OctoInformBeta

OpenAI preprint, September 2026 · §3. Arithmetic and nested partitions · §3

Lemma 3.1 [Grid error]: At an infinite place the mean depth is the raw depth plus the error mean; at a finite place the depth is the raw depth

The grid error has a universal mean and an exponential tail.

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 raw depth of two distinct points z,wz,w of the complex plane is −log⁡|z−w|-\log|z-w|. Take a dyadic grid of the plane, randomly shifted, rotated and dilated, and the smallest of its square cells that contains both points: minus the logarithm of that cell’s side is a second depth of the pair. The question is how far it lies from the raw depth, and how the gap depends on the pair.

The library asks it for any two distinct complex numbers, and then, place by place, for the points of a set PP carried into a number field KK. An injective coordinate Zj:P→KZ_j\colon P\to K, with j∈{1,2}j\in\{1,2\}, and a place vv of KK give two distinct points x,y∈Px,y\in P the raw depth ij(x,y)=−log⁡|Zj(y)−Zj(x)|v,i_j(x,y)=-\log|Z_j(y)-Z_j(x)|_v,(1) where |⋅|v|\cdot|_v is the absolute value of vv. At an infinite place, |α|v=|σv(α)||\alpha|_v=|\sigma_v(\alpha)| for a complex embedding σv\sigma_v of KK inducing vv, and ij(x,y)i_j(x,y) is the raw depth of the complex points σvZj(x)\sigma_vZ_j(x) and σvZj(y)\sigma_vZ_j(y).

Random Dyadic Grids

An axis sequence is a real sequence a=(ai)i≥0a=(a_i)_{i\ge0}. The axis law is the probability law on axis sequences aa under which the terms are independent, a0a_0 is uniform on (0,1](0,1], and an+1a_{n+1} is a fair binary step of length 2n2^n, equal to 00 or 2n2^n with probability 1/21/2 each.✦ The offset of aa at an integer scale mm is the partial sum sm(a)=a0+a1+…+amax⁡(m,0).s_m(a)=a_0+a_1+\dots+a_{\max(m,0)}.(2)✦

For a length l>0l>0 and an offset ss, the bin index of a real number tt is the integer ⌊(t−s)/l⌋\lfloor (t-s)/l\rfloor, the kk with t∈[s+kl,s+(k+1)l)t\in[s+kl,s+(k+1)l).✦

A grid sample is a pair ω=(a(1),a(2))\omega=(a^{(1)},a^{(2)}) of axis sequences✦; under the grid sample law the two are independent, each with the axis law.✦ Two points x=(x1,x2)x=(x_1,x_2) and y=(y1,y2)y=(y_1,y_2) of ℝ2\mathbb R^2 lie in the same cell at scale m∈ℤm\in\mathbb Z, written x∼mωyx\sim^\omega_m y, when their coordinates have equal bin indices for the length 2m2^m and the offsets (2): ⌊xk−sm(a(k))2m⌋=⌊yk−sm(a(k))2m⌋,k=1,2.\Big\lfloor\frac{x_k-s_m(a^{(k)})}{2^m}\Big\rfloor =\Big\lfloor\frac{y_k-s_m(a^{(k)})}{2^m}\Big\rfloor,\qquad k=1,2.(3)✦ The minimal shared scale mω(x,y)m_\omega(x,y) of xx and yy is the infimum of the integers mm with x∼mωyx\sim^\omega_m y✦, and their grid depth is gω(x,y)=−mω(x,y)log⁡2.g_\omega(x,y)=-m_\omega(x,y)\log 2.(4)✦

The Phase and the Grid Error

A phase is a pair (ξ,θ)(\xi,\theta) of real numbers. The phase law is the probability law on phases (ξ,θ)(\xi,\theta) under which ξ\xi is uniform on (0,log⁡2](0,\log 2] and θ\theta, independently, uniform on (0,2π](0,2\pi].✦ A sample Ω=((ξ,θ),ω)\Omega=((\xi,\theta),\omega) is a phase with a grid sample✦; the full law is the product of the phase law and the grid sample law, and ℙ\mathbb P and 𝔼\mathbb E refer to it. The phase acts on ℂ\mathbb C by the rotation scaling Rξ,θ(z)=e−ξeiθz,R_{\xi,\theta}(z)=e^{-\xi}e^{i\theta}z,(5)✦ with ℂ\mathbb C identified with ℝ2\mathbb R^2 through real and imaginary parts.

