Function spaces

Lecture notes, Fall 2008

Mikko Salo

Department of Mathematics and Statistics

University of Helsinki
 Contents

Chapter 1. Introduction 1
Chapter 2. Interpolation theory 5
2.1. Classical results 5
2.2. Abstract interpolation 13
2.3. Real interpolation 16
2.4. Interpolation of Lp spaces 20
Chapter 3. Fractional Sobolev spaces, Besov and Triebel spaces 27
3.1. Fourier analysis 28
3.2. Fractional Sobolev spaces 33
3.3. Littlewood-Paley theory 39
3.4. Besov and Triebel spaces 44
3.5. Hölder and Zygmund spaces 54
3.6. Embedding theorems 60
Bibliography 63

v
 CHAPTER 1

Introduction

In mathematical analysis one deals with functions which are dif-

ferentiable (such as continuously differentiable) or integrable (such as
square integrable or Lp ). It is often natural to combine the smoothness
and integrability requirements, which leads one to introduce various
spaces of functions.
This course will give a brief introduction to certain function spaces
which are commonly encountered in analysis. This will include Hölder,
Lipschitz, Zygmund, Sobolev, Besov, and Triebel-Lizorkin type spaces.
We will try to highlight typical uses of these spaces, and will also give
an account of interpolation theory which is an important tool in their
study.
The first part of the course covered integer order Sobolev spaces in
domains in Rn , following Evans [4, Chapter 5]. These lecture notes
contain the second part of the course. Here the emphasis is on Sobolev
type spaces where the smoothness index may be any real number. This
second part of the course is more or less self-contained, in that we will
use the first part mainly as motivation. Also, we will give new proofs
in more general settings of certain results in the first part.
Let us describe the structure of these notes. Interpolation theory
is covered in Chapter 2. We begin by proving two classical interpola-
tion results for Lp spaces, namely the Riesz-Thorin and Marcinkiewicz
theorems. Next the basic setup of abstract interpolation is discussed.
The main part of the chapter includes a development of the real in-
terpolation method in the setting of quasinormed Abelian groups, and
interpolation results for Banach-valued Lp spaces. The interpolation
results proved here or in Chapter 3 are not the most general available
(stronger results are found in Bergh-Löfström [1] and Triebel [13]).
Rather, the point is to give an idea of what interpolation theory is
about, without spending too much time on technicalities.

1
2 1. INTRODUCTION

Chapter 3 is an introduction to Sobolev, Besov, Triebel-Lizorkin,

and Zygmund spaces in Rn . These spaces are most conveniently defined
via the Fourier transform, and the first matter is to review (without
proofs) some basic facts about Fourier analysis on tempered distribu-
tions. It is then not difficult to define fractional Sobolev spaces using
the Bessel potentials. We discuss Littlewood-Paley theory to motivate
the definition of Besov and Triebel-Lizorkin spaces, which are then
defined in terms of suitable Littlewood-Paley partitions of unity.
Zygmund spaces are defined in terms of finite differences and it is
shown that these spaces coincide with the L∞ -based Besov spaces, thus
giving a Fourier transform characterization of Hölder and Zygmund
spaces. In the course of Chapter 3, we establish the following basic
relations between the various spaces:

H k,p = W k,p ,
s s
B22 = F22 = H s,2 ,
s
Fp2 = H s,p ,
s
B∞∞ = C∗s .

Finally, we discuss embedding theorems for these spaces.

There are several notable omissions in these notes. To retain the
real analytic flavor of the course, we do not discuss the complex in-
terpolation method. Also, in the discussion of the real interpolation
method, we restrict ourselves to the K-functional and do not consider
other related concepts such as the J-functional. In the chapter on
Sobolev and related spaces, one of the most basic results is the Mihlin
multiplier theorem stating the Lp boundedness of certain multipliers on
the Fourier side. We use this result extensively, but do not give a proof
of it (or of the related Littlewood-Paley theorem). All the results that
we need in this context follow from the vector-valued Mihlin theorem
(Theorem 3.4.5), which is a standard consequence of the theory of sin-
gular integrals (see for instance [2]). Also, we mostly restrict ourselves
to Lp based spaces with 1 ≤ p ≤ ∞ or 1 < p < ∞, and do not discuss
Hardy or BMO type spaces.
These notes are intended to be accessible to graduate or advanced
undergraduate students having some background in real and functional
analysis and multivariable calculus (the real analysis part of Rudin [8]
should be more than sufficient).
 1. INTRODUCTION 3

References. We will not follow any particular reference in this

course, the main reference will be these lecture notes. Useful books on
interpolation theory and function spaces are Bergh-Löfström [1] and
Triebel [13]. Function spaces from a harmonic analysis point of view
are discussed in Grafakos [6] and Stein [11].
Notation. We will use multi-index notation: if α = (α1 , . . . , αn )
is in Nn and x = (x1 , . . . , xn ) ∈ Rn , we write
xα := xα1 1 · · · xαnn , Dα := D1α1 · · · Dnαn ,
1 ∂
where Dj := i ∂xj
. We will also use the Japanese bracket
hxi := (1 + |x|2 )1/2 .
The equivalence of two (quasi)norms is written as kuk1 ∼ kuk2 .
This means that there exists a constant C > 0 such that
1
kuk1 ≤ kuk2 ≤ Ckuk1
C
for all relevant functions u. If A and B are two (quasi)normed spaces
or groups, we write A ⊆ B to denote that A is contained in B in the
set-theoretic sense, and the inclusion map i : A → B is continuous. We
also write A = B to denote that A and B are equal as sets, and they
have equivalent (quasi)norms.
 CHAPTER 2

Interpolation theory

2.1. Classical results

We will begin by proving two classical interpolation theorems. The
first was proved by Riesz (1926) and Thorin (1938), and it forms the
basis for the complex interpolation method. The second is due to
Marcinkiewicz (1939) and Zygmund (1956), and it is a precursor of the
real interpolation method.
Both these results concern interpolation of Lp spaces. Their nature
is illustrated by the following special case: if T is a bounded linear
operator Lp0 → Lp0 and Lp1 → Lp1 , where p0 < p1 , then T is also
bounded Lp → Lp for p0 ≤ p ≤ p1 . Thus, it is enough to establish
estimates for two end point values of p, and interpolation will then
give estimates for all the intermediate values of p for free.

2.1.1. Riesz-Thorin theorem. We will prove the theorem for

the spaces Lp (Rn ), although the same proof applies to Lp (X, µ) and
Lq (Y, ν) where X, Y are measure spaces and µ, ν are σ-finite positive
measures.

Theorem 1. (Riesz-Thorin interpolation theorem) Suppose T is a

complex-linear map from Lp0 + Lp1 to Lq0 + Lq1 , 1 ≤ p0 , p1 , q0 , q1 ≤ ∞,
such that

kT f kLq0 ≤ M0 kf kLp0 , f ∈ Lp0 ,

kT f kLq1 ≤ M1 kf kLp1 , f ∈ Lp1 .

For 0 < θ < 1, define p and q by

1 1−θ θ 1 1−θ θ
= + , = + .
p p0 p1 q q0 q1
If 1 < q < ∞ then T maps Lp to Lq , and

(1) kT f kLq ≤ M01−θ M1θ kf kLp , f ∈ Lp .

5
6 2. INTERPOLATION THEORY

Proof. 1. If p0 = p1 = p, then (1) follows from the interpolation

inequality for Lq norms (that is, the Hölder inequality):
kT f kLq ≤ kT f kL1−θ θ 1−θ
q0 kT f kLq1 ≤ M0 M1θ kf kLp .
Thus we may assume p0 6= p1 , and so 1 < p < ∞. Also, we may assume
that M0 , M1 > 0 since otherwise one would have T ≡ 0.
R
2. Write hf, gi = Rn
f g dx. We claim that it is enough to prove
0
(2) |hT (f0 F 1/p ), g0 G1/q i| ≤ M01−θ M1θ
for any measurable f0 , g0 with |f0 |, |g0 | ≤ 1, and for any simple func-
tions F, G ≥ 0 satisfying kF kL1 = kGkL1 = 1.
In fact, since 1 < q < ∞, (1) is equivalent with
|hT f, gi| ≤ M01−θ M1θ kf kLp kgkLq0 .
0
Fix f ∈ Lp , g ∈ Lq with f, g 6= 0, and let

f (x)/|f (x)| if f (x) 6= 0,
f0 (x) :=
0 otherwise.
0 0
Define g0 in a similar way. If F := |f |p /kf kpLp and G := |g|q /kgkqLq0 ,
then the last estimate may be written as
0
|hT (f0 F 1/p ), g0 G1/q i| ≤ M01−θ M1θ .
Now, there exist nonnegative simple functions F̃j with kF̃j kLp = 1 such
that F̃j → F 1/p in Lp . Then Fj := F̃jp ≥ 0 are simple functions with
1/p
kFj kL1 = 1 and Fj → F 1/p in L1 . Using a similar approximation for
G, we see that the theorem will follow from the claim (2).

3. To prove (2), let |f0 |, |g0 | ≤ 1 and let F, G be simple functions

M
X N
X
F := ak χAk , G := bl χBl
k=1 l=1

where ak , bl > 0 and kF kL1 = kGkL1 = 1. Define

1−z 1−z z
+ pz 0 + q0
Φ(z) := hT (f0 F p0 1 ), g0 G q0 1 i,
for z in the strip {z ∈ C ; 0 ≤ Re z ≤ 1}. Here we define az = ez log a
for a > 0, so that |az | = aRe z and |ait | = 1 for t ∈ R.
 2.1. CLASSICAL RESULTS 7

We have
1−z 1−z z
X + pz 0 + q0
q0
p0
Φ(z) = ak 1
bl 1
hT (f0 χAk ), g0 χBl i.
k,l

Thus, Φ is analytic for 0 < Re z < 1 and bounded and continuous for
0 ≤ Re z ≤ 1. By Hölder’s inequality and using the assumptions on T ,
1−it 1−it it
+ pit 0 + q0
|Φ(it)| ≤ kT (f0 F p0 1 )kLq0 kg0 G q0 1 kLq00 ≤ M0
and similarly
−it −it 1+it
+ 1+it 0 + q0
|Φ(1 + it)| ≤ kT (f0 F p0 p1
)kLq1 kg0 G q0 1 kLq10 ≤ M1 .
The claim (2) follows from the three lines theorem below. 

Theorem 2. (Three lines theorem) Assume that Φ is analytic in

0 < Re z < 1 and bounded and continuous in 0 ≤ Re z ≤ 1. Suppose
that for t ∈ R,
|Φ(it)| ≤ M0 , |Φ(1 + it)| ≤ M1 .
Then |Φ(θ + it)| ≤ M01−θ M1θ for 0 < θ < 1 and t ∈ R.

Proof. We only consider the case where M0 , M1 > 0. The function

−(1−z)
Ψ(z) := M0 M1−z Φ(z)
is analytic in 0 < Re z < 1, bounded and continuous in 0 ≤ Re z ≤ 1,
and satisfies |Ψ(it)| ≤ 1 and |Ψ(1 + it)| ≤ 1. It is enough to prove that
(3) |Ψ(θ + it)| ≤ 1
for 0 < θ < 1.
Now, if sup0≤θ≤1 |Ψ(θ + it)| → 0 as t → ±∞, one obtains (3) by
noting that |Ψ| ≤ 1 on all four sides of the rectangle [0, 1] × [−R, R] for
R large, thus |Ψ| ≤ 1 inside the rectangle by the maximum principle.
If Ψ does not decay as Im z → ±∞, we consider the functions
Ψε (z) := exp(ε(z 2 − 1))Ψ(z), ε > 0.
Since (s + it)2 = s2 − t2 + 2ist, we have |Ψε (it)| ≤ 1 and |Ψε (1 + it)| ≤ 1
and also |Ψε (z)| → 0 as Im z → ±∞. Thus (3) holds for Ψε . Since
2 −t2 −1)
|Ψε (θ + it)| = eε(θ |Ψ(θ + it)|
where ε > 0 is arbitrary, we obtain the result by letting ε → 0. 
8 2. INTERPOLATION THEORY

A standard application of the Riesz-Thorin interpolation theorem is

the Hausdorff-Young inequality for Fourier transforms. If f ∈ L1 (Rn ),
the Fourier transform of f is the function F f = fˆ ∈ L∞ (Rn ) given by
Z
ˆ
f (ξ) = e−ix·ξ f (x) dx, ξ ∈ Rn .
Rn

It immediately follows that

(4) kfˆkL∞ ≤ kf kL1 .
Now, if f ∈ L1 ∩ L2 , the Plancherel (or Parseval) identity states that
fˆ ∈ L2 and
(5) kfˆkL2 = (2π)n/2 kf kL2 .
This can be used to extend the Fourier transform to any L2 function,
and the Plancherel identity remains valid.

Theorem 3. (Hausdorff-Young inequality) If 1 ≤ p ≤ 2 then the

0
Fourier transform maps Lp (Rn ) into Lp (Rn ), and
0
kfˆkLp0 ≤ (2π)n/p kf kLp .

Proof. We saw above that the Fourier transform is a well defined

map L1 → L∞ and L2 → L2 , thus it maps L1 + L2 into L2 + L∞ . We
have the norm bounds (4) and (5). Let 0 < θ < 1, and define p by
1 1−θ θ
= + .
p 1 2
Then θ = 2/p0 and conseqently 1/p0 = θ/2 = (1 − θ)/∞ + θ/2.
Now, the Riesz-Thorin theorem implies that the Fourier transform
0
maps Lp to Lp with the right norm bound. Since any p with 1 < p < 2
can be obtained in this way, the result follows. 

Remark. (Hausdorff-Young for Fourier series) The Fourier coeffi-

cients are given by the map L2 ((0, 2π)) → l2 , f 7→ fˆ, where
Z 2π
ˆ 1
f (k) = f (x)e−ikx dx, k ∈ Z.
2π 0
One has the norm bounds
kfˆkl∞ ≤ (2π)−1 kf kL1 ((0,2π)) ,
kfˆkl2 = (2π)−1/2 kf kL2 ((0,2π)) ,
 2.1. CLASSICAL RESULTS 9

the second bound coming from Parseval’s theorem. The Riesz-Thorin

theorem applied to Lp and lp spaces gives that
kfˆk p0 ≤ (2π)−1/p kf kLp ((0,2π))
l

for 1 ≤ p ≤ 2.
2.1.2. Marcinkiewicz theorem. The Riesz-Thorin theorem im-
plies that if a linear operator is known to be bounded on Lp0 and
Lp1 , then it is bounded on Lp for any p between p0 and p1 . The
Marcinkiewicz theorem gives a similar conclusion, but here it is enough
to have weak type bounds at the endpoints p0 and p1 to obtain strong
bounds for intermediate p. The weak type bounds are given in terms
of the distribution function
m(|f | > λ) := m({x ∈ Rn ; |f (x)| > λ})
where m is the Lebesgue measure on Rn .
Motivation. Let f ∈ Lp (Rn ). The elementary identity
Z a
p
a = pλp−1 dλ, a ≥ 0,
0
implies Z ∞
p
|f (x)| = pλp−1 χ{|f |>λ} (x) dλ.
0
Integrating over x, we obtain
Z Z ∞
p
(6) |f | dx = pλp−1 m(|f | > λ) dλ.
Rn 0
This indicates why distribution functions might be useful in studying
Lp norms. Also, if T : Lp (Rn ) → Lq (Rn ) is a bounded linear operator,
then
Ckf kqLp
Z Z
1
m(|T f | > λ) = dx ≤ q |T f |q dx ≤ .
{|T f |>λ} λ {|T f |>λ} λq
In particular, if T is the identity operator, then
λm(|f | > λ)1/p ≤ kf kLp , f ∈ Lp .
Definition. The weak Lp space, written Lp ∞ (Rn ) (0 < p < ∞),
consists of those measurable functions on Rn for which the expression
kf kLp ∞ := sup λm(|f | > λ)1/p
λ>0
∞∞ ∞
is finite. We define L =L .
10 2. INTERPOLATION THEORY

p−ε
We have Lp ⊆ Lp ∞ ⊆ Lloc if p − ε > 0 (exercise). The function
−α
f (x) = |x| , α > 0, is in weak Lp iff p = n/α, since

m(|f | > λ) = m(|x| < λ−1/α ) = Cn λ−n/α

n
and λm(|f | > λ)1/p = Cn,p λ1− αp is uniformly bounded iff αp = n. Note
that k · kLp ∞ is not a norm if 1 ≤ p < ∞.

