Restriction Theorems and Salem Sets

by

Kyle David Hambrook

A thesis submitted in partial fulfillment of the requirements for

the degree of

Doctor of Philosophy

in

The Faculty of Graduate and Postdoctoral Studies

(Mathematics)

The University of British Columbia

(Vancouver)

July 2015

c Kyle David Hambrook, 2015

Abstract

In the first part of this thesis, I prove the sharpness of the exponent range in the
L2 Fourier restriction theorem due to Mockenhaupt [37] and Mitsis [35] (with the
endpoint due to Bak and Seeger [1]) for measures on R. The proof is based on a
random Cantor-type construction of Salem sets due to Laba and Pramanik [31]. The
key new idea is to embed in the Salem set a small deterministic Cantor set that
disrupts the restriction estimate for the natural measure on the Salem set but does
not disrupt the measure’s Fourier decay.
In the second part of this thesis, I prove a lower bound on the Fourier dimension
of
E(Q, Ψ, θ) = {x ∈ R : kqx − θk ≤ Ψ(q) for infinitely many q ∈ Q},

where Q is an infinite subset of Z, Ψ : Z → [0, ∞), and θ ∈ R. This generalizes

theorems of Kaufman [29] and Bluhm [5] and yields new explicit examples of Salem
sets. I also prove a multi-dimensional analog of this result. I give applications of
these results to metrical Diophantine approximation and determine the Hausdorff
dimension of E(Q, Ψ, θ) in new cases.

ii
Preface

The main result in Chapter 2 of this thesis is a stronger version of the main result in
the paper [23], published by Izabella Laba and myself. The proof of the main result
in Chapter 2 is a minor reformulation of the proof of the main result in [23].
Chapter 3 of this thesis is adapted from my paper [22], which has been submitted
for publication. Some of the text of [22] has been used verbatim in this thesis. I am
the sole author of [22].

iii
Table of Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Notation, Conventions, and Basic Definitions . . . . . . . . . . . . . . 4
1.2 Hausdorff and Fourier Dimension; Salem Sets . . . . . . . . . . . . . 7
1.3 Restriction Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 The Sharpness of the Exponents in the Mockenhaupt-Mitsis-Bak-

Seeger Restriction Theorem . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1 Theorem 2.0.3 Implies Theorem 2.0.2 . . . . . . . . . . . . . . . . . . 20
2.2 The Numbers r, s, t, n, Sj , Tj , Nj , and Lj . . . . . . . . . . . . . . . 22
2.3 The Sequences of Sets (Aj )∞ ∞
j=0 and (Pj )j=0 . . . . . . . . . . . . . . . 23
2.4 The Measures µ and µj ; The Functions fl . . . . . . . . . . . . . . . 24
2.5 Existence of the Weak Limit . . . . . . . . . . . . . . . . . . . . . . . 25
2.6 The Ball Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.7 The Crucial Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . 26

iv
 2.8 Proof of the Crucial Inequality. I. . . . . . . . . . . . . . . . . . . . . 28
2.9 Proof of the Crucial Inequality. II. . . . . . . . . . . . . . . . . . . . 32
2.10 Proof of the Crucial Inequality. III. . . . . . . . . . . . . . . . . . . . 34
2.11 The Fourier Decay of µ . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2r c 2r
2.12 The Convergence of kfd l µj k2r to kfl µk2r . . . . . . . . . . . . . . . . . 37
2r
2.13 A Lower Bound on kfd l µj k2r . . . . . . . . . . . . . . . . . . . . . . . 38
2.14 The Divergence of kfc l µkp /kfl kL2 (µ) . . . . . . . . . . . . . . . . . . . 42

3 Explicit Salem Sets and Applications to Metrical Diophantine Ap-

proximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1 Remarks on the Proof of Theorems 3.0.1 and 3.0.2 . . . . . . . . . . . 46
3.2 Motivation: Explicit Salem Sets . . . . . . . . . . . . . . . . . . . . . 47
3.3 Motivation: Metrical Diophantine Approximation . . . . . . . . . . . 50
3.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5 Proof of Theorems 3.0.1 and 3.0.2: The Key Function . . . . . . . . . 59
3.6 Proof of Theorems 3.0.1 and 3.0.2: The Main Lemma . . . . . . . . . 62
3.7 Proof of Theorems 3.0.1 and 3.0.2: The Measure . . . . . . . . . . . . 65
3.8 Proof of Lemma 3.3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Directions for Future Work . . . . . . . . . . . . . . . . . . . . . . . . 71

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

Appendix A Hoeffding’s Inequality . . . . . . . . . . . . . . . . . . . . . 79

Appendix B A Fourier Decay Lemma . . . . . . . . . . . . . . . . . . . 81

v
Acknowledgements

I thank my supervisor, Izabella Laba, for her invaluable guidance throughout my

Ph.D. She taught me how to be a mathematician.
I thank my supervisory committee members, Malabika Pramanik and Akos Mag-
yar, for their advice and their time.
I thank my fiancée, Yumi Kondo, for her emotional support throughout my Ph.D.
Every mathematical breakthrough I have had during my Ph.D. was made sitting by
her side.

vi
Dedication

To Yumi

vii
Chapter 1

Introduction

We will begin this chapter by stating the main results of this thesis so that experts
can quickly discern its contribution. The notation, conventions, and basic definitions
used in this thesis are described in Section 1.1. Section 1.2 discusses the concepts
of Hausdorff dimension, Fourier dimension, and Salem sets and describes the con-
tribution of this thesis to the study of these concepts. Section 1.3 provides detailed
background and motivation for the restriction problem. It also explains the contri-
bution of this thesis to the study of the restriction problem.
This thesis consists of two main parts.
The first part concerns the following L2 Fourier restriction theorem.

Theorem 1.0.1. Let µ be a compactly supported probability measure on Rd . Suppose

there are α, β ∈ (0, d) such that

(A) µ(B(x, r)) . rα ∀x ∈ Rd , r > 0,

(B) µ(ξ)| . (1 + |ξ|)−β/2
|b ∀ξ ∈ Rd .

Then for all p ≥ 2(2d − 2α + β)/β we have

(1.0.1) kfcµkp . kf kL2 (µ) ∀f ∈ L2 (µ).

Mockenhaupt [37] and Mitsis [35] independently proved Theorem 1.0.1 for p >

1
2(2d − 2α + β)/β. The endpoint case p = 2(2d − 2α + β)/β was proved by Bak
and Seeger [1]. Theorem 1.0.1 generalizes the Tomas-Stein theorem, which is the
following.

Theorem 1.0.2 (Tomas-Stein theorem). Let µ be the uniform probability measure

on the unit sphere S d−1 in Rd , d ≥ 2. Then for all p ≥ (2d + 2)/(d − 1) we have

kfcµkp . kf kL2 (µ) ∀f ∈ L2 (µ).

Theorem 1.0.1 implies the Tomas-Stein theorem because (A) and (B) hold with
α = β = d − 1 when µ is the uniform probability measure on S d−1 .
The range of p in the Tomas-Stein theorem is optimal in the sense that the
theorem does not hold if (2d + 2)/(d − 1) is replaced by any smaller number. The
counterexample that shows this is called the Knapp example. In the Knapp example,
f is taken to be the characteristic function of a small (hence almost flat) spherical
cap (cf. [49, Chapter 7]).
Since Theorem 1.0.1 contains the Tomas-Stein theorem as a special case, the
range of p in Theorem 1.0.1 is also optimal. However, one can still ask whether
the range of p in Theorem 1.0.1 is optimal for specific classes of measures. The
Tomas-Stein theorem and analogs for curved hypersurfaces besides the sphere were
known for a long time (cf. [43]). The point of Theorem 1.0.1 is that it also applies to
measures on fractal sets and to measures on the real line R. Therefore, it is natural
to ask whether the range of p in Theorem 1.0.1 is optimal for such measures. This
question is difficult because there is no obvious analog of the Knapp example (with
its almost flat spherical cap) for such measures.
The main result of the first part of this thesis is the following theorem, which
shows that the range of p in Theorem 1.0.1 is optimal for measures µ on R that
satisfy (A) and (B) with β ≤ α.

Theorem 1.0.3. Given any 0 < β ≤ α < 1 and given any p∗ < 2(2 − 2α + β)/β
there is a compactly supported probability measure on R that satisfies (A) and (B)
but fails to satisfy (2.0.1) whenever p ≤ p∗ .

2
 Chapter 2 of this thesis is devoted to Theorem 1.0.3 (and a slightly stronger
variant). Theorem 1.0.3 strengthens a result proved by Izabella Laba and myself in
2013 [23]. The proof of Theorem 1.0.3 is a minor modification of the proof given
in [23], but it gives a stronger result. More details on how the argument in this
thesis differs from the argument in [23] are also given in Chapter 2. Background on
restriction theorems and further motivation for Theorem 1.0.3 is given in Section 1.3
of the current chapter.
The second part of this thesis concerns the explicit construction of Salem sets.
A Salem set is a subset of Rd whose Hausdorff and Fourier dimension are equal.
Every countable set in Rd is Salem set of dimension 0, and Rd itself is a Salem set
of dimension d. The classic example of a non-trivial Salem set is a sphere in Rd ,
which is a Salem set of dimension d − 1. The existence of Salem sets in R of arbi-
trary dimension α ∈ (0, 1) was proved by Salem [41] using a random Cantor-type
construction. Kahane [27] showed that for every α ∈ (0, d) there is a Salem set in
Rd of dimension α by considering the images of compact subsets of R under certain
stochastic processes (see also Chapters 17 and 18 of [28]). Moreover, Kahane [28,
Chapter 18] showed that the image of any compact subset of Rd under fractional
Brownian motion is almost surely a Salem set, indicating that Salem sets are ubiqui-
tous amongst random sets. Recently, other random constructions of Salem sets have
been given by Bluhm [6] and Laba and Pramanik [31]. These random constructions
cannot produce any explicit examples of Salem sets. Kaufman [29] was the first to
find an explicit Salem set of dimension α ∈ / {0, d − 1, d}. The set Kaufman proved
to be Salem is

x ∈ R : kqxk ≤ |q|−τ for infinitely many q ∈ Z ,



where τ > 1 and kzk denotes the distance from z ∈ R to the nearest integer. That
this set has Hausdorff dimension 2/(1 + τ ) is the classical Jarnı́k-Besicovitch theorem
[26], [3]. Kaufman showed the Fourier dimension of this set is also 2/(1 + τ ). In his

3
thesis, Bluhm [5] generalized Kaufman’s result by showing that the set

{x ∈ R : kqxk ≤ ψ(|q|) for infinitely many q ∈ Z}

is a Salem set for any decreasing function ψ : N → (0, ∞). Using a theorem of
Gatesoupe [19], explicit Salem sets in Rd of any dimension α between d − 1 and d
can be obtained by considering radial versions of the sets of Kaufman and Bluhm.
Besides those already mentioned, no other explicit constructions of Salem sets
in R of dimension α 6= 0, 1 or in Rd of dimension α 6= 0, d − 1, d were known until
now. Moreover, even now, no explicit sets in Rd (Salem or otherwise) with Fourier
dimension strictly between 1 and d − 1 are known.
The main results of the second part of this thesis generalize the theorems of
Kaufman [29] and Bluhm [5] and exhibit many new explicit Salem sets in R. To be
more specific, let Q be an infinite subset of Z, let Ψ : Z → [0, ∞), and let θ ∈ R. We
prove a lower bound on the Fourier dimension of the set

{x ∈ R : kqx − θk ≤ Ψ(q) for infinitely many q ∈ Q}

and prove that it is Salem set for many sets Q and functions Ψ (regardless of the
value of θ). We also prove a multi-dimensional analog of this result and exhibit new
explicit sets in Rd with Fourier dimension larger than 1. We give applications of these
results to metrical Diophantine approximation. Chapter 3 of this thesis is devoted
to these results. A paper on these results was submitted for publication in 2014 (see
[22]).

1.1 Notation, Conventions, and Basic Definitions

In this section, we explain the notation and conventions we use that may be less
than universal. We also give some basic definitions.
The expression X . Y stands for “there is a constant C > 0 such that X ≤ CY .”
The expression X & Y is analogous.

4
 If A ⊆ Rd is a finite set, |A| is the cardinality of A. If A ⊆ Rd is an infinite set,
|A| is the Lebesgue outer measure of A.
For any A, B ⊆ R and c ∈ R we define

A + B = {a + b : a ∈ A, b ∈ B} ,
c + A = {c + a : a ∈ A} ,
cA = {ca : a ∈ A} .

The diameter of a set A ⊆ Rd is diam(A) = sup {|x − y| : x, y ∈ A}. For x ∈ Rd

and r > 0, B(x, r) denotes the open ball in the Euclidean norm with center x and
radius r.
We let N = {1, 2, . . .} be the set of positive integers and N0 = {0, 1, 2, . . .} be the
set of non-negative integers. For n ∈ N, let [n] = {0, . . . , n − 1}.
For x ∈ R, kxk = mink∈Z |x − k| is the distance from x to the nearest integer.
The support of a function f : Rd → C is the closure of the set of points x ∈ Rd
where f (x) 6= 0; it is denoted supp(f ).
For k ∈ N, we let C k (Rd ) denote the set of functions f : Rd → C that are k-times
continuously differentiable. We let C ∞ (Rd ) denote the set of functions f : Rd → C
that are infinitely differentiable. For k ∈ N ∪ {∞}, we let Cck (Rd ) denote the set
of functions f ∈ C k (Rd ) with compact support. When no confusion is possible, we
write C k for C k (Rd ) and Cck for Cck (Rd ).
Every measure on Rd we consider will be assumed to be defined on a σ-algebra
on Rd that contains the Borel σ-algebra on Rd .
The support of a measure on Rd is the smallest closed set whose complement has
measure 0. Thus any measurable set disjoint from the support has measure 0. Any
set that contains the support of a measure is said to support that measure. We write
supp(µ) for the support of a measure µ on Rd .
A probability measure on Rd is a measure µ with µ(Rd ) = 1.

5
 Let µ be a measure on Rd . For any µ-measurable function f : Rd → C,
Z 1/p
p
kf kLp (µ) = |f (x)| dµ(x) if 1 ≤ p < ∞,
Rd

kf kL∞ (µ) = inf{C ≥ 0 : |f (x)| ≤ C for µ-almost every x ∈ Rd }.

For p ∈ [1, ∞], the space of µ-measurable functions f : Rd → C with kf kLp (µ) < ∞
is denoted Lp (µ). It is a complete normed vector space with norm k kLp (µ) (provided
we identify functions that are equal µ-almost everywhere).
The Lebesgue measure on Rd will be denoted by λ. We will often use the notation
Lp (Rd ) = Lp (λ) and kf kp = kf kLp (Rd ) = kf kLp (λ) .
The Fourier transform of an L1 (λ) function f : Rd → C is defined by
Z
fb(ξ) = e−2πix·ξ f (x)dx for all ξ ∈ Rd .
Rd

If µ is a measure on Rd and f : Rd → C is an L1 (µ) function, we define

Z
(1.1.1) fcµ(ξ) = e−2πix·ξ f (x)dµ(x) for all ξ ∈ Rd .
Rd

The Fourier transform of a finite complex-valued measure µ on Rd is defined by

Z
µ
b(ξ) = e−2πix·ξ dµ(x) for all ξ ∈ Rd .
Rd

Let V and W be vector spaces with norms k kV and k kW , respectively. A linear

operator T : V → W is called bounded if there is a constant C > 0 such that

kT (v)kW ≤ CkvkV ∀v ∈ V.

A linear operator T : V → W is continuous if and only if it is bounded.

6
1.2 Hausdorff and Fourier Dimension; Salem Sets
For α ≥ 0, the α-dimensional Hausdorff content of a set A ⊆ Rd is
X
H α (A) = inf (diam(B))α ,
B
B∈B

S
where the infimum is over all countable collections B of balls such that A ⊆ B∈B B,
and diam(C) denotes the diameter of a set C ⊆ Rd .
The Hausdorff dimension of a set A ⊆ Rd is the supremum of all α ∈ [0, d] such
that H α (A) > 0. It is denoted by dimH (A).
The following lemma is due to Frostman [18] (cf. [34, Chapter 12], [49, Chapter
8]).

Lemma 1.2.1 (Frostman’s lemma). If A ⊆ Rd is a Borel set, the Hausdorff dimen-

sion of A is the supremum of all α ∈ [0, d] such that

µ(B(x, r)) . rα ∀x ∈ Rd , r > 0

for some probability measure µ on Rd with supp(µ) ⊆ A.

The Fourier dimension of a set A ⊆ Rd is the supremum of all β ∈ [0, d] such that

µ(ξ)| . (1 + |ξ|)−β/2
|b ∀ξ ∈ Rd

for some probability measure µ on Rd with supp(µ) ⊆ A. It is denoted by dimF (A).

