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

Constant-factor hardness of Min-UnCut

For every fixed integer K≥2K \ge 2, there is a deterministic polynomial-time many-one reduction mapping a 3SAT formula φ\varphi to a pair (G,k)(G, k), where GG is an explicit finite simple undirected unweighted graph and k≥1k \ge 1 is a binary-encoded integer, such that φ\varphi satisfiable   ⟹  OPTuncut(G)≤k\implies \mathrm{OPT}_{\mathrm{uncut}}(G) \le k and φ\varphi unsatisfiable   ⟹  OPTuncut(G)>Kk\implies \mathrm{OPT}_{\mathrm{uncut}}(G) > Kk; consequently, approximating Min-UnCut within any fixed factor C>1C > 1 is NP-hard.

For every fixed real C>1C > 1, minimum 2CNF clause deletion is NP-hard to approximate within factor CC, and a deterministic polynomial-time CC-approximation would imply P=NP\mathrm{P} = \mathrm{NP}.

For every rational J≥1J \ge 1, there are effectively selectable integers m,n≥2m, n \ge 2 and positive rational constants σ,δ,η,p,θ<1\sigma, \delta, \eta, p, \theta < 1 with η≤σ\eta \le \sigma, depending only on JJ, such that for any finite weighted family of nonempty affine alphabets and folded bit proofs satisfying εj≤Jbj\varepsilon_j \le J b_j (j=1,2,3j = 1, 2, 3), the one-row conditional function b↦hx(B[Bx−ii←b],c)b \mapsto h_x(B[B_{x_{-i}}^i \leftarrow b], c) has an odd Fourier coefficient of absolute value at least θ\theta with probability at least pp.

Under the inner test assumptions εj≤Jbj\varepsilon_j \le J b_j (j=1,2,3j=1,2,3) and the parameter choices (3.7)–(3.8), the low-degree projected field L=ΠBΠchL = \Pi_B \Pi_c h has Fourier row degree at most RmRm, Gaussian degree at most ss, satisfies E⟨L,h⟩x=E∥L∥2,x2≥1/2\mathbb{E}\langle L, h\rangle_x = \mathbb{E}\|L\|_{2,x}^2 \ge 1/2, and obeys EdX(L)≥1/(4T)\mathbb{E}\mathfrak{d}_X(L) \ge 1/(4T) uniformly in m,nm, n and the affine alphabets.

If no assignment satisfies more than 3/43/4 of the equation list, then for every q≥0q \ge 0 and t≥t0≥1t \ge t_0 \ge 1, the value of the outer game with qq affine hints is at most (1−2−q(1−c))t0(1 - 2^{-q}(1 - c))^{t_0}, and an effective choice of t0t_0 makes this bound strictly less than any prescribed rational s>0s > 0 simultaneously for all t≥t0t \ge t_0.

For any actual first question UU, integers d≥1d \ge 1 and t≥t0≥1t \ge t_0 \ge 1, and L1,…,LdL_1, \dots, L_d drawn independently and uniformly from D(AV)D(A_V), the total variation distance between L((πe∗L1,…,πe∗Ld)∣U)\mathcal{L}((\pi_e^* L_1, \dots, \pi_e^* L_d) \mid U) and Unif(D(AU)d)\mathrm{Unif}(D(A_U)^d) is at most 12(24dt0−1)t02/t\frac{1}{2}\sqrt{(2^{4dt_0} - 1)t_0^2 / t}, uniformly in UU and the equation list.

For every fixed integer J≥20J \ge 20, effective choices of constants give a four-test comparison system constructed from the uniform equation list in Theorem 2.2 (for a fixed completeness error ζ>0\zeta > 0) such that in the YES case there are proof tables of cost at most 44, and in the NO case every collection of proof tables has cost greater than JJ.

