OctoInformBeta

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

Lemma 4.1 [Neighboring direction budget]: For hh supported in a disk, summed over dyadic levels; at one level, by the drop in scalar mass over twelve doublings

Mass-weighted direction energy is controlled by the total mass.

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:\mathbb{C}\to\mathbb{C} be an integrable vector density on the plane, such as the gradient ∇g\nabla g of a compactly supported C1C^1 function gg, and let V=∫ℂ|h|V=\int_{\mathbb{C}}|h| be its total mass. Testing hh against the bilinear hats, the products of two tent functions centered at the vertices of a dyadic grid, gives each vertex of the grid a vector mass, a complex number whose direction is the mass direction at that vertex. From one vertex to the next these directions may turn.

The question is whether the turning is paid for by the total mass: whether the squared distance between the direction at a vertex and the directions chosen at the vertices around it, weighted by the size of the mass at the vertex and summed over the vertices of any number of scales, stays below a fixed multiple of VV.

Hats and Masses

The tent is the function Λ:ℝ→ℝ\Lambda:\mathbb{R}\to\mathbb{R} given by Λ(x)=max⁡(1−|x|,0)\Lambda(x)=\max(1-|x|,0).✦ For a scale r>0r>0 and a vertex p=(p1,p2)p=(p_1,p_2) of ℤ2\mathbb{Z}^2, the bilinear hat φr,p\varphi_{r,p} is the real function on ℂ\mathbb{C} centered at the point r(p1+ip2)r(p_1+ip_2) given by φr,p(x)=Λ(Re⁡x/r−p1)Λ(Im⁡x/r−p2).\varphi_{r,p}(x)=\Lambda\bigl(\operatorname{Re}x/r-p_1\bigr)\, \Lambda\bigl(\operatorname{Im}x/r-p_2\bigr).(1)✦

The vector mass of hh at the vertex pp and scale rr is the complex number hr,p=∫ℂφr,p(y)h(y)dy.h_{r,p}=\int_{\mathbb{C}}\varphi_{r,p}(y)\,h(y)\,dy.(2)✦ The scalar mass ar,pa_{r,p} of hh at a vertex pp and scale rr is the modulus |hr,p||h_{r,p}| of the vector mass hr,ph_{r,p}. The mass direction θr(p)\theta_r(p) at pp is the direction of hr,ph_{r,p}: the unit complex number hr,p/|hr,p|h_{r,p}/|h_{r,p}| when hr,p≠0h_{r,p}\neq0, and 11 when hr,p=0h_{r,p}=0.✦

Levels are counted from a finest scale rr: level k∈ℕk\in\mathbb{N} has the dyadic scale 2kr2^kr, and level 00 is rr itself.✦ For a natural number NN, the index box GNG_N is the set of vertices p∈ℤ2p\in\mathbb{Z}^2 with |p1|≤N|p_1|\le N and |p2|≤N|p_2|\le N.✦

Chosen Directions and Their Energy

For a natural number dd and a vertex q∈ℤ2q\in\mathbb{Z}^2, the window Wd(q)W_d(q) is the square of the (2d+1)2(2d+1)^2 vertices z∈ℤ2z\in\mathbb{Z}^2 with |z1−q1|≤d|z_1-q_1|\le d and |z2−q2|≤d|z_2-q_2|\le d.✦ For a real function ff on ℤ2\mathbb{Z}^2, a maximizing vertex of ff in Wd(q)W_d(q) is a vertex of Wd(q)W_d(q) at which ff attains its maximum over Wd(q)W_d(q), one such vertex being fixed once and for all.✦

At scale rr, 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\bigl(c_r(z)\bigr).(3)✦

The worst vertex wr(q)w_r(q) of a vertex 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).✦ The direction energy of qq at scale rr is the real number Er(q)=|Θr(wr(q))−θr(q)|2=maxz∈W200(q)|Θr(z)−θr(q)|2.E_r(q)=\bigl|\Theta_r(w_r(q))-\theta_r(q)\bigr|^2 =\max_{z\in W_{200}(q)}\bigl|\Theta_r(z)-\theta_r(q)\bigr|^2.(4)✦

The Claim

Theorem 1 (Neighboring direction budget). Let r>0r>0 and R≥0R\ge0 be real numbers and NN a natural number with R/r+1≤NR/r+1\le N, and let h:ℂ→ℂh:\mathbb{C}\to\mathbb{C} be an integrable function that vanishes outside the closed disk of radius RR about the origin. Then for every natural number mm, ∑k=0m−1∑q∈GNa2kr,qE2kr(q)≤576⋅40972∫ℂ|h(y)|dy,\sum_{k=0}^{m-1}\sum_{q\in G_N}a_{2^kr,q}\,E_{2^kr}(q) \le 576\cdot4097^2\int_{\mathbb{C}}|h(y)|\,dy,(5)✦ and at the single scale rr, ∑q∈GNar,qEr(q)≤48⋅40972(∑q∈GNar,q−∑p∈GNa212r,p).\sum_{q\in G_N}a_{r,q}\,E_r(q) \le 48\cdot4097^2\Bigl(\sum_{q\in G_N}a_{r,q}-\sum_{p\in G_N}a_{2^{12}r,p}\Bigr).(6)✦

In the paper, the result reads: there is an absolute constant CC with ∑k=1m∑q∈VkaqEk(q)≤CV\sum_{k=1}^m\sum_{q\in V_k}a_qE_k(q)\le CV, where the paper’s level kk has scale 2−k2^{-k}, VkV_k is the set of vertices of that level, aqa_q and Ek(q)E_k(q) are the scalar mass and the direction energy of a vertex qq there, and V=∫|h|V=\int|h| is the total mass.

What Is Checked

The claim is stated over the scalar masses ar,qa_{r,q} of an integrable density hh, taken against the bilinear hats (1) at the dyadic scales 2kr2^kr, and over the direction energy (4), built from the mass directions θr\theta_r, the chosen directions (3) on windows of radius 300300 and the worst vertices on windows of radius 200200, all summed over the vertices of the index box GNG_N.

The library also states that every real function on ℤ2\mathbb{Z}^2 attains a maximum on every window Wd(q)W_d(q).✦