OctoInformBeta

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

Lemma 2.2 [Balanced kernels]: At r∈[1,2]r\in[1,2] with ur(x)≥ur±δ(x)u_r(x)\ge u_{r\pm\delta}(x), a probability density on [r−δ,r+δ][r-\delta,r+\delta] bounded by 216/δ216/\delta mixes the indicators of B(x,t)B(x,t) into one

Averaged disk indicators that cancel the gradient’s complex moment.

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

For a real function gg on ℂ\mathbb{C}, a point x∈ℂx\in\mathbb{C} and a radius t>0t>0, the disk average ut(x)=Atg(x)u_t(x)=A_tg(x) of gg at xx is the integral of gg over the open disk B(x,t)={y∈ℂ:|y−x|<t}B(x,t)=\{y\in\mathbb{C}:|y-x|<t\}, divided by the area of that disk: ut(x)=1πt2∫B(x,t)g(y)dy.u_t(x)=\frac{1}{\pi t^2}\int_{B(x,t)} g(y)\,dy.(1)✦

Take gg real and C1C^1, a center xx, a number 0<δ≤1/40<\delta\le 1/4, and a radius r∈[1,2]r\in[1,2] at which ur(x)u_r(x) is at least ur−δ(x)u_{r-\delta}(x) and ur+δ(x)u_{r+\delta}(x). The question is whether the indicator of B(x,r)B(x,r) can then be replaced by a weighted average of the indicators of the disks B(x,t)B(x,t), r−δ≤t≤r+δr-\delta\le t\le r+\delta, with a weight bounded by a constant over δ\delta, so that the complex moment vanishes. The complex moment of ∇g\nabla g about xx is the integral over ℂ\mathbb{C} of the gradient of gg paired with the conjugate of y−xy-x, taken against that weight.

Radius Densities and Disk Kernels

Let a,ba,b be real numbers and M≥0M\ge 0. A radius density on [a,b][a,b] bounded by MM is a probability density kk on ℝ\mathbb{R} that vanishes outside [a,b][a,b] and is at most MM everywhere: k∈L1(ℝ),0≤k≤M,∫ℝk(t)dt=1,k(t)=0for t∉[a,b].k\in L^1(\mathbb{R}),\qquad 0\le k\le M,\qquad \int_{\mathbb{R}} k(t)\,dt=1,\qquad k(t)=0 \ \text{for } t\notin[a,b].(2)✦

The tail Tk(u)T_k(u) of kk at u∈ℝu\in\mathbb{R} is the mass that kk puts on radii larger than uu✦, and the disk kernel of kk about a point x∈ℂx\in\mathbb{C} is the function KxK_x on ℂ\mathbb{C} whose value at yy is the tail at the distance |y−x||y-x|: Tk(u)=∫u∞k(t)dt,Kx(y)=Tk(|y−x|).T_k(u)=\int_u^{\infty} k(t)\,dt,\qquad K_x(y)=T_k(|y-x|).(3)✦

Neighborhood and Endpoint Radii

A radius rr is a neighborhood radius of gg at xx if r∈[1,2]r\in[1,2] and ur(x)u_r(x) is the largest of the averages at positive radii within 1/41/4 of rr: ut(x)≤ur(x)for every t>0 with |t−r|≤1/4.u_t(x)\le u_r(x)\quad\text{for every } t>0 \text{ with } |t-r|\le 1/4.(4)✦

A radius rr is an endpoint radius of gg at xx if r∈[1,2]r\in[1,2] and either rr is one of the endpoints 11 and 22 or rr maximizes the average over that interval: ut(x)≤ur(x)for every t∈[1,2].u_t(x)\le u_r(x)\quad\text{for every } t\in[1,2].(5)✦

The Claim

Theorem 1 (Balanced kernels). Let gg be a real C1C^1 function on ℂ\mathbb{C}, let x∈ℂx\in\mathbb{C}, and let 0<δ≤1/40<\delta\le 1/4.

