OctoInformBeta

OpenAI preprint, September 2026 · §6. From separating filtrations to finite chains

Lemma 6.1: Local constancy of the refined triangulation and equivariance of its uniform component weights

Component labels remain in finite graph windows.

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

The labels attached to a refined simplex are complex vectors on the vertices of a graph. They come from the component of a filtration stratum containing the simplex’s barycenter. The parameter of the filtration varies in a topological space XX.

Fix the graph geometry of a simplicial complex KK on a vertex set YY. The graph Γ\Gamma on YY is connected and has finite neighbor sets. Its graph metric dΓd_\Gamma is shortest-path distance, regarded as a real number.✦ Let a group HH act on YY and XX. Write V=ℝ(Y)V=\mathbb R^{(Y)} for the finitely supported real vectors, with the Euclidean norm induced by the coordinate embedding into ℓ2(Y)\ell^2(Y). For h∈Hh\in H, the vertex action h*h_* on VV, and on ℂ(Y)\mathbb C^{(Y)}, reindexes coordinates by y↦hyy\mapsto hy.

For y∈Yy\in Y, let eye_y be the real unit coordinate vector and δy\delta_y its complex counterpart. The standard simplex vertices and their convex hull are given by (1); the geometric realization of KK in VV is denoted by Kˆ\widehat K, with underlying space |K||K|. e(s)={ey:y∈s},κs=conv⁡e(s)(s⊆Y finite).e(s)=\{e_y:y\in s\},\qquad \kappa_s=\operatorname{conv} e(s) \quad(s\subseteq Y\text{ finite}).(1)

A finitely supported probability measure pp on YY is a nonnegative real coefficient function with finite support and total coefficient one.✦ For finitely supported probability measures p,qp,q on YY, a function f:Y→ℝf:Y\to\mathbb R, and a positive integer nn, the expectation 𝔼p[f]\mathbb E_p[f], expectation difference Δp,q(f)\Delta_{p,q}(f), dual transportation distance W1W_1, and normalized distance ρn\rho_n are defined by (2). 𝔼p[f]=∑y∈Yp(y)f(y),Δp,q(f)=𝔼p[f]−𝔼q[f],W1(p,q)=supLip⁡dΓ(f)≤1Δp,q(f),ρn(p,q)=W1(p,q)n.\begin{aligned} \mathbb E_p[f]&=\sum_{y\in Y}p(y)f(y),& \Delta_{p,q}(f)&=\mathbb E_p[f]-\mathbb E_q[f],\\ W_1(p,q)&=\sup_{\operatorname{Lip}_{d_\Gamma}(f)\leq1} \Delta_{p,q}(f),& \rho_n(p,q)&=\frac{W_1(p,q)}{\sqrt n}. \end{aligned}(2) The barycentric probability measure pvp_v of a realization point v∈|K|v\in|K| has coefficient function vv.✦ For vectors v,w∈Vv,w\in V, the realization distance ρ‾n\bar\rho_n between them is ρn(pv,pw)\rho_n(p_v,p_w) when both belong to |K||K|, and is zero otherwise.✦

Subdivisions and Filtrations

For a geometric simplicial complex BB in VV and a subset PP of VV, the vertex restriction B|PB|_P is the subcomplex of BB consisting of the faces whose vertices all lie in PP.✦ The geometric restriction B⌈PB\lceil P of a geometric simplicial complex BB in VV to a subset PP of VV consists of the faces of BB whose convex hulls lie in PP.✦ A geometric simplicial complex BB in VV respects a subset PP of VV if, for each face FF of BB, the intersection conv⁡(F)∩P\operatorname{conv}(F)\cap P is the convex hull of the vertices of FF that belong to PP.✦

A simplicial subdivision LL of a geometric complex BB is a geometric simplicial complex with the same underlying space as BB, each of whose faces is carried by a face of BB. It respects the convex hull of each face of BB, and its vertex restriction to each such hull has finitely many faces. A refinement LL of a geometric complex BB is the complex of a simplicial subdivision of BB.✦

