Annals of Mathematics 182 (2015), 351–389

http://dx.doi.org/10.4007/annals.2015.182.1.9

The proof of the l2 Decoupling Conjecture

By Jean Bourgain and Ciprian Demeter

Abstract

We prove the l2 Decoupling Conjecture for compact hypersurfaces with

positive definite second fundamental form and also for the cone. This has
a wide range of important consequences. One of them is the validity of the
Discrete Restriction Conjecture, which implies the full range of expected
Lpx,t Strichartz estimates for both the rational and (up to N ε losses) the ir-
rational torus. Another one is an improvement in the range for the discrete
restriction theorem for lattice points on the sphere. Various applications
to Additive Combinatorics, Incidence Geometry and Number Theory are
also discussed. Our argument relies on the interplay between linear and
multilinear restriction theory.

1. The l2 Decoupling Theorem

Let S be a compact C 2 hypersurface in Rn with positive definite sec-
ond fundamental form. Examples include the sphere S n−1 and the truncated
(elliptic) paraboloid
P n−1 := {(ξ1 , . . . , ξn−1 , ξ12 + · · · + ξn−1
2
) ∈ Rn : |ξi | ≤ 1/2}.
Unless specified otherwise, we will implicitly assume throughout the whole
paper that n ≥ 2. We will write A ∼ B if A . B and B . A. The implicit
constants hidden inside the symbols . and ∼ will in general depend on fixed
parameters such as p, n and sometimes on variable parameters such as ε, ν.
We will not record the dependence on the fixed parameters.
Let Nδ be the δ neighborhood of P n−1 , and let Pδ be a finitely overlapping
cover of Nδ with curved regions θ of the form
(1) θ = {(ξ1 , . . . , ξn−1 , η + ξ12 + · · · + ξn−1
2
) : (ξ1 , . . . , ξn−1 ) ∈ Cθ , |η| ≤ 2δ},
1/2 1/2
where Cθ runs over all cubes c + [− δ 2 , δ 2 ]n−1 with
δ 1/2 n−1
c∈ Z ∩ [−1/2, 1/2]n−1 .
2

The first author is partially supported by the NSF grant DMS-1301619. The second author
is partially supported by the NSF Grant DMS-1161752.
c 2015 Department of Mathematics, Princeton University.

351
352 JEAN BOURGAIN and CIPRIAN DEMETER

Note that each θ sits inside a ∼ δ 1/2 × · · · × δ 1/2 × δ rectangular box. It is

also important to realize that the normals to these boxes are ∼ δ 1/2 separated.
A similar decomposition exists for any S as above, and we will use the same
notation Pδ for it. We will denote by fθ the Fourier restriction of f to θ.
Our main result is the proof of the following l2 Decoupling Theorem.
Theorem 1.1. Let S be a compact C 2 hypersurface in Rn with positive
definite second fundamental form. If supp(fˆ) ⊂ Nδ , then for p ≥ 2(n+1)
n−1 and
ε > 0,
Ç å1/2
− n−1 + n+1 −ε
X
(2) kf kp .ε δ 4 2p kfθ k2p .
θ∈Pδ
8 4
Theorem 1.1 has been proved in [21] for p > 2 + n−1 − n(n−1) . A standard
−ε
construction is presented in [21] to show that, up to the δ term, the exponent
of δ is optimal. We point out that Wolff [36] has initiated the study of lp
decouplings, p > 2 in the case of the cone. His work provides part of the
inspiration for our paper.
A localization argument and interpolation between p = 2(n+1) n−1 and the
trivial bound for p = 2 proves the subcritical estimate
Ç å1/2
−ε
X
(3) kf kp .ε δ kfθ k2p ,
θ∈Pδ

when 2 ≤ p < 2(n+1)

n−1 . Estimate (3) is false for p < 2. This can easily be seen
by testing it with functions of the form fθ (x) = gθ (x + cθ ), where supp(g“θ ) ⊂ θ
and the numbers cθ are very far apart from each other.
2n
Inequality (3) has been recently proved by the first author for p = n−1 in
[11], using a variant of the induction on scales from [14] and the multilinear
restriction Theorem 6.1.
An argument similar to the one in [21] was used in [18] to prove Theo-
rem 1.1 for p > 2(n+2)
n−1 , by interpolating Wolff’s machinery with the estimate
2n
p = n−1 from [11]. This range is better than the one in [21] due to the use of
multilinear theory as opposed to bilinear theory.1
We mention briefly that there is a stronger form of decoupling, sometimes
referred to as square function estimate, which predicts that
kf kp .ε δ −ε k(
X
(4) |fθ |2 )1/2 kp ,
θ∈Pδ
2n
in the slightly smaller range 2 ≤ p ≤ n−1 .
When n = 2, this easily follows via a
geometric argument. Minkowski’s inequality shows that (4) is indeed stronger

1
While both the bilinear theorem in [32] and the multilinear theorem in [3] are sharp, the
latter one is “morally” stronger.
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 353

2n
than (3) in the range 2 ≤ p ≤ n−1 . This is also confirmed by the lack of any
results for (4) when n ≥ 3. Our methods do not seem to enable any progress
on (4).
It is reasonable to hope that in the subcritical regime (3) one may be able
to replace δ −ε by a constant Cp,n independent of δ. This is indeed known when
n = 2 and p ≤ 4, but it seems in general to be an extremely difficult question.
To the authors’ knowledge, no other examples of 2 < p < 2(n+1)n−1 are known for
when this holds.
In Section 5 we introduce a multilinear version of the decoupling inequal-
ity (2) and show that the multilinear and the linear theories are essentially
equivalent. This in itself is not enough to prove Theorem 1.1, as Theorem 6.1
2n
gives multilinear decoupling only in the range 2 ≤ p ≤ n−1 . To bridge the
gap between n−12n
and 2(n+1)
n−1 , in Section 6 we refine our analysis based on the
multilinear theory. In particular, we set up an induction on scales argument
that makes use of Theorem 6.1 at each step of the iteration, rather than once.
Let us now briefly describe some of the consequences of Theorem 1.1. The
first one we mention is a sharp decoupling for the (truncated) cone
n » » o
C n−1 = (ξ1 , . . . , ξn−1 , ξ12 + · · · + ξn−1
2 ), 1 ≤ ξ12 + · · · + ξn−1
2 ≤2 .

Abusing earlier notation, we let Nδ (C n−1 ) be the δ neighborhood of C n−1 , and

we let Pδ (C n−1 ) be the partition of Nδ (C n−1 ) associated with a given partition
of S n−1 into δ 1/2 caps. More precisely, each θ ∈ Pδ (C n−1 ) is essentially a
1 × δ × δ 1/2 × · · · × δ 1/2 rectangular box.
Theorem 1.2. Assume supp(fˆ) ⊂ Nδ (C n−1 ). Then for each ε > 0,
Ç å1/2
− n−2 n
+ 2p −ε
X 2n
kf kp .ε δ 4 kfθ k2p if p ≥
n−2
θ∈Pδ (C n−1 )
and
Ç å1/2
−ε
X 2n
kf kp .ε δ kfθ k2p if 2 ≤ p ≤ .
n−2
θ∈Pδ (C n−1 )

The proof of Theorem 1.2 is presented in the last section of the paper, and
it turns out to be a surprisingly short application of Theorem 1.1 for the elliptic
paraboloid. It has some striking consequences, some of which were described
in (and provided some of the original motivation for) the work [36] of Wolff.
Examples include progress on the “local smoothing conjecture for the wave
equation” (see [30] and [36]), the regularity for convolutions with arclength
measures on helices [27], and the boundedness properties of the Bergman pro-
jection in tube domains over full light cones; see [20] and [2]. We refer the
interested reader to these papers for details.
354 JEAN BOURGAIN and CIPRIAN DEMETER

Theorem 1.1 immediately implies the validity of the Discrete Restriction

Conjecture in the expected range; see Theorem 2.2 below. This in turn has a
wide range of interesting consequences that are detailed in Section 2. First, we
get the full range of expected Lpx,t Strichartz estimates for both the rational
and (up to N ε losses) irrational tori. Second, we derive sharp estimates on the
additive energies of various sets. These can be rephrased as incidence geometry
problems, and in some cases we are not aware of an alternative approach. While
our theorems successfully address the case of “nicely separated” points, some
intriguing questions are left open for arbitrary points.
A third type of applications includes sharp (up to N ε losses) estimates
for the number of solutions of various Diophantine inequalities. This is rather
surprising given the fact that our methods do not rely on any number theory.
We believe that they provide a new angle by means of our use of induction on
scales and the topology of Rn . Indeed, the Multilinear Restriction Theorem 6.1
that we use repeatedly in the proof of our main Theorem 1.1 relies at its core
on the multilinear Kakeya phenomenon, which has some topological flavor (see
[24], [16]).
Finally, we use Theorem 2.2 to improve the range from [13], [12] in the
discrete restriction problem for lattice points on the sphere.
In forthcoming papers we will develop the decoupling theory for arbitrary
hypersurfaces with nonzero Gaussian curvature, as well as for nondegenerate
curves.
Acknowledgments. The authors are indebted to the anonymous referees
whose comments helped improve the presentation in Section 6. The second
author has benefited from helpful conversations with Nets Katz and Andreas
Seeger. The second author would like to thank his student Fangye Shi for a
careful reading of the original version of the manuscript and for pointing out
a few typos.

2. First applications
In this section we present the first round of applications of our decoupling
theory. Additional applications will appear elsewhere.
2.1. The discrete restriction phenomenon. To provide some motivation we
recall the Stein-Tomas Restriction Theorem; see [34].
Theorem 2.1. Let S be a compact C 2 hypersurface in Rn with nonzero
Gaussian curvature, and let dσ denote the natural surface measure on S. Then
for p ≥ 2(n+1) 2
n−1 and f ∈ L (S, dσ), we have

f‘
dσ . kf kL2 (S) .
Lp (Rn )
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 355

Note that this result only needs nonzero Gaussian curvature. We will use
the notation e(a) = e2πia . For fixed p ≥ 2(n+1)
n−1 , it is an easy exercise to see
that this theorem is equivalent to the statement that
Ñ é1/p
p
1
Z
n
− n−1
X
aξ e(ξ · x) . δ 2p 4 kaξ kl2 (Λ) ,
|BR | BR ξ∈Λ

for each 0 ≤ δ ≤ 1, each aξ ∈ C, each ball BR ⊂ Rn of radius R ∼ δ −1/2 and

each δ 1/2 separated set Λ ⊂ S. Thus, the Stein-Tomas Theorem measures the
average Lp oscillations of exponential sums at spatial scale equal to the inverse
of the separation of the frequencies. It will be good to keep in mind that for
each R & δ −1/2 ,
X
(5) aξ e(ξ · x) ∼ |BR |1/2 kaξ k2
ξ∈Λ L2 (BR )

as can be seen using Plancherel’s Theorem.

The discrete restriction phenomenon consists in the existence of stronger
cancellations at the larger scale R & δ −1 . We prove the following.
Theorem 2.2. Let S be a compact C 2 hypersurface in Rn with positive
definite second fundamental form. Let Λ ⊂ S be a δ 1/2 - separated set, and let
R & δ −1 . Then for each ε > 0,
Ñ é1/p
p
1
Z
n+1
− n−1 −ε
X
(6) aξ e(ξ · x) .ε δ 2p 4 kaξ k2
|BR | BR ξ∈Λ

2(n+1)
if p ≥ n−1 .

