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

Critical mixing in the Sherrington–Kirkpatrick model

For every fixed ε∈(0,1/2)\varepsilon \in (0, 1/2), in probability over the Gaussian disorder at the critical inverse temperature β=1\beta = 1 in the zero-field Gaussian Sherrington–Kirkpatrick model, the worst-start total-variation mixing time for single-site heat-bath dynamics is tmix,J(ε)=n2/3+o(1)t_{\mathrm{mix},J}(\varepsilon) = n^{2/3+o(1)} when every site has rate one, or kmix,J(ε)=n5/3+o(1)k_{\mathrm{mix},J}(\varepsilon) = n^{5/3+o(1)} attempted uniform-site updates, and the relaxation time and inverse classical logarithmic Sobolev constant satisfy RJ=n2/3+o(1)R_J = n^{2/3+o(1)} and CLS,J=n2/3+o(1)C_{\mathrm{LS},J} = n^{2/3+o(1)}.

For every fixed ω>0\omega > 0, there exist cω>0c_\omega > 0 and disorder events of probability tending to one on which, simultaneously for every nonzero real function ff with p=μJ(f≠0)≤1/2p = \mu_J(f \neq 0) \le 1/2, DJ(f)≥cωn−2/3−ωlog⁡(1/p)μJ(f2)\mathcal{D}_J(f) \ge c_\omega n^{-2/3-\omega} \log(1/p) \mu_J(f^2), and in particular the spectral gap for rate-one-per-site heat-bath dynamics is at least cω(log⁡2)n−2/3−ωc_\omega (\log 2) n^{-2/3-\omega}.

For every fixed ρ>0\rho > 0, with probability tending to one over the Gaussian couplings, the critical zero-field Gibbs measure satisfies the classical logarithmic Sobolev inequality Entμ(f2)≤n2/3+ρD0(f)\mathrm{Ent}_\mu(f^2) \le n^{2/3+\rho} \mathcal{D}_0(f) for every f:{−1,1}n→Rf : \{-1, 1\}^n \to \mathbb{R}.

For every realization of the disorder, the entropy-production inverse constant satisfies RJ/2≤CEP,J≤CLS,J/4R_J / 2 \le C_{\mathrm{EP},J} \le C_{\mathrm{LS},J} / 4; consequently, for the critical zero-field Gaussian model, CEP,J=n2/3+o(1)C_{\mathrm{EP},J} = n^{2/3+o(1)} in disorder probability, and KL(νetLJ∥μJ)≤e−t/CEP,JKL(ν∥μJ)≤e−4t/CLS,JKL(ν∥μJ)\mathrm{KL}(\nu e^{t\mathcal{L}_J} \| \mu_J) \le e^{-t/C_{\mathrm{EP},J}} \mathrm{KL}(\nu \| \mu_J) \le e^{-4t/C_{\mathrm{LS},J}} \mathrm{KL}(\nu \| \mu_J) for every initial probability law ν\nu on the cube and every t≥0t \ge 0.

Along the isotropic observation path, for every sufficiently small fixed ϵ>0\epsilon > 0, there are s0,a>0s_0, a > 0 and Klog⁡<∞K_{\log} < \infty such that ∥Σs∥≤(1+ϵ)/s\|\Sigma_s\| \le (1+\epsilon)/s at every deterministic time 0<s≤s00 < s \le s_0 with ns3/2≥Klog⁡log⁡nn s^{3/2} \ge K_{\log} \log n, outside an integrated planted exceptional set intersected with G\mathcal{G} of mass at most e−ans3/2e^{-a n s^{3/2}}; below the logarithmic threshold ns3/2≤Klog⁡log⁡nn s^{3/2} \le K_{\log} \log n, ∥Σs∥≤Cn2/3ℓC\|\Sigma_s\| \le C n^{2/3} \ell^C outside an integrated planted event intersected with G\mathcal{G} of mass at most n−Bn^{-B} for any fixed B>0B > 0.

