OctoInformBeta

OpenAI preprint, September 2026 · Section 2. From bounded blocks to AP without BAP

Section 2 (setting): metric wedges, linearized retractions and AP: Separated bounded components yield the approximation property, while every bounded approximation constant passes to each component

Metric wedges give contractive retractions and approximation.

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

Joining pointed metric spaces at their basepoints gives a space whose components can be retained separately. The approximation questions concern finite-rank approximation of the identity uniformly on compact subsets of its Lipschitz-free space, and bounds on the norms of the approximating operators.

For a pointed metric space (M,d,o)(M,d,o), the Lipschitz space Lip⁡0(M)\operatorname{Lip}_0(M) is the real linear space of functions f:M→ℝf:M\to\mathbb R that vanish at oo and admit a bound |f(x)−f(y)|≤Cd(x,y)|f(x)-f(y)|\leq C d(x,y) for some C≥0C\geq0 and all x,y∈Mx,y\in M.✦ The Lipschitz norm ∥f∥Lip\|f\|_{\operatorname{Lip}} of a function f∈Lip⁡0(M)f\in\operatorname{Lip}_0(M) is given by (1). ∥f∥Lip=supx≠y|f(x)−f(y)|d(x,y).\|f\|_{\operatorname{Lip}} =\sup_{x\ne y}\frac{|f(x)-f(y)|}{d(x,y)}.(1) The evaluation functional δx∈Lip⁡0(M)*\delta_x\in\operatorname{Lip}_0(M)^* and the Lipschitz-free Banach space ℱ(M)\mathcal F(M) are defined by (2); the Dirac map is x↦δxx\mapsto\delta_x, with values in this closed evaluation span. δx(f)=f(x),ℱ(M)=span⁡ℝ{δx:x∈M}¯⊆Lip⁡0(M)*.\delta_x(f)=f(x),\qquad \mathcal F(M)=\overline{\operatorname{span}_{\mathbb R} \{\delta_x:x\in M\}} \subseteq\operatorname{Lip}_0(M)^*.(2)

A real normed space XX has the approximation property (AP) if, for every compact C⊆XC\subseteq X and every ε>0\varepsilon>0, there are a continuous linear operator T:X→XT:X\to X with finite-dimensional range and a real number η\eta satisfying (3). η<ε,∥Tu−u∥≤η(u∈C).\eta<\varepsilon,\qquad \|Tu-u\|\leq\eta\quad(u\in C).(3)✦ For λ∈ℝ\lambda\in\mathbb R, a real normed space XX has the bounded approximation property (BAP) with bound λ\lambda, written λ\lambda-BAP, if it has the approximation property with approximating operators satisfying ∥T∥≤λ\|T\|\leq\lambda.

Definition 1. Let (Mn,dn,on)n∈ℕ(M_n,d_n,o_n)_{n\in\mathbb N} be a countable family of pointed metric spaces. The metric wedge WW is the disjoint union of the sets {n}×(Mn\{on})\{n\}\times(M_n\setminus\{o_n\}), together with a common basepoint oo.✦ The distance dWd_W on the metric wedge WW is defined by (4), where (n,x)(n,x) and (m,y)(m,y) denote points different from oo. dW(o,o)=0,dW(o,(n,x))=dW((n,x),o)=dn(x,on),dW((n,x),(m,y))={dn(x,y),n=m,dn(x,on)+dm(y,om),n≠m.\begin{aligned} d_W(o,o)&=0,\\ d_W(o,(n,x))=d_W((n,x),o)&=d_n(x,o_n),\\ d_W((n,x),(m,y))&= \begin{cases} d_n(x,y),&n=m,\\ d_n(x,o_n)+d_m(y,o_m),&n\ne m. \end{cases} \end{aligned}(4)✦

The block inclusion in:Mn→Wi_n:M_n\to W and block retraction rn:W→Mnr_n:W\to M_n are the maps defined in (5). in(on)=o,in(x)=(n,x)(x≠on),rn(o)=on,rn((m,y))={y,m=n,on,m≠n.\begin{aligned} i_n(o_n)&=o,&i_n(x)&=(n,x)\quad(x\ne o_n),\\ r_n(o)&=o_n,&r_n((m,y))&= \begin{cases}y,&m=n,\\o_n,&m\ne n.\end{cases} \end{aligned}(5) For a finite set A⊆ℕA\subseteq\mathbb N, the finite-block truncation ρA:W→W\rho_A:W\to W is the pointed 11-Lipschitz map defined by (6). ρA(o)=o,ρA((n,x))={(n,x),n∈A,o,n∉A.\rho_A(o)=o,\qquad \rho_A((n,x))= \begin{cases}(n,x),&n\in A,\\o,&n\notin A.\end{cases}(6)