Definition. Let T be an operator (not necessarily linear) from

L (Rn ) to the space of measurable functions on Rn . We say that
p

(a) T is subadditive if for almost every x ∈ Rn ,

|T (f + g)(x)| ≤ |T f (x)| + |T g(x)|,

(b) T is strong type (p, q) if it maps Lp (Rn ) into Lq (Rn ) and

kT f kLq ≤ Ckf kLp , f ∈ Lp , and

(c) T is weak type (p, q), q < ∞, if

 q
Ckf kLp
m(|T f | > λ) ≤ , f ∈ Lp , λ > 0.
λ

Further, T is weak type (p, ∞) if it is strong type (p, ∞).

From the remarks above it follows that any operator of strong type
(p, q) is also of weak type (p, q), and that a weak type (p, q) operator
maps Lp into weak Lq . The interpolation result is as follows. Note that
as in the Riesz-Thorin theorem, we could easily replace Rn by more
general measure spaces on the right and on the left.

Theorem 4. (Marcinkiewicz interpolation theorem) Suppose that

1 ≤ p0 < p1 ≤ ∞, and let T be a subadditive operator from Lp0 + Lp1 to
the space of measurable functions on Rn . Assume that T is weak type
(p0 , p0 ) and (p1 , p1 ). Then T is strong type (p, p) for p0 < p < p1 .

Proof. 1. First assume p1 < ∞. The proof is based on the idea

that any f ∈ Lp , p0 < p < p1 , may be decomposed as f = f0 + f1
where f0 ∈ Lp0 and f1 ∈ Lp1 , by taking

f (x) if |f (x)| > γ,
f0 (x) :=
0 otherwise.
 2.1. CLASSICAL RESULTS 11

Here γ > 0 is a fixed height. We have

Z Z Z
p0 p0 p0 −p
|f0 | dx = |f | dx ≤ γ |f |p dx,
{|f |>γ} {|f |>γ}
Z Z Z
|f1 |p1 dx = |f |p1 dx ≤ γ p1 −p |f |p dx.
{|f |≤γ} {|f |≤γ}

2. Fix λ > 0. If f ∈ Lp , we make the decomposition f = f0 + f1 where

γ = λ. By subadditivity we have
{|T f | > λ} ⊆ {|T f0 | > λ/2} ∪ {|T f1 | > λ/2}.
Using the assumptions on T , the distribution function satisfies
m(|T f | > λ) ≤ m(|T f0 | > λ/2) + m(|T f1 | > λ/2)
 p  p
2C0 kf0 kLp0 0 2C1 kf1 kLp1 1
≤ +
λ λ
 p0 Z  p Z
2C0 p0 2C1 1
≤ |f | dx + |f |p1 dx.
λ {|f |>λ} λ {|f |≤λ}

Note that the result in Step 1 would immediately imply that T is weak
type (p, p).

3. To obtain strong type bounds, we express the Lp norm in terms of

the distribution function as in (6):
Z ∞
p
kT f kLp = pλp−1 m(|T f | > λ) dλ
0
Z ∞Z
p0
≤ p(2C0 ) λp−1−p0 |f |p0 dx dλ
0 {|f |>λ}
Z ∞Z
+ p(2C1 )p1 λp−1−p1 |f |p1 dx dλ.
0 {|f |≤λ}

We estimate the first integral by using Fubini’s theorem:

Z ∞Z Z Z |f (x)|
p−1−p0 p0
λ |f | dx dλ = λp−1−p0 |f |p0 dx dλ
0 {|f |>λ} Rn 0
p−p0
|f (x)|
Z
1
= |f (x)|p0 dx = kf kpLp .
Rn p − p0 p − p0
Similarly, the second integral is equal to
Z Z ∞
1
λp−1−p1 |f |p1 dx dλ = kf kpLp .
Rn |f (x)| p1 − p
12 2. INTERPOLATION THEORY

Both these computations used the assumption p0 < p < p1 . The

theorem is proved in the case p1 < ∞, and the case p1 = ∞ is left as
an exercise. 
The Marcinkiewicz theorem is a basic result in real analysis with
many applications, such as establishing the Lp mapping properties of
the Hilbert transform or the Hardy-Littlewood maximal function. We
will instead give a simple result concerning the Hardy operator, defined
for f ∈ L1loc (R+ ) by
1 x
Z
T f (x) = f (t) dt, x > 0.
x 0
Note that the proof of the Marcinkiewicz theorem works equally well
with Rn replaced by R+ , which is the setting in the next result. The
Hardy operator is not strong type (1, 1), so here it is essential that
weak type bounds at the end points are sufficient.
Theorem 5. (Mapping properties for T ) The Hardy operator maps
L (R+ ) to weak L1 (R+ ), and is a bounded operator Lp (R+ ) → Lp (R+ )
1

for 1 < p ≤ ∞.
Proof. Since |T f (x)| ≤ x1 kf kL1 (R+ ) , we have
1 kf kL1 (R+ )
m(|T f | > λ) ≤ m({x ∈ R+ ; kf kL1 (R+ ) > λ}) = .
x λ
Thus T is weak type (1, 1), and clearly T is strong type (∞, ∞). The
result follows from Marcinkiewicz interpolation. 
Remark. The Marcinkiewicz theorem is true also when T is a
subadditive operator of weak type (p0 , q0 ) and (p1 , q1 ), where p0 6= p1
and q0 6= q1 . Then the conclusion is that T is strong type (p, q) if p
and q are as in the Riesz-Thorin theorem, and if one has the additional
restrictions
(7) p0 ≤ q0 , p1 ≤ q1 .
Although the Riesz-Thorin and Marcinkiewicz theorems look similar,
there are notable differences. In Marcinkiewicz, it is enough to have
weak type bounds at the endpoints, and the result applies to certain
nonlinear operators. In Riesz-Thorin, the restrictions (7) do not ap-
pear, and one can prove Lp bounds in the more general case where T
depends analytically on θ.
 2.2. ABSTRACT INTERPOLATION 13

2.2. Abstract interpolation

We will mostly interpolate Banach spaces. However, for technical
reasons it will be convenient to forget a great part of the structure in
Banach spaces, and to work with quasinormed Abelian groups.

Definition. Let A be an Abelian group that is written additively.

If c ≥ 1 is a constant, a c-quasinorm k · k is a real valued function on
A satisfying
(1) kak ≥ 0 for all a ∈ A, and kak = 0 iff a = 0,
(2) k−ak = kak for a ∈ A, and
(3) ka + bk ≤ c(kak + kbk) for a, b ∈ A.
We call (A, k · k) a quasinormed group.

Definition. A quasinormed vector space is a vector space which is

a quasinormed group and where the quasinorm satisfies kλak = |λ|kak
for all scalars λ.

Example. If 0 < p < ∞, let Lp (Rn ) be the set of all measurable

functions on Rn for which the expression
Z 1/p
p
kf kLp := |f | dx
Rn

is finite. We have
1−p
kf + gkLp ≤ max(1, 2 p )(kf kLp + kgkLp ).
Thus, if we identify functions which agree outside a set of measure zero,
Lp (Rn ) will be a quasinormed vector space if 0 < p < ∞.

The topology on a quasinormed group A is defined as follows: a

set U ⊆ A is open if for all a ∈ U , there is ε > 0 such that the set
{b ∈ A ; kb − ak < ε} is contained in U . The following result shows
that (A, k · k) is in fact a metric space.

Lemma 1. (Quasinormed groups are metric spaces) Suppose that A

is a c-quasinormed group, and let ρ be defined by the identity (2c)ρ = 2.
Then there is a 1-quasinorm k · k∗ on A such that
kak∗ ≤ kakρ ≤ 2kak∗ , a ∈ A.
Also, d(a, b) := kb − ak∗ is a metric defining the topology on A.
14 2. INTERPOLATION THEORY

Proof. If a ∈ A, define
N
X N
X
∗ ρ
kak := inf{ kaj k ; aj = a, N ≥ 1}.
j=1 j=1

Given this definition, we leave the proof as an exercise. 

Since every quasinormed vector space is metrizable, it makes sense

to talk about Cauchy sequences and completeness.

Definition. A complete quasinormed vector space will be called

a quasi-Banach space.

The Lp spaces are quasi-Banach if 0 < p ≤ ∞ (and of course Banach

if 1 ≤ p ≤ ∞). Also, the weak Lp spaces are quasi-Banach, and if p > 1
one can in fact find equivalent norms which make the weak Lp spaces
Banach spaces.
We now move to basic concepts of abstract interpolation.

Definition. Let A0 and A1 be quasinormed groups. We say that

(A0 , A1 ) is an interpolation couple if there is a quasinormed group A
such that A0 and A1 are subgroups of A and A0 , A1 ⊆ A with contin-
uous inclusions.

Remark. More generally, the reference group A could be any

topological group (such as a space of distributions on Rn ).

If (A0 , A1 ) is an interpolation couple, then the sets A0 ∩ A1 and

A0 + A1 are subgroups of A , where

A0 + A1 := {a ∈ A ; a = a0 + a1 for some a0 ∈ A0 , a1 ∈ A1 }.

We define

kakA0 ∩A1 := max(kakA0 , kakA1 ),

kakA0 +A1 := inf (ka0 kA0 + ka1 kA1 ).
a=a0 +a1 ,aj ∈Aj

Lemma 2. If (A0 , A1 ) is an interpolation couple, then A0 ∩ A1 and

A0 + A1 are quasinormed groups.

Proof. Clearly k · kA0 ∩A1 is nonnegative, satisfies kakA0 ∩A1 = 0

iff a = 0, and k−akA0 ∩A1 = kakA0 ∩A1 . The quasi-triangle inequality
 2.2. ABSTRACT INTERPOLATION 15

follows since
ka + bkA0 ∩A1 ≤ max(c0 (kakA0 + kbkA0 ), c1 (kakA1 + kbkA1 ))
≤ max(c0 , c1 ) max(kakA0 + kbkA0 , kakA1 + kbkA1 )
≤ max(c0 , c1 )(kakA0 ∩A1 + kbkA0 ∩A1 ).
For k · kA0 +A1 , nonnegativity and invariance under a 7→ −a are
clear. If kakA0 +A1 = 0, then for any ε > 0 there are aj,ε ∈ Aj with
a = a0,ε + a1,ε and ka0,ε kA0 + ka1,ε kA1 < ε. Thus aj,ε → 0 in Aj as
ε → 0, so also a = a0,ε + a1,ε → 0 in A since Aj ⊆ A continuously.
Finally, if a, b ∈ A0 + A1 and if a = a0 + a1 , b = b0 + b1 with aj , bj ∈ Aj ,
we have
ka + bkA0 +A1 ≤ ka0 + b0 kA0 + ka1 + b1 kA1
≤ max(c0 , c1 )(ka0 k + ka1 k + kb0 k + kb1 k).
Taking the infimum over all such decompositions of a and b, we obtain
ka + bkA0 +A1 ≤ max(c0 , c1 )(kakA0 +A1 + kbkA0 +A1 ).

Now, one has the continuous inclusions
A0 ∩ A1 ⊆ Aj ⊆ A0 + A1 ,
and A0 ∩ A1 and A0 + A1 are the maximal and minimal groups for
which this holds. The idea of interpolation theory is to find groups Aθ
such that
A0 ∩ A1 ⊆ Aθ ⊆ A0 + A1 ,
and Aθ should somehow be intermediate between A0 and A1 .
Also, as in the Riesz-Thorin and Marcinkiewicz theorems, we wish
to prove that boundedness of operators between endpoint spaces would
imply boundedness between intermediate spaces. We restrict our at-
tention to homomorphisms (that is, operators T : A → B such that
T (a + b) = T a + T b and T (−a) = −T a).
Definition. Let A and B be quasinormed groups, and let T : A →
B be a homomorphism. The quasinorm of T is
kT akB
kT k = kT kA→B := sup .
a∈A,a6=0 kakA

We say that T is bounded iff kT k < ∞.

16 2. INTERPOLATION THEORY

We are now ready to give the definition of an interpolation functor,

which incorporates the idea of intermediate groups and boundedness
of operators between them.
Definition. An interpolation functor is a rule F such that
(1) F assigns to every interpolation couple (A0 , A1 ) a quasinormed
group F (A0 , A1 ) such that
A0 ∩ A1 ⊆ F (A0 , A1 ) ⊆ A0 + A1
with continuous inclusions, and
(2) if (A0 , A1 ) and (B0 , B1 ) are any two interpolation couples and
T : A0 + A1 → B0 + B1 is a homomorphism which is bounded
A0 → B0 and A1 → B1 , then T is bounded F (A0 , A1 ) →
F (B0 , B1 ).
Definition. An interpolation functor F is of type θ if
kT kF (A0 ,A1 )→F (B0 ,B1 ) ≤ Cθ kT k1−θ θ
A0 →B0 kT kA1 →B1 .

The functor is said to be exact if one can choose Cθ = 1.

If j = 0, 1, then Fj (A0 , A1 ) := Aj are exact interpolation functors
of type j. The main purpose below will be to introduce nontrivial exact
interpolation functors obtained from real interpolation, and to study
their properties.

2.3. Real interpolation

The real interpolation method is due to Lions and Peetre (late 1950s
and early 1960s). There are several equivalent ways of introducing
the corresponding interpolation functors. We will use here the K-
functional for this purpose.
Definition. Let (A0 , A1 ) be an interpolation couple. We define,
for t > 0 and a ∈ A0 + A1 ,
K(t, a) = K(t, a; A0 , A1 ) := inf (ka0 kA0 + tka1 kA1 ).
a=a0 +a1 ,aj ∈Aj

The idea is that for each fixed t > 0, K(t, · ) is a quasinorm on

A0 + A1 which is equivalent (that is, gives the same topology) to the
usual one. We will define interpolation groups as subgroups of A0 + A1
by imposing restrictions on the behaviour of K(t, a) as t → 0 and
t → ∞. The values t correspond to the cut levels γ in the proof of the
Marcinkiewicz theorem.
 2.3. REAL INTERPOLATION 17

Definition. Let (A0 , A1 ) be an interpolation couple. If 0 < θ < 1

and 0 < q < ∞, and if a ∈ A0 + A1 , we define
Z ∞ 1/q
−θ q dt
(1) kak(A0 ,A1 )θ,q := [t K(t, a)]
0 t
We denote by (A0 , A1 )θ,q the set of all a ∈ A0 + A1 for which the
quantity (1) is finite.
Remark. Here θ is the main interpolation parameter, and q can
be used to fine-tune the interpolation. One could also consider the case
q = ∞, but we will not need to do so.
We shall prove that the rule (A0 , A1 ) 7→ (A0 , A1 )θ,q is an exact
interpolation functor of type θ. The first step is a simple lemma.
Lemma 1. (Properties of the K-functional) Suppose (A0 , A1 ) is an
interpolation couple. If a ∈ A0 + A1 is fixed, then K( · , a) is a positive,
nondecreasing, concave, and continuous function of t > 0. If Aj is
cj -quasinormed, we have
(2) K(t, a + b) ≤ c0 [K(c1 t/c0 , a) + K(c1 t/c0 , b)] , a, b ∈ A0 + A1 .
Proof. It is clear that K( · , a) is positive and nondecreasing. We
leave as an exercise to check that it is concave and continuous. If
a, b ∈ A0 + A1 and if a = a0 + a1 , b = b0 + b1 with aj , bj ∈ Aj , then
K(t, a + b) ≤ ka0 + b0 kA0 + tka1 + b1 kA1
≤ c0 (ka0 kA0 + kb0 kA0 ) + c1 t(ka1 kA1 + kb1 kA1 )
 
c1 t c1 t
≤ c0 ka0 kA0 + ka1 kA1 + kb0 kA0 + kb1 kA1 .
c0 c0
The result follows by taking the infimum over all decompositions of a
and b. 
Next we show that the first requirement for an interpolation functor
is satisfied.
Theorem 2. (Interpolation groups) Suppose (A0 , A1 ) is an inter-
polation couple, 0 < θ < 1, and 0 < q < ∞. Then (A0 , A1 )θ,q is a
quasinormed group, and
A0 ∩ A1 ⊆ (A0 , A1 )θ,q ⊆ A0 + A1
with bounded inclusions.
18 2. INTERPOLATION THEORY

Proof. 1. The first step is to prove that k · k := k · k(A0 ,A1 )θ,q satisfies
the conditions for a quasinorm. Clearly kak ≥ 0 and k−ak = kak,
and if kak = 0 then K(t, a) = 0 for a.e. t > 0, showing that a = 0
as in Lemma 2. For the quasi-triangle inequality we use (1) and the
quasi-triangle inequality on Lq := Lq (R+ , dtt ):

ka + bk = kt−θ K(t, a + b)kLq

1−q
≤ c0 max(1, 2 q )(kt−θ K(c1 t/c0 , a)kLq + kt−θ K(c1 t/c0 , b)kLq )
1−q
≤ c1−θ θ
0 c1 max(1, 2
q )(kak + kbk).

