Arithmetic Structures in Small Subsets of Euclidean Space
Dissertation
Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University
By
Marc Carnovale, M.S.
Graduate Program in Mathematics
The Ohio State University
2019
Dissertation Committee:
Prof. Vitaly Bergelson, Advisor Prof. Alexander Leibman Prof. Krystal Taylor c Copyright by
Marc Carnovale
2019 Abstract
In this thesis we extend techniques from additive combinatorics to the setting of harmonic analysis and geometric measure theory. We focus on studying the distribution of three- term arithmetic progressions (3APs) within the supports of singular measures in Euclidean space. In Chapter 2 we prove a relativized version of Roth’s theorem on existence of 3APs for positive measure subsets of pseudorandom measures on R and show that a positive measure of points are the basepoints for three-term arithmetic progressions within these measures’ supports. In Chapter 3 we combine Mattila’s approach to the Falconer distance
d conjecture with Green and Tao’s arithmetic regularity lemma to show that measures on R with sufficiently small Fourier transform as measured by an Lp-norm have supports with an abundance of three-term arithmetic progressions of various step-sizes. In Chapter 4
d we develop a novel regularity lemma to show that measures on R with sufficiently large dimension, as measured by a gauge function, must either contain non-trivial three-term arithmetic progressions in their supports or else be structured in a specific quantitative manner, which can be qualitatively described as, at infinitely many scales, placing a large amount of mass on at least two distinct cosets of a long arithmetic progression.
ii Aknowlegements
First and foremost, I would like to thank my Ph.D. advisor Vitaly Bergelson, who set me on the road many years ago. His constant support and generosity have had a greater impact on me than I could ever express.
In addition to Vitaly Bergelson, who taught me so much of what I know about math- ematics, I would especially like to thank my M.Sc. advisors IzabellaLaba and Malabika
Pramanik who taught me much of the rest.
Many thanks also to my thesis defense committee members Alexander Leibman and
Krystal Taylor for their time and encouragement.
Finally I would like to thank my parents, family, and friends, who have always been there for me and provided endless support.
iii Vita
2010 ...... B.S. Mathematics Ohio State University. 2013 ...... M.S. Mathematics University of British Columbia. 2013-present ...... Graduate Teaching Associate, Ohio State University.
Publications
Research Publications
M. Carnovale “A relative Roth theorem in dense subsets of pseudorandom fractals.” Online Journal of Analytic Combinatorics, 27–55, June. 2015.
Fields of Study
Major Field: Mathematics
iv Table of Contents
Page
Abstract ...... ii
Vita...... iv
List of Tables ...... vii
1. Introduction ...... 1
1.1 Notation ...... 2 1.2 A selected history of additive combinatorics ...... 4 1.3 A selected history of geometric measure theory ...... 6 1.4 A selected history of harmonic analysis ...... 9 1.4.1 The Restriction Conjecture ...... 10 1.4.2 The multilinear Hilbert transform ...... 12 1.5 Content of this thesis...... 14
2. A relative Roth theorem in dense subsets of sparse pseudorandom fractals . . . 16
2.1 Notation and approach ...... 19 2.2 Existence of the restricted measure on 3APs ...... 21 2.3 The Main Estimate : Proof of Lemma 2.3.2 ...... 25 2.4 3APs in dense subsets : Proof of Theorem 2.0.1 ...... 32 2.5 Existence of the measure on 3APs ...... 33 2.6 Further Remarks ...... 40
3. On the number of 3APs in fractal subsets of Euclidean Space ...... 42
3.0.1 Integer sets ...... 43 3.0.2 Fractal sets ...... 44 3.0.3 Finite fields ...... 50 3.1 Integer results ...... 51 3.1.1 3AP counts of L2 functions ...... 58
v 3.2 Fractal results ...... 62 3.2.1 Proof of Theorem 3.0.7 ...... 63 3.2.2 Proof of Theorem 3.0.8 ...... 73 3.3 Finite field results ...... 77 3.3.1 Proof of Corollary 3.0.11 ...... 77 3.3.2 Gauss sums ...... 78 3.3.3 Circular bounds and the proof of Theorem 3.0.14 ...... 82 3.3.4 Decay and the proof of Theorem 3.0.12 ...... 83
4. 3APs in dense subsets of Euclidean Space ...... 86
4.1 Introduction ...... 86 4.2 Large gauges without 3APs ...... 88 4.3 Almost flat fractals ...... 89 4.4 Counting 3APs in almost-flat fractals ...... 90 4.4.1 Counting 3APs between three progressions ...... 90 4.4.2 The counting lemma ...... 95 4.5 An (L1, U 2) regularity lemma ...... 98 4.5.1 The regularity lemma ...... 98 4.5.2 Bohr sets ...... 104 4.6 Decomposing a large fractal into an almost-flat portion and a small pseu- dorandom error ...... 106 4.7 Counting 3APs in large fractals ...... 108
Bibliography ...... 109
vi List of Tables
Table Page
1.1 Comparison of discrete versus fractal results on 3APs...... 15
vii Chapter 1: Introduction
This thesis explores applications of additive combinatorics to Euclidean harmonic analysis and geometric measure theory through the lens of one very specific family of structures: arithmetic progressions. We are particularly interested in the study of their presence within subsets of Hausdorff dimension less than d of d-dimensional Euclidean space (we will refer to such subsets as ‘fractals’). Such a study was initiated by the seminal 2009 paper ofLaba and Pramanik [35] (see Theorem 1.4.1 in Section 1.4 below).
There are several reasons such a study may be of interest. First, a number of recent advances in Harmonic Analysis have seen connections to questions about the distribution of arithmetic progressions. Second, it is natural to count the number of k-term arithmetic progressions (kAPs for short) in a set A with characteristic function 1A by introducing the multilinear functional
ZZ Λk(f0, . . . , fk−1) := f0(x)f1(x − r) . . . fk−1(x + (k − 1)r) dx dr
and applying it to the tuple (f0, . . . , fk−1) = (1A,..., 1A); it is the consequence of monu- mental advances in combinatorial number theory, additive combinatorics, and ergodic theory that in many contexts these multilinear funtionals are fairly well-understood compared to two to three decades ago. Finally, many of the tools being developed in additive combina- torics to understand discrete problems are of a finitistic Fourier-analytic nature but largely novel to Euclidean harmonic analysis, and extending them to this setting is often non-trivial.
1 It is our belief that important future advances in Euclidean harmonic analysis will continue
to benefit from tools developed in additive combinatorics.
1.1 Notation
Given quantities A and B which may depend on auxiliary parameters ~s = hs1, . . . , ski
we write
A . B
to indicate the existence of a finite positive constant C (independent of these parameters
unless otherwise noted) such that
A ≤ CB.
uniformly in these parameters.
N Similarly, when A = A , B = B , and lim (1+As) = 0 for some power N > 1, we s s s → ∞ Bs will write A B.
We use also the notation A & B when B . A and A B when B A, and we write
A ≈ B when both A . B and B . A hold.
Given an integer N ∈ N, write [N] to denote the set
[N] := {n ∈ N : 0 ≤ n < N} ;
(note that this differs from standard usage where [N] denotes the integers from 1 to N).
d Define the Fourier transform of a function f or Borel measure µ on R by Z fˆ(ξ) = f(x)e−2πiξ·x dx, d R Z µ(ξ) = e−2πiξ·x dµ(x) b d R d for ξ ∈ R .
2 For a function f on ZN := Z /N Z define for ξ ∈ ZN
1 X fˆ(ξ) = f(x)e−2πinξ/N . N x∈ZN ˆ For a function f :[N] → C, we define f(ξ) via first identifying f with a function on Zp
for p a prime between 3N and 6N via the embedding [N] 7→ Zp, n 7→ (n mod p) ∈ Zp and then taking the Fourier transform on this group. That such a prime, p, always exists is
guaranteed by Bertrand’s postulate.
Given functions f, g : G 7→ C on an abelian group, G with Haar measure m, we denote convolution with an asterisk:
Z f ∗ g(x) = f(x − y)g(y) dm(y), G
and extend this to finite Borel measures in the usual way, R f dµ∗ν := R f(x+y) dµ(x) dν(y).
We will denote the Lebesgue measure of a set, E, by |E|.
We will at times write Lˆq to refer to the space of functions whose Fourier transforms
possess finite Lq norms.
d Denote the ball of radius r centered at a point x in R by B(x, r), and let σ denote the surface measure on the unit sphere.
In this thesis, we will very often be interested in positive Borel measures µ on [0, 1]d
satisfying the following Hausdorff dimension and Fourier decay conditions
α (a) µ(B(x, r)) ≤ CH r
and
−β/2 (b) |µb(ξ)| ≤ CF (1 + |ξ|)
for some positive constants CH and CF .
We will refer back to conditions (a) and (b) frequently.
3 1.2 A selected history of additive combinatorics
The study of the distribution of kAPs within various subsets of abelian groups is one of the central questions in combinatorial number theory and additive combinatorics, dating back to a 1927 theorem of van der Waerden, [53], stating that any parition of the natu- ral numbers into finitely many subsets must contain a cell which contains arbitrarily long arithmetic progressions, and a later conjecture of Erd˝osand Tur´an,[14], asking whether any large enough subset, A, of the natural numbers must contain arbitrarily long arithmetic
progressions. Here by large enough is meant that the upper density of the set be positive,
that is
|A ∩ [N]| d(A) := lim sup > 0. (1.1) N → ∞ N
Erd˝osis also known for another conjecture, namely that to guarantee kAPs in a set A
for arbitrary k it is sufficent that
X 1 = ∞. (1.2) n n∈A
It is easily seen that (1.2) is strictly stronger than (1.1); in fact, (1.2) is roughly equivalent to asking whether
|A ∩ [N]| 1 d (A) := (1.3) [N] N & log N
(we refer the reader to [23] for a precise statement of (1.3)).
By the Prime Number Theorem, the density d[N](P) of the set of primes, P, is approx-
1 imately log N , and so an affirmative answer to the question of whether, for any k ∈ N,
1 d[N](A) & log N for sufficiently large N implies that the set A contains a (non-trivial) kAP would in turn imply that statement for the primes, answering an older question of Erd˝oson
whether the primes contain arbitrarily long arithmetic progressions.
4 Using special properties of the prime numbers, in 1939 van der Corput showed that the primes do indeed contain infinitely many 3APs ([52]) but this proof does not extend readily to other sets. The full question of whether the primes contain arbitrarily long arithmetic progressions went unanswered until the 2008 breakthrough paper [26] of Green and Tao.
In 1954 Roth ([45]) gave a proof that (1.1) implies that the set A contains (non-trivial)
3APs, and his proof gives something a little more quantitative than that.
Theorem 1.2.1 (Roth’s Theorem). Let A ⊂ [N] and suppose that
1 d (A) . [N] & log log N
Then provided that N is sufficiently large, the set A contains non-trivial three-term arithmetic progressions.
A key feature of Roth’s proof is that it does not directly show that the original set contains
3APs, but rather proceeds by refining this set into a subset A0 which is either “structured” in the sense that A0 is itself a long arithmetic progression, or “pseudorandom” in the sense that 2 |A0| 0 sup 1 0 (ξ) < (which may be shown to roughly mean that A is approximately ξ6=0 cA N evenly distributed over the long enough sub-progressions of [N].1). In both cases, it is easy to show that the subset contains 3APs, so the main work is in showing this dichotomy that any set must exhibit either “structured” or “pseudorandom” behaviour.
In 2005 Green ([25]) extended Roth’s theorem (as a theorem about the natural numbers
N) to the primes by exploiting deep results that the primes are especially pseudorandom modulo trivial congruence properties. Chapter 2 takes particular inspiration from this result.
In 1975, Szemer´ediextended Roth’s theorem by proving the following.
Theorem 1.2.2 (Szemer´edi’s Theorem, [47]). Let k ∈ N. If A ⊂ [N], |A|/N ≥ δ, and N is sufficiently large depending on k and δ, then A contains kAPs.
1In fact, in Roth’s proof of his theorem as well as in a majority of other proofs, one does not obtain this decomposition in a single step but rather as a result of successive refinements of the original set.
5 Many consider this one of the most significant results of 20th century mathematics. A
second proof was provided by Furstenberg, using ergodic methods, in 1977 ([18]). This led
to many sinificant extensions, some of which remain inaccessible via non-ergodic methods
(c.f. [19],[20],[3],[4]).
1.3 A selected history of geometric measure theory
Geometric measure theory studies the geometry of sets through the viewpoint of measures
supported on those sets. Perhaps the most fundamental geometric data accessible via such
d measures is the Hausdorff dimension, dimH E, of a set E ⊂ R . The usual definition of Hausdorff dimesion goes through certain coverings and outer
measures as follows.
d Let R be the semi-ring of sets generated by the balls in R . Let s ∈ [0, d].
0 s 0 Given a ball B(x, r), set ms(B(x, r)) = r and extend ms to R via additivity. For a given δ > 0, let R0 denote the sub-semi-ring R0 ⊂ R consisting of those elements of
s R with diameter less than δ, and define Hδ to be the measure generated by the pre-measure
0 0 ms on R .
s It is easy to see that for any set E, Hδ(E) is monotonic increasing as δ ↓ 0.
Definition 1.3.1. Define Hs, the s-dimensional Hausdorff measure, to be given by Hs(E) :=
s s supδ>0 Hδ(E) = limδ↓0 Hδ(E).
We also define the Hausdorff dimension of a measure µ as
α dimH µ = sup {α ∈ (0, d): ∃ C such that µ(B(x, r)) ≤ Cr for µ − a.e.x} .
It is a consequence of the following theorem of Frostman that the Hausdorff dimension
of a set is a geometric measure-theoretic quantity.
6 d Theorem (Frostman’s lemma, c.f. [40] Theorem 8.8.). Let E ⊂ R be a Borel set. Then Hs (E) > 0 if and only if there exists a positive Borel measure, µ, supported on E such that
µ (B(x, r)) ≤ rs
d for µ-a.e. x ∈ R .
In fact, there is a weaker notion of dimension for Euclidean measures than that guar- anteed by Frostman’s Lemma which is more analytically tractable and often sufficient in applications. We give the definition below.
Given a real number s and a distribution f, define the Sobolev norm kfk d by W 2,s(R )
Z 2 1/2 s kfk d := f(ξ) (1 + |ξ|) dξ . W 2,s(R ) b
We are particularly interested in the (negative exponent) Sobolev norm
Z 2 2 −(d−α) kfk 2,−(d−α) d = f(ξ) (1 + |ξ|) dξ W (R ) b for α ∈ (0, d).
We now define
2 I (µ) := kµk 2,−(d−α) d , α W (R ) usually termed the α-dimensional energy of µ.
It turns out that Iα has a more direct spatial representation, stemming from the (distribu-
−(d−α) −α tional) Fourier transform cd,α|ξ| of the function x 7→ |x| (where cd,α is an appropriate constant),
ZZ 1 I (µ) = c dµ(x) dµ(y). α d,α |x − y|α
One then defines the capacitary dimension, dimcap(µ), as
dimcap µ = sup {α ∈ (0, d): Iα(µ) < ∞}
7 with the convention that the supremum of the empty set is zero.