(i) Let r∈[1,2]r\in[1,2] be a radius with ur−δ(x)≤ur(x)u_{r-\delta}(x)\le u_r(x) and ur+δ(x)≤ur(x)u_{r+\delta}(x)\le u_r(x). Then there is a radius density kk on [r−δ,r+δ][r-\delta,r+\delta] bounded by 216/δ216/\delta whose disk kernel KxK_x about xx satisfies ∫ℂKx(y)(y−x)¯∇g(y)dy=0.\int_{\mathbb{C}} K_x(y)\,\overline{(y-x)}\,\nabla g(y)\,dy=0.(6)✦

(ii) Let a,ba,b be real, M≥0M\ge 0, and kk a radius density on [a,b][a,b] bounded by MM. At all points y,y′∈ℂy,y'\in\mathbb{C}, the disk kernel KxK_x of kk about xx is an average of indicators of disks centered at xx, Kx(y)=∫ℝk(t)𝟏B(x,t)(y)dt,K_x(y)=\int_{\mathbb{R}} k(t)\,\mathbf{1}_{B(x,t)}(y)\,dt,(7)✦ it takes values between 00 and 11, 0≤Kx(y),0\le K_x(y),(8)✦ Kx(y)≤1,K_x(y)\le 1,(9)✦ it is Lipschitz with constant MM, |Kx(y)−Kx(y′)|≤M|y−y′|,|K_x(y)-K_x(y')|\le M\,|y-y'|,(10)✦ and it equals 11 on the disk B(x,a)B(x,a) and vanishes outside the disk B(x,b)B(x,b): Kx(y)=1if |y−x|<a,K_x(y)=1\quad\text{if } |y-x|<a,(11)✦ Kx(y)=0if |y−x|≥b.K_x(y)=0\quad\text{if } |y-x|\ge b.(12)✦

(iii) If rr is a neighborhood radius of gg at xx, there is a radius density kk on [r−δ,r+δ][r-\delta,r+\delta] bounded by 216/δ216/\delta whose disk kernel KxK_x about xx satisfies ∫ℂKx(y)(y−x)¯∇g(y)dy=0.\int_{\mathbb{C}} K_x(y)\,\overline{(y-x)}\,\nabla g(y)\,dy=0.(13)✦

(iv) If rr is an endpoint radius of gg at xx and [r−δ,r+δ][r-\delta,r+\delta] lies in [1,2][1,2], there is a radius density kk on [r−δ,r+δ][r-\delta,r+\delta] bounded by 216/δ216/\delta whose disk kernel KxK_x about xx satisfies ∫ℂKx(y)(y−x)¯∇g(y)dy=0.\int_{\mathbb{C}} K_x(y)\,\overline{(y-x)}\,\nabla g(y)\,dy=0.(14)✦

In the paper, the result reads: if r∈[1,2]r\in[1,2], 0<δ≤1/40<\delta\le 1/4, ur(x)≥ur−δ(x)u_r(x)\ge u_{r-\delta}(x) and ur(x)≥ur+δ(x)u_r(x)\ge u_{r+\delta}(x), there is a probability density kk on [r−δ,r+δ][r-\delta,r+\delta], bounded by C/δC/\delta, such that K(y)=∫k(t)𝟏B(x,t)(y)dtK(y)=\int k(t)\,\mathbf{1}_{B(x,t)}(y)\,dt satisfies ∫K(y)(y−x)¯dh(y)=0\int K(y)\,\overline{(y-x)}\,dh(y)=0, and moreover 0≤K≤10\le K\le 1, Lip⁡(K)≤C/δ\operatorname{Lip}(K)\le C/\delta, K=1K=1 on B(x,r−δ)B(x,r-\delta) and K=0K=0 outside B(x,r+δ)B(x,r+\delta), with the density chosen measurably in (x,r)(x,r); there CC is a constant, KK is the page’s KxK_x and dh(y)dh(y) stands for ∇g(y)dy\nabla g(y)\,dy.

What Is Checked

The claim is stated over real C1C^1 functions gg on ℂ\mathbb{C} with gradients read as complex numbers, the disk averages (1), the radius densities (2) and their disk kernels (3), built from the tail of the density, and the neighborhood and endpoint radii (4) and (5).

Of the disk kernel alone, with no function gg involved, the library proves the properties (7) to (12), for every radius density and every center.