It is a well-known consequence of Frostman’s lemma that

(1.2.1) dimF (A) ≤ dimH (A)

for all Borel sets A ⊆ Rd (cf. [34, Chapter 12], [49, Chapter 8]).
A set A ⊆ Rd with dimF (A) = dimH (A) is called a Salem set.
Here are some examples of Salem sets. Every countable set in Rd is a Salem set
of dimension 0. More generally, every Borel set in Rd of Hausdorff dimension 0 is

7
a Salem set of dimension 0. The set Rd itself is a Salem set of dimension d. More
generally, every set in Rd that contains a ball is a Salem set of dimension d. Every
sphere in Rd is a Salem set of dimension d − 1.
Here are some examples of non-Salem sets. All k-dimensional planes in Rd with
k < d have Hausdorff dimension k and Fourier dimension 0. The middle-thirds Can-
tor set in R has Hausdorff dimension (ln 2)/(ln 3) and Fourier dimension 0. Körner
[30] showed that for any given 0 ≤ β < α ≤ 1 there is a compact subset of R with
Fourier dimension β and Hausdorff dimension α.
Note that the concepts of Fourier dimension and Salem sets depend on the am-
bient space, unlike the concept of Hausdorff dimension. For example, the real line
R is a Salem set of dimension 1, but a line in R2 (which we can regard as a copy
of R embedded in R2 ) is a non-Salem set with Hausdorff dimension 1 and Fourier
dimension 0.
The existence of Salem sets in R for every dimension α ∈ (0, 1) was first proved
by Salem [41] using a random Cantor-type construction. The existence of Salem sets
in Rd for every dimension α ∈ (0, d) was first proved by Kahane [27] (cf. Chapters
17 and 18 of [28])) who considered the images of compact subsets of R under certain
stochastic processes. Other random constructions of Salem sets have been given by
Bluhm [6] and Laba and Pramanik [31]. These random constructions do not produce
any explicit examples of Salem sets. In fact, it is quite difficult to find explicit Salem
sets in Rd of dimensions other than 0, d − 1, and d. The first person to do so was
Kaufman [29]. He proved that the set

x ∈ R : kqxk ≤ |q|−τ for infinitely many q ∈ Z ,



where τ > 1, is a Salem set of dimension 2/(1 + τ ). Bluhm [5] showed that replacing
|q|−τ by ψ(|q|) for any decreasing function ψ : N → (0, ∞) also yields a Salem set.
No other explicit examples of Salem sets in R with dimension 0 < α < 1 were known
until now. A theorem of Gatesoupe [19] shows that for any Salem set E contained in
[0, 1] of dimension 0 < α < 1, the set of points x ∈ Rd such that |x| ∈ E is a Salem
set of dimension d − 1 + α. As it is not difficult to modify Kaufman’s proof to show

8
that
x ∈ [0, 1] : kqxk ≤ |q|−τ for infinitely many q ∈ Z


is a Salem set of dimension 2/(1 + τ ) when τ > 1, Gatesoupe’s theorem leads to

explicit Salem sets in Rd of every dimension α ∈ (d − 1, d). No explicit examples
of Salem sets in Rd of dimension less than d − 1 are known. Moreover, no explicit
sets (Salem or otherwise) in Rd of Fourier dimension strictly between 1 and d − 1 are
known.
In this thesis we generalize the theorems of Kaufman [29] and Bluhm [5]. To be
more specific, let Q be an infinite subset of Z, let Ψ : Z → [0, ∞), and let θ ∈ R. We
prove a lower bound on the Fourier dimension of the set

{x ∈ R : kqx − θk ≤ Ψ(q) for infinitely many q ∈ Q} ,

and prove that it is Salem for many sets Q and functions Ψ (regardless of the value of
θ). We also prove a multi-dimensional analog of this result and exhibit new explicit
sets in Rd with Fourier dimension larger than 1. We give applications of these results
to metrical Diophantine approximation.

1.3 Restriction Theorems

The restriction problem is one of the most important problems in harmonic analysis.
It is connected to many other important problems in mathematical analysis, includ-
ing the Kakeya conjecture, the Bochner-Riesz conjecture, the estimation of solutions
to the wave, Schroedinger, and Helmholtz equations, and the local smoothing conjec-
ture for partial differential equations (cf. [45]). It is even connected to Montgomery’s
conjecture in analytic number theory on the size of Dirichlet sums (see [9] and [10]).
The restriction problem was originally proposed by Stein in the late 1960s. It
originates from the study of the boundedness of the Fourier transform as a linear
operator between Lp spaces. To motivate it, we begin with the simplest case. Recall

9
that the Fourier transform of a function f ∈ L1 (λ) is defined by
Z
(1.3.1) fb(ξ) = e−2πix·ξ f (x)dx for all ξ ∈ Rd .
Rd

It is easy to see that the Fourier transform operator f 7→ fb is linear. Moreover, the
inequality | Rd e−2πix·ξ f (x)dx| ≤ Rd |f (x)|dx implies
R R

kfbkL∞ (µ) ≤ kf k1

for any measure µ, meaning the Fourier transform is always a bounded linear operator
from L1 (λ) to L∞ (µ).
The restriction problem asks: For which measures µ on Rd and for which values
of p and q is the Fourier transform a bounded linear operator from Lp (λ) to Lq (µ)?
In other words, it asks: For which measures µ on Rd and for which values of p and
q do we have an inequality of the form

(1.3.2) kfbkLq (µ) ≤ Cp,q kf kp

for all functions f in Lp (λ)? Here Cp,q is a constant. The inequality (1.3.2) is called
a restriction inequality.
Before saying anything more, there is a technical issue in (1.3.2) we need to
address. We haven’t defined fb for functions f outside of L1 (λ). In fact, it is not
true that the integral in (1.3.1) exists for all f in Lp (λ) when p > 1. However, the
integral in (1.3.1) is certainly defined when f ∈ Cc∞ (Rd ). A standard application
of Khinchin’s inequality (cf. [49, Chapter 4]) shows that (1.3.2) cannot hold for all
f ∈ Cc∞ (Rd ) when p > 2, so we can restrict attention to p ≤ 2. If for some p and q we
have (1.3.2) for all f ∈ Cc∞ (Rd ), then (since Cc∞ (Rd ) is dense in Lp (λ)) the Fourier
transform can be extended uniquely to all f ∈ Lp (λ) in such a way that (1.3.2) holds
for all f ∈ Lp (λ). Moreover, the extension agrees with (1.3.1) when f ∈ L1 (λ). In
light of this discussion, it is technically cleaner to view the restriction problem as
asking: When does an inequality of the form (1.3.2) hold for all f ∈ Cc∞ (Rd )?

10
 One can view the restriction problem as the problem of generalizing the classi-
cal Hausdorff-Young theorem, which says that the Fourier transform is a bounded
operator from Lp (λ) to Lq (λ) if and only if 1 ≤ p ≤ 2 and q = p0 = p/(p − 1).
To understand the term restriction in this context, note kfbkLq (µ) = kfb1S kLq (µ) ,
where S is the support of the measure µ. So (1.3.2) says the Fourier restriction
operator that maps f to fb1S (i.e., maps f to the restriction of the Fourier transform
of f to the set S) is a bounded operator from Lp (λ) to Lq (µ).
As noted above, the Fourier transform is a bounded operator from L1 (λ) to L∞ (µ)
for any measure µ. If µ is absolutely continuous with respect to Lebesgue measure,
the proof of the Hausdorff-Young theorem can be modified to give a restriction
inequality (1.3.2) with p and q depending on the regularity and support of the density
function of µ.
The restriction problem gets interesting if µ is singular with respect to Lebesgue
measure, so that the support of µ has Lebesgue measure 0. If p > 1, the Fourier
transform of an Lp (λ) function may be infinite on a set of Lebesgue measure 0. If the
intersection of that set with the support of µ has positive µ-measure, the left-hand
side of (1.3.2) will be infinite. Hence f 7→ fb may not be a bounded operator from
Lp (λ) to Lq (µ) for any q when µ is singular.
So which singular measures should we consider? Well, if the support S of µ is
contained in a hyperplane in Rd , we can construct a function that is in Lp (λ) for
every p > 1 and whose Fourier transform is infinite on all of S. For instance, if
ψ : Rd−1 → C is any continuous compactly supported function, then the function f
defined by
ψ(x2 , . . . , xd )
f (x1 , . . . , xd ) =
1 + |x1 |
is in Lp (λ) for every p > 1, but its Fourier transform is infinite everywhere on the

hyperplane ξ ∈ Rd : ξ1 = 0 . Hence we must consider measures µ supported on sets
that are not flat.
Perhaps the most natural singular measure to consider is the uniform probability
measure on S d−1 , the unit sphere in Rd . Indeed, this is the measure Stein considered
when he first proposed the restriction problem in the late 1960s.

11
 To discuss specific results, it will be convenient to restate the restriction problem
in its equivalent dual form. We do so via the following lemma.

Lemma 1.3.1. Let µ be a finite measure on Rd . Let p ∈ [1, ∞), p0 ∈ (1, ∞], and
q, q 0 ∈ [1, ∞] be such that p1 + p10 = 1, 1q + q10 = 1. The following statements are
equivalent.

(1.3.3) g kLq (µ) ≤ Ckgkp ∀g ∈ Cc∞ (Rd ).

kb
0
(1.3.4) khµk
c p0 ≤ Ckhk q0
L (µ) ∀h ∈ Lq (µ).

Proof. We show only that (1.3.3) implies (1.3.4) as the reverse implication is analo-
gous. By Fubini’s theorem, Hölder’s inequality, and (1.3.3), we have
Z Z
g(y)hµ(y)dy
c = gb(x)h(x)dµ(x) ≤ kb
g kLq (µ) khkLq0 (µ) ≤ Ckgkp khkLq0 (µ)
Rd Rd

0
for all g ∈ Cc∞ (Rd ) and h ∈ Lq (µ). By applying the formula
Z 
∞ d
khµk
c p0 = sup g(y)hµ(y)dy : g ∈ Cc (R ), kgkp = 1
c
Rd

(cf. [17, Chapter 6]), we obtain (1.3.4).

0
Since µ is assumed to be finite in Lemma 1.3.1, Lq (µ) ⊆ L1 (µ) and hµ c is well-
0
defined by (1.1.1) for every h ∈ Lq (µ). An analog of Lemma 1.3.1 can be proved
without the hypothesis that µ is finite. However, to avoid complications, we consider
only finite measures from now on.
In light of Lemma 1.3.1, it is clear an equivalent formulation of the restriction
problem is the following. For which measures µ on Rd and for which values of p and
q do we have an inequality of the form

(1.3.5) kfcµkp ≤ Cp,q kf kLq (µ)

for all functions f in Lq (µ)?

12
 Now we discuss some specific results on the restriction problem.
We start with a conjecture of Stein.

Conjecture 1.3.2 (Stein’s restriction conjecture). If µ is the uniform probability

measure on the unit sphere S d−1 in Rd , then (1.3.5) holds for all p > 2d/(d − 1)
when q = ∞.

In 1970, Fefferman and Stein (see [16], [43]) proved the conjecture for d = 2, but
it remains open for d ≥ 3 (and is nonsense for d = 1).
Another foundational result on the restriction problem is the Tomas-Stein theo-
rem of 1975 (see [43], [46], [47]).

Theorem 1.3.3 (Tomas-Stein theorem). If µ is the uniform probability measure on

the unit sphere S d−1 in Rd , then (1.3.5) holds for all p ≥ (2d+2)/(d−1) when q = 2.

Many researchers have studied the restriction problem for hypersurfaces besides
the sphere, the restriction problem for surfaces of lower dimension, and analogs of
the restriction problem in finite fields and in the primes (cf. [2], [12], [20], [36], [38],
[39], [45]).
Recently, there has been active research into the restriction problem for fractals,
where µ is a natural measure on a fractal set.
In 2000-2002, Mockenhaupt and Mitsis independently generalized the Tomas-
Stein restriction theorem to the fractal setting (see [35], [37]) with the following
theorem.

Theorem 1.3.4. Let µ be a compactly supported probability measure on Rd . Suppose

there are α, β ∈ (0, d) such that

(A) µ(B(x, r)) . rα ∀x ∈ Rd , r > 0,

(B) µ(ξ)| . (1 + |ξ|)−β/2
|b ∀ξ ∈ Rd .

Then (1.3.5) holds whenever p ≥ 2(2d − 2α + β)/β and q = 2.

Note the endpoint case p = 2(2d − 2α + β)/β was not proved by Mockenhaupt
or Mitsis; it was proved by Bak and Seeger [1] in 2011. Theorem 1.3.4 also holds for

13
any compactly supported finite measure on Rd , as any finite measure is a constant
multiple of a probability measure. If µ is the uniform probability measure on the
sphere in Rd , then (A) and (B) hold with α = β = d − 1, and so (1.3.5) holds for all
p ≥ (2d + 2)/(d − 1) and q = 2. Therefore Theorem 1.3.4 contains the Tomas-Stein
theorem as a special case.
We now explain how Theorem 1.3.4 is connected to fractals.
There is no consensus in the literature on the definition of a fractal, but we will
not need a formal definition anyway. Informally, one may think of a fractal as a set
in Rd that exhibits self-similarity, structure at arbitrarily small scales, and local and
global irregularity.
The complicated structure of fractal sets is often summarized by various numerical
dimensions, such as Hausdorff dimension and Fourier dimension. For example, the
middle-thirds Cantor set (one of the most famous fractals) has Hausdorff dimension
(ln 2)/(ln 3) and Fourier dimension 0.
Condition (A) is closely connected to Hausdorff dimension. Indeed, recall that
Lemma 1.2.1 says the Hausdorff dimension of a Borel set A ⊆ Rd is the supremum
of all α ∈ [0, d] for which there exists a probability measure µ supported on A such
that condition (A) holds. Consequently, if µ is a probability measure satisfying (A)
for some α ∈ [0, d], then α ≤ dimH (supp(µ)).
Condition (B) concerns the Fourier decay rate of µ and is connected to Fourier
dimension. The original proofs of the Tomas-Stein theorem for the sphere and analogs
of it for other surfaces hinge on deducing an estimate of the Fourier decay rate
of the relevant measure from the curvature of the surface on which the measure
is supported. By the definition of Fourier dimension in Section 1.2, a probability
measure that satisfies condition (B) must satisfy β ≤ dimF (supp(µ)).
Complimentary to the Tomas-Stein theorem is the important fact that the range
of p in the theorem cannot be enlarged. This is shown by the classic example of
Knapp, in which one takes f to be the characteristic function of a small (hence
almost flat) spherical cap (cf. [49, Chapter 7]). As Theorem 1.3.4 contains the
Tomas-Stein theorem as a special case, the range of p in Theorem 1.3.4 also cannot
be enlarged in general. However, the point of Theorem 1.3.4 is that it also applies to

14
measures supported on fractal sets and to measures on R. Therefore, it is interesting
to ask whether the range of p in Theorem 1.3.4 is best possible for such measures.
In 2013, Izabella Laba and I addressed this question (see [23]).
To understand the result of Laba and myself in [23], first observe that for a
probability measure µ on R that satisfies (A) for some α ∈ (0, 1) and satisfies (B)
for every β < α, the inequality (1.3.5) holds whenever p > 2(2 − α)/α and q = 2.
The result Laba and I established in [23] implies the following.
Theorem 1.3.5. Given any 0 < α < 1 and given any p∗ < 2(2 − α)/α there is a
compactly supported probability measure on R that satisfies (A) and satisfies (B) for
every β < α but does not satisfy (1.3.5) when p ≤ p∗ and q = 2.
The construction in [23] of the measure µ is based on the randomized Cantor-
type construction of Salem sets by Laba and Pramanik [31]. Laba and I modified the
construction so that the Cantor set that supports µ contains a smaller deterministic
Cantor set. Care must be taken in including the deterministic Cantor set to avoid
destroying the Fourier decay estimates on µ. The inequality (1.3.5) is shown to be
violated when p ≤ p∗ and q = 2 for any prescribed constant Cp,q by taking f to be
the characteristic function of a finite iteration of the deterministic Cantor set. The
argument ultimately hinges on counting solutions to certain linear equations in the
endpoints of the finite iterations of the larger Cantor set. To ensure the number of
solutions is sufficiently large, the embedded deterministic Cantor set is constructed
with the endpoints of each finite iteration in a generalized arithmetic progression.
In this thesis, we improve the result of [23]. In particular, we prove the following.
Theorem 1.3.6. Given any 0 < β ≤ α < 1 and given any p∗ < 2(2 − 2α + β)/β
there is a compactly supported probability measure on R that satisfies (A) and (B)
but does not satisfy (1.3.5) when p ≤ p∗ and q = 2.
This is a restatement of Theorem 1.0.3. In Chapter 2, we discuss the minor
differences between the proof of Theorem 1.3.6 (which we give in this thesis) and the
proof of the result of [23].
Theorem 1.3.6 establishes the sharpness of range of p in Theorem 1.3.4 for the
class of probability measures µ on R that satisfy (A) and (B) for some 0 < β ≤ α < 1.

15
In particular, Theorem 1.3.6 covers those measures that satisfy (A) and (B) with
α = β. Theorem 1.3.5 does not cover such measures; it only covers those measures
that satisfy (A) for some 0 < α < 1 and satisfy (B) for β arbitrarily close to (but
strictly less than) α. Theorem 1.3.6 establishes sharpness for those measures that
satisfy 0 < β < α < 1. Theorem 1.3.5 does not cover such measures because it
only permits taking p∗ up to the number 2(2 − α)/α, not up to the larger number
2(2 − 2α + β)/β.
Chen [11] modified the argument used by Laba and myself in [23] to eliminate
the dependence of the measure on p∗ . In particular, Chen [11] proved

Theorem 1.3.7. Given any 0 < β ≤ α < 1 there is a compactly supported probability
measure on R that satisfies (A) and (B) but does not satisfy (1.3.5) when p < 2(2 −
2α + β)/β and q = 2.

Evidently, Theorem 1.3.7 implies Theorem 1.3.6. Theorem 1.3.7 is perhaps also
a more satisfying sharpness statement than Theorem 1.3.6. The modification of the
argument of [23] that we present in this thesis to prove Theorem 1.3.6 does not follow
the modification of [11]. However, we do borrow some of his notation.

16
Chapter 2

The Sharpness of the Exponents in

the Mockenhaupt-Mitsis-Bak-
Seeger Restriction Theorem

In this chapter, we study the sharpness of the range of exponents in the following
general L2 restriction theorem of Mockenhaupt [37] and Mitsis [35] (with the endpoint
due to Bak and Seeger [1]).

