On estimates of constants for maximal

functions

ALEXANDER IAKOVLEV

Doctoral Thesis
Stockholm, Sweden 2014
 KTH
TRITA-MAT-A 2014:09 Institutionen för Matematik
ISRN KTH/MAT/A-14/09-SE 100 44 Stockholm
ISBN 978-91-7595-186-7 SWEDEN

Akademisk avhandling som med tillstånd av Kungl Tekniska högskolan

framlägges till offentlig granskning för avläggande av teknologie doktors-
examen i matematik måndagen den 16 juni 2014 kl 13.00 i sal F3, Kungl
Tekniska högskolan, Lindstedtsvägen 26, Stockholm.

⃝
c Alexander Iakovlev, 2014

Tryck: Universitetsservice US AB
 iii

Abstract

In this work we will study Hardy-Littlewood maximal function and

maximal operator, basing on both classical and most up to date works.
In the first chapter we will give definitions for different types of those
objects and consider some of their most important properties. The
second chapter is entirely devoted to an overview of the fundamental
properties of Hardy-Littlewood maximal function, which are strong
(p, p) and weak (1, 1) inequalities. Here we list the most actual results
on this inequalities in correspondence to the way the maximal func-
tion is defined. The third chapter presents the theorem on asymptotic
behavior of the lower bound of the constant in the weak-type (1, 1)
inequality for the maximal function associated with cubes of Rd , then
the dimension d tends to infinity. In the last chapter a method for
computing constant c, appearing in the main theorem of chapter 3, is
given.
iv

Sammanfattning

I detta avhandlingsarbete studeras Hardy-Littlewood maximal funk-

tion och maximal operator, med stöd i både klassiska och mest uppda-
terade studier. I det första kapitlet ges definitioner for olika typer av
dessa objekt och några av deras viktigaste egenskaper granskas. Det
andra kapitlet ägnas helt åt en översikt av de grundläggande egenska-
perna hos Hardy-Little maximal funktion, d.v.s. starka (p, p) och svaga
(1, 1) olikheter. Här presenteras de mest aktuella resultaten av dessa
olikheter i motsvarighet till det sätt som den maximala funktionen de-
finieras på. I det tredje kapitlet presenteras satsen om asymptotiskt be-
teende av den undre gransen av konstanten i den svaga (1, 1) olikheten
för den maximala funktionen associerad med kuber Rd , dådimensionen
d går mot oändligheten. I det sista kapitlet presenteras en metod for
berakning av konstanten c, som forekommer i huvudsatsen i kapitel 3.
Contents

Contents v

Acknowledgements vii

1 Introduction 1
1.1 The origins of maximal function’s problem . . . . . . . . . . . 1
1.2 Definitions and different types of maximal functions . . . . . 3
1.3 Properties of the maximal operator . . . . . . . . . . . . . . . 5
1.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Estimations of the best constants 19

2.1 The case of R . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1

2.2 The case of higher dimension . . . . . . . . . . . . . . . . . . 25

3 Asymptotic estimations of the best constant 35

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Proof of the theorem . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 Proof of the proposition . . . . . . . . . . . . . . . . . . . . . 37
3.5 Final proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.6 Proof of lemma 1 . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.7 Proof of lemma 2 . . . . . . . . . . . . . . . . . . . . . . . . . 51

4 More precise estimation 53

4.1 Case of u < 1/2. . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Numerator of (4.4) . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3 Denominator of (4.4) . . . . . . . . . . . . . . . . . . . . . . . 61
4.4 Case of u ∈ (1/2, 1) . . . . . . . . . . . . . . . . . . . . . . . . 65

v
vi CONTENTS

4.5 Estimating the intersection . . . . . . . . . . . . . . . . . . . 66

4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

References 69
Acknowledgements

I would like to express my sincere gratitude to everyone who helped me.

Special thanks to my supervisor Jan-Olov Strömberg and to my father
Sergey L. Yakovlev as well to all other my family members for their support.

vii
Chapter 1

Introduction

1.1 The origins of maximal function’s problem

In the original paper [18] of G. H. Hardy and J. E. Littlewood published in
1930 the following problem was stated:
Suppose that λ is a real positive constant and a function f ∈ Lλ is analytic
within the unit circle. Denote the maximum of |f | on the radius θ as follows

F (θ) = max |f (reiθ )|. (1.1)

0≤r≤1

Is it true that
∫ π ∫ π
1 1
F λ dθ ≤ A(λ) |f (eiθ )|λ dθ, (1.2)
2π −π 2π −π

where A(λ) is a function of λ only?

The affirmative answer in some cases (λ > 1) to this question is given in
the paper and was formulated first for sums and then for integrals.
It is interesting to remark that the considerations are stated in terms of
sport achievements analysis and described in the following manner:
Suppose that a cricket batsman plays in a given season and scored a given
’stock’ of n innings
a1 , a2 , . . . , an .

1
2 CHAPTER 1. INTRODUCTION

Let αν be the average after the ν-th innings, so that

Aν a1 + a2 + · · · + aν
αν = = . (1.3)
ν ν
Let s(x) be a positive function which increases (in the wide sense) with x,
and let the batsmen’s ’satisfaction’ after the ν-th inning be measured by
sν = s(αν ). (1.4)
Finally, let his total satisfaction for the entire season be
∑
n ∑
n
S= sν = s(αν ). (1.5)
ν=1 ν=1

It is then easily verified that S is a maximum, for a given stock of innings,

when they are played in decreasing order.
A trivial logic and a common sense is enough to solve this problem,
although it has a nontrivial generalization. Let us replace (1.3) by the
following
Aν aν ∗ + aν ∗ +1 + · · · + aν aµ + aµ+1 + · · · + aν
αν = = ∗
= max , (1.6)
ν ν−ν +1 µ≤ν ν−µ+1
where ν ∗ should be chosen as small as possible, and we arrive at the following
result:
Theorem 1. If a1 , a2 , . . ., an are positive, in the wide sense, and given
except in arrangement, s(x) is any increasing function of x, and aν , sν , S
are defined by (1.6), (1.4), and (1.5), then S is maximum when aν , are
arranged in descending order.
Despite this theorem has the same conclusion its proof is far from ob-
vious. To formulate a maximal theorem for integrals we need to introduce
some auxiliary functions.
Suppose that f (x) is positive bounded and measurable on (0, a), and let
m(y) be the measure of the set in which f (x) ≥ y. We define f ∗ (x), for
0 ≤ x ≤ a, by
f ∗ (m(y)) = y (0 ≤ m(y) ≤ a).
The function f ∗ (x) is called the rearrangement of f (x) in decreasing order.
We also need to define an average of f on (x, ξ) as
∫x
1
A(x, ξ) = A(x, ξ, f ) = f (t)dt, A(x, x) = f (x)
x−ξ
ξ
1.2. DEFINITIONS AND DIFFERENT TYPES OF MAXIMAL
FUNCTIONS 3
and
Θ(x) = Θ(x, f ) = max A(x, ξ).
0≤ξ≤x

A theorem for integrals corresponding to Theorem 1 is formulated as

Theorem 2. If s(x) is a continuous and increasing function then
∫a ∫a
s (Θ(x, f )) dx ≤ s (Θ(x, f ∗ )) dx.
0 0

It may seem that this theorem is just a generalization of the previous

one but the proof of it is much more involved.
Three "original" maximum functions of Hardy and Littlewood are defined
later in the text for integrable 2π periodic function (here we keep the original
notations)
∫t
1
M (θ) = M (θ, f ) = max f (θ + x)dx ,
0<|t|≤π t
0
 
∫t
1
M (θ) = M (θ, f ) = max  |f (θ + x)|dx ,
0<|t|≤π t
0
 
∫t
1
N (θ) = N (θ, f ) = max  |f (θ + x)|dx .
|t|≤π t
0

Some basic properties of this functions are formulated and proved in that
paper [18]. There are also several applications of maximal theorems, some
of them will be presented in this work.
The results obtained by Hardy and Littlewood were later generalized on
the case of Rn by Wiener in [38]

1.2 Definitions and different types of maximal

functions
The maximal operator M · is a sublinear operator which maps a locally
integrable function f (x), x ∈ Rd into another function M f (x) that, at each
point x ∈ Rd , is the maximum "average" value that f may have on some pre-
defined collection of sets. The function M f is called the maximal function.
4 CHAPTER 1. INTRODUCTION

One can find alternative notations for the maximal function, for example,
Stein and Shakarchi use the notation f ∗ , which can be a bit confusing.
The maximal operator M · or maximal function M f can be defined in
following ways:

a) in the most general case, for a locally integrable function f (x), x ∈ Rd

and measure µ ∫
1
M f (x) := sup |f | dµ, (1.7)
R∋x µ(R) R

where R ∈ R and R is a chosen collection of sets. The most common

measures to define the maximal function are Lebesgue and Borel mea-
sures.

b) maximal function can also be defined for a measure. Then for the measure
σ ∫
1 σ(R)
M σ(x) := sup dσ = sup . (1.8)
R∋x |R| R R∋x |R|

c) in the case when µ is the Lebesgue measure one can find the following
notation ∫
1
M f (x) := sup |f (y)| dy. (1.9)
R∋x |R| R

The last two cases are practically equivalent but the first one differs sig-
nificantly. The properties of the maximal function depend strongly on the
definition and can differ dramatically. For example, for different collections
R we get different maximal functions. The most important collections are
balls, spheres, rectangles or segments of lines.
The sets R may possess some additional properties. For example, they
can be centered at the point x. In this case the function M f (x) is called
the centered maximal function.
By Hardy-Littlewood maximal operator we usually mean
∫
1
M f (x) := sup |f (y)| dy, (1.10)
r>0 |Q(x, r)| Q(x,r)

where Q(x, r) is an ln ball (cube for n = ∞) with the radius r centered

at x, although there is no conventional notation what is exactly called by
Hardy-Littlewood maximal operator and function.
1.3. PROPERTIES OF THE MAXIMAL OPERATOR 5

In dimension d > 1 the maximal operator can be defined in the following

way
Nd = M1 × M2 × . . . × Md . (1.11)
where × stands for the standard operator product. Here Mi is the one-
dimensional uncentered maximal operator acting on xi ∈ R1 only. This
operator is called then the strong maximal operator.
There are many other special kinds of maximal operators. Let us present
some of them here. Let Ω ∈ Rd be an arbitrary subdomain (i.e. an open
and connected subset). Then the local Hardy-Littlewood maximal operator
MΩ f , is defined by
∫
1
M f (x) := sup |f (y)| dy, where f ∈ L1loc (Ω).
0<r<D(x,Ωc ) |Q(x, r)| Q(x,r)

Here D(x, Ωc ) denotes the Euclidian distance from x to closure of Ω.

For functions f ∈ L1loc (Rd ) one can define a fractional maximal operator
α
M , where 0 < α < d. This operator has applications in potential theory
and partial differential equations. The definition is given by the following
formula ∫
rα
M f (x) := sup |f (y)| dy,
r>0 |Q(x, r)| Q(x,r)

for x ∈ Rd .
The restricted Hardy-Littlewood maximal operator MλQ , λ ≥ 0, is given
by ∫
1
MλQ f (x) := sup |f (y)| dy.
r≥λ |Q(x, r)| Q(x,r)
Again, as it was in the case of the original maximal operator, we could
define the local, fractional or restricted maximal operator also by using cubes
or non-centered balls.
In some cases, the restriction on f be the function from Lp appears to
be too strong and one can allow f to be a distribution.

1.3 Properties of the maximal operator

In this section we will present some properties of maximal operator M · and
maximal function M f . In most cases the way the operator M · is defined
is not important. Any restrictions applied will be mentioned specifically in
this text.
6 CHAPTER 1. INTRODUCTION

Sublinearity
The obvious property of the original maximal operator on locally integrable
functions L1loc is that it is sublinear. This implies that the operator M · as a
mapping L1loc → L1loc possesses two following fundamental properties:

1) sub-addictiveness: for every two functions f, f ′ ∈ L1loc the following in-

equality holds

|M (f + f ′ )(x)| ≤ |M f (x)| + |M f ′ (x)|.

2) homogeneity
|M (kf )(x)| = |k||M f (x)|
for all k ∈ C1 .

This properties can easily be verified.

Mapping properties for maximal operator

The maximal operator is usually considered for the Lebesgue measure as an
operator form Lp (Rd ) to Lp (Rd ) althogh it can be defined on a wider class
of functions. While the size of the maximal function is of main interest, it
is also useful and interesting to study how the maximal operator preserves
the regularity properties of functions. A simple observation is that the set

{x ∈ Rd ; M f (x) > λ}

is open for arbitrary function f ∈ L1loc . In other words, M f is always lower

semi-continuous. Moreover, it is easy to see that if f is continuous, then so
is M f as well (if M f is not equivalent to ∞). Kinnunen [21] proved that
M · is bounded on the Sobolev spaces W 1,p (Rd ). Moreover, he showed the
pointwise inequality
|Di M f (x)| ≤ M (Di f )(x) (1.12)
for all f ∈ W 1,p (Rd ) and almost everywhere, for x ∈ Rd . Here Di f denotes
the weak partial derivative of f in the direction ei . Kinnunen applied this
result to study the Lebesgue points of Sobolev functions.
The boundedness of M · on Sobolev spaces follows from the properties of
maximal operator:
1.3. PROPERTIES OF THE MAXIMAL OPERATOR 7

1) the maximal operator commutes with translations. For h ∈ Rd denote

fh (x) = f (x + h). Then, M (fh )(x) = (M f )h (x), for every x ∈ Rd .

2) M is sublinear, which means that for given locally integrable functions

f and g, the inequality |M f (x) − M g(x)| ≤ M (f − g)(x) holds almost
everywhere.

From these facts we see that

( )
1 fh − f
|(M f )h − M f | ≤ M .
|h| |h|

The boundedness of M · on W 1,p (Rd ) and the pointwise inequality (1.12) are
easily implied by this inequality.

Continuity
In general, boundedness does not need to imply continuity as shown in [3].
However it does imply for operators on Lp = W 0,p , since M · is sublinear.
As shown by Luiro in [25] M · is also continuous as operator from W 1,p (Rd )
into W 1,p (Rd ), when 1 < p < ∞.
For the case of fractional Sobolev spaces one may refer to [22] where
the boundedness of the Hardy-Littlewood operator on the Triebel-Lizorkin
p (Rd ) was proved for 1 < p, q < ∞ and 0 < s < 1. Since it is
spaces Fs,q
p
known that Fs,2 (Rd ) = W s,p (Rd ) the result of [22] can be extended to some
Sobolev spaces.
A number of questions regarding regularity are considered in the works
of Luiro [25], [24] and Korry [22].