Note that since the integral R 1 dm (x) < ∞ for all > 0 (and is infinite for B(0,1) |x|d− d
= 0), the capacitary dimension of Lebesgue measure md is d.
It is straightforward that if dimH µ = α, then dimcap µ ≥ α.
If dimcap µ > α for some α ∈ (0, d) then it can be shown that one may select a new measure, ν, supported on a subset of supp µ such that dimH ν > α. Thus Frostman’s lemma implies the variant of itself where dimH µ is replaced by dimcap µ.
There is another quantity often referred to as a dimension and which is of particular interest within harmonic analysis. We noted that dimH µ > α implies that dimcap µ > α, and we have that
0 dimcap µ = sup {α ∈ (0, d): Iα0 (µ) < ∞} where Z 2 −α0 Iα0 (µ) = |µ(ξ)| |ξ| dξ. d b R −α/2 Thus, informally, |µb| must decay as |ξ| → ∞ like |ξ| on average. If for some β > 0, −β/2 |µb(ξ)| . |ξ| for all (large enough) ξ, then certainly dimcap µ ≥ β, though it is not neces- sary that there be any such β > 0 even when µ is of full Hausdorff dimension. Nonetheless, it is useful to define
n −β/2o dimF µ := sup β ∈ (0, d): |µb(ξ)| . |ξ| ,
referred to as the Fourier dimension of µ.
Many basic questions in geometric measure theory take the form of asking whether large-
ness of the dimension of a set is sufficient to guarantee structure of the set under question.
One of the most important unresolved questions in the field is of this form, where we make
d “largeness” precise by asking, given E ⊂ R , that dimH E > d/2, and make structure precise
8 by asking that the set
∆(E) := {r = |x − y| : x, y ∈ E} have positive Lebesgue measure.
In terms of measures, this may be restated as
d Conjecture 1.3.2 (Falconer, [16]). Suppose that µ is a probability measure on R and that
α µ(B(x, r)) . r for some α > d/2. Then
|{r : there exist x, y ∈ supp(µ) such that |x − y| = r}| > 0.
In Chapter 3 we adopt the viewpoint that the Falconer distance conjecture asks a question about the distribution of 2APs in supp µ under the assumption that dimH µ > d/2. If one replaces dimH by dimF , then it is an observation of Mattila ([39]) that the conclusion of the
Falconer distance conjecture holds true.
1.4 A selected history of harmonic analysis
In their 2009 paper [35],Labaand Pramanik proved:
Theorem 1.4.1 (Laba-Pramanik, 2009). Suppose that µ is a probability measure on [0, 1] satisfying2 conditions (a) and (b) for some β > 2/3 and α sufficiently close to 1 dependng on β and the constants CH ,CF appearing in (a) and (b). Then supp µ contains non-trivial
3APs.
There are a number of results in harmonic analysis where the arithmetic structure influ- ences the analytic or geometric/combinatorial structure. In this section we discuss a few of them, and discuss how each relates, also, to the kind of 3AP counts first studied by Theorem
1.4.1.
2See Section 1.1 for the statements of conditions (a) and (b).
9 1.4.1 The Restriction Conjecture
One of the most significant areas of research in Euclidean harmonic analysis is the study
of the (Fourier) Restriction Conjecture.
The Restriction Conjecture3 relates the concentration of the energy of the Fourier trans-
form of a function f, as measured by some Lbq(σ) norm, fˆ , to its mass (as measured Lq(σ) p by some L norm, kfkLp ), by asking whether
ˆ f . kfkLp( d) . Lq(σ) R
It is not hard to observe that the curvature of the sphere plays a significant role in the
problem. For instance, instead of σ consider the Lebesgue measure mV supported on a flat
d d affine subspace V ⊂ R . Then given f on R , first, we have
f\ dλ d ≤ kfk q0 R d L (λ d ) Lq(R ) R
for q ≥ 2 (this is the classical Hausdorff-Young inequality, and it is known to be sharp in
d d that no further Lp(R ) − Lq(R )-inequality is true in general); and second, it is easy to see that
f\ dλV 6 kfk p d . L (mV ) Lq(R )
d for any (p, q) 6= (1, ∞) for any proper affine subspace V 0 (R .
On the other hand, if one replaces the flat measure mV with the surface measure mS
d where S is any hypersurface of non-vanishing Gaussian curvature in R , then one has the remarkable inequality
f\ dmS ≤ kfk 2 , d L (mS ) Lq(R )
3Stein posed the Restriction Conjecture in the 1970’s, asking whether the following is true.
p d 2d Conjecture 1.4.2 (Restriction Conjecture, Stein). Let f ∈ L (R ) for some p ∈ [1, d+1 ) and suppose that d−1 0 0 p q ≤ d+1 p where p = p−1 . Then
ˆ f . kfkLp( d) . Lq (σ) R
10 d+1 d−1 for q ≥ 2 2−1 , first proven for the d-dimensional unit sphere, S = S , by Tomas ([51]) for d+1 d+1 ˆ q > 2 2−1 and by Stein for the endpoint q = 2 2−1 . This inequality shows that although f is only initially defined Lebesgue-almost-everywhere, it has a well-defined restriction to the
(Lebesgue-measure zero) sphere. This theorem was later generalized by Greenleaf ([28]) to the case of general hypersurfaces of non-vanishing Gaussian curvature.
For a long time the phenomenon of Fourier Restriction was perceived to rely on the geometry of certain smooth manifolds. Then around the year 2000, it was noticed that the structure of a smooth manifold doesn’t play any vital role in the estimate, and in fact the
Stein-Tomas restriction theorem holds for any measure µ satisfying appropriate Hausdorff dimension and Fourier decay bounds (those same ones encapsulated by conditions (a) and
(b) stated in Section 1.1).
Theorem 1.4.3 (Mitsis [42], Mockenhaupt [43], endpoint Lq bound by Bak-Seeger [1]). Let
d µ be a probability measure on R satisfying (a) and (b) for some CH and CF ∈ (0, ∞).
d−α+β 2 Then given any q ≥ 2 2(d−α)+β and any f in L (µ),
f dµ ≤ kfk 2 . d d L (µ) Lq(R )
In [43], Mockenhaupt remarks that (bounds on) the number of long progressions in the
set may play the role of curvature hypotheses in the fractal setting, and this was made
more precise by the following result ofLabaand Hambrook, showing sharpness of the Mitis-
Mockenhaupt-Bak-Seeger result by building a measure whose support contains many long
progressions at small scales.
Theorem 1.4.4 (Laba-Hambrook [29]). For every α0 ∈ (0, 1) there exists an α > α0 such
4 that for all p ∈ [1, α − 2), for all β < α there exists a measure µ satisfying (a) and (b), and
11 a sequence of functions fl, such that
f[l dµ p L (R) → ∞. kflkL2(dµ)
We note also that the Mitsis-Mockenhaupt-Bak-Seeger Stein-Tomas restriction theorem
has the same hypotheses as theLaba-Pramanik 3AP result Theorem 1.4.1 - namely the
conditions (a) and (b). We can rephrase this as saying that the hypotheses which lead
to the conclusion of Theorem 1.4.3 are connected to those which allow for good estimates
on 3AP counts within fractals. In the other direction, we have theLaba-Hambrook result
demonstrating sharpness of the Mitsis-Mockenhaupt restriction theorem in the case of certain
measures with very bad kAP counts. Given these two observations, it is natural to conjecture
that good control on the number of kAPs for k much larger than 3 might allow for L2 − Lp
restriction theorems for exponents p below that guaranteed by Theorem 1.4.3. Thus it seems
desireable to better understand arithmetic behaviour of fractals associated to the counts of
the arithmetic progressions in their support.
1.4.2 The multilinear Hilbert transform
Another of the outstanding open questions in harmonic analysis is that of boundedness
of the multilinear Hilbert transform. The Hilbert transform is defined initially for Schartz
dt R 1 functions f : R → C (and treating t as a tempered distribution) as Hf(x) = f(x − t) t dt,
R 1 and the bilinear Hilbert transform as H(f, g)(x) = f(x−t)g(x+t) t dt. Boundedness of the
Hilbert transform on Lp(R) for p > 1 was established by by Calder´onand Zygmund in 1952 in their seminal study of singular integrals, ([7]). Boundedness of the Bilinear Hilbert transform from Lp1 ( ) × Lp2 ( ) to Lq( ), for 1 + 1 = 1 when q ∈ ( 2 , ∞), was proven by Lacey and R R R p1 p2 q 3 Thiele in the 90’s [36], [37], answering a decades old question of Calder´on.More generally,
R Qk dt p given fi : R → R, i = 1, . . . , k, one defines H(f1, . . . , fk)(x) = i=1 fi(x − it) t . L - boundedness of the trilinear and higher order Hilbert transforms remains entirely open. Tao
12 recently used tools from additive combinatorics4 in order to obtain estimates for truncations5 of the mulitlinear Hilbert transform which blow up logarithmically more slowly than the trivial estimates available via Holder’s inequality ([48]).
One may associate a multilinear form to the Lp norm of the multilinear Hilbert transform via Lp duality as follows:
Z kH(f1, . . . , fk−1)kLp = sup g(x)H(f1, . . . , fk−1)(x) dx kgk 0 Lp =1 ZZ dt = g(x)f (x − t)f (x − 2t) . . . f (x − (k − 1)t) dx . 1 2 k−1 t
If one considers a modification of the tri-linear form associated to the norm of the bilinear
1 Hilbert transform, in which the singular measure t dt is replaced by the Lebesgue measure dt, then one obtains the 3AP-counting functional
ZZ Λ3(f0, f1, f2) := f0(x)f1(x − t)f2(x − 2t) dx dt, (1.4) and so perhaps it is not so surprising that recent advancements on the study of multilinear
Hilbert transforms have used some of the same technology as that appearing in the study of
Λk(f0, . . . , fk−1) and the distribution of arithmetic progressions.
If one replaces the functions fi in (1.4) with φ ∗ µ where φ is an approximate identity and µ is a probabilty measure satisfying (a) and (b), then this is the primary object of study to obtain existence of 3APs within the support of the fractal µ, as in, e.g., Theorem 1.4.1.
This is one manner in which the study of 3APs within singular sets may be seen as a model for importing tools from additive combinatorics to a harmonic analysis setting.
4In particular, the Inverse Conjecture for the Gowers norms in the form of the Arithmetic Regularity Lemma. 5Given , define k Z Y dt H(f1, . . . , fk)(x) := fi(x − it) . c t B(0,) i=1 The bounds in [48] blow up, as ↓ 0, more slowly than would be implied by Holder’s inequality by a logarithmic factor in .
13 1.5 Content of this thesis.
Our primary results in this thesis focus on a delicate analysis of the 3AP counts within
the support of a measure µ, taking the study instigated by [35] as an excellent model case
for applying the tools of additive combinatorics in a Euclidean harmonic analysis setting.
In Chapter 2 we extend the main result ofLaba and Pramanik’s paper “Arithmetic
progressions in sets of fractional dimension” ([35]) to obtain a “relative” version of their
theorem along the lines of the 3AP version of the “relative Szemer´editheorem” of Conlon,
Fox, and Zhao [10] (this relative Szemer´editheorem, extending work of [25] and [26], showed
that dense subsets of appropriately pseudorandom sets, not just dense subsets of the full
space, contain arbitrarily long arithmetic progressions). In [35], the authors proved that any
measure µ on [0, 1] with Hausdorff dimension α ∈ (1 − 0, 1) (here 0 is a small constant) large enough depending on its Fourier dimension β ∈ (2/3, α] and the parameters CH ,CF
of (a) and (b) contains in its support three-term arithmetic progressions (3APs). Here, we
adapt an approach introduced by Green in “Roth’s Theorem in the Primes”, [25], to show
for any δ > 0 that if α is large enough depending on δ (and the other data in the hypotheses)
then µ gives positive measure to the (basepoints of the) non-trivial 3APs contained within
any set A for which µ(A) > δ.
In Chapter 3, we use techniques introduced by Mattila in [40] to study the Falconer
distance conjecture, to explore conditions which guarantee largeness (in terms of bounded
L2 density/Lebesgue measure and Hausdorff measure) of the set of lengths of step-sizes of
3APs which occur within fractal sets. Our main result is a version ofLabaand Pramanik’s
theorem in [35] that relies only on an assumption of a lower bound, δ, on the mass of the
measure µ together with an upper bound, M on the Lq norm of its Fourier transform for some q ∈ (2, 3] depending on the parameters δ and M. In fact, we show that conditions (a) and (b)
14 Comparison of Discrete vs. Fractal Settings Discrete sets on [N] Singular measures on [0, 1]d Roth’s Theorem Theorem 4.1.3 Sparse Pseudorandom Roth’s Theorem Laba-PramanikTheorem 1.4.1 Green’s Relative Roth Theorem Theorem 2.0.1
Table 1.1: Comparison of discrete versus fractal results on 3APs.
imply our assumption on the Lˆq norm of µ, so that out theorem implies theLaba-Pramanik
as a special case. We also obtain analogous statements in the discrete setting.
In Chapter 4 we use a novel regularity lemma to show that any set E ⊂ [0, 1] satisfying
the hypothesis that E have positive generalized Hausdorff measure Hh(E) > 0 with respect
to a sufficiently large gauge function, h, must either contain non-trivial 3APs or possess a
specific structure on its support. The proof uses the regularity lemma to replace the measure
µ with a new measure µe with approximatley the same 3AP count behaviour but which is “structured” in the sense that it looks like a measure built via a Cantor-type construction.
We then use this structure to obtain strong enough counts on the number of 3APs within
the support of µe, as well as their distribution, to guarantee that the set E must contain non-trivial 3APs as well.
This allows us to complete Table 1.1 comparing what is known in the discrete versus
fractal settings and how these results relate to one another.
15 Chapter 2: A relative Roth theorem in dense subsets of sparse pseudorandom fractals
This chapter is primarily work published in [8].
In this chapter we employ the technology of restriction/transference results in additive
combinatorics developed in [25] for the prime numbers to extend the result, Theorem 1.4.1, of
Labaand Pramanik [35] demonstrating three-term arithmetic progressions in certain fractal
sets.
We will be concerned with a measure µ on T satisfying the following conditions for
appropriate α, β, CH ,CF in the case that d = 1
α d (a) |µ|(B(x, r)) ≤ CH r for all x ∈ T
− β d 6 (b) |µb(ξ)| ≤ CF (1 + |ξ|) 2 for all ξ ∈ Z .
In [35], it was shown that a measure µ satisfying (a) and (b) for α sufficiently close
to 1 depending on β > 2/3, CH , and CF must contain in its support 3-term arithmetic
progressions. Consequently, any closed set supporting a measure satisfying (a) and (b) with
appropriate constants must contain 3APs. In the present thesis, we show that in fact, dense
subsets of such sets still contain 3APs, in particular providing a condition for a merely
measurable set to contain progressions. Our main result is
6The largest α and β ∈ (0, 1) for which (a) and b hold are referred to as the Hausdorff and Fourier dimensions of the measure µ. For more on these, see, e.g., [40].
