Fine properties of functions: an introduction

Petru Mironescu

To cite this version:

Petru Mironescu. Fine properties of functions: an introduction. Master. Fine properties of functions:
an introduction, Scoala Normala Superioara din Bucuresti, 2005, pp.80. �cel-00747696�

HAL Id: cel-00747696

https://cel.hal.science/cel-00747696
Submitted on 1 Nov 2012

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Fine properties of functions: an introduction

Petru Mironescu

April 30, 2005

Introduction

These lecture notes, intended as support to an intensive course at Şcoala Normalǎ Superioarǎ din
Bucureşti, cover classical properties of function spaces such as: a.e. differentiability of Lipschitz
functions (Rademacher’s theorem), maximal functions (the Hardy-Littlewood-Wiener theorem),
functions of bounded variation, area and coarea formulae, Hausdorff measure and capacity, isoperi-
metric inequalities, Hardy and bounded mean oscillations spaces (and their duality), trace theory,
precise representatives.
Several textbooks cover part of these topics:
Herbert Federer, Geometric measure theory, Springer, 1969
Vladimir Maz’ja, Sobolev spaces, Springer, 1980
Leon Simon, Lectures on Geometric Measure Theory, Proceedings of the Centre for Mathemat-
ical Analysis, Australian National University, 1983
William P. Ziemer, Weakly differentiable functions, Springer, 1989
Lawrence C. Evans and Ronald F. Gariepy, Measure Theory and Fine Properties of Functions,
Studies in Advanced Mathematics, CRC Press, 1992
Elias M. Stein, Harmonic Analysis: real variable methods, orthogonality, and oscillatory inte-
grals, Princeton University press, 1993
However, there is no single source covering the basic facts the working analyst needs. This is
the main purpose of the notes. These notes are also an invitation to reading the above wonderful
books.
The background required is a good knowledge of standard measure theory: Radon-Nikodym
and Hahn decomposition theorems, Riesz representation theorem. Also, the standard theory of
distributions and basics about Sobolev spaces are presumes known. All other standard tools (Vitali
and Whitney covering theorems, for example) are proved in the text.
Proofs are to be considered cum grano salis: there are a number of typos left. The reader is
acknowledged in advance for his patience in this respect.

Lyon
April 2005

1
Contents

I Basic Tools 4

0 The distribution function 5

0.1 Lorentz spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1 Elementary interpolation 8

2 Hardy’s inequality 10

3 Coverings 12
3.1 The Vitali covering lemma (simplified version) . . . . . . . . . . . . . . . . . . . . 12
3.2 Whitney’s covering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 The maximal function 15

4.1 Maximal inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2 Lebesgue’s differentiation theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Pointwise inequalities for convolutions . . . . . . . . . . . . . . . . . . . . . . . . . 19

5 The Calderón-Zygmund decomposition 20

II Hardy and bounded mean oscillations spaces 22

6 Substitutes of L1 23
6.1 The space L log L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.2 The Hardy space H1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.3 More maximal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
6.4 Transition from one admissible function to another . . . . . . . . . . . . . . . . . 28
6.5 Proof of Theorem 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2
CONTENTS 3

7 Atomic decomposition 34
7.1 Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
7.2 Atomic decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

8 The substitute of L∞ : BM O 42
8.1 Definition of BM O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
8.2 H1 and BM O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.3 BM O functions are almost bounded . . . . . . . . . . . . . . . . . . . . . . . . . 46

9 Lp regularity for the Laplace operator 48

9.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
9.2 Proof of Theorem 17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
9.3 An equation involving the jacobian . . . . . . . . . . . . . . . . . . . . . . . . . . 55

III Functions in Sobolev spaces 58

10 Improved Sobolev embeddings 59

10.1 An equivalent norm in Lorentz spaces . . . . . . . . . . . . . . . . . . . . . . . . . 59
10.2 Properties of f ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
10.3 Rearrangement and convolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
10.4 Improved Sobolev embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
10.5 The limiting case p = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
10.6 The limiting case p = N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

11 Traces 73
11.1 Definition of the trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
11.2 Trace of W 1,p , 1 < p < ∞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
11.3 Trace of W 1,1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
 Part I

Basic Tools

4
Chapter 0

The distribution function

Throughout this course, we consider on RN the usual Lebesgue measure dx. The measure of a set
A ⊂ RN will be simply denoted by |A|.
Let f : RN → C be a measurable function. We consider the distribution function of f ,
F (t) = |{ x ∈ RN ; |f (x)| > t }|. (1)
Clearly, F : [0, ∞) → [0, ∞] is decreasing and thus measurable. F is related to various norms of
f via
Proposition 1. For 1 ≤ p < ∞ we have
Z∞
p
a) kf kLp = p tp−1 F (t)dt;
0
kf kpLp
b) (Chebyshev’s inequality) F (t) ≤ .
tp
Proof. a) We have

Z Z |fZ(x)| Z∞ Z Z∞
kf kpLp = |f (x)|p dx = ptp−1 dtdx = p tp−1 dxdt = p tp−1 F (t)dt. (2)
0 0 {x; |f (x)| > t} 0

b) Chebyshev’s inequality follows from

Z Z
p
kf kLp ≥ |f (x)|p dx ≥ tp dx = tp F (t). (3)
{x; |f (x)| > t} {x; |f (x)| > t}

5
CHAPTER 0. THE DISTRIBUTION FUNCTION 6

By copying the proof of a) above, we obtain the following

Proposition 2. Let Φ : [0, ∞] → [0, ∞], Φ ∈ C 1 , be a non decreasing function s. t. Φ(0) = 0.
Then
Z Z∞
Φ(|f (x)|)dx = Φ0 (t)F (t)dt. (4)
RN 0

0.1 Lorentz spaces

One may read the property a) in Proposition 1 as kf kpLp = pktF 1/p (t)kp p . This
L ((0, ∞); dt/t)
suggests a more general definition: a measurable function f belongs to the Lorentz space Lp,q
(1 ≤ p < ∞, 1 ≤ q ≤ ∞) if
kf kLp,q = ktF 1/p (t)kLq ((0, ∞); dt/t) < ∞. (5)

Despite this notation, k · kLp,q is not a norm (but almost: it is a quasi-norm). When q = p,
the corresponding Lorentz space coincides with Lp . When q = ∞, the corresponding space Lp,∞
is called the weak Lp , also denoted by Lpw (the Marcinkiewicz space). Clearly, a function f
belongs to Lpw if and only if its distribution function F satisfies a Chebyshev type inequality:
C
F (t) ≤ p for each t > 0.
t
It is well known that there is no inclusion relation for the Lp spaces. However, for fixed p, the
Lorentz spaces are monotonic in q:
Proposition 3. Let 1 ≤ q < r ≤ ∞. Then Lp,q ⊂ Lp,r .
Proof: Assume first that r = ∞. If f ∈ Lp,q , then
Z∞ Zs Zs
1
kf kqLp,q = tq−1 F q/p (t)dt ≥ tq−1 F q/p (t)dt ≥ tq−1 F q/p (s)dt = sq F q/p (s). (6)
q
0 0 0

C
Thus F 1/p (s) ≤ , i.e. f ∈ Lp,∞ .
s
Let now r < ∞ and let f ∈ Lp,q . Then, using Hölder’s inequality and the case r = ∞, we find
q/r 1−q/r
ktF 1/p (t)kLr ((0, ∞); dt/t) ≤ ktF 1/p (t)k q ktF 1/p (t)k ∞ < ∞. (7)
L ((0, ∞); dt/t) L ((0, ∞); dt/t)

Incidentally, we proved the stronger statement

kf kLp,r ≤ Ckf kLp,q , 1 ≤ p < ∞, 1 ≤ q ≤ r ≤ ∞. (8)
CHAPTER 0. THE DISTRIBUTION FUNCTION 7

We complete the scale of Lorentz spaces by setting L∞,q = L∞ for all q. The above inequality,
combined with the fact that k · kp,q is a quasi-norm yields immediately the following
Corollary 1. The inclusion Lp,q ⊂ Lp,r is continuous, 1 ≤ q < r ≤ ∞.
Remark 1. One should understand the Lorentz spaces as ”microscopic” versions of the Lp spaces.
We mean that the properties of Lp,q are very close to those of Lp . Here is an example: if Ω is
a bounded set in RN , one may define in an obvious way the spaces Lp,q (Ω). It is easy to prove
that, if p1 < p2 , then Lp2 ,q2 (Ω) ⊂ Lp1 ,q1 (Ω) for all the possible values of q1 , q2 . This is exactly the
inclusion relation we have for the standard Lp spaces.
Chapter 1

Elementary interpolation

Theorem 1. (Marcinkiewicz’ interpolation theorem; simplified version) Let 1 < q < ∞

and let T : L1 ∩ Lq (RN ) → D0 be linear and s. t.
kT f kL1w ≤ C1 kf kL1 , ∀f (1.1)
and
kT f kLqw ≤ Cq kf kLq , ∀ f. (1.2)
(In other words, T extends by density as a continuous operator from L1 into L1w and from Lq into
Lqw .) Then T is a continuous operator from Lp into Lp , for each 1 < p < q, i. e.
kT f kLp ≤ Ckf kLp , ∀ f ∈ L1 ∩ Lq . (1.3)
Before proceeding to the proof of the theorem, let us note that Lq embeds into Lqw , and thus
we have the following consequence, which is the form we usually make use of the above theorem
Corollary 2. Let 1 < q < ∞ and let T : L1 ∩ Lq (RN ) → D0 be linear and s. t.
kT f kL1w ≤ C1 kf kL1 , ∀f (1.4)
and
kT f kLq ≤ Cq kf kLq , ∀ f. (1.5)
p p
Then T extends as a continuous operator from L into L , 1 < p < q.
Proof. Let t > 0 and let f ∈ L1 ∩Lq . We are going to estimate the distribution ( function of T f . For
f (x), if |f (x)| > t
this purpose, we cut f at height t, i. e. we write f = f1 + f2 , where f1 (x) =
0, otherwise
and f2 = f − f1 . Since T f = T f1 + T f2 , we have |T f | > t =⇒ |T f1 | > t/2 or |T f2 | > t/2, and
thus
2C1 2q Cqq
|{|T f | > t}| ≤ |{|T f1 | > t/2}| + |{|T f2 | > t/2}| ≤ kf1 kL1 + q kf2 kqLq . (1.6)
t t
8
CHAPTER 1. ELEMENTARY INTERPOLATION 9

Therefore,
Z Z Z
kT f kpLp =p t p−1
|{|T f | > t}| ≤ 2pC1 t p−2
kf1 kL1 + 2 q
pCqq tp−q−1 kf2 kqLq . (1.7)
(
F (α), if α ≥ t
Next, if F is the distribution function of f , then the distribution function of f1 is
F (t), if α < t,
(
0, if α ≥ t
and the one of f2 is . Thus
F (α) − F (t), if α < t

Z∞ Zt
kf1 kL1 = tF (t) + F (α)dα and kf2 kqLq =q αq−1 F (α)dα − tq F (t). (1.8)
t 0

By combining (1.6) and (1.8) and applying Fubini’s theorem (to interchange the order of integration
over α and t), we find that
2q Cqq
 
p 2C1
kT f kLp ≤ p + kf kqLq . (1.9)
p−1 q−p

Remark 2. We see that the estimate we obtain for the norm of T from Lp into Lp blows up as
p → 1 or p → q. This is not a weakness of the proof. If this norm remains bounded as, say,
p → 1, then T must continuous from L1 into L1 , which may not be the case.
There is a way to improve the estimate (1.9): instead of cutting f at height t, we cut it at
height at, where a > 0 is fixed. The above computations yield this time :
2C1 1−p 2q Cqq q−p
 
p p
kT f kLp ≤ pkf kLp a + a . (1.10)
p−1 q−p

Optimizing the above r. h. s. over a > 0 (it is minimal when a = 1/2(C1 /Cq )1/(q−1) ), we find the
following
Theorem 2. With the notations and under the hypotheses of the preceding theorem, let θ ∈ (0, 1)
1 θ 1−θ
be the (unique) number s. t. = + . Then the norm of T from Lp into Lp satisfies
p 1 q
kT kLp →Lp ≤ cp,q kT kθL1 →L1w kT kL1−θ
p →Lp .
w
(1.11)
This conclusion is reminiscent from the one of the Riesz-Thorin convexity theorem.
Chapter 2

Hardy’s inequality

We present two (equivalent) forms of Hardy’s inequality. The first one generalizes the usual (and
Z∞ 2 Z∞
F (x)
historically first) Hardy’s inequality dx ≤ 4 (F 0 (x))2 dx, F ∈ C0∞ ((0, ∞)). The second
x2
0 0
one will be needed later in the study of the Lorentz spaces.
Theorem 3. (Hardy) Let 1 ≤ p < ∞ and r > 0 and let f : (0, ∞) → R.
Z∞
If |f (x)|p xp−r−1 dx < ∞, then f ∈ L1loc ([0, ∞)).
0
Zx
With F (x) = f (t)dt, we have
0

Z∞  p Z∞
p −r−1 p
|F (x)| x dx ≤ |f (x)|p xp−r−1 dx. (2.1)
r
0 0

Proof. In view of the conclusions, we may assume f ≥ 0. In this case, it suffices to prove (2.1).
 Zx p
We want to apply Jensen’s inequality in order to estimate the integral f (t)dt . We consider,
0
r
on (0, x), the normalized measure µ = x−r/p tr/p−1 dt. Then (with u 7→ up playing the role of the
p
convex function)
 Zx p  p  Zx p  p Zx
p p
f (t)dt = xr f (t)t1−r/p dµ ≤ xr [f (t)t1−r/p ]p dµ, (2.2)
r r
0 0 0

10
CHAPTER 2. HARDY’S INEQUALITY 11

which yields
 Zx p  p−1 Zx
p r(1−1/p)
f (t)dt ≤ x f p (t)tp−r−1+r/p dt. (2.3)
r
0 0

Thus
Z∞  p−1 Z∞ Zx  p Z∞
p −r−1 p −1−r/p p p−r−1+r/p p
F (x)x dx ≤ x f (t)t dtdx = f p (t)tp−r−1 dt, (2.4)
r r
0 0 0 0

by Fubini’s theorem.
Corollary 3. (Hardy) With 1 ≤ p < ∞ and r > 0, we have
Z∞ Z∞ p  p Z∞
p
f (t)dt xr−1 dx ≤ |f (x)|p xp+r−1 dx. (2.5)
r
0 x 0

Proof. We may assume f ≥ 0. We apply the preceding theorem to the map g given by g(t) =
t−2 f (t−1 ) and find that
Z∞  Zx p  p Z∞  p Z∞
−2 −r−1 p p −1 −p−r−1 p
t f (t)dt x dx ≤ f (x )x dx = f p (y)y p+r−1 dx. (2.6)
r r
0 0 0 0

Zx
We next perform, in the integral t−2 f (t)dt, the substitution t = s−1 , next we substitute, in the
0
first integral in (2.6), y = x−1 , and obtain the desired result.
Chapter 3

Coverings

3.1 The Vitali covering lemma (simplified version)

Let F be a finite family of balls in RN .
Lemma 1. (Vitali’s lemma). F contains a subfamily F 0 of disjoint balls such that
X [
|B| ≥ C| B|.
B∈F 0 B∈F

Here, C depends only on the space dimension N , not on the family F.

Proof. Let B1 be the largest ball in F. Let B2 be the largest ball in F that does not intersect B1 , B3
the largest ball in F that does not intersect neither B1 nor B2 , and so on. Let F 0 = { B1 , B2 , ... }.
Note that, for each B ∈ F, there is some j s. t. B ∩Bj 6= ∅. For each ball B in F 0 , let
[ B̃ be the
[ball
having the same center as B and the radius thrice the one of B. We claim that B⊂ B̃.
B∈F B∈F 0
Indeed, let B ∈ F and let j be the smallest integer such that B ∩ Bj 6= ∅. Since B ∩ Bj−1 = ∅,
the radius of B is at most the one of Bj , for otherwise we would have picked B instead of Bj at
step j in the construction of F 0 . Since B ∩ Bj 6= ∅, we find that B ⊂ B̃j .
It follows that [ [ X
| B| ≤ | B̃| ≤ 3N |B|, (3.1)
B∈F B∈F 0 B∈F 0

which is the desired result with C = 3−N .

3.2 Whitney’s covering

Throughout this section, the norm we consider on RN is the k · k∞ one.

12
CHAPTER 3. COVERINGS 13

Let F ⊂ RN be a not empty closed set and let O = RN \ F . If C is a (closed) cube, let l(C)
be its size, i. e. the length of its edges.
Lemma[ 2. (Whitney’s covering lemma) There is a family F of closed cubes s. t.
a) O = C;
C∈F
b) distinct cubes in F have disjoint interiors;
c) c−1 l(C) ≤ dist(C, F ) ≤ c l(C) for each C ∈ F.
Here, c depends only on N .
Proof. We may assume that 0 ∈ F . For j ∈ Z, let Fj be the grid of cubes of size 2j , with sides
parallel to the coordinate axes, s. t. 0 be one of the vertices. Note that each cube C ∈ Fj is
contained in exactly one predecessor C 0 ∈ Fj+1 . In addition, each cube has an ancestor containing
0. Thus, the non increasing sequence dist(C, F ), dist(C 0 , F ), dist(C 00
[ , F ), . . ., becomes 0 starting
with a certain range. We throw away all the cubes contained in Fj s. t. dist(C, F ) ≤ l(C) and
j
call F the family of all kept cubes C s. t. their predecessors C 0 were thrown away.
Note that, by definition, if C ∈ F, then dist(C, F ) > l(C), while there are some y ∈ C 0 and z ∈ F
s. t. ky − zk∞ ≤ 2l(C). Let x ∈ C; then dist(C, F ) ≤ kx − zk∞ ≤ 3l(C), so that c) holds with
c = 3.
Let x ∈ O. If j is sufficiently close to −∞, we have dist(C, F ) > l(C) whenever C ∈ Fj and
x ∈ C. Pick any such j and C and set k = sup{l ∈ N ; dist(C (l) , F ) > l(C (l) )}. Then k is finite
and it is clear from the definition that x ∈ C (k) ∈ F. Thus a) holds.
[ ◦ ◦
Finally, if C, D ∈ Fj are distinct cubes s. t. C ∩ D6= ∅, then one of these cubes is contained
j
in the other one. Assume, e. g., that C ⊂ D. Then C 0 ⊂ D. Therefore, we can not have at the
same time C ∈ F and D ∈ F, for otherwise l(C 0 ) ≥ dist(C 0 , F ) ≥ dist(D, F ) > l(D) ≥ l(C 0 ).
For a cube C, let C∗ be the cube concentric with C and of size 3/2l(C).
Proposition 4. Let F , O and F be as in the proof of the above lemma. Then:
[ ◦
a) O = C ∗;
C∈F
b) we have d−1 l(C∗ ) ≤ dist(C∗ , F ) ≤ d l(C∗ ) for each C ∈ F;
c) if x ∈ C∗ , then e−1 dist(x, F ) ≤ dist(C∗ , F ) ≤ e dist(x, F );
d) each point x ∈ O belongs to at most M cubes C∗ .
Here, d, e and M depend only on N .
Proof. On the one hand, we have dist(C∗ , F ) ≤dist(C, F ) ≤ 3l(C) = 2l(C∗ ). On the other hand,
if x ∈ C∗ , then there is some y ∈ C s. t. kx − yk∞ ≤ 1/2l(C). In addition, dist(y, F ) > l(C).
CHAPTER 3. COVERINGS 14

Thus dist(x, F ) ≥ 1/2l(C) = 1/3l(C∗ ). Thus b) holds with d = 3. Property c) is a straightforward

consequence of b).
Clearly, C∗ ⊂ O, by b), which implies a) with the help of a) of Whitney’s lemma.
If x ∈ C∗ , then (by b) and c)) 2/3(de)−1 dist(x, F ) ≤ l(C), and therefore C∗ ⊂ B(x, r), with
r = de dist(x, F ). Thus, if k is the number of cubes C∗ containing x, we have
[ [ X
(r/2)N = |B(x, r)| ≥ | C∗ | ≥ | C| = |C| ≥ k(1/3(de)−1 dist(x, F ))N ,
C∗ ∩F 6=∅ C∗ ∩F 6=∅ C∗ ∩F 6=∅

whence conclusion d).

