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

The joint Dickman law for consecutive integers

For every fixed a,b∈(0,1)a, b \in (0, 1), lim⁡X→∞1X#{2≤n≤X:P+(n)≤na,P+(n+1)≤nb}=ρ(1/a)ρ(1/b)\lim_{X \to \infty} \frac{1}{X} \#\{2 \le n \le X : P^+(n) \le n^a, P^+(n+1) \le n^b\} = \rho(1/a)\rho(1/b), where the limit is taken over all real XX with ordinary, unweighted counting.

The ordering P+(n)<P+(n+1)P^+(n) < P^+(n+1) has natural density 1/21/2, i.e., lim⁡X→∞1X#{2≤n≤X:P+(n)<P+(n+1)}=12\lim_{X \to \infty} \frac{1}{X} \#\{2 \le n \le X : P^+(n) < P^+(n+1)\} = \frac{1}{2}, and the reverse ordering has the same natural density.

For every fixed c,d∈(0,1)c, d \in (0, 1), lim⁡X→∞1X#{2≤n≤X:P+(n)≤Xc,P+(n+1)≤Xd}=ρ(1/c)ρ(1/d)\lim_{X \to \infty} \frac{1}{X} \#\{2 \le n \le X : P^+(n) \le X^c, P^+(n+1) \le X^d\} = \rho(1/c)\rho(1/d), and for 0≤c,d≤10 \le c, d \le 1, lim⁡X→∞1X#{2≤n≤X:P+(n)>Xc,P+(n+1)>Xd}=(1−D(c))(1−D(d))\lim_{X \to \infty} \frac{1}{X} \#\{2 \le n \le X : P^+(n) > X^c, P^+(n+1) > X^d\} = (1 - D(c))(1 - D(d)).

For every fixed integer J≥2J \ge 2 and every pair of fixed phase vectors (ζk)k=1J−1(\zeta_k)_{k=1}^{J-1} and (ξk)k=1J−1(\xi_k)_{k=1}^{J-1} of absolute value one, the corresponding bin-character labels satisfy lim⁡x→∞1x∑1≤n<xgx(n)‾Fx(n+1)=0\lim_{x \to \infty} \frac{1}{x} \sum_{1 \le n < x} \overline{g_x(n)} F_x(n+1) = 0, where Fx=fx−μF_x = f_x - \mu and the limit is through all integer scales.

For T=T(B)→∞T = T(B) \to \infty, LB→∞L_B \to \infty, any integers aBa_B, fixed 0<s0<s1<∞0 < s_0 < s_1 < \infty, positive integer DD, residue r(modD)r \pmod D, and v≥0v \ge 0, lim⁡B→∞lim sup⁡x→∞1Tx∑s0Tx≤n≤s1Tx∣1LB∑i=aB+1aB+LB1i≡r(modD)Fx(n+i)GB,v(n+i)∣2=0\lim_{B \to \infty} \limsup_{x \to \infty} \frac{1}{Tx} \sum_{s_0 Tx \le n \le s_1 Tx} \left| \frac{1}{L_B} \sum_{i=a_B+1}^{a_B+L_B} \mathbf{1}_{i \equiv r \pmod D} F_x(n+i) G_{B,v}(n+i) \right|^2 = 0.

There is c5>0c_5 > 0, independent of all sufficiently large fixed C∗C_*, such that the nonnegative quadratic energy I2,xI_{2,x} satisfies lim inf⁡B→∞lim inf⁡x→∞I2,xBT≥c5\liminf_{B \to \infty} \liminf_{x \to \infty} \frac{I_{2,x}}{BT} \ge c_5, while its diagonal contribution satisfies lim⁡B→∞lim sup⁡x→∞I2,xdiagBT=0\lim_{B \to \infty} \limsup_{x \to \infty} \frac{I_{2,x}^{\mathrm{diag}}}{BT} = 0.

For any symmetric real kernel Mik(Si,Sk;s)\mathcal{M}_{ik}(S_i, S_k; s) zero on nonpairs with total absolute mass O(B10)O(B^{10}), Eu∥(Kk−i(u+i,s)−Mik(Si(u),Sk(u);s))i,k∥□≤o(1)+E(Si) indep∥(Lik(Si,Sk;s)−Mik(Si,Sk;s))i,k∥□\mathbb{E}_u \|(\mathcal{K}_{k-i}(u+i, s) - \mathcal{M}_{ik}(S_i(u), S_k(u); s))_{i,k}\|_\square \le o(1) + \mathbb{E}_{(S_i)\,\mathrm{indep}} \|(\mathcal{L}_{ik}(S_i, S_k; s) - \mathcal{M}_{ik}(S_i, S_k; s))_{i,k}\|_\square, uniformly for s∈Ss \in \mathcal{S}.

For a sufficiently fine fixed regularity grid and sufficiently small fixed tolerance, the independent-site latent kernel satisfies ∑k≠iELik2≪B−.21\sum_{k \neq i} \mathbb{E}\mathcal{L}_{ik}^2 \ll B^{-.21} (1≤i≤M1 \le i \le M) and E(1M∑i,kLik)2≪1\mathbb{E}\left(\frac{1}{M}\sum_{i,k} \mathcal{L}_{ik}\right)^2 \ll 1.

For every fixed coarse partition, uniformly for 1≤d≤B1 \le d \le B, ∥Ud∥≪1\|U_d\| \ll 1 and ∥(1−Pav)Ud∥≪B−c7\|(1 - P_{\mathrm{av}})U_d\| \ll B^{-c_7} with c7=1/200c_7 = 1/200. Moreover, for every ϵ1>0\epsilon_1 > 0, one can choose m1m_1 such that lim sup⁡B→∞sup⁡∥g∥2≤1∥ug−Pm1ug∥2≤ϵ1\limsup_{B \to \infty} \sup_{\|g\|_2 \le 1} \|u_g - P_{m_1}u_g\|_2 \le \epsilon_1.

