OctoInformBeta

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

§4 (setting): neighboring directions, the label law, R_k and T_k: The direction Θz\Theta_z of the heaviest nearby vertex, the direction energy, and transition rows on palette labels with switching and derivative costs

Neighboring directions and laws of labels across scales.

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

Let gg be a real C1C^1 function with compact support on the plane, identified with ℂ\mathbb C, and let h=∇gh=\nabla g be its gradient, a complex density on ℂ\mathbb C. At a scale rr, integrating hh against the bilinear hats of the grid rℤ2r\mathbb Z^2 gives a vector mass at each vertex, and each nonzero mass points in a direction of the unit circle.

Two kinds of object are built from these masses. In position, each vertex adopts the direction of the heaviest vertex near it, and an energy records how far the adopted directions around a vertex stray from its own. In scale, directions are rounded to a finite palette of labels, and along a sequence of radii from the coarsest down, every point xx of the plane carries a law on the labels at each radius, passed from one radius to the next by transition rows that pull a label toward the labels of the vertices near xx, weighted by their hats at xx. Switching costs and derivative costs measure how much these laws move, from one radius to the next and as xx varies.

Hats, Masses and Windows

The tent is the function Λ:ℝ→ℝ\Lambda:\mathbb R\to\mathbb R, Λ(t)=max⁡{1−|t|,0}\Lambda(t)=\max\{1-|t|,0\}.✦ The vertices are the points p=(p1,p2)p=(p_1,p_2) of ℤ2\mathbb Z^2. For a scale rr, the bilinear hat of a vertex pp is the real function φr,p\varphi_{r,p} on ℂ\mathbb C given by φr,p(x)=Λ(Re⁡xr−p1)Λ(Im⁡xr−p2),\varphi_{r,p}(x)=\Lambda\Bigl(\frac{\operatorname{Re}x}{r}-p_1\Bigr)\, \Lambda\Bigl(\frac{\operatorname{Im}x}{r}-p_2\Bigr),(1)✦ and the tent interpolation at scale rr of a function f:ℤ2→ℝf:\mathbb Z^2\to\mathbb R is the function IrfI_rf on ℂ\mathbb C given by Irf(x)=∑p∈ℤ2φr,p(x)f(p).I_rf(x)=\sum_{p\in\mathbb Z^2}\varphi_{r,p}(x)\,f(p).(2)✦

The vector mass of a complex density hh on ℂ\mathbb C 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 the scalar mass of hh there is its modulus ar,p=|hr,p|a_{r,p}=|h_{r,p}|.✦

The window Wd(q)W_d(q) of radius d∈ℕd\in\mathbb N around a vertex qq is the set of vertices pp with |p1−q1|≤d|p_1-q_1|\le d and |p2−q2|≤d|p_2-q_2|\le d.✦ A maximizing vertex of a real function ff on ℤ2\mathbb Z^2 over Wd(q)W_d(q) is a vertex of Wd(q)W_d(q) at which ff takes its largest value on Wd(q)W_d(q); one such vertex is chosen once and for all, and called the maximizing vertex.✦

Palettes and Laws on Finite Sets

A finite circle palette is a finite set of points of the unit circle in ℂ\mathbb C such that every point of the unit circle lies at distance less than 1/20001/2000 from one of them.✦ The fixed palette 𝒫*\mathcal P_\ast is a finite circle palette chosen once and for all✦, and the quantization Q(z)Q(z) of a point zz of the unit circle is a point of 𝒫*\mathcal P_\ast chosen with |z−Q(z)|<1/2000|z-Q(z)|<1/2000.✦

A law on a finite set II is a function μ:I→ℝ\mu:I\to\mathbb R with μ(i)≥0\mu(i)\ge0 for every i∈Ii\in I and ∑i∈Iμ(i)=1\sum_{i\in I}\mu(i)=1.✦ The point mass δi\delta_i at i∈Ii\in I is the law with δi(i)=1\delta_i(i)=1 and δi(j)=0\delta_i(j)=0 for j≠ij\neq i✦, and the expectation of a function f:I→ℝf:I\to\mathbb R under a law μ\mu is 𝔼μf=∑i∈Iμ(i)f(i)\mathbb E_\mu f=\sum_{i\in I}\mu(i)\,f(i).✦

