OctoInformBeta

OpenAI preprint, September 2026 · §5. Signed finite-radius envelopes (proof of Proposition 1.2) · §5

§5: the signed assembly (proof of Proposition 1.2): For compactly supported C1C^1 functions: cone majorants averaged over a label law, gradient majorants weighted by switch and derivative costs, and a zero-row-sum identity

Local majorants bound the gradient integral of signed finite-radius envelopes.

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

Take a real C1C^1 function gg with compact support on the plane, identified with ℂ\mathbb C, so that gradients are complex numbers. For a point xx and a radius t>0t>0, the signed disk average of gg is the mean of gg itself, not of |g||g|, over the open disk B(x,t)B(x,t): ut(x)=Atg(x)=1πt2∫B(x,t)g(y)dy.u_t(x)=A_tg(x)=\frac{1}{\pi t^2}\int_{B(x,t)}g(y)\,dy.(1)✦ The radius envelope of a family of functions (ut)(u_t) over radii a≤t≤ba\le t\le b is its pointwise supremum on that closed interval.✦ For the disk averages utu_t of gg, the radius envelope over [a,b][a,b] is the signed finite-radius envelope Sa,bgS_{a,b}g of gg, defined by Sa,bg(x)=supa≤t≤but(x),S_{a,b}g(x)=\sup_{a\le t\le b}u_t(x),(2) a centered maximal function of gg truncated to the radii in [a,b][a,b].

The question is how large ∫ℂ|∇Sa,bg(x)|dx\int_{\mathbb C}|\nabla S_{a,b}g(x)|\,dx can be when [a,b][a,b] is a union of dyadic bands of radii [s,2s][s,2s]. Proposition 1.2 of the OpenAI preprint An Endpoint Gradient Bound for the Centered Disk Maximal Operator is assembled from local data, one band at a time: bounds for the gradients of the disk averages at distinguished radii of each band, and a random direction from a finite palette whose law changes from one dyadic level to the next.

Radii and Majorants

Along a sequence of radii (rk)k≥0(r_k)_{k\ge0}, the index kk of rkr_k is its level, and the run envelope from level ii to level jj is Srj,2rigS_{r_j,2r_i}g.✦ A halving sequence is one with rk=2rk+1r_k=2r_{k+1} for every kk, so that level 00 is the coarsest. For a finest scale r>0r>0, the dyadic scale of level jj is 2jr2^jr✦, and the geometric grid of depth mm is the halving sequence rk=2m−krr_k=2^{m-k}r✦. Its index m−jm-j carries the dyadic scale of level jj.

Fix a point xx and a scale s>0s>0. A physical neighborhood radius of gg at xx and scale ss is a radius t∈[s,2s]t\in[s,2s] such that ut(x)≥ut′(x)u_t(x)\ge u_{t'}(x) for every t′>0t'>0 with |t′−t|≤s/4|t'-t|\le s/4.✦ A physical endpoint radius of gg at xx and scale ss is a radius t∈[s,2s]t\in[s,2s] that equals ss or 2s2s or maximizes t′↦ut′(x)t'\mapsto u_{t'}(x) over [s,2s][s,2s].✦

A finite circle palette is a finite set of points of the unit circle in ℂ\mathbb C such that every point of the circle lies at distance less than 1/20001/2000 from one of them.✦ The fixed palette is one finite circle palette chosen once and for all✦, and its points are the palette labels✦. The quantization Q(z)Q(z) of a complex number zz of modulus one is a palette label at distance less than 1/20001/2000 from zz, likewise chosen once and for all.✦

For n∈ℂn\in\mathbb C and a real number ϵ\epsilon, the cone axis is an(ϵ)=32n+ϵ2ina_n(\epsilon)=\tfrac{\sqrt3}{2}\,n+\tfrac{\epsilon}{2}\,in; for a palette label nn, the axes an(1)a_n(1) and an(−1)a_n(-1) are the unit vectors at angles ±30∘\pm30^\circ from nn.✦ The unfavorable cone part of a vector z∈ℂz\in\mathbb C relative to a palette label nn is Un(z)=(−⟨an(1),z⟩)++(−⟨an(−1),z⟩)+,U_n(z)=\big(-\langle a_n(1),z\rangle\big)_++\big(-\langle a_n(-1),z\rangle\big)_+,(3)✦ where ⟨w,z⟩=Re⁡(wz‾)\langle w,z\rangle=\operatorname{Re}(w\bar z) is the Euclidean inner product and y+=max⁡{y,0}y_+=\max\{y,0\}.

