OctoInformBeta

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

Proposition 4.3 [Packing of switch and derivative costs]: Over all dyadic levels and box vertices, mass-weighted switching and derivative costs sum to at most a constant times ∫|h|\int|h|

Mass-weighted label costs are 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, and let V=∫ℂ|h|V=\int_{\mathbb C}|h| be its total mass; the case of interest is h=∇gh=\nabla g for a real C1C^1 function gg with compact support. Fix a finest scale r>0r>0 and an integer m≥0m\ge0; the dyadic scale of level jj is 2jr2^jr, so that level 00 is the finest and each level doubles the one below it.✦ On each grid, whose vertices are the indices p∈ℤ2p\in\mathbb Z^2 of its grid points, hh has a mass and a direction at every vertex. A random label, with values in a finite set of unit directions fixed below as the palette, follows these directions from the coarsest grid down to the finest and switches when it strays from them.

The question is whether the label’s switches and its variation in space, charged at each vertex with the mass of hh there, cost in total at most a constant times VV, with a constant that does not depend on the number of levels.

Masses and Directions on the Grid

For a scale s>0s>0 and an index p=(p1,p2)∈ℤ2p=(p_1,p_2)\in\mathbb Z^2, the grid point of index pp is sp=sp1+isp2∈ℂsp=sp_1+i\,sp_2\in\mathbb C.✦ 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 ss is the function on ℂ\mathbb C given by φs,p(x)=Λ(Re⁡x/s−p1)Λ(Im⁡x/s−p2)\varphi_{s,p}(x)=\Lambda(\operatorname{Re}x/s-p_1)\,\Lambda(\operatorname{Im}x/s-p_2).✦ The vector mass hs,ph_{s,p} and the scalar mass as,pa_{s,p} of hh at pp are the complex number and the nonnegative number hs,p=∫ℂφs,p(y)h(y)dy,as,p=|hs,p|,h_{s,p}=\int_{\mathbb C}\varphi_{s,p}(y)\,h(y)\,dy,\qquad a_{s,p}=|h_{s,p}|,(1) and the direction of hh at pp is the unit complex number θs(p)=hs,p/|hs,p|\theta_s(p)=h_{s,p}/|h_{s,p}|, set to 11 when hs,p=0h_{s,p}=0.✦ The tent interpolation at scale ss of a real function ff on ℤ2\mathbb Z^2 is the function Isf(x)=∑p∈ℤ2φs,p(x)f(p)I_sf(x)=\sum_{p\in\mathbb Z^2}\varphi_{s,p}(x)\,f(p) on ℂ\mathbb C.✦

The box norm of a point x∈ℂx\in\mathbb C is ∥x∥∞=max⁡{|Re⁡x|,|Im⁡x|}\|x\|_\infty=\max\{|\operatorname{Re}x|,|\operatorname{Im}x|\}.✦ For N∈ℕN\in\mathbb N, the index box GNG_N is the set of q∈ℤ2q\in\mathbb Z^2 with |q1|≤N|q_1|\le N and |q2|≤N|q_2|\le N✦; for d∈ℕd\in\mathbb N, the window Wd(q)W_d(q) of radius dd around a vertex qq is the set of 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.✦

The Label Law

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

Fix a scale ss. The chosen vertex of a vertex z∈ℤ2z\in\mathbb Z^2 is a vertex z*z^* of the window W300(z)W_{300}(z), fixed by a choice, at which the scalar mass as,⋅a_{s,\cdot} attains its maximum over that window✦, and the chosen direction at zz is Θs(z)=θs(z*)\Theta_s(z)=\theta_s(z^*).✦ The target label of zz is ns(z)=Q(Θs(z))n_s(z)=Q(\Theta_s(z))✦, and the switch chance of a label nn at zz is the following, where x+=max⁡{x,0}x_+=\max\{x,0\} denotes the positive part of a real number xx: ps(n,z)=πτ(|n−Θs(z)|),πτ(d)=min⁡{1,((d−τ)+τ)2},τ=1100,p_s(n,z)=\pi_\tau\bigl(|n-\Theta_s(z)|\bigr),\qquad \pi_\tau(d)=\min\Bigl\{1,\Bigl(\frac{(d-\tau)_+}{\tau}\Bigr)^2\Bigr\},\qquad \tau=\frac1{100},(2) where πτ\pi_\tau is the thresholded probability at tolerance τ\tau. The switching rate of nn at a point x∈ℂx\in\mathbb C is the interpolated chance p‾s(x,n)=Is[ps(n,⋅)](x)\bar p_s(x,n)=I_s[p_s(n,\cdot)](x).✦