A locally constant subdivision family TT assigns a subdivision T(x)T(x) of Kˆ\widehat K to each x∈Xx\in X, with the face set of T(x)|κsT(x)|_{\kappa_s} locally constant for each original face s∈Ks\in K.✦ A marked subset ZZ of a geometric complex is a union of convex hulls of its faces.✦ For a locally constant subdivision family TT of Kˆ\widehat K over XX, a marked subdivision family ZZ is a family of marked subsets Z(x)Z(x) of T(x)T(x), indexed by x∈Xx\in X, for which the face set of (T(x)⌈Z(x))|κs(T(x)\lceil Z(x))|_{\kappa_s} is locally constant for every original face s∈Ks\in K.✦ An equivariant marked subdivision pair (T,Z)(T,Z) over XX is a pair consisting of a locally constant subdivision family TT of Kˆ\widehat K and a marked subdivision family ZZ for TT, for which h*h_* sends faces at xx to faces at hxhx and marked points at xx to marked points at hxhx, for every h∈Hh\in H.✦

Fix n>0n>0 and R≥0R\geq0. A separating filtration QQ is a sequence of equivariant marked subdivision pairs (Ti,Zi)(T_i,Z_i) with the top and nesting conditions in (3), and with Ti(x)T_i(x) refining Ti+1(x)T_{i+1}(x) for i<ni<n. Tn(x)=Kˆ,Zn(x)=|K|,Zi(x)⊆Zi+1(x)(0≤i<n).T_n(x)=\widehat K,\qquad Z_n(x)=|K|,\qquad Z_i(x)\subseteq Z_{i+1}(x)\quad(0\leq i<n).(3) Its dimension drop condition says that if FF is a face of Ti+1(x)T_{i+1}(x) contained geometrically in Zi+1(x)Z_{i+1}(x), and EE is a face of Ti(x)T_i(x) whose convex hull lies in conv⁡(F)∩Zi(x)\operatorname{conv}(F)\cap Z_i(x), then |E|<|F||E|<|F|.✦ The difference Zi+1(x)\Zi(x)Z_{i+1}(x)\setminus Z_i(x) has component diameter bound 2R2R for ρ‾n\bar\rho_n: every preconnected subset of that difference has ρ‾n(v,w)≤2R\bar\rho_n(v,w)\leq2R for all its points v,wv,w.✦

Components and Weights

At a fixed xx, the complexes Ti(x)T_i(x) and marked sets Zi(x)Z_i(x) form the pointwise filtration QxQ_x.✦ A final triangulation face FF at xx is a face of T0(x)T_0(x).✦ The vertex mean bFb_F of a final triangulation face FF at xx is the vector in VV obtained by averaging the vertices of FF, with value zero when FF is empty.✦ The stratum Si(x)S_i(x), face level ℓx(F)\ell_x(F), and face stratum component Ux(F)U_x(F) are specified in (4); the component is taken in the Euclidean topology of VV. bF=1|F|∑v∈Fv,Si(x)=Zi(x)\⋃j<iZj(x),ℓx(F)=min⁡{i≤n:bF∈Zi(x)},Ux(F)=the connected component of Sℓx(F)(x) containing bF.\begin{aligned} b_F&=\frac{1}{|F|}\sum_{v\in F}v,& S_i(x)&=Z_i(x)\setminus\bigcup_{j<i}Z_j(x),\\ \ell_x(F)&=\min\{i\leq n:b_F\in Z_i(x)\},& U_x(F)&=\text{the connected component of }S_{\ell_x(F)}(x) \text{ containing }b_F. \end{aligned}(4)

The vertex support Vert⁡(U)\operatorname{Vert}(U) of a subset U⊆VU\subseteq V is the subset of YY consisting of vertices occurring in the support of at least one vector in UU.✦ For a final face, write Jx(F)=Vert⁡(Ux(F))J_x(F)=\operatorname{Vert}(U_x(F)) for its component vertices, as in (5). Vert⁡(U)={y∈Y:some v∈U has v(y)≠0},Jx(F)=Vert⁡(Ux(F)).\operatorname{Vert}(U)=\{y\in Y:\text{some }v\in U\text{ has }v(y)\ne0\}, \qquad J_x(F)=\operatorname{Vert}(U_x(F)).(5)