It has been observed in [11] that Theorem 1.1 for a given p implies (6) for
the same p. Here is a sketch of the argument. First, note that the statement
Ç å1/2
whenever supp(fˆ) ⊂ Nδ
X
cp
kf kp . δ kfθ k2p
θ∈Pδ

easily implies that for each g : S → C and R & δ −1 ,

ÇZ å1/p Ñ é1/2
p 2
‘ g’
X
(7) gdσ . δ cp θ dσ ;
BR Lp (wBR )
θ∈Pδ

here gθ = g1θ is the restriction of g to the δ 1/2 - cap θ on S. See Remark 5.2.
Also, throughout the paper we write
ÇZ å1/p
p
kf kLp (wB ) = |f (x)| wBR (x)dx
R
Rn
356 JEAN BOURGAIN and CIPRIAN DEMETER

for weights wBR that are Fourier supported in B(0, R1 ) and satisfy
Ç å−10n
|x − c(BR )|
(8) 1BR (x) . wBR (x) ≤ 1+ .
R
It now suffices to use g = ξ∈Λ aξ σ(U (ξ, τ ))−1 1U (ξ,τ ) in (7), where U (ξ, τ ) is
P

a τ -cap on S centered at ξ, and to let τ → 0.

Using (6) with p = 2(n+1)
n−1 and Hölder’s inequality we determine that
Ñ é1/p
p
1
Z
.ε δ −ε kaξ k2
X
(9) δ ε kaξ k2 .ε aξ e(ξ · x)
|BR | BR ξ∈Λ

for 1 ≤ p ≤ 2(n+1)
n−1 and R & δ −1 . We mention that prior to our current
work, the only known results for (6) and (9) were the ones in the range where
Theorem 1.1 was known.
2.2. Strichartz estimates for the classical and irrational tori. The discrete
restriction phenomenon has mostly been investigated in the special case when
the frequency points Λ come from a lattice. There is extra motivation in
considering this case coming from PDEs, where there is interest in establishing
Strichartz estimates for the Schrödinger equation on the torus. Prior to the
current work, the best known result for the paraboloid
P n−1 (N ) := {ξ := (ξ1 , . . . , ξn ) ∈ Zn : ξn = ξ12 +· · ·+ξn−1
2
, |ξ1 |, . . . , |ξn−1 | ≤ N }
was obtained by the first author [9], [11]. We recall this result below.
Theorem 2.3 (Discrete restriction: the lattice case (paraboloid)). Let
aξ ∈ C and ε > 0. Then
(i) if n ≥ 4, we have
n−1
− n+1 +ε
X
aξ e(ξ · x) .ε N 2 p kaξ kl2 ,
ξ∈P n−1 (N ) Lp (Tn )

2(n+2)
for p ≥ n−1 and
X
aξ e(ξ · x) .ε N ε kaξ kl2
ξ∈P n−1 (N ) Lp (Tn )
2n
for 1 ≤ p ≤ n−1 ;
(ii) if n = 2, 3, then
X
aξ e(ξ · x) .ε N ε kaξ kl2
ξ∈P n−1 (N ) Lp (Tn )

2(n+1)
for p = n−1 .
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 357

The proof of (i) combines the implementation of the Stein-Tomas argu-

ment via the circle method with the inequality (3) proved in [11]. The argument
for (ii) is much easier; it uses the fact that circles in the plane contain “few”
lattice points. It has been conjectured in [9] that (ii) should also hold for n ≥ 4.
This is easily seen to be sharp, up to the N ε term. We will argue below that
our Theorem 2.2 implies this conjecture — in fact, a more general version of it.
The analogous question for the more general irrational tori has been
recently investigated in [10], [17], [19] and [23]. More precisely, fix 12 <
θ1 , . . . , θn−1 < 2. For φ ∈ L2 (Tn−1 ), consider its Laplacian
∆φ(x1 , . . . , xn−1 )
X
= (ξ12 θ1 + · · · + ξn−1
2
θn−1 )φ̂(ξ1 , . . . , ξn−1 )e(ξ1 x1 + · · · + ξn−1 xn−1 )
(ξ1 ,...,ξn−1 )∈Zn−1
Qn−1
on the irrational torus i=1 R/(θi Z). Let also
eit∆ φ(x1 , . . . , xn−1 , t)
X
= φ̂(ξ1 , . . . , ξn−1 )e(x1 ξ1 +· · ·+xn−1 ξn−1 +t(ξ12 θ1 +· · ·+ξn−1
2
θn−1 )).
(ξ1 ,...,ξn−1 )∈Zn−1

We prove
Theorem 2.4 (Strichartz estimates for irrational tori). Let φ ∈ L2 (Tn−1 )
with supp(φ̂) ⊂ [−N, N ]n−1 . Then for each ε > 0, p ≥ 2(n+1)
n−1 and each interval
I ⊂ R with |I| & 1, we have
n−1
− n+1 +ε
(10) keit∆ φkLp (Tn−1 ×I) .ε N 2 p |I|1/p kφk2 ,
and the implicit constant does not depend on I, N and θi .
1/2
θi ξi
Proof. For −N ≤ ξ1 , . . . , ξn−1 ≤ N , define ηi = 4N and aη = φ̂(ξ). A
simple change of variables shows that
1
Z Z
it∆ p
|e φ| . n+1 |y |,...,|y |≤8N
Tn−1 ×I N 1 n−1
y ∈I n N2
p
X
× aη e(y1 η1 + · · · + yn−1 ηn−1 + yn (η12 + ··· + 2
ηn−1 )) dy1 · · · dyn ,
η1 ,...,ηn−1

where IN 2 is an interval of length ∼ N 2 |I|. By periodicity in the y1 , . . . , yn−1

variables we bound the above by
1
Z

N n+1 (N |I|)n−1 BN 2 |I|

p
X
× aη e(y1 η1 + · · · + yn−1 ηn−1 + yn (η12 + ··· + 2
ηn−1 )) dy1 · · · dyn
η1 ,...,ηn−1
358 JEAN BOURGAIN and CIPRIAN DEMETER

for some ball BN 2 |I| of radius ∼ N 2 |I|. Our result will follow once we note
that the points
(η1 , . . . , ηn−1 , η12 + · · · + ηn−1
2
)
1
are ∼ N separated on P n−1 and then apply Theorem 2.2 with R ∼ N 2 |I|. 
Remark 2.5. The diagonal form ξ12 θ1 + · · · + ξn−1
2 θ
n−1 may in fact be re-
placed with an arbitrary definite quadratic form Q(ξ1 , . . . , ξn−1 ) to incorporate
the more general case of flat tori. The case θ1 = · · · = θn−1 = 1 corresponds
to the classical (periodic) torus Tn . When combined with our Theorem 2.4,
Propositions 3.113 and 3.114 from [9] show that in fact (10) holds true with
ε = 0 in the range p > 2(n+1) n
n−1 for T . Similar partial results in the direction of
ε removal are derived for the irrational torus in [23].
2.3. The discrete restriction for lattice points on the sphere. Given inte-
gers n ≥ 3 and λ = N 2 ≥ 1 consider the discrete sphere
Fn,N 2 = {ξ = (ξ1 , . . . , ξn ) ∈ Zn : |ξ1 |2 + · · · + |ξn |2 = N 2 }.
In [8], the first author made the following conjecture about the eigenfunctions
of the Laplacian on the torus and found some partial results.
2n
Conjecture 2.6. For each n ≥ 3, aξ ∈ C, ε > 0 and each p ≥ n−2 , we
have
n−2
−n +ε
X
(11) aξ e(ξ · x) .ε N 2 p kaξ kl2 (Fn,N 2 ) .
ξ∈Fn,N 2 Lp (Tn )

We refer the reader to [13], [12] for a discussion on why the critical index
2n
n−2 for the sphere is different from the one for the paraboloid. The conjecture
2n
has been verified by the authors in [13] for p ≥ n−3 when n ≥ 4 and then
later improved in [12] to p ≥ 7 when n = 4 and p ≥ 14
44
3 when n = 5. The
methods in [8], [13] and [12] include Number Theory of various sorts, Incidence
Geometry and Fourier Analysis. Using Theorem 2.2 we can further improve
our results.
2(n−1)
Theorem 2.7. Let n ≥ 4. Inequality (11) holds for p ≥ n−3 .
Proof. Fix kaξ k2 = 1, and define
X
F (x) = aξ e(x · ξ).
ξ∈Fn,N 2

We start by recalling the following estimate (24) from [13], valid for n ≥ 4 and
n−1
α &ε N 4 +ε :
n−1 2
(12) |{|F | > α}| .ε α−2 n−3 N n−3 .
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 359

By invoking interpolation with the trivial L∞ bound, it suffices to consider the

endpoint p = pn = 2(n−1)
n−3 . Note that kF k∞ ≤ N
Cn . It follows that

Z Z
pn
|F | = n−1 +ε αpn −1 |{|F | > α}|dα
N 4 .ε α≤N Cn
2(n+1)
Z 2(n+1)
n−1
+ε)(pn − n−1 )
+ N( 4 |F | n−1 .

The result will follow by applying (12) to the first term and Theorem 2.2 with
p = 2(n+1)
n−1 to the second term. 
2.4. Additive energies and Incidence Geometry. The proof of Theorem 1.1
in the following sections will implicitly rely on the incidence theory of tubes
and cubes. This theory manifests itself in the deep multilinear Kakeya phe-
nomenon that lies behind Theorem 6.1. It thus should come as no surprise
that Theorem 1.1 has applications to Incidence Geometry.
An interesting question is whether there is a proof of Theorem 2.2 using
softer arguments — or at least if there is such an argument that recovers (6)
for R large enough, depending on Λ. When n = 3 and S = P 2 we can prove
such a result. In fact our result is surprisingly strong, in that the bound |Λ|ε
does not depend on the separation between the points in Λ.
Theorem 2.8. Let Λ ⊂ P 2 be an arbitrary collection of distinct points.
Then for R large enough, depending only on the geometry of Λ and on its
cardinality |Λ|, we have
Ñ é1/4
4
1
Z X
(13) aξ e(ξ · x) .ε |Λ|ε kaξ k2 .
|BR | BR ξ∈Λ

Due to periodicity, this recovers (ii) of Theorem 2.3 for n = 3. To see the
proof we recall some terminology and well-known results.
Given an integer k ≥ 2 and a set Λ in Rn , we introduce its k-energy

Ek (Λ) = |{(λ1 , . . . , λ2k ) ∈ Λ2k : λ1 + · · · + λk = λk+1 + · · · + λ2k }|.

Note the trivial lower bound |Ek (Λ)| ≥ |Λ|k .

