OctoInformBeta

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

Lemma 4.4: Boolean translates of a lattice point receive Gaussian lower bounds, while eligible cells have small individual weights in sufficiently large blocks

Positive line weights select Gaussian cells.

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

The comparison concerns individual Gaussian weights at Boolean translates of a code-lattice point and a sparse maximum along a line in the residue grid.

Let II be a finite coordinate set of cardinality NN, and let pp be a prime. Write k¯\overline{k} for the coordinatewise reduction of k∈ℤIk\in\mathbb Z^I modulo pp, and zˆ\widehat z for the least nonnegative representatives of z∈𝔽pIz\in\mathbb F_p^I. For t∈ℝIt\in\mathbb R^I, the grid anchor ap(z)∈ℝIa_p(z)\in\mathbb R^I, floor vector kp(t)∈ℤIk_p(t)\in\mathbb Z^I and grid residue xp(t)∈𝔽pIx_p(t)\in\mathbb F_p^I are the maps in (1). ap(z)j=zˆjp,kp(t)j=⌊ptj⌋,xp(t)=kp(t)¯.a_p(z)_j=\frac{\widehat z_j}{p},\qquad k_p(t)_j=\lfloor pt_j\rfloor,\qquad x_p(t)=\overline{k_p(t)}.(1) The grid cell Qp,z⊂ℝIQ_{p,z}\subset\mathbb R^I is the half-open box in (2); the scaled integer vector Gp(k)G_p(k) and the Construction A lattice Λp(C)\Lambda_p(C) of a linear code C⊂𝔽pIC\subset\mathbb F_p^I are given there as well. Qp,z={t∈ℝI:0≤tj−ap(z)j<p−1 for all j∈I},Gp(k)=k/p,Λp(C)={Gp(k):k∈ℤI,k¯∈C}.\begin{aligned} Q_{p,z}&=\{t\in\mathbb R^I:0\le t_j-a_p(z)_j<p^{-1}\text{ for all }j\in I\},\\ G_p(k)&=k/p,\qquad \Lambda_p(C)=\{G_p(k):k\in\mathbb Z^I,\ \overline k\in C\}. \end{aligned}(2)

Folded Gaussian Weights

The standard Gaussian densities on ℝ\mathbb R and ℝI\mathbb R^I are (3). γ1(s)=e−s2/22π,γN(y)=∏j∈Iγ1(yj)=(2π)−N/2e−∥y∥2/2.\gamma_1(s)=\frac{e^{-s^2/2}}{\sqrt{2\pi}},\qquad \gamma_N(y)=\prod_{j\in I}\gamma_1(y_j) =(2\pi)^{-N/2}e^{-\|y\|^2/2}.(3) For h>0h>0, the univariate folded density qh:ℝ→ℝq_h:\mathbb R\to\mathbb R and the product folded density Fh:ℝI→ℝF_h:\mathbb R^I\to\mathbb R are defined by (4). qh(s)=hγ1(hs)+hγ1(h(s−1)),Fh(t)=∏j∈Iqh(tj).q_h(s)=h\gamma_1(hs)+h\gamma_1(h(s-1)),\qquad F_h(t)=\prod_{j\in I}q_h(t_j).(4) A Boolean digit has numerical value zero for false and one for true; a Boolean word ε∈{0,1}I\varepsilon\in\{0,1\}^I uses these values coordinatewise.✦ The conditional bit probability πh(s,δ)\pi_h(s,\delta), conditional word probability Ph(t,ε)P_h(t,\varepsilon) and Gaussian atom Ah(t,ε)A_h(t,\varepsilon) are the functions in (5), with δ∈{0,1}\delta\in\{0,1\}. πh(s,δ)=hγ1(h(s−δ))qh(s),Ph(t,ε)=∏j∈Iπh(tj,εj),Ah(t,ε)=hNγN(h(t−ε)).\begin{aligned} \pi_h(s,\delta)&=\frac{h\gamma_1(h(s-\delta))}{q_h(s)},\qquad P_h(t,\varepsilon)=\prod_{j\in I}\pi_h(t_j,\varepsilon_j),\\ A_h(t,\varepsilon)&=h^N\gamma_N(h(t-\varepsilon)). \end{aligned}(5)

