By OpenAISourceGitHub

Manuscripts from openai/math

Fork
Clone
git clone https://openversions.com/git/index/openai-math.git
Cite
Version link
https://openversions.com/index/openai-math/commit/dfd8c46870b48511e697cbb43a3c8c91d375fa2f
Version identifier (commit hash)
dfd8c46870b48511e697cbb43a3c8c91d375fa2f

Fork to create your own copy with credit to the source. Clone to get a Git copy on your computer.

All papers

Claims

Findings

A power saving for square-difference-free sets

There are absolute constants c>0c > 0 and C<∞C < \infty such that, for every integer N≥1N \ge 1 and every square-difference-free set A⊆[N]A \subseteq [N], ∣A∣≤CN1−c|A| \le C N^{1-c}.

Whenever integers r,N≥1r, N \ge 1 satisfy N>(Cr)1/cN > (Cr)^{1/c} (with c,Cc, C from Theorem 1.1 and 0<c<10 < c < 1), every rr-coloring of [N][N] has same-colored elements a,a+d2a, a + d^2 with d≥1d \ge 1, and the chromatic number of the square-difference graph GNG_N satisfies χ(GN)≥C−1Nc\chi(G_N) \ge C^{-1} N^c.

There is an absolute integer PP such that for every prime p≥Pp \ge P, there exist a probability measure μp\mu_p on FpV\mathbb{F}_p^V and a nonnegative submeasure εp≤μp\varepsilon_p \le \mu_p with ηp=εp(FpV)≤p−4\eta_p = \varepsilon_p(\mathbb{F}_p^V) \le p^{-4} such that μp\mu_p is reflection positive, both μp\mu_p and εp\varepsilon_p are square-affine invariant with uniform single-label marginals, μp∘=(μp−εp)/(1−ηp)\mu_p^\circ = (\mu_p - \varepsilon_p)/(1 - \eta_p) is supported on zvj+1−zvj∈Qp∗z_{v_{j+1}} - z_{v_j} \in Q_p^* with uniform pair marginals on Ωp\Omega_p, and both μp\mu_p and μp∘\mu_p^\circ satisfy conditional mixing with exponent γ=1/64\gamma = 1/64.

For every finite set P\mathcal{P} of primes at least PP and real functions g,gv:XP→Rg, g_v : X_{\mathcal{P}} \to \mathbb{R}, ZP(g)≥0\mathcal{Z}_{\mathcal{P}}(g) \ge 0 and ∣ZP((gv)v∈V)∣≤∏v∈VZP(gv)1/d|\mathcal{Z}_{\mathcal{P}}((g_v)_{v \in V})| \le \prod_{v \in V} \mathcal{Z}_{\mathcal{P}}(g_v)^{1/d}.

For a finite set P\mathcal{P} of primes at least PP, ρp∈{μp,μp∘}\rho_p \in \{\mu_p, \mu_p^\circ\}, R≥1R \ge 1, and functions gw:XP→Cg_w : X_{\mathcal{P}} \to \mathbb{C} such that gv^(ξ)=0\widehat{g_v}(\xi) = 0 whenever q(ξ)≤Rq(\xi) \le R for some v∈Vv \in V, ∣∫∏w∈Vgw(zw) d⨂p∈Pρp(z)∣≤R−γ∥gv∥2∏w≠v∥gw∥2(d−1)\left|\int \prod_{w \in V} g_w(z_w) \, d\bigotimes_{p \in \mathcal{P}} \rho_p(z)\right| \le R^{-\gamma} \|g_v\|_2 \prod_{w \neq v} \|g_w\|_{2(d-1)}.

For Q≥1Q \ge 1, an interval II of length L≥10Q6L \ge 10 Q^6, ∥f∥∞≤M\|f\|_\infty \le M, and any subfamily F⊆{U⊆P:qU≤Q}\mathcal{F} \subseteq \{U \subseteq \mathcal{P} : q_U \le Q\}, ∥LI,P,QFf∥2r≤2M(k+1)Ark\|\mathcal{L}_{I,\mathcal{P},Q}^{\mathcal{F}} f\|_{2r} \le 2M(k+1)A_r^k for every integer r≥1r \ge 1, which is at most MQor(1)M Q^{o_r(1)} uniformly in I,f,P,FI, f, \mathcal{P}, \mathcal{F}.

For N≥1N \ge 1, H=Nβ/2H = N^{\beta/2}, D=8d0≤NβD = 8d_0 \le N^\beta (d0d_0 odd squarefree), and F,J:[N]→CF, J : [N] \to \mathbb{C} with ∣F∣,∣J∣≤1|F|, |J| \le 1 supported in residue classes a,b(modD)a, b \pmod D where b−ab-a is a strict square class, x<yx < y on supports, and no y−xy-x is a positive integer square, D∣∑ξ∈Q/Z:q(ξ)≤H,q(ξ) sqfree,(q(ξ),D)=1g(ξ)F^(−ξ)J^(ξ)∣≪H−1/3D \left| \sum_{\xi \in \mathbb{Q}/\mathbb{Z}: q(\xi) \le H, q(\xi) \text{ sqfree}, (q(\xi),D)=1} \mathfrak{g}(\xi) \widehat{F}(-\xi) \widehat{J}(\xi) \right| \ll H^{-1/3}.

