P D D D Z

Download as pdf or txt
Download as pdf or txt
You are on page 1of 39

THE IONESCU–WAINGER MULTIPLIER THEOREM

AND THE ADELES

TERENCE TAO
arXiv:2008.05066v1 [math.CA] 12 Aug 2020

Abstract. The Ionescu–Wainger multiplier theorem establishes


good Lp bounds for Fourier multiplier operators localized to ma-
jor arcs; it has become an indispensible tool in discrete harmonic
analysis. We give a simplified proof of this theorem with more
explicit constants (removing logarithmic losses that were present
in previous versions of the theorem), and give a more general vari-
ant involving adelic Fourier multipliers. We also establish a closely
related adelic sampling theorem that shows that ℓp (Zd ) norms of
functions with Fourier transform supported on major arcs are com-
parable to the Lp (AdZ ) norm of their adelic counterparts.

1. Introduction

This paper will be concerned with the Lp theory of Fourier multiplier


operators on various locally compact abelian groups, such as Zd , Rd ,
and AdZ . In order to treat these groups in a unified fashion we adopt
the following abstract harmonic analysis notation.
Definition 1.1 (Pontryagin duality). An LCA group is a locally com-
pact abelian group G = (G, +) equipped with a Haar measure µG . A
Pontryagin dual of an LCA group G is an LCA group G∗ = (G∗ , +)
with a Haar measure µG∗ and a continuous bihomomorphism (x, ξ) 7→
x · ξ (which we call a pairing) from G × G∗ to the unit circle T = R/Z,
such that the Fourier transform FG : L1 (G) → C(G∗ ) defined by
Z
FG f (ξ) := f (x)e(x · ξ) dµG (x),
G

where e : T → C is the standard character e(θ) := e2πiθ , extends to a


unitary map from L2 (G) to L2 (G∗ ); in particular we have the Plancherel
identity
Z Z
2
|f (x)| dµG (x) = |FG f (ξ)|2 dµG∗ (ξ)
G G∗

2010 Mathematics Subject Classification. 42B15.


1
2 TERENCE TAO

for all f ∈ L2 (G), as well as the inversion formula


Z
−1
FG F (x) = F (ξ)e(−x · ξ) dµG∗ (ξ)
G∗

for all F ∈ L1 (G∗ ) ∩ L2 (G∗ ).

If Ω ⊂ G∗ is measurable, we say that f ∈ L2 (G) is Fourier supported


in Ω if FG f vanishes outside of Ω (modulo null sets). The space of such
functions will be denoted L2 (G)Ω .

If m ∈ L∞ (G∗ ), we define the associated Fourier multiplier operator


Tm : L2 (G) → L2 (G) by the formula
FG Tm f := mFG f
for all f ∈ L2 (G), thus Tm = FG−1 mFG . We refer to m as the symbol
of Tm .

For any finite-dimensional normed vector space V , we extend Tm to an


operator on L2 (G; V ) in the obvious fashion.

We will focus in particular on the Pontryagin dual pairs


(G, G∗ ) = (Zd , Td ), (Rd , Rd ), (AdZ , Rd × (Q/Z)d )
where AZ = R × Ẑ denotes the adelic integers and d ≥ 1 is an integer;
see Appendix A for a more precise description of these pairs. To avoid
technicalities we shall largely restrict attention to smooth symbols m,
although rougher symbols can also be treated by applying suitable
limiting arguments, as our estimates will not depend on any smooth
norms of m. We view the adelic space AdZ = Rd × Ẑd as a simplified
model of the lattice Zd that captures both the “continuous” aspects
of this lattice (via the factor Rd ) and the “arithmetic” aspects of this
lattice (via the factor Ẑd ). The reader may wish to restrict attention
to the one-dimensional case d = 1 as it already captures all of the
key ideas, but the extension to higher values of d requires only minor
notational changes and is also useful in some applications (e.g., [14]),
so we work with general d in this paper.

A central problem in harmonic analysis is to understand the operator


norm kTm kB(Lp (G)) of a Fourier multiplier operator Tm on a Lebesgue
space Lp (G) (restricting Tm initially to some dense subclass such as
the Schwartz-Bruhat space S(G) to avoid technicalities). For p = 2
this norm is just the L∞ (G∗ ) norm of m, but for other choices of p the
situation is considerably more complicated. Our initial focus here will
be on understanding this problem in the case where G = Zd and m is
supported on “major arcs”.
IONESCU–WAINGER AND THE ADELES 3

If m ∈ Cc∞ (Rd ) is a smooth symbol, then Tm is a Fourier multiplier


operator on L2 (Rd ), but we can also define associated Fourier multiplier
operators Tm;α on ℓ2 (Zd ) for various shifts α ∈ Td by the formula
Tm,α := Tmα
where mα ∈ C ∞ (Td ) is the symbol
X
mα (ξ) := m(θ).
θ∈Rd :ξ=α+θ mod Zd

Equivalently, one has


Z
Tm;α f (n) = m(θ)e(−n · (α + θ))FZd f (α + θ) dθ.
Rd

More generally, for any finite set Σ ⊂ Rd , define


X
Tm;Σ := Tm;α , (1.1)
α∈Σ

thus
XZ
Tm;Σ f (n) = m(θ)e(−n · (α + θ))FZd f (α + θ) dθ.
α∈Σ Rd

If the support of m is suitably restricted, then the ℓp (Rd ) multiplier


theory of Tm;α or Tm;Σ is closely tied to the Lp (Zd ) multiplier theory
of Tm . One basic manifestation of this is via the following sampling
principle of Maygar, Stein, and Wainger [9]. For any ξ0 ∈ Rd , let
ξ0 + [−r, r]d denote the closed cube of sidelength 2r centred at ξ0 . We
also define analogous balls (or cubes or “arcs”) α + [−r, r]d ⊂ Td for
α ∈ Td . For any positive integer Q, let
 d
d d 1
T [Q] := {x ∈ T : Qx = 0} = Z/Z
Q
denote the subgroup of Q-torsion points of the torus Td .
Proposition 1.2 (Maygar–Stein–Wainger sampling principle). Let d ≥
1 be an integer, let 1 ≤ p ≤ ∞, and let V be a finite-dimensional Ba-
nach space.

(i) If m ∈ Cc∞ (Rd ) is supported in [− 12 , 12 ]d , then


kTm;0 kB(ℓp (Zd ;V )) ≤ O(1)dkTm kB(Lp (Rd ;V )) . (1.2)
(See Section 1.2 for our conventions on asymptotic notation, as
well as our notation B(W ) for the operator norm on a normed
vector space W .)
4 TERENCE TAO

(ii) More generally, if Q ≥ 1 is an integer, and m ∈ Cc∞ (Rd ) is


1 1 d
supported in [− 2Q , 2Q ] , then
kTm;Td [Q] kB(ℓp (Zd ;V )) ≤ O(1)dkTm kB(Lp (Rd ;V )) .

Proof. Part (i) is [9, Proposition 2.1] and part (ii) is [9, Corollary 2.1],
after noting that all implied constants in the proof are at most expo-
nential in the dimension d. 

In the remarks after [9, Proposition 2.1] the question is posed as to


whether the O(1)d constant in (1.2) can be made independent of d, or
even replaced with 1. Although not the main focus of this paper, in
Appendix B we show that the answer to these questions is negative if p
is sufficiently close to 1 or ∞. Most likely the answer is negative for all
p 6= 2, but we were unable to demonstrate this. (For p = 2 it is easy to
see from Plancherel’s theorem that the O(1)d factor may be deleted.)

This sampling principle lets us control certain Fourier multiplier oper-


ators whose symbol is supported in sets of the form
Td [Q] + [−ε, ε]d
for ε > 0 small enough (in particular, the above proposition applies
when ε ≤ 1/2Q). For applications to discrete harmonic analysis (par-
ticularly involving averaging over “arithmetic” sets such as polynomial
sequences or primes), it would be desirable to have a similar estimate
that could handle symbols supported on the “classical major arcs”
N
[
Td [q] + [−ε, ε]d (1.3)
q=1

for some N ≥ 1 and ε > 0. As it turns out, these classical arcs are
inconvenient to work with directly for the purposes of ℓp multiplier
theory. However, in the remarkable work of Ionescu and Wainger [7],
a more complicated major arc set
M = Σ≤k + [−ε, ε]d
was introduced for which (for suitable choices of parameters) contained
the classical major arc set (1.3) while simultaneously enjoying a satis-
factory ℓp multiplier theory for relatively large values of ε. The Ionescu–
Wainger multiplier theorem has since been indispensable in many re-
sults in discrete harmonic analysis, in particular providing an analogue
of Littlewood-Paley theory adapted to major arcs; see e.g., [18], [6],
[12], [13], [14], [8].

In this note we give a general version of the Ionescu–Wainger theorem


which avoids some logarithmic loss factors present in earlier treatments,
IONESCU–WAINGER AND THE ADELES 5

and quantifies the dependence on various parameters. To describe the


result we need some notation.
Definition 1.3 (Generalized Ionescu–Wainger major arcs). A major
arc parameter set is a quadruplet (d, k, S, ε) where d, k ≥ 1 are integers,
S is a finite collection of pairwise coprime integers, and ε > 0 is a
real number. For any A ⊆ S, we let QA denote the positive integer
QA := q∈A q, and let Σ⊆A ⊆ Td denote the subgroup
Q

Σ⊆A := Td [QA ].
We let ΣA denote the set
[
ΣA := Σ⊆A \ Σ⊆B .
B(A

We define the sets


   
S S
:= {A ⊆ S : |A| = k}; := {A ⊆ S : |A| ≤ k}
k ≤k
and let Σ≤k denote the set
[ [
Σ≤k := Σ⊆A = ΣA .
S S
A∈(≤k ) A∈(≤k )
The major arc set M = M(d,k,S,ε) associated to the parameter set
(d, k, S, ε) is defined as
M := Σ≤k + [−ε, ε]d .

A major arc parameter set is said to be (r, c)-good for some integer
r ≥ 1 and 0 < c < 1 if one has the smallness condition
c
ε< 2rk
(1.4)
2rqmax
for some integer qmax that is greater than or equal to all the elements
of S.

Expanding out the definitions, we see that M consists of all elements


of Td of the form aq + θ mod Zd , where q is the product of at most
k elements of S, a ∈ Zd , and θ ∈ [−ε, ε]d . The major arc sets M
considered here are more general than the ones constructed in [7], which
involved a specific choice of S involving a partition of all the primes up
to a certain threshold. In Section 5 we explain how the major arcs in
[7] become a special case of the ones considered here. The parameter c
is of minor technical importance and the reader may wish to fix it as an
absolute constant (e.g., c = 1/2) for most of the following discussion.
In typical applications one should think of the quantities d, k, r as being
bounded, |S| and qmax as being large, and ε as being quite small.
6 TERENCE TAO

We can now state our first form of the Ionescu–Wainger multiplier


theorem. To simplify the bounds slightly we adopt the notation

Logx := log(2 + x).

Theorem 1.4 (Ionescu–Wainger multiplier theorem, real form). Let


(d, k, S, ε) be a major arc parameter set, and let H be a finite-dimensional
Hilbert space. Let m ∈ Cc∞ (Rd ) be supported on [−ε, ε]d . Then if
(d, k, S, ε) is (r, c)-good for some integer r ≥ 1 and 0 < c < 1, one has

kTm;Σ≤k kB(ℓ2r (Zd ;H)) ≤ Oc (1)d O(rLog1/2 (kr))k kTm kB(L2r (Rd )) ;

more generally, one has


X
k ǫA Tm;ΣA kB(ℓ2r (Zd ;H)) ≤ Oc (1)d O(rLog1/2 (kr))k kTm kB(L2r (Rd )) .
S
A∈(≤k )
S

whenever ǫA is a complex number with |ǫA | ≤ 1 for each A ∈ ≤k .

