Algebraic Wasserstein distances and stable homological invariants of data (Q7020722)

This work further develops the theory of algebraic Wasserstein distances between persistence modules, originally introduced by \textit{P. Skraba} and \textit{K. Turner} [``Wasserstein stability for persistence diagrams'', Preprint, \url{arXiv:2006.16824v5}]. Persistence modules, a central object in topological data analysis, capture the geometry of data at different scales. Distances between persistence modules have been used in clustering problems and are almost always stable with respect to noise in the data. The contribution of this paper is the introduction of Wasserstein stable ranks -- new invariants of persistence modules that are 1-Lipschitz-stable with respect to the Wasserstein distance used to define them. These invariants depend on learnable parameters that can be fine-tuned in downstream machine-learning tasks. Moreover, Wasserstein stable ranks can be computed efficiently from a barcode decomposition of a persistence module.\N\NMore specifically, the authors develop the theory of bar-to-bar monomorphisms and epimorphisms between persistence modules. In applications, every persistence module is a direct sum of indecomposable interval modules (bars). Bar-to-bar morphisms induce a new type of matching between bars of persistence modules, allowing the authors to bridge between algebraic and the more computable combinatorial distances between persistence modules. This matching is fundamentally different from previous constructions, such as that of \textit{U. Bauer} and \textit{M. Lesnick} [J. Comput. Geom. 6, No. 2, 162--191 (2015; Zbl 1405.68398)]. Decomposing every morphism into its kernel and cokernel lets the authors approximate the \(p\)-norm of morphisms by the \(p\)-norms of simpler bar-to-bar morphisms.\N\NFor some time, longer bars in a persistence module were considered the relevant geometric features of data, while short bars were viewed as noise. Recent work shows that shorter bars can carry important geometric information. Determining the scale at which to interpret persistence bars therefore depends on the data and the hypothesis at hand. To automate this choice, the authors show how different parameter values in the definition of Wasserstein stable ranks can (i) weight the importance of long versus short bars and (ii) select more discriminative invariants for classification tasks.