For a δ-separated (δ,d)-Katz–Tao set P on the truncated cone in R3, 1<d≤2, we prove the sharp energy bound \E_{3,δ}(P)\lesssim_{\varepsilon} δ^{-\varepsilon} \max\{δ^{-(3d+2)},δ^{-(5d-1)}\}. The proof applies the Fu–Ren incidence theorem to triples with a fixed sum. As a corollary we also deduce a sharp L6 moment estimate for the Fourier transform of Frostman measures supported on the cone.
Introduction
The main results for the cone
We consider the truncated cone
C={(rcosθ,rsinθ,r):21≤r≤1},
and the key notation associated with it
Q(x,y,t)=t2−x2−y2,e(θ)=(cosθ,sinθ).
A probability measure μ on C is called d-Frostman if there is a constant C such that μ(B(x,r))≤Crd for all 0<r≤2. We use μ(ξ)=∫e−2πix⋅ξdμ(x).
Theorem 1.1. Let 1<d≤2, and let μ be a probability measure supported on C such that
μ(B(x,r))≤Crd,0<r≤2.(1)
For every ε>0 and R≥2,
∫B(0,R)∣μ(ξ)∣6dξ≲ε,CRεmax{R5−3d,R2−d}.(2)
When d≤1 the supporting cone loses its relevance, as the measure could live entirely on a line segment.
The exponents 5−3d, 2−d are both attainable, as explained in Remark 4.3.
In the last section we derive Theorem 1.1 from a discrete energy theorem, whose proof occupies most of this paper.
Key words and phrases. Additive energy, cone, Katz–Tao sets, planar incidences.
The author is partially supported by NSF grant DMS-2349828.
We call a subset P of a metric space a (δ,d,C)-KT set if (and we will call this a KT non-concentration condition; KT is short for Katz–Tao who first showed the relevance of such sets for continuous problems)
#(P∩B(x,ρ))≤C(ρ/δ)d,δ≤ρ≤2.(3)
Let P⊂C be a finite δ-separated set, N=#P. Condition (3) gives N≲Cδ−d. We define the (3,δ)-energy (in previous literature this was sometimes called (6,δ)-energy) by
Both exponents occur for sets with #P≳δ−d, as we will prove below.
The case 0<d≤1 does not have any substance. The trivial upper bound O(N5) for the energy is easily attainable by placing the points on a line inside the cone, in arithmetic progression. This is precisely possible because the Hausdorff dimension of the line is at least as big as the “dimension” d of the points.
From here on we assume that 1<d≤2.
The flat example. The δ-neighborhood of the truncated cone is nearly planar on an angular sector of width ∼δ. Fix θ and let e=e(θ) and e⊥=e′(θ). Points of the form
p(r,y)=(re+ye⊥,r2+y2),21≤r≤43,0≤y≤cδ(5)
lie on a fixed such sector of C. Note that
p(r,y)=(re+ye⊥,r)+O(c2δ).(6)
Take r in a δ-progression and y in a δd/2-progression. Each grid has
n≍δ−1δd/2δ=δ−(d+1)/2
points. The planar grid in (6) has ≳n5 additive 6-tuples: independently choose five points in the grid, and a positive proportion of the choices determine a sixth point within the grid. By (6), these are O(δ)-additive 6-tuples on the cone.
We select J≲δ−(d−1)/2 angular sectors (of width ∼δ) with separation ≲δ(d−1)/2. For each sector use the grid above, and we let P be their union. Then
#P≍Jn≳δ−d,E3,δ(P)≳Jn5=δ−(3d+2).
The Katz–Tao condition follows by counting grid points in a ball. For ρ≤δ a ball meets O(1) grids and has O((ρ/δ)(1+ρ/δd/2)) points. For ρ≥δ it meets O(ρ/δ(d−1)/2) grids and has O(ρδ−(d+1)/2) points in each. Both bounds are O((ρ/δ)d), as needed.
The radial arithmetic progression example. Let
h=δd−1,Rh={21+jh:0≤j≲h−1},
and let Θ be a δ-separated subset of [0,2π], with #Θ≍δ−1. Define
P={(rcosθ,rsinθ,r):r∈Rh,θ∈Θ}.
Then #P≍h−1δ−1=δ−d. Moreover, for δ≤ρ≤1,
#(P∩B(x,ρ))≲(1+ρ/h)ρ/δ≲(ρ/δ)d.
The height coordinate of every triple sum belongs to Rh+Rh+Rh, which contains only O(h−1) elements. At each such height, the two remaining sum coordinates occupy O(δ−2) squares of side O(δ) in the plane. Thus the N3≳δ−3d triple sums occupy at most O(h−1δ−2)=O(δ−(d+1))δ-cubes in R3. Cauchy–Schwarz gives
E3,δ(P)≳δ−(d+1)N6≳δ−(5d−1).(7)
The two lower bounds cross at d=3/2. For 1<d≤3/2 the flat term dominates, while for 3/2≤d≤2 the radial term dominates. Our main theorem gives the corresponding upper bound. In fact, as a by-product of our method, a more refined bound emerges, one that takes into account the size N of the set.