The factor of Oc (1)d O(rLog1/2 (kr))k looks somewhat messy, but the
key point is that it is uniform in the parameters S, ε, H, m. The orig-
inal version of this result in [7], when adapted to this notion of major
arc, gave instead a bound of the form Oc,d,k,r (Logk |S|), whhich was
later refined in [11] to Oc,d,k,r (Log|S|) (see also [12], [13], [14]). The de-
pendence of c will be unimportant in applications as one can typically
take c to equal a constant value such as c = 1/2. The restriction to
even integer exponents 2r will be removed in Theorem 1.7 below (at
the cost of worsening the bounds slightly). We work with finite dimen-
sional Hilbert spaces H here to avoid technical complications, but one
can extend this result to separable Hilbert spaces without difficulty by
a standard limiting argument.

We prove Theorem 1.4 in Section 4, after some preliminaries in Sec-


tions 2, 3. The argument uses the same basic approach as previous
proofs of the Ionescu–Wainger theorem in the literature (particularly
[14, Theorem 2.1]), which we summarize as follows.

(i) One begins by exploiting “Type II superorthogonality” (fol-


lowing the terminology recently introduced by Pierce [19]) of
the terms Tm;ΣA f (arising from “denominator orthogonality”
of
P the rational set Σ≤k ) to2r estimate the ℓ2r (Zd ; H) norm of
d [≤k]
A∈( S ) ǫA Tm;ΣA f by the ℓ (Z ; H
≤k
) of the square function
S
(Tm;ΣA f )A∈( S ) that takes values in H [≤k] := H (≤k) , in the spirit
≤k
of reverse square function estimates of Khintchine type.
IONESCU–WAINGER AND THE ADELES 7

(ii) Using a “nonconcentration estimate”, one estimates this square


function norm by an expression summing over various “sunflow-
ers” in S.
(iii) By exploiting “numerator orthogonality” of the rational set
Σ≤k , one estimates this resulting sum over sunflowers by a more
tractable square function involving the functions Tm;α+Σ⊆A0 f
that are summed over cosets of a fixed finite subgroup Σ⊆A0 =
Td [QA0 ] of Td .
(iv) At this point the symbol m can be disposed of using the Marcinkiewicz–
Zygmund theorem, and then the resulting quantity can be es-
timated using a square function estimate of Rubio de Francia
type [21].

Our main innovations are to eliminate logarithmic losses in (i) using the
probabilistic decoupling trick (cf. [17]), and to obtain efficient bounds
in (ii) by using recent progress [20] on the sunflower conjecture of Erdős
and Rado [4].

We also interpret these results through the lens of adelic harmonic


analysis, following the slogan
Major arc analysis on Zd ≈ Low frequency analysis on AdZ
recently advocated (in the one-dimensional setting d = 1) in [8]. As
reviewed in Appendix A, we have an inclusion homomorphism
ι : Zd → AdZ
and a addition homomorphism
π : Rd × (Q/Z)d → Td
that is Fourier adjoint to ι, with π being given explicitly by
π(θ, α) := α + θ
for θ ∈ Rd and α ∈ (Q/Z)d .

There is a sampling map S : S(AdZ ) → S(Zd ), where S(G) denotes the


Schwartz-Bruhat space on G (as defined in Appendix A), defined by
Sf := f ◦ ι.
As in [8], we say that a compact subset Ω of adelic frequency space
Rd × (Q/Z)d is non-aliasing if the projection map π is injective on Ω.
In [8, (4.6)] it was shown that the sampling map extends to a unitary
transformation
S : L2 (AdZ )Ω → ℓ2 (Zd )π(Ω)
for any non-aliasing compact Ω ⊂ Rd × (Q/Z)d (the argument was
presented for d = 1, but extends to arbitrary dimension). In particular
8 TERENCE TAO

one has an interpolation map


SΩ−1 : ℓ2 (Zd )π(Ω) → L2 (AdZ )Ω
that inverts S; we extend these operators to vector-valued functions
taking values in a finite-dimensional vector space in the obvious fashion.
For instance, if Ω is a set of the form [−ε, ε]d × Σ for some ε > 0 and
some finite Σ ⊂ (Q/Z)d , then Ω is non-aliasing if the elements of Σ
are separated from each other by more than 2ε (in the ℓ∞ metric), and
then π(Ω) = Σ + [−ε, ε]d and every element f of ℓ2 (Zd )π(Ω) then has a
unique Fourier representation of the form
XZ
f (n) = e(−n · (α + θ))FZd f (α + θ) dθ
α∈Σ [−ε,ε]d

for n ∈ Zd , and the interpolated function SΩ−1 f ∈ L2 (AdZ )Ω is then given


by the formula
XZ
−1
SΩ f (x, y) = e(−x · θ − y · α)FZd f (α + θ) dθ
α∈Σ [−ε,ε]d

and then it is clear that f = SSΩ−1 f .

From unitarity we have


kSΩ−1 f kL2 (AdZ ) = kf kℓ2 (Zd )

whenever f ∈ ℓ2 (Zd )π(Ω) . In many cases we can extend this L2 isometry


to an Lp equivalence. For instance, we have
Proposition 1.5 (Quantitative Shannon sampling theorem). Let 1 ≤
p ≤ ∞, and V be a finite-dimensional normed vector space. If Ω is the
(non-aliasing) set Ω := [− Qc , Qc ]d × Td [Q] for some positive integer Q
and 0 < c < 21 then SΩ extends to a bounded invertible linear operator
from ℓp (Zd ; V )π(Ω) to Lp (Zd ; V )Ω with
kSΩ−1 f kLp (AdZ ;V ) = exp(Oc (d))kf kℓp (Zd ;V )

for all f ∈ ℓp (Zd ; V )π(Ω) , or equivalently


kSF kℓp (Zd ;V ) = exp(Oc (d))kF kLp (AdZ ;V )

for all F ∈ Lp (AdZ ; V )Ω .

Proof. See [8, Theorem 4.18], after generalizing from d = 1 to general


d and carefully tracking the dependence on constants. The result also
extends to 0 < p < 1 (after allowing the implied constants to depend
on p as well as c), but we will only need the p ≥ 1 case here. 
IONESCU–WAINGER AND THE ADELES 9

Proposition 1.5 can be used to partially explain the sampling principle,


Proposition 1.2. First observe that if [−ε, ε]d × Σ is a non-aliasing set
then we have the identity
Tm;Σ Sf = STm⊗1Σ f (1.5)
for any m ∈ Cc∞ (Rd ) supported on [−ε, ε]d × Σ and any f ∈ S(AdZ ),
where the tensor product ⊗ is defined in Section 1.2; see [8, Lemma
4.12] (extended to general dimension d in the obvious fashion). Now
suppose that Q ≥ 1 is an integer and m ∈ Cc∞ (Rd ) is supported in
[− Qc , Qc ]d for some 0 < c < 12 . Then we may use (1.5) and basic
Fourier-analytic manipulations to factorize
Tm;Td [Q] f = Tm;Td [Q] SSΩ−1 Tϕ;Td [Q] f
= STm⊗1Td [Q] SΩ−1 Tϕ;Td [Q] f = STm⊗1 SΩ−1 Tϕ;Td [Q] f

where ϕ ∈ Cc∞ (Rd ) is a function of the form


d
Y
ϕ(ξ1 , . . . , ξd ) := ϕ0 (Qξj ),
j=1

1
ϕ 0 ∈ C∞ ′ ′ ′
c (R) is supported on [−c , c ] for some c < c < 2 that equals
′ c′
1 on [−c, c], and Ω := [− cQ , Q ]d × Td [Q]. From Proposition 1.2 applied
to ϕ we have
kTϕ;Td [Q] kB(ℓp (Zd ;H)) ≤ Oc,c′ (1)d
while from working on each fibre Rd × {y} of AdZ = Rd × Ẑd sepa-
rately and using the Marcinkiewicz–Zygmund theorem (Theorem 1.8),
we have
kTm⊗1 kB(Lp (AdZ ;H)) = kTm kB(Lp (Rd ;H)) = kTm kB(Lp (Rd ))
and hence from Proposition 1.5 we have
kTm;Σ kB(ℓp (Zd ;H)) ≤ Oc,c′ (1)d kTm kB(Lp (Rd ))
which recovers a slightly weaker version of Proposition 1.2; in fact with
a bit more effort (applying a smooth partition of unity to m followed
by the triangle inequality) one can in fact recover the full strength
of Propostition 1.2. Admittedly, there is some circularity here since
Proposition 1.2 was used in the proof, but only for the bump function
ϕ and not for arbitrary multipliers m.

It turns out that there is a similar phenomenon for major arcs:


Theorem 1.6 (Major arc sampling). Let (d, k, S, ε) be a major arc
parameter set, which is (r, c) good for some r ≥ 1 and 0 < c < 1. Set
Ω := [−ε, ε]d × Σ≤k . Then for any finite-dimensional Hilbert space H
10 TERENCE TAO

and (2r)′ ≤ p ≤ 2r, the interpolation operator SΩ−1 extends to a bounded


invertible linear operator from ℓp (Zd ; H)π(Ω) to Lp (Zd ; H)Ω , with
kSΩ−1 f kLp (AdZ ;H) = exp (Oc (d) + O(kLog(rLogk))) kf kℓp (Zd ;H)

for all f ∈ ℓp (Zd ; H)π(Ω) , or equivalently


kSF kℓp (Zd ;H) = exp (Oc (d) + O(kLog(rLogk))) kF kLp (AdZ ;H)

for all F ∈ Lp (AdZ ; H)Ω .

We prove this theorem in Section 4, by reusing the machinery used to


establish Theorem 1.4. As a consequence of this sampling theorem,
we can obtain a more general “adelic” version of the Ionescu–Wainger
multiplier theorem, in which one transfers a multiplier on AdZ = Rd ×
Ẑd rather than on Rd to the lattice Zd , or equivalently one allows
the use of a different multiplier on each major arc. More precisely,
given m ∈ L∞ (Rd × (Q/Z)d ) and a finite set Σ, define the multiplier
Tm;Σ : ℓ2 (Zd ) → ℓ2 (Zd ) by the formula
X
Tm;Σ := Tm(·,α);α
α∈Σ

or equivalently
XZ
Tm;Σ f (n) = m(θ, α)e(−n · (α + θ))FZd f (α + θ) dθ.
α∈Σ Rd

Note that the previous definition (1.1) corresponds to the special case
in which the adelic symbol m(θ, α) does not depend on the arithmetic
component α.
Theorem 1.7 (Ionescu–Wainger multiplier theorem, adelic form). Let
(d, k, S, ε) be a major arc parameter set, and let H be a finite-dimensional
Hilbert space. Let m ∈ S(AdZ ) be supported on [−ε, ε]d × (Q/Z)d . Then
if (d, k, S, ε) is (r, c)-good for some r ≥ 1 and 0 < c < 1, one has
kTm;Σ≤k kB(ℓp (Zd ;H)) ≤ Oc (1)d O(rLogk)O(k) kTm kB(Lp (AdZ ))

for any (2r)′ ≤ p ≤ 2r.

In principle this theorem converts the analysis of linear Fourier multi-


pliers on major arcs to that of linear Fourier multipliers on the adelic
space AdZ , which in principle is a simpler setting due to the product
structure on AdZ = Rd × Ẑd . We remark that the method of proof
also extends to bilinear or multilinear Fourier multipliers (as long as
all exponents p involved lie strictly between 1 and ∞), but we do not
IONESCU–WAINGER AND THE ADELES 11

have applications in mind for this extension1 and so we leave it to the


interested reader.