There is a deterministic construction, with all preliminary work depending only on KK, of a finite rational replacement for the four tests such that ∣Costfin(f)−Cost∞(f)∣<1/2|\mathrm{Cost}_{\mathrm{fin}}(f) - \mathrm{Cost}_{\infty}(f)| < 1/2 for every collection of folded proof tables ff, yielding minimum cost <9/2< 9/2 in the YES case and >10K−1/2> 10K - 1/2 in the NO case.

The finite rational comparison system can be transformed deterministically into a simple undirected unweighted graph GG with OPTuncut(G)=D0min⁡fCostfin(f)\mathrm{OPT}_{\mathrm{uncut}}(G) = D_0 \min_f \mathrm{Cost}_{\mathrm{fin}}(f), having OK(M0t)O_K(M_0^t) vertices and edges and constructed in polynomial time (including bit complexity) for fixed KK.

Methods

Folded bit proofs on finite affine spaces over F2\mathbb{F}_2 tested via three inner tests using face-array affine functions B(x)=∑i=1mBx−iiB(x) = \sum_{i=1}^m B_{x_{-i}}^i and Gaussian threshold sign tables Pc+σg,ηλP_{c+\sigma g, \eta\lambda}.

Source excerpts

“For a nonempty affine alphabet A, a folded bit proof is a function f : {P : A → {1, −1}} −→ {1, −1}, f(−P) = −f(P).” · Show in paper

“We take these entries independently and uniformly from D(A) and put B(x) = ∑m i=1 B^ix−ii_{x−i} ∈ D(A), B(a) = (B(x)(a))_{x∈X}.” · Show in paper

Gaussian smoothing, Fourier–Hermite degree projections L=ΠBΠchL = \Pi_B \Pi_c h, and truth-table differentiation to obtain low-degree product-of-faces correlation.

Source excerpts

“FB(c) = Eg,λf(Pc+σg,ηλ(B)), hx(B, c) = √N ∂cxFB(c).” · Show in paper

“Conditional on each ξ, let ΠB project onto Fourier row degree at most Rm, and let Πc project onto Hermite degree at most s. They act on different variables and commute. Define L = ΠBΠch.” · Show in paper

Gauge-identity partitioning, localization to a random subbox Y=∏iYiY = \prod_i Y_i, and iterated Box Cauchy–Schwarz to extract an odd one-row Fourier coefficient uniformly in the affine alphabet.

Source excerpts

“We first choose m large enough to partition the Fourier–Hermite terms of Lx: each nonzero term either has positive degree in cx, or carries equal odd frequencies in two rows B^ix−ii_{x−i} and B^jx−jj_{x−j}.” · Show in paper

“Lemma 4.2 (Box Cauchy–Schwarz). For every real array v, 𝔡Y (v)^{2^m} ≤ Eu0,u1 ∏_{ω∈{0,1}^m} v(u^ω).” · Show in paper

An equation-versus-variable outer game with fresh slots, canonical truth-table encodings, and qq shared affine hints.

Source excerpts

“For an integer q ≥ 0, the game with q affine hints additionally draws independent uniform functions L1, . . . , Lq ∈ D(AV ). The first player receives (U, π∗eL1, . . . , π∗eLq) and the second receives (V, L1, . . . , Lq).” · Show in paper

Deterministic rational approximation of the joint Gaussian sign tables (T0,T1)(T_0, T_1) via cube truncation, grid rounding, and Taylor polynomial antiderivatives.

Source excerpts

“Define the rational polynomial Qr(v) = ∑_{j=0}^{2r} (−1)^j v2jv^{2j} / (2^j j!).” · Show in paper

Path-gadget realization of weighted equality and inequality demands using 4-edge and 3-edge paths with fresh internal vertices, and reduction to minimum 2CNF clause deletion.

Source excerpts

“For every unit of integer multiplicity in (7.10), join its endpoint vertices by a path of length four for an equality demand and length three for an inequality demand.” · Show in paper

“Given G = (V, E), introduce a variable xv for each vertex and, for each edge uv, the two clauses (xu ∨ xv), (¬xu ∨ ¬xv).” · Show in paper