 2 eix·η λ(x, η)ϕ̂(η) dη
Rn
Z ∞ ∞ Z
−n −n
X X
ix·η
= (2π) 2 e (λk (x, η))ϕ̂(η) dη = (2π) 2 eix·η λk (x, η)ϕ̂(η) dη
Rn k=0 k=0 Rn
∞ ∞ Z
−n
X X
= (Tσk Tτ ϕ)(x) = (2π) 2 eix·ξ σk (x, ξ)Td
τ ϕ(ξ) dξ
k=0 k=0 Rn
Z ∞ Z
n n
X
= (2π)− 2 eix·ξ (σk (x, ξ))Td −
τ ϕ(ξ) dξ = (2π) 2 eix·ξ σ(x, ξ)Td
τ ϕ(ξ) dξ = (Tσ Tτ ϕ)(x),
Rn k=0 Rn

which completes the proof.

40
8.3 Elliptic pseudo-differential operators
In this section we will prove that pseudo-differential operators with an elliptic symbol have a so
called approximate inverse. We say that a pseudo-differential operator is m − elliptic if its symbol
is m − elliptic.

Theorem 8.8. Suppose σ ∈ S m is an m-elliptic symbol. Then there exists a symbol τ ∈ S −m such
that

Tτ Tσ = I + R (79)
Tσ Tτ = I + S, (80)

where I is the identity map on S and R and S are pseudo-differential operators with symbols in
k
T
k∈R S . The operator Tτ is called an approximate inverse of Tσ .

Proof. The symbol τ will be constructed by an asymptotic expansion as follows: Let τ0 ∈ S −m be

the function from Proposition 5.16, i.e
(
ψ(ξ)
, |ξ| ≥ R
τ0 (x, ξ) = σ(x,ξ)
0, |ξ| < R.

For integers k > 0 define τk recursively by

X (−i)|µ| µ
τk = −τ0 (∂x σ)(∂ξµ τj ). (81)
µ!
|µ|+j=k
0<|µ|<k
0≤j<k

First we show that τk ∈ S −m−k for all k ∈ N by induction. The base case k = 0 holds by Proposition
5.16. If k > 0 and τj ∈ S −m−j for all j < k, then for all multi-indices µ that satisfy |µ| + j = k,
Proposition 5.10 implies
(∂xµ σ)(∂ξµ τj ) ∈ S m−m−j−|µ| = S −k .
Therefore,
X (−i)|µ| µ
(∂x σ)(∂ξµ τj ) ∈ S −k .
µ!
|µ|+j=k
0≤j<k

So by (81), τk ∈ S −m−k . By Proposition 5.13 there is a symbol τ ∈ S −m such that

∞
X
τ∼ τj .
j=0

Our goal is now to prove that τ satisfies (79) and (80). By the definition of asymptotic expansion,
k−1
X
τk0 := τ − τj ∈ S −m−k (82)
j=0

for any integer k > 0. Now let λ ∈ S 0 be the symbol of the composition Tτ Tσ which by Theorem
8.5 satisfies
X (−i)|µ|
λ− (∂xµ σ)(∂ξµ τ ) ∈ S −k (83)
µ!
|µ|<k

41
for any integer k > 0. By (82),

X (−i)|µ| X k−1
X (−i)|µ| X (−i)|µ|
(∂xµ σ)(∂ξµ τ ) − 1 = (∂xµ σ)(∂ξµ τj ) − 1 + (∂xµ σ)(∂ξµ τk0 ). (84)
µ! j=0
µ! µ!
|µ|<k |µ|<k |µ|<k

For the first sum in the right hand side of (84), we can write

X k−1
X (−i)|µ| k−1
X X k−1
X (−i)|µ|
(∂xµ σ)(∂ξµ τj ) − 1 = στ0 − 1 + σ τj + (∂xµ σ)(∂ξµ τj ). (85)
µ! µ!
|µ|<k j=0 j=1 0<|µ|<k j=0

In the last above sum, we sum over all multi-indices µ that satisfy 0 < |µ| < k and j that satisfy
0 ≤ j ≤ k − 1. We can partition this set of µ and j into those that satisfy |µ| + j = l where 0 < l < k
and those that satisfy |µ| + j ≥ k. So the sum can be rewritten as

X k−1
X (−i)|µ| k−1
X X (−i)|µ| µ X (−i)|µ| µ
(∂xµ σ)(∂ξµ τj ) = (∂x σ)(∂ξµ τj ) + (∂x σ)(∂ξµ τj ).
j=0
µ! µ! µ!
0<|µ|<k l=1 |µ|+j=l |µ|+j≥k
0<|µ|<k 0<|µ|<k
0≤j<k 0≤j<k
(86)
Then by (85), (86) and (81),

X k−1
X (−i)|µ|
(∂xµ σ)(∂ξµ τj ) − 1
µ!
|µ|<k j=0
k−1 k−1
X X X (−i)|µ| µ X (−i)|µ| µ
= στ0 − 1 + σ τj + (∂x σ)(∂ξµ τj ) + (∂x σ)(∂ξµ τj )
j=1
µ! µ!
l=1 |µ|+j=l |µ|+j≥k
0<|µ|<k 0<|µ|<k
0≤j<k 0≤j<k
k−1
X X (−i)|µ| µ
= στ0 − 1 − στ0 (∂x σ)(∂ξµ τj )
µ!
l=1 |µ|+j=l
0<|µ|<k
0≤j<k (87)
k−1
X X (−i)|µ| µ X (−i)|µ| µ
+ (∂x σ)(∂ξµ τj ) + (∂x σ)(∂ξµ τj )
µ! µ!
l=1 |µ|+j=l |µ|+j≥k
0<|µ|<k 0<|µ|<k
0≤j<k 0≤j<k
 
 k−1

 X X (−i)|µ| µ
 X (−i)|µ|
= (στ0 − 1) 1 − µ
(∂x σ)(∂ξ τj ) + (∂xµ σ)(∂ξµ τj ).
 
 µ!  µ!
 l=1 |µ|+j=l  |µ|+j≥k
0<|µ|<k 0<|µ|<k
0≤j<k 0≤j<k

For |ξ| ≥ 2R, σ(x, ξ)τ0 (x, ξ) = 1 by the definition of τ0 so this means the first term
T in the last
equality in (87) has compact support in the ξ-variable and therefore is an element of k∈R S k . Also
∂xµ σ ∈ S m and ∂ξµ τj ∈ S −m−j−|µ| ⊆ S −m−k for j + |µ| ≥ k so the second sum in the last equality in
(87) is an element of S −k . So (87) implies that

X k−1
X (−i)|µ|
(∂xµ σ)(∂ξµ τj ) − 1 ∈ S −k . (88)
j=0
µ!
|µ|<k

42
From (82), (∂xµ σ)(∂ξµ τk0 ) ∈ S −k−|µ| ⊆ S −k for all multi-indices µ so

X (−i)|µ|
(∂xµ σ)(∂ξµ τk0 ) ∈ S −k . (89)
µ!
|µ|<k

By combining (88) and (89) with (84), we see that

X (−i)|µ|
(∂xµ σ)(∂ξµ τ ) − 1 ∈ S −k , (90)
µ!
|µ|<k

and by summing (90) with (83),

λ − 1 ∈ S −k . (91)
Now k was an arbitrary positive integer so
\
λ−1∈ S −k . (92)
0<k∈Z

If f ∈ 0<k∈Z S −k , then for every k 0 ∈ R we can pick a positive integer k such that −k < k 0 and
T
0
then f ∈ S −k ⊆ S k Thus f ∈ k∈R S k , so this argument shows that 0<k∈Z S −k = k∈R S k . Then
T T T
(92) implies that \
λ−1∈ Sk. (93)
k∈R

