Lecture Notes
Uploaded by
chingofchinaLecture Notes
Uploaded by
chingofchinaMA4J0 Lecture Notes
Advanced Real Analysis
Filip Rindler
University of Warwick
Spring Term 2024
Filip Rindler
Mathematics Institute
University of Warwick
Coventry CV4 7AL
United Kingdom
F.Rindler@warwick.ac.uk
http://www.warwick.ac.uk/filiprindler
Copyright ©2024 Filip Rindler.
Licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 Interna-
tional license (CC BY-NC-SA 4.0)
Version 3.5
Introduction
One of the most classical questions in PDE theory is to understand the regularity properties
of solutions to the Poisson equation
−∆u := −(∂11 u + · · · + ∂dd u) = f in Rd , ()
where f ∈ L2 (on the whole space Rd ). It was a big step forward when mathematicians
understood that one should not take () literally, but instead consider its weak formulation:
Here, one looks for u ∈ L2 with the property that ∇u = (∂1 u, . . . , ∂d u) ∈ L2 , where ∇u
is the weak gradient, that is, ∇u = (v1 , . . . , vd ) with
vi ψ dx = − u ∂i ψ dx
for all test functions ψ ∈ C∞
c (R ) (smooth maps with compact support). Then, such a u is
d
called a weak solution of the Poisson equation () if
∇u · ∇ψ dx = f ψ dx (†)
for all ψ ∈ C∞ c (R ). If all the second-order derivatives of u exist in the classical sense
d
and are continuous, then the two formulation are equivalent, as can be seen by a simple
integration by parts. However, (†) makes sense even if only the weak gradient ∇u of u
exists and lies in L2 . Standard results, namely the Lax–Milgram lemma, or even the Riesz
representation theorem in the present case, entail that one may find such a weak solution u
as in (†); details are discussed in MA4A2 Advanced Partial Differential Equations.
On the other hand, in Physics it is common to write the fundamental solution of the
Poisson equation () via the Newtonian potential (with negative sign)
1
|x|2−d if d ∕= 2,
Γ(x) := d(d−2)ω 1
d
− 2π log |x| if d = 2,
where ωd is the volume of the unit ball in Rd . For instance, if d = 3, then Γ(x) = 1/(4π|x|).
If f is smooth with compact support, it can be shown that
u(x) := (Γ f )(x) := Γ(x − y)f (y) dy
i
ii
solves (†) (with “” the convolution). In fact, physicists “compute” without hesitation that
−∆Γ = δ0 ,
the Dirac delta “function” at zero. Then, since δ0 f = f and the Laplace operator com-
mutes with the convolution, we obtain, at least formally,
−∆(Γ f ) = (−∆Γ) f = δ0 f = f.
Thus, Γ f “solves” the Poisson equation () with right-hand side f . One may wonder
if this can be understood in a rigorous fashion if f is not assumed to be smooth and with
compact support.
Other prototypical questions about (†) are the following: Can we say more about u?
Most importantly, does u have second (weak) derivatives? Or even higher derivatives, as-
suming that f has more regularity?
If one starts from the representation of u via the Newtonian potential, one could try to
push second derivatives onto the kernel Γ to obtain, formally
∂ij u(x) = (∂ij Γ f )(x) = (∂ij Γ)(x − y)f (y) dy.
However, ∂ij Γ(z) ∼ |z|−d has a non-integrable singularity at the origin and so it is not even
clear how to make sense of the above integral representation, let alone justify the exchange
of differentiation and integration.
It turns out that if f is assumed to be in Lp for 1 < p < ∞ then we have that also
∇ u ∈ Lp . These so-called Calderón–Zygmund estimates lie at the origin of Real Analysis
2
and Harmonic Analysis and we will develop a systematic theory to investigate all the above
questions (and many more) in this course. It is a crucial principle of the area that many tech-
niques originally developed to investigate convolution operators like the ones above, have
dual counterparts via the Fourier transform. While MA433 Fourier Analysis focusses on the
Fourier side and gave some glimpses of the general picture (e.g., proving Lp -boundedness
properties for the Hilbert transform), we here concentrate on the more abstract aspects, that
is, a distributional approach to convolution and singular integral operators, their general
Calderón–Zygmund theory, Fourier multipliers, Littlewood–Paley theory, as well as some
aspects of Sobolev spaces on unbounded domains. The course culminates with an introduc-
tion to the framework of pseudodifferential operators, which unifies much of the material.
Standard texts in the field are Muscalu–Schlag [8, 9], Grafakos [2, 3], Stein [10, 11],
Stein–Weiss [12], and the monographs by Hörmander [4–7]. These notes have also bene-
fitted from some “institutional memory” in the form of notes on the preceding versions of
this course by José Rodrigo, Vedran Sohinger, Hendrik Weber. I would like to thank Paolo
Bonicatto and Giacomo Del Nin, as well as the participants of the module in Spring 2022,
for many comments and corrections.
iii
These lecture notes are a living document and I would appreciate comments, corrections
and remarks – however small. They can be sent to me via e-mail to:
F.Rindler@warwick.ac.uk
Filip Rindler
January 2024
Contents
1 Distributions 1
1.1 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Localization and differentiation . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Convergence and approximation . . . . . . . . . . . . . . . . . . . . . . . 12
2 Tempered Distributions 15
2.1 Tempered distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Calderón–Zygmund Theory 23
3.1 Weak-type and strong-type bounds . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Convolution operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Boundedness of convolution operators . . . . . . . . . . . . . . . . . . . 30
4 Fourier Multipliers and Littlewood–Paley Theory 37
4.1 The Mikhlin multiplier theorem . . . . . . . . . . . . . . . . . . . . . . . 38
4.2 Dyadic decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 The square function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5 Sobolev Spaces 49
5.1 Homogeneous and inhomogeneous Sobolev spaces . . . . . . . . . . . . . 49
5.2 Inequalities and embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.3 Applications to PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6 Pseudodifferential Operators 59
6.1 General singular integral operators . . . . . . . . . . . . . . . . . . . . . 59
6.2 Symbols and operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.3 Mapping properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Bibliography 71
Index 73
v
Chapter 1
Distributions
The idea of distribution theory is to have a rigorous way to deal with rough objects, e.g.,
functions that have jumps or even more pathological singularities. While this is possible
in astonishing generality, the price one pays is that distributions are generally too rough
for most applications and one usually works in scales of spaces with “intermediate” and
precisely quantified regularity. Nevertheless, distributions form the foundation on which
Real and Harmonic Analysis are built.
1.1 Distributions
Let Ω ⊂ Rd be open. As usual we denote by C(Ω) = C0 (Ω), Ck (Ω), k = 1, 2, . . ., the
spaces of continuous and k-times continuously differentiable functions. As norms in these
spaces we have
uCk := ∂ α u∞ , u ∈ Ck (Ω), k = 0, 1, 2, . . . ,
|α|≤k
where ∞ is the supremum norm. Here, the sum is over all multi-indices α ∈ Nd0 (here
N0 := N ∪ {0}) with |α| := α1 + · · · + αd ≤ k and
∂ α := ∂1α1 ∂2α2 · · · ∂dαd
d
is the α-derivative operator. As usual, we write ∂i for the i’th partial derivative and dx for
the only derivative in R.
Similarly, we define the set C∞ (Ω) of infinitely-often differentiable functions (but this is
not a Banach space anymore). A subscript “c ” indicates that all functions u in the respective
function space (e.g. C∞ c (Ω)) must have their support
supp u := { x ∈ Ω : u(x) ∕= 0 }
compactly contained in Ω (so, supp u ⊂ Ω and supp u compact). For the compact contain-
ment of a A in an open set B we write A ⋐ B, which means that A is compact and A ⊂ B.
1
2 CHAPTER 1. DISTRIBUTIONS
In this way, the previous condition could be written “supp u ⋐ Ω”. We also abbreviate
D := D(Rd ).
The space D(Ω) := C∞ c (Ω; C), that is, the set of all ϕ : Ω → C that are infinitely
differentiable and compactly supported, is called the space of test functions (on Ω).
A sequence (ϕj ) ⊂ D(Ω) is said to converge to zero in D(Ω), in symbols “ϕj → 0 in
D”, if there is a compact set K ⊂ Ω such that supp ϕj ⊂ K for all j ∈ N and ∂ α ϕj ∞ →
0 as j → ∞ for all multi-indices α ∈ Nd0 . We say that (ϕj ) ⊂ D(Ω) converges to
ϕ ∈ D(Ω), in symbols ϕj → ϕ in D(Ω), if ϕj − ϕ → 0 in D(Ω). It is an exercise to see
that this convergence is not induced by a metric.
A distribution on an open set Ω ⊂ Rd is a C-linear map u : D(Ω) → C with the
property that for every compact set K ⊂ Ω there is a constant C = C(K) > 0 and an
integer N = N (K) ∈ N0 such that
u, ϕ ≤ C ∂ α ϕ∞ (1.1)
|α|≤N
for all ϕ ∈ D(Ω) with supp ϕ ⊂ K. Here, 〈u, ϕ〉 = u(ϕ) is the application of the linear
map u on ϕ. We collect all distributions on Ω into the set D′ (Ω), which is a vector space
over C under addition and scalar multiplication of distributions. If Ω = Rd , we simply write
D instead of D(Rd ).
If N can be chosen independent of the set K, then the least N for which (1.1) holds (for
every compact set K ⊂ Ω) is called the order of the distribution u. If N cannot be chosen
uniform in the set K, then we say that the distribution has infinite order.
For a distribution u ∈ D′ (Ω) we define the support of u to be the intersection of all
(relatively) closed sets F ⊂ Ω with the property that for all ϕ ∈ D(Ω \ F ) it holds that
〈u, ϕ〉 = 0. In symbols,
supp u = F ⊂ Ω (relatively) closed : 〈u, ϕ〉 = 0 for all ϕ ∈ D(Ω \ F ) .
The support is always a (relatively) closed subset of Ω.
The following are standard examples:
Example 1.1. Let f ∈ L1loc (Ω) with Ω ⊂ Rd open, that is, f ∈ L1 (K) for all compact
K ⋐ Ω. Then, f ∈ D′ (Ω) via
f, ϕ := f ϕ dx, ϕ ∈ D(Ω).
Indeed, if supp ϕ ⊂ K with K ⊂ Rd compact, then
f, ϕ ≤ f L1 (K) · ϕ∞
and hence the condition in (1.1) is satisfied (with N = 0), so f is a distribution of order 0
and support supp f .
1.1. DISTRIBUTIONS 3
Example 1.2. The distribution δz ∈ D′ for z ∈ Rd , defined via
δz , ϕ := ϕ(z), ϕ ∈ D,
is called the Dirac delta distribution at z. It is a distribution of order 0 and support {z}.
Example 1.3. The distribution ∂ β δz ∈ D′ , for any multi-index β ∈ Nd0 and z ∈ Rd , is
defined via
β
∂ δz , ϕ := (−1)|β| ∂ β ϕ(z), ϕ ∈ D.
Its order is |β| = β1 + · · · + βd and its support is {z}.
Example 1.4. The Hilbert kernel is the distribution KH ∈ D(R) (d = 1) defined as
1 1
KH := · PV , ϕ ∈ D,
π x
where “PV” stands for principal value integral: For ϕ ∈ D,
1 ϕ(y)
KH , ϕ := · lim dy.
π ε↓0 {|y|>ε} y
To see that this is well-defined and indeed defines a distribution, note that
1 ϕ(y) − ϕ(0) ϕ(y)
KH , ϕ = · lim dy + dy .
π ε↓0 {ε<|y|<1} y {|y|≥1} y
By the mean value theorem, for each y with ε < |y| < 1 we have that ϕ(y)−ϕ(0) y = ϕ′ (ζ)
′
for some ζ ∈ (−1, 1). Thus, the first integral is bounded by 2ϕ ∞ . If supp ϕ ⊂ [−R, R]
for some R > 0, then the second term is bounded by 2Rϕ∞ . Thus, KH is a distribution
of order at most 1 and support R. It is an instructive exercise to see that KH is indeed of
order 1 and not of order 0 even though it does not (directly) contain a derivative.
Example 1.5. Set Ω := (0, ∞) and
∞
1
u, ϕ := ∂kϕ , ϕ ∈ D((0, ∞)).
k
k=1
Then, it can be checked easily that u is a distribution of infinite order with support the set
{1, 1/2, 1/3, . . .} (which is relatively closed in (0, ∞)).
Distributions can also be characterized as those linear maps on D(Ω) that are continuous
with respect to convergence (to zero) in D(Ω). Here, and in all of the following, Ω denotes
a generic open subset of Rd .
Lemma 1.6. Let u : D(Ω) → C be a C-linear map. Then, the following are equivalent:
4 CHAPTER 1. DISTRIBUTIONS
(i) u ∈ D′ (Ω);
(ii) for every sequence ϕj → 0 in D(Ω) it holds that 〈u, ϕj 〉 → 0;
(iii) for every sequence ϕj → ϕ in D(Ω) it holds that 〈u, ϕj 〉 → 〈u, ϕ〉.
Proof. Since the equivalence between (ii) and (iii) is obvious, we only need to show that (i)
implies (ii) and that (ii) implies (i).
If u ∈ D′ (Ω) and ϕj → 0 in D(Ω), then by (1.1) there is an N ∈ N0 and a C > 0 such
that
u, ϕj ≤ C ∂ α ϕj ∞ → 0
|α|≤N
as j → ∞. Here we have used the definition of the convergence ϕj → 0 in D(Ω), in
particular that supp ϕj ⊂ K for a fixed compact set K ⊂ Ω.
For the other direction, let u : D(Ω) → C be linear and such that for every sequence
ϕj → 0 in D(Ω) it holds that 〈u, ϕj 〉 → 0. Arguing by contradiction, we assume that
u∈ / D′ (Ω). So, there exists a compact set K ⊂ Ω with the property that for all N ∈ N0
there is a ϕN ∈ D(Ω) \ {0} with supp ϕN ⊂ K and
u, ϕN ≥ N ∂ α ϕN ∞ .
|α|≤N
Set
ϕN
ψN := α
,
N |α|≤N ∂ ϕN ∞
so that
u, ψN ≥ 1. (1.2)
On the other hand, for every multi-index β ∈ Nd0 ,
1 ∂ β ϕN ∞
∂ β ψN ∞ = · α
→ 0,
N |α|≤N ∂ ϕN ∞
hence ψN → 0 in D(Ω). By our assumption on u, we have |〈u, ψN 〉| → 0, contradict-
ing (1.2).
Next, we prove the following useful fact about distributions of order zero:
Lemma 1.7. If u ∈ D′ (Ω) has order zero, then there is a (local) C-valued Radon measure
µ ∈ Mloc (Ω; C) (i.e., a σ-additive function on the Borel σ-algebra with values in C) with
u = µ in the sense that
u, ϕ = ϕ dµ, ϕ ∈ D(Ω).
1.1. DISTRIBUTIONS 5
Proof. By assumption, for any compact set K ⊂ Ω it holds that
u, ϕ ≤ Cϕ∞
for all ϕ ∈ D(Ω) with supp ϕ ⊂ K and C = C(K). In particular, u may be extended to a
bounded linear functional on C0 (K; C) (continuous functions on K). It is well known by
one version of the Riesz representation theorem that this implies that indeed u is given by
integration against a (local) Radon measure µ ∈ Mloc (Ω; C).
The multiplication of a distribution u ∈ D′ (Ω) by a smooth function f ∈ C∞ (Ω; C) is
the distribution f u ∈ D′ (Ω) given as
f u, ϕ := u, f ϕ , ϕ ∈ D.
To check that f u is indeed a distribution, observe that if supp ϕ ⊂ K for a compact set
K ⊂ Ω, then via (1.1) there are C = C(K) > 0 and N = N (K) ∈ N0 such that for any
ϕ ∈ D(Ω) with supp ϕ ⊂ K it holds that
u, f ϕ ≤ C ∂ α (f ϕ)∞ ≤ C ∂ α ϕ∞ ,
|α|≤N |α|≤N
where we have applied the Leibniz rule
α
α
∂ (f ϕ) = ∂ β f · ∂ α−β ϕ
β
β≤α
and absorbed Here,
αany constants
α!
as well as the terms ∂ β f ∞ into the new constant C.
recall that β := β!(α−β)! with α! := α1 ! · · · αd !, and β ≤ α is understood elementwise.
Thus, also f u ∈ D′ (Ω). It is easy to see that the order of f u is less than or equal to the
order of u (as an exercise one should verify that the order of f u may be strictly smaller than
the order of u) and supp(f u) ⊂ supp u ∩ supp f .
Example 1.8. Let f ∈ C∞ (Rd ; C) and let δz ∈ D′ be the Dirac delta distribution at
z ∈ Rd (as in Example 1.2). Then,
f δz , ϕ = δz , f ϕ = f (z)ϕ(z) = f (z)δz , ϕ , ϕ ∈ D,
so that f δz = f (z)δz .
However, one needs to be careful when computing products of distributions with smooth
functions, as is exemplified in the following:
Example 1.9. Let f ∈ C∞ (Rd ; C). We want to compute an explicit expression for the
distribution f ∂i δz , where z ∈ Rd and i = 1, . . . , d. For ϕ ∈ D(Ω) we have
f ∂i δz , ϕ = − δz , ∂i (f ϕ)
= − δz , (∂i f ) ϕ − δz , f ∂i ϕ
= −(∂i f )(z) ϕ(z) − f (z) ∂i ϕ(z)
= −∂i f (z) δz , ϕ + f (z) ∂i δz , ϕ ,
6 CHAPTER 1. DISTRIBUTIONS
so that f ∂i δz = f (z) ∂i δz − ∂i f (z) δz . In particular, even if f (z) = 0, it may be that
f ∂i δz ∕= 0.
Let u ∈ D′ and ψ ∈ D. Then the convolution u ψ of u with ψ is defined via
u ψ, ϕ := u, ψ ϕ , ϕ ∈ D,
where we have set
ψ(x) := ψ(−x), x ∈ Rd .
Note that if u ∈ L1 (identified with the corresponding distribution), then, using Fubini’s
theorem and the commutativity of the convolution,
u ψ, ϕ = u(x) ψ ϕ (x) dx
= u(x)ψ(y − x)ϕ(y) dy dx
= u(x)ψ(y − x) dx ϕ(y) dy
= u ψ](y)ϕ(y) dy.
Thus, our distributional definition agrees with the classical one in this case.
Lemma 1.10. For u ∈ D′ and ψ ∈ D it holds that u ψ ∈ D′ and u ψ is, in fact, given
by integration against a smooth function, namely x → 〈u, ψ(x − )〉.
Proof. It is a standard property of the convolution that for ϕ ∈ D it holds that
∂ α (ψ ϕ)∞ = ψ (∂ α ϕ)∞ ≤ ψL1 · ∂ α ϕ∞ .
The fact that u ψ ∈ D′ is then easy to see using the definition of distributions.
To show that u ψ is given by integration against the function
v(x) := u, ψ(x − ) ,
x ∈ Rd ,
we may first observe that v is smooth: For i ∈ {1, . . . , d} we have
ψ(x + hei − ) − ψ(x − )
→ u, ∂i ψ(x − )
v(x + hei ) − v(x)
= u,
h h
since the difference quotient converges to the derivative in D (this is an easy exercise).
Iterating this argument for higher derivatives, we obtain that v is smooth.
1.2. LOCALIZATION AND DIFFERENTIATION 7
To show that v agrees with u ψ, for any ϕ ∈ D we may write 〈v, ϕ〉 as a Riemann
sum,
v, ϕ = v(x)ϕ(x) dx
= lim εd v(εz)ϕ(εz)
ε↓0
z∈Zd
εd u, ψ(εz − ) ϕ(εz)
= lim
ε↓0
z∈Zd
= lim u, ε ψ(εz − )ϕ(εz)
d
ε↓0
z∈Zd
since only finitely many terms in the sum are nonzero. Then we see that
εd ψ(εz − )ϕ(εz)
hε :=
z∈Zd
is a Riemann sum for the function ψ ϕ. Thus, by the smoothness of ϕ and ψ, we have
hε → ψ ϕ in D, meaning that the difference converges to zero (this is an exercise). We
can now invoke the continuity property of u (see Lemma 1.6) to see that
v, ϕ = u, ψ ϕ = u ψ, ϕ .