We recall the point-line incidence theorem due to Szemerédi and Trotter:
Theorem 2.9 ([31]). There are O(|L| + |P| + (|L||P|)2/3 ) incidences be-
tween any collections L and P of lines and points in the plane.
Up to extra logarithmic factors, the same thing is conjectured to hold if
lines are replaced with circles. Another related conjecture is
Conjecture 2.10 (The unit distance conjecture). The number of unit
distances between N points in the plane is always .ε N 1+ε .
360 JEAN BOURGAIN and CIPRIAN DEMETER

The point-circle and the unit distance conjectures are thought to be rather
difficult, and only partial results are known.
Proof of Theorem 2.8. The following parameter encodes the “additive ge-
ometry” of Λ:
υ := min {|η1 + η2 − η3 − η4 | : ηi ∈ Λ and |η1 + η2 − η3 − η4 | =
6 0} .
|Λ|2
We show that Theorem 2.8 holds if R & υ . Fix such an R. Using restricted
type interpolation it suffices to prove
4
1
Z
e(x · η) dx .ε |Λ0 |2+ε
X
|BR | BR η∈Λ0

for each subset Λ0 ⊂ Λ. See Section 6 in [12] for details on this type of approach.
Expanding the L4 norm we need to prove
1
Z
e((η1 + η2 − η3 − η4 ) · x)dx .ε |Λ0 |2+ε .
X

ηi ∈Λ0
R3 BR

Note that if A 6= 0, then

Z R
e(At)dt ≤ A−1 .
−R

Using this we get that

1 |Λ0 |4
Z
≤ |Λ0 |2 .
X
e((η1 + η2 − η3 − η4 ) · x)dx ≤
ηi ∈Λ0
R3 BR Rυ
|η1 +η2 −η3 −η4 |6=0

Thus it suffices to prove the following estimate for the additive energy:
(14) E2 (Λ0 ) .ε |Λ0 |2+ε .
Assume
(15) η1 + η2 = η3 + η4 ,
with ηi := (αi , βi , αi2 + βi2 ). It has been observed in [9] that given A, B, C ∈ R,
the equality
η1 + η2 = (A, B, C)
implies that for i ∈ {1, 2},
Å ã Å ã
A 2 B 2 2C − A2 − B 2
(16) αi − + βi − = .
2 2 4
Thus the four points Pi = (αi , βi ) corresponding to any additive quadruple (15)
must belong to a circle. As observed in [9], this is enough to conclude (14)
in the lattice case, as circles of radius M contain .ε M ε lattice points. The
bound (14) also follows immediately if one assumes the circle-point incidence
conjecture.
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 361

However, we need a new observation. Note that if (15) holds, then in

fact both P1 , P2 and P3 , P4 are diametrically opposite on the circle (16). Thus
each additive quadruple gives rise to a distinct right angle, the one subtended
by P1 , P2 , P3 (say). Estimate (14) is then an immediate consequence of the
following application of the Szemerédi-Trotter Theorem. 
Theorem 2.11 (Pach, Sharir [26]). The number of repetitions of a given
angle among N points in the plane is O(N 2 log N ).
It has been recognized that the restriction theory for the sphere and the
paraboloid are very similar.2 Consequently, one expects not only Theorem 2.8
to be true also for S 2 , but for a very similar argument to work in that case,
too. If that is indeed the case, it does not appear to be obvious. The same
argument as above shows that an additive quadruple of points on S 2 will belong
to a circle on S 2 and, moreover, the four points will be diametrically opposite
in pairs. There will thus be at least E2 (Λ) right angles in Λ. This is, however,
of no use in this setting, as Λ lives in three dimensions. It is proved in [1] that
a set of N points in R3 has O(N 7/3 ) right angles and, moreover, this bound is
tight in general.
Another idea is to map an additive quadruple to the plane using the
stereographic projection. The resulting four points will again belong to a circle,
so the bound on the energy would follow if the circle-point incidence conjecture
is proved. Unfortunately, the stereographic projection does not preserve the
property of being diametrically opposite and thus prevents the application of
Theorem 2.11. We thus ask
Question 2.12. Is it true that E2 (Λ) .ε |Λ|2+ε for each finite Λ ⊂ S 2 ?
One can ask the same question for P n−1 and S n−1 when n ≥ 4. The right
conjecture seems to be
3n−5
(17) E2 (Λ) .ε |Λ| n−1 +ε .
Interestingly, when Λ ⊂ P 3 this follows from the aforementioned result in [1]
and, in fact, there is no Λε loss this time. However, in the same paper [1] it
is proved that this argument fails in dimensions five and higher: there is a set
with N points in R4 that determines & N 3 right angles. We point out that
Theorem 2.2 implies (17) for subsets of P n−1 and S n−1 when n ≥ 3, in the
1
case when the points Λ are ∼ |Λ|− n−1 separated.
It is also natural to investigate the two dimensional phenomenon for S =
S and S = P 1 .
1

2
A notable difference is the lattice case of the discrete restriction, but that has to do with
a rather specialized scenario.
362 JEAN BOURGAIN and CIPRIAN DEMETER

Question 2.13. Is it true that for each Λ ⊂ S,

(18) E3 (Λ) .ε |Λ|3+ε ?
Surprisingly, this question seems to be harder than its three dimensional
analogue from Theorem 2.8. Note that the case when the points are |Λ|−C
separated follows from Theorem 2.2. We are not aware of an alternative (softer)
argument.
A positive answer to Question 2.13 would have surprising applications to
Number Theory. In particular, it would answer the following question posed
in [5].
Question 2.14. Let N be a positive integer. Does (18) hold when Λ are
the lattice points on the circle N 1/2 S 1 ?
Note that Theorem 2.2 is too weak to answer this question. Indeed, rescal-
ing by N 1/2 , the lattice points in N 1/2 S 1 become N −1/2 -separated points on
O(1)
S 1 . However, it is known that there are O(N log log N ) lattice points on the circle
N 1/2 S 1 .
The analysis in [5] establishes some partial results as well as some intrigu-
ing connections to the theory of elliptic curves; see, for example, [5, Th. 8]. An
easier question with similar flavor is answered in the next subsection.
The best that can be said regarding Question 2.13 with topological based
methods seems to be the following.
Proposition 2.15. Let S be either P 1 or S 1 . For each Λ ⊂ S,
7
E3 (Λ) .ε |Λ| 2 +ε .
Proof. This was observed by Bombieri and the first author [5] when S = S 1 .
The proofs for P 1 and S 1 are very similar; we briefly sketch the details for
S = P 1 . Let N be the cardinality of Λ. It goes back to [9] that if
(19) (x1 , x21 ) + (x2 , x22 ) + (x3 , x23 ) = (n, j),
√
then the point (3(x1 +x2 ), 3(x1 −x2 )) belongs to the circle centered at (2n, 0)
and of radius squared equal to 6j − 2n2 . Note that there are N 2 such points
with (xi , x2i ) ∈ Λ; call this set of points T√. Assume we have Mn such circles
containing roughly 2n points (3(x1 + x2 ), 3(x1 − x2 )) ∈ T in such a way that
(19) is satisfied for some x3 ∈ Λ. Then clearly
X
E3 (S) . Mn 22n .
2n ≤N

It is easy to see that

(20) Mn 2n . N 3 ,
as each point in T can belong to at most N circles.
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 363

The nontrivial estimate is

(21) Mn 23n . N 4 ,
which is an immediate consequence of the Szemerédi-Trotter Theorem for
curves satisfying the following two fundamental axioms: two curves intersect in
O(1) points, and there are O(1) curves passing through any two given points.
The number of incidences between such curves and points is the same as in the
case of lines and points; see, for example, [33, Th. 8.10]. Note that since our
circles have centers on the x axis, any two points in T sitting in the upper (or
lower) half plane determine a unique circle. Combining the two inequalities we
get for each n
7
Mn 22n . N 2 . 
In the case when Λ ⊂ S 1 , the same argument leads to incidences between
unit circles and points. The outcome is the same, since for any two points there
are at most two unit circles passing through them. An interesting observation
is the fact that Question 2.13 has a positive answer if the Unit Distance Con-
jecture is assumed. Indeed, the argument above presents us with a collection
T of N 2 points and a collection of . N 3 unit circles. For 2n . N , let Mn
be the number of such circles with ∼ 2n points. There will be at least Mn 2n
unit distances among the N 2 points and the Mn centers. The Unit Distances
Conjecture forces Mn 2n .ε (Mn + N 2 )1+ε . Since Mn . N 3 , it immediately
follows that Mn 22n .ε N 3+ε , which gives the desired bound on the energy.
It seems likely that in order to achieve the conjectured bound on E3 (Λ),
the structure of T must be exploited, paving the way to algebraic methods.
One possibility is to make use of the fact that T has sumset structure. Another
interesting angle for the parabola is√the following. Recall that whenever (19)
holds, the three points (3(xi + xj ), 3(xi − xj )), (i, j) ∈ {(1, 2), (2, 3), (3, 1)},
belong to the circle centered at (2n, 0) and of radius squared equal to 6j − 2n2 .
One can easily check that in fact they form an equilateral triangle! This poten-
tially opens up the new toolbox of symmetries since, for example, the rotation
by π/3 about the center of any such circle C will preserve C ∩ T .
2.5. Additive energies of annular sets. We start by mentioning a more
general version of Theorem 2.2.
Theorem 2.16. Let S be a C 2 compact hypersurface in Rn with positive
definite second fundamental form. For each θ ∈ Pδ , let Λθ be a collection of
points in θ, and let Λ = ∪θ Λθ . Then for each R- ball BR with R & δ −1 , we
have
Ñ é1/2
2
.ε δ −ε
X X X
aξ e(x · ξ) 2(n+1)
aξ e(x · ξ) 2(n+1)
.
ξ∈Λ L n−1 (B R) θ ξ∈Λθ L n−1 (wBR )
364 JEAN BOURGAIN and CIPRIAN DEMETER

To see why this holds, note first that the case R ∼ δ −1 follows by applying
(the localized version of) Theorem 1.1 to functions whose Fourier transforms
approximate weighted sums of Dirac deltas supported on Λ. The case R & δ −1
then follows using Minkowski’s inequality.
For R > 1, define
AR = {ξ ∈ R2 : R ≤ |ξ| ≤ R + R−1/3 }
and A0R = AR ∩Z2 . We prove the following inequality related to Question 2.14.

Theorem 2.17.
(22) E3 (A0R ) .ε |A0R |3+ε .
Note that this is essentially sharp. The old Van der Corput estimate
|N (R) − πR2 | = o(R2/3 )
for the error term in the Gauss circle problem shows that |A0R | = 2πR2/3 +
o(R2/3 ). It thus suffices to show
X 1
e(ξ · x) .ε R 3 +ε .
ξ∈A0R L6 (T2 )

Subdivide AR into sectors Aα of size ∼ R1/3 × R−1/3 so that each of them fits
inside a rectangle Rα of area < 12 . Applying Theorem 2.16 after rescaling by
R and using periodicity, we get
Ñ é1/2
2
X X X
(23) e(ξ · x) .ε Rε e(ξ · x) ,
ξ∈A0R L6 (T2 ) α ξ∈A0α L6 (T2 )

with A0α = Aα ∩ Z2 .
An elementary observation, which goes back (at least) to Jarnı́k’s work
[25], is the fact that the area determined by a nondegenerate triangle with
vertices in Z2 is half an integer. It follows that the points in each A0α lie on a
line Lα . In fact they must be equidistant, with consecutive points at distant
d, for some d ≥ 1. Now for 1 ≤ 2s . R1/3 , define
Ls = {α : 2s ≤ |A0α | < 2s+1 }.
Let also Ls,m be those α ∈ Ls for which 2m ≤ d < 2m+1 . Note that if α ∈ Ls,m ,
then Lα makes an angle ∼ 2−m 2−s R−1/3 with the long axis of Rα . Thus the
directions of the lines Lα will be distinct for each collection of α ∈ Ls,m whose
corresponding arcs on S 1 are C2−m 2−s R−1/3 -separated. Obviously there are
u, v ∈ A0α such that |u − v| ∼ 2m . Since there are O(22m ) lattice points
with length ∼ 2m , it follows that there can be at most O(22m ) elements α in
Ls,m that are C2−m 2−s R−1/3 -separated. Thus |Ls,m | . 22m R1/3 2−m 2−s . As
2m+s . R1/3 , we conclude that |Ls | . R2/3 2−2s .
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 365

Note that by using Hölder’s inequality with L2 − L∞ endpoints, we have

≤ |A0α |5/6 .
X
e(ξ · x)
ξ∈A0α L6 (T2 )

Using this, the bound on |Ls | and (23) finishes the proof of (22).
2.6. Counting solutions of Diophantine inequalities. In this section we
show how to use the Decoupling Theorem to recover and generalize results
from the literature as well as to prove some new type of results. We do not
aim at providing a systematic study of these problems but rather to explain
the way our methods become useful in this context.
To motivate our first application we consider the system of equations for
k ≥ 2,

nk + nk + nk = nk + nk + nk ,
1 2 3 4 5 6
n 1 + n 2 + n 3 = n 4 + n 5 + n 6 ,

with 1 ≤ ni ≤ N . It is easy to see that there are 6N 3 trivial solutions. The

question here is to determine the correct asymptotic for the number Uk (N ) of
nontrivial solutions. This is in part motivated by connections to the Waring
problem; see [4]. The case k = 3, known as the Segre cubic, has been intensely
studied. Vaughan and Wooley have proved in [35] that U3 (N ) ∼ N 2 (log N )5 ;
see also [15] for a more precise result. For k ≥ 4, Greaves [22] (see also [29])
17
has proved that Uk (N ) = O(N 6 +ε ). All these results follow through the use
of rather delicate Number Theory.
While our methods in this paper cannot produce such fine estimates, they
successfully address the perturbed case. The following result is perhaps a
surprising consequence of Theorem 2.2.
Theorem 2.18. For fixed k ≥ 2 and C, the system

|nk + nk + nk − nk − nk − nk | ≤ CN k−2 ,
1 2 3 4 5 6
n 1 + n 2 + n 3 = n 4 + n 5 + n 6

has O(N 3+ε ) solutions with ni ∼ N .

