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 uniform Hilbert transform estimate for Lipschitz directions

There exist absolute constants a∗∈(0,1/2)a_* \in (0, 1/2) and C∗<∞C_* < \infty such that every Lipschitz unit vector field v:R2→S1v : \mathbb{R}^2 \to S^1 with Lip(v)≤1\mathrm{Lip}(v) \le 1 satisfies sup⁡0<ϵ<a∗∥Hv,a∗ϵf∥L2(R2)≤C∗∥f∥L2(R2)\sup_{0 < \epsilon < a_*} \|H_{v,a_*}^\epsilon f\|_{L^2(\mathbb{R}^2)} \le C_* \|f\|_{L^2(\mathbb{R}^2)} for all f∈S(R2)f \in \mathcal{S}(\mathbb{R}^2), and its symmetric principal value Hv,a∗fH_{v,a_*}f satisfies the same strong L2L^2 bound and the weak-(2,2)(2,2) bound ∣{x:∣Hv,a∗f(x)∣>λ}∣≤C∗2λ−2∥f∥22|\{x : |H_{v,a_*}f(x)| > \lambda\}| \le C_*^2 \lambda^{-2} \|f\|_2^2 for λ>0\lambda > 0, answering Stein's weak-(2,2)(2,2) conjecture affirmatively.

For the fixed absolute constants in (1.2), there is an absolute constant CC such that for every slope field uu with ∣u∣≤1|u| \le 1 and Lip(u)≤CLip\mathrm{Lip}(u) \le C_{\mathrm{Lip}}, every 0<ϵ<a0 < \epsilon < a, every dyadic ww, and every vertical band cutoff PwP_w, the restricted estimate ∣⟨Hu,aϵPwf,g⟩∣≤C(1+∣log⁡(∣F∣/∣E∣)∣)−3∣F∣∣E∣|\langle H_{u,a}^\epsilon P_w f, g \rangle| \le C(1 + |\log(|F|/|E|)|)^{-3} \sqrt{|F||E|} holds whenever F,E⊂I×RF, E \subset I \times \mathbb{R} have positive finite measure, ∣f∣≤1F|f| \le \mathbf{1}_F, ∣g∣≤1E|g| \le \mathbf{1}_E, and ∣I∣≤b0|I| \le b_0.

The uniform restricted estimate on a vertical band in Theorem 1.4 implies the global uniform L2L^2 and weak-(2,2)(2,2) estimates in Theorem 1.1.

If the three uniform model estimates (1.38), (1.39), and (1.40) hold for every admissible finite packet form ΛS\Lambda_{\mathcal{S}} in Definition 1.13, then Theorem 1.4 holds with constants uniform in ww, ϵ\epsilon, and the number of retained scales.

The three model estimates—(i) ∣Λ(f,g)∣≲(1+log⁡(2m))C∣F∣∣E∣|\Lambda(f,g)| \lesssim (1 + \log(2m))^C \sqrt{|F||E|} for m≥1m \ge 1, (ii) ∣Λ(f,g)∣≲L−20∣F∣∣E∣|\Lambda(f,g)| \lesssim L^{-20} \sqrt{|F||E|} for L≥L0L \ge L_0, L/2≤log⁡∣E∣/∣F∣≤2LL/2 \le \log\sqrt{|E|/|F|} \le 2L, and Lipschitz uu on I×RI \times \mathbb{R}, and (iii) ∣Λ(f,g)∣≲∣F∣∣E∣(∣E∣/∣F∣)c|\Lambda(f,g)| \lesssim \sqrt{|F||E|}(|E|/|F|)^c for m=1m=1 and ∣E∣≤∣F∣|E| \le |F|—hold uniformly over every finite subcollection of the packet forms (1.33).

