betaFeedback

WorksMathematics

Simultaneous popular polynomial differences in higher dimensions

Abstract

We extend the simultaneous popular polynomial difference theorem of Conlon, Dong, and the author to polynomial shifts along prescribed directions in FpD\mathbb{F}_p^D. Let P1,…,Pk∈Z[t]P_1,\ldots,P_k\in\mathbb{Z}[t] be linearly independent polynomials with zero constant terms, and let v1,…,vk∈ZD∖{0}v_1,\ldots,v_k\in\mathbb{Z}^D\setminus\{0\}. We show that, for every ε>0\varepsilon>0, every sufficiently large prime pp, and every A⊆FpDA\subseteq\mathbb{F}_p^D of density αα, there exists a nonzero d∈Fpd\in\mathbb{F}_p such that Ex‾∈FpD1A(x‾)∏1≤i≤kωi=11A(x‾+Pi(d)vi)≥α 1+∑i=1kωi−ε \mathbb{E}_{\underline{x}\in\mathbb{F}_p^D} \mathbf{1}_A(\underline{x}) \prod_{\substack{1\leq i\leq k\\ ω_i=1}} \mathbf{1}_A\bigl(\underline{x}+P_i(d)v_i\bigr) \geq α^{\,1+\sum_{i=1}^kω_i}-\varepsilon simultaneously for every ω∈{0,1}kω\in\{0,1\}^k. The main new ingredient is a directional arithmetic regularity decomposition adapted to mixed directional box norms.

Introduction

Popular polynomial difference theorems seek a nonzero parameter for which the normalized count of a prescribed polynomial configuration is at least its expected value in a random set of the same density, up to a small error. A simultaneous version asks for a single parameter that gives these lower bounds for several polynomial configurations at once. Conlon, Dong, and the author [1] established such a result over finite fields. In this paper, we extend their theorem to configurations in FpD\mathbb{F}_p^D, with polynomial shifts along prescribed directions.

To state our result, fix D,k∈ND,k\in\mathbb{N}, and set G:=FpDG:=\mathbb{F}_p^D, where pp is an odd prime. Let V={v1,…,vk}V=\{v_1,\ldots,v_k\} be a collection of nonzero vectors in ZD\mathbb{Z}^D, and let P={P1,…,Pk}\mathcal{P}=\{P_1,\ldots,P_k\} be a collection of linearly independent polynomials in Z[t]\mathbb{Z}[t] with zero constant terms. For d∈Fpd\in\mathbb{F}_p and functions f0,…,fk:G→Cf_0,\ldots,f_k:G\to\mathbb{C}, define

IP,V(f0,…,fk)(d):=Ex∈Gf0(x)∏i=1kfi(x+Pi(d)vi).I_{\mathcal{P},V}(f_0,\ldots,f_k)(d):=\mathop{\mathbb{E}}_{x\in G}f_0(x)\prod_{i=1}^k f_i\left(x+P_i(d)v_i\right).

When all the functions are 1A\mathbf{1}_A, this average gives the normalized count of polynomial configurations x,x+P1(d)v1,…,x+Pk(d)vkx,x+P_1(d)v_1,\ldots,x+P_k(d)v_k contained in AA, with the shifts determined by the common parameter dd. We consider subconfigurations obtained by keeping xx and any subset of the shifted points. Given ω=(ω1,…,ωk)∈{0,1}k\omega=(\omega_1,\ldots,\omega_k)\in\{0,1\}^k, let (P,V)ω(\mathcal{P},V)_\omega denote the subcollection of pairs (Pi,vi)(P_i,v_i) for which ωi=1\omega_i=1, and set

I(P,V)ω(f0,…,fk)(d):=Ex∈Gf0(x)∏1≤i≤kωi=1fi(x+Pi(d)vi).I_{(\mathcal{P},V)_\omega}(f_0,\ldots,f_k)(d):=\mathop{\mathbb{E}}_{x\in G}f_0(x)\prod_{\substack{1\leq i\leq k\\\omega_i=1}}f_i\left(x+P_i(d)v_i\right).

Here and throughout, an empty product is understood to be 1. With this notation, our main result is as follows.

Theorem 1.1. Let D,k∈ND,k\in\mathbb{N}. Let V={v1,…,vk}V=\{v_1,\ldots,v_k\} be a collection of nonzero vectors in ZD\mathbb{Z}^D, and let P={P1,…,Pk}⊂Z[t]\mathcal{P}=\{P_1,\ldots,P_k\}\subset\mathbb{Z}[t] be a collection of linearly independent polynomials with zero constant terms. For every ε>0\varepsilon>0, there exists p0=p0(ε,P,V)p_0=p_0(\varepsilon,\mathcal{P},V) such that for every prime p≥p0p\geq p_0 and every A⊆FpDA\subseteq\mathbb{F}_p^D of density α\alpha, there exists d∈Fp∖{0}d\in\mathbb{F}_p\setminus\{0\} such that

I(P,V)ω(1A,…,1A)(d)≥α1+∑i=1kωi−εI_{(\mathcal{P},V)_\omega}(\mathbf{1}_A,\ldots,\mathbf{1}_A)(d)\geq\alpha^{1+\sum_{i=1}^k\omega_i}-\varepsilon

holds simultaneously for every ω=(ω1,…,ωk)∈{0,1}k\omega=(\omega_1,\ldots,\omega_k)\in\{0,1\}^k.

The same nonzero parameter dd therefore works for every choice of ω\omega. In particular, it gives the lower bound αk+1−ε\alpha^{k+1}-\varepsilon for the full configuration and the lower bound α2−ε\alpha^2-\varepsilon for each of the kk two-term configurations. No linear-independence assumption is imposed on the vectors v1,…,vkv_1,\ldots,v_k.

The study of popular polynomial differences has roots in ergodic theory. Frantzikinakis and Kra [2] proved a Khintchine-type recurrence theorem for rationally independent polynomials, giving a lower bound for recurrence along the full polynomial configuration. In a related direction, Lyall and Magyar [5] investigated a different strengthening of Khintchine’s recurrence theorem, which they called simultaneous polynomial recurrence. Their result shows that a single parameter can satisfy the Khintchine-type lower bound for each of several two-term polynomial configurations. These results suggest asking whether a single parameter can give the corresponding lower bounds for the full configuration and all its subconfigurations simultaneously. Conlon, Dong, and the author [1] established this simultaneous conclusion over Fp\mathbb{F}_p. More recently, Shao and Zhang [7] obtained simultaneous popular difference results over Fp\mathbb{F}_p for certain polynomial configurations with linearly dependent shift polynomials. In a complementary direction, Theorem 1.1 extends the result of [1] to polynomial shifts along prescribed directions in FpD\mathbb{F}_p^D.

The proof of Theorem 1.1 follows the regularity approach developed in [1]. In the one-dimensional setting, a generalized von Neumann estimate derived from Peluse’s polynomial Szemerédi theorem [6] controls contributions involving a component that is small in the U2U^2-norm. Tao’s arithmetic regularity decomposition [8] then reduces the argument for an arbitrary set to estimating a structured contribution and a small error.

In the multidimensional setting, Kuca’s polynomial Szemerédi theorem [4] yields a generalized von Neumann estimate in terms of the mixed directional box norms ∥⋅∥G,⟨vi⟩\|\cdot\|_{G,\langle v_i\rangle}; see Lemma 3.2. To use this estimate, we need a regularity decomposition adapted to each prescribed direction. Proposition 3.3 provides these decompositions and is the main new ingredient in the proof. The resulting structured components need not coincide. We construct them using symmetrized Bohr kernels so that the convolution inequality in Lemma 3.6 gives the required lower bound for their joint contribution.

The remainder of the paper is organized as follows. § 2 introduces the notation and records the preliminary estimates. In § 3, we prove Theorem 1.1 using the directional arithmetic regularity decomposition.

Notation and preliminaries

We collect the notation and the directional Fourier and Bohr-set estimates used in the proof.

Conventions

All averages are normalized. Thus, for a finite nonempty set XX,

Ex∈Xf(x):=1∣X∣∑x∈Xf(x),∥f∥ℓq(X):=(Ex∈X∣f(x)∣q)1/q(1≤q<∞).\mathbb{E}_{x\in X}f(x):=\frac{1}{|X|}\sum_{x\in X}f(x),\qquad\|f\|_{\ell^q(X)}:=\left(\mathbb{E}_{x\in X}|f(x)|^q\right)^{1/q}\quad(1\leq q<\infty).

We omit an averaging domain when it is evident from the variables. For example, Ef\mathbb{E}f denotes the normalized mean of ff on its domain. For a set EE, the function 1E\mathbf{1}_E denotes its indicator. The density of a subset A⊆XA\subseteq X is Ex∈X1A(x)=∣A∣/∣X∣\mathbb{E}_{x\in X}\mathbf{1}_A(x)=|A|/|X|. We also put ∥f∥ℓ∞(X):=max⁡x∈X∣f(x)∣\|f\|_{\ell^\infty(X)}:=\max_{x\in X}|f(x)|. We say that ff is 11-bounded if ∣f(x)∣≤1|f(x)|\leq1 for every xx.

