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.

Latest version

Checks complete

Sync OpenAI Math manuscripts from 4eaee87548250b2682affda7d05d947c05554bae

Committed

All papers

Claims

Findings

The second Kahn–Kalai conjecture

For every n≥2n \ge 2 and every finite simple graph HH with h=e(H)≥1h = e(H) \ge 1 edges and at most nn vertices, pc(n,H)≤min⁡{1,2048e50pE(n,H)(1+log⁡2h)}p_c(n, H) \le \min\{1, 2048\mathrm{e}^{50}p_{\mathrm{E}}(n, H)(1 + \log_2 h)\}, and in particular pc(n,H)≤6144e50pE(n,H)log⁡2np_c(n, H) \le 6144\mathrm{e}^{50}p_{\mathrm{E}}(n, H)\log_2 n.

Let T\mathcal{T} be a disjoint probability tree on a finite set XX with k≥1k \ge 1 levels, where every level is σ\sigma-spread (σ>0\sigma > 0) and its positive integer capacities satisfy ℓi+1≥16ℓi\ell_{i+1} \ge 16\ell_i (1≤i<k1 \le i < k). For r=min⁡{1,16e50σ(1+log⁡2ℓk)}r = \min\{1, 16\mathrm{e}^{50}\sigma(1 + \log_2 \ell_k)\}, PW∼μr(U(P)⊆W for some path P of T)≥23\mathbb{P}_{W \sim \mu_r}(\mathcal{U}(P) \subseteq W \text{ for some path } P \text{ of } \mathcal{T}) \ge \frac{2}{3}.

Given a disjoint probability tree T\mathcal{T} with k≥1k \ge 1 levels, positive integer capacities mj+1≥16mjm_{j+1} \ge 16m_j, and level jj being αj\alpha_j-spread with 0<αj≤a0 < \alpha_j \le a and ρ=e50a<1\rho = \mathrm{e}^{50}a < 1, there is a deterministic procedure that, given W∼μρW \sim \mu_\rho, succeeds with probability at least 3/43/4 and returns a disjoint probability tree T′\mathcal{T}' with the same levels such that level jj has capacity ⌊mj/2⌋\lfloor m_j/2\rfloor and is αj/(1−e−mj)\alpha_j/(1 - \mathrm{e}^{-m_j})-spread, and every path QQ of T′\mathcal{T}' has a path PP of T\mathcal{T} with U(P)⊆W∪U(Q)\mathcal{U}(P) \subseteq W \cup \mathcal{U}(Q).

Suppose ∣Ab∣≤m|A_b| \le m for a positive integer mm, the law ν\nu is aa-spread, and 0<ρ<10 < \rho < 1 with a/ρ≤e−50a/\rho \le \mathrm{e}^{-50}. For W∼μρW \sim \mu_\rho independent of the source, calling the local resampling transition a local failure if Z(Y)<e−10mZ(Y) < \mathrm{e}^{-10m} or ∣T∣>m/2|T| > m/2, we have P(local failure)≤e−9m\mathbb{P}(\text{local failure}) \le \mathrm{e}^{-9m}, and for each fixed target bb and every nonnegative function ff depending only on W∩(X∖Ab)W \cap (X \setminus A_b), E[1{target=b, Z(Y)≥e−10m}f(W)]≤e11mνbEf(W)\mathbb{E}[\mathbf{1}_{\{\text{target}=b,\, Z(Y) \ge \mathrm{e}^{-10m}\}} f(W)] \le \mathrm{e}^{11m}\nu_b \mathbb{E} f(W).

There is a disjoint probability tree on X=([n]2)X = \binom{[n]}{2} whose path unions are copies of HH with its isolated vertices removed, such that every level is σ\sigma-spread (with σ=128pE(n,H)\sigma = 128 p_{\mathrm{E}}(n, H)) and its label sizes ℓ1,…,ℓk\ell_1, \dots, \ell_k satisfy 1≤ℓ11 \le \ell_1, ℓi+1≥16ℓi\ell_{i+1} \ge 16\ell_i, and ℓk≤e(H)\ell_k \le e(H).

Methods

Planted-set and posterior-resampling local transition that replaces a source label AcA_c by a smaller fragment T=Ab∖WT = A_b \setminus W while controlling the change of measure on coordinates outside AbA_b.

Source excerpts

“1. Choose a source c with law ν, and set Y = W ∪ Ac. 2. Choose a target b with conditional probabilities νbρ−|Ab|1{Ab⊆Y } / Z(Y ), Z(Y ) = ∑d∈B νdρ−|Ad|1{Ad⊆Y }. (2.2) 3. Retain the fragment T = Ab \ W.” · Show in paper

Simultaneous tree reduction using a single random set WW at all nodes, controlled by a recursively defined node potential Dv(W)D_v(W) and disjointness of labels along root-to-leaf paths.

Source excerpts

“Set Dv(W) = 0 at leaves and, at a node with outgoing level s, define recursively Dv(W) = (dv(W) + ∑b∈Ch(v) tvb(W)Db(W)) / ϵs. (3.4)” · Show in paper

Graph-specific spread-link extraction via iterative maximization of the normalized containment ratio Rσ(I,S)=σ−∣S∣P(S0⊆I)R_\sigma(I, S) = \sigma^{-|S|}\mathbb{P}(S_0 \subseteq \mathbf{I}) to build a geometrically growing disjoint probability tree.

Source excerpts

“For a current nonempty edge set I ⊆ H, choose an edge subset S ⊆ I maximizing Rσ(I, S) = σ−|S|P(S0 ⊆ I), (4.2) where S0 is any fixed copy of S on [n].” · Show in paper