OctoInformBeta

OpenAI preprint, September 2026 · §4. Folded Gaussian blocks

Lemma 4.1: Grid masses, logarithmic means, and Gaussian rounding with uniform bounds on eligible atoms

Prime scales for folded Gaussian blocks.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

A Gaussian block assigns weights to the Boolean vertices of a unit cube and samples its folded density on a residue grid. The denominator of the grid is chosen on an exponential scale determined by the mean logarithmic density.

Let DD be a nonnegative integer. Identify a Boolean digit with its real value 00 or 11, so that a Boolean word ε\varepsilon is a vertex of [0,1]D[0,1]^D. The standard Gaussian densities γ1:ℝ→ℝ\gamma_1:\mathbb R\to\mathbb R and γD:ℝD→ℝ\gamma_D:\mathbb R^D\to\mathbb R are the functions defined by (1): γ1(x)=e−x2/22π,γD(y)=∏j=1Dγ1(yj)=(2π)−D/2e−∥y∥22/2.\gamma_1(x)=\frac{e^{-x^2/2}}{\sqrt{2\pi}}, \qquad \gamma_D(y)=\prod_{j=1}^D\gamma_1(y_j) =(2\pi)^{-D/2}e^{-\|y\|_2^2/2}.(1) For h>0h>0, the univariate folded density fh:ℝ→ℝf_h:\mathbb R\to\mathbb R and the product density Fh:ℝD→ℝF_h:\mathbb R^D\to\mathbb R are the functions defined by (2): fh(u)=hγ1(hu)+hγ1(h(u−1)),Fh(t)=∏j=1Dfh(tj).f_h(u)=h\gamma_1(hu)+h\gamma_1(h(u-1)), \qquad F_h(t)=\prod_{j=1}^D f_h(t_j).(2) The Gaussian atom associated with a Boolean vertex ε\varepsilon is the function Ah(⋅,ε):ℝD→ℝA_h(\cdot,\varepsilon):\mathbb R^D\to\mathbb R in (3): Ah(t,ε)=hDγD(h(t−ε)).A_h(t,\varepsilon)=h^D\gamma_D\bigl(h(t-\varepsilon)\bigr).(3)✦ For h>0h>0, the conditional bit probability qh:ℝ×{0,1}→ℝq_h:\mathbb R\times\{0,1\}\to\mathbb R and, for t∈ℝDt\in\mathbb R^D, the conditional word probability Ph,t:{0,1}D→ℝP_{h,t}:\{0,1\}^D\to\mathbb R are the functions defined by (4), with words indexed by all Boolean words: qh(u,a)=hγ1(h(u−a))fh(u),Ph,t(ε)=∏j=1Dqh(tj,εj).q_h(u,a)=\frac{h\gamma_1(h(u-a))}{f_h(u)}, \qquad P_{h,t}(\varepsilon)=\prod_{j=1}^D q_h(t_j,\varepsilon_j).(4)

The folded mass M(h)M(h) is the integral of fhf_h on [0,1][0,1].✦ For h>0h>0, the mean logarithmic density μ(h)\mu(h) of fhf_h on [0,1][0,1] is the average of log⁡fh\log f_h with respect to the measure with density fhf_h and total mass M(h)M(h), as in (5): M(h)=∫01fh(u)du,μ(h)=1M(h)∫01fh(u)log⁡fh(u)du.M(h)=\int_0^1 f_h(u)\,du, \qquad \mu(h)=\frac{1}{M(h)}\int_0^1 f_h(u)\log f_h(u)\,du.(5) The block height hbh_b and its logarithmic mean μb\mu_b, for integers b≥2b\geq2, are given in (6); all logarithms are natural: hb=94log⁡b,μb=μ(hb).h_b=\sqrt{\frac94\log b},\qquad \mu_b=\mu(h_b).(6)

For a positive integer pp, let GD,p=(ℤ/pℤ)DG_{D,p}=(\mathbb Z/p\mathbb Z)^D. The anchor tzt_z of a residue vector z∈GD,pz\in G_{D,p} and its grid cell QzQ_z are the point and half-open cube in (7), where z‾j\bar z_j is the representative of zjz_j in {0,…,p−1}\{0,\ldots,p-1\}: (tz)j=z‾jp,Qz={t∈ℝD:0≤tj−(tz)j<1pfor every j}.(t_z)_j=\frac{\bar z_j}{p},\qquad Q_z=\left\{t\in\mathbb R^D: 0\leq t_j-(t_z)_j<\frac1p\ \text{for every }j\right\}.(7) Write ℰD(h,μ,Δ,r;p)⊆GD,p\mathcal E_D(h,\mu,\Delta,r;p)\subseteq G_{D,p} for the eligible grid residues with parameters h,μ,Δ,rh,\mu,\Delta,r. For an integer b≥2b\geq2, the eligible block set Eb,pE_{b,p} is the subset of Gb,pG_{b,p} defined by the parameter choices in (8): Eb,p=ℰb(hb,μb,b0.56,b0.70;p).E_{b,p}=\mathcal E_b(h_b,\mu_b,b^{0.56},b^{0.70};p).(8) For h>0h>0, the discrete folded Gaussian mass ℳD,h,p(E)\mathcal M_{D,h,p}(E) of a set E⊆GD,pE\subseteq G_{D,p} is the grid average in (9): ℳD,h,p(E)=1pD∑z∈EFh(tz).\mathcal M_{D,h,p}(E)=\frac1{p^D}\sum_{z\in E}F_h(t_z).(9)