16 Theorem 2.0.1. Let δ > 0, and suppose that the probability measure µ on T satisfies (b)
1 with β > 2 and (a) with α sufficiently close to 1 depending on β, δ, and the implicit constants
CH and CF in (a) and b.
Then any measurable set A with µ(A) > δ contains 3APs 7 ; indeed, µ gives positive
measure to the set of x such that x, x + r, x + 2r is a non-trivial 3AP contained with A.
Remark 2.0.1. Notice that it follows from the above Theorem that the total measure with
respect to µ of the set of x which are not the basepoint of a 3AP entirely contained within
supp µ is less than δ.
The proof of Proposition 2.4.1 (upon which the proof of the above theorem rests) requires
9−2β that α > 8 . By Remark 2.3.1, in the case that δ = 1 and CH ,CF ≈ 1, it is sufficient to take α > 1 − β/6 for Lemma 2.3.2. Thus if β = α in the full-density case, as is nearly-
achieved by Salem sets, then we need only take α = β > 9/10. It is unclear whether the
dependency of α on the parameter δ < 1 may be removed. For details, see Section 2.6.
This result even applied to the full support of the measure is novel in that we show µ
gives positive measure to the basepoints of non-trivial 3APs contained in its support.
Our approach is based on [25], in which it was shown that 3APs are contained in any
set taking up a positive proportion of the prime numbers. There, the main ingredients were
the pseudorandomness of the prime numbers, a restriction theorem for pseudorandom sets
(though in that paper stated only for the prime numbers), and the properties of Bohr sets.
In the present context, all three ingredients are available. Namely, the Fourier decay condi-
tion (b) plays the role of pseudorandomness, we invoke Theorem 2.0.2 to obtain restriction
estimates, and the properties of Bohr sets carry over unchanged to the continuous setting.
7Note that by requiring that supp(µ) ⊂ [1/3, 2/3] ⊂ [0, 1] ≈ T, or equivalently by dilating, we guarantee that the progressions above are genuine progressions when supp(µ) is embedded in R.
17 One reason to be interested in the arithmetic properties of fractional sets is their impli-
cation in the fine analytic behaviour of singular sets.
Recall from Chapter 1 that in 2002 (respectively, 2000) Mitsis [42] (independently Mock-
enhaupt [43]) obtained the following Stein-Tomas type restriction theorem for fractional
sets.8
d Theorem 2.0.2. Let µ be a compactly supported positive measure on R which obeys (a)
2(2d−2α+β) and (b) for some α, β ∈ (0, d). Then for all p ≥ p◦ = p◦(d, α, β) := β , there is a C(p) > 0 such that
9 kfdµdk`p(Zd) ≤ C(p)kfkL2(µ).
As noted by Mockenhaupt [43], obstructions to restriction results for the sphere rely on
arithmetic properties of the surface measure on the sphere (more specifically, on the existence
of arithmetic progressions in small neighborhoods of pieces of this measure). Recent work
ofLabaand Hambrook [29] made more precise the analogy between arithmetic progressions
in fractional sets and the classical Knapp example showing sharpness of the spherical Stein-
Tomas theorem. They did so by constructing measures µ on R obeying (a) and b which nonetheless contained an abundance of long arithmetic progressions, and demonstrating
that an extension of Theorem 2.0.2 to an improved Lp range fails for such µ via an argument
which makes direct use of the many long arithmetic progressions contained in supp(µ). Thus
an understanding of deep properties of singular measures can be seen to begin with the easier
problem of understanding their arithmetic properties.
This work is particularly motivated by a desire to directly connect to harmonic analysis
a principle in additive combinatorics that a not too small pseudorandom subset of a space
should possess many of the same properties as the full space, and results for subsets of
8Aside from the end point, which was obtained by Bak and Seeger in 2010.
9 In [42] and [43] this theorem was stated with kfdµk p d in place of kfdµk p d . By Lemma 2.3.3, the d L (R ) d ` (Z ) two are equivalent for p ∈ (1, ∞]
18 the entire space should, under suitable hypotheses, hold for subsets of the pseudorandom subspace. The most well-known example of such a phenomenon is the celebrated Green-Tao theorem that the primes contain arbitrarily long arithmetic progressions [26], whose proof entailed first demonstrating that the prime numbers behave in a suitably pseudorandom fashion and second developing a “relative Szemer´edi”theorem for subsets of appropriately pseudorandom subsets of the integers. Their result extends ideas present in Green’s earlier
Roth’s Theorem in the Primes [25], from which we take inspiration in the present chapter.
The principle of such transference from a large global space to a nice but sparser subspace can be seen in earlier work, such as in Stein’s Spherical maximal theorem [46], which finds that the measure on the sphere behaves almost “as well” as the measure on the unit ball for purposes of obtaining a maximal theorem, and similarly in Bourgain’s ergodic theorem along the squares [6]. What many of these results have in common is that they rely on the pseudorandom set being close in a Fourier or spectral sense to the indicator function for the entire space. In [25], this is expressed by a Fourier restriction estimate which underlies the result, but in [26], as further developed in [50], abstracted in [22] and [44], and in the excellent combinatorial strengthening [10], this proximity takes on an arithmetic nature encoded by the Gowers uniformity norms and related objects. Although we do not take the combinatorial viewpoint here, relying instead on a restriction estimate as in [25], it seems that combinatorial methods such as those of [10] would be necessary to extend our result to the case of longer progressions.
In Section 2.1 we set up notation and describe the approach.
2.1 Notation and approach
Throughout, µ will always refer to a (possibly complex) Radon measure supported on
[0, 1] ≈ T. We denote the variation norm of the measure µ by kµk.
19 ∗ For a set A and measure µ, set µA := µ|A the restriction of µ to the set A, and let Aµ denote the set of points of positive density with respect to the measure µ; that is, for each
∗ x ∈ Aµ, lim sup |µ|(A ∩ B(x, r))/|µ|(B(x, r)) > 0.
∗ Note that µ(A/Aµ) = 0 by [40], pg. 91, Remark 1.
Throughout, let (φn) be an approximate identity on T (that is, a sequence of positive functions with L1-norm 1 converging weak∗ to the dirac delta function), with the property
n n that supp(φcn) ⊂ [−2 , 2 ] and φcn|[−2n−1,2n−1] ≡ 1.
The primary additional technical aspect of the argument involves the “progression count- ing functional” Λ3(µ), defined as Z lim φn ∗ µ(x)φn ∗ µ(x − r)φn ∗ µ(x − 2r) dx dr n → ∞
which first appeared in the context of measures, in a slightly different form, in [35], and more
generally the measure ∩3µ defined by Z Z 3 f d ∩ µ := lim f(x, r)φn ∗ µ(x)φn ∗ µ(x − r)φn ∗ µ(x − 2r) dx dr. n → ∞
We develop their existence theory in Section 2.5.
Recall that the quantity α in (a) is a Hausdorff dimension estimate, while the quantity
β of (b) is an estimate on the Fourier dimension of µ.
1 Theorem 2.1.1. Suppose that µ is a measure on T satisfying (a) and (b) with β > 2 and
9−2β α > 8 .
2 Then the finite measure ∩3µ : C(T ) → C given by
Z 3 g 7→ g(x, r) d ∩ µ(x, r) := Λg(µ) (2.1) Z := lim g(x, r)φn ∗ µ(x)φn ∗ µ(x − r)φn ∗ µ(x − 2r) dx dr n → ∞
20 is well-defined.
1 Lemma 2.1.2. Suppose that the probability measure µ satisfies (a) and (b) for some β > 2
9−2β and α > 8 . Then the trivial progressions of step size 0 lie outside the support of the measure ∩3µ, or in other words
Z 3 1{0}(r) d ∩ µ = 0.
In Section 2.5 we show that the measure ∩3µ exists and that it gives no mass to the
degenerate progressions of step size zero; the main tools here are a Littlewood-Paley type
decomposition, the Fourier decay assumption b, and the use of uniformity norm-type esti-
mates to bound progression-counting multilinear functionals. In Section 2.2, we show that
3 the measure ∩ µA exists and that it is supported on the set of (x, r) for which x, x+r, x+2r is
∗ an arithmetic progression contained in Aµ. Section 2.4 contains the proof of Theorem 2.0.1. In Section 2.3, we obtain the proof of Lemma 2.3.2, which is the main ingredient in proving
the results of Section 2.2; the tools here are Bohr sets and the restriction estimate Theo-
rem 2.0.2.
2.2 Existence of the restricted measure on 3APs
In this section, for measurable sets A ⊂ T we demonstrate the existence of the measure
3 1 ∩ (µA) on T × T under the assumption that µ satisfies (a) and (b) with β > 2 and α suffi-
3 ciently close to 1, and we discuss the support properties of the measure ∩ (µA). Throughout
this section, we will be assuming the results of Section 2.5, specifically the truth of Theo-
rem 2.1.1 and Lemma 2.1.2.
1 Lemma 2.2.1. Suppose that µ is a probability measure on T which satisfies (b) for a β > 2
9−2β 3 and a for α > 8 , and that the set A ⊂ T is measurable. Then the Radon measure ∩ µA exists.
21 Proof. Consider the more general measure
3 ∗ d ∩ (µg)(x, r) := w − lim φn ∗ (g dµ)(x)φn ∗ (g dµ)(x − r)φn ∗ (g dµ)(x − 2r) dx dr
for a bounded measurable function g, provided the above weak∗ limit exists. First note that
∞ 3 we have boundedness of g ∈ L (µ) 7→ ∩ (µg) on its domain of definition, since for any
2 f ∈ C(T ) Z 3 | f d ∩ (µg)| Z = lim | f(x, r)φn ∗ (g dµ)(x)φn ∗ (g dµ)(x − r)φn ∗ (g dµ)(x − 2r) dx dr| n → ∞ Z 3 ≤ lim kfkL∞ kgkL∞ φn ∗ µ(x)φn ∗ µ(x − r)φn ∗ µ(x − 2r) dx dr n → ∞
3 3 =kfkL∞ kgkL∞ k ∩ µk
1 and for β > 2 and α close enough to 1 this last is finite by Theorem 2.1.1. Thus it suffices to show that the limit defining the operator
3 g 7→ ∩ (µg)
∞ L (µ) 7→M(T × T)
1 exists on a dense subset. Conditional on β > 2 and α being suffciently close to 1, by taking trigonometric polynomials as this dense subset we again use Theorem 2.1.1 to obtain
3 existence of ∩ (µg) for such g by applying linearity and the fact that for ξ ∈ Z, the measure dµ˜(x) := e2πiξx dµ(x) satisfies
α •| µ˜|(B(x, r)) . r
− β •| µb˜(ξ)| . (1 + |ξ|) 2 .
∞ 3 3 Thus for any g ∈ L , ∩ µg exists, and so in particular does ∩ µA.
3 We turn now to the support properties of the measure ∩ (µA).
22 Lemma 2.2.2. Suppose that µ is a probability measure on T satisfying (a) and (b) with
1 9−2β β > 2 and a value of α > 8 . Suppose further that the set A ⊂ T is measurable with µ(A) > 0. Then we have
3 ∗ ∩ µA(T × T \ {(x, r) ∈ T ×(T \{0}): x, x + r, x + 2r ∈ Aµ} = 0
∗ where we recall from Section 2.1 that Aµ denotes the points of positive density of the set A with respect to the measure µ.
Before proving Lemma 2.2.2, we need the following fact about the projection of ∩3µ to the x-coordinate.10 Interestingly, it seems inaccessible without recourse to restriction esimates
(a direct application of the U 2-techniques of Section 2.5 fails, for instance), leaving open the question of what shape, if any, an analogue would take in the case of longer progressions.
2 R 2 R 3 3 Let D µ denote the measure T f(x) dD µ(x) := f(x) d ∩ µ(x, r) provided that ∩ µ exists.
Proposition 2.2.3. Suppose that µ is a probability measure on T satisfying (a) and (b)
1 with β > 2 and a value of α > 1 − β/4. Suppose further that the set A ⊂ T is measurable
3 with µ(A) > 0, and that the measure ∩ µA exists. Then we have
2 D µA µA.
∞ R 2 R 2 1 Proof. For f ∈ L we will show that f dD µA . ( |f| dµA) 2 , which applied to indicator functions of sets implies absolute continuity.
By Dominated Convergence, it suffices to show the above for the dense subclass of f with fb ∈ `1(Z). Fix such an f.
10 k Q In the discrete case, it is known that the dual function D f := x 7→ E~u ι∈{0,1}k\{0} f(x − ι · u) of a K function f with L2 (m) bounds is in L∞(m)(m-denoting Haar measure), or indeed, that the dual function is continuous (see Lemma 2.6 of [31]). What we show may be thought of as a singular U 2 version of this fact.
23 Taking Fourier transforms,
2 Z Y X f(x) φn ∗ µA(x − ir) dr dx = φcn(ξ)φcn(2ξ)µcA(ξ)µcA(2ξ)fφ\n ∗ µA(ξ). (2.2) i=0 ξ∈Z We have by the triangle inequality that
φcn(ξ)φcn(2ξ)µcA(ξ)µcA(2ξ)fφ\n ∗ µA(ξ) ≤ µcA(ξ)µcA(2ξ) |fb| ∗ |µcA| (ξ) .
1 1 1 Choosing + + = 1 with pi > p◦ (when β > 0, this is possible by taking pi = 3 for p1 p2 p3
each i provided α > 1 − β/4 so that p◦ < 3)
X µcA(ξ)µcA(2ξ) |fb| ∗ |µcA| (ξ) ξ∈Z
≤kµcAk`p1 (Z)kµcAk`p2 (2 Z)k|fb| ∗ |µcA| k`p3 (Z)
1/2 1/2 3/2 . [µ(A)] [µ(A)] kfbk`1 kµcAk`p3 < [µ(A)] kfbk`1 < ∞
where in the first inequality of the last line, we have used the restriction estimate Theorem
2.0.2 and Young’s convolution inequality, and in the second inequality we have applied
Theorem 2.0.2 again.
This shows absolute convergence of the sum in (2.2).
But then evaluating (2.2) in the limit as n → ∞, again applying the restriction estimate
Theorem 2.0.2 we get
Z 2 X f dD µA = µA(ξ)µA(2ξ)fµdA(ξ) c c ξ∈Z
≤kµcAk`p1 (Z)kµcAk`p2 (2 Z)kfµdAk`p3 (Z) Z 1/2 1/2 1/2 2 . [µ(A)] [µ(A)] |f| dµA ≤ kfkL2(µ), as was to be shown.
24 1 3 Proof of Lemma 2.2.2. If β > 2 and α is close enough to 1, then by Lemma 2.2.1 ∩ µA exists. We show first that
3 ∗ c ∩ µA {(x, r) ∈ T × T : x, x + r, x + 2r ∈ Aµ} = 0.
Suppose that E ⊂ T × T is a collection of (x, r) which do not belong to the above set.
∗ Then for each (x, r) ∈ E, one of {x, x + r, x + 2r} ∈/ Aµ. We will decompose E into three sets corresponding to each of these three possibilities, E = E0 ∪ E00 ∪ E000.