1.1. Acknowledgments. The author was partially supported by NSF


grant DMS-1764034 and by a Simons Investigator Award. The author
also thanks Mariusz Mirek for helpful comments and corrections.

1.2. Notation. Random variables will be denoted in boldface, and


deterministic quantities in non-boldface. We use N = {0, 1, . . . } to
denote the natural numbers, and Z+ = {1, 2, . . . } to denote the positive
integers.

We use X = O(Y ) to denote an estimate of the form |X| ≤ CY for


some constant C. We write X ∼ Y if X = O(Y ) and Y = O(X). If
one needs the constant C to depend on parameters, we indicate this
by subscripts, for instance X ≤ Oc (1)d Y denotes the bound X ≤ Ccd Y
for some Cc depending only on c.

If (X, µ) is a measure space, V = (V, kk) is a finite-dimensional normed


vector space, and 1 ≤ p ≤ ∞, we define Lp (X; V ) to denote the
space of measurable functions f : X → V whose norm kf kLp (X;V ) :=
( X kf (x)kpV dµ(x))1/p is finite, up to almost everywhere equivalence,
R
with the usual modifications at p = ∞. We write Lp (X) for Lp (X; C),
and when µ is counting measure we write ℓp for Lp . For any 1 ≤ p ≤ ∞,
we define the dual exponent 1 ≤ p′ ≤ ∞ by 1/p + 1/p′ = 1.

All Hilbert spaces will be over the complex numbers. Given a bounded
linear operator T : V → W between (quasi-)normed vector spaces
V, W , we use kT kB(V →W ) to denote its operator norm; if V = W ,
we abbreviate B(V → V ) as B(V ). We recall

Theorem 1.8 (Marcinkiewicz–Zygmund theorem). [10] Let X, Y be


measure spaces, let 0 < p < ∞, and let T : Lp (X) → Lp (Y ) be a linear
operator. Then for any finite-dimensional Hilbert space H, one has

kT kB(Lp (X;H)→Lp (Y ;H)) ≤ kT kB(Lp (X)→Lp (Y )) .

1For instance, the bilinear estimates considered in [8] typically involve the end-
point space ℓ1 (or even ℓp for some p < 1), and also take values in variational norm
spaces rather than Hilbert spaces, so would not be able to be directly treated by a
bilinear variant of Theorem 1.7.
12 TERENCE TAO

Proof. We normalize kT kB(Lp (X)→Lp (Y )) = 1. Taking orthonormal bases,


it suffices to show that
Z n
!p/2 Z n
!p/2
X X
| T fi |2 ≤ | fi |2
Y i=1 X i=1
p
for any f1 , . . . , fn ∈ L (X). If we let g1 , . . . , gn be independent complex
gaussian variables of mean zero and variance 1, we have from hypothesis
that Z X n Z X n
p
| gi T fi | ≤ | gi fi |p .
Y i=1 X i=1
Taking expectations of both sides and noting that the sum of indepen-
dent gaussians is again a gaussian, we conclude that
Z n
!p/2 Z n
!p/2
X X
Cp | T fi |2 ≤ Cp | fi |2
Y i=1 X i=1
p
where Cp := E|g| with g a complex gaussian of mean zero and variance
1. Since 0 < Cp < ∞, the claim follows. 

If E is a finite set, we use |E| to denote its cardinality. If E, F are


subsets of an additive group G = (G, +) (such as the torus Td ), we
write E + F := {ξ + η : ξ ∈ E, η ∈ F } for their sumset. If ξ ∈ G, we
write ξ +E = E +ξ = E +{ξ} for the translate of ξ by E. These sumset
notions also extend in the obvious fashion to the setting in which one
of the summands lies in G and the other lies in a quotient G/H (for
instance, if one lies in Rd and the other in Td ). We use 1E to denote
the indicator function of a set E, and 1S the indicator of a statement
S, thus for instance 1E (x) = 1x∈E is equal to 1 when x ∈ E, and equal
to 0 otherwise.

We will need the following combinatorial concepts:


Definition 1.9 (Nonces and sunflowers). Let A1 , . . . , An be a collection
of sets.

(i) A nonce of the collection A1 , . . . , An is an element s that belongs


to exactly one of the Ai . A collection is nonce-free if there does
not exist a nonce.
(ii) The collection A1 , . . . , An is a sunflower if there is a set A0
contained in A1 , . . . , An (the core of the sunflower) such that
the petals A1 \A0 , . . . , An \A0 are all disjoint.

Thus for instance the sets {1, 2}, {1, 3}, {2, 4} contain 3 and 4 as nonces,
whereas {1, 2}, {1, 3}, {2, 3} are nonce-free, while {1, 2}, {1, 3}, {1, 4}
IONESCU–WAINGER AND THE ADELES 13

is a sunflower with core {1} and petals {2}, {3}, {4}. The property of
having a nonce is also referred to as the uniqueness property in [7], [14],
[19].

If f : X → C and g : Y → C are functions, we define the tensor product


f ⊗ g : X × Y → C by the formula
(f ⊗ g)(x, y) := f (x)g(y).

If H is a finite-dimensional Hilbert space and S is a finite set, we use


H S for the space of tuples (us )s∈S with us ∈ H with inner product
X
h(us )s∈S , (vs )s∈S i = hus , vs i.
s∈S

For any natural number k, we use H ⊗k to denote the k-fold tensor


product of H with itself, spanned by vectors u1 ⊗· · ·⊗uk , u1 , . . . , uk ∈ H
with
k
Y
hu1 ⊗ · · · ⊗ uk , v1 ⊗ · · · ⊗ vk i = hui , vi i.
i=1

2. Superorthogonality

The (upper) Khintchine inequality asserts that


n
!1/p n
!1/2
X X
E| ǫi zi |p ≤ O(p1/2 ) |zi |2
i=1 i=1

for any 1 ≤ p ≤ ∞ and complex numbers z1 , . . . , zn , where ǫ1 , . . . , ǫn


are independent random signs in {−1, +1} of mean zero. In the case
where p = 2r is an even integer, this inequality can be proven by direct
combinatorial expansion of the left-hand side. As laid out recently in
[19], this latter argument can be abstracted to more general “Type II
superorthogonal systems”. We give the relevant definitions (as well as
an extension to hypersystems) as follows.
Definition 2.1 (Type II superorthogonality). Let S be a finite set, let
X = (X, µ) be a measure space, let r be a positive integer, and let H
be a Hilbert space.

(i) A collection (fs )s∈S of functions fs ∈ L2r (X; H) indexed by S


is said to be a Type II 2r-superorthogonal system if one has
Z Y r
hfsj , fsr+j iH dµ = 0 (2.1)
X j=1
14 TERENCE TAO

whenever s1 , . . . , s2r ∈ S is such that the singleton sets {s1 }, . . . , {s2r }


contain a nonce (as defined in Definition 1.9).
(ii) A collection (fA )A∈A of functions fA ∈ L2r (X; H) indexed by
some family A of subsets of a set S is said to be a Type II
2r-superorthogonal hypersystem if one has
r
Z Y
hfAj , fAr+j iH dµ = 0 (2.2)
X j=1

whenever A1 , . . . , A2r ∈ A is such that the sets A1 , . . . , A2r


contain a nonce.

Note that any 2r-superorthogonal system (fs )s∈S can also be viewed
as a 2r-superorthogonal hypersystem (fA )A∈(S ) by identifying each ele-
1

ment s ∈ S with the associated singleton {s} ∈ S1 . The nomenclature



“Type II” is due to Pierce [19]; there is also a stronger notion of Type I
superorthogonality and a weaker notion of Type III superorthogonality
discussed in that paper, but we will not need these notions here.

Several examples of superorthogonal systems are given in [19]. Our


primary concern will come from functions supported on major arcs,
but we can give another representative example of a superorthogonal
hypersystem here:

Example 2.2 (Polynomials of random variables). Let k, R be posi-


tive integers. Let (Xs )s∈S be R-wise independent random variables
(thus Xs1 , . . . , Xsr are jointly independent for any r ≤ R and distinct
S

s1 , . . . , sr ∈ S), and for each A ∈ ≤k let fA be an complex random
P Q
variable of the form fA = j∈JA cA,j s∈A fA,s,j (Xs ), where JA is a
finite set, cA,j are complex coefficients, and each fA,s,j (Xs ) is a func-
tion of Xs of mean zero; thus for instance f∅ is a constant. Then for
any 1 ≤ r ≤ R/2k, (fA )A∈( S ) is a 2r-superorthogonal hypersystem
≤k
over the ambient sample space of the random variables. Indeed, if
A1 , . . . , A2r contains a nonce s, then the expression in (2.2) expands
into a sum of finitely many terms, each of which consists of the expec-
tation of a product of an expression of the form fA,s,j (Xs ), times at
most 2kr − 1 ≤ R − 1 other expressions depending on other random
variables than Xs , and each of these terms vanishes by the 2R-wise
independent nature of the Xs . If the Xs are scalar random variables,
then any polynomial
P of degree at most k in the Xs can be expressed
in the form A∈( S ) fA for some hypersystem (fA )A∈( S ) as above by
≤k ≤k
removing the expectation from every monomial Xas appearing in the
polynomial and regrouping terms.
IONESCU–WAINGER AND THE ADELES 15

We now give the Khintchine inequalities for superorthogonal systems


and hypersystems.

Theorem 2.3 (Superorthogonal Khintchine inequality). Let k, r ∈ Z+ ,


let X = (X, µ) be a measure space, S a finite set, and H a finite-
dimensional Hilbert space.

(i) (Khintchine for superorthogonal systems) If (fs )s∈S is a Type II


2r-superorthogonal system in L2r (X; H) indexed by S, then
X
k fs kL2r (X;H) ≤ O(r)1/2 k(fs )s∈S kL2r (X;H S ) .
s∈S

(ii) (Khintchine for superorthogonal hypersystems) If (fA )A∈( S ) is


≤k
a Type II 2r-superorthogonal hypersystem in L2r (X; H) then
X
k/2
k fA kL2r (X;H) ≤ O(r) (fA )A∈( S )

≤k L2r (X;H [≤k] )
S
A∈(≤k )

where we adopt the notation


S
H [≤k] := H (≤k) .

Part (i) is standard (see e.g., [19, §3.1]). Part (ii) (without any losses of
Log|S| factors) appears to new; with logarithmic losses one can obtain
a result of this type from [13, Lemma 5.1], an iteration of (i), and the
triangle inequality.