We have thus indeed verified that u ψ is given by integration against the smooth map
v.
1.2 Localization and differentiation
We will now investigate some local properties of distributions. For u ∈ D′ (Ω) we say that
“u = 0 on an open set U ⊂ Ω” if supp u ∩ U = ∅. We further say for u, v ∈ D′ (Ω) that
“u = v on U ” if u − v = 0 on Ω.
One key tool in the study of distributions, and in fact in Real Analysis in general, is
the following: A partition of unity of a compact set K ⊂ Ω subordinate to the open sets
U1 , . . . , Um (m ∈ N) with
K ⊂ U1 ∪ · · · ∪ Um
is a collection of functions ψi ∈ C∞
c (Ui ; [0, 1]) (that is, 0 ≤ ψi ≤ 1) for i = 1, . . . , m such
that
m
m
ψi = 1 on a neighborhood of K and ψi ≤ 1 on Ω.
i=1 i=1
The existence of a partition of unity is proved in other modules in more generality, but it
is straightforward in Rd : Pick for every x ∈ K a compact neighborhood Nx of x with
8 CHAPTER 1. DISTRIBUTIONS
Nx ⊂ Ui for one i = 1, . . . , m. By the compactness of K, finitely many x1 , . . . , xn ∈ K
suffice such that K ⊂ nℓ=1 Nx ℓ
. Then, for every i ∈ {1, . . . , m} there is a function
i ∈ Cc (Ui ; [0, 1]) that is 1 on Nx ⊂Ui Nxℓ (using a direct construction or Urysohn’s
ℓ
lemma). Normalizing ψi := (1−1 ) · · · (1−i−1 )i on Ui , we obtain that ψ1 +· · ·+ψm =
1 − (1 − 1 ) · · · (1 − m ). Since for every x ∈ K there is at least one i ∈ {1, . . . , m} with
i (x) = 1, the existence of a partition of unity is proved.
The next result shows how we can construct a distribution from local pieces:
Theorem 1.11. Let Uλ ⊂ Ω, λ ∈ Λ, be an open cover for Ω, where Λ is any index set.
Suppose that for every λ ∈ Λ we are given a distribution uλ ∈ D′ (Uλ ) such that, for
λ, µ ∈ Λ,
uλ = uµ on Uλ ∩ Uµ . (1.3)
Then, there is a unique distribution u ∈ D′ (Ω) with u = uλ on Uλ for any λ ∈ Λ.
Proof. Let ϕ ∈ D(Ω). By the compactness of supp ϕ, there exist finitely many indices
λ1 , . . . , λm ∈ Λ such that supp ϕ ⊂ Uλ1 ∪ · · · ∪ Uλm . Take a partition of unity {ψi }i=1,...,m
of supp ϕ subordinate to the sets Uλ1 , . . . , Uλm and define
m
u, ϕ := uλ i , ψi ϕ .
i=1
In order to see that this definition is well-defined, that is, it does not depend on the chosen
cover and partition of unity, assume that supp ϕ ⊂ Uµ1 ∪ · · · ∪ Uµn is another finite open
cover of supp ϕ with corresponding partition of unity {ψj }j=1,...,n . We have, using (1.3),
m
m
n m
n n
uλ i , ψi ϕ = uλi , ψj ψi ϕ = uµj , ψj ψi ϕ = uµj , ψj ϕ .
i=1 i=1 j=1 i=1 j=1 j=1
So, our definition did not depend on the choice of open cover.
Fix a compact set K ⊂ Ω and fix a finite (by compactness) open cover Uλ1 , . . . , Uλm
of K as above. For every ϕ ∈ D(Ω) with supp ϕ ⊂ K, we then have
m m
u, ϕ ≤ uλ , ψi ϕ ≤ Ci ∂ α (ψi ϕ)∞ ≤ C ∂ α ϕ∞ .
i
i=1 i=1 |α|≤Ni |α|≤N
Here, the Ni , Ci originate from (1.1) for uλi and we have set N := maxi Ni . The last
inequality holds since we can expand ∂ α (ψi ϕ) via the Leibniz rule and absorb all constants
as well as the terms ∂ β ψi ∞ into the constant C = C(K) > 0 (we have already seen this
argument when defining the multiplication of a distribution). Thus, u satisfies (1.1). Since
the linearity is clear, we have shown that u ∈ D′ (Ω).
Finally, if ϕ ∈ D(Uλ ) for some λ ∈ Λ, the above definition immediately shows that
〈u, ϕ〉 = 〈uλ , ϕ〉 and thus u = uλi on Uλi . This in particular implies for any u ∈ D′ (Ω)
with u = uλ on Uλ for any λ ∈ Λ that u = uλ = u on Uλ for any λ. Thus, u ∈ D′ (Ω) with
u = uλ on Uλ for any λ ∈ Λ, is uniquely identified by this property.
1.2. LOCALIZATION AND DIFFERENTIATION 9
Next, we investigate distributions with support in only one point.
Theorem 1.12. Let u ∈ D′ with supp u = {z}. Then, there is an N ∈ N0 such that
u= c α ∂ α δz .
|α|≤N
for some constants cα ∈ C (α ∈ Nd0 ).
Proof. Without loss of generality let z = 0.
Step 1. We first claim that there is an N ∈ N0 such that for ϕ ∈ D the following implication
holds: If ∂ α ϕ(0) = 0 for |α| ≤ N , then 〈u, ϕ〉 = 0.
To prove this claim, take a cut-off function η ∈ C∞ c (R ; [0, 1]) with η = 1 on B1/2
d
(the open ball centered at the origin with radius 1/2) and supp η ⊂ B1 . For ϕ ∈ D and
0 < ε < 1 set
x
ϕε (x) := ϕ(x)η , x ∈ Rd .
ε
Since ϕε = ϕ on Bε/2 , we have that 〈u, ϕε 〉 = 〈u, ϕ〉. Moreover, since supp ϕε ⊂ B1 ,
there are (ε-independent) N ∈ N0 and C > 0 with
u, ϕ = u, ϕε ≤ C ∂ α ϕε ∞ . (1.4)
|α|≤N
Taylor’s theorem (in its multi-variate version) gives that for any multi-index β ∈ Nd0 it holds
that
1 β+γ
∂ β ϕ(x) = ∂ ϕ(0)xγ + Rγ (x)xγ ,
γ!
|γ|≤N −|β| |γ|=N −|β|+1
where xγ here denotes the map x → xγ = xγ11 · · · xγdd and γ! := γ1 ! · · · γd !, and
1 β+γ
Rγ ∞ ≤ ∂ ϕ∞ .
γ!
Using the assumption that ∂ α ϕ(0) = 0 if |α| ≤ N , this yields for |β| ≤ N and |x| ≤ ε that
|∂ β ϕ(x)| ≤ CN εN −|β|+1 · ∂ γ ϕ∞ .
|γ|=N +1
On the other hand, by the Leibniz rule we have
x
∂ α ϕε (x) = Cα,β,γ ε−|γ| ∂ β ϕ(x) ∂ γ η .
ε
β,γ : α=β+γ
We may combine the last two statements to deduce that
∂ α ϕε ∞ ≤ Cϕ,α εN −|α|+1 .
10 CHAPTER 1. DISTRIBUTIONS
Here, Cϕ,α > 0 is a constant that depends on ϕ and α, but not on ε. Plugging this into (1.4),
we obtain (adjusting constants as we go)
u, ϕ ≤ C εN −|α|+1 ≤ Cε
|α|≤N
and this tends to zero as ε ↓ 0. Hence, the claim is established.
Step 2. Let N be as in Step 1 and take the same cut-off function η as before. We may write
a generic ϕ ∈ D as
∂ α ϕ(0)
ϕ=η· xα + ϕ.
α!
|α|≤N
One verifies directly that ∂ α ϕ(0)
= 0 if |α| ≤ N . Then, via Step 1, we see that 〈u, ϕ〉
=0
and, consequently,
∂ α ϕ(0)
u, ϕ = u, xα η = c α ∂ α δ0 , ϕ ,
α!
|α|≤N |α|≤N
(−1)|α|
with cα := α! u, xα η . This was the claim of the theorem.
The reason distributions were originally introduced is that they can be differentiated to
any order, even though they are potentially very “rough”. The basic idea is the following:
For two smooth maps u ∈ C∞ (Rd ; C) and ϕ ∈ C∞ c (R ; C) = D we have the integration-
d
by-parts formula
β |β|
∂ u · ϕ dx = (−1) u · ∂ β ϕ dx
for any multi-index β ∈ Nd0 ; note that there are no boundary terms because ϕ has compact
support. The central insight is that the right-hand side is defined even if only u ∈ L1loc (Rd ).
So, if there exists a v ∈ L1loc (Rd ; C) with
|β|
v · ϕ dx = (−1) u · ∂ β ϕ dx
for all ϕ ∈ D, then we may consider v to be the “weak” β-derivative of u.
We may generalize this idea even further and define the β-derivative ∂ β u of a distribu-
tion u ∈ D′ (Ω) to be the distribution ∂ β u ∈ D′ (Ω) given as follows:
β
∂ u, ϕ := (−1)|β| u, ∂ β ϕ ϕ ∈ D.
To see that this is indeed a distribution, we estimate from (1.1) as follows: If K ⊂ Ω is
compact, then for all ϕ ∈ D(Ω) with supp ϕ ⊂ K it holds that
β
∂ u, ϕ = u, ∂ β ϕ ≤ C ∂ α ∂ β ϕ∞ ≤ C ∂ α ϕ∞ .
|α|≤N |α|≤N +|β|
Thus, ∂ β u ∈ D′ (Ω). It is easy to see that supp ∂ β u ⊂ supp u and that the order of ∂ β u is
at most the order of u plus |β|.
1.2. LOCALIZATION AND DIFFERENTIATION 11
Example 1.13. We see now that ∂ β δz defined in Example 1.3 is indeed the β-derivative
of the Dirac delta distribution δz from Example 1.2.
Example 1.14. The Heaviside function H ∈ L1loc (R) is defined as
0 if x < 0,
H(x) :=
1 if x > 0.
d
As in Example 1.1 we see that H ∈ D′ (R). To compute its derivative H ′ = dt H, we
observe for ϕ ∈ D(R) that
∞
′
H , ϕ = − H, ϕ′ = − ϕ′ (x) dx = ϕ(0) = δ0 , ϕ ,
0
so that H ′ = δ0 .
Example 1.15. Recall the definition of the Hilbert kernel KH in Example 1.4. We claim
that
d 1
log |x| = PV = πKH .
dx x
Here we note that log |x| ∈ L1loc (R) and hence it defines a distribution on R. To see this
claim, let ϕ ∈ D and compute as follows:
d
log |x|, ϕ = − log |x|, ϕ′
dx
= − ϕ′ (x) log |x| dx
−ε ∞
′ ′
= − lim ϕ (x) log(−x) dx + ϕ (x) log x dx
ε↓0 −∞ ε
−ε ∞
ϕ(x) ϕ(x)
= lim ϕ(ε) − ϕ(−ε) log ε + dx + dx
ε↓0 x x
−∞ ε
1
= lim ϕ(ε) − ϕ(−ε) log ε + PV , ϕ ,
ε↓0 x
where we have integrated by parts in the second-to-last line. The first term in the last line
goes to zero since, by the mean value theorem,
ϕ(ε) − ϕ(−ε) log ε ≤ Cε log ε ↓ 0
as ε ↓ 0. Thus, the claim is established.
Derivatives for distributions obey similar rules to classical derivatives, most importantly
the following product rule:
12 CHAPTER 1. DISTRIBUTIONS
Lemma 1.16. Let u ∈ D′ (Ω) and f ∈ C∞ (Ω; C). Then, for i = 1, . . . , d it holds that
∂i (f u) = (∂i f )u + f ∂i u.
Proof. For ϕ ∈ D(Ω) we have
∂i (f u), ϕ = − u, f ∂i ϕ = − u, ∂i (f ϕ) − (∂i f )ϕ = f ∂i u, ϕ + (∂i f )u, ϕ ,
which already shows the claim.
Example 1.17. For f H ∈ D′ (R), where H is the Heaviside function from Example 1.14
and f ∈ C∞ (R; C), one may thus compute that
d
(f H) = f ′ H + f (0)δ0 .
dx
Example 1.18. With the Laplacian
∆ := ∂11 + ∂22 + . . . + ∂dd
interpreted distributionally, one may compute using polar coordinates (exercise) for d ∕= 2
that
∆(|x|2−d ) = −d(d − 2)ωd δ0 , ωd := Ld (B1 (0)),
whereas for d = 2 we have
∆(log |x|) = 2πδ0 .
1.3 Convergence and approximation
Let (uj ) ⊂ D′ be a sequence of distributions. We say that uj converges to u ∈ D′ (in the
∗
distributional sense or in the weak* sense), in symbols “uj ⇀ u in D′ ”, if
uj , ϕ → u, ϕ
for all ϕ ∈ D.
Example 1.19. Let (δ )δ>0 ⊂ C∞
c (R ) be an approximation of the identity, that is,
d
δ ≥ 0, δ (x) dx = 1, supp δ ⊂ Bδ .
Then, let ϕ ∈ D and observe, by the mean value theorem,
δ (x)ϕ(x) dx − ϕ(0) = δ (x)(ϕ(x) − ϕ(0)) dx
≤C δ (x)δ dx · ∇ϕ∞
Bδ
= Cδ∇ϕ∞
→ 0.
∗
Hence, δ ⇀ δ0 in D′ as δ → 0.
1.3. CONVERGENCE AND APPROXIMATION 13
Example 1.20. Let uj (x) := e−2πijx for x ∈ R and j ∈ N. Then, uj ∈ D′ and for ϕ ∈ D
we integrate by parts to see
uj , ϕ = ϕ(x) e−2πijx dx
1 d −2πijx
= ϕ(x) e dx
−2πij dx
1
= ϕ′ (x) e−2πijx dx.
2πij
Clearly, the last expression converges to zero as j → ∞. Alternatively, one could realize
that 〈uj , ϕ〉 = ϕ(j)
and use the Riemann–Lebesgue lemma for the Fourier series (cf. The-
∗
orem 2.10). Thus, uj ⇀ 0 in D′ even though uj (x) does not converge for any x ∕= 0.
For distributions, differentiation is always a (weakly*) continuous operation:
∗ ∗
Lemma 1.21. Let uj ⇀ u in D′ . Then, for any multi-index α ∈ Nd0 it holds that ∂ α uj ⇀
∂ α u in D′ .
Proof. For ϕ ∈ D,
α
∂ uj , ϕ = (−1)|α| uj , ∂ α ϕ → (−1)|α| u, ∂ α ϕ = ∂ α u, ϕ .
This is the claim.
The next result gives us a very useful density result in D′ :
∗
Theorem 1.22. For any u ∈ D′ there is a sequence (ψk )k ⊂ D with ψk ⇀ u in D′ . In
other words, D is (weakly*) dense in D′ . If u ∈ Lp , where 1 ≤ p < ∞, then ψk → u in Lp .
Proof. Let (δ )δ>0 ⊂ C∞c (R ) be an approximation of the identity as in Example 1.19. Let
d
ϕ ∈ D. It is a standard property of the convolution that supp δ ϕ ⊂ supp ϕ + Bδ and
that
∂ α (δ ϕ) = δ (∂ α ϕ) → ∂ α ϕ
uniformly for any multi-index α ∈ Nd0 as δ ↓ 0. Hence, δ ϕ → ϕ in D.
We now take ηk ∈ C∞ c (R ; [0, 1]) with ηk = 1 on Bk (k ∈ N) and set
d
ψk := (ηk u) 1/k
From Lemma 1.10 it is clear that ψk ∈ D. Moreover, for ϕ ∈ D, it holds that ηk = 1 on
supp 1/k ϕ for k sufficiently large. Consequently,
lim ψk , ϕ = lim u, ηk ( 1/k ϕ) = lim u, 1/k ϕ = u, ϕ .
k→∞ k→∞ k→∞
∗
Thus, ψk ⇀ u in D′ as k → ∞. The additional convergence assertion are standard and we
omit a proof.
Chapter 2
Tempered Distributions
One issue with distributions is that they act only on compactly supported test functions.
However, when we employ the Fourier transform (and we will do so many times in this
course), then we immediately observe that the Fourier transform does not map D into D.
In fact, the only ϕ ∈ D with ϕ ∈ D is the zero map, this is one of the versions of the
Heisenberg–Hardy uncertainty principle. However, if we extend the set of test functions to
the set of rapidly decaying (Schwartz) functions, we obtain a very fruitful Fourier theory
for the corresponding class of distributions.
2.1 Tempered distributions
The vector space S of Schwartz functions is the set of those ϕ ∈ C∞ (Rd ; C) such that for
all α, β ∈ Nd0 it holds that
α,β (ϕ) := xα ∂ β ϕ∞ < ∞.
It is easy to see that D ⊂ S ⊂ C∞ (Rd ; C), with strict inclusions, and that S ⊂ Lp for all
1 ≤ p ≤ ∞ (exercise). Moreover, S is stable under differentiation and multiplication with
polynomials.
We say that a sequence (ϕj ) ⊂ S converges to ϕ ∈ S, in symbols “ϕj → ϕ in S”, if
α,β (ϕj − ϕ) → 0 as j → ∞ for all α, β ∈ Nd0 .
A tempered distribution is a C-linear map u : S → C such that there is a constant
C > 0 and an integer N ∈ N0 for which
u, ϕ ≤ C α,β (ϕ) (2.1)
|α|,|β|≤N
for all ϕ ∈ S. We collect all tempered distributions into the set S ′ , which is a vector space
over C under addition and scalar multiplication of distributions.
For u ∈ S ′ and ϕ ∈ D with supp ϕ ⊂ K, where K ⊂ Rd is any compact set, we have
u, ϕ ≤ C(K) ∂ β ϕ∞ ,
|β|≤N
15
16 CHAPTER 2. TEMPERED DISTRIBUTIONS
where C(K) > 0 depends on K. Thus, we see that u ∈ D′ and hence
S ′ ⊂ D′ .
Since N is fixed, we have that u has finite order (at most N ).
Lemma 2.1. The set S ′ of tempered distributions is stable under differentiation and mul-
tiplication with polynomials.
Proof. For u ∈ S ′ and i = 1, . . . , d we have for any ϕ ∈ S that
∂i u, ϕ = u, ∂i ϕ ≤ C α,β (∂i ϕ) ≤ C α,β (ϕ),
|α|,|β|≤N |α|,|β|≤N +1
so also ∂i u ∈ S ′ . For higher derivatives we may iterate this argument.
Similarly, if ϕ ∈ S, then also xi ϕ ∈ S (where “xi ” here denotes the map x → xi ) and
so,
xi u, ϕ = u, xi ϕ ≤ C α,β (xi ϕ) ≤ C α,β (ϕ)
|α|,|β|≤N |α|,|β|≤N +1
by the Leibniz rule, where C > 0. Thus, xi u ∈ S ′ . The claim for polynomials then follows
by linearity and iteration of the above.
Example 2.2. If u ∈ D′ with supp u compact (which is often denoted as u ∈ E ′ (Rd )),
then u ∈ S ′ . To see this fact take a cut-off function η ∈ C∞
c (R ; [0, 1]) with η = 1 on a
d
neighborhood of supp u and set, for ϕ ∈ S,
u, ϕ := u, ηϕ .
It can be seen easily that this definition does not depend on the choice of η and indeed
constitutes a tempered distribution.
We say that a sequence (uj ) ⊂ S ′ converges to u ∈ S ′ (in the weak* or distributional
∗
sense), in symbols “uj ⇀ u in S ′ ”, if
uj , ϕ → u, ϕ
for all ϕ ∈ S.
2.2 Fourier transform
Define for ϕ ∈ S the Fourier transform ϕ
= Fϕ ∈ S as follows:
ϕ(ξ) := ϕ(x) e−2πiξ·x dx, ξ ∈ Rd .
Rd
2.2. FOURIER TRANSFORM 17
Note that this is well-defined since S ⊂ L1 . We also define the inverse Fourier transform
ψq = F −1 ψ for ψ ∈ S to be
q
ψ(x) := ψ(ξ) e2πiξ·x dξ, x ∈ Rd .
