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

Generalized Star Height at Most Three

For every finite alphabet Σ\Sigma and every regular language L⊆Σ∗L \subseteq \Sigma^*, the generalized star height hΣ(L)h_\Sigma(L) (with complement taken in Σ∗\Sigma^*) is at most 33.

If D′⊆D⊆ED' \subseteq D \subseteq E where EE is a prefix code, Px,QyP_x, Q_y satisfy soundness on D∗D^* and completeness on D∖D′D \setminus D', D,Px,QyD, P_x, Q_y have height at most 11, and the skip graph languages Wx,yW_{x,y} on (D′)∗(D')^* have height at most H≥0H \ge 0, then the graph languages on D∗D^* have height at most max⁡{2,1+H}\max\{2, 1 + H\}.

Boolean affine recovery expresses the graph language for the product of linear parts LeL_e of affine episode updates Ge(v)=vLe+beG_e(v) = v L_e + b_e on D∗D^* as ⋃b∈S(Rx,y+b∩R0,b)\bigcup_{b \in S} (R_{x, y+b} \cap R_{0, b}), preserving any common height bound on the affine graph languages Rx,zR_{x,z}.

For every nonempty finite set AA and finite vector space SS, there exists an integer l≥1l \ge 1 depending only on ∣A∣|A| and ∣S∣|S| such that for every function f:Al→Endlin(S)f: A^l \to \text{End}_{\text{lin}}(S), there is βf:Al→S\beta_f: A^l \to S where for every a∈Ala \in A^l and x∈Sx \in S, some coordinate i∈{1,…,l}i \in \{1, \dots, l\} makes b↦xf(a1,…,ai−1,b,ai+1,…,al)+βf(a1,…,ai−1,b,ai+1,…,al)b \mapsto xf(a_1, \dots, a_{i-1}, b, a_{i+1}, \dots, a_l) + \beta_f(a_1, \dots, a_{i-1}, b, a_{i+1}, \dots, a_l) constant on AA.

For any finite nonempty sets Λ\Lambda (with m=∣Λ∣m = |\Lambda|) and P\mathcal{P}, there exist a finite-dimensional torus T\mathcal{T}, sets Iα,Jα,W⊆TI_\alpha, J_\alpha, W \subseteq \mathcal{T}, points vp∈Tv_p \in \mathcal{T}, and constants ρ,η>0\rho, \eta > 0 such that: (i) for each z∈Tz \in \mathcal{T}, Ez={(α,λ)∈Λ2:z∈Iα+Jλ}\mathcal{E}_z = \{(\alpha, \lambda) \in \Lambda^2 : z \in I_\alpha + J_\lambda\} is either contained in the diagonal or consists of one off-diagonal pair; (ii) if z∉Wz \notin W, Iα∩(z−Jα)I_\alpha \cap (z - J_\alpha) contains a ball of radius ρ\rho for every α\alpha; and (iii) for any η\eta-perturbation (v~p)(\tilde{v}_p) and any a∈Ta \in \mathcal{T}, at most 107log⁡(2m)10^7 \log(2m) distinct pairs (p,p′)(p, p') satisfy a+v~p′−v~p∈Wa + \tilde{v}_{p'} - \tilde{v}_p \in W.

For the total affine updates FeF_e on VV (with linear part Ts,bT_{s,b}) and D⊆ED \subseteq E defined in (42), the languages Px,QyP_x, Q_y satisfy: (1) on any word in E∗E^*, a match of PxQyP_x Q_y beginning at the start of the word consumes exactly its first episode ee and satisfies Fe(x)=yF_e(x) = y; (2) for every e∈E∖De \in E \setminus D and every x∈Vx \in V, some cut of ee gives a match of PxQFe(x)P_x Q_{F_e(x)}; and there is no match on the empty word.

For each complete episode, the number of distinct-slot pairs (p,p′)∈P2(p, p') \in \mathcal{P}^2 (p≠p′p \ne p') satisfying Z(t+p−p′)Z(t + p - p') or A∗(t+p−p′,Qt+p−p′)A_*(t + p - p', Q_{t+p-p'}) is strictly less than μ\mu.

When paired with a genuine stencil prefix [s,k)[s, k), k=s+pk = s + p, of a first episode e∈De \in D in a word of D∗D^*, if the suffix decoder test succeeds, it decodes the actual slot pp and consumes exactly to the actual endpoint bb; conversely, on the actual suffixes of a fixed e∈De \in D, fewer than μ\mu actual slots fail the decoder, so each kind has a copy at which decoding succeeds.

For the total affine updates GeG_e on UU (with linear part F^e\widehat{F}_e) on DD, the languages P~x,Q~y\widetilde{P}_x, \widetilde{Q}_y satisfy: (1) on a word in D∗D^*, every match of P~xQ~y\widetilde{P}_x \widetilde{Q}_y from its beginning consumes exactly the first episode ee and has Ge(x)=yG_e(x) = y; (2) every e∈De \in D admits such a split for every input x∈Ux \in U, with empty skip domain.

The episode language EE, its skipped sublanguage DD, and every suffix language in the outer and inner splits have generalized star height at most 11 over Σ\Sigma, while every prefix language in these splits has height 00.

Methods

Prefix-code episode factorization and the split identity converting sound and complete independent prefix/suffix tests into bounded-height graph language expressions.

Source excerpts

“Lemma 2.1 (Split identity). Let D′ ⊆ D ⊆ E, where E is a prefix code, and give each e ∈ D a deterministic total update Ge of a finite set S.” · Show in paper

Homogeneous linear lift and Boolean affine recovery to remove affine translations from graph languages without increasing generalized star height.

Source excerpts

“Lemma 2.2 (Boolean affine recovery). Let S be a finite vector space and let the episode updates on D ⊆ E be Ge(v) = vLe + be, with Le ∈ Endlin(S) and be ∈ S determined by e.” · Show in paper

“Lemma 2.3 (Homogeneous lift). For a finite vector space S over a finite field F, the map F(v) = vL + b has the linear lift on S ⊕ F Fb(v, c) = (vL + cb, c).” · Show in paper

Shift-table construction via the Erdős–Lovász finite dependency lemma to omit one unknown tuple coordinate using a deterministic affine correction.

Source excerpts

“We use the following form of the finite dependency argument of Erdős and Lovász [6, Section 2] and include its proof.” · Show in paper

Robust tag clock on a finite-dimensional torus using left/right partitions, Type 1/Type 2 batches, and forest-dedicated coordinates for roster values.

Source excerpts

“First coordinates will exclude all competitors of an unequal edge except its reversal. Type batches will exclude the reversal too. Each block also permits a diagonal decomposition away from an explicitly specified obstruction.” · Show in paper

Finite-window local marker priority rule and periodic continuation overlap to force either nearby marker boundaries or a common short-period repetition.

Source excerpts

“This is the finite-word marker principle associated with Krieger’s marker lemma [7, Lemma 2].” · Show in paper

Compilation on finitely many periodic templates aciba c^i b to obtain generalized star height at most one over the original alphabet Σ\Sigma.

Source excerpts

“Lemma 10.1 (Compilation on periodic templates). Fix a finite alphabet Σ, a finite monoid M, and a morphism T : Σ∗ → M. Consider finitely many templates acib, i ≥ 0” · Show in paper