0 ∗ We consider the case of E , the other cases being similar. Recall that µ(Aµ) = 0.
∗ 0 Let B = Aµ be the projection of E onto the x-axis.
3 c 2 c 2 c 3 c Then since ∩ µA(B ×T) = D µA(B ), D µA µA and µA(B ) = 0, we have ∩ µA(B ×
T) = 0. (The cases of E00 and of E000 require the analogous variants of Proposition 2.2.3, which we leave to the reader.)
3 ∗ Thus ∩ (µA) is in fact concentrated on the (x, r) such that x, x + r, x + 2r ∈ Aµ.
3 All that remains now is to show that ∩ (µA)(T ×{0}) = 0. But this is immediate from the obvious inequality
3 3 ∩ (µA) ≤ ∩ (µ)
and Lemma 2.1.2.
2.3 The Main Estimate : Proof of Lemma 2.3.2
Recall that for a given frequency set S ⊂ Z and radius 1 ≥ > 0, a Bohr set B = B(S, ) is defined by
2πiξx B = {x ∈ T : |e − 1| < for all ξ ∈ S}
and (see, e.g., Section 4.4 of [49]) that
|S| . |B|. (2.3)
25 We will need the following bound.
Lemma 2.3.1. Suppose that the measure µ on T satisfies (a) and b. Then if S ⊂ Z is a
−1 finite set of frequencies and B = |B| 1B is the normalized indicator of the Bohr set B of radius and frequency set S, one has
−1 −(1−α)|S|(1−α+ β ) kB ∗ B ∗ µkL∞ . 2
Proof. Set µ1 = B ∗ B ∗ µ.
Let N ∈ N be a large integer to be specified in a moment.
2 2 2 2 Since kBb k`1 = kBbk`2 = kBkL2 < ∞ and |µb1| ≤ |Bb| , we may apply Fourier inversion to decompose
X 2 2πiξx X 2 2πiξx |µ1(x)| ≤ |B ∗ B ∗ µ(x)| ≤ | Bb(ξ) µb(ξ)e | + | Bb(ξ) µb(ξ)e | |ξ|≤N |ξ|>N =I + II.
Taking ψ a mollifier with ψ ≈ N1B(0,N −1), ψb ≈ 1B(0,N), we have
X 2πiξx I ≈ ψbB\∗ B ∗ µ(ξ)e = |ψ ∗ B ∗ B ∗ µ| ≤ ψ ∗ |µ|
1−α .N (2.4)
1−α where in the last inequality, we have used kψ ∗ |µ|kL∞ . N which follows from the definition of ψ and the Hausdorff dimension assumption on the positive measure |µ| (for details see, e.g., Lemma 4.1 of [15]).
Further
X 2 X 2 − β II ≤ |Bb| |µb| ≤ ( |Bb| )kµbk`∞(B(0,N)c ) . N 2 kBkL2 |ξ|>N − β −|S| ≤N 2 (2.5)
26 by Plancherel together with (2.3).
Combining (2.4) and (2.5) we obtain
1−α − β −|S| kµ1kL∞ . N + N 2 .
This is minimized when both terms on the right are equal. Thus we let
β −1 log N = −|S| 1 − α + log , 2
at which point we find
−1 −(1−α)|S|(1−α+ β ) kµ1kL∞ . 2 . (2.6)
Taking α ↑ 1 in the above estimate, we see that if β > 0, kµ1kL∞ . 1 for α sufficiently
close to 1, and a closer inspection yields the bound lim supα → 1 kµ1kL∞ ≤ 2 in the limit as
max(CH ,CF ) ↓ 1.
Lemma 2.3.2. Suppose that the probability measure µ on T satisfies (a) and (b) and that
A ⊂ T satisfies
µ(A) > δ. (2.7)
1 Suppose further that β > 2 . Then there is a lower bound α0 on α ∈ (0, 1), depending on
β, δ, CH , and CF , and a number c = c(δ) > 0 such that if
α > α0 then
Λ3(µA) > c.
27 Proof. Let , η > 0. Let
S = {ξ ∈ Z : |1dAµ(ξ)| ≥ η} be the large spectrum of µA and let
B = {x : |e2πiξx − 1| < ∀ξ ∈ S} be the associated Bohr set of radius .
Set 1 B := 1 . |B| B
By the lower bound (2.3) on the size of a Bohr set
−|S| kBkL∞ . . (2.8)
We first show that |S| can be bounded only in terms of η. This is a consequence of the restriction estimate Theorem 2.0.2. We have for p in the range guaranteed by Theorem 2.0.2 that
X X Z |S|ηp ≤ |µ (ξ)|p ≤ |1 µ(ξ)|p |1 µ(ξ)|p k1 kp ≤ 1 cA dA . dA . A L2(µ) ξ∈S where in the second to last inequality we applied the restriction estimate, and in the inequal- ity before it we invoked Lemma 2.3.3 from the end of this section.
From the above we obtain that for appropriate p,
−p |S| . η . (2.9)
We will compare µA to B ∗ B ∗ µA =: µ1 and find that they contain approximately the same number of 3-term progressions. By assuming α sufficiently close to 1 in terms of δ in
3 2 order to bound kµ1kL∞ , a standard lower bound on ∩ (µ1)(T ) will then give us the result.
28 We next obtain a lower bound on Λ3(µ1). According to Lemma 2.3.1,
−1 −(1−α)|S|(1−α+ β ) kB ∗ B ∗ µAkL∞ ≤ kB ∗ B ∗ µkL∞ . 2 .
0 −1 So by choosing α ≥ α0 := β/(|S| ln( ) − 1), we find that kµ1kL∞ ≤ U for some constant
U ↓ 2 as α ↑ 1,CH ,CF ↓ 1. We suppose that we have done so.
By a version of the dense Roth’s theorem on T due in the discrete case to Varnavides [54] (see also Proposition 2.2 of [22] for a general statement or Lemma 6.1 of [9] for a derivation of the continuous inequality from the discrete), we have
kµ1kL1 Λ3(µ1) ≥ c (2.10) kµ1kL∞ where c(x) increases as x increases. Using the bound on kµ1kL∞ from (2.6) and that kµ1kL1 >
δ since B is normalized, we see that c ≥ c(δ/U) > 0 remains bounded below as a function of δ > 0 as ↓ 0 provided α is chosen as above.
We would then be done if we knew that |Λ3(µA) − Λ3(µ1)| < c, so we now bound this quantity, keeping in mind that we are free to set the parameters , η to be anything we want provided α depends on them as specified.
By writing Λ3(µA) and Λ3(µ1) on the Fourier side, we have
X 2 4 2 |Λ3(µA) − Λ3(µ1)| = µcA(ξ) µcA(−2ξ)(1 − Bb(ξ) Bb(−2ξ) ) (2.11) ξ∈Z X 2 4 2 X 2 4 2 ≤ |µcA(ξ)| |µcA(−2ξ)||1 − Bb(ξ) Bb(−2ξ) | + |µcA(ξ)| |µcA(−2ξ)||1 − Bb(ξ) Bb(−2ξ) | ξ∈S ξ∈ /S 4 2 2−p p+1 k1 − B(ξ) B(−2ξ) k ∞ |S| + kµ k kµ k . b b L (ξ∈S) cA `∞(Sc) cA `p+1 for any p > 0.
∞ 2−p In the last line, we have used Holder’s inequality after pulling out the ` norm of |µcA| . We wish to choose p so that the restriction estimate of Theorem 2.0.2 may be applied to
29 bound kµcAk`p+1 by a constant, but we also want the exponent 2 − p to be positive. This is possible provided that
2 > p ≥ p◦ − 1
where p◦ is the critical exponent in Theorem 2.0.2.
Thus we want p◦ < 3. Observing the formula for p◦ given in Theorem 2.0.2, it suffices to
0 assume that α > 1 − β/4. So set α0 = max(α0, 1 − β/4) and assume α > α0. Choosing such a p in (2.11), and applying Theorem 2.0.2 and the bound
4 2 2 k1 − Bb(ξ) Bb(−2ξ) k`∞(ξ∈S) .
(cf. [25], Lemma 6.7) we obtain
2 2−p |Λ3(µA) − Λ3(µ1)| . |S| + η . (2.12)
By choosing and η appropriately, we can finish the proof by showing that |Λ3(µ1) −
Λ3(µA)| is smaller than the lower bound c of Λ3(µ1). To do this, we examine the estimate
(2.12).
Combining (2.9) with (2.12), we have
2 2−p 2 −p 2−p |Λ3(µA) − Λ3(µ1)| ≤ C |S| + Cη ≤ C η + Cη .
Since we have fixed α depending on , η so that the bound in (2.10) is at least the constant c = c(δ/U) , we may choose first η, and then small enough that the right hand side above
1 is less than, say, 2 c, and thus
1 1 Λ (µ ) > Λ (µ ) − c > c 3 A 3 1 2 2 giving us the result.
30 Remark 2.3.1. A computation shows that in the above theorem we may take η = =
(c/2C)1/(2−p), so that we can set
βηp ln(2)/2 p ln(2)/2 α0 = 1 − = 1 − β =: 1 − γ0(c)β. 0 ln(−1) − ηp ln(2) ln(−1) − p ln(2)
With CH ,CF close enough to 1 that we obtain a sufficient upper bound on kµ1kL∞ (and
thus lower bound on c) that c ∼ c(δ/2) we then have that for any ρ > 0 we may take
α − ρ ≥ max(1 − β/4, 1 − γ(δ)β) where γ(δ) = γ0(c(δ/2)). Since the constant C can be taken to be > 1, c ≤ 1, p < 2, and γ is growing as a function of p, we can calculate that
γ(1) ≤ 1/6 when µ(A) = 1, in which case Lemma 2.3.2 remains valid when α > 1 − β/6
(provided CH ,CF sufficiently close to 1).
Lemma 2.3.3. Suppose that ν is a measure compactly supported in T d. Then for any p ∈ (1, ∞] X Z |ν(ξ)|p ≈ |ν(ξ)|p dξ. b d b d R ξ∈Z Proof. We first prove that for p < ∞
Z X p p |ν| . |ν(ξ)| dξ b d b d R ξ∈Z (if p = ∞, the corresponding inequality is immediate).
Let g be any Schwartz function equal to 1 on the support of ν and for which supp gb ⊂ 1 1 d [− 2 , 2 ] . Then since dν = g dν, Z |ν(ξ)| = |ν ∗ g(ξ)| = | ν(ξ − η)g(η) dη| b b b 1 1 d b b [− 2 , 2 ] Z ≤kgkL∞ |ν(ξ − η)| dη. b 1 1 d b [− 2 , 2 ] R So |ν(ξ)| 1 1 d |ν(ξ − η)| dη. Thus by Holder’s inequality b . [− 2 , 2 ] b Z !p Z Z X p X X p p |ν(ξ)| . |ν(ξ − η)| dη ≤ |ν(ξ − η)| dη = |ν(ξ)| dξ. b 1 1 d b 1 1 d b d b d d [− , ] d [− , ] R ξ∈Z ξ∈Z 2 2 ξ∈Z 2 2
31 Now we prove the converse inequality. Fix a Schwartz function h equal to 1 on supp(ν)
d such that supp(h) ⊂ B(0, 1). Then for any ξ ∈ R
X ˆ νb(ξ) = hνc(ξ) = νb(n)h(ξ − n) d n∈Z by (111) of [55]. By a variant of Young’s inequality, we then have for any 1 ≤ p ≤ ∞ that ˆ kνbkLp ≤ kνbk`p khkL1 . Indeed, Z X ˆ ˆ kνbkL1 = νb(n)h(ξ − n) ≤ kνbk`1 khkL1 d n∈Z X ˆ ˆ kνkL∞ = sup ν(n)h(ξ − n) . kνk`∞ khkL1 b d b b ξ∈R d n∈Z P where we have used the first part of the Lemma to bound d |hˆ(ξ − n)| by khˆk 1 , and n∈Z L interpolating between the two gives for any p ∈ [1, ∞]
kνbkLp . kνbk`p .
2.4 3APs in dense subsets : Proof of Theorem 2.0.1
The following is only a few lines removed from our main result, Theorem 2.0.1.
Proposition 2.4.1. Let δ > 0, and suppose that the probability measure µ on T satisfies
1 (b) with β > 2 and (a) with α sufficiently close to 1 depending on β, δ, and the implicit constants CH and CF in (a) and b.
∗ Then if A is any measureable set with µ(A) > δ, Aµ contains 3APs; indeed, µ gives positive measure to the set of x such that x, x + r, x + 2r is a non-trivial 3AP contained
∗ within Aµ.
3 Proof. By Lemma 2.2.1, the measure ∩ (µA) exists and by Lemma 2.2.2 it is concentrated
∗ 3 inside a set identifiable with the non-trivial 3APs contained in Aµ. By Lemma 2.3.2, ∩ (µA)
32 ∗ is not the trivial measure and hence Aµ must contain 3APs. By Proposition 2.2.3, µ gives positive measure to the collection of base points x for such 3APs.
∗ Proof of Theorem 2.0.1. Let L = Aµ \ A denote the collection of points of density of A not contained in A. Of course, µ(L) = 0. We may thus find an arbitrarily small open set around
0 0∗ L since µ is Radon; in particular, we may find a subset A ⊂ A such that Aµ ⊂ A and such that
µ(A0) > δ
remains valid.
0∗ Applying Proposition 2.4.1, we find that Aµ contains 3APs and that µ gives the collection of their basepoints positive measure. Consequently so too for A.
2.5 Existence of the measure ∩3µ : Proof of Theorem 2.1.1 and Lemma 2.1.2
The goal of this section is to prove Theorem 2.1.1 and Lemma 2.1.2.
We will need the following bound, which follows from the Hausdorff dimension condition
(a) via straightforward arguments (see Lemma 4.1 of [15]).
(1−α)n kφn ∗ µk`∞ . 2 . (2.13)
2 For bounded functions f : T → R and g : T → R, define Z Λg(f) := g(x, r)f(x)f(x − r)f(x − 2r) dx dr
and for three such functions f1, f2, f3
Z Λg(f1, f2, f3) := g(x, r)f1(x)f2(x − r)f3(x − 2r) dx dr.
33 Following Gowers [21], set also
∆1f(x; r) = f(x)f(x − r)
∆2f(x; r, s) = ∆1f(x; r)∆1f(x − s; r) = f(x)f(x − r)f(x − s)f(x − r − s)
∆1g(x, y; r) = g(x, y)g(x, y − r)
∆2g(x, y; r, s) = ∆1g(x, y; r)∆1g(x − s, y − s; r)
=g(x, y)g(x, y − r)g(x − s, y − s)g(x − s, y − r − s).
One easily verifies the following standard identity.
2 Lemma 2.5.1. Let f1 and f2 be bounded functions on T and g be bounded on T . Then
X Λg(f1) − Λg(f2) = Λg(h1, h2, h3) f1,f2 where the symbol P refers to a sum taken over the set f1,f2
{(f1 − f2, f1, f1), (f2, f1 − f2, f1), (f2, f2, f1 − f2)}.
