Intrinsic Multiresolution Persistent Homology: Diffusion-Induced Resolutions and Entropy-Summary Representations
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
In topological data analysis, persistent homology (PH) characterizes topological features across filtration scales. Conventional PH, however, analyzes a single geometric representation across these scales and therefore does not organize features by intrinsic geometric resolution. This creates two coupled gaps: resolutions are fixed in advance rather than data-induced, and separate summaries do not jointly retain resolution and filtration-domain information. We introduce (Diffusion-Induced Geometry for Intrinsic-Resolution Topology), which constructs ordered diffusion-induced geometries and selects each sample's retained range from its diffusion evolution. Its multiresolution entropy summary function (\MESF) integrates the resulting persistence diagrams while retaining diffusion resolution, filtration-domain information, and persistence contribution. We establish conditional stability under discrete diffusion-operator perturbations. Experiments demonstrate progressive topological coarse-graining on synthetic point clouds; on three CATH benchmarks at the Class, Architecture, and Topology levels, + attains the highest mean accuracy among the evaluated persistence-based representations on all benchmarks.
- Primary area:
- learning on graphs and other geometries & topologies