OctoInformBeta

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

Lemma 4.1: Support projections approximate finite-rank operators on bounded unit-separated spaces, with control of norm and anchor errors

Finite-support truncation of finite-rank operators.

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

Finite-support approximation of an operator on a Lipschitz-free space asks for a common finite set of evaluation vectors for its outputs, with control of the error at prescribed anchors. Let (M,d,o)(M,d,o) be a pointed metric space satisfying unit separation: d(x,y)≥1(x≠y).d(x,y)\geq 1 \qquad (x\ne y).(1)

The real Lipschitz space Lip⁡0(M)\operatorname{Lip}_0(M) is the normed linear space of real functions ff on MM satisfying (2) for some nonnegative Lipschitz bound CC, equipped with the Lipschitz norm: f(o)=0,|f(x)−f(y)|≤Cd(x,y)(x,y∈M),C≥0.f(o)=0,\qquad |f(x)-f(y)|\leq C d(x,y) \quad (x,y\in M),\qquad C\geq 0.(2)✦

Evaluation at x∈Mx\in M is the continuous linear functional δx\delta_x on Lip⁡0(M)\operatorname{Lip}_0(M) given by (3): δx(f)=f(x).\delta_x(f)=f(x).(3)✦ The Lipschitz-free space ℱ(M)\mathcal{F}(M) is the Banach space given by the closed evaluation span (4) in Lip⁡0(M)*\operatorname{Lip}_0(M)^*; the Dirac embedding x↦δxx\mapsto\delta_x regards these evaluations as elements of ℱ(M)\mathcal{F}(M): ℱ(M)=span⁡ℝ{δx:x∈M}¯.\mathcal{F}(M)= \overline{\operatorname{span}_{\mathbb R}\{\delta_x:x\in M\}}.(4)

The canonical dual pairing of u∈ℱ(M)u\in\mathcal{F}(M) with f∈Lip⁡0(M)f\in\operatorname{Lip}_0(M) is evaluation of uu at ff, as in (5); here J⊂MJ\subset M is finite and the coefficients aja_j are real: ⟨u,f⟩=u(f),⟨∑j∈Jajδj,f⟩=∑j∈Jajf(j).\langle u,f\rangle=u(f),\qquad \left\langle\sum_{j\in J}a_j\delta_j,f\right\rangle =\sum_{j\in J}a_j f(j).(5)✦

For j∈Mj\in M, the Lipschitz point indicator χj∈Lip⁡0(M)\chi_j\in\operatorname{Lip}_0(M) is the function specified by (6): χj(x)={1,x=j and j≠o,0,otherwise.\chi_j(x)= \begin{cases} 1,&x=j\text{ and }j\ne o,\\ 0,&\text{otherwise}. \end{cases}(6)✦ The coefficient functional cj∈ℱ(M)*c_j\in\mathcal{F}(M)^* is the canonical dual pairing against χj\chi_j, as in (7): cj(u)=⟨u,χj⟩.c_j(u)=\langle u,\chi_j\rangle.(7)✦ For a finite set F⊂MF\subset M, the finite-support projection PF:ℱ(M)→ℱ(M)P_F:\mathcal{F}(M)\to\mathcal{F}(M) is the bounded linear operator defined by (8): PFu=∑j∈Fcj(u)δj.P_Fu=\sum_{j\in F}c_j(u)\delta_j.(8)

The Claim

Theorem 1. Suppose that d(x,o)≤Bd(x,o)\leq B for every x∈Mx\in M, where B≥0B\geq 0. Let λ≥1\lambda\geq 1 be real, let A⊂MA\subset M be a finite set of anchors, and let T0:ℱ(M)→ℱ(M)T_0:\mathcal{F}(M)\to\mathcal{F}(M) be a bounded finite-rank operator with ∥T0∥≤λ\|T_0\|\leq\lambda and ∥T0δa−δa∥<1/2\|T_0\delta_a-\delta_a\|<1/2 for every a∈Aa\in A. There are a finite set S⊂MS\subset M and a bounded finite-rank operator T:ℱ(M)→ℱ(M)T:\mathcal{F}(M)\to\mathcal{F}(M) satisfying (9): o∈S,∥T∥≤2λ,PST=T,∥Tδa−δa∥<1(a∈A).o\in S,\qquad \|T\|\leq 2\lambda,\qquad P_ST=T,\qquad \|T\delta_a-\delta_a\|<1\quad(a\in A).(9)✦

On the same bounded space, let U:ℱ(M)→ℱ(M)U:\mathcal{F}(M)\to\mathcal{F}(M) be any bounded finite-rank operator and let ε>0\varepsilon>0. There is a finite set Fε⊂MF_\varepsilon\subset M for which (10) holds: ∥PGU−U∥<εfor every finite G⊂M containing Fε.\|P_GU-U\|<\varepsilon \qquad\text{for every finite }G\subset M\text{ containing }F_\varepsilon.(10)✦

On an arbitrary pointed metric space satisfying unit separation, let F⊂MF\subset M be finite and let u∈span⁡ℝ{δj:j∈F}u\in\operatorname{span}_{\mathbb R}\{\delta_j:j\in F\}. The finite-support projection satisfies (11): PFu=u.P_Fu=u.(11)✦

In the paper, the result reads: for an integer p≥1p\geq 1 and its constructed block MpM_p, a bounded finite-rank operator T0T_0 on ℱ(Mp)\mathcal{F}(M_p) with ∥T0∥≤p\|T_0\|\leq p and error less than 1/21/2 at every anchor admits a finite set SS containing oo and a bounded finite-rank operator TT with range in span⁡ℝ{δj:j∈S}\operatorname{span}_{\mathbb R}\{\delta_j:j\in S\}, norm at most 2p2p, and error less than 11 at every anchor.

What Is Checked

The reduction and operator norm approximation are stated over bounded, unit-separated pointed metric spaces and bounded finite-rank operators on their closed evaluation spans; the assertion about a finite evaluation span is stated under unit separation alone. The evaluation functionals have norm at most d(x,o)d(x,o), and the point indicators have Lipschitz norm at most 11.