Theorem 1.2. Let 1<d≤2 and let P⊂C be a δ-separated (δ,d,C)-KT set, with N=#P. For every ε>0,
E3,δ(P)≲ε,C,C∗Nδ−εmax{δ−(2d+2),δ−(4d−1)}.(8)
In particular, (4) holds and is sharp up to δ−ε for sets of maximal cardinality.
The conjectures for the parabolic cylinder and the parabola
Consider the truncated parabolic cylinder
Y={(x,x2,z):x∈[0,1],z∈[0,1]}⊂R3.
The cylinder is often a good model for the cone, as both surfaces have zero Gaussian curvature. The analogous conjecture we make for a (δ,d)-KT set S⊂Y, 1<d≤2, is
E3,δ(S)≲εδ−εmax{δ−(3d+2),δ−(5d−1)}.(9)
The two exponents are the same as for the cone, and their justification is also similar. The first obstruction comes from δ-wide strips of the cylinder, which are nearly planar at scale δ. For the second, we use a set which projects to an arithmetic progression in the x-coordinate. Both constructions yield the same exponents as above, via the same type of counting. The cylinder conjecture is at least as hard as the corresponding 3-energy conjecture for the parabola, introduced in [6].
Conjecture 1.3 (3-energy on the parabola). Let 0<s≤1, and let A⊂{(x,x2):0≤x≤1} be a δ-separated (δ,s,C)-KT set. Then for every ε>0,
E3,δ(A)≲ε,C,C∗δ−εmax{δ−3s,δ−(5s−1)}.(10)
Indeed, let A⊂{(x,x2):0≤x≤1} be a maximal (δ,s)-KT set, where s=d−1, and let Z be a δ-arithmetic progression of length comparable to δ−1. Then S=A×Z, with the coordinates ordered as above, is a (δ,s+1)-KT set. Each approximate additive 6-tuple in A can be paired with ≳δ−5 exact additive 6-tuples in Z. Thus
E3,δ(S)≳δ−5E3,cδ(A)
for a fixed c=O(1). Applying (9) with d=s+1 would give
E3,δ(A)≲εδ−εmax{δ−3s,δ−(5s−1)},0<s≤1,(11)
which is precisely the conjectured parabola numerology. This gives a concrete sense in which the cylinder problem is harder. We do not know a comparable implication from the cone estimate to the parabola estimate.
Orponen [6] initiated the study of the δ-energy problem for the parabola, deriving a power improvement over the (comparatively speaking, elementary) fourth-moment estimate, using Furstenberg-set theory. Subsequent progress is described in [2], where O’Regan and the author prove the bound E3,δ(A)≲εδ−7s/2−ε for s≤2/3. We refer to that paper for the intervening results and their references. More recently, Orponen and Shmerkin [7, Theorem 6.9] proved the sharp triple-sumset dimension formula dimH(A+A+A)=3s for every s-Ahlfors–David regular subset A of the parabola with 0<s≤21. This Hausdorff-dimension result does not by itself imply the discretized energy estimate in Conjecture 1.3.
There is a contrast with familiar Fourier-analytic estimates. The cone and parabolic cylinder share similar critical exponents for a multitude of problems: they have the same critical ℓ2L6 decoupling exponent, and comparable critical restriction and L4 square-function phenomena. The cone is usually (both conceptually and technically) much harder, notably in reverse square-function estimates, see [4] and [1], because its curved and flat directions interact under rescaling. In our context however, the cone energy admits a full resolution, whereas the cylinder conjecture contains the unsolved parabola energy problem.
Remark 1.4 (Energy and the ring problem). For the parabola, Orponen’s improvement over the elementary energy exponent is connected to the discretized ring problem and to Furstenberg-set estimates. This is one motivation for the Furstenberg-theoretic methods used in [2]. We do not know an analogous implication from improved energy on the cone to a ring or sum–product theorem. The radial variable and the angular variable do not combine into the same one-dimensional additive–multiplicative structure. Our proof below uses planar incidences of lower complexity than the Furstenberg-set theorems known to have non-trivial sum-product consequences.
A warm-up: exact energy (=0) and Szemerédi–Trotter
We explain how the argument works in the technically simpler case when δ=0. Let P⊂C be finite, without any non-concentration assumptions. For each angle θ, the segment
Lθ={(rcosθ,rsinθ,r):21≤r≤1}
is called a generator of the truncated cone. Let M be the largest number of points of P on any generator, and write
E3(P)=#{(p1,…,p6)∈P6:p1+p2+p3=p4+p5+p6}.
We prove the discrete analog of Theorem 1.2.
Proposition 2.1. For every finite P⊂C,
E3(P)≲N13/3+NM4.
Remark 2.2. The estimate in Proposition 2.1 serves only as a warm-up for the transformation and -to some extent- for the incidence mechanism used below. Jing and Wu [5] recently proved that, for a finite set P on a fixed irreducible algebraic surface of degree at least two (such as the cone), its fourth energy obeys
E2(P)≲εMN2+ε,
where M≥1 bounds the number of points on any line contained in the surface. Applying Young’s convolution inequality gives
E3(P)=∥1P∗1P∗1P∥22≤N2∥1P∗1P∥22≲εMN4+ε.