Rd
ψq ∈ S follows directly from the definition of the Schwartz functions and
The fact that ϕ,
the following standard rules for the Fourier transform:
(i) Differentiation: F[∂j ϕ](ξ) = 2πiξj ϕ(ξ).
(ii) Multiplication: F[−2πixj ϕ](ξ) = ∂j ϕ(ξ).
(iii) Translation: F[ϕ( − v)](ξ) = e−2πiv·ξ ϕ(ξ),
where v ∈ Rd .
(iv) Modulation: F[e2πiv·x ϕ](ξ) = ϕ(ξ
− v), where v ∈ Rd .
(v) Scaling: ϕ
λ (ξ) = ϕ(λξ),
where ϕλ (x) := λ−d ϕ(x/λ) for λ > 0.
(vi) Almost self-inverse: F[F[ϕ]] = ϕ,
where ϕ(x)
:= ϕ(−x).
(vii) Plancherel’s relation: ϕ
L2 = ϕL2 (isometry on L2 ).
(viii) Parseval’s relation: ϕ·ψ dx = ϕ· dx = ϕ· ψ dξ.
ψ dξ (unitarity) or ϕ·ψ
2 2
(ix) Gaussian fixed point: F[e−π| | ](ξ) = e−π|ξ| .
(x) Convolution: F[ϕ ψ] = ϕ
· ψ.
Here, for ϕ, ψ ∈ S, we have defined the convolution of ϕ and ψ via
(ϕ ψ)(x) := ϕ(x − y)ψ(y) dy, x ∈ Rd ,
which itself lies in S as an easy argument based on differentiation under the integral shows.
By Lemma 1.10 it is clear that this definition agrees with the distributional one.
We can now extend the domain of the Fourier transform to S ′ by duality: The Fourier
transform of u ∈ S ′ is given as
, ϕ := u, ϕ
u , ϕ ∈ S.
This definition is consistent with the classical definition by Parseval’s relation (viii).
Lemma 2.3. For u ∈ S ′ it holds that u
∈ S ′.
18 CHAPTER 2. TEMPERED DISTRIBUTIONS
Proof. Clearly, u
is linear on S. Moreover, for ϕ ∈ S,
u, ϕ = u, ϕ
≤C
α,β (ϕ).
|α|,|β|≤N
By the rules above for the Fourier transform,
= ξ α ∂ β ϕ
α,β (ϕ) ∞ ≤ C F[∂ α (xβ ϕ)]∞ ,
where the constant absorbs the multiples of |2πi|. Thus, also using the Leibniz rule (and
adjusting the constant C)
≤C
α,β (ϕ) F[∂ α (xβ ϕ)] ≤C F[xα ∂ β ϕ] .
∞ ∞
|α|,|β|≤N |α|,|β|≤N |α|,|β|≤N
We further observe that for any ψ ∈ S it holds that
|ψ(ξ)| ≤ |ψ(x)| dx
dx
≤ sup (1 + |x|)d+1 |ψ(x)|
x∈Rd (1 + |x|)d+1
≤ C sup (1 + |x|)d+1 |ψ(x)| .
x∈Rd
Applying this for ψ = xα ∂ β ϕ and combining with the above, we obtain (exercise)
u, ϕ ≤ C sup (1 + |x|)d+1 |xα ∂ β ϕ(x)| ≤ C α,β (ϕ).
d
|α|,|β|≤N x∈R |α|,|β|≤N +d+1
∈ S ′.
This shows that also u
Example 2.4. If u ∈ L1 , then u
as defined above coincides with the classical definition of
the Fourier transform.
Example 2.5. Let δ0 ∈ S ′ be the Dirac delta distribution at the origin, which is also a
tempered distribution since it has compact support. To compute its Fourier transform, let
ϕ ∈ S and observe
= ϕ(x) dx = , ϕ ,
δ0 , ϕ = δ0 , ϕ
so that the constant-one function is seen to be the Fourier transform of δ0 , i.e., δ0 = .
Example 2.6. Since u
=u , where u ∈ S ′ is the distribution defined via 〈
u, ϕ〉 = 〈u, ϕ〉
with ϕ(x)
:= ϕ(−x), we obtain from the last example that = δ0 .
2.2. FOURIER TRANSFORM 19
Example 2.7. Let KH = π1 · PV x1 be the Hilbert kernel (see Example 1.4), which is easily
seen to be a tempered distribution. Then, for ϕ ∈ S,
H , ϕ = KH , ϕ
K
1
ϕ(ξ)
= · lim dξ
π ε→0 {|ξ|>ε} ξ
e−2πixξ
= lim ϕ(x) dx dξ
ε→0 {ε<|ξ|<1/ε} πξ
e−2πixξ
= lim ϕ(x) dξ dx.
ε→0 {ε<|ξ|<1/ε} πξ
Now observe that (using sin z = Im eiz = (eiz − e−iz )/(2i))
e−2πixξ 1 e−2πixξ 1 e2πixξ
dξ = dξ − dξ
{ε<|ξ|<1/ε} πξ 2 {ε<|ξ|<1/ε} πξ 2 {ε<|ξ|<1/ε} πξ
sin(2πxξ)
= −i dξ
{ε<|ξ|<1/ε} πξ
sin(|x|ξ)
= −i sgn(x) dξ,
{ε/(2π)<|ξ|<1/(2πε)} πξ
where the signum function on R given as
−1 if x < 0,
sgn(x) := 0 if x = 0,
+1 if x > 0.
One can now show (exercise) that
sin(|x|ξ)
lim dξ = 1.
ε→0 {ε/(2π)<|ξ|<1/(2πε)} πξ
Thus,
H , ϕ =
K ϕ(x)(−i sgn(x)) dx,
H is given by the locally integrable function −i sgn.
whereby K
H = −i sgn. Taking the Fourier
Example 2.8. In the previous example we showed that K
transform again, we get
H = − 1 1
−i s
gn = K · PV .
π ξ
Hence,
1 1
s
gn = · PV .
πi ξ
20 CHAPTER 2. TEMPERED DISTRIBUTIONS
Example 2.9. For the Heaviside function H from Example 1.14 we have H = (1+sgn)/2,
so that the previous examples imply
1 1 1
H(ξ) = δ0 + · PV .
2 2πi ξ
Note that here we have written H(ξ) is not a function;
on the left-hand side, even though H
this is a common notation to fix the variable.
We end this section with a result that is called the Riemann–Lebesgue lemma:
Theorem 2.10. Let u ∈ L1 . Then, u
∈ L1loc ∩ C (that is, u
is given by integration against
a continuous Lloc -map) and
1
|
u(ξ)| → 0 as |ξ| → ∞.
Proof. If u ∈ L1 , then |
u(ξ)|, ξ ∈ Rd , is uniformly bounded by uL1 , so u
∈ L1loc .
Fix ε > 0 and let ϕ ∈ S with u − ϕL1 < ε/2, which exists by Theorem 1.22. Since
∈ S, we may now find a M > 0 large enough such that |ϕ(ξ)|
ϕ ≤ ε/2 for ξ ∈ Rd with
|ξ| > M . Then, for |ξ| > M it holds that
ε ε
|
u(ξ)| ≤ |
u(ξ) − ϕ(ξ)|
+ |ϕ(ξ)|
≤ u − ϕL1 + |ϕ(ξ)|
< + = ε.
2 2
This implies that |
u(ξ)| → 0 as |ξ| → ∞.
The continuity follows from a straightforward argument on the continuity of the integral
(exercise).
Finally, one may check that analogous rules to the ones above hold for the Fourier
transform in S ′ (where they make sense). We will show the convolution rule explicitly in
the next section.
2.3 Convolution
Let u ∈ S ′ and ψ ∈ S. Then the convolution of u with ψ is defined as before,
u ψ, ϕ := u, ψ ϕ , ϕ ∈ S.
Lemma 2.11. For u ∈ S ′ and ψ ∈ S it holds that u ψ ∈ S ′ and u ψ is, in fact, given
by integration against a smooth function, namely x → 〈u, ψ(x − )〉.
Proof. We have for any ϕ ∈ S that
u ψ, ϕ ≤ C α,β (ψ ϕ).
|α|,|β|≤N
2.3. CONVOLUTION 21
For α, β ∈ Nd0 with |α|, |β| ≤ N it holds that
α,β (ψ ϕ) = xα ∂ β (ψ ϕ)∞
≤ (1 + | |)N (∂ β ψ)
ϕ
≤ C (1 + | |) |∂ ψ| (1 + | |)N |ϕ| ∞ ,
∞
N β
where in the last line we have used the elementary fact that
(1 + |x|)N ≤ C(1 + |x − y|)N · (1 + |y|)N , y ∈ Rd .
We can now apply Young’s inequality for convolutions (in the case of exponents 1 and ∞)
to estimate
α,β (ψ ϕ) ≤ C (1 + | |)N ∂ β ψL1 · (1 + | |)N |ϕ|∞ .
Since the above L1 -norms are bounded uniformly for any β as above, we can combine this
with the above estimates to obtain
u ψ, ϕ ≤ C (1 + | |)N |ϕ| .
∞
It is now elementary to see that the right-hand side can be estimated as in (2.1). Hence,
u ψ ∈ S ′ . The other statements have already been shown in Lemma 1.10.
We may then also record the following convolution rule (in analogy to the convolution
rule for the Fourier transform in S):
Lemma 2.12. Let u ∈ S ′ and ψ ∈ S. Then, F[u ψ] = u
· ψ.
ψ (exercise) and we may thus compute
Proof. For any ϕ ∈ S we have F[ϕ · ψ] = ϕ
= u, ψ ϕ
F[u ψ], ϕ = u ψ, ϕ = u, ψ ϕ
= u, F[ψ · ϕ] = u ϕ .
· ψ,
This already shows the claim.
We also show that (distributional) differential operators commute with convolutions.
Lemma 2.13. Let u ∈ S ′ and ψ ∈ S. Then, for any multi-index β ∈ Nd0 it holds that
∂ β (u ψ) = (∂ β u) ψ = u (∂ β ψ).
Proof. Let ϕ ∈ S and observe, using the classical rules for the convolution,
β
∂ (u ψ), ϕ = (−1)|β| u, ψ (∂ β ϕ)
= (−1)|β| u, ∂ β (ψ ϕ)
= (∂ β u) ψ, ϕ .
22 CHAPTER 2. TEMPERED DISTRIBUTIONS
This shows the first equality. On the other hand,
β
∂ (u ψ), ϕ = (−1)|β| u, ψ (∂ β ϕ)
ϕ
= (−1)|β| u, (∂ β ψ)
= u, ∂ βψ ϕ
= u (∂ β ψ), ϕ .
This shows the second equality.
Combining the above results we can now come back to one of the fundamental prob-
lems mentioned in the introduction, namely solutions to the Poisson equation (on the whole
space).
Example 2.14. According to Example 1.18, for the Newtonian potential
1
|x|2−d if d ∕= 2,
Γ(x) := d(d−2)ω
1
d
− 2π log |x| if d = 2,
we have
−∆Γ = δ0 .
Then, for ψ ∈ S,
−∆(Γ ψ) = (−∆Γ) ψ = δ0 ψ = ψ.
So, u := Γ ψ solves the Poisson equation −∆u = ψ. Using Lemma 2.11, one can then
write
u(x) = Γ(x − y)ψ(y) dy.
One may further ask whether this integral exists (in some sense) also for ψ ∈ Lp , 1 < p <
∞. This is quite a delicate matter, which we will investigate in subsequent chapters.
Chapter 3
Calderón–Zygmund Theory
One class of operators that is encountered in many problems in Analysis is the following:
A (singular integral) convolution operator T is a C-linear map that is defined for ϕ ∈ S
via
T ϕ := K ϕ ∈ S ′ ,
where the kernel K ∈ S ′ is a tempered distribution that coincides with a locally integrable
function K = K(x) ∈ L1loc (Rd \ {0}; C) away from the origin. This statement is to be
understood in the sense that K = K(x) on Rd \ {0}, as defined in Section 1.2, where
we write “K(x)” for the function that the distribution K agrees with away from the origin
(this abuse of notation should not cause any confusion). It is important to realize that K
is not determined solely by the function K(x). From Lemma 2.11 we know that T ϕ is in
S ′ ∩ C∞ (Rd ; C) for any ϕ ∈ S.
The goal of Calderón–Zygmund theory is to establish boundedness properties of such
convolution operators T = K with respect to Lp -norms (and other norms), which then
allows one to extend the domain of T by a simple density argument.
3.1 Weak-type and strong-type bounds
Let u : Rd → C be measurable. For 1 ≤ p ≤ ∞ we define the Lebesgue norms uLp and
Lorentz quasi-norms uLp,∞ of u as follows:
1/p
p
uLp := |u(x)| dx , 1 ≤ p < ∞,
uL∞ := ess sup |u(x)|,
x∈Rd
1/p
uLp,∞ := sup λ · Ld x ∈ Rd : |u(x)| > λ , 1 ≤ p < ∞,
λ>0
uL∞,∞ := uL∞ .
23
24 CHAPTER 3. CALDERÓN–ZYGMUND THEORY
The Lorentz quasi-norms satisfy are positive definite and positively 1-homogeneous (namely,
α Lp,∞ = |α| · Lp,∞ for α ∈ R), but the triangle inequality holds only up to a dimen-
sional constant (exercise),
u + vLp,∞ ≤ Cd,p (uLp,∞ + vLp,∞ ).
The vector spaces of measurable maps, quotiened by equality almost everywhere, corre-
sponding to these norms or quasi-norms are denoted by Lp := Lp (Rd ; C) and Lp,∞ :=
Lp,∞ (Rd ; C), respectively; spaces on domains are defined analogously. Let us remark that
for the Lorentz spaces the usual Banach space theorems remain valid despite the weaker
notion of norm, which justifies to call these spaces “quasi-Banach spaces”.
A (not necessarily linear) operator T between Lp and Lq , where 1 ≤ p, q ≤ ∞, satisfy-
ing
T [u]Lq ≤ CuLp , u ∈ Lp ,
for a constant C > 0 is said to be of strong type (p, q). If instead T maps Lp to Lq,∞ and
T [u]Lq,∞ ≤ CuLp , u ∈ Lp ,
for a 1 ≤ p ≤ ∞ and a constant C > 0 then it is said to be of weak type (p, q). If T is
defined first only for ϕ ∈ S (or ϕ ∈ D) and satisfies one of the above bounds for such u,
then it is also said to be of strong or weak type (p, q). We can then always extend it to Lp
by a standard density argument; this will often be done implicitly.
Lemma 3.1. If T is of strong type (p, q), where 1 ≤ p, q ≤ ∞, then it is also of weak type
(p, q).
Proof. Assume 1 ≤ q < ∞, as the case q = ∞ is trivial. For u ∈ Lp , by Markov’s
inequality,
1/q
d
1/q uqLq
sup λ · L d
x ∈ R : |u(x)| > λ ≤λ = uLq .
λ>0 λq
Hence, Lq,∞ ≤ Lq . This already shows the claim.
An operator T taking values in the C-valued measurable maps on Rd (or another mea-
sure space) is called sublinear if, for almost every x ∈ Rd ,
|T [u + v](x)| ≤ |T [u](x)| + |T [v](x)|, |T [αu](x)| = |α| · |T [u](x)|,
where u, v are from the domain of T and α ∈ C. A linear operator T is sublinear, as is |T |.
A cornerstone of Harmonic Analysis is the following Marcinkiewicz interpolation
theorem (sometimes called the “real method of interpolation”):
3.1. WEAK-TYPE AND STRONG-TYPE BOUNDS 25
Theorem 3.2. Let 1 ≤ p0 < p1 ≤ ∞. Let T be a sublinear operator defined on
Lp0 + Lp1 := { u0 + u1 : ui ∈ Lpi , i = 0, 1 }
and taking values in the C-valued measurable maps on Rd . Assume furthermore that T is
of weak type (pi , pi ) for i = 0, 1. Then T is of strong type (p, p) for any p0 < p < p1 .
Proof. We will only prove this theorem in the case p1 < ∞; the case p1 = ∞ is similar.
Let u ∈ Lp and fix δ > 0 (to be determined later). Set, for α > 0,
0 0 if |u(x)| ≤ δα,
uα (x) :=
u(x) if |u(x)| > δα,
u(x) if |u(x)| ≤ δα,
u1α (x) :=
0 if |u(x)| > δα.
Then,
u = u0α + u1α , where u0α ∈ Lp0 , u1α ∈ Lp1 .
Indeed,
u0α pL0p0 = |u(x)|p0 −p · |u(x)|p dx ≤ (δα)p0 −p upLp
{|u|>δα}
and, similarly, u1α pL1p1 ≤ (δα)p1 −p upLp .
Define the distribution function of a measurable map v : Rd → C as
dv (α) := Ld x ∈ Rd : |v(x)| > α , α > 0,
for which the following identity holds:
∞
p
vLp = p αp−1 dv (α) dα. (3.1)
0
By the sublinearity of T we have |T [u]| ≤ |T [u0α ]| + |T [u1α ]| and thus
dT [u] (α) ≤ dT [u0α ] (α/2) + dT [u1α ] (α/2).
Noting that dv (λ) ≤ vpLp,∞ /λp , and applying the weak-type (pi , pi ) bound for T , say
with constants Ai (i = 0, 1), we then obtain
1
T [u0α ]p0p ,∞ + 1
T [u1α ]p1p ,∞
dT [u] (α) ≤ L L 1
(α/2)p0 0
(α/2)p1
Ap00 p0 Ap11
≤ |u(x)| dx + |u(x)|p1 dx.
(α/2)p0 {|u|>δα} (α/2)p1 {|u|≤δα}
26 CHAPTER 3. CALDERÓN–ZYGMUND THEORY
Via (3.1) we thus get
∞
T [u]pLp ≤ p(2A0 ) p0
αp−1−p0
|u(x)|p0 dx dα
0 {|u|>δα}
∞
p1 p−1−p1
+ p(2A1 ) α |u(x)|p1 dx dα.
0 {|u|≤δα}
Using Fubini’s theorem,
∞
p−1−p0
α |u(x)|p0 dx dα
0 {|u|>δα}
|u|/δ
p0
= |u(x)| αp−1−p0 dα dx
0
1 1
= · |u(x)|p0 · |u(x)|p−p0 dx
p − p0 δ p−p0
1 1
= · p−p0 upLp
p − p0 δ
and, likewise,
∞
1
αp−1−p1 |u(x)|p1 dx dα = · δ p−p1 upLp .
0 {|u|≤δα} p1 − p
Combining all the above, we arrive at
(2A0 )p0 1 (2A1 )p1 p1 −p
T [u]pLp ≤p · p−p0 + ·δ upLp .
p − p0 δ p1 − p
Now choose δ > 0 optimally, namely such that
(2A0 )p0
= (2A1 )p1 δ p1 −p .
δ p−p0
Upon plugging this into the preceding estimate and performing some computations, we see
that indeed T [u]Lp ≤ AuLp for
1 1 1 1
1/p p − p1 p0 − p
p p 1 − 1
p0 p1
1 − 1
p0 p1
A := 2 + A0 A1 .
p − p0 p1 − p
This concludes the proof in the case p1 < ∞.
There are many more interpolation results between operators, most notably the “com-
plex method of interpolation”, see for instance [2, Section 1.3] for some pointers to standard
results.
3.2. CONVOLUTION OPERATORS 27
3.2 Convolution operators
Let us now turn back to the theory of convolution operators.
Example 3.3. The canonical example of a convolution operator is given by the Hilbert
transform, which for ϕ ∈ S(R) is defined as
Hϕ := KH ϕ,
1
where KH := π · PV x1 is the Hilbert kernel from Example 1.4. Written out, this means
1 ϕ(x − y)
(Hϕ)(x) = · lim dy, x ∈ R.
π ε→0 {|y|>ε} y
The Hilbert transform and its multi-variate counterparts, the Riesz transforms (introduced
in Example 4.4), are very important in a variety of contexts, most notably in the theory of
PDEs, probability theory, and complex analysis.
We can often easily infer that such an operator is of strong type (2, 2) from the following
criterion:
Lemma 3.4. If T is a convolution operator so that its kernel K has as its Fourier transform
L∞ < ∞, then T is of strong type
a tempered distribution given by a bounded map, i.e., K
(2, 2) and can be uniquely extended to map L to itself.
