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

Randomized quasipolynomial-time mean-payoff games

There is a uniform randomized algorithm that computes the complete zero-threshold winning set U={v∈V:there is a Max strategy σ such that for every Min strategy τ, MPw(πv,σ,τ)≥0}U = \{v \in V : \text{there is a Max strategy } \sigma \text{ such that for every Min strategy } \tau,\ \mathrm{MP}_w(\pi_{v,\sigma,\tau}) \ge 0\} of a finite mean-payoff game with arbitrary signed integer edge weights encoded in binary, correctly with probability at least 7/87/8, using at most 2C(log⁡2(L+2))22^{C(\log_2(L+2))^2} bit operations on every random tape for an absolute constant C>0C > 0, where LL is the total explicit binary input length.

A polynomial-time check certifies the winning regions and positional strategies for both players or reports failure using 2O((log⁡2(L+2))2)2^{O((\log_2(L+2))^2)} bit operations on every tape and succeeding with probability at least 7/87/8; repeating independent trials until a certificate is accepted terminates almost surely, never returns an incorrect answer, and has expected bit cost 2O((log⁡2(L+2))2)2^{O((\log_2(L+2))^2)}.

For the output pp of `Solve` on an allowed integer box [a,b][a, b] with positive integer reference width B≤B∗B \le B_* and positive rational support weights and budgets, each fixed admissible subsolution yy satisfies Pr⁡(p≥y)≥1−1/16\Pr(p \ge y) \ge 1 - 1/16 and separately each fixed admissible supersolution zz satisfies Pr⁡(p≤z)≥1−1/16\Pr(p \le z) \ge 1 - 1/16; consequently, on [0,B∗][0, B_*] with unit weights and budgets k=m=nk = m = n, `Solve` recovers the exact box solution xx with probability at least 7/87/8.

The algorithm has a uniform implementation in a standard bit-operation model such that, for every input of bit length LL and every random tape, the running time is at most 2C(log⁡2(L+2))22^{C(\log_2(L+2))^2} for an absolute constant C>0C > 0, including preprocessing, every recursive execution, intermediate arithmetic, and random-bit generation.

If xx solves the constant box [0,B∗][0, B_*] with B∗=8nWB_* = 8nW, then the original game's exact zero-threshold winning set is U={i∈V:xi>B∗/2}U = \{i \in V : x_i > B_*/2\}, with a positional Max strategy guaranteeing original mean payoff at least zero from UU and a positional Min strategy ensuring upper limiting average at most −1/n-1/n outside UU against arbitrary history-dependent strategies.

For every permitted `Solve` input, the execution and all its subroutine executions terminate on every random tape, with outputs in their respective input boxes, at most 8ns8ns tightening updates (16ns+216ns + 2 calls to `Ask`), at most K=8s(g+1)+1K = 8s(g+1)+1 solver invocations along a main continuation chain, and at most g+1g+1 orientation invocations per `Orient` trial returning at most 2g+22g+2 sets of μ\mu-measure at most 1/21/2.

Under the lower-rank comparison guarantees, `Orient` (and `OrientPlus` with failure probability at most ε=2−J\varepsilon = 2^{-J}) satisfies Pr⁡(χ=Y or ∃H∈H:α(Sy∩H)>ηk)≥1−1/16\Pr(\chi = Y \text{ or } \exists H \in \mathcal{H} : \alpha(S_y \cap H) > \eta k) \ge 1 - 1/16 for every fixed admissible subsolution yy with α(Iy)>ηk\alpha(I_y) > \eta k, and separately Pr⁡(χ=Z or ∃H∈H:β(Sz∩H)>ηm)≥1−1/16\Pr(\chi = Z \text{ or } \exists H \in \mathcal{H} : \beta(S_z \cap H) > \eta m) \ge 1 - 1/16 for every fixed admissible supersolution zz with β(Iz)>ηm\beta(I_z) > \eta m.

The minimization variable-selection rule printed in Truffet's arXiv:2603.26423v4 (dated August 29, 2026) loses the optimum on the explicit instance min⁡x1\min x_1 subject to max⁡(x1,x2)≥0\max(x_1, x_2) \ge 0, x1≥max⁡(x2+1,−2)x_1 \ge \max(x_2 + 1, -2), where the true minimum is 00 but the rule produces 11.

Methods

Integer cycle perturbation t(e)=(n+1)w(e)+1t(e) = (n+1)w(e) + 1 and clipped box operator Ta,b(p)=clip[a,b]F(p)T_{a,b}(p) = \mathrm{clip}_{[a,b]} F(p) on finite integer boxes [a,b][a, b].

Source excerpts

“Apply the integer cycle perturbation used by Björklund and Vorobyov [4, Section 9], and define t(e) = (n + 1)w(e) + 1, W = max {1, max_{e∈E} |t(e)|}, B∗ = 8nW. (2.1)” · Show in paper

“It is exactly the unique fixed point of Ta,b(p) = clip[a,b] F(p).” · Show in paper

Weighted support budgets (α,β,k,m)(\alpha, \beta, k, m) and scalar rank parameter R=km∑i∈V1/(αiβi)R = km \sum_{i \in V} 1/(\alpha_i \beta_i) with rank-reducing reweighting and budget scaling.

Source excerpts

“Define the positive rational parameter R and a probability measure µ on V by R = km \sum_{i∈V} 1/(α_i β_i), µ(S) = (\sum_{i∈S} 1/(α_i β_i)) / (\sum_{i∈V} 1/(α_i β_i)). (3.2)” · Show in paper

Recursive solver (`Solve`) combining a tightening loop (`Ask`), randomized orientation via moving centers (`Orient`, `OrientPlus`), and an unamplified main continuation.

Source excerpts

“Procedure 4.3 (The solver). Given (a, b; B, α, β, k, m), compute R, µ from Equation (3.2) and execute these steps.” · Show in paper

“Procedure 4.4 (Orientation). For (a, b; B, α, β, k, m) with R ≥ 1, compute µ and proceed as follows.” · Show in paper

Polynomial-time potential certification check p=clip[0,B∗]F(p)p = \mathrm{clip}_{[0,B_*]} F(p) and positional strategy extraction from the two potential bands.

Source excerpts

“Run the root solver to obtain p ∈ [0, B∗] and check p = clip[0,B∗] F(p). Accept precisely when this equality holds.” · Show in paper