Since the symbol of the identity I on S is 1 ∈ S 0 , (93) and the definition of λ implies that the
pseudo-differential operator R defined by

R := Tτ Tσ − I = Tλ − I (94)

has symbol in k∈R S k . This proves (79). To prove (80), we first prove the existence of another
T
symbol κ ∈ S −m such that
Tσ Tκ = I + Q, (95)
k
T
where Q has symbol in k∈R S and then we will show that this implies (80). The construction of
κ and the proof of (95) is very similar to the above proof of (79) but with slight modifications so
the identical parts will not be repeated: Define κ0 = τ0 . For integers k > 0, define κk recursively by
X (−i)|µ| µ
κk = −κ0 (∂x κj )(∂ξµ σ). (96)
µ!
|µ|+j=k
0<|µ|<k
0≤j<k

The definition is nearly the same as τk but we switch the places of ∂x and ∂ξ . As with τk we prove
that κk ∈ S −m−k by induction. The base case k = 0 holds. So assume k > 0 and that the result
is true for all j < k. Then (∂xµ κj ) ∈ S −m−j and (∂ξµ σ) ∈ S m−|µ| so (∂xµ κj )(∂ξµ σ) ∈ S −j−|µ| = S −k
whenever |µ| + j = k. So by (96), κk ∈ S −m−k which P∞ completes the induction proof. Then by
Proposition 5.13, let κ ∈ S −m be the symbol with k=0 κk as its asymptotic expansion and let
π ∈ S 0 be the symbol of the product Tσ Tκ . Now fix any integer k > 0. By the definition of
asymptotic expansion,
k−1
X
κ0k := κ − κj ∈ S −m−k , (97)
j=0

43
and by Theorem 8.5,
X (−i)|µ|
π− (∂xµ κ)(∂ξµ σ) ∈ S −k . (98)
µ!
|µ|<k

By (97),

X (−i)|µ| X k−1
X (−i)|µ| X (−i)|µ|
(∂xµ κ)(∂ξµ σ) − 1 = (∂xµ κj )(∂ξµ σ) − 1 + (∂xµ κ0k )(∂ξµ σ). (99)
µ! j=0
µ! µ!
|µ|<k |µ|<k |µ|<k

As was done with τ , we can rewrite the first sum in the last equality of (99):

X k−1
X (−i)|µ|
(∂xµ κj )(∂ξµ σ) − 1
µ!
|µ|<k j=0
k−1 k−1
X X X (−i)|µ| µ X (−i)|µ| µ
= κ0 σ − 1 + κj σ + (∂x κj )(∂ξµ σ) + (∂x κj )(∂ξµ σ)
j=1
µ! µ!
l=1 |µ|+j=l |µ|+j≥k
0<|µ|<k 0<|µ|<k
0≤j<k 0≤j<k
k−1
X X (−i)|µ| µ
= κ0 σ − 1 − κ0 σ (∂x κj )(∂ξµ σ)
µ!
l=1 |µ|+j=l
0<|µ|<k
0≤j<k (100)
k−1
X X (−i)|µ| µ X (−i)|µ| µ
+ (∂x κj )(∂ξµ σ) + (∂x κj )(∂ξµ σ)
µ! µ!
l=1 |µ|+j=l |µ|+j≥k
0<|µ|<k 0<|µ|<k
0≤j<k 0≤j<k
 
 k−1

 X X (−i)|µ| µ 
 X (−i)|µ|
= (κ0 σ − 1) 1 − µ
(∂x κj )(∂ξ σ) + (∂xµ κj )(∂ξµ σ).

 µ!  µ!
 l=1 |µ|+j=l  |µ|+j≥k
0<|µ|<k 0<|µ|<k
0≤j<k 0≤j<k

The above expression is in S −k by the same kind of argument that was made for (87). And
X (−i)|µ|
(∂xµ κ0k )(∂ξµ σ) ∈ S −k (101)
µ!
|µ|<k

by (97) and the same kind of argument that was made for (89). So by combining (100) and (101)
with (99),
X (−i)|µ|
(∂xµ κ)(∂ξµ σ) − 1 ∈ S k . (102)
µ!
|µ|<k

Then by summing (102) with (98), we have π − 1 ∈ S k . Since k > 0 was an arbitrary integer, this
implies \
π−1∈ Sk. (103)
k∈R

Similarly to the previous proof of (79), this implies (95). Now by (79) and (95),
Tσ Tτ = Tσ Tτ I = Tσ Tτ (Tσ Tκ − Q) = Tσ Tτ Tσ Tκ − Tσ Tτ Q = Tσ (I + R)Tκ − Tσ Tτ Q
(104)
= Tσ Tκ + Tσ RTκ − Tσ Tτ Q = I + Q + Tσ RTκ − Tσ Tτ Q = I + S,

44
where
S = Q + Tσ RTκ − Tσ Tτ Q. (105)
k
T
By Theorem (8.5), the composition of a pseudo-differential operator with symbol in k∈R S with
any other pseudo-differential operator yields an operator with symbol in k∈R S k . Therefore S has
T

symbol in k∈R S k since R and Q have their symbols in k∈R S k . This proves (80).
T T

T Approximate
k
inverses of elliptic pseudo-differential operators are unique modulo
k∈R S :
Proposition 8.9. Suppose Tσ is anT elliptic pseudo-differential operator and Tτ and Tκ are approx-
imate inverses of Tσ . Then τ − κ ∈ k∈R S k .
Proof. By the definition of an approximate inverse given in Theorem 8.8, there are pseudo-differential
operators R and S with symbols in k∈R S k such that
T

Tσ Tτ = I + R
Tκ Tσ = I + S
Therefore,
Tτ −κ = Tτ − Tκ = ITτ − Tκ I = (Tκ Tσ − S)Tτ − Tκ (Tσ Tτ − R) = Tκ R − STτ
and Tκ R − STτ has its symbol in k∈R S k by Theorem 8.5 and because R and S have their symbols
T

in k∈R S k . Hence τ − κ ∈ k∈R S k .

T T

By Theorem 8.8, it can be of interest to know when a pseudo-differential operator is elliptic. So

in the following propositions we show some ways to determine if an operator is elliptic other than
using the definition directly. For example, in the following proposition we prove that the composition
of two elliptic operators is elliptic.
Proposition 8.10. If Tσ is m1 -elliptic and Tτ is m2 -elliptic, then Tσ Tτ is m1 + m2 -elliptic.
Proof. By Theorem 8.5, the symbol λ ∈ S m1 +m2 for the composition Tσ Tτ satisfies
λ − στ ∈ S m1 +m2 −1 .
So if we let r = λ − στ , then there is a constant D > 0 such that
|r(x, ξ)| ≤ D(1 + |ξ|)m1 +m2 −1 . (106)
m1 m2
Because σ ∈ S and τ ∈ S , there are constants C1 , C2 > 0 and R > 0 such that
|σ(x, ξ)| ≥ C1 (1 + |ξ|)m1 , |ξ| ≥ R (107)
m2
|τ (x, ξ)| ≥ C2 (1 + |ξ|) , |ξ| ≥ R. (108)
Hence (106), (107), (108) and the reverse triangle inequality imply that
|λ(x, ξ)| = |σ(x, ξ)τ (x, ξ) + r(x, ξ)| ≥ |σ(x, ξ)||τ (x, ξ)| − |r(x, ξ)|
≥ C1 C2 (1 + |ξ|)m1 +m2 − D(1 + |ξ|)m1 +m2 −1 (109)
−1 m1 +m2
≥ (C1 C2 − D(1 + |ξ|) )(1 + |ξ|) , |ξ| ≥ R.
0
Since C1 , C2 , D > 0 we can choose a constant R ≥ R such that
C1 C2 − D(1 + R0 )−1 > 0.
Then by (109),
|λ(x, ξ)| ≥ (C1 C2 − D(1 + |ξ|)−1 )(1 + |ξ|)m1 +m2
≥ (C1 C2 − D(1 + R0 )−1 )(1 + |ξ|)m1 +m2 , |ξ| ≥ R0 ,
which shows that λ is m1 + m2 -elliptic.

