OctoInformBeta

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

Lemma 4.2 [Good labels near a controlled vertex]: If Er(q)≤(1/1000)2E_r(q)\le(1/1000)^2 and |x−rq|∞≤50r|x-rq|_\infty\le 50r, the label laws at xx give no mass to labels bad for θr(q)\theta_r(q)

Bad labels get no mass near a low-energy vertex.

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 h:ℂ→ℂh\colon\mathbb{C}\to\mathbb{C} be a vector density, that is, an integrable function, in the main argument the gradient ∇g\nabla g of a real C1C^1 function gg with compact support. Averaged against the piecewise bilinear tent functions centered at the vertices of a square grid, hh has a direction at each vertex, a point of the unit circle. A random label, a direction taken from a fixed finite palette, is followed at a fixed point xx from coarse scales to fine ones: at each scale it stays put while it agrees with the directions near xx up to a small tolerance, and otherwise it jumps to one of them.

The question is which labels can carry mass at a point xx near a vertex qq when the directions around qq all lie close to the direction of mass at qq.

Masses and Directions on the Grid

Let r>0r>0 be a scale. The grid point of index p∈ℤ2p\in\mathbb{Z}^2 is the point rp=rp1+irp2rp=rp_1+i\,rp_2 of ℂ\mathbb{C}✦, and the max norm of x∈ℂx\in\mathbb{C} is |x|∞=max⁡(|Re⁡x|,|Im⁡x|)|x|_\infty=\max(|\operatorname{Re}x|,|\operatorname{Im}x|)✦. The tent is the function Λ(t)=max⁡(1−|t|,0)\Lambda(t)=\max(1-|t|,0) on ℝ\mathbb{R}✦, and the bilinear hat of index pp at scale rr is the function 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)✦ The tent interpolation at scale rr of a real function ff on ℤ2\mathbb{Z}^2 is the function Irf(x)=∑p∈ℤ2φr,p(x)f(p)I_rf(x)=\sum_{p\in\mathbb{Z}^2}\varphi_{r,p}(x)f(p) on ℂ\mathbb{C}, a finite sum at each point.✦

For a vector density hh on ℂ\mathbb{C}, the vector mass of hh at a vertex pp 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 there is ar,p=|hr,p|a_{r,p}=|h_{r,p}|✦. The direction of a vector vv of a normed space, with fallback ee, is v/∥v∥v/\|v\| when v≠0v\neq0 and ee when v=0v=0.✦ The mass direction θr(q)\theta_r(q) at a vertex qq is the direction of hr,qh_{r,q} with fallback 11.✦

The grid window of radius d∈ℕd\in\mathbb{N} 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.✦ The maximizing vertex of a real function ff on ℤ2\mathbb{Z}^2 in Wd(q)W_d(q) is a vertex of Wd(q)W_d(q) at which ff attains its maximum over the window, chosen once and for all.✦ The chosen vertex cr(z)c_r(z) of a vertex zz is the maximizing vertex of the scalar mass p↦ar,pp\mapsto a_{r,p} in W300(z)W_{300}(z)✦, and the chosen direction at zz is the mass direction at its chosen vertex, Θr(z)=θr(cr(z))\Theta_r(z)=\theta_r(c_r(z))✦. The worst vertex wr(q)w_r(q) of qq is the maximizing vertex of z↦|Θr(z)−θr(q)|2z\mapsto|\Theta_r(z)-\theta_r(q)|^2 in W200(q)W_{200}(q)✦, and the direction energy of qq is the largest squared disagreement between θr(q)\theta_r(q) and the chosen directions near qq: Er(q)=|Θr(wr(q))−θr(q)|2=maxz∈W200(q)|Θr(z)−θr(q)|2.E_r(q)=|\Theta_r(w_r(q))-\theta_r(q)|^2=\max_{z\in W_{200}(q)}|\Theta_r(z)-\theta_r(q)|^2.(2)✦

Labels and Their Laws

A finite circle palette is a finite set of complex numbers of modulus one such that every point of the unit circle lies at distance less than 1/20001/2000 from one of them.✦ One such palette 𝒩\mathcal{N} is fixed✦, and its elements are the labels✦. The quantization Q(z)Q(z) of a complex number zz of modulus one is a label with |z−Q(z)|<1/2000|z-Q(z)|<1/2000, chosen once and for all.✦

For n∈ℂn\in\mathbb{C} and real ss, the cone coordinate ℓn,s\ell_{n,s} is the real-linear functional on ℂ\mathbb{C} given by ℓn,s(z)=13⟨n,z⟩+s⟨in,z⟩\ell_{n,s}(z)=\frac{1}{\sqrt3}\langle n,z\rangle+s\langle in,z\rangle, where ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle is the real inner product of ℂ=ℝ2\mathbb{C}=\mathbb{R}^2.✦ A point z∈ℂz\in\mathbb{C} lies in the good sector of nn when ℓn,1(z)≥0\ell_{n,1}(z)\ge0 and ℓn,−1(z)≥0\ell_{n,-1}(z)\ge0; for |n|=1|n|=1 this sector is the closed cone of half-angle 30∘30^\circ about nn.✦ A label nn is good for a direction θ\theta when θ\theta lies in the good sector of nn, and bad for it otherwise.