For a finite field KK, a finite KK-vector space VV, a subset U⊂VU\subset V and functions F,f:V→ℝF,f:V\to\mathbb R, the sparse normalized weight FUF_U and the line maximum ℳvf\mathcal M_vf in direction v∈Vv\in V are given by (6). FU(z)={F(z)/|K|,z∈U,0,z∉U,(ℳvf)(x)=maxa∈Kf(x−av).F_U(z)=\begin{cases}F(z)/|K|,&z\in U,\\0,&z\notin U,\end{cases} \qquad (\mathcal M_vf)(x)=\max_{a\in K}f(x-av).(6) For U⊂𝔽pIU\subset\mathbb F_p^I and v∈𝔽pIv\in\mathbb F_p^I, the sparse Gaussian line weight gp,h,U,vg_{p,h,U,v} and its code lattice Λp,v\Lambda_{p,v} are defined by (7). gp,h,U,v(x)=ℳv((Fh∘ap)U)(x),Λp,v=Λp(span⁡𝔽p{v}).g_{p,h,U,v}(x)=\mathcal M_v\bigl((F_h\circ a_p)_U\bigr)(x),\qquad \Lambda_{p,v}=\Lambda_p(\operatorname{span}_{\mathbb F_p}\{v\}).(7)

Blocks and Eligibility

A folded Gaussian block B=(b,p,h,U,v)B=(b,p,h,U,v) consists of a natural dimension bb, a prime modulus pp, a positive height hh, a usable residue set U⊂𝔽pbU\subset\mathbb F_p^b and a nonzero direction v∈𝔽pbv\in\mathbb F_p^b.✦ Its block group is 𝔽pb\mathbb F_p^b✦, and its real vector space is ℝb\mathbb R^b.✦ For y,ℓ,λ∈ℝby,\ell,\lambda\in\mathbb R^b, the block density gBg_B, lattice ΛB\Lambda_B, grid residue xB(y)x_B(y) and Gaussian weight WB(y,ℓ)W_B(y,\ell) are given by (8); the Boolean vertex translate Tλ(ε)T_\lambda(\varepsilon) is defined there for every ε∈{0,1}b\varepsilon\in\{0,1\}^b. gB=gp,h,U,v,ΛB=Λp,v,xB(y)=xp(y/h),WB(y,ℓ)=hbpγb(y−hℓ),Tλ(ε)=λ+ε.\begin{aligned} g_B&=g_{p,h,U,v},\qquad \Lambda_B=\Lambda_{p,v},\qquad x_B(y)=x_p(y/h),\\ W_B(y,\ell)&=\frac{h^b}{p}\gamma_b(y-h\ell),\qquad T_\lambda(\varepsilon)=\lambda+\varepsilon. \end{aligned}(8)

For positive block dimensions bb, the Gaussian height hbh_b is the prescribed scale in (9). The folded mass M(h)M(h) is the integral of qhq_h on the unit interval✦, and the mean log density μ(h)\mu(h) is its normalized density-weighted logarithmic mean, also given in (9)✦; all logarithms are natural. hb=94log⁡b,M(h)=∫01qh(s)ds,μ(h)=1M(h)∫01qh(s)log⁡qh(s)ds.h_b=\sqrt{\tfrac94\log b},\qquad M(h)=\int_0^1q_h(s)\,ds,\qquad \mu(h)=\frac{1}{M(h)}\int_0^1q_h(s)\log q_h(s)\,ds.(9) For a nonzero modulus pp, the eligible grid residue set ℰ(h,m,D,r;p)⊂(ℤ/pℤ)I\mathcal E(h,m,D,r;p)\subset(\mathbb Z/p\mathbb Z)^I is the finite residue set selected at height hh, reference mean mm, deviation parameter DD and threshold rr. The eligible block residues Eb(p)⊂(ℤ/pℤ)bE_b(p)\subset(\mathbb Z/p\mathbb Z)^b are its specialization in (10). Eb(p)=ℰ(hb,μ(hb),b56/100,b70/100;p).E_b(p)=\mathcal E\bigl(h_b,\mu(h_b),b^{56/100},b^{70/100};p\bigr).(10)✦

