Persistent Entropy Transform: An entropy-based descriptor for topological data analysis

Abstract

Persistent entropy provides a compact summary of persistence diagrams, but discards geometric information inherent to the data. This limitation creates a gap between scalar summaries, which are computationally efficient but geometrically coarse, and directional topological transforms, which are expressive but high-dimensional and computationally demanding. In this work, we introduce the Persistent Entropy Transform (PET), a novel directional topological descriptor which can be interpreted as an entropy-based compression of directional topological transforms. We establish basic theoretical properties of PET. In particular, we prove translation invariance, scale invariance under positive uniform scalings, and orthogonal equivariance. Empirically, we use synthetic shapes to assess directional sensitivity, consistency with rotational equivariance, robustness under controlled perturbations, and dependence on directional sampling density. In addition, we test the novel tool on two real time-series benchmark datasets, the TwoLead- ECG and the MIT-BIH, to provide a proof of concept showing that PET embeddings can be used succesfully as compact feature vectors on real time-series benchmarks. These experiments support and validate PET as a compact and computationally tractable descriptor.