OctoInformBeta

OpenAI preprint, September 2026 · §3. A Hilbert kernel with chart extensions

Lemma 3.1: Positivity of the profile constant c*c^* and explicit error bounds for the scaled profile integral and the base kernel distance

Approximating Euclidean distance with a truncated Hilbert kernel.

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 points θ\theta and η\eta together with a finite family of vectors in a real inner-product space EE with norm ∥⋅∥\|\cdot\|, and fix scale bounds 0<a≤b0 < a \le b. To approximate a constant multiple of the Euclidean distance ∥η−θ∥\|\eta - \theta\| by a kernel integrated over scales in [a,b][a, b], one combines a Gaussian decay term in ∥η−θ∥\|\eta - \theta\| with trigonometric perturbations along the vector family. The question is which positive constant multiplies ∥η−θ∥\|\eta - \theta\|, and what explicit error bounds hold for truncating the underlying profile integral and for the base kernel distance on [a,b][a, b].

Definitions

The profile function ϕ:ℝ→ℝ\phi : \mathbb{R} \to \mathbb{R} is the real-valued function mapping each t∈ℝt \in \mathbb{R} to ϕ(t)=1−e−1/(2t2).\phi(t) = 1 - e^{-1/(2t^2)}.(1)✦

On pairs (r,t)∈ℝ2(r, t) \in \mathbb{R}^2, the scaled profile ψ:ℝ2→ℝ\psi : \mathbb{R}^2 \to \mathbb{R} is the two-variable function defined by ψ(r,t)=1−e−r2/(2t2).\psi(r, t) = 1 - e^{-r^2/(2t^2)}.(2)✦

From the profile function ϕ\phi, the profile constant c*∈ℝc^* \in \mathbb{R} is the real number given by c*=2∫0∞ϕ(t)dt.c^* = 2 \int_0^\infty \phi(t)\,dt.(3)✦

For a real inner-product space EE with inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle and norm ∥⋅∥\|\cdot\|, a finite index set II, a vector family g:I→Eg : I \to E, real weights q,ε∈ℝq, \varepsilon \in \mathbb{R}, a scale σ∈ℝ\sigma \in \mathbb{R}, points θ,η∈E\theta, \eta \in E, and coefficient functions α,β:I→ℝ\alpha, \beta : I \to \mathbb{R}, the kernel pairing Pg,q,ε(σ;θ,η,α,β)∈ℝP_{g,q,\varepsilon}(\sigma; \theta, \eta, \alpha, \beta) \in \mathbb{R} is the real scalar defined by Pg,q,ε(σ;θ,η,α,β)=qe−∥η−θ∥2/(2σ2)+εσ∑i∈I((1+αiβi)cos⁡(⟨gi,η−θ⟩/σ)+(αi−βi)sin⁡(⟨gi,η−θ⟩/σ)).P_{g,q,\varepsilon}(\sigma; \theta, \eta, \alpha, \beta) = q e^{-\|\eta - \theta\|^2/(2\sigma^2)} + \frac{\varepsilon}{\sigma} \sum_{i \in I} \Bigl( (1 + \alpha_i \beta_i)\cos\bigl(\langle g_i, \eta - \theta\rangle / \sigma\bigr) + (\alpha_i - \beta_i)\sin\bigl(\langle g_i, \eta - \theta\rangle / \sigma\bigr) \Bigr).(4)✦

The Hilbert kernel Kg,q,ε,a,b(θ,η;α,β)∈ℝK_{g,q,\varepsilon,a,b}(\theta, \eta; \alpha, \beta) \in \mathbb{R} for a vector family g:I→Eg : I \to E in a real inner-product space EE indexed by a finite set II, real weights q,ε∈ℝq, \varepsilon \in \mathbb{R}, integration bounds a,b∈ℝa, b \in \mathbb{R}, points θ,η∈E\theta, \eta \in E, and coefficient functions α,β:I→ℝ\alpha, \beta : I \to \mathbb{R} is the real number defined by Kg,q,ε,a,b(θ,η;α,β)=∫abPg,q,ε(σ;θ,η,α,β)dσ.K_{g,q,\varepsilon,a,b}(\theta, \eta; \alpha, \beta) = \int_a^b P_{g,q,\varepsilon}(\sigma; \theta, \eta, \alpha, \beta)\,d\sigma.(5)✦