A law on a finite set is a probability distribution on it: a nonnegative mass at each element, the masses summing to 11.✦ Here δi\delta_i is the point mass at ii✦, and 𝔼μf=∑iμ(i)f(i)\mathbb E_\mu f=\sum_i\mu(i)f(i) is the expectation of a real function ff under a law μ\mu.✦ Given a finite set of labels, an old label n0n_0, a target label t(z)t(z) and a chance p(z)∈[0,1]p(z)\in[0,1] at every vertex zz, the spatial transition row at scale ss and point xx is the law on the labels with masses δn0(j)+Is[z↦p(z)(δt(z)(j)−δn0(j))](x).\delta_{n_0}(j)+I_s\bigl[z\mapsto p(z)\bigl(\delta_{t(z)}(j)-\delta_{n_0}(j)\bigr)\bigr](x).(3)✦ The scale transition row Ps(x;n,⋅)P_s(x;n,\cdot) from a palette label nn is the spatial transition row with old label nn, targets ns(z)n_s(z) and chances ps(n,z)p_s(n,z).✦ The initial law μsin(x)\mu^{\mathrm{in}}_s(x) is the spatial transition row with old label Q(1)Q(1), targets ns(z)n_s(z) and chance 11 at every vertex.✦

Along a sequence of radii r0,r1,r2,…r_0,r_1,r_2,\dots, the label marginal at a point xx is the sequence of laws on ℒ\mathcal L given by μ0(x)=μr0in(x)\mu_0(x)=\mu^{\mathrm{in}}_{r_0}(x) and μk+1(x)(j)=∑n∈ℒμk(x)(n)Prk+1(x;n,j).\mu_{k+1}(x)(j)=\sum_{n\in\mathcal L}\mu_k(x)(n)\,P_{r_{k+1}}(x;n,j).(4)✦

Switching and Derivative Costs

The row derivative cost of a label nn at scale ss and point xx is Γs(x,n)=s∑j∈ℒ|∇xPs(x;n,j)|\Gamma_s(x,n)=s\sum_{j\in\mathcal L}|\nabla_xP_s(x;n,j)|.✦ Along the radii (rk)(r_k), the switching cost RkR_k and the derivative cost TkT_k are the real functions on ℂ\mathbb C given by R0=1R_0=1, T0(x)=r0∑j∈ℒ|∇xμ0(x)(j)|T_0(x)=r_0\sum_{j\in\mathcal L}|\nabla_x\mu_0(x)(j)| and Rk+1(x)=𝔼μk(x)p‾rk+1(x,⋅),Tk+1(x)=𝔼μk(x)Γrk+1(x,⋅).R_{k+1}(x)=\mathbb E_{\mu_k(x)}\,\bar p_{r_{k+1}}(x,\cdot),\qquad T_{k+1}(x)=\mathbb E_{\mu_k(x)}\,\Gamma_{r_{k+1}}(x,\cdot).(5)

Local Suprema and the Cell Cost

The square neighborhood of a vertex q∈ℤ2q\in\mathbb Z^2 at scale ss is the closed square Ds(q)={x∈ℂ:∥x−sq∥∞≤25s}D_s(q)=\{x\in\mathbb C:\|x-sq\|_\infty\le25s\}(6)✦ about the grid point sqsq, and the local supremum of a real function ff on ℂ\mathbb C at qq is supDs(q)f=sup⁡{f(x):x∈Ds(q)}.\sup_{D_s(q)}f=\sup\{f(x):x\in D_s(q)\}.(7)✦ The causal dyadic average of a real sequence c=(ck)k≥0c=(c_k)_{k\ge0} is the sequence Flc=∑k=0l2k−lck,l≥0.F_lc=\sum_{k=0}^{l}2^{k-l}c_k,\qquad l\ge0.(8)✦