In the last step we changed the integration variable t to c0 t/c1 .

2. It immediately follows that (A0 , A1 )θ,q is a quasinormed group: if

a, b ∈ (A0 , A1 )θ,q then kak, kbk < ∞, so by Step 1 also ka + bk < ∞
and a + b ∈ (A0 , A1 )θ,q .

3. To show that A0 ∩ A1 ⊆ (A0 , A1 )θ,q with bounded inclusion, note

that for a ∈ A0 ∩ A1 we have K(t, a) ≤ kakA0 and K(t, a) ≤ tkakA1 , so

K(t, a) ≤ min(1, t)kakA0 ∩A1 .

Thus, if a ∈ A0 ∩ A1 , we have

kak ≤ kt−θ min(1, t)kLq (R+ ,dt/t) kakA0 ∩A1 ≤ Cθ,q kakA0 ∩A1 .

4. Finally, we prove that (A0 , A1 )θ,q ⊆ A0 +A1 with bounded inclusion.

This is the case t = 1 in the more general identity

K(t, a) ≤ Cθ,q tθ kak(A0 ,A1 )θ,q , a ∈ (A0 , A1 )θ,q , t > 0.

To show the last identity, since K( · , a) is nondecreasing we have

Z ∞ 1/q
−θ −θq ds
t K(t, a) = Cθ,q s K(t, a) ≤ Cθ,q kak(A0 ,A1 )θ,q .
t s
The result is proved. 

It remains to show the second requirement for an interpolation func-

tor, involving boundedness of operators. The definition of (A0 , A1 )θ,q
is set up so that this will be easy.
 2.3. REAL INTERPOLATION 19

Theorem 3. (Interpolation of operators) Let (A0 , A1 ) and (B0 , B1 )

be interpolation couples, and let T : A0 + A1 → B0 + B1 be a homo-
morphism which is bounded Aj → Bj with quasinorm Mj (j = 0, 1). If
0 < θ < 1 and 0 < q < ∞, then T is bounded (A0 , A1 )θ,q → (B0 , B1 )θ,q
and
kT k(A0 ,A1 )θ,q →(B0 ,B1 )θ,q ≤ M01−θ M1θ .
Proof. First assume M0 > 0, and let a ∈ (A0 , A1 )θ,q . Then we
have T a ∈ B0 + B1 , and
K(t, T a) ≤ inf (kT a0 kB0 + tkT a1 kB1 )
a=a0 +a1 ,aj ∈Aj

M1 t
≤ M0 inf (ka0 kA0 + ka1 kA1 )
a=a0 +a1 ,aj ∈Aj M0
= M0 K(M1 t/M0 , a).
Thus
kT ak(B0 ,B1 )θ,q = kt−θ K(t, T a)kLq (R+ ,dt/t)
≤ M0 kt−θ K(M1 t/M0 , a)kLq (R+ ,dt/t)
≤ M01−θ M1θ kak(A0 ,A1 )θ,q
by changing the variable t to M0 t/M1 .
If M0 = 0, we can repeat the above argument with M0 replaced by
ε > 0. The result is obtained by letting ε → 0. 
There is a density result which will be useful when interpolating Lp
and Sobolev spaces. It is usually proved via the J-method, which is an
equivalent way of introducing the spaces (A0 , A1 )θ,q in real interpola-
tion. Due to lack of time, we will not give the proof here.
Theorem 4. (Density of A0 ∩ A1 ) If (A0 , A1 ) is an interpolation
couple and 0 < θ < 1, 0 < q < ∞, then A0 ∩ A1 is dense in (A0 , A1 )θ,q .
Proof. See Bergh-Löfström [1, Section 3.11]. 
We end the section by stating some standard properties of real
interpolation, and by giving a simple example.
Theorem 5. (Elementary properties of real interpolation) Suppose
(A0 , A1 ) is an interpolation couple and 0 < θ < 1, 0 < q < ∞. Then
(1) (A0 , A1 )θ,q = (A1 , A0 )1−θ,q ,
(2) (A0 , A1 )θ,q ⊆ (A0 , A1 )θ,q̃ if q ≤ q̃,
20 2. INTERPOLATION THEORY

(3) if A0 = A1 then (A0 , A1 )θ,q = A0 = A1 ,

1−θ
(4) kak(A0 ,A1 )θ,q ≤ Cθ,q kakA 0
kakθA1 for a ∈ A0 ∩ A1 .
(5) if A0 and A1 are complete, then so is (A0 , A1 )θ,q .
Proof. Exercise. 
Example. Let A0 = R×{0} and A1 = R2 , with the usual additive
structure and with norms k(u, 0)kA0 = |u|, k(u, v)kA1 = |u| + |v|. To
determine (A0 , A1 )θ,q , note that if a = (u, v) ∈ A1 is written as a =
a0 + a1 with aj ∈ Aj , then a0 = (s, 0) and a1 = (u − s, v). Thus
K(t, a) = inf (|s| + t(|u − s| + |v|)) = min(1, t)|u| + t|v|
s∈R

since s 7→ |s| + t|u − s| is a continuous piecewise linear function going

to infinity as s → ±∞, which reaches its minimum at s = 0 or s = u.
Now, if v 6= 0 then
Z ∞ 1/q
−θ q dt
kak(A0 ,A1 )θ,q ≥ [t t|v|] = ∞.
1 t
Thus (A0 , A1 )θ,q = A0 = R × {0} for all θ and q.

2.4. Interpolation of Lp spaces

2.4.1. Interpolation of powers. When interpolating Lp spaces,
instead of the Lp norm it will be convenient to consider the quantity
kf kpLp which does not involve the pth root. We will next formalize this
idea.
Let (A, k · k) be a quasinormed group. If p > 0, then by the in-
equality
(x + y)p ≤ max(1, 2p−1 )(xp + y p ), x, y ≥ 0,
the quantity k · kp is also a quasinorm defining the topology on A.
We denote by (A)p the group A equipped with quasinorm k · kp . The
interpolation result for powers of groups is as follows.
Theorem 1. (Interpolation of powers) Let 0 < p0 , p1 < ∞ and let
0 < η < 1. If (A0 , A1 ) is an interpolation couple, then
((A0 )p0 , (A1 )p1 )η,1 = ((A0 , A1 )θ,p )p
with equivalent norms, where
p1
(1) p := (1 − η)p0 + ηp1 , θ := η.
p
 2.4. INTERPOLATION OF Lp SPACES 21

Proof. 1. We will use the functional

K∞ (t, a) = K∞ (t, a; A0 , A1 ) := inf max(ka0 kA0 , tka1 kA1 ).
a=a0 +a1 ,aj ∈Aj

This satisfies
K∞ (t, a) ≤ K(t, a) ≤ 2K∞ (t, a),
and thus kak(A0 ,A1 )θ,q ∼ kt−θ K∞ (t, a)kLq (R+ ,dt/t) .

2. If a ∈ A0 + A1 = (A0 )p0 + (A1 )p1 , then

Z ∞
ds
(2) kak((A0 )p0 ,(A1 )p1 )η,1 ∼ s−η K∞ (s, a; (A0 )p0 , (A1 )p1 ) .
0 s
We claim that
(3) K∞ (s, a; (A0 )p0 , (A1 )p1 ) = (K∞ (t, a; A0 , A1 ))p0
if s is defined by
(4) s := tp1 (K∞ (t, a; A0 , A1 ))p0 −p1 .
To prove (3), we temporarily denote the left and right hand sides by
K∞ (s) and K∞ (t). Then, using (4),
K∞ (s) = inf max(ka0 kpA00 , ska1 kpA11 )
a=a0 +a1
 p  p 
p0 ka0 kA0 0 tka1 kA1 1
= K∞ (t) inf max , .
a=a0 +a1 K∞ (t) K∞ (t)
Given ε > 0, we have K∞ (t) ≤ max(ka0,ε kA0 , tka1,ε kA1 ) ≤ (1 + ε)K∞ (t)
for some aj,ε ∈ Aj . Then
 
ka0,ε kA0 tka1,ε kA1
1 ≤ max , ≤ 1 + ε.
K∞ (t) K∞ (t)
The claim (3) follows.

3. We wish to go back to (2) and make the change of variables (4). To

this end, we claim that

s = s(t) in (4) is a strictly increasing bijective map R+ → R+
which is Lipschitz continuous on bounded intervals.
We start by proving the Lipschitz continuity. If 0 < t0 ≤ t1 , then
K∞ (t1 ) ≤ tt01 K∞ (t0 ) and
t1 − t0
(5) 0 ≤ K∞ (t1 ) − K∞ (t0 ) ≤ K∞ (t0 ).
t0
22 2. INTERPOLATION THEORY

Thus, if I := [1/R, R] ⊆ R+ with R large, then

 
|K∞ (t1 ) − K∞ (t0 )| ≤ R sup K∞ (t) |t1 − t0 |, t0 , t1 ∈ I.
t∈I

This shows that t 7→ K∞ (t) is Lipschitz continuous on bounded inter-

vals in R+ , and then so is s = s(t).
Whenever the derivative exists, we have
s0 (t)
 
t 0 1
= p1 + (p0 − p1 ) K∞ (t) .
s(t) K∞ (t) t
By (5), again at points where the derivative exists,
0 K∞ (t)
0 ≤ K∞ (t) ≤ .
t
Thus, we have
min(p0 , p1 ) s0 (t) max(p0 , p1 )
(6) ≤ ≤ .
t s(t) t
Consequently s(t) is strictly increasing, and it maps R+ onto itself.

4. By Step 3, we may change variables in (2) according to (4) (by

restricting the integral to the interval [1/R, R] and letting R → ∞).
Using (3) and the estimate (6), we obtain
Z ∞
dt
kak((A0 )p0 ,(A1 )p1 )η,1 ∼ t−p1 η K∞ (t)p .
0 t
The right hand side is kakp(A0 ,A1 )θ,p as required. 

Remark. (Exercise) Let p0 , p1 > 0. If 0 < η < 1 is given, then (1)

implies that
1 1−θ θ p0
(7) = + , 1 − θ = (1 − η) .
p p0 p1 p
Conversely, if 0 < θ < 1 is given, then (7) defines p and η so that (1)
is satisfied.
2.4.2. Interpolation of sequence spaces. We are now ready to
interpolate the sequence spaces lp . More generally, we will consider the
weighted Banach-valued sequence spaces lps (A) where 0 < p < ∞ and s
is a real number. The reason for writing p as a lower index is that later
on, s will correspond to the smoothness index in Sobolev type spaces
and p will denote the base Lp space. This notation will be consistent
with the notation for Sobolev type spaces.
 2.4. INTERPOLATION OF Lp SPACES 23

Definition. Let A be a Banach space. We denote by lps (A) the set

of all sequences a = (ak )∞
k=0 where ak ∈ A, for which the quasinorm

∞
!1/p
X
kaklps (A) := (2ks kak kA )p
k=0

is finite.

It is easy to see that lps (A) is a quasi-Banach space (Banach if p ≥ 1),

and that lp0 (A) is just the usual lp space of A-valued sequences.

Theorem 2. (Interpolation of lps (A) spaces) Let 0 < p0 , p1 < ∞

and let s0 , s1 be real numbers. Also let A0 and A1 be Banach spaces
such that (A0 , A1 ) is an interpolation couple. If 0 < θ < 1, then

(8) (lps00 (A0 ), lps11 (A1 ))θ,p = lps ((A0 , A1 )θ,p )

where p and s are defined by

1 1−θ θ
= + , s = (1 − θ)s0 + θs1 .
p p0 p1
Proof. 1. To make the formulas shorter, we omit writing A0 and A1 .
By Theorem 1 and the remark after it, we have

((lps00 )p0 , (lps11 )p1 )η,1 = ((lps00 , lps11 )θ,p )p

where 0 < η < 1 is such that p = (1 − η)p0 + ηp1 . Thus, it is enough

to study the space ((lps00 )p0 , (lps11 )p1 )η,1 .

2. First assume that a = (ak )∞ k=0 is a finite sequence with values in

A0 ∩ A1 , such that ak = 0 for k > N . Then
Z ∞
dt
kak((lp0 )p0 ,(lp1 )p1 )η,1 =
s s t−η K(t, a; (lps00 )p0 , (lps11 )p1 )
0 1
0 t
Z ∞ h i dt
= t−η inf kbkpls00 + tkckpls11
0 a=b+c p0 p1 t
Z ∞ N
X  dt
= t−η inf (2ks0 kbk k)p0 + t(2ks1 kck k)p1
0 a=b+c
k=0
t
24 2. INTERPOLATION THEORY

where b ∈ lps00 and c ∈ lsp11 . We may interchange the sum and infimum
and use Fubini’s theorem, and the last expression is equal to
N Z ∞
X  dt
t−η (2ks0 kbk k)p0 + t(2ks1 kck k)p1

= inf
k=0 0 ak =bk +ck t
N Z ∞
X dt
= t−η 2ks0 p0 K(t2k(s1 p1 −s0 p0 ) , ak ; (A0 )p0 , (A1 )p1 )
k=0 0 t
N
X
= 2k[(1−η)s0 p0 +ηs1 p1 ] kak k((A0 )p0 ,(A1 )p1 )η,1 ,
k=0

where the last equality follows by changing t to 2−k(s1 p1 −s0 p0 ) t in the

integral.
By Theorem 1 and the remark after it, we have

((A0 )p0 , (A1 )p1 )η,1 = ((A0 , A1 )θ,p )p ,

and by (7) it also follows that

(1 − η)s0 p0 + ηs1 p1 = (1 − θ)s0 p + θs1 p = sp.

Thus, using the result in Step 1, we have proved that

(9) kak(lps0 (A0 ),lps1 (A1 ))θ,p = kaklps ((A0 ,A1 )θ,p )
0 1

for all finite sequences a with values in A0 ∩ A1 .