Theorem 2.0.1. Let µ be a compactly supported probability measure on Rd . Suppose

there are α, β ∈ (0, d) such that

(A) µ(B(x, r)) . rα ∀x ∈ Rd , r > 0,

(B) µ(ξ)| . (1 + |ξ|)−β/2
|b ∀ξ ∈ Rd .

Then for all p ≥ 2(2d − 2α + β)/β we have

(2.0.1) kfcµkp . kf kL2 (µ) ∀f ∈ L2 (µ).

We will consider the following question. Is the range of p in Theorem 2.0.1 sharp?
In other words, does the theorem still hold if 2(2d−2α+β)/β is replaced by a smaller
number p∗ ?

17
 This question, as stated, is too general. The Knapp example, where f is the
characteristic function of a small spherical cap and µ is the uniform probability
measure on the unit sphere in Rd , shows the range of p is sharp (cf. [49, Chapter 7]).
However, we can consider the sharpness for specific sets of measures. The novelty
of Theorem 2.0.1 compared to earlier restriction theorems is that it also applies to
measures on fractal sets and to measures on R. Therefore, it is natural to consider
the sharpness of the range of p in Theorem 2.0.1 for such measures.
In this chapter, we will show the range of p in Theorem 2.0.1 is sharp for measures
µ on R that satisfy (A) and (B) with β ≤ α. In particular, I will prove the following
theorem.

Theorem 2.0.2. Given any 0 < β ≤ α < 1 and given any p∗ < 2(2 − 2α + β)/β
there is a compactly supported probability measure on R that satisfies (A) and (B)
but fails to satisfy (2.0.1) whenever p ≤ p∗ .

Theorem 2.0.2 will be deduced from the following stronger result.

Theorem 2.0.3. Let 0 < β0 ≤ α0 < 1 with α0 = ln a/ ln c and β0 = (2 ln a−ln b)/ ln c

for some a, b, c ∈ N. Given any p∗ < 2(2−2α0 +β0 )/β0 , there is a compactly supported
probability measure on R that satisfies the following.

(2.0.2) µ(B(x, r)) . rα0 ∀x ∈ R, ∀r > 0.

(2.0.3) µ(ξ)| . (1 + |ξ|)−β0 /2 (ln(1 + |ξ|))1/2
|b ∀ξ ∈ R.

There is a sequence of functions fl ∈ L2 (µ) such that

kfcl µkp
(2.0.4) lim =∞ ∀p ≤ p∗ .
l→∞ kfl kL2 (µ)

In section 2.1, we show that Theorem 2.0.3 implies Theorem 2.0.2. The rest of
the chapter constitutes the proof of Theorem 2.0.3.
In 2013, Izabella Laba and I [23] proved (implicitly) the following theorem, which
is the special case of Theorem 2.0.3 where α0 = β0 .

18
Theorem 2.0.4. Let 0 < α0 < 1 with α0 = ln a/ ln c for some a, c ∈ N. Given
any p∗ < 2(2 − α0 )/α0 , there is a compactly supported probability measure on R that
satisfies the following.

(2.0.5) µ(B(x, r)) . rα0 ∀x ∈ R, ∀r > 0.

(2.0.6) µ(ξ)| . (1 + |ξ|)−α0 /2 (ln(1 + |ξ|))1/2
|b ∀ξ ∈ R.

There is a sequence of functions fl ∈ L2 (µ) such that

kfcl µkp
(2.0.7) lim =∞ ∀p ≤ p∗ .
l→∞ kfl kL2 (µ)

Theorem 2.0.4 can be deduced from Theorem 2.0.3 by taking a = b.

Just as Theorem 2.0.3 implies Theorem 2.0.2, Theorem 2.0.4 implies the following
theorem, which says the the range of p in Theorem 2.0.1 is sharp for measures µ on
R that satisfy (A) for some α ∈ (0, 1) and satisfy (B) for every β < α.

Theorem 2.0.5. Given any 0 < α < 1 and given any p∗ < 2(2 − α)/α there is a
compactly supported probability measure on R that satisfies (A) and satisfies (B) for
every β < α but fails to satisfy (2.0.1) whenever p ≤ p∗ .

Let us compare what Theorem 2.0.2 and Theorem 2.0.5 tell us.
Theorem 2.0.2 establishes the sharpness of range of p in Theorem 2.0.1 for the
class of probability measures µ on R that satisfy (A) and (B) for some 0 < β ≤ α < 1.
In particular, it covers those measures that satisfy (A) and (B) with α = β. Theorem
2.0.5 does not cover such measures. Theorem 2.0.5 only covers those measures that
satisfy (A) for some 0 < α < 1 and satisfy (B) for all β < α.
Furthermore, Theorem 2.0.5 does not establish the sharpness of the range of
p in Theorem 2.0.1 for the class of probability measures µ on R that satisfy (A)
and (B) for some 0 < β < α < 1, while Theorem 2.0.2 does. It is true that
Theorem 2.0.5 produces, for any prescribed 0 < β < α < 1 and p∗ < 2(2 − α)/α, a
probability measure µ that satisfies (A) and (B) but does not satisfy (2.0.1) when
p ≤ p∗ . However, establishing sharpness for the class of measures under consideration

19
requires exhibiting, for any prescribed 0 < β < α < 1 and p∗ < 2(2 − 2α + β)/β, a
probability measure µ that satisfies (A) and (B) but does not satisfy (2.0.1) when
p ≤ p∗ . Theorem 2.0.5 does not allow us to take p∗ all the way up to 2(2 − 2α + β)/β.
It only permits taking p∗ up to the number 2(2 − α)/α, which is strictly less than
2(2 − 2α + β)/β when β < α.
The proof of Theorem 2.0.3 is a straightforward modification of the proof of
Theorem 2.0.4 given in [23]. It is based ultimately on the randomized Cantor-type
construction of Salem sets by Laba and Pramanik [31]. We modify the construction
so that the Cantor set that supports µ contains a smaller deterministic Cantor set.
The functions f` are the characteristic functions of the finite iterations of this de-
terministic Cantor set. We must be careful that the deterministic Cantor set is not
so large that it destroys the Fourier decay estimates furnished by the randomness of
the larger Cantor set. The divergence of kfc ` µkp /kf` kL2 (µ) relies on counting solutions

to linear equations in the endpoints of the finite iterations of the larger Cantor set.
The deterministic Cantor set is chosen with the endpoints of its finite iterations in
a multi-scale arithmetic progression so that this count is sufficiently large to ensure
divergence.
The difference between the proof of Theorem 2.0.3 and the proof of Theorem 2.0.4
given in [23] is the presence of β0 = (2 ln a − ln b)/ ln c in addition to α0 = ln a/ ln c.
This is handled by simply using β0 instead of α0 in all estimates related to the
Fourier decay of µ. The form of β0 is chosen so that the length (s = nα0 −β0 /2 in the
terminology to be introduced in Section 2.2) of the single-scale arithmetic progression
(which underlies the multi-scale arithmetic progression) is an integer.

2.1 Theorem 2.0.3 Implies Theorem 2.0.2

In this section, we show that Theorem 2.0.3 implies Theorem 2.0.2.
Let 0 < β ≤ α < 1 and p∗ < 2(2 − 2α + β)/β be given. Choose  > 0 small

20
enough that α +  < 1 and

2(2 − 2(α + ) + (β + ))

(2.1.1) p∗ < .
β+

Choose c ∈ N large enough that

1 
<
ln c 2
Let a ∈ N be the smallest positive integer that

 ln a
(2.1.2) α+ < .
2 ln c

Then

 ln(a − 1) ln a 1 ln a 
(2.1.3) α+ ≥ > − > − .
2 ln c ln c ln c ln c 2

Therefore, by combining (2.1.2) and (2.1.3), we obtain


(2.1.4) α<α+ < α0 < α + ,
2

where
ln a
α0 = .
ln c
Let b ∈ N be the largest positive integer such that

ln b
(2.1.5) < 2α0 − β.
ln c

Then

ln(b + 1) ln b 1 ln b 
(2.1.6) 2α0 − β ≤ < + < + .
ln c ln c 2 ln c ln c 2

Therefore, by combining (2.1.5) and (2.1.6), we obtain


(2.1.7) β < β0 < β + < β + ,
2

21
where
ln b 2 ln a − ln b
β0 = 2α0 − = .
ln c ln c
Since β ≤ α and α + /2 < α0 , it follows from (2.1.7) that

β0 < α0 .

By (2.1.1), (2.1.4), and (2.1.7), we have

2(2 − 2(α + ) + (β + )) 2(2 − 2α0 + β0 )

p∗ < < .
β+ β0

Theorem 2.0.3 now gives a compactly supported probability measure µ on R and

a sequence of functions fl ∈ L2 (µ) such that (2.0.2), (2.0.3), and (2.0.4) hold. It is
clear that (2.0.4) implies the failure of (2.0.1) for every p ≤ p∗ . Since β < β0 , it is
also clear that (2.0.3) implies (B). Since α < α0 , (2.0.2) implies

µ(B(x, r)) . rα ∀x ∈ R, 0 < r < 1.

Since µ is a probability measure, we also have

µ(B(x, r)) ≤ µ(R) = 1 ≤ rα ∀x ∈ R, r ≥ 1.

This establishes (A) and completes the proof of Theorem 2.0.2 assuming Theorem
2.0.3.

2.2 The Numbers r, s, t, n, Sj , Tj , Nj , and Lj

We now begin the proof of Theorem 2.0.3. We start by defining some important
numbers.
Let 0 < β0 ≤ α0 < 1 with α0 = ln a/ ln c and β0 = (2 ln a − ln b)/ ln c for some
a, b, c ∈ N. Let p∗ < 2(2 − 2α0 + β0 )/β0 . Choose r ∈ N such that rβ0 > 1 and
2r ≥ p∗ . Define n = c2m , t = a2m = nα0 , and s = bm = nα0 −β0 /2 , where m is a fixed

22
positive integer chosen so large that

2(2 − 2α0 + β0 ) 4 ln r
(2.2.1) p∗ < − .
β0 β0 ln n

Define Sj = sj , Tj = tj , Nj = nj , and Lj+1 = ln(400(j + 1)Nj+1 ) for j ∈ N0 .

2.3 The Sequences of Sets (Aj )∞ ∞

j=0 and (Pj )j=0

In this section, we describe the sequences of sets (Aj )∞ ∞

j=0 and (Pj )j=0 .
Each Aj will be the set of left endpoints of the intervals in the j-th iteration of
the Cantor set supporting µ. Each Pj will be the set of left endpoints of the intervals
in the j-th iteration of the smaller deterministic Cantor set we embed within the
Cantor set supporting µ.
Let (Pj )∞
j=0 be the sequence of sets defined by

1 [ 1
P0 = {0} , Pj+1 = Pj + {1, . . . , s} = a+ {1, . . . , s} .
Nj+1 a∈P
Nj+1
j

We will construct inductively a sequence of sets (Aj )∞

j=0 of the form

• A0 = {0}
S
• Aj+1 = a∈Aj a + Aj+1,a
1
• Aj+1,a ⊆ [n]
Nj+1
• |Aj+1,a | = Tj+1

• Aj ⊆ Pj .
Note

1 1
|Aj | = Tj , Aj ⊆ [Nj ], Aj,a ⊆ {1, . . . , s}
Nj Nj

for all j ∈ N. The details of the construction of (Aj )∞

j=0 will be given in Section 2.8.

23
2.4 The Measures µ and µj ; The Functions fl
In this section, we define the probability measures µ and µj (j ∈ N0 ) on R and the
functions fl : R → C (l ∈ N0 ).
For j ∈ N0 , define

Ej = Aj + [0, Nj−1 ]

and define µj to be the uniform probability measure on Ej , i.e.,

1 N X
dµj = 1Ej dx = j 1 −1 dx.
|Ej | Tj a∈A [a,a+Nj ]
j

Define µ to be the weak limit of (µj )∞

j=0 , i.e.,

µ = lim µj .
j→∞

We prove the weak limit exists in Section 2.5. A good reference on weak convergence
is [4]. Note
\∞
supp(µ) = Ej .
j=1

Let j ∈ N0 and a ∈ Aj . Then

1
µj ([a, a + Nj−1 ]) = ,
Tj

and in fact
1
µk ([a, a + Nj−1 ]) = ∀k ∈ N0 , k ≥ j.
Tj
Consequently,
1
µ([a, a + Nj−1 ]) = .
Tj
For l ∈ N0 , define
Fl = Pl + [0, Nl−1 ) and fl = 1Fl .

24
2.5 Existence of the Weak Limit
In this section, we prove the existence of the weak limit

µ = lim µj .
j→∞

For all other sections in this chapter, Fj is the set defined in Section 2.4. However,
in this section, Fj will be the cumulative distribution function of µj , i.e.,

Fj (t) = µj ((−∞, t]), ∀t ∈ R.

Fix j ∈ N0 . For a ∈ Aj , we have Fj+1 (a) = Fj (a) by the definitions of µj+1 and
µj . For t ∈ R, let a(t) be the smallest element of Nj−1 [Nj ] such that a(t) ≤ t. Then

|Fj+1 (t) − Fj (t)| = |µj+1 ([a(t), t]) − µj ([a(t), t])|

≤ µj+1 ([a(t), t]) + µj ([a(t), t])
≤ µj+1 ([a(t), a(t) + N −j ]) + µj ([a(t), a(t) + N −j ])
2
≤ .
Tj

Therefore

2
kFj+1 − Fj k∞ ≤ ∀j ∈ N0 .
Tj
P∞ −1
Since Tj = njα0 for j ∈ N0 , we have j=0 Tj < ∞. It follows that (Fj )∞ j=1 is
a uniformly convergent sequence of continuous cumulative distribution functions.
Therefore the limit F is a continuous cumulative distribution function of a probability
measure µ. Hence (µj )∞ j=1 converges weakly to µ.

25
2.6 The Ball Condition
In this section, we will show that µ(|I|) . |I|α0 for all intervals I ⊆ R. This is
equivalent to (2.0.2) in Theorem 2.0.3.
Let I be an interval in R. If |I| > 1, then

µ(I) ≤ µ(R) = 1 ≤ |I|α0

since µ is a probability measure. Now suppose |I| ≤ 1. Choose j ∈ N0 such that

−1
Nj+1 ≤ |I| ≤ Nj−1 . Assume I intersects supp(µ) (otherwise µ(I) = 0). Note

∞
\ ∞ [
\
supp(µ) = Ej = [a, a + Nj−1 ].
j=1 j=1 a∈Aj

Then I intersects an interval [a, a + Nj−1 ] for some a ∈ Aj . There are at most two
such intervals, say J1 and J2 , because |I| ≤ Nj−1 and Aj ⊆ Nj−1 [Nj ]. Therefore

µ(I) = µ(I ∩ J1 ) + µ(I ∩ J2 ) ≤ µ(J1 ) + µ(J2 )

1 1 2nα0
= + = α0 ≤ 2nα0 |I|α0 .
Tj Tj Nj+1

2.7 The Crucial Inequality

In this section, we state the crucial Fourier analytic inequality, explain the steps in
its proof, and explain its use.
Here is the crucial Fourier analytic inequality:
∞
X
−β0 /2
(2.7.1) |f\
l µj+1 (k) − fl µj (k)| . |k|
d (ln(e + |k|))1/2
j=0

for all k ∈ Z with k 6= 0 and all l ∈ N0 .

In the next three sections, we prove we can construct the sets (Aj )∞
j=0 so that
(2.7.1) holds.

26
 First we show we can construct the sets (Aj )∞
j=0 so that

1 X −β /2 1/2
(2.7.2) Ya (k) . Nj+10 Lj+1
Tj a∈Aj ∩Fl

for all k ∈ Z and all j, l ∈ N0 , j ≥ l. Here

 
1 X 1 X
Ya (k) = e−2πika  e−2πikb − e−2πikb  .
t b∈A n
j+1,a b∈[n]/Nj+1

Next we show
 
N j+1 1 X
(2.7.3) |f\
l µj+1 (k) − fl µj (k)| ≤ 2 min 1,
d Ya (k)
|k| Tj a∈Aj ∩Fl

for all k ∈ Z, k 6= 0 and all j, l ∈ N with j ≥ l.

Finally we show that
∞  
X Nj+1 −β /2 1/2
(2.7.4) min 1, Nj+10 Lj+1 . |k|−β0 /2 (ln(e + |k|))1/2
j=0
|k|

for all k ∈ Z, k 6= 0.
Evidently (2.7.1) follows upon combining (2.7.2), (2.7.3), and (2.7.4).
Once (2.7.1) is established, we will use it to prove

(2.7.5) µ(ξ)| . (1 + |ξ|)−β0 /2 (ln(e + |ξ|))1/2

|b ∀ξ ∈ R.

and

2r c 2r
(2.7.6) lim kfd
l µj k2r = kfl µk2r ∀l ∈ N0 .
j→∞

Note (2.7.5) is (2.0.3) in Theorem 2.0.3.

27
2.8 Proof of the Crucial Inequality. I.
In this section, we will define (Aj )∞
j=0 so that (2.7.2) holds. Recall that (2.7.2) is the
inequality

1 X −β /2 1/2
Ya (k) . Nj+10 Lj+1
Tj a∈Aj ∩Fl

for all k ∈ Z and all j, l ∈ N0 , j ≥ l. Here (as elsewhere)

 
1 X −2πikb 1 X
Ya (k) = e−2πika  e − e−2πikb  .
t b∈A n
j+1,a b∈[n]/Nj+1

We will define (Aj )∞

j=0 inductively. Fix j ∈ N0 . Assume Aj containing Pj is given.
S
We will choose Aj+1 = a∈Aj a + Aj+1,a by choosing Aj+1 , a for each a ∈ Aj .
We will need the following version of Hoeffding’s inequality.

