OctoInformBeta

OpenAI preprint, September 2026 · §4. Directions that change slowly in scale and position · §4

Lemma 4.6 [Spatial comparison]: At level jj, for xx and yy near shared grid vertices with |x−y|≤104rj|x-y|\le 104r_j, coarser levels weighted by 2s−j2^{s-j}

Nearby switch costs and energies bound the costs at a point.

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

Fix a function h:ℂ→ℂh\colon\mathbb C\to\mathbb C, the vector density; in the main argument it is the gradient ∇g\nabla g of a real C1C^1 function gg with compact support. Over a sequence of radii r0,r1,…r_0,r_1,\dots, indexed from the coarsest level 00, every point xx of the plane carries a random label: a unit direction from a finite palette, drawn at the coarsest level from the grid vertices around xx, and then, from one level to the next, kept or switched toward the directions of the mass of hh seen at nearby vertices of the finer grid. Two costs measure this label law at xx and level kk: how likely a switch is, and how fast the transition rows change as xx moves, in units of the radius.

The question is how these costs at xx are controlled by what happens at a nearby point yy: by the switch costs at yy and by the direction energies of grid vertices close to both points, with the coarser levels weighted by powers of 12\frac12.

Grids, Masses and Directions

The box norm of a point xx of the plane is ∥x∥∞=max⁡(|Re⁡x|,|Im⁡x|)\|x\|_\infty=\max(|\operatorname{Re}x|, |\operatorname{Im}x|), and the tent is the function Λ(t)=max⁡(1−|t|,0)\Lambda(t)=\max(1-|t|,0) on ℝ\mathbb R.✦ For a scale rr and a vertex p=(p1,p2)p=(p_1,p_2) of ℤ2\mathbb Z^2, the grid point of pp is rp=rp1+irp2rp=rp_1+i\,rp_2✦, and the bilinear hat of pp is the function φr,p\varphi_{r,p} on the plane given by φr,p(x)=Λ(Re⁡xr−p1)Λ(Im⁡xr−p2).\varphi_{r,p}(x)=\Lambda\Big(\frac{\operatorname{Re}x}{r}-p_1\Big)\, \Lambda\Big(\frac{\operatorname{Im}x}{r}-p_2\Big).(1)✦ The tent interpolation at scale rr of a real function ff on ℤ2\mathbb Z^2 is the function Irf(x)=∑pφr,p(x)f(p)I_rf(x)=\sum_{p}\varphi_{r,p}(x)f(p) on the plane, to which only the finitely many vertices whose hats do not vanish at xx contribute.✦

The vector mass of hh at a vertex pp and scale rr is the complex number hr,p=∫ℂφr,p(y)h(y)dyh_{r,p}=\int_{\mathbb C}\varphi_{r,p}(y)\,h(y)\,dy✦, and its scalar mass is the modulus ar,p=|hr,p|a_{r,p}=|h_{r,p}|.✦ The mass direction θr(p)\theta_r(p) at pp is the direction of the vector mass: the unit complex number hr,p/|hr,p|h_{r,p}/|h_{r,p}|, and 11 when hr,p=0h_{r,p}=0.✦

The window of radius dd around a vertex qq is the square Wd(q)={z∈ℤ2:|z1−q1|≤d,|z2−q2|≤d}W_d(q)=\{z\in\mathbb Z^2:|z_1-q_1|\le d,\ |z_2-q_2|\le d\} of (2d+1)2(2d+1)^2 vertices.✦ A maximizing vertex of a real function ff on Wd(q)W_d(q) is a vertex of the window at which ff attains its maximum over the window, chosen once and for all.✦ The chosen vertex z*z^* of a vertex zz is the maximizing vertex of the scalar mass p↦ar,pp\mapsto a_{r,p} on W300(z)W_{300}(z)✦, and the chosen direction at zz is Θr(z)=θr(z*)\Theta_r(z)=\theta_r(z^*)✦. The worst vertex of qq is the maximizing vertex of z↦|Θr(z)−θr(q)|2z\mapsto|\Theta_r(z)-\theta_r(q)|^2 on W200(q)W_{200}(q).✦ The direction energy Er(q)E_r(q) of a vertex qq at scale rr is the value of this function at the worst vertex of qq, that is, its maximum over W200(q)W_{200}(q): Er(q)=maxz∈W200(q)|Θr(z)−θr(q)|2.E_r(q)=\max_{z\in W_{200}(q)}\,|\Theta_r(z)-\theta_r(q)|^2.(2)✦

