We extend the simultaneous popular polynomial difference theorem of Conlon, Dong, and the author to polynomial shifts along prescribed directions in FpD. Let P1,…,Pk∈Z[t] be linearly independent polynomials with zero constant terms, and let v1,…,vk∈ZD∖{0}. We show that, for every ε>0, every sufficiently large prime p, and every A⊆FpD of density α, there exists a nonzero d∈Fp such that Ex∈FpD1A(x)∏1≤i≤kωi=11A(x+Pi(d)vi)≥α1+∑i=1kωi−ε simultaneously for every ω∈{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, with polynomial shifts along prescribed directions.
To state our result, fix D,k∈N, and set G:=FpD, where p is an odd prime. Let V={v1,…,vk} be a collection of nonzero vectors in ZD, and let P={P1,…,Pk} be a collection of linearly independent polynomials in Z[t] with zero constant terms. For d∈Fp and functions f0,…,fk:G→C, define
When all the functions are 1A, this average gives the normalized count of polynomial configurations x,x+P1(d)v1,…,x+Pk(d)vk contained in A, with the shifts determined by the common parameter d. We consider subconfigurations obtained by keeping x and any subset of the shifted points. Given ω=(ω1,…,ωk)∈{0,1}k, let (P,V)ω denote the subcollection of pairs (Pi,vi) for which ωi=1, and set
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∈N. Let V={v1,…,vk} be a collection of nonzero vectors in ZD, and let P={P1,…,Pk}⊂Z[t] be a collection of linearly independent polynomials with zero constant terms. For every ε>0, there exists p0=p0(ε,P,V) such that for every prime p≥p0 and every A⊆FpD of density α, there exists d∈Fp∖{0} such that
I(P,V)ω(1A,…,1A)(d)≥α1+∑i=1kωi−ε
holds simultaneously for every ω=(ω1,…,ωk)∈{0,1}k.
The same nonzero parameter d therefore works for every choice of ω. In particular, it gives the lower bound αk+1−ε for the full configuration and the lower bound α2−ε for each of the k two-term configurations. No linear-independence assumption is imposed on the vectors v1,…,vk.
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. More recently, Shao and Zhang [7] obtained simultaneous popular difference results over Fp 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.
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 U2-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⟩; 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 X,
We omit an averaging domain when it is evident from the variables. For example, Ef denotes the normalized mean of f on its domain. For a set E, the function 1E denotes its indicator. The density of a subset A⊆X is Ex∈X1A(x)=∣A∣/∣X∣. We also put ∥f∥ℓ∞(X):=maxx∈X∣f(x)∣. We say that f is 1-bounded if ∣f(x)∣≤1 for every x.
Throughout the paper, p denotes an odd prime. Let D∈N, and denote G=FpD. Linear independence of polynomials in Z[t] is understood over Q. Integer vectors and polynomials are reduced modulo p when they are used over Fp. For v∈G, write ⟨v⟩:={tv:t∈Fp}. For a∈Fp, write ep(a):=exp(2πia/p). For f:Fp→C, our normalized Fourier transform is
f(ξ):=Ex∈Fpf(x)ep(−ξx).
For x,h∈G, define the multiplicative derivative
△hf(x):=f(x)f(x+h).
We write △h1,…,hsf:=△h1⋯△hsf for the iterated multiplicative derivative, with the operators applied by composition.
We use Landau’s notation O, o, and Vinogradov’s notation ≪, ≫. 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∈G, ξ∈Fp, and f:G→C, define the Fourier coefficient of f along v by
f(x;v;ξ):=En∈Fpf(x+nv)ep(−nξ).(1)
If x′=x+sv, then
f(x′;v;ξ)=ep(sξ)f(x;v;ξ),
so the modulus of (1) is constant on cosets of ⟨v⟩.
For subspaces H1,…,Hs≤G, define the Gowers box norm along H1,…,Hs by
The usual Gowers U2-norm is the special case ∥f∥U2(G):=∥f∥G,G. The notation ∥f∥G,⟨v⟩ denotes the box norm with H1=G and H2=⟨v⟩. 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} and f:G→C. If f is 1-bounded, then
∥f∥G,⟨v⟩4≤ξ∈FpsupEx∈G∣f(x;v;ξ)∣2.(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/Z, write ∥t∥T for the distance to the nearest integer. If Γ⊆Fp∖{0} and 0<ρ≤1/2, define
B(Γ,ρ):={b∈Fp:pξbT≤ρ for every ξ∈Γ}.
Its rank is ∣Γ∣, and we adopt the convention B(∅,ρ)=Fp.