Proof. Apply Theorem 2.2 to the curve

{(ξ, ξ k ) : |ξ| ∼ 1},

the points
®Ç Å ãk å ´
n n
Λ= , : n∼N
N N
366 JEAN BOURGAIN and CIPRIAN DEMETER

and δ = N −2 . We get that

Ç Å ãk å 6
1 n n
Z Z X
e x +y dxdy .ε N 3+ε .
N4 |x|≤N 2 |y|≤N 2 n∼N
N N
Upon rescaling and using periodicity we get
Z Z 6
X
k−3 k
N e(xn + yn ) dxdy
|x|≤N |y|≤N 2−k n∼N
(24) Z Z 6
X
= N k−2 e(xn + ynk ) dxdy .ε N 3+ε .
|x|≤1 |y|≤N 2−k n∼N
Let now φ : R → [0, ∞) be a Schwartz function with positive Fourier transform
satisfying φ(ξ)
b & 1 for |ξ| ≤ 1. Define φN (y) = φ(N k−2 y). A standard
argument allows us to replace the cutoff |y| ≤ N 2−k with φN (y) in (24). It
suffices then to note that
Z Z 6
X
N k−2 e(xn + ynk ) φN (y)dxdy
|x|≤1 R n∼N
X Ä Ä ää
= φb N 2−k nk1 + nk2 + nk3 − nk4 − nk5 − nk6 . 
ni ∼N
n1 +n2 +n3 =n4 +n5 +n6

Note also that our method proves that

Z Z 6
X
N k−2 e(xn + ynk ) dxdy .ε N 3+ε
|x|≤1 |y−c|≤N 2−k n∼N
for each c ∈ R. The difficulty in proving this for k ≥ 3 using purely num-
ber theoretic methods comes from estimating the contribution of the minor
arcs. When k = 2, the left-hand side is at least cN 3 log N , which shows that
one cannot dispense with the N ε term. This can be seen by evaluating the
contribution from the major arcs; see, for example, [9, p. 118].
Our second application generalizes the result from [28] (k = 4) to k ≥ 4.
Its original motivation lies in the study of the Riemann zeta function on the
critical line (cf. [6], [7]) and also in getting refinements of Heath-Brown’s variant
of Weyl’s inequality; see [28].
Theorem 2.19. For k ≥ 4 and 0 ≤ λ ≤ 1, we have
Z Z λ X 6
e(xn + yn ) dxdy .ε λN 3+ε + N 4−k+ε .
2 k
|x|≤1 0 n∼N
In particular, the system

|nk + nk + nk − nk − nk − nk | ≤ CN k−1 ,
1 2 3 4 5 6
n2 + n2 + n2 = n2 + n2 + n2
1 2 3 4 5 6

has O(N 3+ε ) solutions with ni ∼ N .

 THE PROOF OF THE l2 DECOUPLING CONJECTURE 367

Proof. The estimate on the number of solutions follows by using λ =

N 1−k . Note that it suffices to prove that
Z Z 6
X
e(xn2 + ynk ) dxdy .ε N 4−k+ε
|x|≤1 J n∼N

for each interval J with length N 1−k .

We apply Theorem 2.16 to the curve
{(ξ 2 , ξ k ) : |ξ| ∼ 1},
the points ®ÇÅ ã2 Å ãk å ´
n n
Λ= , : n∼N ,
N N
R−1 = δ = N −1 and BN = [M N, (M + 1)N ] × N k J with
M ∈ {−N, . . . , 0, . . . , N − 1}.
Summing over M , due to periodicity we get
Ç å
n2 nk
e x 2 + y0 k
0
X

n∼N
N N L6 (|x0 |≤N 2 , y 0 ∈N k J)
Ñ Ç å 2 é1/2
n2 nk
e x0 2 + y 0 k
X X
.ε N ε .
α n∈Iα
N N L6 (|x0 |≤N 2 , y 0 ∈N k J)

Here Iα = [nα , nα +N 1/2 ] are intervals of length N 1/2 that partition the integers
n ∼ N . It follows after a change of variables that
X
e(xn2 + ynk )
n∼N L6 (|x|≤1, y∈J)
(25) Ñ é1/2
2
X X
.ε N ε e(xn2 + ynk ) .
α n∈Iα L6 (|x|≤1, y∈J)

Next note that for y ∈ J,

X
e(xn2 + ynk )
n∈Iα
N 1/2 Ç Ç å å
X
2 k(k − 1) k−2
= cm,J,nα e m x+ nα y + m(2xnα + knk−1
α y) + O(1),
m=1
2
with |cm,J,nα | = 1. To estimate the first term we change variables to

x0 = x + k(k−1) nk−2 y,
2 α
y 0 = (2k − k 2 )nk−1 y.
α
368 JEAN BOURGAIN and CIPRIAN DEMETER

We get
X
e(xn2 + ynk )
n∈Iα L6 (|x|≤1, y∈J)
N 1/2
− k−1 1−k
 
cm,J,nα e(x0 m2 + 2x0 nα m + my 0 )
X
. nα 6 +O N 6

m=1 L6 (BC )
N 1/2 +nα
− k−1 1−k
 
cm,J,nα e(x0 m2 + my 0 )
X
= nα 6 +O N 6

m=1+nα L6 (BC )

for some ball BC of radius C = O(1). By the result in Theorem 2.3, this can
1 1−k 1 1−k
further be seen to be O(N 4 + 6 +ε ). We conclude that (25) is O(N 2 + 6 +ε ),
as desired. 
There are further number theoretical consequences of the decoupling the-
ory that will be investigated elsewhere.

3. Norms and wave packet decompositions

We will use C to denote various constants that are allowed to depend on
the fixed parameters n, p, but never on the scale δ. We will denote by | · | both
the Lebesgue measure on Rn and the cardinality of finite sets.
This section and the next one is concerned with introducing some of the
tools that will be used in the proof of Theorem 1.1 from Section 6. For 2 ≤
p ≤ ∞, we define the norm
Ñ é1/2
X
kf kp,δ = kfθ k2p ,
θ∈Pδ

where fθ is the Fourier restriction of f to θ. We note the following immediate

consequence of Hölder’s inequality,
2
p
1− 2
(26) kf kp,δ ≤ kf k2,δ kf k∞,δp ,

and the fact that if supp(fˆ) ⊂ Nδ , then

kf k2,δ ∼ kf k2 .
Definition 3.1. Let N be a real number greater than 1. An N -tube T
is an N 1/2 × · · · × N 1/2 × N rectangular parallelepiped in Rn that has dual
orientation to some θ = θ(T ) ∈ Pδ . We call a collection of N -tubes separated
if no more than C tubes with a given orientation overlap.
Let φ : Rn → R be given by
φ(x) = (1 + |x|2 )−M
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 369

for some M large enough compared to n, whose value will become clear from
the argument. Define φT = φ ◦ aT , where aT is the affine function mapping T
to the unit cube in Rn centered at the origin.
Definition 3.2. An N -function is a function f : Rn → C such that
X
f= fT ,
T ∈T (f )

where T (f ) consists of finitely many separated N -tubes T and, moreover,

|fT | ≤ φT ,

kfT kp ∼ |T |1/p , 1 ≤ p ≤ ∞
and
T ) ⊂ θ(T ).
supp(fc
For θ ∈ P1/N , let T (f, θ) denote the N -tubes in T (f ) dual to θ . An N -function
is called balanced if |T (f, θ)| ≤ 2|T (f, θ0 )| whenever T (f, θ), T (f, θ0 ) 6= ∅.
The k · kp,δ norms of N -functions are asymptotically determined by their
plate distribution over the sectors θ.
Lemma 3.3. For each N -function f and for 2 ≤ p ≤ ∞,
!1/2
n+1 X 2
(27) kf kp,1/N ∼ N 2p |T (f, θ)| p .
θ

If the N -function is balanced, then

n+1 1
− p1
(28) kf kp,1/N ∼ N 2p M (f ) 2 |T (f )|1/p ,

where M (f ) is the number of sectors θ for which T (f, θ) 6= ∅.

Proof. It suffices to prove (27) when T (f ) = T (f, θ) for some θ. We
first observe the trivial estimates kf k1 . |T ||T (f )|, kf k∞ . 1 and kf k2 ∼
|T |1/2 |T (f )|1/2 . Applying Hölder’s inequality twice we get
2(p−1) 2−p
1/p 0
kf k2 p
kf k1 p ≤ kf kp ≤ kf k1 kf k1/p
∞ ,

which is exactly what we want. 

The crucial role played by balanced N -functions is encoded by

Lemma 3.4.
(i) Each N -function f can be written as the sum of O(log |T (f )|) balanced
N -functions.
370 JEAN BOURGAIN and CIPRIAN DEMETER

(ii) For each balanced N -function f and 2 ≤ p ≤ ∞, we have the converse

of (26), namely,
2
p p
1− 2
(29) kf kp,1/N ∼ kf k2,1/N kf k∞,1/N .

Proof. Note that (i) is immediate by using dyadic ranges. Also, (ii) will
follow from (28). 

In the remaining sections we will use the fact that the contribution of f to
various inequalities comes from logarithmically many N -functions. The basic
mechanism is the following.
Lemma 3.5 (Wave packet decomposition). Assume f is Fourier supported
in Nδ . Then for each dyadic 0 < λ . kf k∞,δ , there is an N = δ −1 -function
fλ such that
X
f= λfλ ,
λ.kf k∞,δ

and for each 2 ≤ p < ∞, we have

n+1
(30) λp N 2 |T (fλ )| ≤ kλfλ kpp,δ . kf kpp,δ .
Proof. Using a partition of unity, write
f˜θ ,
X
f=
θ∈Pδ

with f˜θ = fθ ∗ Kθ Fourier supported in 109

θ with kKθ k1 . 1. Consider a
˜
windowed Fourier series expansion for each fθ :
f˜θ = hf˜θ , ϕT iϕT ,
X

T ∈Tθ

where ϕT are L2 normalized Schwartz functions Fourier localized in θ such

that
|T |1/2 |ϕT | . φT .
The tubes in Tθ are separated. Note that by Hölder’s inequality,
1
aT := hf˜θ , ϕT i . kf˜θ k∞ . kfθ k∞ ≤ kf k∞,δ .
|T |1/2
It is now clear that we should take
aT λ−1 |T |1/2 ϕT .
X X
fλ =
θ T ∈Tθ : |aT |∼λ

To see (30) note that the first inequality follows from (27) and the fact
that k · klp/2 ≤ k · kl1 . To derive the second inequality, it suffices to prove that
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 371

for each θ,

hf˜θ , ϕT iϕT
X
. kfθ kp .
T ∈Tθ : |aT |∼λ p

Using (27) and the immediate consequence of Hölder’s inequality |aT |p .

|f˜θ |p |T |−1/2 ϕT , we get
R

p
hf˜θ , ϕT iϕT
X X
. λp |T ||{T ∈ Tθ : |aT | ∼ λ}| . |T | |aT |p
T ∈Tθ : |aT |∼λ p T ∈Tθ
Z Z Z
|f˜θ |p |f˜θ |p .
X
. φT . |fθ |p . 
T ∈Tθ

4. Parabolic rescaling
1
Proposition 4.1. Let δ ≤ σ < 2 and Kp ( σδ ) be such that
Ö è1/2
Å ã
δ X
kf kp ≤ Kp kfθ k2p
σ θ∈P δ
σ

for each f with Fourier support in N δ . Then for each f with Fourier support
σ
in Nδ and for each τ ∈ Pσ , we have
Ñ é1/2
Å ã
δ X
kfτ kp . Kp kfθ k2p .
σ θ∈Pδ : θ∩τ 6=∅

Proof. Let a = (a1 , . . . , an−1 ) be the center of the σ 1/2 -cube Cτ ; see (1).
We will perform the parabolic rescaling via the affine transformation

Lτ (ξ1 , . . . , ξn ) = (ξ10 , . . . , ξn0 )

Ç Pn−1 Pn−1 2 å
ξ1 − a1 ξn−1 − an−1 ξn − 2 i=1 ai ξi + a i=1 i
= 1/2
,..., , .
σ σ 1/2 σ
Note that
n−1 n−1
!
2
ξn0 ξi0 −1
X X
− =σ ξn − ξi2 .
i=1 i=1

It follows that Lτ maps the Fourier support Nδ ∩τ of fτ to N δ ∩([− 21 , 12 ]n−1 ×R).