This is stronger than the N13/3 term of Proposition 2.1 for bounded M, though the bounds have different advantages when M is large. We present the elementary Szemerédi–Trotter proof solely to introduce the nonlinear incidence transformation needed in the δ-problem.
In preparation for the proof of Proposition 2.1, we distinguish the two possible cases for the triple sum of points on the cone. If pi=ri(e(θi),1) with 21≤ri≤1 and q=p1+p2+p3, then the definition of Q gives
Here we used ∣e(θ)∣=1 and e(θi)⋅e(θj)=cos(θi−θj). Subtracting gives (12). Thus Q(q)≥0.
We call a sum qzero if Q(q)=0 and nonzero if Q(q)>0. Since every summand in (12) is nonnegative, a sum is zero exactly when all three points lie on one generator. Nonzero means that at least two of their angular directions differ. Only the nonzero case requires incidence geometry, while the zero case is trivial. We start with the nonzero case.
For q∈P+P+P write
n(q)=#{(p1,p2,p3)∈P3:p1+p2+p3=q}.
Then E3(P)=∑qn(q)2 and ∑qn(q)=N3. The distinguished feature of our argument is that we will derive uniform bounds for n(q), when Q(q)>0. This allows us to fix q. We will apply two transformations. The first one normalizes the geometry of points, to lay the ground for uniform estimates for n(q). The second one recasts triple sums contributing to n(q) as planar incidence estimates between points and lines.
A fixed sum and the linear Lorentz rescaling
Write q=(q1,q2,t). Since the cone and its truncation are invariant under rotations in the horizontal (x,y)-plane, we may rotate the coordinates so that q=(s,0,t), where s=q12+q22≥0. This rotation preserves distances and additive relations. For a nonzero sum Q(q)>0, we have t>s. Set h=t2−s2 and consider the Lorentz rescaling
Bq(x,y,z)=(htx−sz,y,htz−sx),Aq=h−1Bq.(13)
The map Aq is linear and satisfies Aqq=(0,0,1). Moreover, Q(Aqp)=h−2Q(p), so Aq maps the upper cone to itself. By linearity, if p1+p2+p3=q, then
Aqp1+Aqp2+Aqp3=(0,0,1).
Apply Aq to all three points. Their images still lie on the upper cone, and their sum is now (0,0,1). We may therefore write
Aqpi=ri′(e(θi′),1).
The primed radii and angles generally differ from the original ones; in particular, the transformed radii need not belong to [1/2,1]. For points in a triple with sum q, however, they satisfy absolute bounds. Indeed, writing the original points as pi=ri(e(θi),1), the formula for Aq gives
Here we used ri≥1/2, t=r1+r2+r3≤3, and 0≤s<t, hence t+s≤6. On the other hand, the normalized sum equation gives
r1′+r2′+r3′=1,i=1∑3ri′e(θi′)=0.
Thus, for each i, the triangle inequality implies
ri′=j=i∑rj′e(θj′)≤j=i∑rj′=1−ri′,
so ri′≤1/2. We have proved that every point in a contributing triple to q has
121≤ri′≤21.
Remark 2.3. We record a more refined observation. This is not needed in the discrete argument presented in this section, but will be needed when we deal with δ-approximate incidences. If the transformed radii are ordered as r(1)′≥r(2)′≥r(3)′, then 2r(2)′+1/12≤1. In particular, at least two of the three transformed radii belong to [1/12,11/24]⊂[1/24,1/2−1/24].
To simplify notation, from now on we omit the primes and write pi=ri(e(θi),1) for the transformed points.
The nonlinear change of coordinates
At this point, we could recast a triple contributing to q as an incidence between points and ellipses having two degrees of freedom. The desired estimate would be a consequence of the generalized Szemerédi–Trotter theorem due to Pach and Sharir, [8]. We instead follow an alternative approach. The role of the nonlinear map V is to straighten ellipses into lines. This perspective will be useful later when we estimate the δ-energy, and apply an incidence theorem available for straight tubes.
For the transformed sum q=(0,0,1), direct calculation gives
For 0<r<1/2, choose a continuous angular coordinate θ on a fixed interval of length less than 2π and define the non-linear
V(p)=V(r,θ)=1−2r2re(θ/2),p=r(e(θ),1).(15)
The notation V(p) means that the two parameters r and θ are the radial and angular coordinates of the cone point p; thus V(pi)=V(ri,θi). The angle is defined modulo 2π, but e(θ/2) changes sign when θ is increased by 2π. We therefore use finitely many angular intervals on which θ is chosen continuously, making V single-valued.
Since r1+r2+r3=1 and all three radii are positive, at most one of them can be greater than or equal to (or better said, equal to) 1/2. Thus every triple has at least two points whose transformed radii belong to (0,1/2), where V is well-defined. We apply V only to these two points, which we choose to be p1,p2, and split the count according to the three possible choices of pair. The third point is uniquely determined by the sum equation. Later, in § 3.3, we will take the two selected radii in the compact sub-interval [1/48,1/2−1/48], which is needed for the quantitative distortion estimates.