Eigenfunctions of maximal operator

In his work [23], Korry proved the existence of fixed points of Hardy-
Littlewood maximal operator, i.e. the existence of functions of the type
M f (x) = f (x) for f ∈ Lp (Rd ), d ≥ 3 and d/(d − 2) < p ≤ ∞. Such fixed
points are positive super harmonic functions, for example, inf{1, |x|d−2 }.
It is also proved that the strong centered maximal operator over parallel-
ograms with sided parallel to the axes has no fixed points in Lp for every
1 ≤ p < +∞. These results have been extended to rearrangement invariant
8 CHAPTER 1. INTRODUCTION

spaces in [26]. Following these lines of research, Colzani and Làzaro [10] stud-
ied the same problem for uncentered maximal operator in R1 . They claimed
that since the maximal function of a non constant function is larger than the
function, this maximal operator has no non constant fixed points, although
it is possible to find eigenvalues λ > 1 and their corresponding eigenfunc-
tions. Indeed, since the operator commutes with dilations and reflections,
an homogeneity argument shows that the functions |x|−a with 0 < a < 1
are eigenfunctions, with eigenvalues which are the values of the maximal
function at the point 1. Moreover, since the operator commutes with trans-
lations, also the translated homogeneous functions are eigenfunctions. How-
ever, there are also other eigenfunctions. For example supn∈Z |x − n|−a is
an eigenfunction with the same eigenvalue as |x|−a . Their result from [10]
is formulated in what follows:
Theorem. Let f (x) be an eigenfunction of the uncentered maximal operator
with eigenvalue λ and with a single peak at a point c. Also, let 1 < λ, p < +∞
be related by equation (p − 1)λp − pλp−1 − 1 = 0. Then

1) If c = a or c = b then f (x) = γ|x − c|(1−λ)/λ .

2) If a < c < b and |x − c| < min(c − a, b − c), then f (x) = γ|x − c|−1/p ,

where γ is an arbitrary constant.

Properties of maximal functions

Some properties of the maximal functions are partially inherited from the
properties of the maximal operators. Nevertheless, there is a number of
general properties which are worth mentioning.
Let f ∈ Lploc (Rd ), p ≥ 1. Then,

a) M f is measurable,

b) M f < ∞ almost everywhere,

c) for all α > 0, M f satisfies the following inequality

α |{x : Md f (x) ≥ α}| ≤ A∥f ∥L1 , (1.13)

where A is a constant which depends only on the dimension d.

1.3. PROPERTIES OF THE MAXIMAL OPERATOR 9

d) If f ∈ Lp (Rd ), with 1 < p ≤ ∞, then M f ∈ Lp (Rd ) and

∥M f ∥p ≤ Ap ∥f ∥Lp . (1.14)

The last four facts are in fact the Hardy-Littlewood (or sometimes more
general Hardy-Littlewood-Wiener) maximal theorem. The original Hardy-
Littlewood maximal theorem (for d = 1) is presented in the beginning of
this text.
The inequality (1.13) is called weak type inequality because it is weaker
than the corresponding inequality for L1 norm. However in recent years
evidence has been mounting to the effect that not only weak type (1, 1)
inequalities are formally stronger than strong (p, p) inequalities for 1 < p <
∞ (since the latter are implied by the former via interpolation) but they are
also stronger in a substantial way, meaning that the strong type may hold for
all p > 1 while the weak-type (1, 1) may fail. This is the case, for instance,
with the uncentered maximal function associated to the standard Gaussian
measure and Euclidean balls. It is shown in [30] that this maximal function
is not of weak type (1, 1), while it is of strong-type (p, p) for all p > 1; cf.
[13] (for cubes the strong (p, p) type follows from a more general result in
[9] cf. Th. 1).
The function g(λ) = |{x : Md f (x) ≥ λ}| defined to be the measure of
this set is called the distribution function of f (x).
The first property can easily be verified for the Lebesgue measure if we
observe thet the set

Eλ = {x ∈ Rd ; M f (x) > λ}

is open. If x′ ∈ Eλ then there exists a ball B such that x′ ∈ B and

∫
1
|f (y)|dy > λ.
µ(B) B

Now any point x close enough to x′ will also belong to B, hence x ∈ Eλ as

well.
The proof of inequality (1.13) is based on the following Vitali-type cov-
ering lemma:
Lemma. Let B = {B1 , B2 , . . . , Bn } be a finite collection of open balls in Rd ,
not necessarily disjoint. Then there is a disjoint sub-collection Bi1 , Bi2 , . . . , Bik
10 CHAPTER 1. INTRODUCTION

of B that satisfies
∪
N ∑
k
µ( Bl ) ≤ 3d µ(Bij ).
l=1 j=1

(For a proof of this lemma one may refer, for example, to [33]).
For every x ∈ Eλ there exists a ball Bx that contains x and such that
∫
1
|f (y)|dy > λ.
µ(Bx ) Bx

Therefore, for each ball Bx we have

∫
1
µ(Bx ) < |f (y)|dy. (1.15)
λ Bx
∪
Fix a compact subset K of Eλ . Since K is covered by x∈Eλ Bx we
∪
may select a finite subcover of K, say K ⊂ N l=1 Bl . The covering lemma
guarantees the existence of a sub-collection Bi1 , Bi2 , . . . , Bik of disjoint balls
with
∪
N ∑
k
µ( Bl ) ≤ 3d µ(Bij ). (1.16)
l=1 j=1

Since the balls Bi1 , Bi2 , . . . , Bik are disjoint and satisfy (1.15) as well as
(1.16), we find that
(N ) k ∫
∪ ∑
k
3d ∑
µ(K) ≤ µ Bl ≤ 3d m(Bij ) ≤ |f (y)|dy
l=1 j=1
λ j=1 Bij

∫ ∫
3d 3d
∪k |f (y)|dy ≤ |f (y)|dy.
λ B
j=1 ij
λ Rd

One can obtain the strong-type inequality from the weak-type inequality.
Using a property of the distribution function and integration by parts one
gets ∫ ∫ ∫ ∞ ∞
(M f )p dx = − λp dg(λ) = p λp−1 g(λ)dλ.
Rd 0 0
In particular, due to (1.13)
∫ ∫ ( ∫ )
∞ ∞ 2A
∥M f ∥pp =p λp−1
µ(Eλ )dλ ≤ p λ p−1
|f (x)|dx dλ.
0 0 λ |f |≤λ/2
1.3. PROPERTIES OF THE MAXIMAL OPERATOR 11

The double integral is evaluated by interchanging the orders of integration

and integrating first with respect to λ. The inner integral is then
∫ 2|f (x)| ( )
1
λ p−2
dλ = |2f (x)|p−1 ,
0 p−1
since p > 1. So the outer integral gives
∫ ∫
2Ap
|f ||2f | p−1
dx = (Ap ) p
|f |p dx,
p−1 Rd Rd

which proves the conclusion.

The proof of weak-type inequalities can be quite a challenge for more
general measure. But some generalizations are known. For example one
can use Besicovitch covering lemma to proof the existence of weak-type
inequality for the Borel measure.
Remark. In general an operator is said to be of weak-type (p, p) if for
any f ∈ Lp (Rd ) we have

αp |{x : Md f (x) ≥ α}| ≤ Cd ∥f ∥Lp , (1.17)

with the constant Cd that depends only on dimension d and of strong-type

(p, p) if for all f ∈ Lp (Rd ) we have

∥M f ∥p ≤ Cp,d ∥f ∥p .

Here Cp,d depends only on d and p.

It should be noted that in some works the weak-type (1, 1) operators are
defined more precisely as acting from L1 (Rd ) into L1,∞ (Rd ). Here L1,∞ is
defined as a metric space X with a measure µ as follows:

L1,∞ (µ) := {f : X → C : ∥f ∥L1,∞ (µ) := sup t·µ{ x ∈ X : |f (x)| > t} < +∞}.
t>0

Note that the ’norm’ of L1,∞ (µ) is not actually a norm since it does not
satisfy the triangle inequality.
Remark. In the case of the Lebesgue measure, the Hardy-Littlewood
maximal function the strong-type (p, p) inequality implies the weak-type
(p, p) due to the following
∫ ∫ ( )p
(M f (x))p c∥f ∥p
|Eλ | = |{x : M f (x) > λ}| = χEλ (x)dx ≤ ≤ ,
λp λ
12 CHAPTER 1. INTRODUCTION

where p ≥ 1.
Proposition. The maximal function M f (x) of a function f ∈ L1 (Rd )
is not in L1 (Rd ), unless f = 0 almost everywhere in Rd .
This is true, because M f (x) decays too slowly at infinity for its integral
to converge. To show this, let a > 0 and suppose that |x| ≥ a. Then, by
considering the average of |f | at x over a ball of radius r = 2|x| and using
the fact that B2x (x) ⊂ Ba (0), we see that
∫ ∫
1 C
M f (x) ≥ |f (y)|dy ≥ |f (y)|dy,
B2|x| (x) B2|x| (x) |x|d Ba (0)

where C > 0. The function |x|1 d is not integrable on Rd \ Ba (0) so if M f is

integrable then ∫
C
|f (y)|dy = 0
|x|d Ba (0)
for every positive a, which implies f = 0 almost everywhere in Rd .
Moreover, as the following example shows, the maximal function of an
integrable function does not need not even be locally integrable [20].
The remarkable property of Hardy-Littlewood maximal operator is that
in the strong-type (p, p) inequality one can choose a constant Cp , indepen-
dent on dimension d. This question will be considered in more details in the
next chapter.

1.4 Applications
The maximal function has a lot of applications in different areas of mathe-
matics, mostly in analysis, but some applications can be found in stochastics
and other branches of mathematics. Here we list the most well-known and
the most useful ones.

Differentiation theorems
One of the most important applications of the maximal function is to prove
the Lebesgue differentiation theorem.
Theorem. (Lebesgue differentiation theorem) If f is an integrable function
on Rd then
∫
1
lim f (y)dy = f (x), for almost every x. (1.18)
|B|→0,x∈B |B| B
1.4. APPLICATIONS 13

Proof. It suffices to show that for each α > 0 the set

∫
1
Eα = {x : lim sup f (y)dy − f (x) > 2α}
|B|→0,x∈B |B| B

has the zero measure. This assertion then guarantees that the set E =
∪∞
n=1 E1/n has the zero measure , and the limit in (1.18) holds for all points
of E c .
We fix α, and we will use the fact that for each ϵ > 0 we may select
a continuous function g of compact support with ∥f − g∥L1 (Rd ) < ϵ. The
continuity of g implies
∫
1
lim g(y)dy = g(x), for all x.
|B|→0,x∈B |B| B

1 ∫
Since we may write the difference |B| B f (y)dy − f (x) as
∫ ∫
1 1
(f (y) − g(y))dy + g(y)dy − g(x) + g(x) − f (x)
|B| B |B| B

we find that
∫
1
lim sup | f (y)dy − f (x)| ≤ M (f − g)(x) + |g(x) − f (x)|.
|B|→0,x∈B |B| B

Consequently, if
Fα = {x : M (f − g)(x) > α},
and
Gα = {x : |f (x) − g(x)| > α},
then Eα ⊂ (Fα ∪Gα ) because if two arbitrary variables u1 and u2 are positive
then u1 + u2 > 2α only if ui > α for at least one ui . On the other hand
Tchebyshev’s inequality yields
1
|Gα | ≤ ∥f − g∥L1 (Rd ) ,
α
and on the other hand, the weak type estimate for the maximal function
gives
A
|Fα | ≤ ∥f − g∥L1 (Rd ) .
α
14 CHAPTER 1. INTRODUCTION

The function g was selected such that ∥f − g∥L1 (Rd ) < ϵ, hence we get

A 1
|Eα | ≤ ϵ + ϵ.
α α

Since ϵ is arbitrary, we must have |Eα | = 0 and the proof of theorem is

complete.
There is a number of other theorems, such as Rademacher differentiation
theorem that can be proven using maximal function.

Poisson kernels and Fatou teorem

Maximal functions are widely used to estimate and prove almost everywhere
convergence, for example for Poisson integrals.
We define the Poisson integral in an upper half space {(x, y) ∈ Rn+1 |y >
0} as follows. Let f ∈ L2 (Rd ) and let fˆ be it’s Fourier transform. Then,
∫
u(x, y) = fˆ(t)e−2πit·x e−2π|t|y dt, y > 0 (1.19)
t∈Rn

is called the Poisson integral of function f and

∫
y
e−2πit·x e−2π|t|y dt = cn
t∈Rn (y 2 + |x|2 )(n+1)/2

is called the Poisson kernel. Here cn stands for

Γ ((n + 1)/2)
cn = .
π n+1 /2

The following theorem describes some properties of the Poisson integral.

Theorem. Suppose f ∈ Lp (Rd ), 1 < p < ∞, and let u(x, y) be it’s poisson
integral as defined in (1.19). Then

a) supy>o |u(x, y)| ≤ M f (x), where M f is the Hardy-Littlewood maximal

function defined in (1.10).

b) limy→0 u(x, y) = f (x), for almost every x.

c) If 1 ≤ p ≤ ∞, u(x, y) converges to f (x) in the Lp (Rd ) norm, as y → 0.

1.4. APPLICATIONS 15

The proof of the theorem is usually contained in a proof of a more gen-

eral theorem, valid for a large class of approximations to the identity.
Theorem. Let φ be an integrable function on Rd , and set φϵ (x) =
ϵ−d φ(x/ϵ). Suppose there∫
exists at least one integrable function ψ(x) =
sup|y|≤|x| |φ(y)| and let Rd |φ(x)|dx = A < ∞. Then

a) supϵ>0 |f ∗ φϵ | ≤ AM f (x), f ∈ Lp (Rd ), 1 ≤ p ≤ ∞.

∫
b) If in addition Rd φ(x)dx = 1, then limϵ→0 (f ∗ φϵ )(x) = 0, almost every-
where.

c) If 1 ≤ p < ∞, then ∥f ∗ φϵ ∥Lp → 0 as ϵ → 0.

Here and in what follows ∗-product means the convolution.

Poisson integrals provide us with an extension of solutions to a Laplace
equation with given Dirichlet boundary conditions for unit circle. Thus, it
lies in the very base of the entire theory of harmonic analysis.
Having the above in hand one can prove the following fundamental the-
orem of harmonic analysis
Theorem. (Fatou’s theorem) Suppose u is harmonic and bounded in Rd+1 + .
Then u has a non-tangential limits at almost every point on the boundary
(Rd ) of Rd+1
+ .
We also have the following result
Corollary Suppose u is harmonic and bounded in Rd+ , and 1 ≤ p ≤ ∞.
If supy>0 ∥u(·, y)∥Lp (Rd ) < ∞ , then u is the Poisson integral of a function
f ∈ Lp (Rd ) if p > 1. If p = 1, then u is the Poisson integral of a finite
measure.

