Persistent Intersection Homology for the Analysis of Discrete Data.

Bastian Rieck, Markus Banagl, Filip Sadlo, Heike LeittearXiv 2019 · DOI · OpenReview

Abstract

Topological data analysis is becoming increasingly relevant to support the analysis of unstructured data sets. A common assumption in data analysis is that the data set is a sample—not necessarily a uniform one—of some high-dimensional manifold. In such cases, persistent homology can be successfully employed to extract features, remove noise, and compare data sets. The underlying problems in some application domains, however, turn out to represent multiple manifolds with different dimensions. Algebraic topology typically analyzes such problems using intersection homology, an extension of homology that is capable of handling configurations with singularities. In this paper, we describe how the persistent variant of intersection homology can be used to assist data analysis in visualization. We point out potential pitfalls in approximating data sets with singularities and give strategies for resolving them.

Introduction

Most Topological Data Analysis (TDA) relies on the manifold hypothesis: the assumption that high-dimensional data points are sampled from a single, smooth underlying manifold. Under this assumption, standard persistent homology is effective for identifying "holes" or cycles in data across multiple scales.

However, real-world data often contains singularities that violate this hypothesis. Examples include a wedge sum of two circles touching at a single point or a "pinched torus" where a ring has been collapsed to a point. In these instances, the space is not a manifold at the singular point. Standard TDA tools, such as the Vietoris-Rips complex, tend to "thicken" the data, effectively smoothing over these singularities and making them invisible to analysis. This paper introduces Persistent Intersection Homology (PIH) to explicitly detect and account for these non-manifold structures.

Methodology

The authors implement PIH by incorporating two key concepts from algebraic topology into a data-driven framework:

  1. Stratification: The space $X$ is partitioned into a sequence of subspaces (strata) such that each component is a manifold. For instance, a pinched torus is stratified into a 0-dimensional singular point (the pinch) and a 2-dimensional surface.
  2. Perversity Functions: These functions act as a "rulebook" to determine how much a simplicial chain can deviate from the singular strata, allowing researchers to tune the sensitivity to singularities.
The pinched torus (a) and its stratification (b), illustrating how non-manifold spaces are decomposed into manifold-like pieces.

Since raw point clouds lack predefined singularities, the authors propose three strategies for discrete stratification:

  • Dimensionality-based: Using local PCA to identify points where the intrinsic dimension changes (e.g., where a 1D "flare" meets a 2D "core").
  • Density-based: Using kernel density estimation to flag high-concentration or outlier points as potentially singular.
  • Curvature-based: Fitting high-dimensional spheres to local patches to detect areas of extreme curvature.

Experimental Evaluation

The method was evaluated on synthetic and benchmark datasets, including a wedge of spheres, a pinched torus, and the Isomap "Synthetic Faces" dataset.

A sample data set (a) and its corresponding persistence diagram (b), demonstrating the extraction of topological features from discrete points.

Key Results

  • Detecting Splits: In the wedge of spheres ($S^2 \vee S^2$), standard persistent homology identifies only one connected component ($\beta_0 = 1$). PIH, using density-based stratification, correctly identifies the singular point as a "split," yielding $\beta_0 = 2$ and revealing the underlying structure.
  • Non-Manifold Detection: For the pinched torus, PIH produces a persistence diagram significantly different from standard homology, providing a quantitative "red flag" that the manifold hypothesis has failed.
  • Computational Efficiency: By optimizing the "admissibility condition" check—the most computationally expensive part of the algorithm—the authors' implementation in the Aleph library runs efficiently on modern hardware.

Discussion and Future Work

A critical insight of this work is the Vietoris-Rips Pitfall: standard TDA tools often "heal" singularities by thickening points into balls, rendering intersection homology ineffective unless specific stratifications are enforced. PIH restores Poincaré duality for non-manifolds, providing more robust topological signatures.

The "Synthetic Faces" experiment confirmed that PIH yields results nearly identical to standard persistent homology when the data is truly a manifold, suggesting it is a "safe" extension that adds power without distorting smooth data.

Future work will address the sensitivity of results to stratification noise and extend the framework to handle higher-dimensional singular strata (e.g., singular curves or surfaces) beyond the isolated points focused on in this study.

Summary updated

Abstract

Topological data analysis is becoming increasingly relevant to support the analysis of unstructured data sets. A common assumption in data analysis is that the data set is a sample---not necessarily a uniform one---of some high-dimensional manifold. In such cases, persistent homology can be successfully employed to extract features, remove noise, and compare data sets. The underlying problems in some application domains, however, turn out to represent multiple manifolds with different dimensions. Algebraic topology typically analyzes such problems using intersection homology, an extension of homology that is capable of handling configurations with singularities. In this paper, we describe how the persistent variant of intersection homology can be used to assist data analysis in visualization. We point out potential pitfalls in approximating data sets with singularities and give strategies for resolving them.
BibTeX
@misc{rieck2019persistent,
  author = {Bastian Rieck and Markus Banagl and Filip Sadlo and Heike Leitte},
  title = {{Persistent Intersection Homology for the Analysis of Discrete Data.}},
  year = {2019},
  eprint = {1907.13485},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.13485},
}
Figure