Proof. We begin with (i). We may index S = {1, . . . , n}. The desired
estimate may be rewritten as
Z n
X Z X n
2r r
k fs kH dµ ≤ O(r) ( kfs k2H )r dµ.
X s=1 X s=1

From the binomial theorem, the Cauchy-Schwarz inequality, and the


triangle inequality, for any u, v ∈ H we have
2r  
!
X 2r j 2r−j
ku + vk2r 2r
H = kukH + 2rRehv, uikukH
2r−2
+O kvkH kukH .
j=2
j

Observe for any odd 2j + 1 between 1 and 2r that


   1/2  1/2
2r 2r 2r

2j + 1 2j 2j + 2
16 TERENCE TAO

(since k! ∼ ((k − 1)!(k + 1)!)1/2 for any k ∈ Z+ ), and hence by Young’s


inequality we may restrict the j summation here to even integers, thus
r  
!
X 2r
ku + vk2r 2r
H = kukH + 2rRehv, uikukH
2r−2
+O kvk2j 2r−2j
H kukH
j=1
2j
and in particular
r  
X 2r
ku + vk2r
H ≤ 2r−2
2rRehv, uikukH + C 1j≥1
kvk2j 2r−2j
H kukH
j=0
2j
for some absolute constant C > 1. As a special case, for any v1 , . . . , vn ∈
H, one has
2r 2r−2j
n r   n
X X 2r X
vs ≤ ReZ + C 1j≥1 kv1 k2j v

H s
2j


s=1 H j=0 s=2 H
Qr
where Z is a linear combination of expressions of the form j=1hvsj , vsr+j i
where {s1 }, . . . , {s2r } contains a nonce. Iterating this identity n times,
we conclude that
X 2r
n ∗  
X 2r
vs ≤ ReZ + ′
C 1j1 ≥1 +···+1jn ≥1
kv1 k2j 1 2jn
H . . . kvn kH

2j1 , . . . , 2jn

s=1
H

where Z is also a linear combination of expressions of the form rj=1 hvsj , vsr+j i

Q

with {s1 }, . . . , {s2r } containing a nonce, and ∗ denotes a sum over


P
tuples (j1 , . . . , jn ) ∈ Nn with j1 + · · · + jn = r. Applying this with vi :=
fi (x), integrating in X, and using the Type II 2r-superorthogonality
hypothesis (2.1) as well as the bound C 1j1 ≥1 +···+1jn ≥1 ≤ C r , we conclude
that
n ∗  Z
2r
Z X X
k 2r
fs kH dµ ≤ C r
kf1 k2j 1 2jn
H . . . kfn kH dµ.
X s=1
2j1 , . . . , 2jn X
On the other hand, we have
n ∗  Z
r
Z X X 2j1
( 2 r
kfs kH ) = kf1 kH . . . kfn k2j n
H dµ.
X s=1 j1 , . . . , jn X
To finish the proof it will suffice to establish the inequality
   
2r r r
≤ (2r)
2j1 , . . . , 2jn j1 , . . . , jn
for any j1 , . . . , jn ≥ 0 summing to r. But this follows from the combina-
torial observation that given a partition of {1, . . . , 2r} into n classes of
cardinality 2j1 , . . . , 2jn respectively, one can remove j1 elements from
the first class, then j2 elements from the second class, and so forth until
one is left with a partition of r elements of {1, . . . , 2r} into n classes
of cardinality j1 , . . . , jn . There are at most (2r)r ways to remove these
IONESCU–WAINGER AND THE ADELES 17

r

elements in the order indicated, and j1 ,...,j n
ways to partition the re-
maining elements, with the original partition being recoverable from
this data. This gives (i).

Now we prove (ii). By the triangle inequality it suffices to establish


S
this result with ≤k replaced by Sk . The case k = 0 is trivial, so
suppose k ≥ 1. We apply the probablistic decoupling method (cf.,
[17]), which can be viewed as a substitute for the random partitioning
lemma in [13, Lemma 5.1] that avoids logarithmic losses. We form a
random partition S = S1 ⊎ · · · ⊎ Sk by setting Si := {s ∈ S : is = i},
where is , s ∈ S are independent random variables  drawn uniformly at
S
random from {1, . . . , k}. Observe that if A ∈ k , then A takes the form
A = {s1 , . . . , sk } with si ∈ Si for i = 1, . . . , k with probability precisely
k!
kk
(this is the probability that the tuple (is )s∈A forms a permutation
of {1, . . . , k}). Thus we have
X kk X
fA = E f{s1 ,...,sk }
k! s
A∈( S
) 1 ∈S1 ,...,sk ∈Sk
k

and hence by the triangle inequality



X kk X
k fA kL2r (X;H) ≤ E f{s1 ,...,sk } .

k!
s1 ∈S1 ,...,sk ∈Sk

A∈(S
k) L2r (X;H)

Using the Taylor expansion

k0 k1 kk kk
ek = + +···+ +··· ≥
0! 1! k! k!
it will thus suffice to establish the deterministic inequality

X
≤ O(r)k/2 (f{s1 ,...,sk } )s1 ∈S1 ,...,sk ∈Sk L2r (X;H S1 ×···×Sk )

f{s1 ,...,sk }



s1 ∈S1 ,...,sk ∈Sk L2r (X;H)

whenever S = S1 ⊎ · · · ⊎ Sk is a partition of S. By induction, it suffices


to establish the bound
!

X

f{s 1 ,...,s k }


si ∈Si ,...,sk ∈Sk
s1 ∈S1 ,...,si−1 ∈Si−1 L2r (X;H S1 ×···×Si−1 )
 

X
1/2 

≤ O(r) f{s1 ,...,sk } 

si+1 ∈Si+1 ,...,sk ∈Sk
s1 ∈S1 ,...,si ∈Si L2r (X;H S1 ×···×Si )
18 TERENCE TAO

for all 1 ≤ i ≤ k. But this follows by applying part (i) to the Hilbert
space H S1 ×···×Si−1 and the functions
 

f~si := 
X
f{s1 ,...,sk } 
si+1 ∈Si+1 ,...,sk ∈Sk
s1 ∈S1 ,...,si−1 ∈Si−1

for si ∈ Si , which one can verify to form a 2r-superorthogonal system


in L2r (X; H S1 ×···×Si−1 ). 

As a sample application of Theorem 2.3, we can specialize to the situ-


ation in Example 2.2 to conclude
Corollary 2.4 (Hoeffding-type inequality). Let the notation and hy-
potheses be as in Example 2.2. If we have the bound
X
|fA |2 ≤ σ 2 (2.3)
S
A∈(≤k ):A6=∅
almost surely for some σ > 0, then one has
 

X
fA − f∅ ≥ λσ  ≤ O(1)k exp(−ckλ2/k ) + exp(−cR)

P 
 
S

A∈(≤k )
(2.4)
for all λ > 0 and some absolute constant c > 0.

See [16], [22] for some previous Hoeffding-type inequalities for sums of
R-wise independent random variables.

Proof. We may normalize f∅ = 0. By shrinking λ if necessary we can


also assume that λ ≤ (R/k)k/2 (otherwise the first term on the right-
hand side is dominated by the second). We can also assume that λ ≥
C k and R ≥ Ck for a large constant C, as the bound is trivial otherwise.
Let 1 ≤ r ≤ R/2k be an integer to be chosen later. By Markov’s
inequality one has
  2r

X
 X
P  fA ≥ λσ  ≤ λ−2r σ −2r E fA .

S
S

A∈(≤k ) A∈(≤k )
Applying Theorem 2.3(ii) to the 2r-superorthogonal hypersystem (fA )A∈( S ) ,
≤k
we obtain
2r  r

X
 X
E fA ≤ O(r)kr E  |fA |2  .

S
S
A∈(≤k ) A∈(≤k )
IONESCU–WAINGER AND THE ADELES 19

Combining this with the preceding inequality and (2.3), we conclude


that  


 X
P  fA ≥ λσ  ≤ (O(r)/λ2/k )kr .

S

A∈(≤k )

If we set r := ⌊cλ2/k ⌋ for a sufficiently small absolute constant c > 0,


we obtain the claim. 

We now discuss the sharpness of the estimates in Corollary 2.4. The


first example below shows that the first term on the right-hand side of
(2.4) is reasonably sharp; the second example shows the second term
in (2.4) only has a small amount of room for improvement.
Example 2.5. Let n, k be positive integers, and let Xi,j for i =
1, . . . , k and j = 1, . . . , n be independent Bernoulli random variables
taking values in {−1, +1} of mean zero. Then the random variable
Qk Pn P
i=1 j=1 Xi,j can be expanded in the form A fA where fA = X1,j1 . . . Xk,jk
when A is of the form {(1, j1 ), . . . , (k, jk )} and fA = 0 otherwise. One
then easily verifies that (2.3) holds with σ = nk/2 , and that
 

X
P  fA ≥ λσ  = 2−nk
 
S

A∈(≤k )
when λ = nk/2 and S = {1, . . . , k} × {1, . . . , n}. Here one can take R
to be arbitrary. This shows that the first-term on the right-hand side
of (2.4) cannot be improved except possibly for the O(1)k factor or in
the explicit value of c. Modifications of this example can also be used
to illustrate the sharpness of Theorems 2.3; we leave the details to the
interested reader.
Example 2.6. Let R be a natural number, let p be a prime greater
than R, and let P be a random polynomial of degree at most R − 1
with coefficients in Z/pZ, drawn uniformly among all such polyno-
mials. Then the random variables P(i) for i = 1, . . . , p are R-wise
independent, since the Lagrange interpolation formula shows that for
any distinct i1 , . . . , iR , the map from polynomials P to evaluations
(P(i1 ), . . . , P(iR )) is a bijection. We then have
p !
X
P (1P(i)=0 − 1/p) = p − 1 = p−R


i=1

Comparing this with (2.4) with σ 2 = p, k = 1, and λ = p−1


√ , and taking
p
p comparable to CR log R, we see that the exp(−cR) term in (2.4)
cannot be improved to more than exp(−CR log R) for some constant C.
20 TERENCE TAO

One can construct similar


Qk examples
Pp for higher values of k by considering
the random variable j=1 i=1 (1Pj (i)=0 − 1/p) where P1 , . . . , Pk are
independent copies of P; we leave the details to the interested reader.

3. Sunflower bound

If k, r ∈ Z+ , let Sun(k, r) denote the smallest natural number with


the property that any family of Sun(k, r) distinct sets of cardinality at
most k contains r distinct elements A1 , . . . , Ar that form a sunflower
(as defined in Definition 1.9). The celebrated Erdős-Rado theorem [4]
asserts that Sun(k, r) is finite; in fact Erdős and Rado gave the bounds
(r − 1)k ≤ Sun(k, r) ≤ (r − 1)k k! + 1.
The sunflower conjecture asserts in fact that the upper bound can be
improved to Sun(k, r) ≤ O(1)k r k . This remains open at present; the
best bound known currently (in the regime where k, r are both large)
is
Sun(k, r) ≤ O(rLog(kr))k (3.1)
for all k, r ∈ Z+ , due to Rao [20] (building upon a recent breakthrough
of Alweiss, Lovett, Wu, and Zhang [1]).

We can give a probabilistic version of the Erdős-Rado theorem:


Lemma 3.1 (Probabilistic Erdős-Rado theorem). Let k, r ∈ Z+ , let S
be a finite set, and let Abe a random subset of S of cardinality k (i.e.,
a random element of Sk ). Let A1 , . . . , Ar be r independent copies of
A. Then with probability at least (4Sun(k, r))−r , A1 , . . . , Ar form a
sunflower.

Proof. If there is a set A ∈ Sk with P(A = A) ≥ (4Sun(k, r))−1, then




with probability at least (4Sun(k, r))−r we have A1 = · · · = Ar = A.


Since A, . . . , A is a sunflower, this gives the claim.

Now suppose that P(A = A) < (4Sun(k, r))−1 for all A ∈ Sk . We form

2Sun(k, r) independent samples A1 , . . . , A2Sun(k,r) of A. Consider the
event E that these samples only consist of at most Sun(k, r) distinct
sets. On this event, there are at most m≤Sun(k,r) 2Sun(k,r) ≤ 2Sun(k,r)
P
m
ways to a maximal collection Ai1 , . . . , Aim of distinct samples for some
m ≤ Sun(k, r); if we fix these indices i1 , . . . , im , we see from hypothesis
that each of the other samples Ai has a probability at most 1/4 of
matching one of these distinct samples. From the union bound, we
conclude that
P(E) ≤ 2Sun(k,r) (1/4)Sun(k,r) ≤ 1/2.
IONESCU–WAINGER AND THE ADELES 21

If we now condition to the complement of E, the samples A1 , . . . , A2Sun(k,r)


necessarily contain a sunflower, hence by symmetry and then undoing
the conditioning we see that A1 , . . . , Ar is a sunflower with probability
at least  −1
1 2Sun(k, r)
≥ 2−1 (2Sun(k, r))−r ,
2 r
giving the claim. 