X
Proposition 5. With the above notations, there is, in O, a partition of the unit 1 = ϕC s. t.:
C∈F
a) for each C, supp ϕC ⊂ C∗ ;
b) |∂ α ϕC (x)| ≤ Cα |C|−|α|/N ≤ Cα0 dist(x, F )−|α| when x ∈ supp ϕC .
Here, the constants Cα do not depend on O, x and C.
Proof. Fix a function ζ ∈ C ∞ (RN ; [0, 1]) s. t. ζ = 1 in B(0, 1/2) and supp ζ ⊂ B(0, 3/4). If
C ∈ F is of size 2l and center x, set ζC = ζ((·
X− x)/l). Note that supp ζC ⊂ C∗ and that ζC = 1
α −|α|
in C. Moreover, |∂ ϕC | ≤ Cα l . Set ϕ = ζC , which satisfies 1 ≤ ϕ ≤ M , by a) in Whitney’s
lemma and d) in the above proposition. Finally, set ϕC = ζC /ϕ. Properties a) and b) follow
immediately by combining the conclusions of Whitney’s lemma and of the above proposition.
Chapter 4

The maximal function

If f is locally integrable, we define the (uncentered) maximal function of f ,

Z
1
Mf (x) = sup{ |f (y)|dy ; B ball containing x }. (4.1)
|B|
B

In this definition, one may consider cubes instead of balls. This will affect the value of Mf , but
not its size. E.g., if we consider instead
Z
0 1
M f (x) = sup{ |f (y)|dy ; Q cube containing x }, (4.2)
|Q|
Q

then we have C −1 M0 f ≤ Mf ≤ CM0 f , where C is the ratio of the volumes of the unit cube
and of the unit ball. Thus the integrability properties of Mf remain unchanged if we change the
definition. Similarly, one may consider balls centered at x; this definition yields the centered
maximal function.
A basic property of Mf is that it is lower semi continuous, i.e. the level sets { x ; Mf (x) > t }
are open.

4.1 Maximal inequalities

When f ∈ L∞ , we clearly have Mf ∈ L∞ . However, it is not obvious whether, for 1 ≤ p < ∞ and
f ∈ Lp , the maximal function has some integrability properties or even whether it is finite a.e.
Theorem 4. (Hardy-Littlewood-Wiener) Let 1 ≤ p ≤ ∞ and f ∈ Lp . Then:
a) Mf is finite a.e.;
b) if 1 < p ≤ ∞, then Mf ∈ Lp and kMf kLp ≤ Ckf kLp ;

15
CHAPTER 4. THE MAXIMAL FUNCTION 16

Ckf kL1
c) if p = 1, then Mf ∈ L1w and kMf kL1w ≤ Ckf kL1 , i.e. |{ x ; Mf (x) > t }| ≤ for each
t
t > 0.
Here, C denotes a constant independent of f .
Proof. When p = ∞, the statement is clear and we may take C = 1. Let next p = 1. We fix some
t > 0. Let O = { x ; Mf (x) > t }, which is an open set. Thus |O| = sup{ |K| ; K compact ⊂ O }.
Let K be any compact in O. Z From the definition of Mf , for each x ∈ K there is some ball B
1
containing x such that |f (y)|dy > t. These balls cover K, so that we may extract a finite
|B|
B
covering. Using Vitali’s lemma, we may find a finite family F 0 = {Bj } such that
Z
1 X
Bj ∩ Bk = ∅ for j 6= k, |f (x)|dx > t, |Bj | ≥ C|K|. (4.3)
|Bj | j
Bj
Thus Z XZ X
kf kL1 ≥ |f (x)|dx = |f (x)|dx ≥ t |Bj | ≥ Ct|K|. (4.4)
j j
Bj
[
Bj
j

Ckf kL1
Taking the supremum over K in the above inequality, we find that |O| ≤ , i.e. the property
t
c). Letting t → ∞, we find a) for p = 1.
p
( now prove a) for 1 < p < ∞. Let f ∈ L . We split f as f = f1 + f2 , where f1 (x) =
We
f (x), if |f (x)| ≥ 1
and f2 = f − f1 . Then f1 ∈ L1 and f2 ∈ L∞ . Since Mf ≤ Mf1 + Mf2 ,
0, if |f (x)| < 1
we obtain a).
Finally, we prove b) for 1 ( < p < ∞. Let t > 0. We use a splitting of f similar to the above one:
f (x), if |f (x)| ≥ t/2 t
f = f1 + f2 , with f1 (x) = and f2 = f − f1 . Then Mf2 ≤ kf2 kL∞ ≤ . It
0, if |f (x)| < t/2 2
t
follows that Mf (x) > t ⇒ Mf1 (x) > . Therefore
2
t 2Ckf1 kL1
|{ x ; Mf (x) > t }| ≤ |{ x ; Mf1 (x) > }| ≤ , (4.5)
2 t
using c). We find that
Z∞ Z∞
kMf kpLp = p tp−1 |{ x ; Mf (x) > t }|dt ≤ C tp−2 kf1 kL1 dt. (4.6)
0 0
CHAPTER 4. THE MAXIMAL FUNCTION 17

If F is the distribution function of f , then we have

Z∞ Z∞
t t
kf1 kL1 = |{ x ; |f1 (x)| > s }|ds = F ( ) + F (s)ds. (4.7)
2 2
0 t/2
Thus
Z∞ Z∞ Z∞ Z∞ Z∞
1 t
tp−2 kf1 kL1 dt = tp−1 F ( )dt + tp−2 F (s)dsdt = Cp tp−1 F (t)dt = Ckf kpLp , (4.8)
2 2
0 0 0 t/2 0

by Fubini’s theorem.
2p−1 1
The constant in the last line of computations equals + . We obtain the following
p p(p − 1)
C
Corollary 4. For 1 < p ≤ 2 we have kMf kLp ≤ kf kLp for some constant C depending only
p−1
on N .
Remark 3. The maximal
Z function is never in L1 (except when f = 0). Indeed, if Zf =
6 0, there is
1 C
some R > 0 s. t. |f | > 0. Then, for |x| ≥ R, we have Mf (x) ≥ |f | ≥ N
|B(x, 2|x|)| |x|
B(0,R) B(0,R)
1
and thus M f 6∈ L .
Remark 4. By the above remark, if f ∈ L1 , f 6= 0, the best we can hope is that Mf ∈ L1loc .
1 1
However, this may not be true. Indeed, let f : R → R, f (x) = 2 χ[0,1/2] . Then f ∈ L .
x ln x
Zx
1 1
However, for x ∈ (0, 1/2), Mf (x) ≥ f (t)dt = , so that Mf 6∈ L1loc .
x x| ln x|
0

4.2 Lebesgue’s differentiation theorem

Theorem 5. (Lebesgue-Besicovitch) If f ∈ L1loc , then for a.e. x ∈ RN we have
CHAPTER 4. THE MAXIMAL FUNCTION 18
Z
1
lim f (y)dy = f (x).
r→0 |B(x, r)|
B(x, r)
Proof. We start by recalling the following simple measure theoretic
X
Lemma 3. (Borel-Cantelli) Let (An ) be a sequence of measurable sets such that |An | < ∞.
\ [ n
Then |limAn | = 0, where limAn = Am .
n m≥n
Z
1
Let f (x, r) = f (y)dy. The conclusion of the theorem being local, it suffices to
|B(x, r)|
B(x, r)
prove it with f replaced by f ϕ for any compactly supported smooth function ϕ. We may thus
assume that f ∈ L1 . Let n ≥ 1 and let fn be a smooth compactly supported function such
1
that kf − fn kL1 ≤ n . Let also gn = f − fn . Since fn is uniformly continuous, there is some
2
1
δn such that |fn (x, r) − fn (x)| ≤ for r ≤ δn and x ∈ RN . Thus, if for some r ≤ δn we have
n
2 1 1
|f (x, r) − f (x)| > , then we must have |gn (x, r) − gn (x)| > , so that either |gn (x)| > or
n n 2n
1 1
|gn (x, r)| > . In the latter case, we have Mgn (x) > . Therefore
2n 2n
2 1 1
{ x ; |f (x, r) − f (x)| > for some r ≤ δn } ⊂ An = { x ; |gn (x)| > or Mgn (x) > }. (4.9)
n 2n 2n
Cn
By the maximal and Chebysev’s inequalities, we find that |An | ≤ . If x 6∈ limAn , then clearly
2n X
n
lim f (x, r) = f (x). The theorem follows from the above lemma, since < ∞.
r→0
n
2n

The same argument yields the following variants of the differentiation theorem:
Theorem 6. If f ∈ L1loc , then for a.e. x ∈ RN we have
Z
1
lim f (y)dy = f (x). (4.10)
x ∈ Q, |Q| → 0 |Q|
Q

Here, we may choose the Q0 s to be balls or cubes (or, more generally, balls for some norm).
CHAPTER 4. THE MAXIMAL FUNCTION 19

Theorem 7. If f ∈ L1loc , then for a.e. x ∈ RN we have

Z
1
lim |f (y) − f (x)|dy = 0. (4.11)
x ∈ Q, |Q| → 0 |Q|
Q

We end this section with two simple consequences of the maximal inequalities and of the above
theorem:
Corollary 5. Let f ∈ L1loc . Then Mf ≥ |f | a.e.
Corollary 6. Let 1 < p ≤ ∞. Then kf kLp ≤ kMf kLp ≤ Ckf kLp .

4.3 Pointwise inequalities for convolutions

Proposition 6. Let ϕ be such that |ϕ| ≤ g for some g ∈ L1 , g non increasing and radially
symmetric. Then
|f ∗ ϕ(x)| ≤ kgkL1 Mf (x). (4.12)
Proof. Since |f ∗ ϕ| ≤ |f | ∗ g, it suffices to prove the proposition for |f | and g. We start with a
special case: we assume g to be piecewise constant; the general case will follow by approximation,
using e.g. the Beppo Levi theorem. We assume thus that there isX a sequence of radii r1 < r2 < ...
and a sequence of non negative numbers a1 , a2 , ... such that g = ak on B(0, rj ). Then
k≥j
Z X Z X
|f (x − y)||g(y)|dy = aj |f (x − y)|dy ≤ aj |B(0, rj )|Mf (x) = kgkL1 Mf (x),
j j
RN B(0, rj )
(4.13)
which is the desired estimate.
x
Let ϕ ∈ S(RN ). For t > 0, let ϕt (x) = t−N ϕ( ). As a consequence of the above proposition,
t
we derive the following
Corollary 7. We have
|f ∗ ϕt (x)| ≤ CMf (x). (4.14)
Here, C depends only on ϕ, not on t or f .
C
Proof. Since ϕ ∈ S, we have |ϕ(x)| ≤ g(x) = . Then clearly ϕt ≤ gt . Since g is in L1
1 + |x|N +1
and decreasing, so is gt . Moreover, we have kgt kL1 = kgkL1 . The corollary follows now from the
above proposition.
Chapter 5

The Calderón-Zygmund decomposition

kf kL1
If f ∈ L1 , then the set where f is large is relatively small, i.e. |{ x ; |f (x)| > t }| ≤ . The
t
following result provides a nice covering of this set.
Theorem 8. (The Calderón-Zygmund decomposition) Let f ∈ L1 (RN ) and t > 0. Then
there is a sequence of disjoint
[cubes (Cn ) such that:
N
a) |f (x)| ≤ t a. e. in R \ ( Cn );
n Z
−1 1
b) for each n we have C t ≤ |f (x)|dx ≤ Ct;
|Cn |
Cn
X Ckf kL1
c) |Cn | ≤ .
n
t

Here, C depends only on the space dimension N , not on f or t.

Proof. The construction looks like the Whitney decomposition. Fix some l > 0 such that lN >
kf kL1
. We cover RN with disjoint cubes of size l. We call F1 the family of all these cubes. We
t
bisect the cubes in F1 and call F2 the family of cubes obtained in this way. We keep bisecting
and obtain in the same way the families Fj , j ≥ 2. We start by throwing Zall the cubes in F1 .
1
For j ≥ 2, we keep a cube C in Fj if all its ancestors were thrown and |f (x)|dx > t. Let
|C|
[ C
F = (Cn ) be the family of all kept cubes and A = Cn . If x 6∈ A, then all the cubes containing
n
x were thrown. Thus |f (x)| ≤ t a.e. in RN \ A, by the Lebesgue differentiation theorem. Let now

20
CHAPTER 5. THE CALDERÓN-ZYGMUND DECOMPOSITION 21

C ∈ F. If j ≥ 3, then the (unique) cube Q in Fj−1 containing C was thrown, so that

2N
Z Z Z
1 1
|f (x)|dx ≤ |f (x)|dx = |f (x)|dx ≤ 2N t. (5.1)
|C| |C| |Q|
C Q Q

This inequality holds also for j = 2, by our choice of l. Thus b) holds with C = 2N .
Finally, c) follows from
XZ X
kf kL1 ≥ |f (x)|dx ≥ C −1 |Cn |t. (5.2)
n n
Cn

A variant of the above theorem is the following

X
Theorem 9. Let f ∈ L1 (RN ) and t > 0. Let (Cn ) as above. Then f = g + hn , where:
[ n
a) g ∈ L1 , |g| ≤ Ct a. e. and g = f in RN \ ( Cn );
n
b) supp hn ⊂ Cn ; Z
c) for each n we have hn (x)dx = 0;
Cn Z
1
d) for each n we have |hn (x)|dx ≤ Ct;
|Cn |
X Cn
e) kgkL1 + khn kL1 ≤ Ckf kL1 .
n

f (x),Z
 if x 6∈ A Z

1 1
Proof. Let g(x) = f (y)dy, if x ∈ Cn and hn (x) = f (x) − f (y)dy for x ∈ Cn . It

 |C n | |Cn |

Cn Cn
is easy to check that this decomposition has all the desired properties.
 Part II

Hardy and bounded mean oscillations

spaces

22
Chapter 6

Substitutes of L1

6.1 The space L log L

As we have already seen, if f ∈ L1 , we can expect at best Mf ∈ L1loc , but even this could be
wrong if we only assume f ∈ L1 . We present below a necessary and sufficient condition for having
Mf ∈ L1loc . R
A measurable function f belongs to L log L iff |f | ln(1 + |f |) < ∞. The space L log Lloc is
defined as the set of measurable functions s. t. f|K ∈ L log L for each compact K.
Theorem 10. Let f ∈ L1 . Then Mf ∈ L1loc ⇐⇒ f ∈ L log Lloc .
R
Remark R 5. Set Φ(t) = t ln(1 + t), t ≥ 0. If F is the distribution function of f , then |f | ln(1 +
|f |) = Φ0 (t)F (t)dt. It is easy to see that Φ0 (λs) ≤ max{1, λ}Φ0 (s) when λ, s > 0. Thus
Z Z Z Z
0 0
|λf | ln(1 + |λf |) = Φ (t)F (t/|λ|)dt = |λ| Φ (λs)F (s)ds ≤ Cλ |f | ln(1 + |f |) (6.1)

for each λ ∈ R. On the other hand, we have |{|f + g| > t}| ≤ |{|f | > t/2}| + |{|f | > t/2}|.
Therefore, if F is the distribution function of f and G the one of g, we have, with h = f + g,
Z Z Z Z
0
|h| ln(1 + |h|) ≤ Φ (t)(F (t/2) + G(t/2))dt ≤ 4 |f | ln(1 + |f |) + 4 |g| ln(1 + |g|). (6.2)

Thus L log L is a vector space. Similarly, L log Lloc is a vector space.

Proof. Assume that f ∈ L log Lloc . Fix a compact K ⊂ RN . Let F be the distribution function
Z of
f and let G be the distribution function of Mf|K . Note that G ∈ L∞ . Since kMf kL1 = G(t)dt,
Z∞
it suffices to check that G(t)dt < ∞. By combining (4.5) and (4.7), we find that, with some
2

23
CHAPTER 6. SUBSTITUTES OF L1 24

Z∞
universal constant C, we have G(t) ≤ C(F (t/2) + 1/t F (s)ds). Integrating this inequality, we
t/2
obtain
Z∞ Z∞ Z∞ Z
0
G(t)dt ≤ C (ln s + 2)F (s)ds ≤ c Φ (s)F (s)ds = c |f | ln(1 + |f |).
2 1 0
Z
Conversely, assume that Mf ∈ L1loc . 1
Since f ∈ L , that is F (s)ds < ∞, it suffices to prove
Z∞
that (Φ0 (s) − Φ0 (1))F (s)ds < ∞. Let, for a fixed t > 0, O = {Mf > t}. Note that c) of the
1 [
6 RN . Let O =
Hardy-Littlewood-Wiener theorem implies that O = C be a Whitney covering
C∈F
of O. Recall that, if C ∈ F, then there is an x ∈ RN \ O
Z s. t. dist(x,ZC) ≤ 3l(C), and thus
√
C ⊂ B(x, (3 + N )l(C)). Since Mf (x) ≤ t, we find that |f | ≤ |f | ≤ c l(C)N t,
√
C B(x,(3+ N )l(C))
Z
that is |f | ≤ ctG(t). Note that c does not depend on t. Since {|f | > t} ⊂ {Mf > t}, we
{Mf >t}
Z
find that |f | ≤ ctG(t). Invoking again (4.7), we obtain
{|f |>t}

Z∞ Z∞ Z
F (s)ds ≤ tF (t) + F (s)ds = |f | ≤ ctG(t), (6.3)
t t {|f |>t}

so that
Z∞ Z∞ Z∞ Z∞ Z∞
kMf kL1 ≥ G(t)dt ≥ c−1 F (s)/tdsdt = c−1 ln s F (s)ds ≥ d (Φ0 (s) − Φ0 (1))F (s)ds,
1 1 t 1 1
(6.4)
for some constant d.

6.2 The Hardy space H1

There is a different way to come around the difficulty that Mf is never in L1 . Maximal functions
are especially interesting because they provide pointwise estimates for convolutions. Instead of
CHAPTER 6. SUBSTITUTES OF L1 25

asking Mf to be in L1 , one could ask upper bounds for convolutions

Z convolutions to be in L1 . Here
it is how it works. Fix a smooth map Φ ∈ S(RN ) s. t. Φ 6= 0. Set, for t > 0, Φt = t−N Φ(·/t).
For u ∈ S 0 , let MΦ u = sup |u ∗ Φt |. We define, for 1 ≤ p ≤ ∞,
t>0

p
HΦ = {u ∈ S 0 ; MΦ u ∈ Lp }.
Z
Note that we may assume, without loss of generality, that (H1) Φ = 1. On the other hand,
Z
MΦ u = MΦs u, since f ∗ (us )t = f ∗ ust . The condition Φ = 1 reads also Φ̂(0) = 1; replacing,
if necessary, Φ by Φs for appropriate s, we may assume that (H2) 1/2 ≤ |Φ̂(ξ)| ≤ 3/2 for |ξ| ≤ 2.
We will always implicitly assume that the different test functions Φ, Ψ we will consider below are
admissible, in the sense that they satisfy (H1) and (H2).
This definition brings nothing new when 1 < p ≤ ∞.
p
Proposition 7. For 1 < p ≤ ∞, we have HΦ = Lp and kukLp ∼ kMΦ ukLp .
Proof. Recall that, if u ∈ Lp , then |u ∗ Φt | ≤ CMu, and thus MΦ u ∈ Lp . Conversely, assume that
MΦ u ∈ Lp . Then the family (u ∗ Φt )t is bounded in Lp and thus contains a sequence (u ∗ Φtn ) with
tn → 0, weakly-*
Z convergent in Lp . Since, on the other hand, u ∗ Φt → u in S 0 (here, we use the
assumption Φ = 1), we find that u ∈ Lp . Now, if u ∈ Lp , then (possibly along a subsequence
(tn )), u ∗ Φt → u a. e. and thus |u| ≤ MΦ u ≤ CMu, which together with the maximal theorem
implies the equivalence of norms.
1
We next note some simple properties of HΦ .
1
Proposition 8. a) u 7→ kMΦ ukL1 is a norm on HΦ ;
1 1
b) HΦ ⊂ L , with continuous inclusion;
1
c) HΦ is a Banach space.
Proof. The only property to be checked for a) is that kMΦ ukL1 = 0 =⇒ u = 0. If kMΦ ukL1 = 0,
then u ∗ Φt = 0 for each t; by taking the limit in S 0 as t → 0, we find that u = 0.
1
If u ∈ HΦ , then the family (u ∗ Φt )t is bounded in L1 and thus contains a sequence (u ∗ Φtn ) with
tn → 0, weakly-* convergent to some Radon measure µ. As above, this implies that u = µ, and
thus u is a Radon measure. We will prove that |µ| is absolutelyZ continuous with respect to the
Lebesgue measure. Let ε > 0. Then there is a δ > 0 s. t. |MΦ u| < ε whenever A is a Borel
A
set s. t. |A| < δ. If B is a Borel set s. t. |B| < δ, then there is an open set O containing B s. t.
CHAPTER 6. SUBSTITUTES OF L1 26

|O| < δ. Then

Z Z
|µ|(B) ≤ |µ|(O) = sup{| ϕ dµ| ; ϕ ∈ C0 (O), |ϕ| ≤ 1} ≤ MΦ u < ε, (6.5)
O

since Z Z Z Z
| ϕ dµ| = lim | u ∗ Φtn ϕ| ≤ MΦ u|ϕ| ≤ MΦ u. (6.6)
O
1
Thus u ∈ L . Moreover, |u| ≤ MΦ u, since the Lebesgue-Besicovitch differentiability theorem
implies, for a. e. x ∈ RN , that
Z Z
1 1 1
|u(x)| = lim |u| = lim |µ|(B) ≤ lim MΦ u = MΦ u(x); (6.7)
x∈B, |B|→0 |B| x∈B, |B|→0 |B| x∈B, |B|→0 |B|
B B

here, the limit is taken over all the balls. In conclusion, kukL1 ≤ kMΦ ukL1 , which implies b).
1
In order to prove that HΦ is a Banach
X space, it suffices to check
Xthat an absolutely convergent
X
1
series has a sum in HΦ . Assume that kMΦ fn kL1 < ∞. Then kfn kL1 < ∞ and thus fn
Xk X∞
1
converges in L to some f . Clearly, MΦ (f − fn ) ≤ MΦ fn → 0 as k → ∞ and thus
0 k+1
X
1
fn = f in HΦ .
1
There are two problems with the definition of HΦ . The first one is that this space depends, in
principle, on Φ. The second one is that it is not clear at all how to check that a given function
1
belongs to HΦ . A partial answer to the second question will be given in the next chapter. The
answer to the first question is given by the following
1 1
Theorem 11. (Fefferman-Stein) Let Φ, Ψ be two admissible functions. Then HΦ = HΨ and
1
kMΦ f kL1 ∼ kMΨ f kL1 for each f ∈ L .
In view of this result, we may define H1 = HΦ 1
for some admissible Φ, and endow it with the
norm kMΦ f kL1 .
The proof of the theorem is long and difficult; the remaining part of this chapter is devoted to
it.