Pointwise convergence of operators

There is a number of important problems in analysis and in other areas in
which a sequence (or generalized sequence) of operators arise in a natural
ways. Those are, for example partial sums of Fourier series where each
partial sum can be represented as a convolution Dk ∗ f , where Dk is the
Dirichlet kernel, defined as

sin(k + 1/2)x
Dk (x) = ,
2π sin(x/2)
16 CHAPTER 1. INTRODUCTION

or Cezaro sums, Fk ∗ f , where Fk is the Fejér kernel

1 sin(k + 1/2)x
Fk = ,
2π(k + 1) 2π sin(x/2)
and many others. The most natural question in these situations is to find
out whether, or under which additional non trivial conditions on f or on
the operators Tk , the corresponding sequence Tk f (x) converges for every x
or almost every x and which are the properties of their limit T .
To answer this question let us consider the properties of the set A(f, λ)
defined for a linear operator T as follows

A(f, λ) = {x ∈ Rd : lim sup |Tp f (x) − Tq f (x)| > λ}.

p,q→∞

Let us take for example a function f ∈ L1 (Rd ). Then for λ > 0 we would like
to prove that |A(f, λ)| = 0. This would give us the convergence of {Tk f } at
almost every x ∈ Rd . Assume that for some function g we know that {Tk g}
converges for every x ∈ Rd . Then if h = f − g we have the following

|A(f, λ)| = |{x ∈ Rd : lim sup |Tp h(x) − Tq h(x)| > λ}| = |A(h, λ)|,
p,q→∞

hence the problem is reduced to prove that A(h, λ) is of small measure if h

is of small L1 (Rd )-norm.
However, the set A(f, λ) has a rather unhandy structure and so one can
think of substituting it with one easier to handle. It is clear that

|A(f, λ)| ≤ |{x : sup |Tk f (x) > λ/2|}| = |A∗ (f, λ/2)|
k

and that the operator T ∗ defined by T ∗ f = supk Tk f has a rather simple

structure. The operator T ∗ is called the maximal operator associated to
{Tk }.
Hence after consecutive reductions, our initial problem has been trans-
formed into following: prove that |A∗ (f, λ)| → 0 as ∥f ∥1 → 0. This will as
well give us our desired almost everywhere convergence of {Tk f }.
The very important example of this technique is a mollification and ap-
proximation to identity. These are very useful tools in analysis that provides
us with a sequence of functions or operators which are more regular than
the original ones.The mollification procedure can be described as follows.
Let Φ be a fixed nonnegative function with the following properties:
1.4. APPLICATIONS 17

1) Φ is compactly supported in Rd ,
∫
2) Rd Φ(x) dx = 1,

3) limϵ→0 Φϵ (x) = limϵ→0 ϵ−d Φ(x/ϵ) = δ(x).

Then Φ is called mollifier and fulfills the following inequality:

sup |f ∗ Φϵ (x)| ≤ M f (x).

ϵ

Obviously, the properties of the function f ∗Φϵ depend on both f and Φϵ right
up to ϵ = ∞, since due to the third property we have that limϵ→∞ f ∗Φϵ = f .
The functions Φϵ can be used as kernels to the approximation to the
identity. This procedure can be summarized as follows

lim Tϵ (x) = lim T ∗ Φϵ (x) = T (x).

ϵ→0 ϵ→0
Chapter 2

Estimations of the best

constants in weak-type and
strong-type inequalities for
the maximal function

Within a wide range of arguments, the maximal functions satisfy maximal

inequalities with different constants for each case. We will define a specific
maximal function for every case and present the results obtained up to date.
The problem of existence of such inequalities will be considered as well.

2.1 The case of R1

We start with the one dimensional case because there is not so much free-
dom to choose the collection of sets of R. Apparently, they will always be
segments, but the result will strongly depend on the measure we choose. On
the other hand, for some cases it is possible to present a sharp estimate, i.e.
the best constant. The sharp estimate means that it is impossible to choose
any constant smaller (or bigger) than the one presented.

19
20 CHAPTER 2. ESTIMATIONS OF THE BEST CONSTANTS

Weak-type, uncentered maximal function

Probably, the easiest case we have is the uncentered maximal function which
is defined as follows
∫
1
M f (x) := sup |f | dµ. (2.1)
x∈I µ(I) I

Here the supremum is taken over all intervals I containing x. It was proven
by Bernal in [4], that the best constant that appears in the weak-type in-
equality (1.13) is equal to 2 exactly if µ is the Lebesgue measure. Moreover
the best constant was found for an arbitrary positive Borel measure as well.
The proof is done by using a specific covering lemma. The proof itself is
short and elegant so we present here.
We start from introducing a new, so called ’right’ maximal function M+ ,
associated to a specific set I of intervals of the type [x, x + h), for h > 0.
Let λ > 0 and A be the set of all x with M+ f (x) > λ.
The set A can be seen to be a countable union of pairwise disjoint open
intervals (an , bn ) and of point cn , each cn being a certain bm . Let B be any
finite union of [dn , bn ), an < dn < bn , and of {cm }. Clearly, if the following
is proven
∥f ∥L1
µ(B) ≤ (2.2)
λ
then the weak type estimate for M+ with constant 1 will be established.
Proposition. M+ is of a weak-type (1, 1) with constant 1.
Proof. If x ∈ B, there is hx > 0 so that
∫
|f |dµ > λµ(I(x)),
I(x)

with I(x) = [x, x + hx ). Denote by I the set of all I(x), x ∈ B.

The covering lemma is now
Lemma. There is a countable collection {Ik }k , with Ik ∈ I pairwise
disjoint and such that ∪
B ⊂ Ik .
Once the lemma is proved, the proof of (2.2) follows most usual steps
∪ ∑ ∥f ∥L1
µ(B) ≤ µ( Ik ) = µ(Ik ) ≤
λ
2.1. THE CASE OF R1 21

and that will be the end of the proof.

Proof. Let T be the family of all nonempty countable subsets T of I that
are made up by pairwise disjoint intervals and have the following property: if
∪ ∪
x ∈ B \ T , then x ≤ sup T . Since {Iα } belongs to T if α = min{dk , cj },
the set T is not empty. Also it is clear that the inclusion order of T is
inductive, so that we can consider a maximal element T0 . T0 is easily seen
to cover B.
End of proof.
If one would consider the ’left’ function M− associated correspondingly
with the sets (x − k, x], k > 0 then the result will be the same.
Proposition. M− is of weak type (1, 1) with constant 1.
Now consider M defined by (2.1) . By easy convergent arguments, we can
restrict the supremum to intervals of the form (x − k, x + h), k, h > 0. We
have
M ≤ M+ + M− ,

so we obtain the following

Corollary. M is of weak-type (1, 1) with constant 4.
and
Corollary. For continuous measures, if µ({x}) = 0 for each x ∈ R then M
is of weak-type (1, 1) with constant 2.

Later this result was confirmed by Melas [27] who claimed that the same
constant can be obtained by considering a single delta function. The moti-
vation of replacing continuous functions with sums of Dirac deltas will be
provided later in this text.
As one can see form the result of Bernal, the constant in the inequality
may vary, depending on the classes of functions. For example in the work of
Grafakos, Montgomery-Smith and Motrunich [15] the following result, based
on research of José Barrionuevo, was mentioned. For a class of positive
functions increasing on (−∞, c) and decreasing on (c, ∞) the best constant
is equal to 1 and this is sharp.
Osȩkowski, in his work [29] considered a general case of (q, p) weak-type
inequalities for maximal functions associated to general Borel measures. For
the one dimentional case he formulated the following theorem.
Theorem. For any locally integrable function f : R1 → R1 , any Borel
subset A of R1 , any 1 ≤ p < ∞, q ∈ (0, p] and the Borel measure µ the
22 CHAPTER 2. ESTIMATIONS OF THE BEST CONSTANTS

following inequality holds

∥Mµ f ∥Lq,∞ (A,µ) ≤ Cp ∥f ∥Lp (R,µ) µ(A)1/q−1/p ,

where
(p − 1)(2p/(p−1) − 1) ( )−1/p
Cp = (p − 1)(2p/(p−1) − 2) ,
p
when 1 < p < ∞ and C1 = 2.
If µ is a the Lebesgue measure then the constant Cp is the best possible.
The proof provided in [29] is rather brief but technical.

Weak-type, centered maximal functions

The search for the best constant in
C
|{M f > λ}| ≤ ∥f ∥L1 , (2.3)
λ

was made by Melas in [27].

His approach uses as a starting point the discretization technique intro-
duced by M. de Guzmàn [17] and sharpened by M. Trinidad Menàrguez-F.
and Soria (see Theorem 1 in [37]). For this technique the corresponding
maximal function ∫
1 x+h
M σ(x) = sup |dσ|.
h>0 2h x−h

should be defined for any finite measure σ on R1 . Then, the best constant
C in inequality (2.3) is equal to the corresponding best constant in the
inequality
∫
1 x+h
|{M µ(x) < λ}| = sup dµ
h>0 2h x−h
∑
where λ > 0 and µ runs through all measures of the form di=1 δti where
n ≥ 1 and t1 , . . . , tn ∈ R1 . This technique allows one to apply arguments
of combinatorial nature to get information on bounds for this constant. In
(see [27]) the following estimates were
√
11 + 61 5
1.5675208 . . . = ≤ C ≤ = 1.66 . . . (2.4)
12 3
2.1. THE CASE OF R1 23

and also made the conjecture that the lower bound in (2.4) is actually the
exact value of C. In [27] Melas has also found the best constant in a re-
lated but more general covering problem on the real line. This implies the
following improvement of the upper bound in (2.4)
√
11 + 61
C= = 1.57735 . . . .
12
None of these however tells us what the exact value of C is. In his paper
the author proved that the above conjecture is correct thus settling the
problem of the computation of the best constant C completely. The author
formulated his result as follows.
Theorem 1. For the centered Hardy-Littlewood maximal operator M ·, for
every measure µ of the form k1 δy1 +· · ·+kn δyn , where ki > 0, for i = 1, . . . , n
and y1 < · · · < yn and for every λ > 0 we have
√
11 + 61
λ|{M µ > λ}| ≤ ∥µ∥
12
and this inequality is sharp.
In view of the discretization technique described above, Theorem 1 im-
plies the following result.
Corollary 1. For every f ∈ L1 (R1 ) and for every λ > 0 we have
√
11 + 61
λ|{M f > λ}| ≤ ∥f ∥1
12
and this inequality is sharp. Hence
√
11 + 61
C= = 1.5675208...
12
is the largest solution of the quadratic equation

12C 2 − 22C + 5 = 0.

The author also generalized the above to general finite Borel measure.
Theorem 2. For any finite Borel measure σ on R1 and for any λ > 0
we have √
11 + 61
λ|{M σ > λ}| ≤ ∥σ∥.
12
24 CHAPTER 2. ESTIMATIONS OF THE BEST CONSTANTS

Theorem 3. For any measure µ that is a positive linear combination

of Dirac deltas and for any λ > 0 we have
√
11 + 61
λ|{M µ > λ}| < ∥µ∥.
12

Strong-type, uncentered maximal function

This case was studied in [15] and the result was presented in the form of the
operator norm of the maximal operator associated with the the Lebesgue
measure.
Theorem. For 1 < p < ∞, the operator norm of M1 : Lp (R1 ) → Lp (R1 )
is the unique positive solution of the equation

(p − 1)xp − pxp−1 − 1 = 0.

The proof is based on an idea that is somewhat alike to the one in the proof
of Bernal, for the best constant of the weak-type inequality and based on
introduction of two auxiliary "left" and "right" maximal functions
∫ x
1
(M− f )(x) = sup f (t)dt
a<x x − a a

and ∫
1 b
(M+ f )(x) = sup f (t)dt.
b>x b − x x

For the proof of the next result, known as the "Sunrise lemma", we refer the
reader to Lemma (21.75) (i), Ch VI in [19].

Strong-type, centered maximal function

In paper [11] the fundamental research on properties of Hardy-Littlewood
maximal operator was undertaken. The behavior of eigenfunctions and
eigenvalues was studied. The interesting result is formulated below. The
constant Cp in the strong-type (p, p) inequality for the centered maximal
function can be expressed as follows
p−1 p−1
(y + 1) p + (y − 1) p
Cp = p−1 ,
2y p
2.2. THE CASE OF HIGHER DIMENSION 25

where y > 1 is the unique solution of

( )p ( )p
y y
1− (y + 1) − 1 + (y − 1) = 0.
p p
Here y represents the radius (half length) of the interval where the actual
maximal value is reached.
This formula supposes to give the best constant for a strong-type (p, p)
inequality for all p and f from the corresponding space Lp (R1 ). How-
ever in the work [16] mentioned earlier, Grafakos, Montgomery-Smith and
Motrunich doubted that (Lp ). They restricted this result onto the class
of "peak shaped" functions: functions f on R1 , which are convex except
at one point (where f is allowed to be discontinuous). Moreover they also
marked a specific point in their proof where generalization should brake.
They also have shown that the result presented for a constant in one dimen-
sional weak-type inequality for this class of functions is two times smaller
than for a general case, thus the problem of presenting the best constant in
case of general f remains open.

2.2 The case of higher dimension

This case is much more complicated partially due to the freedom to choose
the collection of sets for averaging. In certain cases only the existence of
inequalities for maximal operator is proved.

Strong-type, uncentered maximal function

If the dimension d ≥ 2, the uncentered maximal function associated to an
arbitrary measure is not bounded on Lp (Rd ), for 1 < p < ∞ in general. To
see this, like in [14], select closed balls B1 , B2 , . . . so that the origin is on
the boundary of each ball and such that for every Bi , i = 1, 2, ..., there is a
point xi ∈ Bi ∪j̸=i Bj. Set
∞
∑
µ= δxi ,
i=0
where x0 = 0 and δxi denotes Dirac delta at xi . Let χ be the characteristic
function of B1 . Clearly ∥χ∥Lp (µ) = ∥χ∥p,µ ≤ 21/p , but
∫
1 1
M χ(xi ) ≥ χdµ ≥ , for all i = 1, 2, . . .
µ(Bi ) Bi 2
26 CHAPTER 2. ESTIMATIONS OF THE BEST CONSTANTS

In this context maximal function is defined as follows. For a locally

integrable function f on Rd let
∫
1
M f (x) := sup |f (y)| dµ, (2.5)
Q∋x |Q| Q

and hence
1
∥M f ∥p,µ ≥ µ(Rd )1/p = ∞.
2
A similar counterexample for the strong-type inequality for the maximal
operator was given by Fefferman [12].
The result for the case of strong maximal function was provided by
Grafakos and Montgomery-Smith in [15]. In their work they studied the
norm of the uncentered maximal operator in one dimension and have proved
the following result.
Theorem. For 1 < p < ∞, the operator norm of M : Lp (R1 ) → Lp (R1 ) is
the unique positive solution of the equation

(p − 1) xp − p xp−1 − 1 = 0.

Then, they generalized their result to the case of d-dimensional strong

maximal function (1.11) and proved that, for 1 < p < ∞, the operator
norm of the d-dimensional strong maximal function Nd : Lp (Rd ) → Lp (Rd )
is (cd )p , where cd is the norm of the one dimensional maximal function.

