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 quadratic bound for Jacobsthal's function

There is an absolute constant C>0C > 0 such that Jacobsthal's function satisfies h(k)≤Ck2(log⁡log⁡(3k))2h(k) \le C \frac{k^2}{(\log\log(3k))^2} for every integer k≥1k \ge 1, giving an affirmative answer to Jacobsthal's quadratic-bound question.

There are absolute constants c>0c > 0 and z0z_0 such that for every real z≥z0z \ge z_0 and every choice of a residue class ap(modp)a_p \pmod{p} for each prime p≤zp \le z, #{1≤n≤Y:n≢ap(modp) for every p≤z}≥cYV0B2\#\{1 \le n \le Y : n \not\equiv a_p \pmod{p} \text{ for every } p \le z\} \ge c \frac{Y V_0}{B^2}, where L=log⁡zL = \log z, Y=⌊z2/L2⌋Y = \lfloor z^2 / L^2 \rfloor, w=L/(log⁡L)2w = L / (\log L)^2, B=L/log⁡wB = L / \log w, and V0=∏p≤w(1−1/p)V_0 = \prod_{p \le w} (1 - 1/p).

At the root of the reference tree, B2Pe(r0,B)eγ→−a⋆+2IMg>0.02\frac{B^2 P_{\mathrm{e}}(r_0, B)}{e^\gamma} \to -a_\star + \frac{2I}{M_g} > 0.02, so there is an absolute constant cL>0c_L > 0 such that B2Pe(r0,B)≥cLB^2 P_{\mathrm{e}}(r_0, B) \ge c_L for all sufficiently large zz.

For a candidate edge in a full regular box, outside a set of normalized harmonic box measure at most σ\sigma, every occurrence of ∣Npqu−Npq/u∣>δYV0pqu|N_{pqu} - N_{pq}/u| > \delta \frac{Y V_0}{pqu} has a rational A/DA/D in a fixed whole-box list A\mathcal{A} of size at most FF (with D≤RZ10D \le R\mathcal{Z}^{10} and ∣A∣≤SqRZ10|A| \le S_q R\mathcal{Z}^{10}) such that pp aligns with A/DA/D and DqE(A/D)≤RZ10D q_E(A/D) \le R\mathcal{Z}^{10}.

Except for a set of regular compact endpoints of total harmonic mass at most ρB−2\rho B^{-2}, every bad compact endpoint ∣Nd/μd−1∣>ε|N_d/\mu_d - 1| > \varepsilon has a witnessing candidate and a rational A/DA/D on that candidate's fixed list satisfying D≤wCsD \le w^{C_s} and ∣A∣≤YwCs+2|A| \le Y w^{C_s + 2} such that all prime factors of dd with exponent x(t)>Csx(t) > C_s align to A/DA/D.

The stops of Definition 10.2 satisfy ∑d stoppedSd(x(Pd))≥∑d stoppedμdPe(rd,x(Pd))\sum_{d \text{ stopped}} S_d(x(P_d)) \ge \sum_{d \text{ stopped}} \mu_d P_{\mathrm{e}}(r_d, x(P_d)), and the regular included hard endpoints with rd≤Kr_d \le K having no stopped prefix have harmonic mass at most (δ+o(1))B−2(\delta + o(1))B^{-2}.

The linear-sieve deficits a(s)=1−f(s)a(s) = 1 - f(s) and c(s)=F(s)−1c(s) = F(s) - 1 are positive, satisfy a(s)+c(s)=c0(s)a(s) + c(s) = c_0(s), a(s)≍c0(s)a(s) \asymp c_0(s), c(s)≍c0(s)c(s) \asymp c_0(s) for s≥1s \ge 1, and obey the future-integral identities sa(s)=∫s−1∞c(t) dts a(s) = \int_{s-1}^\infty c(t)\,dt (s≥2s \ge 2) and sc(s)=∫s−1∞a(t) dts c(s) = \int_{s-1}^\infty a(t)\,dt (s≥1s \ge 1).

For any nonzero polynomial P∈R[X1,X2]P \in \mathbb{R}[X_1, X_2] of degree at most D0≥1D_0 \ge 1 and S≥1S \ge 1, the number of lattice points in [0,S]2[0, S]^2 lying on at least one nonlinear irreducible component of P=0P = 0 (over C\mathbb{C}) is O(D04(1+S2/3))O(D_0^4(1 + S^{2/3})).

Methods

Decreasing-prime tree expansion with an upper-node admission rule r−x≥max⁡(2x,x+2)r - x \ge \max(2x, x+2) and a prefix-free lower-node stopping rule.

Source excerpts

“Nodes of even length are called lower nodes, and nodes of odd length are called upper nodes. From a lower node include every available child. From an upper node include a child with exponent x = x(p) exactly when r − x ≥ H(x), H(x) = max(2x, x + 2), (3.3) where r is the parent’s gap. Other children are omitted.” · Show in paper

Normalization of harmonic path transitions into a regenerative probability kernel using derivative weights ϕe(s)=s2f′(s)\phi_{\mathrm{e}}(s) = s^2 f'(s) and ϕo(s)=−s2F′(s)\phi_{\mathrm{o}}(s) = -s^2 F'(s) of the linear-sieve functions, combined with the key renewal theorem.

Source excerpts

“Define ϕe(s) = s2f′(s) (s ≥ 2), ϕo(s) = −s2F′(s) (s ≥ 0.95).” · Show in paper

“At a state of type i, gap r, cutoff b, and ratio s = r/b, make a draw t with probability density Ki(s, dt) = W0(t)ϕi′(t) / ϕi(s) 1{t≥mi(s)} dt. (4.16)” · Show in paper

Total-variation coupling of the killed prime-path Markov chain with the continuous harmonic path chain over logarithmic ratio bins, supplemented by discrete marginal majorants.

Source excerpts

“This chain can be coupled with the continuous chain (4.16), killed at the first draw with x ≤ ℓ or t > S, so that the probability of failure is at most exp(−c3 √log w).” · Show in paper

Inverse-sieve method for a witnessing edge combining the additive large sieve, pigeonhole polynomial construction, Jarník's convex-arc lattice-point bound, and Gallagher's larger sieve pair-counting.

Source excerpts

“It combines polynomial interpolation with Jarník’s convex-arc lattice-point argument [18] and the pair-counting principle of Gallagher’s larger sieve [7].” · Show in paper

Variance estimate for aligned progression families using Brun-Titchmarsh for small moduli H≤R4/5H \le R^{4/5} and modular hyperbola completion with Kloosterman sums for large moduli R4/5<H≤RZ11R^{4/5} < H \le R\mathcal{Z}^{11}.

Source excerpts

“For H ≤ R4/5, Brun–Titchmarsh controls each residue pattern after we fix m0: the possible primes lie in a progression of modulus HT0T1. ... For larger H, we instead count integer pairs (p, m0) on a modular hyperbola and then sieve the p-coordinate.” · Show in paper

Owner-and-tag stopping rule that averages exact survivor counts over a large prime bin of multiplicity one.

Source excerpts

“A search bin has an owner if some eligible rational aligns to at least a 1 − η fraction of its primes, by prime count. The owner is that rational. Along a descending-prime path, its tag prime is the first search-bin factor that aligns to the owner of its bin, if such a factor occurs. Its tag rational is this owner and is retained for the rest of the path.” · Show in paper