6.3 More maximal functions

Let f ∈ L1 (we will always consider such f ’s, in view of the preceding proposition) and let Φ ∈ S.
We set
CHAPTER 6. SUBSTITUTES OF L1 27

F (x, t) = |f ∗ Φt (x)|
F ∗ (x) = MΦ f (x) = sup{F (x, t) ; t > 0}
Fa∗ (x) = sup{F (y, t) ; t > 0, |x − y| < at} (here, a > 0 is fixed)
tM
∗
Fa,ε,M (x) = sup{F (y, t) ; a−1 |x − y| < t < ε−1 } (here, ε, M are fixed positive
(ε + t + ε|y|)M
constants).
Finally, let, for α, β ∈ NN , pα,β be the semi norm pα,β (ϕ) = sup |xα ∂ β ϕ|, which is finite for
ϕ ∈ S. We consider any finite family F of such semi norms and set
F F (x) = sup{Mϕ f (x) ; ϕ ∈ S, pα,β (ϕ) ≤ 1, ∀ pα,β ∈ F}
tM
F
Fa,ε,M (x) = sup{|f ∗ ϕ(y, t)| M
; a−1 |x − y| < t < ε−1 , pα,β (ϕ) ≤ 1, ∀ pα,β ∈ F}
(ε + t + ε|y|)
We note the following elementary properties
F ∗ ≤ Fa∗ ≤ Fb∗ if 0 < a < b
(*) F ∗ ≤ cΦ,F F F
∗
limε→0 Fa,ε,M = Fa∗
∗
limε→0 Fa,ε,M ≥ FF
if F ⊃ F0 , then FF ≤ FF0 .
The main theorem is an immediate consequence of (*) and of the following
Theorem 12. There is a finite family F0 s. t.
Z Z Z
−1
c ∗ F0
F ≤ F ≤ c F ∗.

Here, c depends on Φ and F0 , but not on f . An immediate consequence of the theorem is the
following
Corollary 8. If F ⊃ F0 , then, with a constant c independent of f , we have
Z Z Z
−1
c F ≤ F ≤ c F ∗.
∗ F

We end this section with a simple statement we will need in the proof of Theorem 5.
Lemma 4. With a constant c depending only on N , we have
Z Z
∗ N ∗
Fb,ε,M ≤ c(b/a) Fa,ε,M , 0 < a < b.

∗
Proof. Set, for α > 0, Oa = {Fa,ε,M > α} and define similarly Ob . Then x ∈ Ob iff there are y, t s. t.
tM
|y −x| < bt, t < ε−1 and F (y, t) > α. It follows immediately that x ∈ B(y, bt) ⊂ Ob
(ε + t + ε|y|)M
and that B(y, at) ⊂ Oa . Let now K ⊂ Ob be a fixed compact. Since the balls B(y, bt) cover Ob ,
CHAPTER 6. SUBSTITUTES OF L1 28

we may find a finite collection of such balls that cover K. In addition, Vitali’s lemma implies
that we may find a finite collection of such balls, say (B(yi , bti )), mutually disjoint and s. t.
P
|B(yi , bti )| ≥ c|K|, where K depends only on N . Since
P aN< b, the corresponding
P balls (B(yi , ati ))
are mutually disjoint and contained in Oa . Thus bN ti ≥ c|K|, while aN tN
i ≤ c|Oa |. We
N
find that |K| ≤ c(b/a) |Oa |; by taking the sup over K, we find that
∗ ∗
|{Fb,ε,M > α}| ≤ c(b/a)N |{Fa,ε,M > α}|. (6.8)
The conclusion of the lemma follows by integrating the above inequality over α > 0.

6.4 Transition from one admissible function to another

Lemma 5. Let F0 be a finite collection of semi norms. Then there is another finite collection F
of semi norms s. t., for each ϕ ∈ S, sup{pα,β (ϕ) ; pα,β ∈ F0 } ≤ c sup{pα,β (ϕ̂) ; pα,β ∈ F}.
Proof. We have, for each α, β,
Z
−N
|x ∂ ϕ(x)| = (2π) | eıx·ξ (ı∂)α [(ıξ)β ϕ̂](ξ)dξ| ≤ c sup(1 + |ξ|)N +1 |(ı∂)α [(ıξ)β ϕ̂]| ≤ c sup pγ,δ (ϕ̂),
α β
Fα,β
[
for some appropriate family Fα,β . We may thus take F = Fα,β .
pα,β ∈F0

Remark 6. We may, of course, reverse the roles of ϕ and ϕ̂ in the above lemma.
Lemma 6. Let Φ be an admissible function, let M > 0 and let F0 be a finite collection of
semi norms. Then there is another finite collection F s. t. we may write each ϕ ∈ S as ϕ =
X∞
Φ2−k ∗ η k . Here, the series is convergent in S, the functions η k (which depend on ϕ) belong to
k=0
S and satisfy
pα,β (η k ) ≤ c2−kM sup{pγ,δ (ϕ) ; pγ,δ ∈ F}, ∀ pα,β ∈ F0 . (6.9)
Here, c and L do not depend on ϕ.
Proof. Assume (6.9) proved, for the moment. Since pα,β (Φ2−k ) = 2k(N −|α|+|β|) pα,β (Φ), we find, by
taking M sufficiently large, that the series giving ϕ is convergent in S.
In view of the preceding lemma and of the remark that follows it, it suffices to prove, for each
pα,β and for an appropriate F, the estimate

pα,β (ηbk ) ≤ c2−kM sup{pγ,δ (ϕ̂) ; pγ,δ ∈ F}. (6.10)

Fix a function ζ ∈ C0∞ s. t. ζ(ξ) = 1 when |ξ| ≤ 1 and ζ(ξ) = 0 when |ξ| ≥ 2. Define ζ0 = ζ and,
for k ≥ 1, ζk (ξ) = ζ(2−k ξ) − ζ(2−k+1 ξ). Clearly, for k ≥ 1, ζk (ξ) = 0 unless if 2k−1 ≤ |ξ| ≤ 2k+1 ,
CHAPTER 6. SUBSTITUTES OF L1 29

∞
X
and, for each ξ, ζk (ξ) = 1 (despite the series that appears here, for each ξ the sum contains at
k=0
−k
most three non vanishing terms). Thus, noting that Φ
[2−k (ξ) = Φ̂(2 ξ), we find that
∞ ∞ ∞
X X ϕ̂(ξ)ζk (ξ) X
k
ϕ̂(ξ) = ϕ̂(ξ)ζk (ξ) = Φ
[ 2−k (ξ) ≡ Φ
[ 2−k (ξ)Ψ (ξ). (6.11)
k=0 k=0 Φ
[ 2−k (ξ) k=0

We first note that Ψk is well-defined. Indeed, Φ being admissible, we have 1/2 ≤ |Φ

[2−k (ξ)| ≤ 3/2
if ξ ∈supp ζk . Moreover, ζk being compactly supported, so is Ψ . Finally, Ψ ∈ C0∞ , and thus
k k

Ψk = ηbk for some η k ∈ S. It remains to establish (6.10), i. e.

pα,β (Ψk ) ≤ c2−kM sup{pγ,δ (ϕ̂) ; pγ,δ ∈ F} (6.12)
for some appropriate F.
Set Ψ = 1/Φ̂ ∈ C ∞ (B(0, 2)). Since for ξ ∈supp ζk we have 2−k ξ ∈ B(0, 2), we find, for such ξ,
−|β|k β
|∂ β (1/Φ
[2−k )(ξ)| = 2 |∂ Ψ(2−k ξ)| ≤ cβ .
Similarly, |∂ β ζk | ≤ cβ . Therefore, for k ≥ 1 and ξ ∈supp ζk
X
|ξ α ∂ β Ψk (ξ)| ≤ c |∂ γ ϕ̂(ξ)| ≤ c(1 + |ξ|)−M sup pα,β (ϕ̂) ≤ c2−kM sup pα,β (ϕ̂),
pα,β ∈F pα,β ∈F
γ≤β

provided we choose F properly. A similar conclusion holds for k = 0, completing the proof of the
lemma.
We set C0 = {x ; |x| ≤ 2} and, for j ∈ N∗ , Cj = {x ; 2j−1 ≤ |x| < 2j }.
Corollary 9. For each M > 0, we may choose, in the above lemma, F is sufficiently rich in order
to have Z
|η k | ≤ c2−M (k+j) (6.13)
Cj

provided pα,β (ϕ) ≤ 1 when pα,β ∈ F.

Proof. If F0 is sufficiently rich and ϕ satisfies pα,β (ϕ) ≤ 1 when pα,β belongs to the corresponding
F, then sup(1 + |x|)N +M |η k (x)| ≤ c2−kM . In this case,
Z Z
|η | ≤ (1 + |x|)−N −M sup(1 + |x|)N +M |η k (x)| ≤ c2−M (k+j) .
k
(6.14)
Cj
Cj Cj
CHAPTER 6. SUBSTITUTES OF L1 30

6.5 Proof of Theorem 12

∗ F
Proof. Step 1. F1,ε,M controls F2,ε,M
Lemma 7. Assume that F is sufficiently rich and that 0 < ε < 1. Then there is a constant c,
which may depend on F and M , but not on f or ε, s. t.
Z Z
F ∗
F2,ε,M ≤ c F1,ε,M . (6.15)

Proof. Fix an ε > 0. By the transition lemma, we have

XZ XZ
k −N
|f ∗ ϕt (z)| ≤ |f ∗ Φ2−k t (z − y)||ηt (y)|dy = t |f ∗ Φ2−k t (z − y)||η k (y/t)|dy. (6.16)
k k
−1 −1
If 2 |z − x| < t < ε and y ∈ tCj (the sets Cj were defined in the preceding section), we find
that |y − x| < 2k+j+2 2−k t, while 2−k t < ε−1 . Thus
(ε + 2−k t + ε|z − y|)M ∗
|f ∗ Φ2−k t (z − y)| = F (z − y, 2−k t) ≤ F2k+j+2 ,ε,M (x), (6.17)
(2−k t)M
so that
(ε + 2−k t + ε|z| + 2j t)M ∗
|f ∗ Φ2−k t (z − y)| ≤ F2k+j+2 ,ε,M (x). (6.18)
(2−k t)M
We find that
X (ε + 2−k t + ε|z| + ε2j t)M Z
−N
|f ∗ ϕt (z)| ≤ t F2∗k+j+2 ,ε,M (x) |η k (y/t)|. (6.19)
k,j
(2−k t)M
tCj

Therefore, for each fixed L > 0 we have, if F is sufficiently rich,

X (ε + 2−k t + ε|z| + ε2j t)M
|f ∗ ϕt (z)| ≤ c −k t)M
F2∗k+j+2 ,ε,M (x)2−L(k+j) . (6.20)
k,j
(2
F
The definition of F2,ε,M implies that, with A = {ϕ ; pα,β (ϕ) ≤ 1, ∀ ϕ ∈ F},

F tM X
F2,ε,M (x) = sup |f ∗ ϕt (z)| ≤ c 2M (j+2k) F2∗k+j+2 ,ε,M (x)2−L(k+j)
ϕ∈A; 2−1 |z−x|<t<ε−1 (ε + t + ε|z|)M k,j
(6.21)
(here, we use the assumption ε < 1). Recalling that L is arbitrary, we take L = 2M + N + 1 and
find that
Z X Z X Z Z
F −M j−(N +1)(k+j) ∗ −M j−(k+j) ∗ ∗
F2,ε,M (x) ≤ c 2 F2k+j+2 ,ε,M ≤ c 2 F1,ε,M ≤ c F1,ε,M .
k,j k,j
(6.22)
CHAPTER 6. SUBSTITUTES OF L1 31

∗
Step 2. MΦ f controls F1,ε,M
∗
Lemma 8. If M > N , then F1,ε,M ∈ L1 .
Proof. We note that |f ∗ Φt | ≤ kf kL1 kΦt kL∞ ≤ ct−N . Thus

∗ tM −N 1
F1,ε,M (x) ≤ c sup ≤ c ε sup , (6.23)
|y−x|<t<ε−1 (ε + t + ε|y|)M |y−x|<ε−1 (1 + |y|)
M

and the latter function belongs to L1 .

Lemma 9. Assume that M > N and that ε < 1. Then, with some constant c that may depend
on M , but not on ε or f , we have
Z Z
∗
F1,ε,M ≤ c MΦ f. (6.24)

Proof. For each x, there are t and y s. t. |x − y| < t < ε−1 and

∗ tM 3 ∗
F1,ε,M (x) ≥ F (y, t) M
≥ F1,ε,M (x).
(ε + t + ε|y|) 4
Let δ be a small constant to be fixed later. We claim that, if δ is sufficiently small and F is
sufficiently rich (i. e., as in the preceding step), then
tM tM F 1 ∗
|z − y| < δt =⇒ |F (y, t) M
− F (z, t) M
| ≤ cδF2,ε,M (x) + F1,ε,M (x).
(ε + t + ε|y|) (ε + t + ε|z|) 4
(6.25)
The above implication is an immediate consequence of the following inequalities
tM tM 1 tM
| − |≤ , (6.26)
(ε + t + ε|y|)M (ε + t + ε|z|)M 4 (ε + t + ε|y|)M
respectively
tM F
|F (y, t) − F (z, t)| M
≤ cδF2,ε,M (x). (6.27)
(ε + t + ε|z|)
Inequality (6.26) is elementary and left to the reader (it works when δ < (4/3)1/M − 1). As for
(6.27), we start by noting that  
∂f ∗ Φt 1 ∂Φ
= f∗ , (6.28)
∂xj t ∂xj t
CHAPTER 6. SUBSTITUTES OF L1 32

and thus  
|y − z| ∂Φ
|f ∗ Φt (y, t) − f ∗ Φt (z, t)| ≤ c sup sup |f ∗ |(w). (6.29)
t 1≤j≤N |w−z|<δt ∂xj t

Assuming, without loss of generality, that δ < 1, we find that

tM F (ε + t + ε|w|)M F
|F (y, t) − F (z, t)| M
≤ cδF2,ε,M (x) sup M
≤ 2M cδF2,ε,M (x), (6.30)
(ε + t + ε|z|) |w−z|<δt (ε + t + ε|z|)

whence (6.27).
For each x, one of the two happens:
F 1 ∗
either (i) cδF2,ε,M (x) ≤ F1,ε,M (x),
4
F 1 ∗
or (ii) cδF2,ε,M (x) > F1,ε,M (x).
4
Let A, respectively B, be the set of points s. t. (i), respectively (ii) holds. If x ∈ A, then we have
1 ∗
F (z, t) ≥ F1,ε,M (x) whenever |z − y| < δt, and thus
4
p p 1q ∗
MΦ f (z) ≥ F (z, t) ≥ F1,ε,M (x)
2
for each such z. Thus
Z
q
∗ c p
F1,ε,M (x) ≤ MΦ f (z). (6.31)
|{|z − y| < δt}|
{|z−y|<δt}

Noting that {|z − y| < δt} ⊂ {|z − x| < 2t}, we find that, in case (i),
Z
q
∗ cδ p p
F1,ε,M (x) ≤ MΦ f (z) ≤ M( MΦ f )(x). (6.32)
|{|z − x| < 2t}|
{|z−x|<2t}

Therefore, Z Z Z
∗
p 2
F1,ε,M ≤c (M( MΦ f )) ≤ c MΦ f, (6.33)
A A

by the maximal inequalities. (Here, the different constants may depend on δ.)
Concerning the set B, we have
Z Z Z
∗ F 0 ∗
F1,ε,M ≤ 4cδ F2,ε,M ≤ c δ F1,ε,M . (6.34)
B B
CHAPTER 6. SUBSTITUTES OF L1 33

We finally fix δ sufficiently small in order to have (6.26), (6.27), δ < 1 and c0 δ < 1/2. Then
Z Z Z
∗ 1 ∗ 1 ∗
F1,ε,M ≤ F1,ε,M + F1,ε,M , (6.35)
2 2
B B A

so that Z Z
∗
F1,ε,M ≤c MΦ f, (6.36)

by combining (6.33) and (6.35).

Step 3. Conclusion Z Z
By letting ε → 0 in Lemma 9, we find that F1∗
≤ c MΦ f . Next, letting ε → 0 in Lemma
Z Z
9 yields F F ≤ c F1∗ . Finally, it suffices to note that MΦ f ≤ cF F .
Chapter 7

Atomic decomposition

7.1 Atoms
For the moment, we do not even know if H1 contains a non zero function! In this section, we will
give examples of functions in H1 : the atoms. In the next section, we will show that this example
is ”generic”. To motivate the definition of atoms, we start with the following simple
Z
1
Proposition 9. If f ∈ H , then f = 0.
Z Z
Proof. Argue by contradiction and assume, e. g., that f = 1. Pick some R > 0 s. t. f>
B(0,R)
Z
2/3 and |f | < 1/3. Let Φ ∈ C0∞ be s. t. Φ = 1 in B(0, 1) and 0 ≤ Φ ≤ 1. For x ∈ RN s.
RN \B(0,R)
t. |x| > R, let t = |x| + R, so that t ∼ |x|. Then
Z Z
−N 1
MΦ f (x) ≥ f ∗ Φt (x) ≥ t f − t−N |f | ≥ t−N ≥ c|x|−N , (7.1)
3
B(0,R) RN \B(0,R)

and thus MΦ f 6∈ L1 .
Z
1 1
Remark 7. H is a strict subspace of {f ∈ L ; f = 0}. To see this, it suffices to modify
1
the example in Remark 4 as follows: set f1 : R → R, f1 (x) = χ[0,1/2] and let f (x) =
Z x ln2 x
f1 (x) − f1 (3 − x). Then f ∈ L1 and f = 0. However, if we pick Φ ∈ C0∞ s. t. 0 ≤ Φ ≤ 1, supp

34
CHAPTER 7. ATOMIC DECOMPOSITION 35

Φ ⊂ [0, 2] and Φ = 1 in [0, 1], then, for x ∈ [0, 1/2] we have

Zx
1
MΦ f (x) ≥ x−1 f1 (x)dx = , (7.2)
x| ln x|
0

and thus MΦ f 6∈ L1 .
Definition 1. An atom is a function a : RN → R s. t.:
(i) supp a ⊂ B, where B is a ball;
−1
(ii) |a|
Z ≤ |B| ;
(iii) a = 0.

We may replace balls by cubes, in this definition, since if a is an atom with respect to a ball
B, then cN a is an atom with respect to any minimal cube containing B and conversely; here, cN
depends only on N .
Proposition 10. If a is an atom, then MΦ f ∈ L1 and kMΦ f kL1 ≤ c for some constant depending
only on Φ.
Proof. Since MΦ f ≤ cMf ≤ kf kL∞ , we have MΦ f ≤ c|B|−1 . Therefore,
Z
MΦ f ≤ c|B ∗ ||B|−1 = c2N ; (7.3)
B∗

here, B ∗ is the ball concentric to B and twice larger.

If x 6∈ B ∗ , we use the information (iii) and find that, with R the radius of B, we have
Z Z
−1
|f ∗ Φt (x)| = | a(y)[Φt (x − y) − Φt (x)]dy| ≤ |B| sup |∇(Φt )(x − z)| |y|dy. (7.4)
z∈B
B B

Taking into account the fact that |x − z| ∼ |x| and the inequality |∇Φ(x)| ≤ c|x|−N −1 , we obtain
cR cR
|f ∗ Φt (x)| ≤ N +1 , and thus MΦ f (x) ≤ N +1 . Integrating the latter inequality, we find that
|x| |x|
Z
MΦ f ≤ c (7.5)
RN \B ∗

and the desired inequality follows from (7.3) and (7.5).