3. We can now prove (8). We first claim that


the set of finite A0 ∩ A1 -valued sequences is dense
in both spaces in (8).
By Theorem 4 in §2.3, A0 ∩ A1 is dense in (A0 , A1 )θ,p and thus the
claim is true for lps ((A0 , A1 )θ,p ). In a similar fashion, lps00 (A0 ) ∩ lps11 (A1 ) =
lps00 (A0 ∩ A1 ) ∩ lps11 (A0 ∩ A1 ) is dense in (lps00 , lps11 )θ,p so the claim holds also
for the latter space.
Let now a ∈ (lps00 , lps11 )θ,p . This means that a ∈ lps00 + lps11 and that
kak(lps0 ,lps1 )θ,p is finite. Now, there is a sequence (a(N ) ) of finite A0 ∩ A1 -
0 1
valued sequences such that a(N ) → a in (lps00 , lps11 )θ,p . It follows from (9)
that (a(N ) ) is a Cauchy sequence in lps ((A0 , A1 )θ,p ), and by completeness
there exists ã ∈ lps ((A0 , A1 )θ,p ) with a(N ) → ã in lps ((A0 , A1 )θ,p ). It
 2.4. INTERPOLATION OF Lp SPACES 25

remains to show that a = ã, but this follows from uniqueness of limits
since
min(s ,s )
(lps00 , lps11 )θ,p , lps ((A0 , A1 )θ,p ) ⊆ lmax(p00 ,p11 ) (A0 + A1 )
continuously. We have proved that (lps00 , lps11 )θ,p ⊆ lps ((A0 , A1 )θ,p ), and
the other inclusion follows similarly. 

2.4.3. Interpolation of Lp spaces. Finally, we prove the inter-

polation result for Banach valued Lp spaces. First we need to recall
some concepts related to vector-valued integration.

Definition. Let (X, Σ, µ) be a measure space where µ is a positive

σ-finite measure, and let (A, k · k) be a Banach space.
(1) A simple function is any function s : X → A having the form
s(x) = N
P
j=1 aj χBj (x), where aj ∈ A and where Bj are disjoint
subsets of X with finite measure.
(2) A function f : X → A is strongly measurable if there exists a
sequence (sk ) of simple functions such that sk (x) → f (x) for
µ-a.e. x as k → ∞.
(3) If 0 < p < ∞, the space Lp (A) = Lp (X, Σ, µ; A) consists of
those strongly measurable functions f : X → A for which the
quasinorm
Z 1/p
p
kf kLp (A) := kf (x)k dµ(x)
X

is finite. We identify elements of Lp (X; A) which agree µ-a.e.

The last definition is valid since kf k is measurable if f is strongly

measurable. The spaces Lp (A) are quasi-Banach (Banach if p ≥ 1),
and simple functions are dense.

Theorem 3. (Interpolation of Lp spaces) Let (X, Σ, µ) be a σ-finite

measure space, let 0 < p0 , p1 < ∞, and let A0 and A1 be Banach spaces
so that (A0 , A1 ) is an interpolation couple. If 0 < θ < 1, then

(Lp0 (A0 ), Lp1 (A1 ))θ,p = Lp ((A0 , A1 )θ,p )

where p is defined by
1 1−θ θ
= + .
p p0 p1
26 2. INTERPOLATION THEORY

Proof. 1. The proof is analogous to the proof of Theorem 2, and we

will give a sketch. Again, we omit writing A0 and A1 . Theorem 1 and
the remark after it imply that
((Lp0 )p0 , (Lp1 )p1 )η,1 = ((Lp0 , Lp1 )θ,p )p
where 0 < η < 1 is such that p = (1 − η)p0 + ηp1 .

2. Let f be a simple function with values in A0 ∩ A1 . By Step 1, we

have
Z ∞
p dt
kf k(Lp0 ,Lp1 )θ,p ∼ t−η inf p (kf0 kpL0p0 + tkf1 kpL1p1 )
0 f =f0 +f1 ,fj ∈L j t
Z ∞ Z
dt
t−η kf0 (x)kpA00 + tkf1 (x)kpA11 dµ(x) .
 
∼ inf p
0 f =f0 +f1 ,fj ∈L j X t
We wish to interchange the infimum and the integral over X. In fact,
we claim that
Z
kf0 (x)kpA00 + tkf1 (x)kpA11 dµ(x)
 
inf
f =f0 +f1
X
fj ∈Lpj
Z
kf0 (x)kpA00 + tkf1 (x)kpA11 dµ(x).
 
= inf
f (x)=f0 (x)+f1 (x)
X
fj (x)∈Aj

The measurability of the integrand on the right is part of the claim.

The proof is left as an exercise. Now, Fubini’s theorem implies that
kf kp(Lp0 ,Lp1 )θ,p
Z Z ∞
dt
t−η kf0 (x)kpA00 + tkf1 (x)kpA11 dµ(x)
 
∼ inf
X 0 f (x)=f0 (x)+f1 (x) t
fj (x)∈Aj
Z Z ∞
dt
∼ t−η K(t, f (x); (A0 )p0 , (A1 )p1 ) dµ(x)
t
ZX 0 Z
∼ kf (x)k((A0 )p0 ,(A1 )p1 )η,1 dµ(x) ∼ kf (x)kp(A0 ,A1 )θ,p dµ(x).
X X
In the last step we used Theorem 1.

3. We have seen that the quasinorms on (Lp0 , Lp1 )θ,p and Lp ((A0 , A1 )θ,p )
are equivalent on simple functions with values in A0 ∩ A1 . One now
argues that such functions are dense in both of the spaces involved,
and the result follows as in the proof of Theorem 2. 
 CHAPTER 3

Fractional Sobolev spaces, Besov and Triebel

spaces

The purpose in this chapter is to extend the theory of integer order

Sobolev spaces W k,p to the case of fractional Sobolev spaces H s,p , Besov
s s
spaces Bpq , and Triebel-Lizorkin spaces Fpq . Here, k is a nonnegative
integer, s is any real number, and 1 < p, q < ∞ (or in some cases
0 < p, q ≤ ∞). With these conventions, the various spaces satisfy the
relations
W k,p = H k,p ,
H s,2 = B22
s s
= F22 ,
H s,p = Fp2
s
.
Because of the first identity, we will write W s,p := H s,p .
There are many situations where it is useful to go beyond the W k,p
spaces. We will describe a few such situations, in the setting of the
Dirichlet problem

−∆u = f in U,
u=g on ∂U
where U is a bounded open set in Rn with C 1 boundary.

(1) If f = 0 and g ∈ Lp (∂U ), then there is a unique solution

(
W 1/p,p (U ) if p ≥ 2,
u∈ 1/p
Bp2 (U ) if 1 < p < 2.

/ W 1,p (U ).
In general u ∈

(2) If one looks for solutions u ∈ W 1,p (U ), then the natural space
1−1/p 1−1/p
for g is Bpp (∂U ) = Fpp (∂U ). It can be shown that the
latter space is the image of W 1,p (U ) under the trace operator.

27
28 3. FRACTIONAL SOBOLEV SPACES

(3) If f ∈ Lp (U ) and g = 0, and if ∂U is C 2 , it can be shown that

there is a unique solution u ∈ W 2,p (U ). However, if ∂U is only
C 1 , then u ∈
/ W 2,p (U ) in general (for p = 2 the sharp result is
that u ∈ W 3/2,2 (U )).
The above results may be found in Jerison-Kenig [5].
Besides PDE, fractional Sobolev type spaces are useful in harmonic
analysis and approximation theory, where it is often possible to prove
sharper and more general statements by choosing the right spaces.
s s
Also, the scales of spaces Bpq and Fpq include many other spaces which
are common in analysis, such as Hölder, Zygmund, Hardy, and BMO
spaces.
Below, we will exclusively consider function spaces in Rn . The
theory for function spaces in bounded domains or lower dimensional
manifolds can be developed based on the corresponding theory for Rn .

3.1. Fourier analysis

It would be possible to define fractional Sobolev spaces via real
interpolation, for instance by W θ,2 (Rn ) := (L2 (Rn ), W 1,2 (Rn ))θ,2 for
0 < θ < 1. We will instead begin with an intrinsic definition based
on the Fourier transform, and we will later show that the interpolation
definition coincides with the intrinsic one.

Motivation. Let f ∈ L2 ((0, 2π)) be a 2π-periodic function (that

is, an L2 function on the torus T 1 := R/2πZ). One has the Fourier
series
∞
X
(1) f (x) = fˆ(k)eikx ,
k=−∞

R 2π
with fˆ(k) = 1
2π 0
f (x)e−ikx dx. Also, by Parseval’s theorem

kfˆkl2 = (2π)−1/2 kf kL2 ((0,2π)) .

1 d
If D = i dx
, differentiating (1) formally gives
∞
X
m
D f (x) = k m fˆ(k)eikx .
k=−∞
 3.1. FOURIER ANALYSIS 29

This suggests the following notion of a fractional derivative: if s ≥ 0,

then for any f ∈ L2 ((0, 2π)) we define
∞
X
s
|D| f (x) := |k|s fˆ(k)eikx ,
k=−∞

provided that the last series converges in some sense. If f is such

that (|k|s fˆ(k)) is square summable, then |D|s f is an L2 function by
Parseval’s theorem. This motivates the definition

W s,2 (T 1 ) := {f ∈ L2 ((0, 2π)) ; (|k|s fˆ(k)) ∈ l2 }, s ≥ 0,

which is the right definition for Sobolev spaces on the torus.

A similar idea works for Sobolev spaces in Rn , but one needs to re-
place Fourier series by the Fourier transform. We have already defined
the Fourier transform in §2.1 for functions in L1 (Rn ), but here a more
general definition is required.
In the following, we will give a brief review of the Fourier transform
in the general setting of tempered distributions. For proofs we refer to
Rudin [9, Chapter 7]. We begin by considering a test function space
designed for the purposes of Fourier analysis.

Definition. The Schwartz space S (Rn ) is the set of all infinitely

differentiable functions f : Rn → C such that the seminorms

kf kα,β := kxα Dβ f (x)kL∞ (Rn )

j=1 is a sequence in S ,
are finite for all multi-indices α, β ∈ Nn . If (fj )∞
we say that fj → f in S if kfj − f kα,β → 0 for all α, β.

It follows from the definition that a smooth function f is in S (Rn )

iff for all β and N there exists C > 0 such that

|Dβ f (x)| ≤ Chxi−N , x ∈ Rn .

Based on this, Schwartz space is sometimes called the space of rapidly

decreasing test functions.

Example. Any function in Cc∞ (Rn ) is in Schwartz space, and func-

tions like e−γ|x| , γ > 0, also belong to S . The function e−γ|x| is not in
2

Schwartz space because it is not smooth at the origin, and also hxi−N
is not in S because it doesn’t decrease sufficiently rapidly at infinity.
30 3. FRACTIONAL SOBOLEV SPACES

There is a metric defining the topology on S , and then S is a

complete metric space. The operations f 7→ P f and f 7→ Dβ f are
continuous maps S → S , if P is any polynomial and β any multi-
index. More generally, let
OM (Rn ) := {f ∈ C ∞ (Rn ) ; for all β there exist C, N > 0
such that |Dβ f (x)| ≤ ChxiN }.
It is easy to see that the map f 7→ af is continuous S → S if a ∈ OM .
Definition. If f ∈ S (Rn ), then the Fourier transform of f is the
function F f = fˆ : Rn → C defined by
Z
ˆ
f (ξ) := e−ix·ξ f (x) dx, ξ ∈ Rn .
Rn

The importance of Schwartz space is based on the fact that it is

invariant under the Fourier transform.
Theorem 1. (Fourier inversion formula) The Fourier transform
is an isomorphism from S (Rn ) onto S (Rn ). A Schwartz function f
may be recovered from its Fourier transform by the inversion formula
Z
−1 ˆ
f (x) = F f (x) = (2π) −n
eix·ξ fˆ(ξ) dξ, x ∈ Rn .
Rn

After introducing the Fourier transform on nicely behaving func-

tions, it is possible to define it in a very general setting by duality.
Definition. Let S 0 (Rn ) be the set of continuous linear functionals
S (Rn ) → C. The elements of S 0 are called tempered distributions,
and their action on test functions is written as
hT, ϕi := T (ϕ), T ∈ S 0, ϕ ∈ S .
j=1 is a sequence in S and if T ∈ S , we say that Tj → T in
If (Tj )∞ 0 0

S 0 if hTj , ϕi → hT, ϕi for all ϕ ∈ S (Rn ).

More precisely, S 0 is the dual space of S , and it is equipped with
the weak-* topology. The elements in S 0 are somewhat loosely also
called distributions of polynomial growth, which is justified by the fol-
lowing examples.
Examples. 1. Let u : Rn → C be a measurable function, such
that for some C, N > 0 one has
|u(x)| ≤ ChxiN , for a.e. x.
 3.1. FOURIER ANALYSIS 31

Then Tu ∈ S 0 (Rn ), where

Z
hTu , ϕi := u(x)ϕ(x) dx, ϕ ∈ S (Rn ).
Rn

We will always identify a function u with the tempered distribution Tu

defined in this way.
2. In a similar way, any function u ∈ Lp (Rn ) with 1 ≤ p ≤ ∞ is a
tempered distribution (with the identification u = Tu ).
3. Let µ be a positive Borel measure in Rn such that
Z
hxi−N dµ(x) < ∞
Rn

for some N > 0. Then Tµ is a tempered distribution, where

Z
hTµ , ϕi := ϕ(x) dµ(x), ϕ ∈ S (Rn ).
Rn

In particular, the Dirac measure δx0 at x0 ∈ Rn is in S 0 , and

hδx0 , ϕi = ϕ(x0 ), ϕ ∈ S (Rn ).

4. TheRfunction eγx is not in S 0 (R) if γ 6= 0, since it is not possible to

define R eγx ϕ(x) dx for all Schwartz functions ϕ. However, eγx cos(eγx )
belongs to S 0 since it is the distributional derivative (see below) of the
bounded function sin(eγx ) ∈ S 0 .

We now wish to extend some operations, defined earlier for Schwartz

functions, to the case of tempered distributions. This is possible via
duality. For instance, let a ∈ OM (Rn ). If u, ϕ ∈ S (Rn ) we have
hau, ϕi = hu, aϕi.
Since ϕ 7→ aϕ is continuous on S , we may for any T ∈ S 0 define the
product aT as the tempered distribution given by
haT, ϕi := hT, aϕi, ϕ ∈ S.
Similarly, motivated by the corresponding property for functions in S ,
if T ∈ S 0 then the (distributional) derivative Dβ T is the tempered
distribution given by
hDβ T, ϕi := (−1)|β| hT, Dβ ϕi, ϕ ∈ S.
32 3. FRACTIONAL SOBOLEV SPACES

Finally, we can define the Fourier transform of any T ∈ S 0 as the

tempered distribution F T = T̂ with

hT̂ , ϕi := hT, ϕ̂i, ϕ ∈ S.

We state the basic properties of these operations as a theorem.

Theorem 2. (Operations on S 0 ) If a ∈ OM and β is any multi-

index, then the multiplication T 7→ aT and the distributional derivative
T 7→ Dβ T are continuous maps S 0 → S 0 . Also, the Fourier transform
T 7→ T̂ is an isomorphism from S 0 onto S 0 , and one has the inversion
formula
hT, ϕi = (2π)−n hT̂ , ϕ̂(− · )i, ϕ ∈ S.
If T ∈ S then these operations coincide with the usual ones on S .

The preceding result shows that one may differentiate or Fourier

transform any tempered distribution (thus, also any Lp function or
measurable polynomially bounded function) any amount of times, and
the result will be a tempered distribution. We will give some examples
of this very general (and useful) phenomenon.

Examples. 1. Let u(x) := |x|, x ∈ R. Since u is continuous and

polynomially bounded, we have u ∈ S 0 (R). We claim the one has the
distributional derivatives

u0 = sgn(x),
u00 = 2δ0 .