Write ∇ut(x)\nabla u_t(x) for the gradient at xx of y↦ut(y)y\mapsto u_t(y), the radius tt held fixed. Let s>0s>0 be a scale, H=(Hn)H=(H^n) a family of nonnegative functions on ℂ\mathbb C indexed by the palette labels, and JJ a nonnegative function on ℂ\mathbb C; write H(x)H(x) for the function n↦Hn(x)n\mapsto H^n(x) on the palette labels. The pair (H,J)(H,J) is a pair of majorants of gg at scale ss if, for almost every xx,

  1. Un(∇ut(x))≤Hn(x)U_n(\nabla u_t(x))\le H^n(x) for every palette label nn and every physical neighborhood radius tt of gg at xx and scale ss;

  2. |∇ut(x)|≤J(x)|\nabla u_t(x)|\le J(x) for every physical endpoint radius tt of gg at xx and scale ss;

  3. |ut(x)−ut′(x)|≤sJ(x)|u_t(x)-u_{t'}(x)|\le s\,J(x) for all radii tt and t′t' in (0,4s](0,4s].

The Label Law and Its Costs

A law on a finite set ι\iota is a probability distribution on it: a mass μ(i)≥0\mu(i)\ge0 for each i∈ιi\in\iota, with ∑iμ(i)=1\sum_i\mu(i)=1.✦ The point mass δi\delta_i is the law with mass 11 at ii and 00 elsewhere.✦ The expectation of a real function ff on ι\iota under a law μ\mu is 𝔼μf=∑iμ(i)f(i)\mathbb E_\mu f=\sum_i\mu(i)f(i).✦

The tent is the function Λ(y)=max⁡{1−|y|,0}\Lambda(y)=\max\{1-|y|,0\} on ℝ\mathbb R.✦ For a scale r>0r>0 and a vertex p=(p1,p2)p=(p_1,p_2) of ℤ2\mathbb Z^2, the hat of pp at scale rr is the function φr,p(x)=Λ(Re⁡x/r−p1)Λ(Im⁡x/r−p2)\varphi_{r,p}(x)=\Lambda(\operatorname{Re}x/r-p_1)\,\Lambda(\operatorname{Im}x/r-p_2) on ℂ\mathbb C✦, and the tent interpolation 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 ℂ\mathbb C✦.

Fix a complex function hh on ℂ\mathbb C and a scale r>0r>0. The vector mass of hh at a vertex pp is 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 |hr,p||h_{r,p}|✦. The mass direction at pp is the unit complex number θ(p)=hr,p/|hr,p|\theta(p)=h_{r,p}/|h_{r,p}|, or 11 when hr,p=0h_{r,p}=0.✦

The window of radius dd around a vertex qq is the set Wd(q)W_d(q) of vertices zz with |z1−q1|≤d|z_1-q_1|\le d and |z2−q2|≤d|z_2-q_2|\le d✦, and a maximizing vertex of a real function on ℤ2\mathbb Z^2 in Wd(q)W_d(q) is a vertex of the window where the function attains its maximum over it, chosen once and for all✦. The chosen vertex c(z)c(z) of a vertex zz is the maximizing vertex of the scalar mass in W300(z)W_{300}(z).✦ The chosen direction at zz is Θ(z)=θ(c(z))\Theta(z)=\theta(c(z))✦, and the target label of zz is its quantization T(z)=Q(Θ(z))T(z)=Q(\Theta(z))✦.