X X
Corollary 10. Let f = λk ak , where each ak is an atom and |λk | < ∞. Then f ∈ H1 and
X
kf kH1 ≤ c |λk |.
CHAPTER 7. ATOMIC DECOMPOSITION 36

More generally, we could weaken condition (ii) in the definition of an atom as follows
Definition 2. Let 1 < q ≤ ∞. A q atom is a function satisfying (i), (iii) and
(ii’) kakLq ≤ |B|1/q−1 .
Thus, the usual atoms are ∞ atoms.
c
Proposition 11. If a is a q atom, then kakH1 ≤ cq . If, in addition, q ≤ 2, then cq ≤ .
q−1
Proof. We may assume that q < ∞. We repeat the reasoning in the preceding proposition. On
the one hand, we have
Z Z Z 1/q
∗ 1−1/q q
|MΦ f | ≤ c |Mf | ≤ c|B | |Mf | ≤ cq ; (7.6)
B∗ B∗

c
here, we use Hölder’s inequality and the maximal theorem. In addition, we see that cq ≤ if
q−1
q ≤ 2.
When x 6∈ B ∗ , we find that
Z Z 1/q0
c c q0 cq R
MΦ f (x) ≤ |f (y)||y|dy ≤ kf kLq |y| ≤ ; (7.7)
|x|N +1 |x|N +1 |x|N +1
B B

here, cq remains bounded when q ≤ 2. Thus

Z
MΦ f ≤ c (7.8)
RN \B ∗

with c independent of q ≤ 2. We conclude by combining (7.6) and (7.8).

7.2 Atomic decomposition

The following result tells that the atoms represent ”generic”H1 functions.
X
Theorem 13. (Coifman-Latter) Let f ∈ H1 . Then we may write f = λk ak , where each ak
X
is an atom and |λk | ∼ kf kH1 .
X X
Proof. It suffices to write, in the sense of distributions, f = λk ak , with |λk | ≤ ckf kH1 .
X
Indeed, if we are able to do this, then on the one hand the series λk ak is convergent in H1 ,
CHAPTER 7. ATOMIC DECOMPOSITION 37

thus in D0 , andXtherefore its sum Xhas to be f , by uniqueness of the limit. On the other hand, we
always have k λ k ak k H 1 ≤ c |λk |.
We fix a large family F of semi norms as in the preceding section. Let F F be the corresponding
maximal function, i. e.,
F F (x) = F F f (x) = sup{MΦ f (x) ; pα,β (Φ) ≤ 1, ∀ pα,β ∈ F}.
Let, for j ∈ Z, Oj = {F F > 2j }; clearly, Oj is an open set and Oj+1 ⊂ Oj . In addition, Oj 6= RN ,
since F F ∈ L1 . Set fj = f χOj .
Lemma 10. As j → ∞, fj → 0 in L1 . As j → −∞, fj − f → 0 in L∞ .
Z
Proof. We have kfj kL1 = |f | → 0 as j → ∞, since |Oj | → 0 as j → ∞. On the other hand,
Oj

kfj − f kL∞ = sup |f | ≤ sup Mf ≤ c sup F F ≤ c2j → 0 (7.9)

RN \Oj RN \Oj RN \Oj

as j → ∞.
∞
X
Corollary 11. Set gj = fj − fj+1 . Then gj = f in the distribution sense.
−∞

Let(Ckj ) be a Whitney covering of Oj and let ϕjk be the corresponding partition of the unit in
Oj . Recall that, with 1 < a < b depending only on N , we have
(i) Ckj∗ ⊂ Oj (where Ckj∗ is the cube concentric with Ckj and having a times its size);
(ii) Ckj∗∗ 6⊂ Oj (where Ckj∗∗ is the cube concentric with Ckj and having b times its size);
(iii) at most M cubes Ckj∗ meet at some point, where M depends only on N ;
(iv) supp ϕjk ⊂ Ckj∗ ;
(v) |∂ α ϕjk | ≤ cα size(Ckj )−α ;
(vi) ϕjk ≥ 1/M in Ckj .
The last
Z property implies
(vii) ϕjk ∼ size(Ckj ).
X j
We have fj = f ϕk (the series that appears is well-defined, at least in the sense of distribu-
tions, since on each compact
Z there are finitely many non vanishing terms). We define the coefficient
X X j j
cjk by the condition (f − cjk )ϕjk = 0. We have fj = (f − cjk )ϕjk + Rj , where Rj = ck ϕ k .
∞
X
Lemma 11. We have (Rj − Rj+1 ) = 0 in the sense of distributions.
−∞
CHAPTER 7. ATOMIC DECOMPOSITION 38

Proof. We have Z Z Z Z
|cjk ϕjk | = |cjk ϕjk | =| f ϕjk | ≤ |f |, (7.10)
Ckj∗

and thus XZ Z
kRj kL1 ≤ |f | ≤ M |f |; (7.11)
Ckj∗ Oj
X j j X j
here, M is the constant in (iii). The series ck ϕk and f ϕk being convergent in L1 , we have
XZ j j XZ j
Z X
j
Z
ck ϕ k = f ϕk = f ϕk = f. (7.12)
Oj

Thus,
j=k Z Z Z
X
lim (Rj − Rj+1 ) = lim f= f= f = 0, (7.13)
k→∞ k→∞
j=−k
O−k \Ok+1 {F F >0}

since f is vanishing in the set {F F = 0}.

X∞ X X 
j j j+1 j+1
Corollary 12. We have f = (f −ck )ϕk − (f −cl )ϕl in the sense of distributions.
−∞ k l
X
Using the fact that ϕj+1
l = ϕj+1
l ϕjk (since Oj+1 ⊂ Oj ), we may further decompose the
k
general term of the above series as follows
X X X X X j j+1
(f −cjk )ϕjk − (f −cj+1
l )ϕj+1
l = (f −cjk )ϕjk − [(f −cj+1
l )ϕjk −cjk,l ]ϕj+1
l − ck,l ϕl ; (7.14)
k l k k,l k,l
Z
here, the coefficients cjk,l are chosen s. t. [(f − cj+1
l )ϕjk − cjk,l ]ϕj+1
l = 0.

Z Actually, the last sum in (7.14) vanishes. The reason is that, for fixed l, we have, with c =
ϕj+1
l 6= 0,

X Z X Z X Z
c cjk,l = cjk,l ϕj+1
l = (f − cl )ϕk ϕl = (f − cj+1
j+1 j j+1
l )ϕj+1
l = 0; (7.15)
k k k

commuting the series with the integral in the above computations is justified by the fact that,
when l is fixed, we have only finitely many non vanishing terms.
CHAPTER 7. ATOMIC DECOMPOSITION 39

∞ X
X
Thus f = bjk , where
j=−∞ k
X X j+1 j j+1
bjk = (f − cjk )ϕjk − [(f − cj+1
l )ϕjk − cjk,l ]ϕj+1
l = f ϕjk χRN \Oj+1 − cjk ϕjk + [cl ϕk ϕl + cjk,l ϕj+1
l ].
l l
(7.16)
Let C > 0 be a large constant to be specified later. We set, with lkj the size of Ckj , λjk = C(lkj )N 2j
and ajk = (λjk )−1 bjk , so that
X∞ X
f= λjk ajk ; (7.17)
j=−∞ k

this is going to be the atomic decomposition of f . Clearly, the functions ajk satisfy, by construction,
the cancellation property (iii) required in the definition of an atom. It remains to establish three
facts: a) that the support of ajk is contained
X in some ball B; b) that |ajk | ≤ |B|−1 (here, the choice
of the constant C will count); c) that |λjk | ≤ Ckf kH1 . These information are easily obtained
j,k
by combining the conclusions of the following lemmata.
Lemma 12. There is a constant b > 0 depending only on N s. t. supp bjk ⊂ Bkj , where Bkj is the
ball concentric with Ckj and of radius b lkj .
Proof. If bjk (x) 6= 0, then either x ∈ supp ϕjk ⊂ Ckj∗ , or there is some l s. t. supp ϕjk intersects
supp ϕj+1
l and s. t. ϕj+1
l (x) 6= 0. In the latter case, we have, on the one hand, x ∈ Clj+1∗ . On the
other hand, if y ∈ supp ϕjk ∩ supp ϕj+1 l ⊂ Ckj∗ ∩ Clj+1∗ , then

llj+1 ≤ c1 dist (y, RN \ Oj+1 ) ≤ c1 dist (y, RN \ Oj ) ≤ c2 lkj . (7.18)

In both cases, we may find b s. t. the conclusion of the lemma holds.
Lemma 13. We have |cjk | ≤ c2j , and |cjk,l | ≤ c2j . Here, c depends only on N and F.
Z Z Z
j j j
Proof. By definition, we have ck = f ϕk / ϕk . As already noted, we have ϕjk ∼ (lkj )N . Let
x ∈ Ckj∗∗ \ Oj ; thus F F f (x) ≤ 2j , by the definition of Oj . Set t = lkj and ϕ(z) = ϕjk (x − zt).
Clearly, Z Z Z Z Z
j j j N j
ck = f ϕk / ϕk = t f (y)ϕt (x − y)dy/ ϕk ∼ f (y)ϕt (x − y)dy, (7.19)

so that
|cjk | ≤ cF F f (x) sup{pα,β (ϕ) ; pα,β ∈ F} ≤ c2j sup{pα,β (ϕ) ; pα,β ∈ F}. (7.20)
CHAPTER 7. ATOMIC DECOMPOSITION 40

Therefore, it suffices to prove that pα,β (ϕ) ≤ cα,β , with cα,β depending only on N . We first note
that, since x ∈ Ckj∗∗ and supp ϕjk ⊂ Ckj∗ , there is some constant R > 0 depending only on N s. t.
supp ϕ ⊂ B(0, R). Thus we may consider only the case α = 0. Now
sup |∂ β ϕ| = t|β| sup |∂ β ϕjk | ≤ t|β| cβ (lkj )−|β| = cβ , (7.21)
by the properties of Whitney’s partition of the unit.
The argument for cjk,l is similar. Since
Z Z Z Z
j j j+1 j+1 j+1
ck,l = f ϕk ϕl / ϕl − cl ϕk ϕl / ϕj+1
j j+1
l , (7.22)

we find that Z Z
|cjk,l | ≤ |cj+1
l | +| f ϕjk ϕj+1
l / ϕj+1
l |; (7.23)

the latter term appears only if ϕjk ϕj+1

l 6≡ 0.
j j+1
It suffices to prove that, when ϕk ϕl 6≡ 0, we have
Z Z
j j+1
| f ϕk ϕl / ϕj+1l | ≤ c2j . (7.24)

To this purpose, we pick an x ∈ Clj+1∗∗ \ Oj+1 and define, with t = llj+1 , Φ(z) = ϕjk (x − zt)ϕj+1
l (x −
j j+1 j+1
zt). Since supp ϕk ϕl ⊂ supp ϕl , we have (with the same R as above) supp Φ ⊂ B(0, R). In
addition,
tN
Z Z
| f ϕk ϕl / ϕj+1
j j+1
l | = Z |f ∗ Φt (x)| ≤ c2j sup{pα,β (Φ) ; pα,β ∈ F}. (7.25)
j+1
ϕl

Therefore, it suffices to prove that p0,β (Φ) ≤ cβ , with cβ depending only on N . Using the properties
of the Whitney decomposition, we have
|∂ β Φ(z)| ≤ ct|β| (lkj + llj+1 )−|β| ≤ cβ , (7.26)
since we have already noted that llj+1 ≤ clkj if the supports of ϕjk and ϕj+1
l do intersect.
Lemma 14. With some constant c depending only on N and of the family F of semi norms, we
have |f | ≤ c2j in the support of bjk .
Proof. In view of the second equality in (7.16), we have
X j+1
|bjk | ≤ |f |χRN \Oj+1 + |cjk | + (|cl | + |cjk,l |) ≤ c 2j + |f |χRN \Oj+1 . (7.27)
l

The desired conclusion is obtained by noting that |f | ≤ cF F f a. e., and thus |f | ≤ c 2j in

RN \ Oj+1 .
CHAPTER 7. ATOMIC DECOMPOSITION 41

By combining the above results, we find immediately that ajk are atoms, provided we chose C
sufficiently large (depending only on N and F).
We may now conclude as follows: we have
X j X X X
|λk | ≤ c 2j (lkj )N = c 2j |Ckj | = c 2j |Oj |. (7.28)
j,k j,k j

On the other hand,

j
Z ∞ Z2
X ∞
X
F F F
kF f kL1 = |{F f > α}|dα ≥ |{F f > α}|dα ≥ 2j−1 |Oj |. (7.29)
−∞ −∞
2j−1

If we take F sufficiently rich, we find, by combining (7.28) with (7.29), that

X j
kf kH1 ∼ kF F f kL1 ≥ c |λk |. (7.30)

The proof of the theorem is complete.

Corollary 13. On H1
X X
kf k = inf{ |λk | ; f = λk ak , the a0k s are atoms}

is a norm equivalent to the usual ones.

Chapter 8

The substitute of L∞: BM O

8.1 Definition of BM O
Definition 3. A function f ∈ L1loc belongs to BM O (=bounded mean oscillation) if
Z Z
1 1
kf kBM O = sup{ |f − f | ; C cube with sides parallel to the axes} < ∞. (8.1)
|C| |C|
C C

Despite the notation, k · kBM O is not a norm, since kf kBM O = 0 when f is a constant. However,
it is easy to see that, if we identify two functions in BM O when their difference is constant (a.
e.), then k · kBM O is a norm on the quotient space (still denoted BM O). Z
1
It will be convenient to denote fC the average of f on C, i. e., fC = f.
|C|
C

Proposition 12. a) BM O is a Banach space.

b) For each cube C and each constant m, we have
Z Z
1
|f − m| ≥ |f − fC |. (8.2)
2
C C

c) We may replace cubes by balls; the space remains the same and the norm is replaced by an
equivalent one. Similarly, we may consider cubes in general position.
d) If C ⊂ Q are parallel cubes of sizes l ≤ L, then |fC − fQ | ≤ c(1 + ln(L/l))kf kBM O .
e) If Ψ is a Lipschitz function of Lipschitz constant k, then kΨ ◦ f kBM O ≤ 2kkf kBM O .
Warning: In e), we do not identify two functions if there difference is constant.
X
Proof. a) Let fn be an absolutely convergent series in BM O. Let C be a cube. The series
X X
(fn |C − (fn )C ) is absolutely convergent (thus convergent) in L1 . Set f C = (fn |C − (fn )C ).

42
CHAPTER 8. THE SUBSTITUTE OF L∞ : BM O 43
Z
Then f C = 0 and
C
Z X 1 Z
1 C
X
|f | ≤ |fn − (fn )C )| ≤ kfn kBM O . (8.3)
|C| |C|
C C

We now cover RN with an increasing sequence of cubes (Ck ). We set f (x) = f Ck (x) − (f Ck )C0 if
x ∈ Ck . We claim that the definition is correct (in the sense that it does not depend on the choice
of Ck ). This follows immediately from the equality
X
(f C )C0 = [fn |C − (fn )C0 ], (8.4)

valid whenever C0 ⊂ C.
b) We have Z Z
|C||m − fC | = | (m − f )| ≤ |m − f |, (8.5)
C C

and thus Z Z Z Z
|f − fC | ≤ |f − m| + |m − fC | ≤ 2 |m − f |. (8.6)
C C C C

c) We prove the assertion concerning balls. The proof of the other statement is analog. Let B be
a ball and let C, Q be cubes s. t. C is inscribed in B and B is inscribed in Q. Then
Z Z Z Z
1 1 1 c
|f − fB | ≤ |f − fQ | ≤ |f − fQ | ≤ |f − fQ | (8.7)
2|B| |B| |B| |Q|
B B Q Q

and similarly
Z Z Z Z
1 1 1 c
|f − fC | ≤ |f − fB | ≤ |f − fB | ≤ |f − fB |, (8.8)
2|C| |C| |C| |B|
C C B B

so that the supremum over the balls and the supremum over the cubes are equivalent quantities.
d) We have
Z Z Z
1 1 1
|fC − fQ | = | (f − fQ )| ≤ |f − fQ | ≤ |f − fQ | ≤ (L/l)N kf kBM O , (8.9)
|C| |C| |C|
C C Q

which implies the desired estimate when L/l ≤ 2. If L/l > 2, let j ∈ N∗ be s. t. L ∈ [2j l, 2j+1 l)
and consider a sequence C0 , . . . , Cj+1 of cubes s. t. C0 = C, Cj+1 = Q and the size of each cube
CHAPTER 8. THE SUBSTITUTE OF L∞ : BM O 44

is at most the double of the size of its predecessor. Then

l=j
X
|fC − fQ | ≤ |fl − fl+1 | ≤ c jkf kBM O ≤ c0 ln(L/l)kf kBM O . (8.10)
l=0

e) We have
Z Z Z
1 2 2k
|Ψ ◦ f − (Ψ ◦ f )C | ≤ |Ψ ◦ f − Ψ(fC )| ≤ |f − fC | ≤ 2kkf kBM O . (8.11)
|C| |C| |C|
C C C

Remark 8. The space BM O is not trivial: L∞ functions are in BM O and, if f = g + const,

then kf kBM O ≤ 2kgkL∞ . However, BM O is not reduced to L∞ functions. Here is an example: let
f : RN → R, f (x) = ln |x|. Then f ∈ BM O. Indeed, let B be a ball of radius R and center x. If
|x| ≤ 2R, then there is a ball B ∗ of radius ρ ∼ R, containing B and centered at the origin. Then
Z Z Z Z
1 2 2 c
|f − fB | ≤ |f − ln ρ| ≤ |f − ln ρ| ≤ ∗ |f − ln ρ|. (8.12)
|B| |B| |B| |B |
B B B∗ B∗

Now it is easy to see that the last integral is finite and independent of ρ.
R
Assume now that |x| > 2R. Then | ln |y| − ln |z|| ≤ c whenever y, z ∈ B, and therefore
|x|
Z Z Z
1 1 cR
|f − fB | = 2
| (f (y) − f (z))dy|dz ≤ ≤ c. (8.13)
|B| |B| |x|
B B B

We emphasize the following consequence of our above computation

Z
1
lim | ln |y| − ln |y|B(x,R) |dy = 0, ∀ R > 0. (8.14)
|x|→∞ |B(x, R)|
B(x,R)

8.2 H1 and BM O
1
Theorem 14. (Fefferman) BM O is the dual Z of H in the following sense:
a) if f ∈ BM O, then the functional T (g) = f g, initially defined on the set of finite combinations
of atoms, satisfies |T (g)| ≤ ckf kBM O and thus gives raise (by density) to a unique element of (H1 )∗
of norm ≤ ckf kBM O . Z
b) Conversely, let T ∈ (H1 )∗ . Then there is some f ∈ BM O s. t. T (g) = f g whenever g is a
finite combination of atoms. In addition, kf kBM O ≤ ckT k(H1 )∗ .
CHAPTER 8. THE SUBSTITUTE OF L∞ : BM O 45
Z
Remark 9. Since atoms are bounded and compactly supported, f g makes sense when f ∈ L1loc
and g is an atom. Moreover, the definition is correct when f ∈ BM O, in the sense that if we
replace f by f + const, then the value of the integral does not change, since atoms have zero
integral.
Z
Proof. a) Assume first that f is bounded. Then T (g) = f g is well-defined and continuous in
X
H1 , since the inclusion H1 ⊂ L1 is continuous. If g = λk ak is an atomic decomposition of g
X
s. t. |λk | ≤ ckgkH1 and each ak is supported in some Bk , then
Z Z Z
X X X 1
|T (g)| ≤ |λk || f ak | = |λk || (f − fBk )ak | ≤ |λk | |f − fBk | ≤ ckf kBM O kgkH1
|Bk |
Bk
(8.15)
and a) follows. 
n,
 if f (x) ≥ n
When f is arbitrary, we apply (8.15) to the truncated function fn (x) = f (x), if |f (x) < n .

