Embeddings of persistence diagrams into Hilbert spaces (Q2209745): Difference between revisions

Persistence diagrams capture the evolving homology of a one-parameter family of topological spaces. One can define metrics, the so-called \(p\)-Wasserstein distances, on the space of persistence diagrams. Since persistence diagrams do not admit an inner product structure, in order to apply kernel methods, the standard approach is to map the data into Hilbert space in a controlled way. For these methods to be effective, the map (called a feature map) should change the original metric as little as possible. The paper under review confirms a result that computational topologists have observed in practice: that any feature map from persistence diagrams into a Hilbert space must distort the metric. More specifically, in the case \(1\le p\le \infty\), the authors combine results of [\textit{J. H. Wells} and \textit{L. R. Williams}, Embeddings and extensions in analysis. Springer-Verlag, Berlin (1975; Zbl 0324.46034)] and [\textit{K. Turner} and \textit{G. Spreemann}, Abel Symp. 15, 459--490 (2020; Zbl 1450.62141)] to show that there is no isometric embedding of the space of persistence diagrams in a \(p\)-Wasserstein metric into a Hilbert space. Moreover, in the case of \(p=\infty\) -- the so-called bottleneck distance, which is most often used in practice -- the authors show that the space of persistence diagrams does not even embed into a Hilbert space in a coarsely uniform way. The upshot of this result is that the distortion caused by a feature map is not uniformly controllable. They conclude that the space of persistence diagrams in the bottleneck distance has generalized roundness 0 and infinite asymptotic dimension. Finally, they show that any separable bounded metric space can be isometrically embedded into the space of persistence diagrams with the bottleneck distance.