45
 By the following proposition, one only needs to looks at the highest degree terms in a partial
differential operator to determine if it is elliptic.
α
P
Proposition 8.11. Let P (x, D) = |α|≤m aα (x)D be a partial differential operator where all
∞ n β
P aα ∈ C (R ) and the derivatives ∂ aα are bounded for all multi-indices β. Let
the coefficients
Pm (x, ξ) = |α|=m aα (x)ξ α . Then P (x, D) is m-elliptic if and only if Pm (x, ξ) is m-elliptic.

Proof. =⇒ : Assume P (x, D) is m-elliptic. Its symbol is P (x, ξ) = |α|≤m aα (x)ξ α so there exist
P
C > 0 and R > 0 such that
|P (x, ξ)| ≥ C(1 + |ξ|)m , |ξ| ≥ R. (110)
Since |α|≤m−1 aα (x)ξ α ∈ S m−1 , there exist D > 0 such that
P

X
aα (x)ξ α ≤ D(1 + |ξ|)m−1 , ξ ∈ Rn . (111)
|α|≤m−1

By the reverse triangle inequality, (110) and (111) imply that

X X
|Pm (x, ξ)| = P (x, ξ) − aα (x)ξ α ≥ |P (x, ξ)| − aα (x)ξ α
|α|≤m−1 |α|≤m−1 (112)
D
≥ (C − )(1 + |ξ|)m , |ξ| ≥ R.
1 + |ξ|

Choose R0 ≥ R such that C − D

1+R0 > 0 and let C 0 = C − D
1+R0 Then by (112),

D
|Pm (x, ξ)| ≥ (C − )(1 + |ξ|)m ≥ C 0 (1 + |ξ|)m , |ξ| ≥ R0 .
1 + |ξ|

Thus Pm (x, ξ) is m-elliptic, since Pm (x, ξ) ∈ S m also holds.

⇐= : Assume Pm (x, ξ) is m-elliptic. Then there exist C > 0 and R > 0 such that

|Pm (x, ξ)| ≥ C(1 + |ξ|)m , |ξ| ≥ R.

As before, there exist D > 0 such that (111) holds. So by the reverse triangle inequality,

X X
|P (x, ξ)| = Pm (x, ξ) + aα (x)ξ α ≥ |Pm (x, ξ)| − aα (x)ξ α
|α|≤m−1 |α|≤m−1
D
≥ (C − )(1 + |ξ|)m , |ξ| ≥ R.
1 + |ξ|

In the same way as the converse proof, this shows that P (x, ξ) is m-elliptic.

For partial differential operators with constant coefficients, one only needs to look at the roots
to determine if it is elliptic.

Proposition 8.12. Let P (D) = |α|≤m aα Dα where m ≥ 1 and all coefficients aα ∈ C are constant.
P

Then P (D) is m-elliptic if and only if 0 is the only root of Pm (ξ).

46
 aα ξ α is m-elliptic if and only if 0
P
Proof. By Proposition 8.11, it suffices to prove Pm (ξ) = |α|=m
is the only root of Pm (ξ).

=⇒ : Assume Pm (ξ) is m-elliptic. Then there exist C > 0 and R > 0 such that
|Pm (ξ)| ≥ C(1 + |ξ|)m , |ξ| ≥ R.
In particular, Pm (ξ) 6= 0 for |ξ| ≥ R. So if ξ 6= 0, then choosing r > 0 such that |rξ| ≥ R implies
that Pm (rξ) 6= 0. But
X X X
Pm (rξ) = aα (rξ)α = aα r|α| ξ α = rm aα ξ α = rm Pm (ξ),
|α|=m |α|=m |α|=m

so Pm (ξ) 6= 0 as well. Therefore Pm has no nonzero roots. And 0 is clearly a root since m ≥ 1 so
Pm only contains positive powers of ξ.

⇐= : Suppose 0 is the only root of Pm (ξ). Since the unit sphere in Rn is compact, every continuous
function has a minimum there so we can let C = min|ξ|=1 |P (ξ)|. Then C > 0 by assumption. As
before, Pm (rξ) = rm Pm (ξ) for all r > 0. So for all |ξ| ≥ 1,
ξ |ξ|m C
|P (ξ)| = |ξ|m P ( ) ≥ C|ξ|m = C (1 + |ξ|)m = (1 + |ξ|)m
|ξ| (1 + |ξ|) m (1 + |ξ|−1 )m
C
≥ (1 + |ξ|)m ,
2m
which proves that Pm (ξ) is m-elliptic.
If an elliptic symbol σ(x, ξ) is independent of the x-variable, then we can find an explicit formula
for the symbol of the approximate inverse instead of just an asymptotic expansion. An example of
when this is useful is if we have an elliptic partial differential operator with constant coefficients
(which we just classified in Proposition 8.12).
Proposition 8.13. Suppose that σ ∈ S m is an m-elliptic symbol which is independent of the x-
variable. Let τ ∈ S −m be the symbol defined by Proposition 5.16, i.e
(
ψ(ξ)
, |ξ| ≥ R
τ (ξ) = σ(ξ)
0, |ξ| < R.

Then Tτ is an approximate inverse for Tσ .

Proof. Since both τ and σ are independent of x, all their partial derivatives with respect to x are
0. So by Theorem 8.5, the asymptotic expansion of the symbol λ1 for Tσ Tτ and the symbol λ2 for
Tτ Tσ is στ for both. Hence, \
λ1 − στ, λ2 − στ ∈ Sk.
k∈R

For |ξ| ≥ 2RT we have σ(ξ)τ (ξ) = ψ(ξ) = 1 so the function 1 − στ has compact support, and is
therefore in k∈R S k . So for i = 1, 2,
\
λi − 1 = (λi − στ ) + (στ − 1) ∈ Sk.
k∈R

Thus,
Tσ Tτ = Tλ1 = I + R
Tτ Tσ = Tλ2 = I + S,

47
 S k . This
T
where R and S are pseudo-differential operators with symbols λ1 − 1 and λ2 − 1 in k∈R
proves that Tτ is an approximate inverse for Tσ .

48
8.4 Pseudo-differential operators on Lp (Rn )
Now we will prove that under ceratin conditions, a pseudo-differential operator Tσ can be extended
from an operator on S to an operator on Lp (Rn ). However, the proof will use Lemma 8.15 and
8.16 whose proofs require more advanced distribution theory and are not included in this text.

Theorem 8.14. For any σ ∈ S 0 and p > 1, Tσ : S → S can be extended to a bounded linear
operator Tσ : Lp (Rn ) → Lp (Rn )
Since S is a dense subspace of Lp (Rn ) which is a Banach space, it suffices by Theorem 3.6 to
prove that the operator Tσ : S → S is bounded in Lp -norm. So the goal is to prove that there
exists a constant C > 0 such that

kTσ ϕkp ≤ Ckϕkp , ∀ϕ ∈ S . (113)

The first lemma needed is essentially a special case of (113) when σ only depends on ξ. It is
a modification of a theorem stated in [6, Theorem 11.8, p. 77] but was originally proved by L.
Hörmander in [1, Theorem 2.5].