−n, if f (x) ≤ −n
Noting that fn = Ψn ◦ f , where Ψn is Lipschitz of Lipschitz constant 1, we find that
Z
| fn g| ≤ ckf kBM O kgkH1 . (8.16)

When g is a finite combination of atoms, we have |fn g| ≤ |f g| ∈ L1 and fn g → f g a. e. Thus

Z Z
| f g| = lim | fn g| ≤ ckf kBM O kgkH1 , (8.17)

by dominated convergence. This implies a) in full generality.

b) Conversely, let T ∈ (H1 )∗ . Let B be a ball and let XB be the space of L2 functions supported in
1
B having zero integral. If g ∈ XB , then g is a 2 atom. Thus kgkH1 ≤ ckgkL2 |B|1/2 . It
kgkL2 |B|1/2
1/2
follows that T restricted to XB defines a linear continuous functionalZ of norm ≤ ckT k|B| . Thus,
there is some f B ∈ XB s. t. kf B kL2 ≤ ckT k|B|1/2 and T (g) = f B g when g ∈ XB . We now
cover RN with an increasing sequence of balls Bn and set f (x) = f Bn (x) − (f Bn )B0 if x ∈ Bn . This
B
definition is correct. Indeed, if j > k, then f Bk and f|Bjk − (f Bj )Bk yield the same functional T|XB
and thus must coincide. Therefore, f Bj and f Bk differ only by a constant in Bj (and thus in B0 ),
which implies that the definition of f is correct. Another obvious consequence of our argument is
that, on each ball B, f|B and f B differ with a constant. In other words, f B = f − fB .
CHAPTER 8. THE SUBSTITUTE OF L∞ : BM O 46
Z
We claim that, when g is a finite combination of atoms, we have T (g) = f g. Indeed, there is
Z Z
Bn
some n s. t. supp g ⊂ Bn . Since g ∈ XBn for such n, we find that T (g) = f g = f g.
It remains to prove that f ∈ BM O and that kf kBM O ≤ ckT k(H1 )∗ . This follows from the fact
that, if B is any ball, then we have
1 1 B 1
|f − fB | = |f | ≤ p kf B kL2 ≤ ckT k. (8.18)
|B| |B| |B|

8.3 BM O functions are almost bounded

Strictly speaking, the assertion in the title is not even nearly true, as shows the example x 7→ ln |x|.
However, we will see that, on compacts, BM O functions are in each Lp , p < ∞, and even better.
Proposition 13. Let f ∈ BM O. Then, for each ball B, f ∈ Lp (B) and kf − fB kLp (B) ≤
cp |B|1/p kf kBM O . In addition, we have cp ≤ c p when p ≥ 2.
Proof. We copy the proof of b) in the preceding theorem. When p = 1, the conclusion is trivial, so
that we may assume that 1 < p < ∞. We may also assume that fB = 0. Let q be the conjugate
q
exponent of p. It is straightforward Z that, if g ∈ L (B), then kg − gB kLq (B) ≤ 2kgkLq (B) . On
1
the other hand, if g ∈ Lq (B) and g = 0, then 1−1/q
g is a q atom and thus kgkH1 ≤
|B| kgkLq
B
1−1/q c
Cq |B| kgkL . Here, Cq satisfies Cq ≤
q ≤ c p when q ≤ 2 (and thus p ≥ 2). Thus
q−1
Z
kf kLp (B) = sup{ f (g − gB ) ; kgkLq (B) ≤ 1} ≤ 2c Cq kf kBM O |B|1−1/q = cp kf kBM O |B|1/p . (8.19)

Theorem 15. (John-Nirenberg) There are constants c1 , c2 > 0 s. t.

|{x ∈ B |f − fB | > α}| ≤ c1 |B| exp(−c2 α/kf kBM O ). (8.20)

Proof. It is immediate that, if the conclusion holds for f , it also holds for a multiple of f . We
may therefore assume that kf kBM O = 1. It is also clear that, if the conclusion holds for α ≥ 2c e,
where c is the constant in the preceding proposition, then we may adjust the constants s. t. (8.20)
CHAPTER 8. THE SUBSTITUTE OF L∞ : BM O 47

holds for each α. We may therefore assume that α ≥ 2c e. We assume also that fB = 0. Let
α
p= ≥ 2. Then
ce
kf kpLp (B) (cp)p |B|
|{x ∈ B ; |f | > α}| ≤ ≤ = |B| exp(−c α/e). (8.21)
αp αp

Theorem 16. (John-Nirenberg) There are constants C, k > 0 s. t. if f ∈ BM O and kf kBM O ≤

1, then Z
1
exp(C|f − fB |) ≤ k. (8.22)
|B|
B

Remark 10. The normalization condition kf kBM O ≤ 1 is necessary. Indeed, if exp f ∈ L1 , there
is no reason to have exp(2f ) ∈ L1 . On the other hand, the constant C cannot be arbitrary large,
as shown by the example x 7→ ln |x|.
Proof. We may assume that fB = 0. Then
∞ ∞
(cCp)p
Z
1 1 X p X
(exp(C|f |) − 1) = 1 + C kf kpLp (B) ≤ 1 + ≤k<∞ (8.23)
|B| |B| p=1 p=1
p!
B

provided C < (ce)−1 , as it is easily seen using Stirling’s formula.

Chapter 9

Lp regularity for the Laplace operator

9.1 Preliminaries
Let E bethe fundamental solution of the Laplace operator in RN , N ≥ 2,
 1 ln |x|,
 if N = 2
E(x) = 2π 1 . If f ∈ C0∞ , then u = E ∗ f is a (classical) solution
−
 , if N ≥ 3
(N − 2)|S N −1 ||x|N −2
of the equation (*) ∆u = f . If f ∈ Lp0 for some 1 ≤ p ≤ ∞, we may still define u = E ∗ f and we
then have u ∈ Lploc . Indeed, if K is a compact and L = supp f , let Φ ∈ C0∞ be s. t. Φ = 1 in the
compact K − L. Then, in K, we have E ∗ f = (ΦE) ∗ f , and thus
kE ∗ f kLp (K) ≤ k(ΦE) ∗ f kLp ≤ kΦEkL1 kf kLp ≤ CK,L kf kLp , (9.1)
using Young’s inequality and the fact that E ∈ L1loc .
In addition, u still satisfies (*), this time in the distribution sense. The reason is that we may
approximate f with a sequence (fn ) ⊂ C0∞ s. t. fn → f in L1 and supp fn ⊂ L0 , with L0 a compact
independent of n. Then (9.1) with p = 1 and L replaced by L0 implies that E ∗ fn → E ∗ f in L1loc ,
and thus in D0 . Since we also have ∆(E ∗ fn ) = fn → f in D0 , we find that ∆(E ∗ f ) = f .
Let now 1 ≤ j, k ≤ N and consider the operator T = Tp : Lp0 → D0 , T f = ∂j ∂k (E ∗ f ). Note that
the definition does not depend on p, in the sense that, if f ∈ Lp0 ∩ Lq0 , then Tp f = Tq f .
We start by noting some simple properties of T that will be needed in the next section.
Lemma 15. If f ∈ L20 , then  
−1 ξj ξk ˆ
Tf = F f . (9.2)
|ξ|2
Consequently, T has a continuous extension to L2 , given by the r. h. s. of (9.2).

48
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 49

In addition, T is self-adjoint in L2 , i. e.
Z Z
T f g = f T g, ∀ f, g ∈ L2 . (9.3)

Proof. The r. h. s. of (9.2) is continuous from L2 into L2 (and thus into D0 ), by Plancherel’s
theorem. On the other hand, if L is a fixed compact, the l. h. s. is continuous from L2L (the space
of L2 functions supported in L) into D0 . Therefore, it suffices to prove the equality when f ∈ C0∞ .
Since (1 + |x|2 )−N E ∈ L1 , we have E ∈ S 0 , and thus ∂j ∂k E ∈ S 0 . Therefore, Tcf = ∂\ ˆ
j ∂k E f , and
ξ ξ
j k
it suffices to prove that ∂\ j ∂k E = . We write E = E1 + E2 , where E1 = ΦE, E2 = (1 − Φ)E,
|ξ|2
Φ ∈ C0∞ and Φ = 1 near the origin. Then ∂\ \ \ ∞ 2
j ∂k E = ∂j ∂k E1 + ∂j ∂k E2 ∈ C + L , since ∂j ∂k E ∈ E ,
0

while ∂j ∂k E2 ∈ L2 . On the other hand, ∆∂j ∂k E = ∂j ∂k δ, and thus |ξ|2 ∂\ j ∂k E = ξj ξk . Thus

ξ ξ
j k
X
∂\j ∂k E = + cα ∂ α δ. The coefficients cα must be zero, since ∂\ j ∂k E ∈ C
∞
+ L2 , whence the
|ξ|2
|α|≤2
first conclusion of the lemma.
As for (9.3), it follows from Plancherel’s theorem:
Z Z Z Z Z Z
−N −N ξj ξk ˆ −N ˆξj ξk −N ˆ
T f g = (2π) T f ĝ = (2π)
c f ĝ = (2π) f 2 ĝ = (2π) f T g = f T g.
c
|ξ|2 |ξ|
(9.4)
 
1 δj,k N xj xk
Lemma 16. Assume that f ∈ Lp0 and let x 6∈ supp f . Then, with K(x) = N −1 − ,
|S | |x|N |x|N +2
we have Z
T f (x) = K(x − y)f (y)dy. (9.5)

In addition, K satisfies
C|y|
|K(x − y) − K(x)| ≤ , if |y| < 1/2|x|. (9.6)
|x|N +1
Proof. If L = supp f and O is a relatively compact open set s. t. O ∩ L = ∅, then the (pointwise)
derivatives of E(x − y)f (y) with respect to x satisfy
|∂xα (E(x − ·)f (·)| ≤ cα |f (·)| ∈ L1 , x ∈ O, (9.7)
and thus E ∗ f ∈ C ∞ (O). Moreover, we may differentiate twice under the integral sign in the
formula of E ∗ f to obtain, in O, both
Z the pointwise and the distributional derivative ∂j ∂k (E ∗ f )
through the formula ∂j ∂k (E ∗ f ) = ∂j ∂k E(x − y)f (y)dy. Here, ∂j ∂k E stands for the pointwise
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 50

derivative. Finally, we have ∂j ∂k E = K, whence the first conclusion.

To prove the inequality (9.6), we note that |DK(z)| ≤ C|z|−N −1 and thus, for x, y s. t. |y| < 1/2|x|,
we have
C|y|
|K(x − y) − K(x)| ≤ |y| sup |DK(z)| ≤ C|y| sup |z|−N −1 ≤ N +1 . (9.8)
z∈[x−y,x] z∈[x−y,x] |x|

Lemma 17. Let Φ ∈ C0∞ . Then

a) (T Φ)t = T Φt , for each t > 0.
b) T Φ ∈ L∞ .
c) D(T Φ) ∈ L∞ .
C|y|
d) If |y| < 1/2|x|, then |T Φ(x − y) − T Φ(x)| ≤ .
|x|N +1
Proof. a) Actually, this holds under the sole assumption that Φ ∈ L2 . It suffices to check that
\
(T Φ)t = Td
Φt . This equality follows from
\ [ ξj ξk ξj ξk c
(T Φ)t (ξ) = (T Φ)(tξ) = 2
Φ̂(tξ) = Φt (ξ) = Td
Φt (ξ). (9.9)
|ξ| |ξ|2
b) Since |TcΦ| ≤ |Φ̂|, we find that TcΦ ∈ L1 and thus T Φ ∈ L∞ .
\
c) Similarly, D(T \
Φ) = ıξ TcΦ, and thus D(T Φ) ∈ L1 , which implies that D(T Φ) ∈ L∞ .
d) Let R > 0 be s. t. Φ = 0 outside B(0, R). If |x| ≤ 3R, then the conclusion follows from b).
Assume |x| > 3R. Then both x and x − y are outside the support of Φ and thus
Z
C|y|
|T Φ(x−y)−T Φ(x)| = | (K(x−y−z)−K(x−z))Φ(z)dz| ≤ C sup |K(x−y−z)−K(x−z)| ≤ N +1 ;
|z|≤R |x|
(B(0,R)
(9.10)
here, we rely on the inequality (9.6) and we take into account the fact that |x − y − z| ∼ |x − z| ∼
|x|.
In the next section, we will prove the following
Theorem 17. a) (Calderón-Zygmund) For p = 1, the operator T , initially defined on L1loc , has
a continuous extension from L1 into L1w .
b) (Fefferman-Stein) When restricted to H1 , the extension of T to L1 maps continuously H1
into H1 .
c) (Calderón-Zygmund) For 1 < p < ∞, the operator T , initially defined on Lploc , has a
continuous extension from Lp into Lp .
d) (Spanne-Peetre-Stein) T maps BM O0 continuously into BM O and thus L∞ 0 continuously
into BM O.
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 51

The most widely used form of the above result sais that a solution u of ∆u = f ”gains two
derivatives with respect to f ”:
Corollary 14. Assume that ∆u = f in the distribution sense.
2,p
a) If f ∈ Lploc for some 1 < p < ∞, then u ∈ Wloc .
1 2,1
b) If f ∈ H , then u ∈ Wloc .

Proof. Let K be a compact in RN and let Φ ∈ C0∞ be s. t. Φ = 1 in an open neighborhood O

of K. Set g = Φf ∈ Lp0 and let v = E ∗ g, which satisfies ∆v = f in O. Then ∆(u − v) = 0 in
O, and thus u − v ∈ C ∞ (O), by Weyl’s lemma. Now v ∈ Lploc , since g ∈ Lp0 , and the second order
derivatives of v are in Lp if f ∈ Lploc and 1 < p < ∞, respectively in L1 if f ∈ H1 . In addition,
it is easy to see that the distributional
Z first order derivatives of E ∗ f are computed according to
the formula ∂j (E ∗ g)(x) = (∂xj E)(x − y)f (y)dy, where ∂xj E stands for the pointwise derivative
(this is obtained using an integration by parts when f ∈ C0∞ ; the general case is obtained by
approximation, with the help of Young’s inequality). Since ∂xj E ∈ L1loc and (in all the cases)
g ∈ Lp0 , we find that ∂j v ∈ Lploc . Therefore, u ∈ Wloc
2,p
.
The above results are optimal, in the following sense:
Proposition 14. T does not map L10 into L1 and does not map L∞ ∞
0 into L .

Proof. We fix a compact L in RN . We already noted that T maps continuously LpL into D0 . We
claim that, if T : LpL → Lp , then T has to be continuous. Indeed, let fn → f in LpL be s. t.
T fn → g in Lp . Since T fn → T f in D0 , we find that T f = g, and thus T has closed graph. Thus
T is continuous.
Let now p = 1. We argue by contradiction. Let L be a ball containing the origin. We consider a
sequence (fn ) ⊂ C0∞ s. t. kfn kL1 ≤ C, supp fn ⊂ L and fn → δ in D0 and set un = E ∗ fn . Then
un → E in D0 and kD2 un kL1 ≤ C. On the other hand, Dun = (DE) ∗ fn (where DE ∈ L1loc is
the pointwise derivative of E), and thus kDun kL1 (L) ≤ C. Consequently, the sequence (Dun ) is
bounded in W 2,1 (L). The Sobolev embeddings imply that (Dun ) is bounded also in LN/(N −1) (L).
Since Dun → DE in D0 , we find that DE ∈ LN/(N −1) (L); thus |DE|N/(N −1) is integrable near
the origin. However, if we compute the (pointwise or distributional) gradient DE, we see that
|DE(x)| ∼ |x|−(N −1) , a contradiction.
We next consider the case p = ∞. Argue again by contradiction. Recall that there is a function
u : RN → R, u 6∈ C 2 , s. t. f = ∆u (computed in the distributional sense) be continuous (example
due to Weierstrass). We may assume, e. g., that u 6∈ C 2 (B(0, 1)). Let g = Φf , where Φ ∈ C0∞ ,
Φ = 1 in B(0, 1), supp Φ ⊂ B(0, 2). Then g ∈ L∞ 0 , and thus T f ∈ L
∞
(for all j, k). Let (gn ) ⊂ C0∞
be s. t. gn → g uniformly, supp gn ⊂ B(0, 2). Then T gn → T g uniformly. Since T gn ∈ C ∞ , this
implies that T g is continuous. Thus E ∗ g ∈ C 2 . Since ∆(E ∗ g) = ∆u in B(0, 1), Weyl’s lemma
implies that u ∈ C 2 (B(0, 1)), a contradiction.
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 52

9.2 Proof of Theorem 17

Proof. The plan of the proof is the following: a) we prove that T maps L1 into L1w ; this will rely on
the Calderón-Zygmund decomposition. b) Marcinkiewicz’ interpolation theorem, combined with
the continuity of T from L2 into L2 , will imply the result when 1 < p < 2. c) For H1 , the result is
obtained via the atomic decomposition. d) The remaining cases, i. e. 2 < p ≤ ∞ or BM O0 , will
be obtained by duality; we will exploit the fact that T is a symmetric operator.
Step 1. Continuity from L1 into L1w
It suffices to prove the following estimate
C
|{|T f | > t}| ≤ kf kL1 , ∀ t > 0, ∀ f ∈ L1 ∩ L2 . (9.11)
t
Indeed, assume (9.11) proved, for the moment. Let f ∈ L1 and consider a sequence (fn ) ⊂ L1 ∩ L2
s. t. fn → f in L1 . Then (9.11) applied to fn − fm implies that |{|T fn − T fm | > t}| → 0 when t is
fixed and m, n → ∞.Thus (T fn ) is a Cauchy sequence in measure, and thus it converges in measure
to some g. In particular, this means that g does not depend on the sequence (fn ), that g = T f
if f happens to be in L1 ∩ L2 and that f 7→ g is linear. Possibly after passing to a subsequence
C
(fnk ), we have T fnk → g a. e., and thus |{|g| > t}| ≤ lim inf k |{|T fnk | > t}| ≤ kf kL1 . This
t
implies that f 7→ g is the desired extension of T . (We needed this argument since L1w is not a
normed space.)
We return to the proof of (9.11). Let t > X 0. We write, as in Theorem 9 (with α replaced
1 2
by t), a function f ∈ L ∩ L as f = g + hn . We first note that g ∈ L2 , since g ∈ L1
X
and |g| ≤ Ct. We also claim that the series hn is convergent in L2 . Indeed, the functions
X
hn are mutually orthogonal in L2 , and thus it suffices to prove that khn k2L2 < ∞. Since
X X
khn kL2 = kf − fCn kL2 (Cn ) ≤ kf kL2 (Cn ) , we find that khn k2L2 ≤ kf k2L2 (Cn ) ≤ kf k2L2 , whence
the claim. X X
This allows us to write T f = T (g + hn ) = T g + T hn . Thus
X
|{|T f | > t}| ≤ |{|T g| > t/2}| + |{|T hn | > t/2}|. (9.12)
Z Z
2
On the one hand, we have |g| ≤ Ct |g| ≤ Ctkf kL1 , by the properties of the Calderón-
Zygmund decomposition. Thus
C C
|{|T g| > t/2}| ≤ 2
kgk2L2 ≤ kf kL1 . (9.13)
(t/2) t
On the other hand, let, for each n, Cn∗ be the cube concentric with Cn and twice bigger than it.
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 53

Then, with A = RN \ Cn∗ , we have

S

X [ X X CX
|{|T hn | > t/2}| ≤ | Cn∗ | + |{x ∈ A ; |T hn | > t/2}| ≤ C |Cn | + kT hn kL1 (A) .
t
(9.14)
We denote by xn the center of Cn and by ln its size. For x ∈ A, we have
Z Z
T hn (x) = K(x − y)hn (y)dy = [K(x − y) − K(x − xn )]hn (y)dy, (9.15)

and thus Z
C Cln
|T hn (x)| ≤ N +1
|y − xn ||hn |dy ≤ khn kL1 . (9.16)
|x − xn | |x − xn |N +1
Integrating the above inequality and summing over n, we find that
X X
kT hn kL1 (A) ≤ C khn kL1 ≤ Ckf kL1 , (9.17)
by the properties of the Calderón-Zygmund decomposition. X
We conclude the first step by combining (9.12), (9.13), (9.14), (9.17) and the fact that |Cn | ≤
C
kf kL1 .
t
Step 2. Continuity in Lp , 1 < p < ∞
We know that T , when defined in L1 ∩ L2 , is continuous from L1 into L1w and from L2 into L2 .
Marcinkiewicz’ interpolation theorem implies that T has a unique extension continuous from Lp
into Lp when 1 < p < 2. Let now 2 < p < ∞. Part c) of the theorem follows if we prove that
kT f kLp ≤ Ckf kLp whenever f ∈ Lp ∩ L2 . For such an f , we have, with q < 2 the conjugate
exponent of p,
Z Z Z
kT f kLp = sup Tf g = sup Tf g = sup f T g ≤ Ckf kLp ;
g∈Lq ; kgkLq ≤1 g∈Lq ∩L2 ; kgkLq ≤1 g∈Lq ∩L2 ; kgkLq ≤1
(9.18)
here, we use the continuity of T in Lq .
Step 3. Continuity in H1
In view of the properties of the atomic decomposition, it suffices to prove, with a constant C
independent of a, the estimate
kT akH1 ≤ C, ∀ atom a. (9.19)
Z
Let a be an atom supported in B = B(x, R). Let Φ ∈ C0∞ be s. t. Φ = 1 and supp Φ ⊂ B(0, 1).
For each x, we have MΦ a(x) ≤ CMa(x), and thus
Z Z
MΦ a ≤ C Ma ≤ kMakL2 |B(x, 2R)|1/2 ≤ CkakL2 |B|1/2 ≤ C. (9.20)
B(x,2R) B(x,2R)
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 54