We next turn the equality
p1+p2+p3=q=(0,0,1)
into an incidence between a point (namely V(p2)) and a line parametrized by V(p1). The role of p3 will be solely coming from the fact that it belongs to the cone. The fact that it is in P is irrelevant.
Since 0<r1,r2<1/2, the prefactor is nonzero. Hence Q(q−p1−p2)=0 if and only if
V(p1)⋅V(p2)=±1.(16)
Fix p1 and write V(p1)=(a,b) and V(p2)=(x,y). Since V(p1)=0, equation (16) says
ax+by=1orax+by=−1
Each equation describes a line in the (x,y)-plane. Thus, for a fixed choice of sign, p1 determines a line ℓp1, and every compatible p2 gives an incidence V(p2)∈ℓp1.
We check that no multiplicities are introduced. First, the map p↦V(p) is injective. Indeed,
∣V(p)∣2=1−2r2r
and the right-hand side is strictly increasing for 0<r<1/2. Hence V(p) determines r. Its direction determines θ/2, and therefore θ, once a continuous choice of angle has been fixed. Second, for a fixed sign the line ax+by=±1 determines (a,b) uniquely: the right-hand side is fixed and nonzero. Hence the map from p1 to its line is injective as well. We count the two signs separately, so that writing u⋅v=1 and (−u)⋅v=−1 does not identify two lines within the same family.
Nonzero sums
Lemma 2.4.For every q with Q(q)>0, n(q)≲N4/3.
Proof. We recall the reduction from the previous two subsections. Since Q(q)>0, after rotating the horizontal coordinates we may write q=(s,0,t) with h=t2−s2>0. Thus the Lorentz map Aq is well-defined, maps the upper cone to itself, and satisfies Aqq=(0,0,1). By linearity, a triple with p1+p2+p3=q satisfies Aqp1+Aqp2+Aqp3=(0,0,1). Writing the transformed points as ri(e(θi),1), comparison of the third coordinates gives r1+r2+r3=1. In particular, at least two radii are strictly smaller than 1/2. This is needed in order to make the use of V possible. Split according to the choice of the pair. For each such pair, (16) gives incidences between at most N distinct points and at most N distinct lines. Szemerédi–Trotter gives O(N4/3+N) incidences. Once the pair is fixed, the third point is uniquely determined. There are only three choices of pair and finitely many charts.
It follows immediately that
Q(q)>0∑n(q)2≤(Q(q)>0supn(q))q∑n(q)≲N13/3.(17)
:::
Zero sums
Lemma 2.5.The contribution to E3(P) from Q(q)=0 is O(NM4).
Proof. By (12), Q(q)=0 forces all three points of a contributing triple to lie on the same generator. Two triples with the same zero sum lie on the same generator as each other. If a generator contains m points, its energy is at most m5 because five points determine the sixth. Summing over generators gives ∑m5≤M4∑m=NM4.
Proof of Proposition 2.1. Combine (17) and Lemma 2.5.
Remark 2.6 (Exact and thickened energy). For a δ-separated maximal-size KT set one always has M≲δ−1, and the exact theorem gives E3(P)≲δ−13/3+δ−(d+4). This is not the sharp thickened exponent. In the latter problem a δ-wide plate, not merely a generator, behaves as a plane at tolerance O(δ). Ordinary Szemerédi–Trotter is therefore not the needed incidence input.
The δ-energy estimate
This section is devoted to proving Theorem 1.2. All implicit constants may depend on C and C∗. Partition the bounded region containing P+P+P into cubes of side cδ, with c>0 fixed. Let n(q) be the number of triples with sum in the cube q. We may write
E3,δ(P)≲q∑n(q)2,q∑n(q)=N3.(18)
For a triple (p1,p2,p3), its angular diameter is
where distS1 is the shorter angular distance. We say that the triple has angular diameter σ when this number lies between σ/2 and σ. Identity (12), together with 1/2≤ri≤1, shows that Q(p1+p2+p3)≍σ2: the largest angular separation contributes ∼σ2, while the other two contribute O(σ2).
Assume now σ2≫δ. For a sum cube q and for a triple satisfying
p1+p2+p3=q+O(δ),diamang(p1,p2,p3)≍σ
we may still conclude that
Q(q)≍σ2,
since Q(q)=Q(p1+p2+p3)+O(δ). The scales below C0δ, with C0 a sufficiently large fixed constant, are included in the near-zero contribution of Lemma 3.3: their sums satisfy ∣Q(q)∣=O(δ). We apply Lorentz normalization only when σ≥C0δ. We work dyadically in this remaining range C0δ≤σ≲1.
Lemma 3.1 (The sector determined by a sum cube). Fix δ≪σ≤1. There is a boundedly overlapping family Tσ of angular arcs τ of length ≲σ and a rule assigning each sum cube q with ∣Q(q)∣≲σ2 to one arc τ(q)∈Tσ, with the following property. If
∣Q(q)∣≲σ2,p1+p2+p3=q+O(δ),
then the directions of p1,p2,p3 all lie in τ(q).
When ∣Q(q)∣=O(δ), we may use a family Tδ consisting of arcs τ of length ≲δ.