Lemma 8.15. Suppose σ ∈ S 0 is a symbol depending only on the ξ-variable and k is an integer
such that k > n2 and p > 1. If Cσ > 0 is a constant such that for all multi-indices |α| ≤ k,

|∂ α σ(ξ)| ≤ Cσ (1 + |ξ|)−|α| , (114)

then there exists a constant D > 0 such that

kTσ ϕkp ≤ Cσ Dkϕkp .

Remark. The constant Cσ in Lemma 8.15 is indexed by the symbol σ to signify the dependence on
it. In most other theorems in this text, the constants also depend on the specific functions involved
but in this lemma the constant D is actually independent of the symbol σ, as long as it satisfies the
assumptions given. For this reason Cσ is indexed so that it is clear what depends on σ and what
does not.
The existence of a constant Cσ which satisfy (114) is trivial. Just take Cσ = max|α|≤k (Cα ) where
Cα are the constants satisfying
|∂ α σ(ξ)| ≤ Cα (1 + |ξ|)−|α| ,
given in the definition of a symbol.
This lemma allows us to under write a pseudo-differential operator without taking the Fourier
transform of ϕ. A proof can be found in [6, Lemma 10.10, p. 84-85] but E. M. Stein provides more
detailed proof in [5, Section VI.4, p. 241-245].
Lemma 8.16. Let σ ∈ S 0 . Then there exists a function K : Rn × (Rn \ {0}) such that
Z
n
(Tσ ϕ)(x) = (2π)− 2 K(x, x − z)ϕ(z) dz
Rn

for all ϕ ∈ S and x ∈ Rn such that ϕ is 0 in an open neighborhood around x. Moreover, for every
integer k > n there exists a constant Ck > 0 such that

|K(x, z)| ≤ Ck |z|−k , (x, z) ∈ Rn × (Rn \ {0}).

The next lemma can also be thought of a special case of (113) when σ has compact support in
the x-variable. In the proof of Theorem 8.5 the method was to partition σ into a family of functions

49
with compact support in ξ. Here we will do something similar but with x instead. The partition is
constructed as follows: For each m ∈ Zn , let Qm be the hypercube defined by
1
Qm = {x ∈ Rn | |xi − mi | ≤ ∀i ∈ {1, . . . , n}}. (115)
2
By Proposition 4.2, we can take a function ψ ∈ C0∞ (Rn ) equal to 1 on Q0 and define for each
m ∈ Zn ,
σm (x, ξ) = ψ(x − m)σ(x, ξ).
Then σm ∈ S 0 by Proposition 5.10.ix) and has compact support in the x-variable and is equal to σ
on Qm × Rn We also note that (Qm )m∈Z n is a cover of Rn .
Lemma 8.17. Let σ ∈ S 0 . Then there exist a constant C > 0, such that
Z
|(Tσ ϕ)(x)|p dx ≤ Ckϕkpp , ∀ϕ ∈ S , ∀m ∈ Zn .
Qm

Proof. First notice that by the definition of σm ,

Z Z
n n
(Tσm ϕ)(x) = (2π)− 2 eix·ξ ψ(x − m)σ(x, ξ)ϕ̂(ξ) dξ = (2π)− 2 eix·ξ ψ(x − m)σ(x, ξ)ϕ̂(ξ) dξ
Rn Rn
Z
−n
= ψ(x − m)(2π) 2 eix·ξ σ(x, ξ)ϕ̂(ξ) dξ = ψ(x − m)(Tσ ϕ)(x).
Rn
(116)
Since ψ(x − m) = 1 for all x ∈ Qm , it follows that
Z Z Z Z
|(Tσ ϕ)(x)|p dx = |ψ(x − m)(Tσ ϕ)(x)|p dx = |(Tσm ϕ)(x)|p dx ≤ |(Tσm ϕ)(x)|p dx.
Qm Qm Qm Rn
(117)
Because σm has compact support in the x-variable, σm (·, ξ) ∈ S for each ξ ∈ Rn so we can take
the Fourier transform and inverse Fourier transform of σm in the x-variable. Hence by the Fourier
inversion theorem,
Z
−n
(Tσm ϕ)(x) = (2π) 2 eix·ξ σm (x, ξ)ϕ̂(ξ) dξ
Rn
Z Z Z Z
= (2π)−n eix·ξ eix·η σ̂m (η, ξ) dη ϕ̂(ξ) dξ = (2π)−n eix·η eix·ξ σ̂m (η, ξ)ϕ̂(ξ) dξ dη.
Rn Rn Rn Rn
(118)
We will apply Lemma 8.15 to the function σ̂m (η, ·) but first the assumptions of that lemma needs
to be checked. We saw before that σm ∈ S 0 and σm has compact support in x which we will denote
by Sm . Hence for all multi-indices α, ∂ξα σm ∈ S −|α| ⊆ S 0 and ∂ξα σm and has compact support in x
contained in Sm . Therefore ∂ξα σm is bounded so
Z Z
sup |∂ξα σm (x, ξ)| dx ≤ sup |∂ξα σm (x, ξ)| dx < ∞.
ξ∈Rn Rn (x,ξ)∈Rn ×Rn Sm

Then by Theorem 3.8, σ̂m (η, ·) ∈ C ∞ (Rn ) for all η ∈ Rn and

Z Z
n n
∂ξα σ̂m (η, ξ) = (2π)− 2 ∂ξα e−iη·x σm (x, ξ) dx = (2π)− 2 e−iη·x ∂ξα σm (x, ξ) dx. (119)
Rn Rn

For all multi-indices α and β, we can since σ ∈ S 0 and ψ ∈ C0∞ (Rn ) choose constants Cα,β and
Dα such that
|∂xβ ∂ξα σ(x, ξ)| ≤ Cα,β (1 + |ξ|)−|α|
|∂xβ ψ(x)| ≤ Dβ .

50
Therefore Leibniz’s formula implies
|∂xβ ∂ξα σm (x, ξ)| = |∂xβ (ψ(x − m)∂ξα σ(x, ξ))|
X β 
≤ |∂xγ ψ(x − m)|∂xβ−γ ∂ξα σ(x, ξ)|
γ (120)
γ≤β
X β 
≤ Dγ Cα,β−γ (1 + |ξ|)−|α| ≤ Cα,β0
(1 + |ξ|)−|α| ,
γ
γ≤β

where
X β 
0
Cα,β = Dγ Cα,β−γ .
γ
γ≤β
0
The important point here is that the constant Cα,β is independent of m ∈ Zn . If we just wanted an
arbitrary constant in the estimate (120) one could just have picked the one given in the definition
of a symbol, since σm ∈ S 0 .
Now by (119), (120) and Proposition 5.8
Z
−n
β α
|η ∂ξ σ̂m (η, ξ)| = (2π) 2 (−iη)β e−iη·x ∂ξα σm (x, ξ) dx
Rn
Z Z
β −iη·x α
= ∂x (e )∂ξ σm (x, ξ) dx = e−iη·x ∂xβ ∂ξα σm (x, ξ) dx
Rn Rn
Z Z (121)
β α β α
≤ |∂x ∂ξ σm (x, ξ)| dx = |∂x ∂ξ σm (x, ξ)| dx
Rn Sm
Z
0
≤ Cα,β (1 + |ξ|)−|α| dx.
Sm