For the packet forms ΛS(f,g)\Lambda_{\mathcal{S}}(f,g) with H=CJ\mathcal{H} = \mathbb{C}^J, J≤CbinmJ \le C_{\mathrm{bin}} m, ∥g(z)∥H≤1E(z)\|g(z)\|_{\mathcal{H}} \le \mathbf{1}_E(z), and DE(s)≤γD_E(s) \le \gamma for all s∈Ss \in \mathcal{S}, ∣ΛS(f,g)∣≤C(1+log⁡(2m))Cγ1/4∣E∣1/2∥f∥L2(H)|\Lambda_{\mathcal{S}}(f,g)| \le C(1 + \log(2m))^C \gamma^{1/4} |E|^{1/2} \|f\|_{L^2(\mathcal{H})}; in the scalar case with ∣f∣≤1F|f| \le \mathbf{1}_F, ∣g∣≤1E|g| \le \mathbf{1}_E, and 0<∣E∣<∣F∣<∞0 < |E| < |F| < \infty, ∣ΛS(f,g)∣≤C∣F∣∣E∣(∣E∣/∣F∣)9/20|\Lambda_{\mathcal{S}}(f,g)| \le C\sqrt{|F||E|}(|E|/|F|)^{9/20}, both uniformly for every subcollection.

For the popular tops in (2.30), the weighted top count n(z)=∑T∈ThT(z)n(z) = \sum_{T \in \mathcal{T}} h_T(z) satisfies the weak-L2L^2 estimate ∥n∥L2,∞(D)≤CLCS\|n\|_{L^{2,\infty}(\mathcal{D})} \le C L^C \sqrt{S}.

There is a fixed exponent CDC_D such that for every fixed J>0J > 0 and all sufficiently large LL, there is a measurable exceptional set Ecount⊂R2\mathcal{E}_{\mathrm{count}} \subset \mathbb{R}^2 with ∣Ecount∣≤Se−JL|\mathcal{E}_{\mathrm{count}}| \le S e^{-JL} outside which, on Z0=Fconc∖EcountZ_0 = F_{\mathrm{conc}} \setminus \mathcal{E}_{\mathrm{count}}, every ρ\rho-separated list L\mathcal{L} of significant tops at zz of diameter at most BρB\rho and precision at most ρ/B2\rho/B^2 has cardinality #L≤D/100\#\mathcal{L} \le D/100 with D=LCDD = L^{C_D}.

Let λ1,…,λd∈R\lambda_1, \dots, \lambda_d \in \mathbb{R} with d≤Dd \le D satisfy ∣λi−λj∣≥Δ>0|\lambda_i - \lambda_j| \ge \Delta > 0 for i≠ji \ne j, and let ν\nu be any positive probability measure on R\mathbb{R} with supp ν^⊂(−ϵ0Δ,ϵ0Δ)\mathrm{supp}\,\widehat{\nu} \subset (-\epsilon_0 \Delta, \epsilon_0 \Delta). Then for every finite p≥4p \ge 4 and every complex (or diagonally acted finite-dimensional Hilbert-valued) coefficient sequence, (∫R∣∑i=1daieiλit∣p dν(t))1/p≤(∑i=1d∣ai∣r∗)1/r∗\left(\int_{\mathbb{R}} |\sum_{i=1}^d a_i e^{i\lambda_i t}|^p \, d\nu(t)\right)^{1/p} \le (\sum_{i=1}^d |a_i|^{r_*})^{1/r_*} with constant one, where r∗=(1−(α−1)/p)−1r_* = (1 - (\alpha - 1)/p)^{-1} and α=2+cph/log⁡(2D)\alpha = 2 + c_{\mathrm{ph}}/\log(2D).

The popular adjoint sum P=∑sϕsP = \sum_s \phi_s satisfies ∫F∥P∥≲L−20N∣F∣\int_F \|P\| \lesssim L^{-20} N |F|.

Methods

Localization to small squares, rotation to slope form Hu,aϵH_{u,a}^\epsilon, and replacement of variable truncation endpoints at uniform L2L^2 cost.