Lemma 2.8.1. If X1 , . . . , Xn are independent complex-valued random variables with

EXi = 0 and |Xi | ≤ 1 for i = 1, . . . , n, then

n
!
1X
P Xi ≥ u ≤ 4 exp(−nu2 /4)
n i=1

for all u ≥ 0.

This version of Hoeffding’s inequality for complex-valued random variables is

deduced from the standard version for real-valued random variables in Appendix A.
−1
[n]. There are nt sets


Let Bj+1 be the collection of all size tj+1 subsets of Nj+1
−1
[n] is contained in exactly n−1


in Bj+1 . Moreover, each fixed b ∈ Nj+1 t−1
sets in Bj+1 .
For each a ∈ Aj , choose a set Bj+1,a independently and uniformly at random

28
from Bj+1 . For each a ∈ Aj and k ∈ Z, define the random variable
 
1  1 X −2πikb 1 X
Xa (k) = e−2πiak  e − e−2πikb  .

2 t b∈B n −1
j+1,a b∈Nj+1 [n]

For each a ∈ Aj , we have

 
X 1 X X −2πiξb
E e−2πiξb  = n

e
b∈Bj+1,a t B∈Bj+1 b∈B
 
1 X n − 1 −2πiξb
= n
e
t −1
t−1
b∈Nj+1 [n]

t X
= e−2πiξb .
n −1
b∈Nj+1 [n]

It follows that EXa (k) = 0 for all a ∈ Aj and k ∈ Z. Clearly |Xa (k)| ≤ 1 for all
a ∈ Aj and k ∈ Z.
Fix k ∈ [Nj+1 ]. The random variables (Xa (k))a∈Aj are independent because of
how we choose the sets (Bj+1,a )a∈Aj . Note

|Aj ∩ Fl | = sl tj−l = (Sl /Tl )Tj ∀l ∈ N0 , l ≤ j.

Define ul,j ≥ 0 by

u2l,j = 4(Sl /Tl )−1 Tj−1 ln(400(j + 1)Nj+1 ) ∀l ∈ N0 , l ≤ j.

Lemma 2.8.1 implies

 
1 X 1
P Xa (k) ≥ ul,j  ≤ 4 exp(−(Sl /Tl )Tj u2l,j /4) = .
(Sl /Tl )Tj a∈Aj ∩Fl
100(j + 1)Nj+1

29
Then
 
1 X 1
P Xa (k) ≥ ul,j for some k ∈ [Nj+1 ], l ∈ {0, . . . , j} ≤ .
Tj Sl−1 a∈Aj ∩Fl
100

Therefore there is some realization of the randomly chosen sets (Bj+1,a )a∈Aj such
that

1 X
(2.8.1) Xa (k) < ul,j
(Sl /Tl )Tj a∈Aj ∩Fl

for all k ∈ [Nj+1 ] and all l ∈ N0 , l ≤ j. Fix such a realization. Since Xa (k) is periodic
with period Nj+1 when a ∈ Aj , the inequality (2.8.1) holds for all k ∈ Z.
We cannot simply define Aj+1,a = Bj+1,a because then
[
Aj+1 = a + Aj+1,a
a∈Aj

may not contain

[
−1
Pj+1 = a + Nj+1 {1, . . . , s} .
a∈Pj

We define Aj+1,a for a ∈ Aj as follows. For a ∈ / Pj , define Aj+1,a = Bj+1,a . For

a ∈ Pj , define Aj+1,a to be the set obtained by first removing s arbitrary elements
−1
of Bj+1,a and then replacing them by the elements of Nj+1 {1, . . . , s}. Then Pj+1 ⊆
Aj+1 .
Recall that for a ∈ Aj and k ∈ Z we have
 
1 X 1 X
Xa (k) = e−2πika  e−2πikb − e−2πikb  ,
t b∈B n
j+1,a b∈[n]/Nj+1
 
1 X 1 X
Ya (k) = e−2πika  e−2πikb − e−2πikb  .
t b∈A n
j+1,a b∈[n]/Nj+1

30
 e−2πikb differs from
P
For a ∈ / Pj , we have Ya (k) = Xa (k). For a ∈ Pj , b∈Aj+1,a
−2πikb
P
b∈Bj+1,a e by at most s terms, hence

1 X X 2s
|Ya (k) − Xa (k)| = e−2πikb − e−2πikb ≤ .
t b∈Aj+1,a b∈Bj+1,a
t

Therefore, for all l ∈ N0 , l ≤ j, we have

1 X 1 X 1 X
Ya (k) − Xa (k) ≤ |Ya (k) − Xa (k)|
Tj a∈Aj ∩Fl
Tj a∈Aj ∩Fl
Tj a∈Aj ∩Fl

1 X
= |Ya (k) − Xa (k)|
Tj a∈P
j

|Pj | 2s
≤
Tj t
2Sj+1
= .
Tj+1

Combining this with (2.8.1) yields

1 X 2Sj+1
Ya (k) < (Sl /Tl )1/2 (4Tj−1 ln(400(j + 1)Nj+1 ))1/2 + .
Tj a∈Aj ∩Fl
Tj+1

Recall that s = nα0 −β0 /2 , t = nα0 , Sj = sj , Tj = tj , Nj = nj , and Lj+1 = ln(400(j +

1)Nj+1 ). It follows that

1 X −β /2 1/2
Ya (k) . Nj+10 Lj+1 .
Tj a∈Aj ∩Fl

31
2.9 Proof of the Crucial Inequality. II.
In this section, we prove (2.7.3), i.e.,

 
N j+1 1 X
|f\
l µj+1 (k) − fl µj (k)| ≤ 2 min 1,
d Ya (k)
|k| Tj a∈Aj ∩Fl

for all k ∈ Z, k 6= 0 and all j, l ∈ N with j ≥ l. Recall

 
1 X 1 X
Ya (k) = e−2πika  e−2πikb − e−2πikb  .
t b∈A n
j+1,a b∈[n]/Nj+1

Fix j, l ∈ N0 with j ≥ l. For a ∈ Aj and x ∈ R, define

ga,j (x) = N j 1[a,a+Nj−1 ] (x) = N j 1[0,1] (Nj (x − a)).

Then

1 X
dµj (x) = ga,j (x)dx,
Tj a∈A
j

and so

1 X 1 X
fl dµj (x) = fl (x)ga,j (x)dx = ga,j (x)dx.
Tj a∈A Tj a∈Aj ∩Fl
j

For ξ ∈ R,

gc
a,j (ξ) = e 1[0,1] (ξ/Nj ).
−2πiaξ d

Therefore

1 d
1[0,1] (ξ/Nj )
X
(2.9.1) fd
l µj (ξ) = e−2πiaξ .
Tj a∈A ∩F j l

32
Likewise

1 d
1[0,1] (ξ/Nj+1 )
X
(2.9.2) f\
l µj+1 (ξ) = e−2πiaξ .
Tj a∈A ∩F j+1 l

We now rewrite (2.9.1) and (2.9.2). First, by the geometric sum formula,

−2πi(ξ/Nj )1d
[0,1] (ξ/Nj ) = e
−2πiξ/Nj
− 1 = e−2πi(ξ/Nj+1 )n − 1
X
= (e−2πi(ξ/Nj+1 ) − 1) e−2πi(ξ/Nj+1 )k
k∈[n]

= −2πi(ξ/Nj+1 )1d
X
[0,1] (ξ/Nj+1 ) e−2πi(ξ/Nj+1 )n
k∈[n]
1
= −2πi(ξ/Nj )1d
X
[0,1] (ξ/Nj+1 ) e−2πiξb .
n −1
b∈Nj+1 [n]

Therefore

1
1d
[0,1] (ξ/Nj ) = 1[0,1] (ξ/Nj+1 )
X
d e−2πiξb .
n −1
b∈Nj+1 [n]

Applying this to (2.9.1) gives

1 d 1
1
X X
(2.9.3) fdµ
l j (ξ) = [0,1] (ξ/Nj+1 ) e−2πiaξ e−2πiξb .
Tj a∈A ∩F
nj+1 −1
j l b∈Nj+1 [nj+1 ]

On the other hand, since

[
Aj+1 = a + Aj+1,a ,
a∈Aj

33
we have

1
1d
X X
(2.9.4) f\
l µj+1 (ξ) = [0,1] (ξ/Nj+1 ) e−2πiξ(a+b)
Tj+1 a∈Aj ∩Fl b∈Aj+1,a
1 d 1 X −2πiξb
1[0,1] (ξ/Nj+1 )
X
= e−2πiξa e .
Tj t b∈A
a∈A j ∩Fl j+1,a

Fix k ∈ Z, k 6= 0. Comparing (2.9.3) and (2.9.4) yields

1
l µj+1 (k) − fl µj (k) = 21[0,1] (k/Nj+1 )
X
(2.9.5) f\ d d Ya (k).
Tj a∈Aj ∩Fl

Since

|1d
[0,1] (k/Nj+1 )| ≤ 1[0,1] (0) = 1
d

and

e−2πik/Nj+1 − 1 1
|1d
[0,1] (k/Nj+1 )| = ≤ ,
−2πik/Nj+1 π|k|/Nj+1

it follows from (2.9.5) that

 
Nj+1 1 X
|f\
l µj+1 (k) − fl µj (k)| ≤ 2 min 1,
d Ya (k) .
|k| Tj a∈Aj ∩Fl

2.10 Proof of the Crucial Inequality. III.

In this section, we prove (2.7.4), i.e.,
∞  
X Nj+1 −β /2 1/2
min 1, Nj+10 Lj+1 . (ln(e + |k|))1/2 |k|−β0 /2
j=0
|k|

for all ∀k ∈ Z, k 6= 0.

34
 Fix k ∈ Z, k 6= 0. Let j∗ be the smallest non-negative integer such that Nj∗ +1 >
|k|. Then Nj∗ ≤ |k|. Our strategy is to the split the sum above into a sum over
0 ≤ j < j∗ and a sum over j ≥ j∗ and estimate each separately. For any j ∈ N0 ,

Lj+1 = ln(400(j + 1)Nj+1 ) ≤ 10 ln(Nj+1 ).

If 0 ≤ j < j∗ , then
Lj+1 ≤ 10 ln(Nj∗ ) ≤ 10 ln |k|.

If j ≥ j∗ , then

Lj+1 ≤ 10 ln(Nj−j∗ +1 Nj∗ ) ≤ 10 ln(Nj−j∗ +1 |k|)

≤ 20 ln(Nj−j∗ +1 ) ln(e + |k|) = 20(j − j∗ + 1) ln(n) ln(e + |k|).

For the sum over 0 ≤ j < j∗ , we have

 1−β0 /2
1 X 1−β0 /2 1/2 1 X Nj∗ 1/2
Nj+1 Lj+1 = j −j−1
Lj+1
|k| 0≤j<j |k| 1≤j<j n ∗
∗ ∗
X 1
≤ (10 ln |k|)1/2 |k|−β0 /2 (j −j−1)(1−β 0 /2)
1≤j<j
2 ∗
∗
∞
X 1
= (10 ln |k|)1/2 |k|−β0 /2 .
j=0
2j(1−β0 /2)

For the sum over j ≥ j∗ , we have

X −β /2 1/2
X
−β0 /2 1/2
Nj+10 Lj+1 = Nj∗ +1 nj−j∗ Lj+1
j≥j∗ j≥j∗
X
≤ (20 ln(n) ln(e + |k|))1/2 |k|−β0 /2 2−(j−j∗ )β0 /2 (j − j∗ + 1)
j≥j∗
X∞
= (20 ln(n) ln(e + |k|))1/2 |k|−β0 /2 2−jβ0 /2 (j + 1).
j=0

35
Combining the estimates yields
∞  
X Nj+1 −β /2 1/2
min 1, Nj+10 Lj+1 . (ln(e + |k|))1/2 |k|−β0 /2 .
j=0
|k|

2.11 The Fourier Decay of µ

In this section, we use the crucial inequality (2.7.1) to prove the inequality (2.7.5),
which is

µ(ξ)| . (1 + |ξ|)−β0 /2 (ln(e + |ξ|))1/2

|b ∀ξ ∈ R.

This is (2.0.3) in Theorem 2.0.3.

Since µj converges to µ weakly, we have that µbj converges to µ
b pointwise. There-
fore
∞
X
(2.11.1) µ
b(k) = µb0 (k) + j+1 (k) − µ
(µd bj (k)) ∀k ∈ Z, k 6= 0.
j=0

Since dµ0 = 1[0,1] dx, we have

e−2πik − 1 1
(2.11.2) |µb0 (k)| = ≤ ∀k ∈ Z, k 6= 0.
−2πik |k|

Since f0 = 1[0,1] , we have f0 µj = µj for all j ∈ N0 , so we can apply (2.7.1) and

(2.11.2) to the series in (2.11.1) to obtain

µ(k)| . (ln(e + |k|))1/2 |k|−β0 /2

|b ∀k ∈ Z, k 6= 0.

Since (1 + |k|)β0 /2 ≤ 2|k|β0 /2 for k ∈ Z, k 6= 0, and since µ

b(0) = 1, it follows that

µ(k)| . (ln(e + |k|))1/2 (1 + |k|)−β0 /2

|b ∀k ∈ Z.

36
Then, by applying Lemma B.0.1 in Appendix B, we arrive at

µ(ξ)| . (1 + |ξ|)−β0 /2 (ln(e + |ξ|))1/2

|b ∀ξ ∈ R.

2r c 2r
2.12 The Convergence of kfd
l µj k2r to kfl µk2r

In this section, we use the crucial inequality (2.7.1) to prove (2.7.6), i.e., to prove

2r c 2r
lim kfd
l µj k2r = kfl µk2r ∀l ∈ N0 .
j→∞

Fix l ∈ N0 . Clearly fl = 1Fl is a bounded Borel measurable function. Since µ is

continuous (i.e., the cumulative distribution function of µ is continuous), the finite
set of points where fl is discontinuous has µ-measure 0. Since µj converges to µ
weakly, we have Z Z
lim gdµj → gdµ
j→∞ R R

for all bounded Borel measurable functions g : R → C for which the set of points
at which g is discontinuous has µ-measure 0. Applying this to g(x) = e−2πiξx fl (x)
gives that fdl µj converges to fl µ pointwise as j → ∞. Of course, fl µi converges to
c d
l µ pointwise as i → ∞ also. Therefore, for all k ∈ Z and j ∈ N0 ,
fc
∞
X
l µ(k) − fl µj (k) =
fc l µi+1 (k) − fl µi (k)).
(f\
d d
i=j

Then (2.7.1) yields

−β0 /2
|fc
l µ(k) − fl µj (k)| . |k|
d (ln(e + |k|))1/2

for all k ∈ Z, k 6= 0 and all j ∈ N0 , j ≥ l. Since (1 + |k|)β0 /2 ≤ 2|k|β0 /2 for

k ∈ Z, k 6= 0, and since

l µ(0) − fl dµl (0) = µ(Fl ) − µl (Fl ) = 0,

fc [

37
it follows that

−β0 /2
|fc
l µ(k) − fl µj (k)| . (1 + |k|)
d (ln(e + |k|))1/2

for all k ∈ Z and j ∈ N0 with j ≥ l. By applying Lemma B.0.1 from Appendix B to

the measure fl (µ − µj ), we obtain

−β0 /2
|fc
l µ(ξ) − fl µj (ξ)| . (1 + |ξ|)
d (ln(e + |ξ|))1/2

for all ξ ∈ R and j ∈ N0 with j ≥ l. Then since r ∈ N was chosen so that rβ0 > 1
and since
l µ(ξ) − fl µj (ξ)) → 0
lim (fc d
j→∞

pointwise for every ξ ∈ R, the dominated convergence theorem implies

lim kfc
l µ(ξ) − fl µj (ξ)k2r = 0.
d
j→∞

It follows by the triangle inequality that

2r c 2r
lim kfd
l µj k2r = kfl µk2r .
j→∞

2r
2.13 A Lower Bound on kfd
l µj k2r

In this section, we show that

 2r
2r Sl Nl
(2.13.1) kfd
l µj k2r ≥ Cr l+1
∀j, l ∈ N0 , j ≥ l,
Tl r Sl

where Cr > 0 is a constant depending only on r.