Strong-type, centered maximal function

The general case of a uniformly scaled arbitrary convex body was initially
considered by Bourgain in [7] for f ∈ L2 (Rd ) where the following result was
obtained:
Proposition. Define the scaling Bρ of an arbitrary convex body B centered
at x ∈ B as
Bρ = {yρ : yρ = x + ρ(y − x) ∀y ∈ B]}
Then the maximal function MB can be defined as
∫
1
MB f (x) := sup |f (y)| dµ, (2.6)
ρ |Bρ | Bρ

The operator norm of MB : L2 (Rd ) → L2 (Rd ) fulfills

∥MB ∥2→2 ≤ D,
2.2. THE CASE OF HIGHER DIMENSION 27

where D is a constant independent on B and the dimension d.

The main arguments of the proof rely on geometrical properties of convex
bodies, namely Brunn’s theorem [6] and techniques from Fourier analysis,
to estimate maximal functions.
In the paper [8] Bourgain observed that the previous estimate implies
automatically that
∥MB ∥p→p ≤ D, 2 ≤ p ≤ ∞.
It appears that using estimations for diadic maximal operator the range of
the parameter p can be extended to p > 3/2.
Further improvement of this result could be found in the work of Müller
[28], where the general case of f ∈ Lp , p > 1 is considered. There he
introduces linear invariants σ(·) and Q(·) of B and proves the following
theorem
Theorem. Let p > 1, then for all f ∈ Lp (Rd )

∥MB f ∥p ≤ C(p, σ(B), Q(B))∥f ∥p ,

where the constant C = C(p, σ(B), Q(B)) is independent on d and grows

with σ and Q. For the case of p > 3/2 the constant C may be chosen
independent on σ and Q.
Stein, in his work [32] has summed up similar results.
Theorem. Let maximal function be defined by
∫
1
MR f (x) := sup f (x − y) dµ(y) ,
R⊃R,R∋x µ(R) R

where R ranges over a suitable collection R of sets. Then inequality

∥MR f ∥p ≤ Ap ∥f ∥p (2.7)

holds in one of the following cases:

a) R is the collection of spheres centered at the origin; dµ is the uniform

surface measure; and d ≥ 3, with p > d/(d − 1).

b) R is the collection of initial segments {γ(t), 0 ≤ t ≤ h} of a smooth curve

t → γ(t), with γ(0) = 0, and γ having non zero "curvature" at the origin;
here dµ is the arc-length, d ≥ 1 and p > 1 .
28 CHAPTER 2. ESTIMATIONS OF THE BEST CONSTANTS

c) R is the collection of rectangles (in R2 ) containing the origin, which make

an angle θk with a fixed direction, where{θk } is a sequence of numbers
tending rapidly to zero; here p > 1 .
Part a) was proved by Stein himself in [31]. There, a spherical averaging
function is considered
∫
Mt f (x) = f (x − ty)dσ(y).
|y|=1

Here σ is the usual measure on a sphere and

M(f ) = sup Mt f (x)
t>o

is the maximal spherical function. The following result is known as Stein’s

spherical theorem.
Theorem. Let f be a function of Schwartz space S(Rd ) and suppose d ≥ 3.
Then the inequality
∥M(f )∥p ≤ Ap ∥f ∥p (2.8)
holds whenever d/(d − 1) < p ≤ ∞. If p ≤ d/(d − 1) the inequality (2.8) is
not valid.
It is easy to see that when d = 1, only the trivial case (p = ∞) of (2.8)
holds. When d = 2 and 1 < p ≤ 2 the inequality fails. The case d = 2 (and
2 < p) remains open.
To see that (2.8) cannot hold if p ≤ d/(d − 1),
{
|x|−d+1 (log(1/|x|)−1 , when 0 < |x| ≤ 1/2
F (x) =
0, when |x| > 1/2.

Then F ∈ Lp (Rd ), if p ≤ d/(d − 1), but supt>0 |Mt (F )(x)| = ∞ everywhere.

Thus, a simple limiting argument shows that inequality (2.8) cannot hold
for those p.
The proof of a the first part of the theorem is more involved. One can
refer to the original paper for more details. Note that the analogue of this
theorem with a sphere replaced by a cube is also false.
For part b) one could refer to [35]. There the following theorem is stated
and proved.
Theorem. If γ(t) is well-curved then for maximal function defined as
∫ h
Mf (x) = sup |f (x − γ(t))|dt
0<h≤1 0
2.2. THE CASE OF HIGHER DIMENSION 29

the following inequality holds

∥Mf ∥p ≤ Ap ∥f ∥p . (2.9)

By γ being well-curved we mean:

• In R2 , any curve γ(t) with non-vanishing curvature at the origin is
well-curved.

• In R3 , any curve γ(t) with non-vanishing curvature and torsion at the

origin is well-curved.

• In Rd , any analytic curve γ(t) is well curved.

Part c) was considered in the work [36] of J.-O. Strömberg where the
following is claimed. Let R(ϕ) be the family of rectangles S in R2 such that
the angle between the longest side of S and the x1 -axis is ϕ. We call ϕ the
direction of S.
For a given a set Φ of directions we define the maximal function MΦ f
by ∫
1
MΦ f (x) = sup |f (y)|dy,
S µ(S) S
where the supremum is taken over all rectangles S in the families R(ϕ),
ϕ ∈ Φ, containing the point x. Here µ is the Lebesgue measure in R2 .
Only countable sets Φ are considered with {ϕi }∞i=l satisfying the following
condition: |ϕj − ϕi | > C|ϕi − ϕ∞ |, for some constant C > 0. Here ϕ∞ =
limi→∞ ϕi . Such sets Φ are called exponential.
For this type of the maximal function the two following theorems hold.
Theorem 1. Let χE be the characteristic function of the measurable set
E in R2 . Then

µ{x : MΦ χE > α} ≤ Cα−2 (1 + log(1/α))2 µ(E)

for all 0 < α ≤ 1. The constant C depends only on Φ.

and
Theorem 2. Let ϵ > 0 and let f be a function on R2 . Then
∫
µ{x : MΦ f > α} ≤ C |f (y)/α|2 (1 + log + |f (y)/α|)4+ϵ dy
R2

for all α > 0, if the integral is finite. The constant C depends only on ϵ and
Φ.
30 CHAPTER 2. ESTIMATIONS OF THE BEST CONSTANTS

These weak type estimates are slightly weaker (with some logaritms)
than the weak-type estimates of type (2, 2) and will thus imply that the
maximal operators are bounded in Lp for p larger than 2.
The proof of the first theorem involves some geometric-type covering
arguments while the second theorem follows from the first one.
The idea of generalization of the result obtained in part a) of the theorem
to maximal function associated with balls can be found in the original Stein’s
paper [31]. There he conjectures the following:
Theorem. For Hardy-Littlewood maximal function in Rd the following
inequality holds
∥M (f )∥p ≤ Ap ∥f ∥Lp ,
where Ap is independent on the dimension d.
The proof, in details appears in the appendix of [34], where the rotation
method of Calderón-Zygmund is used by considering special weighted max-
imal functions of the following type
∫
| |y|≤r f (x − y)|y|m dy|
Mk,m (f )(x) = sup ∫ .
r>0 |y|≤r |y|m dy

In the work of Grafakos and Kinnunen [14], there were derived the best
constant for a strong-type (p, p) inequality for the maximal function associ-
ated to balls and the Borel measure as a solution of the following equation

(p − 1)xp − pxp−1 − (cp − 1) = 0,

where cp is a Besicovitch constant [5].

Weak-type, uncentered maximal function

Let us note that the definition of the d-dimensional strong maximal function
(1.11) is equivalent to
∫ b1 ∫
a1 · · · abdd f (y1 , . . . yd )dy1 . . . dyd
Nd f (x) = sup . . . sup .
a1 <x<b1 ad <x<bd (b1 − a1 ) · · · (bd − ad )

Osȩkowski in [29] generalizes his one dimensional result for a general

(q, p) weak-type inequality for higher dimensions and presents it the follow-
ing form
Theorem. Let Mµ f (x) be a strong maximal function defined as above then
2.2. THE CASE OF HIGHER DIMENSION 31

i) If d ≥ 2, then, in general, Mµ does not map L1 (Rd , µ) into L1,∞ (Rd , µ)

ii) if 1 < p < ∞, then for any f : Rd → R we have

∥Mµ f ∥Lp,∞ (Rd ,µ) ≤ Cp cd−1

p ∥f ∥Lp (Rd ,µ)

if µ is the Lebesgue measure on Rd then the constant has the optimal

order O((p − 1)1−d ) as p → 1.

iii) if 1 < p < ∞, then for any f : Rd → R we have

∥Mµ f ∥Lp,∞ (Rd ,µ) ≤ cd−1

p ∥f ∥Lp,∞ (Rd ,µ)

If µ is the Lebesgue measure on Rd then the constant is the best possible

The constants Cp and cp refer to the one dimensional case and have the
following meaning. The constant Cp represents the best possible constant
for the weak-type (q, p) inequality associated with Borel measure and has
the form

(p − 1)(2p/(p−1) − 1) ( )−1/p
Cp = (p − 1)(2p/(p−1) − 2)
p

and cp is the Lp -norm of the one-dimensional strong maximal operator de-

rived by Grafakos and Montgomety-Smoth in [15] and is the unique positive
solution of the equation

(p − 1)xp − pxp−1 − 1 = 0.

Weak-type, centered maximal function

The obvious upper bound equal to 3d that arises in the proof of maximal the-
orem appears to be too pessimistic. One can find much more precise result
in the work [34] of Stein and Strömberg where they proved the following

1) For dilate Dr of an arbitrary open, bounded, convex, and symmetric set

B ∈ Rd , defined as Dr = {x : r−1 x ∈ D}, r > 0, the centered maximal
function fulfills
c
|{x : M f (x) > λ}| ≤ d log d∥f ∥L1 .
λ
32 CHAPTER 2. ESTIMATIONS OF THE BEST CONSTANTS

2) For a ball Br ∈ Rd centered at the origin with radius r > 0 the corre-
sponding inequality is
cd
|{x : M f (x) > λ}| ≤ ∥f ∥L1 .
λ
The first part of the theorem was proved using covering-type arguments,
while the second part part by using of a majorization and results obtained
on heat-diffusion semi-group. We refer to original paper [34] for details.
In his monograph [17] Guzman devoted the entire Chapter 4 to study
how maximal operators interact with sums of Dirac deltas. The central
result of his work is formulated in the following theorem.
Theorem. Let {kj }∞ j=1 ⊂ L (R ) be an ordinary sequence of functions,
1 d

{Kj }j=1 be the sequence of convolution operators associated to it and K ∗ be

∞

the corresponding maximal operator. Then K ∗ is of weak-type (1, 1) if and

only if K ∗ is a weak-type (1, 1) over finite sums of Dirac deltas.
In other words (forgetting Dirac deltas) K ∗ is of weak-type (1, 1) if and
only if there exists c > 0 such that for each finite set of different points
a1 , a2 , . . . , aH ∈ Ω and for each λ ≥ 0 we have
∑
H
H
|{x ∈ Ω : sup | kj (x − ah )| > λ}| ≤ c .
j h=1
λ
Here Ω stands for either n-dimensional torus Tn or Rn (but could, as well,
be a locally compact group).
Later the result based on this theorem was stated and proved by Trinidad
and Soria [37] in the form of more defined theorem
Theorem. Let {kj } ⊂ L1 (Rd ) be a sequence of positive functions; forf ∈
L1 (Rd ) we define K ∗ f (x) = supj |kj ∗ f (x)|. Then the conditions (a) and (b)
below are equivalent
a) K ∗ is of weak type (1, 1); that is,
C
|{x ∈ Rd : K ∗ f (x) > λ}| ≤ ∥f ∥L1
λ
for certain constant C independent on f ∈ L1 (Rd ) and λ > 0.
b) There is a constant C > 0 such that for each finite set of not necessarily
different points {αj }N
j=1 , and each λ > 0 we have

∑
N
C
|{x ∈ Rd : sup | kj (x − αk )| > λ}| ≤ N.
j k=1
λ
2.2. THE CASE OF HIGHER DIMENSION 33

We may also take the same constant in a) and b).

The methods based on this result led to a consecutive development of
estimations for the weak-type inequality.
Aldaz actually proved that the constant cd → ∞ when d → ∞ using
this approach, an ingenious geometrical approach and uniformly distributed
grid of Dirac deltas.
In this work, in Chapters 3 and 4, we will present a significant qualitative
improvement to the result given by Aldaz and prove the following assymp-
totical behavior of cd = Cd1/4 , where the constant C can be obtained in the
following form

∫ 1 √
√ a 1
C = sup πd1/4 a−1/2 erf( √ ) dt.
a 2(1 − t)t (1 − t)t
0
Chapter 3

Asymptotic estimations of
the best constant for
weak-type inequality in Rd

3.1 Introduction
In this chapter we will present a technique which provides us with an esti-
mation for asymptotic behavior of the constant cd . The estimation is given
in two parts. First we will prove that cd = Cd1/4 . In the second part we
will use this result to estimate the constant C.

Notations
In this chapter we will use the following definitions. As previously we use
Rd for the real d dimensional space, we will also use k and j with some
occasional subindexes to define the dimension of some subspaces of Rd .
By a cube Q(x, r) we mean a closed l∞ ball of radius r with the center
at x in Rd , in other words Q(x, r) is a closed cube centered at x with sides
parallel to coordinate axes, and the side length 2r.
Definition. Let f (x) be a function of L1loc (Rd ). Then the maximal
function Md f (x) associated with the cube Q(x, r) is
∫
1
Md f (x) := sup |f (y)| dy. (3.1)
r>0 |Q(x, r)| Q(x,r)

35
 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
36 CONSTANT
From now on we explicitly indicate the dimension for which the maximal
function if defined. Here |Q(x, r)| stands for the Lebesgue measure in Rd .
and it will be convenient to use the following representations for this quan-
tity. ∫ ∫
|Q(x, r)| = 1dy = χQ(x,r) (y)dy. (3.2)
Q(x,r) Rd

As proviously χQ denotes the characteristic function of the set Q.

A fundamental property of the Hardy-Littlewood maximal function Md
is that it satisfies the weak-type (1, 1) inequality, i.e. there exists a constant
A > 0 such that for all α > 0 and for all f ∈ L1

α |{x : Md f (x) ≥ α}| ≤ A∥f ∥L1 , (3.3)

Let us denote by Ad the best (i.e. lowest) constant A satisfying this inequal-
ity in Rd .

3.2 Results
The main result of this chapter is formulated in the following theorem:
Theorem. If Md is the maximal function defined in (3.1) and the weak-type
inequality (3.3) holds then there exists a constant c > 0, independent of d
such that Ad ≥ cd1/4 .
In Chapter 4 we will present a technique of obtaining a numerical value
of the constant c.

