Persistent homology for metric measure spaces, and robust statistics for hypothesis testing and confidence intervals

Andrew J. Blumberg, Itamar Gal, Michael A. Mandell, Matthew Pancia

This paper, titled "Robust Statistics, Hypothesis Testing, and Confidence Intervals for Persistent Homology on Metric Measure Spaces," represents a significant step in bridging the gap between algebraic topology and classical statistics. It provides a rigorous framework for making Topological Data Analysis (TDA) resilient to noise and capable of providing statistical guarantees.

TL;DR

The authors transform persistent homology from a "fragile" qualitative tool into a robust statistical framework by analyzing the distribution of barcodes generated from subsamples. This approach allows researchers to perform hypothesis testing and calculate confidence intervals for topological features, effectively distinguishing true underlying structures from sampling noise.

What Problem Does This Paper Address?

Persistent Homology (PH) is a powerful tool for finding "holes" and "handles" in high-dimensional data. However, it has two major weaknesses:

  1. Sensitivity to Noise: A single outlier or a tiny "spurious" handle can drastically change a barcode, making PH "fragile."
  2. Lack of Statistical Context: Standard PH treats all topological features equally, regardless of whether they represent a high-probability region of the data or a low-probability fluke.

The authors address these issues by moving from "metric spaces" (just points and distances) to "metric measure spaces" (points, distances, and a probability measure). This allows them to ask not just "is there a hole?" but "how likely is it that this hole is a significant feature of the underlying distribution?"

Methods

The core innovation is the definition of the empirical measure $\Phi^n_k$. Instead of computing one massive barcode for an entire dataset, the authors:

  • Take many subsamples of a fixed size $n$.
  • Compute the $k$-th persistent homology barcode for each subsample.
  • Treat the resulting collection of barcodes as a probability distribution.

They prove a Stability Theorem (1.2), showing that if two underlying spaces are close (in the Gromov-Prohorov sense), their resulting barcode distributions will also be close.

To make this practical for data science, they introduce:

  • Homological Distance ($HD$): A way to measure the distance between two barcode distributions.
  • Median Homological Distance ($MHD$): A "robust" version that uses the median to ignore the influence of a small proportion of corrupted or noisy samples.
  • Inference Tools: They adapt standard non-parametric tests, like the Kolmogorov-Smirnov and $\chi^2$ tests, to compare these barcode distributions.

Datasets

The authors evaluate their framework using both synthetic and real-world data:

  • Synthetic Annulus & "Friendly Circles": Used to demonstrate how the method can distinguish between a simple circle and a circle with a "diameter linkage" (a figure-8) even when noise is added.
  • Spheres and Tori: Used to test the ability to detect higher-dimensional features (Betti 1 and Betti 2) under Gaussian and uniform noise.
  • Natural Images Dataset: A well-known TDA benchmark consisting of $3 \times 3$ high-contrast patches from outdoor scenes. The authors use their confidence interval method to validate previous claims that this data lives on a manifold resembling a Klein bottle.

Key Insights and Ideas

  • Subsample Size as Scale: The size of the subsample ($n$) acts as a scale parameter. Smaller $n$ captures local features, while larger $n$ reveals more complex global topology.
  • Robustness via Medians: By using the median of the barcode distances rather than the mean, the authors create a "breakdown point." This means a certain percentage of the data can be complete garbage (noise) without ruining the topological estimate.
  • Topology is Borel Measurable: A key theoretical "aha" moment is proving that the map from a metric space to its barcode is continuous and Borel measurable, which is the "legal" requirement for doing probability theory on barcodes.

Key Results

  • Theoretical Stability: They provide explicit bounds showing that the distribution of barcodes is a "robust invariant." Specifically, Theorem 5.6 proves that $\Phi^n_k$ is uniformly robust.
  • Noise Resilience: In experiments with the "Friendly Circles," the method correctly identified the two-circle structure even when 90% of the points were replaced with uniform noise—a scenario where standard PH would fail completely.
  • Validation of Natural Image Topology: The authors successfully calculated 95% and 99% confidence intervals for the natural images dataset, providing the first formal statistical validation for the "5-cycle" (Klein bottle) hypothesis in that specific data context.

Limitations and Future Work

  • Computational Cost: Computing thousands of barcodes for subsamples is computationally expensive, though the authors note it is highly parallelizable.
  • Choosing $n$: Selecting the optimal subsample size $n$ remains a heuristic process. If $n$ is too small, you miss the topology; if $n$ is too large, the "fragility" of PH returns.
  • Future Directions: The authors suggest further work on "bootstrap" resampling methods to provide even tighter confidence intervals and exploring different metrics (like Wasserstein distance) on the barcode space.

Summary updated

Abstract

We study distributions of persistent homology barcodes associated to taking subsamples of a fixed size from metric measure spaces. We show that such distributions provide robust invariants of metric measure spaces, and illustrate their use in hypothesis testing and providing confidence intervals for topological data analysis.
BibTeX
@inproceedings{blumberg2014persistent,
  author = {Andrew J. Blumberg and Itamar Gal and Michael A. Mandell and Matthew Pancia},
  title = {{Persistent homology for metric measure spaces, and robust statistics for hypothesis testing and confidence intervals}},
  booktitle = {Found. Comput. Math},
  year = {2014},
  url = {https://openreview.net/forum?id=uZy7bgnnl6},
}
Figure