We use the following standard estimate, which can be found in [TV06, Lemma 4.20]:
p∣B(Γ,ρ)∣≥ρ∣Γ∣.(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, define the normalized convolution and involution by
For a real number u, put (u)+:=max{u,0}. Let ∅=Γ⊆Fp∖{0} and 0<ρ≤1/4. Following Tao’s smoothing approach [8], we use the normalized Lipschitz cutoff associated with B(Γ,ρ):
For Γ=∅, we set μ∅,ρ=1. For fixed Γ and ρ, write μ:=μΓ,ρ. The structured components will be formed using the symmetrized kernel μ∗μ. The following lemma records the Fourier and translation estimates needed for these averages.
Lemma 2.2. Let μ:=μΓ,ρ for some Γ⊆Fp∖{0} and 0<ρ≤1/4. Then, the functions μ and μ∗μ are nonnegative and have mean one. Moreover, we have
∥μ∥L2(Fp)≪∣Γ∣,ρ1,(4)
μ∗μ(ξ)=∣μ(ξ)∣2,(5)
∣1−μ(ξ)∣≪ρ(ξ∈Γ∪{0}).(6)
For 0<σ≤1/2, if b∈B(Γ,σ), then
∥τb(μ∗μ)−μ∗μ∥L1(Fp)≪∣Γ∣,ρσ.(7)
Proof of Lemma 2.2. If Γ=∅, then μ=1 and all assertions are immediate. Now, we suppose that Γ=∅. The cutoff before normalization takes values in [0,1] and is at least 1/2 on B(Γ,ρ/2). By (3), its normalized mean is bounded below by a positive constant depending only on ∣Γ∣ and ρ, and this gives (4).
By construction, μ is nonnegative and has mean one, and the same holds for μ∗μ. The convolution identity for the Fourier transform gives (5).
For ξ∈Γ∪{0} and t∈suppμ, we have ∥ξt/p∥T≤ρ, and hence ∣1−ep(−ξt)∣≤2πρ. Since μ is nonnegative and has mean one, it follows that
Finally, let b∈B(Γ,σ). The triangle inequality on T implies that the unnormalized cutoff changes by at most σ/ρ under translation by b. Using the lower bound for the normalizing denominator, we obtain
∥τbμ−μ∥L1(Fp)≪∣Γ∣,ρσ.
Since ∥μ∥L1(Fp)=1, Young’s inequality therefore gives
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}, 0<ρ,ρ′≤1/4, b,ξ∈Fp, and let v∈G∖{0}. Write
Indeed, a change of variables gives τbvTμ∗μvh=Tτb(μ∗μ)vh, and the bound follows from ∣h∣≤1.
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.
A generalized von Neumann estimate
We first establish an L2(Fp) 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∈N. Let W={w1,…,wr∈ZD∖{0}}, and R={P1,…,Pr∈Z[t]} be a collection of linearly independent polynomials with zero constant terms. There exist p1=p1(W,R), and c=c(R)>0 such that for every prime p≥p1, we have
The following consequence applies uniformly to every subconfiguration. The bound can be obtained from any one of the functions gi, i∈S.
Lemma 3.2. Let D,k∈N, and denote G=FpD. Let V={v1,…,vk∈ZD∖{0}}, and P={P1,…,Pk∈Z[t]} be a collection of linearly independent polynomials with zero constant terms. There exist p0=p0(V,P), and c=c(P)>0 such that for every prime p≥p0, we have
Since there are only finitely many subfamilies, we fix one exponent c>0, one implicit constant, and one prime threshold that work for all of them. For each fixed h, apply Theorem 3.1 to the subfamily indexed by S. Its error is uniform in h, since every multiplicative derivative remains 1-bounded. We obtain
Taking the minimum over i 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), uniformly in frequency, and the remaining components have small L2(G) norms.
Proposition 3.3. Let D,k∈N, and denote G=FpD. Let 0<ε<1 and let η1,η2:[1,∞)→[2,∞) be nondecreasing growth functions. Let p be prime, let v1,…,vk∈G∖{0}, and let f:G→[0,1]. Then there are finite sets Γi⊆Fp∖{0}, a radius σ∈(0,1/4], and parameters ρ1,ρ2∈(0,1/2] such that
The parameters ρ1,ρ2 and the radius σ are bounded away from zero in terms of ε, k, η1, and η2 alone. For every 1≤i≤k, put μi:=μΓi,σ. There is a decomposition
f=fstr(i)+fpsr(i)+fsml(i)
with the following properties:
The functions fstr(i), fpsr(i), and fsml(i) are all 1-bounded;
All bounds are independent of p, f, D, and the particular nonzero vectors vi.