σ
Also, for each τ 0 ∈ P δ , we have that Lτ (θ) = τ 0 for some θ ∈ Pδ with θ ∩ τ 6= ∅.
σ
Thus
kfτ kpp = kgkpp (det(Lτ ))1−p ,
372 JEAN BOURGAIN and CIPRIAN DEMETER

where g is the Lτ dilation of fτ Fourier supported in N δ ∩ ([− 21 , 12 ]n−1 × R).

σ
By invoking the hypothesis we get that
Ö è1/2
Å ã
δ X
kgkp . Kp kgτ 0 k2p .
σ τ 0 ∈P δ
σ

We are done if we use the fact that

kfθ kpp = kgτ 0 kpp (det(Lτ ))1−p
whenever Lτ (θ) = τ 0 . 

5. Linear versus multilinear decoupling

The material in this section is an application of the Bourgain-Guth induc-
tion on scales [14], and it is most closely connected to the argument in [11].
Let g : P n−1 → C. For a cap τ on P n−1 , we let gτ = g1τ be the (spatial)
restriction of g to τ . We denote by π : P n−1 → [−1/2, 1/2]n−1 the projection
map.
Definition 5.1. We say that the caps τ1 , . . . , τn on P n−1 are ν-transverse
if the volume of the parallelepiped spanned by any unit normals vi at τi is
greater than ν.
We denote by Cp,n (δ, ν) the smallest constant such that
n
!1/n
’
Y
|gτ dσ|
i
i=1 Lp (Bδ−1 )
 Ü ê1/2 1/n
n
2
g’
Y X 
≤ Cp,n (δ, ν)  θ dσ p
 

 L (wB −1 ) 
i=1 θ: δ 1/2 -cap δ
θ⊂τi

for each ν-transverse caps τi ⊂ P n−1 , each δ −1 ball Bδ−1 and each g : P n−1 → C.
Let also Kp,n (δ) be the smallest constant such that
Ñ é1/2
2
‘ g’
X
gdσ ≤ Kp,n (δ) θ dσ
Lp (Bδ−1 ) Lp (wB )
θ:δ 1/2 -cap δ −1

for each g : P n−1 → C and each δ −1 ball Bδ−1 .

Remark 5.2. As before, the norm kf kLp (wB ) refers to the weighted Lp
R
integral
ÅZ ã1/p
|f (x)|p wBR (x)dx
Rn
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 373

for some weight satisfying (8). It is important to realize that there are such
weights which, in addition, are Fourier supported in B(0, R−1 ). Note also that
‘
if g is supported on P n−1 and if w −1 ‘
BR is supported in B(0, R ), then (gdσ)wBR
has Fourier support inside NR−1 . This simple observation justifies the various
(entirely routine) localization arguments that follow, as well as the interplay be-
tween Fourier transforms of functions and Fourier transforms of measures sup-
(1)
ported on P n−1 . In particular, let Kp,n (δ) be the smallest constant such that
Ñ é1/2
2
‘ g’
X
(1)
gdσ ≤ Kp,n (δ) θ dσ
Lp (wB ) Lp (wB )
δ −1 θ:δ 1/2 -cap δ −1

(1)
for each g : P n−1 → C and each δ −1 ball Bδ−1 . Then Kp,n (δ) ∼n,p Kp,n (δ).
(2) (3) (4)
Also, if Kp,n (δ), Kp,n (δ), Kp,n (δ) are the smallest constants such that
Ñ é1/2
X
(2)
kf kLp (Rn ) ≤ Kp,n (δ) kfθ k2Lp (Rn ) ,
θ∈Pδ
Ñ é1/2
X
(3)
kf kLp (Bδ−1 ) ≤ Kp,n (δ) kfθ k2Lp (Rn ) ,
θ∈Pδ
Ñ é1/2
X
(4)
kf kLp (wB ) ≤ Kp,n (δ) kfθ k2Lp (wB )
δ −1 δ −1
θ∈Pδ

for each f Fourier supported in Nδ and each δ −1 ball Bδ−1 , then

(2) (3) (4)
Kp,n (δ), Kp,n (δ), Kp,n (δ) ∼n,p Kp,n (δ).
The same observation applies to the family of constants related to Cp,n (δ, ν)
from the multilinear inequality.
Note that due to Hölder’s inequality,
Cp,n (δ, ν) ≤ Kp,n (δ).
We will show that the reverse inequality essentially holds true.
Theorem 5.3. Assume one of the following holds:
(i) n = 2;
0
(ii) n ≥ 3 and Kp,d (δ 0 ) .ε δ −ε for each δ 0 , ε > 0 and each 2 ≤ d ≤ n − 1.
Then for each 0 < ν ≤ 1, there is ε(ν) with limν→0 ε(ν) = 0 and Cν such that
Kp,n (δ) ≤ Cν δ −ε(ν) Cp,n (δ, ν)
for each δ.
We prove the case n = 3 and will indicate the modifications needed for
n ≥ 4. The argument will also show how to deal with the case n = 2.
374 JEAN BOURGAIN and CIPRIAN DEMETER

Remark 5.4. If Q1 , Q2 , Q3 ∈ [−1/2, 1/2]2 , the volume of parallelepiped

spanned by the unit normals to P 2 at π −1 (Qi ) is comparable to the area of
the triangle ∆Q1 Q2 Q3 .
The key step in the proof of Theorem 5.3 for n = 3 is the following.
Proposition 5.5. Assume Kp,2 (δ) .ε δ −ε for each ε > 0. Then for each ε,
there is Cε such that for each R > 1 and K ≥ 1,
ñÇ å1/2
‘ kg’
X
ε 2
gdσ ≤ Cε K α dσk p
L (wBR )
Lp (wBR )
α⊂P 2
α: 1 -cap
K
Ç å1/2 ô
kg’
X
+ dσk2 p
β L (wBR )
β⊂P 2
β: 1 -cap
K 1/2
Ç å1/2
+ K 10 Cp,3 (R−1 , K −2 ) kg’
X
2
∆ dσkLp (wB ) .
R
∆⊂P 2
∆: 1 -cap
R1/2

Proof. Following the standard formalism (see, for example, [14, §§2–5])
we will regard |g’α dσ| as being essentially constant on each ball BK . Denote
by cα (BK ) this value, and let α∗ be the cap that maximizes it.
The starting point in the argument is the observation in [14] that for each
BK , there exists a line L = L(BK ) in the (ξ1 , ξ2 ) plane such that if
ß ™
C
SL = (ξ1 , ξ2 ) : dist((ξ1 , ξ2 ), L) ≤ ,
K
then for x ∈ BK ,
‘
|gdσ(x)|
(31) ≤ C max |g’
α dσ(x)|
α
3
!1/3
|g÷
Y
(32) + K4 max
α1 ,α2 ,α3 αi dσ(x)|
K −2 −transverse i=1

g’
X
(33) + α dσ(x) .
α⊂π −1 (SL )∩P 2

To see this, we distinguish three scenarios. First, if cα (BK ) ≤ K −2 cα∗ (BK )

for each α with dist(π(α), π(α∗ )) ≥ 10
K , then (31) suffices, as
‘
X
|gdσ(x)| ≤ cα (BK ).
α
If not, there is α∗∗ with dist(π(α∗∗ ), π(α∗ )) ≥ 10
K and cα∗∗ (BK ) ≥ K −2 cα∗ (BK ).
The line L is determined by α∗ , α∗∗ .
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 375

Second, if there is α∗∗∗ such that π(α∗∗∗ ) intersects the complement of SL

and cα∗∗∗ (BK ) ≥ K −2 cα∗ (BK ), then (32) suffices. Indeed, note that α∗ , α∗∗ ,
α∗∗∗ are K −2 transverse by Remark 5.4.
Otherwise the sum of (31) and (33) will suffice.
The only nontrivial case to address is the one corresponding to this latter
scenario. Cover π −1 (SL ) ∩ P 2 by pairwise disjoint strips U of length ∼ K 11/2 .
An application of Minkowski’s inequality shows that
!1/2
2
g’ −1
g’
X X
α dσ . Kp,2 (K ) U dσ .
Lp (wBK )
α:π(α)⊂SL Lp (BK ) U

We are of course relying on the fact that π −1 (L) is a parabola with principal
curvature equal to 1. Note however that since we are dealing with the third
scenario
!1/2 !1/2
2
g’ kg’ + g÷
X X
U dσ . βdσk2 p L (wBK ) α∗ dσ .
Lp (BK ) Lp (wBK )
U β: 1 -cap:
K 1/2
π(β)⊂SL

We conclude that in either case,

ñÇ å1/2
2
‘ g’
X
ε
gdσ ≤ Cε K α dσ
Lp (BK ) Lp (wBK )
α⊂P 2
α: 1 -cap
K
Ç ô
2
g’
X
1/2
+ dσ
β )
Lp (wBK )
β⊂P 2
β: 1 -cap
K 1/2
Ç 3 å1/3
|g÷
Y
4
+K max
α1 ,α2 ,α3 αi dσ(x)| .
K −2 -transverse i=1 Lp (wBK )

It suffices now to raise to the pth power and sum over BK ⊂ BR using
‘
Minkowski’s inequality. Also, the norm gdσ can be replaced by the
p L (BR )
weighted norm ‘
gdσ via the localization argument described in Re-
Lp (wBR )
mark 5.2. 

Rescaling gives the following.

0
Proposition 5.6. Let τ be a δ cap. Assume Kp,2 (δ 0 ) .ε δ −ε for each
ε > 0 and δ 0 . Then for each ε, there is Cε such that for each R > δ −2 and
K ≥ 1,
376 JEAN BOURGAIN and CIPRIAN DEMETER

ñÇ å1/2
2
g’ g’
X
ε
τ dσ ≤ Cε K α dσ
Lp (wBR ) α⊂τ
Lp (wBR )
α: δ -cap
K
Ç å1/2 ô
2
g’
X
+ βdσ
Lp (wBR )
β⊂τ
β: δ -cap
K 1/2
Ç å1/2
2
2 −1 −2
g’
X
10
+ K Cp,3 ((Rδ ) ,K ) ∆ dσ .
Lp (wBR )
∆⊂τ
∆: 1 -cap
R1/2

Proof. Note that if γ ⊂ [−1/2, 1/2]2 , then

Z
gÿ
π −1 γ dσ(x1 , x2 , x3 ) = πg(ξ1 , ξ2 )e(ξ1 x1 + ξ2 x2 + (ξ12 + ξ22 )x3 )dξ1 dξ2 .
γ

Let a = (a1 , a2 ). Changing variable to ξi = ai + δξi0 and letting

πg a,δ (ξ 0 ) = πg(a + δξ 0 ),
γ 0 = δ −1 (γ − a),

we get

Ÿ
gÿ 2 a,δ 2
π −1 γ dσ(x1 , x2 , x3 ) = δ gπ −1 γ 0 dσ(δ(x1 + 2a1 x3 ), δ(x2 + 2a2 x3 ), δ x3 ) .

