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

An <i>L</i><sup>3</sup> bound for the trilinear Hilbert transform

There is a finite constant CC such that ∥T(f1,f2,f3)∥L1(R)≤C∏j=13∥fj∥L3(R)\|T(f_1, f_2, f_3)\|_{L^1(\mathbb{R})} \le C \prod_{j=1}^3 \|f_j\|_{L^3(\mathbb{R})} for all Schwartz functions f1,f2,f3∈S(R)f_1, f_2, f_3 \in \mathcal{S}(\mathbb{R}), where T(f1,f2,f3)(x)=p.v.∫Rf1(x−t)f2(x−2t)f3(x−3t)dttT(f_1, f_2, f_3)(x) = \mathrm{p.v.} \int_{\mathbb{R}} f_1(x-t)f_2(x-2t)f_3(x-3t)\frac{dt}{t}, and consequently TT has a unique bounded trilinear extension from L3(R)×L3(R)×L3(R)L^3(\mathbb{R}) \times L^3(\mathbb{R}) \times L^3(\mathbb{R}) to L1(R)L^1(\mathbb{R}).

For some q∈(2,3)q \in (2, 3), the least constants CN(q)C_N(q) in the local maximal-function bound ∑I∈I∣I∣∣HI(f0,f1,f2,f3)∣≤CN∫R∏j=03Mqfj\sum_{I \in \mathcal{I}} |I||H_I(f_0, f_1, f_2, f_3)| \le C_N \int_{\mathbb{R}} \prod_{j=0}^3 M_q f_j are bounded independently of the number of consecutive depths NN.

For a suitable fixed q∈(2,3)q \in (2, 3) and every sufficiently small fixed c>0c > 0, there is an A<∞A < \infty, independent of NN, such that for any finite rooted subtree with root I∗I_* lying in at most NN consecutive depths and satisfying ∥fj∥q,I≤1\|f_j\|_{q,I} \le 1 (0≤j≤30 \le j \le 3, II in the subtree), ∑I a summation cell∣I∣∣HI(f)∣≤(A+cCN)∣I∗∣\sum_{I \text{ a summation cell}} |I| |H_I(f)| \le (A + c C_N)|I_*|.

For inherited mask families P1,…,PmP^1, \dots, P^m under a common budget D≥2D \ge 2 on a finite dyadic tree with root JJ, and finitely many bounded linear rows Bν:L2(J;H)→YνB_\nu : L^2(J; \mathcal{H}) \to \mathcal{Y}_\nu satisfying ∑ν∥Bνz∥Yν2≤∥z∥L2(J;H)2\sum_\nu \|B_\nu z\|_{\mathcal{Y}_\nu}^2 \le \|z\|_{L^2(J;\mathcal{H})}^2 and Bν=Bν1u(ν)B_\nu = B_\nu \mathbf{1}_{u(\nu)}, an absolute constant CC satisfies (∑ν∥BνPu(ν)m⋯Pu(ν)1z∥2)1/2≤(C(m+1)Clog⁡(2D))C(m+1)∥z∥L2(J;H)\left(\sum_\nu \|B_\nu P_{u(\nu)}^m \cdots P_{u(\nu)}^1 z\|^2\right)^{1/2} \le (C(m+1)^C \log(2D))^{C(m+1)} \|z\|_{L^2(J;\mathcal{H})}, uniformly in the row cells, total number of labels, and number of spatial depths.

If R,h,u,vR, h, u, v occupy the four slots of HIH_I in any order and satisfy ∥R∥2,I≤C1\|R\|_{2,I} \le C_1, ∥h∥q,I≤1\|h\|_{q,I} \le 1, ∥u∥2,I,∥v∥2,I≤1\|u\|_{2,I}, \|v\|_{2,I} \le 1, and ∣HI(R,h,u,v)∣>δ|H_I(R, h, u, v)| > \delta for q>2q > 2, C1≥1C_1 \ge 1, and 0<δ<1/20 < \delta < 1/2, then RR has a normalized moment against a bounded quadratic chart atom φ\varphi of magnitude at least D−1D^{-1}, where the complexity of φ\varphi is at most D≤exp⁡(Cq(1+log⁡C1+log⁡(1/δ))C)D \le \exp(C_q (1 + \log C_1 + \log(1/\delta))^C) with an absolute exponent CC independent of qq.