With the angular tolerance τ=1/100\tau=1/100, the switch chance of a palette label nn at a vertex zz is the thresholded probability of the distance from nn to Θ(z)\Theta(z): pn(z)=min⁡{1,((|n−Θ(z)|−τ)+τ)2}.p_n(z)=\min\Big\{1,\Big(\frac{(|n-\Theta(z)|-\tau)_+}{\tau}\Big)^2\Big\}.(4)✦

Let ι\iota be a finite set and n0∈ιn_0\in\iota an old label, and give each vertex zz a target η(z)∈ι\eta(z)\in\iota and a chance p(z)∈[0,1]p(z)\in[0,1]. The spatial transition row at scale rr and point xx is the law on ι\iota with masses R(j)=δn0(j)+Ir[z↦p(z)(δη(z)(j)−δn0(j))](x).R(j)=\delta_{n_0}(j)+I_r\big[z\mapsto p(z)\big(\delta_{\eta(z)}(j)-\delta_{n_0}(j)\big)\big](x).(5)✦ The scale transition row Rr,x(n)R_{r,x}(n) from a palette label nn is the spatial transition row with old label nn, targets TT and chances pnp_n, all at scale rr.✦ The initial law of hh at scale rr and point xx is the spatial transition row at scale rr and point xx with old label Q(1)Q(1), targets TT and chance 11 at every vertex.✦ The switch rate of a palette label nn at xx is λr,x(n)=Irpn(x)\lambda_{r,x}(n)=I_rp_n(x)✦, and its row derivative is Dr,x(n)=r∑j|∇xRr,x(n)(j)|D_{r,x}(n)=r\sum_j|\nabla_xR_{r,x}(n)(j)|, with the gradient taken in xx and the sum over all palette labels jj✦.

Along a sequence of radii (rk)k≥0(r_k)_{k\ge0}, the label marginal of hh is the sequence of laws μk(x)\mu_k(x) on the palette labels in which μ0(x)\mu_0(x) is the initial law at scale r0r_0 and μk+1(x)(j)=∑nμk(x)(n)Rrk+1,x(n)(j).\mu_{k+1}(x)(j)=\sum_n\mu_k(x)(n)\,R_{r_{k+1},x}(n)(j).(6)✦ The switch costs σk\sigma_k and the derivative costs κk\kappa_k of hh along (rk)(r_k) are the functions on ℂ\mathbb C given by σ0(x)=1,σk+1(x)=𝔼μk(x)λrk+1,x,κ0(x)=r0∑j|∇xμ0(x)(j)|,κk+1(x)=𝔼μk(x)Drk+1,x.\begin{aligned} \sigma_0(x)&=1, &\sigma_{k+1}(x)&=\mathbb E_{\mu_k(x)}\lambda_{r_{k+1},x},\\ \kappa_0(x)&=r_0\sum_j\big|\nabla_x\mu_0(x)(j)\big|, &\kappa_{k+1}(x)&=\mathbb E_{\mu_k(x)}D_{r_{k+1},x}. \end{aligned}(7) For an integer m≥0m\ge0, called the endpoint level, the boundary cost of hh along (rk)(r_k) is βkm=σk+σk+1\beta^m_k=\sigma_k+\sigma_{k+1} for k≠mk\ne m and βmm=σm+1\beta^m_m=\sigma_m+1. The causal dyadic filter of a real sequence (ci)(c_i) is the sequence (Fc)k=∑i=0k2i−kci(Fc)_k=\sum_{i=0}^k2^{i-k}c_i.✦ The point cost PkmP^m_k of hh at level kk for the endpoint level mm is Pkm(x)=βkm(x)+(Fκ(x))k.P^m_k(x)=\beta^m_k(x)+\big(F\kappa(x)\big)_k.(8)✦

A row jet on a finite set ι\iota with values in a real inner product space EE is a law π\pi on ι\iota together with vectors ∇πi∈E\nabla\pi_i\in E, one for each i∈ιi\in\iota, whose sum is zero.✦

