OctoInformBeta

OpenAI preprint, September 2026 · §2. Cancellation at disks and three local bounds · §2

Lemma 2.1 [Local estimates]: When the gradient splits locally into error pieces of price AA, majorants of integral at most CACA on sets of measure at most one

Three local bounds on disk averages and their gradients.

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

The plane is the complex plane ℂ\mathbb{C}, identified with ℝ2\mathbb{R}^2 by a fixed isometry, so that gradients are complex numbers and, for e,w∈ℂe,w\in\mathbb{C}, e⋅w=Re⁡(e‾w)e\cdot w=\operatorname{Re}(\bar e\,w) is the Euclidean inner product. The signed disk average of a real function gg on ℂ\mathbb{C} at a point xx and radius t>0t>0 is 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)✦ where B(x,t)B(x,t) is the open disk of center xx and radius tt.

The estimates concern a real C1C^1 function gg on ℂ\mathbb{C} whose gradient, locally, is a sum of an integrable part with a one-sided direction constraint and of finitely many small-scale integrable parts; the two kinds of parts are made precise below as cone densities and error pieces.

Neighborhood and Endpoint Radii

A neighborhood radius of a real function gg at a point xx is a radius r∈[1,2]r\in[1,2] that maximizes t↦ut(x)t\mapsto u_t(x) within distance 14\tfrac14 of itself: ut(x)≤ur(x)for every t>0 with |t−r|≤14.u_t(x)\le u_r(x)\qquad\text{for every } t>0 \text{ with } |t-r|\le\tfrac14.(2)✦ An endpoint radius of gg at xx is a radius r∈[1,2]r\in[1,2] that is 11 or 22 or maximizes t↦ut(x)t\mapsto u_t(x) over [1,2][1,2]: r∈{1,2}orut(x)≤ur(x) for every t∈[1,2].r\in\{1,2\}\qquad\text{or}\qquad u_t(x)\le u_r(x)\ \text{ for every } t\in[1,2].(3)✦ Write I(x)I(x) and J(x)J(x) for the sets of neighborhood and endpoint radii of gg at xx.

Cone Densities and Error Pieces

A cone density for a vector e∈ℂe\in\mathbb{C} and a constant κ∈ℝ\kappa\in\mathbb{R} is an integrable function a:ℂ→ℂa:\mathbb{C}\to\mathbb{C} with κ|a(y)|≤e⋅a(y)for almost every y∈ℂ.\kappa\,|a(y)|\le e\cdot a(y)\qquad\text{for almost every } y\in\mathbb{C}.(4)✦

An error piece with constant Cs∈ℝC_s\in\mathbb{R} is an integrable function β:ℂ→ℂ\beta:\mathbb{C}\to\mathbb{C} together with a size s∈(0,1]s\in(0,1] and a center c∈ℂc\in\mathbb{C} such that, when s<1s<1, β\beta has integral zero and vanishes almost everywhere outside the closed disk of center cc and radius CssC_ss: s<1⇒∫ℂβ(y)dy=0andβ(y)=0 for almost every y with |y−c|>Css.s<1\ \Longrightarrow\ \int_{\mathbb{C}}\beta(y)\,dy=0\quad\text{and}\quad \beta(y)=0\ \text{ for almost every } y \text{ with } |y-c|>C_ss.(5)✦

The mass of an error piece (β,s,c)(\beta,s,c) is m=∫ℂ|β(y)|dym=\int_{\mathbb{C}}|\beta(y)|\,dy.✦ The price at exponent α\alpha of a finite family (βd,sd,cd)d(\beta_d,s_d,c_d)_d of error pieces with masses mdm_d is A=∑dsdαmd,A=\sum_d s_d^{\alpha}\,m_d,(6)✦ and the theorem takes α=1100\alpha=\frac1{100}.

The Claim

Theorem 1 (Local estimates). Let gg be a real C1C^1 function on ℂ\mathbb{C}, let Q⊂ℂQ\subset\mathbb{C} be a set of measure at most one, and let (βd,sd,cd)d(\beta_d,s_d,c_d)_d be a finite family of error pieces with constant Cs≥0C_s\ge0, of price AA at α=1100\alpha=\frac1{100}.

