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

Almost-everywhere Fourier convergence in L log L

For every complex-valued function f∈Llog⁡L(T)f \in L\log L(\mathbb{T}), there is a measurable set NfN_f of measure zero such that the ordinary symmetric Fourier partial sums SNf(x)S_N f(x) converge to f(x)f(x) for all x∈T∖Nfx \in \mathbb{T} \setminus N_f along the full sequence N=0,1,2,…N = 0, 1, 2, \dots, establishing the classical Llog⁡LL\log L sufficiency conjecture.

For the ordinary symmetric partial sums S∗f=sup⁡N≥0∣SNf∣S^* f = \sup_{N\ge 0} |S_N f|, there is an absolute constant CC such that sup⁡λ>0λμ{S∗f>λ}≤C∥f∥Φ\sup_{\lambda > 0} \lambda \mu\{S^* f > \lambda\} \le C \|f\|_\Phi for every f∈Llog⁡L(T;C)f \in L\log L(\mathbb{T}; \mathbb{C}), where ∥f∥Φ\|f\|_\Phi is the Luxemburg norm for Φ(t)=tlog⁡(2+t)\Phi(t) = t\log(2+t).

There are absolute constants CC and d∗d_* such that whenever d≥d∗d \ge d_* and ρ\rho is a nonnegative periodic function with ∫01ρ=1\int_0^1 \rho = 1 and ρ≤ed\rho \le e^d almost everywhere, ∫Gρsup⁡N≥0∣SNρ(x)∣ dx≤Cd\int_{G_\rho} \sup_{N\ge 0} |S_N \rho(x)|\,dx \le Cd, where Gρ={Mρ≤e4h}G_\rho = \{M\rho \le e^{4h}\} and h=log⁡log⁡dh = \log\log d.

There are absolute constants CC and d0d_0 such that for d≥d0d \ge d_0, h=log⁡log⁡dh = \log\log d, any half-open interval I⊂RI \subset \mathbb{R} with 1/2≤∣I∣≤11/2 \le |I| \le 1, any measurable ρ:I→[0,∞)\rho : I \to [0, \infty) with ρ≤ed\rho \le e^d a.e. and mtop=∣I∣−1∫Iρ≤2m_{\mathrm{top}} = |I|^{-1}\int_I \rho \le 2, any J≥1J \ge 1, Q≥0Q \ge 0, and Fj+1\mathcal{F}_{j+1}-measurable multipliers ϵj\epsilon_j with ∣ϵj∣≤1|\epsilon_j| \le 1, ∫Imax⁡u∈Z,∣u∣≤Q∣HI,Juρ(x)∣dx∣I∣≤Cd\int_I \max_{u\in\mathbb{Z}, |u|\le Q} |\mathcal{H}_{I,J}^u \rho(x)| \frac{dx}{|I|} \le Cd, uniformly in I,ρ,J,QI, \rho, J, Q, and the multipliers.

Under the hypotheses of Lemma 3.1 (sumset bound log⁡∣nS∣≤D0dhP0log⁡d\log |nS| \le D_0 d h^{P_0} \log d for 1≤n≤d21 \le n \le d^2 and uniform moment bound −log⁡EλHq(X1+⋯+XT−X1′−⋯−XT′)≤E-\log \mathbb{E}_\lambda H_q(X_1+\dots+X_T-X'_1-\dots-X'_T) \le E with d/h2≤E≤D0dd/h^2 \le E \le D_0 d), there exist x0∈Sx_0 \in S and a nonempty subset S′⊂SS' \subset S such that log⁡∣S∣−log⁡∣S′∣≤C(D0)d2Eh2\log |S| - \log |S'| \le C(D_0) \frac{d^2}{Eh^2} and ∫exp⁡(−L∑x∈S′v(x)(1−Re⁡χg(x−x0)(θ))) dω(θ)≥e−CE\int \exp(-L \sum_{x\in S'} v(x)(1 - \operatorname{Re} \chi_{g(x-x_0)}(\theta)))\,d\omega(\theta) \ge e^{-CE} for every probability measure vv on S′S'.

Every bounded complex-valued periodic function has almost-everywhere convergence of its symmetric Fourier partial sums along the full sequence.

In the finite dyadic procedure, there are fixed absolute choices of the constants such that for all sufficiently large dd, the procedure is finite in every finite spatial tree, preserves the exact splitting identity, and satisfies the live budgets for kbk_b, tt, F(A+)−F(A+Sp)F(A^+)-F(A+S_p), membership/small-mass norms ∥Bmem∥L2(P)+∥X∗∥L2(P)≤Ch\|B_{\mathrm{mem}}\|_{L^2(\mathbf{P})} + \|X_*\|_{L^2(\mathbf{P})} \le Ch, P(transfer)≤Ch−10\mathbf{P}(\mathrm{transfer}) \le Ch^{-10}, live plateau weights ∑e liveU(pe)αe≤Ch\sum_{e\text{ live}} \mathbf{U}(p_e)\alpha_e \le Ch, and global label Fisher sum ∑live pU(p)αeJp≤Cd\sum_{\text{live } p} \mathbf{U}(p)\alpha_e J_p \le Cd.

Methods

Entropy compression via finite Gibbs variational duality, block thinning to exact iid marginals, and a two-order context refinement search to extract a common phase-alignment region.

Source excerpts

“The optimization below uses finite Gibbs variational duality: subtracting linear constraints from an entropy functional produces a normalized exponential law.” · Show in paper

“The nested partition below uses increasing indices within a subblock to control skips, decreasing subblocks within a macroblock to control completed decreases, and increasing macroblocks to control the remaining skips.” · Show in paper

Thermal labels with densities proportional to e−Ae^{-A} and Griffiths-Ginibre positive-correlation inequalities to bound channel updates, form tightening, and likelihood evidence.

Source excerpts

“The covariance argument below belongs to the Griffiths–Ginibre positive-correlation method [10]. We include the circle calculation, together with the form-order consequences needed for the label updates.” · Show in paper

Broad capture of high frequencies using trigonometric polynomial masks and rounded-coordinate snapshots to bound representative counts.

Source excerpts

“Trim a distance δp from both ends of p ′ , and convolve its indicator on the unit circle with the positive normalized kernel proportional to (sin(πNx)/(N sin(πx)))^{2s}” · Show in paper

“Use a fixed partition of the unit circle into O(Dmax/b6 ) arcs, rounding each χn(Y ), n ∈ D, to a chosen point of its arc with error at most c5b 6/Dmax.” · Show in paper

Delayed transform estimates along plateau paths using predictable sign inequalities, two-step block vector norms, and Fisher checkpoints.

Source excerpts

“We use the standard exponential-supermartingale proof of a maximal inequality, based on Ville’s stopping argument [23, Chapter V, first section, Theorem 1, pp. 84–85].” · Show in paper

“We prove the norm estimate using both signs of a two-step block.” · Show in paper

Translation and scale averaging of delayed Haar shifts (Petermichl-Hytönen dyadic representation) to recover the periodic Hilbert transform and symmetric Fourier partial sums.

Source excerpts

“The averaging follows the dyadic representation method of Petermichl and Hytönen [18, 12]; Section 10 includes the kernel calculation and all limiting steps.” · Show in paper