In fact, if ϕ ∈ S (R), one has

Z 0 Z ∞
0 0 0
hu , ϕi = −hu, ϕ i = xϕ (x) dx − xϕ0 (x) dx
−∞ 0
Z
= sgn(x)ϕ(x) dx = hsgn(x), ϕi,
R

using integration by parts and the rapid decay of ϕ. Similarly,

Z 0 Z ∞
00 0 0 0
hu , ϕi = −hu , ϕ i = ϕ (x) dx − ϕ0 (x) dx
−∞ 0
= 2ϕ(0) = h2δ0 , ϕi.
 3.2. FRACTIONAL SOBOLEV SPACES 33

2. The Fourier transform of δ0 is the constant 1, since

Z ∞
hδb0 , ϕi = hδ0 , ϕ̂i = ϕ̂(0) = ϕ(x) dx
−∞
= h1, ϕi.
If u ∈ L2 (Rn ) then u is a tempered distribution, and the Fourier
transform û is another element of S 0 . The Plancherel theorem (which
is the exact analog of Parseval’s theorem for Fourier series) states that
in fact û ∈ L2 , and that the Fourier transform is an isometry on L2 up
to a constant.
Theorem 3. (Plancherel theorem) The Fourier transform is an
isomorphism from L2 (Rn ) onto L2 (Rn ), and one has
kûkL2 = (2π)n/2 kukL2 .
We end this section by noting the identities
(Dα u)ˆ = ξ α û,
(xα u)ˆ = (−Dξ )α û,
which hold for Schwartz functions u by a direct computation, and re-
main true for tempered distributions u by duality. This shows that
the Fourier transform converts derivatives into multiplication by poly-
nomials, and vice versa. This will be our route for defining fractional
Sobolev spaces: it is easy to define fractional derivatives of u by mul-
tiplying û with fractional powers of polynomials on the Fourier side.

3.2. Fractional Sobolev spaces

We are now in a position to give the Fourier transform definition of
fractional Sobolev spaces.
Motivation. We can characterize W k,2 (Rn ) by using the Fourier
transform and the Plancherel theorem by
u ∈ W k,2 (Rn ) ⇔ Dα u ∈ L2 (Rn ) for all |α| ≤ k
⇔ ξ α û ∈ L2 (Rn ) for all |α| ≤ k.
Thus, a function is in W k,2 (Rn ) iff its Fourier transform, multiplied by
any polynomial of degree ≤ k, is in L2 (Rn ). In fact, one polynomial is
sufficient: we have
u ∈ W k,2 (Rn ) ⇔ hξik û ∈ L2 (Rn ).
34 3. FRACTIONAL SOBOLEV SPACES

To see this, it is enough to note that |ξ α | ≤ |ξ||α| ≤ hξik for |α| ≤ k,

and that X
hξi2k = (1 + ξ12 + . . . + ξn2 )k = cγ ξ γ .
|γ|≤2k

This shows that hξik û ∈ L2 iff ξ α û ∈ L2 for |α| ≤ k, as desired.

Here and later, it will be convenient to introduce a notation for
multipliers on the Fourier side.
Definition. Let m(ξ) ∈ OM (Rn ). We define the Fourier multi-
plier operator
m(D)u := F −1 {m(ξ)û(ξ)}
for any tempered distribution u ∈ S 0 (Rn ).
The maps u 7→ û 7→ mû 7→ F −1 {mû} are continuous on S 0 , so
the operator m(D) is a well defined, continuous linear operator on S 0 .
Since m(ξ) = hξis is in OM for any s ∈ R, also the next definition is a
valid one.
Definition. Let 1 < p < ∞ and let s be a real number. The
fractional Sobolev space H s,p (Rn ) is the set of all tempered distributions
u ∈ S 0 (Rn ) such that
hDis u ∈ Lp (Rn ).
The norm is given by kukH s,p (Rn ) := khDis ukLp (Rn ) .
More explicitly, the space H s,p (Rn ) consists of those tempered dis-
tributions u in Rn , for which the tempered distribution F −1 {hξis û(ξ)}
happens to be an Lp function. The operator hDis is a Bessel potential,
and it is an isometric isomorphism from H s,p (Rn ) onto Lp (Rn ). The
spaces H s,p are also called Bessel potential spaces for this reason.
Theorem 1. (Sobolev spaces as function spaces) If s is a real num-
ber and 1 < p < ∞, then H s,p (Rn ) is a Banach space and S (Rn ) is a
dense subset.
Proof. If u, v ∈ H s,p and λ, µ are scalars, then hDis (λu + µv) =
λhDis u + µhDis v is an Lp function, showing that λu + µv ∈ H s,p . Thus
H s,p is a vector space.
To show that k · kH s,p is a norm, all the other properties are clear
except that fact that kukH s,p = 0 implies u = 0. But if kukH s,p = 0
then hDis u = 0 as an Lp function, thus also as a tempered distribution.
 3.2. FRACTIONAL SOBOLEV SPACES 35

Taking Fourier transforms, we obtain hξis û = 0, which implies û = 0

upon multiplying by the function hξi−s ∈ OM . Thus u = 0.
For completeness, let (uj ) be a Cauchy sequence in H s,p . It follows
that (hDis uj ) is a Cauchy sequence in Lp , and there exists ũ ∈ Lp such
that
hDis uj → ũ in Lp .
Let u := hDi−s ũ. Then u ∈ H s,p , and uj → u in H s,p as required.
The Schwartz space is a subset of H s,p , since F −1 {hξis fˆ} belongs
to S for f ∈ S (recall that the Fourier transform and multiplication
by OM functions are continuous maps S → S ). To show density, let
u ∈ H s,p . Then u0 := hDis u is an Lp function, and since 1 < p < ∞
there exists a sequence (vj ) ⊆ Cc∞ (Rn ) with vj → u0 in Lp . Let
uj := hDi−s vj .
Since vj ∈ S , we have uj ∈ S . Also,
kuj − ukH s,p = khDis uj − hDis ukLp = kvj − u0 kLp → 0
as j → ∞, showing that S is dense in H s,p . 
We next wish to show that H k,p (Rn ) = W k,p (Rn ) if k is a nonnega-
tive integer. The case p = 2 follows from the argument in the beginning
of this section. In general, if u ∈ H k,p and |α| ≤ k, we have
Dα u = Dα hDi−k (hDik u)
where hDik u ∈ Lp by definition. If one could show that Dα hDi−k maps
Lp to Lp if |α| ≤ k, then it would follow that Dα u ∈ Lp as required.
The following Fourier multiplier result will be used several times
below. For a proof, see [2].

Theorem 2. (Mihlin multiplier theorem) Assume that m = m(ξ)

is a bounded C ∞ complex valued function in Rn , which satisfies
|Dα m(ξ)| ≤ M hξi−|α| , ξ ∈ Rn ,
for |α| ≤ n2 + 1. Then m(D) is a bounded linear operator Lp (Rn ) →
 

Lp (Rn ) for 1 < p < ∞, and

km(D)ukLp ≤ Cp M kukLp .

In the next proof, notice how Fourier multipliers make it possible

to manipulate derivatives in an efficient way.
36 3. FRACTIONAL SOBOLEV SPACES

Theorem 3. (The space H s,p for s integer) If 1 < p < ∞ and if k

is a nonnegative integer, then H k,p (Rn ) = W k,p (Rn ).
Proof. 1. If u ∈ H k,p (Rn ), then Dα u = m(D)(hDik u) where
ξα
m(ξ) := .
hξik
If |α| ≤ k then m satisfies the conditions of the Mihlin multiplier
theorem. Consequently Dα u ∈ Lp for |α| ≤ k, showing that u ∈ W k,p .
2. Let u ∈ W k,p (Rn ). Then Dα u ∈ Lp for |α| ≤ k, and we need to
prove that hDik u ∈ Lp . Let χ(ξ) ∈ C ∞ (R) with 0 ≤ χ ≤ 1, χ = 0 for
|ξ| ≤ 1, and χ = 1 for |ξ| ≥ 2. We write
n
X
hDi u = F
k −1
{hξi û} = F
k −1
{m̃(ξ)(1 + χ(ξj )|ξj |k )û}
j=1

where
hξik
m̃(ξ) := Pn k
.
1+ j=1 χ(ξj )|ξj |
This shows that
n
X
k
hDi u = m̃(D)u + F −1 {m̃(ξ)χ(ξj )|ξj |k û}
j=1
n
X
= m̃(D)u + m̃(D)mj (Dj )(Djk u)
j=1

where

χ(ξj ), ξj ≥ 0,
mj (ξj ) := χ(ξj )|ξj |k ξj−k = k
(−1) χ(ξj ), ξj < 0.

3. We claim that m̃(ξ) satisfies the conditions of the Mihlin multiplier

theorem. Note that m̃ is smooth because χ = 0 near 0. Also,
n
X
k 2 k/2 k
hξi ≤ (1 + n max|ξj | ) ≤ C(1 + max|ξj | ) ≤ C(1 + |ξj |k ),
j j
j=1

which shows that m̃ is bounded. The derivatives satisfy |Dα m̃(ξ)| ≤

Chξi−|α| , so the Mihlin condition holds. We obtain
n
!
X
khDik ukLp ≤ C kukLp + kmj (Dj )(Djk u)kLp .
j=1
 3.2. FRACTIONAL SOBOLEV SPACES 37

4. Clearly mj (ξj ) satisfies the conditions in Mihlin’s theorem for n = 1.

We have

kmj (Dj )(Djk u)kLp (Rn ) = k kmj (Dj )(Djk u)kLp (R) kLp (Rn−1 )
≤ Ck kDjk ukLp (R) kLp (Rn−1 ) = CkDjk ukLp (Rn ) .

Since Djk u ∈ Lp for all j, we obtain hDik u ∈ Lp as required. 

Definition. We will sometimes write W s,p (Rn ) := H s,p (Rn ).

Note that u ∈ W k,p iff u ∈ W k−1,p and Dj u ∈ W k−1,p for all j.

Mihlin’s theorem allows to prove a similar result for H s,p .

Theorem 4. (Monotonicity of H s,p spaces) Let 1 < p < ∞ and let

s be real. Then
0
H s ,p (Rn ) ⊆ H s,p (Rn ) for s0 ≥ s,

and one has

u ∈ H s,p (Rn ) ⇔ u ∈ H s−1,p (Rn ), Dj u ∈ H s−1,p (Rn ) (1 ≤ j ≤ n).

Proof. Exercise. 

The preceding result implies in particular that H s,p is a subspace

of Lp if s ≥ 0, and that W k,p ⊆ H s,p if k ≥ s. If s is negative then H s,p
will contain distributions which are not functions.

Examples. 1. We have

δ0 ∈ H s,2 iff s < −n/2.

Indeed, since δb0 = 1, one has δ0 ∈ H s,2 iff F −1 {hξis } ∈ L2 iff hξis ∈ L2
by Plancherel. The last condition holds iff s < −n/2.

2. If 1 < p < ∞, a more careful analysis of F −1 {hξis } shows that

δ0 ∈ H s,p iff s < −n/p0 .

3. Consider the function u : R → R with


1 − |x| for |x| < 1,
u(x) :=
0 for |x| ≥ 1.
We claim that u ∈ H s,2 (R) iff s < 3/2. This can be proved by taking
the Fourier transform, or alternatively by considering distributional
38 3. FRACTIONAL SOBOLEV SPACES

derivatives:

 1 for −1 < x < 0,
0
u = −1 for 0 < x < 1,
0 for |x| ≥ 1,


u00 = δ−1 − 2δ0 + δ1 .

By part 1 in this example, u00 ∈ H s,2 (R) for s < −1/2. Then by
Theorem 4, u0 is in H s,2 (R) for s < 1/2, and consequently u ∈ H s,2 (R)
if s < 3/2. The ’only if’ part follows by showing that u00 ∈
/ H −1/2,2 (R)
(exercise).

We will continue to develop the theory of fractional Sobolev spaces

(embedding theorems, interpolation, equivalent characterizations) in
the more general setting of Triebel spaces later. Here, we will conclude
the section by determining the dual space of H s,p (in the dual pairing
of S and S 0 ).

Theorem 5. (Duality of H s,p spaces) Let 1 < p < ∞ and let s be

a real number. Then
0
(H s,p (Rn ))0 = H −s,p (Rn ).
More precisely, given a continuous linear functional T : H s,p → C,
0
there is a unique u ∈ H −s,p such that T (ϕ) = hu, ϕi for ϕ ∈ S .
0
Conversely, if u ∈ H −s,p , then u (acting on S ) has a unique extension
as a continuous linear functional on H s,p .

Proof. Let T : H s,p → C be a continuous linear functional. Then

|T (v)| ≤ CkhDis vkLp for all v ∈ H s,p , and we may define a functional
T̃ : Lp → C, T̃ (w) := T (hDi−s w).
This satisfies |T̃ (w)| ≤ CkwkLp , and the duality for Lp spaces implies
0
that there is a unique function u0 ∈ Lp such that
T̃ (w) = hu0 , wi, w ∈ Lp .
0
We let u := hDis u0 . Then u ∈ H −s,p and for ϕ ∈ S (expressed as
ϕ = ψ̂ for ψ ∈ S ) one has
hu, ϕi = hhDis u0 , ψ̂i = hhξis û0 , ψi = hû0 , hξis ψi = hu0 , (hξis ψ)ˆi
= hu0 , hDis ϕi
 3.3. LITTLEWOOD-PALEY THEORY 39

since F {hDis ϕ} = hξis F 2 {ψ} = F 2 {hξis ψ}. This shows that

hu, ϕi = T̃ (hDis ϕ) = T (ϕ)

as required.
0 0
Conversely, if u ∈ H −s,p , then u = hDis u0 with u0 ∈ Lp . Similarly
as above, we have for ϕ ∈ S

hu, ϕi = hhDis u0 , ϕi = hu0 , hDis ϕi

and consequently

|hu, ϕi| ≤ ku0 kLp0 khDis ϕkLp = kukH −s,p0 kϕkH s,p .

Thus u has an extension (by continuity) which is a continuous linear

functional on H s,p . 

3.3. Littlewood-Paley theory

For the discussion of Besov and Triebel spaces, it will be useful to
introduce some more properties for the Fourier transform in Lp . We
have seen many times that the L2 theory is simpler than the Lp theory,
the reason being that in the L2 case one can exploit orthogonality and
the Plancherel theorem. We will next introduce a partial substitute for
these properties in the Lp case.

Motivation. We will, as usual, use Fourier series in 1D to gain

intuition on the problem. If f is an L2 function in (0, 2π), one has the
Parseval theorem
Z 2π ∞
X
2
|f (x)| dx = 2π |fˆ(k)|2 .
0 k=−∞

We would like to express the Lp norm of f in terms of its Fourier

coefficients. For simplicity, we consider the case p = 4 and the function

N
X
f (x) = ak eikx .
k=0
40 3. FRACTIONAL SOBOLEV SPACES

To study the L4 norm of f , we compute

Z 2π N
Z 2π X !2 N !2
X
|f (x)|4 dx = ak eikx āl e−ilx dx
0 0 k=0 l=0
X Z 2π
= ak1 ak2 āl1 āl2 ei(k1 +k2 −l1 −l2 )x dx
0≤k1 ,k2 ,l1 ,l2 ≤N 0
X
= 2π ak1 ak2 āl1 āl2 .
0≤k1 ,k2 ,l1 ,l2 ≤N
k1 +k2 =l1 +l2

For given k1 and k2 , there are many l1 and l2 satisfying k1 +k2 = l1 +l2 ,
and it is hard to see any orthogonality in the above expression. How-
ever, in the special case where the Fourier series is sparse (or lacunary)
in the sense that

ak = 0 unless k = 2p for some integer p,

it follows that
Z 2π X
|f (x)|4 dx = 2π ak1 ak2 āl1 āl2
0 p q
kj =2 j ,lj =2 j
2p1 +2p2 =2q1 +2q2
X X
= 2π |ak |4 + 4π |ak1 |2 |ak2 |2 .
k=2p kj =2pj ,k1 6=k2

The last equality follows since 2p1 + 2p2 = 2q1 + 2q2 implies {p1 , p2 } =
{q1 , q2 }. Finally, we obtain
N
!2 Z N
!2
X 2π X
2π |ak |2 ≤ |f (x)|4 dx ≤ 4π |ak |2 ,
k=0 0 k=0

which can be written as

∞
!1/2
X
kf kL4 ∼ |fˆ(k)|2 .
k=−∞

j j+1
The above example indicates that the functions ei2 x and ei2 x are
somehow orthogonal in an Lp sense (this has to do with the fact that
the first function is roughly constant on the period of the second one).
A similar property remains true for eikx and eilx if k ∈ ∆j and l ∈ ∆j+1 ,
 3.3. LITTLEWOOD-PALEY THEORY 41

where

 {0}, j = 0,
j−1 j−1 j
∆j = {2 ,2 + 1, . . . , 2 − 1}, j > 0,
−∆−j , j < 0.


