OctoInformBeta

OpenAI preprint, September 2026 · §6. From masked energies to the maximal transform

§6 (setting): odd Gaussian windows and the smooth kernel: Window ratios and the triple-convolution signed kernel at positive scale, with the profile identity for ordered positive scales away from zero

Defining the smooth kernel via odd Gaussian windows.

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 scale s>0s > 0 and a point t∈ℝt \in \mathbb{R}. Weighting a variance-ss Gaussian density by t/st/\sqrt{s} produces an odd window at scale ss along with its ratio to a rescaled Gaussian, while integrating a signed kernel over squared scales a,b∈ℝa, b \in \mathbb{R} against ds/sds/s produces a smooth kernel. The problem is to identify the triple convolution of the odd window with the signed kernel and to evaluate the smooth kernel away from t=0t = 0 via a smooth kernel profile.

Gaussian Density

For real variance σ∈ℝ\sigma \in \mathbb{R}, the variance-σ\sigma Gaussian density gσ:ℝ→ℝg_\sigma \colon \mathbb{R} \to \mathbb{R} is the real-valued function given for t∈ℝt \in \mathbb{R} by (1): gσ(t)=(2πσ)−1/2exp⁡(−t2/(2σ)).g_\sigma(t) = (2\pi\sigma)^{-1/2} \exp\bigl(-t^2/(2\sigma)\bigr).(1)✦

The Claim

Definition 1. For a real scale s∈ℝs \in \mathbb{R}, the odd window ks:ℝ→ℝk_s \colon \mathbb{R} \to \mathbb{R} is the function defined for t∈ℝt \in \mathbb{R} by (2): ks(t)=tsgs(t).k_s(t) = \frac{t}{\sqrt{s}}\, g_s(t).(2)✦ At scale s∈ℝs \in \mathbb{R}, the odd window ratio ρs:ℝ→ℝ\rho_s \colon \mathbb{R} \to \mathbb{R} is the real-valued function specified at t∈ℝt \in \mathbb{R} by (3): ρs(t)=ks(t)g(11/10)s(t).\rho_s(t) = \frac{k_s(t)}{g_{(11/10)s}(t)}.(3)✦ Given s∈ℝs \in \mathbb{R}, the signed kernel ψs:ℝ→ℝ\psi_s \colon \mathbb{R} \to \mathbb{R} is the function on ℝ\mathbb{R} determined by (4) for t∈ℝt \in \mathbb{R}: ψs(t)=t3/27−ts/3ssg3s(t).\psi_s(t) = \frac{t^3/27 - ts/3}{s\sqrt{s}}\, g_{3s}(t).(4)✦ If s>0s > 0 and a,b,c∈ℝa, b, c \in \mathbb{R}, then the odd window ksk_s and the signed kernel ψs\psi_s satisfy the triple-convolution identity (5): ∫ℝ2ks(a−z1)ks(b−z2)ks(c+z1+z2)dz1dz2=ψs(a+b+c).\int_{\mathbb{R}^2} k_s(a - z_1)\, k_s(b - z_2)\, k_s(c + z_1 + z_2)\, dz_1\, dz_2 = \psi_s(a + b + c).(5)✦ Over arbitrary squared scale endpoints a,b∈ℝa, b \in \mathbb{R}, the smooth kernel I(a,b;⋅):ℝ→ℝI(a, b; \cdot) \colon \mathbb{R} \to \mathbb{R} is the function given for t∈ℝt \in \mathbb{R} by (6): I(a,b;t)=∫abψs(t)sds.I(a, b; t) = \int_a^b \frac{\psi_s(t)}{s}\, ds.(6)✦ At linear scale endpoints ε,R>0\varepsilon, R > 0, the smooth kernel Iε,R:ℝ→ℝI_{\varepsilon,R} \colon \mathbb{R} \to \mathbb{R} is the function on ℝ\mathbb{R} given for t∈ℝt \in \mathbb{R} by Iε,R(t)=I(ε2,R2;t)I_{\varepsilon,R}(t) = I(\varepsilon^2, R^2; t). The heat kernel constant κ∈ℝ\kappa \in \mathbb{R} is the real scalar defined by (7): κ=−23g3(0).\kappa = -\frac{2}{3}\, g_3(0).(7)✦ Using κ\kappa, the smooth kernel profile φ:ℝ→ℝ\varphi \colon \mathbb{R} \to \mathbb{R} is the real-valued function given for v∈ℝv \in \mathbb{R} by (8): φ(v)=κ−2(v29−13)g3(v).\varphi(v) = \kappa - 2\left(\frac{v^2}{9} - \frac{1}{3}\right) g_3(v).(8)✦ Whenever a,b,t∈ℝa, b, t \in \mathbb{R} satisfy 0<a≤b0 < a \le b and t≠0t \neq 0, the smooth kernel evaluates in terms of φ\varphi as (9): I(a,b;t)=φ(t/a)−φ(t/b)t.I(a, b; t) = \frac{\varphi(t/\sqrt{a}) - \varphi(t/\sqrt{b})}{t}.(9)✦

In the paper, the result reads: for s>0s > 0 and 0<ε<R0 < \varepsilon < R, ks(t)=(t/s)gs(t)k_s(t) = (t / \sqrt{s})\, g_s(t), ψ=k1*k1*k1\psi = k_1 * k_1 * k_1, and Iε,R(t)=∫ε2R2s−1/2ψ(t/s)ds/sI_{\varepsilon,R}(t) = \int_{\varepsilon^2}^{R^2} s^{-1/2} \psi(t / \sqrt{s})\, ds/s, where ψ\psi stands for ψ1\psi_1.

What Is Checked

The claim is stated over real scale parameters σ,s∈ℝ\sigma, s \in \mathbb{R} and squared scale endpoints a,b∈ℝa, b \in \mathbb{R} on the real line, defining from the Gaussian density gσg_\sigma the odd window ksk_s, the odd window ratio ρs\rho_s, the signed kernel ψs\psi_s, the smooth kernel I(a,b;⋅)I(a, b; \cdot), the heat kernel constant κ\kappa, and the smooth kernel profile φ\varphi, with s>0s > 0 and arbitrary real shifts for the triple-window integral and 0<a≤b0 < a \le b with t∈ℝ\{0}t \in \mathbb{R} \setminus \{0\} for the profile identity.