In particular,

2− p4 Ÿ
gÿ
π −1 γ dσ =δ gπa,δ
−1 γ 0 dσ ,
Lp (wBR ) Lp (wCR )

where CR is a ∼ δR ×δR ×δ 2 R cylinder. Cover CR with balls Bδ2 R . The result

now follows by applying Proposition 5.5 to g a,δ (with a the center of π(τ )) on
each Bδ2 R and then summing using Minkowski’s inequality. 

We are now ready to prove Theorem 5.3 for n = 3. Let K = ν −1/2 . Iterate
Proposition 5.6 staring with scale δ = 1 until we reach scale δ = R−1/2 . Each
iteration lowers the scale of the caps from δ to at least K δ1/2 . Thus we have to
iterate ∼ logK R times. Since

Cp,3 ((δ 2 R)−1 , K −2 ) . Cp,3 (R−1 , ν),

 THE PROOF OF THE l2 DECOUPLING CONJECTURE 377

for each ε > 0, we get

‘
gdσ
Lp (wBR )
Ö è1/2
2
≤ (CCε K ε )logK R K 10 Cp,3 (R−1 , ν) g’
X
∆ dσ
Lp (wBR )
1
∆: -cap
R1/2
Ö è1/2
2
= R−2 logν (CCε )+ε ν −5 Cp,3 (R−1 , ν) g’
X
∆ dσ .
Lp (wBR )
1
∆: -cap
R1/2

The result follows since C, Cε doe not depend on ν.

To summarize, the proof of Theorem 5.3 for n = 3 relied on the hypoth-
esis that the contribution coming from caps living near the intersection of P 2
with a plane is controlled by Kp,2 (δ) = O(δ −ε ). When n ≥ 4, the hypothesis
Kp,d (δ) = O(δ −ε ) for 2 ≤ d ≤ n − 1 plays the same role; it controls the con-
tribution coming from caps living near lower dimensional elliptic paraboloids
with principal curvatures equal to 1. And, of course, no such hypothesis is
needed when n = 2. The statement and the proof of Proposition 5.6 for these
values of n will hold without further modifications.

6. Proof of Theorem 1.1 for the paraboloid

In this section we prove Theorem 1.1 for P n−1 . We first consider the
open range p > 2(n+1)
n−1 , and in the end of the section we prove the result for the
endpoint. We use notation from the previous section such as Kp,n (δ), Cp,n (δ, ν)
and δ = N −1 .
Proposition 4.1 shows that Kp,n (δ) . Kp,n (δ 1/2 )2 . Let
log Kp,n (δ)
γ = lim inf .
δ→0 log(δ −1 )
It follows that for each ε,
δ −γ . Kp,n (δ) .ε δ −γ−ε .
Write γ = n−1 n+1
4 − 2p + α. We have to show that α = 0.
The following multilinear restriction estimate from [3] will play a key role
in our proof.
Theorem 6.1. Let τ1 , . . . , τn be ν-transverse caps on P n−1 , and assume
f“i is supported on the δ-neighborhood of τi . Then we have
Ç n å1/n Ç n å1/n
− 12 +ε
Y Y
|fi | 2n
.ε,ν N kfi kL2 .
i=1 L n−1 (BN ) i=1
378 JEAN BOURGAIN and CIPRIAN DEMETER

Using Plancherel’s identity this easily implies that

Cp,n (δ, ν) .ε,ν δ −ε

2n
for p = n−1 . Combined with the Bourgain-Guth induction on scales this
further leads to
Kp,n (δ) .ε δ −ε
2n
for 2 ≤ p ≤ n−1 . These inequalities were proved in [11]. We will not rely on
them in our argument below.
We now present the first step of our proof, which shows how to interpolate
the k · kp,δ norms.
Proposition 6.2. Let τ1 , . . . , τn be ν-transverse caps on P n−1 , and as-
sume f“i is supported on the δ-neighborhood of τi . Then for each n−1
2n
≤ p ≤ ∞,
we have
Ç n å1/n Ç n å1
n−1 n2 +n n
Y − 2p(n−1) +ε Y
(34) |fi | .ε,ν N 4 kfi k p(n−1) ,δ
n
i=1 Lp (BN ) i=1

and also
(35)
ñ n Ç å1/2 ô1/n Ç n å1/n
n
Y X − (n−1)p +ε Y
|fi,θ |2 .ε,ν N kfi k p(n−1) ,δ .
n
i=1 θ∈Pδ Lp (B N) i=1
n−1
Proof. Let us start with the proof of (34). Let λn = N 4 and F =
Qn
( i=1 |fi |)1/n . Note that by Cauchy-Schwartz, we have
n
!1
Y n
kF k∞ ≤ λn kfi k∞,δ .
i=1

By combining this with Theorem 6.1 and Hölder’s inequality, we find that
n Ç å! n1
1− 2n ε− n Y 2n
1− 2n
(36) kF kLp (BN ) .ε,ν λn (n−1)p N (n−1)p kfi k2 (n−1)p
kfi k∞,δ(n−1)p .
i=1

Finally, to get (34), we use the wave packet decomposition and the fact
that
2n 2n
1−
kf k2(n−1)p kf k∞,δ(n−1)p ∼ kf k p(n−1) ,δ
n

if f is a balanced N -function. We can assume kfi k p(n−1) ,δ = 1 for each i. As

n
in Lemma 3.5, write
X
fi = λfi,λ .
λ.kfi k∞,δ
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 379

We use the triangle inequality to estimate the left-hand side of (34). In

the following C will denote a large enough constant depending on n, p. As
kfi,λ k∞ . 1, we have that

kfi,λ kLp (BN ) ≤ N C .

As the right-hand side in (34) is & N −C , it follows that the contribution coming
from λfi,λ with λ . N −C is well controlled.
On the other hand, recall that by Bernstein’s inequality, kfi k∞,δ . N C .
This shows that it suffices to consider O(log δ −1 ) many terms in the triangle
inequality. Each of these terms is dealt with by using (36), Lemma 3.4 and (30).
The proof of (35) is very similar. First, a randomization argument and
Theorem 6.1 imply that
ñ n Ç å1/2 ô1/n Ç n å1/n
− 12 +ε
Y X Y
2
|fi,θ | 2n
.ε,ν N kfi k2 .
i=1 θ∈Pδ L n−1 (BN ) i=1

Combining this with the trivial inequality

ñ n Ç å1/2 ô1/n Ç n å1/n
Y X Y
2
|fi,θ | ≤ kfi k∞,δ
i=1 θ∈Pδ L∞ (BN ) i=1

and then with Hölder’s inequality gives

ñ n Ç å1/2 ô1/n
Y X
2
|fi,θ |
i=1 θ∈Pδ Lp (BN )
Ç n Ç 2n
åå1/n
n
− (n−1)p +ε Y (n−1)p
1− 2n
.ε,ν N kf k2 kfi k∞,δ(n−1)p .
i=1

Then (35) follows from interpolation, as explained before. 

At this point it is useful to introduce the local norms for g : P n−1 → C

and arbitrary balls B
Ç å1/2
2
‘ g’
X
gdσ = dσ
θ .
p,δ,B Lp (wB )
θ: δ 1/2 -cap

2n
Fix n−1 < p < ∞. To simplify notation we also introduce the following
quantities. First, define

2 n(np − 2n − p − 2)
ξ= and η= .
(p − 2)(n − 1) 2p(n − 1)2 (p − 2)
380 JEAN BOURGAIN and CIPRIAN DEMETER

For a fixed 0 ≤ β ≤ 1, consider the inequality

Ç n å1/n
g‘
Y
i dσ
i=1 Lp (BN )
(37) Ç n å 1−β Ç n åβ
n n
g‘ g‘
Y Y
.ε,ν Aβ (N )N ε i dσ i dσ p(n−1)
p,δ,BN n
,δ,BN
i=1 i=1

for arbitrary ε > 0, ν, N , gi and BN as before.

We prove the following result.
Proposition 6.3.
n−1 n2 +n
−
(a) Inequality (37) holds true for β = 1 with A1 (N ) = N 4 2p(n−1) .
(b) Moreover, if we assume (37) for some β ∈ (0, 1], then we also have (37)
for βξ with
γ
Aβξ (N ) = Aβ (N 1/2 )N βη+ 2 (1−βξ) .

Proof. The proof of (a) is an immediate consequence of Remark 5.2 and

(34). We next focus on proving (b). By using Hölder’s inequality on the N 1/2 -
ball ∆
2 2
1− (p−2)(n−1)
g‘
i dσ p(n−1) 1/2 ≤ g‘
i dσ g‘
i dσ
(p−2)(n−1)
,
n
,δ ,∆ p,δ 1/2 ,∆ 2,δ 1/2 ,∆
we get
Ç n å1/n
|g‘
Y
i dσ|
i=1 Lp (∆)
(38) Ç n Ç å1
1−ξβ ξβ n
g‘ g‘
Y
ε 1/2
.ε,ν N Aβ (N ) i dσ i dσ .
p,δ 1/2 ,∆ 2,δ 1/2 ,∆
i=1

Consider a finitely overlapping cover of BN with balls ∆ of radius N 1/2 . Note

that
Ç n å1/n Ç Ç n å1/n p å1/p
’ g‘
Y X Y
(39) |g dσ|
i . i dσ .
i=1 Lp (BN ) ∆ i=1 Lp (∆)

We will use (38) to bound each k( ni=1 |g‘ 1/n k p Q

i dσ|) L (∆) . After raising to the
th
p power, the right-hand side of (38) is summed using Hölder’s inequality
n
!ξβ n ! 1−ξβ
1−ξβ n
X ξβp Y p X p Y X p
(40) b∆ n
a∆,i ≤ b∆ a∆,i ,
∆ i=1 ∆ i=1 ∆

with
a∆,i = g‘
i dσ 1
p,δ 2 ,∆
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 381

and
1
n
!
np
p
g‘
Y
b∆ = i dσ 1 .
2,δ 2 ,∆
i=1

To sum the factors ap∆,i we invoke first Minkowski’s inequality then Propo-
sition 4.1 to get
p p p
g‘ . g‘ . Kp,n (δ 1/2 )p g‘
X
(41) i dσ 1 i dσ 1 i dσ .
p,δ 2 ,∆ p,δ 2 ,BN p,δ,BN
∆

We next show how to sum the factors bp∆ . Rather than using the n-linear
Hölder’s inequality followed by Minkowski’s inequality as we did with the terms
ap∆,i , we first transform bp∆ to make it amenable to another application of
Theorem 6.1.
To this end we recall the standard formalism (see, e.g., [14, §§2–4]) that for
each δ 1/2 -cap θ, |g’
θ dσ| is essentially constant on each ∆. Thus, in particular,
it is easy to see that
n p n Ç å 1 p
2 2n
g‘ g÷
X Y n
X Y X
i dσ . dσ
i,θ
p,δ,∆
∆⊂BN i=1 ∆⊂BN i=1 θ: δ 1/2 -cap Lp (w∆ )
θ⊂τi
(42) Ç å 1
n p
2 2n
g÷
Y X
. i,θ dσ .
i=1 θ: δ 1/2 -cap Lp (wBN )
θ⊂τi

Note also that by orthogonality followed by Hölder’s inequality, for each ∆,

n 1
( 2 − p1 )
(43) g‘
i dσ 1 . g‘
i dσ .N2 g‘
i dσ .
2,δ 2 ,∆ 2,δ,∆ p,δ,∆

Now (43), (42) and (35) lead to

n
!1/n
n
g‘
X p − (n−1) +ε n p
( −1)
Y
(44) b∆ .ε,ν N N 2 2
i dσ (n−1)p .
n
,δ,BN
∆ i=1

To end the argument simply invoke estimates (39), (40), (41) and (44). 