Throughout the paper, pp denotes an odd prime. Let D∈ND\in\mathbb{N}, and denote G=FpDG=\mathbb{F}_p^D. Linear independence of polynomials in Z[t]\mathbb{Z}[t] is understood over Q\mathbb{Q}. Integer vectors and polynomials are reduced modulo pp when they are used over Fp\mathbb{F}_p. For v∈Gv\in G, write ⟨v⟩:={tv:t∈Fp}\langle v\rangle:=\{tv:t\in\mathbb{F}_p\}. For a∈Fpa\in\mathbb{F}_p, write ep(a):=exp⁡(2πia/p)e_p(a):=\exp(2\pi ia/p). For f:Fp→Cf:\mathbb{F}_p\to\mathbb{C}, our normalized Fourier transform is

f^(ξ):=Ex∈Fpf(x)ep(−ξx).\widehat{f}(\xi):=\mathbb{E}_{x\in\mathbb{F}_p}f(x)e_p(-\xi x).

For x,h∈Gx,h\in G, define the multiplicative derivative

△hf(x):=f(x)f(x+h)‾.\triangle_h f(x) := f(x)\overline{f(x+h)}.

We write △h1,…,hsf:=△h1⋯△hsf\triangle_{h_1,\ldots,h_s}f := \triangle_{h_1}\cdots\triangle_{h_s}f for the iterated multiplicative derivative, with the operators applied by composition.

We use Landau’s notation OO, oo, and Vinogradov’s notation ≪\ll, ≫\gg. A subscript records the data on which the implicit constant may depend.

Directional Fourier coefficients and box norms

We introduce the Fourier coefficients and box norms used to control pseudorandomness in each prescribed direction.

For a nonzero v∈Gv\in G, ξ∈Fp\xi\in\mathbb{F}_p, and f:G→Cf:G\to\mathbb{C}, define the Fourier coefficient of ff along vv by

f^(x;v;ξ):=En∈Fpf(x+nv)ep(−nξ).(1)\widehat{f}(x;v;\xi):=\mathbb{E}_{n\in\mathbb{F}_p}f(x+nv)e_p(-n\xi). \tag*{(1)}

If x′=x+svx'=x+sv, then

f^(x′;v;ξ)=ep(sξ)f^(x;v;ξ),\widehat{f}(x';v;\xi)=e_p(s\xi)\widehat{f}(x;v;\xi),

so the modulus of (1) is constant on cosets of ⟨v⟩\langle v\rangle.

For subspaces H1,…,Hs≤GH_1,\ldots,H_s\leq G, define the Gowers box norm along H1,…,HsH_1,\ldots,H_s by

∥f∥H1,…,Hs:=(Ex∈GEh1∈H1⋯Ehs∈Hs△h1,…,hsf(x))1/2s.\|f\|_{H_1,\ldots,H_s}:=\left(\mathbb{E}_{x\in G}\mathbb{E}_{h_1\in H_1}\cdots\mathbb{E}_{h_s\in H_s}\triangle_{h_1,\ldots,h_s}f(x)\right)^{1/2^s}.

The usual Gowers U2U^2-norm is the special case ∥f∥U2(G):=∥f∥G,G\|f\|_{U^2(G)}:=\|f\|_{G,G}. The notation ∥f∥G,⟨v⟩\|f\|_{G,\langle v\rangle} denotes the box norm with H1=GH_1=G and H2=⟨v⟩H_2=\langle v\rangle. This is the mixed directional box norm used in [3].

The following estimate relates the directional Fourier coefficients to the box norm appearing in the counting estimate; see [3 Lemma 2.2].

Lemma 2.1. Let v∈G∖{0}v\in G\setminus\{0\} and f:G→Cf:G\to\mathbb{C}. If ff is 11-bounded, then

∥f∥G,⟨v⟩4≤sup⁡ξ∈FpEx∈G∣f^(x;v;ξ)∣2.(2)\|f\|_{G,\langle v\rangle}^{4}\leq\sup_{\xi\in\mathbb{F}_p}\mathbb{E}_{x\in G}|\widehat{f}(x;v;\xi)|^2. \tag*{(2)}

Bohr sets

Bohr sets describe the shifts under which the structured components are almost invariant. We record their definition and a standard lower bound for their size.

For t∈T:=R/Zt\in\mathbb{T}:=\mathbb{R}/\mathbb{Z}, write ∥t∥T\|t\|_{\mathbb{T}} for the distance to the nearest integer. If Γ⊆Fp∖{0}\Gamma\subseteq\mathbb{F}_p\setminus\{0\} and 0<ρ≤1/20<\rho\leq1/2, define

B(Γ,ρ):={b∈Fp:∥ξbp∥T≤ρ for every ξ∈Γ}.B(\Gamma,\rho):=\left\{b\in\mathbb{F}_p:\left\|\frac{\xi b}{p}\right\|_{\mathbb{T}}\leq\rho\text{ for every }\xi\in\Gamma\right\}.

Its rank is ∣Γ∣|\Gamma|, and we adopt the convention B(∅,ρ)=FpB(\varnothing,\rho)=\mathbb{F}_p.

We use the following standard estimate, which can be found in [TV06, Lemma 4.20]:

∣B(Γ,ρ)∣p≥ρ∣Γ∣.(3)\frac{|B(\Gamma,\rho)|}{p}\geq\rho^{|\Gamma|}. \tag*{(3)}

Directional Bohr kernels

We use symmetrized Bohr kernels to construct the structured components in the regularity decomposition. The kernel estimates below control their Fourier multipliers and their change under translations.

For functions ϕ,ψ:Fp→C\phi,\psi:\mathbb{F}_p\to\mathbb{C}, define the normalized convolution and involution by

(ϕ∗ψ)(t):=Es∈Fpϕ(s)ψ(t−s),ϕ~(t):=ϕ(−t).(\phi*\psi)(t):=\mathbb{E}_{s\in\mathbb{F}_p}\phi(s)\psi(t-s),\qquad\widetilde{\phi}(t):=\phi(-t).

For v∈Gv\in G, b∈Fpb\in\mathbb{F}_p, and h:G→Ch:G\to\mathbb{C}, put

Tϕvh(x):=Et∈Fpϕ(t)h(x−tv),τbϕ(t):=ϕ(t+b),τbvh(x):=h(x+bv).T_\phi^v h(x):=\mathbb{E}_{t\in\mathbb{F}_p}\phi(t)h(x-tv),\qquad\tau_b\phi(t):=\phi(t+b),\qquad\tau_b^v h(x):=h(x+bv).

For a real number uu, put (u)+:=max⁡{u,0}(u)_+ := \max\{u,0\}. Let ∅≠Γ⊆Fp∖{0}\varnothing\ne\Gamma\subseteq\mathbb{F}_p \setminus\{0\} and 0<ρ≤1/40 < \rho\le1/4. Following Tao’s smoothing approach [8], we use the normalized Lipschitz cutoff associated with B(Γ,ρ)B(\Gamma,\rho):

μΓ,ρ(t):=(1−ρ−1max⁡ξ∈Γ∥ξtp∥T)+Es∈Fp(1−ρ−1max⁡ξ∈Γ∥ξsp∥T)+.\mu_{\Gamma,\rho}(t) := \frac{\left(1-\rho^{-1}\max_{\xi\in\Gamma}\left\|\frac{\xi t}{p}\right\|_{\mathbb{T}}\right)_+}{\mathbb{E}_{s\in\mathbb{F}_p}\left(1-\rho^{-1}\max_{\xi\in\Gamma}\left\|\frac{\xi s}{p}\right\|_{\mathbb{T}}\right)_+}.

For Γ=∅\Gamma=\varnothing, we set μ∅,ρ=1\mu_{\varnothing,\rho}=1. For fixed Γ\Gamma and ρ\rho, write μ:=μΓ,ρ\mu:= \mu_{\Gamma,\rho}. The structured components will be formed using the symmetrized kernel μ∗μ~\mu* \widetilde{\mu}. The following lemma records the Fourier and translation estimates needed for these averages.