Fix j, l ∈ N0 with j ≥ l. Recall from (2.9.1) that

1 d
1[0,1] (ξ/Nj )
X
fd
l µj (ξ) = e−2πiξa
Tj a∈A ∩F j l

38
for all ξ ∈ R. Therefore

1 \
1[− 12 , 12 ] (ξ/Nj )e−πξ/Nj
X
fd
l µj (ξ) = e−2πiξa .
Tj a∈A ∩F j l

Then
2r
1 2r
1\
X
|fd
l µj (ξ)|
2r
= e−2πiξa [− 21 , 21 ] (ξ/Nj )e
−πiξ/Nj
Tj2r a∈Aj ∩Fl
2r
1  2r
1\
X
= 2r e−2πiξa [− 21 , 21 ] (ξ/Nj )
Tj a∈Aj ∩Fl

1
1
X Pr P2r
= e−2πiξ( s=1 as − s=r+1 as ) \
∗2r
[− 12 , 12 ]
(ξ/Nj ).
Tj2r a1 ,...,a2r ∈Aj ∩Fl

Here we have used

e−πiξ/Nj − eπiξ/Nj sin(πξ/Nj )

1\
[− 21 , 21 ] (ξ/Nj ) = = ∈ R,
−2πiξ/Nj πξ/Nj
 2r
and the formulas |z| 2r r
= z z and fbr ∗2r , where f ∗2r is the 2r-fold convolution
= fd
of f . Therefore
Z Z
kfl µj k2r
2r = d 2r
|fl µj (ξ)| dξ = Nj |fd 2r
l µj (Nj ξ)| dξ
R Z R
Nj
e−2πiξNj ( s=1 as − s=r+1 as ) 1\
X Pr P2r
∗2r
= 2r [− 1 1 (ξ)dξ.
, ]
Tj a ,...,a ∈A ∩F R 2 2
1 2r j l

By Fourier inversion,

r 2r
!!
Nj
1[−
X X X
∗2r
kfl µj k2r
2r = 2r 1 1
, ]
Nj as − as .
Tj a1 ,...,a2r ∈Aj ∩Fl
2 2
s=1 s=r+1

39
Since 1∗2r
[− 1 , 1 ]
≥ 0, all the terms in the sum are non-negative. Therefore
2 2

Nj ∗2r
kfl µj k2r
2r ≥ 1 1 1 (0)Mj,l,r ,
Tj2r [− 2 , 2 ]

where ( )
r
X 2r
X
Mj,l,r = (a1 , . . . , a2r ) ∈ (Aj ∩ Fl )2r : as = as .
s=1 s=r+1

Take Cr = 1∗2r
[− 12 , 21 ]
(0).
To prove (2.13.1) now, it remains to establish the lower bound
2r
Tj2r

Sl Nl
(2.13.2) Mj,l,r ≥ l+1
.
Nj Tl r Sl

Define ( r )
X
(Aj ∩ Fl )⊕r = as : a1 , . . . , ar ∈ Aj ∩ Fl .
s=1

We have
X
Mj,l,r = (g(b))2 ,
b∈(Aj ∩Fl )⊕r

where ( )
r
X
g(b) = (a1 , . . . , ar ) ∈ (Aj ∩ Fl )r : as = b
s=1

So by the Cauchy-Schwarz inequality

 2
X
(2.13.3)  g(b) ≤ |(Aj ∩ Fl )⊕r |Mj,l,r .
b∈(Aj ∩Fl )⊕r

40
Clearly
 r
X
r Sl
(2.13.4) g(b) = |(Aj ∩ Fl ) | = Tj
Tl
b∈(Aj ∩Fl )⊕r

To estimate |(Aj ∩ Fl )⊕r |, note each a ∈ Aj ∩ Fl has a digit representation of the

form
l j
X a(k) X a(k)
a= + ,
k=1
Nk k=l+1
Nk

where a(k) ∈ {1, . . . , s} for k = 1, . . . , l and a(k) ∈ [n] for k = l + 1, . . . , j. Therefore

each a ∈ Aj ∩ Fl is of the form

l
X a(k) a∗
a= + ,
k=1
Nk Nl

where a(k) ∈ {1, . . . , s} for k = 1, . . . , l and a∗ ∈ [Nj /Nl ]. Therefore each b ∈

(Aj ∩ Fl )⊕r is of the form

l
X b(k) b∗
b= +
k=1
Nk Nl

where b(k) ∈ {r, . . . , rs} for k = 1, . . . , l and b∗ ∈ [rNj /Nl ]. Hence

rNj Nj
(2.13.5) |(Aj ∩ Fl )⊕r | ≤ (rs)l = rl+1 Sl .
Nl Nl

Now (2.13.2) follows by combining (2.13.3), (2.13.4), and (2.13.5).

41
2.14 The Divergence of kfc
l µkp /kfl kL2 (µ)

In this section, we show that

kfcl µkp
(2.14.1) lim = ∞ ∀p ≤ p∗ .
l→∞ kfl kL2 (µ)

This is (2.0.4) in Theorem 2.0.3.

Let l ∈ N0 and 1 ≤ p ≤ p∗ be given. Recall that r ∈ N was chosen so that
p∗ ≤ 2r. Since

Sl
|fc
l µ(ξ)| ≤ µ(Fl ) = ∀ξ ∈ R,
Tl

we have
Z  2r−p
2r 2r−p c Sl
kfc
l µk2r = |fc
l µ(ξ)| |fl µ(ξ)|p dξ ≤ kfc p
l µkp .
R Tl

Therefore
 p−2r
p Sl 2r
kfc
l µkp ≥ kfc
l µk2r .
Tl

Since

Sl
kfl k2L2 (µ) = µ(Fl ) = ,
Tl

it follows that
p (p/2)−2r
kfcl µkp

Sl 2r
≥ kfc
l µk2r .
kfl kpL2 (µ) Tl

42
By combining (2.7.6) and (2.13.1), we have
 2r
2r Sl Nl
kfc
l µk2r ≥ Cr l+1
.
Tl r S l

Therefore p/2
p
kfcl µkp

Sl Nl
≥ Cr
kfl kpL2 (µ) Tl l+1
r Sl

Recall that s = nα0 −β0 /2 , t = nα0 , Sl = sl , Tl = tl , and Nl = nl . Note also that

ln r
rl = nl ln n . Therefore we have
p
kfcl µkp 1 1 ln r
p ≥ Cr nl(− 4 pβ0 +1−α0 + 2 β0 − ln n ) .
kfl kL2 (µ)

Recall now that n was defined so that

2(2 − 2α0 + β0 ) 4 ln r
p∗ < − .
β0 β0 ln n

Using that p ≤ p∗ and rearranging, we get

1 1 ln r
− pβ0 + 1 − α0 + β0 − > 0.
4 2 ln n

Therefore
p
kfcl µkp
lim p = ∞.
l→∞ kfl k 2
L (µ)

This proves (2.14.1) and completes the proof of Theorem 2.0.3.

43
Chapter 3

Explicit Salem Sets and

Applications to Metrical
Diophantine Approximation

In this chapter, we generalize theorems of Kaufman [29] and Bluhm [5] to construct
new explicit examples of Salem sets in R. We also prove a multi-dimensional analog
of our result and obtain new explicit sets in Rd with large Fourier dimension. We
give applications of our results to metrical Diophantine approximation.
In this chapter only, for x ∈ Rd ,

Xd
|x| = max |xi |, |x|2 = ( |xi |2 )1/2 .
1≤i≤d
i=1

Let Q be an infinite subset of Z, let Ψ : Z → [0, ∞), and let θ ∈ R. Define

E(Q, Ψ, θ) to be the set of all x ∈ R such that

kqx − θk ≤ Ψ(q) for infinitely many q ∈ Q.

Recall that, for x ∈ R, kxk = mink∈Z |x − k| is the distance from x to the nearest in-
teger. We will always assume Ψ is bounded. Since kxk ≤ 1/2 for all x ∈ R, assuming

44
Ψ is bounded results in no loss of generality. We will also always assume Ψ(0) = 1.
This assumption is imposed only to avoid tedious notation. Since redefining Ψ at
finitely many points does not change the set E(Q, Ψ, θ), assuming Ψ(0) = 1 results
in no loss of generality.
For M > 0, define

Q(M ) = {q ∈ Q : M/2 < |q| ≤ M } ,

(M ) = min Ψ(q).
q∈Q(M )

If Q(M ) is empty, (M ) is undefined.

The main result of this chapter is the following theorem.

Theorem 3.0.1. Suppose there is a number a ≥ 0, an increasing function h :

(0, ∞) → (0, ∞), and an unbounded set M ⊆ (0, ∞) such that

(3.0.1) |Q(M )|(M )a h(M ) ≥ M a ∀M ∈ M.

Then there is a Borel probability measure µ supported on E(Q, Ψ, θ) such that

 
−a ln |ξ|
(3.0.2) |b
µ(ξ)| . |ξ| exp h(4|ξ|) ∀ξ ∈ R, |ξ| > e.
ln ln |ξ|

Note the hypotheses M ⊆ (0, ∞) and (3.0.1) imply Q(M ) is non-empty and (M )
is defined for all M ∈ M.
We also have a higher-dimensional version of Theorem 3.0.1.
Let m, n ∈ N, let Q be an infinite subset of Zn , let Ψ : Zn → [0, ∞), and let
θ ∈ Rm . Define E(m, n, Q, Ψ, θ) to be the set of all points

(x11 , . . . , x1n , . . . , xm1 , . . . , xmn ) ∈ Rmn

such that n
X
max k qj xij − θi k ≤ Ψ(q) for infinitely many q ∈ Q.
1≤i≤m
j=1

45
Clearly E(1, 1, Q, Ψ, θ) = E(Q, Ψ, θ). As above, we will always assume Ψ is bounded
and Ψ(0) = 1, and these assumptions result in no loss of generality.
For M > 0, define

Q(M ) = {q ∈ Q : M/2 < |qj | ≤ M ∀1 ≤ j ≤ n} ,

(M ) = min Ψ(q).
q∈Q(M )

If Q(M ) is empty, (M ) is undefined.

Theorem 3.0.2. Suppose there is a number a ≥ 0, an increasing function h :

(0, ∞) → (0, ∞), and an unbounded set M ⊆ (0, ∞) such that

(3.0.3) |Q(M )|(M )a h(M ) ≥ M a ∀M ∈ M.

Then there is a Borel probability measure µ supported on E(m, n, Q, Ψ, θ) such that

 
−a ln |ξ|
(3.0.4) |b
µ(ξ)| . |ξ| exp h(4|ξ|) ∀ξ ∈ Rmn , |ξ| > e.
ln ln |ξ|

Note the hypotheses M ⊆ (0, ∞) and (3.0.3) imply Q(M ) is non-empty and (M )
is defined for all M ∈ M.
We give some remarks on the proofs of Theorems 3.0.1 and 3.0.2 in Section 3.1.
Sections 3.2 and 3.3 discuss motivations for Theorems 3.0.1 and 3.0.2. Section 3.4
contains applications of Theorems 3.0.1 and 3.0.2 to metrical Diophantine approxi-
mation. The combined proof of Theorems 3.0.1 and 3.0.2 constitutes Sections 3.5,
3.6, and 3.7. Section 3.8 contains the proof of Lemma 3.3.1.

3.1 Remarks on the Proof of Theorems 3.0.1 and

3.0.2
Since Theorem 3.0.2 is a generalization of Theorem 3.0.1, we give a single unified
proof. The proof follows the methods of Kaufman [29] and Bluhm [5, 7]. We also

46
found Wolff’s variant of Kaufman’s proof in [49] instructive. Kaufman [29] and
Bluhm [7] considered E(Z, Ψτ , 0), where Ψτ (q) = |q|−τ and τ > 1. In [5], Bluhm
considered E(Z, Ψ, 0) with Ψ(q) = ψ(|q|) and ψ : N → (0, ∞) decreasing. We needed
relatively straightforward modifications in our proof to handle the mn-dimensional
setting, more general functions Ψ, and the case θ 6= 0. A more substantial modifica-
tion was needed to handle general sets Q 6= Z. We now explain this modification.
It is fairly easy to see that the proofs of Bluhm, Kaufman, and Wolff can be
adapted to work for any set Q containing the set P of prime numbers. Their proofs
use two properties of the primes. The first property is about their density:

M
(3.1.1) |P ∩ (M/2, M ]| & ∀M ∈ N.
ln M

The second property is about their arithmetic structure:

ln n
(3.1.2) |{p ∈ P : p divides n, p ≥ M }| ≤ ∀n, M ∈ N, 2 ≤ M ≤ n.
ln M

In our proof, we use the divisor bound of Wigert [48] (see Section 3.5) as a substitute
for (3.1.2). This allows us to handle more general sets Q that may not have any
special arithmetic structure. The cost is that we have the factor exp(ln |ξ|/ ln ln |ξ|)
rather than ln |ξ| in Theorems 3.0.1 and 3.0.2. The analog in our proof of the density
property (3.1.1) is essentially the assumption (3.0.1) (though (3.0.1) is a combined
assumption about the density of Q and the size of Ψ).

3.2 Motivation: Explicit Salem Sets

The first motivation for our main result is the construction of explicit Salem sets and
explicit sets with non-zero Fourier dimension. We start by recalling some definitions
and basic facts mentioned in Chapter 1.
Let A ⊆ Rd . For α ≥ 0, the α-dimensional Hausdorff content of A is
X
H α (A) = inf (diam(B))α ,
B
B∈B

47
 S
where the infimum is over all countable collections B of balls such that A ⊆ B∈B B.
The Hausdorff dimension of A, denoted dimH (A), is the supremum all of α ∈ [0, d]
such that H α (A) > 0. The Fourier dimension of A, denoted dimF (A), is the supre-
mum of all β ∈ [0, d] such that

µ(ξ)| . (1 + |ξ|)−β/2
|b ∀ξ ∈ Rd .

for some probability measure µ on Rd with supp(µ) ⊆ A. If A is a Borel set,

Frostman’s lemma implies

(3.2.1) dimF (A) ≤ dimH (A)

(cf. [34, Chapter 12], [49, Chapter 8]).

A set A ⊆ Rd with dimF (A) = dimH (A) is called a Salem set. Every Borel set in
Rd of Hausdorff dimension 0 is a Salem set, every set in Rd that contains a ball is a
Salem set of dimension d, and every sphere in Rd is a Salem set of dimension d − 1.
Salem [41] proved the existence of Salem sets in R of arbitrary dimension α ∈
(0, 1) using a random Cantor-type construction. Kahane [27] showed that for every
α ∈ (0, d) there is a Salem set in Rd of dimension α by considering the images of
compact subsets of R under certain stochastic processes (cf. Chapters 17 and 18 of
[28]). Moreover, Kahane [28, Chapter 18] showed that the image of any compact
subset of Rd under fractional Brownian motion is almost surely a Salem set, indi-
cating that Salem sets are ubiquitous amongst random sets. Recently, other random
constructions of Salem sets have been given by Bluhm [6] and Laba and Pramanik
[31]. These random constructions do not produce explicit examples of Salem sets.
Kaufman [29] was the first to find an explicit Salem set of dimension α ∈ /
{0, d − 1, d}. The set Kaufman proved to be Salem is E(Z, Ψτ , 0), where Ψτ (q) =
|q|−τ and τ > 1. If τ ≤ 1, then E(Z, Ψτ , 0) = R by Dirichlet’s approximation theo-
rem, and so E(Z, Ψτ , 0) is a Salem set of dimension 1. We are concerned with τ > 1.

48
An easy and well-known argument (which we give in Section 3.8) gives

2
dimH E(Z, Ψτ , 0) ≤ .
1+τ

Since E(Z, Ψτ , 0) is a Borel set, (3.2.1) implies

dimH E(Z, Ψτ , 0) ≥ dimF (Z, Ψτ , 0).

Kaufman showed that for every τ > 1 there is a probability measure µ with support
contained in E(Z, Ψτ , 0) ∩ [0, 1] such that

µ(ξ)| . |ξ|−1/(1+τ ) ln |ξ| ∀ξ ∈ R, |ξ| ≥ 2.

|b

which implies
2
dimF E(Z, Ψτ , 0) ≥ ,
1+τ
and hence that E(Z, Ψτ , 0) is a Salem set. See [7] for a detailed variation of Kaufman’s
argument. In his thesis, Bluhm [5] showed that E(Z, Ψ, 0) is Salem for any Ψ with
Ψ(q) = ψ(|q|) and ψ : N → (0, ∞) decreasing. Technically, the results of Bluhm and
Kaufman are for E(N, Ψ, 0), not E(Z, Ψ, 0), but it is easy to adapt their proofs to
E(Z, Ψ, 0).
If A ⊆ R is a set of Fourier dimension α ∈ [0, 1], then it is easy to see the product
set Ad ⊆ Rd has Fourier dimension at least α by considering product measures.
A theorem of Gatesoupe [19] implies that if A ⊆ [0, 1] has Fourier dimension α ∈

[0, 1], then x ∈ Rd : |x|2 ∈ A has Fourier dimension at least d − 1 + α. Moreover,
Gatesoupe’s theorem implies that if A ⊆ [0, 1] is a Salem set of dimension α ∈ [0, 1],

then x ∈ Rd : |x|2 ∈ A is a Salem set in Rd of dimension d − 1 + α. Combining
Gatesoupe’s and Kaufman’s results yields explicit examples of Salem sets in Rd of
dimension α for every α ∈ (d − 1, d). Of course, the unit sphere in Rd and Rd itself
are explicit Salem sets with dimensions d − 1 and d, respectively. Explicit examples
of sets (Salem or otherwise) in Rd with Fourier dimension α ∈ (1, d−1) are unknown.
Theorems 3.0.1 and 3.0.2 generalize the theorems of Kaufman [29] and Bluhm

49
[5]. Theorem 3.0.1 gives many new explicit Salem sets in R. Some particular Salem
sets produced by Theorem 3.0.1 are discussed in Section 3.4. Theorem 3.0.2 gives
new examples of explicit sets in Rd with Fourier dimension α > 1.