In the converse direction, we can find a collection A1 , . . . , ASun(k,r)−1


of distinct sets of cardinality k, such that no distinct r elements in
this collection form a sunflower. If A1 , . . . , Ar are drawn uniformly
from this collection, then the probability that they form a sunflower
is then precisely (Sun(k, r) − 1)1−r (the probability that all the Ai
coincide). Thus the bound of (4Sun(k, r)−r in the above lemma cannot
be dramatically improved.

Lemma 3.1 lets us control square functions:


Corollary 3.2 (Sunflower bound on square function). Let S be a fi-
nite set, let k ∈ Z+ , let X be a measure space, and let H be a finite-
dimensional Hilbert space. Let (fA )A∈(S) be a finite collection of func-
k
tions fA ∈ L2r (X; H). Then
2
2r X X ∗∗ Y r
(f ) ≤ (4Sun(k, r))r kf k

A A∈(S) 2r A0 ∪Ai H
k L (X;H [k] ) 2
S
A0 ∈(≤k ) i=1 L (X)

and conversely
2
X X ∗∗ Y r 2r
kfA0 ∪Ai kH ≤ (fA )A∈(S ) ,

k L2r (X;H [k] )
S
A0 ∈(≤k ) i=1 L2 (X)
P∗∗
where  denotes the sum over tuples (A1 , . . . , Ar ) of sets A1 , . . . , Ar ∈
S\A0
k−|A0 |
that are pairwise disjoint (or equivalently, that A0 ∪A1 , . . . , A0 ∪
S
A form a sunflower), and H [k] := H (k ) .
r

See [7, Lemma 2.3] for a version of this result in the k = 1 case;

Proof. We begin with the first inequality. Expanding out both sides,
it suffices to establish the pointwise estimate
 r
∗∗ Yr
X 2  r
X X
 kfA (x)kH  ≤ (4Sun(k, r)) kfA0 ∪Ai (x)k2H
A∈(S S i=1
k)
A0 ∈(≤k )
(3.2)
22 TERENCE TAO

for all x.

Fix x. We may normalise the left-hand side of (3.2) to equal 1. We


can then view the sequence (kfA (x)k2H )A∈(S ) as the probability density
k
function for a random subset A of S of cardinality k, and the inequality
then can be written as
1 ≤ (4Sun(k, r))r P(A1 , . . . , Ar form a sunflower).
The claim now follows from Lemma 3.1. The second inequality simi-
larly follows from the trivial bound
P(A1 , . . . , Ar form a sunflower) ≤ 1.


4. Proof of main theorems

Let (d, k, S, ε) be a major arc parameter set. We now explore the


additive structure of the major arcs associated to this set. We first
observe from the Chinese remainder theorem that
ΣA1 + ΣA2 = ΣA1 ⊎A2 (4.1)
S

whenever A1 , A2 ⊆ S are disjoint. For A0 ∈ ≤k , we also define the
set [
Σ(A0 ) := ΣA .
S\A0
A∈(≤k−|A )
0|

From (4.1) we then have the inclusion


ΣA0 + Σ(A0 ) ⊆ Σ≤k . (4.2)

Let H be a finite dimensional Hilbert space. Define a major arc system


adapted to (d, k, S, ε) taking values in H to be a collection (fα )α∈Σ≤k of
functions fα ∈ ℓ2 (Zd ; H) with Fourier support in α + [−ε, ε]d for each
α ∈ Σ≤k , thus
d
fα ∈ ℓ2 (Zd ; H)α+[−ε,ε] .
For any Σ ⊆ Σ≤k , we define
X
fΣ := fα .
α∈Σ

Lemma 4.1 (Orthogonality properties). Let (fα )α∈Σ≤k be a major arc


system adapted to a major arc parameter set (d, k, S, ε), taking values
in a Hilbert space H. Suppose that the parameter set (d, k, S, ε) is
(r, c)-good for some r ∈ Z+ and 0 < c < 1.
IONESCU–WAINGER AND THE ADELES 23

(i) The major arcs α + [−ε, ε]d , α ∈ Σ≤k are disjoint. (Indeed, the
α ∈ Σ≤k are at least 2ε/c-separated in the ℓ∞ metric.)
(ii) (Denominator orthogonality) The hypersystem (fΣA )A∈( S ) is
≤k
Type II 2r-superorthogonal.
S

(iii) (Numerator orthogonality) If A1 , . . . , Ar ∈ ≤k form a sun-
flower with core A0 and petals A1 \A0 , . . . , Ar \A0 , then the func-
tions
Yr
fΣA0 +αi ∈ ℓ2 (Zd ; H ⊗r )
i=1

for α1 ∈ ΣA1 \A0 , . . . , αr ∈ ΣAr \A0 are pairwise orthogonal in the


Hilbert space L2 (Td ; H ⊗r ), where we use the product notation
r
Y
fi (x) := f1 (x) ⊗ · · · ⊗ fr (x).
i=1

Proof. Note from Definition 1.3 that the coefficients of every element of
k
Σ≤k is a rational number with denominator at most qmax . In particular,
if α, α are two distinct elements of Σ≤k , then α, α differ in ℓ∞ metric
′ ′

by at least qk1 . The claim (i) now follows (with room to spare) from
max
(1.4).

Now we prove (ii). From inspecting the Fourier transform, it suffices


to show that
Xr 2r
X
(αj + θj ) − (αj + θj ) 6= 0
j=1 j=r+1

in Td whenever αj ∈ ΣAj and θj ∈ [−ε, ε]d for j = 1, . . . , 2r. As the


A1 , . . . , A2r contain a nonce, there exists an A ⊆ S which contains all
but exactly one of the A1 , . . . , A2r . This implies that QA (α1 + · · · +
αr − αr+1 − · · · − α2r ) has precisely one non-zero term, and hence the
point α1 + · · ·+ αr − αr+1 − · · ·− α2r ∈ Td is non-zero. Observe that the
coordinates of this point consist of rational numbers of denominator at
1
most q2rk . The claim now follows from (1.4) and the triangle inequality.
max

Now we prove (iii). Inspecting the Fourier transform, it suffices to show


that
Xr X2r
(α0,j + αj + θj ) − (α0,j + αj + θj ) 6= 0
j=1 j=r+1

in Td whenever α0,j ∈ ΣA0 and θj ∈ [−ε, ε]d for j = 1, . . . , 2r, and


(α1 , . . . , αr ), (αr+1 , . . . , α2r ) ∈ ΣA1 \A0 × · · · × ΣAr \A0
24 TERENCE TAO

are distinct. Multiplying by QA0 to cancel the α0,j factors, it suffices


to show that
r
X 2r
X
QA0 (αj + θj ) − QA0 (αj + θj ) 6= 0.
j=1 j=r+1

0|A |
Note that QA0 ≤ qmax . On the other hand, from the sunflower hy-
pothesis and the Chinese remainder theorem we see that the point
QA0 (α1 + · · · + αr − αr+1 − · · · − α2r ) is non-zero. The coordinates of
1
this point consist of rational numbers of denominator at most 2r(k−|A 0 |)
,
qmax
and the claim now follows from (1.4) and the triangle inequality. 

We can exploit these orthogonality properties to obtain a description


of the ℓ2r norm of a sum fΣ≤k associated to a major arc system, as well
as a companion result that will be useful in the sequel.
Theorem 4.2 (Applying orthogonality). Let (d, k, S, ε) be a major arc
parameter set which is (r, c)-good for some r ∈ Z+ and 0 < c < 1. Let
H be a finite-dimensional Hilbert space.

(i) (Description of ℓ2r norm) If (fα )α∈Σ≤k is a major arc system


adapted to (d, k, S, ε), then we have
O(rLog1/2 (kr))−k kfΣ≤k kℓ2r (Zd ;H)
 1/2r
2r
 X
≤ (fα+ΣA0 )α∈Σ(A0 ) 2r d Σ(A ) 

ℓ (Z ;H 0 )
S
A0 ∈(≤k )
≤ Oc (1)d O(1)k kfΣ≤k kℓ2r (Zd ;H) .
(4.3)
(ii) (Rubio de Francia type estimate) Let ϕ0 ∈ Cc∞ (R) be a bump
function supported on [−1, 1], and let ϕ ∈ Cc∞ (Rd ) be the symbol
d  
Y ξj
ϕ(ξ1 , . . . , ξd ) = ϕ0 .
j=1
ε

Then for any 2 ≤ p ≤ ∞ and f ∈ ℓp (Zd ; H), one has the


inequality
 1/p
p
 X
(Tϕ;α+ΣA0 f )α∈Σ(A0 ) p d Σ(A )  ≤ Oϕ0 (1)d O(1)k kf kℓp (Zd ;H) .


ℓ (Z ;H 0 )
S
A0 ∈(≤k )
(4.4)
IONESCU–WAINGER AND THE ADELES 25

Proof. We begin with (ii), as this will be used in the proof of (i). By
interpolation it suffices to establish the claims for p = 2, ∞. For p = 2
the claim follows from Lemma 4.1(i) and Plancherel’s theorem, noting
that each Tϕ;α+ΣA0 f has Fourier transform supported in ΣA0 + α +
[−ε, ε]d , and each β ∈ Σ≤k k
 has at most 2 representations of the form
S
β = α + α0 with A0 ∈ ≤k , α0 ∈ ΣA0 , α ∈ Σ(A0 ) . For p = ∞, it suffices
by translation invariance to show that

≤ Oϕ0 (1)d O(1)k kf kℓ∞ (Zd ;H)

Tϕ;α+ΣA0 f (0) α∈Σ Σ

(A0 ) H (A0 )

S
for any f ∈ L∞ (Td ; H) and A0 ∈ ≤k

. By the inclusion-exclusion
k
formula, and conceding a factor of 2 , it suffices to show that
 
≤ Oϕ0 (1)d kf kℓ∞ (Zd ;H)

Tϕ;α+Σ ′ f (0)
⊆A
0 α∈Σ(A ) Σ
(A )
0 H 0

for any A′0 ⊆ A0 . By duality, this bound is equivalent to the assertion


that


X

≤ Oϕ0 (1)d k(cα )α∈Σ(A0 ) kH Σ(A0 )


cα Tϕ;α+Σ⊆A′ δ0

α∈Σ(A ) 0 1 d
0 ℓ (Z ;H)

for any sequence cα ∈ H, α ∈ Σ(A0 ) , where δ0 is the Kronecker delta


function. Observe that the integrand on the left-hand side is actually
supported on (QA′0 Z)d . By Cauchy-Schwarz (and (1.4)), it then suffices
to show that


d/2
X

≤ Oϕ0 (1)d ε−d/2 QA′ k(cα )α∈Σ(A0 ) kH Σ(A0 )

cα wTϕ;α+Σ⊆A′ δ0
0
0
α∈Σ(A ) 2 d
0 ℓ (Z ;H)

where w is the weight function


d
Y
w(n1 , . . . , nd ) := (1 + ε2 n2j ).
j=1

From Lemma 4.1(i) we see that the functions wT∗ϕ;α+Σ⊆A′ δ0 have dis-
0
joint Fourier supports and are thus pairwise orthogonal. Each such
function can be split into QdA′ orthogonal components wT∗ϕ;α+α0 δ0 with
0
α0 ∈ Σ⊆A′0 . By the Pythagorean theorem, it thus suffices to establish
the bound
kwT∗ϕ;α+α0 δ0 kℓ2 (Zd ) ≤ Oϕ0 (1)d εd/2
for each α ∈ Σ(A0 ) , α0 ∈ Σ⊆A′0 . The magnitude of the expression inside
the norm of the left-hand side does not actually depend on α + α0 , so
we may assume that α + α0 = 0. The left-hand side then factors as
26 TERENCE TAO

a tensor product and it now suffices to establish the claim for d = 1,


that is to say to show that
X
(1 + ε2 n2 )2 ε2 |ϕ̂0 (εn)|2 ≤ Oϕ0 (ε)
n∈Z

which follows from the rapid decrease of ϕ̂ (and noting from (1.4) that
ε ≤ 1). This completes the proof of (ii).