The Label Law

A finite circle palette is a finite set of points of the unit circle such that every point of the unit circle lies at distance less than 1/20001/2000 from one of them.✦ One such palette 𝒫\mathcal P is fixed once and for all✦; its points are the palette labels.✦ The quantization Q(z)Q(z) of a complex number zz of modulus one is a label at distance less than 1/20001/2000 from zz, chosen once and for all✦, and the target label of a vertex zz at scale rr is tr(z)=Q(Θr(z))t_r(z)=Q(\Theta_r(z)).✦

A law on a finite set ι\iota is a probability vector on it: a nonnegative mass at each element, the masses summing to 11.✦ The point mass δn\delta_n at an element nn gives mass 11 to nn and 00 to every other element.✦ The expectation of a real function ff on ι\iota under a law μ\mu is 𝔼μf=∑nμ(n)f(n)\mathbb E_\mu f=\sum_{n}\mu(n)f(n)✦, and the ℓ1\ell^1 distance of two real functions aa and bb on ι\iota is ∥a−b∥1=∑n|a(n)−b(n)|\|a-b\|_1=\sum_{n}|a(n)-b(n)|.✦

The thresholded probability of a distance dd at a tolerance τ>0\tau>0 is Pτ(d)=min⁡{1,((d−τ)+/τ)2}P_\tau(d)=\min\{1,((d-\tau)_+/\tau)^2\}.✦ The switch chance of a label nn at a vertex zz and scale rr is pr(n,z)=Pτ(|n−Θr(z)|),τ=1100.p_r(n,z)=P_\tau\big(|n-\Theta_r(z)|\big),\qquad \tau=\tfrac1{100}.(3)✦

Given an old element n0n_0 of a finite set ι\iota, a target t(z)∈ιt(z)\in\iota and a chance p(z)∈[0,1]p(z)\in[0,1] at each vertex zz of ℤ2\mathbb Z^2, the spatial transition row at scale rr and point xx is the law on ι\iota with masses j↦δn0(j)+∑z∈ℤ2φr,z(x)p(z)(δt(z)(j)−δn0(j)).j\longmapsto\delta_{n_0}(j)+\sum_{z\in\mathbb Z^2}\varphi_{r,z}(x)\,p(z)\, \big(\delta_{t(z)}(j)-\delta_{n_0}(j)\big).(4)✦ The transition row Kr(x;n,⋅)K_r(x;n,\cdot) from a label nn is the spatial transition row on 𝒫\mathcal P with old label nn, targets trt_r and chances pr(n,⋅)p_r(n,\cdot).✦ The initial law νr(x)\nu_r(x) at a point xx and scale rr is the spatial transition row on 𝒫\mathcal P with old label Q(1)Q(1), targets trt_r and chance 11 at every vertex.✦

The label marginals along the radii at a point xx are the laws μk(x)\mu_k(x) on 𝒫\mathcal P defined from the coarsest level by μ0(x)=νr0(x),μk+1(x)(j)=∑n∈𝒫μk(x)(n)Krk+1(x;n,j).\mu_0(x)=\nu_{r_0}(x),\qquad \mu_{k+1}(x)(j)=\sum_{n\in\mathcal P}\mu_k(x)(n)\,K_{r_{k+1}}(x;n,j).(5)✦

The Costs

The switch rate of a label nn at a point xx and scale rr is the interpolated switch chance ρr(x,n)=Ir[pr(n,⋅)](x)\rho_r(x,n)=I_r\big[p_r(n,\cdot)\big](x)✦, and the row derivative of nn at xx and scale rr is the number Dr(x,n)=r∑j∈𝒫|∇xKr(x;n,j)|D_r(x,n)=r\sum_{j\in\mathcal P}|\nabla_xK_r(x;n,j)|✦. The switch cost and the derivative cost are the sequences of functions on the plane given by R0(x)=1,Rk+1(x)=𝔼μk(x)ρrk+1(x,⋅),R_0(x)=1,\qquad R_{k+1}(x)=\mathbb E_{\mu_k(x)}\,\rho_{r_{k+1}}(x,\cdot),(6)✦ T0(x)=r0∑j∈𝒫|∇xνr0(x)(j)|,Tk+1(x)=𝔼μk(x)Drk+1(x,⋅).T_0(x)=r_0\sum_{j\in\mathcal P}\big|\nabla_x\,\nu_{r_0}(x)(j)\big|,\qquad T_{k+1}(x)=\mathbb E_{\mu_k(x)}\,D_{r_{k+1}}(x,\cdot).(7)✦