We can now present the final stage of the proof of Theorem 1.1 for
2(n+1)
p > n−1 . Recall from the beginning of this section that this amounts to
proving that α = 0. Since p > 2(n+1) 1
n−1 , we have that ξ < 2 . We begin with a
general discussion that applies in every dimension n and then conclude with
an inductive argument.
A simple computation reveals that the inequality α > 0 is equivalent with
1−ξ n−1 n2 + n 2η
γ > − + .
1 − 2ξ 4 2p(n − 1) 1 − 2ξ
382 JEAN BOURGAIN and CIPRIAN DEMETER

It follows that if α > 0, we can choose s0 ∈ N large enough and ε0 small enough
so that
Ç å
1−ξ ξ(2ξ)s0 n−1 n2 + n
γ − > − + 2 s 0 ε0
1 − 2ξ 1 − 2ξ 4 2p(n − 1)
(45)
2η n
+ (1 − (2ξ)s0 ) + (2ξ)s0 .
1 − 2ξ (n − 1)p
Choose now ν0 > 0 small enough such that ε0 > ε(ν0 ), with ε(ν0 ) as in
Theorem 5.3. Note that s0 , ε0 and ν0 depend only on the fixed parameters
p, n, α. As a result, we follow our convention and do not record the dependence
on them when using the symbol ..
Proposition 6.3 implies that for each s ≥ 0,
s
Aξs (N ) = N ψ(ξ ) ,
with
1 γ
(46) ψ(ξ s+1 ) = ψ(ξ s ) + (1 − ξ s+1 ) + ηξ s .
2 2
Iterating (46) gives
Å ã −s
1 η γ 2 − ξs
(47) ψ(ξ s ) = ψ(1) + γ(1 − 2−s ) + 2 − .
2s ξ 2 ξ −1 − 2
Note that by Hölder’s inequality,
Ç n å1 Ç n å1
n n n
g‘ g‘
Y Y
i dσ p(n−1) . i dσ N (n−1)p .
n
,δ,BN p,δ,BN
i=1 i=1
Combining this with (37) for β = ξ s , we get
nξs
(48) Cp,n (δ, ν0 ) .ε,s δ −ε Aξs (N )N (n−1)p .
To finish the argument, we first consider n = 2. Since (48) (with s = s0 )
holds for arbitrarily small δ and ε, using Theorem 5.3, we get
nξ s0
(49) γ − ε0 ≤ ψ(ξ s0 ) + .
(n − 1)p
Combining (47) and (49), we find
Ç å
1−ξ ξ(2ξ)s0 2η n
γ − ≤ ψ(1) + 2s0 ε0 + (1 − (2ξ)s0 ) + (2ξ)s0 ,
1 − 2ξ 1 − 2ξ 1 − 2ξ (n − 1)p
which contradicts (45), if α > 0. Thus α = 0 and Theorem 1.1 is proved for
n = 2 and p > 6.
The higher dimensional proof follows by induction. Assume that n ≥ 3 and
that Theorem 1.1 was proved for all 2 ≤ d ≤ n − 1 when p > 2(d+1)
d−1 . To prove
2(n+1) 2(n+1) 2n
Theorem 1.1 in Rn for p > n−1 , it suffices to prove it for n−1 <p< n−2 .
Note that in this range we have p < 2(d+1)
d−1 for each 2 ≤ d ≤ n − 1 and
thus Theorem 5.3 is applicable, due to the induction hypothesis. To finish the
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 383

argument, one applies the same reasoning as in the case n = 2 to conclude

that α > 0 leads to a contradiction.
It remains to see why Theorem 1.1 holds for the endpoint p = pn =
2(n+1)
n−1 . Remark 5.2 shows that it suffices to investigate the best constant in
the localized inequality
Ç å1/2
X
(3)
(50) kf kLp (BN ) ≤ Kp,n (δ) kfθ k2Lp (Rn )
θ∈Pδ
2(n+1)
for each N -ball BN . It suffices now to invoke Theorem 1.1 for p > n−1
together with
n
−n
kf kLpn (BN ) . kf kLp (BN ) N pn p (by Hölder’s inequality),
n+1
− n+1
kfθ kLp (Rn ) . N 2p 2pn kfθ kLpn (Rn ) (by Bernstein’s inequality)
and then to let p → pn .

7. Extension to other hypersurfaces

Let S be a compact C 2 hypersurface in Rn with positive definite second
fundamental form. Recall that we have proved Theorem 1.1 for P n−1 . By a
linear transformation, the proof extends to elliptic paraboloids of the form
{(ξ1 , . . . , ξn−1 , θ1 ξ12 + · · · + θn−1 ξn−1
2
) ∈ Rn : |ξi | ≤ 1/2},
with θi ∈ [C −1 , C]. The implicit bounds will of course depend on C > 0.
We now show how to extend the result in Theorem 1.1 to S as above. It
suffices to prove the result for p = 2(n+1)
n−1 . We can assume that all the principal
−1
curvatures of S are in [C , C].
The following argument is sketched in [21] and was worked out in detail
for conical surfaces in [27]. For δ < 1, as before let Kp (δ) be the smallest
constant such that for each f with Fourier support in Nδ , we have
Ç å1/2
X
kf kp ≤ Kp (δ) kfθ k2p .
θ∈Pδ
Fix such an f . First, note that
Ç å1/2
2 X
kf kp ≤ Kp (δ ) 3 kfτ k2p .
τ ∈P 2
δ3
Second, our assumption on the principal curvatures of S combined with Tay-
lor’s formula shows that on each τ ∈ P 32 , S is within δ from a paraboloid with
δ
similar principal curvatures. By invoking Theorem 1.1 for this paraboloid,
combined with parabolic rescaling (Proposition 4.1), we get
Ç å1/2
−ε
X
kfτ kp .ε δ kfθ k2p .
θ∈Pδ :θ⊂τ
384 JEAN BOURGAIN and CIPRIAN DEMETER

For each ε > 0, we conclude the existence of Cε such that for each δ < 1,
2
Kp (δ) ≤ Cε δ −ε Kp (δ 3 ).
By iteration this immediately leads to Kp (δ) .ε δ −ε .

8. Proof of Theorem 1.2

To simplify notation we present the argument for n = 3. Define the
extension operator
Z  » 
ER g(x1 , x2 , x3 ) = g(ξ1 , ξ2 )e x1 ξ1 + x2 ξ2 + x3 ξ12 + ξ22 dξ1 dξ2
R
for a subset R of the annulus
n » o
A1 = (ξ1 , ξ2 ) : 1 ≤ ξ12 + ξ22 ≤ 2 .

It will suffice to prove that

Ç å1/2
X
ε
(51) kEA1 gkL6 (BN ) .ε N kES gk2L6 (wB ) ,
N
S⊂A1

where the sum is over a tiling of A1 into sectors S of length 1 and aperture
N −1/2 . The idea behind the proof is rather simple; we will apply the decoupling
2
inequality from Theorem 1.1 to the parabola (ξ, ξ2 ). The observation that
makes this application possible is the fact that
»  ξ22
ξ12 + ξ22 − ξ1 −
2
is “small” if ξ1 is “close” to 1 and ξ2 is “close” to 0. It remains to quantify the
meaning of “small” and “close.”
The key step is the following partial decoupling for a “significant” subset
of the cone.
Proposition 8.1. Fix ν, µ > 0 such that 2µ + ν ≤ 1 and 2µ ≥ ν. Given
intervals I ⊂ [1, 2] and J ⊂ (−π/2, π/2) of lengths N −ν and N −µ respectively,
consider the sector
ß » Å ã ™
ξ2
S = (ξ1 , ξ2 ) : ξ12 + ξ22 ∈ I, arctan ∈J
ξ1
of length N −ν and aperture N −µ . For each ε > 0, we have
Ç å1/2
kES 0 gk2L6 (wB )
X
νε
kES gkL6 (BN ) .ε N ,
N
S 0 ⊂S

where the sum runs over a tiling of S into sectors S 0 of length N −ν and aperture
ν
∼ N −µ− 2 .
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 385

Proof. Due to rotational and radial symmetry it suffices to assume I =

[1, 1 + N −ν ] and J = [N −µ , 2N −µ ]. Moreover, we may also assume that ξ1 > 0,
which implies, in particular, that
|1 − ξ1 | . N −ν , |ξ22 | ∼ N −2µ .
Note that for each function F = F (x1 , x2 , x3 ), we have
1
(52) kF kL6 (BN ) ∼ 1/3 kkF (x1 − y3 , x2 + y2 , x3 + y3 )kL6y ,y (BN ) kL6x ,x ,x (BN ) .
N 2 3 1 2 3

We apply this to
F (x1 , x2 , x3 ) = ES g(x1 , x2 , x3 ).
Fix (x1 , x2 , x3 ) ∈ BN , and evaluate the inner L6 norm in (52) using the change
of variables » 
(ξ1 , ξ2 ) 7→ (η, ξ2 ) := ξ12 + ξ22 − ξ1 , ξ2 ,
whose Jacobian J(η, ξ2 ) is nonzero. We get
Z  » 
g(ξ1 , ξ2 )e x1 ξ1 + x2 ξ2 + x3 ξ12 + ξ22
S
 » 
× e y2 ξ2 + y3 ξ12 + ξ22 − ξ1 dξ1 dξ2
L6y2 ,y3 (BN )
Z
= h(η, ξ2 )e(y2 ξ2 + y3 η)dηdξ2
L6y2 ,y3 (BN )

for some appropriate function h. Note that if (ξ1 , ξ2 ) ∈ S,

 » 
ξ22 ξ22 ξ1 + ξ12 + ξ22 − 2
η− =  »  . N −2µ−ν .
2 2 ξ1 + ξ12 + ξ22

It follows that the support of h is inside the δ ∼ N −2µ−ν neighborhood of the

parabola ®Ç 2 å ´
ξ2
, ξ2 , ξ2 ∼ N −µ .
2
We can now invoke the parabolic rescaling Proposition 4.1 and Theorem 1.1
with f = h b to conclude that
Z
h(η, ξ2 )e(y2 ξ2 + y3 η)dηdξ2
L6y2 ,y3 (BN )
Ç Z 2 å1/2
X
εν
.ε N h(η, ξ2 )e(y2 ξ2 + y3 η)dηdξ2 .
−µ− ν ξ2 ∈H L6y2 ,y3 (wBN )
|H|=N 2

The conclusion of our proposition follows now by changing back to the original
variables, using Minkowski’s inequality and (52). 
386 JEAN BOURGAIN and CIPRIAN DEMETER

By iterating this proposition we get the following result.

1
Proposition 8.2. Fix ν = M , with M ≥ 2 an integer. Given intervals
ν
I ⊂ [1, 2] and J ⊂ (−π/2, π/2) of lengths N −ν and N − 2 respectively, consider
the sector
ß » Å ã ™
ξ2
S = (ξ1 , ξ2 ) : ξ12 + ξ22 ∈ I, arctan ∈J
ξ1
ν
of length N −ν and aperture N − 2 . We have
Ç å1/2
X
ε
kES gkL6 (BN ) .ε N kE∆ gk2L6 (wB ) ,
N
∆⊂S

where the sum runs over a tiling of S into sectors ∆ of length N −ν and aperture
1
∼ N−2 .
Proof. Repeatedly apply Proposition 8.1 with µ = µj = j ν2 staring with
j = 1 until j = M − 1. 

