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.

All papers

Claims

Findings

Bounded-Step Walks on Gaussian Primes

For every finite real DD there is a finite BDB_D such that every connected component of GDG_D has at most BDB_D vertices, independent of the starting prime and including primes on the coordinate axes; consequently every sequence of distinct Gaussian primes whose successive distances are at most DD has at most BDB_D terms.

For every D≥1D \ge 1 there is a finite set PD\mathcal{P}_D of rational primes congruent to 11 modulo 44, depending only on DD, such that A(PD)\mathcal{A}(\mathcal{P}_D) contains no infinite sequence of distinct points with successive distances at most DD.

If D≥1D \ge 1 and P\mathcal{P} is a finite set of rational primes congruent to 11 modulo 44 such that A(P)\mathcal{A}(\mathcal{P}) contains no infinite self-avoiding walk with steps at most DD, then every component CC of GDG_D satisfies #C≤max⁡{Q2,∣E∣+(KD−1)∣E∣Q2}\#C \le \max\{Q^2, |E| + (K_D - 1)|E|Q^2\}, where Q=∏p∈PpQ = \prod_{p \in \mathcal{P}} p, KD=#{v∈Z[i]:∣v∣≤D}K_D = \#\{v \in \mathbb{Z}[i] : |v| \le D\}, and EE is the set of associates of the selected Gaussian factors.

Under the forward difference-endpoint sampling transition on a fixed self-avoiding DD-step walk with q=⌊U/L∗⌋q = \lfloor U/L_* \rfloor and b=⌊g2U/L∗⌋b = \lfloor g^2 U/L_* \rfloor in the asymptotic regime DD fixed, T→∞T \to \infty, q≤kq \le k, and gCU/log⁡T→∞g^C U / \log T \to \infty for every fixed C>0C > 0, uniformly over walks, batches, and starting-time laws, fV(b)bL∗≥12(1+fZ(q)qL∗)−OD(g)\frac{f_V(b)}{b L_*} \ge \frac{1}{2}\left(1 + \frac{f_Z(q)}{q L_*}\right) - O_D(g).

In the one-step coverage transfer experiment, if at every aa in the support of AA all but a fraction at most η′\eta' of the signed coordinates have (τ′,β′)(\tau', \beta') coverage for ϕc(y(a)+h1)\phi_c(y(a) + h_1) under the law conditional on A=aA = a, and hypotheses (5.3)-(5.6) hold, then all but a fraction at most η\eta of the coordinates have (τ,β)(\tau, \beta) coverage for ϕc(Z)\phi_c(Z).

For all sufficiently large MM, uniformly over the batches, 0≤i≤J0 \le i \le J, and all exact start times a≥0a \ge 0 at AiA_i, all but a fraction at most ηi\eta_i of the signed primes satisfy #{u∈Z[i]/(π):bi,a,π(u)<τi/p}<p1−βi\#\{u \in \mathbb{Z}[i]/(\pi) : b_{i,a,\pi}(u) < \tau_i/p\} < p^{1-\beta_i}.

There is a constant c1>0c_1 > 0, depending only on the fixed schedule parameters, such that for all sufficiently large MM and every batch in (6.1), 1LTECI(XT′;WLT(t∗)∣OT′)≥c1log⁡T\frac{1}{L_T} \mathbb{E}_{\mathcal{C}} I(X'_T; W_{L_T}(t_*) \mid O'_T) \ge \frac{c_1}{\log T}, uniformly over all infinite self-avoiding DD-step walks in A(PM)\mathcal{A}(\mathcal{P}_M).

For every fixed collection C\mathcal{C} of prime choices, ∑TI(XT′;WLT(t∗)∣OT′)LT≤log⁡KD+o(1)\sum_T \frac{I(X'_T; W_{L_T}(t_*) \mid O'_T)}{L_T} \le \log K_D + o(1), where the error tends to zero uniformly over the choices and the walks as M→∞M \to \infty.

Methods

Finite periodic sieving reduction: reducing the uniform bound on connected component sizes of the Gaussian-prime graph G_D to the nonexistence of an infinite self-avoiding D-step walk in a periodic sieved set A(P_D) via translation invariance and edge-degree counting around exceptional primes.

Source excerpts

“Proposition 2.1 (From a periodic obstruction to a uniform bound). Fix D >= 1 and a finite set P of rational primes congruent to 1 modulo 4. Suppose that A(P) contains no infinite self-avoiding walk with steps at most D.” · Show in paper

Geometric difference-set sampling and torus-degree lower bound: lower-bounding the number of distinct differences |E - E| of a walk segment via a degree-(-1) map from the parameter torus to R^2/Lambda, and sampling an endpoint of a uniformly chosen difference to enrich residue entropy.

Source excerpts

“Homotoping both paths to their straight segments while keeping their endpoints fixed shows that this map has degree -1, with the target oriented by a, b. In particular it is onto.” · Show in paper

“Conditional on tau = t, choose h uniformly from the set E_t - E_t, including zero, and then choose one ordered pair (X, Y) in E_t^2 with X - Y = h by a fixed deterministic rule.” · Show in paper

Random signed prime-factor separation: ruling out collisions in residue reductions of difference sets using independent fair sign choices over split rational primes, determinant divisibility, Hoeffding's inequality, and the Boolean cube isoperimetric inequality.

Source excerpts

“Lemma 4.2 (A randomly signed product avoids a thin rectangle).” · Show in paper

“We combine it with the separation of their sign classes using the cube's isoperimetric inequality.” · Show in paper

Backward coverage transfer with shared continuation data: strengthening (tau', beta') coverage to (tau, beta) coverage across checkpoints by conditioning on a single shared vector of independent continuation displacements and applying a short-list Fano-type entropy bound.

Source excerpts

“Theorem 5.2 (One-step coverage transfer).” · Show in paper

“We begin at A_J using the particularly accurate marginal estimate, then verify every numerical hypothesis of Theorem 5.2 for the passage from Ai+1A_{i+1} to A_i.” · Show in paper

Common-law entropy-decrement telescope: measuring zero-avoidance information costs across geometrically spaced prime batches at a single smoothed terminal time t_* (with an independent uniform offset) and bounding their sum by log K_D + o(1) via dyadic word partitioning and conditional block subadditivity.

Source excerpts

“After all transitions in all windows have been performed, add one independent uniform offset in {0, ..., N - 1}, where N = ceil(T_+^{10}). Denote the resulting random time by t_*.” · Show in paper

“Lemma 8.1 (A common-law entropy telescope).” · Show in paper