3.3 Proof of the theorem

In the proof of the theorem we will use a discrete infinite measure µ which
is constructed as Dirac deltas placed at each point of the integer lattice
Z1 . Then, for d-dimensional space, µd is defined as the standard Cartesian
product µd = µ1 × µ1 × ... × µ1 . This will allow us to prove the theorem in
three steps. At the first stage we will prove the following proposition
Proposition: Let Q0 be the unit cube [0, 1]d in Rd . Then there exists α0
such that
α0 |{x : x ∈ Q0 , Md µd (x) > α0 }| > c0 d1/4 , (3.4)
where c0 is a constant independent of d.
3.4. PROOF OF THE PROPOSITION 37

The proof of this proposition is essential for the proof of the theorem formu-
lated above. As an immediate consequence of the proposition we get that
µd can be restricted to a cube Q(x, r0 ) of radius r0 = ρd1/4 , where ρ is such
that r0 ∈ N and the result will still holds.
At the second stage we will be able to restrict the infinite measure µd
to a significantly large cube to get a finite measure and then proceed by
applying the following lemma.
Lemma 1. For every dimension d there exists a closed cube K(R) ∈ Rd
with sidelength R ∈ N such that if we restrict µd to K(R) and N is the
number of translated unit cubes Q0 + n where n ∈ Nd inside K(R) then
N
lim = 1.
d→∞ (R + 1)d

Note that (R+1)d is equal to the number of points within K(R) with integer
coordinates.
The finite measure µd |K(R) from Lemma 1 then can be used in the mollifi-
cation procedure [2] to obtain a function f ∈ L1 with the desired property.
This will be done at the third stage of the proof, applying the following
lemma:
∑
Lemma 2. Let c be a given constant. If for some finite sum µ = ni=1 δxi
of Dirac deltas in Rd the inequality α|{x : Md µ(x) ≥ α}| > cn is satisfied
for some α > 0, then there exists a function f ∈ L1 (Rd ) and a constant
β > 0 for which the inequality β|{x : Md f (x) ≥ β}| > c∥f ∥L1 holds.
Technically the most complicated stage is the proof of the proposition. This
will be the subject of the next section. We will give the proofs of the Lemmas
1 and 2 in the appendix.

3.4 Proof of the proposition

Notations
Let us consider u ∈ (0, 1/2) and an interval I ⊂ R. Let us call y ∈ I
to be centered at level u (more briefly centered) if it belongs to the closed
subinterval with the same center as I and with the length (1 − 2u)|I| and to
be off center (at level u) otherwise. In particular for I = [0, 1] the centered
values are those belonging to the interval [u, 1 − u]. The role of u in the
 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
38 CONSTANT

2 with k=0,1,2
Figure 3.1: Sets Eu,k

proof is to serve as a parameter for which a discrete set of values will be

chosen to describe which cubes should be considered when estimating the
value of the maximal function at a given point.
Definition. Let Q0 be the unit cube [0, 1]d in Rd and let Eu,k d be a set
consisting of x ∈ Q0 such that k coordinates of the vector x = {x1 , ..., xd }
are off-centered at level u whereas remaining d − k coordinates are centered.

It is clear that for each particular choice of k off-centered coordinates

d has the measure (2u)k (1 − 2u)d−k . For the
the corresponding subset of Eu,k
d with combinatorial arguments one gets
entire Eu,k
( )
d
|Eu,k
d
| = (2u) (1 − 2u)
k d−k
, (3.5)
k
( )
where kd stands for the binomial coefficient. To illustrate this construction
one can refer to figures 3.1 and 3.2.
d and let s be a positive integer s > 0.
Let x be a point in the set Eu,k
Then the cube Q(x, s − u) contains (2s − 1)k (2s)d−k integer grid points. For
this cube the quantity
∫
1
M (d, u, k, s) = dµd
|Q(x, s − u)| Q(x,s−u)

can easily be computed explicitly

(2s − 1)k (2s)d−k

M (d, u, k, s) = . (3.6)
2d (s − u)d
3.4. PROOF OF THE PROPOSITION 39

3 with k=0,1,2,3
Figure 3.2: Sets Eu,k

The inequality
Md µd (x) ≥ M (d, u, k, s)

is an obvious consequence of the definition of the maximal function applied

to the measure µd . If now we restrict x onto the set Eu,k
d then the value of

M (d, u, k, s) can be taken as the constant α0 in the proposition and the left
hand side of (3.4) in this case takes the explicit form

M (d, u, k, s)|{x ∈ Eu,k

d
: Md µd (x) ≥ M (d, u, k, s)}| = M (d, u, k, s)|Eu,k
d
|.
(3.7)
In order to expand (3.7) onto entire cube Q0 we consider the following
d
construction. Let us take sets Eu(k),k with k ∈ ∆ = [kmin , kmax ] and u(k)
such that u(j) < u(k) if j < k. It is apparent that the sum of these sets
∪
d
Q∆ = Eu(k),k
k∈∆
 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
40 CONSTANT
will obey the exhausting condition

lim lim Q∆ = Q0 .
kmin →0 kmax →d

Then, in order to expand (3.7) onto the set Q∆ one needs to take the minimal
value of all M (d, u(k), k, s) terms, i.e.

M0 = min M (d, u(k), k, s).

k∈∆

As the result the extension of the relation (3.7) onto Q∆ reads as

M0 {x ∈ Q∆ : Md µd (x) ≥ M0 } = M0 |Q∆ | , (3.8)

where M0 now serves as the constant α0 of the proposition. This observation

is the basis for all estimations which are made in what follows. In order to
prevent overloading of formulae below we will often suppress in notations
the dependence of u on k implying it implicitly.
As the next step we will estimate the value of |Q∆ |. Let us make use of
the representation (3.2) and use the Schwarz inequality in the form
 2
∫ ∑ ∫ ∫ ∑ ∑
 χQ 
χE d dx ≤ χQ∆ dx
2
χE d χE d dx. (3.9)
∆ u,k u,k u,j
k∈∆ k∈∆ j∈∆

Taking into account the property of the characteristic function χ2Q∆ = χQ∆ ,
the obvious identity ∑ ∑
χQ∆ χE d = χE d
u,k u,k
k∈∆ k∈∆
and properly rearranging the terms in the double sum of the integrand in
the second integral of the right hand side of (3.9) we arrive at the inequality
( )2
∫∑
(∫ ) k χE d dx
u,k
|Q∆ | = χQ∆ dx ≥ ∫ ∑ ∫∑ . (3.10)
k χE d dx + 2 k>j χE d ∩Eu,j
d dx
u,k u,k

Here it is assumed that the k and j sums are taken over the interval ∆.
Introducing (3.10) into (3.8) we obtain the inequality
( )2
∫∑
M02 k χE d dx
u,k
M0 |Q∆ | ≥ ∫∑ ∫∑ . (3.11)
M0 k χE d dx + 2M0 k>j χE d ∩Eu,j
d dx
u,k u,k
3.4. PROOF OF THE PROPOSITION 41

In we now take into account the monotonic character of M (d, u, k, s) as

function of u and k then this inequality can be strengthened to the form
( )2
∫∑
k Mk χE d dx
u,k
M0 |Q∆ | ≥ (M0 /M1 )2 ∫ ∑ ∫∑ . (3.12)
k Mk χE d dx + 2 k>j Mk χE d ∩Eu,j
d dx
u,k u,k

Here for shortening of notations we have introduced the following: M1 =

maxk∈∆ M (d, u, k, s) and Mk = M (d, u, k, s).
In what follows we will deal with this representation. The right hand
side of (3.12) will be estimated by making use of the appropriate estimations
for every term of this fraction. For that purpose we will look for s0 such that
M (d, u, k, s0 ) = sups∈R+ M (d, u, k, s). In principal s0 should be integer, i.e.
s0 ∈ N, however technically the case of real value of s0 such that s0 ∈ R+
can be easier analyzed. From now on we will consider parameter s to be real
valued. The latter will allow us using the differential calculus for computing
extrema.
In order to be able to use the same estimations for s0 real and integer we
will need sups∈R+ M (d, u, k, s) and sups∈Z+ M (d, u, k, s) to have the same
order of d. That will be a restriction on some of our parameters. Formally
this restriction can be written as
sups∈R+ M (d, u, k, s)
< e. (3.13)
sups∈Z+ M (d, u, k, s)

In general it is enough to have any constant independent of d in the right

hand side of this inequality. This special choice of the bound is made for
computational convenience and will be used later.

Estimations for s ∈ R
We start by finding s0 that provides the maximal value for Mk . The problem
reduces to find the solution of the equation
∂ (d − k)u + (k − 2ud)s
log M (d, u, k, s) = = 0,
∂s (2s − 1)s(s − u)
which yields
s0 = (d − k)u/(2ud − k). (3.14)
 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
42 CONSTANT
This expression introduces the following restriction for k, i.e. since the value
of s0 is positive, k should not exceed 2ud, i.e. k < 2ud. Calculating Mk at
s = s0 leads us to the following result

k k (d − k)d−k
M (d, u, k, s0 ) = . (3.15)
dd (2u)k (1− 2u)d−k

We can now fix the parameter s in all above and following formulae as s = s0 .

Let us estimate the integral
∫
Mk χE d dx
u,k

from (3.12), when d ≫ 1 and k ≫ 1, recalling that

∫
d
Mk χE d dx = M (d, u, k, s0 )|Eu,k |.
u,k

Whereas the first factor in the right hand side has the relatively simple
form (3.15), the second one contains the binomial coefficient (3.5). This
coefficient can be evaluated for large values of d and k by using the Stirling
formula [1] ( )n
√ n
n! = 2πn (1 + O(1/n)).
e
In our construction we will restrict k to be in the interval [ζd, ξd], where
0 < ζ < ξ < 1 which implies that the lower bound for k, namely kmin ,
in our computations
( )
will also be considered large. Therefore, the binomial
coefficient kd in (3.5) can be approximated as
( ) √
d d dd
≈√ ,
k 2π(d − k)k k k (d − k)d−k

which yields √
d
d
M (d, u, k, s0 )|Eu,k | ≈√ . (3.16)
2π(d − k)k
Consequently, for the sum of such terms we get
∫ ∑ √
∑ ∑ d
Mk χE d dx = d
M (d, u, k, s0 )|Eu,k | ≈ √ . (3.17)
k∈∆
u,k
k∈∆ k∈∆
2π(d − k)k
3.4. PROOF OF THE PROPOSITION 43

For large d, the sum in the right hand side of (3.17) can be transformed to
an expression containing the Riemann integral sum, i.e.
√
∑ d ∑ 1/d
√ = d1/2 √
k∈∆
2π(d − k)k k∈∆
2π(1 − k/d)k/d

for the integral

∫ξ
dτ
I(ζ, ξ) = √ , (3.18)
2π(1 − τ )τ
ζ

where ζ = kmin /d and ξ = kmax /d. This implies that

∫ ∑
Mk χE d dx ≈ d1/2 I(ζ, ξ). (3.19)
u,k
k∈∆

The integral I(a, b) in the right hand side of (3.19) converges for all a
and b such that (a, b) ∈ (0, 1). Now it is obvious that, since the integrand
in (3.18) is positive, then extending or narrowing of the integration interval
(a, b) relatively to (ζ, ξ) will lead us to the upper or lower bounds. So that
we can state that there exist constants B1 < B2 which are independent of d
such that ∫ ∑
B1 d1/2 ≤ Mk χE d dx ≤ B2 d1/2 .
u,k
k∈∆

Let us notice that in any case B2 can be taken as I(0, 1). These inequalities
allow us to estimate the numerator term and the first term in denominator
of (3.12).
What is left is the estimation of the second term in the denominator of
the fraction in (3.12). Recalling that u = u(k) and u(k) > u(j) if k > j and
setting for shortening of notations u(k) = u and u(j) = v, the measure of
d and E d can easily be evaluated as
the intersection of two sets Eu,k v,j
( )( )
d k
|Eu,k
d
∩ d
Ev,j | = (2v) (2u − 2v)
j k−j
(1 − 2u)
d−k
.
k j

After regrouping the factors we arrive at the expression

( ) ( )
d k
|Eu,k
d
∩ d
Ev,j | = (2u) (1 − 2u)
k d−k
(2w)j (1 − 2w)k−j ,
k j
 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
44 CONSTANT
where w = v/2u. Using (3.16) for {d, k, j} ≫ 1 we get
√
k
|Eu,k ∩ Ev,j | ≈ |Eu,k |
d d d √
M −1 (k, w, j, s0 )
2π(k − j)j

and consequently
√ √
d k
∩
d
M (d, u, k, s0 )|Eu,k ≈√ d
Ev,j | √ M −1 (k, w, j, s0 ).
2π(d − k)k 2π(k − j)j
(3.20)
Introducing representations derived in (4.29) into the right hand side of
(3.12) we obtain
( √ )2
∑
k
√ d
2π(d−k)k
M0 |Q∆ | ≥ ∑ √ ∑ √ √ .
k
√ d
+ 2 k>j √ d √ k M −1 (k, w, j, s0 )
2π(d−k)k 2π(d−k)k 2π(k−j)j
(3.21)
Now we rewrite the double sum in the form
√ √
∑ d kM −1 (k, w, j, s0 )
√ √ = (3.22)
k>j
2π(d − k)k 2π(k − j)j
√
∑ d ∑ M −1 (k, w, j, s0 )
k−1
√ √ . (3.23)
k∈∆
2π(d − k) j=kmin
2π(k − j)j

As in the case of the sum (3.17) the double sum (3.22) can be transformed
into the integral sum of some double integral. This transformation will be
the subject of the rest of this subsection and finally will provide us with the
desired estimation for the second term in the denominator of (3.21).
First we estimate M −1 (k, w, j, s0 ). For computational convenience we
introduce a new function L(k, w, j) defined as follows

L(k, w, j) = log M (k, w, j, s0 ).