Proof. We show the argument in the case δ≪σ≤1. The other case follows by simple modifications. Write pi=ri(e(θi),1), R=∑iri≍1, and x=∑irie(θi). Thus S:=p1+p2+p3=(x,R). Since S=q+O(δ) and δ≲σ2, Q(S)=Q(q)+O(δ)≲σ2. The identity
i=1∑3rie(θi)−Rx2=R−R∣x∣2=RQ(S)(19)
and ri≥1/2 imply ∣e(θi)−x/R∣≲σ for each i. If σ is sufficiently small, then ∣x∣2=R2−Q(S) gives ∣x∣≲1, and hence
e(θi)−∣x∣x≲σ.
Writing v=(q1,q2), we also have ∣v−x∣≲δ, so ∣v∣≲1 and
distS1(θi,argv)≲σ,i=1,2,3.(20)
It follows that there is L≲σ so that each contributing triple θ1,θ2,θ3 lies inside an L-arc. Cover the circle with a boundedly overlapping collection Tσ of 3L/2-arcs, so that each L-arc lies inside some τ∈Tσ.
Thus Pτcone consists of the original cone points whose angles belong to τ.
Lemma 3.2. For δ≲σ≲1,
τ∈TσmaxNτ≲δ−dσd−1,τ∈Tσ∑Nτ≲N.(21)
Proof. An angular sector of aperture O(σ) on the truncated cone is covered by O(σ−1) balls of radius O(σ). Apply (3) in each ball. The second assertion follows from bounded overlap.
Lemma 3.3 (Near-zero sums). Triples whose sum cube has ∣Q(q)∣=O(δ) contribute O(Nδ−2d−2) to E3,δ(P).
Proof. Apply Lemma 3.1. For each output cube q with ∣Q(q)∣=O(δ), the lemma places all triples counted by n(q) in the same sector τ(q). Within this sector, five points determine the sixth up to O(1) possibilities, since P is δ-separated. By bounded overlap the contribution is at most
Consider a sum cube centered at q such that Q(q)≍σ2≫δ. By Lemma 3.1, all triples contributing to n(q) use points in the single sector τ=τσ(q). Let q=(q1,q2,t). Rotate the horizontal coordinates so that (q1,q2) becomes (s,0), where s=(q12+q22)1/2. This rotation preserves the value of Q(q). It also preserves the angular diameter, in fact all angular parameters change by the same constant. Consequently, the comparison Q(q)≍σ2, obtained before rotation, remains valid afterward. Since q=(s,0,t) and t>0, we have
t2−s2=Q(q)≍σ2>0,h:=(t2−s2)1/2=Q(q)≍σ.
We now use Aq from (13). If
p1+p2+p3=q+O(δ)(22)
then
Aqp1+Aqp2+Aqp3=(0,0,1)+O(δ/σ2).(23)
Indeed, the symmetric matrix representing Aq has eigenvalues (t+s)/h2, (t−s)/h2, and 1/h; the first two have eigenvectors (1,0,−1) and (1,0,1). Hence its Euclidean operator norm is
∥Aq∥op=∣v∣=1sup∣Aqv∣=h2t+s≲σ−2.
Thus an additive error O(δ) becomes O(δ/σ2) under Aq.
As explained in Remark 2.3, we may assume that the radii of at least two of the three points Aqp1,Aqp2,Aqp3 satisfy, say (assuming the error term O(δ/σ2) is small enough)
1/48≤ri′,rj′≤1/2−1/48.
Any three points on the circle belong to an arc of length at most 4π/3, as follows by considering the three consecutive angular gaps. Fix a collection of 24 open arcs γ of length 3π/2, with equally spaced starting points. Every arc of length 4π/3 is contained in at least one of these γ. It follows that any triple Aq(p1),Aq(p2),Aq(p3) will have their polar angles θi′ inside some γ. The only relevance is that the length of γ is <2π, as used below.
Let τ be the arc associated with q via Lemma 3.1. We may and will assume it is centered at the origin. Indeed, after rotating the horizontal direction of q, the original angles of all points in Pτcone satisfy ∣θ∣≲σ. This is the angular hypothesis used in Lemma 3.4 below. We now bound n(q) using this sector and the finite family of transformed-angle charts γ.
The planar map for a fixed $\gamma. We aim to do planar incidence geometry, so we shift the discussion to the plane. Recall our notation p=r(e(θ),1), Aqp=r′(e(θ′),1). We define the planar map
T=Tq:(r,θ)⟼V(r′,θ′)=1−2r′2r′e(θ′/2).
Here V is the nonlinear map from Section 2. On the fixed chart we require 1/48≤r′≤1/2−1/48 and the transformed angle θ′∈γ. Define the planar points
The points of Pτ lie in R2 and have cardinality O(Nτ). They arise from the points Pτcone on the cone-sector via two transformations, whose combined action is encoded by the useful planar map T. They are first mapped via the linear map Aq, then via the nonlinear map V (which has a linearizing effect, leading to lines). Since the domain of V will need θ′ in an interval shorter than 2π, we restrict Pτ to an arc γ. This is a rather unimportant technicality.
The next lemma is instrumental in determining the KT non-concentration of the points Pτ.