We introduce the dyadic (or Littlewood-Paley) decomposition of f ,

given in terms of
X
Sj f := fˆ(k)eikx .
k∈∆j

We denote by Sf the vector

Sf := (Sj f )j∈Z .

The Littlewood-Paley square function is the map

∞
!1/2
X
x 7→ kSf (x)kl2 = |Sj f (x)|2 .
j=−∞

The heart of Littlewood-Paley theory is the fact that the Lp norm of

a function is comparable to the Lp norm of the corresponding square
function.

Theorem 1. (Littlewood-Paley theorem for Fourier series) If f is

an Lp function in (0, 2π), and if 1 < p < ∞, then Sf ∈ Lp (l2 ) and

kf kLp ∼ kSf kLp (l2 ) .

Proof. See [3]. 

The theorem is saying that the Littlewood-Paley pieces Sj f of an

p
L function f are somewhat independent of each other. For instance,
one can multiply each such piece by ±1 and the resulting function will
still be in Lp . When proving Lp results in harmonic analysis it is often
enough to prove the result for a single Littlewood-Paley piece, and the
general case follows by summing over all such pieces.
The purpose in the remainder of this section is to discuss the coun-
terpart of Theorem 1 for the Fourier transform in Rn . This will involve
a smooth Littlewood-Paley decomposition of a function f in Lp (Rn ).

Definition. Let η = η(ξ) be a fixed radial function in Cc∞ (Rn )

such that 0 ≤ η ≤ 1, η = 1 for |ξ| ≤ 1/4, and η = 0 for |ξ| ≥ 1/2. We
42 3. FRACTIONAL SOBOLEV SPACES

define nonnegative functions ϕ, ϕj by ϕ(ξ)2 = η(ξ/2) − η(ξ), and

ϕj (ξ)2 = ϕ(2−j ξ)2 (j ≥ 1),
∞
X
2
ϕ0 (ξ) = 1 − ϕj (ξ)2 .
j=1

We collect some properties of the functions ϕj .

Theorem 2. (Littlewood-Paley partition of unity) The functions
ϕj are smooth in Rn , they satisfy 0 ≤ ϕj ≤ 1, and they are supported
in frequency annuli of radius ∼ 2j :
supp(ϕj ) ⊆ {2j−2 ≤ |ξ| ≤ 2j }, supp(ϕ0 ) ⊆ {|ξ| ≤ 1}.
Also, they form a partition of unity in the sense that
∞
X
1= ϕ2j .
j=0

Proof. Since η is radial and nonnegative, it is not hard to see

that ϕ(ξ) := (η(ξ/2) − η(ξ))1/2 is C ∞ and that 0 ≤ ϕ ≤ 1. Also, ϕ is
supported in {1/4 ≤ |ξ| ≤ 1}, which shows the support condition for
ϕj if j ≥ 1. We have
N
X N
X  −j−1
2
ξ) − η(2−j ξ) = η(2−N −1 ξ) − η(2−1 ξ).

ϕj (ξ) = η(2
j=1 j=1

The last function is supported in {1/2 ≤ |ξ| ≤ 2N } and is equal to 1

in {1 ≤ |ξ| ≤ 2N −1 }. This shows that ϕ0 is supported in the unit ball,
as required. 
Definition. If f ∈ S 0 (Rn ) and j ≥ 0, we define
Sj f := ϕj (D)f,
Sf := (Sj f )∞
j=0 .

The functions Sj f are called the dyadic (or Littlewood-Paley) pieces

of f . They are always C ∞ functions in Rn , as Fourier transforms of
compactly supported distributions (see Rudin [9, Chapter 7]). The
square function in this case is the map
∞
!1/2
X
x 7→ kSf (x)kl2 = |Sj f (x)|2 .
j=0
 3.3. LITTLEWOOD-PALEY THEORY 43

The Littlewood-Paley theorem states that the Lp (Rn ) norm of f is

comparable to the Lp (Rn ) norm of the square function.
Theorem 3. (Littlewood-Paley theorem in Rn ) Let 1 < p < ∞. If
f ∈ Lp (Rn ), then Sf ∈ Lp (l2 ) and
kf kLp ∼ kSf kLp (l2 ) .
If p = 2 then kf kL2 = kSf kL2 (l2 ) .
Proof. We prove the case p = 2, which is a direct consequence of
the Plancherel theorem: if f ∈ L2 (Rn ) then
∞ Z
X
−n ˆ 2
2
kf kL2 = (2π) kf kL2 = (2π) −n
ψj (ξ)2 |fˆ(ξ)|2 dξ
j=0
∞ Z
X Z
= |Sj f (x)|2 dx = kSf (x)k2l2 dx
j=0

= kSf k2L2 (l2 ) .

The case p 6= 2 is a nontrivial result. A standard proof, as in [2], is
based on the theory of (Hilbert-valued) singular integrals, where it is
shown that operators such as S map L1 (Rn ) into weak L1 (l2 ). Then a
version of the Marcinkiewicz interpolation theorem gives that
kSf kLp (l2 ) ≤ Ckf kLp
for 1 < p < 2, and the case where 2 < p < ∞ follows by duality.
For the converse inequality, we use the identity kf k2L2 = kSf k2L2 (l2 ) ,
which gives upon polarization that
Z Z X ∞
f ḡ dx = Sj f Sj g dx.
Rn Rn j=0

Then
Z
kf kLp = sup{ f ḡ dx ; kgkLp0 = 1}
Z X
= sup{ Sj f Sj g dx ; kgkLp0 = 1}

≤ sup{kSf kLp (l2 ) kSgkLp0 (l2 ) ; kgkLp0 = 1}

≤ C sup{kSf kLp (l2 ) kgkLp0 ; kgkLp0 = 1}
≤ CkSf kLp (l2 ) .

44 3. FRACTIONAL SOBOLEV SPACES

3.4. Besov and Triebel spaces

3.4.1. Definition of Besov and Triebel spaces. The discussion
of Littlewood-Paley theory in §3.3 motivates the definition of Besov and
Triebel spaces. As always, we begin by considering the L2 case.

Motivation. If f ∈ S (Rn ), let Sj f = ϕj (D)f be the Littlewood-

Paley pieces of f as in §3.3. Since hξi ∼ 2j on supp(ϕj ), the H s,2 norm
of f can be expressed as
kf k2H s,2 = kF −1 {hξis fˆ}k2L2 = (2π)−n khξis fˆk2L2
X∞ Z ∞
X
2s ˆ
= (2π)−n 2 2
ϕj (ξ) hξi |f (ξ)| dξ ∼ 22js kϕj (ξ)fˆk2L2
j=0 Rn j=0
∞
X
∼ 22js kϕj (D)f k2L2
j=0
!1/2 2
∞
X
∼ 22js |ϕj (D)f (x)|2 .
j=0
L2

This shows that

(1) kf kH s,2 ∼ k(ϕj (D)f )kl2s (L2 ) ∼ k(ϕj (D)f )kL2 (l2s ) ,
where lps are the weighted sequence spaces introduced in §2.4.

The Besov and Triebel spaces are obtained by replacing the expo-
nents 2 by p and q in the norms appearing in (1). For later purposes,
we will need to use more general Littlewood-Paley partitions of unity
in the definition of the spaces.

Definition. If N is a positive integer, let ΨN be the set of all

sequences (ψj )∞
j=0 of functions which satisfy
(a) ψj ∈ Cc∞ (Rn ) and ψj ≥ 0 for all j,
(b) ψj is supported in {2j−N ≤ |ξ| ≤ 2j+N } for j ≥ 1, and ψ0 is
supported in {|ξ| ≤ 2N },
(c) there exists c > 0 such that ∞
P
j=0 ψj (ξ) ≥ c, and
(d) for any multi-index α there exists Cα > 0 such that
|Dα ψj (ξ)| ≤ Cα hξi−|α| , j = 1, 2, . . . .
S∞
We let Ψ = N =1 ΨN .
 3.4. BESOV AND TRIEBEL SPACES 45

Example. It is easy to check that (ϕj )∞ j=0 is a sequence in Ψ. We

can generate more sequences in Ψ by choosing a nonnegative smooth √
−N N
function
√ ψ supported in {2 ≤ |ξ| ≤ 2 } with ψ > 0 for 1/ 2 ≤
−j
|ξ| ≤ 2, and by letting ψj (ξ) := ψ(2 ξ) for j ≥ 1 and choosing ψ0 in
a suitable way. The normalized functions
P0 −k
ψj (ξ) k=−∞ ψ(2 ξ)
χj (ξ) := ∞
P −k
(j ≥ 1), χ0 (ξ) := ∞ P −k
,
k=−∞ ψ(2 ξ) k=−∞ ψ(2 ξ)

yield a sequence (χj ) ∈ Ψ with ∞

P
j=0 χj ≡ 1. Further, if (ψj ) ∈ Ψ is
given, we may use the above idea to find a sequence (ψ̃j ) ∈ Ψ satisfying
ψ̃j = 1 on supp(ψj ) for all j ≥ 0.
We will use the constructions in the previous example several times
below. After these preparations, we are ready to give the definition of
Besov and Triebel spaces.
Definition. Suppose that (ψj )∞j=0 is a sequence in Ψ, let s be a
real number, and let 1 < p, q < ∞.
(a) The Besov space Bpq s
(Rn ) is the set of all f ∈ S 0 (Rn ) for which
(ψj (D)f ) ∈ lqs (Lp ). The norm is
∞
!1/q
X
kf kBpq
s := k(ψj (D)f )kls (Lp ) =
q
[2js kψj (D)f kLp ]q .
j=0

s
(b) The Triebel space (or Triebel-Lizorkin space) Fpq (Rn ) consists of
those f ∈ S 0 (Rn ) for which (ψj (D)f ) ∈ Lp (lqs ). The norm is
∞
!1/q
X  js q
kf kFpq
s := k(ψj (D)f )kLp (ls ) =
q
2 |ψj (D)f | .
j=0
Lp
s
Remarks. 1. More precisely, the Besov space Bpq consists of those
tempered distributions f for which each ψj (D)f is an Lp function, and
the sum [2js kψj (D)f kLp ]q converges.
P

2. For the Triebel spaces, we note that if f ∈ S 0 then any ψj (D)f is a

C ∞ function as the inverse Fourier transform of a compactly supported
distribution. It follows that the map
∞
!1/q
X  js q
x 7→ 2 |ψj (D)f (x)|
j=0
46 3. FRACTIONAL SOBOLEV SPACES

is nonnegative and measurable (as the qth root of the supremum of

s
continuous functions). Then f is in Fpq iff the last function is in Lp (Rn ).
3. One could also consider the case 0 < p, q ≤ 1, but we will restrict
our attention to 1 < p, q < ∞ for simplicity. Later we will discuss the
Besov spaces also when p = q = ∞.
We now prove that the spaces defined above do not depend on the
choice of the sequence (ψj ). This will allow considerable flexibility when
proving statements about these spaces. Below, we will usually give the
details for Besov spaces and leave the Triebel case as an exercise.
Theorem 1. (Independence of partition of unity) Let (ψj ), (ψ̃j ) be
two sequences in Ψ. Then for all admissible s, p, q,
k(ψj (D)f )klqs (Lp ) ∼ k(ψ̃j (D)f )klqs (Lp ) ,
k(ψj (D)f )kLp (lqs ) ∼ k(ψ̃j (D)f )kLp (lqs ) .
s s
Thus, any sequence (ψj ) ∈ Ψ in the definition of Bpq and Fpq will result
in the same spaces with equivalent norms.
Proof. 1. Let f ∈ S 0 be such that (ψj (D)f ) ∈ lqs (Lp ). We wish
to show that (ψ̃j (D)f ) ∈ lqs (Lp ), and
k(ψ̃j (D)f )klqs (Lp ) ≤ Ck(ψj (D)f )klqs (Lp )
with C independent of f .
2. Choose N so that (ψj ), (ψ̃j ) ∈ ΨN . We write
X ∞
ψ̃j (D)f = ψ̃j (D)χ(D)ψk (D)f
k=0
where !−1
∞
X
χ(ξ) := ψk (ξ) .
k=0
∞
Here χ is a bounded C function, and χ and ψ̃j satisfy the conditions
in Mihlin’s theorem with bounds independent of j by (a)–(d) above.
Since ψ̃j (ξ)ψk (ξ) = 0 if |j − k| ≥ 2N , we have (with ψk := 0 for k < 0)
X
2js kψ̃j (D)f kLp ≤ C 2js kψk (D)f kLp
j−2N ≤k≤j+2N
X
≤C 2ks kψk (D)f kLp .
j−2N ≤k≤j+2N
 3.4. BESOV AND TRIEBEL SPACES 47

It follows that
∞ h
X iq
k(ψ̃j (D)f )kqlqs (Lp ) = 2js kψ̃j (D)f kLp
j=0
∞
" #q
X X
≤C 2ks kψk (D)f kLp
j=0 j−2N ≤k≤j+2N
∞
X X q
2ks kψk (D)f kLp

≤C
j=0 j−2N ≤k≤j+2N
∞
X  js q
≤C 2 kψj (D)f kLp .
j=0

This proves the result for the Besov case.

3. The Triebel case is left as an exercise. It uses the vector-valued
Mihlin multiplier theorem, which will be stated (for q = 2) below. 

3.4.2. Properties of Besov and Triebel spaces. We begin by

s s
proving that Bpq and Fpq are Banach spaces. To this end, we first state
an abstract result which describes how the function spaces sit inside
lqs (Lp ) and Lp (lqs ).

Definition. Let A and B be normed spaces. We say that a

bounded linear operator R : A → B is a retraction is there is another
bounded linear operator S : B → A with
RSb = b for all b ∈ B.
Then S is called a coretraction for R.

Lemma 2. If A is Banach and B is normed, and if S : B → A is a

coretraction for R : A → B, then B is Banach and S is an isomorphism
from B onto a closed subspace of A.

Proof. If Sb = 0 then b = RSb = 0, showing that S is injective.

We claim that SR : A → A is a projection and the range of S is equal
to the range of SR. Clearly (SR)2 a = S(RS)Ra = SRa for a ∈ A and
SR(A) ⊆ S(B). Also, any Sb can be written as Sb = SR(Sb), so that
S(B) ⊆ SR(A). Then S(B) is closed as the range of a projection.
If (bj ) is a Cauchy sequence in B, then (Sbj ) is Cauchy in A and
there exists Sb ∈ S(B) with Sbj → Sb in A. Applying R, we see that
48 3. FRACTIONAL SOBOLEV SPACES

bj → b in B and that B is complete. The map S : B → S(B) is an

isomorphism by the open mapping theorem. 

s s s s
Theorem 3. (Bpq and Fpq as function spaces) Bpq and Fpq are Ba-
nach spaces, and S ⊆ Bpq ⊆ S and S ⊆ Fpq ⊆ S with continuous
s 0 s 0

inclusions. Further, S is dense in Bpq

s s
and Fpq .

s
Proof. 1. We first show that Bpq is a normed space. Let (ψj ) be
P∞ s
a sequence in Ψ with j=0 ψj = 1. If f, g ∈ Bpq then

kλf + µgkBpq
s = kλ(ψj (D)f ) + µ(ψj (D)g)kls (Lp )
q

≤ |λ|kf kBpq
s + |µ|kgkB s ,
pq

s
showing that Bpq is a vector space and k · kBpq
s obeys the triangle in-
equality. Also the other conditions for a norm are satisfied, for instance
s
if f is in Bpq and kf kBpq
s = 0, then ψj (D)f = 0 for all j which implies

ˆ P ˆ
that f = j ψj (ξ)f = 0 and f = 0.

s
2. To show that Bpq is complete, we define the maps

s
S : Bpq → lqs (Lp ), f 7→ (ψj (D)f ),
∞
X
R: lqs (Lp ) → s
Bpq , (gj ) 7→ ψ̃j (D)gj ,
j=0

where (ψ̃j ) is a sequence in Ψ with ψ̃j = 1 on supp(ψj ) for all j ≥ 0.