For N≥NtrN \ge N_{\mathrm{tr}}, A⊆[N]A \subseteq [N] square-difference-free, B⊆PNB \subseteq \mathcal{P}_N with qB≤Nτq_B \le N^\tau, N′=⌈N/D2⌉N' = \lceil N/D^2 \rceil, λ=D2N′/N\lambda = D^2 N'/N, and gv=(L[N],PN,Qfsv)(zB,v,⋅)g_v = (\mathcal{L}_{[N],\mathcal{P}_N,Q} f_{s_v})(z_{B,v}, \cdot), ∣ZS((gv)v∈V)∣≤λdmax⁡A′⊆[N′] sq-diff-freeY(N′,A′)+CtrN−σ0|\mathcal{Z}_{\mathcal{S}}((g_v)_{v \in V})| \le \lambda^d \max_{A' \subseteq [N'] \text{ sq-diff-free}} Y(N', A') + C_{\mathrm{tr}} N^{-\sigma_0}.

There are absolute constants CC and N0N_0 such that whenever N≥N0N \ge N_0, A⊆[N]A \subseteq [N] has no positive square difference, and s\mathbf{s} is good, ∣∫∏v∈VL[N],PN,Qfsv(zv) d⨂p∈PN(μp−εp)∣≤CN−σ0\left| \int \prod_{v \in V} \mathcal{L}_{[N],\mathcal{P}_N,Q} f_{s_v}(z_v) \, d\bigotimes_{p \in \mathcal{P}_N} (\mu_p - \varepsilon_p) \right| \le C N^{-\sigma_0}.

There are absolute constants a0>0a_0 > 0 and C0<∞C_0 < \infty such that, for every positive integer NN and every square-difference-free set A⊆[N]A \subseteq [N], Y(N,A)≤C0N−a0Y(N, A) \le C_0 N^{-a_0}.

Methods

Constructing finite-field reflection-positive probability laws µ_p with a designated 24-cycle of strict square differences outside an exceptional submeasure ε_p of mass ≤ p−4p^{-4} by mixing measures with marked blocks and list-entry couplings.

Source excerpts

“For S ⊆ [ℓ] with s = |S| ≤ t0 + 1, regard the blocks in S as marked; blocks outside S are called clean. ... Define the finite measure νp,S by the identity ∫ G(z) dνp,S(z) = EY Ej IS(Y )WS(j)G((YuS(v),jv )v∈V ).” · Show in paper

Proving a finite signed Hölder inequality (chessboard method) for the multilinear functional Z_P by applying a finite sorting sequence of reflection folds.

Source excerpts

“Concatenating the sequences for all ℓ blocks gives fold maps ϕ1, . . . , ϕT , with T = ℓ(h − 1)2, such that ϕT ◦ · · · ◦ ϕ1(v) = v∗ for every v ∈ V, where v∗ has the identity word in every block.” · Show in paper

Bounding arbitrary even moments of Fourier lifts uniformly in the interval, function, prime set, and retained denominator family via a conditional square-sum bound, transversal coloring, and coordinatewise copy-swapping symmetrization.

Source excerpts

“We establish a conditional square-sum bound before estimating moments. ... Swapping independent copies coordinate by coordinate is a standard symmetrization device” · Show in paper

Approximating a signed divisor-sieve kernel supported on actual positive integer squares by a rational-frequency model kernel using Weyl differencing on minor arcs and Stieltjes partial summation on major arcs.

Source excerpts

“We first define the model and actual-square kernels. We then compare their Fourier transforms on minor and major arcs, and finish with the exact bilinear Fourier identity.” · Show in paper

Comparing the product-space multilinear functional across scales by specializing a small set of prime coordinates and splitting the resulting residue class into progressions with square step D^2.

Source excerpts

“Splitting the exact fiber into square-step progressions. Let a mod D combine s and zB. ... Each A′j is square-difference-free: a difference m2 > 0 between its elements would give the difference D2m2 = (Dm)2 between the corresponding elements of A.” · Show in paper

Partitioning [N] into ordered intervals to isolate an ascending edge along a 24-cycle of good residue assignments, truncating all other slots in conductor, and applying the ordered square-kernel estimate.

Source excerpts

“In every remaining assignment there is a directed cycle edge u → v whose first interval precedes its second. ... Retain these two lifts in full. In each of the other d − 2 slots, truncate the lift to exact conductors at most T = Nτ.” · Show in paper

Expanding the product of prime measures by inclusion-exclusion into nonexceptional and exceptional parts to derive a weighted recurrence and close a strong induction for power-saving decay.

Source excerpts

“The finite product expansion gives the following identity of signed measures: ⊗p∈PN µp − ⊗p∈PN (µp − εp) = ∑∅̸=B⊆PN (−1)|B|+1κB.” · Show in paper