3.3 Motivation: Metrical Diophantine Approxi-

mation
The second motivation for our main result comes from metrical Diophantine approxi-
mation, where there is considerable interest in the Hausdorff dimension of E(Q, Ψ, θ).
For τ ∈ R, define Ψτ : Z → [0, ∞) by Ψτ (q) = |q|−τ . The classical Jarnı́k-
Besicovitch theorem [26], [3] is that
 
2
dimH E(Z, Ψτ , 0) = min ,1 .
1+τ

In the setting of restricted Diophantine approximation, where Q is not necessarily

equal to Z, Borosh and Fraenkel [8] showed that
 
1 + ν(Q)
dimH E(Q, Ψτ , 0) = min ,1 ,
1+τ

where ( )
X
ν(Q) = inf ν≥0: |q|−ν < ∞ .
q∈Q
q6=0

Eggleston [15] previously obtained this result for certain sets Q with ν(Q) = 0 or
ν(Q) = 1.
There are also several results for more general functions Ψ. For Ψ of the form
Ψ(q) = ψ(|q|) with ψ : N → (0, ∞) decreasing, Dodson [14] showed that
 
2
dimH E(Z, Ψ, 0) = min ,1 ,
1+λ

50
where
− ln ψ(M )
λ = lim inf .
M →∞ ln M
(HS)
Hinokuma and Shiga [24] considered the non-monotone function Ψτ (q) = |sin q||q|−τ
and proved that  
(HS) 2
dimH E(Z, Ψτ , 0) = min ,1 .
1+τ
Dickinson [13] considered restricted Diophantine approximation with a function Ψ
satisfying Ψ(q) = ψ(|q|) with ψ : N → (0, ∞) and

− ln ψ(M ) − ln ψ(M )
λ = lim inf = lim sup .
M →∞ ln M M →∞ ln M

Dickinson deduced from the result of Borosh and Fraenkel above that
 
1 + ν(Q)
dimH E(Q, Ψ, 0) = min ,1 .
1+λ

Rynne [40] proved a very general result that implies all of those above. Suppose only
that Ψ : Z → [0, ∞) is positive for all q ∈ Q. Let
(  η )
X Ψ(q)
η(Q, Ψ) = inf η≥0: |q| <∞ .
q∈Q
|q|
q6=0

Rynne showed that

dimH E(Q, Ψ, 0) = min {η(Q, Ψ), 1} .

The main result in the case of inhomogeneous Diophantine approximation (i.e,

the case where θ is non-zero) is due to Levesley [32]. Levesley showed that if Ψ(q) =
ψ(|q|) with ψ : N → (0, ∞) decreasing, and if

− ln ψ(M )
λ = lim inf ,
M →∞ ln M

51
then  
2
dimH E(Z, Ψ, θ) = min ,1 .
1+λ
By an adaptation of Dickinson’s argument from [13], the assumption that ψ is de-
creasing can be replaced by the assumption that

− ln ψ(M ) − ln ψ(M )
λ = lim inf = lim sup .
M →∞ ln M M →∞ ln M

The main content of the formulas above is the lower bounds they give on the Haus-
dorff dimension of E(Q, Ψ, θ). The ≤-half of all the formulas for dimH E(Q, Ψ, θ)
above are implied by the following lemma whose proof is well-known and straight-
forward. For completeness, we give the proof in Section 3.8.

Lemma 3.3.1.

dimH E(Q, Ψ, θ) ≤ min {η(Q, Ψ), 1} .

Because of (3.2.1), the Fourier analytic method of Theorem 3.0.1 stands as an

alternative to the usual methods of proving lower bounds on the Hausdorff dimension
of E(Q, Ψ, θ). In fact, Theorem 3.0.1 implies or implies special cases of all the results
for dimH E(Q, Ψ, θ) above (details are given in Section 3.4). Moreover, Theorem 3.0.1
allows us to calculate the Hausdorff dimension of E(Q, Ψ, θ) in cases that (as far as
we know) have not been treated previously in the literature, such as the case where
θ 6= 0 and Q 6= N, Z.
One particular advantage of the Fourier analytic method of Theorem 3.0.1 is the
ease with which it handles the inhomogeneous case θ 6= 0. In the proof of Theorem
3.0.1 it is trivial to accommodate θ 6= 0, while Levelsey’s proof of his result for θ 6= 0
is a non-trivial extension of Dodson’s proof for θ = 0.
There are analogs of the formulas above for dimH E(m, n, Q, Ψ, θ). For example,
Rynne [40] proved

dimH E(m, n, Q, Ψ, 0) = min {m(n − 1) + η(Q, Ψ), mn}

52
when Ψ : Zn → [0, ∞) is positive for all q ∈ Q and
(  η )
X Ψ(q)
η(Q, Ψ) = inf η≥0: |q|m <∞ .
q∈Q
|q|
q6=0

Theorem 3.0.2 (via (3.2.1)) provides a lower bound on dimH E(m, n, Q, Ψ, θ), but
it does not reach the true value of dimH E(m, n, Q, Ψ, θ) for any known case with
mn > 1.

3.4 Applications
In this section we will present several consequences of Theorem 3.0.1 that give new
families of explicit Salem sets and imply formulas for dimH E(Q, Ψ, θ) discussed in
Section 3.3. We will also present a typical consequence of Theorem 3.0.2 that yields
explicit sets in Rmn (mn > 1) with Fourier dimension larger than 1.

Theorem 3.4.1. Assume Ψ is of the form Ψ(q) = ψ(|q|) with ψ : N → (0, ∞) a

decreasing function. Assume there is an increasing function h : (0, ∞) → (0, ∞)
such that

(3.4.1) |Q(M )| ≥ M/h(M ) ∀M ∈ N

and
ln h(x)
lim =0
ln x
x→∞

Then E(Q, Ψ, θ) is a Salem set of dimension min {2/(1 + λ), 1}, where

− ln(ψ(M ))
(3.4.2) λ = lim inf .
M →∞ ln M

Proof. Since Ψ is bounded, λ ≥ 0.

0
Let λ0 < λ and δ > 0. By (3.4.2), ψ(M ) < M −λ for all large M ∈ N. It follows

53
that  (2+δ)/(1+λ0 )
X Ψ(q) X
|q| . |q|−(1+δ) < ∞.
q∈Q
|q| q∈Q
q6=0 q6=0

Therefore, by applying Lemma 3.3.1 and then letting λ0 → λ and δ → 0, we have

 
2
dimH E(Q, Ψ, θ) ≤ min ,1 .
1+λ

If λ = ∞, this argument shows dimH E(Q, Ψ, θ) = dimF E(Q, Ψ, θ) = 0.

Assume 0 ≤ λ < ∞. Let λ0 > λ. By (3.4.2), there is an infinite set M ⊆ N
0
such that ψ(M ) > M −λ for all M ∈ M. Since ψ is decreasing, it follows that
0
(M ) > M −λ for all M ∈ M. Combining this with (3.4.1), we see that (3.0.1) holds
with a = 1/(1 + λ0 ). Since λ0 > λ is arbitrary, Theorem 3.0.1 gives
 
2
dimF E(Q, Ψ, θ) ≥ min ,1 .
1+λ

Theorem 3.4.1 implies the result of Dodson [14] for dimH E(Z, Ψ, 0) discussed
in Section 3.3. Theorem 3.4.1 also implies the formula for dimH E(Z, Ψ, θ) due to
Levesley [32] mentioned in Section 3.3.

Theorem 3.4.2. Assume Ψ is of the form Ψ(q) = ψ(|q|) with ψ : N → (0, ∞).
Assume

− ln ψ(M ) − ln ψ(M )
(3.4.3) λ = lim inf = lim sup
M →∞ ln M M →∞ ln M

and
X
(3.4.4) |q|−1 = ∞.
q∈Q
q6=0

Then E(Q, Ψ, θ) is a Salem set of dimension min {2/(1 + λ), 1}.

54
Proof. Since Ψ is bounded, λ ≥ 0.
The same argument as in the proof of Theorem 3.4.1 shows
 
2
dimH E(Q, Ψ, θ) ≤ min ,1 .
1+λ

If λ = ∞, the argument shows dimH E(Q, Ψ, θ) = dimF E(Q, Ψ, θ) = 0.

Assume 0 ≤ λ < ∞. Seeking a contradiction, suppose

|Q(M )| < M/ ln2 (M )

for all large M ∈ N. Then

X ∞
X X ∞
X X
−1 −1
|q| = |q| . 2−k 1
q∈Q k=0 q∈Q(2k ) k=0 q∈Q(2k )
q6=0
∞ ∞
X 1 1 X 1
. = < ∞,
k=0
ln2 (2k ) ln2 (2) k=0 k 2

which contradicts (3.4.4). So there is an infinite set M ⊆ N such that

|Q(M )| ≥ M/ ln2 (M ) ∀M ∈ M.

0
Let λ0 > λ be given. By (3.4.3), ψ(M ) > M −λ for all large M ∈ N. Therefore
0
(M ) > M −λ for all large M ∈ N. After removing finitely elements of M, we have

0
(M ) > M −λ ∀M ∈ M.

Then (3.0.1) holds with a = 1/(1 + λ0 ) and h(x) = ln2 (x). Since λ0 > λ is arbitrary,
Theorem 3.0.1 gives
 
2
dimF E(Q, Ψ, θ) ≥ min ,1 .
1+λ

55
 Theorem 3.4.2 implies the result of Dickinson [13] for dimH E(Q, Ψ, 0) discussed
in Section 3.3 in the case ν(Q) = 1. Consequently, it also implies the results of
Borosh and Fraenkel [8] and Eggleston [15] in the case ν(Q) = 1. Theorem 3.4.2
implies the variation of the result of Levesley [32] for dimH E(Z, Ψ, θ) mentioned in
Section 3.3 that uses Dickinson’s argument from [13].
As far as we know, Theorems 3.4.1 and 3.4.2 represent the first calculation of
dimH E(Q, Ψ, θ) in the case where θ 6= 0 and Q 6= N, Z.
(HS) (HS)
Theorem 3.4.3. Suppose Ψτ : Z → (0, ∞) is defined by Ψτ (q) = | sin q||q|−τ .
(HS)
Then E(Z, Ψτ , θ) is a Salem set of dimension of min {2/(1 + τ ), 1}.
Proof. For every  > 0,
 2/(1+τ )+
X Ψ(q) X
|q| = |q|−1−(1+τ ) | sin q|2/(1+τ )+ < ∞.
q∈Z
|q| q∈Z
q6=0 q6=0

So, by Lemma 3.3.1,

 
2
dimH E(Z, Ψ(HS)
τ , θ) ≤ min ,1 .
1+τ

Let Q = {q ∈ N : | sin q| ≥ 1/2}. Clearly

1
min Ψ(HS)
τ (q) ≥ M −τ ∀M ∈ N.
q∈Q(M ) 2

It is also easy to see that

M
|Q(M )| ≥
4π
for all large M (for instance, by noting Q contains the nearest integer(s) to (2k+1)π/2
for every k ∈ N). It follows that (3.0.1) holds with a = 1/(1 + τ ) and h(x) = 4π.
(HS) (HS)
Therefore Theorem 3.0.1 and the fact E(Q, Ψτ , θ) ⊆ E(Z, Ψτ , θ) implies
 
2
dimF E(Z, Ψ(HS)
τ , θ) ≥ min ,1 .
1+τ

56
 (HS)
Theorem 3.4.3 implies the formula for dimH E(Z, Ψτ , 0) due to Hinokuma and
Shiga [24] mentioned in Section 3.3.
Finally, we give a typical consequence of Theorem 3.0.2 that yields explicit sets
in Rmn with Fourier dimension larger than 1. The Hausdorff dimension of the sets
is also determined for comparison.

Theorem 3.4.4. Assume Ψ : Zn → [0, ∞) is of the form Ψ(q) = ψ(|q|) with ψ :

N → (0, ∞) and

− ln ψ(M ) − ln ψ(M )
(3.4.5) λ = lim inf = lim sup .
M →∞ ln M M →∞ ln M

Assume h : (0, ∞) → (0, ∞) is an increasing function such that

ln h(x)
(3.4.6) lim = 0.
x→∞ ln x

Assume there is an unbounded set M ⊆ (0, ∞) such that

(3.4.7) |Q(M )|h(M ) ≥ M n ∀M ∈ M.

Then  
m+n
dimH E(m, n, Q, Ψ, 0) = min m(n − 1) + , mn
1+λ
and  
2n
dimF E(m, n, Q, Ψ, 0) ≥ min , mn .
1+λ
Proof. Since Ψ is bounded, λ ≥ 0.
0
Let λ0 < λ and δ > 0. By (3.4.5), ψ(M ) < M −λ for all large M ∈ N. It follows
that  (m+n+δ)/(1+λ0 ) X
X Ψ(q)
|q|m
. |q|−(n+δ) < ∞.
q∈Q
|q| q∈Q
q6=0 q6=0

57
Since λ0 < λ and δ > 0 are arbitrary, we have
 η
X
m Ψ(q) m+n
|q| < ∞ ∀η > .
q∈Q
|q| 1+λ
q6=0

0
Let λ0 > λ and δ > 0. By (3.4.5), ψ(M ) > M −λ for all large M ∈ N. Choose a
sequence (Mk )k∈N of numbers in M such that Mk ≥ 2Mk−1 . Then
 (m+n−δ)/(1+λ0 )
X Ψ(q) X X X
m
|q| & |q|−(n−δ) & |q|−(n−δ)
q∈Q
|q| q∈Q k∈N q∈Q(Mk )
q6=0 q6=0
−(n−δ)
X X
& Mk |Q(Mk )| & Mkδ (h(Mk ))−1
k∈N k∈N
X δ/2
& Mk =∞
k∈N

Since λ0 > λ and δ > 0 are arbitrary, we have

 η
X
m Ψ(q) m+n
|q| = ∞ ∀η < .
q∈Q
|q| 1+λ
q6=0

By the result of Rynne [40] for dimH E(m, n, Q, Ψ, 0) discussed in Section 3.3, we
have  
m+n
dimH E(m, n, Q, Ψ, 0) = min m(n − 1) + , mn .
1+λ
If λ = ∞, this argument shows dimH E(m, n, Q, Ψ, 0) = m(n − 1).
0
Assume 0 ≤ λ < ∞. Let λ0 > λ. By (3.4.5), ψ(M ) > M −λ for all large M ∈ N.
0
Therefore (M ) > M −λ for all large M ∈ N. After removing finitely many elements
0
of M, we have (M ) > M −λ for all M ∈ M. Combining this with (3.4.7), we see
that (3.0.3) holds with a = n/(1 + λ0 ). As λ0 > λ is arbitrary, Theorem 3.0.2 implies
 
2n
dimF E(m, n, Q, Ψ, 0) ≥ min , mn .
1+λ

58
3.5 Proof of Theorems 3.0.1 and 3.0.2: The Key
Function
We start with two basic definitions.
Suppose f : Rd → C. If f ∈ L1 (Rd ), the Fourier transform of f is defined by
Z
f (ξ) =
b f (x)e−2πix·ξ dx ∀ξ ∈ Rd .
R
R
If f ∈ L1 ([0, 1]d ) (i.e., [0,1]d |f (x)|dx < ∞) and f is periodic for the lattice Zd (i.e.,
f (x) = f (x + k) for all x ∈ Rd and k ∈ Zd ), the Fourier transform of f is defined by
Z
f (ξ) =
b f (x)e−2πix·ξ dx ∀ξ ∈ Rd .
[0,1]d

There is no ambiguity with these definitions; if f ∈ L1 (Rd ) and f is periodic for the
lattice Zd , then fb = 0 using either definition.
Now we begin the construction.
Let K be a positive integer with K > mn + a. Let φ : Rm → R be a non-negative
R
C K function with supp(φ) ⊆ [−1, 1]m and Rm φ = 1. Then there is a C1 > 0 such
that

(3.5.1) |φ(ξ)|
b ≤ C1 (1 + |ξ|)−K ∀ξ ∈ Rm .

For  > 0 and x ∈ Rm , let φ (x) = −m φ(−1 x), and

X
Φ (x) = φ (x − k).
k∈Zm

Note Φ is C K , is periodic for the lattice Zm , and satisfies

c (k) = φb (k) = φ(k)

Φ b ∀k ∈ Zm .

Therefore
X
Φ (x) = φ(k)e
b 2πikx
∀x ∈ Rm
k∈Zm

59
with uniform convergence.
For q ∈ Zn , θ ∈ Rm , and x = (x11 , . . . , x1n , . . . , xm1 , . . . , xmn ) ∈ Rmn , define xq
by identifying x with the m × n matrix whose ij-entry is xij , and define
X
Φq,θ (x) = Φ (xq − θ) = φ(k)e
b 2πik·(xq−θ)
.
k∈Zm

Note Φq,θ is C K and is periodic for the lattice Zmn .

Assume from now on that qj 6= 0 for all 1 ≤ j ≤ n.
If mn = 1, then for ` ∈ Z we have
XZ
 2πik(xq−θ) −2πi`x
Φ
d
q,θ (`) = φ(k)e
b e dx
k∈Z [0,1]
X Z
−2πikθ b
= e φ(k) e2πix(kq−`) dx
k∈Z [0,1]
( −1 `θ
e−2πiq b −1 `) if q −1 ` ∈ Z,
φ(q
=
0 otherwise.

In general, for ` = (`11 , . . . , `1n , . . . , `m1 , . . . , `mn ) ∈ Zmn we have

XZ
 2πik·(xq−θ) −2πi`·x
(3.5.2) Φ
d
q,θ (`) = φ(k)e
b e dx
k∈Zm [0,1]mn
X Z
−2πik·θ b
= e φ(k) e2πi(k·(xq−θ)−`·x) dx
k∈Zm [0,1]mn

X m Y
Y n Z
−2πik·θ b
= e φ(k) e2πixij (ki qj −`ij ) dxij
k∈Zm i=1 j=1 [0,1]
(
−2πiq1−1 `1 ·θ b −1 `1 ) if q ∈ Q(`),
e φ(q 1
e
=
0 otherwise,

where
e = q ∈ Zn : q −1 `1 = · · · = q −1 `n ∈ Zm

Q(`) 1 n

and `j = (`1j , . . . , `mj ) for all 1 ≤ j ≤ n.

60
 For M > 0, define

1 X (M )
FM (x) = Φq,θ (x) ∀x ∈ Rmn .
|Q(M )|
q∈Q(M )

This is the key function in the proof.