Lemma 2.2. Let μ:=μΓ,ρ\mu:= \mu_{\Gamma,\rho} for some Γ⊆Fp∖{0}\Gamma\subseteq\mathbb{F}_p \setminus\{0\} and 0<ρ≤1/40 < \rho\le1/4. Then, the functions μ\mu and μ∗μ~\mu* \widetilde{\mu} are nonnegative and have mean one. Moreover, we have

∥μ∥L2(Fp)≪∣Γ∣,ρ1,(4)\|\mu\|_{L^2(\mathbb{F}_p)} \ll_{|\Gamma|,\rho} 1, \tag*{(4)}
μ∗μ~^(ξ)=∣μ^(ξ)∣2,(5)\widehat{\mu* \widetilde{\mu}}(\xi)=|\widehat{\mu}(\xi)|^2, \tag*{(5)}
∣1−μ^(ξ)∣≪ρ(ξ∈Γ∪{0}).(6)|1-\widehat{\mu}(\xi)|\ll\rho\qquad(\xi\in\Gamma\cup\{0\}). \tag*{(6)}

For 0<σ≤1/20<\sigma\le1/2, if b∈B(Γ,σ)b\in B(\Gamma,\sigma), then

∥τb(μ∗μ~)−μ∗μ~∥L1(Fp)≪∣Γ∣,ρσ.(7)\|\tau_b(\mu* \widetilde{\mu})-\mu* \widetilde{\mu}\|_{L^1(\mathbb{F}_p)}\ll_{|\Gamma|,\rho}\sigma. \tag*{(7)}

Proof of Lemma 2.2. If Γ=∅\Gamma=\varnothing, then μ=1\mu=1 and all assertions are immediate. Now, we suppose that Γ≠∅\Gamma\ne\varnothing. The cutoff before normalization takes values in [0,1][0,1] and is at least 1/21/2 on B(Γ,ρ/2)B(\Gamma,\rho/2). By (3), its normalized mean is bounded below by a positive constant depending only on ∣Γ∣|\Gamma| and ρ\rho, and this gives (4).

By construction, μ\mu is nonnegative and has mean one, and the same holds for μ∗μ~\mu* \widetilde{\mu}. The convolution identity for the Fourier transform gives (5).

For ξ∈Γ∪{0}\xi\in\Gamma\cup\{0\} and t∈supp⁡μt\in\operatorname{supp}\mu, we have ∥ξt/p∥T≤ρ\|\xi t/p\|_{\mathbb{T}}\le\rho, and hence ∣1−ep(−ξt)∣≤2πρ|1-e_p(-\xi t)|\le2\pi\rho. Since μ\mu is nonnegative and has mean one, it follows that

∣1−μ^(ξ)∣≤Et∈Fpμ(t)∣1−ep(−ξt)∣≤2πρ.|1-\widehat{\mu}(\xi)|\le\mathbb{E}_{t\in\mathbb{F}_p}\mu(t)|1-e_p(-\xi t)|\le2\pi\rho.

This proves (6).

Finally, let b∈B(Γ,σ)b\in B(\Gamma,\sigma). The triangle inequality on T\mathbb{T} implies that the unnormalized cutoff changes by at most σ/ρ\sigma/\rho under translation by bb. Using the lower bound for the normalizing denominator, we obtain

∥τbμ−μ∥L1(Fp)≪∣Γ∣,ρσ.\|\tau_b\mu-\mu\|_{L^1(\mathbb{F}_p)}\ll_{|\Gamma|,\rho}\sigma.

Since ∥μ~∥L1(Fp)=1\|\widetilde{\mu}\|_{L^1(\mathbb{F}_p)}=1, Young’s inequality therefore gives

∥τb(μ∗μ~)−μ∗μ~∥L1(Fp)≤∥τbμ−μ∥L1(Fp)≪∣Γ∣,ρσ,\|\tau_b(\mu* \widetilde{\mu})-\mu* \widetilde{\mu}\|_{L^1(\mathbb{F}_p)} \le\|\tau_b\mu-\mu\|_{L^1(\mathbb{F}_p)} \ll_{|\Gamma|,\rho}\sigma,

and this proves (7). □

We now express these kernel properties in terms of directional averaging operators. The resulting identities will be used to compare successive averages and control their changes under translation.

Let Γ,Γ′⊆Fp∖{0}\Gamma,\Gamma'\subseteq\mathbb{F}_p\setminus\{0\}, 0<ρ,ρ′≤1/40<\rho,\rho'\le1/4, b,ξ∈Fpb,\xi\in\mathbb{F}_p, and let v∈G∖{0}v\in G\setminus\{0\}. Write

μ:=μΓ,ρ,μ′:=μΓ′,ρ′.\mu:=\mu_{\Gamma,\rho},\qquad\mu':=\mu_{\Gamma',\rho'}.

If h:G→Ch:G\to\mathbb{C}, then

Tμ∗μ~ vh^(x;v;ξ)=∣μ^(ξ)∣2h^(x;v;ξ).(8)\widehat{\mathcal{T}_{\mu*\widetilde{\mu}}^{\,v}h}(x;v;\xi)=|\widehat{\mu}(\xi)|^2\widehat{h}(x;v;\xi). \tag*{(8)}

and fibrewise Parseval gives

∥Tμ′∗μ~′vh−Tμ∗μ~vh∥L2(G)2=∑ξ∈FpEx∈G∣h^(x;v;ξ)∣2(∣μ′^(ξ)∣2−∣μ^(ξ)∣2)2.(9)\left\|T_{\mu'*\widetilde{\mu}'}^v h-T_{\mu*\widetilde{\mu}}^v h\right\|_{L^2(G)}^2 =\sum_{\xi\in\mathbb{F}_p}\mathbb{E}_{x\in G}\left|\widehat{h}(x;v;\xi)\right|^2\left(\left|\widehat{\mu'}(\xi)\right|^2-\left|\widehat{\mu}(\xi)\right|^2\right)^2. \tag*{(9)}

The translation estimate passes to the smoothed function: if 0≤h≤10\le h\le1, then

∥τbvTμ∗μ~vh−Tμ∗μ~vh∥L∞(G)≤∥τb(μ∗μ~)−μ∗μ~∥L1(Fp).(10)\left\|\tau_b^vT_{\mu*\widetilde{\mu}}^v h-T_{\mu*\widetilde{\mu}}^v h\right\|_{L^\infty(G)} \leq\left\|\tau_b(\mu*\widetilde{\mu})-\mu*\widetilde{\mu}\right\|_{L^1(\mathbb{F}_p)}. \tag*{(10)}

Indeed, a change of variables gives τbvTμ∗μ~vh=Tτb(μ∗μ~)vh\tau_b^vT_{\mu*\widetilde{\mu}}^v h=T_{\tau_b(\mu*\widetilde{\mu})}^v h, and the bound follows from ∣h∣≤1|h|\leq1.

Proof of Theorem 1.1

Following the approach developed in [1], we construct a regularity decomposition in each prescribed direction. The generalized von Neumann estimate controls the pseudorandom contributions. A Bohr-set estimate provides many parameters for which the structured components are almost invariant under the corresponding polynomial shifts, and a convolution inequality gives the required lower bounds for their contributions. We then select one nonzero parameter that works for every choice of ω\omega.

A generalized von Neumann estimate

We first establish an L2(Fp)L^2(\mathbb{F}_p) estimate for the polynomial counting averages in terms of the mixed directional box norms. This estimate will allow us to discard the pseudorandom contributions after applying the regularity decomposition. We begin with Kuca’s finite-field multidimensional polynomial Szemerédi theorem [4 Theorem 1.2].

Theorem 3.1. Let D,r∈ND,r\in\mathbb{N}. Let W={w1,…,wr∈ZD∖{0}}\mathcal{W}=\{w_1,\ldots,w_r\in\mathbb{Z}^D\setminus\{0\}\}, and R={P1,…,Pr∈Z[t]}\mathcal{R}=\{P_1,\ldots,P_r\in\mathbb{Z}[t]\} be a collection of linearly independent polynomials with zero constant terms. There exist p1=p1(W,R)p_1=p_1(\mathcal{W},\mathcal{R}), and c=c(R)>0c=c(\mathcal{R})>0 such that for every prime p≥p1p\geq p_1, we have

Ex∈FpDEd∈Fpf0(x)∏i=1rfi(x+Pi(d)wi)=Ex∈FpDf0(x)∏i=1rEni∈Fpfi(x+niwi)+OR(p−c),\mathbb{E}_{x\in\mathbb{F}_p^D}\mathbb{E}_{d\in\mathbb{F}_p}f_0(x)\prod_{i=1}^{r}f_i(x+P_i(d)w_i)=\mathbb{E}_{x\in\mathbb{F}_p^D}f_0(x)\prod_{i=1}^{r}\mathbb{E}_{n_i\in\mathbb{F}_p}f_i(x+n_iw_i)+O_{\mathcal{R}}(p^{-c}),

for all 11-bounded functions f0,…,fr:FpD→Cf_0,\ldots,f_r:\mathbb{F}_p^D\to\mathbb{C}.

The following consequence applies uniformly to every subconfiguration. The bound can be obtained from any one of the functions gig_i, i∈Si\in S.