Now we prove (i). By Lemma 4.1(ii) and Theorem 2.3(ii), we have



kfΣ≤k kℓ2r (Zd ;H) ≤ O(r)k/2 (fΣA )A∈( S ) .

2r d [≤k] ≤k ℓ (Z ;H )

If we then apply Corollary 3.2 with the sunflower bound (3.1), and
apply the triangle inequality to sum in k, we conclude that
 1/2r
 X
kfΣ≤k kℓ2r (Zd ;H) ≤ O(rLog1/2 (kr))k  XA 0  (4.5)

S
A0 ∈(≤k)

where
∗∗∗
X r
Y
XA 0 := k fΣA0 ∪Ai k2ℓ2 (Zd ;H ⊗r )
i=1

and ∗∗∗
P
S\A0
 denotes a sum over tuples (A1 , . . . , Ar ) of disjoint sets A1 , . . . , Ar ∈
≤k−|A0 |
. For A0 , A1 , . . . , Ar as above, we can split
r
Y X r
Y
fA0 ∪Ai = fαi +ΣA0 .
i=1 α1 ∈A1 ,...,αr ∈Ar i=1

From Lemma 4.1(iii) and the Pythagorean theorem, we may thus write
∗∗∗
X X r
Y
XA 0 = k fαi +ΣA0 k2ℓ2 (Zd ;H ⊗r ) .
α1 ∈ΣA1 ,...,αr ∈ΣAr i=1
P∗∗∗
We drop the hypothesis of disjointness in the sum to obtain the
upper bound
X X r
Y
XA 0 ≤ k fαi +ΣA0 k2ℓ2 (Zd ;H ⊗r )
S\A0
A1 ,...,Ar ∈(≤k−|A
0|
) α1 ∈ΣA1 ,...,αr ∈ΣAr i=1

which by the Fubini–Tonelli theorem can be rearranged as


2r
XA0 ≤ (fα+ΣA0 )α∈Σ(A0 ) .

Σ(A )
ℓ2r (Zd ;H 0 )

This gives the first inequality in (4.3).


IONESCU–WAINGER AND THE ADELES 27

1+c
Now we establish the second inequality in (4.3). Let c′ := 2
, so that
c < c′ < 1. Let ϕ ∈ Cc∞ (Rd ) be a multiplier of the form
d
Y
ϕ(ξ1 , . . . , ξd) := ϕ0 (ξj /ε), (4.6)
j=1

where ϕ0 ∈ Cc∞ (R) is a fixed real even bump function (depending only
on c) supported on [−c′ /c, c′ /c] that equals 1 on [−1, 1]. From Lemma

4.1(i) (with c replaced by c′ , and ε replaced by cc ε) we have
fα+ΣA0 := Tϕ;α+ΣA0 fΣ≤k
and the claim now follows from (ii) (setting p = 2r). 

Now we can prove Theorem 1.4. Let the notation and hypotheses be as
in that theorem. We normalize kTm kB(L2r (Rd )) = 1 and kf kℓ2r (Zd ;H) = 1
(we can also assume by limiting arguments that f ∈ ℓ2 (Zd ; H) to avoid
technicalities), and our task is to show that


X
ǫA Tm;ΣA f ≤ Oc (1)d O(rLog1/2 (kr))k .


S

A∈(≤k ) 2r d
ℓ (Z ;H)

Applying Theorem 4.2(i) to the hypersystem (ǫA Tm;ΣA f )A∈( S ) , it suf-


≤k
fices to show that
 1/2r
2r

 X
(Tm;α+ΣA0 f )α∈Σ(A0 ) 2r d Σ(A )  ≤ Oc (1)d+k . (4.7)


ℓ (Z ;H 0 )
S
A0 ∈(≤k )
With ϕ as in (4.6), we may use Lemma 4.1(i) to factor
Tm;α+ΣA0 f = Tm;α+ΣA0 Tϕ;α+ΣA0 f.
Next, from the Magyar–Stein–Wainger sampling principle (Proposition
1.2) we have
kTm;ΣA0 F kℓ2r (Zd ;H) ≤ O(1)dkF kℓ2r (Zd ;H)
for any F ∈ ℓ2r (Zd ; H), hence by the Marcinkiewicz–Zygmund theorem
(Theorem 1.8) one has

d
(Tm;ΣA0 Fα )α∈Σ(A0 ) ≤ O(1) (F )

Σ(A ) α α∈Σ(A0 ) Σ(A )
ℓ2r (Zd ;H 0 ) ℓ2r (Zd ;H 0 )

for any Fα ∈ ℓ2r (Zd ; H), which by the modulation symmetries of the
Fourier transform imply that

d
(Tm;α+ΣA0 Fα )α∈Σ(A0 ) ≤ O(1) (Fα )α∈Σ(A0 ) .

Σ(A ) Σ(A )
ℓ2r (Zd ;H 0 ) ℓ2r (Zd ;H 0 )
28 TERENCE TAO

Putting all this together, we reduce to showing that


 1/2r
2r
 X
(Tϕ;α+ΣA0 f )α∈Σ(A0 ) 2r d Σ(A )  ≤ Oc (1)d O(1)k .


ℓ (Z ;H 0 )
S
A0 ∈(≤k )
But this follows from Theorem 4.2(ii). This concludes the proof of
Theorem 1.4.

Now we observe an arithmetic analogue of Theorem 4.2, in which the


spatial scale parameter ε becomes irrelevant:
Theorem 4.3 (Applying orthogonality, arithmetic limit). Let (d, k, S, ε)
be a major arc parameter set. Let H be a finite-dimensional Hilbert
space. For each α ∈ Σ≤k ,Plet fα ∈ S(AdZ ) have Fourier support in
Rd × {α}, and define fΣ := α∈Σ fα as before. Then for every positive
integer r, we have
O(rLog1/2 (kr))−k kfΣ≤k kL2r (AdZ ;H)
 1/2r
2r
 X
≤ (fα+ΣA0 )α∈Σ(A0 ) 2r d Σ(A ) 

L (AZ ;H 0 )
S
A0 ∈(≤k )
≤ O(1)d+k kfΣ≤k kL2r (AdZ ;H) .
(4.8)
Also, we have
kT1Σ≤k kB(L2r (Ẑd ;H)) ≤ O(1)d O(rLog1/2 (kr))k . (4.9)

Proof. One can establish (4.8) by direct repetition of the proof of The-
orem 4.2, but we shall instead deduce this theorem as a limiting case
of Theorem 4.2 (basically by sending ε to zero). By splitting AdZ into
fibres {x} × Ẑd for x ∈ Rd and using the Fubini–Tonelli theorem, it suf-
fices to establish the analogous claim for Ẑd , that is to say to establish
the bound
O(rLog1/2 (kr))−k kfΣ≤k kL2r (Ẑd ;H)
 1/2r
2r
 X
≤ (fα+ΣA0 )α∈Σ(A0 ) 2r d Σ(A ) 

L (Ẑ ;H 0 )
S
A0 ∈(≤k )
≤ O(1)d+k kfΣ≤k kL2r (Ẑd ;H)
(4.10)
where for each α ∈ Σ≤k , fα ∈ S(Ẑd ; H) has Fourier support in {α},
that is to say fα (y) = cα e(−y · α) for some cα ∈ H. Let 0 < ε < 1 be a
IONESCU–WAINGER AND THE ADELES 29

small parameter, let ϕ ∈ S(Rd ) be a Schwartz function whose Fourier


transform is supported in [−1, 1]d with normalization kϕkL2r (Rd ) = 1,
and let fα,ε ∈ S(Zd ; H) be the functions
fα,ε (n) := fα (ι̂(n))ϕ(εn) = cα e(−n · α)ϕ(εn)
where ι̂ : Zd → Ẑd is the canonical embedding. Then (fα,ε′ )α∈Σ≤k is a
major arc system adapted to (d, k, S, ε). For ε small enough, this set
of parameters is (r, 1/2)-good, and so we see from Theorem 4.2 that
O(rLog1/2 (kr))−k kfΣ≤k ,ε kℓ2r (Zd ;H)
 1/2r
2r

 X
≤ (fα+ΣA0 ,ε )α∈Σ(A0 ) 2r d Σ(A ) 

L (Z ;H 0 )
S
A0 ∈(≤k )
≤ O(1)d+k kfΣ≤k ,ε kℓ2r (Zd ;H)
(4.11)
where for any Σ ⊆ Σ≤k we denote
X
fΣ,ε (n) := fα,ε (n) = fΣ (ι̂(n))ϕ(εn).
α∈Σ

The functions fΣ ◦ ι̂ are all periodic with period QS . By Riemann


integrability one then has
ε1/2r kfΣ≤k ,ε kℓ2r (Zd ;H) → kfΣ≤k kℓ2r (Ẑd ;H)
and similarly

1/2r
ε (fα+ΣA0 ,ε )α∈Σ(A0 ) → (fα+ΣA0 )α∈Σ(A0 )

Σ(A ) Σ(A )
L2r (Zd ;H 0 ) L2r (Ẑd ;H 0 )

as ε′ → 0 for any A0 . Multiplying (4.11) by ε1/2r and taking the limit


ε → 0, we obtain the claim (4.8).

We now prove (4.9). Let F ∈ L2 (Ẑd ; H), then we have T1Σ≤k F = FΣ≤k ,
P
where Fα (y) := e(−y ·α)FẐd F (α) and FΣ := α∈Σ Fα for any Σ ⊆ Σ≤k .
By (4.8) we then have
 1/2r
2r
 X
kT1Σ≤k F kL2r (Ẑd ;H) ≤ O(rLog1/2 (kr))k  (Fα+ΣA0 )α∈Σ(A0 ) 2r d Σ(A ) 

L (Ẑ ;H 0 )
S
A0 ∈(≤k )
so it will suffice to establish the bound
 1/p
p

 X
(Fα+ΣA0 )α∈Σ(A0 ) p d Σ(A )  ≤ O(1)d kF kLp (Ẑd ;H)


L (Ẑ ;H 0 )
S
A0 ∈(≤k )
30 TERENCE TAO

for all 2 ≤ p ≤ ∞. By interpolation it suffices to establish this for p = 2


and p = ∞. The claim p = 2 is immediate from Bessel’s inequality.
For p = ∞ it suffices by translation invariance to show that
X
( kFα+ΣA0 (0)k2H )1/2 ≤ O(1)dkF kL∞ (Ẑ;H)
α∈Σ(A0 )

which by duality is equivalent to the assertion that



Z X X

X
dµ d (y) ≤ O(1)d ( kcα k2H )1/2


c α e(−y · (α + α 0 )) Ẑ
Ẑd α∈Σ α0 ∈ΣA0 α∈Σ(A )
(A )
0 0
H
for any cα ∈ H for α ∈ Σ(A0 ) .

Observe that the integrand vanishes unless the projection of y to (Z/QA0 Z)d
vanishes, thus the integrand is supported on a set of measure Q−d A0 . By
Cauchy-Schwarz, it thus suffices to show that
2
Z X X

X
d d
kcα k2H .


c α e(−y · (α + α ))
0
dµ Ẑd (y) ≤ O(1) QA0
d
Ẑ α∈Σ
(A )
0
α0 ∈ΣA0 α∈Σ(A ) 0
H

But this is immediate from Plancherel’s theorem since |ΣA0 | = QdA0 . 