Note FM is C K and periodic for the lattice Zmn . Since supp(φ) ⊆ [−1, 1]m and
(M ) = min Ψ(q), we have
q∈Q(M )

n
X
mn
supp(FM ) ⊆ {x ∈ R : max k xij qj − θi k ≤ Ψ(q) for some q ∈ Q(M )}.
1≤i≤m
j=1

If q ∈ Q(M ) and 0 < |`| ≤ M/2, then q ∈

/ Q(`).
e So by (3.5.2) we have

(3.5.3) M (`) = 0 for 0 < |`| ≤ M/2.

Fc

As φ ≥ 0 and φ(0)
b = 1, (3.5.2) also implies

mn
(3.5.4) |Fc
M (`)| ≤ FM (0) = 1 ∀` ∈ Z
c .

Choose i0 ∈ {1, . . . , m} and j0 ∈ {1, . . . , n} such that |`| = |`j0 | = |`i0 j0 |. By

(3.5.1) and (3.5.2), for all ` ∈ Zmn we have

1 X
|Fc
M (`)| ≤ C1 |φ((M
b )q1−1 `1 )|
|Q(M )|
q∈Q(M )∩Q(`)
e

1 X
≤ C1 (1 + (M )|q1 |−1 |`1 |)−K
|Q(M )|
q∈Q(M )∩Q(`)
e

1 X
= C1 (1 + (M )|qj0 |−1 |`j0 |)−K
|Q(M )|
q∈Q(M )∩Q(`)
e

|Q(M ) ∩ Q(`)|
e
≤ C1 (1 + (M )M −1 |`|)−K .
|Q(M )|

61
To bound |Q(M ) ∩ Q(`)|,
e first note that for ` 6= 0 we have
 
n `i j `i j qj
e ⊆
Q(`) (q1 , . . . , qn ) ∈ Z : 0 0 ∈ Z, qj = 0 0 ∀1 ≤ j ≤ n .
qj0 `i0 j0

It follows that
|Q(M ) ∩ Q(`)|
e ≤ |Q(`)|
e ≤ 2τ (|`|),

where τ (v) denotes the number of positive integer divisors of the integer v. The
divisor bound of Wigert [48] (cf. [21, p. 262]) says
ln τ (v)
lim sup = ln 2.
v→∞ ln v/ ln ln v
For ζ > 0, define  
ζ ln x
fζ (x) = exp ∀x > e.
ln ln x
So for any ζ > ln 2 there is a vζ ∈ N such that

τ (v) ≤ fζ (v) ∀v ∈ N, v ≥ vζ .

Therefore

fζ (|`|)
(3.5.5) |Fc
M (`)| ≤ 2C1 (1 + (M )M −1 |`|)−K ∀` ∈ Zmn , |`| ≥ vζ .
|Q(M )|

3.6 Proof of Theorems 3.0.1 and 3.0.2: The Main

Lemma
Define (
1 if |ξ| ≤ e
g(ξ) =
|ξ|−a f1 (|ξ|)h(4|ξ|) if |ξ| > e

Lemma 3.6.1. For every δ > 0, M0 > 0, and χ ∈ CcK (Rmn ), there is an M∗ =
M∗ (δ, M0 , χ) ∈ M such that M∗ ≥ M0 and

|χF
[ M∗ (ξ) − χ
b(ξ)| ≤ δg(ξ) ∀ξ ∈ Rmn .

62
 The proof will show M∗ can be taken to be any sufficiently large element of M.

Proof. Since χ ∈ CcK (Rmn ), there is a C2 > 0 such that

(3.6.1) χ(ξ)| ≤ C2 (1 + |ξ|)−K

|b ∀ξ ∈ Rmn .

For every p > mn, we have

X
(3.6.2) sup (1 + |ξ − `|)−p < ∞.
ξ∈Rmn
`∈Zmn

Since FM is C K and periodic for the lattice Zmn , we have

X
2πi`·x
FM (x) = Fc
M (`)e ∀x ∈ Rmn
`∈Zmn

with uniform convergence. Since χ ∈ L1 (Rmn ), multiplying by χ and taking the

Fourier transform yields
X Z X
χFM (ξ) =
[ FM (`)
c χ(x)e2πi(`−ξ)·x dx = Fc χ(ξ − `)
M (`)b
`∈Zmn Rmn `∈Zmn

for all ξ ∈ Rmn . Then by (3.5.4) and (3.5.3) we have

X
(3.6.3) χF M (ξ) − χ
b(ξ) = b(ξ − `)Fc
χ M (`).
[
|`|≥M/2

Case 1: |ξ| ≤ M/4. If |`| ≥ M/2, then |ξ − `| ≥ M/4 ≥ |ξ|. Hence by (3.5.3),
(3.6.1), (3.6.2) and (3.6.3) we have
X
|χF
[ M (ξ) − χ
b(ξ)| ≤ C2 (1 + |ξ − `|)−K
|`|≥M/2

≤ C2 (1 + |ξ|)−a (1 + M/4)−(K−a−mn)/2
X
(1 + |ξ − `|)−mn−(K−a−mn)/2
|`|≥M/2

≤ δg(ξ)

63
for M sufficiently large.
Case 2: |ξ| ≥ M/4. Using (3.6.3), write
X X
χF M (ξ) − χ
b(ξ) = b(ξ − `)Fc
χ M (`) + b(ξ − `)Fc
χ M (`) = S1 + S2 .
[
|`|≥M/2 |`|≥M/2
|`|≤|ξ|/2 |`|>|ξ|/2

If |`| ≤ |ξ|/2, then |ξ − `| ≥ |ξ|/2 ≥ M/8. Hence by (3.5.3), (3.6.1), and (3.6.2) we
have
X
|S1 | ≤ C2 (1 + |ξ − `|)−K
|`|≥M/2
|`|≤|ξ|/2
X
≤ C2 (1 + |ξ|/2)−a (1 + M/8)−(K−a−mn)/2 (1 + |ξ − `|)−mn−(K−a−mn)/2
|`|≥M/2
|`|≤|ξ|/2
δ
≤ g(ξ)
2

for M sufficiently large.

Fix ln 2 < ζ < 1. By (3.0.1), (3.5.5), (3.6.1), (3.6.2) and because K > a we have

X fζ (|`|)
|S2 | ≤ 2C1 C2 (1 + (M )M −1 |`|)−K (1 + |ξ − `|)−K
|Q(M )|
|`|≥M/2
|`|>|ξ|/2
X ((M )M −1 )−a
≤ 2C1 C2 |`|−a fζ (|`|) (1 + |ξ − `|)−K
|Q(M )|
|`|≥M/2
|`|>|ξ|/2
X
≤ 2C1 C2 (|ξ|/2)−a fζ (|ξ|/2)h(M ) (1 + |ξ − `|)−K
|`|≥M/2
|`|>|ξ|/2
δ
≤ g(ξ)
2

for all sufficiently large M ∈ M.

64
3.7 Proof of Theorems 3.0.1 and 3.0.2: The Mea-
sure
R
Let χ0 ∈ C K (Rmn ) with Rmn χ0 (x)dx = 1, supp(χ0 ) = [−1, 1]mn , and χ0 (x) > 0 for
all |x| < 1. With the notation of Lemma 3.6.1, define

M1 = M∗ (2−1 , 1, χ0 ), Mk = M∗ (2−k−1 , 2Mk−1 , χ0 FM1 · · · FMk−1 ) ∀k ≥ 2.

For k ∈ N0 , define the measure µk by

dµk = χ0 FM1 · · · FMk dx.

By Lemma 3.6.1,

−k−1
(3.7.1) |c
µk (ξ) − µd
k−1 (ξ)| ≤ 2 g(ξ) ∀ξ ∈ Rmn , k ∈ N.

Since g(ξ) is bounded , (3.7.1) implies (c

µk )k∈N0 is a Cauchy sequence in the supremum
norm. Therefore, since each µ ck is a continuous function, lim µ ck is a continuous
k→∞
function. By (3.7.1) we have
∞ ∞
X X 1
(3.7.2) | lim µ
ck (ξ) − µb0 (ξ)| ≤ |c
µk (ξ) − µd
k−1 (ξ)| ≤ g(ξ) 2−k−1 = g(ξ)
k→∞
k=1 k=1
2

for all ξ ∈ Rmn . Since µb0 (0) = g(0) = 1, it follows from (3.7.2) that

1/2 ≤ | lim µ
ck (0)| ≤ 3/2.
k→∞

Therefore, by Lévy’s continuity theorem, (µk )k∈N0 converges weakly to a non-trivial

finite measure µ with µb = lim µ
ck and
k→∞

∞
\ ∞
\
supp(µ) = supp(µk ) = supp(χ0 ) ∩ supp(FMk ).
k=1 k=1

65
Because Mk ≥ 2Mk−1 and because (as we noted in Section 3.5)
n
X
mn
supp(FM ) ⊆ {x ∈ R : max k xij qj − θi k ≤ Ψ(q) for some q ∈ Q(M )},
1≤i≤m
j=1

we have
supp(µ) ⊆ E(m, n, Q, Ψ, θ).

Since χ0 ∈ CcK and K > a, we have |µb0 (ξ)| . (1 + |ξ|)−a for all ξ ∈ Rmn . Combining
this with (3.7.2) gives
µ(ξ)| . g(ξ) ∀ξ ∈ Rmn .
|b

By multiplying µ by a constant, we can make µ a probability measure. This completes

the proof of Theorems 3.0.1 and 3.0.2.

3.8 Proof of Lemma 3.3.1

Lemma 3.8.1 (Restatement of Lemma 3.3.1).

dimH E(Q, Ψ, θ) ≤ min {η(Q, Ψ), 1} ,

where (  η )
X Ψ(q)
η(Q, Ψ) = inf η≥0: |q| <∞ .
q∈Q
|q|
q6=0

Proof. Since dimH E(Q, Ψ, θ) ≤ dimH R ≤ 1, we only need to prove

dimH E(Q, Ψ, θ) ≤ η(Q, Ψ).

66
Note E(Q, Ψ, θ) is invariant under translation by integers. Therefore
[
dimH E(Q, Ψ, θ) = dimH E(Q, Ψ, θ) ∩ ([0, 1] + k)
k∈Z
[
= dimH (E(Q, Ψ, θ) − k) ∩ [0, 1]
k∈Z
[
= dimH dimH E(Q, Ψ, θ) ∩ [0, 1]
k∈Z

= dimH E(m, n, Q, Ψ, θ) ∩ [0, 1].

So it suffices to prove

dimH E(Q, Ψ, θ) ∩ [0, 1] ≤ η(Q, Ψ).

Therefore, according to the definition of Hausdorff dimension, it will suffice to show

that for all  > 0 and all η > η(Q, Ψ) there is a countable collection I of intervals
that covers E(Q, Ψ, θ) ∩ [0, 1] and satisfies
X
(diam(I))η < .
I∈I

Let  > 0 and η > η(Q, Ψ). Define C = |θ| + supΨ(q). Observe that
q∈Z

\ [ [
E(Q, Ψ, θ) ∩ [0, 1] = {x ∈ [0, 1] : |xq − θ − k| ≤ Ψ(q)} .
N ∈N q∈Q k∈Z
|q|≥N |k|≤C+|q|

Let N ∈ N. Then
[ [
E(Q, Ψ, θ) ∩ [0, 1] ⊆ {x ∈ [0, 1] : |xq − θ − k| ≤ Ψ(q)}
q∈Q k∈Z
|q|≥N |k|≤C+|q|
[ [
⊆ Iq,k ,
q∈Q k∈Z
|q|≥N |k|≤C+|q|

67
where
Iq,k = [(θ + k − Ψ(q))/|q|, (θ + k + Ψ(q))/|q|].

We have
 η
X X
η
X X
η Ψ(q)
(diam(Iq,k )) = 2
q∈Q k∈Z q∈Q k∈Z
|q|
|q|≥N |k|≤C+|q| |q|≥N |k|≤C+|q|
 η
X
η Ψ(q)
≤ (2(C + |q|) + 1)2
q∈Q
|q|
|q|≥N
 η
X Ψ(q)
. |q| .
q∈Q
|q|
|q|≥N

The last sum converges because η > η(Q, Ψ). So, by taking N sufficiently large, we
can make the sum less than .

68
Chapter 4

Conclusion

In this chapter, we summarize the contributions of this thesis and discuss directions
for future work.

4.1 Contributions
In this section, we summarize the contributions of this thesis.
The first major contribution of this thesis concerns the L2 Fourier restriction
theorem due to Mockenhaupt [37] and Mitsis [35] (with the endpoint due to Bak and
Seeger [1]). We state the theorem again here for convenience.

Theorem 4.1.1. Let µ be a compactly supported probability measure on Rd . Suppose

there are α, β ∈ (0, d) such that

(A) µ(B(x, r)) . rα ∀x ∈ Rd , r > 0,

(B) µ(ξ)| . (1 + |ξ|)−β/2
|b ∀ξ ∈ Rd .

Then for all p ≥ 2(2d − 2α + β)/β we have

(4.1.1) kfcµkp . kf kL2 (µ) ∀f ∈ L2 (µ).

Restriction theorems are some of the main objects of study in harmonic analysis.

69
They represent an important way to understand the nature of the Fourier transform
and the geometry of sets and measures. Restriction theorems are connected to many
other important problems in mathematical analysis, including the Kakeya conjecture,
the Bochner-Riesz conjecture, the estimation of solutions to the wave, Schroedinger,
and Helmholtz equations, and the local smoothing conjecture for partial differential
equations (cf. [45]).
Once a restriction theorem such as Theorem 4.1.1 has been established, it is
natural to ask whether it can be improved. In this thesis, we considered whether
we can enlarge the range of p in the theorem by replacing 2(2d − 2α + β)/β with a
smaller number. We considered this question only for a particular class of measures.
If we do not limit the class of measures considered, the well-known Knapp example
(cf. [49, Chapter 7]) shows the range of p cannot be enlarged. Whereas the classic
restriction theorems established before Theorem 4.1.1 applied only to smooth curved
surfaces (like spheres) in Rd , d ≥ 2, the novel aspect of Theorem 4.1.1 is that it
applies to measures supported on fractal sets and to measures on R. In this thesis,
we have considered the optimality of the range of p in Theorem 4.1.1 for measures
on R. In particular, we proved the following theorem.

Theorem 4.1.2. Given any 0 < β ≤ α < 1 and given any p∗ < 2(2 − 2α + β)/β
there is a compactly supported probability measure on R that satisfies (A) and (B)
but does not satisfy (4.1.1) for every p ≤ p∗ .

The theorem is important and interesting because there is no obvious analog of

the Knapp example (with its small, almost flat, spherical cap) available to prove the
optimality of the range of p in Theorem 4.1.1 when considering measures on the real
line.
The second major contribution of this thesis is the construction of new explicit
Salem sets in R and new explicit sets in Rd with large Fourier dimension.
There are many random constructions of Salem sets (cf. [41], [27], [28], [6], [31]).
Moreover, Kahane [28, Chapter 18] showed that the image of any compact set in
Rd under fractional Brownian motion is almost surely a Salem set, which illustrates
that Salem sets are ubiquitous as random sets. These random constructions do not

70
produce any explicit examples of Salem sets. Moreover, besides the classic examples
of Salem sets in Rd of dimensions 0, d − 1, and d, explicit examples of Salem sets are
very difficult to find. The first (and essentially only) construction of explicit Salem
sets with dimensions other than 0, d − 1, or d is due to Kaufman [29], who showed
that
x ∈ R : kqxk ≤ |q|−τ for infinitely many q ∈ Z


is a Salem set of dimension 2/(1 + τ ) when τ > 1. Bluhm [7] generalized this by
showing that replacing |q|−τ by ψ(|q|) for any decreasing function ψ : N → (0, ∞)
still gives a Salem set. In this thesis, we further generalized the results of Kaufman
and Bluhm. For Q ⊆ Z infinite, Ψ : Z → [0, ∞), and θ ∈ R, we proved a lower
bound on the Fourier dimension of

E(Q, Ψ, θ) = {x ∈ R : kqx − θk ≤ Ψ(q) for infinitely many q ∈ Q} .

Moreover, we showed that this set is Salem for many choices of Q and Ψ (regardless
of the value of θ). Explicit examples of sets in Rd with Fourier dimension larger than
1 are also hard to find. We gave in this thesis some new explicit examples of such
sets. Furthermore, we applied our results to compute the Hausdorff dimension of
E(Q, Ψ, θ) in new cases, which is of interest in metrical Diophantine approximation.

4.2 Directions for Future Work

In this section, we discuss directions for future work related to the content of this
thesis.
We first note that Chen [11] has obtained the following improved version of The-
orem 4.1.2 by modifying the argument of Laba and myself from [23].

Theorem 4.2.1. Given any 0 < β ≤ α < 1, there is a compactly supported proba-
bility measure on R that satisfies (A) and (B) but does not satisfy (4.1.1) for every
p ≤ 2(2d − 2α + β)/β.