To get an estimation for L(k, w, j) from below we will construct the Taylor
expansion with respect to j around minimal value of this function. From
here on we will treat both j and k as real parameters which will allow us to
calculate derivatives and determine dependence for other parameters. This
procedure does not affect the accuracy since at every moment the transition
3.4. PROOF OF THE PROPOSITION 45

from R to Z can be performed. Then the value of j0 that provides the

minimum to L(k, w, j) can be found from the following equation
( )
∂ j(1 − 2w)
L(k, w, j) = log =0
∂j 2w(k − j)

and has the form j0 = 2wk. By direct computation one gets L(k, w, j0 ) = 0.
The second derivative of L(k, w, j) with respect to j has the following form

∂2 k
L(k, w, j) = ≥ 0,
∂j 2 j(k − j)

which shows that the Taylor expansion up to the first nonzero term of second
order can be written as follows
1 (j0 − j)2
L(k, w, j) = + O((j0 − j)3 ). (3.24)
2 2wk(1 − 2w)

The value of j0 − j in the numerator can be represented as

( )
u(j) k j
j0 − j = 2wk − j = k − j = u(j) − .
u(k) u(k) u(j)

By the Mean Value theorem we obtain

( )
k j ∂ k
− = (k − j),
u(k) u(j) ∂k u(k) k=k̃
( )
where k̃ ∈ (j, k). The value ∂k
∂ k
u(k) can be evaluated using the following
considerations. By Implicit Function theorem we define u(k) as the solution
of the equation
L(d, u, k) = L0 , (3.25)
where L0 = log(M0 ) is a constant. Then, differentiating (3.25) and using

d ∂ ∂ ∂u(k)
L(d, u(k), k) = L(d, u(k), k) + L(d, u(k), k)
dk ∂k ∂u ∂k
we arrive at the representation
∂
∂u(k) L(d, u(k), k)
= − ∂k
∂
,
∂k ∂u L(d, u(k), k)
 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
46 CONSTANT
or more explicitly ( )
k(1−2u)
∂u(k) log 2u(d−k)
=− 2du−k
.
∂k (1−2u)u
( )
∂ k
Consequently for ∂k u(k) we get
( ) ( )
∂ k 1 k(1 − 2u) k(1 − 2u)
= + log . (3.26)
∂k u(k) u u(2ud − k) 2u(d − k)
To simplify (3.26) will expand the logarithm with help of Taylor series in
the vicinity of k0 = 2ud. Due to the singularity in the prelogarithmic term
at k = k0 the expansion should be taken up to the quadratic term
( )
k(1 − 2u) k − 2ud (4u − 1)(k − 2ud)2
log = + + O((k − 2ud)3 ).
2u(d − k) 2ud(1 − 2u) 2(2ud)2 (1 − 2u)2
Substitution into (3.26) leads us to
( )
∂ k 2ud − k
= + O((k − 2ud)2 ). (3.27)
∂k u(k) − 2u)
4u2 d(1
The denominator of the right hand side of (3.24) can be estimated using
the following consideration. The value w was set as w = u(j)/2u(k). We
can express u(k) with the help of k0 as follows
k0 k0 − k + k k k0 − k
u(k) = = = + . (3.28)
2d 2d 2d 2d
From our previous considerations k ∈ [ζd, ξd], so we can conclude that
k
∼ const.
2d
The restriction (3.13) in the form
( )
M (d, u, k, s0 )
log < 1. (3.29)
M (d, u, k, s∗ )
can be used to estimate k0 − k as follows. Since s∗ ∈ Z is the point such that
M (d, u, k, s∗ ) = sups∈Z+ M (d, u, k, s), it is obvious that |s0 − s∗ | < 1. Now if
we will show that for any s̃ ∈ (min(s0 , s∗ ), max(s0 , s∗ )) the inequality holds
∂
log M (d, u, k, s)|s=s̃ (s0 − s∗ ) < 1, (3.30)
∂s
3.4. PROOF OF THE PROPOSITION 47

then (3.29) will hold true due to the Mean Value theorem. The explicit form
of (3.30) reads
(k0 − k)(s0 − s̃)(s0 − s∗ )
< 1.
(2s̃ − 1)s̃(s̃ − u)
One can strengthen the last inequality removing the positive factor in the
numerator (s0 − s̃)(s0 − s∗ ) < 1 which yields

k0 − k
< 1. (3.31)
(2s̃ − 1)s̃(s̃ − u)

Since |s̃ − s0 | < 1 and s0 ≫ 1 we can replace the denominator by 2s30 .

Inserting the explicit value (3.14) of s0 into inequality (3.31) we get it in the
following form
(k0 − k)4
< 1.
2u3 (d − k)3
Thus we have arrived at the estimation

k0 − k < (d − k)3/4 (2u3 )1/4 .

It was shown before that d − k ∼ d1 and u ∼ d0 then we may conclude that

k0 − k < k̂1 d3/4 (3.32)

with k̂1 being a constant. This inequality, being introduced to (3.14) will
generate the fact that s0 ∼ d1/4 which is consistent with our assumption
s0 ≫ 1.
Now we can rewrite (3.28) as

k
u(k) = + O(d−1/4 ).
2d
The latter allows us to use the following asymptotic relation
j
w≃
2k
as d → ∞. Thus we can write asymptotic equality for the denominator of
(3.24)

j j
2j0 (k − j0 ) = 4wk(k − 2wk) ≃ 2 k(k − k) = 2j(k − j) (3.33)
k k
 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
48 CONSTANT
Taking into account (3.27) and (3.33), the formula (3.24) for large d can be
rewritten in the form
( )2 [ ]
k j (k0 − k)2
L(k, w, j) ≃ (k − j)2 . (3.34)
2j(k − j) 2d (2ud)2 (2u)2 (1 − 2u)2 k=k̃

In order to get the final estimation from below for L(k, w, j) we proceed as
follows. The inequalities
2ud < d, 2u(1 − 2u) < 1/4
can be used for the denominator, and the numerator can be estimated with
the help of inequality
j > kmin .
For the estimation of L we are dealing with the smaller value of k0 − k
chosen the better. However, the estimation should be valid for all range of
k0 − k including the worst case of k0 − k ∼ d3/4 , i.e. when it is close to its
upper bound (3.32). Hence we will use the inequality k0 − k > k̂0 d3/4 where
k̂0 < k̂1 is a constant. This shows that
kkmin k̂0
L(k, w, j) > 2 (k − j).
d5/2
With this inequality we are ready to perform the final evaluation of the
double sum (3.21). At the first stage for the sum with respect to j in (3.22)
we get
∑
k
M −1 (k, w, j, s0 ) ∑ exp(−2 kkmin
k−1
d5/2
k̂0
(k − j))
√ < √ .
j=kmin
2π(k − j)j j=kmin
2π(k − j)kmin

Replacing j with kmin in the denominator in the right hand side makes this
inequality even stronger. Changing the summation index in the right hand
side we obtain
√ √ ( )
∑
k−1
M −1 (k, w, j, s0 ) k ∑
k−k min
1 k k 2 kmin k̂0 r 1
√ < √ exp −2
j=kmin
2π(k − j)j kmin r=1 2π r d5/2 k k
(3.35)
and we can see that the sum in the right hand side of (3.35) can now be
considered when k ≫ 1 as the Riemann integral sum for the integral
∫ 1−kmin /k 1 k 2 kmin k̂0
J(1 − kmin /k, 0) = √ exp(−2 τ )dτ.
0 2πτ d5/2
3.4. PROOF OF THE PROPOSITION 49

That implies that

√ 
∑
k−1
M −1 (k, w, j, s 0) d5/4 (1 − kmin /k)k 2 k min k̂0 
√ < √ Φ ,
j=kmin
2π(k − j)j 2kmin k k̂0 d5/2

√ ∫
where Φ(x) stands for error function [1] Φ(z) = 2/ π 0z exp{−x2 }dx. It is
possible to show that the argument of the error function has the order of
d1/4 . Therefore one can extend this argument of the error function to +∞
and taking into account that Φ(+∞) = 1, we end up with
√ √
∑ d ∑ M −1 (k, w, j, s0 )
k−1
d5/4 ∑ d
√ √ < √ .
k∈∆
2π(d − k) j=kmin 2π(k − j)j 2kmin k̂0 k∈∆ 2π(d − k)k
1/2

The sum on the right hand side was already estimated in (3.19) as
√
∑ d
√ < d1/2 I(0, 1).
k∈∆
2π(d − k)k

This gives that

√
∑ d ∑ M −1 (k, w, j, s0 )
k−1
d3/4
√ √ < I(0, 1).
k∈∆
2π(d − k) j=kmin 2π(k − j)j 1/2
2ζ k̂0

Here we have used the fact that kmin = ζd. So, the upper bound for the
second term in the denominator of (3.11) can be written in the form
∫ ∑
M0 χE d ∩Eu,j
d dx < B3 d3/4 ,
u,k
k>j

where B3 is some constant. Finally we obtain the estimate for M0 |Q∆ |

B12 d
M0 |Q∆ | > ∼ c0 d1/4 ,
B2 d1/2 + 2B3 d3/4

which finalizes the proof of the proposition.

 CHAPTER 3. ASYMPTOTIC ESTIMATIONS OF THE BEST
50 CONSTANT
3.5 Final proof
Collecting results from the previous sections we have proven by direct con-
struction that the constant M0 can be taken as α0 of the proposition. We
have shown that inequality Md µd > M0 is fulfilled on the set Q∆ ⊂ Q0 .
Moreover, we have proven that there exists a constant c0 , which is indepen-
dent of d, such that
M0 |Q∆ | > c0 d1/4 .
Thus, the proposition has been proven.
Now we can use Lemma 1 in order to extend the proposition onto the
class of finite measures, i.e. the restriction of the measure µd onto a large
cube K(R) with sidelength R ≥ θd5/4+ε leads us to the inequality

α0 |{x : Md (µd |K(R) )(x) ≥ α0 }| > c0 d1/4 (R + 1)d .

The final result follows from the Lemma 2, i.e. there exists a constant β > 0
and f ∈ L1 such that

β|{x : Md f (x) ≥ β}| > c0 d1/4 ∥f ∥L1 . (3.36)

3.6 Proof of lemma 1

Let us consider some x inside Q0 = [0, 1]d . Then the proposition holds true
not only for infinite measure µd but for µd |Q(x,r0 ) where r0 > s0 − u or more
explicitly r0 = ρd1/4 , r0 ∈ Z as well, with some appropriate ρ independent
of d. Thus for a significantly large cube K(R) the proposition will hold true
for any translated unit cube Q0 + n where n ∈ Zd within a smaller cube
K(R − r0 ) with the same center as K(R). The number of unit cubes that
fits into K(R − r0 ) is N = (R − r0 )d . Thus asymptotically as d → ∞
( )d
N (r0 − 1)
d
∼ 1− .
(R + 1) (R + 1)

If we now set R ≥ θd5/4+ε where ε > 0 and θ is some constant independent

of d, then asymptotically when d → ∞ we will obtain
( )d ( )d
(r0 − 1) ρ
1− ≥ 1 − 1+ε ∼ 1.
(R + 1) θd
3.7. PROOF OF LEMMA 2 51

Now taking into account the fact that N < (R + 1)d we arrive at the desired
result
N
lim = 1.
d→∞ (R + 1)d

3.7 Proof of lemma 2

Although the original proof of this Lemma can be found in [2] we give it
∑
here for completeness and unification of notations. Suppose µ = ni=1 δxi
and α > 0 are such that the inequality

α|{M µ ≥ α}| > cn

holds. Let us prove that for every ϵ > 0 there exists f ∈ L1 (Rd ) with
∥f ∥1 = n and {M µ ≥ α} ⊂ {M f > (1 + ϵ)−1 α}. Let l(Q) denotes the side
length of the cube Q. For each x ∈ {M µ ≥ α}, select a cube Q(x, r) with

N (Q(x, r))
> α.
|Q(x, r)|

Here N (Q(x, r)) stands for a number of point masses contained in Q(x, r).
We make additional assumption that l(Q(x, r)) ≥ α−1/d . This is always
possible to fulfill since each Q(x, r) must contain at least one point mass.
Since α is strictly positive we can define a as

a = sup{l(Q(x, r)) : x ∈ {M µ ≥ α}} < ∞.

r

Given any ϵ > 0, choose δ(ϵ) such that for every y ∈ [α−1/d , a], (y + δ)d /y d <
1 + ϵ. Let Ei be a closed cube centered at xi with the sidelength δ and
∑
define f = δ −d n1 χEi . Replacing each Q(x, r) by the cube Q(x, r + δ/2),
we see that {M µ ≥ α} ⊂ {M f > (1 + ϵ)−1 α}. Pick ϵ > 0 so small that
α|{M µ ≥ α}| > (1 + ϵ)cn. Since ||f ||1 = n, we get

α|{M f ≥ α(1 + ϵ)−1 }| ≥ α|{M µ ≥ α}| > (1 + ϵ)c||f ||1 .

The conclusion now follows by setting β = α(1 + ϵ)−1 .

Chapter 4

More precise estimation

To estimate the constant c0 in (3.36) we will use the technique which was
developed above, although improved at some critical points.

4.1 Case of u < 1/2.

We will use the same idea as in the previous proof, but we will be more
precise in our estimations. Consider that k changes within some interval
∆ = [kmin , kmax ], where kmin = ζdθ and kmax = d − ζdθ where θ < 1 and ζ
is positive constant. As it can be seen in above proof this assumption does
not break our estimation. Let us now divide [kmin , kmax ] into N pieces of
equal lengths and denote those pieces by ∆n = (kn , kn+1 ), and |∆n | = Λ for
all n. These intervals fulfill the following condition

∪
N
∆n = ∆ = [kmin , kmax ].
n=1

Obviously, for large d the interval [kmin > 1, kmax < d] can be enlarged op
to [1, d], thus in our further considerations we will usually address ∆ as [1, d]
and Λ as d/N .
We can define sets Fn as
∪
Fn = Ekd .
k∈∆n

53
54 CHAPTER 4. MORE PRECISE ESTIMATION

It is apparent that the following two relations hold true

∪
N
Fn = Q∆ (4.1)
n=1

and consequetly
∪
Ekd = Q∆ . (4.2)
k∈∆

For now we will only demand Λ to fulfill the following condition Λ/d → 0
as d → ∞. Some additional conditions on the value of λ will be discussed
below.
If we define M (d, u, k, s) the same way as in it was done in subsection
3.4.1 ∫
1
M (d, u, k, s) = dµd (4.3)
|Q(x, s − u)| Q(x,s−u)
and set now a constant α in weak type inequality as some value M∗ then
the Schwartz inequality with some additional algebraic transformations will
give us
∑
(M∗ n |Fn |)2
M∗ |Q∆ | ≥ ∑ ∑ ∩ . (4.4)
M∗ n |Fn | + 2 n>m M∗ |Fn Fm |
Here M∗ is chosen to fulfill

M∗ = M (d, u(k), k, s∗ ) (4.5)

for some specially chosen s∗ . We will use this fact later for obtaining u(k)
as an implicit function of k and s∗ . In our work s∗ denotes an integer value
where the function M (d, u, k, s) reaches the largest value possible, namely

s∗ = arg max M (d, u, k, s). (4.6)

s∈Z

It is apparent that s∗ is a function of d and k but since an actual form of

s∗ is never used in this work the dependence on that parameters will not
appear in our notation.
We aim to transform sums in (4.4) into the Riemann sums and then
estimate their values as integrals. For that let us study the arguments of
sums that appear in the numerator and in the denominator of this fraction.
4.2. NUMERATOR OF (4.4) 55

4.2 Numerator of (4.4)

For the numerator of inequality (4.4) we can use the same technique as for
(4.4) itself and obtain the following relation
∑
(M∗ k∈∆n |Ekd |)2
M∗ |Fn | ≥ ∑ ∑ ∩ . (4.7)
M∗ k∈∆n |Ekd | + 2M∗ k>j |Ekd Ejd |

Let us again estimate sums in the numerator and in the denominator sepa-
rately.