For pointed metric spaces (M,d,o)(M,d,o) and (N,dN,b)(N,d_N,b), a pointed 11-Lipschitz map f:M→Nf:M\to N has Lipschitz pullback f*:Lip⁡0(N)→Lip⁡0(M)f^*: \operatorname{Lip}_0(N)\to\operatorname{Lip}_0(M) and linearization fˆ:ℱ(M)→ℱ(N)\widehat f:\mathcal F(M)\to\mathcal F(N), the continuous linear maps defined by (7). f*g=g∘f,⟨fˆu,g⟩=⟨u,f*g⟩(u∈ℱ(M),g∈Lip⁡0(N)).f^*g=g\circ f,\qquad \langle\widehat f u,g\rangle=\langle u,f^*g\rangle \quad\bigl(u\in\mathcal F(M),\ g\in\operatorname{Lip}_0(N)\bigr).(7) The linearized block inclusions In:ℱ(Mn)→ℱ(W)I_n:\mathcal F(M_n)\to\mathcal F(W), block retractions Pn:ℱ(W)→ℱ(Mn)P_n:\mathcal F(W)\to\mathcal F(M_n) and block truncations SA:ℱ(W)→ℱ(W)S_A:\mathcal F(W)\to\mathcal F(W) are the continuous linear maps defined in (8). In=inˆ:ℱ(Mn)→ℱ(W),Pn=rnˆ:ℱ(W)→ℱ(Mn),SA=ρAˆ:ℱ(W)→ℱ(W).I_n=\widehat{i_n}:\mathcal F(M_n)\to\mathcal F(W),\qquad P_n=\widehat{r_n}:\mathcal F(W)\to\mathcal F(M_n),\qquad S_A=\widehat{\rho_A}:\mathcal F(W)\to\mathcal F(W).(8)

The Claim

Theorem 1. For the metric wedge WW, each block inclusion satisfies (9), and each block retraction satisfies (10). dW(in(x),in(y))=dn(x,y)(x,y∈Mn).d_W(i_n(x),i_n(y))=d_n(x,y)\quad(x,y\in M_n).(9)✦ dn(rn(u),rn(v))≤dW(u,v)(u,v∈W).d_n(r_n(u),r_n(v))\leq d_W(u,v)\quad(u,v\in W).(10)✦ For every pointed 11-Lipschitz map f:M→Nf:M\to N, its linearization satisfies (11) and (12). fˆ(δx)=δf(x)(x∈M).\widehat f(\delta_x)=\delta_{f(x)}\quad(x\in M).(11)✦ ∥fˆ∥≤1.\|\widehat f\|\leq1.(12)✦ The block linearizations satisfy (13) for every n∈ℕn\in\mathbb N. PnIn=id⁡ℱ(Mn).P_n I_n=\operatorname{id}_{\mathcal F(M_n)}.(13)✦ Let EE and YY be real normed spaces, and let I:E→YI:E\to Y and P:Y→EP:Y\to E be continuous linear contractions with PI=id⁡EPI=\operatorname{id}_E. If YY has λ\lambda-BAP, then (14) holds. E has λ-BAP.E\text{ has }\lambda\text{-BAP}.(14)✦ Let C⊆ℱ(W)C\subseteq\mathcal F(W) be compact and let ε>0\varepsilon>0. There is a finite A⊆ℕA\subseteq\mathbb N such that (15) holds for every finite G⊆ℕG\subseteq\mathbb N containing AA. ∥SGu−u∥<ε(u∈C).\|S_Gu-u\|<\varepsilon\quad(u\in C).(15)✦ Suppose distinct points of each MnM_n are at distance at least one, and there are numbers Bn≥0B_n\geq0 such that dn(x,on)≤Bnd_n(x,o_n)\leq B_n for every n∈ℕn\in\mathbb N and x∈Mnx\in M_n. Then (16) holds. ℱ(W) has AP.\mathcal F(W)\text{ has AP}.(16)✦

In the paper, the result reads: identify the block basepoints, keep each block metric and set dW((n,x),(m,y))=dn(x,on)+dm(y,om)d_W((n,x),(m,y))=d_n(x,o_n)+d_m(y,o_m) across distinct blocks; the pointed 11-Lipschitz inclusions and retractions induce contractions InI_n and PnP_n with PnIn=idP_nI_n=\operatorname{id}; retaining finitely many blocks gives contractive projections converging uniformly on compact subsets of ℱ(W)\mathcal F(W), finite-block approximations give AP for separated bounded blocks, and compression transfers every BAP bound to each block.

What Is Checked

The metric assertions are stated over countable families of pointed metric spaces. The AP assertion assumes separation by one and a finite radius bound for each component. Linearization concerns pointed 11-Lipschitz maps, and transfer of a BAP bound concerns contractive retracts of real normed spaces.

Evaluation at xx is a bounded functional with ∥δx∥Lip⁡0(M)*≤d(x,o)\|\delta_x\|_{\operatorname{Lip}_0(M)^*}\leq d(x,o). The adjoint of Lipschitz pullback sends the closed evaluation span in Lip⁡0(M)*\operatorname{Lip}_0(M)^* into the closed evaluation span in Lip⁡0(N)*\operatorname{Lip}_0(N)^*.✦