We leave as an exercise to show that the sum defining R converges
s
in Bpq , and that S and R are bounded linear operators (this uses the
argument in Theorem 1). We have that

X X
RSf = ψ̃j (D)ψj (D)f = ψj (D)f = f,

so S is a coretraction. Since lqs (Lp ) is Banach, Lemma 2 shows that

s
Bpq is a Banach space.
 3.4. BESOV AND TRIEBEL SPACES 49

3. Next we show that S ⊆ Bpq

s
. If f ∈ S then
kψj (D)f kLp ≤ Ckhxi2n ψj (D)f kL∞
≤ CkF −1 {hDξ i2n [ψj (ξ)fˆ(ξ)]}kL∞
≤ CkhDξ i2n [ψj (ξ)fˆ(ξ)]kL1
≤ Ckhξi2n hDξ i2n [ψj (ξ)fˆ(ξ)]kL∞
X
≤C khξi2n Dα ψj (ξ)Dβ fˆ(ξ)kL∞
|α|,|β|≤2n
X
≤ C2−j(s+1) khξi2n+s+1 Dβ fˆ(ξ)kL∞ .
|β|≤2n

Thus k(2js kψj (D)f kLp )klq ≤ C |β|≤2n khξi2n+s+1 Dβ fˆ(ξ)kL∞ , showing
P

that S is continuously embedded in Bpq s

. The embedding Bpq s
⊆ S 0 is
an exercise.
4. It remains to show that S is dense in Bpq
s s
. If f ∈ Bpq , we first note
that the functions
XN
fN := ψj (D)f
j=0
s
converge to f in Bpq as N → ∞. This follows from the argument in
the proof of Theorem 1. Thus it is enough to approximate each fN by
Schwartz functions.
Fix ε > 0, and let R and S be as in Step 2. If fN is given, there is
some N 0 > 0 such that ψj (D)fN = 0 for j > N 0 . Since each ψj (D)fN
is in Lp , we may find ϕj ∈ S such that
kψj (D)fN − ϕj kLp ≤ 2−j(s+1) ε, 0 ≤ j ≤ N 0.
Set ϕj := 0 for j > N 0 , and let ϕ := R(ϕj ) =
P
ψ̃j (D)ϕj . Since the
sum is finite, we have ϕ ∈ S , and
kfN − ϕkBpq
s = kRSfN − R(ϕj )kB s ≤ Ck(ψj (D)fN ) − (ϕj )kls (Lp )
pq q

1/q
N0
!
X  js q
= 2 kψj (D)fN − ϕj kLp ≤ Cq ε.
j=0

This completes the proof. 

The following result gives some simple relations between the Besov
and Triebel spaces.
50 3. FRACTIONAL SOBOLEV SPACES

s s
Theorem 4. (Monotonicity of Bpq and Fpq ) Let s be a real number
and 1 < p, q < ∞.
(a) The Besov spaces satisfy
s2 s1
Bpq ⊆ Bpq if s1 ≤ s2 ,
s s
Bpq 1
⊆ Bpq 2
if q1 ≤ q2 .
These identities are true for the F -spaces as well.
s s
(b) If p = q, one has Bpp = Fpp .
(c) One has
s s s
Bpq ⊆ Fpq ⊆ Bpp if 1 < q ≤ p < ∞,
s s s
Bpp ⊆ Fpq ⊆ Bpq if 1 < p ≤ q < ∞.

Proof. 1. The first two identities follow immediately from the

inclusions lqs2 ⊆ lqs1 and lq1 ⊆ lq2 :
∞
!1/q
X  js1 q
kf kBpq
s1 = 2 kψj (D)f kLp
j=0

∞
!1/q
X  js2 q
≤ 2 kψj (D)f kLp = kf kBpq
s2 ,

j=0

∞
!1/q1
X q1
2js kψj (D)f kLp

kf kBpq
s
1
=
j=0

∞
!1/q2
X  js q
≤ 2 kψj (D)f kLp 2 = kf kBpq
s ,
2
j=0

since if (xj ) ∈ lq1 then

X q2 −q1 X X q2 /q1
|xj |q2 ≤ (sup|xj |q1 ) q1
|xj |q1 ≤ |xj |q1 .

The proof for the F -spaces is similar.

2. If f ∈ S then
X Z X
kf kpBpp
s = 2jsp
kψj (D)f kpLp = 2jsp |ψj (D)f |p dx
p
X  p 1/p
= 2js |ψj (D)f | = kf kFpp
s .

Lp
 3.4. BESOV AND TRIEBEL SPACES 51

3. In (c), the second inclusion on the first line and first inclusion on
the second line follow from (a) and (b). To show the first inclusion, if
s
f ∈ Bpq and 1 < q ≤ p < ∞ we have
X  q 1/q X q 1/q
kf kFpq
s = 2js |ψj (D)f | = 2js |ψj (D)f |
Lp Lp/q
X 1/q
≤ 2jsq kψj (D)f kqLp = kf kBpq
s .

The proof of the last inclusion is analogous. 

Next we consider the relation of Besov and Triebel spaces to the

Bessel potential spaces. Note that kf kFp20 = k(ψj (D)f )kLp (l2 ) , so the

Littlewood-Paley theorem implies that Fp2 = Lp with equivalent norms.

0

To prove a corresponding result for s 6= 0, we need a generalization of

Mihlin’s theorem. See [2] for a proof.

Theorem 5. (Vector-valued Mihlin multiplier theorem) Suppose

m(ξ) := (mjk (ξ))∞
j,k=0 is a matrix of C
∞
complex valued functions on
n
R , which for any multi-index α satisfies
∞
!1/2
X
|Dα mjk (ξ)|2 ≤ Cα hξi−|α| .
j,k=0

Then the operator m(D), defined for sequences F = (fj )∞

j=0 of Schwartz
functions with fj = 0 for j large, by
∞
X
(m(D)F )j := mjk (D)fk , j = 0, 1, 2, . . . ,
k=0

has a bounded extension Lp (l2 ) → Lp (l2 ) for any 1 < p < ∞, and

km(D)F kLp (l2 ) ≤ CkF kLp (l2 ) , F ∈ Lp (l2 ),

where C only depends on p, n, and finitely many Cα .

The inequality k(ϕj (D)f )kLp (l2 ) ≤ Ckf kLp in the Littlewood-Paley
theorem is an immediate consequence of the last result, if one takes
mj0 (ξ) := ϕj (ξ) and mjk (ξ) := 0 for k ≥ 1, and applies m(D) to
F := (f, 0, 0, . . .). The following proof uses a similar idea.
52 3. FRACTIONAL SOBOLEV SPACES

Theorem 6. (Fractional Sobolev spaces as Triebel spaces) If s is a

real number and 1 < p < ∞, then
s
Fp2 = H s,p ,
s s
F22 = B22 = H s,2 .
Proof. 1. We show that kf kFp2 s ≤ Ckf kH s,p for any Schwartz
s
function f . Let f0 := hDi f , let (ψj ) be a sequence in Ψ, and define
mj0 (ξ) := 2js hξi−s ψj (ξ),
mjk (ξ) := 0 (k ≥ 1),
F := (f0 , 0, 0, . . .).
Since |Dα mj0 (ξ)| ≤ Cα hξi−|α| χsupp(ψj ) (ξ) by the properties of (ψj ), the
matrix m = (mjk ) satisfies the conditions in the vector-valued Mihlin
theorem. Consequently
js
kf kFp2
s = k(2 ψj (D)f )kLp (l2 ) = k(mj0 (D)f0 )kLp (l2 ) = km(D)F kLp (l2 )

≤ CkF kLp (l2 ) = Ckf0 kLp = Ckf kH s,p .

2. Let us prove the converse inequality, kf kH s,p ≤ Ckf kFp2 s for f ∈ S .

P
If (ψk ) is a sequence in Ψ with ψk = 1, we have
X
kf kH s,p = khDis f kLp = k ψk (D)hDis f kLp .
k
p
The last sum converges in L by the Littlewood-Paley theorem for
instance. We choose another sequence (ψ̃k ) in Ψ with ψ̃k = 1 on
supp(ψk ), and define
m0k (ξ) := 2−ks hξis ψk (ξ),
mjk (ξ) := 0 (j ≥ 1),
F := (2ks ψ̃k (D)f )∞
k=0 .

Again, m satisfies the conditions in Mihlin’s theorem. Since ψk (D) =

ψk (D)ψ̃k (D), we have
X
kf kH s,p ≤ k m0k (D)Fk kLp = km(D)F kLp (l2 )
k
≤ CkF kLp (l2 ) = Ckf kFp2
s .

s ∼ kf kH s,p for f ∈ S . The standard

3. We have proved that kf kFp2
s
density argument then shows that Fp2 = H s,p with equivalent norms.
 3.4. BESOV AND TRIEBEL SPACES 53

The second identity in the theorem is a consequence of the first one

s s
and the fact that Fpp = Bpp . 
Finally, we consider the interpolation of various function spaces.
The following abstract result will be used to pass from interpolation
results on Lp and lqs spaces to similar results for Besov spaces.

Lemma 7. Let (A0 , A1 ) and (B0 , B1 ) be two interpolation couples

of Banach spaces, and assume that R : A0 + A1 → B0 + B1 and S :
B0 + B1 → A0 + A1 are linear operators such that R is a retraction
Aj → Bj with coretraction S : Bj → Aj (j = 0, 1). If F is any
interpolation functor, then S is a coretraction F (B0 , B1 ) → F (A0 , A1 )
which is an isomorphism onto a closed subspace of F (A0 , A1 ).

Proof. Since R is bounded Aj → Bj , the definition of interpola-

tion functor implies that R is bounded F (A0 , A1 ) → F (B0 , B1 ). Simi-
larly, S is bounded F (B0 , B1 ) → F (A0 , A1 ). We also have RSb = b for
b ∈ B0 and b ∈ B1 , thus also for b ∈ F (B0 , B1 ). Then R is a retraction
on F (A0 , A1 ) with coretraction S, and the result follows from Lemma
2 above. 
The following interpolation results are not the sharpest possible (for
better results see Triebel [13]), but they give an idea of what can be
done.

Theorem 8. (Interpolation of function spaces) Let s0 , s1 be real

numbers, and let 1 < p0 , p1 , q0 , q1 < ∞. If 0 < θ < 1, let s, p, q be
defined by
1 1−θ θ 1 1−θ θ
s = (1 − θ)s0 + θs1 , = + , = + .
p p0 p1 q q0 q1

(a) The Besov spaces satisfy

s0 s1 s
(Bpq 0
, Bpq ) = Bpq
1 θ,q
if p0 = p1 = p,
(Bps00q0 , Bps11q1 )θ,q = Bpq
s
if p = q.

(b) The Triebel spaces satisfy

(Fps00q0 , Fps11q1 )θ,p = Bpp
s
if s0 6= s1 ,
(Fps0 q0 , Fps1 q1 )θ,p = Bpp
s
if s0 = s1 = s and p = q,
(Fps0 r , Fps1 r )θ,p = Fpr
s
if s0 = s1 = s and 1 < r < ∞.
54 3. FRACTIONAL SOBOLEV SPACES

(c) The fractional Sobolev spaces satisfy

(H s0 ,p , H s1 ,p )θ,r = Bpr
s
if s0 6= s1 , p0 = p1 = p, 1 < r < ∞,
(H s,p0 , H s,p1 )θ,p = H s,p if s0 = s1 = s.
Proof. We prove (a). Consider the coretraction S given in the
proof of Theorem 3,
s s
S : Bpjj qj → lqjj (Lpj ), f 7→ (ψk (D)f )∞
k=0 (j = 0, 1).
By Lemma 7, S is an isomorphism from (Bps00q0 , Bps11q1 )θ,q onto a closed
subspace of (lqs00 (Lp0 ), lqs11 (Lp1 ))θ,q . But by Theorems 2 and 3 in §2.4, we
have with equivalent norms
(lqs00 (Lp0 ), lqs11 (Lp1 ))θ,q = lqs ((Lp0 , Lp1 )θ,q ) = lqs (Lp )
if either p0 = p1 = p or p = q. In these cases, since S is an isomorphism
we have
kf k(Bps0 q ,Bps1 q )θ,q ∼ kSf klqs (Lp ) ∼ kf kBpq
s .
0 0 1 1

The second and third identities in (b) follow in a similar way from
a retraction argument. However, for the first identity one needs a more
general interpolation result for the lqs spaces than the one proved in
§2.4; see Triebel [13]. For the first identity in (c) we also refer to [13],
and the second identity in (c) follows from the last identity in (b). 

3.5. Hölder and Zygmund spaces

In this section, we will see how the Hölder spaces fit into the frame-
work of Besov and Triebel spaces. First, we extend the definition of
Besov spaces to the case p = q = ∞. We use a sequence (ψj )∞ j=0 in Ψ
P∞ −j
such that j=0 ψj = 1, ψj (ξ) = ψ(2 ξ) for j ≥ 1, and ψj (−ξ) = ψj (ξ)
for j ≥ 0.
s
Definition. If s is a real number, the space B∞∞ (Rn ) consists of
those f ∈ S 0 (Rn ) for which the sequence (2js kψj (D)f kL∞ ) is bounded.
The norm is
kf kB∞∞
s := sup 2js kψj (D)f kL∞ .
j≥0

s
We will show that if 0 < s < 1, the space B∞∞ is precisely the space
of Hölder continuous functions of exponent s. However, the equality
breaks down if s = 1. To obtain a characterization for integer s, we
will need to consider second order differences.
 3.5. HÖLDER AND ZYGMUND SPACES 55

Definition. The Zygmund space C∗s (Rn ) for s > 0 is defined as

follows: if s = k + γ where k is a nonnegative integer and 0 < γ < 1,
we define C∗s (Rn ) := C k,γ (Rn ) and
X |∂ α f (x + h) − ∂ α f (x)|
kf kC∗s := kf kC k,γ = kf kC k + sup .
x,h |h|γ
|α|=k

Further, if s = k where k is a positive integer, then C∗k (Rn ) is the set

of all f ∈ C k−1 (Rn ) for which the following norm is finite:
X |∂ α f (x + h) − 2∂ α f (x) + ∂ α f (x − h)|
kf kC∗k := kf kC k−1 + sup .
x,h |h|
|α|=k

Example. The space C∗1 consists of those continuous and bounded

functions f for which |f (x + h) − 2f (x) + f (x − h)| ≤ C|h| for all x, h.
Clearly one has the inclusions
C 1 ⊆ C 0,1 ⊆ C∗1 .
These inclusions are strict: an example of a function in C∗1 which is not
Lipschitz continuous is given by the lacunary Fourier series
∞
k
X
2−k ei2 x .
k=1

Below, it will be useful to write ψj (D)f as the convolution of f

against a function. If m ∈ Cc∞ (Rn ) and f ∈ L1 (Rn ) we have
Z
ˆ
m(D)f (x) = F {m(ξ)f (ξ)} = (2π)
−1 −n
eix·ξ m(ξ)fˆ(ξ) dξ
Rn
Z  Z 
−n i(x−y)·ξ
= (2π) e m(ξ) dξ f (y) dy
Rn Rn
Z
= m̌(x − y)f (y) dy.
Rn

Thus m(D)f = m̌ ∗ f , where we have written m̌ := F −1 m. If m ∈

OM and f ∈ S 0 , with correct interpretations (see Schwartz [10]) it
is possible to write m(D)f = m̌ ∗ f where m̌ is a rapidly decreasing
distribution.

Theorem 1. (Hölder spaces as Besov spaces) If s > 0 is not an

integer, we have
C∗s = B∞∞
s
.
56 3. FRACTIONAL SOBOLEV SPACES

Proof.R 1. Let us first assume f ∈ C∗γ = C 0,γ with 0 < γ < 1. If

j ≥ 1 then ψ̌j (y) dy = ψj (0) = 0 and
Z Z
|ψj (D)f (x)| = ψ̌j (y)f (x − y) dy = ψ̌j (y)[f (x − y) − f (x)] dy
Z
≤ kf kC 0,γ |y|γ |ψ̌j (y)| dy.

