OctoInformBeta

OpenAI preprint, September 2026 · §3. Sheets and extremal derivatives

The variance identity: ⟨∥G−a∥2⟩\langle\|G-a\|^2\rangle through ⟨∥G∥2⟩\langle\|G\|^2\rangle, aa and ⟨G⟩\langle G\rangle; ∥φ(t)−φ(s)∥2≤|I|2⟨∥g∥2⟩I\|\varphi(t)-\varphi(s)\|^2\le|I|^2\langle\|g\|^2\rangle_I for φ=∫g\varphi=\int g and s,t∈Is,t\in I

The variance identity and an energy bound on segments.

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

Let (X,μ)(X,\mu) be a probability space and EE a real Hilbert space, with inner product ⟨u,v⟩\langle u,v\rangle. The mean of a function G:X→EG\colon X\to E is the vector ⟨G⟩=∫XGdμ\langle G\rangle=\int_X G\,d\mu of EE, and its second moment is the number ⟨∥G∥2⟩=∫X∥G∥2dμ\langle\|G\|^2\rangle=\int_X\|G\|^2\,d\mu. The question is how far GG strays from a fixed vector a∈Ea\in E in mean square, as measured by ⟨∥G−a∥2⟩\langle\|G-a\|^2\rangle, once the second moment, the vector aa and the gap ⟨G⟩−a\langle G\rangle-a between the mean and aa are known.

Let SS be the set under study, whose points are written (p,w)(p,w) with pp in the base plane, and ff a bi-Lipschitz map from SS to a finite-dimensional real space. For each ww, the sheet of ww is the map Fw:Ωw→ℝmF_w\colon\Omega_w\to\mathbb{R}^m, Fw(p)=f(p,w)F_w(p)=f(p,w), on the open set Ωw\Omega_w of base points pp with (p,w)∈S(p,w)\in S. A derivative column GG of a sheet FF is one of its partial derivatives FxF_x, FyF_y in the horizontal and vertical directions, taken over a rectangle QQ of its domain with μ\mu the normalized Lebesgue measure on QQ, so that the mean is ⟨G⟩Q\langle G\rangle_Q. The second moment of GG is set against AA, the largest first coordinate of the closure KK of the pairs (∥Fx(p)∥2,∥Fy(p)∥2)(\|F_x(p)\|^2,\|F_y(p)\|^2) over all sheets FF and the points pp of Ωw\Omega_w at which both partial derivatives exist.

The second question is in one variable: how large can the increment ∥φ(t)−φ(s)∥\|\varphi(t)-\varphi(s)\| of a function φ\varphi on the real line be between two points s,ts,t of an interval, when φ\varphi is an indefinite integral of a function gg and the energy of gg on the interval, the mean of ∥g∥2\|g\|^2 over it, is given?

Means on a Segment

A segment II is a closed interval [c,d][c,d] of the real line with c<dc<d✦; its length is |I|=d−c|I|=d-c.✦ The uniform law on II is the probability measure μI=|I|−1λ|I\mu_I=|I|^{-1}\lambda|_I, where λ\lambda is Lebesgue measure on the line.✦ The mean over II of a function hh integrable on II is ⟨h⟩I=∫hdμI=1|I|∫cdh(x)dx.\langle h\rangle_I=\int h\,d\mu_I=\frac{1}{|I|}\int_c^d h(x)\,dx.(1)

A bounded measurable field on a measurable space XX with values in a real normed space EE is a strongly measurable map g:X→Eg\colon X\to E that is bounded: there is a constant CC with ∥g(x)∥≤C\|g(x)\|\le C for all x∈Xx\in X.✦

The Claim

Theorem 1. Let (X,μ)(X,\mu) be a probability space, EE a real Hilbert space, G:X→EG\colon X\to E a function such that GG and ∥G∥2\|G\|^2 are integrable, and a∈Ea\in E. Then GG satisfies the variance identity ⟨∥G−a∥2⟩=⟨∥G∥2⟩−∥a∥2−2⟨a,⟨G⟩−a⟩,\langle\|G-a\|^2\rangle=\langle\|G\|^2\rangle-\|a\|^2-2\langle a,\langle G\rangle-a\rangle,(2)✦ and for every real number AA the deficit of its second moment below AA satisfies A−⟨∥G∥2⟩≤A−∥a∥2+2∥a∥∥⟨G⟩−a∥.A-\langle\|G\|^2\rangle\le A-\|a\|^2+2\|a\|\,\|\langle G\rangle-a\|.(3)✦ If moreover ⟨∥G∥2⟩≤M\langle\|G\|^2\rangle\le M for a real number MM, then ⟨∥G−a∥2⟩≤M−∥a∥2+2∥a∥∥⟨G⟩−a∥.\langle\|G-a\|^2\rangle\le M-\|a\|^2+2\|a\|\,\|\langle G\rangle-a\|.(4)✦

Theorem 2. Let EE be a real Banach space, I=[c,d]I=[c,d] a segment, g:ℝ→Eg\colon\mathbb{R}\to E a bounded measurable field, and φ:ℝ→E\varphi\colon\mathbb{R}\to E an indefinite integral of gg: φ(t)−φ(s)=∫stg(x)dx\varphi(t)-\varphi(s)=\int_s^t g(x)\,dx for all real numbers ss and tt. Then for all s,t∈Is,t\in I the increment of φ\varphi is bounded by the energy of gg on II: ∥φ(t)−φ(s)∥2≤|I|2⟨∥g∥2⟩I.\|\varphi(t)-\varphi(s)\|^2\le|I|^2\,\langle\|g\|^2\rangle_I.(5)✦

What Is Checked

The identity (2) and the bounds (3) and (4) are stated over an arbitrary probability space and a function GG from it to a real Hilbert space with GG and ∥G∥2\|G\|^2 integrable; the bound (5) is stated over a segment II with its length |I||I| and its uniform law μI\mu_I, a bounded measurable field gg on the line with values in a real Banach space, and an indefinite integral φ\varphi of gg.