Persistent Cost of Lipschitz Maps

Francisco J. Gozzi, Manuela A. Cerdeiro, Pablo E. Riera · 2025

Read Persistent Cost of Lipschitz Maps by Francisco J. Gozzi, Manuela A. Cerdeiro, Pablo E. Riera (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-07674.txt.

Introduction

Francisco Gozzi, Manuela Cerdeiro, and Pablo Riera price how much a gentle map distorts shape data. A Lipschitz map that never stretches distances induces a homomorphism between persistence modules, the summaries topological data analysis uses to describe holes across scales. Information lost or created by the map needs a measure.

Persistent cost counts the persistence carried by the kernel and cokernel of that homomorphism. The cost bounds the interleaving distance between persistence summaries of the two spaces. A purely metric upper bound and a stability proof under a Gromov-Hausdorff type distance complete the picture. Practitioners gain guarantees when compressing shapes.

Catalog record

CREATORFrancisco J. Gozzi, Manuela A. Cerdeiro, Pablo E. Riera
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-07674

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Persistent Cost of Lipschitz Maps Francisco J. Gozzi, Manuela A. Cerdeiro, Pablo E. Riera Abstract A 1-Lipschitz map between compact metric spaces f : X → Y f:X\to Y induces a homomorphism on the level of persistent d-homology of the associated Vietoris-Rips simplicial filtrations. This homomorphism naturally bounds the Bottleneck distance between said modules in terms of the maximum size of its kernel and cokernel persistence modules, a stable invariant which we refer to as the persistent cost associated to f f . We provide metric upper bounds on the persistent cost of a 1-Lipschitz map and revisit its stability. Introduction Persistent homology is a fundamental tool in the emerging field of topological data analysis (TDA) that extracts a vast amount of metric information by encoding homological features across scales. The richness of this output coupled with its explainability has yielded significant insights across a spectrum of disciplines, including biology and time series analysis, among others, c.f. [ 13 , 15 , 8 ] , and been integrated into learning pipelines, c.f., [ 9 , 10 ] . We outline the standard pipeline for topological data analysis. Given a metric space X , the Vietoris–Rips filtration (\VR_{s}(X))_{s\geq 0} is constructed by including simplices whose vertex set has diameter at at most s\geq 0 . As the parameter increases, the complex evolves from a discrete set to the full power set of X , encoding geometric information at multiple scales. Applying the homology functor in a fixed dimension d yields a persistence module: a sequence of vector spaces connected by linear maps that track the birth, persistence and death of homological features. While the TDA community has primarily focused on analyzing individual point clouds, the functorial nature of these constructions naturally leads us to consider mappings between datasets. Indeed, we advocate for a treatment of maps between spaces that parallels that of the spaces themselves, enabling us to address the geometry of functions from a persistent homology perspective. In this context we shall consider the natural case of non-expanding, i.e., 1-Lipschitz, maps f:X\to Y between Vietoris-Rips filtrations of finite metric spaces which yields a homomorphism of point-wise finite dimensional persistence modules, f_{*}:\operatorname{H}_{d}(\VR_{{}_{\bullet}}(X))\to\operatorname{H}_{d}(\VR_{{}_{\bullet}}(Y)). A homomorphism of persistence modules defines kernel, cokernel, and image modules via point-wise linear algebra, along with naturally induced maps. The “size” of the modules \Ker(F) and \Coker(F) , measured by their interleaving distance to the trivial module \Delta , serves as a measure of how far the map F is from a persistent isomorphism. This leads us to define the persistent cost of a 1 -Lipschitz function as the maximum interleaving distance of kernel and cokernel persistence modules to the diagonal, \begin{array}[]{rcl}C(f)&=&\max\big\{d_{I}(\Ker(f),\Delta)\penalty\ ,\penalty\ d_{I}(\Coker(f),\Delta)\big\}\end{array} (1) This persistent cost bounds the interleaving distance between the modules associated to X and Y , by the Induced Matching Theorem of U. Bauer and M. Lesnick in [ 1 ] (see also [ 2 ] ). Building on this result, we contribute a metric upper bound for the persistent cost of non-expanding map f:X\to Y , showing that C(f) is bounded above by its distortion and the Hausdorff distance between its image and the codomain, f(X)\subseteq Y . Specifically, we establish the second inequality in the following sequence. d_{I}\left(\operatorname{H}_{d}(\VR_{{}_{\bullet}}(X)),\operatorname{H}_{d}(\VR_{{}_{\bullet}}(Y))\right)\penalty\ \leq\penalty\ C(f)\penalty\ \leq\penalty\ \dist(f)+2\cdot d_{H}(f(X),Y) (2) From a purely algebraic perspective, as we range over all possible persistence module homomorphisms F_{\bullet}:V_{\bullet}\to W_{\bullet} , the infimum of the associated persistent cost attains the interleaving distance between the given modules. However, our interest lies in geometrically defined persistence homomorphisms, i.e. those induced by 1 -Lipschitz maps between the underlying spaces via the homology of their Vietoris-Rips filtrations. In this geometric context, we provide an example of two finite metric spaces that yield identical persistence diagrams while any 1-Lipschitz map is bounded away from zero in its associated persistent cost. A prior concern is the stability of kernel and cokernel persistence modules which was addressed in [ 5 ] . We provide a self-contained exposition of this fact that was claimed without proof. In doing so, we introduce a Gromov-Hausdorff type distance between maps f:X\to Y , f^{\prime}:X^{\prime}\to Y^{\prime} , of compact metric spaces that may be of independent interest. Our results are developed within the framework of point-wise finite dimensional persistence modules and finite metric spaces, and naturally extended to q-tame modules that arrise from compact metric spaces. The paper is organized as follows. A first section provides a self-contained exposition of the algebraic notions regarding persistence theory. We then move on to the geometric of spaces, maps and distances. The last two subsection are focused on our contributed results on metric upper bounds for the cost, Theorem 2.11 and stability, Theorem 2.16 . 1 Algebraic Persistence The main references for this section are the comprehensive book by S. Oudot in [ 14 ] and the papers on the Induced Matching Theorem [ 1 , 2 ] by U. Bauer and M. Lesnick. 1.1 Persistence Modules A persistence module is a linear representation of \mathbb{R} , or a given subset T\subseteq\mathbb{R} , regarded as a poset. In other words, it is a functor V:(\mathbb{R},\leq)\to\mathrm{Vect}_{\mathbb{F}}, (3) where \mathrm{Vect}_{\mathbb{F}} denotes the category of vector spaces over a field \mathbb{F} with linear transformations as homomorphisms. This, a persistence module is more than just a one-parameter family of vector spaces; as a representation of a poset, it carries “inner persistence” or “transition” maps \varphi_{s<t}:V_{s}\xrightarrow{\penalty\ \penalty\ }V_{t}, for each pair s<t\in\mathbb{R} , which satisfy \varphi\circ\varphi=\varphi, the last equation being read as \varphi_{s_{2}<s_{3}}\circ\varphi_{s_{1}<s_{2}}=\varphi_{s_{1}<s_{3}} . As a first and computationally relevant approach to the subject, the reader may consider the case of finite-dimensional spaces indexed over a finite set 0=\epsilon_{0}<\epsilon_{2}<\dots<\epsilon_{r} , giving rise to V_{\epsilon_{0}}\to V_{\epsilon_{1}}\to\cdots\to V_{\epsilon_{r}}. (4) The key idea of persistence is that the dimension of V_{s} may be preserved (via an isomorphism) in V_{t} for t>s . However, indexing over the real numbers has several advantages. It is less cumbersome, as it avoids the problem of refining different indexing sets. It is also more general, since any finitely indexed persistence module can be easily extended to a \mathbb{R} -indexed one. We will first develop the theory within the setting of point-wise finite dimensional (p.f.d.) modules indexed over \mathbb{R} , as in definition ( 3 ). A fundamental result states that such persistence modules admit a unique decomposition into interval modules [ 7 , 14 ] . Specifically, V\cong\bigoplus_{I\in\mathcal{I}}\mathbb{F}^{I}, (5) where \mathcal{I} is a collection of intervals I\subseteq\mathbb{R} each defining an “interval module” as follows: (\mathbb{F}^{I})_{(t)}=\begin{cases}\mathbb{F},&t\in I,\\ 0,&\text{otherwise},\end{cases} with inner persistence maps: \varphi_{s<t}=\begin{cases}Id,&s,t\in I,\\ 0,&\text{otherwise}.\end{cases} The field \mathbb{F} will play no role in the discussion, usually being \mathbb{Z}_{p} for p prime, \mathbb{Q} or \mathbb{R} in practice. The multiset of intervals \mathcal{I} is called the persistence diagram (or barcode ). Each interval \mathcal{I} is represented buy its endpoints (b_{i},d_{i})\in\bar{\mathbb{R}}^{2} . A barcode represents each interval as a horizontal line from b to d , as shown in Figure 1 . b d 1 2 3 4 5 t Figure 1: A persistence diagram and barcode representation of the same decomposition. In order to define a distance between persistence diagrams D_{1},D_{2} , we may begin by setting a partial matching between their persistent features. If a pair (b_{1},d_{1})\in D_{1} and (b_{2},d_{2})\in D_{2}

Read the complete text