Numerator of (4.7)
d | as an argument of the
In the numerator of (4.7) we have M (d, u, k, s∗ )|Eu,k
sum. Since this value is rather hard to estimate we will estimate the product
d | instead and then the ratio M (d, u, k, s )/M (s, u, k, s ),
M (d, u, k, s0 )|Eu,k ∗ 0
where s0 = arg maxs∈R M (d, u, k, s). Recall that M (d, u, k, s0 ) was obtained
in the following form

k k (d − k)d−k
M (d, k, u, s0 ) = . (4.8)
dd (2u)k (1 − 2u)d−k

The explicit value of |Ekd | is given by

( )
d
|Ekd | = (2u) (1 − 2u)
k d−k
. (4.9)
k
( )
Thus, applying Stirling’s formula for binomial coefficient kd we arrive at a
rather compact formula for approximation of the product M (d, u, k, s0 )|Ekd |.
The resulting expression will have the following form
√
d
M (d, u, k, s0 )|Ekd | ≈ . (4.10)
2πk(d − k)

Now let us transform the sum in the numerator of (4.7) into

∑ ∑ M∗
M∗ |Ekd | = M (d, u, k, s0 )|Ekd |. (4.11)
k∈∆n k∈∆n
M (d, u, k, s0 )
56 CHAPTER 4. MORE PRECISE ESTIMATION

This sum on a narrow segment ∆n can be approximated as

⟨ √ ⟩
∑ M∗ ∑ M∗ d
M (d, u, k, s0 )|Ekd | ≈ √ ,
k∈∆n
M (d, u, k, s0 )
n k∈∆
−
M (d, u, k, s0 )
2π(d k)k
(4.12)
where ⟨M∗ /M (d, u, k, s0 )⟩ stands for an average value on an interval |s∗ −
s0 | ≤ 1/2. The last procedure is possible because both M (d, u, k, s0 )|Ekd |
and M∗ /M (d, u, k, s0 ) are changing slowly within the interval of summation
∆n of parameter k. This property of M (d, u, k, s0 )|Ekd | was already shown
in our previous consideration. Now let us prove it for ⟨M∗ /M (d, u, k, s0 )⟩
and get an approximation for this value. For that we will introduce a new
function
L(d, u, k, s) = log(M (d, u, k, s)). (4.13)
Let us expand L(d, u, k, s∗ ) into power series with respect to parameter s in
the vicinity of s0 which is the solution to the following equation
∂
L(d, u, k, s)|s0 = 0. (4.14)
∂s
This yields
(d − k)u
s0 = . (4.15)
2ud − k
Then the power series, up to the second order of (s∗ − s0 ) will have only two
terms, namely

1 ∂2
L(d, u, k, s∗ ) = L(d, u, k, s0 ) + L(d, u, k, s)|s0 (s∗ − s0 )2 + O((s∗ − s0 )3 ).
2 ∂s2
(4.16)
The latter can be written explicitly as
k − 2ud
L(d, u, k, s∗ ) = L(d, u, k, s0 ) + (s∗ − s0 )2 + O((s∗ − s0 )3 ).
2(2s0 − 1)s0 (s0 − u)
(4.17)
If we substitute s0 with its explicit value then we arrive at

(k − 2ud)4
L(d, u, k, s∗ ) = L(d, u, k, s0 )− (s∗ −s0 )2 +O((s∗ −s0 )3 ).
2(d − k)kdu2 (1 − 2u)2
(4.18)
Note that exp(L(d, u, k, s∗ )−L(d, u, k, s0 )) provides us with the desired value
of M (d, u, k, s∗ )/M (d, u, k, s0 ).
4.2. NUMERATOR OF (4.4) 57

Let us estimate the value of 2ud − k. Recall that k0 = 2ud is a solution

to the following equation

∂
L(d, u, k, s0 )|k0 = 0. (4.19)
∂k
Note that L(d, u, k0 , s0 ) = 0, thus power series for L(d, u, k, s0 ) with respect
to k will have the form
1
L(d, u, k, s0 ) = (k − k0 )2 + O((k − k0 )3 ). (4.20)
4ud(1 − 2u)

Beside the fact that k0 − k is a rather big parameter one can construct the
∂n
power series with respect to (k0 − k)/d and since ∂k n L(d, u, k, s0 ) = o(d
1−n )

the power series converges. This leads us to the following approximation

(k − k0 )2 ≈ L(d, u, k, s0 )4ud(1 − 2u). (4.21)

If we apply this to (4.16) then we arrive at

8d
L(d, u, k, s∗ ) − L(d, u, k, s0 ) ≈ L2 (d, u, k, s0 )(s∗ − s0 )2 (4.22)
(d − k)k

Now to get an average value of the ratio M (d, u, k, s∗ )/M (d, u, k, s0 ) we

will use the following consideration. Since s0 changes slowly with k we can
consider s0 (k) as a continuous function of a real parameter and without loss
of generality consider s∗ to be closest integer to s0 . Therefore the average
value takes the form
⟨ ⟩ ∑
M (d, u, k, s∗ ) 1 8dL2 (d, u, k, s0 )
= exp(− (s∗ − s0 )2 ). (4.23)
M (d, u, k, s0 ) Λ k:|s −s |<1/2 (d − k)k
0 ∗

Here the summation runs over all k such that |s0 − s∗ | < 1/2 holds. The
sum on the right hand side is a Riemann sum for the following integral

∫1/2
8dL2 (d, u, k, s0 )
exp(− (s∗ − s0 )2 )d(s∗ − s0 ). (4.24)
(d − k)k
−1/2

Evaluating this integral one arrives at

58 CHAPTER 4. MORE PRECISE ESTIMATION

∫1/2
2dL2 (d, u, k, s0 )
exp(− (s∗ − s0 )2 )d((s∗ − s0 )2 ) =
(d − k)k
−1/2
√ √
d
π erf( 2(d−k)k L(d, u, k, s0 ))
√ . (4.25)
2d
(d−k)k L(d, u, k, s0 )

Note that the expression on the right hand side varies slowly with k, which
allows us to use this value for an entire narrow interval ∆n even if the value
s∗ does not remain unchanged. Since we are planning to use this expression
in a Riemann’s sum we can choose k = kn .
Now we can rewrite the sum in the numerator of (4.12) in the following
form
⟨ ⟩ ∑ √
∑ M∗ d
M∗ |Ek | ≈
d √ . (4.26)
k∈∆n
M (d, u, k, s0 ) k∈∆
n
2π(d − k)k

Introducing an estimate obtained in (4.25) we get

√ √
d √
∑ π erf( 2(d−kn )kn L(d, u, kn , s0 )) ∑ d
M∗ |Ekd | ≈ √ √ .
2d 2π(d − k)k
k∈∆n (d−kn )kn L(d, u, kn , s0 ) k∈∆n
(4.27)

Denominator of (4.7)
If we introduce a new notation w = u(j)/u(k) then the Lebesgue measure
|Eu,k
d ∩ E d | of intersection of sets in the denominator can be transformed
v,j
into the following expression
( ) ( )
d k
|Eu,k
d
∩ d
Ev,j | = (2u) (1 − 2u)
k d−k
(2w)j (1 − 2w)k−j , (4.28)
k j

and consequently
√ √
d k
d
M (d, u, k, s0 )|Eu,k ∩ d
Ev,j | ≈√ √ M −1 (k, w, j, s0 ).
2π(d − k)k 2π(k − j)j
(4.29)
4.2. NUMERATOR OF (4.4) 59

Then the double sum will take the form

√
∑ 1 ∑ d ∑ exp(−L(k, w, j))
k−1
d
M (d, u, k, s0 )|Eu,k ∩ Ev,j
d
| ≈ √ .
k>j
2π k∈∆ (d − k) j=k (k − j)j
n n
k,j∈∆n
(4.30)
Here the short notation L(d, u, k) = L(d, u, k, s0 (k)) is used. Now let us
decompose L(k, w, j) into power series with respect to j in the vicinity of
j0 = 2wk and in analogy to (4.20) we will arrive at

(j0 − j)2 k(j0 − j)2

L(k, w, j) ≈ ≈ . (4.31)
4wk(1 − 2w) 2j(k − j)
To estimate the value of j0 − j we transform it as
( )
u(j) k j
j0 − j = 2wk − j = k − j = u(j) − . (4.32)
u(k) u(k) u(j)
The Mean Value theorem gives us
( )
k j d k
− = (k − j). (4.33)
u(k) u(j) dk u(k) k=k̃

We can obtain ∂k u(k) using Implicit Function theorem. To do that consider

L(d, u, k, s∗ ) for fixed d and s∗ . For chosen d and fixed s∗ the parameter k
can vary within some limits while s∗ remains the closest integer to s0 and so
does u(k). At the same time under these conditions L(d, u(k), k, s∗ ) remains
constant with respect to k. Correspondingly we have
d ∂ du ∂
0= L(d, u, k, s∗ ) = L(d, u, k, s∗ ) + L(d, u, k, s∗ ). (4.34)
dk ∂u dk ∂k
This yields
du ∂k L(d, u, k, s∗ )
=− . (4.35)
dk ∂u L(d, u, k, s∗ )
The corresponding patrial derivatives of L(d, u, k, s∗ ) have the form
( )
∂ 2s∗ − 1
L(d, u, k, s∗ ) = log (4.36)
∂k 2s∗
and
∂ d
L(d, u, k, s∗ ) = . (4.37)
∂u s∗ − u
60 CHAPTER 4. MORE PRECISE ESTIMATION

Thus we arrive at the following expression for du/dk

( )
du s∗ − u 2s∗ − 1
=− log . (4.38)
dk d 2s∗
To make the above expression simpler let us expand log into a power series
with respect to 1/s∗ and obtain
du 4s∗ − 4u + 1
≈ . (4.39)
dk 8ds∗
Now we can write down
( ) ( )
d k 1 k(4u − 1)
= 2 2ud − k + . (4.40)
dk u 2u d 4s∗
From the equation (4.15) we get
2ud − k ≈ (d − k)u/s0 , (4.41)
since both s0 and s∗ are rather big (∼ d1/4 ) and |s0 − s∗ | ≤ 1/2 then
asymptotically 1/s0 = 1/s∗ which gives us
( )
d k 2ud − k
= . (4.42)
dk u 4(d − k)u2
This value is changing slowly with k hence one can use only zeroth term in
the power series in vicinity of k and due to the fact that Λ is a sufficiently
narrow interval in comparison to d we can effectively assume k̃ = k in (4.33)
and u(j)/u(k) = 1 to transform (4.31) in the following form
( )2
k 2ud − k
L(k, w, j) ≈ (k − j)2 . (4.43)
2j(k − j) 4(d − k)u
The inner sum of (4.30) now becomes
√ ( ( )2 )
∑
k
k k 2ud − k
exp − (k − j).
j=kn
2π(k − j)j 2j 4(d − k)u

This sum can be transformed into a Riemann sum. To do that let us intro-
duce a new parameter t = (k − j)/Λ. Then the above sum is modified as
shown ( )
(k−kn )/Λ
∑ ( )
Λ 1 1 2ud − k 2 1
√ √ exp − tΛ ,
2π t=1/Λ t 32 (d − k)u Λ
4.3. DENOMINATOR OF (4.4) 61

and it has the following integral as its limit

∫ n )/Λ
(k−k ( ( )2 )
Λ 1 1 2ud − k
√ √ exp − tΛ dt.
2π t 32 (d − k)u
0

The value of Λ can be chosen such that the argument of the exponential
tends to minus infinity as d goes to infinity. Thus evaluation of this integral
gives √
8(d − k)u2 π 4(d − k)u
√ = .
π(2ud − k) 2 2ud − k
Now let us substitute the right hand side of (4.21) instead of 2ud − k and
use d − k ≈ d(1 − 2u) to get
√
4(d − k)u2 d(1 − 2u)u
≈2 .
2ud − k L(d, u, k, s0 )
Thus an approximation for the second term in the denominator of (4.7) will
look as follows √ √
1∑ d d(1 − 2u)u
.
π k (d − k)k L(d, u, k, s0 )
Since we have shown earlier that the second term dominates the first one
we can consider the entire denominator of (4.7) looking like
1 ∑ 1
√ √ .
2π k∈∆n L(d, u, k, s0 )

One can apply the same trick as in (4.26) factoring an average of ratio
⟨M∗ /M (d, u, k, s0 )⟩ out of the sum. Now we can rewrite (4.7) with estima-
tions we have made for the right hand side
√ ∑ √
√ √ d
π erf( 2(d−kdn )kn L(d, u, kn , s0 )) k∈∆n 2π(d−k)k
M∗ |Fn | ≥ √ ∑ .
2d √1 L−1/2 (d, u, k, s0 )
(d−kn )kn L(d, u, kn , s0 ) 2π k∈∆n

4.3 Denominator of (4.4)

For the term in the denominator of (4.4)
∑ ∩
M0 n>m |Fn Fm |
∑ (4.44)
M0 n |Fn |
62 CHAPTER 4. MORE PRECISE ESTIMATION

we will use the technique given below. Consider the sum

∑
M0 |Fn ∩ Fm |,
n>m

where
∪ ∪ ∪
|Fn ∩ Fm | = |( Ekd ) ∩ ( Ejd )| = | (Ekd ∩ Ejd )|.
k∈∆n j∈∆m k∈∆n ,j∈∆m

¯ n = (kn + ϵd7/8 , kn+1 −

Each interval ∆n = (kn , kn+1 ) we will replace with ∆
ϵd7/8 ) thus
∑
N
lim χ∆
¯ n = χ∆
ϵ→0
n=1
in measure. Hence we can make an estimation k − j ∼ d. Now we can
transform the same sum as (4.29)
√ √
d k
M (d, u, k, s0 )|Eu,k ∩ Ev,j | ≈ √
d d √ M −1 (k, w, j, s0 ).
2π(d − k)k 2π(k − j)j
¯ n and j ∈ ∆
For new k ∈ ∆ ¯ n−1 the inner sum

∑
k−1
exp(−L(k, w, j))
√
j=kn +ϵd
(k − j)j

can be approximated as
∑
k−1
exp(−ad1/2 )
j=kn +ϵd
Dd
∑
which is negligibly small in comparison to M∗ n |Fn | ∼ d1/4 that we have
for the denominator of (4.44).
Now let us estimate the rest of the sum. To do this note that without
loss of generality we can reduce our double sum
∑
|Fn ∩ Fm |
n>m

to the single sum

∑
N −1
|Fn ∩ Fn+1 |
n=1
4.3. DENOMINATOR OF (4.4) 63

and furthermore to
   
∑
N −1 ∪ ∩ ∪
 Ekd   Ejd  .
n=1 k∈(kn+1 −ϵd7/8 ,kn+1 ) j∈(kn+1 +ϵd7/8 ,kn+1 )

Obviously the measure of intersection of any two sets cannot be larger than
the measure of any of those sets, thus
   
∑
N −1 ∪ ∩ ∪
 Ekd   Ejd  <
n=1 k∈(kn+1 −ϵd7/8 ,kn+1 ) j∈(kn+1 ,kn+1 +ϵd7/8 )
 
∑
N −1 ∪
 Ekd  . (4.45)
n=1 k∈(kn+1 −ϵd7/8 ,kn+1 )