A law on a finite set AA is a probability distribution on AA: a function μ:A→[0,∞)\mu\colon A\to[0,\infty) with ∑a∈Aμ(a)=1\sum_{a\in A}\mu(a)=1.✦ The point mass δa\delta_a at a∈Aa\in A is the law with δa(b)=1\delta_a(b)=1 for b=ab=a and δa(b)=0\delta_a(b)=0 otherwise.✦ Given an old element o∈Ao\in A, a target T(z)∈AT(z)\in A and a chance ϖ(z)∈[0,1]\varpi(z)\in[0,1] at each vertex zz, the spatial transition row at scale rr and point xx is the law RR on AA that keeps oo with weight one minus the interpolated chance at xx and hands the rest to the targets of the vertices near xx, in proportion to their hats and chances: R(j)=δo(j)+Irfj(x),fj(z)=ϖ(z)(δT(z)(j)−δo(j)).R(j)=\delta_o(j)+I_rf_j(x),\qquad f_j(z)=\varpi(z)\bigl(\delta_{T(z)}(j)-\delta_o(j)\bigr).(3)✦

The thresholded probability of a distance dd at tolerance τ\tau is Pτ(d)=min⁡{1,((d−τ)+/τ)2}P_\tau(d)=\min\{1,((d-\tau)_+/\tau)^2\}.✦ For a vector density hh, the switch chance of a label nn at a vertex zz is σr,n(z)=Pτ(|n−Θr(z)|)\sigma_{r,n}(z)=P_\tau(|n-\Theta_r(z)|) with τ=1/100\tau=1/100✦, and the target label of zz is Tr(z)=Q(Θr(z))T_r(z)=Q(\Theta_r(z))✦. The scale transition row Kr,x(m,⋅)K_{r,x}(m,\cdot) from a label mm at a point xx is the row (3) on 𝒩\mathcal{N} with old label mm, targets TrT_r and chances σr,m\sigma_{r,m}.✦ The initial label law νr,x\nu_{r,x} is the row (3) with old label Q(1)Q(1), targets TrT_r and chance 11 at every vertex; it gives each label jj the mass ∑z:Tr(z)=jφr,z(x)\sum_{z:\,T_r(z)=j}\varphi_{r,z}(x).✦

A sequence of radii (rk)k≥0(r_k)_{k\ge0} assigns a scale rkr_k to each level k≥0k\ge0 and descends from the coarsest level k=0k=0. The label marginal of hh at level kk and point xx is the law μk(x)\mu_k(x) on 𝒩\mathcal{N} given by the initial law at level 00 and, at each later level, by moving the previous marginal through the scale transition rows: μ0(x)=νr0,x,μk+1(x)(j)=∑m∈𝒩μk(x)(m)Krk+1,x(m,j).\mu_0(x)=\nu_{r_0,x},\qquad \mu_{k+1}(x)(j)=\sum_{m\in\mathcal{N}}\mu_k(x)(m)\,K_{r_{k+1},x}(m,j).(4)✦

The Claim

Theorem 1. Let hh be a vector density on ℂ\mathbb{C}, let (rk)k≥0(r_k)_{k\ge0} be a sequence of radii and kk a level with rk>0r_k>0, let q∈ℤ2q\in\mathbb{Z}^2 be a vertex with Erk(q)≤(1/1000)2E_{r_k}(q)\le(1/1000)^2, and let x∈ℂx\in\mathbb{C} satisfy |x−rkq|∞≤50rk|x-r_kq|_\infty\le50r_k. Then every label nn bad for θrk(q)\theta_{r_k}(q) has no mass in the label marginal at level kk: μk(x)(n)=0.\mu_k(x)(n)=0.(5)✦ Likewise, at a single scale r>0r>0: if q∈ℤ2q\in\mathbb{Z}^2 has Er(q)≤(1/1000)2E_r(q)\le(1/1000)^2 and x∈ℂx\in\mathbb{C} has |x−rq|∞≤50r|x-rq|_\infty\le50r, then every label nn bad for θr(q)\theta_r(q) has no mass in the scale transition row from any label mm, Kr,x(m,n)=0,K_{r,x}(m,n)=0,(6)✦ nor in the initial label law, νr,x(n)=0.\nu_{r,x}(n)=0.(7)✦

In the paper, the result reads: set c*=(τ/10)2c_*=(\tau/10)^2; if q∈Vkq\in V_k, Ek(q)≤c*E_k(q)\le c_* and |x−q|∞≤50rk|x-q|_\infty\le50r_k, then (q,nk)(q,n_k) is good with PxP_x-probability one, where the paper’s level kk is the page’s index k−1k-1, VkV_k is the set of grid vertices at that level, nkn_k is the label at that level and PxP_x is the law of the labels at xx.

At τ=1/100\tau=1/100 the constant c*c_* is (1/1000)2(1/1000)^2, and the label at a level is good with probability one exactly when its marginal vanishes on bad labels, as in (5).

What Is Checked

The claim is stated over the vector masses, mass directions, chosen directions and direction energies (2) of a vector density on the grid of each scale, and over the initial label law, the scale transition rows and the label marginals (4), which the fixed palette, the quantization, the switch chances and the spatial transition rows (3) build from them.

The library proves that a finite circle palette exists✦, that every real function on a grid window attains its maximum there✦, and that every switch chance lies between 00 and 11.