Let rr be a scale, xx a point of ℂ\mathbb C, i0i_0 an element of II, t:ℤ2→It:\mathbb Z^2\to I an assignment of targets to the vertices and s:ℤ2→[0,1]s:\mathbb Z^2\to[0,1] an assignment of switch chances. The spatial transition row from i0i_0 is the law on II Sr,x(i0;t,s)=δi0+∑z∈ℤ2φr,z(x)s(z)(δt(z)−δi0).S_{r,x}(i_0;t,s)=\delta_{i_0}+\sum_{z\in\mathbb Z^2}\varphi_{r,z}(x)\,s(z)\, \bigl(\delta_{t(z)}-\delta_{i_0}\bigr).(3)✦

The Definition

Definition 1. Let hh be a complex density on ℂ\mathbb C and rr a scale; the objects θq\theta_q, w(z)w(z), Θz\Theta_z, E(q)E(q), nzn_z and p(n,z)p(n,z) below depend on hh and rr, which are left out of the notation. The mass direction θq\theta_q of a vertex qq is the direction of the vector mass hr,qh_{r,q}: θq={hr,q/|hr,q|if hr,q≠0,1if hr,q=0.\theta_q=\begin{cases}h_{r,q}/|h_{r,q}| & \text{if } h_{r,q}\neq0,\\ 1 & \text{if } h_{r,q}=0.\end{cases}(4)✦ The chosen vertex w(z)w(z) of a vertex zz is the maximizing vertex of the scalar mass over W300(z)W_{300}(z), and the chosen direction Θz\Theta_z of zz is the mass direction there: w(z)∈W300(z),ar,w(z)=maxw∈W300(z)ar,w,w(z)\in W_{300}(z),\qquad a_{r,w(z)}=\max_{w\in W_{300}(z)}a_{r,w},(5)✦ Θz=θw(z).\Theta_z=\theta_{w(z)}.(6)✦ The worst vertex v(q)v(q) of a vertex qq is the maximizing vertex of z↦|Θz−θq|2z\mapsto|\Theta_z-\theta_q|^2 over W200(q)W_{200}(q)✦, and the direction energy of qq is the value there: E(q)=|Θv(q)−θq|2=maxz∈W200(q)|Θz−θq|2.E(q)=|\Theta_{v(q)}-\theta_q|^2=\max_{z\in W_{200}(q)}|\Theta_z-\theta_q|^2.(7)✦

The palette labels are the points of the fixed palette; their set is 𝒩=𝒫*.\mathcal N=\mathcal P_\ast.(8)✦ The target label of a vertex zz is the quantization of its chosen direction: nz=Q(Θz)∈𝒩.n_z=Q(\Theta_z)\in\mathcal N.(9)✦ The thresholded probability with tolerance τ\tau is the function on ℝ\mathbb R Pτ(d)=min⁡{1,(τ−1(d−τ)+)2},P_\tau(d)=\min\bigl\{1,\bigl(\tau^{-1}(d-\tau)_+\bigr)^2\bigr\},(10)✦ and, with τ=1/100\tau=1/100, the switch chance of a label n∈𝒩n\in\mathcal N at a vertex zz is p(n,z)=Pτ(|n−Θz|).p(n,z)=P_\tau\bigl(|n-\Theta_z|\bigr).(11)✦

At a point x∈ℂx\in\mathbb C, the initial label law πr0(x)\pi^0_r(x), with mass πr0(x;j)\pi^0_r(x;j) at a label jj, and the transition row πr(x;n,⋅)\pi_r(x;n,\cdot) from a label nn, with mass πr(x;n,j)\pi_r(x;n,j) at jj, are the spatial transition rows (3) toward the target labels πr0(x)=Sr,x(Q(1);z↦nz,z↦1),\pi^0_r(x)=S_{r,x}\bigl(Q(1);\,z\mapsto n_z,\,z\mapsto1\bigr),(12)✦ πr(x;n,⋅)=Sr,x(n;z↦nz,z↦p(n,z)).\pi_r(x;n,\cdot)=S_{r,x}\bigl(n;\,z\mapsto n_z,\,z\mapsto p(n,z)\bigr).(13)✦ The switching rate of nn at xx is the tent interpolation of its switch chances, and the derivative cost of the row from nn is the radius times the sum, over the labels jj, of the norms of the gradients in xx of the masses that the row gives to jj: ρr(x;n)=Ir(p(n,⋅))(x),\rho_r(x;n)=I_r\bigl(p(n,\cdot)\bigr)(x),(14)✦ Dr(x;n)=r∑j∈𝒩|∇xπr(x;n,j)|.D_r(x;n)=r\sum_{j\in\mathcal N}\bigl|\nabla_x\pi_r(x;n,j)\bigr|.(15)✦