The remaining patterns in the two-earliest-increments expansion (with at least three BB arguments and optionally one UU argument) can be restricted, with total absolute error at most e−k∣R∣e^{-k}|R|, to a common active collection whose path multiplicity satisfies M≤DbaseCM \le D_{\mathrm{base}}^C, leaving ∑I active∣I∣min⁡(η,∣HI(z)∣)\sum_{I \text{ active}} |I| \min(\eta, |H_I(z)|) to be estimated for t≥12t \ge 12.

For an active pattern in (11), there are poly(k)\mathrm{poly}(k) fixed search leaves such that, up to total absolute length-weighted error at most Ce−k∣R∣C e^{-k}|R|, each leaf contributes ∑G currentHI(Z0,G,s,Z1,G,s,Z2,G,s,Z3,G,s)\sum_{G \text{ current}} H_I(Z_{0,G,s}, Z_{1,G,s}, Z_{2,G,s}, Z_{3,G,s}), where Zl,G,s=L‾l,G,sBl,szlZ_{l,G,s} = \overline{L}_{l,G,s} \mathcal{B}_{l,s} z_l is a positive smeared ball filter composed with a product of cumulative filters or complements of length O(log⁡L)O(\log \mathcal{L}).

Assuming the base, localization, helper, and linear-side square bound (17) hypotheses, for p∈(2/3,1)p \in (2/3, 1), after common first-crossing cuts of arbitrarily small fixed total relative length, terms with at least three side factors have total length-weighted pp-th power cost at most PC∣R∣P^C |R|, and terms with exactly two side factors and whole current coarse complements have freeze, large-complement, and restoration errors of pp-th power cost at most PC∣R∣P^C |R|, transferred main terms of absolute cost at most ηPC∣R∣\eta P^C |R|, and coupled near-band terms obeying (1+CN(q))CqP−1/2G0(K+2)C(q−2)∣R∣(1 + C_N(q)) C_q P^{-1/2} G_0 (K+2)^{C(q-2)} |R|.

For each basic-BB role, the exact subtotal Vl,G,bV_{l,G,b} of the linear-side modes in Z~l,G,b\widetilde{Z}_{l,G,b} satisfies ∑G,b∥Vl,G,b∥2≤G0∣R∣\sum_{G,b} \|V_{l,G,b}\|^2 \le G_0 |R| uniformly in the branch offsets, verifying the linear-side square bound (17); furthermore, after removing summands on first stopping cells of arbitrarily small fixed total relative length, the distinct representatives AjG\mathcal{A}_j^G of successful components of a fixed role, height, type, and dyadic amplitude bin [σj,2σj)[\sigma_j, 2\sigma_j) satisfy ∑G∣AjG∣≤min⁡(D∗,PCσj−2)\sum_G |\mathcal{A}_j^G| \le \min(D_*, P^C \sigma_j^{-2}).

There are absolute constants ε>0\varepsilon > 0 and C4<∞C_4 < \infty, with 1−ε>2/31 - \varepsilon > 2/3, such that the terms containing at least three core roles in the localized expansion of (11) have total contribution at most CqηεPC4∣R∣+O(e−k∣R∣)C_q \eta^{\varepsilon} P^{C_4} |R| + O(e^{-k}|R|).

Under the hypotheses of the finite sparse-count lemma (pair determination, side incidence budget, scale-wise separation, annular progression relations, and pure-shift gap), for 0<δ≤10 < \delta \le 1 and ℓ=1+log⁡2(2+(d+1)2ν/μ)\ell = 1 + \log_2(2 + (d+1)^2 \nu / \mu), one has ∑s:Ns≤δM2Ns≤CℓL29δ1/2080M2\sum_{s: N_s \le \delta M^2} N_s \le C \ell L^{29} \delta^{1/2080} M^2, with constants and exponents independent of dd, the real frequencies, the number of scales, and all label cardinalities.

Methods

Outer reduction of the continuous trilinear Hilbert transform to a uniform local maximal-function estimate via a compactly supported Gevrey kernel decomposition W(X,S)=∏j=03(∂S−j∂X)[ω(X)ϕ(S)]W(X,S) = \prod_{j=0}^3 (\partial_S - j\partial_X)[\omega(X)\phi(S)] and shifted dyadic lattices.