2
Proof. By Lemma 2.12 and Plancherel’s relation we have that for ϕ ∈ S it holds that
· ϕ
T ϕL2 = F[K ϕ]L2 = K L∞ · ϕ
L2 ≤ K L∞ · ϕL2 .
L2 = K
Thus, T ϕ ∈ L2 and T L2 →L2 ≤ K L∞ . Since T is linear we may extend it uniquely to
all of L with the same operator norm bound.
2
Example 3.5. Consider again the Hilbert transform from Example 3.3. We computed in
Example 2.7 that K H = −i sgn, which is bounded. The previous result then gives that the
Hilbert transform is of strong type (2, 2).
In the following we want to investigate the boundedness properties of a convolution
operator T on other Lp spaces. To this aim we first write the distributional convolution via
an integral representation for any
u ∈ L2c := u ∈ L2 : ess supp u compact ,
where
ess supp u := Rd \ U ⊂ Rd open : u = 0 a.e. in U
is the essential support of u.
28 CHAPTER 3. CALDERÓN–ZYGMUND THEORY
Lemma 3.6. Let T be a (singular integral) convolution operator with kernel K ∈ S ′ that
agrees with a locally integrable function K(x) ∈ L1loc (Rd \ {0}; C) away from the origin.
Assume that T is of strong type (2, 2). Then,
T u(x) = K(x − y)u(y) dy (3.2)
for every u ∈ L2c and almost every x ∈ Rd \ ess supp u.
Proof. Let first u ∈ D and x ∈ Rd \ supp u. Then, by Lemma 1.10 it holds that
T u(x) = 〈K, u(x − )〉.
Let now η ∈ C∞c (R ; [0, 1]) with η = 0 on a neighborhood of the origin and η = 1 on a
d
neighborhood of supp u(x − ). We then have ηK ∈ L1 and thus
T u(x) = ηK, u(x − ) = η(y)K(y)u(x − y) dy = K(y)u(x − y) dy.
A change of variables then shows (3.2) in this case.
For general u ∈ L2c , let Y ⊂ Rd be a compact set containing a neighborhood of
ess supp u. Then, we take a sequence (uj ) ⊂ D with supp uj ⊂ Y (e.g., by mollifica-
tion) and uj → u in L2 , hence also uj → u in L1 . Let now x ∈ X, where X ⊂ Rd \ Y is a
compact set. Denoting by X−Y the indicator function of the set X − Y := { x − y : x ∈
X, y ∈ Y } (that is, X−Y (z) = 1 if z ∈ X − Y and 0 otherwise), we obtain by virtue of
Young’s inequality for convolutions that
x → K(x − y)[u(y) − uj (y)] dy ≤ K X−Y L1 (X) · u − uj L1 ,
L1 (X)
where we have used that we may restrict the integral to y ∈ Y and that (for x ∈ X)
|K(x − y) Y (y)| = |K(x − y) x−Y (x − y)| ≤ |K(x − y) X−Y (x − y)|.
Since K X−Y ∈ L1 and u − uj L1 → 0 as j → ∞, the first part of the proof implies that
T uj (x) = K(x − y)uj (y) dy → K(x − y)u(y) dy
in L1 , and then also, up to a subsequence (not relabelled), for almost every x ∈ X. On
the other hand, T uj → T u in L2 since T is of strong type (2, 2). Since L2 -convergence
implies pointwise almost everywhere convergence on the compact set X, we see that (3.2)
also holds for general u ∈ L2c and almost every x ∈ X. Finally, let X approach Rd \ Y and
Y approach ess supp u to see that (3.2) in fact holds for all x ∈ Rd \ ess supp u.
The condition (3.2) by itself also entails that a convolution operator agrees with a locally
integrable function away from the origin:
3.2. CONVOLUTION OPERATORS 29
Lemma 3.7. Let T = K be convolution operator with kernel an arbitrary tempered
distribution K ∈ S ′ such that (3.2) holds for a locally integrable function K(x) ∈ L1loc (Rd \
{0}; C) (which is not a-priori assumed to be related to K). Then, K agrees with K(x) away
from the origin.
Proof. Let ϕ ∈ S with supp ϕ ⊂ Rd \ {0}. Let (δ )δ>0 ⊂ C∞ c (B1 ) be a family of
∗ ′
mollifiers, for which it holds that δ ⇀ δ0 in S as δ ↓ 0 (see Example 1.19). Thus,
T ϕ, δ = K ϕ, δ = K, ϕ δ → K, ϕ
since ϕ δ → ϕ in S as δ ↓ 0. On the other hand, T ϕ ∈ C∞ (Rd ; C) (see Lemma 2.11),
so for a cut-off function η ∈ C∞c (R ; [0, 1]) with η = 1 on B1 we have ηT ϕ ∈ S and
d
T ϕ, δ = δ , ηT ϕ → δ0 , ηT ϕ = T ϕ(0).
Furthermore, 0 ∈ Rd \ supp ϕ, whereby (3.2) implies
T ϕ(0) = K(−y)ϕ(y) dy.
Hence,
〈K, ϕ〉
=
K(y)ϕ(y) dy
for all ϕ ∈ S with supp ϕ ⊂ Rd \ {0}, which implies the claim of the lemma via a change
of variables.
Remark 3.8. Before we come to the central results of this chapter, it is worthwhile to
pause and discuss a question of terminology: Some texts on Harmonic Analysis (e.g., [8,9])
work not with the convolution operators defined above, but with concrete singular integral
operators. One such type of operators is defined as follows: Let K ∈ L1loc (Rd \ {0}) satisfy
the following conditions:
(i) |K(x)| ≤ C|x|−d ;
(ii) K(x) dx = 0 for all 0 < r < s < ∞;
{r<|x|<s}
(iii) |K(x)−K(x−y)| dx ≤ B for all y ∈ Rd \{0}, where B > 0 is a constant.
{|x|>2|y|}
Then, the Calderón–Zygmund (singular integral) operator with kernel K is given as
TK ϕ(x) := lim K(x − y)ϕ(y) dy, x ∈ Rd ,
ε→0 {ε≤|x−y|≤ε−1 }
for ϕ ∈ S. In contrast, note that our (singular integral) convolution operators may also have
a part at the origin, e.g., a Dirac mass. Indeed, a number of natural operators, most notably
30 CHAPTER 3. CALDERÓN–ZYGMUND THEORY
the Fourier multiplier operators we will encounter in Chapter 4, are convolution operators,
but not always Calderón–Zygmund (singular integral) operators in the above sense as they
may indeed contain a part at the origin; see for instance [2, Proposition 2.4.7]. On the other
hand, more general classes of kernels (not necessarily satisfying (i) or (ii)) can be considered
as convolution-type operators; see, e.g., Section 4.3.2 in [2]. One may, however, observe
that in all of the following the essential difficulty is estimating the part of a convolution
operator that can be expressed via singular integral, so all these approaches yield more or
less the same theory. Later, in Chapter 6 we will also define more general singular integral
operators, where the kernel may depend on x as well.
3.3 Boundedness of convolution operators
The result in Lemma 3.4 for the strong-type (2, 2) bound is very simple, but it can be used
as one “anchor point” to invoke the Marcinkiewicz interpolation theorem. The other one
will be a weak-type (1, 1) bound. This will require some preparation, most notably the
following Calderón–Zygmund decomposition at height λ > 0:
Lemma 3.9. Let u ∈ L1 and λ > 0. Then, one can decompose u as
u=g+b
with g, b ∈ L1 such that
d
ess supp g ⊂ R \ Q , gL1 ≤ uL1 (supp g) , gL∞ ≤ λ
Q∈Q
and
b= Q u,
Q∈Q
where Q is an (at most countable) collection of disjoint dyadic cubes of the form
Q = [z2ℓ , (z + 1)2ℓ ) := [z1 2ℓ , (z1 + 1)2ℓ ) × · · · × [zd 2ℓ , (zd + 1)2ℓ )
with ℓ ∈ Z, z ∈ Zd , such that
λ ≤ − |u| dx ≤ 2d λ for every Q ∈ Q, (3.3)
Q
and
1
Ld Q ≤ uL1 . (3.4)
λ
Q∈Q
3.3. BOUNDEDNESS OF CONVOLUTION OPERATORS 31
Proof. Define for ℓ ∈ Z the collections of dyadic cubes
Dℓ := [z2ℓ , (z + 1)2ℓ ) : z ∈ Zd .
Pick ℓ0 ∈ Z so large that
− |u| dx ≤ λ for every Q ∈ Dℓ0 .
Q
Let Q ∈ Dℓ0 and let Q′ ∈ Dℓ0 −1 be a child of Q, that is Q′ ⊂ Q. We distinguish two cases:
First, if
− |u| dx > λ,
Q′
then
λ < − |u| dx ≤ 2 − |u| dx ≤ 2d λ.
d
Q′ Q
We collect all such Q′ in the collection Qℓ0 −1 .
If, on the other hand,
− |u| dx ≤ λ,
Q′
then we consider for every such Q′ its children Q′′ and repeat the above procedure, where
∈ Dℓ (ℓ = ℓ0 − 1, ℓ0 − 2, . . .) with − |u| dx > λ in
in each step we collect the cubes Q Q
the collection Qℓ .
We set
0 −1
ℓ
Q := Qℓ .
ℓ=−∞
Then, for this Q the first condition (3.3) holds. Let now x0 ∈ Rd \ Q∈Q Q. Then, there
is a decreasing sequence Q1 ⊃ · · · ⊃ Qj ⊃ Qj+1 ⊃ · · · of nested dyadic cubes containing
x0 such that
− |u| dx ≤ λ.
Qj
We now employ Lebesgue’s differentiation theorem for uncentered cubes (for instance,
see [1, Theorem 1.34], which however treats the case of balls, not cubes, but the proof
is the same for cubes). This standard result implies that
lim − |u| dx = |u(x0 )|
j→∞ Qj
32 CHAPTER 3. CALDERÓN–ZYGMUND THEORY
for almost every x0 ∈ Rd . Thus, for such x0 , we have |u(x0 )| ≤ λ. Since furthermore
d
L (Q \ Q) = 0,
Q∈Q
we have indeed decomposed u = g + b with
b= Q u, g := u − b,
Q∈Q
for which
d
ess supp g ⊂ R \ Q , gL1 ≤ uL1 (supp g) , gL∞ ≤ λ
Q∈Q
and (3.3) hold. Finally, we observe
d 1 1
L Q ≤ Ld (Q) ≤ |u| dx ≤ uL1 ,
λ Q λ
Q∈Q Q∈Q Q∈Q
which shows (3.4).
We can now obtain the following weak-type (1, 1) bound:
Lemma 3.10. Let T be a (singular integral) convolution operator with kernel K ∈ S ′
that agrees with a locally integrable function K ∈ L1loc (Rd \ {0}; C) away from the origin.
Assume that T is of strong type (2, 2) with operator norm A > 0, and that K(x) satisfies
Hörmander’s (cancellation) condition,
|K(x) − K(x − y)| dx ≤ B for all y ∈ Rd \ {0} (3.5)
{|x|>2|y|}
with B > 0 a constant. Then, T is also of weak type (1, 1) with bound
T uL1,∞ ≤ C(A2 + B + 1)uL1 , u ∈ L1 ,
where C = C(d) > 0 is a dimensional constant.
Proof. First, take u ∈ D. Let λ > 0 and let u have the Calderón–Zygmund decomposition
u = g + b at height λ, as in the preceding Lemma 3.9. With the collection Q of dyadic
cubes from that lemma set
u1 := g + − u dx Q ,
Q∈Q Q
u2 := uQ , where uQ := u − − u dx Q.
Q∈Q Q
3.3. BOUNDEDNESS OF CONVOLUTION OPERATORS 33
From the properties of this decomposition it follows that u = u1 + u2 and that the following
estimates hold:
u1 L1 ≤ uL1 , u1 L∞ ≤ 2d λ, u2 L1 ≤ 2uL1 .
We now estimate, using Markov’s inequality,
Ld x ∈ Rd : |T u(x)| > λ
d λ λ
≤L d
x ∈ R : |T u1 (x)| > +L d d
x ∈ R : |T u2 (x)| >
2 2
4 λ
≤ 2 T u1 2L2 + Ld x ∈ Rd : |T u2 (x)| > .
λ 2
For the first term we have, using the strong-type (2,2) bound for T ,
4 2 4A2 2 4A2 4 · 2d · A2
T u 1 L 2 ≤ u 1 L 2 ≤ u 1 L ∞ · u1 L1 ≤ uL1 .
λ2 λ2 λ2 λ
√
For the second term, let Q∗ be the dilation of Q by the factor 2 d (and the same center).
Then,
d λ
L d
x ∈ R : |T u2 (x)| >
2
λ
≤ Ld Q∗ + L d x ∈ Rd \ Q∗ : |T u2 (x)| >
2
Q∈Q Q∈Q
2
≤C Ld (Q) + |T u2 | dx
λ Rd \Q∈Q Q∗
Q∈Q
C 2
≤ uL1 + |T uQ | dx,
λ λ Rd \Q∗
Q∈Q
where we have used (3.4) as well as Markov’s inequality. For any Q ∈ Q denote by yQ the
center of Q. If x ∈ Rd \ Q∗ , then, by Lemma 3.6,
T uQ (x) = K(x − y)uQ (y) dy = K(x − y) − K(x − yQ ) uQ (y) dy
Q Q
because uQ has zero average. Hence, using Fubini’s theorem,
|T uQ | dx ≤ |K(x − y) − K(x − yQ )| · |uQ (y)| dy dx
Rd \Q∗ Rd \Q∗ Q
= |K(x − y) − K(x − yQ )| dx · |uQ (y)| dy.
Q Rd \Q∗
Now,
Rd \ Q ∗ ⊂ x ∈ Rd : |x − yQ | > 2|y − yQ | for any y ∈ Q.
34 CHAPTER 3. CALDERÓN–ZYGMUND THEORY
So, from Hörmander’s condition (3.5) we infer (changing variables to x
:= x − yQ , y :=
y − yQ )
|T uQ | dx ≤ |K(
x − y) − K( x · |uQ (
x)| d y + yQ )| dy
Rd \Q∗ Q−yQ {|
x|>2|
y |}
≤B |uQ (y)| dy
Q
≤ 2B Q uL1 .
Combining all the above estimates, we arrive at
CA2 C CB
Ld x ∈ Rd : |T u(x)| > λ ≤ uL1 + uL1 + uL1
λ λ λ
C(A2 + B)
≤ uL1 ,
λ
where C = C(d) > 0 is a dimensional constant. Taking the supremum over all λ > 0,
we obtain the claim for u ∈ D. The extension to u ∈ L1 then follows by a simple density
argument.
We observe, en passant, a stronger condition than (3.5), which is often easier to verify:
Lemma 3.11. Let K ∈ C1loc (Rd \ {0}; C) satisfy the strong Hörmander condition: There
exists a C > 0 with
C
|∇K(x)| ≤ , x ∈ Rd \ {0}. (3.6)
|x|d+1
Then, K satisfies Hörmander’s condition (3.5).
Proof. Let x, y ∈ Rd with |x| > 2|y|. Then, x − ty ∕= 0 for all t ∈ [0, 1]. By the mean
value theorem,
C|y|
|K(x) − K(x − y)| ≤
|x|d+1
for some constant C > 0, which depends on C and the dimension. Then,
1
|K(x) − K(x − y)| dx ≤ C|y| d+1
dx ≤ B
{|x|>2|y|} {|x|>2|y|} |x|
for some B > 0 (depending on C and the dimensions) since the integral can be estimated
by a constant times |y|−1 (exercise). This is (3.5).
We can now state the main result of this chapter, the Calderón–Zygmund theorem:
3.3. BOUNDEDNESS OF CONVOLUTION OPERATORS 35
Theorem 3.12. Let T be a (singular integral) convolution operator with kernel K ∈ S ′
that agrees with a locally integrable function K ∈ L1loc (Rd \ {0}; C) away from the origin.
Assume that T is of strong type (2, 2) and that K(x) satisfies Hörmander’s condition (3.5).
Then, T is of strong type (p, p) for any 1 < p < ∞ and of weak type (1, 1).
Proof. We have already shown the weak-type (1, 1) bound in Lemma 3.10. Since we also
assumed that T is of strong type (2, 2), we obtain the strong type (p, p) bound for 1 < p < 2
directly from the Marcinkiewicz Interpolation Theorem 3.2.
For 2 < p < ∞ we may argue by duality: For ϕ, ψ ∈ S it holds that
T ϕ, ψ = T t ψ, ϕ ,
where T t is the (singular integral) convolution operator with kernel K := K(− ) (as a
tempered distribution). Since T also of strong type (2, 2) and satisfies Hörmander’s condi-
t
tion (3.5) (this is easy to see), we obtain that it is of strong type (q, q) for any 1 < q < 2.
Let p be the conjugate exponent to 1 < q < 2, that is 1/p + 1/q = 1, and denote by Aq
the operator norm of T t as an operator mapping Lq to itself. Then, by Hölder’s inequality,
T ϕ, ψ = T t ψ, ϕ ≤ T t ψLq · ϕLp ≤ Aq ψLq · ϕLp .
Taking the supremum over all ψ ∈ Lq with ψLq ≤ 1 and using that Lq is the dual space
′
to Lq , we obtain that T maps Lp to itself as a bounded operator. Since p runs through the
interval (2, ∞) as q runs through (1, 2), we obtain also for 2 < p < ∞ strong-type bounds
for our original T .
Example 3.13. For the Hilbert transform H = KH from Example 3.3 we have seen
in Example 3.5 that it is of strong type (2, 2). Its kernel, the Hilbert kernel KH = π1 ·
PV x1 , agrees with the locally integrable function KH (x) = πx
1
away from the origin, which
satisfies
′ 1
KH (x) = − , x ∈ R \ {0}.
πx2
So, |KH′ (x)| ≤ C|x|−2 , whereby K satisfies the strong Hörmander condition (3.6) and
H
hence also the normal Hörmander condition (3.5) by virtue of Lemma 3.11. Thus, by the
preceding Calderón–Zygmund theorem, H is of strong type (p, p) for any 1 < p < ∞ and
of weak type (1, 1).
Chapter 4
Fourier Multipliers and
Littlewood–Paley Theory
Consider a map u : Rd → R satisfying the Poisson equation
−∆u = −(∂11 u + · · · + ∂dd u) = f
in the weak sense (see Example 2.14), where, for the moment, we do not worry too much
about the smoothness of u and f . If we Fourier transform this equation and apply the
standard rules of the Fourier transform (in particular, F[∂i u](ξ) = 2πiξi u
(ξ)), then
4π 2 (ξ12 + · · · + · · · ξd2 ) u (ξ) = f(ξ),
(ξ) = 4π 2 |ξ|2 u ξ ∈ Rd .
Let us multiply this equation by ξi ξj /|ξ|2 (i, j ∈ {1, . . . , d}). Then, we obtain
ξi ξj
F[∂i ∂j u](ξ) = 4π 2 ξi ξj u
(ξ) = f (ξ).
|ξ|2
Taking the inverse Fourier transform, this suggests the formal relation
ξi ξj
q f,
∂i ∂j u = m where m(ξ) := .
|ξ|2
In the heuristic that multiplying the Fourier transform with ξik corresponds to (a multiple of)
k-times the i’th derivative, this “Fourier multiplier” m above corresponds to no derivatives.
Hence, one may wonder if the convolution operator with kernel m q has good mapping prop-
erties, e.g., L to L . This is indeed true for 1 < p < ∞, as we will see in this section via
p p
the so-called Littlewood–Paley theory. This yields that solutions u have Lp -bounded second
derivatives (note that a-priori the PDE only implies that the sum ∆u = ∂11 u + · · · + ∂dd u
of second derivatives is as smooth as f is).
37
38 CHAPTER 4. FOURIER MULTIPLIERS AND LITTLEWOOD–PALEY
4.1 The Mikhlin multiplier theorem
Let m : Rd \ {0} → C with at most polynomial growth. Then, we define the corresponding
(Fourier) multiplier operator for ϕ ∈ S as follows:
Tm ϕ(x) := m(ξ)ϕ(ξ) e2πix·ξ dξ, x ∈ Rd .
This definition is well-defined for every x since ξ → m(ξ)ϕ(ξ)
is integrable. Moreover,
Tm ϕ = F −1 [m(ξ)ϕ(ξ)]
q ϕ
=m
is well-defined as an element of S ′ .