We postpone the proofs of the following lemmas to the end of this section.
Lemma 2.5.2. Let f be a function on T which satisfies
ˆ − β |f(ξ)| ≤ CF (1 + |ξ|) 2 for all ξ ∈ Z (2.14)
supp(fˆ) ⊂ B(0, 2n+1) \ B(0, 2n) (2.15) for some β > 0.
Then for all η ∈ Z
−(2β−1)n |∆d2f(ξ; η)| . 2 and
supp ∆d2f(0; ·) ⊂ B(0, 2n+4).
34 2 Lemma 2.5.3. Let f, h1, h2 : T → R and g : T → R be bounded functions. Then Z
g(x, r)h1(x)h2(x − r)f(x − 2r) dx dr
Z Z 2 !1/4 1 1 ≤kh1kL∞ kh2kL∞ ∆ g(2x − r, x − r; −s)∆ f(r; 2s) dr dx ds .
Lemma 2.5.4. Let f be a function on T which satisfies
ˆ − β |f(ξ)| ≤ CF (1 + |ξ|) 2 for all ξ ∈ Z . (2.16)
supp(fˆ) ⊂ B(0, 2n+1) \ B(0, 2n) (2.17)
∞ 2 for some β > 0. Then for any h1, h2 ∈ L (T) and g : T → R, for all p ∈ [1, ∞]
− 1 (2β−1− 2 )n 1 4 p0 2 4 |Λg(h1, h2, f)| 2 kh1kL∞ kh2kL∞ k∆ g\(0, ξ; η)k 1 p . . `ξ`η
Lemma 2.5.5. Suppose that µ is a measure on T satisfying (b) with β > 0. Then for any
2 g : T → R and p ∈ [1, ∞], for all N > m ∈ N, N−1 ! − 1 (2β−1− 2 )n 1 X 4 p0 2 2 4 Λg(φN ∗ µ) − Λg(φm ∗ µ) . 2 max kφn+i ∗ µkL∞ k∆dg(0, ξ; η)k 1 p ι=0,1 `ξ`η n=m
1 P P p p where the mixed norm kG(ξ, η)k 1 p := |G(ξ, η)| . `ξ`η ξ η
Using the above lemmas we can complete the proof of Theorem 2.1.1 and Lemma 2.1.2.
Proof of Theorem 2.1.1. It suffices to show that the functional defined by the limit in (2.1)
2 exists for g in a dense subclass of C(T ) and is bounded on that subclass. We take as our subclass the collection of all trigonometric polynomials, and by linearity it suffices to check
existence and boundedness for a monomial of the form g(x, r) = e2πi(xξ0+rη0).
To show existence of the limit, it suffices to show that given > 0 there is some m such
that for all large N,
Λg(φN ∗ µ) − Λg(φm ∗ µ) < . (2.18)
35 To this end, we apply Lemma 2.5.5 with p = 1. By our assumption on g, it is easy to see
that
2 k∆dgk`1 = 1.
Inserting this and the bound (2.13) into Lemma 2.5.5 gives
N−1 N−1 X 2(1−α)(n+1) −(2β−1)n/4 X −n( 2β−1 −2(1−α)) Λg(φN ∗ µ) − Λg(φm ∗ µ) . 2 2 ≈ 2 4 (2.19) n=m n=m
2β−1 1 9−2β and the exponent −( 4 − 2(1 − α)) will be negative when β > 2 and α > 8 . Since the sum (2.19) is a geometric series and remains finite in the limit N ↑ ∞, we can
take m large enough to guarantee that (2.18) holds, as desired.
Thus Z lim g(x, r)φn ∗ µ(x)φn ∗ µ(x − r)φn ∗ µ(x − 2r) dx dr n → ∞
exists for g a trigonometric polynomial. In particular, the limit exists for for g ≡ 1.
Finally, we complete the proof by showing that the above limit is bounded where defined.
Indeed, setting
M := Λ1(|µ|)
we have
Z | lim g(x, r)φn ∗ µ(x)φn ∗ µ(x − r)φn ∗ µ(x − 2r) dx dr| n → ∞ Z ≤kgkL∞ lim φn ∗ |µ|(x)φn ∗ |µ|(x − r)φn ∗ |µ|(x − 2r) dx dr n → ∞
≤MkgkL∞
using that the limit defining Λg(|µ|) exists for the trigonometric polynomial g ≡ 1.
1 1 Proof of Lemma 2.1.2. Choose a compactly supported Schwartz function g with [− 2 , 2 ] ⊂
−1 supp g ⊂ [−1, 1]. Let > 0 and set g(r) = g((4 ) r).
It is not hard to see that ∆kg is rapidly decaying since g is Schwartz.
36 We have
3 ∩ µ({|r| ≤ }) ≤ |Λg (µ)|. (2.20)
We will show that
p−1 2 p Λg (µ) .
for some p > 1 to be specified later.
R 3 Since Λg(φn ∗ µ) → g d ∩ µ by Theorem 2.1.1, we can send N → ∞ and set m = 0 in
the bound from Lemma 2.5.5, which gives
2β−1− 2 ! ∞ p0 Z − 4 −2(1−α) n 1 3 X 2 4 | g d ∩ µ − Λg (φ0 ∗ µ)| . 2 k∆[g(0; ·)k`p n=0
1 9−2β for some p > 1. Since β > 2 , provided that α > 8 we may find p > 1 such that the sum in the right hand side above is finite. Fix such a p. We then have
Z 1 1 3 2 4 2 4 g d ∩ µ . Λg (φ0 ∗ µ) + k∆[g(0; ·)k`p = C +k∆[g(0; ·)k`p
2 We are now done, since for any function G on [0, 1]2 with Gˆ ∈ L1(R ), for p > 1 if
−1 G(~x) = G((4 ) ~x) then applying Lemma 2.3.3 and a change of variables,
2 p−1 ˆ ˆ p ˆ kGk`p . kGkLp = kGkLp
and G = ∆2g(0; ·) is such a function.
All that remains is to prove Lemmas 2.5.2, 2.5.3, 2.5.4, and 2.5.5.
Proof of Lemma 2.5.2. The (trivial) identity
1 ˆ ˆ ∆df(η1; τ) = f(τ)f(η1 + τ)
together with the Fourier decay condition (2.14) and the support condition (2.15) give
−β/2 −β/2 1 n n+1 |∆df(η1; τ)| . 1{2 ≤|τ|,|τ+η1|≤2 } (1 + |τ|) (1 + |η1 + τ|) (2.21)
37 One readily verifies that
2 X 1 1 ∆df(ξ; η) = ∆df(−η1; τ)∆df(−ξ − η1; τ − η2) τ∈Z which shows with (2.21) that η ∈ B(0, 2n+4), and it follows by Cauchy-Schwarz that
s s 2 X 1 2 X 1 2 |∆df(ξ; η)| ≤ |∆df(η1; τ)| |∆df(ξ + η1; τ)| . τ∈Z τ∈Z
Applying (2.21) to the above gives
1 ! 2 X −β −β 2 n n+1 |∆df(ξ; η)| . 1{2 ≤|τ|,|τ+η1|≤2 } (1 + |τ|) (1 + |η1 + τ|) τ∈Z 1 ! 2 X −β −β n n+1 + 1{2 ≤|τ|,|τ+ξ+η1|≤2 } (1 + |τ|) (1 + |ξ + η1 + τ|) τ∈Z −(2β−1)n .2 .
Proof of Lemma 2.5.3. We have
Z
g(x, r)h1(x)h2(x − r)f(x − 2r) dx dr
1 Z Z 2 2 ≤kh1kL∞ | g(x, r)h2(x − r)f(x − 2r) dr| dx
1 Z 2 =kh1kL∞ g(x, r)g(x, s + r)h2(x − r)h2(x − r − s)f(x − 2r)f(x − 2r − 2s) dr dx ds
1 Z 2 1 1 1 =kh1kL∞ 4 g(x, r; −s)4 h2(x − r; s)4 f(x − 2r; 2s) dr dx ds
Z Z 2 !1/4 1 1 ≤kh1kL∞ kh2kL∞ 4 g(x + r, r; −s)4 f(x − r; 2s) dr dx ds
Z Z 2 !1/4 1 1 =kh1kL∞ kh2kL∞ ∆ g(2x − r, x − r; −s)∆ f(r; 2s) dr dx ds
38 where in the last line we have applied the change of variables r 7→ −r + x.
Proof of Lemma 2.5.4. By Lemma 2.5.3,
Z
g(x, r)h1(x)h2(x − r)f(x − 2r) dx dr
Z Z 2 !1/4 1 1 ≤kh1kL∞ kh2kL∞ ∆ g(2x − r, x − r; −s)∆ f(r; 2s) dr dx ds
Z 1/4 2 2 =kh1kL∞ kh2kL∞ ∆ g(2x − r, x − r; −s)∆ f(r; 2s, t) dx dr ds dt .
Applying Plancherel gives
ZZ 2 1 1 2 ∆ g(x + r, x; s, t)∆ f(r; s, t) dr dt ds dx 2 2
X 2 2 =2 ∆dg(0, 2ξ; η1/2, η2)∆df(ξ; η) 2 ξ∈Z,η∈Z 1 p0 2 X 2 p0 k∆ g(0, ξ; η1/2, η2)k 1 p sup |∆ f(ξ; η)| (2.22) . d `ξ`η d ξ∈Z 2 η∈Z by Holder’s inequality.
By Lemma 2.5.2, this is bounded by
2 2 −(2β−1− 0 )n k∆ g(0, ξ; η1/2, η2)k 1 p 2 p d `ξ`η and two applications of Lemma 2.3.3 in order to move to the continuous context, change variables, and back, give the result.
39 Proof of Lemma 2.5.5. We decompose into telescoping series and invoke Lemma 2.5.1
N−1 X Λg(φN ∗ µ) − Λg(φm ∗ µ) = Λg(φn+1 ∗ µ) − Λg(φn ∗ µ) n=m N−1 X X = Λg(h1, h2, h3) φ ∗µ,φ ∗µ n=m n+1 n N−1 X X ≤ Λg(h1, h2, h3) . (2.23) φ ∗µ,φ ∗µ n=m n+1 n
We next calculate a bound on each individual term in the above sum. By a linear change
11 of variables, it suffices to suppose that h3 = φn+1 ∗ µ − φn ∗ µ := µn.
We can apply Lemma 2.5.4, which gives
2β−1− 2 p0 1 − 4 n 2 4 |Λg(h1, h2, µn)| 2 kh1kL∞ kh2kL∞ k∆dg(0, ξ; η)k 1 p . . `ξ`η
Inserting this into the right hand side of (2.23) gives the result.
2.6 Further Remarks
It is interesting to note that if the dependence of the Hausdorff dimension α in Theo- rem 2.0.1 on the measure δ of the set A were dropped, we would in fact have that µ-almost every point x was the starter of a three-term arithmetic progression contained in the support of µ. It is not a priori clear whether this dependence is necessary, or whether for some α,
β strictly less than 1, and some CH , CF , it might be true that whenever µ satisfied (a) and b that every A for which µ(A) > 0 must contain 3APs. In the discrete context, Green’s result applies to any dense subset of the primes, regardless of density, but the primes have
“dimension” 1 within the natural numbers. On the other hand Conlon, Fox, and Zhao [10]
11The change of variables may result in a different g. However, since the change of variable is linear, it will affect the bounds we obtain by at most a constant which doesn’t affect the conclusion of the theorem.
40 have developed a combinatorial approach strong enough to obtain a relative Szemer´edithe-
orem (kAPs for k ≥ 3) for subsets of pseudorandom sets of integers having “dimension”
less than 1 within the integers, and their result applies for any dense subset of a sufficiently
pseudorandom set, which gives reasonably strong evidence that such a result may be possible
in our context.
There is some further weak evidence the dependence of α on δ may be unnecessary, in
the form of a result of Conlon and Gowers [22] (see also [34] for the k = 3 case) that as
N ↑ ∞, with probability converging to 1, every random subset E of [N] large enough for the statement to not trivially fail has the property that all size δ|E| subsets of E contain kAPs. The natural cutoff for 3APs in this result of Conlon and Gowers is that the set
1 E ⊂ [N] satisfy |E| ≈ N 2 . Thus it is natural to conjecture that the appropriate bound on
1 the Hausdorff dimension α of a measure µ on T in order that it contain 3APs be α > 2 .
1 Though we have achieved results for any β > 2 , we require a far larger α. In a different direction, it is natural to ask whether the collection of r for which ther exist x such that x, x + r, x + 2r lies in supp(µ), or indeed in A for a set A of positive µ-measure, be large in Hausdorff dimension. We obtain results along these lines in Chapter 3.
In [9], we applied a notion of “higher order” Fourier dimension to extend the results of
[35] to longer progressions. It would be natural to extend the results of the present thesis to this setting, namely, to demonstrate that given a measure of (k−1)st order Fourier dimension
sufficiently close to 1, that any subset of sufficiently large measure must contain kAPs. The
obvious approach to such a result would be to adapt the methods of [10].
41 Chapter 3: On the number of 3APs in fractal subsets of Euclidean Space
This chapter is primarily work joint with Steven Senger. √ In the discrete setting, a set E ⊂ [N] := [0,...,N − 1] of ≈ N 1−1/ log N elements12
need not contain non-trival three term arithmetic progressions, as shown by Behrend in [2].
However if E contains ≈ N 1−(log log N)4/(log N)2 elements then it must contain (non-trivial)
3APs (Bloom 2016, [5]). If on the other hand the set possesses some pseudorandomness, such as enough Fourier decay that
4 4 X 1 3 k1E − δEk 2 = Eb(ξ) ≤ δ U 2 E ξ6=0∈ZN
then much smaller densities, δE, suffice for the set to contain non-trivial 3APs - such a set
− 1 E contains non-trivial 3APs once δE is greater than N 2 (this follows from Gowers’ proof,
[21]).
In this chapter we explore applications of techniques developed in relation to the Falconer
Distance Conjecture (a question on the distribution of the two-point configurations within a
set) to questions of 3AP counts in both the finite and fractal settings. We obtain largeness
results for the set of lengths of step-sizes of 3APs occurring within certain sets and obtain a
generalization (Theorem 3.0.8) ofLabaand Pramanik’s result in [35] with an assumption on
12 Given numbers A and B depending on some parameter or set of parameters t, we use the notation A . B denote the existence of an unspecified constant C, independent of t, for which the inequality A ≤ CB holds. We use the notation A ≈ B to denote the simultaneous inequalities A . B and B . A.
42 q the L norm of µb in place of a Fourier decay assumption, which relies heavily on techniques and tools from the discrete world (primarily Bohr sets and the Arithmetic Regularity Lemma
of Green and Tao ([27])).
3.0.1 Integer sets
The integer version of a question of Croot asks
Question 3.0.1 (Croot, question 7.6 of [12]). Given C, D > 0, is it true that for all suffi-
ciently large N a subset E of [N] satisfying
|E| 1 ≥ N logC N
and
Eb ≤ logD N `1 must contain a non-trivial three-term arithmetic progression?