  1. Let κ>0\kappa>0, and let aa be a cone density for a unit vector ee and κ\kappa. There are a constant C>0C>0, depending only on CsC_s and κ\kappa, and a measurable F:ℂ→[0,∞]F:\mathbb{C}\to[0,\infty] with ∫QF(x)dx≤CA\int_QF(x)\,dx\le CA, such that at every point xx where ∇g=a+∑dβd\nabla g=a+\sum_d\beta_d almost everywhere on B(x,3)B(x,3), (−e⋅∫B(x,r)∇g(y)dy)+≤F(x)for every r∈I(x).\Bigl(-e\cdot\int_{B(x,r)}\nabla g(y)\,dy\Bigr)_+\le F(x)\qquad\text{for every } r\in I(x).(7)✦

  2. There are a constant C>0C>0, depending only on CsC_s, and a measurable F:ℂ→[0,∞]F:\mathbb{C}\to[0,\infty] with ∫QF(x)dx≤CA\int_QF(x)\,dx\le CA, such that at every point xx where ∇g=∑dβd\nabla g=\sum_d\beta_d almost everywhere on B(x,3)B(x,3), |∫B(x,r)∇g(y)dy|≤F(x)for every r∈J(x).\Bigl|\int_{B(x,r)}\nabla g(y)\,dy\Bigr|\le F(x)\qquad\text{for every } r\in J(x).(8)✦

  3. Suppose Cs=2C_s=2, ∇g\nabla g is bounded on ℂ\mathbb{C}, and ∇g=∑dβd\nabla g=\sum_d\beta_d almost everywhere on the closed disk of radius 88 about 00. There are an absolute constant C>0C>0 and a measurable F:ℂ→[0,∞]F:\mathbb{C}\to[0,\infty] with ∫QF(x)dx≤CA\int_QF(x)\,dx\le CA such that |ut(x)−ut′(x)|≤F(x)for |x|≤2 and 0<t,t′≤4.|u_t(x)-u_{t'}(x)|\le F(x)\qquad\text{for } |x|\le2 \text{ and } 0<t,t'\le4.(9)✦

In the paper, the result reads: if QQ is a unit square and on Ω={y:dist⁡(y,Q)<10}\Omega=\{y:\operatorname{dist}(y,Q)<10\} the measure ∇gdx\nabla g\,dx is a finite sum ∑iviσi+∑dβd\sum_iv_i\sigma_i+\sum_d\beta_d of finite absolutely continuous measures on ℝ2\mathbb{R}^2, with each σi≥0\sigma_i\ge0, each viv_i a constant unit vector, and each βd\beta_d of mass md=|βd|(ℝ2)m_d=|\beta_d|(\mathbb{R}^2) given a number sd∈(0,1]s_d\in(0,1] such that βd(ℝ2)=0\beta_d(\mathbb{R}^2)=0 and βd\beta_d is supported in a ball of radius at most CssdC_ss_d whenever sd<1s_d<1, then, with S=∑iσiS=\sum_i\sigma_i, A=∑dsdαmdA=\sum_ds_d^{\alpha}m_d and I(x)I(x), J(x)J(x) as above, (i) ∫Qsup⁡r∈I(x)(−e⋅∇ur(x))+dx≤CA\int_Q\sup_{r\in I(x)}(-e\cdot\nabla u_r(x))_+\,dx\le CA for every unit vector ee with e⋅vi≥κ>0e\cdot v_i\ge\kappa>0 for all ii, (ii) ∫Qsup⁡r∈J(x)|∇ur(x)|dx≤C(A+S(ℝ2))\int_Q\sup_{r\in J(x)}|\nabla u_r(x)|\,dx\le C(A+S(\mathbb{R}^2)), and (iii) ∫Qsup⁡0<t,t′≤4|ut(x)−ut′(x)|dx≤C(A+S(ℝ2))\int_Q\sup_{0<t,t'\le4}|u_t(x)-u_{t'}(x)|\,dx\le C(A+S(\mathbb{R}^2)), where CC depends only on CsC_s, and in (i) also on κ\kappa.

The theorem takes the measure ∑iviσi\sum_iv_i\sigma_i as a cone density in (i) and as error pieces of size one in (ii) and (iii). The integrals ∫B(x,r)∇g(y)dy\int_{B(x,r)}\nabla g(y)\,dy in the box are πr2∇ur(x)\pi r^2\,\nabla u_r(x).

What Is Checked

The three bounds are stated over the signed disk averages (1), the neighborhood and endpoint radii (2) and (3), cone densities (4), and finite families of error pieces (5) with their masses and prices (6), for real C1C^1 functions on ℂ\mathbb{C} and sets of measure at most one.

Every integrable function h:ℂ→ℂh:\mathbb{C}\to\mathbb{C} is an error piece of size one with every constant CsC_s: its unit piece has density, size and center β=h,s=1,c=0.\beta=h,\qquad s=1,\qquad c=0.(10)✦