The Claim

Theorem 1. For all sufficiently large integers bb, there is a prime pp in the logarithmic window (10): bμb−4b0.56≤log⁡p≤bμb−3b0.56.b\mu_b-4b^{0.56}\leq\log p\leq b\mu_b-3b^{0.56}.(10)✦ For all sufficiently large integers bb, every positive integer pp satisfying log⁡p≥bμb−4b0.56\log p\geq b\mu_b-4b^{0.56} has the grid mass bounds (11): |ℳb,hb,p(Gb,p)−1|≤7b−1/20,|ℳb,hb,p(Eb,p)−1|≤7b−1/20.\begin{aligned} \bigl|\mathcal M_{b,h_b,p}(G_{b,p})-1\bigr|&\leq 7b^{-1/20},\\ \bigl|\mathcal M_{b,h_b,p}(E_{b,p})-1\bigr|&\leq 7b^{-1/20}. \end{aligned}(11)✦ For every h>0h>0, the logarithmic mean has the lower bound (12) and the upper bound (13): log⁡h−log⁡2π−12M(h)≤μ(h).\log h-\log\sqrt{2\pi}-\frac{1}{2M(h)}\leq\mu(h).(12)✦ μ(h)≤log⁡2h2π.\mu(h)\leq\log\frac{2h}{\sqrt{2\pi}}.(13)✦ Let h>0h>0, let s,t∈[0,1]Ds,t\in[0,1]^D, and let δ∈ℝ\delta\in\mathbb R satisfy |tj−sj|≤δ|t_j-s_j|\leq\delta in every coordinate. Put ξ=Dh2δ\xi=Dh^2\delta. For every Boolean word ε\varepsilon, the Gaussian rounding bounds are (14): e−ξFh(s)≤Fh(t)≤eξFh(s),e−ξFh(s)Ph,s(ε)≤Ah(t,ε)≤eξFh(s)Ph,s(ε).\begin{aligned} e^{-\xi}F_h(s)&\leq F_h(t)\leq e^{\xi}F_h(s),\\ e^{-\xi}F_h(s)P_{h,s}(\varepsilon) &\leq A_h(t,\varepsilon) \leq e^{\xi}F_h(s)P_{h,s}(\varepsilon). \end{aligned}(14)✦ There is an absolute constant c>0c>0 such that, for all sufficiently large integers bb and every prime pp satisfying log⁡p≥bμb−4b0.56\log p\geq b\mu_b-4b^{0.56}, each z∈Eb,pz\in E_{b,p}, t∈Qzt\in Q_z, and Boolean word ε\varepsilon satisfy (15): Fhb(tz)pPhb,tz(ε)≤e−cb0.70,Ahb(t,ε)p≤e−cb0.70.\begin{aligned} \frac{F_{h_b}(t_z)}pP_{h_b,t_z}(\varepsilon)&\leq e^{-cb^{0.70}},\\ \frac{A_{h_b}(t,\varepsilon)}p&\leq e^{-cb^{0.70}}. \end{aligned}(15)✦

In the paper, the result reads: for all sufficiently large bb there is a prime pbp_b with bμb−4b0.56≤log⁡pb≤bμb−3b0.56b\mu_b-4b^{0.56}\leq\log p_b\leq b\mu_b-3b^{0.56}, the uniform expectations over z∈Gb,pbz\in G_{b,p_b} satisfy 𝔼zFhb(tz)=1+O(b−1/20)\mathbb E_z F_{h_b}(t_z)=1+O(b^{-1/20}) and 𝔼zFhb(tz)𝟏Eb,pb(z)=1+O(b−1/20)\mathbb E_z F_{h_b}(t_z)\mathbf 1_{E_{b,p_b}}(z)=1+O(b^{-1/20}), and μb=log⁡hb+O(1)\mu_b=\log h_b+O(1); with ηb=bhb2/pb\eta_b=bh_b^2/p_b, every t∈Qzt\in Q_z and Boolean word ε\varepsilon satisfy e−ηbFhb(tz)≤Fhb(t)≤eηbFhb(tz)e^{-\eta_b}F_{h_b}(t_z)\leq F_{h_b}(t)\leq e^{\eta_b}F_{h_b}(t_z) and e−ηb(Fhb(tz)/pb)Phb,tz(ε)≤(hbb/pb)γb(hb(t−ε))≤eηb(Fhb(tz)/pb)Phb,tz(ε)e^{-\eta_b}(F_{h_b}(t_z)/p_b)P_{h_b,t_z}(\varepsilon) \leq(h_b^b/p_b)\gamma_b(h_b(t-\varepsilon)) \leq e^{\eta_b}(F_{h_b}(t_z)/p_b)P_{h_b,t_z}(\varepsilon), while there is an absolute constant c>0c>0 such that, for eligible zz, every t∈Qzt\in Q_z and Boolean word ε\varepsilon have both (Fhb(tz)/pb)Phb,tz(ε)(F_{h_b}(t_z)/p_b)P_{h_b,t_z}(\varepsilon) and (hbb/pb)γb(hb(t−ε))(h_b^b/p_b)\gamma_b(h_b(t-\varepsilon)) at most e−cb0.70e^{-cb^{0.70}}.

What Is Checked

The statements are over finite Boolean cubes with Gaussian atom weights and finite residue grids, with the eligible sets ℰD(h,μ,Δ,r;p)\mathcal E_D(h,\mu,\Delta,r;p).