We cannot answer this difficult question. Instead, we study the easier question of how much pseudorandomness as measured by the `q norm of Eb is necessary to guarantee 3APs, as q varies in [2, 3]. We prove
|E| 1 − 2 Theorem 3.0.1. Let E ⊂ [N] be a set with density δ = N N which satisfies a bound
Eb ≤ δM `q where ! T − 1 C q ≤ q(M, δ) := 2 + min 3 1 − , 1 T >1 T (2) δ T (2) δ T C2c M ln C2c M (see Definition 3.0.4 for the definition of the L2 Varnavides function c(2).) Then provided N
is sufficiently large depending on q and M the set E contains non-trivial 3APs.
43 Alternatively, for any q ∈ [2, 3] and δ > 0 there is an M = M(δ, q) ∈ [0, ∞] such that
for all sufficiently large N a set E ⊂ [N] of density ≥ δ with Eb < M contains a non- `q √ trivial 3AP, and M is monotonic decreasing in its second argument with M(δ, 3) = 3 2δ and
M(δ, 2) = ∞.
3.0.2 Fractal sets
At a rough level, an analog of Hausdorff dimension for subsets of integers E ⊂ [N] is
the smallest α for which |E| ≤ N α. Then rephrasing the discussion at the start of the
present chapter, informally Behrend’s result shows that even a Hausdorff dimension of 1 is
insufficient to gaurantee 3APs, but the corollary of Gowers’ proof shows that good enough
Fourier decay does yield results.
The above phrasing strongly suggests the question: What about in the fractal setting?
In 2008 Keleti constructed a full dimensional subset of [0, 1] containing no 3APs ([33]). In
2009,Labaand Pramanik ([35]) proved that a probability measure µ on [0, 1] satisfying
α µ(B(x, r)) ≤ CH r (3.1)
−β/2 |µb(ξ)| ≤ CF |ξ| (3.2)
for β > 2/3 and α sufficiently large depending on α and β must contain 3APs in its support.
In 2016, Carnovale showed that under these conditions, µ must give positive measure to
the set of starting points for 3APs in its support ([8]). This is the content of Chapter 2.
Also in 2016,Laba, Pramanik, and Henriot ([30]) showed, amongst other things, the 2009
Laba-Pramanik result holds when β is taken to be any positive number. Following these
results, we conjecture the following.
d Conjecture 3.0.2. Suppose that the compactly supported probability measure µ on R sat- isfies conditions (3.1) and (3.2) and that α > 2d/3. Then under appropriate quantitative
44 assumptions on the parameters involved, the set of r for which there exists an x and a u with
|u| = r such that x, x + u, x + 2u is a 3AP contained within supp µ has positive Lebesgue measure.
Definition 3.0.3. Let X be a compact abelian group with Haar measure m. We define the
∞ 3 multilinear functional Λ3 :(L (G)) → C via the formula ZZ Λ3(f0, f1, f2) = f0(x)f1(x − r)f2(x − 2r) dm(x) dm(r).
Given a single f ∈ L∞(G), we define
Λ3(f) = Λ3(f, f, f).
Definition 3.0.4. The Varnavides function c : (0, ∞) →(0, ∞) is given by
n o c(t) = inf Λ (f): f ≥ 0, kfk = 1, kfk ≤ t . 3 L1(T) L∞(T)
The L2 Varnavides function is the function c(2) : (0, ∞) →(0, ∞) given by
n o c(2)(t) = inf Λ (f): f ≥ 0, kfk ≥ t, kfk = 1 . 3 L1(T) L2(T)
The following is the celebrated Varnavides’ Theorem, from [54].
Theorem 3.0.5. [Varnavides’ Theorem.] Let 0 < δ < 1. Then there exists a constant,
N = N(δ) > 0 such that for any subset E ⊂ {0, 1,...,N}d, of size |E| > δN d, there are
2d & N three-term arithmetic progressions in E.
A discretization argument together with Theorem 3.0.5 gives that the Varnavides function c above does indeed map (0, ∞) to (0, ∞), while the same statement for the function c(2) will be given as Corollary 3.1.1 in Section 3.1. We summarize the equivalent functional form of the first inequality below.
45 d d Corollary 3.0.6. Let X ∈ T , ZN . Then for any M, δ > 0 there exists a constant
c = c(M, δ) > 0 with the following property. Let f : X →[0,M] satisfy kfkL1(X) ≥ δ. Then
Λ3(f) ≥ c.
Throughout, given an element r ∈ T, we will write
|r| = |r| = min (r mod 1, (1 − r) mod 1) , T
d and similarly for r = (r1, . . . , rd) ∈ T we define
1 d ! 2 X 2 |r| = |r| d = |r | . T i T i=1
To state our main result, we define relaxations of conditions (3.1) and (3.2) suited for
our purposes:
(a) Iα(|µ|) < ∞
− β d (b) ||cµ|(ξ)| ≤ CF (1 + |ξ|) 2 for all ξ ∈ Z ,
where we remind the reader of the Riesz potential,
ZZ −α X 2 −(d−α) Iα(µ) := |x − y| dµ(x)dµ(y) ≈ |µb(ξ)| |ξ| . d ξ∈Z We need to define the following notation for the set of r > 0 for which there exists at
least one 3AP of difference with length r in a given set,
43(E) := {r > 0 : x, x + u, x + 2u ∈ E, |u| = r}.
Note that studying the size of 43(supp µ) as a notion of “largeness” for the parameter
set
{r : there exists x such that x, x + r, x + 2r ∈ supp µ}
46 seems natural in light of Falconer’s Distance Conjecture, which conjectures that the
42(supp µ) := {r : there exists x and a u with |u| = r such that x, x + u ∈ supp(µ)} have positive Lebesgue measure whenever the Hausdorff dimension, α, of µ is greater than d/2 and d ≥ 2. A sharpness result of Falconer ([16]) shows that dimension d/2 is best possible, and Mattila ([39]) showed that a weaker Fourier decay condition than in (b) guarantees the truth of the conclusion of the Falconer Distance Conjecture for a measure µ when α + β > d
(an approach which has been central to progress since). As necessarily the set of step-sizes of 3APs within supp µ is contained within the set of differences between points in supp µ, optimal size estimates on the former under a given set of hypotheses is strictly harder than for the latter.
d Theorem 3.0.7. Let µ be a Radon probability measure on T satisfying (a) and (b). Then if
d − 1 > (4d − 3β)/2,
(2β + d − α)/β < 3,
47 13 and α is sufficiently close to d depending on the quantities CF , β, and Iα(µ), the bilinear
distance set 43(supp µ) has positive Lebesgue measure.
2 q We also prove the following L -type bound under an assumption that µb ∈ ` for some q ∈ [2, 3].
Theorem 3.0.8. Suppose that µ is a positive measure on [0, 1]d with
(c) µ([0, 1]d) ≥ δ
(d) kµbkLq ≤ M
for some δ > 0, M < ∞ and q ∈ [2, 3]. Then there is a number q0 = q0(M, δ) ∈ (2, 3] with ! T − 1 C1 q0(M, δ) = 2 + min 3 1 − , T >1 T (2) δ T (2) δ T C2c M ln C2c M (2) 2 for some constants C1 and C2 and where c is the L Varnavides function, such that if q ≤ q0 the set
3 ∆ (supp µ) = {r ∈ R : x, x + u, x + 2u ∈ supp µ, |u| = r} has positive s-dimensional Hausdorff measure for all s ≤ 1 satisfying
1 d(6 − 2q) s < + , 2 2q
13Throughout this chapter we will often use phrases of the form “α is sufficiently close to d depending on the parameters [β, C, X, Y , etc.]” in situations where satisfyingly explicit bounds are unavailable. In the case at hand, for instance, what can be said is the following. In Theorem 3.0.7, provided that 2 |µb(ξ)| α ≥ α0,Iα(µ) ≤ Iα0 , sup β ≤ CF d ξ∈Z |ξ| for some bounds 0 < α0,Iα0 < ∞, and p N(d−α0) c0 := c Iα0 CF 2
where c is the Varnavides function, it suffices that 12C I F α0 −(β0−(d−α0))N/2 2 < c0 1 − 2−(β0−2(d−α0))/2 for some N ∈ N. We will see in the proof that by increasing α0 towards d such an N may always be found provided that β0 > 2(d − α0).
48 and positive Lebesgue measure when d ≥ 2.
Further, when d ≥ 2 the natural measure on this set derived from µ (see the definition
of δ (µ) below) possesses a weighted L2 density with respect to the one-dimensional Lebesgue
measure.
Laba-Pramanik type results like Theorem 3.0.7 follow directly from Theorem 3.0.8, as
the following corollary makes precise. This is one indication of the naturality of the `q-norm
condition on µb.
Corollary 3.0.9. Suppose that the positive measure µ satisfies (a) and (b) where α is
sufficiently close to d depending on the parameters Iα(µ), CF , and β. Then supp µ contains
non-trivial 3APs, and indeed the set
3 ∆ (supp µ) = {r ∈ R : x, x + u, x + 2u ∈ supp µ, |u| = r}
has positive Lebesgue measure if d ≥ 2 and positive s-dimensional Hausdorff measure for any
1 d(2β+4α−4d) s ≤ 1 satisfying s < 2 + 4(β+d−α) .
β+d−α Proof. By Lemma 3.2.3, we have that for q1 = 2 β
q1 q1−2 −(q1−2)β/2 d−α q1−2 kµk q ≤ I (µ)C sup |ξ| |ξ| = I (µ)C , b ` 1 α F α F ξ
or
β d−α p β+d−α β+d−α kµk q ≤ I (µ) C . (3.3) b ` 1 α F
Consider the function q0(M) := q(M, 1) of Theorem 3.0.8. If we set
M = M(Iα(µ),CF , α, β)
to be the right-hand-side of (3.3), then we see that as Iα(µ) stays bounded above by a p parameter CH , and as d − α ↓ 0, M ↓ Iα(µ) and so q0(M) nears some q0 ∈ (2, 3).
49 β+d−α Meanwhile, q1 = 2 β ↓ 2, so if d − α is small enough, q1 ≤ q0 and the conclusion of Theorem 3.0.8 holds for the measure µ.
3.0.3 Finite fields
A very similar strategy to that of Section 3.2, following Iosevich and Rudnev’s adaptation,
[32], of Mattila’s paper [39], yields analogs of Theorem 3.0.7 in finite fields. This setting has
the advantage of exposing a clear dependence of the bounds in these theorems on the total
number of 3APs contained in the set under question (as made precise by Corollary 3.3.16
and Lemma 3.3.10), a feature obscured in the continuous setting.
d Let Fq denote the d-dimensional vector space over the finite field with q elements, and
d for a subset E ⊂ Fq , let
d 3AP (E) =: (x, r) ∈ Fq : x, x + r, x + 2r ∈ E .
Many expectations regarding 3APs in finite fields were recently shattered by the following
result ([11], [13]).
Theorem 3.0.10. [Croot-Lev-Pach-Ellenberg-Gijswijt] There is a γ ∈ (0, 1) such that if d
d γd is sufficiently large and E ⊂ Fq with |E| ≥ q then E contains a non-trivial 3AP.
Building on this work, Fox and Lovasz have recently proved some powerful results in [17],
which we discuss further below. One consequence of their work is the following corollary,
which gives us some quantitative bounds on the number of 3APs present in a sufficiently
d large subset of Fq .
d αd Corollary 3.0.11. There is a cq > 0 such that if E ⊂ Fq with |E| > q then for d sufficiently large
|3AP (E)| ≥ q−cq(1−α)dq2d.
50 In line with this, we have the following bound on the number of distances arising from
the step-sizes in 3AP (E).
d αd Theorem 3.0.12. Let E ⊂ Fq with |E| ≥ q . Then if c α > q 1 + cq
where cq is the constant from Corollary 3.0.11 and d is sufficiently large
3 4 (E) & q.
In a private communication, Hans Parshall pointed out that we cannot hope to replace
& with = in Theorem 3.0.12 using size conditions alone. This is a consequence Lemma 17
d in [38], which shows that there exist dense subsets of Fq not containining a given isometric copy of a fixed“non-spherical” configuration.
It is well-known that if the set E is Fourier-pseudorandom, then E contains the “ex-
pected” number of three-term arithmetic progressions. Following Iosevich and Rudnev in
[32], we define the following class of sets:
d Definition 3.0.13. Say that E is a Salem set if for all non-zero m ∈ Fq
−d 1 2 Eb(m) . q |E| .
d d/2 Theorem 3.0.14. If E ⊂ Fq is a Salem set and |E| & q then
3 4 (E) & q.
3.1 Integer results
d 2 A non-zero positive measure µ on T with µb ∈ ` must necessarily contain within its sup- port non-trivial three-term arithmetic progressions: the qualitative statement follows from
51 the observation that since µ ∈ L2, supp µ is a set of positive Lebesgue measure, together with the statement that such sets contain non-trivial 3APs (and indeed affine images of any
finite point configuration), a consequence of the Lebesgue density theorem and the pigeon- hole principle. The quantitative statement is the following consequence of the arithmetic regularity lemma, whose discussion we postpone until Subsection 3.1.1.
d d Corollary 3.1.1. Suppose that X = T or ZN , and that f : X →[0, ∞) satisfies kfkL1 ≥
(2) δ, kfkL2 < M < ∞. Then there exists a number c = c (M, δ) such that
Λ3(f) ≥ c.
3 On the other hand, if µb ∈ ` with sufficiently small norm, then again lower bounds on
Λ3(µ) are available.
d d Lemma 3.1.2. Suppose that µ is a positive measure on T or ZN with mass kµk ≥ δ and √ 3 kµbk`3 ≤ 1 + cδ. Then 3 Λ3(µ) ≥ (1 − c)δ .
Proof. We have
√ 3 3 3 3 X 3 3 X 3 1 + cδ ≥ kµbk`3 = µb(0) + |µb(ξ)| ≥ δ + |µb(ξ)| ξ6=0 ξ6=0 so √ 3 kµbk`3(ξ6=0) ≤ cδ.
Further, note by Holder that !2/3 !1/3 X 2 X 3 X 3 3 µb(ξ) µb(−2ξ) ≤ |µb(ξ)| |µb(2ξ)| ≤ kµbk`3(ξ6=0) . ξ6=0 ξ6=0 ξ6=0 Then
3 X 2 3 X 2 3 Λ3(µ) − δ = µb(ξ) µb(−2ξ) − δ = µb(ξ) µb(−2ξ) ≤ cδ . ξ ξ6=0
52 d In summary: If µ is a positive, non-zero measure on T and kµbk`p = M, then if M is
sufficiently small and p = 3 then Λ3(µ) > 0, while if p = 2 then no bound whatsoever on
M is required to conclude that Λ3(µ) > 0. It is natural to ask whether one can interpolate
between these results for p ∈ (2, 3). Indeed we can, as Proposition 3.1.3 below shows.