Lemma 3.4 (The map T). Let q=(s,0,t) and h=(t2−s2)1/2≍σ. The intermediate polar coordinates of Aqp satisfy
r′=h2r(t−scosθ),tan2θ′=ht+stan2θ.(24)
Suppose ∣θ∣≲σ and 1/48≤r′≤1/2−1/48. For a Euclidean ball of radius w in the target V-plane, the preimage under T is contained in a region of radial width O(w) and angular width O(σw). All constants are uniform in σ.
Differentiating shows ∣∂r/∂r′∣≲1 and ∣∂r/∂θ′∣≲1: indeed, the θ-derivative of the coefficient is O(σ−1), which is multiplied by dθ/dθ′=O(σ).
It remains to take into account V. On 1/48≤r′≤1/2−1/48,
V(r′,θ′)=f(r′)e(θ′/2),f(r′)=1−2r′2r′.
Indeed,
f′(r′)=2r′(1−2r′)3/21,
so both f and f′ are bounded above and below by positive constants on this fixed radial interval. To see explicitly why V has a uniformly bi-Lipschitz inverse on its image, take two points uj=V(rj′,θj′)=ρje(θj′/2), where ρj=f(rj′) and θj′∈γ. Then
For the last comparison, the mean value theorem controls ∣ρ1−ρ2∣ by ∣r1′−r2′∣ in both directions, while ∣θ1′−θ2′∣≤3π/2 ensures that ∣sin((θ1′−θ2′)/4)∣≍∣θ1′−θ2′∣. The constants are independent of q and σ. Equivalently, the inverse can be written as
r′=2(1+∣u∣2)∣u∣2,θ′=2Argγ/2(u),
where Argγ/2 denotes the branch of the argument taking values in the lifted interval γ/2. This explains the role of restricting the transformed angle to an arc γ of length less than 2π: it allows a consistent single-valued angular branch (indeed, V(r′,θ′+2π)=−V(r′,θ′)) and gives the uniform chord–angle comparison above. The finite collection of such charts ensures that no contributing triple is lost. Consequently, the inverse image of a w-ball under V has r′-width and θ′-width O(w). The preceding derivative bounds for the inverse Lorentz coordinate map then imply radial width O(w) and angular width O(σw) in the original coordinates. □
The transformed incidence scale. Recall that
∥Aq∥op=∣v∣=1sup∣Aqv∣=h2t+s≍σ−2.
Thus an additive error O(δ) becomes O(δ/σ2). More precisely, Q(Aqv)=h−2Q(v) and Aqq=(0,0,1), so
Q((0,0,1)−Aqpi−Aqpj)=h−2Q(q−pi−pj)=O(δ/σ2).
For each a∈Pτ define
La±={v∈R2:a⋅v=±1}
and let
Lτ±={La±:a∈Pτ}.
The identity (16) relating Q((0,0,1)−Aqpi−Aqpj) to V(Aqpi)⋅V(Aqpj) shows that a compatible selected pair gives a point of Pτ within O(δ/σ2) of a line in Lτ+ or Lτ−. Set
η:=δ/σ2.
Lemma 3.5 (Rescaled Katz–Tao constant). Let τ∈Tσ, σ≫δ. The planar point collection Pτ and the line collections Lτ± are (η,d,Kσ)-KT sets, where
Kσ≲σ−d−1.(25)
Proof. By Lemma 3.4, the preimage under T of a w-ball with w≥η has radial width O(w) and angular width O(σw). It can be covered by O(σ−1) balls of radius O(σw). Since σw≥ση=δ/σ≥δ, the KT hypothesis for the original points on the cone (which translates into a similar condition in the (r,θ) variables) gives
#{a∈Pτ:a∈B(v,w)}≲σ−1(σw/δ)d=σ−d−1(w/η)d.
The map u↦{v:u⋅v=±1} is bi-Lipschitz whenever ∣u∣ is bounded above and away from zero. Both requirements hold, since u=V(r′,θ′) and 1/48≤r′≤1/2−1/48. Therefore the same KT bound transfers from Pτ to Lτ±. □
The incidence bound for one sum cube
We introduce the replacement for the Szemerédi–Trotter theorem. The Fu–Ren result takes advantage of the key features of our point-tube incidences derived in Lemma 3.5: non-concentration of both points and tubes.
Theorem 3.6 (Fu–Ren [3], Theorem 1.5). Let 1<d≤2 and 0<η<1. Let P be a collection of points and L a collection of η-neighborhoods of unit line segments in the plane. Assume that the point collection and the line-parameter collection satisfy the (η,d) Katz–Tao bounds with constants KP and KL, respectively. Then, for every ϵ>0,
Lemma 3.7 (Fixed-sum fibre). For a sum cube q with Q(q)≲σ2≫δ, let τ=τ(q) be its assigned sector from Lemma 3.1. Then, letting Nτ=#Pτcone
n(q)≲ϵδ−c−ϵσ−2cdNτ2−2c,c=max{2d−1,2}1.(27)
Proof. By Lemma 3.1, all triples counted by n(q) lie in Pcone, and the angular hypothesis of Lemma 3.4 holds on this sector after rotation. Normalize q by Aq. Every normalized triple satisfies
r1′2+r2′2+r3′2=1+O(η).