Lemma 3.2. Let D,k∈ND,k\in\mathbb{N}, and denote G=FpDG=\mathbb{F}_p^D. Let V={v1,…,vk∈ZD∖{0}}\mathcal{V}=\{v_1,\ldots,v_k\in\mathbb{Z}^D\setminus\{0\}\}, and P={P1,…,Pk∈Z[t]}\mathcal{P}=\{P_1,\ldots,P_k\in\mathbb{Z}[t]\} be a collection of linearly independent polynomials with zero constant terms. There exist p0=p0(V,P)p_0=p_0(\mathcal{V},\mathcal{P}), and c=c(P)>0c=c(\mathcal{P})>0 such that for every prime p≥p0p\geq p_0, we have

Ed∈Fp∣Ex∈Gg0(x)∏j∈Sgj(x+Pj(d)vj)∣2≤min⁡i∈S∥gi∥G,⟨vi⟩2+OP(p−c)(11)\mathbb{E}_{d\in\mathbb{F}_p}\left|\mathbb{E}_{x\in G}g_0(x)\prod_{j\in S}g_j(x+P_j(d)v_j)\right|^2 \leq\min_{i\in S}\|g_i\|_{G,\langle v_i\rangle}^2+O_{\mathcal{P}}(p^{-c}) \tag*{(11)}
≤min⁡i∈S(sup⁡ξ∈FpEx∈G∣gi^(x;vi;ξ)∣2)1/2+OP(p−c),(12)\leq\min_{i\in S}\left(\sup_{\xi\in\mathbb{F}_p}\mathbb{E}_{x\in G}\left|\widehat{g_i}(x;v_i;\xi)\right|^2\right)^{1/2}+O_{\mathcal{P}}(p^{-c}), \tag*{(12)}

for every nonempty S⊆{1,…,k}S\subseteq\{1,\ldots,k\} and all 11-bounded functions g0:G→Cg_0:G\to\mathbb{C} and gi:G→Cg_i:G\to\mathbb{C} for i∈Si\in S.

Proof of Lemma 3.2. Expanding the square and writing the second averaging variable as x+hx+h gives

Ed∈Fp∣Ex∈Gg0(x)∏j∈Sgj(x+Pj(d)vj)∣2=Ex,h∈GEd∈Fp△hg0(x)∏j∈S△hgj(x+Pj(d)vj).\mathbb{E}_{d\in\mathbb{F}_p}\left|\mathbb{E}_{x\in G}g_0(x)\prod_{j\in S}g_j(x+P_j(d)v_j)\right|^2 =\mathbb{E}_{x,h\in G}\mathbb{E}_{d\in\mathbb{F}_p}\triangle_hg_0(x)\prod_{j\in S}\triangle_hg_j(x+P_j(d)v_j).

Since there are only finitely many subfamilies, we fix one exponent c>0c>0, one implicit constant, and one prime threshold that work for all of them. For each fixed hh, apply Theorem 3.1 to the subfamily indexed by SS. Its error is uniform in hh, since every multiplicative derivative remains 1-bounded. We obtain

Ed∈Fp∣Ex∈Gg0(x)∏j∈Sgj(x+Pj(d)vj)∣2=Ex,hΔhg0(x)∏j∈SEnj∈FpΔhgj(x+njvj)+OP(p−c).\mathbb{E}_{d\in\mathbb{F}_p}\left|\mathbb{E}_{x\in G}g_0(x)\prod_{j\in S}g_j\left(x+P_j(d)v_j\right)\right|^2 =\mathbb{E}_{x,h}\Delta_hg_0(x)\prod_{j\in S}\mathbb{E}_{n_j\in\mathbb{F}_p}\Delta_hg_j\left(x+n_jv_j\right)+O_P\left(p^{-c}\right).

Fix i∈Si\in S. Bounding all factors except the ith by 1, and then using Cauchy–Schwarz yields

Ed∈Fp∣Ex∈Gg0(x)∏j∈Sgj(x+Pj(d)vj)∣2≤∣Ex,hΔhg0(x)∏j∈SEnj∈FpΔhgj(x+njvj)∣+OP(p−c)≤(Ex,h∣Eni∈FpΔhgi(x+nivi)∣2)1/2+OP(p−c)=∥gi∥G,⟨vi⟩2+OP(p−c).\begin{aligned} \mathbb{E}_{d\in\mathbb{F}_p}\left|\mathbb{E}_{x\in G}g_0(x)\prod_{j\in S}g_j\left(x+P_j(d)v_j\right)\right|^2 \leq\left|\mathbb{E}_{x,h}\Delta_hg_0(x)\prod_{j\in S}\mathbb{E}_{n_j\in\mathbb{F}_p}\Delta_hg_j\left(x+n_jv_j\right)\right|+O_P\left(p^{-c}\right) \\ &\leq\left(\mathbb{E}_{x,h}\left|\mathbb{E}_{n_i\in\mathbb{F}_p}\Delta_hg_i\left(x+n_iv_i\right)\right|^2\right)^{1/2}+O_P\left(p^{-c}\right) \\ &=\|g_i\|_{G,\langle v_i\rangle}^2+O_P\left(p^{-c}\right). \end{aligned}

Taking the minimum over ii proves (11), and Lemma 2.1 gives (12).

Directional arithmetic regularity decomposition

The preceding estimate motivates a decomposition adapted to each prescribed direction. Following Tao’s arithmetic regularity argument [8], we construct these decompositions with structured components defined using symmetrized Bohr kernels. The directional Fourier coefficients of the pseudorandom components are small in L2(G)L^2(G), uniformly in frequency, and the remaining components have small L2(G)L^2(G) norms.

Proposition 3.3. Let D,k∈ND,k\in\mathbb{N}, and denote G=FpDG=\mathbb{F}_p^D. Let 0<ε<10<\varepsilon<1 and let η1,η2:[1,∞)→[2,∞)\eta_1,\eta_2:[1,\infty)\to[2,\infty) be nondecreasing growth functions. Let pp be prime, let v1,…,vk∈G∖{0}v_1,\ldots,v_k\in G\setminus\{0\}, and let f:G→[0,1]f:G\to[0,1]. Then there are finite sets Γi⊆Fp∖{0}\Gamma_i\subseteq\mathbb{F}_p\setminus\{0\}, a radius σ∈(0,1/4]\sigma\in(0,1/4], and parameters ρ1,ρ2∈(0,1/2]\rho_1,\rho_2\in(0,1/2] such that

max⁡1≤i≤k∣Γi∣=Oε,k,η1,η2(1),ρ1≤1η1(max⁡i∣Γi∣+ε−1),ρ2≤1η2(ρ1−1).(13)\max_{1\leq i\leq k}|\Gamma_i|=O_{\varepsilon,k,\eta_1,\eta_2}(1),\qquad \rho_1\leq\frac{1}{\eta_1\left(\max_i|\Gamma_i|+\varepsilon^{-1}\right)},\qquad \rho_2\leq\frac{1}{\eta_2\left(\rho_1^{-1}\right)}. \tag*{(13)}

The parameters ρ1,ρ2\rho_1,\rho_2 and the radius σ\sigma are bounded away from zero in terms of ε\varepsilon, kk, η1\eta_1, and η2\eta_2 alone. For every 1≤i≤k1\leq i\leq k, put μi:=μΓi,σ\mu_i:=\mu_{\Gamma_i,\sigma}. There is a decomposition

f=fstr(i)+fpsr(i)+fsml(i)f=f_{\mathrm{str}}^{(i)}+f_{\mathrm{psr}}^{(i)}+f_{\mathrm{sml}}^{(i)}

with the following properties:

  1. The functions fstr(i)f_{\mathrm{str}}^{(i)}, fpsr(i)f_{\mathrm{psr}}^{(i)}, and fsml(i)f_{\mathrm{sml}}^{(i)} are all 1-bounded;

2.

fstr(i)=Tviμi∗μi~f,0≤fstr(i)≤1,Efstr(i)=Ef;f_{\mathrm{str}}^{(i)}=T_{v_i}^{\mu_i*\widetilde{\mu_i}}f,\qquad 0\leq f_{\mathrm{str}}^{(i)}\leq1,\qquad \mathbb{E}f_{\mathrm{str}}^{(i)}=\mathbb{E}f;
  1. if b∈B(Γi,ρ1)b\in B(\Gamma_i,\rho_1), then

∥Tbvifstr(i)−fstr(i)∥L∞(G)≤ε;(14)\left\|T_b^{v_i}f_{\mathrm{str}}^{(i)}-f_{\mathrm{str}}^{(i)}\right\|_{L^\infty(G)}\leq\varepsilon; \tag*{(14)}

4.

sup⁡ξ∈Fp(Ex∈G∣fpsr(i)^(x;vi;ξ)∣2)1/2≤ρ2;(15)\sup_{\xi\in\mathbb{F}_p}\left(\mathbb{E}_{x\in G}\left|\widehat{f_{\mathrm{psr}}^{(i)}}(x;v_i;\xi)\right|^2\right)^{1/2}\leq\rho_2; \tag*{(15)}

(5)

∑i=1k∥fsm(i)∥L2(G)2≤ε2.(16)\sum_{i=1}^{k} \lVert f_{\mathrm{sm}}^{(i)} \rVert_{L^2(G)}^2 \le\varepsilon^2. \tag*{(16)}

In particular,

∥fpsr(i)∥L2(G,⟨vi⟩)2≤ρ2.\lVert f_{\mathrm{psr}}^{(i)} \rVert_{L^2(G,\langle v_i\rangle)}^2 \le\rho_2.