Proof of Proposition 3.3. We construct nested frequency sets and smooth f in each direction using symmetrized Bohr kernels. A pigeonhole argument then selects two consecutive stages for which the sum of the squared L2(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≤k, fibrewise Parseval gives
ξ∈Fp∑Ex∈G∣f(x;vi;ξ)∣2=∥f∥L2(G)2≤1.(17)
We construct the scales and frequency sets inductively, following [8]. Choose an integer L≥2 sufficiently large in terms of k and ε that 2k/(L−1)≤ε2/3. Choose a sufficiently small absolute constant c0>0. We construct positive scales
δ0:=min{41,c0kε},δ0>δ1>⋯>δ3L>0,
together with nested frequency sets
∅=Γi,0⊆Γi,1⊆⋯⊆Γi,L⊆Fp∖{0}(1≤i≤k).
Each scale is chosen as a fixed positive function of the fixed inputs and the previously chosen scales and cardinality bounds.
Set R0:=0. For 0≤j<L, suppose that δ0,…,δ3j, the sets Γi,j, and a bound Rj, depending only on ε, k, η1, η2, and j, have been chosen, with
imax∣Γi,j∣≤Rj.
Define
μi,j:=μΓi,j,δ3j,
and choose δ3j+1∈(0,δ3j) sufficiently small that Lemma 2.2 gives translation error at most ε on B(Γi,j,δ3j+1) whenever ∣Γi,j∣≤Rj, and so that
δ3j+1≤η1(Rj+ε−1)−1.
Next choose positive scales δ3j+3<δ3j+2<δ3j+1 sufficiently small that
δ3j+2≪min{21,η2(δ3j+1−1)−1}.(18)
When j≥1, impose in addition that
kδ3j+24≤3ε2.
We include frequencies at which a directional Fourier coefficient of f has large L2(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
The first new set has cardinality at most δ3j+2−2 by (17). For each 0≤ℓ≤j, Lemma 2.2 and Plancherel bound the number of frequencies satisfying ∣μi,ℓ(ξ)∣≥δ3j+2 by ORℓ,δ3ℓ(δ3j+2−2). We may therefore choose Rj+1=Oε,k,η1,η2,j(1) which bounds the cardinalities of all the sets Γi,j+1, independently of f, p, D, 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=1∑kξ∈Γi,j∑Ex∈Gf(x;vi;ξ)2
are nondecreasing in j and lie in [0,k]. Since the frequency sets are nested, each frequency is counted at most twice in the sum below. Hence,
The use of Γi,j+2 allows us to control both kernels μi,j and μi,j+1: this set contains every nonzero frequency at which either kernel has a Fourier coefficient of magnitude at least δ3j+5. Fix this index j. We define the three components using the smoothings at stages j and j+1:
Since 0≤f≤1 and the kernels are nonnegative and have mean one, both smoothings in (21) take values in [0,1]. Hence the structured component takes values in [0,1], while the other two components take values in [−1,1]. In particular, all three components are 1-bounded.
On the first range, both multipliers are 1+O(δ3j) by (6). On the second range, their squared difference is at most 1, since both multipliers lie in [0,1]. On the third range, the selection rule at stage j+1 gives
∣μi,j(ξ)∣,∣μi,j+1(ξ)∣<δ3j+5,
so the squared difference of the multipliers is at most δ3j+54. Summing over i and using (17) and (20), we obtain
where C is an absolute constant. The last inequality follows from δ3j≤δ0, the choice of c0, and the condition imposed at stage j+1. This proves (16).
where C is a sufficiently large absolute constant. This constant is absorbed by the choice of scales in (18).
(The structured component.) Set
Γi:=Γi,j,σ:=δ3j,ρ1:=δ3j+1,μi:=μi,j.
The choice of δ3j+1 and (10) give (14). Moreover, since maxi∣Γi∣≤Rj and η1 is nondecreasing, the choice of δ3j+1 gives
ρ1=δ3j+1≤η1(Rj+ε−1)1≤η1(maxi∣Γi∣+ε−1)1.
The choice of ρ2, together with (18), gives ρ2≤η2(ρ1−1)−1. Thus, both scale inequalities in (13) hold.
Finally, since each kernel μi∗μi has mean one, we have Efstr(i)=Ef. The bounds on ∣Γi∣, σ−1, ρ1−1, and ρ2−1 follow from the construction. Applying Lemma 2.1 to the 1-bounded function fpsr(i) and using (15) gives the claimed box-norm estimate. □
Remark 3.4. In Tao’s one-dimensional argument, the structured term is smoothed with a cutoff μ, whereas the ith directional component here uses μi∗μi. Replacing μi by ∣μi∣2 changes the regularity estimates only by fixed absolute factors. The symmetrized kernel is needed when estimating the structured contribution, because Tμi∗μivi=TμiviTμivi allows Lemma 3.6 to handle structured components from different directions.
Polynomial images in prescribed Bohr sets
We want to find a common parameter d for which every structured component is almost invariant under its prescribed shift. The following lemma counts the parameters satisfying Pi(d)∈B(Γi,ri) for all i simultaneously.
