OctoInformBeta

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

Lemma 6.3: For 0<r≤R0 < r \le R, splitting the smooth kernel at scale endpoints r2r^2 and R2R^2 with a Gaussian envelope of variance 6r26r^2 at scale rr

Hard-kernel decomposition and Gaussian endpoint-error bounds.

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 real scale parameters 0<r≤R0 < r \le R and a real point tt. Integrating a signed Gaussian-weighted kernel between the squared scales r2r^2 and R2R^2 produces a smooth kernel, and the question is how it splits into a constant multiple of a hard-truncated Hilbert kernel on r<|t|≤Rr < |t| \le R and a difference of endpoint errors at scales rr and RR, each bounded by a multiple of a Gaussian density.

Kernels and Error Profiles

For each real variance σ\sigma, the Gaussian density gσ:ℝ→ℝg_\sigma \colon \mathbb{R} \to \mathbb{R} is the real-valued function defined for t∈ℝt \in \mathbb{R} by (1): gσ(t)=(2πσ)−1/2exp⁡(−t22σ).g_\sigma(t) = (2\pi\sigma)^{-1/2} \exp\!\left(-\frac{t^2}{2\sigma}\right).(1)✦ The signed kernel κ:ℝ×ℝ→ℝ\kappa \colon \mathbb{R} \times \mathbb{R} \to \mathbb{R} is the function of two real variables (s,t)(s, t) given by (2): κ(s,t)=t3/27−ts/3ssg3s(t).\kappa(s,t) = \frac{t^3/27 - ts/3}{s\sqrt{s}}\, g_{3s}(t).(2)✦ Given real scale endpoints aa and bb, the smooth kernel ℐa,b:ℝ→ℝ\mathcal{I}_{a,b} \colon \mathbb{R} \to \mathbb{R} is the real-valued function defined for t∈ℝt \in \mathbb{R} by (3): ℐa,b(t)=∫abκ(s,t)sds.\mathcal{I}_{a,b}(t) = \int_a^b \frac{\kappa(s,t)}{s}\,ds.(3)✦ For real scale parameters rr and RR, the smooth kernel over squared scale endpoints Ir,R:ℝ→ℝI_{r,R} \colon \mathbb{R} \to \mathbb{R} is the function defined for t∈ℝt \in \mathbb{R} by Ir,R(t)=ℐr2,R2(t)I_{r,R}(t) = \mathcal{I}_{r^2, R^2}(t).

The heat kernel constant c0c_0 is the real number −23g3(0)-\frac{2}{3} g_3(0).✦ Define the smooth kernel profile ψ:ℝ→ℝ\psi \colon \mathbb{R} \to \mathbb{R} as the real function given at v∈ℝv \in \mathbb{R} by (4): ψ(v)=c0−2(v29−13)g3(v).\psi(v) = c_0 - 2\left(\frac{v^2}{9} - \frac{1}{3}\right) g_3(v).(4)✦ Write e:ℝ→ℝe \colon \mathbb{R} \to \mathbb{R} for the kernel error profile, the function on ℝ\mathbb{R} satisfying (5) for each v∈ℝv \in \mathbb{R}: e(v)=ψ(v)−c0𝟏|v|>1.e(v) = \psi(v) - c_0 \mathbf{1}_{|v| > 1}.(5)✦

For real scale parameters rr and RR, the hard-truncated Hilbert kernel Hr,R:ℝ→ℝH_{r,R} \colon \mathbb{R} \to \mathbb{R} is the real function given for t≠0t \neq 0 by (6) with Hr,R(0)=0H_{r,R}(0) = 0: Hr,R(t)=𝟏r<|t|−𝟏R<|t|t.H_{r,R}(t) = \frac{\mathbf{1}_{r < |t|} - \mathbf{1}_{R < |t|}}{t}.(6)✦ At any real scale parameter rr, the endpoint error Er:ℝ→ℝE_r \colon \mathbb{R} \to \mathbb{R} is the real-valued function satisfying (7) for t≠0t \neq 0, with Er(0)=0E_r(0) = 0: Er(t)=e(t/r)t.E_r(t) = \frac{e(t/r)}{t}.(7) The error envelope constant KK is a chosen positive real number such that |e(v)|≤K|v|g6(v)|e(v)| \le K |v| g_6(v) for all v∈ℝv \in \mathbb{R}.✦

The Claim

Theorem 1. If r,R,t∈ℝr, R, t \in \mathbb{R} satisfy 0<r≤R0 < r \le R, then the smooth kernel decomposes according to (8): Ir,R(t)=c0Hr,R(t)+Er(t)−ER(t).I_{r,R}(t) = c_0 H_{r,R}(t) + E_r(t) - E_R(t).(8)✦ Moreover, if r,t∈ℝr, t \in \mathbb{R} satisfy r>0r > 0, then the endpoint error satisfies the Gaussian bound (9): |Er(t)|≤Kg6r2(t).|E_r(t)| \le K g_{6r^2}(t).(9)✦

In the paper, the result reads: there are an absolute constant CC and functions ErE_r (r>0r > 0) such that, for t≠0t \neq 0, apart from endpoint values, Iε,R(t)=(c0/t)𝟏{ε<|t|<R}+Eε(t)−ER(t)I_{\varepsilon,R}(t) = (c_0 / t)\mathbf{1}_{\{\varepsilon < |t| < R\}} + E_\varepsilon(t) - E_R(t) with c0=−2/(36π)c_0 = -2 / (3\sqrt{6\pi}), and |Er(t)|≤(C/r)(1+|t|/r)−2|E_r(t)| \le (C / r)(1 + |t|/r)^{-2} for t≠0t \neq 0, where Ir,RI_{r,R} with r=εr = \varepsilon stands for Iε,RI_{\varepsilon,R}.

What Is Checked

The decomposition is stated for real scales 0<r≤R0 < r \le R and t∈ℝt \in \mathbb{R}, and the endpoint bound for real r>0r > 0 and t∈ℝt \in \mathbb{R}, over the smooth kernel Ir,RI_{r,R} built from the signed kernel κ\kappa and Gaussian densities gσg_\sigma, the heat kernel constant c0c_0, the hard-truncated Hilbert kernel Hr,RH_{r,R}, the endpoint error ErE_r built from the smooth kernel profile ψ\psi and the kernel error profile ee, and the error envelope constant KK. The library also proves the Gaussian bound for the error profile: there exists a positive real constant KK such that for all v∈ℝv \in \mathbb{R}, |e(v)|≤K|v|g6(v)|e(v)| \le K |v| g_6(v).✦