All bounds are independent of pp, ff, DD, and the particular nonzero vectors viv_i.

Proof of Proposition 3.3. We construct nested frequency sets and smooth ff in each direction using symmetrized Bohr kernels. A pigeonhole argument then selects two consecutive stages for which the sum of the squared L2(G)L^2(G) norms of the differences between the corresponding smoothings is small. These smoothings determine the three components.

We first recall that for every 1≤i≤k1 \le i \le k, fibrewise Parseval gives

∑ξ∈FpEx∈G∣f^(x;vi;ξ)∣2=∥f∥L2(G)2≤1.(17)\sum_{\xi\in\mathbb{F}_p} \mathbb{E}_{x \in G} \lvert\widehat{f}(x;v_i;\xi) \rvert^2 = \lVert f \rVert_{L^2(G)}^2 \le1. \tag*{(17)}

We construct the scales and frequency sets inductively, following [8]. Choose an integer L≥2L \ge2 sufficiently large in terms of kk and ε\varepsilon that 2k/(L−1)≤ε2/32k/(L-1) \le\varepsilon^2/3. Choose a sufficiently small absolute constant c0>0c_0 > 0. We construct positive scales

δ0:=min⁡{14,c0εk},δ0>δ1>⋯>δ3L>0,\delta_0 := \min\left\{\frac{1}{4},c_0\frac{\varepsilon}{\sqrt{k}}\right\}, \qquad\delta_0 > \delta_1 > \cdots> \delta_{3L} > 0,

together with nested frequency sets

∅=Γi,0⊆Γi,1⊆⋯⊆Γi,L⊆Fp∖{0}(1≤i≤k).\varnothing= \Gamma_{i,0} \subseteq\Gamma_{i,1} \subseteq\cdots\subseteq\Gamma_{i,L} \subseteq\mathbb{F}_p \setminus\{0\} \qquad(1 \le i \le k).

Each scale is chosen as a fixed positive function of the fixed inputs and the previously chosen scales and cardinality bounds.

Set R0:=0R_0 := 0. For 0≤j<L0 \le j < L, suppose that δ0,…,δ3j\delta_0,\ldots,\delta_{3j}, the sets Γi,j\Gamma_{i,j}, and a bound RjR_j, depending only on ε\varepsilon, kk, η1\eta_1, η2\eta_2, and jj, have been chosen, with

max⁡i∣Γi,j∣≤Rj.\max_i \lvert\Gamma_{i,j} \rvert\le R_j.

Define

μi,j:=μΓi,j,δ3j,\mu_{i,j} := \mu_{\Gamma_{i,j},\delta_{3j}},

and choose δ3j+1∈(0,δ3j)\delta_{3j+1} \in(0,\delta_{3j}) sufficiently small that Lemma 2.2 gives translation error at most ε\varepsilon on B(Γi,j,δ3j+1)B(\Gamma_{i,j},\delta_{3j+1}) whenever ∣Γi,j∣≤Rj\lvert\Gamma_{i,j}\rvert\le R_j, and so that

δ3j+1≤η1(Rj+ε−1)−1.\delta_{3j+1} \le\eta_1(R_j+\varepsilon^{-1})^{-1}.

Next choose positive scales δ3j+3<δ3j+2<δ3j+1\delta_{3j+3}<\delta_{3j+2}<\delta_{3j+1} sufficiently small that

δ3j+2≪min⁡{12,η2(δ3j+1−1)−1}.(18)\delta_{3j+2} \ll\min\left\{\frac{1}{2},\eta_2(\delta_{3j+1}^{-1})^{-1}\right\}. \tag*{(18)}

When j≥1j \ge1, impose in addition that

kδ3j+24≤ε23.k\delta_{3j+2}^4 \le\frac{\varepsilon^2}{3}.

We include frequencies at which a directional Fourier coefficient of ff has large L2(G)L^2(G) norm, to control the pseudorandom component. We also include frequencies at which an earlier kernel has a large Fourier coefficient, to control the difference between successive smoothings. We therefore set

Γi,j+1:=Γi,j∪{ξ≠0:(Ex∈G∣f^(x;vi;ξ)∣2)1/2≥δ3j+2}∪⋃ℓ=0j{ξ≠0:∣μ^i,ℓ(ξ)∣≥δ3j+2}.(19)\Gamma_{i,j+1}:=\Gamma_{i,j}\cup\left\{\xi\ne0:\left(\mathbb{E}_{x\in G}\left|\widehat{f}(x;v_i;\xi)\right|^2\right)^{1/2}\ge\delta_{3j+2}\right\}\cup\bigcup_{\ell=0}^{j}\left\{\xi\ne0:\left|\widehat{\mu}_{i,\ell}(\xi)\right|\ge\delta_{3j+2}\right\}. \tag*{(19)}

The first new set has cardinality at most δ3j+2−2\delta_{3j+2}^{-2} by (17). For each 0≤ℓ≤j0\le\ell\le j, Lemma 2.2 and Plancherel bound the number of frequencies satisfying ∣μ^i,ℓ(ξ)∣≥δ3j+2\left|\widehat{\mu}_{i,\ell}(\xi)\right|\ge\delta_{3j+2} by ORℓ,δ3ℓ(δ3j+2−2)O_{R_\ell,\delta_{3\ell}}(\delta_{3j+2}^{-2}). We may therefore choose Rj+1=Oε,k,η1,η2,j(1)R_{j+1}=O_{\varepsilon,k,\eta_1,\eta_2,j}(1) which bounds the cardinalities of all the sets Γi,j+1\Gamma_{i,j+1}, independently of ff, pp, DD, and the directions. Repeating this procedure constructs the entire hierarchy. Since there are finitely many stages and each scale depends only on the fixed inputs and the previously chosen bounds and scales, all cardinalities and inverse scales are bounded in terms of the fixed inputs alone.

We now choose the stage of the decomposition. The quantities

∑i=1k∑ξ∈Γi,jEx∈G∣f^(x;vi;ξ)∣2\sum_{i=1}^{k}\sum_{\xi\in\Gamma_{i,j}}\mathbb{E}_{x\in G}\left|\widehat{f}(x;v_i;\xi)\right|^2

are nondecreasing in jj and lie in [0,k][0,k]. Since the frequency sets are nested, each frequency is counted at most twice in the sum below. Hence,

∑j=0L−2∑i=1k∑ξ∈Γi,j+2∖Γi,jEx∈G∣f^(x;vi;ξ)∣2≤2k.\sum_{j=0}^{L-2}\sum_{i=1}^{k}\sum_{\xi\in\Gamma_{i,j+2}\setminus\Gamma_{i,j}}\mathbb{E}_{x\in G}\left|\widehat{f}(x;v_i;\xi)\right|^2\le2k.

The choice of LL therefore gives an index 0≤j≤L−20\le j\le L-2 such that

∑i=1k∑ξ∈Γi,j+2∖Γi,jEx∈G∣f^(x;vi;ξ)∣2≤ε23.(20)\sum_{i=1}^{k}\sum_{\xi\in\Gamma_{i,j+2}\setminus\Gamma_{i,j}}\mathbb{E}_{x\in G}\left|\widehat{f}(x;v_i;\xi)\right|^2\le\frac{\varepsilon^2}{3}. \tag*{(20)}

The use of Γi,j+2\Gamma_{i,j+2} allows us to control both kernels μi,j\mu_{i,j} and μi,j+1\mu_{i,j+1}: this set contains every nonzero frequency at which either kernel has a Fourier coefficient of magnitude at least δ3j+5\delta_{3j+5}. Fix this index jj. We define the three components using the smoothings at stages jj and j+1j+1:

fstr(i):=Tμi,j∗μ~i,jvif,f_{\mathrm{str}}^{(i)}:=T_{\mu_{i,j}*\widetilde{\mu}_{i,j}}^{v_i}f,
fsml(i):=Tμi,j+1∗μ~i,j+1vif−Tμi,j∗μ~i,jvif,f_{\mathrm{sml}}^{(i)}:=T_{\mu_{i,j+1}*\widetilde{\mu}_{i,j+1}}^{v_i}f-T_{\mu_{i,j}*\widetilde{\mu}_{i,j}}^{v_i}f,
fpsr(i):=f−Tμi,j+1∗μ~i,j+1vif.(21)f_{\mathrm{psr}}^{(i)}:=f-T_{\mu_{i,j+1}*\widetilde{\mu}_{i,j+1}}^{v_i}f. \tag*{(21)}

Since 0≤f≤10\le f\le1 and the kernels are nonnegative and have mean one, both smoothings in (21) take values in [0,1][0,1]. Hence the structured component takes values in [0,1][0,1], while the other two components take values in [−1,1][-1,1]. In particular, all three components are 1-bounded.