Now we can prove Theorem 1.6. To abbreviate the notation we write


X / Y for
X ≤ exp(Oc (d) + O(kLog(rLogk)))Y
and X ≈ Y for X / Y / X.

We first establish the claim in the case p = 2r. By a limiting ar-


2 d π(Ω)
gument
P we may assume that f ∈ ℓ (Z ; H) , thus we can write
f = α∈Σ≤k fα where
Z
fα (n) := e(−n · (α + θ))FZd f (α + θ) dθ.
[−ε,ε]d

We then have SΩ−1 f =


P
Fα where
α∈Σ≤k
Z
Fα (x, y) := e(−y · α) e(−x · θ)FZd f (α + θ) dθ.
[−ε,ε]d
P P
Writing fΣ := α∈Σ fα and FΣ := α∈Σ Fα for any Σ ⊆ Σ≤k , we see
from Theorem 4.2(i) (and bounding rLog1/2 (kr) = exp(O(Log(rLogk))))
that
 1/2r
 X
kf kℓ2r (Zd ;H) ≈  k(fα+ΣA0 )α∈Σ(A0 ) k2r2r d Σ(A0 ) 

ℓ (Z ;H )
S
A0 ∈(≤k )
IONESCU–WAINGER AND THE ADELES 31

and similarly from Theorem 4.3 that


 1/2r
 X
kSΩ−1 f kL2r (AdZ ;H) ≈  k(Fα+ΣA0 )α∈Σ(A0 ) k2r2r d Σ(A0 ) 

L (AZ ;H )
S
A0 ∈(≤k )
so it will suffice to show that
k(fα+ΣA0 )α∈Σ(A0 ) kℓ2r (Zd ;H Σ(A0 ) ) ≈ k(Fα+ΣA0 )α∈Σ(A0 ) kL2r (Ad ;H Σ(A0 ) )
Z
S

for each A0 ∈ ≤k . From expanding the definitions, we see that
d ×Σ
(Fα+ΣA0 )α∈Σ(A0 ) ∈ L2 (AdZ ; H Σ(A0 ) )[−ε,ε] A0

and
(fα+ΣA0 )α∈Σ(A0 ) = S(Fα+ΣA0 )α∈Σ(A0 ) .
The claim now follows from Proposition 1.5 and Lemma 4.1(i).

Now we establish Theorem 1.6 for general (2r)′ ≤ p ≤ 2r. We begin


with the upper bound
kSΩ−1 f kLp (AdZ ;H) / kf kℓp (Zd ;H)
for f ∈ ℓp (Zd ; H)π(Ω) . With ϕ as in (4.6), we can write
SΩ−1 f = SΩ−1′ Tϕ;Σ≤k f
′ ′
where Ω := [− cc ε, cc ε]d × Σ≤k . Note that the right-hand side is well
defined for all f in ℓp (Zd ; H) (with no restriction on the Fourier support
on f ). Thus it will suffice to show that
kSΩ−1′ Tϕ;Σ≤k kB(ℓp (Zd ;H)→Lp (AdZ ;H)) / 1 (4.12)
for all (2r)′ ≤ p ≤ 2r. Now that the Fourier restriction has been
removed, interpolation becomes available, and it suffices to establish
this bound for p = 2r, (2r)′. For p = 2r the claim follows from the
p = 2r case of Theorem 1.6 already established (with c replaced by c′ ),
noting from Theorem 1.4 that
kTϕ;Σ≤k kB(ℓp (Zd ;H)) / 1 (4.13)
for p = 2r (indeed, this estimate holds for all (2r)′ ≤ p ≤ 2r by duality
and interpolation).

For p = (2r)′, we apply duality to write the estimate in the equivalent


form
kTϕ;Σ≤k SkB(L2r (AdZ ;H)→ℓ2r (Zd ;H)) / 1. (4.14)
From (1.5) we have
Tϕ;Σ≤k S = STϕ⊗1Σ≤k .
From Theorem 4.3 one has
kTϕ⊗1Σ≤k kB(L2r (AdZ ;H)) = kTϕ kB(L2r (Rd ;H)) kT1Σ≤k kB(L2r (Ẑd ;H)) / 1
32 TERENCE TAO

and the claim now follows from the p = 2r case of Theorem 1.6 already
established (with c replaced by c′ ).

Now we obtain the lower bound


kSΩ−1 f kLp (AdZ ;H) ' kf kℓp (Zd ;H)

for f ∈ ℓp (Zd ; H)π(Ω) . This is equivalent to


kSF kℓp (Zd ;H) / kF kLp (AdZ ;H)

for F ∈ Lp (Zd ; H)Ω . For such F we have SF = STϕ⊗1Σ≤k F , so it


suffices to show that
kSTϕ⊗1Σ≤k kB(Lp (AdZ ;H)→ℓp (Zd ;H)) / 1

for all (2r)′ ≤ p ≤ 2r. By interpolation it suffices to establish this


bound for p = 2r, (2r)′ . For p = 2r the claim follows from (4.14). For
p = (2r)′ we dualize to
kTϕ⊗1Σ≤k SΩ−1′ kB(ℓ2r (Zd ;H)→L2r (AdZ ;H)) / 1

and the claim now follows from (4.12), (1.5). This concludes the proof
of Theorem 1.6 for general p.

Now we can prove Theorem 1.7. Let the notation and hypotheses be
as in that theorem, and as before let ϕ be the function (4.6). Then by
(1.5) (and Lemma 4.1(i)) we can factorize
Tm;Σ≤k = Tm;Σ≤k Tϕ;Σ≤k = STm SΩ−1′ Tϕ;Σ≤k .
The claim now follows from (4.13) and Theorem 1.6.

5. The Ionescu–Wainger major arc construction

We now describe the specific choice of major arcs that essentially ap-
pears in the original work [7] of Ionescu and Wainger, as well as in
many subsequent works.
Lemma 5.1. Let 0 < ρ < 1 be a parameter, and set k := ⌊ 2ρ ⌋ + 1.
Suppose that N ≥ 2k . Then there exists a set S of pairwise coprime
natural numbers such that for any d ∈ Z+ , and ε > 0, the major arc
parameter set (d, k, S, ε) obeys the following properties:

(i) One has Td [q] ⊆ Σ≤k for all natural numbers 1 ≤ q ≤ N.


(ii) One has Σ≤k ⊆ Td [Q] for some Q ≤ 3N .
IONESCU–WAINGER AND THE ADELES 33

ρ/2
(iii) All elements of S are bounded by C kN for some absolute con-
stant C > 1. In particular, (d, k, S, ε) will be (r, 12 )-good when-
ever
1
ε< . (5.1)
4rC 2rk2 N ρ/2
(iv) Σ≤k is the union of finitely many subgroups of Td , each of the
2 ρ/2
form Td [q] for some q ≤ O(1)k N . In particular, |Σ≤k | ≤
2 ρ/2
O(1)dk N .

Typically ρ (and hence k) and r will be fixed in applictaions. For


N sufficiently large depending on ρ, r, d, the condition (5.1) can be
2 ρ/2
simplified to ε ≤ exp(−N ρ ), and the bounds q ≤ O(1)k N , |Σ≤k | ≤
2 ρ/2
O(1)dk N in (iv) can be similarly simplified to q, |Σ≤k |S≤ exp(N ρ ).
The main point here is we can cover the Farey sequence 1≤q≤N Td [q]
by good major arcs whose width ε can be as large as exp(−N ρ ).

Proof. We set S equal to


Y ⌊ log N ⌋ log N
S := { p log p } ∪ {p⌊ log p ⌋ : N ρ/2 < p ≤ N}
p≤N ρ/2

where p is always understood to be restricted to the primes. Clearly


the elements of S are pairwise coprime. To prove (i), we have to show
that every natural number 1 ≤ q ≤ N is a factor of a product of at
most k distinct elements from S. But by the fundamental theorem of
arithmetic we can write q = pa11 . . . pamm for some primes 1 < p1 < · · · <
pm ≤ N and 1 ≤ ai ≤ ⌊ log N
log pi
⌋. At most ⌊ ρ2 ⌋ = k − 1 of these primes can
log N
exceed N ρ/2 . One can then write q as a factor of p≤N ρ p⌊ log p ⌋ times
Q
log N
at most k − 1 terms of the form p⌊ log p ⌋ , giving the claim.

The product QS of all the elements of S is equal to


Y ⌊ log N ⌋
p log p = lcm(1, . . . , N) ≤ 3N
p≤N

where the latter inequality is established in [5]. Since Σ≤k ⊆ Td [QS ],


this gives (ii).
log N
For (iii), we trivially bound p⌊ log p ⌋ by N, and note from the prime num-
ber theorem that the number of primes less than N ρ/2 is O(N ρ/2 / log N ρ/2 ) =
O(kN ρ/2 / log N), giving (iii) as claimed (noting from the hypothesis
ρ/2 ρ/2 ρ/2
N ≥ 2k that N ρ/2 ≥ 21/2 and hence N ≤ N O(N / log N ) = O(1)kN ).
34 TERENCE TAO

Finally to prove (iv), note from definition that Σ≤k is the union of Td [q]
where q is the product of at most k elements of S, and the claim now
follows from (iii). 

As a particular corollary of this construction, we can prove a sampling


theorem for the classical major arcs.
Corollary 5.2 (Classical major arc sampling). Let d, N ∈ Z+ , ε > 0,
and set
[N
d
Ω := [−ε, ε] × Td [Q].
q=1

Let 0 < ρ < 1 and 1 < p < ∞ be such that


ε < exp(−C max(p, p′ )ρ−2 N ρ/2 ) (5.2)
for a sufficiently large absolute constant C. Then for any finite-dimensional
Hilbert space H, one has
kSΩ−1 f kLp (AdZ ;H) = exp(O(d + ρ−1 Log(max(p, p′ )Logρ−1 )))kf kℓp (Zd ;H)

for all f ∈ ℓp (Zd ; H)π(Ω) .

Proof. Set r to be the first natural number such that (2r)′ ≤ p ≤ 2r,
then r ∼ max(p, p′ ). Let k := ⌊ ρ2 ⌋ + 1 ∼ ρ−1 , and let S be the set
constructed by Lemma 5.1, then Ω ⊆ [−ε, ε]d × Σ≤k . If the constant
C in (5.2) is large enough, Lemma 5.1(iii) ensures that (d, k, S, ε) is
(r, 21 )-good, and the claim now follows from Theorem 1.6. 

Appendix A. Abstract harmonic analysis

We define the Pontryagin dual pairs (G, G∗ ) of LCA groups used in


this paper.

(i) If G = R with Lebesgue measure µR = dx, then G∗ = R∗ = R


with Lebesgue measure µR∗ = dξ is a Pontryagin dual, with
pairing x · ξ := xξ mod 1.
(ii) If G = Z with counting measure µZ , then G∗ = T with Lebesgue
measure µT = dξ is a Pontryagin dual, with pairing x · ξ := xξ.
(iii) If G = Z/QZ is a cyclicR group for some Q ∈ Z+ with normal-
ized counting measure Z/QZ f (x) dµZ/QZ (x) := Ex∈Z/QZ f (x),
then the dual cyclic group G∗ = T[Q] = Q1 Z/Z with counting
measure µT[Q] is a Pontryagin dual, with pairing x · ξ := xξ.
IONESCU–WAINGER AND THE ADELES 35

(iv) If G = Ẑ := lim Z/QZ is the compact group of profinite integers


←−
with Haar probability measure (using the projection maps from
Z/QZ to Z/qZ whenever q divides Q), then the discrete group
G∗ = Ẑ∗ = Q/Z of “arithmetic frequencies” with counting mea-
sure µQ/Z is a Pontragin dual, with pairing x · ( aq mod 1) :=
xa mod q
q
.
(v) If G1 , G2 are LCA groups with Pontryagin duals G∗1 , G∗2 , then
the product G1 × G2 (with product Haar measure) is an LCA
group with Pontryagin dual G∗1 ×G∗2 and pairing (x1 , x2 )·(ξ1 , ξ2 ) :=
x1 · ξ1 + x2 · ξ2 . In particular, if G = AdZ = Rd × Ẑd is the dth
power of the adelic integers AZ := R × Ẑ2 (with the product
Haar measure µAdZ := µdR × µdẐ ), then adelic frequency space
G∗ = (AdZ )∗ = Rd × (Q/Z)d is a Pontryagin dual (with product
measure µR×Q/Z := µdR × µdQ/Z and the indicated pairing.