The reason Proposition 5.8 could be used is because ∂xβ ∂ξα σm (·, ξ) has compact support and is
therefore in SR as mentioned before and all derivatives of e−iη·x are bounded by a polynomial.
In (121) , Sm dx is finite but it is also bounded by a constant independent of m. The reason is
because Sm , being the x-support of σm , is contained in the compact support of ψ(x − m), by the
definition of σm . But the support of ψ(x − m) is just a translation of the support of the function
ψ(x), so integrating over it gives the same result. Therefore
Z Z
dx ≤ dx < ∞, (122)
Sm S
00
where S is the support of ψ(x). By combining (121) and (122) we see that there is a constant Cα,β
independent of m such that
00
|η β ||∂ξα σ̂m (η, ξ)| ≤ Cα,β (1 + |ξ|)−|α| . (123)

Because β is arbitrary, η β in (123) may be replaced by an arbitrary polynomial P (η), since by the
00
triangle inequality |P (η)| is bounded by a sum β≤deg(P ) |aβ ||η β |. So then one just replaces Cα,β
P
in
00 2 n
P
(123) by Cα := β≤deg(P ) |aβ |Cα,β . In particular, for the polynomial (1 + |η| ) there is a constant
Cα such that
(1 + |η|2 )n |∂ξα σ̂m (η, ξ)| ≤ Cα (1 + |ξ|)−|α| ,
so
|∂ξα σ̂m (η, ξ)| ≤ Cα (1 + |ξ|)−|α| (1 + |η|2 )−n .
This proves that for each fixed η ∈ Rn , σ̂m (η, ·) ∈ S 0 . Let C = max|α|≤n Cα . Then

|∂ξα σ̂m (η, ξ)| ≤ C(1 + |ξ|)−|α| (1 + |η|2 )−n

51
for all |α| ≤ n. Then the assumptions of Lemma 8.15 holds for σ̂m (η, ·) so there exists a constant
D, independent of m and η (in fact Theorem 8.15 says that D is even independent of σ but here the
independence of m and η is sufficient) such that

kTσ̂m (η,·) ϕkp ≤ DC(1 + |η|2 )−n kϕkp , ∀ϕ ∈ S . (124)

By definition, Z
n
(Tσ̂m (η,·) ϕ)(x) = (2π)− 2 eix·ξ σ̂m (η, ξ)ϕ̂(ξ) dξ, (125)
Rn

so by (117), (118),(124), (125) and the Minkowski inequality,

Z 1/p Z 1/p
|(Tσ ϕ)(x)|p dx ≤ |(Tσm ϕ)(x)|p dx
Qm Rn
Z Z Z p 1/p
−n ix·η ix·ξ
= (2π) e e σ̂m (η, ξ)ϕ̂(ξ) dξ dη dx
Rn Rn Rn
Z Z p 1/p
n
= (2π)− 2 eix·η (Tσ̂m (η,·) ϕ)(x) dη dx
Rn Rn
Z Z 1/p
−n p
≤ (2π) 2 |(Tσ̂m (η,·) ϕ)(x)| dx dη
n Rn
ZR Z
n n
= (2π)− 2 kTσ̂m (η,·) ϕkp dη ≤ (2π)− 2 DC(1 + |η|2 )−n kϕkp dη
Rn Rn
Z
n
= (2π)− 2 DC (1 + |η|2 )−n dηkϕkp = C 0 kϕkp ,
Rn

(1 + |η|2 )−n dη is finite. This proves the lemma.

R
since the integral Rn

Lemma 8.18. For any x, y ∈ C and p > 0,

|x + y|p ≤ 2p (|x|p + |y|p ).

Proof. Let M = max(|x|, |y|). Then M p = max(|x|p , |y|p ) and M p ≤ |x|p + |y|p . By the triangle
inequality
|x + y|p ≤ (|x| + |y|)p ≤ (2M )p ≤ 2p (|x|p + |y|p ).

Now we turn to the proof of Theorem 8.14.

Proof. For each m ∈ Zn , let Q0m be the hypercube in Rn with center m and twice the length of Qm
from (115), so that it has side length 2. Take a function ψ ∈ C0∞ (Rn ) such that 0 ≤ ψ ≤ 1 and is equal
to 1 on Q00 and let Sm be the compact support of ψ(x − m). Then Qm ⊆ Q0m ⊆ Sm and the size of
Qm , Q0m and Sm is invariant of m ∈ Zn . Let ϕ ∈ S be arbitrary and define ϕm,1 (x) = ψ(x−m)ϕ(x)
and ϕm,2 (x) = (1 − ψ(x − m))ϕ(x). Then ϕ = ϕm,1 + ϕm,2 so Tσ ϕm,1 + Tσ ϕm,2 . Hence by Lemma
8.18 and Lemma 8.17 and the fact that ϕm,1 has its support in Sm ,
Z Z Z
p p p p
|Tσ ϕ(x)| dx ≤ 2 |Tσ ϕm,1 (x)| + 2 |Tσ ϕm,2 (x)|p dx
Qm Qm Qm
Z Z Z
p p p p p p p
≤ 2 Ckϕm,1 kp + 2 |Tσ ϕm,2 (x)| dx ≤ 2 C |ϕm,1 (x)| dx + 2 |Tσ ϕm,2 (x)|p dx (126)
Qm Sm Qm
Z Z
p p p
≤2 C |ϕ(x)| dx + 2 |Tσ ϕm,2 (x)|p dx.
Sm Qm

52
The last inequality follows from the fact that |ϕm,1 (x)| ≤ |ψ(x − m)||ϕ(x)| ≤ |ϕ(x)|. By definition,
ϕm,2 is 0 on Q0m . So for every x ∈ Qm and integer k > n it follows from Lemma 8.16 that there is
a constant Ck > 0 such that
Z Z
n n
|Tσ ϕm,2 (x)| = (2π)− 2 K(x, x − z)ϕm,2 (z) dz = (2π)− 2 K(x, x − z)ϕm,2 (z) dz
Rn Rn \Q0m
Z Z
n n
≤ (2π)− 2 |K(x, x − z)||ϕm,2 (z)| dz ≤ (2π)− 2 Ck |x − z|−k |ϕm,2 (z)| dz.
Rn \Q0m Rn \Q0m
(127)
For each x ∈ Qm and z ∈ Rn \ Q0m we have that |x − z| ≥ 21 since each of the coordinates in x and
z differ with more than 12 by the definition of Qm and Q0m . Therefore
√ √ k
( n + 1 + |x − z|)k n+1 √
k
= + 1 ≤ (2( n + 1) + 1)k := Dk . (128)
|x − z| |x − z|
So by (127) and (128),
|ϕm,2 (z)|
Z
n
|Tσ ϕm,2 (x)| ≤ (2π)− 2 Ck Dk √ dz. (129)
Rn \Q0m ( n + 1 + |x − z|)k