Recall the convenient notation “≲”, which means that the left-hand side is estimated by
the right-hand side up to an (omitted) constant. Further, “∼” means that the left-hand side
and right-hand side are proportional up to constants, that is, “≲” and “≳” hold.
The crucial idea of Littlewood–Paley theory is to “localize” the Fourier transform in
dyadic annuli around |ξ| ∼ 2k . The localization relies on the following dyadic lemma:
Lemma 4.1. There is a radial η ∈ C∞
c (R ; [0, 1]) with supp η ⊂ R \ {0} such that, for
d d
a constant C > 0,
∞
η(2−k ξ) ≤ C, ξ ∈ Rd \ {0},
k=−∞
where atmost two terms in this sum are nonzero for any given ξ ∕= 0. In fact, one may
require ∞ −k
k=−∞ η(2 ξ) = 1 on R \ {0}.
d
Proof. Take a radial χ ∈ C∞
c (R ; [0, 1]) with χ(ξ) = 1 for |ξ| ≤ 1 and χ(ξ) = 0 for
d
|ξ| ≥ 2. Then, for
η(ξ) := χ(ξ) − χ(2ξ)
we have all the claimed properties. In particular, for any N ∈ N,
N
η(2−k ξ) = χ(2−N ξ) − χ(2N +1 ξ).
k=−N
For given ξ ∕= 0 we may choose N large enough (depending on ξ) so that χ(2−N ξ) = 1
and χ(2N +1 ξ) = 0, so the claims hold.
The following is known as the Mikhlin multiplier theorem and is of great use in many
branches of Analysis:
Theorem 4.2. Let m : Rd \ {0} → C with the property that
|∂ α m(ξ)| ≤ B|ξ|−|α| , ξ ∈ Rd \ {0}, (4.1)
4.1. THE MIKHLIN MULTIPLIER THEOREM 39
for all multi-indices α ∈ Nd0 with |α| ≤ d + 2 and a constant B > 0. Then, the correspond-
ing Fourier multiplier operator Tm is of strong type (p, p) for 1 < p < ∞ and of weak type
(1, 1). The operator Tm can thus be extended to map Lp , 1 < p < ∞, to itself and L1 to
L1,∞ .
Proof. We will show that Tm ϕ as defined in (4.1) for ϕ ∈ S satisfies a strong type (p, p)
bound for 1 < p < ∞ and the weak type (1, 1) bound. The extension result then follows
by density.
Step 1. Let η be as in Lemma 4.1 and set, for k ∈ Z,
mk (ξ) := η(2−k ξ)m(ξ), ξ ∈ Rd .
This multiplier is supported in an annulus around |ξ| ∼ 2k . As usual, we denote by Tmk the
Fourier multiplier operator with multiplier mk . Then, using Plancherel’s theorem, we see
that
Tm ϕ = Tm k ϕ
k
with the convergence taking place in L2 . Denote by Kk := m
q k the inverse Fourier transform
of mk . Since mk ∈ L , we have
1
Kk (x) = mk (ξ) e2πix·ξ dξ. (4.2)
Step 2. We next claim the following estimates for all x ∈ Rd \ {0} and k ∈ Z:
|Kk (x)| ≲ B2kd , (4.3)
|Kk (x)| ≲ B2−2k |x|−(d+2) , (4.4)
|∇Kk (x)| ≲ B2 k(d+1)
, (4.5)
|∇Kk (x)| ≲ B2−k |x|−(d+2) . (4.6)
In (4.3)–(4.6) the constant depends only on the dimension d.
First, the bound (4.3) follows directly from (4.2) since the support of mk has measure
of order 2kd , up to a dimensional constant. For (4.5) we use the same argument applied to
the representation
∇Kk (x) = mk (ξ)(2πiξ) e2πix·ξ dξ, (4.7)
which follows from the rules for the Fourier transform. The factor ξ gives the additional
term of order 2k in the estimate.
For (4.4) we first note the formula
−ix
· ∇ξ e2πix·ξ = e2πix·ξ .
2π|x|2
40 CHAPTER 4. FOURIER MULTIPLIERS AND LITTLEWOOD–PALEY
Applying this d + 2 times to (4.2) and integrating by parts, we obtain
d+2
−ix
Kk (x) = mk (ξ) · ∇ξ e2πix·ξ dξ
2π|x|2
d+2
ix
= · ∇ξ mk (ξ) e2πix·ξ dξ.
2π|x|2
Thus,
1
|Kk (x)| ≲ · B2−k(d+2) · 2kd = B2−2k |x|−(d+2) ,
|x|d+2
where the second term comes from applying (4.1) to ∇d+2 ξ mk (ξ) and that mk is supported
around ξ ∼ 2 ; the third term is again an estimate on the size of the support of mk (ξ). This
k
is (4.4).
The remaining estimate (4.6) follows in a similar way, but now using (4.7) and recog-
nizing that (4.1) implies
|∂ α [ξm(ξ)]| ≲ B|ξ|−|α|+1 , ξ ∕= 0,
for all multi-indices α ∈ Nd0 with |α| ≤ d + 2.
Step 3. We now estimate for x ∈ Rd \ {0},
|Kk (x)| ≲ B 2kd + 2−2k |x|−(d+2) ≲ B|x|−d
k k : 2k <|x|−1 k : 2k ≥|x|−1
where we applied (4.3) in the first sum and (4.4) in the second sum, and also used that a
geometric series is comparable to its largest term (with the factor of comparison depending
on the ratio of the terms). Similarly, we obtain
|∇Kk (x)| ≲ B|x|−(d+1) .
k
With these estimates, the Weierstraß M-test implies that the series
x → Kk (x), x ∈ Rd \ {0},
k
converges in C1loc (Rd \ {0}; C) to a function K ∈ C1loc (Rd \ {0}; C). By construction,
|K(x)| ≲ B|x|−d and |∇K(x)| ≲ B|x|−(d+1) . (4.8)
Step 4. Next, we will show that K is in fact the restriction of a tempered distribution to
Rd \ {0}. Since m ∈ S ′ (this follows from (4.1)), also m
q ∈ S ′ and thus
q ϕ ∈ S ′,
Tm ϕ = m ϕ ∈ S.
4.1. THE MIKHLIN MULTIPLIER THEOREM 41
Moreover, Tm u is also defined for u ∈ L2 and bounded from L2 to L2 by Plancherel’s
relation (cf. Lemma 3.4 for a similar argument). Thus, T is of strong type (2, 2). We will in
the following show that
Tm u(x) = K(x − y)u(y) dy (4.9)
for every u ∈ L2c and almost every x ∈ Rd \ ess supp u, where K is given above. Then,
Lemma 3.7 will imply that K is the restriction of the tempered distribution m
q away from
the origin.
So, let u ∈ L2c and fix ψ ∈ D with supp ψ ∩ ess sup u = ∅. Starting from (4.2), we
compute
Kk (x − y)u(y) dy ψ(x) dx
= mk (ξ) e2πi(x−y)·ξ u(y) ψ(x) dξ dy dx
2πix·ξ
= mk (ξ)
u(ξ) e dξ ψ(x) dx
= (Tmk u)(x) ψ(x) dx. (4.10)
Since the (essential) supports of ψ and u are disjoint, they have a strictly positive distance
and hence (by the same arguments as in Step 2),
|Kk (x − y)| ≤ C if x ∈ supp ψ and y ∈ ess supp u.
k
Moreover, (x, y) → u(y) ψ(x) is integrable in Rd × Rd . Thus, we may use the dominated
convergence theorem to see
Kk (x − y)u(y) dy ψ(x) dx
k
= Kk (x − y)u(y) dy ψ(x) dx
k
= K(x − y)u(y) dy ψ(x) dx (4.11)
On the other hand, we observe by the definition of the distributional Fourier transform and
the dominated convergence theorem that
N
N
Tmk u, ψ = mk (ξ) q dξ →
u(ξ)ψ(ξ) m(ξ) q dξ = Tm , ψ
u(ξ)ψ(ξ)
k=−N k=−N
42 CHAPTER 4. FOURIER MULTIPLIERS AND LITTLEWOOD–PALEY
as N → ∞. Hence,
(Tmk u)(x) ψ(x) dx = (Tm u)(x) ψ(x) dx (4.12)
k
Thus, combining (4.10), (4.11), (4.12) we have shown that
K(x − y)u(y) dy ψ(x) dx = (Tm u)(x) ψ(x) dx.
Varying ψ, we obtain (4.9).
Step 5. We have so far shown that Tm is convolution operator with a kernel m,
q which agrees
with a locally integrable function K(x) away from the origin. Moreover, Tm is of strong
type (2, 2) and the kernel K(x) satisfies the strong Hörmander condition, see Lemma 3.11,
as was proved in (4.8). Thus, said lemma yields that also the standard Hörmander con-
dition (3.5) holds. The claim of our theorem then follows from the Calderón–Zygmund
Theorem 3.12. In fact, it is only (4.9) that is used in the proof of Theorem 3.12, so the
invokation of Lemma 3.7 in Step 4 was in fact redundant.
Example 4.3. In Example 2.7 we showed that K H = −i sgn, which satisfies (4.1) (exer-
cise). Thus, the Hilbert transform H = KH introduced in Example 3.3 is of strong type
(p, p) for any 1 < p < ∞ and of weak type (1, 1). We have already seen this result in
Example 3.13 as a consequence of the Calderón–Zygmund Theorem 3.12.
Example 4.4. Define for j = 1, . . . , d the j’th Riesz transform Rj as the Fourier multi-
plier operator with multiplier −iξj /|ξ|, that is,
−1 iξj
Rj ϕ := F −
ϕ(ξ) , ϕ ∈ S.
|ξ|
Then, the Mikhlin Multiplier Theorem 4.2 yields that Rj extends to an operator that is
bounded from Lp to itself for 1 < p < ∞ (and Rj is also of weak type (1, 1)).
Example 4.5. Another important class of Fourier multipliers are those m ∈ C∞ (Rd \{0})
that are positively homogeneous of degree 0, meaning that
m(λξ) = m(ξ), λ > 0, ξ ∈ Rd \ {0}.
It can be checked easily that such m satisfy (4.1).
4.2 Dyadic decomposition
In all of the following we assume that we are given an η ∈ C∞
c (R ; [0, 1]) with supp η ⊂
d
R \ {0} such that with
d
ηk (ξ) := η(2−k ξ) ξ ∈ Rd , k ∈ Z,
4.2. DYADIC DECOMPOSITION 43
it holds that
∞
ηk (ξ) = 1, ξ ∈ Rd \ {0},
k=−∞
where at most two terms in this sum are nonzero for any given x ∕= 0. The existence of such
an η is guaranteed by Lemma 4.1.
We define the dyadic operators
∆k u := Tηk u
as the Fourier multiplier operator with multiplier ηk , where u ∈ S or u more general if we
consider the extension of ∆k to, say, Lp spaces via the Mikhlin Multiplier Theorem 4.2.
The idea here is to split the frequency domain into “almost orthogonal” pieces. Note that
∞
Id = ∆k
k=−∞
as an identity of operators on S (or Lp , where 1 ≤ 1 < ∞). Furthermore, we define the
following operators for k ∈ Z:
∆≤k := ∆ℓ , ∆≥k := ∆ℓ .
ℓ≤k ℓ≥k
The following result (sometimes associated with Bernstein) shows that for frequency-
localized maps, all Lp -norms can be estimated by one another.
Lemma 4.6. Let 1 ≤ p ≤ q ≤ ∞ and k ∈ Z. Then, for all u ∈ Lp :
kd 1 − 1
∆k uLq ≲ 2 p q ∆k uLp , (4.13)
kd p1 − 1q
∆≤k uLq ≲ 2 ∆≤k uLp . (4.14)
Here, the proportionality constants depend on p, q and d.
Proof. It suffices to show the estimates for ϕ ∈ S, then they also hold for u ∈ Lp by density.
Let η be another function like in Lemma 4.1 (with C = 2) with the property that η = 1
on supp η. For the corresponding dyadic operators ∆ k this implies
k ∆k = ∆k .
∆
Denote by K k := F −1 [ k , so that ∆
ηk ] the kernel corresponding to ∆ k ϕ = K k ϕ (ϕ ∈ S).
Since ηk ∈ S, the kernel K k coincides with a Schwartz function, K k ∈ S and the Fourier
transform can be understood pointwise. By the scaling properties of the Fourier transform,
k (x) = 2kd K(2
K k x), x ∈ Rd ,
44 CHAPTER 4. FOURIER MULTIPLIERS AND LITTLEWOOD–PALEY
:= F −1 [
where K η ]. We then have
k ) (∆k ϕ).
k ∆k ϕ = 2kd K(2
∆k ϕ = ∆
1 1 1
Applying Young’s inequality for convolutions with 1 ≤ r ≤ ∞ such that 1 + q = r + p
we obtain the estimate
k )Lr · ∆k ϕLp ≲ 2kd− kd
∆k ϕLq ≤ 2kd K(2 r ∆ ϕ p ,
k L
where for the second inequality we employed a simple scaling argument (exercise) to see
that
k )Lr = 2− kd
K(2 Lr .
r K
1 1
As kd − kdr = kd p − q , the first inequality (4.13) follows.
The second inequality (4.14) follows from the first together with some straightforward
estimates (exercise).
4.3 The square function
Denote by η̄ the indicator function of the annulus with radius between 1/2 and 1, that is,
1 if 12 < |ξ| ≤ 1,
η̄(ξ) =
0 otherwise.
Consider the corresponding multiplier operators ∆ ¯ k := Tη̄(2−k ξ) . Note that the multipliers
¯
are bounded, so the multiplier operator ∆k are bounded from L2 to itself. For u ∈ L2 , we
¯ k u and ∆
have that ∆ ¯ ℓ u are orthogonal for k ∕= ℓ since
¯ ¯ ¯ ¯
∆k u, ∆ℓ u = ∆k u · ∆ℓ u dx = η̄(2−k ξ) u(ξ) · η̄(2−ℓ ξ)
u(ξ) dξ = 0.
Then, by standard facts about L2 -orthogonality,
∞
u2L2 = ¯ k u2 2 = S̄u2 2 ,
∆ u ∈ L2 ,
L L
k=−∞
where we have defined the (rough) “square function” of u via
∞
1/2
S̄u := ¯ k u|2
|∆ .
k=−∞
The last equality above follows from an exchange of integral and sum, which is permitted
by the monotone convergence theorem.
4.3. THE SQUARE FUNCTION 45
The above relation thus equates the L2 -norm with the infinite sum over the L2 -norms
of the frequency-localized maps ∆ ¯ k u. One may ask if an analogous relationship holds in
L . This turns out to be false if d ≥ 2, but only because our cut-off function η̄ is too rough.
p
In fact, this is related to deep facts concerning the ball multiplier u → F −1 [ B u] for any
ball B, which turns out to be bounded in L only if p = 2 (this is a famous result due to
p
Fefferman). We have also seen in the Mikhlin multiplier theorem that some smoothness of
multipliers seems to be required.
Hence, given a fixed family (∆k )k∈Z of dyadic operators for a function η as in Lemma 4.1,
for u ∈ Lp , where 1 < p < ∞, we define the square function Su of u as follows:
∞
1/2
Su(x) := |(∆k u)(x)|2 , x ∈ Rd .
k=−∞
Then, the following Littlewood–Paley theorem holds true:
Theorem 4.7. Let 1 < p < ∞. For any u ∈ Lp it holds that Su ∈ Lp and
SuLp ∼ uLp , (4.15)
where the constant only depends on d and p.
For the proof we will employ a technique called randomization, which is also very
useful in other contexts. To set this up, we will need the following probabilistic result,
which is called Khinchin’s inequality:
Lemma 4.8. Let N ∈ N and let ak ∈ C, where k ∈ {−N, −N + 1, . . . , N − 1, N }. Also
let σk be independent random variables taking the values −1 and +1 with probability 1/2
each. Then, for 1 ≤ p < ∞,
N p 1/p 1/2
N
E σk a k ∼ |ak |2 ,
k=−N k=−N
where E denotes the expectation operator. Moreover, with uk ∈ Lp for k ∈ {−N, −N +
1, . . . , N − 1, N } it holds that
p 1/p
N 1/2
N
E
σk uk
∼ |uk | 2 .
k=−N Lp
k=−N
Lp
In both statements the constants depend on p only.
Proof. Since the proof of the first assertion is purely probabilistic, we omit it here; see for
instance [8, Lemma 5.5]. The second assertion follows from the first with ak := uk (x)
(x ∈ Rd ) by raising both sides to the power p, integrating in x, noticing that the integral in
x commutes with the expectation operator E, and finally taking the p’th root.
46 CHAPTER 4. FOURIER MULTIPLIERS AND LITTLEWOOD–PALEY
Proof of Theorem 4.7. We will first show the estimates in (4.15) for u ∈ S and then extend
this to general maps by density.
Step 1. Let σk , k ∈ Z, be independent random variables taking the values −1 and +1 with
probability 1/2 each. We first observe that for the ηk (k ∈ Z) defined at the beginning of
this section, it holds that
|∂ α ηk (ξ)| ≲ 2−k|α| · |(∂ α η)(2−k ξ)| ≲ |ξ|−|α| , ξ ∈ Rd \ {0},
for any multi-index α ∈ Nd0 . Note that the constant hidden in the “≲” here depends on α,
but we will in fact only use finitely many such α, so this does not matter. Thus, the operators
N
σk ∆k
k=−N
satisfy the assumptions of the Mikhlin Multiplier Theorem 4.2 uniformly in N (i.e., the
constants do not depend on N ). Hence, for any ϕ ∈ S and 1 < p < ∞,
N
σ k ∆ k ϕ ≲ ϕLp
k=−N Lp
with a constant only depending on d and p (but not on N ). Raising this to the p’th power,
taking the expectation, and then taking the p’th root, we obtain
p 1/p
N
E
σk ∆k ϕ ≲ ϕLp
k=−N Lp
Thus, one part of Khinchine’s inequality for functions in Lemma 4.8 shows that
1/2
N p
|∆k ϕ|2 ≲ ϕLp .
k=−N p
L
We may now let N → ∞ to obtain via the monotone convergence theorem that
SϕLp ≲ ϕLp ,
with the constant depending only on d and p. This is the upper inequality in (4.15).
Step 2. Let now η be another function like in Lemma 4.1 (with C = 2) such that η = 1
on supp η. Denote the corresponding dyadic operators by ∆ k and the square function of
ψ ∈ S relative to η by Sψ. We observe that
k ∆k = ∆k .
∆
4.3. THE SQUARE FUNCTION 47
Thus, writing Id = ∞ k=−∞ ∆k (on S), for any ϕ, ψ ∈ S, we have for the L -scalar product
2
(ϕ, ψ) := ϕ · ψ dx that
∞ ∞ ∞
ϕ, ψ = ∆k ϕ, ψ = kψ =
∆k ϕ, ∆ (∆k ϕ)(x) · (∆ k ψ)(x) dx,
k=−∞ k=−∞ k=−∞
k is self-adjoint (which is easy to check). Now, for N ∈ N,
where we have also used that ∆
by the Cauchy–Schwarz inequality for sums, and the Hölder inequality for integrals,
N
k ψ)(x) dx
(∆k ϕ)(x) · (∆
k=−N
N
≤ (∆k ϕ)(x) · (∆k ψ)(x) dx
k=−N
N
N
2 1/2 2 1/2
≤ (∆k ϕ)(x) ·
(∆k ψ)(x) dx
k=−N k=−N
≤ Sϕ, Sψ
Lq ,
≤ SϕLp · Sψ
where 1/p + 1/q = 1. Since this holds independently of N , we get
Lq ≤ CSϕLp · ψLq ,
ϕ, ψ ≤ SϕLp · Sψ
where in the second inequality
we have used Step 1, but for S in place of S (one checks that
this also holds with only k ηk ≤ 2 instead of “= 1”). Thus, Since L is the dual space to
q
Lp , we have shown that ϕ ∈ Lp and
ϕLp ≲ SϕLp ,
again with a constant that only depends on d and p.