We consider now an x outside B(x, 2R) and estimate Φt ∗ (T a)(x). We have (we take a, Φ real,
here)
Z Z Z
Φt ∗(T a)(x) = Φt (x−y)T a(y)dy = (T Φt )(x−y)a(y)dy = [(T Φ)t (x−y)−(T Φ)t (x−x)]a(y)dy.
B
(9.21)
We next note that, when y ∈ B, we have |x − y| < 1/2|x − x|. We intend to make use of the
decay properties of T Φ. To this purpose, we distinguish two possibilities concerning the size of t:
(i) t > |x − x| and (ii) t ≤ |x − x|. In case (i), we use the fact that T Φ is Lipschitz, and find that

|(T Φ)t (x − y) − (T Φ)t (x − x)| ≤ Ct−N −1 |y − x|, (9.22)

and thus Z
C Cl Cl
|Φt ∗ (T a)(x)| ≤ |y − x||a(y)|dy ≤ ≤ . (9.23)
tN +1 tN +1 |x − x|N +1
B

In case (ii), we make use of Lemma 17 d), and obtain

C
|(T Φ)t (x − y) − (T Φ)t (x − x)| ≤ |y − x|, (9.24)
|x − x|N +1
which gives Z
C Cl
|Φt ∗ (T a)(x)| ≤ |y − x||a(y)|dy ≤ . (9.25)
|x − x|N +1 |x − x|N +1
B

(9.23) combined with (9.25) yields

Cl
MΦ a(x) ≤ when x 6∈ B(x, 2R). (9.26)
|x − x|N +1
Integration of (9.26) over RN \ B(x, 2R) combined with (9.20) gives the needed conclusion
kMΦ akL1 ≤ C.
Step 4. Continuity of T in BM O0
We note that BM O0 ⊂ L20 , by the John-Nirenberg inequalities. We also note that the vector space
V spanned by the atoms is contained in L2 . Thus, for f ∈ BM O0 , we have
Z Z
kT f kBM O ∼ sup Tf g = sup f T g ≤ Ckf kBM O ; (9.27)
g∈V,kgkH1 ≤1 g∈V,kgkH1 ≤1

here, we used the duality between H1 and BM O, the density of V in H1 and the continuity of T
from H1 into H1 .
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 55

9.3 An equation involving the jacobian

We consider, in R2 , the following equation that appears in Geometry

∆u = det(Df, Dg), f, g ∈ H 1 (R2 ). (9.28)

Theorem 18. a) (Wente) Equation (9.28) has one and only one distribution solution u ∈ C(R2 )
vanishing at infinity, i. e., s. t. lim u(x) = 0. In addition, Du ∈ L2 and
|x|→∞

1/2 1/2
kDukL2 ≤ CkDf kL2 kDgkL2 . (9.29)
b) (Coifman-Lions-Meyer-Semmes) In addition, we have D2 u ∈ H1 .
Proof. The main argument in the proof is that
det(Df, Dg) ∈ H1 and k det(Df, Dg)kH1 ≤ CkDf kL2 kDgkL2 . (9.30)

Assuming (9.30) proved for the moment, we reason as follows: let h = det(Df, Dg). Consider
sequences (fn ), (gn ) ⊂ C0∞ s. t. fn → f , gn → g in H 1 . Then hn = det(Dfn , Dgn ) → h in H1 , by
(9.30). Let un = E ∗ hn , which is a solution of ∆un = hn . We claim that un ∈ C∞ (the space of
continuous functions vanishing at infinity) and that (un ) is a Cauchy sequence for the sup norm.
Indeed, let, for fixed n, R = Rn > 0 be s. t. hn (y) = 0 if |y| > R. Then
Z Z
1 1
|un (x)| = | ln |x − y|hn (y)dy| = | [ln |x − y| − (ln | · |)B(x,R) ]hn (y)dy|, (9.31)
2π 2π
and thus
Z Z
|un (x)| ≤ Cn | ln |x − y| − (ln | · |)B(x,R) |dy = Cn | ln y − (ln | · |)B(x,R) |dy → 0 as |x| → ∞.
B(0,R) B(x,R)
(9.32)
1
On the other hand, we have ln ∈ BM O and thus, using the H -BM O duality,
Z
1
|un (x)−um (x)| = | ln |y|(hn (x−y)−hm (x−y))dy| ≤ Ck(hn −hm )(x−·)kH1 = Ckhn −hm kH1 ,
2π
(9.33)
1
since the H norm is translation invariant. (Similarly, we have |un | ≤ Ckhn kH1 .)
To summarize, the sequence (un ) is Cauchy in C∞ , and thus converges to some u ∈ C∞ . This u
is a distribution solution of (9.28). It is also the only solution of (9.28) in C∞ , for if v is another
solution, their difference w is, by Weyl’s lemma, a harmonic function vanishing at infinity, thus
constant, by the maximum principle.
We now turn to the proof of b) and c) (assuming, again, (9.30) already proved). Note that c),
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 56

at least when f, g ∈ C0∞ , follows by combining (9.30) and the Fefferman-Stein regularity result
concerning the equation ∆u = f with f ∈ H1 . The general case is obtained by approximation,
as above. Similarly, it suffices to establish b) when f, g ∈ C0∞ . Formally, estimate b) is clear, as
shown by the following (wrong, in principle) computation:
Z Z Z
|Du| = − u∆u = − uh ≤ kukL∞ khkL1 ≤ CkukL∞ khkH1 ≤ CkDf k2L2 kDgk2L2 . (9.34)
2

The point
Z is that this computation can be transformed into
Z a rigorous one as follows: set F (r) =
1 1
|u|2 dl. Then lim F (r) = 0 and F 0 (r) = u · ur dl. Thus, along a subsequence
2πr r→∞ πr
|x|=r |x|=r
0
rn → ∞, we must have rn F (rn ) → 0 (argue by contradiction; otherwise, we have F (r) ≥ C ln r
for large r). Then, for large n, we have ∆u = 0 outside B(0, rn ) and thus
Z Z Z Z Z
2 2
|Du| = lim |Du| = lim{ u · ur − u∆u} = − u∆u ≤ CkDf k2L2 kDgk2L2 . (9.35)
n n
B(0,rn ) |x|=r

The only part of the proof left open is

Proof of (9.30)
In view of the conclusion we want
Z to obtain, we may assume that f, g ∈ C0∞ . Let Φ ∈ C0∞ be
supported in B(0, 1) and s. t. Φ = 1. Then, with h = det(Df, Dg), we have
Z Z
−1
Φt ∗ h(x) = Φt (y)h(x − y)dy = t f (x − y) det((DΦ)t )(y), Dg(x − y))dy, (9.36)
Z
as shown by an integration by parts. Next, if k, l ∈ C0∞ , then det(Dk, Dl) = (again, this follows
by an integration by parts), and thus
Z
−1
Φt ∗ h(x) = t [f (x − y) − fB(x,t) ] det((DΦ)t )(y), Dg(x − y))dy. (9.37)

Using the Hölder inequality and the inequality |(DΦ)t | ≤ Ct−2 together with the fact that Φt
vanishes outside B(0, t), we find that
 Z 1/4  Z 3/4
−3 4 4/3
|Φt ∗ h(x)| ≤ t |f − fB(x,t) | |Dg| . (9.38)
B(x,t) B(x,t)

Applying Lemma 18 below to the function given by v(y) = f (x − ty), we find that
kf − fB(x,t) kL4 (B(x,t) ≤ CkDf kL4/3 (B(x,t) , (9.39)
CHAPTER 9. LP REGULARITY FOR THE LAPLACE OPERATOR 57

and thus (9.38) yields

 Z 3/4  Z 3/4
1 4/3 1 4/3
|Φt ∗ h(x)| ≤ C |Df | |Dg| . (9.40)
|B(x, t)| |B(x, t)|
B(x,t) B(x,t)

Recalling the definition of the maximal function, we obtain

MΦ h(x) ≤ C(M|Df |4/3 (x))3/4 (M|Dg|4/3 (x))3/4 , (9.41)

and the Cauchy-Schwarz inequality implies that
Z 1/2  Z 1/2
4/3 3/2 4/3 3/2
kMΦ hkL1 ≤ C (M|Df | ) (M|Dg| ) , (9.42)

which may be rewritten as

3/4 3/4 3/4 3/4
kMΦ hkL1 ≤ CkM|Df |4/3 kL3/2 kM|Dg|4/3 kL3/2 ≤ Ck|Df |4/3 kL3/2 k|Df |4/3 kL3/2 = CkDf kL2 kDgkL2 ;
(9.43)
that is, khkH1 ≤ CkDf kL2 kDgkL2 , as claimed at the beginning of the proof.
We next recall the following Sobolev embedding and the corresponding Poincaré inequality
Lemma 18. W 1,4/3 (R2 ) is embedded into L4 and, with B = B(0, 1) and v ∈ W 1,4/3 (B), we have

kv − vB kL4 (B) ≤ CkDvkL4/3 (B) . (9.44)

Proof. The above Sobolev embedding will be proved, in a slightly better form, in the next chapter.
We present a proof of (9.44), which is less standard. The starting point is the usual Poincaré
inequality
kv − vB kL4/3 (B) ≤ CkDvkL4/3 (B) . (9.45)
We may assume, with no loss of ( generality, that vB = 0. We extend v by reflections in a neigh-
v(x), if x ∈ B
borhood of B by setting ṽ(x) = 2
. The new function ṽ is in W 1,4/3
v(x/|x| ), if 1 < |x| < 3/2
and satisfies kṽkL4/3 ≤ CkvkL4/3 and kDṽkL4/3 ≤ CkDvkL4/3 . Next let Φ ∈ C0∞ be s. t. Φ = 1 in
B and supp Φ ⊂ B(0, 3/2). Set w = Φṽ ∈ W 1,4/3 (R2 ). Then

kvkL4 (B) ≤ kwkL4 ≤ CkDwkL4/3 ≤ C(kṽkL4/3 + kDṽkL4/3 ) ≤ C(kvkL4/3 + kDvkL4/3 ) ≤ CkDvkL4/3 .

(9.46)
 Part III

Functions in Sobolev spaces

58
Chapter 10

Improved Sobolev embeddings

The usual form of the Sobolev embeddings states that W 1,p (RN ) ⊂ LN p/(N −p) , provided 1 ≤ p < N .
In this chapter, we will improve the conclusion to W 1,p (RN ) ⊂ LN p/(N −p),p (when 1 < p < N );
NP
this is slightly better, since > p, and thus LN p/(N −p),p ⊂ LN p/(N −p) .
N −p

10.1 An equivalent norm in Lorentz spaces

Let f : RN → C be a measurable function and let F : (0, ∞) → [0, ∞] be its distribution function.
Intuitively, we may think of F as a bijection of (0, ∞) into itself. Then, if p, q < ∞ and if f ∗ = F −1
(which is decreasing), we may (formally) compute the Lp,q quasi-norm as follows:
Z Z Z
q ∗ q−1 ∗0 −1
q−1 q/p q/p
kf kLp,q = t F (t)dt = − s (f ) (s)f (s)ds = p sq/p−1 (f ∗ )q (s)ds; (10.1)

here, the - sign at the beginning of the computation comes from the fact that F is decreasing. The
second equality is obtained through the change of variable F (t) = s, the third one arises after an
integration by parts.
The above equality maybe rewritten as
kf kLp,q = kt1/p F kLq ((0,∞);dt/t) ∼ kt1/p f ∗ kLq ((0,∞);dt/t) . (10.2)
In this section, we will see that this formula is right!...provided we interpret it accurately.
Definition 4. The non increasing rearrangement f ∗ : [0, ∞) → [0, ∞] of f is defined through
the formula
f ∗ (t) = sup{s > 0 ; F (s) ≤ t}. (10.3)
We note that, when F is a bijection, we have f ∗ = F −1 .
The elementary results we gather below explain, in particular, why f ∗ is called the non increasing
rearrangement of f .

59
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 60

Proposition 15. a) F is continuous from the right.

b) F (f ∗ (t)) ≤ t everywhere (in other words, inf = min in the definition of f ∗ ).
c) f ∗ is non increasing and continuous from the right.
d) f and f ∗ are equally distributed, i. e., |{x ∈ RN ; |f (x)| > t}| = |{s ∈ (0, ∞) ; f ∗ (s) > t}| for
each t > 0.
e) f ∗ depends continuously on f in the following sense: if (fn ) is a sequence of functions s. t.
|fn (x)| % |f (x)| for a. e. x ∈ RN , then fn∗ (t) % f ∗ (t) for each t >0.
X ∗ X X
f ) We have (f + g)∗ (2t) ≤ f ∗ (t) + g ∗ (t). More generally, fj ( tj ) ≤ (fj )∗ (tj ).
[
Proof. a) follows from the equality {|f | > t} = {|f | > t + 1/n}, which implies that F (t) =
lim F (t + 1/n).
b) Let s = f ∗ (t). Then F (s + ε) ≤ t for ε > 0, and thus F (s) ≤ t.
c) The fact that f ∗ is non increasing is clear from the definition. Concerning the second assertion,
it suffices to prove that f ∗ (t + 0) ≥ f ∗ (t). Let tn & t. Then F (f ∗ (tn )) ≤ tn , which implies that
F (f ∗ (t + 0)) ≤ tn and thus F (f ∗ (t + 0)) ≤ t, that is f ∗ (t + 0) ≥ f ∗ (t).
d) Since f ∗ is non increasing, we have |{s ∈ (0, ∞) ; f ∗ (s) > t}| = τ , where τ is uniquely defined
by f ∗ (s) > t if s < τ and f ∗ (s) ≤ t if s > τ . In view of the conclusion we want, it suffices to check
that f ∗ (s) > t if s < F (t) and that f ∗ (s) ≤ t if s > F (t). If s < F (t), then F (f ∗ (s)) ≤ s < F (t)
and thus f ∗ (s) > t. On the other hand, if s > F (t), then t ≥ f ∗ (s), by definition of f ∗ (s).
e) We note that |f | ≤ |g| =⇒ f ∗ ≤ g ∗ ; therefore, the sequence (fn∗ ) is non decreasing and
h(t) := lim fn∗ (t) ≤ f ∗ (t) for each t. Hence, it suffices to prove that h(t) ≥ f ∗ (t), i. e., that
F (h(t)) ≤ t. We note that, for each s, we have Fn (s) → F (s), since the set {|f | > s} is the union
of the non decreasing sequence ({|fn | > s}). Thus Fn (fn∗ (t)) ≤ t =⇒ Fn (h(t)) ≤ t =⇒ F (h(t)) ≤ t,
as needed.
f) Let s = f ∗ (t) and τ = g ∗ (t). Then |{|f | > s}| ≤ t and |{|g| > τ }| ≤ t. Since {|f + g| >
s + τ } ⊂ {|f | > s} ∪ {|g| > τ }, we find that |{|f + g| > s + τ }| ≤ 2t, i. e., (f + g)∗ (2t) ≤ s + τ =
f ∗ (t) + g ∗ (t).
We next justify the equality (10.1).
Proposition 16. For 1 ≤ p < ∞ and 1 ≤ q ≤ ∞, we have kf kLp,q ∼ kt1/p f ∗ kLq ((0,∞);dt/t) . For
p = ∞, we have kf kL∞ = kf ∗ kL∞ .
Proof. We start with the case p = ∞; we will prove the equality of the quasi-norms. Indeed,
kf kL∞,q = kf kL∞ = inf{s ; F (s) = 0} = inf{s ; F (s) ≤ 0} = f ∗ (0) = kf ∗ kL∞ , (10.4)
since f ∗ is non increasing and continuous from the right.
Let now p < ∞ and q = ∞; once again, we will prove the equality of the quasi-norms.
” ≤ ” Let C = kf kLp,∞ = sup tF 1/p (t). Let t > 0. With s = f ∗ (t), we want to prove that
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 61

t1/p s ≤ C. If s = 0, there is nothing to prove. If s = ∞, then F (τ ) > t for each τ , and then
C = ∞. If s ∈ (0, ∞, then F (s − ε) > t for small ε > 0, and thus
Cs
t1/p s < f 1/p (s − ε)s ≤ , (10.5)
s−ε
and the desired conclusion follows by letting ε → 0.
” ≥ ” With C = sup t1/p f ∗ (t), we will prove that tF 1/p (t) ≤ C for each t > 0. If F (t) = 0,
there is nothing to prove. If F (t) = ∞, then f ∗ (s) ≥ t for each s, and thus C = ∞. Finally, if
u = F (t) ∈ (0, ∞), let, for small ε > 0, uε = u − ε > 0. Then F (t) > uε and thus f ∗ (uε ) > t. We
find that
tF 1/p (t) ≤ f ∗ (u − ε)u1/p , (10.6)
and we conclude by letting ε → 0.
Finally, we consider the case 1 ≤ p, q < ∞. In view of the preceding proposition, it suffices to
prove the equality pkf kqLp,q = kt1/p f ∗ kqLq ((0,∞);dt/t) when f is a step function; the general case will
follow by monotone convergence, by approximating an arbitrary function f with a sequence (fn )
s. t. each fn is a step function and |fn | % |f |. In addition, since the quantitiesPwe consider do
not distinguish between f and |f |, we may assume that f ≥ 0. Let then f = an χAn , where
a1 > a2 > . . . > ak > 0 and the sets An are measurable and mutually disjoint. Set bn = |An |,
cl = b1 + . . . + bl , c0 = 0 and ck+1 = ∞. Then, with a0 = ∞ and ak+1 = 0, we have F (t) = cl if
t ∈ [al+1 , al ). On the other hand, f ∗ (t) = al+1 if t ∈ [cl , cl+1 ). Then
k Z k
X p X q/p
pkf kqLp,q =p t q−1 q/p
(cl ) dt = (cl ) [(al )q − (al+1 )q ] (10.7)
l=0
q l=1
[al+1 ,al )

and
k Z k−1
X pX
kt1/p f ∗ kqLq ((0,∞);dt/t) = q/p−1
t (al+1 ) = q
(al+1 )q [(cl+1 )q/p − (cl )q/p ], (10.8)
l=0
q l=0
[cl ,cl+1 )

so that the two quantities are equal (since c0 = 0 and ak+1 = 0).

10.2 Properties of f ∗
Lemma 19. For each t > 0 we have (with F the distribution function of f )
F (f ∗
Z (t)) Z∞
∗ ∗ ∗
f (s)ds = F (f (t))f (t) + F (s)ds. (10.9)
0 f ∗ (t)
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 62

Z∞ Zt
In particular, F (s)ds ≤ f ∗ (s)ds.
f ∗ (t) 0
(
|f (x)|, if |f (x)| > f ∗ (t)
Proof. Let g(x) = , whose distribution function G is given by
0, if |f (x) ≤ f ∗ (t)
(
F (s), if s ≥ f ∗ (t)
. Let τ ≥ F (f ∗ (t)). Then G(s) ≤ τ for each s and thus g ∗ (τ ) = 0. On
F (f ∗ (t)), if s < f ∗ (t)
the other
Z hand, ifZτ < F (f ∗ (t)), then clearly g ∗ (τ ) = f ∗ (τ ). The equality kgkL1 = kg ∗ kL1 reads
then G(s)ds = g ∗ (s)ds, which is precisely the desired equality.

Although it is actually part of the preceding proof, we emphasize for later use the following
(
f (x), if |f (x)| > α
Corollary 15. Let, for α > 0, fα (x) = . Then
0, otherwise
Zt
kff ∗ (t) kL1 ≤ f ∗ (s)ds. (10.10)
0

Lemma 20. Let F be the distribution function of f . Then

fZ∗ (t) Z
F (s) Zt Z∞
∗ ∗ ∗
g (u)du ds ≤ f (t) g (s)ds + f ∗ (u)g ∗ (u)du. (10.11)
0 0 0 t

Proof. Let I be the l. h. s. of (10.11). Fubini’s theorem implies that

Z∞
I = g ∗ (u)|{s ; s < f ∗ (t) and u < F (s)}|du. (10.12)
0
∗
Note that u < F (s) =⇒ f (u) > s, and therefore
(
∗ ∗ ∗ (0, f ∗ (t)), if u ≤ t
{s ; s < f (t) and u < F (s)} ⊂ (0, min(f (u), f (t))) = . (10.13)
(0, f ∗ (u)), if u > t
Thus
Zt Z∞
∗ ∗
I≤ g (u)f (t)du + f ∗ (u)g ∗ (u)du, (10.14)
0 t
whence the result.
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 63

Lemma 21. Let A ⊂ RN be a measurable set of measure t. Then

Z Zt
|f | ≤ f ∗ (s)ds. (10.15)
A 0

Proof. We may replace f by f χA (assuming thus f supported in A), since in this way the l. h. s.
of (10.15) remains unchanged, while the r. h. s. is not increased. In this case, we have F (s) ≤ t
for each s, and thus g ∗ (s) = 0 if s ≥ t. Therefore,
Z Z∞ Zt
|f | = kf kL1 = kf ∗ kL1 = f ∗ (s)ds = f ∗ (s)ds. (10.16)
A 0 0

10.3 Rearrangement and convolutions

The reason we considered f ∗ is that it is related convolution products. We start with some
elementary, though tricky, results linking these objects.
Lemma 22. Let f be s. t. |f | ≤ α and f = 0 outside a set E s. t. |E| = t. Let h = f ∗ g. Then
Zt
|f ∗ g| ≤ α g ∗ (s)ds. (10.17)
0

Proof. We have
Z Z Zt
|f ∗ g(x)| ≤ |f (y)||g(x − y)|dy ≤ α |g| ≤ α g ∗ (s)ds, (10.18)
E x−E 0

since |x − E| = t.
Lemma 23. Let f ∈ L∞ and set α = kf kL∞ . Then
F (t)
Zα Z
|f ∗ g| ≤ g ∗ (u)du dt. (10.19)
0 0
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 64

Proof. Using a monotone convergence argument, we may assume that f is a step function. Each
Xk
step function may be written as f = aj χAj , where aj > 0 and 0 < |Ak | < . . . < |A1 | < ∞.
j=1
With bj = a1 + ... + aj , we have f = bj in Aj \ Aj+1 . We set b0 = 0 and A0 = RN . Note that
α = bk . X
Since |f ∗ g| ≤ aj χAj ∗ |g|, the preceding lemma implies that
j

X Z|Aj |
|f ∗ g| ≤ aj g ∗ (t)dt. (10.20)
0

On the other hand, we have F (t) = |Aj | if t ∈ [bj−1 , bj ) and thus

F (t)
Zα Z bj |Aj | Z|Aj | |Aj |
XZ Z X X Z
g ∗ (u)du dt = g ∗ (u)du dt = (bj −bj−1 ) g ∗ (u)du = aj g ∗ (t)dt. (10.21)
0 0 bj−1 0 0 0

Lemma 24. (O’Neil) Let h = f ∗ g. Then

Zt Zt Z∞
∗ 3 ∗ ∗
h (3t) ≤ f (s)ds g (s)ds + f ∗ (s)g ∗ (s)ds. (10.22)
t
0 0 t

Proof. We may assume that f, g ≥ 0. We split f = f1 + f2 , g = g1 + g2 and h = h1 + h2 + h3 .