Notice that for any x ∈ Qm ,

√
v v
u n u n
uX uX 1 n
|x − m| ≤ t 2
(xi − mi ) ≤ t ( )2 ≤ . (130)
i=1 i=1
2 2

So for x ∈ Qm and z ∈ Rn \ Q0m , the triangle inequality implies

√
√
√ √ √ n n
n + 1 + |x − z| ≥ n + 1 + |m − z| − |m − x| ≥ n + 1 + |m − z| − = + 1 + |m − z|.
2 2
(131)
Hence by (129), (131) and Hölder’s inequality,
|ϕm,2 (z)|
Z
−n
|Tσ ϕm,2 (x)| ≤ (2π) Ck Dk
2 √
n
√ dz
k/2 ( n + 1 + |m − z|)k/2
2 + 1 + |m − z|)
Rn \Q0m (
2
!1/p Z !1/q
|ϕm,2 (z)|p
Z
−n 1
≤ (2π) 2 Ck Dk √
n
dz √
n
dz ,
pk/2 qk/2
2 + 1 + |m − z|) 2 + 1 + |m − z|)
Rn \Q0m ( Rn \Q0m (
(132)
where q > 1 satisfies p1 + 1q = 1 We already assumed that k > n but if we also choose k such that
qk/2 > n then (132) implies
!1/p
p
|ϕ
Z
m,2 (z)|
|Tσ ϕm,2 (x)| ≤ Ck0 √
n
dz , (133)
pk/2
2 + 1 + |m − z|)
Rn \Q0m (

where !1/q
Z
−n 1
Ck0 = (2π) 2 Ck D k √
n
dz , (134)
Rn \Q0m ( 2 + 1 + |m − z|)qk/2
which is now finite. The constant Ck0 is independent of m because
Z Z
1 1
√
n
dz = √
n
dz.
qk/2 qk/2
2 + 1 + |z − m|) 2 + 1 + |z|)
Rn \Q0m ( Rn \Q00 (

53
Since our x ∈ Qm was arbitrary, it follows from (133) that
|ϕm,2 (z)|p
Z Z Z
|Tσ ϕm,2 (x)|p dx ≤ Ck0p √
n
dz dx
pk/2
2 + 1 + |m − z|)
Qm Qm Rn \Q0m (
(135)
|ϕm,2 (z)|p |ϕm,2 (z)|p
Z Z Z
0p 0p
= Ck √
n
dz dx = Ck √
n
dz.
pk/2 pk/2
2 + 1 + |m − z|) 2 + 1 + |m − z|)
Rn \Q0m ( Qm Rn \Q0m (

Because (Ql )l∈Z n cover Rn , we obtain

|ϕm,2 (z)|p |ϕm,2 (z)|p
Z XZ
√
n
dz ≤ √
n
dz. (136)
Rn \Q0m ( 2 + 1 + |m − z|)pk/2 l∈Zn Ql ( 2 + 1 + |m − z|)pk/2

For z ∈ Ql , then similiarly to (131), the triangle inequality implies

√ √ √ √
n n n n
+ 1 + |m − z| ≥ + 1 + |m − l| − |z − l| ≥ + 1 + |m − l| − = 1 + |m − l|. (137)
2 2 2 2
Therefore (135), (136) and (137) imply
Z Z
1
|Tσ ϕm,2 (x)|p dx ≤ Ck0p
X
|ϕm,2 (z)|p dz
Qm l∈Zn
(1 + |m − l|)pk/2 Ql
Z (138)
1
≤ Ck0p
X
|ϕ(z)|p dz,
l∈Zn
(1 + |m − l|)pk/2 Ql

where the last inequality follows from the fact that

|ϕm,2 (z)| = |1 − ψ(z − m)||ϕ(z)| ≤ |ϕ(z)|.

So by (126) and (138)

Z Z Z
p p p p
|Tσ ϕ(x)| dx ≤ 2 C |ϕ(x)| dx + 2 |Tσ ϕm,2 (x)|p dx
Qm Sm Qm
Z
1
Z (139)
p 0p
X
p p
≤2 C |ϕ(x)| dx + 2 Ck pk/2
|ϕ(z)|p dz.
Sm l∈Zn
(1 + |m − l|) Q l

Therefore,
Z X Z
|Tσ ϕ(x)|p dx = |Tσ ϕ(x)|p dx
Rn m∈Z n Qm
X Z Z (140)
1
|ϕ(x)|p dx + 2p Ck0p
X X
≤ 2p C |ϕ(z)|p dz.
m∈Z n Sm m∈Z n l∈Zn
(1 + |m − l|)pk/2 Ql

The sets Sm are compact and cover Rn since Qm ⊆ Sm . Moreover, each Sm is just a translation of
S0 by m. Since S0 is bounded there exists a constant M > 0 such that |x| ≤ M for all x ∈ S0 . Then
|x − m| ≤ M for all x ∈ Sm . This implies that Sm and Sm0 are disjoint if |m0 − m| > 2M . Let K
be the number of points m0 ∈ Zn such that |m0 | ≤ 2M . Then for each fixed m ∈ Zn , at most K of
the Sm0 overlap with Sm , the ones that satisfy |m0 − m| ≤ 2M . This means that for each x ∈ Rn , x
is in at most K of the sets Sm .
The above argument implies that each x ∈ Rn is included in at most K of the integrals over Sm
in (140), so
X Z Z
p
|ϕ(x)| dx ≤ K |ϕ(x)|p dx = Kkϕkpp . (141)
m∈Z n Sm Rn

54
Now we turn to the second sum in (140). Notice that for each l ∈ Zn the change of variable m−l 7→ m
implies
X 1 X 1
pk/2
= pk/2
.
m∈Zn
(1 + |m − l|) m∈Zn
(1 + |m|)

Therefore,
Z Z
X X 1 X X 1
|ϕ(z)|p dz = |ϕ(z)|p dz
n
m∈Z l∈Z n
(1 + |m − l|)pk/2 Q l n
l∈Z m∈Z n
(1 + |m − l|)pk/2
Q l
! !
XZ X 1 X 1 XZ
= |ϕ(z)|p dz pk/2
= pk/2
|ϕ(z)|p dz (142)
l∈Z n Q l m∈Z n
(1 + |m − l|) m∈Z n
(1 + |m|) l∈Z n Ql
!Z !
X 1 X 1
= pk/2
|ϕ(z)|p dz = pk/2
kϕkpp .
m∈Zn
(1 + |m|) R n
m∈Zn
(1 + |m|)

If we take k large enough so that pk/2 > n, then

X 1
Ek := < ∞.
m∈Z n
(1 + |m|)pk/2

So by (140), (141) and (142),

Z
kTσ ϕkpp = |Tσ ϕ(x)|p dx ≤ (2p CK + 2p Ck0p Ek )kϕkpp .
Rn

This shows that Tσ is bounded in Lp -norm which is what we wanted to prove.

55
9 Sobolev spaces
Now we will give a brief introduction to Sobolev spaces, which will be used in the next section for
PDE:s.
For any s ∈ R, let σs be the function given by
σs (ξ) = (1 + |ξ|2 )−s/2 , ξ ∈ Rn .
It can be shown that σs ∈ S −s . Now let Js be the pseudo-differential operator with symbol σs . In
Theorem 8.14 we proved that some pseudo-differential operators can be extended to an operator on
Lp (Rn ). Here we will prove that the operator Js can be extended even further to an operator on
S 0 , the space of tempered distributions.
To construct this extension, we define σs u for each u ∈ S 0 as the tempered distribution given
by
(σs u)(ϕ) = u(σs ϕ), ϕ ∈ S .
Next we define Js as an operator S 0 → S 0 by
Js u = F −1 (σs Fu), u ∈ S 0, (143)
where F : S 0 → S 0 is the Fourier transform on S 0 given by Definition 7.6. We want to prove that
this actually is an extension of Js : S → S . Note that
Z Z Z 
−n ix·ξ
Js ϕ(x)ψ(x) dx = (2π) 2 e σs (ξ)ϕ̂(ξ) dξ ψ(x) dx
Rn n Rn
Z Z R  Z
−n ix·ξ
= (2π) 2 σs (ξ)ϕ̂(ξ) e ψ(x) dx dξ = σs (ξ)ϕ̂(ξ)ψ̌(ξ) dξ
Rn Rn Rn
Z Z  (144)
−n −iξ·x
= (2π) 2 σs (ξ) e ϕ(x) dx ψ̌(ξ) dξ
Rn Rn
Z Z  Z
−n −iξ·x
= (2π) 2 e σs (ξ)ψ̌(ξ) dξ ϕ(x) dx = σds ψ̌(x)ϕ(x) dx.
Rn Rn Rn