There are several interesting ways Theorems 4.1.2 and 4.2.1 could possibly be

71
extended. We briefly discuss two of them.
First, one could try to obtain a version of Theorem 4.1.2 or 4.2.1 with β > α. Since
the Hausdorff dimension of a Borel set in R is the supremum of all α ∈ [0, 1] for which
the set supports some probability measure µ satisfying (A), it is tempting to think
that the supremum of the numbers α ∈ [0, 1] for which a given probability measure
µ satisfies (A) is the Hausdorff dimension of the support of µ. This is wrong. The
Hausdorff dimension of supp(µ) can be strictly larger than the supremum of α ∈ [0, 1]
for which µ satisfies (A). Moreover, since the Fourier dimension of Borel set in R is
the supremum of β ∈ [0, 1] for which the set supports some probability measure µ
satisfying (B), it is tempting to think that the supremum of the numbers β ∈ [0, 1]
for which a given probability measure µ satisfies (B) is the Fourier dimension of
the support of µ. This is also wrong. The Fourier dimension of supp(µ) can be
strictly larger than the supremum of β ∈ [0, d] for which µ satisfies (B). Finally, since
the Fourier dimension of a Borel set is always less than or equal to the Hausdorff
dimension of the set, it is tempting to think a probability measure µ on R that
satisfies (B) for some β must satisfy (A) for some α ≥ β. This, again, is wrong.
A probability measure µ may satisfy (B) for some β and fail to satisfy (A) for any
α ≥ β. In other words, there are probability measures µ such that the largest α
and β in (0, d) for which (A) and (B) hold obey β > α. Neither Theorem 4.1.2 nor
Theorem 4.2.1 give the sharpness of the range of p in Theorem 4.1.1 for the class of
such measures. It is likely a substantial new idea is needed to obtain such a sharpness
result as it appears the current method cannot be modified to handle β > α.
Second, it would be interesting to obtain a version of Theorem 4.1.2 or 4.2.1 for
measures on Rd and 0 < β ≤ α < d. The current method of proof does not seem to
generalize easily to this situation. However, it seems likely that such a theorem can
be obtained for α and β in a restricted range.
Concerning the results in this thesis on the construction of explicit Salem sets,
we pose three questions that are interesting for future research.
What is the Fourier dimension of E(Q, Ψτ , 0) when ν(Q) < 1? For example,
consider Q as the set of squares (so that ν(Q) = 1/2) or the set of powers of 2 (so that
ν(Q) = 0). Theorem 3.0.1 says nothing when ν(Q) < 1. One natural conjecture is

72
that E(Q, Ψτ , 0) is a Salem set, meaning its Fourier dimension is min{(1+ν(Q))/(1+
τ ), 1}.
What is the Fourier dimension of E(m, n, Z, Ψτ , 0)? Theorem 3.0.2 implies the
Fourier dimension is at least min {2n/(1 + τ ), mn}. It is natural to conjecture that
the Fourier dimension is exactly min {2n/(1 + τ ), mn}. The verification of this
conjecture would make E(m, n, Z, Ψτ , 0) the first explicit example of a set in Rd
(d ≥ 2) with Fourier dimension strictly between 1 and d − 1. Another natural con-
jecture is that E(m, n, Z, Ψτ , 0) is a Salem set, meaning its Fourier dimension is
min{m(n − 1) + (m + n)/(1 + τ ), mn}.
What is the Fourier dimension of E(m, n, Q, Ψ, θ) when no additional restrictions
are placed on the parameters? This is the most general question and therefore the
most challenging.

73
Bibliography

[1] J.-G. Bak, A. Seeger, Extensions of the Stein-Tomas theorem, Math. Res. Lett.
18 (2011), no. 4, 767–781.

[2] A. Banner, Restriction of the Fourier transform to quadratic submanifolds,

Princeton University Thesis, 2002.

[3] A. S. Besicovitch, Sets of fractional dimension (IV); on rational approximation

to real numbers, J. London Math. Soc., 9 (1934), 126-131.

[4] P. Billingsley, Convergence of Probability Measures, 2ed, Wiley (1999).

[5] C. Bluhm, Zur Konstruktion von Salem-Mengen, Ph. D. Dissertation, Erlangen,

1996.

[6] C. Bluhm, Random recursive construction of Salem sets, Ark. Mat. 34 (1996),
51–63.

[7] C. Bluhm, On a theorem of Kaufman: Cantor-type construction of linear fractal

Salem sets, Ark. Mat. 36 (1998), 307–316.

[8] I. Borosh, A. S. Fraenkel, A generalization of Jarnı́ks theorem on Diophantine

approximations, Indag. Math. 34 (1972), 193-201.

[9] J. Bourgain, Remarks on Montgomery’s conjectures on Dirichlet sums, Geomet-

ric aspects of functional analysis (1989-90), 153-165, Lecture Notes in Math.,
1469, Springer, Berlin, 1991.

74
[10] J. Bourgain, On the distribution of Dirichlet sums, J. Anal. Math. 60 (1993),
21-32.

[11] X. Chen, Sets of Salem type and sharpness of the L2 -Fourier restriction theorem,
preprint, 2013. http://arxiv.org/pdf/1305.5584v2.pdf, visited 2015-04-26.

[12] M. Christ, Restriction of the Fourier transform to submanifolds of low codimen-

sion, Thesis, U. Chicago, 1982.

[13] H. Dickinson, A note on the theorem of Jarnı́k-Besicovitch, Glasgow Math. J.

39 (1997), no. 2, 233-236.

[14] M. M. Dodson, Hausdorff dimension, lower order and Khintchine’s theorem in

metric Diophantine approximation, J. Reine Angew. Math. 432 (1992), 69–76.

[15] H. G. Eggleston, Sets of fractional dimensions which occur in some problems of

number theory, Proc. London Math. Soc. 54 (1951), 42–93.

[16] C. Fefferman, Inequalities for strongly singular convolution operators, Acta

Math. 124 (1970), 9–36.

[17] G. B. Folland, Real Analysis: Modern Techniques and Their Applications, 2nd
ed., Wiley (1999).

[18] O. Frostman, Potentiel déquilibre et capacité des ensembles avec quelques appli-
cations à la théorie des fonctions, Meddel. Lunds Univ. Math. Sem. 3 (1935)
1-118.

[19] M. Gatesoupe, Sur un théorème de R. Salem, Bull. Sci. Math. (2) 91 (1967),
125-127.

[20] B. Green, Roth’s theorem in the primes, Annals of Mathematics, 161 (2005),
1609-1636.

[21] G. H. Hardy, E. M. Wright, An Introduction to the Theory of Numbers, 4th ed.,

Oxford: Clarendon Press, 1975.

75
[22] K. Hambrook, Explicit Salem Sets and applications to metrical Diophantine
approximation, preprint. http://www.math.ubc.ca/~hambrook/research/
Hambrook-Explicit-v2.pdf, visited 2015-04-26.

[23] K. Hambrook, I. Laba, On the sharpness of Mockenhaupt’s restriction theorem,

Geometric and Functional Analysis 23 (2013), no. 4, 1262-1277.

[24] T. Hinokuma, H. Shiga, A remark on theorem of Jarnı́k, Ryukyu Math. J. 5

(1992), 1-6.

[25] W. Hoeffding, Probability inequalities for sums of bounded random variables,

Journal of the American Statistical Association 58 (1963), no. 301, 13–30.

[26] V. Jarnı́k, Diophantischen Approximationen und Hausdorffsches Mass, Mat.

Sbornik, 36 (1929), 371–382.

[27] J.-P. Kahane, Images d’ensembles parfaits par des séries de Fourier gaussiennes,
C. R. Acad. Sci. Paris Sr. A-B 263 (1966) A678–A681.

[28] J.-P. Kahane, Some Random Series of Functions, 2nd ed., Cambridge Univ.
Press, 1985.

[29] R. Kaufman, On the theorem of Jarnı́k and Besicovitch, Acta Arith. 39 (1981),
265–267.

[30] T. Körner, Hausdorff and Fourier dimension, Studia Math. 158 (2011), 37–50.

[31] I. Laba, M. Pramanik, Arithmetic progressions in sets of fractional dimension,

Geom. Funct. Anal. 19 (2009), no. 2, 429–456.

[32] J. Levesley, A general inhomogeneous Jarnk-Besicovitch theorem, J. Number

Theory 71 (1998), 65-80.

[33] W. Littman, Lp -Lq -estimates for singular integral operators arising from hyper-
bolic equations, Partial Differential Equations (Proc. Sympos. Pure Math., Vol.
XXIII, Univ. California, Berkeley, Calif., 1971), pp. 479-481. Amer. Math. Soc.,
Providence, R.I., 1973.

76
[34] P. Mattila, Geometry of sets and measures in Euclidean spaces, Cambridge Stud-
ies in Advanced Mathematics, vol. 44, Cambridge University Press, 1995.

[35] T. Mitsis, A Stein-Tomas restriction theorem for general measures, Publ. Math.
Debrecen 60 (2002), 89–99.

[36] G. Mockenhaupt, Bounds in Lebesgue spaces of Oscillatory integrals, Habilita-

tionsschrift, U. of Siegen, 1996.

[37] G. Mockenhaupt, Salem sets and restriction properties of Fourier transforms,

Geom. Funct. Anal. 10 (2000), 1579–1587.

[38] G. Mockenhaupt and T. Tao, Restriction and Kakeya phenomena for finite
fields, Duke Math. J. 121 (2004), no. 1, 35–74.

[39] E. Prestini, Restriction theorems for the Fourier transform to some manifolds
in Rn , Harmonic Analysis in Euclidean Spaces (Proc. Sympos. Pure Math. 35,
Amer. Math. Soc., 1979, vol I), 101-109.

[40] B. P. Rynne, The Hausdorff dimension of sets arising from Diophantine ap-
proximation with a general error function, J. Number Theory 71 (1998), no. 2,
166–171.

[41] R. Salem, On singular monotonic functions whose spectrum has a given Haus-
dorff dimension, Ark. Mat. 1 (1951), 353–365.

[42] E. M. Stein, Maximal functions. I. Spherical means, Proc. Nat. Acad. Sci. U.S.A.
73 (1976), no. 7, 2174–2175.

[43] E. M. Stein, Harmonic Analysis, Princeton Univ. Press, 1993.

[44] R. Strichartz, Convolutions with kernels having singularities on a sphere, Trans.

Amer. Math. Soc. 148 (1970), 461–471.

[45] T. Tao, Some Recent Progress on the Restriction Conjecture, Fourier Analy-
sis and Convexity, 217-243, Appl. Numer. Harmon. Anal., Birkhuser Boston,
Boston, MA, 2004.

77
[46] P. A. Tomas, A restriction theorem for the Fourier transform, Bull. Amer. Math.
Soc. 81 (1975), 477–478.

[47] P. A. Tomas, Restriction theorems for the Fourier transform, in Harmonic Anal-
ysis in Euclidean Spaces (Proc. Sympos. Pure Math. 35, Amer. Math. Soc.,
1979, vol I), 111–114.

[48] S. Wigert, Sur l’ordre de grandeur du nombre des diviseurs d’un entier, Ark.
Mat. 3 (1906/7), 1-9.

[49] T. Wolff, Lectures on Harmonic Analysis, Amer. Math. Soc., Providence, R.I.
(2003).

78
Appendix A

Hoeffding’s Inequality

The standard version of Hoeffding’s inequality for real-valued random variables is

the following.

Lemma A.0.1. If X1 , . . . , Xn are independent real-valued random variables with

EXi = 0 and |Xi | ≤ 1 for i = 1, . . . , n, then

n
!
1X
P Xi ≥ u ≤ 2 exp(−nu2 /2)
n i=1

for all u ≥ 0.

For the proof of Lemma A.0.1, see [25].

The following version of Hoeffding’s inequality for complex-valued random vari-
ables can be deduced from Lemma A.0.1.

Lemma A.0.2. If X1 , . . . , Xn are independent complex-valued random variables with

EXi = 0 and |Xi | ≤ 1 for i = 1, . . . , n, then

n
!
1X
P Xi ≥ u ≤ 4 exp(−nu2 /4)
n i=1

for all u ≥ 0.

79
 √
Proof. For z ∈ C, if |z| = (|Re(z)|2 + |Im(z)|2 )1/2 ≥ u, then |Re(z)| ≥ u/ 2 or
√
|Im(z)| ≥ u/ 2. Therefore

n
! n
! n
!
1X 1X u 1X u
P Xi ≥ u ≤P Re(Xi ) ≥ √ +P Im(Xi ) ≥ √ .
n i=1 n i=1 2 n i=1 2

Applying Lemma A.0.1 on the right-hand side completes the proof.

80
Appendix B

A Fourier Decay Lemma

Lemma B.0.1. Let µ be a complex-valued measure on R with compact support.

Suppose there are positive constants b and c1 such that

(B.0.1) µ(k)| ≤ c1 (ln(e + |k|))1/2 (1 + |k|)−b

|b ∀k ∈ Z.

Then there is a positive constant c2 such that

(B.0.2) µ(ξ)| ≤ c1 c2 (ln(e + |ξ|))1/2 (1 + |ξ|)−b

|b ∀ξ ∈ R.

Proof. If the support of µ is contained in an interval [−M, M ], then the measure ν

defined by Z Z
f (x)dν(x) = 1
f ( 4M x + 12 )dµ(x) ∀f ∈ L1 (µ)
R R

has support contained in [ 41 , 34 ] and satisfies νb(ξ) = µ 1

b( 4M ξ)e−πiξ for all ξ ∈ R. There-
fore it suffices to prove the lemma under the assumption that the support of µ is
contained in [ 41 , 43 ].
Let γ : R → C be a C ∞ function that is equal to 1 on supp(µ) and has supp(γ) ⊆
[0, 1]. For a ∈ [−1, 1], define

γa (x) = e−2πiax γ(x) ∀x ∈ R.

81
We claim that for every n ∈ N there is a constant Cn,γ > 0 depending only on n and
γ such that

(B.0.3) |γba (k)| ≤ Cn,γ (1 + |k|)−n ∀k ∈ Z, a ∈ [−1, 1].

We now prove this claim. Fix n ∈ N. Note that for any compactly supported C ∞
function f : R → C, integration by parts yields

(n) (k) = (2πik)n fb(k) ∀k ∈ Z.

fd

Therefore

(n)
(B.0.4) γa (k) = (2πik)n γba (k) ∀k ∈ Z, a ∈ [−1, 1].
d

By direct computation, the n-th derivative of γa is

n  
−2πiax
X n
γa(n) (x) =e (−2πia)j γ (n−j) (x) ∀x ∈ R.
j=0
k

0
Therefore there exists a constant Cn,γ > 0 depending only on n and γ such that

0
|γa(n) (x)| ≤ Cn,γ ∀x ∈ R, a ∈ [−1, 1].

(n)
Since supp(γa ) ⊆ [0, 1] for every a ∈ [−1, 1], it follows that
Z
(n) 0
|γa (k)| ≤ |γa(n) (x)|dx ≤ Cn,γ ∀k ∈ Z, a ∈ [−1, 1].
d
[0,1]

Then (B.0.4) yields

0
Cn,γ
|γba (k)| ≤ |k|−n ∀k ∈ Z, k 6= 0, a ∈ [−1, 1].
(2π)n

82
This implies (B.0.3) because (1 + |k|)n ≤ 2|k|n for k ∈ Z, k 6= 0 and

|γba (0)| ≤ kγa k1 = kγk1 < ∞ ∀a ∈ [−1, 1].

Now that the claim is proved, we can complete the proof of the lemma. Let
m ∈ Z and a ∈ [−1, 1] be given. Since γa is C ∞ with supp(γa ) ⊆ [0, 1], we have
X
γa (x) = γba (k)e2πikx ∀x ∈ R,
k∈Z

where the convergence is uniform for x ∈ R. Then, as γ is equal to 1 on supp(µ), we

have
Z Z
−2πiax −2πimx
µ
b(m + a) = e e dµ(x) = e−2πiax γ(x)e−2πimx dµ(x)
ZR Z RX
= γa (x)e−2πimx dµ(x) = γba (k)e2πikx e−2πimx dµ(x)
R R k∈Z
X Z X
= γba (k) e−2πi(m−k)x dµ(x) = µ(m − k).
γba (k)b
k∈Z R k∈Z

Write
X X
µ
b(m + a) = µ(m − k) +
γba (k)b µ(m − k).
γba (k)b
|k|≤ 21 |m| |k|> 12 |m|

We bound each sum separately. Fix a positive integer n > 1 + b and define Cb =
−n+b
. If |k| ≤ 21 |m|, then 21 |m| ≤ |m − k| ≤ 23 |m|, and so (B.0.1) and
P
k∈Z (1 + |k|)
(B.0.3) give
X X
µ(m − k)| ≤ c1 Cn,γ (ln(e + 32 |m|))1/2 (1 + 21 |m|)−b
|γba (k)||b (1 + |k|)−n
|k|≤ 21 |m| |k|≤ 21 |m|

≤ c1 Cb Cn,γ (ln(e + 23 |m|))1/2 (1 + 21 |m|)−b .

83
Define Cb0 = supx≥0 (ln(e + x))1/2 (1 + x)−b . Then by (B.0.1) and (B.0.3) we have
X X
µ(m − k)| ≤ c1 Cb0 Cn,γ
|γba (k)||b (1 + |k|)−n
|k|> 21 |m| |k|> 21 |m|
X
≤ c1 Cb0 Cn,γ (1 + 21 |m|)−b (1 + |k|)−n+b
|k|> 12 |m|

≤ c1 Cb Cb0 Cn,γ (1 + 21 |m|)−b

≤ c1 Cb Cb0 Cn,γ (ln(e + 23 |m|))1/2 (1 + 21 |m|)−b .

Combining the bounds on the two sums yields

µ(m + a)| ≤ 2c1 Cb Cb0 Cn,γ (ln(e + 23 |m|))1/2 (1 + 12 |m|)−b

|b
3
≤ c1 Cb Cb0 Cn,γ 2b+ 2 (ln(e + |m + a|))1/2 (1 + |m + a|)−b

for every m ∈ Z and a ∈ [−1, 1]. This proves (B.0.2) because each ξ ∈ R can be
written as ξ = m + a for some m ∈ Z and a ∈ [−1, 1].

It is possible to prove a more general version of Lemma B.0.1 (cf. [28, p. 252]).

84