Step 3. It remains to extend the definition of the square function to u ∈ Lp , for which we
take a sequence (ϕj ) ⊂ S with ϕj → u in Lp . By the results of Steps 1 and 2, we know that
Sϕj Lp ∼ ϕj Lp ,
where the constant depends only on d and p. The claim will follow by letting j → ∞
in this relation if we can show that Sϕj → Su in Lp . For this, using the the inverse
triangle
inequality for ℓ2 -norm (recall that for a sequence (ak )k the ℓ2 -norm is (ak )k ℓ2 :=
( k |ak |2 )1/2 , which can be shown to satisfy the triangle inequality), we observe
Sϕj (x) − Su(x) = ∆k ϕj (x) ∞ 2 − ∆k u(x) ∞
k=−∞ ℓ k=−∞ ℓ2
∞
≤ ∆k ϕj (x) − ∆k u(x) k=−∞ ℓ
2
= S(ϕj − u)(x).
48 CHAPTER 4. FOURIER MULTIPLIERS AND LITTLEWOOD–PALEY
Applying Fatou’s lemma to the definition of the square function, we then obtain
Sϕj (x) − Su(x) ≤ lim inf S(ϕj − ϕi )(x).
i→∞
We already know that
S(ϕj − ϕi )Lp ≲ ϕj − ϕi Lp → 0.
Thus, applying Fatou’s lemma again,
Sϕj (x) − SuLp ≲ lim inf ϕj − ϕi Lp .
i→∞
Letting j → ∞, we conclude Su ∈ Lp and Sϕj → Su in Lp . This finishes the proof.
Chapter 5
Sobolev Spaces
As we have seen before, the differential operator −∆ = −(∂11 +· · ·+∂dd ) is a Fourier mul-
tiplier operator with multiplier 4π 2 |ξ|2 . The fact that this multiplier has quadratic growth
corresponds to the fact that −∆ is a combination of second-order derivatives. We may now
define (−∆)s/2 for s ∈ R as the Fourier multiplier operator with multiplier (4π 2 |ξ|2 )s/2 .
It thus represents a combination of (fractional) “derivatives” of order s. One may check
easily that if s ∈ N is even, then this definition agrees with s/2 times the application of the
operator −∆.
If α ∈ Nd0 is a multi-index of order |α| = s ∈ N, then, for any ϕ ∈ S,
∂ α ϕ = F −1 (2π)s (−iξ)α ϕ(ξ)
iξ α
= F −1 − · (4π 2 |ξ|2 )s/2 · ϕ(ξ)
|ξ|
= R1α1 · · · Rdαd (−∆)s/2 ϕ,
where by R1 , . . . , Rd are the Riesz transforms defined in Example 4.4. Since we already
know that the Ri (i = 1, . . . , d) are bounded from Lp to itself, where 1 < p < ∞, one
could say that u has “regularity of s (fractional) Lp -derivatives” if (−∆)s/2 u ∈ Lp . This
idea gives rise to the so-called Sobolev spaces, which form a regularity scale for maps
adapted to the Lp -theory. Note, however, that for s < 0 the multiplier (4π 2 |ξ|2 )s/2 has a
singularity at ξ = 0.
5.1 Homogeneous and inhomogeneous Sobolev spaces
Let s ∈ R and 1 < p < ∞. We define the Bessel potential of order s as the Fourier
multiplier operator given via
Js ϕ := (Id − ∆)−s/2 ϕ := F −1 (1 + 4π 2 |ξ|2 )−s/2 ϕ(ξ)
, ϕ ∈ S.
Th Bessel potential operator can in fact be applied to tempered distributions u ∈ S ′ as well:
The multiplier (1 + 4π 2 |ξ|2 )−s/2 is smooth and has at most polynomial growth. Hence, its
49
50 CHAPTER 5. SOBOLEV SPACES
∈ S ′ is well-defined in S ′ (by duality) and we can set
product with u
Js u := (Id − ∆)−s/2 u := F −1 (1 + 4π 2 |ξ|2 )−s/2 u
(ξ) , u ∈ S ′.
Then, the (inhomogeneous) Sobolev space of order s ∈ R and exponent p ∈ (1, ∞) is
Ws,p = Ws,p (Rd ; C) := u ∈ S ′ : J−s u ∈ Lp .
On this space we define the norm
uWs,p := J−s uLp , u ∈ Ws,p ,
which turns Ws,p into a Banach space (exercise).
It is convenient to also define the following variant of Sobolev spaces: The Riesz poten-
tial of order s ∈ R (not to be confused with the Riesz transform!) is the Fourier multiplier
operator given via
Is ϕ := (−∆)−s/2 ϕ := F −1 (4π 2 |ξ|2 )−s/2 ϕ(ξ)
, ϕ ∈ S.
Note that if s > 0, the Riesz potential Is ϕ may nor may not exist as a tempered distribution,
depending on whether (4π 2 |ξ|2 )−s/2 ϕ(ξ)
represents a tempered distribution or not.
To define the corresponding Sobolev space, let us denote by S ′ /P the quotient space of
S ′ modulo the (multi-variate) polynomials P (⊂ S ′ ). Every equivalence class [u] ∈ S ′ /P
for u ∈ S ′ is of the form
[u] = u+p : p∈P .
Polynomials p ∈ P have the property that supp p ⊂ {0} (exercise); in fact, this property
characterizes the polynomials (exercise). Hence we obtain for v = w mod P, i.e., v, w ∈
[u] ∈ S ′ /P, that supp( ⊂ {0}. As a consequence,
v − w)
v=w mod P ⇐⇒ 〈 ϕ〉 for all ϕ ∈ S with supp ϕ ⊂ Rd \ {0}.
v , ϕ〉 = 〈w,
We now fix a cut-off function η ∈ C∞
c (B2 ; [0, 1]) with η = 1 on B1 and define for u ∈ S
′
and ϕ ∈ S,
ξ
s s
|ξ| u
, ϕ := lim u , 1 − η |ξ| ϕ(ξ) ,
ε→0 ε
provided this limit exists. From the facts mentioned above, we obtain for u ∈ [u] ∈ S ′ /P
that if this limit exists for all ϕ ∈ S, then the Fourier multiplier |ξ|s descends to a quotient
= [|ξ|s u
and |ξ|s [u] ] ∈ S ′ /P. For this reason we may identify equivalence classes with
arbitrary representatives and write Is u for the application of the Fourier multiplier operator
with multiplier (4π 2 |ξ|2 )−s/2 = (4π 2 )s/2 |ξ|−s , understood in the above sense, to u = [u] ∈
S ′ /P.
5.1. HOMOGENEOUS AND INHOMOGENEOUS SOBOLEV SPACES 51
.
Let Lp be the space of equivalence classes [u] ∈ S ′ /P such that there is a (unique)
v ∈ [u] with v ∈ Lp . This space of “Lp -maps modulo polynomials” is equipped with the
norm
[u]L. p := vLp if v ∈ [u] ∩ Lp .
Then, we define the homogeneous Lp -Sobolev space of order s and exponent p as
. . .
Ws,p = Ws,p (Rd ; C) := u ∈ S ′ /P : I−s u ∈ Lp ,
which we equip with the norm
.
. s,p := I−s u . p ,
uW u ∈ Ws,p .
L
We here identify the equivalence class [u] ∈ S ′ /P with. any of its representatives and in-
clude the existence of I−s u in the above definition of Ws,p . The fact that W
. s,p is indeed
. .
a norm turning Ws,p into a Banach space uses the above properties of Lp (exercise).
Finally, for the exponent p = 2 one often uses the following special notation:
. .
Hs := Ws,2 , Hs := Ws,2 .
These spaces are Hilbert spaces with inner products
u, v Hs := J−s u, J−s v L2 , u, v ∈ Hs ,
and
.
u, v H. s := I−s u, I−s v L2 , u, v ∈ Hs ,
respectively.
Example 5.1. Let δ0 ∈ S ′ be the Dirac delta distribution. Then, via Plancherel’s relation
we have for s ∈ R that
δ0 Hs = (Id − ∆)s/2 δ0 L2 = (1 + 4π 2 |ξ|2 )s/2 L2 .
The last expression is finite if and only if s < −d/2, so δ0 ∈ Hs for all such s.
Example 5.2. One can show (exercise) that u = [−1,1] ∈ Hs for any s < 1/2 and that
u = (−∞,∞) ∈/ Hs for any s ∈ R.
We can now easily prove that Sobolev spaces are indeed characterized by the existence
of weak derivatives:
.
Lemma 5.3. Let u ∈ Wk,p , where k ∈ N and 1 < p < ∞. Then,
.
∂ α u ∈ Lp for any multi-index α ∈ Nd0 with |α| = k.
52 CHAPTER 5. SOBOLEV SPACES
.
Moreover, for any u ∈ Ws,p it holds that
. k,p ∼
uW ∂ α uL. p , (5.1)
|α|=k
where the proportionality constants only depend on k, p, d. Conversely, ′
. k,p if u ∈ S with
∂ uLp < ∞ for any multi-index α ∈ N0 with |α| = k, then u ∈ W .
α d
Proof. If α ∈ Nd0 is a multi-index of order |α| = k, then for any ϕ ∈ S,
∂ α ϕ = F −1 (2π)k (−iξ)α ϕ(ξ)
iξ α
= F −1 − · (4π 2 |ξ|2 )k/2 · ϕ(ξ)
|ξ|
= R1α1 · · · Rdαd (−∆)k/2 ϕ,
where by R1 , . . . , Rd are the Riesz transforms defined in Example 4.4, which are of strong
type (p, p). Thus,
∂ α ϕLp ≲ I−k ϕL. p = ϕW
. k,p .
. .
Thus, by density, we have for u ∈ Wk,p established the existence of ∂ α u in Lp and the “≳”
part of (5.1).
For the other direction, which also implies the last assertion of the lemma, we observe
that
k ξα
2 2 k/2 k
(4π |ξ| ) = (2π) ξα · k ,
α |ξ|
|α|=k
ξα
where we recall that αk = α1 !···α k!
d!
. The multiplier |ξ| k satisfies the assumptions of the
Mikhlin Multiplier Theorem 4.2; in fact, the corresponding multiplier operator is given as
R1α1 · · · Rdαd , up to constants. Hence, using also another density argument, we may conclude
the “≲” part of (5.1) (details as exercise).
In order to understand the properties of Sobolev maps, we will employ the Littlewood–
Paley methods. So, for a given function η as in Lemma 4.1, we denote the associated dyadic
operators by ∆k , ∆≤k , ∆≥k (k ∈ Z) as usual. The following results, known together as
Bernstein’s inequalities, show how the dyadic operators and the Riesz potentials interact.
The idea here is that ∆k u for u ∈ Lp is frequency-localized around |ξ| ∼ 2k (e.g., u(x) =
e(2k ix)), so that taking s (fractional) derivatives acts approximately like multiplication with
ds
2ks (e.g., for u(x) = e(2k ix) and s ∈ N we have dx k s
s u(x) = (2 i) u(x)).
5.1. HOMOGENEOUS AND INHOMOGENEOUS SOBOLEV SPACES 53
Lemma 5.4. Let s ≥ 0, k ∈ Z, and 1 < p < ∞. Then, for u ∈ Lp the following estimates
hold:
∆≥k uLp ≲ 2−ks I−s ∆≥k uLp , (5.2)
∆≤k I−s uLp ≲ 2 ∆≤k uLp ,
ks
(5.3)
∆k I−s uLp ∼ 2ks ∆k uLp , (5.4)
−ks
∆k Is uLp ∼ 2 ∆k uLp , (5.5)
where the proportionality constants depend on s, p and d.
Proof. We only show the first inequality (5.2); the proof of the inequalities (5.3)–(5.5) is
similar and left as an exercise. Using that the dyadic operators commute with the Riesz
potentials as well as the density of D in Lp (see Theorem 1.22), our claim (5.2) is equivalent
to
ks
2 Is ∆≥k ϕ p ≲ ∆≥k ϕLp (5.6)
L
for all ϕ ∈ D.
Let η the function generating the dyadic operators ∆k (see Lemma 4.1). Take χ ∈
C∞ (Rd ; [0, 1]) with χ = 1 on the support of ℓ≥1 ηℓ and χ = 0 near the origin. We see
that for all multi-indices α ∈ Nd0 there is a constant Cα > 0 such that
Cα
|∂ α χ(ξ)| ≤ , ξ ∈ Rd \ {0}. (5.7)
|ξ||α|
Set
χℓ (ξ) := χ(2−k ξ), ξ ∈ Rd ,
and observe that χk = 1 on the support of ℓ≥k ηℓ . Denote by Tχk the Fourier multiplier
operator associated with χk . Then,
2ks Is ∆≥k = 2ks Is Tχk ∆≥k =: Tm ∆≥k ,
where Tm is the multiplier operator with
m(ξ) := 2ks (2π)−s |ξ|−s χ(2−k ξ), ξ ∈ Rd \ {0}.
In the following, we will show that for some Bα > 0 it holds that
|∂ α m(ξ)| ≤ Bα |ξ|−|α| , ξ ∈ Rd \ {0}, (5.8)
for all multi-indices α ∈ Nd0 with |α| ≤ d + 2 (note that the dependence of Bα on α does
not matter since we only require (5.8) for finitely many multi-indices α). Then, (5.6) and
hence (5.2) will follow from the Mikhlin Multiplier Theorem 4.2.
54 CHAPTER 5. SOBOLEV SPACES
It is easy to see (exercise) that
|∂ β (|ξ|−s )| ≤ Cβ |ξ|−s−|β| .
Moreover, using (5.7), for any multi-index γ ∈ Nd0 , we obtain the estimate
1 Cγ 2k|γ| Cγ
|∂ γ χk (ξ)| ≤ · |∂ γ χ(2−k ξ)| ≤ · = .
2k|γ| 2k|γ| |ξ||γ| |ξ||γ|
Thus, also employing that |ξ| ≳ 2k on supp χk by construction,
ks β
2 ∂ (|ξ|−s )∂ γ χk (ξ) ≲ 2ks |ξ|−s−|β| |ξ|−|γ| ≲ (2ks |ξ|−s )|ξ|−(|β|+|γ|) ≲ |ξ|−(|β|+|γ|) ,
where the proportionality constants depend on β, γ. By the Leibniz rule, this immediately
implies (5.8) and the proof of (5.2) is thus complete.
5.2 Inequalities and embeddings
We will now explore the relationship between different Sobolev spaces. We start by record-
ing the following fact on the difference between the homogeneous and inhomogeneous
Sobolev spaces:
Theorem 5.5. Let s ≥ 0 and 1 < p < ∞. Then,
. s,p ,
uWs,p ∼ uLp + uW u ∈ Ws,p , (5.9)
where the proportionality constants depend on s, p and d.
Proof. For the inequality “≲” in (5.9) we observe that the multiplier
(1 + 4π 2 |ξ|2 )s/2
m(ξ) := , ξ ∈ Rd \ {0},
1 + (4π 2 |ξ|2 )s/2
satisfies the Mikhlin condition |∂ α m(ξ)| ≤ Bα |ξ|−|α| for all multi-indices α ∈ Nd0 with
|α| ≤ d + 2 and some constants Bα > 0. Thus, by the Mikhlin Multiplier Theorem 4.2,
J−s uLp = Tm [u + I−s u]Lp ≲ u + I−s uLp ≤ uLp + I−s uLp .
As uW. s,p = I−s u . p = I−s uLp (since u ∈ Ws,p ⊂ Lp ), this shows “≲” in (5.9).
L
For the other direction, we can use a similar argument using the multipliers
(4π 2 |ξ|2 )s/2
m0 (ξ) := (1 + 4π 2 |ξ|2 )−s/2 and m1 (ξ) := .
(1 + 4π 2 |ξ|2 )s/2
We leave the details as an exercise.
The next theorem is a central result in the theory of Sobolev maps and goes under the
name of Gagliardo–Nirenberg–Sobolev inequality:
5.2. INEQUALITIES AND EMBEDDINGS 55
Theorem 5.6. Let s > 0, 1 < p < q < ∞, and θ ∈ (0, 1) such that
1 1 sθ
= − . (5.10)
q p d
Then,
θ.
.
uLq ≲ u1−θ
Lp · u s,p , u ∈ Lp ∩ Ws,p , (5.11)
W
where the proportionality constant depends on s, p, q and d.
.
Note that the condition “u ∈ Lp ∩ Ws,p ”. means that we have selected the (unique)
element from the equivalence class of u ∈ Ws,p that is also in Lp . Furthermore, since
uW . s,p = I−s u . p = I−s uLp one can also understand the preceding theorem as a
L
result about the Riesz potential, independently of the theory of Sobolev spaces.
Like with any inequality involving norms of Lp -maps, one should check that (5.11) is
scale-invariant under the following two natural transformations of the map u: Set uλ,µ (x) :=
λu(x/µ) for any λ, µ > 0. Then, for t ≥ 0 observe (exercise) that
d
−t
I−t uλ,µ Lp = λµ p I−t uLp .
Applying (5.11) to uλ,µ , we thus obtain
d d d
uθ. s,p
(1−θ) −s θ
λµ uLq ≲ λ
q (1−θ)
µ p u1−θ
Lp ·λ µ θ p
W
Note that λ occurs with exponent 1 on both sides of the inequality. Likewise, for µ the
exponent on the right-hand side is
d d d d
(1 − θ) + − s θ = − sθ =
p p p q
by (5.10), which agrees with the exponent of µ on the left-hand side. Hence, we cannot
improve the conclusion (5.11) by choosing λ, µ > 0 in a particular way and the statement
is optimal with respect to the exponents.
Proof. We may assume without loss of generality that u ∕= 0. Moreover, by the scaling
calculations preceding this proof, we may pass from u to uλ,µ with
d 1
−1
uLpsp I−s uLs p
λ := d , µ := 1 .
I−s uLpsp uLs p
. s,p = I−s uLp = 1.
to see that we may suppose furthermore uLp = 1 and uW
56 CHAPTER 5. SOBOLEV SPACES
Denoting once again a family of dyadic operators by ∆k , k ∈ Z, we have by (4.13) in
Lemma 4.6 and (5.10) that
∞
uLq ≤ ∆k uLq
k=−∞
∞
1
kd − 1q
≲ 2 p ∆k uLp
k=−∞
∞
= 2ksθ ∆k uLp .
k=−∞
For k ≥ 0, the third Bernstein inequality (5.4) from Lemma 5.4 yields
∆k uLp ∼ 2−ks ∆k I−s uLp ≲ 2−ks I−s uLp = 2−ks .
Here, we used that ∆k vLp ≲ vLp for any v ∈ Lp with the constant depending only on
p, d (this follows by an easy application of Mikhlin’s theorem as in the proof of Lemma 4.6).
On the other hand, for k < 0, we have
∆k uLp ≲ uLp = 1.
Thus,
−1
∞
uLq ≲ 2ksθ + 2ks(θ−1) ≲ 1.
k=−∞ k=0
Hence, (5.11) has been established for this u and the general conclusion follows by a scaling,
as explained above.
We can now state one of the most important theorems in the theory of Sobolev maps,
the Sobolev embedding theorem:
Theorem 5.7. Let s ≥ 0, 1 < p ≤ q < ∞ such that
1 s 1
− < . (5.12)
p d q
Then,
uLq ≲ uWs,p , u ∈ Ws,p ,
where the proportionality constants depend on s, p, q and d.
Proof. We may assume that s > 0 and p < q since the other cases are trivial. In this situa-
tion, (5.12) is equivalent to the existence of a θ ∈ (0, 1) for which (5.10) holds. The proof is
then immediate by combining the preceding Gagliardo–Nirenberg–Sobolev inequality with
Theorem 5.5.
5.3. APPLICATIONS TO PDES 57
Corollary 5.8. For all s, s′ ∈ R with s < s′ and for all 1 < p < ∞ it holds that
′
Ws ,p ↩→ Ws,p ,
that is,
′
uWs,p ≲ uWs′ ,p , u ∈ Ws ,p .
Proof. This follows by applying the previous theorem with p := q := p, s := s′ − s, and
:= J−s u. Then,
u
uWs,p = J−s uLp ≲ J−s uWs′ −s,p = uWs′ ,p ,
which is the claim.
5.3 Applications to PDEs
Let us now examine how we can apply the Sobolev space theory developed so far to our
prototypical Poisson equation and related (elliptic) PDEs.