If we identify of Js ϕ and σd
s ψ̌ by their corresponding tempered distributions, then (144) shows that

−1
s ψ̌) = Fϕ(σs ψ̌) = (σs Fϕ)(ψ̌) = (F
Js ϕ(ψ) = ϕ(σd (σs Fϕ))(ψ).
Therefore,
Js ϕ = F −1 (σs Fϕ), ϕ ∈ S. (145)
By comparing (143) with (145), we see that that Js as defined by (143) is an extension of Js defined
as a pseudo-differential operator on S .
The following property of Js will be useful.
Proposition 9.1. For any s, t ∈ R, Js Jt = Js+t .
Proof. By (143),
Js Jt u = F −1 (σs F(F −1 (σt Fu))) = F −1 (σs (σt Fu)).
By the definition of σs it follows that σs σt = σs+t . Hence for any ϕ ∈ S ,
(σs (σt Fu))(ϕ) = (σt Fu)(σs ϕ) = Fu(σt σs ϕ) = Fu(σs+t ϕ) = (σs+t Fu)(ϕ),
so σs (σt Fu) = σs+t Fu. It follows that
Js Jt u = F −1 (σs+t Fu) = Js+t u,
which completes the proof.

56
 Now we can define the Sobolev space H s,p . We assume from now on that p > 1. Observe that
H s,p is also defined for p = 1, but as boundedness results for pseudo-differential operators in general
only holds for p > 1, the case p = 1 is ignored.
Definition 9.2. For any p > 1 and s ∈ R we define the Sobolev space H s,p to be the set of
distributions u ∈ S 0 satisfying J−s u ∈ Lp (Rn ). Define the norm on k · ks,p on H s,p by

kuks,p = kJ−s ukp , u ∈ Hs,p .

From the linearity of J−s and the fact that Lp (Rn ) is a vector space, it follows that H s,p is a
vector space as well. By the fact that σ0 = 1 and (143), it follows that J0 is the identity on S 0 . So
by definition, H 0,p = Lp (Rn ).

Proposition 9.3. For any s, t ∈ R, Jt maps H s,p onto H s+t,p . Moreover, it is an isometry in the
respective norms, i.e
kuks,p = kJt uks+t,p , u ∈ H s,p .
Proof. If u ∈ H s,p , then J−s u ∈ Lp (Rn ). Since J−s−t Jt u = J−s u, this shows that Jt u ∈ H s+t,p . We
also have that
kuks,p = kJ−s ukp = kJ−s−t Jt ukp = kJt uks+t,p .
so Jt is an isometry. To prove surjectivity, assume v ∈ H s+t,p . Then J−s−t v ∈ Lp (Rn ). Since
J−s J−t v = J−s−t v , this shows that J−t v ∈ H s,p . And Jt (J−t v) = v so we are done.
Proposition 9.4. H s,p is a Banach space.
Proof. By Proposition 9.3, Js is a linear isomorphism between H 0,p = Lp (Rn ) and H s,p and an
isometry. Since Lp (Rn ) is a Banach space, it follows that H s,p is as well.

Proposition 9.5. S is dense in H s,p .

Proof. Let u ∈ H s,p and  > 0 be arbitrary. Since J−s u ∈ Lp (Rn ) and S is dense in Lp (Rn ), we can
take ϕ ∈ S such that kJ−s u − ϕkp < . By Proposition 9.3, kJ−s u − ϕkp = kJs (J−s u − ϕ)ks,p =
ku − Js ϕks,p . So ku − Js ϕks,p < . which proves that S is dense in H s,p .
Proposition 9.6. If s ≤ t, then H t,p ⊆ H s,p .

Proof. Suppose u ∈ H t,p . Because s − t ≤ 0, Jt−s has symbol in S s−t ⊆ S 0 . By Theorem 8.14,
Jt−s maps Lp (Rn ) into itself. Since we have J−t u ∈ Lp (Rn ) by definition, it therefore follows that
J−s u = Jt−s J−t u ∈ Lp (Rn ), so u ∈ H s,p .
We can now extend pseudo-differential operators to H s,p . By the following proposition, Tσ can
s,p s,p
S S
be thought of as a linear map s∈R H → s∈R H .
Proposition 9.7. For any symbol σ ∈ S m and s ∈ R and p > 1, Tσ : S → S can be extended to
a bounded linear operator Tσ : H s,p → H s−m,p .
Proof. By Theorem 8.5, Jm−s Tσ Js has symbol in S s−m+m−s = S 0 . Therefore by Theorem 8.14, it
is a bounded linear operator in Lp -norm. So for any ϕ ∈ S ,

kTσ ϕks−m,p = kJm−s Tσ ϕkp = kJm−s Tσ Js J−s ϕkp

≤ CkJ−s ϕkp ≤ Ckϕks,p .

This shows that Tσ : S → S is a bounded linear operator with respect to the norms of H s,p and
H s−m,p . By Theorem 3.6 and Proposition 9.5, Tσ : S → S can be extended to a bounded linear
operator Tσ : H s,p → H s−m,p .

57
 Using the extension given by Proposition 9.7, we obtain two corollaries:
Corollary 9.7.1. If σ ∈ m∈Rn S m , then Tσ maps s∈R H s,p into s∈R H s,p
T S T

Corollary 9.7.2. For any symbol σ, Tσ maps s∈R H s,p into s∈R H s,p
T T

For p = 2 and s ≥ 0 we have an alternate definition of H s,2 .

Proposition 9.8. For any s ≥ 0,
 Z 
H = u ∈ L2 (Rn ) |
s,2 2 s
(1 + |ξ| ) |û(ξ)| dξ < ∞ 2
Rn

and Z 1/2
kuks,2 = (1 + |ξ|2 )s |û(ξ)|2 dξ .
Rn

Proof. If u ∈ H s,2 , then J−s u ∈ L2 (Rn ) and u ∈ L2 (Rn ) since s ≥ 0. By Plancherel’s formula,

kJ−s uk2 = kJ[

−s uk2 .

By the Fourier inversion theorem, the definition of a pseudo-differential operator and the fact that
(1 + |ξ|2 )s/2 is the symbol of J−s , we have
2 s/2
−s ϕ(ξ) = (1 + |ξ| )
J[ ϕ̂(ξ), ϕ ∈ S.

Since S is dense in L2 (Rn ) and the Fourier transform is a bounded and hence a continuous linear
operator on L2 (Rn ), it follows that
2 s/2
−s u(ξ) = (1 + |ξ| )
J[ û(ξ).