Here ψ̌j (y) = F −1 {ψ(2−j · )} = 2jn ψ̌(2j y), so

Z
−jγ
|ψj (D)f (x)| ≤ 2 kf kC 0,γ |y|γ |ψ̌(y)| dy ≤ C2−jγ kf kC 0,γ .

Since kψ0 (D)f kL∞ ≤ kψ̌0 kL1 kf kL∞ we get 2jγ kψj (D)f kL∞ ≤ Ckf kC 0,γ
for all j ≥ 0, so that C 0,γ ⊆ B∞∞
γ
.
2. Let now f ∈ C∗k+γ = C k,γ with k ∈ Z+ and 0 < γ < 1. We use the
Taylor expansion with remainder in integral form,
X ∂ α f (x) X Z 1 ∂ α f (x − ty)
α
f (x−y) = (−y) + k (1−t)k−1 (−y)α dt.
α! 0 α!
|α|<k |α|=k
R
Since (−y)α ψ̌j (y) dy = Dα ψj (0) = 0 for j ≥ 1, we have
Z
ψj (D)f (x) = ψ̌j (y)f (x − y) dy
X Z 1 α
k−1 ∂ f (x − ty)
Z
= ψ̌j (y) k (1 − t) (−y)α dt dy
0 α!
|α|=k
Z X Z 1 ∂ α f (x − ty) − ∂ α f (x)
= ψ̌j (y) k (1 − t)k−1 (−y)α dt dy.
0 α!
|α|=k

Taking absolute values, we obtain

Z 
|ψj (D)f (x)| ≤ C |y| |ψ̌j (y)| dy kf kC k,γ ≤ C2−j(k+γ) kf kC k,γ .
k+γ

Thus C k,γ ⊆ B∞∞

k+γ
.
s
3. Assume now that f ∈ B∞∞ where 0 < s < 1. We need to show that
f ∈ C and |f (x + h) − f (x)| ≤ C|h|s . Since
0
P
ψj = 1, we have
∞
X
kf kL∞ ≤ kψj (D)f kL∞ ≤ Cs kf kB∞∞
s .
j=0

More precisely, the sum ∞ 0

P
j=0 ψj (D)f converges in C , and since it also
converges in S 0 to f we must have f ∈ C 0 .
 3.5. HÖLDER AND ZYGMUND SPACES 57

Let now h ∈ Rn with |h| ≤ 1 (if |h| > 1 the claim is trivial). Choose
k so that 2−k−1 < |h| ≤ 2−k . Also, choose a sequence (χj )∞ j=0 with
χj = 1 on supp(ψj ), and χj (ξ) = χ(2 ξ) for j ≥ 1 where χ ∈ Cc∞ (Rn ).
−j

Let fj := ψj (D)f = χj (D)ψj (D)f . Then for j ≥ 1,

Z
fj (x + h) − fj (x) = [χ̌j (x + h − y) − χ̌j (x − y)]ψj (D)f (y) dy
Z
= [χ̌j (y + h) − χ̌j (y)]ψj (D)f (x − y) dy
Z
= [χ̌(y + 2j h) − χ̌(y)]ψj (D)f (x − 2−j y) dy.
R1
Using that χ̌(y + z) − χ̌(y) = 0 ∇χ̌(y + tz) · z dt, we estimate
C2j |h|, j ≤ k,
Z 
j
|χ̌(y + 2 h) − χ̌(y)| dy ≤
C, j > k.
It follows that
k
X ∞
X
|f (x + h) − f (x)| ≤ |fj (x + h) − fj (x)| + |fj (x + h) − fj (x)|
j=0 j=k+1
k
X ∞
X
j
≤C 2 |h|kψj (D)f kL∞ + C kψj (D)f kL∞
j=0 j=k+1
∞
k
!
X X
≤ C|h|s kf kB∞∞
s 2j(1−s) |h|1−s + |h|−s 2−js
j=0 j=k+1
s
≤ C|h| kf kB∞∞
s ,
since |h| ∼ 2−k .
s
4. Finally, assume f ∈ B∞∞ where s = k + γ and k ∈ Z+ , 0 < γ < 1.
k
We first show that f ∈ C . If (χj ) is as in Step 3, we have for j ≥ 1
χj (D)Dα f = χ̌j ∗ Dα f = Dα χ̌j ∗ f
= 2j|α| 2jn (Dα χ̌)(2j · ) ∗ f.
Therefore, if |α| ≤ k then
∞
X ∞
X
α α
kD f kL ≤ ∞ kψj (D)D f kL ≤
∞ kχj (D)Dα ψj (D)f kL∞
j=0 j=0
∞
X
≤C 2j|α| kψj (D)f kL∞ ≤ Ckf kB∞∞
s .
j=0
58 3. FRACTIONAL SOBOLEV SPACES

The fact that Dα f ∈ C 0,γ for |α| = k follows by repeating the argument
in Step 3 with fj := ψj (D)Dα f . 
In the next result, we consider the case of integer s.

Theorem 2. (Zygmund spaces as Besov spaces) If s > 0 is an

integer, then C∗s = B∞∞
s
.

Proof. 1. We shall only prove Rthat C∗1 = B∞∞ 1

. Assume that f is
1
a function in C∗ . If j ≥ 1 we have ψ̌j (y) dy = 0 and ψ̌j (−y) = ψ̌j (y)
(because ψj (−ξ) = ψj (ξ)), so that
Z
|ψj (D)f (x)| = ψ̌j (y)f (x − y) dy
Z
1
= ψ̌j (y)[f (x − y) − 2f (x) + f (x + y)] dy
2
Z
1
≤ kf kC∗1 |y||ψ̌j (y)| dy ≤ C2−j kf kC∗1 .
2
Also |ψ0 (D)f (x)| ≤ kψ̌0 kL1 kf kL∞ ≤ Ckf kC∗1 , so we have C∗1 ⊆ B∞∞
1
.
1 1−ε
2. Let now f ∈ B∞∞ . Then f ∈ B∞∞ ⊆ C 0 by Theorem 1. We need
to show that
|f (x + h) − 2f (x) + f (x − h)| ≤ C|h|, x, h ∈ Rn .
It is enough to show this for |h| ≤ 1. As in the proof of Theorem
1, choose k so that 2−k−1 < |h| ≤ 2−k , and choose (χj )∞ j=0 ∈ Ψ with
−j
χj = 1 on supp(ψj ), χj (ξ) = χ(2 ξ) for j ≥ 1. Letting fj := ψj (D)f =
χj (D)ψj (D)f , we have
fj (x + h) − 2fj (x) + fj (x − h)
Z
= [χ̌(y + 2j h) − 2χ̌(y) + χ̌(y − 2j h)]ψj (D)f (x − 2−j y) dy.

We use the estimate

C22j |h|2 , j ≤ k,
Z 
j j
|χ̌(y + 2 h) − 2χ̌(y) + χ̌(y − 2 h)| dy ≤
C, j > k.
The second part of this follows from the triangle inequality, and the
first part is a consequence of the Taylor expansion
X Z 1 ∂ α h(tz) α
h(z) = h(0) + ∇h(0) · z + 2 (1 − t) z dt
0 α!
|α|=2
 3.5. HÖLDER AND ZYGMUND SPACES 59

applied to h(z) := χ̌(y + z) − 2χ̌(y) + χ̌(y − z). We obtain

k
X
|f (x + h) − 2f (x) + f (x − h)| ≤ |fj (x + h) − 2fj (x) + fj (x − h)|
j=0
∞
X
+ |fj (x + h) − 2fj (x) + fj (x − h)|
j=k+1
k
X ∞
X
2j 2
≤C 2 |h| kψj (D)f kL∞ + C kψj (D)f kL∞
j=0 j=k+1
∞
k
!
X X
≤ C|h|kf kB∞∞
1 2j |h| + |h|−1 2−j
j=0 j=k+1

≤ C|h|kf kB∞∞
1 ,

since |h| ∼ 2−k . 

Definition. We will write C∗s := B∞∞

s
for s ≤ 0.

Finally, we remark that it is possible to characterize the Besov

s
spaces Bpq in terms of difference quotients also when p, q < ∞. Define

ωp (t, f ) := sup kf ( · + y) − f ( · )kLp ,

|y|<t

ωp2 (t, f ) := sup kf ( · + y) − 2f ( · ) + f ( · − y)kLp .

|y|<t

The following result may be found, also in a more general form, in

Bergh-Löfström [1].

Theorem 3. (Finite difference characterization of Besov spaces)

Let s > 0 and 1 < p, q < ∞. If s = k + γ where k is a nonnegative
integer and 0 < γ < 1, then
n Z ∞ 1/q
q dt
X
−γ k
kf kBpq
s ∼ kf kLp + [t ωp (t, Dj f )] .
j=1 0 t

If s = k + 1 where k is a nonnegative integer, then

n Z ∞ 1/q
q dt
X
−1 2 k
kf kBpq
s ∼ kf kLp + [t ωp (t, Dj f )] .
j=1 0 t
60 3. FRACTIONAL SOBOLEV SPACES

3.6. Embedding theorems

We will finish the chapter with a number of embedding theorems
for the different spaces. The following Lp estimates for convolutions
will be useful for this purpose. Part (c) is called Young’s inequality.

Theorem 1. (Lp estimates for convolution) Let 1 ≤ p ≤ ∞.

(a) If f ∈ L1 and g ∈ Lp , then f ∗ g ∈ Lp and
kf ∗ gkLp ≤ kf kL1 kgkLp .
0
(b) If f ∈ Lp and g ∈ Lp , then f ∗ g ∈ L∞ and
kf ∗ gkL∞ ≤ kf kLp kgkLp0 .
(c) If f ∈ Lp and g ∈ Lq , where 1 ≤ q ≤ p0 , then f ∗ g ∈ Lr and
kf ∗ gkLr ≤ kf kLp kgkLq
1 1
if r is defined by r
= p
− q10 .

Proof. (a) It is proved in [7] (or Rudin [8]) that the convolution
of two L1 functions is an L1 function, and kf ∗ gkL1 ≤ kf kL1 kgkL1 . If
f ∈ L1 and g ∈ L∞ , we have the trivial bound kf ∗gkL∞ ≤ kf kL1 kgkL∞ .
Thus, if f ∈ L1 we have a map T : L1 + L∞ → L1 + L∞ , T g = f ∗ g.
Since
T : L1 → L1 , kT k ≤ kf kL1 ,
T : L∞ → L∞ , kT k ≤ kf kL1 ,
interpolation gives that T maps Lp to Lp with kT gkLp ≤ kf kL1 kgkLp .
(b) This is Hölder’s inequality:
Z
|f ∗ g(x)| = f (x − y)g(y) dy ≤ kf (x − · )kLp kgkLp0 = kf kLp kgkLp0 .

(c) If f ∈ Lp , by (a) and (b) the map T : g 7→ f ∗ g satisfies

T : L1 → Lp , kT k ≤ kf kLp ,
p0
T : L → L∞ , kT k ≤ kf kLp .

Let 1 < q < p0 , and let 0 < θ < 1 be such that 1q = 1−θ1
+ pθ0 . Then
θ = qp0 and 1−θ
p
+∞θ
= p1 − q10 = 1r . Thus interpolation shows that
T : Lq → Lr with kT gkLr ≤ kf kLp kgkLq , as required. 
 3.6. EMBEDDING THEOREMS 61

Theorem 2. (Embeddings for function spaces) Assume that s ≥ s1

and 1 < p, p1 < ∞, and
n n
s − = s1 − .
p p1
Then for any 1 < r < ∞ one has
s
Bpr ⊆ Bps11r ,
s
Fpr ⊆ Fps11r ,
H s,p ⊆ H s1 ,p1 . .
Further, if t is any real number and 1 < p, r < ∞, then
t t−n/p
Bpr ⊆ C∗ ,
t t−n/p
Fpr ⊆ C∗ ,
t−n/p
H t,p ⊆ C∗ .

Proof. 1. We prove the first embedding for Besov spaces. If s = s1

then p = p1 and the claim is trivial. We may assume s > s1 , and then
p < p1 < ∞. Let (ψj )∞ ∞
j=0 and (χj )j=0 be two sequences in Ψ with
χj = 1 on supp(ψj ), and χj (ξ) = χ(2−j ξ) for j ≥ 1. Then
kψj (D)f kLp1 = kχj (D)ψj (D)f kLp1 = kχ̌j ∗ ψj (D)f kLp1
≤ kχ̌j kLq kψj (D)f kLp
by Young’s inequality, if q is chosen so that p11 = p1 − q10 . If j ≥ 1 we
have
0
kχ̌j kLq = k2jn χ̌(2j · )kLq = C2jn/q = C2j(s−s1 )
since
1 1 1 s − s1
0
= − = .
q p p1 n
If j = 0 then kψ0 (D)f kLp1 ≤ kχ̌0 kLq kψ0 (D)f kLp = Ckψ0 (D)f kLp .
Consequently, we obtain
kf kBps1 r = k(2js1 kψj (D)f kLp1 )klr ≤ Ck(2js kψj (D)f kLp )klr = Ckf kBpr
s .
1

This holds for Schwartz functions f , and the first embedding follows
since Schwartz functions are dense.
2. For the second embedding for Besov spaces, we note that
2j(t−n/p) kψj (D)f kL∞ ≤ 2j(t−n/p) kχ̌j kLp0 kψj (D)f kLp ≤ C2jt kψj (D)f kLp ,
62 3. FRACTIONAL SOBOLEV SPACES

since kχ̌j kLp0 = C2jn/p . Consequently

kf kC t−n/p ∼ sup 2j(t−n/p) kψj (D)f kL∞ ≤ Ckf kBpr
t
∗
j≥0

for any r.
3. The first embedding for Triebel spaces uses the Hardy-Littlewood-
Sobolev inequality, and we refer to Triebel [13] for the proof. The
second embedding however follows directly from the corresponding em-
bedding for Besov spaces, upon noting that by Theorem 4 in §3.4, for
any r there is some r1 such that
t t
Fpr ⊆ Bpr1
⊆ C∗t−n/p .

4. The embedding results for fractional Sobolev spaces follow from the
corresponding results for Triebel spaces, using the fact that H t,p = Fp2
t

by Theorem 6 in §3.4. 
s s
Remark. Considering the spaces Bpq , Fpq , and H s,p , the number
n
s − p is called the differential dimension. For the Zygmund space
s
C∗s = B∞∞ , the differential dimension is s. The embeddings considered
here can be memorized by noting that the differential dimension is
preserved while the smoothness index becomes smaller.
 Bibliography

[1] J. Bergh and J. Löfström, Interpolation spaces, Springer, 1976.

[2] J. Duoandikoetxea, Fourier analysis, AMS, 2001.
[3] R. Edwards and G. Gaudry, Littlewood-Paley and multiplier theory, Springer-
Verlag, Berlin, 1977.
[4] L. C. Evans, Partial differential equations, AMS, 1998.
[5] D. Jerison and C. E. Kenig, The inhomogeneous Dirichlet problem in Lipschitz
domains, J. Funct. Anal. 130 (1995), p. 161–219.
[6] L. Grafakos, Classical and modern Fourier analysis, Pearson Education, 2004.
[7] I. Holopainen, Reaalianalyysi 1, lecture notes (in Finnish), available at the
webpage http://www.helsinki.fi/∼iholopai/ReAn02.pdf.
[8] W. Rudin, Real and complex analysis, 3rd edition, McGraw-Hill, 1986.
[9] W. Rudin, Functional analysis, 2nd edition, McGraw-Hill, 1991.
[10] L. Schwartz, Théorie des distributions, Hermann, 1950.
[11] E. Stein, Singular integrals and differentiability properties of functions, Prince-
ton University Press, 1970.
[12] E. Stein, Harmonic analysis, Princeton University Press, 1993.
[13] H. Triebel, Interpolation theory, function spaces, differential operators, North-
Holland, Amsterdam, 1978.

63