Example 5.9. Let u ∈ Lp be a weak solution to the Poisson equation
−∆u = f in Rd
with right-hand side f ∈ Lp for 1 < p < ∞.. Then, by the very definition of the homo-
geneous Sobolev spaces, we know that u ∈ W2,p . We have seen at the beginning of this
chapter that
∂ij u = Ri Rj (−∆)u = Ri Rj f.
Since the Riesz transforms Ri (see Example 4.4) are bounded from Lp to itself, we obtain
that all second partial derivatives of u exist and are in Lp . Higher derivatives exist if f
has higher derivatives (in Lp ). Note that the solution to the Poisson equation has already
been expressed via the Newtonian potential (that is, the Riesz potential of order 2) in Exam-
ple 2.14 and one could alternatively obtain derivative estimates by studying the properties
of this convolution integral. This “direct approach”, however, is not very practical when we
are dealing with PDE operators for which we do not know explicit solution potentials.
Example 5.10. Assume now that we are instead trying to understand a more general “el-
liptic” PDE
Lu := − aij ∂ij u = f in Rd ,
ij
where f ∈ Lp , 1 < p < ∞, and (aij )ij ∈ Rd×d sym is a symmetric, positive definite matrix (if
(aij )ij = Id, then this is again the Poisson equation). If we Fourier transform this equation
we get
4π 2 aij ξi ξj u(ξ) = f(ξ), ξ ∈ Rd .
ij
58 CHAPTER 5. SOBOLEV SPACES
From the properties of (aij )ij it follows that the symbol
L(ξ) := 4π 2 aij ξi ξj
ij
is positive and the ellipticity condition
L(ξ) ≥ c|ξ|2 , ξ ∈ Rd .
holds for some c > 0. Let us define an approximate solution operator as follows:
Gf := F −1 L(ξ)−1 (1 − η(ξ))f(ξ) ,
where η ∈ C∞ c (R ; [0, 1]) is a fixed cut-off function with η = 1 on a neighborhood of
d
the origin. Note that G is a Fourier multiplier operator satisfying the assumptions of the
Mikhlin Multiplier Theorem 4.2, so that G : Lp → Lp is bounded. Moreover, for v ∈ Lp ,
GLv = LGv = v − ηq v =: (Id + E)v,
where we remark that ηq ∈ S and thus EvWk,p ≲ vLp for any k ∈ N0 (use Lemma 5.3
and the properties of the convolution, cf. see Lemma 2.11). Then,
(Id − ∆)u = (Id − ∆)GLu − (Id − ∆)Eu
= (Id − ∆)Gf − (Id − ∆)Eu
=: T0 f + T1 u,
where both T0 := (Id−∆)G and T1 := −(Id−∆)E satisfy the assumptions of the Mikhlin
Multiplier Theorem 4.2 (exercise). Thus, we obtain the regularity estimate
uW2,p ≲ f Lp + uLp ,
where higher-order derivatives of u are estimated by lower-order ones (in this case, only
u itself) and the data of the PDE (here, f ). This kind of argument is quite common. It
draws its strength from the fact that it allows one to obtain estimates by simple, almost
“algebraic”, manipulations of the operators occurring in the PDE. Note in particular that it
was not necessary to explicitly “solve” the PDE.
Chapter 6
Pseudodifferential Operators
Suppose we are now trying to analyze the variable-coefficient elliptic PDE
Lu(x) := − aij (x)∂ij u(x) = f (x) in Rd ,
ij
where f ∈ Lp and a : Rd → Rd×d sym is a symmetric, positive definite matrix field that is
smooth in x (when a = Id, then this is again the Poisson equation). We would like to
mimic the analysis of Example 5.10. However, this does not work because the coefficients
are x-dependent, which makes the definition of the approximate solution operator G unclear.
Moreover, differential operators with x-dependent coefficients do not commute with Riesz
potentials or Fourier multiplier operators.
However, for the regularity properties of this PDE we expect roughly the same results
as in the constant-coefficient case because a is assumed to be smooth. So we desire a notion
of “x-dependent Fourier multiplier operator”. It turns out to be useful to also systematize
the notion of “order” of an operator at the same time. This leads to the so-called pseudod-
ifferential operators, which include Fourier multiplier operators, classical multiplication
operators, potentials, and many other operators. This class of operators appears frequently
in applications, even if one starts with a constant-coefficient differential operator in the first
place. Also, from a theoretical perspective, the theory of pseudodifferential operators nicely
combines all the techniques we have seen so far and thus constitutes a natural culmination
of this course.
6.1 General singular integral operators
Before we come to the heart of the matter, we need a generalization of the theory of convo-
lution operators: We call an operator T : Lq → Lq , where 1 < q < ∞, a (general) singular
integral operator if it satisfies the following three conditions:
1. (SI1) Boundedness: T is of strong type (q, q) with operator norm from Lq to itself
bounded by B > 0.
59
60 CHAPTER 6. PSEUDODIFFERENTIAL OPERATORS
2. (SI2) Integral representation: There exists a function
K ∈ L1loc ((Rd × Rd ) \ D; C), where D := { (x, x) : x ∈ Rd },
called the kernel (function) of T , such that for all
u ∈ Lqc := u ∈ Lq : ess supp u compact
it holds that
T u(x) = K(x, y)u(y) dy for a.e. x ∈ Rd \ ess supp u. (6.1)
3. (SI3) Hörmander’s condition: There is a constant c > 1 such that if Q ⊂ Rd is a
cube with center yQ and Q∗ is its dilation by c (scaling by c with the same center yQ ),
then
|K(x, yQ ) − K(x, y)| dx ≤ B for all y ∈ Q. (6.2)
Rd \Q∗
Note that in the classical case of a (singular integral) convolution operator T = K
with a kernel K ∈ S ′ such that K = K(x) ∈ L1loc (Rd \ {0}; C) away from the origin,
we may set K(x, y) := K(x − y) as the generalized kernel. We then observe that the first
condition with q = 2 can be established for instance via Lemma 3.4, the second condition
corresponds to the result of Lemma 3.6, √and the third condition follows from the classical
Hörmander condition (3.5) (with c = 2 d), as was shown in the proof of Lemma 3.10.
We then have the following generalized Calderón–Zygmund theorem:
Theorem 6.1. Let T be a (general) singular integral operator operator satisfying the con-
ditions (SI1)–(SI3) above. Then, T is of strong type (p, p) for any 1 < p ≤ q and of weak
type (1, 1).
Proof. By the same interpolation argument as in the proof of Theorem 3.12, but now be-
tween exponents 1 and q (using (SI1)), it only remains to establish the weak-type (1, 1)
bound. The same proof as in Lemma 3.10 applies with very minor modifications. Indeed,
given u ∈ Lp ∩ Lq (for now) with corresponding Calderón–Zygmund decomposition, we
proceed line by line as in the proof of that lemma, except that we now define Q∗ to be the
dilation by the constant c > 1 as in (SI3). From (SI2),
T uQ (x) = K(x, y)uQ (y) dy = K(x, y) − K(x, yQ ) uQ (y) dy.
Q Q
Hence,
|T uQ (x)| dx ≤ |K(x, y) − K(x, yQ )| · |uQ (y)| dy dx
Rd \Q∗ Rd \Q∗ Q
= |K(x, y) − K(x, yQ )| dx · |uQ (y)| dy.
Q Rd \Q∗
6.2. SYMBOLS AND OPERATORS 61
The generalized Hörmander condition (6.2) in (SI3) immediately yields
|T uQ (x)| dx ≤ B |uQ (y)| dy ≤ 2B Q uL1 .
Rd \Q∗ Q
One may then conclude as before.
6.2 Symbols and operators
A function a ∈ C∞ (Rd × Rd ; C) is said to be a (smooth) symbol of order m ∈ R, written
as a ∈ Sm , if
|∂xβ ∂ξα a(x, ξ)| ≤ Aα,β (1 + |ξ|)m−|α| , (x, ξ) ∈ Rd × Rd , (6.3)
for all multi-indices α, β ∈ Nd0 and a constant Aα,β ≥ 0. One can also observe the following
closure properties of the symbol classes Sm (exercise):
(i) If ϕ ∈ S, then ϕ ∈ Sm for all m ∈ R.
(ii) If a ∈ Sm (m ∈ R), then b(x, ξ) := ∂xβ ∂ξα a(x, ξ) ∈ Sm−|α| for all α, β ∈ Nd0 .
(iii) If a ∈ Sm and b ∈ Sn , then a + b ∈ Smax(m,n) and a · b ∈ Sm+n .
(iv) If a ∈ Sm is bounded away from zero, then 1/a ∈ S−m .
For a symbol a ∈ Sm , we define the pseudodifferential operator Ta = a(x, D) for
ϕ ∈ S via
Ta ϕ(x) = a(x, D)ϕ := a(x, ξ)ϕ(ξ) e2πix·ξ dξ, x ∈ Rd .
To show that Ta ϕ ∈ S and hence Ta : S → S, observe that for the operator
Lξ := (1 + 4π 2 |x|2 )−1 (Id − ∆ξ )
it holds that Lξ e2πix·ξ = e2πix·ξ . Thus, via repeated integration by parts,
2πix·ξ
Ta ϕ(x) = (Lξ )N a(x, ξ)ϕ(ξ) e dξ.
Using (6.3) and choosing N large enough, one can then verify easily that xα Ta ϕ∞ ≤ Cα
for every multi-index α ∈ Nd0 . The same argument applies also for xα ∂ β Ta ϕ∞ ≤ Cα,β
for all α, β ∈ Nd0 . Hence, Ta ϕ ∈ S.
Example 6.2. If f ∈ C∞
c (R ), then the associated multiplication operator
d
Mf ϕ := f · ϕ, ϕ ∈ S,
is a pseudodifferential operator of order 0.
62 CHAPTER 6. PSEUDODIFFERENTIAL OPERATORS
Example 6.3. Let m ∈ C∞ (Rd ; C) be a Fourier multiplier satisfying
|∂ α m(ξ)| ≤ Bα (1 + |ξ|)−|α| , ξ ∈ Rd ,
for all multi-indices α ∈ Nd0 with a constant Bα ≥ 0. Then, the associated Fourier multi-
plier operator is a pseudodifferential operator of order 0. Note, however, that we required
infinite smoothness and a version of the condition (4.1) from Mikhlin’s Multiplier Theo-
rem 4.2 for all multi-indices α (we do not discuss the delicate matter of “rough” symbols
for pseudodifferential operators here).
Example 6.4. Let
Lu(x) := − aij (x)∂ij u(x) in Rd ,
ij
where a : Rd → Rd×d sym is a symmetric, positive definite matrix field that is smooth in x.
Then, L is a pseudodifferential operator of order 2 with symbol the characteristic polyno-
mial
aL (x, ξ) := L(x, ξ) := 4π 2 aij (x)ξi ξj ,
ij
which lies in S2 if (6.3) is satisfied. Higher-order and non-homogeneous operators also fall
into this framework.
Example 6.5. The Bessel potential or order s ∈ R, i.e., Js := (Id − ∆)−s/2 is a pseu-
dodifferential operator of order −s (the order of a “potential” is the inverse of the order of
a “differential operator”).
One extremely useful feature of the class of pseudodifferential operators is that there is
an associated symbolic calculus:
Theorem 6.6. Let a ∈ Sm and b ∈ Sn . Then, there is a symbol c ∈ Sm+n such that for
the associated pseudodifferential operators Ta , Tb it holds that
Tc = Ta ◦ Tb .
Moreover,
c∼a·b mod Sm+n−1 ,
that is, c − a · b ∈ Sm+n−1 .
Proof. We assume additionally that a and b have (x-uniformly) compact support, that is,
there is a compact set K ⊂ Rd with a(x, ξ) = b(x, ξ) = 0 for x ∈ Rd \ K and any ξ ∈ Rd .
6.2. SYMBOLS AND OPERATORS 63
The other case can be reduced to this one by means of a cut-off and approximation argument
(exercise). Owing to the compact support property, may write for ϕ, ψ ∈ S,
Ta ψ(x) = a(x, η) e2πi(x−y)·η ψ(y) dy dη,
Tb ϕ(y) = b(y, ξ) e2πi(y−z)·ξ ϕ(z) dz dξ.
Thus,
(Ta (Tb ϕ))(x) = a(x, η) e2πi(x−y)·η (Tb ϕ)(y) dy dη
= a(x, η)b(y, ξ) e2πi(x−y)·η e2πi(y−z)·ξ ϕ(z) dz dξ dy dη
= c(x, ξ) e2πi(x−z)·ξ ϕ(z) dz dξ
with
c(x, ξ) := a(x, η)b(y, ξ) e2πi(x−y)·(η−ξ) dy dη
= a(x, ξ + η)b(y, ξ) e2πi(x−y)·η dy dη
= a(x, ξ + η) b(η, ξ) e2πix·η dη,
where
b(η, ξ) := b(y, ξ) e−2πiy·η dy, (η, ξ) ∈ Rd × Rd ,
denotes the Fourier transform of b with respect to the first argument. We now use Taylor’s
theorem to write
a(x, ξ + η) = a(x, ξ) + R(x, ξ, η).
Then,
c(x, ξ) = a(x, ξ)b(x, ξ) + R(x, ξ, η) b(η, ξ) e2πix·η dη.
The following estimates follow from (6.3) and Taylor’s theorem (exercise):
|b(η, ξ)| ≲ (1 + |η|)−ℓ (1 + |ξ|)n for all ℓ ∈ N0 ,
|R(x, ξ, η)| ≲ |η|(1 + |ξ|) m−1
if |ξ| ≥ 2|η|,
|R(x, ξ, η)| ≲ |η| if m ≤ 1,
|R(x, ξ, η)| ≲ (1 + |η|) (1 + |ξ|)
m m−1
if m > 1.
64 CHAPTER 6. PSEUDODIFFERENTIAL OPERATORS
So, for r := c − ab, the first estimate implies
|r(x, ξ)| ≤ R(x, ξ, η) b(η, ξ) dη ≤ (1 + |ξ|)n (1 + |η|)−ℓ |R(x, ξ, η)| dη.
Now split the domain of integration into the set {|ξ| ≥ 2|η|} and its complement. For the
first such integral we obtain, using the second estimate above,
(1 + |η|) −ℓ
|R(x, ξ, η)| dη ≲ (1 + |ξ|) m−1
(1 + |η|)−ℓ |η| dη.
{|ξ|≥2|η|} {|ξ|≥2|η|}
For the second integral, if m ≤ 1, then, using the third estimate above,
(1 + |η|) −ℓ
|R(x, ξ, η)| dη ≲ (1 + |η|)−ℓ |η| dη,
{|ξ|<2|η|} {|ξ|≤2|η|}
whereas if m > 1, using the fourth estimate,
(1 + |η|) −ℓ
|R(x, ξ, η)| dη ≲ (1 + |ξ|) m−1
(1 + |η|)−ℓ+m dη.
{|ξ|<2|η|} {|ξ|≤2|η|}
Choosing ℓ large enough, we obtain in every case that all these integrals are bounded by a
constant multiple times (1 + |ξ|)m−1 . Hence,
|r(x, ξ)| ≲ (1 + |ξ|)m+n−1 .
The same argument gives for every multi-indices α, β ∈ Nd0 that
|∂xβ ∂ξα r(x, ξ)| ≲ (1 + |ξ|)m+n−1−|α| , (x, ξ) ∈ Rd × Rd ,
since ∂xβ a is a symbol of order m and ∂ξα b is a symbol of order n − |α|. Thus, r = c − ab
is indeed a symbol of order m + n − 1 and the proof is complete in the case of compactly
supported a, b.
An immediate consequence of the preceding theorem is the following statement about
commutators:
Corollary 6.7. Let a ∈ Sm and b ∈ Sn . Then, the commutator
[Ta , Tb ] := Ta ◦ Tb − Tb ◦ Ta
is a pseudodifferential operator of order m + n − 1.
6.3. MAPPING PROPERTIES 65
6.3 Mapping properties
We now turn to the central question of the Lp -boundedness of pseudodifferential opera-
tors. For pseudodifferential operators of order 0, the scheme of proof is the same as for
Fourier multiplier operators: First, we establish the boundedness on L2 (which here is more
involved than a mere application of Parseval’s relation). Second, we realize pseudodiffer-
ential operator of order 0 as singular integral operators and then invoke the generalized
Calderón–Zygmund theory to obtain a strong-type (p, p) bound. Finally, we show the map-
ping properties of pseudodifferential operators of arbitrary order relative to the scale of
Sobolev spaces. The end result will be the following theorem:
Theorem 6.8. Let a ∈ Sm (m ∈ R) be a symbol of order m and let Ta = a(x, D) be
the pseudodifferential operator associated to the symbol a. Then, for every s ∈ R and
1 < p < ∞, the operator Ta is bounded from Ws,p to Ws−m,p .
We will carry out the proof according to the strategy outlined above.
Lemma 6.9. For a ∈ S0 , the operator Ta is of strong type (2, 2).
Proof. We again assume that the symbol a(x, ξ) has uniformly compact support in x. Then,
with
a(λ, ξ) := a(x, ξ) e−2πix·λ dx,
(λ, ξ) ∈ Rd × Rd ,
the Fourier inversion formula in S yields
a(x, ξ) = a(λ, ξ) e2πiλ·x dλ, (x, ξ) ∈ Rd × Rd .
Then, for ϕ ∈ S,
Ta ϕ(x) = e2πix·ξ dξ
a(x, ξ)ϕ(ξ)
= a(λ, ξ) e2πiλ·x ϕ(ξ)
e2πix·ξ dξ dλ
= T λ ϕ(x) dλ
with the definition
T λ ϕ(x) := e2πiλ·x [Ta(λ, ) ϕ](x), Ta(λ, ) ϕ := F −1
a(λ, ξ)ϕ(ξ) .
Using the rules for the Fourier transform (or directly integration by parts), for any multi-
index α ∈ Nd0 ,
α
(2πiλ) a(λ, ξ) = ∂xα a(x, ξ) e−2πix·λ dx.
66 CHAPTER 6. PSEUDODIFFERENTIAL OPERATORS
Thus, using the compact support of a and (6.3),
a(λ, )∞ ≲ B(1 + |λ|)−(d+1) .
As a consequence of Plancherel’s relation, we thus have
a(λ, )∞ ≲ B(1 + |λ|)−(d+1) .
T λ L2 →L2 ≤
It remains to observe that
Ta L2 →L2 ≤ T λ L2 →L2 dλ ≲ (1 + |λ|)−(d+1) dλ < ∞.
Thus, Ta is indeed of strong type (2, 2) if a has compact support. The general case (without
the compact support property of a) is treated by a splitting and approximation argument
using the singular integral representation of Ta (see below); details can be found in [11,
Section VI.2].
Lemma 6.10. For a ∈ S0 , the operator Ta is of strong type (p, p) for all 1 < p < ∞.
Proof. The proof proceeds via the singular integral representation of Ta .
Step 1. Similar to the dyadic decomposition of Lemma 4.1, fix a radial cut-off function
χ ∈ C∞ c (R ; [0, 1]) with χ(ξ) = 1 for |ξ| ≤ 1 and χ(ξ) = 0 for |ξ| ≥ 2. Then, for
d
η(ξ) := χ(ξ) − χ(2ξ) and ηk (ξ) := η(2−k ξ) (k ∈ Z, ξ ∈ Rd ) it holds that
∞
χ+ ηk = 1 on Rd .
k=1
We then split (with convergence of bounded operators from L2 to itself)
∞
T a = T a0 + T ak
k=1
with the truncated symbols
a0 (x, ξ) := a(x, ξ)χ(ξ), ak (x, ξ) := a(x, ξ)ηk (ξ),
which are smooth and have compact support in ξ (for frozen x). Thus, for ϕ ∈ S,
Tak ϕ(x) = κk (x, x − y)ϕ(y) dy = κk (x, z)ϕ(x − z) dz, x ∈ Rd ,
where
κk (x, z) := ak (x, ξ) e2πiz·ξ dξ.
6.3. MAPPING PROPERTIES 67
It can be seen that these integrals are convergent and finite. We claim that
β α
∂x ∂z κk (x, z) ≲ 2k(d−M +|α|) |z|−M , (x, z) ∈ Rd × (Rd \ {0}), (6.4)
for all k ∈ N0 , all multi-indices α, β ∈ Nd0 , and all M ∈ N0 . To see this, first observe for
any multi-index γ ∈ Nd0 ,
(−2πiz)γ ∂xβ ∂zα κk (x, z) = ∂ξγ (2πiξ)α ∂xβ ak (x, ξ) e2πiz·ξ dξ.