We then see that

Z 1/2
2 s 2
kuks,2 = kJ−s uk2 = kJ[
−s uk2 = (1 + |ξ| ) |û(ξ)| dξ < ∞,
Rn

because kJ−s uk2 < ∞. Conversely, if u ∈ L2 (Rn ) and (1 + |ξ|2 )s |û(ξ)|2 dξ < ∞, then similarly
R
Rn
Z 1/2
2 s 2
∞> (1 + |ξ| ) |û(ξ)| dξ = kJ[
−s uk2 = kJ−s uk2 ,
Rn

so u ∈ H s,2 by definition.
If s ≥ 0 is large enough, then every function in H s,2 is differentiable up to a certain order in the
following sense:
n
Proposition 9.9. If u ∈ H s,2 and s > 2 + k where k ∈ N, then there is a function v ∈ C k (Rn )
such that v = u almost everywhere and

lim v(x) = 0.
|x|→∞

Proof. Since u ∈ H s,2 , we have ũ := J−k u ∈ H s−k,2 . By Hölder’s inequality,

Z 1/2 Z 1/2
ˆ 1≤ 2 −s+k 2 s−k ˆ 2
kũk (1 + |ξ| ) dξ (1 + |ξ| ) |ũ(ξ)| dξ . (146)
Rn Rn

58
The first integral on the right hand side is convergent since s − k > n2 and the second is convergent
ˆ ∈ L1 (Rn ). Therefore ṽ := ũˇˆ
by Proposition 9.8. So ũ ∈ C(Rn ) by the Riemann-Lebesque lemma.
s−k,2 2 n ˇˆ
Since ũ ∈ H ⊂ L (R ), we have that ṽ = ũ = ũ almost everywhere by the Fourier inversion
theorem. So kṽ − ũks−k,2 = 0. From this it also follows that ṽ ∈ H s−k,2 since ũ ∈ H s−k,2 . Now let
v = Jk ṽ. Then
kv − uks,2 = kJk ṽ − Jk ũks,2 = kṽ − ũks−k,2 = 0,
since Jk is an isometry from H s−k,2 to H s,2 . So v = u almost everywhere. To prove that v ∈ C k (Rn ),
notice that for all multi-indices α that satisfy |α| ≤ k,
Z
n
|∂xα v(x)| = |∂xα Jk ṽ(x)| ≤ (2π)− 2 |∂xα eix·ξ |(1 + |ξ|2 )−k/2 |ṽ(ξ)|
ˆ dξ
R n
Z Z
n n
≤ (2π)− 2 |ξ||α| (1 + |ξ|2 )−k/2 |ṽ(ξ)|
ˆ dξ ≤ (2π)− 2 (1 + |ξ|)k (1 + |ξ|2 )−k/2 |ṽ(ξ)|
ˆ dξ
Rn Rn
Z
n n
≤ (2π)− 2 Ck ˆ
|ṽ(ξ)| dξ ≤ (2π)− 2 Ck kṽk1 < ∞.
Rn

The inequality kṽk1 < ∞ follows by the same reasoning as in (146) and the constant Ck comes from
Proposition 3.12. So all partial derivatives of v up to order k exist, hence v ∈ C k (Rn ) (The fact
that the above integral is convergent for all multi-indices |α| ≤ k shows that it was allowed to take
the derivative inside the integral). Since v ∈ H s,p ⊆ H s−k,p and is continuous, we have in a similar
way as before that v̂ ∈ L1 (Rn ) and v̂ˇ = v so by the Riemann-Lebesgue lemma,

lim v(x) = 0.
|x|→∞

59
10 Applications to elliptic pseudo-differential equations
We use the theory developed so far to prove two results for the solutions u to pseudo-differential
equations of the form Tσ u = f where f ∈ H s,p and σ ∈ S m . We say that u ∈T t∈R H t,p is
S
an exact solution if Tσ u = f and an approximate solution if Tσ u = f modulo t∈R H t,p , i.e.
t,p
T
Tσ u − f ∈ t∈R H .
The first theorem says that if we are given a solution, then we can conclude which Sobolev space
it is in. The second theorem proves the existence of approximate solutions.

Theorem 10.1. Suppose f ∈ H s,p and σ ∈ S m is an m-elliptic symbol. If u ∈ t∈R H t,p is an

S
approximate solution to Tσ u = f , then u ∈ H s+m .
Proof. By the definition of approximate solution, h := Tσ u − f ∈ t∈R H t,p . Since σ is m-elliptic,
T
there exists by Theorem 8.8 a symbol τ ∈ S −m and a pseudo-differential operator R with symbol in
k
T
k∈R S such that
Tτ Tσ = I + R.
Hence,
Tτ (f + h) = Tτ Tσ u = u + Ru.
By Theorem 9.7, Tτ (f + h) ∈ H s+m and by Corollary 9.7.1, Ru ∈ H s+m as well. So

u = Tτ (f + h) − Ru ∈ H s+m .

Since every exact solution is also approximate, Theorem 10.1 holds for exact solutions as well.
Theorem 10.2. Suppose f ∈ H s,p and σ ∈ S m is an m-elliptic symbol. Then there is an ap-
proximate
T solution u to the pseudo-differential equation Tσ u = f . Moreover, the solution is unique
modulo t∈R H t,p .
Proof. By Theorem 8.8, pick τ ∈ S −m such that

Tτ Tσ = I + R
Tσ Tτ = I + S,

where R and S are pseudo-differential operators with symbols in k∈R S k . We let u = Tτ f . Then
T
u ∈ H s+m by Theorem 9.7 and
Tσ u = Tσ Tτ f = f + Sf.
t,p
T
Since Sf ∈ t∈R H by Corollary 9.7.1, this shows that u is an approximate solution.
Now for any two approximate solutions u and v, we have that

Tσ u = f + g
Tσ v = f + h,

H t,p . Therefore,
T
where g, h ∈ t∈R

u = (Tτ Tσ − R)u = Tτ (f + g) − Ru = Tτ f + Tτ g − Ru = Tτ (Tσ v − h)

+ Tτ g − Ru = Tτ Tσ v − Tτ h + Tτ g − Ru = v + Rv − Tτ h + Tτ g − Ru.

H t,p , so v = u modulo H t,p .

T T
By Corollary 9.7.1 and 9.7.2 again, Rv, Ru, Tτ h, Tτ g ∈ t∈R t∈R

60
 By Theorem 10.1 and 10.2 weP can now draw some conclusions about solutions to PDE:s: If
f ∈ Lp (Rn ), p > 1 and P (x, D) = α≤m aα (x)Dα is an m-elliptic differential operator, (which are
classified by Proposition 8.11 and 8.12), then if u ∈ t∈R H t,p is a solution to P (x, D)u = f , we
S
also know that u ∈ H m,p ⊆ Lp (Rn ). Moreover, if m > n2 + k and p = 2, then by Proposition 9.9, we
know that the solution u is equal to a function in C k (Rn ) almost everywhere.
We also know that every
T elliptic PDE P (x, D)u = f has an approximate solution u ∈ Lp (Rn )
t,p
and it is unique modulo t∈R H . If p = 2, then by the definition of an approximate solution and
Proposition 9.9, we have that P (x, D)u and f differ by a function that is C ∞ (Rn ) almost everywhere
and converges to 0 at infinity. And if v and u are two approximate solutions, then v and u also differ
by a function that is C ∞ (Rn ) almost everywhere and converges to 0 at infinity.

References
[1] L. Hörmander. Estimates for translation invariant operators in Lp spaces. Acta Math., 104:93–
140, 1960.

[2] E. Kreyszig. Introductory functional analysis with applications. John Wiley & Sons, New York-
London-Sydney, 1978.
[3] W. Rudin. Principles of mathematical analysis. McGraw-Hill Book Co., New York-Auckland-
Düsseldorf, third edition, 1976. International Series in Pure and Applied Mathematics.

[4] W. Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, third edition, 1987.
[5] E. M. Stein. Harmonic analysis: real-variable methods, orthogonality, and oscillatory integrals,
volume 43 of Princeton Mathematical Series. Princeton University Press, Princeton, NJ, 1993.
[6] M. W. Wong. An introduction to pseudo-differential operators, volume 6 of Series on Analysis,
Applications and Computation. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, third
edition, 2014.

61