The Claim

Theorem 1. Let gg be a real C1C^1 function with compact support on ℂ\mathbb C, hh a complex function on ℂ\mathbb C, (rk)k≥0(r_k)_{k\ge0} a halving sequence of positive radii, and m≥0m\ge0 an integer. For every level k≥0k\ge0, let (Hk,Jk)(H_k,J_k) be a pair of majorants of gg at scale rkr_k, with every HknH^n_k and JkJ_k integrable when k≤mk\le m. Then, for the label marginal μk\mu_k and the point costs PkmP^m_k of hh along (rk)(r_k), the run envelope Srm,2r0gS_{r_m,2r_0}g from level 00 to level mm satisfies ∫ℂ|∇Srm,2r0g(x)|dx≤12(∑k=0m∫ℂ𝔼μk(x)Hk(x)dx+4∑k=0m∫ℂPkm(x)Jk(x)dx).\begin{aligned} \int_{\mathbb C}|\nabla S_{r_m,2r_0}g(x)|\,dx\le12\Big(&\sum_{k=0}^m\int_{\mathbb C} \mathbb E_{\mu_k(x)}H_k(x)\,dx\\ &+4\sum_{k=0}^m\int_{\mathbb C}P^m_k(x)\,J_k(x)\,dx\Big). \end{aligned}(9)✦

Theorem 2. Let gg be a real C1C^1 function with compact support on ℂ\mathbb C, r>0r>0 a finest scale, m≥0m\ge0 an integer, and (rk)(r_k) the geometric grid of depth mm. For every level j≥0j\ge0, let (Hj,Jj)(H_j,J_j) be a pair of majorants of gg at the dyadic scale 2jr2^jr, with every HjnH^n_j and JjJ_j integrable. Then, for the label marginal μk\mu_k and the point costs PkmP^m_k of h=∇gh=\nabla g along the grid, the run envelope Sr,2m+1rgS_{r,2^{m+1}r}g from level 00 to level mm of the grid satisfies ∫ℂ|∇Sr,2m+1rg(x)|dx≤12(∑j=0m∫ℂ𝔼μm−j(x)Hj(x)dx+4∑j=0m∫ℂPm−jm(x)Jj(x)dx).\begin{aligned} \int_{\mathbb C}|\nabla S_{r,2^{m+1}r}g(x)|\,dx\le12\Big(&\sum_{j=0}^m\int_{\mathbb C} \mathbb E_{\mu_{m-j}(x)}H_j(x)\,dx\\ &+4\sum_{j=0}^m\int_{\mathbb C}P^m_{m-j}(x)\,J_j(x)\,dx\Big). \end{aligned}(10)✦

Theorem 3. Let ι\iota be a finite set, EE a real inner product space, and π\pi with the vectors (∇πi)i∈ι(\nabla\pi_i)_{i\in\iota} a row jet on ι\iota with values in EE. For every family (vi)i∈ι(v_i)_{i\in\iota} of vectors of EE and every vector c∈Ec\in E, ∑i∈ι⟨∇πi,vi⟩=∑i∈ι⟨∇πi,vi−c⟩.\sum_{i\in\iota}\langle\nabla\pi_i,v_i\rangle=\sum_{i\in\iota}\langle\nabla\pi_i,v_i-c\rangle.(11)✦

What Is Checked

The claims (9), (10) and (11) are stated over the signed disk averages (1) of a real C1C^1 function with compact support and their radius and run envelopes, the dyadic scales and the geometric grid, the physical neighborhood and endpoint radii, the unfavorable cone parts (3) along the axes of the labels of the fixed palette, the label marginal (6) built from the hats, masses, directions, targets, switch chances and transition rows of hh, the costs (7) and point costs (8) with the causal dyadic filter, and row jets on a finite set. Of these constructions, the library proves that a finite circle palette exists✦ and that a real function on a window of ℤ2\mathbb Z^2 attains its maximum there✦. It also proves that every switch chance lies between 00 and 11.