The assumption (6.3) implies that the integrand is bounded by a multiple of 2k(−|γ|+|α|)
(exercise) and, if k ≥ 1, it is supported in the annulus B2k+1 \ B2k−1 in the variable ξ. Thus,
γ β α
z ∂x ∂z κk (x, z) ≲ 2k(d−|γ|+|α|) .
Taking the supremum over all multi-indices γ with |γ| = M , one obtains (6.4).
Step 2. For fixed x ∈ Rd define
κ(x, ) := κ0 (x, ) + κk (x, ) ∈ S ′
∞
k=1
with the sum understood with respect to weak* convergence in S ′ (it is an exercise to see
that this sum indeed converges in S ′ ). We will show that κ(x, ) agrees with a locally
integrable function away from the origin and that
β α
∂x ∂z κ(x, z) ≲ |z|−(d+|α|+ℓ) , (x, z) ∈ Rd × (Rd \ {0}), (6.5)
for all multi-indices α, β ∈ Nd0 and all ℓ ∈ N0 . Let first 0 < |z| ≤ 1. Then we estimate
using (6.4) as follows:
∞
β α
∂x ∂z κk (x, z) ≤ ∂xβ ∂zα κk (x, z) + ∂xβ ∂zα κk (x, z)
k=0 2k ≤|z|−1 2k >|z|−1
= O(|z|−(d+|α|+ℓ) ),
where for the first sum in (6.4) we chose M = 0 if d + |α| > 0 or M > d + |α| + ℓ if
d + |α| ≤ 0, and for the second sum we choose M > d + |α| + ℓ (details as exercise).
If, on the other hand, |z| > 1 then from (6.4) with M > d + |α| + ℓ we obtain
∞
β α
∂x ∂z κk (x, z) = O(|z|−M ) = O(|z|−(d+|α|+ℓ) )
k=0
since |z| > 1. Hence, (6.5) holds in all cases.
Step 3. We now want to apply Theorem 6.1. Combining Steps 1 and 2, we see that for
ϕ ∈ S and x ∈/ supp ϕ, the integral representation
T ϕ(x) = K(x, y)ϕ(y) dy
68 CHAPTER 6. PSEUDODIFFERENTIAL OPERATORS
holds, where
K = K(x, y) := κ(x, x − y)
lies in L1loc ((Rd × Rd ) \ D; C) (with D := { (x, x) : x ∈ Rd } the diagonal) by (6.5). This
is (SI1).
From Lemma 6.9 we already know that Ta is of strong type (2, 2) and so the above
representation extends to u ∈ L2c , namely
Ta u(x) = K(x, y)u(y) dy for a.e. x ∈ Rd \ ess supp u,
by similar arguments as in the proof of Lemma 3.6. This is (SI2).
Finally, (6.5) also implies
α β
∂x ∂y K(x, y) ≲ |x − y|−(d+|α|+|β|) , x ∕= y. (6.6)
√
Thus, if Q ⊂ Rd is a cube with center yQ and Q∗ is its dilation by c := 2 d, then, by a
similar argument as in Lemma 3.11 (also see the proof of Theorem 3.12)
|K(x, yQ ) − K(x, y)| dx ≤ B for all y ∈ Q,
Rd \Q∗
and a constant B > 0. This is (SI3). We conclude by Theorem 6.1 that Ta is of strong type
(p, p) for 1 < p < 2.
The strong type (p, p) bound for 2 < p < ∞ follows by a duality argument as in the
proof of Theorem 3.12 (the dual operator has the kernel K(x, y) := K(y, x), which satisfies
the same estimates as K itself by the symmetry of (6.6) in x and y).
Proof of Theorem 6.8. By the symbolic calculus of Theorem 6.6, there exists a symbol c ∈
Ss with
(Id − ∆)(s−m)/2 ◦ Ta = Tc .
Then, define the 0’th-order symbol
c(x, ξ)
b(x, ξ) := ∈ S0 ,
(1 + 4π 2 |ξ|)s/2
so that Tc = Tb ◦ (Id − ∆)s/2 . The operator Tb is of strong type (p, p), where 1 < p < ∞,
by Lemma 6.10. Then, for u ∈ Ws,p ,
Ta uWs−m,p = (Id − ∆)(s−m)/2 [Ta u]Lp
= Tb [(Id − ∆)s/2 u]Lp
≲ (Id − ∆)s/2 uLp
= uWs,p .
This is the claim.
6.3. MAPPING PROPERTIES 69
Example 6.11. We can now extend the analysis of Example 5.10 to the variable-coefficient
PDE
Lu(x) := − aij (x)∂ij u(x) = f (x) in Rd ,
ij
understood as usual in the weak sense. Here, f ∈ Lp , 1 < p < ∞, and (aij )ij : Rd → Rd×d
is a symmetric, uniformly positive definite matrix field. The pseudodifferential symbol of
L is
L(x, ξ) := 4π 2 aij (x)ξi ξj ,
ij
which lies in S2 if (6.3) holds. For a fixed cut-off function η ∈ C∞
c (R ; [0, 1]) with η = 1
d
on a neighborhood of the origin, we set
G := TL(x,ξ)−1 (1−η(ξ)) ,
which is well-defined by the uniform positive definiteness of the matrix (aij )ij and is a
pseudodifferential operator of order −2.
According to the symbolic calculus of Theorem 6.6, for the symbol of the pseudodiffer-
ential operator Ta := GL we have
a∼1−η ∼1 mod S−1 .
Thus,
GL = Id + E
with E a pseudodifferential operator or order −1 (i.e., E is “smoothing of order 1”). Then,
for u ∈ Lp ,
(Id − ∆)u = (Id − ∆)GLu − (Id − ∆)Eu
= (Id − ∆)Gf − (Id − ∆)Eu
=: T0 f + T1 u,
where T0 := (Id − ∆)G is a pseudodifferential operator of order 0 and T1 := −(Id − ∆)E
is a pseudodifferential operator of order 1. This implies the regularity estimate
uW2,p ≲ f Lp + uW1,p .
Note that we now have the W1,p -norm of u on the right-hand side, and not the Lp -norm as
in the constant-coefficient case.
Bibliography
[1] L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of Functions, 2nd
ed., Studies in Advanced Mathematics, CRC Press, 2015.
[2] L. Grafakos, Classical Fourier Analysis, 3rd ed., Graduate Texts in Mathematics, vol.
249, Springer, 2014.
[3] , Modern Fourier Analysis, 3rd ed., Graduate Texts in Mathematics, vol. 250,
Springer, 2014.
[4] L. Hörmander, The Analysis of Linear Partial Differential Operators I: Distribution
Theory and Fourier Analysis, Classics in Mathematics, Springer, 2003, Reprint of the
second (1990) edition.
[5] , The Analysis of Linear Partial Differential Operators II: Differential Oper-
ators with Constant Coefficients, Classics in Mathematics, Springer, 2005, Reprint of
the 1983 original.
[6] , The Analysis of Linear Partial Differential Operators III: Pseudo-Differential
Operators, Classics in Mathematics, Springer, 2007, Reprint of the 1994 edition.
[7] , The Analysis of Linear Partial Differential Operators IV: Fourier Integral
Operators, Classics in Mathematics, Springer, 2009, Reprint of the 1994 edition.
[8] C. Muscalu and W. Schlag, Classical and multilinear harmonic analysis. Vol. I, Cam-
bridge Studies in Advanced Mathematics, vol. 137, Cambridge University Press, 2013.
[9] , Classical and multilinear harmonic analysis. Vol. II, Cambridge Studies in
Advanced Mathematics, vol. 138, Cambridge University Press, 2013.
[10] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton
University Press, 1970.
[11] , Harmonic Analysis, Princeton University Press, 1993.
[12] E. M. Stein and G. Weiss, Introduction to Fourier analysis on Euclidean spaces,
Princeton Mathematical Series, vol. 32, Princeton University Press, 1971.
71
Index
approximate solution operator, 58 Fourier transform
approximation of the identity, 12 in S, 16
in S ′ , 17
Bernstein’s inequalities, 52 fundamental solution, i
Bessel potential, 49
Gagliardo–Nirenberg–Sobolev inequal-
Calderón–Zygmund decomposition, 30 ity, 54
Calderón–Zygmund operator, 29
Calderón–Zygmund theorem, 34 Hörmander condition
generalized, 60 for convolution operators, 32
characteristic polynomial, 62 strong, 34
commutator, 64 Hörmander’s condition
convergence for singular integral operators, 60
in D, 2 Heaviside function, 11
in D′ , 12 Hilbert kernel, 3
in S, 15 Hilbert transform, 27
in S ′ , 16
convolution integral representation, 27
in D′ , 6 interpolation
in S, 17 complex method, 26
in S ′ , 20 real method, 24
convolution operator, 23
kernel, 23, 60
derivative of a distribution, 10 Khinchine’s inequality, 45
Dirac delta distribution, 3
distribution, 2 Laplacian, 12
distribution function, 25 Lebesgue norm, 23
dyadic lemma, 38 Leibniz rule, 5
dyadic operators, 43 Littlewood–Paley theorem, 45
Lorentz quasi-norm, 23
elliptic PDE, 57
ellipticity condition, 58 Marcinkiewicz interpolation theorem, 24
essential support, 27 Mikhlin multiplier theorem, 38
multiplication of a distribution, 5
Fourier multiplier operator, 38 multiplication operator, 61
73
74 INDEX
multiplier operator, 38 singular integral operator, 59
Sobolev embedding theorem, 56
Newtonian potential, 22 Sobolev space
order, 2 homogeneous, 51
inhomogeneous, 50
Parseval’s relation, 17 square function, 45
partition of unity, 7 strong type, 24
Plancherel’s relation, 17 sublinear operator, 24
principal value integral, 3 support, 2
product rule, 11 symbol, 58, 61
pseudodifferential operator, 61 symbolic calculus in Sm , 62
randomization, 45 tempered distribution, 15
regularity estimate, 58, 69 test function, 2
Riemann–Lebesgue lemma, 20
Riesz potential, 50 uncertainty principle, 15
Riesz transform, 42
weak derivative, 10
scale-invariance, 55 weak formulation of a PDE, i
Schwartz functions, 15 weak gradient, i
signum function, 19 weak solution, i
singular integral convolution operator, 23 weak type, 24
You might also like
- B5a: Techniques of Applied Mathematics: Derek E. Moulton Based On Notes by Andreas M Unch November 21, 2013No ratings yetB5a: Techniques of Applied Mathematics: Derek E. Moulton Based On Notes by Andreas M Unch November 21, 201359 pages
- Mathematical Aspects of String Theory 600e3e2717No ratings yetMathematical Aspects of String Theory 600e3e271744 pages
- Abbas C. - Functional Analysis-Math 920 (Spring 2003) (2003)No ratings yetAbbas C. - Functional Analysis-Math 920 (Spring 2003) (2003)135 pages
- AGM For Elliptic Curves: DTU (Copenhagen) September 2005No ratings yetAGM For Elliptic Curves: DTU (Copenhagen) September 200540 pages
- Algebraic Topology: Or257/teaching/notes/at PDFNo ratings yetAlgebraic Topology: Or257/teaching/notes/at PDF80 pages
- Maximum Principles and Geometric ApplicationsNo ratings yetMaximum Principles and Geometric Applications54 pages
- Differential Equations: A Visual Introduction For Beginners: by Dan UmbargerNo ratings yetDifferential Equations: A Visual Introduction For Beginners: by Dan Umbarger55 pages
- Course 231: Equations of Mathematical Physics: Notes by Chris BlairNo ratings yetCourse 231: Equations of Mathematical Physics: Notes by Chris Blair17 pages
- QUANTUM NOTES: Matrix Analysis - Linear Algebra On Complex Scalar Product SpacesNo ratings yetQUANTUM NOTES: Matrix Analysis - Linear Algebra On Complex Scalar Product Spaces31 pages
- Solving nonlinear PDE by geometry in matrix spaceNo ratings yetSolving nonlinear PDE by geometry in matrix space63 pages
- V22003 4th Semester Project Source FileNo ratings yetV22003 4th Semester Project Source File22 pages
- Complex Analysis and Conformal Mapping: C 2018 Peter J. OlverNo ratings yetComplex Analysis and Conformal Mapping: C 2018 Peter J. Olver84 pages
- Linear Elliptic Equations of Second Order: Lecture NotesNo ratings yetLinear Elliptic Equations of Second Order: Lecture Notes86 pages
- Geometry of Curves and Surfaces: Weiyi Zhang Mathematics Institute, University of WarwickNo ratings yetGeometry of Curves and Surfaces: Weiyi Zhang Mathematics Institute, University of Warwick78 pages
- Quantization and Arithmetic 1st Edition André Unterberger (Auth.) download100% (1)Quantization and Arithmetic 1st Edition André Unterberger (Auth.) download78 pages
- Kiselev, Roquejoffre, Ryzhik. Appetizers in Nonlinear PDEs (Notes, 2017)(195s)No ratings yetKiselev, Roquejoffre, Ryzhik. Appetizers in Nonlinear PDEs (Notes, 2017)(195s)195 pages
- Uppsala Universitet Matematiska InstitutionenNo ratings yetUppsala Universitet Matematiska Institutionen35 pages
- Quantization and Arithmetic 1st Edition André Unterberger (Auth.) - The special ebook edition is available for download now100% (1)Quantization and Arithmetic 1st Edition André Unterberger (Auth.) - The special ebook edition is available for download now75 pages
- (Ebook) Quantization and Arithmetic by André Unterberger (auth.) ISBN 3764387904 2024 Scribd Download100% (2)(Ebook) Quantization and Arithmetic by André Unterberger (auth.) ISBN 3764387904 2024 Scribd Download81 pages
- Get Quantization and Arithmetic 1st Edition André Unterberger (Auth.) free all chapters100% (2)Get Quantization and Arithmetic 1st Edition André Unterberger (Auth.) free all chapters81 pages
- Electronics Engineering Mastery Test 7 ECE Pre-Board PDFNo ratings yetElectronics Engineering Mastery Test 7 ECE Pre-Board PDF29 pages
- Chap 03 Current Electricity (Formula Sheet)No ratings yetChap 03 Current Electricity (Formula Sheet)5 pages
- Sheet of Emi Student Copy With Ans 1631552042685No ratings yetSheet of Emi Student Copy With Ans 163155204268561 pages
- Chapter 4 DET50063 - BRAKING OF DC MOTORNo ratings yetChapter 4 DET50063 - BRAKING OF DC MOTOR21 pages
- Diffraction and Interference Question BankNo ratings yetDiffraction and Interference Question Bank32 pages
- St. Joseph'S Ss Kakindu: Uace Resource Mock ExaminationsNo ratings yetSt. Joseph'S Ss Kakindu: Uace Resource Mock Examinations10 pages
- 12th - Physics - VOL 1 - EM - WWW - Tntextbooks.online100% (1)12th - Physics - VOL 1 - EM - WWW - Tntextbooks.online352 pages
- Solo Leveling, Vol. 4 (Chugong) (Z-Lib - Org) - Pagenumber - Watermark - Compressed100% (1)Solo Leveling, Vol. 4 (Chugong) (Z-Lib - Org) - Pagenumber - Watermark - Compressed321 pages
- Ssip 2022 Physical Sciences Learner Guide Session 5 11No ratings yetSsip 2022 Physical Sciences Learner Guide Session 5 1156 pages
- Topic 6 Sensors Learning Outcomes: at The End of This Topic, You Should Be Able To100% (2)Topic 6 Sensors Learning Outcomes: at The End of This Topic, You Should Be Able To78 pages
- Lab 1. Understanding The Work of Microscopes: Campbell Biology, Reece Et Al, 2016) )No ratings yetLab 1. Understanding The Work of Microscopes: Campbell Biology, Reece Et Al, 2016) )7 pages
- Bee302 Electrical Measurements InstrumentationNo ratings yetBee302 Electrical Measurements Instrumentation2 pages
- Power Quality Pblms Created by Drives and PQ Impact On DrivesNo ratings yetPower Quality Pblms Created by Drives and PQ Impact On Drives5 pages
- B5a: Techniques of Applied Mathematics: Derek E. Moulton Based On Notes by Andreas M Unch November 21, 2013B5a: Techniques of Applied Mathematics: Derek E. Moulton Based On Notes by Andreas M Unch November 21, 2013
- Abbas C. - Functional Analysis-Math 920 (Spring 2003) (2003)Abbas C. - Functional Analysis-Math 920 (Spring 2003) (2003)
- AGM For Elliptic Curves: DTU (Copenhagen) September 2005AGM For Elliptic Curves: DTU (Copenhagen) September 2005
- Differential Equations: A Visual Introduction For Beginners: by Dan UmbargerDifferential Equations: A Visual Introduction For Beginners: by Dan Umbarger
- Course 231: Equations of Mathematical Physics: Notes by Chris BlairCourse 231: Equations of Mathematical Physics: Notes by Chris Blair
- QUANTUM NOTES: Matrix Analysis - Linear Algebra On Complex Scalar Product SpacesQUANTUM NOTES: Matrix Analysis - Linear Algebra On Complex Scalar Product Spaces
- Complex Analysis and Conformal Mapping: C 2018 Peter J. OlverComplex Analysis and Conformal Mapping: C 2018 Peter J. Olver
- Linear Elliptic Equations of Second Order: Lecture NotesLinear Elliptic Equations of Second Order: Lecture Notes
- Geometry of Curves and Surfaces: Weiyi Zhang Mathematics Institute, University of WarwickGeometry of Curves and Surfaces: Weiyi Zhang Mathematics Institute, University of Warwick
- Quantization and Arithmetic 1st Edition André Unterberger (Auth.) downloadQuantization and Arithmetic 1st Edition André Unterberger (Auth.) download
- Kiselev, Roquejoffre, Ryzhik. Appetizers in Nonlinear PDEs (Notes, 2017)(195s)Kiselev, Roquejoffre, Ryzhik. Appetizers in Nonlinear PDEs (Notes, 2017)(195s)
- Quantization and Arithmetic 1st Edition André Unterberger (Auth.) - The special ebook edition is available for download nowQuantization and Arithmetic 1st Edition André Unterberger (Auth.) - The special ebook edition is available for download now
- (Ebook) Quantization and Arithmetic by André Unterberger (auth.) ISBN 3764387904 2024 Scribd Download(Ebook) Quantization and Arithmetic by André Unterberger (auth.) ISBN 3764387904 2024 Scribd Download
- Get Quantization and Arithmetic 1st Edition André Unterberger (Auth.) free all chaptersGet Quantization and Arithmetic 1st Edition André Unterberger (Auth.) free all chapters
- Geometric functions in computer aided geometric designFrom EverandGeometric functions in computer aided geometric design
- Multiple Integrals, A Collection of Solved ProblemsFrom EverandMultiple Integrals, A Collection of Solved Problems
- Electronics Engineering Mastery Test 7 ECE Pre-Board PDFElectronics Engineering Mastery Test 7 ECE Pre-Board PDF
- St. Joseph'S Ss Kakindu: Uace Resource Mock ExaminationsSt. Joseph'S Ss Kakindu: Uace Resource Mock Examinations
- 12th - Physics - VOL 1 - EM - WWW - Tntextbooks.online12th - Physics - VOL 1 - EM - WWW - Tntextbooks.online
- Solo Leveling, Vol. 4 (Chugong) (Z-Lib - Org) - Pagenumber - Watermark - CompressedSolo Leveling, Vol. 4 (Chugong) (Z-Lib - Org) - Pagenumber - Watermark - Compressed
- Ssip 2022 Physical Sciences Learner Guide Session 5 11Ssip 2022 Physical Sciences Learner Guide Session 5 11
- Topic 6 Sensors Learning Outcomes: at The End of This Topic, You Should Be Able ToTopic 6 Sensors Learning Outcomes: at The End of This Topic, You Should Be Able To
- Lab 1. Understanding The Work of Microscopes: Campbell Biology, Reece Et Al, 2016) )Lab 1. Understanding The Work of Microscopes: Campbell Biology, Reece Et Al, 2016) )
- Power Quality Pblms Created by Drives and PQ Impact On DrivesPower Quality Pblms Created by Drives and PQ Impact On Drives