Proposition 3.1.3 (Mass lemma from `ˆp bounds). Let δ, M > 0. Then there exists a
q = q(M, δ) and a number c = c(q)(M, δ) > 0 with ! T − 1 C q(M, δ) = 2 + min 3 1 − , 1 T >1 T (2) δ T (2) δ T C2c M ln C2c M (2) δ 2 for some constants C1,C2 and where c ( M ) is the L Varnavides function, such that the following holds. Suppose that µ is a positive measure on [0, 1]d or [N]d with kµk ≥ δ and kµbk`q ≤ M. Then (q) Λ3(µ) ≥ c (M, δ).
Equivalently, for any q ∈ [2, 3] there is an M = M(q) and a c = c(q)(M, δ) > 0 such that
if kµk ≥ δ and kµbk`q ≤ M then Λ3(µ) ≥ c, with M(q) ↑ ∞ as q ↓ 2.
Theorem 3.0.1 follows as a corollary.
Proof of Theorem 3.0.1. Set µ = 1E . Then kµk ≤ M and kµk = 1, so by Proposition 3.1.3, δ b `q
(q) Λ3(µ) ≥ c (M, 1)
independent of N.
On the other hand, the trivial 3APs within E contribute
2 1 X Y 1 |E| N µ(x)µ(x)µ(x) = = = δ−2N −1 = o (1) N 2 N 2 δ3 |E|2 N x∈[N] i=0
− 1 since δ N 2 .
For large enough N, this will be less than c(q)(M, 1), guaranteeing the existence of a non-trivial 3AP within E.
53 The proof of Proposition 3.1.3 will use the following lemma, whose proof we delay until
the end of the present section.
Lemma 3.1.4. For all ∈ (0, 1) and C > 1 there exists a q = q(, C) ∈ (2, 3) such that for
d all borel measures f with fˆ ∈ `q(Z ) there exists a decomposition
f = g + h such that
• g ≥ 0,
•k gkL1 = kfk,
ˆ ˆ • gˆ , h . f ,
ˆ • gˆ `2 . C f `q ,
and
ˆ • h `3 . .
Further, if we specify an upper bound M for the `q norm of fˆ, we may take ! T − 1 2 ln C q = 2 + min 3 1 − , . (3.4) T >1 T M T M T ln Proof of Proposition 3.1.3. Set f = µ.
Fix a number > 0 depending on M and δ to be specified later.
We use Lemma 3.1.4 with C ≈ 4 to find a q = q(M, ) ∈ (2, 3) and a decomposition
f = g + h
54 where ! T − 1 C0 q = 2 + min 3 1 − , (3.5) T >1 T M T M T ln g ≥ 0
ˆ ˆ |gˆ| , h . f ,
ˆ kgˆk`2 ≤ 4 f ≤ 4M, `q
ˆ ˆ kgˆk`3 ≤ f ≤ f ≤ 4M, `3 `q ˆ h . . (3.6) `3
Rescaling g by (4M)−1, Corollary 3.1.1 gives
δ Λ (g) ≥ (4M)3 c(2) . (3.7) 3 4M
We have
X ˆ ˆ ˆ ˆ ˆ ˆ |Λ3(f) − Λ3(g)| . |Λ3(f0, f1, f − g)| ≤ f0(ξ) h(ξ) f1(−2ξ) ≤ h f0 f1 `3 `3 `3 where f0, f1 ∈ {f, g} are chosen to maximize the estimate on the right-hand side of the first inequality. Thus applying (3.6)
2 |Λ3(f) − Λ3(g)| . M whence using (3.7) we have
3 (2) 2 Λ3(f) ≥ 4M c (δ/4M) − O M .
So choosing
= C004Mc(2) (δ/4M) for sufficiently small constant C0 gives the result with
(q) 3 (2) c (M, δ) & M c (δ/4M) .
55 1 Proof of Lemma 3.1.4. Fix λ ∈ (0, kfkL1 ) and a number η ∈ (0, 2 ) to be specified later. Let q ∈ (2, 3) be a number to be specified later, and suppose that
fˆ ≤ M. `q
Set
Eλ = {|f| ≥ λ} .
It would be natural to letg ˆ be the restriction of fˆ to the set where it is large, but it seems unlikely that the resulting g would be a positive function. Instead, a variant of this
approach replacing 1Eλ by the Fourier transform of the corresponding Bohr set works. Let
2πix·ξ B = B (Eλ, η) = x : e − 1 ≤ η∀ξ ∈ Eλ be the Bohr set with frequency set Eλ and radus η, and let
φ = 1B/|B|.
We recall the following facts about Bohr sets
|E | (i) |B| & η λ (see, e.g, Lemma 4.4 in [49])
p−1 −|Eλ| −|Eλ| p and consequently kφkL∞ . η , and more generally kφkLp . η , which com- bined with the Hausdorff-Young inequality gives
−|E | 1 φb . η λ p `p
for p ≥ 2.
2 (ii) 1 − φb(ξ) . η for all ξ ∈ Eλ (see, e.g., the proof of Lemma 6.7 in [25]).
Set
g = φ ∗ f,
56 h = f − g.
We have q −q ˆ |Eλ| ≤ λ f `q so that by (i) for p ≥ 2
q q −λ−qkfˆk 1 −( M ) 1 . φb . η `q p ≤ η λ p (3.8) `p
This gives by Holder’s inequality
1/2 1/2 2 2 2 q X − M q−2 ˆ ˆ ˆ ( λ ) 2 kgˆk`2 = φb f f ≤ f φb 2 . Mη . (3.9) `q ` q−2
Meanwhile by (ii), the triangle inequality, and that φb ≤ 1
3 X 3 3 hˆ = 1 − φb(ξ) fˆ(ξ) `3 ξ 3 3 3 3 X ˆ X ˆ = 1 − φb(ξ) f(ξ) + 1 − φb(ξ) f(ξ) c Eλ Eλ 3 3 6 X ˆ X ˆ .η f(ξ) + f(ξ) c Eλ Eλ 3 X q ≤η6 fˆ + λ3−q fˆ(ξ) q ` c Eλ
c ˆ where in the last line we have used that q < 3 and that on Eλ, f ≤ λ. Thus
q 6 3−q 3−q ˆ ˆ 3 2 q/3 h . η 3 M + λ 3 f ≤ η M + λ 3 M . (3.10) `3 `q
Setting
λ = Mη
(3.9) and (3.10) become
δ q q−2 (3−q)q q−2 −(( λ ) 2 ) −(η 2 ) kgˆk`2 . η M = η M (3.11)
3−q 3−q ˆ 2 3 q/3 h . η M + (ηM) M . η 3 M, (3.12) `3
57 where in (3.12) we have used the bounds η < 1 and q ≥ 1.
Using these bounds we now must check that g and h fulfill the requirements of this lemma
for some choice of q ∈ (2, 3). Since g = φ ∗ f, obviously g ≥ 0 and kgkL1 = kfkL1 , and
ˆ ˆ h . f , so we need to check that for an appropriate choice of constant η ∈ (0, 1) and q ∈ (2, 3) the desired norm bounds are attained.
Solving
(3.12) ≤
gives 3 3−q η ≤ M So we may take T η = M
3 T −1 for a choice of T ≥ 3−q , or q ≤ 3 1 − T . Solving
(3.11) ≤ CM
for q yields that we must (and that it is sufficient to) take ! T − 1 2 ln C T − 1 2 ln C q ≤ 2 + min 3 1 − , = 2 + min 3 1 − , , T η−1 ln η−1 T −T −T M ln M where the minimum is taken over all T , showing that we may always take q > 2. This
completes the proof.
3.1.1 3AP counts of L2 functions
Define |f(x) − f(y)| kfkLip = kfkL∞ + sup x6=y |x − y| and Z Z 2 !1/4
kfk 2 = f(x)f(x − r) dx dr . U
58 d d Given numbers M and N, say that θ ∈ T is (M,N)-irrational if for all q ∈ Z with kqk ≤ M, kq · θk ≥ M . `1 T N An analysis of the proof of Theorem 1.2 of [27] reveals that only L2-bounds on the function f are needed, and further that the same proof works over [N]d, so that a slight modification yields the following.
2 d Lemma 3.1.5 (L Regularity Lemma in [N] ). Suppose that f :[N] →[0, ∞) with kfkL1 ≥
2 δ > 0, f ∈ L with kfkL2 ≤ 1. Let F : N → N be an increasing function and let > 0. Then there exists an M ∈ N and a decomposition f = fstr + fU 2 + fL2 such that
kfU 2 kU 2 ≤ F(M),
kfL2 kL2 ≤ , and
fstr(n) = F (n/N, n mod q, θn) where
d F : [0, 1] × Z /q Z × T → [0, 1],
d q, d, kF kLip ≤ M, and θ ∈ T is (F(M),N)-irrational.
We find it convenient to use Herglotz’s Theorem in the proof of Corollary 3.1.1.
Definition 3.1.6. A continuous function ψ on a locally compact abelian group G is positive- definite if for all finite complex-valued sequences an
X ψ(i − j)aiaj ≥ 0. i,j Theorem 3.1.7. [Herglotz’s Theorem] Let f be a positive finite Borel measure on a locally compact abelian group. Then fˆ is positive-definite on the Pontryagin dual Gˆ of G. Con- verseley, if ψ is a positive-definite function on a locally compact abelian group, then ψ is the
Fourier transform of a positive finite Borel measure on the dual group.
59 Proof of Corollary 3.1.1. We deal with the case that X = [N]d first. Let > 0. Use the
Regularity Lemma to write f = fstr + fU 2 + fL2 , where
kfU 2 kU 2 ≤ F(M),
kfL2 kL2 ≤ ,
and
fstr(n) = F (n/N, n mod q, θn)
where
d F : [0, 1] × Z /q Z × T → [0, 1],
d q, d, kF kLip ≤ M, and θ ∈ T is (F(M),N)-irrational. Then by following exactly the proof in [27] of their Theorem 6.1 (Szemeredi’s Theorem),
replacing their use of the Regularity Lemma for 1-bounded functions by Lemma 3.1.5, one
obtains the claim.
d Now if instead X = [0, 1]d, we embed [0, 1]d ⊂ [−1, 2]d ,→ T , noting that this embedding preserves the 3AP count (abusing notation, we continue to write f for the resulting function
d defined on T ). Let > 0. Let N be such that
1 2
X ˆ f(ξ) ≤ d ξ∈Z ,|ξ|>N/100
and write fN = φN ∗ f were φ is a mollifier satisfying φcN |B(0,N/2) ≡ 1, supp φcN ⊂ B(0,N), so that kf − fN kL2 ≤ . We have that
3 |Λ3(f) − Λ3(fN )| .
60 since
X ˆ 2 ˆ ˆ d ˆ |Λ3(f) − Λ3(fN )| = f(ξ) f(−2ξ) − fN (ξ) fN (−2ξ) d ξ∈Z
X ˆ 2 ˆ ˆ d ˆ X ˆ 2 ˆ ˆ d ˆ ≤ f(ξ) f(−2ξ) − fN (ξ) fN (−2ξ) + f(ξ) f(−2ξ) − fN (ξ) fN (−2ξ)
|ξ|≤N/2 |ξ|>N/2
3 3 X ˆ 2 ˆ ˆ ˆ 3 ≤0 + 2 f(ξ) f(−2ξ) . f ≤ f ≤ . `3(|ξ|>N/2) `2(|ξ|>N/2) |ξ|>N/2
−1 d Let p ∈ (10N, 20N) be a prime and g denote the function on (p Zp) given by
X ˆ 2πiξ·x g(x) = fN (ξ)e . d ξ∈B(0,N)⊂(Z /p Z)
−1 d d If we embed (p Zp) within T in the natural way, g(x) = fN (x); this identification may be sidestepped, however - note that
ˆ 1.ˆg is positive-definite since by Herglotz’s Theorem fN is,
2. thus g ≥ 0 again by Herglotz’s Theorem,
ˆ 3. kgkL1 =g ˆ(0) = fN (0) = kfkL1 ,
4. kgkL2 = kfN kL2 ≤ kfkL2 ≤ 1 by Plancherel’s theorem,
5.Λ 3(g) = Λ3(fN ) by the identity
X 2 Λ3(g) = gˆ(ξ) gˆ(−2ξ) ξ ˆ applied first to g, then to fN , and using thatg ˆ = fN under the identification Zp =
p−1 p−1 − 2 ,..., 0,..., 2 ⊂ Z.
By the conclusion of this lemma over [p]d, we therefore obtain that
3 (2) Λ3(f) + O( ) = Λ3(fN ) = Λ3(g) ≥ c (kfkL1 ) .
61 3.2 Fractal results
d Let µ be a Radon measure supported in T or [0, 1]d. Whenever this integral exists, define
ZZ σ (µ)(ρ) := µb(2η)µb(η − ρθ)µb(η + ρθ) dσ(θ) dη, where the σ(θ) inside the integral is the Lebesgue measure on the unit sphere. Also, define
ZZ
ς (µ)(ρ) = µb(2η)µb(η − ρθ)µb(η + ρθ) dσ(θ) dη, and ZZ Λ3(f) := f(x)f(x − r)f(x − 2r)dxdr, when these integrals exist.
In what follows, we will freely use (φn)n∈N as an approximate identity with supp φcn ⊂
B(0, 2n). Finally, let δ (µ) be the measure on R defined as
Z Z Y g(r) dδ (µ)(r) := lim g(|u|) φn ∗ µ(x − iu) dx du n → ∞ i∈[3] provided that the right hand side above defines a continuous linear functional on C(R) (as
d it does, e.g., if µ has a density function f ∈ C(R )).
d We note that the measures on [0, 1]d may be identified with (a subset of) those on T via
d the embedding given by first mapping [0, 1]d to the unit cube in R , followed by identifying
d d the cube [−2, 2] with T in the natural way. This identification preserves the energies Iα(µ)
p and the L norms kµbkLp up to multiplicative constants, and also preserves the 3AP counts as measured by Λ3(µ) and the (up to rescaling) set of lengths of 3APs as measured by δ (µ)
d when the definition of δ (µ) is extended to T in the obvious way. In this section we will perform this identification without further comment.
62 3.2.1 Proof of Theorem 3.0.7
The proof of Theorem 3.0.7 follows from Lemma 3.2.1 (Support), Lemma 3.2.6 (Mass), and Lemma 3.2.10 (Non-singularity), below.
Support Lemma
We state the following without proof.
d Lemma 3.2.1. [Support lemma.] Let µ be a probability measure on T and suppose that δ (µ) exists and has finite total mass. Then
supp(δ (µ)) ⊂ {r > 0 : x, x + u, x + 2u ∈ supp(µ), |u| = r} .
Mass Lemma
In order to prove the Mass Lemma, we need the following technical results.
d Lemma 3.2.2. Let µ be a probability measure on [0, 1] satisfying Iα(µ) < ∞, and (φn)n∈N
n an approximate identity with supp φcn ∈ B(0, 2 ). Then for all n
p n(d−α)/2 kφn ∗ µkL∞ ≤ Iα(µ)2 .