The radial observation preceding the definition of T, based on Remark 2.3, gives a uniform positive lower bound on all three radii and places at least two in [1/48,1/2−1/48]. We split into the three possible choices of this pair.
For the selected pair, the thickened condition Q(q−pi−pj)=O(δ) becomes Q((0,0,1)−Aqpi−Aqpj)=O(η). Equations (14)–(15) then turn it into an O(η)-thickened point–line incidence, between Nτ points and Nτ lines. Lemma 3.5 and (26) show that the number of admissible pairs is at most
η−c−εKσ2cNτ2−2c≲εδ−c−εσ−2cdNτ2−2c.
Once the two original points are fixed, the third lies in an original O(δ)-ball, giving O(1) possibilities by separation. Notice that it is the original δ, not the larger normalized thickness η, that is used in this last step.
Proof of Theorem 1.2. For 1<d≤3/2, c=1/2 and Lemma 3.8 reads
Eσ≲εNδ−3d−1/2−εσ2d−3.
Since 2d−3≤0, the largest value is at σ≍δ, and equals Nδ−2d−2−ε.
For 3/2<d≤2, c=(2d−1)−1 and the same lemma reads
Eσ≲εNδ−(4d−1)−εσ4d−6.
Here 4d−6>0, so the largest value is at σ≲1. The near-zero contribution in Lemma 3.3 is O(Nδ−2d−2), which is bounded by Nδ−(4d−1) for d≥3/2. There are O(log(1/δ)) dyadic scales σ, and we absorb this loss into δ−ε.
L⁶ Fourier decay for fractal measures
We now prove Theorem 1.1, which we recall below. The derivation from the energy bound is rather standard. One decomposes μ into 0,1-valued densities μn that are locally constant at scale δ. The different contributions are evaluated using the triangle inequality. The centers of the δ-balls supporting μn satisfy KT non-concentration (3), albeit with constants that are not O(1). Random sampling is typically used in the literature to go from large to O(1) constants. The even-ness of 6 allows for the sixth moment of μn to be expressed in terms of the 3-energy of the centers. Rather than pursuing this somewhat technical approach, we follow a more elegant random sampling that could be of independent interest.
Theorem 4.1. Let 1<d≤2, and let μ be a probability measure supported on C such that
μ(B(x,r))≤Crd,0<r≤2.(29)
For every ε>0 and R≥2,
∫B(0,R)∣μ(ξ)∣6dξ≲ε,CRεmax{R5−3d,R2−d}.(30)
We first pass from point sets to measures at scale δ.
Lemma 4.2. Under the hypotheses of Theorem 4.1, for 0<δ≪1 and any fixed A>0,
Proof. Choose a maximal δ-separated set {xj} in suppμ and partition suppμ into measurable cells Dj associated with this net. Each Dj is contained in B(xj,δ). Write mj=μ(Dj). The points xj lie on C, are δ-separated, and satisfy mj≤C1δd by (29). Set aj=mj/(C1δd)≤1.
We follow a probabilistic approach. Independently for each k=1,…,6 and each j, select the center xj with probability aj. Denote the resulting six random sets by Sk, and put S=⋃k=16Sk.
We first prove the following: for every D>0 there is a constant KD such that, with probability at least 1−OD(δD),
#(S∩B(x,ρ))≤KDlog(1/δ)(ρ/δ)d,δ≤ρ≤2.(32)
Here are the details. For each j let Yj=1{xj∈S}. The variables Yj are independent as j varies, and
P(Yj=1)=1−(1−aj)6≤6aj.
Let rℓ=2ℓδ, 0≤ℓ≤L, where L is the smallest integer such that rL≥2. For each center xj of the original δ-separated net (whether or not xj is selected into S), put
Xj,ℓ:=#(S∩B(xj,2rℓ))=i:xi∈B(xj,2rℓ)∑Yi.
Since the cells Di are disjoint and Di⊂B(xi,δ), while δ≤rℓ, we have
When 3rℓ>2, the last inequality follows instead from μ(B)≤1 and rℓ>2/3. Put tℓ=Klog(1/δ)(rℓ/δ)d, with K to be chosen. The Chernoff bound for sums of independent Bernoulli variables gives
P(Xj,ℓ≥tℓ)≤(tℓeλj,ℓ)tℓ≤2−tℓ≤δKlog2,
provided K is sufficiently large in terms of C,d. There are O(log(1/δ)) choices of ℓ and O(δ−3) original centers xj, by separation and the boundedness of suppμ⊂C. Taking the union bound over these pairs (j,ℓ), the probability that any estimate fails is at most
O(δ−3log(1/δ)δKlog2)=OD(δD),
if K is large enough depending on D.
Fix D>18, and choose K=KD sufficiently large above. Let G denote the event that Xj,ℓ≤tℓ simultaneously for every center xj of the original net and every dyadic radius rℓ. The preceding union bound shows that
P(Gc)≲DδD.