For the finite-feature matrix M\mathcal{M} in (8.17) and Eik=Lik−Mik\mathcal{E}_{ik} = \mathcal{L}_{ik} - \mathcal{M}_{ik}, sup⁡gi,hk:ΩB→[−1,1]1M∣∑i,kESi,Sk[Eik(Si,Sk;s)gi(Si)hk(Sk)]∣≤Ce−c3C∗+Cηϵ1+o(1)\sup_{g_i, h_k : \Omega_B \to [-1, 1]} \frac{1}{M} \left| \sum_{i,k} \mathbb{E}_{S_i, S_k}[\mathcal{E}_{ik}(S_i, S_k; s) g_i(S_i) h_k(S_k)] \right| \le C e^{-c_3 C_*} + C_\eta \epsilon_1 + o(1), and deterministically sup⁡i∑k∣Mik∣≪η,m11\sup_i \sum_k |\mathcal{M}_{ik}| \ll_{\eta, m_1} 1 and sup⁡i∑k∣Mik∣2≪η,m1T−1\sup_i \sum_k |\mathcal{M}_{ik}|^2 \ll_{\eta, m_1} T^{-1}.

For every fixed ϵ2>0\epsilon_2 > 0, E∥(Eik(Si,Sk;s))∥□≤E0+ϵ2+o(1)\mathbb{E}\|(\mathcal{E}_{ik}(S_i, S_k; s))\|_\square \le E_0 + \epsilon_2 + o(1) uniformly for ss in the fixed compact range, where E0=Ce−c3C∗+Cηϵ1E_0 = C e^{-c_3 C_*} + C_\eta \epsilon_1.

With C∗C_*, η\eta, C6C_6, and the coarse partition fixed, the main-kernel energy satisfies lim⁡B→∞lim sup⁡x→∞∣QB,xmain∣=0\lim_{B \to \infty} \limsup_{x \to \infty} |\mathcal{Q}_{B,x}^{\mathrm{main}}| = 0.

Methods

Bin-character encoding of large prime factors and finite Fourier inversion reducing the joint Dickman law to mixed decorrelation of two independently chosen multiplicative labels.

Source excerpts

“We first replace the largest prime factor by finitely many prime counts. For a fixed integer J ≥ 2, divide the primes in (x1/Jx^{1/J}, x] into bins Bk,xB_{k,x} = (xk/Jx^{k/J}, x(k+1)/Jx^{(k+1)/J}], 1 ≤ k < J. ... Finite Fourier inversion reduces the joint law to the mixed decorrelation 1/x ∑_{n<x} \overline{g_x(n)} F_x(n + 1) −→ 0.” · Show in paper

Pointwise tensor Vandermonde interpolation of centered bin-count functions by real, nonnegative, 1-bounded completely multiplicative functions, combined with real and complex short-interval theorems (Matomäki–Radziwiłł and Matomäki–Radziwiłł–Tao) to control weighted short averages.

Source excerpts

“Here the finite count range gives a pointwise tensor interpolation of the centered function of all bin counts by completely multiplicative functions. ... Choose J + 1 distinct numbers in (0, 1]. The Vandermonde matrix whose entries are their powers of orders 0, . . . , J is invertible.” · Show in paper

Profinite-integer correlation profiling and rough-divisor amplification with independent fair-split coupling to convert a nonzero mixed correlation into a positive quadratic energy.

Source excerpts

“For an integer n, the amplifier defines D_B(n) as a nonnegative weighted count of factorizations n = am and n + 1 = cl_a, restricted by e^B < c < e2Be^{2B} and Tc < a < 2Tc. ... Cauchy–Schwarz in (c, m) removes the unit-modulus factor g_x(m) and squares this inner sum.” · Show in paper

Independent-root comparison in expected cut norm, coupling actual auxiliary-prime divisibility sets at block endpoints to independent 1/p site models and separating extra-root fluctuations.

Source excerpts

“Proposition 5.3 (Independent-root comparison). Let MikM_{ik}(S_i, S_k; s) be a symmetric real kernel, zero on nonpairs, with total absolute mass O(B10B^{10}) uniformly in its arguments. Then E_u ||(Kk−iK_{k−i}(u + i, s) − MikM_{ik}(S_i(u), S_k(u); s))_{i,k}||_□ ≤ o(1) + E(Si)indepE_{(S_i) indep} ||(LikL_{ik}(S_i, S_k; s) − MikM_{ik}(S_i, S_k; s))_{i,k}||_□.” · Show in paper

Smoothing channels on logarithmic and residue space via fair prime splits to flatten residue dependence in operator norm and approximate logarithmic outputs on a fixed coarse partition.

Source excerpts

“The site model admits a useful smoothing operation: choose a fair subset of the primes at one site and record the logarithm and a residue of its product. We prove that this operation suppresses residue dependence in operator norm and that its logarithmic outputs are uniformly approximable on a fixed coarse partition.” · Show in paper

Small-column sampling and McDiarmid's bounded-differences inequality applied to a McShane-extended Lipschitz function to upgrade integrated cut comparisons to expected cut norm.

Source excerpts

“We use the McShane extension [19] in its infimum form [3, Section 3, Equation (3.2)], followed by clamping ... McDiarmid’s bounded-differences inequality [18] therefore gives, for every a > 0, P(Z̃ − E[Z̃ | SJ′S_{J'}] > a | SJ′S_{J'} = t) ≤ exp(−2a^2/(|R'| c_B^2)).” · Show in paper