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

A counterexample to Hadwiger's conjecture

There are finite nonempty simple graphs GG of arbitrarily large order mm for which h(G)<26m75+23<m2≤χf(G)≤χ(G)h(G) < \frac{26m}{75} + \frac{2}{3} < \frac{m}{2} \le \chi_f(G) \le \chi(G), disproving Hadwiger's conjecture and its fractional-coloring weakening χf(G)≤h(G)\chi_f(G) \le h(G).

There are arbitrarily large integers mm for which an mm-vertex finite simple graph GG satisfies α(G)≤2\alpha(G) \le 2 and cm(G)<m100\mathrm{cm}(G) < \frac{m}{100}.

For every finite nonempty graph GG of order mm, h(G)≤m+4 cm(G)+23h(G) \le \frac{m + 4\,\mathrm{cm}(G) + 2}{3}.

The graphs in Theorem 1.1 disprove the sharp connected-matching conjecture of Füredi, Gyárfás, and Simonyi that a graph with independence number two and order 4t−14t - 1 has a connected matching of size tt.

For the construction parameters specified in Section 2 and Appendix A, for all sufficiently large nn, every probability law σ\sigma on units satisfying σ1≤Mμ\sigma_1 \le M\mu, σ2≤Mμ\sigma_2 \le M\mu, and σ≤2DNμ2\sigma \le 2^{DN}\mu^2 has the property that two independent units drawn from σ\sigma have all four cross holes with probability at least 2−100gN2^{-100gN}.

The hole relation on Ωn\Omega_n is symmetric, has no loops, and is triangle-free, so any graph GG formed on sampled positions o1,…,om∈Ωno_1, \dots, o_m \in \Omega_n satisfies α(G)≤2\alpha(G) \le 2 and χ(G)≥⌈m/2⌉\chi(G) \ge \lceil m/2 \rceil deterministically.

Assuming the conclusion of Theorem 3.1 at a given sufficiently large nn, for m=2C0gNm = 2^{C_0 g N}, the sampled graph GG defined in Section 2 satisfies α(G)≤2\alpha(G) \le 2 and cm(G)<m/100\mathrm{cm}(G) < m/100 with probability 1−exp⁡(−Ω(m))1 - \exp(-\Omega(m)).

Methods

Tensor and frame realization of a triangle-free hole relation over F2\mathbb{F}_2 using moment matrices v(s,z)v(s,z)Tv(s,z)v(s,z)^{\mathsf{T}} on cut-profile spaces X\mathcal{X} and injective linear maps UoU_o with linear functionals uou_o.

Source excerpts

“Let X and V be finite-dimensional vector spaces over F_2. Fix a linear functional a \in X^* and a symmetric bilinear form T on X, and write r(x) = a + T(x, \cdot) \in X^*.” · Show in paper

“For a raw vertex o, define U_o x = (Po,eP_{o,e} x_e Qo,eQ_{o,e}^T)_{e \in E}, u_o(\xi) = \sum_{e \in E} \langle Yo,eY_{o,e} Xo,eX_{o,e}^T, \xi_e \rangle.” · Show in paper

Entropy-minimizing container and fingerprint method combined with Csiszár's information projection and Ford–Fulkerson max-flow/min-cut capacities to transfer distribution supersaturation to finite random graphs.

Source excerpts

“We use a finite form of the information-projection argument of Csiszar [8].” · Show in paper

“The next lemma is the max-flow/min-cut principle of Ford and Fulkerson [12] with capacities chosen to encode the three density bounds.” · Show in paper

“We now adapt the independent-set fingerprint method of Kleitman and Winston [17]; see also Samotij [28, Section 2]. Here the progress measure is the minimum relative entropy of a feasible law.” · Show in paper

Iterative peeling into pinned leaves with bounded-dimensional pin spaces and uniform exact-image probability bounds.

Source excerpts

“We partition almost all of a capped unit law into leaves by storing unusually likely exact-image constraints in bounded-dimensional pin spaces. The remaining directions then satisfy a uniform image-probability bound.” · Show in paper

Low-rank Boolean moment decomposition using commuting multiplication operators over F2\mathbb{F}_2 and uniform mixing forms.

Source excerpts

“The low-rank decomposition below uses the quotient and multiplication operators familiar from flat extensions of moment matrices; see Laurent and Mourrain [20, Section 2, especially Lemma 2.1].” · Show in paper

Deterministic gradient realization via linear solvability on quotient spaces, base-coordinate compression, and channel-space factorization.

Source excerpts

“The construction has three steps. We first solve a relaxed linear problem, allowing an arbitrary bilinear correction on the unfrozen quotients. We then bound the rank of that correction, and finally factor it through the available channel coordinates.” · Show in paper

Weak product comparison via finite-partition energy-increment regularity with twisted cut norms and four-copy Cauchy–Schwarz estimates.

Source excerpts

“We use the finite-partition energy-increment method underlying weak regularity; see Frieze and Kannan [14]. The additional Gram conditioning needed here is retained explicitly in the following self-contained proof.” · Show in paper