We are left with proving that inequality (51) follows from this proposition.
3ν
First, note that A1 can be tiled using ∼ N 2 sectors S of length N −ν and
ν
aperture N − 2 . Call this collection Lν . Thus, by using the Cauchy-Schwartz
inequality and invoking the above proposition, we get for arbitrary ε > 0
Ç å1/2
3ν X
kEA1 gkL6 (BN ) . N 4 kES gk2L6 (BN )
S∈Lν
(53) Ç å1/2
εν+ 3ν
X
.ε N 4 kE∆ gk2L6 (wB ) ,
N
∆⊂A1

where the sum runs over a tiling of A1 into sectors ∆ of length N −ν and
1
aperture ∼ N − 2 . We also observe the following easy inequality:

(54) kE∆ gkL6 (wB ) . kES 0 gkL6 (wB ),

N N

where S 0 is the sector in A1 of length 1 and aperture N −1/2 containing ∆.

Note that no curvature is involved in this estimate, as the N −1 neighborhood
of ∆ is essentially a rectangular parallelepiped. Since each such S 0 contains
N ν sectors ∆, by combining (53) and (54) we conclude that
Ç å1/2
εν+ 3ν + ν2
kES 0 gk2L6 (wB )
X
kEA1 gkL6 (BN ) .ε N 4 .
N
S 0 ⊂A1

Inequality (51) now follows by choosing ν appropriately small.

 THE PROOF OF THE l2 DECOUPLING CONJECTURE 387

References
[1] R. Appelbaum and M. Sharir, Repeated angles in three and four dimensions,
SIAM J. Discrete Math. 19 (2005), 294–300. MR 2178103. http://dx.doi.org/10.
1137/S0895480104443941.
[2] D. Békollé, A. Bonami, G. Garrigós, and F. Ricci, Littlewood-Paley de-
compositions related to symmetric cones and Bergman projections in tube do-
mains, Proc. London Math. Soc. 89 (2004), 317–360. MR 2078706. Zbl 1079.
42015. http://dx.doi.org/10.1112/S0024611504014765.
[3] J. Bennett, A. Carbery, and T. Tao, On the multilinear restriction and
Kakeya conjectures, Acta Math. 196 (2006), 261–302. MR 2275834. Zbl 1203.
42019. http://dx.doi.org/10.1007/s11511-006-0006-4.
[4] K. D. Boklan, The asymptotic formula in Waring’s problem, Mathematika
41 (1994), 329–347. MR 1316613. Zbl 0815.11050. http://dx.doi.org/10.1112/
S0025579300007439.
[5] E. Bombieri and J. Bourgain, A problem on sums of two squares, Internat.
Math. Res. Notices 2014, published online: March 18, 2014. http://dx.doi.org/
10.1093/imrn/rnu005.
[6] E. Bombieri and H. Iwaniec, On the order of ζ( 12 + it), Ann. Scuola Norm.
Sup. Pisa Cl. Sci. 13 (1986), 449–472. MR 0881101. Zbl 0615.10047. Available
at http://www.numdam.org/item?id=ASNSP 1986 4 13 3 449 0.
[7] E. Bombieri and H. Iwaniec, Some mean-value theorems for exponential
sums, Ann. Scuola Norm. Sup. Pisa Cl. Sci. 13 (1986), 473–486. MR 0881102.
Zbl 0615.10046. Available at http://www.numdam.org/item?id=ASNSP 1986
4 13 3 473 0.
[8] J. Bourgain, Eigenfunction bounds for the Laplacian on the n-torus, Internat.
Math. Res. Notices 1993, no. 3, 61–66. MR 1208826. Zbl 0779.58039. http:
//dx.doi.org/10.1155/S1073792893000066.
[9] J. Bourgain, Fourier transform restriction phenomena for certain lattice sub-
sets and applications to nonlinear evolution equations. I. Schrödinger equa-
tions, Geom. Funct. Anal. 3 (1993), 107–156. MR 1209299. Zbl 0787.35097.
http://dx.doi.org/10.1007/BF01896020.
[10] J. Bourgain, On Strichartz’s inequalities and the nonlinear Schrödinger equa-
tion on irrational tori, in Mathematical Aspects of Nonlinear Dispersive Equations,
Ann. of Math. Stud. 163, Princeton Univ. Press, Princeton, NJ, 2007, pp. 1–20.
MR 2331676. Zbl 1169.35054.
[11] J. Bourgain, Moment inequalities for trigonometric polynomials with spectrum
in curved hypersurfaces, Israel J. Math. 193 (2013), 441–458. MR 3038558.
Zbl 1271.42039. http://dx.doi.org/10.1007/s11856-012-0077-1.
[12] J. Bourgain and C. Demeter, New bounds for the discrete Fourier restriction
to the sphere in 4D and 5D, Internat. Math. Res. Notices 2014, published online:
March 13, 2014. http://dx.doi.org/10.1093/imrn/rnu036.
388 JEAN BOURGAIN and CIPRIAN DEMETER

[13] J. Bourgain and C. Demeter, Improved estimates for the discrete Fourier
restriction to the higher dimensional sphere, Illinois J. Math. 57 (2013), 213–
227. MR 3224568. Zbl 06322990. Available at http://projecteuclid.org/euclid.
ijm/1403534493.
[14] J. Bourgain and L. Guth, Bounds on oscillatory integral operators based on
multilinear estimates, Geom. Funct. Anal. 21 (2011), 1239–1295. MR 2860188.
Zbl 1237.42010. http://dx.doi.org/10.1007/s00039-011-0140-9.
[15] R. de la Bretèche, Répartition des points rationnels sur la cubique de Segre,
Proc. Lond. Math. Soc. 95 (2007), 69–155. MR 2329549. Zbl 1126.14025. http:
//dx.doi.org/10.1112/plms/pdm001.
[16] A. Carbery and S. I. Valdimarsson, The endpoint multilinear Kakeya the-
orem via the Borsuk-Ulam theorem, J. Funct. Anal. 264 (2013), 1643–1663.
MR 3019726. Zbl 1267.42014. http://dx.doi.org/10.1016/j.jfa.2013.01.012.
[17] F. Catoire and W.-M. Wang, Bounds on Sobolev norms for the defocusing
nonlinear Schrödinger equation on general flat tori, Commun. Pure Appl. Anal. 9
(2010), 483–491. MR 2600446. Zbl 1189.35301. http://dx.doi.org/10.3934/cpaa.
2010.9.483.
[18] C. Demeter, Incidence theory and discrete restriction estimates. arXiv 1401.
1873.
[19] S. Demirbas, Local well-posedness for 2-D Schrödinger equation on irrational
tori and bounds on Sobolev norms. arXiv 1307.0051.
[20] G. Garrigós and A. Seeger, On plate decompositions of cone multipliers,
Proc. Edinb. Math. Soc. 52 (2009), 631–651. MR 2546636. Zbl 1196.42010. http:
//dx.doi.org/10.1017/S001309150700048X.
[21] G. Garrigós and A. Seeger, A mixed norm variant of Wolff’s inequality for
paraboloids, in Harmonic Analysis and Partial Differential Equations, Contemp.
Math. 505, Amer. Math. Soc., Providence, RI, 2010, pp. 179–197. MR 2664568.
Zbl 1202.42021. http://dx.doi.org/10.1090/conm/505/09923.
[22] G. Greaves, Some Diophantine equations with almost all solutions trivial, Math-
ematika 44 (1997), 14–36. MR 1464372. Zbl 0880.11031. http://dx.doi.org/10.
1112/S002557930001192X.
[23] Z. Guo, T. Oh, and Y. Wang, Strichartz estimates for Schrödinger equations
on irrational tori, Proc. Lond. Math. Soc. 109 (2014), 975–1013. MR 3273490.
Zbl 1303.35099. http://dx.doi.org/10.1112/plms/pdu025.
[24] L. Guth, The endpoint case of the Bennett-Carbery-Tao multilinear Kakeya
conjecture, Acta Math. 205 (2010), 263–286. MR 2746348. Zbl 1210.52004. http:
//dx.doi.org/10.1007/s11511-010-0055-6.
[25] V. Jarnı́k, Über die Gitterpunkte auf konvexen Kurven, Math. Z. 24 (1926),
500–518. MR 1544776. Zbl 51.0153.01. http://dx.doi.org/10.1007/BF01216795.
[26] J. Pach and M. Sharir, Repeated angles in the plane and related problems,
J. Combin. Theory Ser. A 59 (1992), 12–22. MR 1141318. Zbl 0749.52014.
http://dx.doi.org/10.1016/0097-3165(92)90094-B.
 THE PROOF OF THE l2 DECOUPLING CONJECTURE 389

[27] M. Pramanik and A. Seeger, Lp regularity of averages over curves and

bounds for associated maximal operators, Amer. J. Math. 129 (2007), 61–103.
MR 2288738. Zbl 1161.42009. http://dx.doi.org/10.1353/ajm.2007.0003.
[28] O. Robert and P. Sargos, Un théorème de moyenne pour les sommes
d’exponentielles. Application à l’inégalité de Weyl, Publ. Inst. Math. (Beograd)
67(81) (2000), 14–30. MR 1761299. Zbl 1006.11046.
[29] C. M. Skinner and T. D. Wooley, On the paucity of non-diagonal solutions
in certain diagonal Diophantine systems, Quart. J. Math. Oxford Ser. 48 (1997),
255–277. MR 1458583. Zbl 0885.11026. http://dx.doi.org/10.1093/qmath/48.2.
255.
[30] C. D. Sogge, Propagation of singularities and maximal functions in the plane,
Invent. Math. 104 (1991), 349–376. MR 1098614. Zbl 0754.35004. http://dx.
doi.org/10.1007/BF01245080.
[31] E. Szemerédi and W. T. Trotter, Jr., Extremal problems in discrete
geometry, Combinatorica 3 (1983), 381–392. MR 0729791. Zbl 0541.05012.
http://dx.doi.org/10.1007/BF02579194.
[32] T. Tao, A sharp bilinear restrictions estimate for paraboloids, Geom. Funct.
Anal. 13 (2003), 1359–1384. MR 2033842. Zbl 1068.42011. http://dx.doi.org/
10.1007/s00039-003-0449-0.
[33] T. Tao and V. Vu, Additive Combinatorics, Cambridge Stud. Adv. Math. 105,
Cambridge Univ. Press, Cambridge, 2006. MR 2289012. Zbl 1127.11002. http:
//dx.doi.org/10.1017/CBO9780511755149.
[34] P. A. Tomas, A restriction theorem for the Fourier transform, Bull. Amer.
Math. Soc. 81 (1975), 477–478. MR 0358216. Zbl 0298.42011. http://dx.doi.
org/10.1090/S0002-9904-1975-13790-6.
[35] R. C. Vaughan and T. D. Wooley, On a certain nonary cubic form and related
equations, Duke Math. J. 80 (1995), 669–735. MR 1370112. Zbl 0847.11052.
http://dx.doi.org/10.1215/S0012-7094-95-08023-5.
[36] T. Wolff, Local smoothing type estimates on Lp for large p, Geom. Funct.
Anal. 10 (2000), 1237–1288. MR 1800068. Zbl 0972.42005. http://dx.doi.org/
10.1007/PL00001652.
(Received: April 16, 2014)
(Revised: August 19, 2014)
Institute for Advanced Study, Princeton, NJ
E-mail : [email protected]

Indiana University, Bloomington, IN

E-mail : [email protected]