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 slowing down in the Sherrington–Kirkpatrick model

At the critical inverse temperature β=1\beta = 1 for zero-field Gaussian Sherrington–Kirkpatrick heat-bath dynamics, for every deterministic sequence tn≥0t_n \ge 0 with tn=o(n2/3)t_n = o(n^{2/3}), μW{x:dtnct(W,x)>1/4}→P1\mu_W\{x : d^{\mathrm{ct}}_{t_n}(W, x) > 1/4\} \xrightarrow{\mathbb{P}} 1, and for every deterministic integer sequence kn≥0k_n \ge 0 with kn=o(n5/3)k_n = o(n^{5/3}), μW{x:dkndt(W,x)>1/4}→P1\mu_W\{x : d^{\mathrm{dt}}_{k_n}(W, x) > 1/4\} \xrightarrow{\mathbb{P}} 1, with convergence in probability over the disorder.

For every positive deterministic sequence Mn→∞M_n \to \infty, with probability tending to one both ∥Σ∥\|\Sigma\| and Rlin(μW)\mathcal{R}_{\mathrm{lin}}(\mu_W) belong to [n2/3/Mn,Mnn2/3][n^{2/3}/M_n, M_n n^{2/3}].

For every fixed ϵ>0\epsilon > 0, P{n2/3−ϵ≤tmix≤eϵn}→1\mathbb{P}\{n^{2/3-\epsilon} \le t_{\mathrm{mix}} \le e^{\epsilon n}\} \to 1 and P{n5/3−ϵ≤kmix≤eϵn}→1\mathbb{P}\{n^{5/3-\epsilon} \le k_{\mathrm{mix}} \le e^{\epsilon n}\} \to 1.

On an eigenvalue-only event Gn(Ln)\mathcal{G}_n(L_n) with P(Gn(Ln)c)≤Cexp⁡(−cLn1/4)\mathbb{P}(\mathcal{G}_n(L_n)^c) \le C \exp(-c L_n^{1/4}), the top three GOE eigenvalues satisfy λj=2+o(s0)\lambda_j = 2 + o(s_0) (1≤j≤31 \le j \le 3), max⁡i∣λi∣≤3\max_i |\lambda_i| \le 3, and the first two resolvent traces at s0≤s≤Ms_0 \le s \le M satisfy 1n∑i(2+s−λi)−1=R(s)+o(s)\frac{1}{n}\sum_i (2+s-\lambda_i)^{-1} = R(s) + o(\sqrt{s}) and 1n∑i(2+s−λi)−2≤CM/s\frac{1}{n}\sum_i (2+s-\lambda_i)^{-2} \le C_M / \sqrt{s}.

For the augmented GOE matrix WW and a unit top eigenvector vv, for every positive deterministic sequence bn→0b_n \to 0, μ{x:∣vTx∣≤bnn1/3}→0\mu\{x : |v^T x| \le b_n n^{1/3}\} \to 0 in probability.

For Qn=Eμ⊗2q(x,y)2Q_n = \mathbb{E}_{\mu^{\otimes 2}} q(x, y)^2, the sequence n2/3Qnn^{2/3} Q_n is bounded in probability.

For every fixed ϵ>0\epsilon > 0, P{tmix≤eϵn,kmix≤eϵn}→1\mathbb{P}\{t_{\mathrm{mix}} \le e^{\epsilon n}, k_{\mathrm{mix}} \le e^{\epsilon n}\} \to 1.

Methods

Gaussian diagonal augmentation: adding an independent N(0,2/n)N(0, 2/n) diagonal to the zero-diagonal SK coupling matrix to form a GOE matrix without changing the Gibbs measure or heat-bath dynamics, then integrating out the auxiliary diagonal.

Source excerpts

“Following the augmentation used by Comets [10, Equation (1.1)], add an independent Gaussian diagonal to the coupling matrix. Its energy is constant on the cube, so the Gibbs law and both chains stay the same, while the augmented matrix is GOE.” · Show in paper

“In the realized-start proof we use that direction to test an intrinsic total-variation distance, and only then integrate out the diagonal.” · Show in paper

GOE tridiagonalization combined with dyadic block summation-by-parts quadratic-form estimates and a deterministic Hermite resolvent identity to control leading eigenvalues and resolvent traces at a slowly diverging edge scale s0=Lnn−2/3s_0 = L_n n^{-2/3}.

Source excerpts

“The proof has three parts. Tridiagonalization isolates independent noise; a simultaneous quadratic-form estimate compares that noise with a deterministic confinement energy; and a Hermite identity evaluates the deterministic resolvent.” · Show in paper

Exact Gaussian conditioning on the sphere: representing the quadratically tilted spherical measure σΛ,a\sigma_{\Lambda, a} as independent centered Gaussian coordinates conditioned on their squared norm, analyzed via radial-density estimates.

Source excerpts

“Conditioning independent centered Gaussian coordinates on their squared radius gives an exact representation of a quadratic tilt of uniform sphere measure.” · Show in paper

Haar averaging over the orthogonal group with a cubic variational deficit for large overlaps and disjoint outward-bin monotonicity comparison for central overlaps to transfer spherical estimates to the Ising cube.

Source excerpts

“Let O be Haar distributed on the orthogonal group, independent of the ordered spectrum, and put W = OΛOT.” · Show in paper

“We will use a variational estimate for large overlaps and a direct comparison of the unweighted overlap laws near zero.” · Show in paper

Stationary mean-square increment bounds and half-space sign-persistence testing for realized initial configurations, paired with spin-deletion cavity reweighting for linear inverse Rayleigh quotients.

Source excerpts

“Choose a projection threshold above its root-mean-square increment scale, use its sign as a test for each fixed starting state, and average only the resulting intrinsic bad-state indicator.” · Show in paper

“We use the standard spin-deletion cavity reweighting; see Talagrand [29] and Chen [9].” · Show in paper

Poincaré inequality comparison with a fixed subcritical ($eta < 1$) SK model using least-squares characterizations of variance and Dirichlet forms.

Source excerpts

“For the upper bound we combine the high-probability norm estimate of Proposition 3.1 with the following fixed-temperature, zero-field functional inequality.” · Show in paper