More explicitly: an element of AdZ is of the form (x, y), where x =


(x1 , . . . , xd ) ∈ Rd and y = (y1 , . . . , yd ) ∈ Ẑd , thus yd mod Q is an
element of Z/QZ for any positive integer Q (with the compatibility
conditions yd mod q = (yd mod Q) mod q whenever q divides Q), and
if (ξ, η) = (ξ1 , . . . , ξd , aq1 mod 1, . . . , aqd mod 1) is an element of the dual
group Rd × (Q/Z)d , then
(x, y) · (ξ, η) = x · ξ + y · η
a1 y1 mod q + · · · + ad yd mod q
= x1 ξ1 + · · · + xd ξd + .
q
We have the canonical inclusion ι : Zd → AdZ defined by
ι(n) := (n, (n mod Q)Q∈Z+ )
and the projection map π : Rd × (Q/Z)d → Td defined by
π(θ, α) := α + θ;
the two maps enjoy the Fourier adjoint relationship
n · π(θ, α) = ι(n) · (θ, α)
for all n ∈ Z and (θ, α) ∈ Rd × (Q/Z)d .
d

We define the following Schwartz-Bruhat spaces S(G) on various LCA


groups3 G:

(i) S(Rd ) is the space of Schwartz functions on Rd .


2The adelic integers AZ should not be confused with the larger ring AQ = AZ ⊗Z Q
of adelic numbers, which we will not use in this paper.
3For a definition of Schwartz-Bruhat spaces on arbitrary LCA groups, see [3],
[15].
36 TERENCE TAO

(ii) S(Zd ) is the space of rapidly decreasing functions on Zd , and


S(Td ) is the space of smooth functions on T.
(iii) S(Ẑd ) is the space of locally constant functions f on Ẑd , or
equivalently those functions of the form f (x) = fQ (x mod Q)
for some Q ∈ Z+ and fQ : (Z/QZ)d → C. S((Q/Z)d ) is the
space of finitely supported functions on (Q/Z)d .
(iv) S(AdZ ) is the space of functions of the form f (x, y) = fQ (x, y mod
Q) for some Q ∈ Z+ and fQ : Rd × (Z/QZ)d that is Schwartz in
the R variable. S(Rd × (Q/Z)d ) is the space of functions sup-
ported on Rd × Σ for some finite set Σ ⊂ (Q/Z)d and Schwartz
in the Rd variable.

Appendix B. Losses in the sampling principle

In this appendix we demonstrate some losses in the Maygar–Stein–


Wainger sampling principle (Proposition 1.2) that demonstrate that
the O(1)d factor in (1.2) is necessary, at least when p is close to 1 or
∞.
Proposition B.1 (Losses in the sampling principle). Let 1 ≤ p ≤ ∞
be sufficiently close to 1 or ∞. Then there is a constant C > 1 such
that for any d ∈ Z+ , there exists a multiplier m ∈ Cc∞ (Rd ) is supported
in [− 21 , 12 ]d for which
kTm;0 kB(ℓp (Zd )) > C d kTm kB(Lp (Rd )) . (B.1)

Proof. By duality we may assume p sufficiently close to 1.

We first observe that it will suffice to establish the claim for d = 1, since
if we can find a one-dimensional multiplier m ∈ Cc∞ (R) supported in
[− 21 , 21 ] with
kTm;0 kB(ℓp (Z)) > CkTm kB(Lp (R))
then by taking tensor products (and many applications of the Fubini–
Tonelli theorem) we see that m⊗d ∈ Cc∞ (Rd ) is supported in [− 12 , 12 ]d
with Tm⊗d the d-fold tensor product of Tm , and Tm⊗d ;0 similarly the
d-fold tensor product of Tm;0 , and hence (by many further applications
of Fubini–Tonelli)
kTm⊗d ;0 kB(ℓp (Zd )) = kTm;0 kdB(ℓp (Zd )) > C d kTm kdB(Lp (R)) = C d kTm⊗d kB(Lp (Rd ))
giving the claim.

It remains to establish the claim for d = 1. Suppose for contradiction


that the claim fails, then there exists a sequence of p > 1 converging
IONESCU–WAINGER AND THE ADELES 37

to 1 such that we have


kTm;0 kB(ℓp (Z)) ≤ kTm kB(Lp (R))
for all m ∈ Cc∞ (R) supported on [−1/2, 1/2]. From the Riesz-Thorin
theorem, the B(Lp (R)) norm is continuous in p, thus on taking limits
we conclude that
kTm;0 kB(ℓ1 (Z)) ≤ kTm kB(L1 (R)) .
But the multiplier operator Tm is convolution with FR−1 m, hence
kTm kB(L1 (R)) = kFR−1 mkL1 (R) .
Similarly, from the Poisson summation formula Tm;0 is convolution with
the restriction of FR−1m to the integers, thus we have
kTm;0 kB(ℓ1 (Z)) = kFR−1 m|Z kℓ1 (Z) ,
thus Z
X
|FR−1m(n)| ≤ |FR−1 m(ξ)| dξ
n∈Z R

for all m ∈ Cc (R) supported on [−1/2, 1/2]. Since the class of m is
invariant under modulations we conclude that
X Z
−1
|FR m(n + θ)| ≤ |FR−1 m(ξ)| dξ
n∈Z R

for all 0 ≤ θ ≤ 1. However we have


Z 1X Z
−1
|FR m(n + θ)|dθ = |FR−1m(ξ)| dξ
0 n∈Z R

and the integrand on the left-hand side is continuous in θ, thus we have


X Z
−1
|FR m(n + θ)| = |FR−1 m(ξ)| dξ
n∈Z R

for all m ∈ Cc∞ (R) supported on [−1/2, 1/2] and 0 ≤ θ ≤ 1. In


particular Z
X
−1
|FR m(n)| = |FR−1 m(ξ)| dξ. (B.2)
n∈Z R

If we formally set m = 1[−1/2,1/2] , then FR−1 m(ξ) = sin(πξ)


πξ
, and the
left-hand side evaluates to 1 and the right-hand side is logarithmically
divergent. Of course, the indicator function 1[−1/2,1/2] is not smooth,
but by applying a standard mollification to this function one can then
easily construct a counterexample to (B.2), giving the claim. 

With some numerical effort one could refine the above argument to
obtain an explicit range p ∈ [1, 1 + ε] ∪ [(1 + ε)′ , ∞] of exponents p
and an explicit constant C > 1 for which the above proposition holds,
but we will not do so here. It seems likely that this proposition in fact
38 TERENCE TAO

holds for all p 6= 2 (presumably with a constant C that converges to


1 as p → 2), but constructing a concrete counterexample may require
locating a multiplier m for which both the continuous operator norm
kTm kB(Lp (R)) and the discrete operator norm kTm;0 kB(ℓp (Z)) are explic-
itly computable (or at least bounded above and below with extreme
precision), and examples of such multipliers are rare. Some variant
of the continuous and discrete Hilbert transforms seem like the most
promising candidates for this task; see [2] for some recent progress in
this direction.

References

[1] R. Alweiss, S. Lovett, K. Wu, J. Zhang, Improved bounds for the sunflower
lemma, STOC 2020: Proceedings of the 52nd Annual ACM SIGACT Sympo-
sium on Theory of Computing, June 2020, 624–630.
[2] R. Bañuelos, M. Kwaśnicki, On the ℓp -norm of the discrete Hilbert transform,
Duke Math. J. 168 (2019), 471–504.
[3] F. Bruhat, Distributions sur un groupe localement compact et applications à
l’t́ude des représentations des groupes ℘-adiques, Bull. Soc. Math. France 89
(1961), 43–75.
[4] P. Erdős, R. Rado, Intersection theorems for systems of sets, Journal of London
Mathematical Society 35 (1960), 85–90.
[5] D. Hanson. On the product of the primes, Canad. Math. Bull. 15 (1972), 33–37.
[6] A. Ionescu, E. Stein, A. Magyar, S. Wainger, Discrete Radon transforms and
applications to ergodic theory, Acta Math. 198 (2007), no. 2, 231–298.
[7] A. D. Ionescu, S. Wainger, Lp boundedness of discrete singular Radon trans-
forms, J. Amer. Math. Soc. 19 (2005), no. 2, pp. 357–383.
[8] B. Krause, M. Mirek, T. Tao, Pointwise ergodic theorems for non-conventional
bilinear polynomial averages, preprint.
[9] A. Magyar, E. M. Stein, S. Wainger, Discrete analogues in harmonic analysis:
spherical averages, Ann. Math. 155 (2002), pp. 189–208.
[10] J. Marcinkiewicz, A. Zygmund, Quelques inegalités pour les opérations
linéaires, J. Marcinkiewicz Collected Papers edited by A. Zygmund, Warsaw
(1964), 541–546.
[11] M. Mirek, Square function estimates for discrete Radon transforms, Analysis
& PDE 11 (2018), no. 3, 583–608.
[12] M. Mirek, E. M. Stein, B. Trojan, ℓp (Zd )-estimates for discrete operators of
Radon type: variational estimates, Invent. Math. 209 (2017), no. 3, 665–748.
[13] M. Mirek, E. M. Stein, B. Trojan, ℓp (Zd )-estimates for discrete operators of
Radon type: maximal functions and vector-valued estimates, J. Funct. Anal.
277 (2019), no. 8, 2471–2521.
[14] M. Mirek, E. M. Stein, P. Zorin-Kranich, Jump inequalities for translation-
invariant operators of Radon type on Zd , Adv. Math. 365 (2020), 107065, 57
pp.
[15] M. S. Osborne, On the Schwartz–Bruhat space and the Paley–Wiener theorem
for locally compact abelian groups, J. Functional Analysis 19 (1975), pp. 40–
49.
[16] B. Pass, S. Spektor, On Khintchine Type Inequalities for k-wise independent
Rademacher random variables, Statist. Probab. Lett. 132 (2018), 35–39.
IONESCU–WAINGER AND THE ADELES 39

[17] V. de la Peña, E. Giné, Decoupling, From dependence to independence. Ran-


domly stopped processes. U-statistics and processes. Martingales and beyond.
Probability and its Applications (New York). Springer-Verlag, New York, 1999.
[18] L. Pierce, Discrete fractional Radon transforms and quadratic forms, Duke
Math. J. 161 (2012), no. 1, 69–106.
[19] L. Pierce, On superorthogonality, preprint. arXiv:2007.10249
[20] A. Rao, Coding for Sunflowers, Discrete Analysis, 2020:2, 8 pp.
[21] J. L. Rubio de Francia, A LittlewoodPaley inequality for arbitrary intervals,
Revista Matematica Iberoamericana 1 (1985), 1–14.
[22] J. Schmidt, A. Siegel, A. Srinivasan, Chernoff-Hoeffding bounds for applica-
tions with limited independence, Proceedings of the Fourth Annual ACM-SIAM
Symposium on Discrete Algorithms (Austin, TX, 1993), 331340, ACM, New
York, 1993.

UCLA Department of Mathematics, Los Angeles, CA 90095-1555.

E-mail address: tao@math.ucla.edu

You might also like