Along a sequence of radii r0,r1,r2,…r_0,r_1,r_2,\dots, indexed from the coarsest level, the label marginals are the laws ζk(x)\zeta_k(x) on 𝒩\mathcal N, with mass ζk(x;n)\zeta_k(x;n) at nn, given by ζ0(x)=πr00(x),ζk+1(x;j)=∑n∈𝒩ζk(x;n)πrk+1(x;n,j),\zeta_0(x)=\pi^0_{r_0}(x),\qquad \zeta_{k+1}(x;j)=\sum_{n\in\mathcal N}\zeta_k(x;n)\,\pi_{r_{k+1}}(x;n,j),(16)✦ the switching costs are the functions RkR_k on ℂ\mathbb C given by R0(x)=1,Rk+1(x)=𝔼ζk(x)ρrk+1(x;⋅),R_0(x)=1,\qquad R_{k+1}(x)=\mathbb E_{\zeta_k(x)}\,\rho_{r_{k+1}}(x;\cdot),(17)✦ and the derivative costs are the functions TkT_k on ℂ\mathbb C given by T0(x)=r0∑j∈𝒩|∇xπr00(x;j)|,Tk+1(x)=𝔼ζk(x)Drk+1(x;⋅).T_0(x)=r_0\sum_{j\in\mathcal N}\bigl|\nabla_x\pi^0_{r_0}(x;j)\bigr|,\qquad T_{k+1}(x)=\mathbb E_{\zeta_k(x)}\,D_{r_{k+1}}(x;\cdot).(18)✦

In the paper, the result reads: w(z)w(z) maximizes awa_w over |w−z|∞≤300rk|w-z|_\infty\le300r_k, Θz=θw(z)\Theta_z=\theta_{w(z)} and nz∈𝒩n_z\in\mathcal N with |nz−Θz|≤τ/20|n_z-\Theta_z|\le\tau/20, and Ek(q)=max⁡|z−q|∞≤200rk|Θz−θq|2E_k(q)=\max_{|z-q|_\infty\le200r_k}|\Theta_z-\theta_q|^2; the law PxP_x on 𝒩m\mathcal N^m of the labels (n1,…,nm)(n_1,\dots,n_m), where mm is the number of levels, has the initial row π1(x)=∑zφz(x)δnz\pi_1(x)=\sum_z\varphi_z(x)\delta_{n_z} and the transition rows πk(x;n,⋅)=δn+∑zφz(x)pk(n,z)(δnz−δn)\pi_k(x;n,\cdot)=\delta_n+\sum_z\varphi_z(x)p_k(n,z)(\delta_{n_z}-\delta_n) with pk(n,z)=P(|n−Θz|)p_k(n,z)=P(|n-\Theta_z|) and P(d)=min⁡{1,τ−2(d−τ)+2}P(d)=\min\{1,\tau^{-2}(d-\tau)_+^2\}, its marginals are the ζk\zeta_k, its switch rates are Rk(x)=Px(nk≠nk−1)R_k(x)=P_x(n_k\neq n_{k-1}) with R1=Rm+1=1R_1=R_{m+1}=1, and the TkT_k are the scaled costs of the row derivatives, the paper’s level kk being the page’s index k−1k-1.

What Is Checked

The definitions are stated over a complex density hh on ℂ\mathbb C, a scale rr or a sequence of radii r0,r1,…r_0,r_1,\dots, the vertices of ℤ2\mathbb Z^2 with their bilinear hats and windows, and the fixed palette 𝒫*\mathcal P_\ast with its quantization QQ.

The library proves that a finite circle palette exists✦ and that every real function on ℤ2\mathbb Z^2 attains its largest value on each window Wd(q)W_d(q).✦ It also proves that every switch chance (11) lies between 00 and 11.