For every outcome in G, we first verify that the non-concentration estimate (32) holds for arbitrary balls. Take a ball B(x,ρ) with δ≤ρ≤2. If S∩B(x,ρ) is empty, there is nothing to prove. Otherwise, choose one of its selected points xj∈S∩B(x,ρ). For every z∈S∩B(x,ρ), ∣z−xj∣≤∣z−x∣+∣x−xj∣<2ρ, and hence
S∩B(x,ρ)⊂S∩B(xj,2ρ)⊂S∩B(xj,2rℓ),
where ℓ is chosen so that ρ≤rℓ<2ρ. The simultaneous estimates for the original centers therefore give
#(S∩B(x,ρ))≤Xj,ℓ≤tℓ≤2dKlog(1/δ)(ρ/δ)d.
Absorbing 2d into KD proves (32) simultaneously for all centers and radii on G.
Let Γ consist of (deterministic) ordered six-tuples of cell indices whose representatives satisfy ∣xj1+xj2+xj3−xj4−xj5−xj6∣≤(A+6)δ. We first relate this discrete sum to the measure on the left of (31). Write
FA,δ:={(p1,…,p6)∈C6:k=1∑3pk−k=4∑6pk≤Aδ}.
If (p1,…,p6)∈FA,δ and pk∈Djk, then ∣pk−xjk∣≤δ for each k, and hence
Independence of the six selections, even when some indices coincide, gives
EZ=Γ∑aj1⋯aj6.
As mj=C1δdaj, (33) now yields
μ⊗6(FA,δ)≤Γ∑mj1⋯mj6=(C1δd)6EZ.(34)
For each outcome, Z is bounded by the energy E3,δ(S), where we take the fixed tolerance constant C∗=A+6 in the definition of that energy; the underlying separation scale remains δ. Split the expectation using the good event G:
EZ=E(Z1G)+E(Z1Gc).
On G, the Katz–Tao estimate (32) holds for S. Consequently, for each individual outcome ω∈G, Theorem 1.2, applied to the δ-separated set S(ω)⊂C with tolerance constant C∗=A+6, gives the uniform pointwise upper bound
On Gc we use only the deterministic estimate Z≤#Γ≲δ−18, as there are O(δ−3) original net points. Thus
E(Z1Gc)≲δ−18P(Gc)≲DδD−18.
Our fixed choice D>18 makes this term bounded (indeed vanishing) as δ→0. Together with the bound on G, this yields
EZ≲ε,C,Aδ−εmax{δ−(3d+2),δ−(5d−1)}.
Multiplying by (C1δd)6 in (34) proves (31).
Proof of Theorem 4.1. Choose a nonnegative, smooth, compactly supported function g on R3, of sufficiently small support and with ∫g=1. Put ϕ=g∗g, where g(x)=g(−x). Then ϕ≥0, ϕ=∣g∣2≥c>0 on B(0,1), and ϕ is supported in a fixed ball B(0,A). With ν=μ∗μ∗μ and ν(E)=ν(−E), Fourier inversion gives
Apply Lemma 4.2 with δ=R−1. The last expression is
≲εR3−6d+εmax{R3d+2,R5d−1}=Rεmax{R5−3d,R2−d},
as required.
Remark 4.3. We close by presenting the sharpness of the exponents in Theorem 4.1. For each R≥2, set δ=R−1 and choose a maximal-size (δ,d)-KT set P that maximizes the energy bound. Let μ be normalized surface measure on the union of small δ-neighborhoods of P within the cone. The KT condition makes μ a d-Frostman measure with constant independent of R. A standard positive-kernel argument then converts the discrete energy lower bound into the corresponding Fourier lower bound. Thus the exponents are sharp uniformly over d-Frostman measures, allowing the measure to depend on R.
Use of AI. The mathematical results in this paper are mostly due to the author. ChatGPT (GPT-6, OpenAI) was used to sharpen the exposition, to generate the map V and the alternative proof to Theorem 4.1 presented here.
Department of Mathematics, Indiana University, 831 East 3rd St., Bloomington, IN 47405, USA.
[1]A. Cushman, C. Demeter and S. Wu, Diameter-free reverse inequalities and superorthogonality, arXiv:2609.13105, 2026.arxiv.org/abs/2609.13105
[2]C. Demeter and W. O’Regan, Incidence estimates for quasi-product sets and applications, arXiv:2511.15899, 2025.arxiv.org/abs/2511.15899
[3]Y. Fu and K. Ren, Incidence estimates for α-dimensional tubes and β-dimensional balls in ℝ², Journal of Fractal Geometry, 11 (2024), no. 1-2, 1-30
[4]L. Guth, H. Wang and R. Zhang, A sharp square function estimate for the cone in ℝ³, Annals of Mathematics 192 (2020), 551–581.arxiv.org/abs/1909.10693
[5]Y. Jing and S. Wu, Near diagonal additive energy bound for points on algebraic surfaces, arXiv:2608.14467, 2026.arxiv.org/abs/2608.14467
[6]T. Orponen, Additive properties of fractal sets on the parabola, arXiv:2205.02770.arxiv.org/abs/2205.02770
[8]J. Pach and M. Sharir, On the number of incidences between points and curves, Combin. Probab. Comput. 7 (1998), 121–127DOI