OctoInformBeta

OpenAI preprint, September 2026 · Section 4. Finite-rank operators and signed measures

Section 4 (setting): affine signed weights and label bumps: Finite signed pairings and distance cutoffs for nonexpansive label maps

Affine weights and Lipschitz label tests.

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 (M,d,o)(M,d,o) be a pointed metric space. The setting concerns finite signed weights on MM and Lipschitz functions concentrated near a prescribed label.

The Lipschitz space Lip⁡0(M)\operatorname{Lip}_0(M) is the real vector space of functions f:M→ℝf:M\to\mathbb R with f(o)=0f(o)=0 that admit a constant C≥0C\geq0 satisfying |f(x)−f(y)|≤Cd(x,y)|f(x)-f(y)|\leq C d(x,y) for all x,y∈Mx,y\in M.✦ The Lipschitz norm ∥f∥Lip\|f\|_{\mathrm{Lip}} of f∈Lip⁡0(M)f\in\operatorname{Lip}_0(M) is the supremum in (1). ∥f∥Lip=supx≠y|f(x)−f(y)|d(x,y).\|f\|_{\mathrm{Lip}} =\sup_{x\ne y}\frac{|f(x)-f(y)|}{d(x,y)}.(1) For x∈Mx\in M, evaluation δx\delta_x on Lip⁡0(M)\operatorname{Lip}_0(M) is the bounded linear functional given by δx(f)=f(x)\delta_x(f)=f(x). The Lipschitz-free space ℱ(M)\mathcal F(M) is the closed evaluation span in the dual, as in (2). ℱ(M)=span⁡ℝ{δx:x∈M}¯⊆Lip⁡0(M)*.\mathcal F(M) =\overline{\operatorname{span}_{\mathbb R}\{\delta_x:x\in M\}} \subseteq\operatorname{Lip}_0(M)^*.(2)✦ For f∈Lip⁡0(M)f\in\operatorname{Lip}_0(M), the canonical dual pairing is the continuous linear functional u↦u(f)u\mapsto u(f) on ℱ(M)\mathcal F(M); on a finite sum u=∑xaxδxu=\sum_x a_x\delta_x its value is ∑xaxf(x)\sum_x a_xf(x).✦

Assume that MM is unit-separated, meaning that d(x,y)≥1d(x,y)\geq1 for distinct points. The Lipschitz point indicator ιj∈Lip⁡0(M)\iota_j\in\operatorname{Lip}_0(M) at j∈Mj\in M is the function in (3). For j∈Mj\in M, the coefficient functional cjc_j on ℱ(M)\mathcal F(M) associated with the point indicator ιj\iota_j is the functional given by cj(u)=u(ιj)c_j(u)=u(\iota_j). ιj(x)={1,x=j and j≠o,0,otherwise.\iota_j(x)= \begin{cases} 1,&x=j\text{ and }j\ne o,\\ 0,&\text{otherwise}. \end{cases}(3)✦

The Claim

Definition 1. Let S⊆MS\subseteq M be finite and contain oo, and let u∈ℱ(M)u\in\mathcal F(M). The affine signed weights μu:S→ℝ\mu_u:S\to\mathbb R are defined by (4). μu(j)=cj(u)+{1−∑t∈Sct(u),j=o,0,j≠o.\mu_u(j)=c_j(u)+ \begin{cases} 1-\displaystyle\sum_{t\in S}c_t(u),&j=o,\\ 0,&j\ne o. \end{cases}(4)✦ The total variation ∥μ∥TV\|\mu\|_{\mathrm{TV}} of real weights μ:S→ℝ\mu:S\to\mathbb R is the sum in (5). For a subset E⊆SE\subseteq S, the variation |μ|(E)|\mu|(E) of real weights μ:S→ℝ\mu:S\to\mathbb R on EE is the sum in (6). ∥μ∥TV=∑j∈S|μ(j)|.\|\mu\|_{\mathrm{TV}}=\sum_{j\in S}|\mu(j)|.(5)✦ |μ|(E)=∑j∈E|μ(j)|.|\mu|(E)=\sum_{j\in E}|\mu(j)|.(6)✦ The signed measure pairing ⟨μ,f⟩\langle\mu,f\rangle of real weights μ:S→ℝ\mu:S\to\mathbb R with a function f:S→ℝf:S\to\mathbb R is the finite sum in (7). ⟨μ,f⟩=∑j∈Sμ(j)f(j).\langle\mu,f\rangle=\sum_{j\in S}\mu(j)f(j).(7)✦

Theorem 1. Let (N,ρ)(N,\rho) be a metric space, let h:M→Nh:M\to N be 11-Lipschitz, and fix an anchor a∈Na\in N. Let L≥0L\geq0 and K>0K>0. Define the label bump function ba:M→ℝb_a:M\to\mathbb R by (8). ba(t)=Lmax⁡{1−ρ(a,h(t))K,0}.b_a(t)=L\max\left\{1-\frac{\rho(a,h(t))}{K},0\right\}.(8)✦ Its differences satisfy (9) for all x,y∈Mx,y\in M. |ba(x)−ba(y)|≤LKd(x,y).|b_a(x)-b_a(y)|\leq\frac{L}{K}d(x,y).(9)✦ If L≤KL\leq K, the bump is nonexpansive: its Lipschitz constant Lip⁡(ba)\operatorname{Lip}(b_a) satisfies (10). Lip⁡(ba)≤1.\operatorname{Lip}(b_a)\leq1.(10)✦

In the paper, the result reads: for an operator TT supported on a finite set SS containing oo and a source vertex xx, the nonbasepoint coefficients of TδxT\delta_x, with the remaining mass assigned to oo, form a signed measure μx\mu_x of mass one, with |μ|(E)=∑t∈E|μ(t)||\mu|(E)=\sum_{t\in E}|\mu(t)| and ∥μ∥TV=|μ|(S)\|\mu\|_{\mathrm{TV}}=|\mu|(S), and for the label map hh into a finite box with the supremum norm, the bump bx(t)=Lmax⁡{1−∥h(x)−h(t)∥∞/K,0}b_x(t)=L\max\{1-\|h(x)-h(t)\|_\infty/K,0\} takes values in [0,L][0,L], equals LL at its anchor, and has Lipschitz constant at most L/K≤1L/K\leq1.

What Is Checked

The affine weights are stated over the coefficient functionals on the Lipschitz-free space of a unit-separated pointed metric space and a finite set containing the basepoint; the bump estimates concern a nonexpansive map into an arbitrary metric label space. Evaluation at xx has operator norm at most d(x,o)d(x,o), and the point indicators used in the coefficient functionals have Lipschitz constant at most 11.