1 Proof. Since φn ∗ µ has an ` Fourier transform, we have by the triangle inequality and
Cauchy-Schwarz
X 2πiξ·x |φn ∗ µ(x)| = φcn(ξ)µb(ξ)e
X −(d−α)/2 (d−α)/2 ≤ |µb(ξ)| |ξ| |ξ| |ξ|≤2n p n(d−α)/2 ≤ Iα(µ)2 .
d Lemma 3.2.3. Suppose that the measure µ on T satisfies (b) for β > 0. Then
q q−2 d−α kµbk`q ≤ Iα(µ) sup |µb(ξ)| |ξ| . ξ
63 In particular, µ ∈ `q for q = 2 + 2 d−α = 2 β+d−α . b β β
Proof. We have
X q |µb| d ξ∈Z X 2 −(d−α) q−2 d−α = |µb| |ξ| |µb| |ξ| d ξ∈Z q−2 d−α ≤Iα(µ) sup |µb(ξ)| |ξ| ξ
From here forward, we define J := J(d−2)/2, the Bessel function of order (d − 2)/2. We
record some well-known Bessel function estimates in the following lemma, which may be
found in [39] (2.1),(2.2), and the first displayed formula on page 216.
Lemma 3.2.4. Consider the Bessel function J = J(d−2)/2.
1. For all ξ ∈ R
−1/2 Jb(ξ) ≤ |ξ| .
2. For all ξ ∈ R
(d−2)/2 Jb(ξ) ≤ |ξ| .
d 3. There is a constant C so that the following holds. Let g ∈ C(R) and G ∈ C(R ) satisfy G(u) = g(|u|). Then
Z ∞ −(d−2)/2 Gb(ξ) = C |ξ| rd/2J(r |ξ|)g(r) dr. 0
d Lemma 3.2.5. For any f, g ∈ C(R ) with gˆ ∈ L1 and supp(g) ⊂ [0, ∞). Z Z Z g(r) dδ (f)(r) = O(1) g(r)rd/2 ρd/2J(rρ)σ (f)(ρ) dρ dr.
Further, if fb ∈ L1, Z δ (f)(r) = O(1)rd/2 ρd/2J(rρ)σ (f)(ρ) dρ.
64 Proof. Let G(u) = g(|u|). Then using Lemma 3.2.4
Z Z Y g dδ (f) = G(u) f(x − ir) dx dr i∈[3] Z = Gb(ξ)fˆ(η + ξ)fˆ(η − ξ)fˆ(2η) dη dξ Z Z ∞ =O(1) |ξ|−(d−2)/2 rd/2J(r |ξ|)g(r) dr fˆ(η + ξ)fˆ(η − ξ)fˆ(2η) dη dξ 0 Z ∞ Z =O(1) g(r) |r|d/2 |ξ|−(d−2)/2 J(r |ξ|)fˆ(η + ξ)fˆ(η − ξ)fˆ(2η) dη dξ dr 0 Z ∞ ZZZ =O(1) g(r) |r|d/2 sd−1s−(d−2)/2J(rs)fˆ(η + sθ)fˆ(η − sθ)fˆ(2η) dη dσ(θ) ds dr 0 Z ∞ Z =O(1) g(r) |r|d/2 sd/2J(rs)σ (f)(s) ds dr. 0
Under the hypothesis that fb ∈ L1, it is immediate that σ (f) is bounded, and thus that the
above gives the desired pointwise estimate.
Lemma 3.2.6. [Mass lemma.] Let µ be a probability measure satisfying (a) and (b) and suppose that the measure δ (µ) exists. Then if
(2β + d − α)/β < 3
and α is sufficiently close to d depending on the quantities CF , β, and Iα(µ), the measure
δ (µ) has positive mass.
Proof. Let (φn) be an approximate identity satisfying the hypotheses of Lemma 3.2.2, so that for any n ∈ N
p n(d−α) kφn ∗ µkL∞ ≤ Iα(µ)2
and kφn ∗ µkL1 = 1.
Then by Varnavides’ Theorem there is a monotonic decreasing c = c (kφn ∗ µkL∞ ) ≥
p n(d−α) c( Iα(µ)2 ) (so that c is in particular independent of any other properties of the mea-
sure µ), such that
Λ3(φn ∗ µ) ≥ c.
65 Now fix N ∈ N. Consider the difference
|kδ (µ)k − Λ3(φN ∗ µ)| = lim |Λ3(φn ∗ µ) − Λ3(φN ∗ µ)| n → ∞ X ≤ |Λ3(φn+1 ∗ µ) − Λ3(φn ∗ µ)| n≥N
Let
Fn = {(f, f, f − g) , (f, f − g, g) , (f − g, g, g)}
where f = φn+1 ∗ µ and g = φn ∗ µ, so that each triple in Fn has exactly one entry equal to
(φ − φ ) ∗ µ. Using the identity Λ (φ ∗ µ) − Λ (φ ∗ µ) = P Λ (f , f , f ), n+1 n 3 n+1 3 n (f0,f1,f2)∈Fn 3 0 1 2 Q Q P Q Q which follows from the identity ai − bi = j i
|Λ3(φn+1 ∗ µ) − Λ3(φn ∗ µ)|
≤12 sup |Λ3(f0, f1, f2)| . f~∈Fn
Taking the Fourier transform, applying Holder’s inequality and using the support properties of φdn+1 − φcn µb (which is equal to one of the fi), we have
X |Λ3(f0, f1, f2)| = fb0(η)fb1(η)fb2(2η) d η∈Z 2 2 1 Y ˆ Y 1/3 d−α 3 ≤ fi ≤ Iα(µ) µ(ξ) |ξ| `3(B(0,2n+1)\B(0,2n−1)) b `∞(|ξ|≈2n) i=0 i=0 where in the last line we have used Lemma 3.2.3. Together with (b) this gives
− β−2(d−α) n |Λ3(f0, f1, f2)| ≤ CF Iα(µ)2 2 .
This yields that
−(β−(d−α))N/2 −(β−2(d−α))/2 |kδ (µ)k − Λ3(φN ∗ µ)| ≤ 12CF Iα(µ)2 / 1 − 2 . (3.13)
66 Fix a lower bound α0 for α and an upper bound I0 for Iα(µ) such that the constant √ n(d−α0) c0 := c( I02 ) coming from Varnavides Theorem is positive. Fix also a lower bound
β0 for β. We may suppose that α0 and β0 satisfy the condition (2β0 + d − α0)/β0 < 3.
Let N be large enough that the expression
−(β0−(d−α0))N/2 −(β0−2(d−α0))/2 12CF I02 / 1 − 2 , which upper-bounds the right-hand-side of (3.13) is strictly less than c0, which is possible since β0 > 2(d − α0). Choose α > α0 sufficiently close to d that the Varnavides bound c satisfies
p n(d−α) c( Iα(µ)2 ) ≥ c0, which is possible since c is decreasing in its argument. Putting these together, we obtain from the reverse triangle inequality that
kδ (µ)k = |Λ3(φN ∗ µ) − (Λ3(φN ∗ µ) − kδ (µ)k)|
≥Λ3(φN ∗ µ) − |kδ (µ)k − Λ3(φN ∗ µ)| > c0 − c0 = 0 showing that kδ (µ)k > 0.
β Remark 3.2.1. From the proof we see that we may take α > d − 2 provided that for some
N ∈ N and c = c(·) the lower bound function coming from Varnavides Theorem 3.0.5 the strict inequality 12C I p N(d−α0) F α0 −(β0−(d−α0))N/2 c Iα CF 2 > 2 0 (1 − 2−(β0−2(d−α0))/2) is satisfied for the bounds
α ≥ α0,
Iα(µ) ≤ Iα0 , 2 |µb(ξ)| sup β ≤ CF . d ξ∈Z |ξ| 67 Non-singularity Lemma
We first establish existence of the measure δ (µ).
d Lemma 3.2.7. [Existence lemma.] Let µ be a Radon probability measure on T . Suppose
q that µb ∈ ` for some q ≤ 3. Then the measure δ (µ) exists. d+β−α In particular, if µ satisfies (a) and (b) then this holds when q = 2 β provided this value is less than or equal to 3.
Proof. Let (φn) be an approximate identity. That δ (µ) exists follows from: the absolute
convergence of the sums 2 X φ\n ∗ µ(ξ) φ\n ∗ µ(−2ξ) d ξ∈Z and more generally for any fixed η0, η1 of
X φ\n ∗ µ(ξ − η0) φ\n ∗ µ(ξ − η1) φ\n ∗ µ(−2ξ) , d ξ∈Z the observation that this convergence guarantees the existence of the limits
2 ZZ Y lim f(x, r) φn ∗ µ(x − ir) dx dr (3.14) n → ∞ i=0
for any trigonometric monomial f(x, r) = e2πi(x·(η0−η1)+r·η1), hence for any trigonometric
polynomial f, hence for any continuous function f, and hence the existence of a measure
d d d ∩3µ on T × T represented by the action on C(T ) given by (3.14), and finally by the
observation that for a function g on R,
Z ∞ Z g(r) δ (µ)(r) = g(|u|) d ∩3 µ(x, u). 0
The following technical estimates will be used in proving the Non-singularity lemma.
68 Lemma 3.2.8. If ρ > d(1 − 1/q) then
Z d−1−ρ 3 r ς (µ)(r) dr . kµbkLq .
In particular, if µ sastisfies (b) and ρ > (4d − 3β)/2 then
Z d−1−ρ r ς (µ)(r) dr . 1.
1 Proof. Suppose first that µb ∈ L . Converting from polar coordinates we have Z rd−1−ρς (µ)(r) dr ZZ −ρ ≤ |ξ| µb(ξ + η)µb(ξ − η)µb(η) dξ dη.
Using the Brascamp-Lieb inequality we have that this is bounded above by
≤ |ξ|−ρ kµk3 . Lp b Lq
The first statement of the result then follows upon taking ρ > d(1−1/q) = d/q0 and applying
a limiting argument to the measures φn ∗ µ. Using the Fourier decay of µ, if 1/p + 3/q = 2,
q > 2d/β, and p > d/ρ this is finite precisely when ρ > (4d − 3β)/2, proving the second
statement.
d Proposition 3.2.9. Suppose f ∈ C(R ) is a positive function with fˆ ∈ L1. Let ρ < d − 1. Then Z ∞ d−1−ρ d−1−ρ δ (f)(s) . s r |σ (f)(r)| dr. 0
Proof. By Lemma 3.2.5 we have
Z ∞ δ (f)(r) = O(1)rd/2 sd/2J(rs)σ (f)(s) ds. 0
Split this integral as
Z 1/r Z ∞! d/2 d/2 δ (f)(r) . r + s J(rs) |σ (f)(s)| ds =: δ (f)1 (r) + δ (f)2 (r). 0 1/r
69 Then using Lemma 3.2.4
δ (f)1 (r) Z 1/r d/2 (d−2)/2 d/2 .r (sr) r |σ (f)(s)| ds 0 Z 1/r =rd−1 sd−1 |σ (f)(s)| ds 0 Z 1/r =rd−1 sρsd−1−ρ |σ (f)(s)| ds 0 Z 1/r ≤rd−1−ρ sd−1−ρ |σ (f)(s)| ds. 0
Similarly,
δ (f)2 (r) Z ∞ d/2 −1/2 d/2 .r (sr) r |σ (f)(s)| ds 1/r Z ∞ = (sr)(d−1)/2 |σ (f)(s)| ds 1/r Z ∞ = r(d−1)/2sρ−(d−1)/2sd−1−ρ |σ (f)(s)| ds 1/r Z ∞ ≤ rd−1sρsd−1−ρ |σ (f)(s)| ds 1/r Z ∞ ≤rd−1−ρ (rs)ρsd−1−ρ |σ (f)(s)| ds 1/r Z ∞ ≤rd−1−ρ sd−1−ρ |σ (f)(s)| ds. 1/r
d Lemma 3.2.10. [Non-singularity lemma.] Let µ be a Radon probability measure on R .
q Suppose that either µ satisfies (b) or that µb ∈ L . Let d−1 > (4d−3β)/2 or d−1 > d(1−1/q), respectively. Then the measure δ (µ) exists, and further δ (µ) ∈ L∞.
Proof. Existence follows from Lemma 3.2.7.
Let fn = φn ∗ µ. Let ρ satisfy either d − 1 > ρ > (4d − 3β)/2 or d − 1 > ρ > d(1 − 1/q)
as appropriate.
70 By Lemma 3.2.8 and the hypotheses on µ and ρ,
Z ∞ d−1−ρ r |σ (fn)(r)| dr < ∞ 0 uniformly in n, so by Lemma 3.2.9
sup δ (fn)(s) < ∞. n,s
Noting also that for functions g and f withg, ˆ fˆ ∈ L1
2 Z ZZ Y g(s) δ (f)(s) = g(|r|) f(x − ir) dx dr, i=0 we have by the definition of the measure δ (µ) that for any g ∈ L1(R) withg ˆ ∈ L1
2 Z ∞ Z Y g δ (µ) = lim g(|r|) fn(x − ir) dx dr n → ∞ 0 i=0 Z Z = lim g(s)δ (fn)(s) ds |g(s)| ds = kgk 1 n → ∞ . L and so δ (µ) ∈ L∞.
We also include here the following weaker result, valid when d = 1.
d q Lemma 3.2.11. Let µ be a Radon probability measure on [0, 1] . Suppose that µb ∈ L for some q ≤ 3. Then the measure δ (µ) exists, and further
s 3 H 4 (µ) & kδ (µ)k for any s ≤ 1 satisfying 1 d(6 − 2q) s < + . 2 2q
d+β−α In particular, if µ satisfies (a) and (b) then this holds with q = 2 β .
Proof. As in Lemma 3.2.10, existence follows from Lemma 3.2.7.
71 d Given a, R ∈ R and an > 0, let χ = 1B(a,R+) − 1B(a,R) be the indicator function of the annulus about the point a with inner radius R and outer radius R + . Then according to
Falconer’s Lemma 2.1 in [16] for any s ∈ (0, 1)
d−1 d−1 − d−1 −1 d−1 d−1 − d−1 s −(1−s) |χb(ξ)| . min R ,R 2 |ξ| 2 min(, |ξ| ) . min R ,R 2 |ξ| 2 |ξ| . (3.15)
By Plancherel’s theorem applied to the integral of the product of the functions F0(x, r) :=
χ(r)φn ∗ µ(x),F1(x, r) := φn ∗ µ(x − r)φn ∗ µ(x − 2r) Z Z χ δ (µ) = lim F0(x, r)F1(x, r) dx dr n → ∞ ZZ = lim χ(ξ − 2η)φ\n ∗ µ(ξ + η)φ\n ∗ µ(ξ)φ\n ∗ µ(η) dξ dη. n → ∞ b
By the Brascamp-Lieb inequality,
Z 3 3 χ δ (µ) ≤ lim kχkLp φ\n ∗ µ = kχkLp kµkLq n → ∞ b Lq b b
where 1 3 + = 2. p q
q 1 d(6−2q) So let p = 2q−3 . If we choose s < 2 + 2q then by (3.15) we have
d−1 s 2 kχbkLp . R .
This gives Z s d−1 δ (µ) ([R,R + ]) = χ δ (µ) . R 2 , and the Falconer type argument follows:
Suppose that ([R ,R + ]) is a covering of supp δ (µ) contained in B(0, 2). Then i i i i∈N