For a finite vertex set AA, its uniform vertex vector uAu_A is defined by (6). The component weight ux(F)u_x(F) is the uniform vector on the component vertices Jx(F)J_x(F).✦ uA=∑y∈A1|A|δy∈ℂ(Y),u⌀=0.u_A=\sum_{y\in A}\frac{1}{|A|}\delta_y\in\mathbb C^{(Y)}, \qquad u_{\varnothing}=0.(6)✦ Fix y0∈Yy_0\in Y. The extended face weight wxw_x is the function on finite subsets of VV defined by (7). wx(F)={ux(F),F is a face of T0(x),δy0,otherwise.w_x(F)= \begin{cases} u_x(F),&F\text{ is a face of }T_0(x),\\ \delta_{y_0},&\text{otherwise}. \end{cases}(7)✦

For a vertex a∈Ya\in Y, a nonnegative integer rr, and a finite vertex set s⊆Ys\subseteq Y, the finite graph neighborhoods Nr(a)N_r(a) and vertex label window Wn,R(s)W_{n,R}(s) are defined by (8), where ⌊t⌋+\lfloor t\rfloor_+ is the greatest nonnegative integer at most tt for t≥0t\geq0. N0(a)={a},Nr+1(a)={a}∪⋃b∼ΓaNr(b),m=⌊2Rn⌋++n,Wn,R(s)=⋃a∈sNm(a).\begin{aligned} N_0(a)&=\{a\},& N_{r+1}(a)&=\{a\}\cup\bigcup_{b\sim_\Gamma a}N_r(b),\\ m&=\lfloor2R\sqrt n\rfloor_++n,& W_{n,R}(s)&=\bigcup_{a\in s}N_m(a). \end{aligned}(8) For an original face ss, let ℱs(x)\mathcal F_s(x) be the face set of T0(x)|κsT_0(x)|_{\kappa_s}.

The Claim

Theorem 1. Let QQ be a separating filtration in the graph geometry above, with n>0n>0 and R≥0R\geq0. If FF is a final face at xx and its vertex mean belongs to κs\kappa_s for a finite set s⊆Ys\subseteq Y, the support bound is (9). Jx(F)⊆Wn,R(s).J_x(F)\subseteq W_{n,R}(s).(9)✦ For every final face FF at xx, (10) holds. Jx(F) is finite.J_x(F)\text{ is finite}.(10)✦ Fix an original face s∈Ks\in K and x∈Xx\in X. There is a neighborhood NcN_c of xx on which (11) holds for every y∈Ncy\in N_c, with the component equality applying to every common final face FF at xx and yy whose convex hull lies in κs\kappa_s. ℱs(y)=ℱs(x),Uy(F)=Ux(F).\mathcal F_s(y)=\mathcal F_s(x),\qquad U_y(F)=U_x(F).(11)✦ There is a neighborhood NwN_w of xx on which (12) holds for every y∈Nwy\in N_w, with FF ranging over ℱs(x)\mathcal F_s(x) in the weight equality. ℱs(y)=ℱs(x),wy(F)=wx(F).\mathcal F_s(y)=\mathcal F_s(x),\qquad w_y(F)=w_x(F).(12)✦ For h∈Hh\in H, x∈Xx\in X, and a final face FF at xx, the equivariance identity is (13). whx(h*F)=h*wx(F).w_{hx}(h_*F)=h_*w_x(F).(13)✦

In the paper, the result reads: for every original simplex κ\kappa there is a finite Wκ⊆YW_\kappa\subseteq Y, independent of the filtration parameter ε\varepsilon and of xx, such that Jx(F)⊆WκJ_x(F)\subseteq W_\kappa whenever bF∈κb_F\in\kappa, and the triangulation over κ\kappa together with all assigned vectors ux(F)u_x(F) is locally constant in xx, with these assignments equivariant.

What Is Checked

The assertions are stated over the standard coordinate realization of KK with its normalized graph transportation distance, and an equivariant separating filtration of positive length with nonnegative radius. Every realization point admits a barycentric probability measure whose support is a face of KK.✦ The barycenter of each final face meets a marked level i≤ni\leq n.✦ The coordinate embedding V→ℓ2(Y)V\longrightarrow\ell^2(Y) is injective.✦