  1. (The small component.) By (9), we have

∥fsml(i)∥L2(G)2=∑ξ∈FpEx∈G∣f^(x;vi;ξ)∣2(∣μ^i,j+1(ξ)∣2−∣μ^i,j(ξ)∣2)2.\left\|f_{\mathrm{sml}}^{(i)}\right\|_{L^2(G)}^2=\sum_{\xi\in\mathbb{F}_p}\mathbb{E}_{x\in G}\left|\widehat{f}(x;v_i;\xi)\right|^2\left(\left|\widehat{\mu}_{i,j+1}(\xi)\right|^2-\left|\widehat{\mu}_{i,j}(\xi)\right|^2\right)^2.

Split the frequency sum into

{0}∪Γi,j,Γi,j+2∖Γi,j,Fp∖(Γi,j+2∪{0}).\{0\}\cup\Gamma_{i,j},\qquad\Gamma_{i,j+2}\setminus\Gamma_{i,j},\qquad\mathbb{F}_p\setminus\left(\Gamma_{i,j+2}\cup\{0\}\right).

On the first range, both multipliers are 1+O(δ3j)1+O(\delta_{3j}) by (6). On the second range, their squared difference is at most 11, since both multipliers lie in [0,1][0,1]. On the third range, the selection rule at stage j+1j+1 gives

∣μ^i,j(ξ)∣,∣μ^i,j+1(ξ)∣<δ3j+5,|\widehat{\mu}_{i,j}(\xi)|,|\widehat{\mu}_{i,j+1}(\xi)|<\delta_{3j+5},

so the squared difference of the multipliers is at most δ3j+54\delta_{3j+5}^{4}. Summing over ii and using (17) and (20), we obtain

∑i=1k∥f^sml(i)∥L2(G)2≤Ckδ3j2+ε23+kδ3j+54≤ε2.\sum_{i=1}^{k}\left\|\widehat{f}_{\mathrm{sml}}^{(i)}\right\|_{L^{2}(G)}^{2}\le Ck\delta_{3j}^{2}+\frac{\varepsilon^{2}}{3}+k\delta_{3j+5}^{4}\le\varepsilon^{2}.

where CC is an absolute constant. The last inequality follows from δ3j≤δ0\delta_{3j}\le\delta_{0}, the choice of c0c_{0}, and the condition imposed at stage j+1j+1. This proves (16).

  1. (The pseudorandom component.) By (8), we have

f^psr(i)(x;vi;ξ)=(1−∣μ^i,j+1(ξ)∣2)f^(x;vi;ξ).\widehat{f}_{\mathrm{psr}}^{(i)}(x;v_i;\xi)=\left(1-|\widehat{\mu}_{i,j+1}(\xi)|^{2}\right)\widehat{f}(x;v_i;\xi).

If ξ∉Γi,j+1∪{0}\xi\notin\Gamma_{i,j+1}\cup\{0\}, then we have

(Ex∈G∣f^(x;vi;ξ)∣2)1/2<δ3j+2\left(\mathbb{E}_{x\in G}|\widehat{f}(x;v_i;\xi)|^{2}\right)^{1/2}<\delta_{3j+2}

by (19). If ξ∈Γi,j+1∪{0}\xi\in\Gamma_{i,j+1}\cup\{0\}, then (6) gives

1−∣μ^i,j+1(ξ)∣2≪δ3j+3.1-|\widehat{\mu}_{i,j+1}(\xi)|^{2}\ll\delta_{3j+3}.

Since δ3j+3<δ3j+2\delta_{3j+3}<\delta_{3j+2}, (15) holds with

ρ2:=Cδ3j+2,\rho_{2}:=C\delta_{3j+2},

where CC is a sufficiently large absolute constant. This constant is absorbed by the choice of scales in (18).

  1. (The structured component.) Set

Γi:=Γi,j,σ:=δ3j,ρ1:=δ3j+1,μi:=μi,j.\Gamma_i:=\Gamma_{i,j},\qquad\sigma:=\delta_{3j},\qquad\rho_1:=\delta_{3j+1},\qquad\mu_i:=\mu_{i,j}.

The choice of δ3j+1\delta_{3j+1} and (10) give (14). Moreover, since max⁡i∣Γi∣≤Rj\max_i|\Gamma_i|\le R_j and η1\eta_1 is nondecreasing, the choice of δ3j+1\delta_{3j+1} gives

ρ1=δ3j+1≤1η1(Rj+ε−1)≤1η1(max⁡i∣Γi∣+ε−1).\rho_1=\delta_{3j+1}\le\frac{1}{\eta_1(R_j+\varepsilon^{-1})}\le\frac{1}{\eta_1(\max_i|\Gamma_i|+\varepsilon^{-1})}.

The choice of ρ2\rho_2, together with (18), gives ρ2≤η2(ρ1−1)−1\rho_2\le\eta_2(\rho_1^{-1})^{-1}. Thus, both scale inequalities in (13) hold.

Finally, since each kernel μi∗μ~i\mu_i*\widetilde{\mu}_i has mean one, we have Efstr(i)=Ef\mathbb{E}f_{\mathrm{str}}^{(i)}=\mathbb{E}f. The bounds on ∣Γi∣|\Gamma_i|, σ−1\sigma^{-1}, ρ1−1\rho_1^{-1}, and ρ2−1\rho_2^{-1} follow from the construction. Applying Lemma 2.1 to the 1-bounded function fpsr(i)f_{\mathrm{psr}}^{(i)} and using (15) gives the claimed box-norm estimate. □\square

Remark 3.4. In Tao’s one-dimensional argument, the structured term is smoothed with a cutoff μ\mu, whereas the iith directional component here uses μi∗μ~i\mu_i*\widetilde{\mu}_i. Replacing μ^i\widehat{\mu}_i by ∣μ^i∣2|\widehat{\mu}_i|^2 changes the regularity estimates only by fixed absolute factors. The symmetrized kernel is needed when estimating the structured contribution, because Tμi∗μ~ivi=TμiviTμ~iviT_{\mu_i*\widetilde{\mu}_i}^{v_i}=T_{\mu_i}^{v_i}T_{\widetilde{\mu}_i}^{v_i} allows Lemma 3.6 to handle structured components from different directions.

Polynomial images in prescribed Bohr sets

We want to find a common parameter dd for which every structured component is almost invariant under its prescribed shift. The following lemma counts the parameters satisfying Pi(d)∈B(Γi,ri)P_i(d) \in B(\Gamma_i,r_i) for all ii simultaneously.

Lemma 3.5. Let k∈Nk \in\mathbb{N} and P={P1,…,Pk∈Z[t]}\mathcal{P} = \{P_1,\ldots,P_k \in\mathbb{Z}[t]\} be a collection of linearly independent polynomials with zero constant terms. For 1≤i≤k1 \le i \le k, let Γi⊆Fp∖{0}\Gamma_i \subseteq\mathbb{F}_p \setminus\{0\} and 0<ri≤1/20 < r_i \le1/2. Then we have

Ed∈Fp∏i=1k1B(Γi,ri)(Pi(d))=∏i=1k∣B(Γi,ri)∣p+OP,{∣Γi∣}i=1k(p−1/2(log⁡p)∑i=1k∣Γi∣).\mathbb{E}_{d\in\mathbb{F}_p}\prod_{i=1}^{k}1_{B(\Gamma_i,r_i)}(P_i(d))=\prod_{i=1}^{k}\frac{|B(\Gamma_i,r_i)|}{p}+O_{\mathcal{P},\{|\Gamma_i|\}_{i=1}^{k}}\left(p^{-1/2}(\log p)^{\sum_{i=1}^{k}|\Gamma_i|}\right).

Proof of Lemma 3.5. The proof is analogous to that of [1 Lemma 3.4] and is omitted. □\square

A convolution inequality

Almost invariance allows us to remove the polynomial shifts from the structured contribution at the cost of a small error. The remaining factors are different directional smoothings of the same function. The following convolution inequality gives the required lower bound for the average of their product with the original function.

Lemma 3.6. Let m≥0m \ge0, let v1,…,vm∈FpDv_1,\ldots,v_m \in\mathbb{F}_p^D, and let f:FpD→[0,∞)f:\mathbb{F}_p^D \to[0,\infty) satisfy Ef=α\mathbb{E}f=\alpha. For every 1≤i≤m1 \le i \le m, let μi:Fp→[0,∞)\mu_i:\mathbb{F}_p \to[0,\infty) satisfy Eμi=1\mathbb{E}\mu_i=1. Then we have

Ex∈FpDf(x)∏i=1mTμi∗μi~vif(x)≥αm+1.(22)\mathbb{E}_{x\in\mathbb{F}_p^D}f(x)\prod_{i=1}^{m}T_{\mu_i*\widetilde{\mu_i}}^{v_i}f(x)\ge\alpha^{m+1}. \tag*{(22)}

Proof of Lemma 3.6. The cases m=0m=0 and α=0\alpha=0 are immediate. By homogeneity, we may therefore assume that m≥1m\ge1 and Ef=1\mathbb{E}f=1. We first treat the case in which ff is strictly positive.

