OctoInformBeta

OpenAI preprint, September 2026 · §2. Finite coordinates and normalized increments

§2 (setting): edges and normalized gradients: Positive pairwise coordinate distances give unit normalized pp-energy for p>0p>0

Normalizing coordinate increments on vertex pairs.

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

A scalar coordinate of a finite configuration assigns a real number to each vertex. Its increments along oriented pairs can be normalized by the distances between the corresponding points.

Let n,kn,k be nonnegative integers. For a real exponent pp, the coordinate distance ρp\rho_p on ℝk\mathbb R^k is the function in (1). ρp(x,y)=(∑a=1k|xa−ya|p)1/p.\rho_p(x,y)=\left(\sum_{a=1}^k |x_a-y_a|^p\right)^{1/p}.(1)✦

Definition 1. The standard pair index PnP_n is the set of pairs of vertices oriented by increasing index: Pn={(i,j)∈{1,…,n}2:i<j}.P_n=\{(i,j)\in\{1,\ldots,n\}^2:i<j\}.(2)✦

Incidence and Normalization

Let UU be a finite vertex set and EE a finite edge set, with endpoint maps s,t:E→Us,t:E\to U. The incidence matrix B∈ℝE×UB\in\mathbb R^{E\times U} is the matrix Bei=𝟏{i=s(e)}−𝟏{i=t(e)}.B_{ei}=\mathbf 1_{\{i=s(e)\}}-\mathbf 1_{\{i=t(e)\}}.(3)✦

Definition 2. For a real edge scale δ:E→ℝ\delta:E\to\mathbb R, with the convention 0−1=00^{-1}=0, the normalized gradient Gδ:ℝU→ℝEG_\delta:\mathbb R^U\to\mathbb R^E is the linear map whose matrix, in terms of the incidence matrix in (3), is Gδ=diag⁡((δe−1)e∈E)B.G_\delta=\operatorname{diag}\bigl((\delta_e^{-1})_{e\in E}\bigr)B.(4)✦

The Claim

Theorem 1. Let E,UE,U be finite sets, let s,t:E→Us,t:E\to U and δ:E→ℝ\delta:E\to\mathbb R, and let z∈ℝUz\in\mathbb R^U. The normalized gradient in (4) satisfies, for every e∈Ee\in E, (Gδz)e=zs(e)−zt(e)δe.(G_\delta z)_e=\frac{z_{s(e)}-z_{t(e)}}{\delta_e}.(5)✦

Let p>0p>0 and y1,…,yn∈ℝky_1,\ldots,y_n\in\mathbb R^k. On the pair index PnP_n in (2), take s(i,j)=is(i,j)=i and t(i,j)=jt(i,j)=j, and suppose that δ(i,j)=ρp(yi,yj)>0\delta_{(i,j)}=\rho_p(y_i,y_j)>0, with ρp\rho_p as in (1). For each coordinate a∈{1,…,k}a\in\{1,\ldots,k\}, let z(a)∈ℝnz^{(a)}\in\mathbb R^n be the scalar coordinate function zi(a)=yiaz^{(a)}_i=y_{ia}. Then, for every e∈Pne\in P_n, ∑a=1k|(Gδz(a))e|p=1.\sum_{a=1}^k\bigl|(G_\delta z^{(a)})_e\bigr|^p=1.(6)✦

What Is Checked

The library states (5) over finite oriented edge sets with arbitrary real edge scales under the zero inverse convention, and (6) over finite real coordinate configurations with p>0p>0 and positive pairwise coordinate distances.