The Claim

Theorem 1. Let II be finite, pp prime, h>0h>0, U⊂𝔽pIU\subset\mathbb F_p^I and v∈𝔽pIv\in\mathbb F_p^I, and abbreviate g=gp,h,U,vg=g_{p,h,U,v}. Suppose that τ∈ℝI\tau\in\mathbb R^I satisfies g(xp(τ))>0g(x_p(\tau))>0. There exist z∈Uz\in U and λ∈Λp,v\lambda\in\Lambda_{p,v} satisfying the Gaussian cell bound (11) for every ε∈{0,1}I\varepsilon\in\{0,1\}^I and j∈Ij\in I: τ−λ∈Qp,z,g(xp(τ))=Fh(ap(z))p,λ+ε∈Λp,v,|h(τj−λj−εj)|≤h,e−Nh2/pFh(ap(z))pPh(ap(z),ε)≤Ah(τ−λ,ε)p.\begin{aligned} \tau-\lambda&\in Q_{p,z},\qquad g(x_p(\tau))=\frac{F_h(a_p(z))}{p},\\ \lambda+\varepsilon&\in\Lambda_{p,v},\qquad |h(\tau_j-\lambda_j-\varepsilon_j)|\le h,\\ e^{-Nh^2/p}\frac{F_h(a_p(z))}{p}P_h(a_p(z),\varepsilon) &\le\frac{A_h(\tau-\lambda,\varepsilon)}{p}. \end{aligned}(11)✦ For a folded Gaussian block B=(b,p,h,U,v)B=(b,p,h,U,v) and y∈ℝby\in\mathbb R^b with gB(xB(y))>0g_B(x_B(y))>0, there exist z∈Uz\in U and λ∈ΛB\lambda\in\Lambda_B satisfying the block cell bound (12) for every ε∈{0,1}b\varepsilon\in\{0,1\}^b and j∈{1,…,b}j\in\{1,\ldots,b\}: gB(xB(y))=Fh(ap(z))p,Tλ(ε)∈ΛB,|yj−hTλ(ε)j|≤h,e−bh2/pgB(xB(y))Ph(ap(z),ε)≤WB(y,Tλ(ε)).\begin{aligned} g_B(x_B(y))&=\frac{F_h(a_p(z))}{p},\qquad T_\lambda(\varepsilon)\in\Lambda_B,\\ |y_j-hT_\lambda(\varepsilon)_j|&\le h,\\ e^{-bh^2/p}g_B(x_B(y))P_h(a_p(z),\varepsilon) &\le W_B(y,T_\lambda(\varepsilon)). \end{aligned}(12)✦ There is a constant c>0c>0 with the following property for all sufficiently large bb. Suppose that pp is prime and log⁡p≥bμ(hb)−4b56/100\log p\ge b\mu(h_b)-4b^{56/100}, and let z∈Eb(p)z\in E_b(p) and t∈Qp,zt\in Q_{p,z}. The eligible atom bounds (13) hold for every ε∈{0,1}b\varepsilon\in\{0,1\}^b: Fhb(ap(z))pPhb(ap(z),ε)≤e−cb70/100,Ahb(t,ε)p≤e−cb70/100.\frac{F_{h_b}(a_p(z))}{p}P_{h_b}(a_p(z),\varepsilon) \le e^{-cb^{70/100}},\qquad \frac{A_{h_b}(t,\varepsilon)}{p}\le e^{-cb^{70/100}}.(13)✦

In the paper, the result reads: for its modulus pbp_b, code CbC_b, usable set UbU_b, line maximum gbg_b and error parameter ηb\eta_b, let Lb=ℤb+Cb/pbL_b=\mathbb Z^b+C_b/p_b, t∈ℝbt\in\mathbb R^b, k=⌊pbt⌋k=\lfloor p_bt\rfloor and x=kmod⁡pbx=k\bmod p_b; if gb(x)>0g_b(x)>0, choose a maximizing d∈Cbd\in C_b with z=x−d∈Ubz=x-d\in U_b in standard representatives, so that λ=(k−z)/pb∈Lb\lambda=(k-z)/p_b\in L_b, τ=t−λ=(z+{pbt})/pb∈Qpb,z\tau=t-\lambda=(z+\{p_bt\})/p_b\in Q_{p_b,z} and gb(x)=Fhb(apb(z))/pbg_b(x)=F_{h_b}(a_{p_b}(z))/p_b, and every Boolean word ε∈{0,1}b\varepsilon\in\{0,1\}^b satisfies λ+ε∈Lb\lambda+\varepsilon\in L_b, tj−λj−εj∈[−1,1]t_j-\lambda_j-\varepsilon_j\in[-1,1] at each coordinate and (hbb/pb)γb(hb(t−λ−ε))≥e−ηbgb(x)Phb(apb(z),ε)(h_b^b/p_b)\gamma_b(h_b(t-\lambda-\varepsilon))\ge e^{-\eta_b}g_b(x)P_{h_b}(a_{p_b}(z),\varepsilon), with this individual weight at most e−cb70/100e^{-cb^{70/100}}.

What Is Checked

The cell statements (11) and (12) concern finite coordinate spaces, prime moduli and positive heights with arbitrary usable residue sets; the block direction is nonzero. The eligible atom statement (13) concerns the prescribed height and eligible sets under the stated lower bound on the modulus. The Construction A set in (2) is an additive subgroup of ℝI\mathbb R^I and is closed under multiplication by integers.