Here, (
f (x), if f (x) > f ∗ (t)
(i) f is cut at height f ∗ (t), i. e., we set f1 (x) = and f2 = f − f1 ;
0, if f (x) ≤ f ∗ (t)
(ii) similarly, g is cut at height g ∗ (t);
(iii) h1 = f2 ∗ g, h2 = f1 ∗ g2 and h3 = f1 ∗ g1 .
We start by noting that f2 ≤ f , and thus the distribution function of f2 is dominated by the one
of f . Lemma 23 implies that
fZ∗ (t) Z
F (s) Zt Z∞
h1 ≤ g ∗ (u)du ds ≤ f ∗ (t) g ∗ (s)ds + f ∗ (u)g ∗ (u)du, (10.23)
0 0 0 t
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 65

by Lemma 10.11.
Concerning h2 , the inequality kh2 kL∞ ≤ kf1 kL1 kg2 kL∞ combined with Corollary 15 yields
Zt
h2 ≤ g ∗ (t) f ∗ (s)ds. (10.24)
0

We next note that h3 satisfies

Zt Zt
kh3 kL1 ≤ kf1 kL1 kg1 kL1 ≤ f ∗ (s)ds g ∗ (s)ds. (10.25)
0 0

To conclude, we start with the inequality h∗ (3t) ≤ (h1 )∗ (t) + (h2 )∗ (t) + (h3 )∗ (t). For h1 and h2 ,
we use the fact that k ∗ (t) ≤ kk ∗ kL∞ = kkkL∞ . For h3 , we rely on the inequality
Zt Z∞
1 1 1 1
k ∗ (t) ≤ k ∗ (s)ds ≤ k ∗ (s)ds = kk ∗ kL1 = kkkL1 . (10.26)
t t t t
0 0

We find that
Zt Zt Zt Zt Z∞
∗ 1 ∗ ∗ ∗ ∗ ∗ ∗
h (3t) ≤ f (s)ds g (s)ds + f (t) g (s)ds + g (t) f (s)ds + f ∗ (s)g ∗ (s)ds; (10.27)
t
0 0 0 0 t

Zt
1 ∗
we complete the proof noting that f (t) ≤ f ∗ (s)ds and a similar inequality holds for g.
t
0

1 1 1
Theorem 19. (O’Neil; simplified version) Let 1 < p, q, r < ∞ be s. t. + = 1 + . If f ∈ Lp
p q r
and g ∈ Lq,w , then f ∗ g ∈ Lr,p .
Remark 11. This statement is to be compared with the usual Young inequality, which asserts that
f ∗ g ∈ Lr if f ∈ Lp and g ∈ Lq . Our hypothesis is weaker, since Lq ⊂ Lq,w , while the conclusion
is stronger, since Lr,p ⊂ Lr (because p < r).
Proof. Let h = f ∗ g. We have to prove that kt1/r h∗ (t)kLp ((0,∞);dt/t) < ∞. Clearly, this is equivalent
to proving that kt1/r h∗ (3t)kLp ((0,∞);dt/t) < ∞. In view of the preceding lemma, this amounts to
proving the following:
Zt Zt
(i) kt1/r−1 f ∗ (s)ds g ∗ (s)dskLp ((0,∞);dt/t) < ∞;
0 0
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 66

Z∞
(ii) kt1/r
f ∗ (s)g ∗ (s)dskLp ((0,∞);dt/t) < ∞.
t
The fact that g ∈ Lq,w is equivalent to the boundedness of the map t 7→ t1/q g ∗ (t), and thus
Zt
g ∗ (t) ≤ Ct−1/q . It follows that g ∗ (s)ds ≤ C 0 t1−1/q , and therefore (i) and (ii) reduce to
0
Zt
(i’) kt1/r−1/q f ∗ (s)dskLp ((0,∞);dt/t) < ∞;
0
Z∞
(ii) kt1/r s−1/q f ∗ (s)dskLp ((0,∞);dt/t) < ∞.
t
To deal with (i’), we apply to f ∗ the first Hardy’s inequality (Theorem 3) with r replaced by
p − 1 > 0 and find that
Zt Z∞  Zt p Z∞
kt1/r−1/q f ∗ (s)dskpLp ((0,∞);dt/t) = t−p f ∗ (s)ds dt ≤ C (f ∗ (s))p ds = Ckf ∗ kpLp < ∞,
0 0 0 0
(10.28)
∗
since kf k = kf k .
Lp Lp
Concerning (ii’), the second Hardy’s inequality (Corollary 3) with r replaced by p/r and f replaced
by s 7→ s−1/q f ∗ (s) in order to obtain
Z∞ Z∞  Z∞ p Z∞
kt1/r s−1/q f ∗ (s)dskpLp ((0,∞);dt/t) = tp/r−1 s−1/q f ∗ (s)ds dt ≤ C (f ∗ (s))p ds < ∞.
t 0 t 0
(10.29)

Corollary 16. Set, with p, q, r as in O’Neil’s theorem, a = N/q. If f ∈ Lp , then f ∗ |x|−a ∈ Lr,p .
Proof. It suffices to prove that |x|−a ∈ Lq,w . This is obvious, since |{|x|−a > t}| = Ct−N/a =
Ct−q .

10.4 Improved Sobolev embeddings

In the remaining part of this chapter, we assume that N ≥ 2.
We start with a simple
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 67

Lemma 25. Let u ∈ C0∞ (RN ). Then

|Du(y)|
Z
1
|u(x)| ≤ dy. (10.30)
|S N −1 | |x − y|N −1
Proof. Let v ∈ S N −1 . Then
Z∞ Z∞
d
u(x) = −[u(x + tv)]t=∞
t=0 = − (u(x + tv))dt = − (Du)(x + tv) · vdt, (10.31)
dt
0 0

and therefore
Z∞
|u(x)| ≤ |Du(x + tv)|dt. (10.32)
0
N −1
Integrating this inequality over v ∈ S we find that
Z Z∞
N −1
|S ||u(x)| ≤ |Du(x + tv)|dtdsv . (10.33)
S N −1 0

We conclude by noting that the change of variables y = x + tv, t > 0, v ∈ S N −1 yields

Z∞
|Du(y)|
Z Z
dy = |Du(x + tv)|dtdsv . (10.34)
|x − y|N −1
S N −1 0

We next recall the following well-known result

Theorem 20. (converse to the dominated convergence) Let 1 ≤ p ≤ ∞. If fn → f in Lp ,
then there is a subsequence (fnk ) and a function g ∈ Lp s. t. fnk → f a. e. and |fnk | ≤ g.
Proof. After passing, if necessary, to a subsequence, we may assumeX that fn → f a. e. Consider
−k
a subsequence (fnk ) s. t. kfnk − fnk+1 kLp ≤ 2 and set g = |fn0 | + |fnk − fnk+1 |. Then
k≥0
X
kgkLp ≤ kfn0 kLp + kfnk − fnk+1 kLp < ∞ (10.35)
k≥0

and, clearly, |fnk | ≤ g for each k.

CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 68

Np
Theorem 21. (O’Neil) Let 1 < p < N and set p∗ = . If u ∈ W 1,p (RN ), then u ∈
N −p
∗
Lp ,p (RN ).
Proof. The strategy consists in proving the following generalization of (10.30)
|Du(y)|
Z
1
|u(x)| ≤ N −1 dy, ∀ u ∈ W 1,p (RN ). (10.36)
|S | |x − y|N −1
Assume (10.36) proved, for the moment. Corollary 16 avec a = N −1 implies that |Du|∗|x|−(N −1) ∈
∗ ∗
Lp ,p . Since |u| ≤ C|Du| ∗ |x|−(N −1) a. e., we obtain that u ∈ Lp ,p .
It remains to prove (10.36). This is done by approximation. Consider a sequence (un ) ⊂ C0∞ s.
t. un → u in W 1,p . Possibly after passing to a subsequence, we may assume that un → u and
p −(N −1) ∗
Dun → Z Du outside a null set A and that |Dun | ≤ g ∈ L . Since g ∗ |x| ∈ Lp ,p ⊂ Lp , we find
g(y)
that dy < ∞ for x outside a null set B. When x 6∈ A ∪ B, we find, by dominated
|x − y|N −1
convergence, that
|Dun (y)| |Du(y)|
Z Z
1 1
|u(x)| = lim |un (x)| ≤ lim inf N −1 −1
dy = N −1 dy. (10.37)
|S | |x − y|N |S | |x − y|N −1
This completes the proof of the theorem.

10.5 The limiting case p = 1

When p = 1, Theorem 19 is no longer true. To see this, we choose f = χB ∈ L1 (here,Z B is the
1
unit ball) and g(x) = |x|−α , which belongs to Lq,w if αq = N . We have f ∗ g(x) = dy.
|x − y|α
B
If |x| ≥ 2, we have |x − y| ∼ |x| when |y| ≤ 1 and thus f ∗ g(x) ∼ |x|−α when |x| ≥ 2. Therefore,
Z
1
kf ∗ gkLq,1 ≥ kf ∗ gkLq ≥ C = ∞. (10.38)
|x|N
{|x|≥2}

A remarkable fact is that the conclusion of Theorem 21 still holds; the proof requires an argument
that does not involves convolution products. We start with one essential ingredient which is the
isoperimetric inequality. We will not need the sharp (i. e., with the best constant) version, so that
we will simply prove the following
Theorem 22. (weak form of the isoperimetric inequality) Let O be a smooth bounded domain in
RN and let Σ be its boundary. Then
|O| ≤ C|Σ|N/(N −1) . (10.39)
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 69
Z
Proof. Let ρ ∈ C0∞ be s. t. ρ ≥ 0, ρ = 1 and supp ρ ⊂ B(0, 1). We apply the Sobolev inequality
kukLN/(N −1) ≤ CkDukL1 to the function u = χO ∗ ρε and find that
XZ Z XZ Z
kχO ∗ ρε kLN/(N −1) ≤ C | ∂j ρε (x − y)dy|dx = C | nj ρε (x − y)dsy |dx; (10.40)
j O j Σ
RN RN

here, nj = nj (y) is the jth component of the outer normal n to O at y. Thus

Z Z
kχO ∗ ρε kLN/(N −1) ≤ C ρε (x − y)dxdsy = C|Σ|. (10.41)
Σ RN

On the other hand, we have χO ∈ LN/(N −1) and thus kχO ∗ ρε kLN/(N −1) → kχO kLN/(N −1) as ε → 0.
This leads us to
|O|(N −1)/N = kχO kLN/(N −1) ≤ C|Σ|, (10.42)
which ends the proof.
Theorem 23. We have
kukLN/(N −1),1 ≤ CkDukL1 , ∀ u ∈ W 1,1 (RN ). (10.43)
Proof. The strategy of the proof is the following: we first prove the inequality (10.43) when
u ∈ C0∞ ; the general case will be obtained from this one by passing to the limits.
Let u ∈ C0∞ ; Sard’s theorem insures that fact that, for a. e. t > 0, all the points x s. t. |u(x)| = t
satisfy ∇u(x) 6= 0; in other words, the set Σt = {|u| = t} is a smooth hyper surface. For any such
t, set Ot = {|u| > t}, which is a bounded open set. We claim that (*) Ot is a smooth domain
with boundary Σt . Indeed, it is obvious that ∂Ot ⊂ Σt . On the other hand, if x ∈ Σt and we set
v = ∇u(x), then Taylor’s formula implies that u(x + sv) − t has the sign of s when s is close to
0. Thus on the one hand x ∈ ∂Ot , on the other hand Ot is locally on one side of Σt , which is the
same as (*).
With F the distribution function of u, we have
F (t) = |Ot | ≤ C|Σt |N/(N −1) , (10.44)
by the weak isoperimetric inequality. Thus, with Ht = {u = t}, we have
Z Z∞ Z Z
(N −1)/N
kukLN/(N −1),w = F (t)dt ≤ C |Σt |dt = C |Ht |dt = C |Du|; (10.45)
0 R

the last equality follows from the coarea formula we will prove later.
We next turn to a general u ∈ W 1,1 . Consider a sequence (un ) ⊂ C0∞ s. t. un → u in W 1,1 and
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 70

pointwise outside an exceptional zero measure set B. We claim that the corresponding distribution
functions, F and Fn , satisfy F (t) ≤ lim inf Fn (t) for each t. Indeed, let A = {|u| > t}, An = {|un | >
t}. If x ∈ A \ B, then x ∈ An for sufficiently large n. Put it otherwise, A \ B ⊂ lim inf(An \ B),
and thus
F (t) = |A| = |A \ B| ≤ lim inf |An \ B| = lim inf |An | = lim inf Fn (t). (10.46)
Fatou’s lemma implies than that
Z Z Z
(N −1)/N (N −1)/N
kukLN/(N −1),1 = F (t)dt ≤ lim inf Fn (t)dt ≤ C |Du|. (10.47)

We have thus obtained (10.43) in full generality.

10.6 The limiting case p = N

The conclusion of Theorem 19 is wrong when 1/p + 1/q = 1 (and p 6= 1). Indeed, let f (x) =
χB(0,1/2) |x|−N/p | ln |x||−β , where β > 1/p. Then f ∈ Lp . Let also g(x) = |x|−N/q ∈ Lq,w . We claim
that f ∗ g 6∈ L∞ if β is well-chosen. We start by noting that Fatou’s lemma Zimplies that f ∗ g(0) ≤
lim inf f ∗ g(x). Therefore, f ∗ g 6∈ L∞ if f ∗ g(0) = ∞. Since f ∗ g(0) = |x|−N | ln |x||−β dx,
x→0
{|x|≤1/2}
we find that f ∗ g(0) = ∞ if β ≤ 1.
Consequently, we may not use Theorem 19 in the proof of Theorem 21 when p = N . Actually,
when p = N , the expected conclusion of Theorem 21, namely W 1,N ⊂ L∞ , is wrong: it is easy to
see that the function given by f (x) = χB(0,1/2) | ln |x||α , where 0 < α < 1 − 1/N , belongs to W 1,N ,
but not to L∞ . However, we will see that each function in W 1,N is ”almost” bounded. We start
with a simple (and non optimal) result.
Z
1
Proposition 17. There are constants c, C > 0 s. t. exp(c|u − uB |) ≤ C for each u ∈
|B|
B
W 1,N (RN ) s. t. kDukLN ≤ 1.
Proof. The above estimate follows immediately from Theorem 16 and the following result.
Proposition 18. We have, for some C depending only on N ,
Z
1
|u − uB | ≤ CkDukLN , ∀ u ∈ W 1,N . (10.48)
|B|
B
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 71

Proof. It suffices to prove (10.48) when u ∈ C0∞ ; the general case is obtained by passing to the
limits in (10.48) when B is kept fixed. If B = B(x, R), then
Z Z Z Z Z
1 1 1
|u − uB | = | (u(y + z) − u(y))dz|dy ≤ |u(y + z) − u(y)|dy dz.
|B| |B|2 |B|2
B B B(0,R) B B(0,R)
(10.49)
Applying Taylor’s formula in integral form, we find that
Z Z Z Z1 Z Z Z1
1 1 C
|u−uB | ≤ |(Du)(y+tz)||z|dt dz dy ≤ |(Du)(y+tz)|dt dz dy.
|B| |B|2 R2N −1
B B B(0,R) 0 B B(0,R) 0
(10.50)
For each t and z, Hölder’s inequality implies that
Z
C
|(Du)(y + tz)|dy ≤ kDukLN |B(x, 2R)|(N −1)/N ≤ kDukLN , (10.51)
RN −1
B

so that
Z Z Z1
1 CkDukLN
|u − uB | ≤ |(Du)(y + tz)|dt dz ≤ CkDukLN . (10.52)
|B| RN
B B(0,R) 0

We may actually replace |u| by |u|N/(N −1) in the preceding exponential integrability result.
The statement we give below includes the assumption that supp u ⊂ B. This is not a crucial
assumption; if we want to remove it, it suffices to apply the theorem when B is replaced by B ∗
(the ball concentric with B and twice larger) and u is replaced by ϕu, where ϕ is a cutoff function
supported in B ∗ and that equals 1 in B. However, the resulting inequality is less elegant.
Theorem 24. (Trudinger) Let u be a W 1,N function supported in B. If kDukLN ≤ 1, then
Z
1
exp c|u|N/(N −1) ≤ C,


(10.53)
|B|
B

where c, C > 0 depend only on N .

Proof. We may assume that B = B(0, R). We start by noting that (10.30) is valid for u as above.
Indeed, u being compactly supported, it belongs to W 1,2N/3 ; we may therefore rely on (10.36).
Let now f = |Du|, which belongs to LN and is supported in B. Set g = f ∗ |x|−(N −1) . In view of
CHAPTER 10. IMPROVED SOBOLEV EMBEDDINGS 72
Z
1
exp c g N/(N −1) ≤ C provided that kf kLN ≤ 1. The key


(10.30), it suffices to prove that
|B|
B
result in proving this estimate is the following inequality
g(x) ≤ C(δMf (x) + (ln(2R/δ))(N −1)/N ), ∀ x ∈ B, ∀ δ ∈ (0, R]. (10.54)

Assume (10.54) proved, for the moment. We consider, for x ∈ B, the two following possibilities:
(i) if x is s. t. Mf (x) ≤ R−1 , we choose δ = R and find that g(x) ≤ C;
(ii) if Mf (x) > R−1 , we choose δ = 1/Mf (x) and find that g(x) ≤ C(1 + ln(RMf (x))(N −1)/N
).
Thus, in any event, we have g(x)N/(N −1) ≤ C(1 + (ln(RMf (x))+ ), so that exp c g N/(N −1)(x) ≤
C1 (1 + (RMf (x))cC2 ). Choosing c s. t. cC2 = N , we find that
Z Z Z
1 N/(N −1)
≤ C (1 + R Mf ) ≤ C(1 + Mf N ) ≤ C(1 + kf kN
N N


exp c g LN ) ≤ C, (10.55)
|B|
B B

by the maximal inequalities.

It thus remains to establish (10.54). We have
f (x − y) f (x − y)
Z Z
g(x) = dy + dy = I1 + I2 . (10.56)
|y|N −1 |y|N −1
B(0,δ) B(x,R)\B(0,δ)

To estimate I1 , we note that I1 = f ∗ h(x), where h(y) = χB(0,δ) |y|−(N −1) . Since h is integrable,
radial and non increasing, we have I1 ≤ Mf (x)khkL1 = CδMf (x). We complete the proof of
(10.54) by noting that Hölder’s inequality combined with the fact that kf kLN = 1 yields
 Z (N −1)/N  Z (N −1)/N
−N −N
I2 ≤ |y| ≤ |y| = C(ln(2R/δ))(N −1)/N . (10.57)
B(x,R)\B(0,δ) B(0,2R)\B(0,δ)
Chapter 11

Traces

11.1 Definition of the trace

We discuss here the properties of the ”restrictions” of Sobolev maps to hyper surfaces, e. g., to the
boundary of a smooth domain. There is a standard reduction procedure which allows to replace
a ”smooth” (at least Lipschitz) hyper surface with a hyperplane; this is done by flattening locally
the coordinates. Since this part works without any problem and we want to insist on the analytic
part, we will simply consider in this chapter maps defined in the whole RN and consider properties
of their trace on the hyperplane H = {x = (x0 , xN ) ∈ RN ; xN = 0}, which we identify with RN −1 .
We start by recalling the following
Proposition 19. The map u → 7 u|H , initially defined from C0∞ (RN ) into C0∞ (RN −1 ), extends
uniquely by density to a linear map (called trace map) u 7→ tr u from W 1,p (RN ) into Lp (RN ),
for 1 ≤ p < ∞.
There also is a statement concerning W 1,∞ (we will see it in the last part), but in that case the
trace is not defined in the same way, since C0∞ is not dense in W 1,∞ .
Proof. Fix a function ϕ ∈ C0∞ (R) s. t. ϕ(0) = 1 and supp ϕ ⊂ (−1, 1). If u ∈ C0∞ , then
v = uϕ(xN ) ∈ C0∞ (RN −1 ×(−1, 1)) and u|H = v|H . In addition, it is clear that kvkW 1,p ≤ CkukW 1,p .
It therefore suffices to prove that kv|H kLp ≤ CkvkW 1,p . This follows from
Z Z Z1 p Z
0 p 0
|v(x , 0)| dx = 0
∂N v(x , t)dt dx ≤0
|Dv|p ≤ kDvkpLp . (11.1)
H H 0 H×(0,1)

When 1 < p < ∞, the above result is not sharp, in the following sense: if f is an arbitrary map
in Lp (RN −1 ), we can not always find a map u ∈ W 1,p s. t. tr u = f . In other words, the trace

73
CHAPTER 11. TRACES 74

map is not surjective between the spaces we consider.