Source excerpts

“Lemma 1.3. For |t| CLip < 1, define Ath(x, y) = h(x − t, y − tu(x, y)). For 1 ≤ p < ∞, ∥Ath∥p ≤ (1 − |t| CLip) −1/p ∥h∥p , (1.3) and At is a contraction on L∞. Consequently, replacing either symmetric endpoint at scale r by a measurable endpoint between cr and Cr, with c, C > 0 fixed and max(1, C)rCLip < 1/2, costs a bounded Lp operator, independently of r.” · Show in paper

Lipschitz commutator estimate ∥[b,T]∂h∥2≤CA(m)∥b′∥∞∥h∥2\|[b, T]\partial h\|_2 \le C A(\mathfrak{m}) \|b'\|_\infty \|h\|_2 and randomized Littlewood-Paley square-function reassembly across vertical frequency bands.

Source excerpts

“Lemma 1.7. Let b ∈ W1,∞(R), and let T = m(D), where m is smooth away from zero and A(m) = max 0≤r≤4 sup ζ̸=0 |ζ| r |m (r) (ζ)| < ∞. Then ∥[b, T]∂h∥2 ≤ CA(m) ∥b ′∥ ∞ ∥h∥2 . (1.8)” · Show in paper

“Randomize the signs independently and use their orthogonality. For f = 1I×Rf, Equation (1.15) and the triangle inequality in the random-sign L 2 space give ( ∑ w∈W ∥1I×RPwHϵ u,af∥ 2 2 )1/2 ≤ C ∥f∥2 + ( ∑ w∈W ∥1I×RHϵ u,aPwf∥ 2 2 )1/2 . (1.16)” · Show in paper

Narrow vertical frequency bin decomposition into a finite-dimensional Hilbert space H=CJ\mathcal{H} = \mathbb{C}^J and amplitude-level decomposition.

Source excerpts

“Choose real smooth cutoffs Qj , 1 ≤ j ≤ J ≤ Cm, supported in intervals of length C/W centered at ηj ≍ w −1 , with bounded overlap and ∑J j=1 Qj (η) 2 = 1 on supp p(wη).” · Show in paper

“Lemma 1.12 (Amplitude levels after bin projection).” · Show in paper

Wave-packet representation via Fourier series in normalized coordinates and the three-grid adjacent-grid principle.

Source excerpts

“Proposition 1.14 (Packet representation). The diagonal form ∑ j ⟨Au,LQejfj , gj ⟩ is a sum, with absolutely summable scalar coefficients, of averages of forms (1.33).” · Show in paper

Size-density-tree decomposition and assignment of high-density tiles to popular tops.

Source excerpts

“Lemma 2.3 (Size selection).” · Show in paper

“Lemma 2.4 (Density selection).” · Show in paper

Every-line concentration using Fourier/martingale 3-variation comparison (Lépingle's inequality) and local John-Nirenberg moment bounds.

Source excerpts

“Lemma 3.3. Suppose Fk ∈ L 2 (R; H) has Fourier support in {ν : ca2 k ≤ |ν| ≤ Ca2 k}, with fixed positive ca, Ca. Then ∥V3(Fk)∥L2(R) ≤ Cca,Ca ( ∑ k ∥Fk∥ 2 2 )1/2 . (3.12)” · Show in paper

“Proposition 3.5. For every slope t, every collection satisfying (3.1), every line Yt = y0, every horizontal interval I of length l on that line, and every b ≥ 2, the normalized L b (I) norms of Et, St, and V3 ( ∑ ℓs=2k ϕs ) , V3 ( ∑ ℓs=2k qT(s)ϕs ) are at most CbC(1 + log B) C.” · Show in paper

Shifted-block angular hierarchy, mass-stopping path expansion, and clock-selected phase averaging on marked spatial cells.

Source excerpts

“Lemma 5.7 (Expansion into top paths).” · Show in paper

“Lemma 5.10 (Fourier-supported averaging).” · Show in paper

Analytic three-lines interpolation of shift parameters and angular-grid boundary averaging.

Source excerpts

“Lemma 7.2 (Interpolation of the shift parameters). Equation (7.1) also holds at ζ c = (µ/K, . . . , µ/K).” · Show in paper

“Lemma 7.3 (Expected boundary loss). Let ϵ = 4G−1/4 . For z ∈ Z0, Eϑ∥Pq(z) − Wϑ(z)∥ ≤ L Cϵn(z). (7.5)” · Show in paper