The Claim

Theorem 1. The profile constant c*c^* is strictly positive: c*>0.c^* > 0.(6)✦ For all real numbers r,a,b∈ℝr, a, b \in \mathbb{R} satisfying r≥0r \ge 0 and 0<a≤b0 < a \le b, |2∫abψ(r,t)dt−c*r|≤2a+r2b.\left| 2 \int_a^b \psi(r, t)\,dt - c^* r \right| \le 2a + \frac{r^2}{b}.(7)✦ Given a real inner-product space EE, a finite index set II, a vector family g:I→Eg : I \to E, real parameters q,ε≥0q, \varepsilon \ge 0 and 0<a≤b0 < a \le b, and points θ,η∈E\theta, \eta \in E, letting 00 denote the identically zero coefficient function on II, |Kg,q,ε,a,b(θ,θ;0,0)+Kg,q,ε,a,b(η,η;0,0)−2Kg,q,ε,a,b(θ,η;0,0)−c*q∥η−θ∥|≤q(2a+∥η−θ∥2b)+4ε|I|ln⁡ba.\bigl| K_{g,q,\varepsilon,a,b}(\theta, \theta; 0, 0) + K_{g,q,\varepsilon,a,b}(\eta, \eta; 0, 0) - 2 K_{g,q,\varepsilon,a,b}(\theta, \eta; 0, 0) - c^* q \|\eta - \theta\| \bigr| \le q\left(2a + \frac{\|\eta - \theta\|^2}{b}\right) + 4\varepsilon |I| \ln\frac{b}{a}.(8)✦

In the paper, the result reads: for a large positive integer mm and a map pp on EE into a normed space, the constant c*c^* (written c*c_* in the paper) satisfies c*=2∫0∞(1−e−1/(2t2))dt>0c^* = 2 \int_0^\infty (1 - e^{-1/(2t^2)})\,dt > 0 and |∥p(θ)−p(η)∥2−c*m∥θ−η∥|≤m−37\bigl|\|p(\theta) - p(\eta)\|^2 - c^* \sqrt{m}\,\|\theta - \eta\|\bigr| \le m^{-37} whenever θ,η∈E\theta, \eta \in E satisfy ∥θ∥,∥η∥≤3m\|\theta\|, \|\eta\| \le 3\sqrt{m}. The library answers the paper’s bound through (6) and (7) together with (8), using the kernel difference Kg,q,ε,a,b(θ,θ;0,0)+Kg,q,ε,a,b(η,η;0,0)−2Kg,q,ε,a,b(θ,η;0,0)K_{g,q,\varepsilon,a,b}(\theta, \theta; 0, 0) + K_{g,q,\varepsilon,a,b}(\eta, \eta; 0, 0) - 2 K_{g,q,\varepsilon,a,b}(\theta, \eta; 0, 0) in place of ∥p(θ)−p(η)∥2\|p(\theta) - p(\eta)\|^2 while keeping qq, ε\varepsilon, aa, bb, θ\theta, and η\eta unspecialized.

What Is Checked

Claims (6), (7), and (8) are stated over the profile constant c*c^* defined from the profile function ϕ\phi, together with the scaled profile ψ\psi at parameters r≥0r \ge 0 and 0<a≤b0 < a \le b and the Hilbert kernel KK built from the kernel pairing PP for a finite vector family g:I→Eg : I \to E in a real inner-product space EE, weights q,ε≥0q, \varepsilon \ge 0, integration bounds 0<a≤b0 < a \le b, points θ,η∈E\theta, \eta \in E, and zero coefficient functions.