Since the ratio (ϵd7/8 )/(kn+1 − kn ) tends to 0 as d → ∞ so does

∑N −1 (∪ )
n=1 k∈(kn+1 −ϵd7/8 ,,k n+1 )
Ekd
∑N (∪ ) → 0.
d
n=1 k∈(kn ,kn+1 ) Ek

Hence we arrive at the following approximation of the entire expression (4.4)

∑
M∗ |Q∆ | ≥ M∗ |Fn |.
n

Now we can substitute the result we have obtained for M∗ |Fn | to get the
following inequality
( √ )2
√ √ ∑
d
∑ π erf( 2(d−kn )kn L(d, u, kn , s0 )) k∈∆n
√ d
2π(d−k)k
M∗ |Q∆ | ≥ √ ∑ .
2d √1 L−1/2 (d, u, k, s0 )
n (d−kn )kn L(d, u, kn , s0 ) 2π k∈∆n
(4.46)
Now we would like to rewrite the sum in the right hand side of (4.46) as a
Riemann sum. To do that we take into account the following observation -
recall that N and d tend to infinity independently, namely for any fixed N
we assume we can use a limit expression with d → ∞ of the argument of
the sum. On the other hand side we have shown that all the factors on the
64 CHAPTER 4. MORE PRECISE ESTIMATION

right hand side of the sum change slowly with k ∈ ∆n thus for this sum we
can use the following simplifications
√ √
∑ d d
√ ≈√ Λ (4.47)
k∈∆n
2π(d − k)k 2π(d − kn )kn

and
1 ∑ −1/2 1
√ L (d, u, k, s0 ) ≈ √ L−1/2 (d, u, kn , s0 ) Λ. (4.48)
2π k∈∆n 2π

Therefore the entire right hand side of (4.46) can be rewritten in such a way
that the inequality transforms to
√ √
d
√ ∑ erf( 2(d−kn )kn L(d, u, kn , s0 )) d d
M∗ |Q∆ | ≥ π . (4.49)
n L1/2 (d, u, kn , s0 ) (d − kn )kn N

We need also take into account the result (3.34) that can be significantly
simplified to
L(d, u, k, s) ≃ ad1/2 .
Note that (3.3) demands our function M (d, u, k, s) to be larger some certain
value α which in our case will be chosen as log(ad1/2 ) and hence we can
assume the value L(d, u, kn , s) dose not change with n. Now we can replace
the sum on the right hand side of (4.49) with the following integral
( )√
√ 1/4 −1/2 ∫
1
a 1
πd a erf √ dt, (4.50)
2(1 − t)t (1 − t)t
0

where the value of a should be chosen such that the integral (4.50) reachs
its maximum. Solving this maximal problem analytically appears to be
too complicated. On the other hand the numerical computation gives the
following result

∫ 1 ( )√
√ a 1
max πd1/4 a−1/2 erf √ dt = 6.553... (4.51)
a 2(1 − t)t (1 − t)t
0

where a = 0.36.
4.4. CASE OF U ∈ (1/2, 1) 65

4.4 Case of u ∈ (1/2, 1)

In this case we need to modify the concept of centered and off-centered
intervals. It is natural to consider that centered sets are those (1 − u, u) and
off-centered are (0, 1 − u) ∪ (u, 1). The function M (d, u, k, s) defined earlier
(3.6) should be modified in the corresponding manner and takes the form

(2s − 1)k (2s − 2)d−k

M (d, u, k, s) = .
2d (s − u)d

For computational convenience it is useful to make a substitution

r = s − 1, u′ = 1 − u.

Then M (d, u, k, s) will be transformed into M (d, u′ , k, r)

(2r + 1)k (2r)d−k

M (d, u′ , k, r) = .
2d (r + u′ )d

Note that u′ ∈ (0, 1/2) and instead of redefining sets Eu,kd we can use E d
u′ ,k
defined as before and the norm of those sets will coincide with our previous
result, i.e. ( )
′ k ′ d−k d
|Eu,k | = |Eu′ k | = (2u ) (1 − 2u )
d d
.
k
The remarkable fact is that the function M (d, u′ , k, r) has the maximum
with respect to the real valued parameter r at the point

(d − k)u′
r0 =
k − 2u′ d
and the corresponding maximal value is

k k (d − k)d−k
M (d, u′ , k, r0 ) = .
dd (2u′ )k (1 − 2u′ )d−k

This perfectly matches the result obtained for u ∈ (0, 1/2) up to the change
of notations. As the result we get
( )
k k (d − k)d−k d
|Eud′ ,k |M (d, u′ , k, r0 ) = .
dd k
66 CHAPTER 4. MORE PRECISE ESTIMATION

From this we can conclude that all the estimations made for u ∈ (0, 1/2)
and s ∈ R hold also for the case u ∈ (1/2, 1). Consider now a generalization
of formula (4.7) onto the case of u ∈ (1/2, 1)
∑
(M∗ k∈∆n |Eud′ ,k |)2
M∗ |Fn | ≥ ∑ ∑ ∩ . (4.52)
M∗ k∈∆n |Eud′ ,k | + 2M∗ k>j |Eud′ ,k Eud′ ,j |
The value log(M (d, u, k, s0 )) plays the crucial role in our method. In par-
ticular this shows that the value ⟨M∗ /M (d, u, k, s0 )⟩ is unaffected whether u
is larger or smaller than 1/2. This fact guarantees that the entire estimation
will remain unchanged.

4.5 Estimating the intersection

Now we will show that the influence of intersections of the collections of sets
d for u < 1/2 and u > 1/2 is negligibly small and thus those two cases
Ek,u
can be treated separately. Taking into account the observation made above
that they produce the identical approximation this will result in factor 2 in
our final formula.
d and E d . When k = j
Let us consider the intersection between sets Eu,k u′ ,j
′
it is apparent that u (k) < u(k) and we will have the following representation
for the norm of intersection
( ) ( )k
′ k d u′
|Eu,k
d
∩ Eud′ ,k | = (2u ) (1 − 2u) d−k
= |Eu,k
d
|. (4.53)
k u
Let us now estimate the value (u′ /u)k using following considerations. From
estimations of the value of s0 we know that 2ud − k > 0 and k − 2u′ d > 0.
We also know that 2ud − k = O(d3/4 ) and so does k − 2u′ d. From that we
can conclude that u − u′ has the magnitude of the order d−1/4 . Hence we
can write
( )k
u′ u′ cd−1/4
= exp(k log( )) ≈ exp(k log(1 − )) ≈ exp(−2cd3/4 ),
u u u
which yields
M∗ |Eu,k
d
∩ Eud′ ,k | → 0 as d → ∞.
For the case of k ̸= j let us show that M∗ |Eu,k
d ∩ E d | has the same
u′ ,j
behavior, namely
M∗ |Eu,k
d
∩ Eud′ ,j | → 0 as d → ∞.
4.6. CONCLUSION 67

For that, let us first consider k > j. As it was discussed earlier the corre-
sponding sets will have intersection of measure zero if u(k) ≤ u′ (j). In the
case of u(k) > u′ (j) we can use the same representation as (4.29) and get
√ √
d k
M∗ |Eu,k ∩ Eu′ ,j | ≈ √
d d √ exp(−L(k, w, j)). (4.54)
2π(d − k)k 2π(k − j)j

Here w is defined as u′ (j)/2u(k). Using representation (4.31) let us estimate

the value of (2wk − j)2 . We get
( )
k j
|2wk − j| = u′ (j) − ′ .
u(k) u (j)
We can add and subtract 2d within the parenthesis above and get
( ) ( )
k j k − 2u(k)d j − 2u′ (j)d
− = − = O(d3/4 ).
u(k) u′ (j) u(k) u′ (j)
Note that (k − 2u(k)d)/u(k) and (j − 2u′ (j)d)/u′ (j) have different signs and
thus in this case their contributions simply add up in the previous formula.
Therefore L(k, w, j) is never small for k > j.
In the case of k < j the only nonzero measure of intersection takes place
when u′ (j) > u(k) which implies that k − j > Cd3/4 and
√ √
d j
M∗ |Eu,k ∩ Eu′ ,j | ≈ √
d d √ exp(−L(j, w, k)), (4.55)
2π(d − j)j 2π(j − k)k

where w = u(k)/2u′ (j). Then, considering the value

( )
j k
|2wj − k| = u(k) − (4.56)
u′ (j) u(k)
and applying the same argument as for case k > j we conclude that (2wk −
j)2 > Cd3/2 .
Thus we have shown that M∗ |Eu,k
d ∩ E d | tends to 0 exponentially as
u′ ,k
d → ∞.

4.6 Conclusion
Gathering the results obtained by numerically calculating the integral in
(4.51) and in the previous section we arrive at our final result.
68 CHAPTER 4. MORE PRECISE ESTIMATION

The asymptotic of the best constant Ad in (3.3) fulfills Ad > cd1/4 where
( )√
√ 1/4 −1/2 ∫
1
a 1
c = max 2 πd a erf √ dt ≈ 13.105...
a 2(1 − t)t (1 − t)t
0
References

[1] M. Abramowitz, I. Stegun , Handbook of Mathematical functions,

Dover Publ. Inc., New York, 1972.
[2] J.M. Aldaz, "A remark on the centered n-dimensional Hardy-Littlewood
maximal function". Czechoslovak Math. J. 50 (125) (2000), no. 1, 103-
112.
[3] F.J. Almgren and E.H. Lieb, "Symmetric decreasing rearrangement is
sometimes continuous", J. Amer. Math. Soc., 2 (1989), 683-773.
[4] A. Bernal "A note on the one-dimensional maximal function", Proc.
Roy. Soc. Edinburgh, 111A (1989), 325-328.
[5] A.S. Besicovitch, "A general form of the covering principle and relative
differentiation of additive functions", Proce. of the Cambr. Phil. Soc.,
41-02 (1945), 103-110.
[6] T. Bonnesen, W. Fenchel, "Theorie der konvexen Körper", Julius
Springer, Berlin, 1934. English translation: "Theory of convex bodies",
BCS Associates, Moscow, ID, (1987).
[7] J. Bourgain, "On high dimensional maximal functions associated to
convex bodies", Amer. J. Math., 108 (1986), 1467-1476.
[8] J. Bourgain, "On the Lp -bounds for maximal functions associated to
convex bodies in Rn ", IsraelJ.Math., 54 (1986), 257-265.
[9] O.N. Capri, N.A. Fava, "Strong differentiability with respect to product
measures", Studia Math., 78 (1984), 173-178.
[10] L. Colzani, J.P. Làzaro, "Eigenfunctions of the Hardy-Littlewood max-
imal operator", Colloq. Math., 118 (2010), 379-389.

69
70 REFERENCES

[11] R. Dror, S. Ganguli and R.S. Strichartz, "A search for best constants
in the Hardy-Littlewood maximal theorem", J. Fourier Anal. Appl.2,
(1996), 473-486.

[12] R. Fefferman, "Strong differentiation with respect to measures", Amer.

J. Math., 103 (1981).

[13] L. Forzani, R. Scotto, P. Sjögren, W. Urbina, "On the Lp bounded-

ness of the non-centered Gaussian Hardy-Littlewood maximal func-
tion", Proc. Amer. Math. Soc., 130 (2002), 73-79.

[14] L. Grafakos, J. Kinnunen, "Sharp inequalities for maximal functions

associated with general measures", Proc. Roy. Soc. Edinburgh, 128A
(1998), 717-723.

[15] L. Grafakos, S. Montgomery-Smith, "Best constants for uncentered

maximal functions", Bull. London Math. Soc., 29 (1997), 60-64.

[16] L. Grafakos, S. Montgomery-Smith, O. Motrunich, "A sharp esti-

mate for the Hardy-Littlewood maximal function", Studia Math., 134.1
(1999), 57-67.

[17] M. de Guzmán, "Real variable methods in Fourier Analysis", North-

Holland Pub. Co., 1981.

[18] G.H. Hardy, J.E. Littlewood, "A maximal theorem with function-
theoretic applications", Acta. Math., 54 (1930), 81-116.

[19] E. Hewitt, K. Stromberg, "Real and Abstract Analysis", Springer-

Verlag, New York, NY, (1965).

[20] J.K. Hunter, "Measure theory", Lecture notes, https://www.math.ucda

vis.edu/~hunter/measure_theory/measure_notes.pdf

[21] J. Kinnunen, "The Hardy-Littlewood maximal function of a Sobolev-

function", Israel J.Math., 100 (1997), 117-124.

[22] S. Korry, "Boundedness of Hardy-Littlewood maximal operator in

the framework of Lizorkin-Triebel spaces", Rev. Mat. Univ. Complut.
Madrid, 15 (2002), 401-416.
REFERENCES 71

[23] S. Korry, "Fixed points of the Hardy-Littlewood maximal operator",

Coll. Math., 52 (2001), 289-294.

[24] H. Luiro, "On the regularity of the Hardy-Littlewood maximal operator

on subdomains of Rn ", Proc. of the Edinburgh Math. Soc., 53 (2010),
211-237

[25] H. Luiro, "Regularity properties of maximal operators", Rep. Univ.

Jyväskylä Dept. Math. Stat., 114, 2008.

[26] J. Martin, J. Soria, "Characterization of rearrangement invariant spaces

with fixed points for the Hardy-Littlewood maximal operator", Annales
Academ. Sci. Fenn. Math., 31 (2006), 39-46.

[27] A.D. Melas, "The best constant for the centered Hardy-Littlewood max-
imal inequality", Annals of Math., 157 (2003), 647-688.

[28] D. Múller "A geometric bound for maximal functions associated to con-
vex bodes" Pacif. j. of Math., 142-2, (1990).

[29] A. Osȩkowski, "Best constants in the weak-type estimates for uncen-

tered Maximal operators", Glasg. Math. J., 54-3, (2012), 655-663.

[30] P. Sjörgen "A remark on the maximal function for measured in Rn ",
Amer. J. Math. 105 (1983), 1231-1233.

[31] E.M. Stein, "Maximal functions: Spherical means", Proc. Natl. Acad.
Sci. USA, 73-7, July 1976, Math, 2174-2175.

[32] E.M. Stein "The development if square functions if the work of A. Zyg-
mund", Bull. of the amer. math. soc., 7-2, September 1982.

[33] E.M. Stein, R. Shakarchi "Real analysis", Princeton univ. press (2005).

[34] E. Stein, J.O. Strömberg, "Behavior of maximal functios in Rn for large

n", Arkiv für mat., 21 (1), 259-269.

[35] E.M. Stein, S. Wainger, "Problems in harmonic analysis related to cur-

vature", Bull. Amer. Math. Soc., 84, 1239-1295.

[36] J.O. Strömberg ,"Weak estimates for maximal functions with rectangles
in certain directions", Inst. Mittag-Leffler, 10, Ark. Mat.15 (1977), 229-
240.
72 REFERENCES

[37] M. Trinidad Menarguez, F. Soria, "Weak type (1, 1) inequalities of max-

imal convolution operators", Rend. Circ. Mat. Palermo, 41 (1992), 342-
352.

[38] N. Wiener, "The ergodic theorem", Duke Math. J., 5-1 (1939), 1-18.