For vertices q0,q1,…q_0,q_1,\dots of ℤ2\mathbb Z^2, the energy rate at a point yy is the sequence ci(y)=Ri(y)+Eri(qi).c_i(y)=R_i(y)+E_{r_i}(q_i).(8)✦

The causal dyadic filter FF takes a real sequence cc to the sequence FcFc with (Fc)0=c0(Fc)_0=c_0 and (Fc)i+1=ci+1+12(Fc)i(Fc)_{i+1}=c_{i+1}+\frac12(Fc)_i; its terms are (Fc)j=∑s=0j2s−jcs.(Fc)_j=\sum_{s=0}^{j}2^{s-j}c_s.(9)✦

The comparison constant is C0=1280000(|𝒫|+1)C_0=1280000\,(|\mathcal P|+1), where |𝒫||\mathcal P| is the number of labels✦, and the spatial constant is C1=208C02C_1=208\,C_0^2✦.

The Claim

Theorem 1 (Spatial comparison). Let h:ℂ→ℂh\colon\mathbb C\to\mathbb C and real numbers r0,r1,…r_0,r_1,\dots be given, with their label marginals μk\mu_k and costs RkR_k and TkT_k, and let xx and yy be points of the plane.

  1. Let j≥0j\ge0 and let q0,…,qjq_0,\dots,q_j be vertices of ℤ2\mathbb Z^2 such that, for every i≤ji\le j, ri>0r_i>0, ∥x−riqi∥∞<199ri\|x-r_iq_i\|_\infty<199\,r_i and ∥y−riqi∥∞<199ri\|y-r_iq_i\|_\infty<199\,r_i. Then ∥μj(x)−μj(y)∥1≤C0|x−y|∑i=0jci(y)ri.\|\mu_j(x)-\mu_j(y)\|_1\le C_0\,|x-y|\sum_{i=0}^{j}\frac{c_i(y)}{r_i}.(10)✦ If moreover ri=2ri+1r_i=2r_{i+1} for every i<ji<j and |x−y|≤104rj|x-y|\le104\,r_j, then Rj(x)+Tj(x)≤C1(Fc(y))j.R_j(x)+T_j(x)\le C_1\,\big(Fc(y)\big)_j.(11)✦

  2. Let i≥0i\ge0 and let qq be a vertex of ℤ2\mathbb Z^2 such that ri+1>0r_{i+1}>0, ∥x−ri+1q∥∞<199ri+1\|x-r_{i+1}q\|_\infty<199\,r_{i+1} and ∥y−ri+1q∥∞<199ri+1\|y-r_{i+1}q\|_\infty<199\,r_{i+1}. Then Ri+1(x)≤C0(Ri+1(y)+Eri+1(q))+∥μi(x)−μi(y)∥1R_{i+1}(x)\le C_0\big(R_{i+1}(y)+E_{r_{i+1}}(q)\big)+\|\mu_i(x)-\mu_i(y)\|_1(12)✦ and Ti+1(x)≤C0(Ri+1(y)+Eri+1(q))+C0∥μi(x)−μi(y)∥1.T_{i+1}(x)\le C_0\big(R_{i+1}(y)+E_{r_{i+1}}(q)\big)+C_0\,\|\mu_i(x)-\mu_i(y)\|_1.(13)✦

In the paper, the result reads: for a particle, a level 1≤l≤m1\le l\le m and a point x∈D(ql)x\in D(q_l), and with cs=Rs(y)+Es(qs)c_s=R_s(y)+E_s(q_s), one has Rk(x)+Tk(x)≤C∑s=1k2s−kcsR_k(x)+T_k(x)\le C\sum_{s=1}^{k}2^{s-k}c_s for 1≤k≤min⁡(l+1,m)1\le k\le\min(l+1,m), the bound on TkT_k holding almost everywhere, with CC uniform over particles, ll and xx; the paper’s levels 1,…,k1,\dots,k are the levels 0,…,k−10,\dots,k-1 here.

What Is Checked

The claim is stated over the hats and tent interpolation of the grids rℤ2r\mathbb Z^2, the vector masses of hh with their directions, the chosen directions and direction energies of the vertices, the fixed palette with its quantization and target labels, the transition rows, initial laws and label marginals built from the switch chances, the switch and derivative costs, the energy rate, the causal dyadic filter and the constants C0C_0 and C1C_1.

The library proves that a finite circle palette exists✦ and that every real function on a window Wd(q)W_d(q) attains its maximum there.✦ It also proves 0≤pr(n,z)≤10\le p_r(n,z)\le1 for every label nn, vertex zz and scale rr.