We estimate each factor separately after taking logarithms. Fix 1≤i≤m1\le i\le m. Since μi\mu_i is nonnegative and has mean one, Jensen’s inequality gives

log⁡(Tμi∗μi~vif(x))=log⁡(TμiviTμi~vif(x))≥[Tμivi(log⁡(Tμi~vif))](x).\log\left(T_{\mu_i*\widetilde{\mu_i}}^{v_i}f(x)\right)=\log\left(T_{\mu_i}^{v_i}T_{\widetilde{\mu_i}}^{v_i}f(x)\right)\ge\left[T_{\mu_i}^{v_i}\left(\log(T_{\widetilde{\mu_i}}^{v_i}f)\right)\right](x).

Multiplying by ff and averaging over xx, we obtain

Eflog⁡(Tμi∗μi~vif)≥EfTμivi(log⁡(Tμi~vif))=E(Tμi~vif)log⁡(Tμi~vif),\mathbb{E}f\log\left(T_{\mu_i*\widetilde{\mu_i}}^{v_i}f\right)\ge\mathbb{E}fT_{\mu_i}^{v_i}\left(\log(T_{\widetilde{\mu_i}}^{v_i}f)\right) =\mathbb{E}(T_{\widetilde{\mu_i}}^{v_i}f)\log(T_{\widetilde{\mu_i}}^{v_i}f),

where the last equality follows by a change of variables. The last expression is nonnegative by Jensen’s inequality for the convex function u↦ulog⁡uu\mapsto u\log u, since ETμi~vif=1\mathbb{E}T_{\widetilde{\mu_i}}^{v_i}f=1.

We now combine these bounds. Since ff is nonnegative and Ef=1\mathbb{E}f=1, Jensen’s inequality with weights ff gives

log⁡(Ef∏i=1mTμi∗μi~vif)≥Eflog⁡(∏i=1mTμi∗μi~vif)=∑i=1mEflog⁡(Tμi∗μi~vif)≥0.\log\left(\mathbb{E}f\prod_{i=1}^{m}T_{\mu_i*\widetilde{\mu_i}}^{v_i}f\right)\ge\mathbb{E}f\log\left(\prod_{i=1}^{m}T_{\mu_i*\widetilde{\mu_i}}^{v_i}f\right) =\sum_{i=1}^{m}\mathbb{E}f\log\left(T_{\mu_i*\widetilde{\mu_i}}^{v_i}f\right) \ge0.

Exponentiating proves the result when ff is strictly positive and has mean one. For a general nonnegative function of mean one, apply the preceding argument to (f+δ)/(1+δ)(f+\delta)/(1+\delta) and let δ↓0\delta\downarrow0. This proves (22). □\square

Remark 3.7. The symmetrization is essential in general: inequality (22) need not hold with TμiviT_{\mu_i}^{v_i} in place of Tμi∗μi~viT_{\mu_i*\widetilde{\mu_i}}^{v_i}. Indeed, take D=m=1D=m=1, v1=1v_1=1, and set

f=1{0},μ=p2(1{1}+1{−1}).f=1_{\{0\}},\qquad\mu=\frac{p}{2}\left(1_{\{1\}}+1_{\{-1\}}\right).

Then μ\mu is nonnegative and has mean one, but

Ex∈Fpf(x)Tμ1f(x)=0<1p2=(Ef)2.\mathbb{E}_{x\in\mathbb{F}_p} f(x)T_{\mu}^{1}f(x)=0<\frac{1}{p^2}=(\mathbb{E}f)^2.

Completion of the proof

We now combine the preceding results to obtain the required lower bound for every choice of ω\omega with the same nonzero parameter dd. We first find many parameters satisfying the Bohr conditions and then remove those belonging to the exceptional sets.

Proof of Theorem 1.1. Recall that α=Ex∈G1A(x)\alpha=\mathbb{E}_{x\in G}1_A(x). We may assume that 0<ε<10<\varepsilon<1. Fix

0<ϑ≪kε.0<\vartheta\ll_k\varepsilon.

Choose nondecreasing growth functions η1,η2\eta_1,\eta_2, depending only on ϑ\vartheta, P\mathcal{P}, and kk, such that η1(t)≥t\eta_1(t)\ge t and η2(t)−1≪P,kϑ2−kt\eta_2(t)^{-1}\ll_{\mathcal{P},k}\vartheta^{2-kt}, with the implicit constant sufficiently small.

Apply Proposition 3.3 to 1A1_A, with parameter ϑ\vartheta and growth functions η1,η2\eta_1,\eta_2. We obtain Γi\Gamma_i, parameters ρ1,ρ2\rho_1,\rho_2, normalized nonnegative kernels μi\mu_i, and decompositions in the prescribed directions

1A=fstr(i)+fpsr(i)+fsml(i)(1≤i≤k).1_A=f_{\mathrm{str}}^{(i)}+f_{\mathrm{psr}}^{(i)}+f_{\mathrm{sml}}^{(i)} \qquad(1\le i\le k).

In particular, we have

fstr(i)=Tμi∗μ~iνi1A,0≤fstr(i)≤1,f_{\mathrm{str}}^{(i)}=T_{\mu_i*\widetilde{\mu}_i}^{\nu_i}1_A,\qquad0\le f_{\mathrm{str}}^{(i)}\le1,
∥fpsr(i)∥G,⟨vi⟩≤ρ2,∑i=1k∥fsml(i)∥L2(G)2≤ϑ2,∥Tbνifstr(i)−fstr(i)∥L∞(G)≤ϑ(b∈B(Γi,ρ1)).\lVert f_{\mathrm{psr}}^{(i)}\rVert_{G,\langle v_i\rangle}\le\rho_2,\qquad \sum_{i=1}^{k}\lVert f_{\mathrm{sml}}^{(i)}\rVert_{L^2(G)}^2\le\vartheta^2,\qquad \lVert T_b^{\nu_i}f_{\mathrm{str}}^{(i)}-f_{\mathrm{str}}^{(i)}\rVert_{L^\infty(G)}\le\vartheta \quad(b\in B(\Gamma_i,\rho_1)).

By (13) and the choice of η1\eta_1, max⁡i∣Γi∣≤ρ1−1\max_i|\Gamma_i|\le\rho_1^{-1}, so ∑i=1k∣Γi∣≤kρ1−1\sum_{i=1}^{k}|\Gamma_i|\le k\rho_1^{-1}. Since 0<ρ1<10<\rho_1<1, the choice of η2\eta_2 therefore gives

ρ2≤η2(ρ1−1)−1≪P,kϑ2ρ1kρ1−1≤ϑ2ρ1∑i=1k∣Γi∣.(23)\rho_2\le\eta_2(\rho_1^{-1})^{-1}\ll_{\mathcal{P},k}\vartheta^2\rho_1^{k\rho_1^{-1}}\le\vartheta^2\rho_1^{\sum_{i=1}^{k}|\Gamma_i|}. \tag*{(23)}

Proposition 3.3 also gives bounds for max⁡i∣Γi∣\max_i|\Gamma_i| and ρ1−1\rho_1^{-1} which depend only on ϑ\vartheta, P\mathcal{P}, and kk, and are independent of pp, AA, DD, and the directions.

We first collect the parameters for which every structured component is almost invariant under its prescribed shift. Set

B∗:={d∈Fp:Pi(d)∈B(Γi,ρ1) for every 1≤i≤k}.B^*:=\{d\in\mathbb{F}_p:P_i(d)\in B(\Gamma_i,\rho_1)\text{ for every }1\le i\le k\}.

By (3), we have

∏i=1k∣B(Γi,ρ1)∣p≥ρ1∑i=1k∣Γi∣.\prod_{i=1}^{k}\frac{|B(\Gamma_i,\rho_1)|}{p}\ge\rho_1^{\sum_{i=1}^{k}|\Gamma_i|}.

Since the ranks of B(Γi,ρ1)B(\Gamma_i,\rho_1) and ρ1−1\rho_1^{-1} are bounded independently of pp, Lemma 3.5 implies that, for all sufficiently large primes, we have

∣B∗∣p≥12ρ1∑i=1k∣Γi∣.(24)\frac{|B^*|}{p}\ge\frac{1}{2}\rho_1^{\sum_{i=1}^{k}|\Gamma_i|}. \tag*{(24)}

We next remove parameters for which a term containing a pseudorandom component is too large. Terms containing a small component will be bounded uniformly in dd using their L2(G)L^2(G) norms. We therefore exclude parameters only for terms containing a pseudorandom component and no small component. Let ∅≠S⊆{1,…,k}\varnothing\ne S\subseteq\{1,\ldots,k\}. For each i∈Si\in S, let gig_i be either fstr(i)f_{\mathrm{str}}^{(i)} or fpsr(i)f_{\mathrm{psr}}^{(i)}, and suppose that gj=fpsr(j)g_j=f_{\mathrm{psr}}^{(j)} for some j∈Sj\in S. Applying Lemma 3.2 to gjg_j gives

Ed∈Fp∣Ex∈G1A(x)∏i∈Sgi(x+Pi(d)vi)∣2≪Pρ2+p−c.\mathbb{E}_{d\in\mathbb{F}_p}\left|\mathbb{E}_{x\in G}1_A(x)\prod_{i\in S}g_i\bigl(x+P_i(d)v_i\bigr)\right|^2\ll_{\mathcal{P}}\rho_2+p^{-c}.