Source excerpts

“To recover the continuous kernel, take an odd compactly supported Gevrey function \phi with \phi^{(3)}(0) \neq 0, and a compactly supported Gevrey function \omega such that \sum_{m\in\mathbb{Z}} \omega(X - m) = 1. Set W(X, S) = \prod_{j=0}^3 (\partial_S - j\partial_X)[\omega(X)\phi(S)].” · Show in paper

Private orthogonal coordinates Φv(y)=(φv(y),ϵvev)\Phi_v(y) = (\varphi_v(y), \sqrt{\epsilon_v}\mathbf{e}_v) in H=C⊕ℓ2(V)\mathcal{H} = \mathbb{C} \oplus \ell^2(\mathcal{V}) combined with a Rademacher-Menshov binary-prefix estimate to compress products of inherited local projections uniformly in spatial depth.

Source excerpts

“Work in the value Hilbert space \mathcal{H} = \mathbb{C} \oplus \ell^2(\mathcal{V}), and define the column on its birth cell by \Phi_v(y) = (\varphi_v(y), \sqrt{\epsilon_v}\mathbf{e}_v), y \in J_v.” · Show in paper

“Its proof uses the classical Rademacher–Menshov binary-decomposition mechanism; see [9, Section 3, Lemma 1] for a vector-valued presentation.” · Show in paper

Continuous bounded degree-two inverse theorem transferring the Leng-Sah-Sawhney discrete U3[L]U^3[L] inverse theorem to continuous intervals via cell averaging, Malcev coordinates, Fejér kernel approximation, and exact floor interpolation.

Source excerpts

“For L \ge 1, define the cell averages g_L(j) = L \int_{j/L}^{(j+1)/L} g(y) dy, 0 \le j < L, and extend them by zero to \mathbb{Z}/(16L\mathbb{Z}).” · Show in paper

“We finish with an exact floor interpolation, which is necessary because the chosen atom may oscillate on the sampling scale.” · Show in paper

Two-earliest-increments expansion with staggered accuracies kt+1=ktAk_{t+1} = k_t^A and delayed high fits on finer partitions justified by predictor moments.

Source excerpts

“Choose an absolute A > 1 and large initial k_1, and set kt+1k_{t+1} = k_t^A. Apply Lemma 5.4 at level t with accuracy e−kte^{-k_t}; denote its projection in the updated position by gt,Ig_{t,I}.” · Show in paper

“A predictor for the multiplier \psi_a is the augmented column \Phi_p = (\overline{\psi_a}, \sqrt{\epsilon} \mathbf{e}_p), 0 < \epsilon \le 1.” · Show in paper

Curvature-based binary localization using an ambient graph of normalized quadratic curvature tags, randomized geometric delays for colorings with controlled changes, and historical ball masks.

Source excerpts

“Use the ambient graph of all hypothetical curvature tags satisfying the fixed basic bounds, not just the currently present vertices.” · Show in paper

“In each of d_0 independent primary rounds, give every original vertex z a geometric delay D_z, truncated to [0, d_0], with untruncated tail \mathbb{P}(D_z \ge n) = 2^{-n}.” · Show in paper

Short-block comparison with CN(q)C_N(q) via a slow lattice and fast torus decomposition, quadratic randomization Uj,γU_{j,\gamma} on the fast torus, and Rubio de Francia mixed-norm extrapolation.

Source excerpts

“A slow lattice and a fast torus. Write r_s for the interval length at depth s, and choose \delta = \beta r_t, where \beta > 0 is a sufficiently small fixed dyadic number depending only on the support margin.” · Show in paper

“On Fourier coefficients in the fast variable define the unitary maps \widehat{Uj,γU_{j,\gamma}F}(n, m) = e(\gamma m^2 / c_j)\widehat{F}(n, m), \gamma \in \mathbb{R}/(3\mathbb{Z}).” · Show in paper

Single first-match mode expansion with random branch offsets, PRE and CURRENT frequency graphs, and hereditary representatives.

Source excerpts

“At its first match, on a cell I_0, make one expansion choice and retain that choice on every relevant descendant of I_0.” · Show in paper

“At height h, join two modes in the PRE graph if the first bound below holds, and in the CURRENT graph if the second holds, for some integral n with |n|_\infty \le E_h” · Show in paper

Finite sparse-count argument using counting blocks, difference covers via four-step paths, and minimal-growth subset iteration.

Source excerpts

“Counting blocks and difference covers. Restrict to the scales with N_s \le \delta M^2.” · Show in paper

“The minimal-growth subset argument. Choose a nonempty X_i \subseteq Z_i minimizing r_i = |X_i - B_i|/|X_i|.” · Show in paper