In this chapter, we will determine the image of the trace map.
Definition 5. For 0 < s < 1 and 1 ≤ p < ∞, we define
|f (x) − f (y)|p
Z Z
s,p s,p N p N
W = W (R ) = {f ∈ L (R ) ; dx dy < ∞}, (11.2)
|x − y|N +sp
RN RN

equipped with the norm

1/p
|f (x) − f (y)|p
Z Z
kf kW s,p = kf kLp + dx dy . (11.3)
|x − y|N +sp
RN RN

We let the reader check that W s,p is a Banach space.

The main result of this chapter states that tr W 1,p (RN ) = W 1−1/p,p (RN −1 ). We start with some
preliminary results.
Lemma 26. C ∞ (RN ) ∩ W s,p (RN ) is dense into W s,p (RN ) for 0 < s < 1 and 1 ≤ p < ∞.
Z
∞
Proof. Let ρ be a standard mollifier (i. e., ρ ∈ C0 , ρ ≥ 0, ρ = 1, supp ρ ⊂ B(0, 1)). We will
prove that, if f ∈ W s,p , then fε = f ∗ ρε → f in W s,p . Clearly, fε → f in Lp . It remains to prove
that, with gε = fε − f , we have
|gε (x) − gε (y)|p |gε (x + h) − gε (x)|p
Z Z Z Z
Iε = dx dy = dx dh → 0 as ε → 0. (11.4)
|x − y|N +sp |h|N +sp
RN RN RN RN

In order to estimate Iε , we start by noting that

Z
gε (x + h) − gε (x) = (f (x + h − y) − f (x + h) − f (x − y) + f (x))ρε (y)dy. (11.5)
B(0,ε)

Using in addition the fact that ρε ≤ Cε−N , we find that

Z
C
|gε (x + h) − gε (x)| ≤ N |f (x + h − y) − f (x + h) − f (x − y) + f (x)|dy. (11.6)
ε
B(0,ε)

We next consider the two following cases:

(i) if |h| < ε, we have
Z
C
|gε (x + h) − gε (x)| ≤ N (|f (x + h − y) − f (x − y)| + |f (x) − f (x + h)|dy; (11.7)
ε
B(0,ε)
CHAPTER 11. TRACES 75

(ii) if |h| ≥ ε, we use the inequality

Z
C
|gε (x + h) − gε (x)| ≤ N (|f (x + h − y) − f (x + h)| + |f (x) − f (x − y)|dy. (11.8)
ε
B(0,ε)

C
Thus Iε ≤ (J1 + J2 + J3 + J4 ), where
εN p
Z Z  Z p
J1 = |f (x + h − y) − f (x − y)|dy |h|−(N +sp) dh dx; (11.9)
RN {|h|<ε} B(0,ε)
Z Z  Z p
J2 = |f (x + h) − f (x)|dy |h|−(N +sp) dh dx; (11.10)
RN {|h|<ε} B(0,ε)
Z Z  Z p
J3 = |f (x + h − y) − f (x + h)|dy |h|−(N +sp) dh dx; (11.11)
RN {|h|≥ε} B(0,ε)
Z Z  Z p
J4 = |f (x − y) − f (x)|dy |h|−(N +sp) dh dx. (11.12)
RN {|h|≥ε} B(0,ε)

We will prove that εN p Jj → 0, j = 1, . . . , 4. The only ingredient we use in the proof is

|f (x + y) − f (x)|p
Z Z
lim dy = 0; (11.13)
ε→0 |y|N +sp
RN B(0,ε)

this follows easily by dominatedconvergence.

Z p
We start with J2 . Noting that |f (x + h) − f (x)|dy = CεN p |f (x + h) − f (x)|p , we find
B(0,ε)
that
|f (x + h) − f (x)|p
Z Z
−N p
ε J2 = C dh → 0. (11.14)
|h|N +sp
RN B(0,ε)

For J1 , Hölder’s inequality with exponents p and p0 implies that

 Z p Z
p−1
|f (x + h − y) − f (x − y)|dy ≤ |B(0, ε)| |f (x + h − y) − f (x − y)|p dy, (11.15)
B(0,ε) B(0,ε)
CHAPTER 11. TRACES 76

and thus
Z Z Z
−N p −N
ε J1 ≤ Cε |f (x + h − y) − f (x − y)|p dy |h|−(N +sp) dh dx. (11.16)
RN {|h|<ε} B(0,ε)

For fixed y and h, the change of variables x − y = z leads to

Z Z
−N p
ε J1 ≤ C |f (z + h) − f (z)|p |h|−(N +sp) dh dz → 0. (11.17)
RN {|h|<ε}

We next estimate J3 ; the computation for J4 is similar and will be omitted. Inequality (11.15)
implies that
Z Z Z
−N p −N
ε J3 ≤ Cε |f (x + h − y) − f (x + h)|p |h|−(N +sp) dy dh dx. (11.18)
RN {|h|≥ε} B(0,ε)

In this integral, we fix y and h and make the change of variables x + h = z. Next we integrate in
h and find that
|f (z − y) − f (z)|p |f (z − y) − f (z)|p
Z Z Z Z
−N p
ε J3 ≤ C dy dz ≤ C dy dz → 0. (11.19)
εN +sp |y|N +sp
RN B(0,ε) RN B(0,ε)

Lemma 27. If u ∈ C(RN ) ∩ W 1,p , then tr u = u|H .

Proof. Let ρ be a standard mollifier and set uε = ρ(ε·)(u ∗ ρε ). Clearly, uε ∈ C0∞ and uε → u in
W 1,p . Thus uε|H = tr uε → tr u in Lp (and thus in D0 ). On the other hand, uε|H converges to u|H
uniformly on compacts (and thus in D0 ), whence the conclusion.
The same argument leads to the following variant
Lemma 28. Assume that u ∈ W 1,p is continuous in a neighborhood of H. Then tr u = u|H .
Lemma 29. Let u ∈ C(RN )∩C 1 (RN \H). Assume that the pointwise differential Du of u satisfies
Du ∈ Lp (RN ). Then Du is also the distributional differential of u.
Z Z
Proof. We have to prove that Dj uϕ = − u∂j ϕ, ϕ ∈ C0∞ , j = 1, . . . , N . When j ≤ N − 1,
Z
this follows simply by Fubini’s theorem. Assume j = N . We integrate by parts DN uϕ in the
set {x ∈ RN ; |xN | > ε} and next let ε → 0. We find that
Z  Z Z Z  Z
DN uϕ = lim uϕdsx −
0 uϕdsx −
0 u∂N ϕdx =− u∂N ϕdx. (11.20)
ε→0
{xN =−ε} {xN =ε} {|xN |>ε}
CHAPTER 11. TRACES 77

11.2 Trace of W 1,p , 1 < p < ∞

Theorem 25. (Gagliardo) Let p ∈ (1, ∞).
a) If u ∈ W 1,p (RN ), then tr u ∈ W 1−1/p,p (RN −1 ) and ktr ukW 1−1/p,p ≤ CkukW 1,p .
b) Conversely, let f ∈ W 1−1/p,p (RN −1 ). Then there is some u ∈ W 1,p (RN ) s. t. tr u = f . In
addition, we may pick u s. t. kukW 1,p ≤ Cktr ukW 1−1/p,p .
Remark 12. Let T : W 1,p (RN ) → W 1−1/p,p (RN −1 ), T u = tr u. T is linear, and the above theorem
implies that T is continuous and surjective. Then the last conclusion in b) follows from the open
map theorem (each surjective linear continuous map between two Banach spaces has a bounded
right inverse). However, we will see during the proof a stronger conclusion: we will construct in
b) a linear right inverse, i. e. , the map f 7→ u in b) will be linear.
Proof. a) By density, it suffices to prove that
ku|H kW 1−1/p,p ≤ CkukW 1,p ∀ u ∈ C0∞ . (11.21)
We start by noting that we already know that ku|H kLp ≤ CkukW 1,p ; thus it suffices to establish,
with f (x0 ) = u(x0 , 0), the inequality
|f (x0 + h0 ) − f (x0 )|p 0 0
Z Z Z
I= dh dx ≤ C |Du(x)|p dx. (11.22)
|h0 |N +p−2
RN −1 RN −1 RN

The starting point is the inequality

|f (x0 + h0 ) − f (x0 )| ≤ |f (x0 + h0 ) − u(x0 + h0 /2, |h0 |/2)| + |f (x0 ) − u(x0 + h0 , |h0 |/2)|, (11.23)
which implies that I ≤ C(I1 + I2 ), where
|f (x0 + h0 ) − u(x0 + h0 /2, |h0 |/2)|p |f (x0 ) − u(x0 + h0 /2, |h0 |/2)|p
Z Z Z Z
I1 = , I2 = .
|h0 |N +p−2 |h0 |N +p−2
RN −1 RN −1 RN −1 RN −1
(11.24)
0 0 0 0 0
If we perform, in I1 , the change of variables x + h = y , next we change h into −h , we see that
I1 = I2 , and thus
|f (x0 ) − u(x0 + h0 /2, |h0 |/2)|p 0 0
Z Z
I≤C dh dx . (11.25)
|h0 |N +p−2
RN −1 RN −1

Changing h0 into 2k 0 and applying the Leibniz-Newton formula, we find that

Z Z  Z|k0 | p
I≤C |Du(x + t(k /|k |), t)| |k 0 |−(N +p−2) dk 0 dx0 .
0 0 0
(11.26)
RN −1 RN −1 0
CHAPTER 11. TRACES 78

Expressing k 0 in polar coordinates, we find that

Z Z Z∞  Zs p
I≤C 0
|Du(x + tω, t)|dt s−p ds dsω dx0 . (11.27)
RN −1 S N −2 0 0

Applying, for fixed x0 and ω, Hardy’s inequality in to the double integral in s and t, we find that
Z Z Z∞
I≤C |Du(x0 + tω, t)|p dt dsω dx0 . (11.28)
RN −1 S N −2 0

Integrating, in the above inequality, first in x0 , next in ω, we find that

Z Z∞ Z
I≤C 0
|Du(x , t)| dt dx =p 0
|Du(x)|p dx ≤ CkDukpLp . (11.29)
RN −1 0 RN +

b) It suffices to construct a linear map f 7→ u, f ∈ C ∞ (RN −1 ) ∩ W 1−1/p,p , u ∈ W 1,p (RN ), s. t. tr

u = f and kukW 1,p ≤ Ckf kW 1−1/p,p . We fix a standard mollifier ρ in RN −1 and an even function
ϕ ∈ C ∞ (R) s. t. ϕ(0) = 1, 0 ≤ ϕ ≤ 1 and supp ϕ ⊂ (−1/2, 1/2). We define, for t 6= 0, v(x0 , t) =
f ∗ ρ|t| (x0 ) and u(x0 , t) = v(x0 , t)ϕ(t). We extend u to RN by setting u(x0 , 0) = f (x0 ). Clearly, the
map f 7→ u is linear and u ∈ C ∞ (RN \ H). In addition, u ∈ C(RN ) when f is continuous. We
also note that, for a fixed t 6= 0, Young’s inequality implies that kf ∗ ρ|t| kLp ≤ kf kLp , and thus
kukLp ≤ kf kLp . Since u is even with respect to xN , it suffices to prove, in view of Lemmata 27
and 29, that the usual differential Du of u satisfies
Z∞
|f (x0 + y 0 ) − f (x0 )|p 0 0
Z Z Z
0 p
|Du(x , t)| dt dx ≤ C 0
N +p−2
dy dx + Ckf kpLp . (11.30)
|y|
RN −1 0 RN −1 RN −1

For 1 ≤ j ≤ N −1, we have |∂j u| ≤ |∂j v|. On the other hand, |∂N u| ≤ C|v|χRN −1 ×(−1/2,1/2) +|∂N v|.
Since k|v|χRN −1 ×(−1/2,1/2) kLp ≤ kukLp , it suffices to prove the estimate
Z∞
|f (x0 + y 0 ) − f (x0 )|p 0 0
Z Z Z
0 p 0
|Dv(x , t)| dt dx ≤ C dy dx . (11.31)
|y|N +p−2
RN −1 0 RN −1 RN −1
Z
Let 1 ≤ j ≤ N − 1. Since ∂j ρ = 0, we have
Z Z
0 −N 0 0 0 0 −N
∂j v(x , t) = t f (y )(∂j ρ)((x − y )/t)dy = t [f (y 0 ) − f (x0 )](∂j ρ)((x0 − y 0 )/t)dy 0 , (11.32)
CHAPTER 11. TRACES 79

so that Z
0 C
|∂j v(x , t)| ≤ N |f (x0 + y 0 ) − f (x0 )|dy 0 . (11.33)
t
B(0,t)
Z Z
d 0 0
We next claim that [ρt (x )]dx = 0. This follows from the fact that ρt ≡ 1. Thus
dt
Z Z
0 0 0 d 0 0 0 C
|∂N v(x , t)| = | [f (y ) − f (x )] [ρt (x − y )]dy | ≤ N |f (x0 + y 0 ) − f (x0 )|dy 0 , (11.34)
dt t
B(0,t)
Z
d C
since ρt ≤ Ct−N . We find that |Dv(x0 , t)| ≤ N |f (x0 + y 0 ) − f (x0 )|dy 0 , and therefore it
dt t
B(0,t)
suffices to establish the estimate
Z Z∞  Z p
|f (x0 + y 0 ) − f (x0 )|p 0 0
Z Z
0 0 0 0 −N p 0
I= |f (x + y ) − f (x )|dy t dt dx ≤ C dy dx .
|y|N +p−2
RN −1 0 B(0,t) RN −1 RN −1
(11.35)
This is done as in the proof of lemma 26: Hölder’s inequality applied to the integral over B(0, t)
implies that
Z Z∞ Z
I≤C |f (x0 + y 0 ) − f (x0 )|p dy 0 t−N −p+1 dt dx0 . (11.36)
RN −1 0 B(0,t)

Fubini’s theorem yields

Z∞
|f (x0 + y 0 ) − f (x0 )|p 0 0
Z Z Z Z
−N −p+1 0 0
I≤C t dt dx dy = C dy dx . (11.37)
|y|N +p−2
RN −1 RN −1 |y 0 | RN −1 RN −1

Corollary 17. Let f ∈ W 1−1/p,p (RN ) and set , for t 6= 0, u(x0 , t) = f ∗ ρ|t| (x0 )ϕ(t). Then u ∈ W 1,p
and tr u = f .

11.3 Trace of W 1,1

We start with some auxiliary results needed in the proof of the fact that the trace of W 1,1 = L1 .
Lemma 30. Let u ∈ W 1,p ∩ W 1,q . Then the two traces of u (one in W 1,p , the other one in W 1,q ),
coincide.
CHAPTER 11. TRACES 80

Proof. If ρ is a standard mollifier, then uε = ρ(ε·)u ∗ ρε converges (as ε → 0) to u both in W 1,p

and in W 1,q . Since, for uε ∈ C0∞ , both traces coincide, we obtain the result by passing to the
limits.
The same argument leads to the following result.
Lemma 31. Let u ∈ W 1,p . For λ 6= 0 and x0 ∈ RN −1 , we have tr u(λ · −x0 ) = (tr u)(λ · −x0 ).
Lemma 32. Let f be the characteristic function of a cube in RN −1 . Then f ∈ W 1−1/p,p for
1 < p < 2.
Proof. We may assume that C = (−l, l)N . If we consider in RN −1 the k · k∞ norm, then
dx0 dy 0
Z Z
p
kf kW 1−1/p,p ∼ . (11.38)
|x0 − y 0 |N +p−2
|x0 |<l |y 0 |>l

If |x0 | < l and |y 0 | > l, then y 0 ∈ RN −1 \ B(x0 , l − |x0 |), and therefore
Z∞
dy 0 dz 0
Z Z
≤ =C r−p = C(l − |x0 |)1−p . (11.39)
|x0 − y 0 |N +p−2 |z 0 |N +p−2
|y 0 |>l |z 0 |>l−|x0 | l−|x0 |

Since p < 2, we find that

dx0
Z Z
kf kpW 1−1/p,p ≤C 0
(l − |x |) 1−p
≤C < ∞. (11.40)
(l − x1 )p−1
|x0 |<l {|xj |≤x1 <l, j=1,...,N −2}

1
Lemma 33. Let C be a cube of size l in RN −1 and set a = χC . Then there is a map u ∈ W 1,1
|C|
s. t. tr u = a and
kukL1 ≤ c l and kDukL1 ≤ c. (11.41)
Proof. We start with the case where C is the unit cube (or any other cube of size 1). We fix a
p ∈ (1, 2). Since a ∈ W 1−1/p,p , we have a = tr u0 for some u ∈ W 1,p . In addition, Corollary 17
implies that we may assume u0 compactly supported. Thus u ∈ W 1,1 and tr u0 = a (computed in
W 1,1 ). Let now C be an arbitrary cube, which we may assume with sides parallel to the unit cube
Q. Let C = x0 + (0, l)N −1 . Set u = l−(N −1) u0 (l−1 (· − x0 )). Then u ∈ W 1,1 and tr u = a. Inequality
(11.41) follows from the identities kukL1 = lku0 kL1 and kDukL1 = kDu0 kL1 .
Theorem 26. (Gagliardo) Let f ∈ L1 (RN −1 ). Then there is some u ∈ W 1,1 (RN ) s. t. tr u = f
and kukW 1,1 ≤ Ckf kL1 .
CHAPTER 11. TRACES 81

Remark 13. This time, the map f 7→ u we construct is not linear.

X
Proof. The main ingredient is the following: if f ∈ L1 , then we may write, in L1 , f = λn an ,
where:
1
(i) each an is of the form an = χC ;
|Cn | n
(ii) each
X Cn is of size at most 1;
(iii) |λn | ≤ Ckf kL1 .
Assuming that this can be achieved, here is the end of the proof: the preceding lemma implies
that each an is the trace of some un ∈X W 1,1 s. t. kun kW 1,1 ≤ C. The linearity of the trace and
property (iii) imply that the map u = λn un ∈ W 1,1 satisfies tr u = f and kukW 1,1 ≤ Ckf kL1 .
X
It remains to perform the decomposition f = λn an . For each j ∈ N, let Fj be the grid of cubes
−j
of size 2 , with sides parallel to the coordinate axes and having the origin among the edges. We
define the linear map Tj : L1 → L1 , Tj f (x) = fC if x ∈ C ∈ Fj . Clearly, Tj is of norm 1. We
claim that, for each f ∈ L1 , we have Tj f → f in L1 as j → ∞. This is clear when f ∈ C0∞ ;
the case of a general f follows by approximation using the Xfact that kTj k = 1. We may thus find
an increasing sequence of indices, (jk ), s. t. kfj0 kL1 + kfjk+1 − fjk kL1 ≤ Ckf kL1 . We claim
X 1
that fjk+1 − fjk = λkn akn , where each akn is of the form χC for some cube of size at most 1
X |C|
and |λkn | = kfjk+1 − fjk kL1 . Indeed, fjk+1 − fjk is constant on each cube C ∈ Fjk , and thus
X X 1
fjk+1 −fjk = (fjk+1 −fjk )|C χC , so that fjk+1 −fjk = λC χC , with λC = (fjk+1 −fjk )|C |C|.
C∈F C∈F
|C|
jk jk

We find that
X Z X X
kfjk+1 − fjk kL1 = |fjk+1 − fjk | = |C||(fjk+1 − fjk )|C | = |λC |. (11.42)
C∈Fjk C C∈Fjk C∈Fjk

X 1
Similarly, we may write fj0 = λ0n a0n , where each a0n is of the form χC for some cube of size
X |C|
at most 1 and |λ0n | = kfj0 kL1 .
XX
Finally, we write f = λkn akn , and this decomposition has the properties (i)-(iii).
k n