Markov’s inequality yields

1p∣{d:∣Ex∈G1A(x)∏i∈Sgi(x+Pi(d)vi)∣>ϑ}∣≪Pϑ−2(ρ2+p−c).\frac{1}{p}\left|\left\{d:\left|\mathbb{E}_{x\in G}1_A(x)\prod_{i\in S}g_i\left(x+P_i(d)v_i\right)\right|>\vartheta\right\}\right|\ll_P\vartheta^{-2}\left(\rho_2+p^{-c}\right).

There are Ok(1)O_k(1) choices of SS and of the functions gig_i. Let EE be the set of parameters for which at least one of these averages has modulus greater than ϑ\vartheta. By the union bound, ∣E∣/p≪P,kϑ−2(ρ2+p−c)|E|/p\ll_{P,k}\vartheta^{-2}(\rho_2+p^{-c}). Using (23) to control the first term and taking pp sufficiently large to control the second, we obtain

∣E∣p≤14ρ1∑i=1k∣Γi∣.(25)\frac{|E|}{p}\leq\frac{1}{4}\rho_1\sum_{i=1}^{k}|\Gamma_i|. \tag*{(25)}

Combining (24) and (3.15 gives

∣B∗∖E∣p≥14ρ1∑i=1k∣Γi∣.\frac{|B^*\setminus E|}{p}\geq\frac{1}{4}\rho_1\sum_{i=1}^{k}|\Gamma_i|.

The right-hand side is bounded below by a positive constant depending only on ϑ\vartheta, P\mathcal{P}, and kk. Hence, for all sufficiently large primes, ∣B∗∖E∣>1|B^*\setminus E|>1, and we may choose

d∈(B∗∖E)∖{0}.(26)d\in(B^*\setminus E)\setminus\{0\}. \tag*{(26)}

Having fixed dd, we verify the required bound for every choice of ω\omega. Fix ω∈{0,1}k\omega\in\{0,1\}^k and put S:={i:ωi=1}S:=\{i:\omega_i=1\}. If S=∅S=\varnothing, then

I(P,V)ω(1A,…,1A)(d)=E1A=α.I_{(\mathcal{P},\mathcal{V})_\omega}(1_A,\ldots,1_A)(d)=\mathbb{E}1_A=\alpha.

This gives the required bound when S=∅S=\varnothing. Suppose now that S≠∅S\ne\varnothing. Leave 1A(x)1_A(x) undecomposed and apply the iith directional decomposition to 1A(x+Pi(d)vi)1_A(x+P_i(d)v_i):

I(P,V)ω(1A,…,1A)(d)=Ex∈G1A(x)∏i∈S(fstr(i)+fpsr(i)+fsml(i))(x+Pi(d)vi).(27)I_{(\mathcal{P},\mathcal{V})_\omega}(1_A,\ldots,1_A)(d)=\mathbb{E}_{x\in G}1_A(x)\prod_{i\in S}\left(f_{\mathrm{str}}^{(i)}+f_{\mathrm{psr}}^{(i)}+f_{\mathrm{sml}}^{(i)}\right)\left(x+P_i(d)v_i\right). \tag*{(27)}

We first consider the term containing only structured components. Since d∈B∗d\in B^*, we have Pi(d)∈B(Γi,ρ1)P_i(d)\in B(\Gamma_i,\rho_1) for every ii. Telescoping the product and using 1-boundedness gives

∣Ex∈G1A(x)∏i∈Sfstr(i)(x+Pi(d)vi)−Ex∈G1A(x)∏i∈Sfstr(i)(x)∣≤∑i∈S∥τPi(d)vifstr(i)−fstr(i)∥L∞(G)≤∣S∣ϑ≤kϑ.\left|\mathbb{E}_{x\in G}1_A(x)\prod_{i\in S}f_{\mathrm{str}}^{(i)}(x+P_i(d)v_i)-\mathbb{E}_{x\in G}1_A(x)\prod_{i\in S}f_{\mathrm{str}}^{(i)}(x)\right| \leq\sum_{i\in S}\|\tau_{P_i(d)}^{v_i}f_{\mathrm{str}}^{(i)}-f_{\mathrm{str}}^{(i)}\|_{L^\infty(G)} \leq|S|\vartheta\leq k\vartheta.

Applying Lemma 3.6 to f=1Af=1_A and the kernels μi\mu_i, i∈Si\in S, gives

Ex∈G1A(x)∏i∈Sfstr(i)(x)≥α∣S∣+1.\mathbb{E}_{x\in G}1_A(x)\prod_{i\in S}f_{\mathrm{str}}^{(i)}(x)\geq\alpha^{|S|+1}.

Combining these two estimates gives

Ex∈G1A(x)∏i∈Sfstr(i)(x+Pi(d)vi)≥α∣S∣+1−kϑ.(28)\mathbb{E}_{x\in G}1_A(x)\prod_{i\in S}f_{\mathrm{str}}^{(i)}(x+P_i(d)v_i)\geq\alpha^{|S|+1}-k\vartheta. \tag*{(28)}

Every other term in (27) contains a small or a pseudorandom component. If a term contains fsml(j)f_{\mathrm{sml}}^{(j)}, then, by Cauchy–Schwarz and 1-boundedness of all other factors, its modulus is at most

∥fsml(j)∥L1(G)≤∥fsml(j)∥L2(G)≤ϑ.\|f_{\mathrm{sml}}^{(j)}\|_{L^1(G)}\leq\|f_{\mathrm{sml}}^{(j)}\|_{L^2(G)}\leq\vartheta.

If it contains no small component but contains a pseudorandom component, then it is one of the terms used to define EE, and its modulus is at most ϑ\vartheta because d∉Ed\notin E. There are at most 3∣S∣−1≤3k3^{|S|}-1\leq3^k terms other than the term containing only structured components. Combining these estimates with (28), we obtain

I(P,V)ω(1A,…,1A)(d)≥α∣S∣+1−Ok(ϑ)≥α∣S∣+1−ε.I_{(\mathcal{P},\mathcal{V})_\omega}(1_A,\ldots,1_A)(d)\geq\alpha^{|S|+1}-O_k(\vartheta)\geq\alpha^{|S|+1}-\varepsilon.

The parameter dd in (26) is independent of ω\omega, so this holds simultaneously for all ω∈{0,1}k\omega\in\{0,1\}^k. ∎

Acknowledgements

The author thanks David Conlon and Dingding Dong for helpful discussions on popular polynomial differences that motivated the present work.

Statement of AI use

The main ideas of this paper, including the directional arithmetic regularity decomposition in Proposition 3.3, were developed by the author. After the initial argument was completed, GPT 5.6 Sol pointed out the need to symmetrize the smoothing kernels defining the structured components. This modification enables the lower bound in Lemma 3.6; see also Remarks 3.4 and 3.7. GPT-6 Astra was subsequently used to assist with proofreading, polish the writing, and improve the exposition.

Department of Mathematics, California Institute of Technology, Pasadena, CA 91125, USA

Email address: [email protected]

References

  1. [1]David Conlon, Dingding Dong, and Guo-Dong Hong. Simultaneous popular polynomial differences over finite fields. Preprint available at arXiv:2607.10051 [math.NT].arxiv.org/abs/2607.10051
  2. [2]Nikos Frantzikinakis and Bryna Kra. Ergodic averages for independent polynomials and applications. J. London Math. Soc., 74(1):131–142, 2006.DOI
  3. [3]Guo-Dong Hong. On weighted multilinear polynomial averages in finite fields. Preprint available at arXiv:2507.14414 [math.NT].arxiv.org/abs/2507.14414
  4. [4]Borys Kuca. Multidimensional polynomial patterns over finite fields: bounds, counting estimates and Gowers norm control. Adv. Math., 448:109700, 2024.arxiv.org/abs/2304.10793
  5. [5]Neil Lyall and Ákos Magyar. Simultaneous polynomial recurrence. Bull. Lond. Math. Soc., 43(4):765–785, 2011.DOI
  6. [6]Sarah Peluse. On the polynomial Szemerédi theorem in finite fields. Duke Math. J., 168(5):749–774, 2019.DOI
  7. [7]Xuancheng Shao and Weiwen Zhang. Simultaneous popular differences for polynomial and linear progressions in finite fields. Preprint available at arXiv:2610.28226 [math.NT].arxiv.org/abs/2610.28226
  8. [8]Terence Tao. A proof of Roth’s theorem. Blog post available at https://terrytao.wordpress.com/2014/04/24/a-proof-of-roths-theorem/, 2014.

Paper details

vlnc:W_jpgsp6rt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents