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

The Mézard–Parisi formula for diluted spin glasses

Under hypotheses (2.1)–(2.4), the thermodynamic limit exists and satisfies lim⁡N→∞FN=inf⁡r≥0Φr\lim_{N\to\infty} F_N = \inf_{r\ge 0} \Phi_r, where the right-hand side is an infimum over finite hierarchies.

Under the hypotheses of Theorem 2.1, FN≤Br(ζ,m)F_N \le \mathcal{B}_r(\zeta, \mathbf{m}) for every NN and every admissible finite-level trial law, and consequently FN≤inf⁡r≥0ΦrF_N \le \inf_{r\ge 0} \Phi_r.

For every bounded permutation-invariant interaction law and bounded field law, with the functionals defined in Section 2, lim inf⁡N→∞FN≥inf⁡r≥0Φr\liminf_{N\to\infty} F_N \ge \inf_{r\ge 0} \Phi_r.

For every fixed labeled topology SS, lim⁡L→∞lim sup⁡N→∞AL,ηVar⁡(RS)=0\lim_{L\to\infty} \limsup_{N\to\infty} \mathcal{A}_{L,\eta} \operatorname{Var}(R_S) = 0 (η>0\eta > 0), where only the b(S)b(S) branching depths are averaged, and for a singleton Var⁡(RS)→0\operatorname{Var}(R_S) \to 0 as N→∞N \to \infty at every fixed LL.

For the perturbation family of Section 6.1, along every sequence from Lemma 6.2, ∑E∈E(T,a∗,C)cEE[fRE]−Ef∑E∈E({a∗},a∗,C)cEE[RE]→0\sum_{E \in \mathfrak{E}(\mathcal{T}, a_*, \mathcal{C})} c_E \mathbb{E}[f R_E] - \mathbb{E}f \sum_{E \in \mathfrak{E}(\{a_*\}, a_*, \mathcal{C})} c_E \mathbb{E}[R_E] \to 0 for every fixed old tree, either kind of anchor, every bounded polynomial ff in its multioverlaps, every finite number of ordered added colors, and every prescribed constraint C\mathcal{C}.

The variational formula (2.13) holds for the Viana–Bray and diluted even-spin models with symmetric couplings (Corollary 10.1) and for soft random even-KK SAT with nonnegative clause weights (Corollary 10.2), and for each model the zero-temperature limit g=lim⁡N→∞N−1Emax⁡σEN(σ)g = \lim_{N\to\infty} N^{-1} \mathbb{E}\max_\sigma \mathcal{E}_N(\sigma) exists and satisfies g=lim⁡β→∞β−1inf⁡r≥0inf⁡ζ,mBr,β(ζ,m)g = \lim_{\beta\to\infty} \beta^{-1} \inf_{r\ge 0} \inf_{\zeta,\mathbf{m}} \mathcal{B}_{r,\beta}(\zeta,\mathbf{m}).

Methods

Finite-tree power-mean insertion calculus (Lemma 3.1), logarithmic expansion (Lemma 3.2), and canceled-anchor expansion (Lemma 3.3) on finite sequences of arbitrary probability kernels.

Source excerpts

“Lemma 3.1 (Insertion calculus). Suppose a leaf partition function is iterated by power means, with exponents m_1, \dots, mL−1m_{L-1}, and K_1, \dots, KL−1K_{L-1} are its tilted kernels. Let K_L be its terminal Gibbs kernel. Multiplying each spin weight by A changes the iterated root logarithm by \log Y_0, where (3.3) uses the kernels before insertion.” · Show in paper

“Lemma 3.3 (Canceled-anchor expansion). Let H be a bounded function of the variables on the old tree, fixed as the insertion parameters vary. For A_t = 1 + \sum_{j=1}^k t_j D_j, with bounded real path functions D_j, one has \partial_{t_1} \cdots \partial_{t_k} E_{\mathcal{T}}^{A_t} \frac{H}{A_ta∗t^{a_*}} \Big|_{t=0} = \sum_E c_E E_{\mathcal{T}_E} \left[ H \prod_{j=1}^k D_jaj(E)j^{a_j(E)} \right]. (3.11)” · Show in paper

Guerra–Franz–Leone–Panchenko–Talagrand interpolation with a constant shift c=θ(1,…,1)c = \theta(1,\dots,1) and λ\lambda-regularization to prove the upper bound without integrability of log⁡a\log a.

Source excerpts

“Normalization. We remove a before taking expectations, so no integrability of \log a is needed. Put D_* = b \prod_\ell f_\ell(1) and c = \theta(1, \dots, 1). Subtract c from each insertion increment.” · Show in paper

Sublinear Poisson single-site perturbations with two parameters (tν,uν)(t_\nu, u_\nu), Efron–Stein variance bounds, and convex-function derivative concentration to select deterministic parameter vectors.

Source excerpts

“The two parameters have different roles: differentiation in u produces a concentrated score, while the values of t probe an insertion's Taylor expansion. Taking u smaller than every fixed power of t will force each limiting Taylor coefficient to vanish.” · Show in paper

“Lemma 6.2 (Deterministic parameter selection). For every fixed L \ge 2 and \varepsilon > 0, there are sizes N_j \uparrow \infty and deterministic vectors \mathbf{u}_j such that (i) \Delta_{N_j}(\mathbf{u}_j) \le \underline{F} + 2\varepsilon; (ii) for every fixed type \nu, eν,Nje_{\nu,N_j}(\mathbf{u}_j) \longrightarrow 0 in the reservoir; (iii) in the reservoir, the expectations of every finite product of multioverlaps on every fixed finite prescribed tree of depth L converge along this sequence.” · Show in paper

Shifted-copy induction and depth-averaged decorrelation using Lemma 3.5 (charging fresh unary splits) and Lemma 7.3 (averaged martingale continuity in depth).

Source excerpts

“The two copies have aligned depths but distinct branching vertices: each vertex supplies its own factor \delta. This is the reason for the shift. The full signed coefficient has an additional factor -\delta from separation at d; we divide by J and leave that factor in the identity. Equation (7.20) then controls every projection error after division.” · Show in paper

Recursive law encoding of a uniformly chosen reservoir site's finite hierarchy into a deterministic trial law ζL,N∈ML\zeta_{L,N} \in \mathcal{M}_L and empirical-pattern comparison via total variation.

Source excerpts

“Lemma 8.1 (Recursive law encoding). For each fixed N, L \ge 2, and selected parameter vector, with the reservoir root disorder averaged, there is a deterministic trial law \zeta_{L,N} \in \mathcal{M}_L whose single-label spin process has the same expected spin products on every prescribed tree as a uniformly chosen site of the reservoir.” · Show in paper