The hierarchy depth of two complex numbers z,wz,w under a sample Ω\Omega is DΩ(z,w)=gω(Rξ,θ(z),Rξ,θ(w))−ξ,D_\Omega(z,w)=g_\omega\big(R_{\xi,\theta}(z),R_{\xi,\theta}(w)\big)-\xi,(6)✦ and their grid error is the excess of the hierarchy depth over the raw depth, ΔΩ(z,w)=DΩ(z,w)+log⁡|z−w|.\Delta_\Omega(z,w)=D_\Omega(z,w)+\log|z-w|.(7)✦

For real numbers tt and θ\theta, the polar mean g‾(t,θ)\bar g(t,\theta) is the mean grid depth, over the grid sample law, between the origin and the point of logarithmic radius tt and angle θ\theta, plus tt: g‾(t,θ)=𝔼ωgω((0,0),(etcos⁡θ,etsin⁡θ))+t.\bar g(t,\theta)=\mathbb E_\omega\,g_\omega\big((0,0),(e^t\cos\theta,e^t\sin\theta)\big)+t.(8)✦ The error mean κ\kappa is the universal constant κ=12πlog⁡2∫0log⁡2∫02πg‾(−ξ,θ)dθdξ.\kappa=\frac{1}{2\pi\log 2}\int_0^{\log 2}\!\!\int_0^{2\pi}\bar g(-\xi,\theta)\,d\theta\,d\xi.(9)✦

Depths at a Place

An ultrametric depth on a set PP is a symmetric function d:P×P→ℝd\colon P\times P\to\mathbb R such that any three distinct points x,y,zx,y,z of PP satisfy d(x,z)≥min⁡{d(x,y),d(y,z)}.d(x,z)\ge\min\{d(x,y),d(y,z)\}.(10)✦

Let vv be an infinite place of KK. A sample Ω\Omega is good at vv when the hierarchy depths DΩ(σvZj(x),σvZj(y))D_\Omega(\sigma_vZ_j(x),\sigma_vZ_j(y)), for x,y∈Px,y\in P, satisfy (10) on every triple of distinct points of PP.✦ The archimedean depth djd_j at vv under a sample Ω\Omega is the function on P×PP\times P given by dj(x,y)={DΩ(σvZj(x),σvZj(y))if Ω is good at v and x≠y,0otherwise.d_j(x,y)=\begin{cases} D_\Omega\big(\sigma_vZ_j(x),\sigma_vZ_j(y)\big) &\text{if $\Omega$ is good at $v$ and $x\ne y$,}\\ 0&\text{otherwise.} \end{cases}(11)✦

The Claim

Theorem 1. For any two points x,y∈ℝ2x,y\in\mathbb R^2 and almost every grid sample ω\omega, x∼nωyfor some n∈ℕ.x\sim^\omega_n y\quad\text{for some } n\in\mathbb N.(12)✦ For distinct z,w∈ℂz,w\in\mathbb C, the grid error ΔΩ(z,w)\Delta_\Omega(z,w) of a sample Ω\Omega drawn from the full law is integrable, 𝔼|ΔΩ(z,w)|<∞,\mathbb E\,|\Delta_\Omega(z,w)|<\infty,(13)✦ it has mean 𝔼ΔΩ(z,w)=κ,\mathbb E\,\Delta_\Omega(z,w)=\kappa,(14)✦ for every u≥0u\ge0 its tail satisfies ℙ(|ΔΩ(z,w)|>u)≤4e−u,\mathbb P\big(|\Delta_\Omega(z,w)|>u\big)\le 4e^{-u},(15)✦ and its absolute value has mean 𝔼|ΔΩ(z,w)|≤4.\mathbb E\,|\Delta_\Omega(z,w)|\le 4.(16)✦ Let KK be a number field and Zj:P→KZ_j\colon P\to K injective. At an infinite place vv of KK, the archimedean depth of distinct x,y∈Px,y\in P has mean 𝔼dj(x,y)=ij(x,y)+κ.\mathbb E\,d_j(x,y)=i_j(x,y)+\kappa.(17)✦ At a finite place vv of KK, the depth djd_j is the function on P×PP\times P that vanishes on the diagonal and is the raw depth off it: dj(x,y)=ij(x,y)(x≠y).d_j(x,y)=i_j(x,y)\qquad(x\ne y).(18)✦

In the paper, the result reads, with Δj=dj−ij\Delta_j=d_j-i_j the grid error at a place vv: at every complex place the depths are almost surely finite on distinct pairs; the law of Δj(x,y)\Delta_j(x,y) is independent of the pair, the place and the configuration, with finite mean an absolute constant κ\kappa, and ℙ(|Δj|>u)≤C0e−b0u\mathbb P(|\Delta_j|>u)\le C_0e^{-b_0u} for u≥0u\ge0, where C0C_0 and b0b_0 are constants; at finite places Δj=0\Delta_j=0.

What Is Checked

The claim is stated over the random dyadic grid: the axis law, the grid sample law, the phase law and the full law; the cells (3) and the grid depth (4); the hierarchy depth (6), the grid error (7) and the constant κ\kappa of (9); and, at the places of KK, the raw depth (1) and the archimedean depth (11). The library proves that the depth at a finite place, and the archimedean depth under every sample, are ultrametric depths on PP in the sense of (10).