For an integer m≥0m\ge0 and radii (rk)(r_k), the cell cost of a vertex qq at an index 0≤l≤m0\le l\le m is the number Kl(q)=supDrl(q)Rl+{supDrl(q)Rl+1,l<m,1,l=m,+Fl((supDrl(q)Tk)k≥0).K_l(q)=\sup_{D_{r_l}(q)}R_l +\begin{cases}\sup_{D_{r_l}(q)}R_{l+1},&l<m,\\[2pt] 1,&l=m,\end{cases} +\;F_l\Bigl(\bigl(\sup_{D_{r_l}(q)}T_k\bigr)_{k\ge0}\Bigr).(9)✦

With LL the number of palette labels, put Ccmp=1280000(L+1)C_{\mathrm{cmp}}=1\,280\,000\,(L+1)✦ and Csp=208Ccmp2C_{\mathrm{sp}}=208\,C_{\mathrm{cmp}}^2✦. The packing constant is the number Cpack=8Csp(2560004+2000000⋅576⋅40972)+8Csp⋅576⋅40972+130Csp+1,C_{\mathrm{pack}}=8C_{\mathrm{sp}}\bigl(2\,560\,004+2\,000\,000\cdot576\cdot4097^2\bigr) +8C_{\mathrm{sp}}\cdot576\cdot4097^2+130\,C_{\mathrm{sp}}+1,(10)✦

The Claim

Theorem 1. Let r>0r>0 and 𝜚≥0\varrho\ge0 be real numbers, let NN be a natural number with 𝜚/r+1≤N\varrho/r+1\le N, and let m≥0m\ge0 be an integer. Let h:ℂ→ℂh:\mathbb C\to\mathbb C be an integrable function that vanishes outside the closed disk {|y|≤𝜚}\{|y|\le\varrho\}, and let KlK_l be the cell costs (9) of hh along the radii rl=2m−lrr_l=2^{m-l}r. Then ∑j=0m∑q∈GNa2jr,qKm−j(q)≤Cpack∫ℂ|h(y)|dy.\sum_{j=0}^{m}\sum_{q\in G_N}a_{2^jr,q}\,K_{m-j}(q)\le C_{\mathrm{pack}}\int_{\mathbb C}|h(y)|\,dy.(11)✦

In the paper, the result reads: for every real g∈Cc∞(ℝ2)g\in C_c^\infty(\mathbb R^2) and every m≥1m\ge1, ∑l=1m∑q∈Vlaq[sup⁡x∈D(q)(Rl(x)+Rl+1(x))+∑k=1l2k−less sup⁡x∈D(q)Tk(x)]≤CV\sum_{l=1}^{m}\sum_{q\in V_l}a_q\bigl[\sup_{x\in D(q)}(R_l(x)+R_{l+1}(x)) +\sum_{k=1}^{l}2^{k-l}\operatorname{ess\,sup}_{x\in D(q)}T_k(x)\bigr]\le CV, where aqa_q is the scalar mass (1) of ∇g\nabla g at qq at the scale rlr_l, VlV_l is the paper’s set of grid vertices at level ll, of scale rl=2−lr_l=2^{-l} in its indexing, D(q)={x:∥x−q∥∞≤25rl}D(q)=\{x:\|x-q\|_\infty\le25r_l\}, V=∫|∇g|V=\int|\nabla g|, and CC is an absolute constant independent of mm.

What Is Checked

The claim (11) is stated over an integrable vector density hh vanishing outside a disk of radius 𝜚\varrho and the index box GNG_N with N≥𝜚/r+1N\ge\varrho/r+1, on which it weighs the scalar masses (1) against the cell costs (9) of the label law along the radii 2m−lr2^{m-l}r; these combine the switching and derivative costs (5) through the local suprema (7) over the squares (6) and the causal dyadic average (8), and the constant is the explicit number (10). About these constructions the library proves that every switch chance ps(n,z)p_s(n,z) lies between 00 and 11 and that a real function on ℤ2\mathbb Z^2 attains its maximum on every window Wd(q)W_d(q).✦ It also proves that a finite circle palette exists.✦