Lemma 3.5. Let k∈N and P={P1,…,Pk∈Z[t]} be a collection of linearly independent polynomials with zero constant terms. For 1≤i≤k, let Γi⊆Fp∖{0} and 0<ri≤1/2. Then we have
Proof of Lemma 3.5. The proof is analogous to that of [1 Lemma 3.4] and is omitted. □
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≥0, let v1,…,vm∈FpD, and let f:FpD→[0,∞) satisfy Ef=α. For every 1≤i≤m, let μi:Fp→[0,∞) satisfy Eμi=1. Then we have
Ex∈FpDf(x)i=1∏mTμi∗μivif(x)≥αm+1.(22)
Proof of Lemma 3.6. The cases m=0 and α=0 are immediate. By homogeneity, we may therefore assume that m≥1 and Ef=1. We first treat the case in which f is strictly positive.
We estimate each factor separately after taking logarithms. Fix 1≤i≤m. Since μi is nonnegative and has mean one, Jensen’s inequality gives
where the last equality follows by a change of variables. The last expression is nonnegative by Jensen’s inequality for the convex function u↦ulogu, since ETμivif=1.
We now combine these bounds. Since f is nonnegative and Ef=1, Jensen’s inequality with weights f gives
Exponentiating proves the result when f is strictly positive and has mean one. For a general nonnegative function of mean one, apply the preceding argument to (f+δ)/(1+δ) and let δ↓0. This proves (22). □
Remark 3.7. The symmetrization is essential in general: inequality (22) need not hold with Tμivi in place of Tμi∗μivi. Indeed, take D=m=1, v1=1, and set
f=1{0},μ=2p(1{1}+1{−1}).
Then μ is nonnegative and has mean one, but
Ex∈Fpf(x)Tμ1f(x)=0<p21=(Ef)2.
Completion of the proof
We now combine the preceding results to obtain the required lower bound for every choice of ω with the same nonzero parameter d. 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). We may assume that 0<ε<1. Fix
0<ϑ≪kε.
Choose nondecreasing growth functions η1,η2, depending only on ϑ, P, and k, such that η1(t)≥t and η2(t)−1≪P,kϑ2−kt, with the implicit constant sufficiently small.
Apply Proposition 3.3 to 1A, with parameter ϑ and growth functions η1,η2. We obtain Γi, parameters ρ1,ρ2, normalized nonnegative kernels μi, and decompositions in the prescribed directions
Since the ranks of B(Γi,ρ1) and ρ1−1 are bounded independently of p, Lemma 3.5 implies that, for all sufficiently large primes, we have
p∣B∗∣≥21ρ1∑i=1k∣Γi∣.(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 d using their L2(G) norms. We therefore exclude parameters only for terms containing a pseudorandom component and no small component. Let ∅=S⊆{1,…,k}. For each i∈S, let gi be either fstr(i) or fpsr(i), and suppose that gj=fpsr(j) for some j∈S. Applying Lemma 3.2 to gj gives
There are Ok(1) choices of S and of the functions gi. Let E be the set of parameters for which at least one of these averages has modulus greater than ϑ. By the union bound, ∣E∣/p≪P,kϑ−2(ρ2+p−c). Using (23) to control the first term and taking p sufficiently large to control the second, we obtain
The right-hand side is bounded below by a positive constant depending only on ϑ, P, and k. Hence, for all sufficiently large primes, ∣B∗∖E∣>1, and we may choose
d∈(B∗∖E)∖{0}.(26)
Having fixed d, we verify the required bound for every choice of ω. Fix ω∈{0,1}k and put S:={i:ωi=1}. If S=∅, then
I(P,V)ω(1A,…,1A)(d)=E1A=α.
This gives the required bound when S=∅. Suppose now that S=∅. Leave 1A(x) undecomposed and apply the ith directional decomposition to 1A(x+Pi(d)vi):
We first consider the term containing only structured components. Since d∈B∗, we have Pi(d)∈B(Γi,ρ1) for every i. Telescoping the product and using 1-boundedness gives
Every other term in (27) contains a small or a pseudorandom component. If a term contains fsml(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.
If it contains no small component but contains a pseudorandom component, then it is one of the terms used to define E, and its modulus is at most ϑ because d∈/E. There are at most 3∣S∣−1≤3k 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.
The parameter d in (26) is independent of ω, so this holds simultaneously for all ω∈{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
[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]Nikos Frantzikinakis and Bryna Kra. Ergodic averages for independent polynomials and applications. J. London Math. Soc., 74(1):131–142, 2006.DOI
[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]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
[6]Sarah Peluse. On the polynomial Szemerédi theorem in finite fields. Duke Math. J., 168(5):749–774, 2019.DOI
[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]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.