For any fixed finite field radius L0L_0, there is C=C(L0)<∞C = C(L_0) < \infty such that for every ϵ>0\epsilon > 0 there are constants ptop>0p_{\mathrm{top}} > 0 and s0>0s_0 > 0 for which, with probability tending to one, simultaneously for every pmin⁡(n)≤p≤ptopp_{\min}(n) \le p \le p_{\mathrm{top}} with pmin⁡(n)=(log⁡n/n)1/3p_{\min}(n) = (\sqrt{\log n}/n)^{1/3}, Pp∇2ϕp(h)Pp⪯1+ϵp2PpP_p \nabla^2 \phi_p(h) P_p \preceq \frac{1+\epsilon}{p^2} P_p for all h∈Pph \in P_p with ∣αp(h)∣≤s0p2|\alpha_p(h)| \le s_0 p^2, and Pp∇2ϕp(h)Pp⪯Cp2PpP_p \nabla^2 \phi_p(h) P_p \preceq \frac{C}{p^2} P_p for all h∈Pph \in P_p with ∥h∥≤L0p2Mp\|h\| \le L_0 p^2 M_p.

Methods

Gaussian stochastic localization with isotropic observations Ys=sX+BsY_s = sX + B_s and GOE diagonal completion Jii∼N(0,2/n)J_{ii} \sim N(0, 2/n) to track posterior covariance and relative entropy.

Source excerpts

“We use the orthogonal invariance of a GOE completion. Adjoin independent Jii ∼ N(0, 2/n) to J◦ and call the completed matrix J. The added energy is 1/2 ∑i Jii because x2i = 1; it is constant on the cube.” · Show in paper

“The isotropic experiment is Ys = sX + Bs, with Bs standard Brownian motion. Its innovation Is := Ys − ∫0s mu du is Brownian in the observation filtration.” · Show in paper

Conditional Kac–Rice root counting combined with root-preserving rotation orbits and signed cancellation between spin and Gaussian replica integrals.

Source excerpts

“The Kac–Rice formula in Section 8 expresses planted expectations of sums over regular solutions as an integral against the observation density and the absolute Jacobian determinant.” · Show in paper

“The next comparison transfers continuous-prior estimates to spins on most orientations of each retained root orbit. Its essential quantitative point is the signed cancellation between two spin and two Gaussian replica integrals.” · Show in paper

Spectral-cap observation scheme with precision Bp=(p2−d)+B_p = (p^2 - d)_+ and inductive transport of a radial covariance deficit using Brascamp–Lieb upper bounds, variational Fisher lower bounds, and Haar transverse compression.

Source excerpts

“Let d = 2I − J. At scale p, observe with precision Bp = (p2 − d)+ and write Jp = J − Bp.” · Show in paper

“The Brascamp–Lieb upper bound retains the radial correction through the inverse in (16.14); the variational Fisher lower bound controls the opposite side and the mixed block. Averaging the rank-one correction regenerates the radial deficit by (16.20). Meanwhile the still hidden eigenframe inside the flat parent space is Haar. Its compression makes the transverse error small enough at the child scale.” · Show in paper

Small-field Thouless–Anderson–Palmer (TAP) expansion and auxiliary potential initialization with finite-step gradient error contraction.

Source excerpts

“We obtain this correction from the small-field TAP expansion, initially for an auxiliary potential. The auxiliary potential differs from the true cap potential by a bounded Euclidean gradient error. A finite number of cap steps contracts this error, after which a covariance comparison permits us to use the true potential at every remaining scale.” · Show in paper

Support-profile truncation and hypercontractive smoothing to convert functional inequalities into worst-start total-variation mixing bounds.

Source excerpts

“The support-profile proof is the finite-state truncation mechanism of [22, Theorem 1.1 and Lemma 2.1]. The classical logarithmic Sobolev proof uses the hypercontractive smoothing of [15, Theorems 3.5 and 3.7], followed by spectral-gap decay.” · Show in paper