On the Effectiveness of Persistent Homology

Renata Turkes, Guido Montufar, Nina Otter

TL;DR

While Persistent Homology (PH) is a cornerstone of Topological Data Analysis, the reasons behind its broad success have remained somewhat mysterious. This paper demonstrates that PH is not just a tool for counting holes; it is remarkably effective at detecting geometric features like curvature and convexity, often outperforming deep learning models like PointNet—especially when data is noisy or scarce.

What Problem Does This Paper Address?

In the world of data science, Persistent Homology is often treated as a "black box" that provides useful features for machine learning. We know it works for everything from analyzing brain artery trees to classifying leaf morphology, but we lack a clear understanding of what it is actually seeing. Is it just capturing global topology, or is it sensing fine-grained geometry?

The authors aim to bridge this gap by investigating PH’s effectiveness in three fundamental shape analysis tasks:

  1. Topology: Detecting the number of holes in a shape.
  2. Local Geometry: Estimating the curvature of a surface.
  3. Global Geometry: Distinguishing between convex and concave shapes.

By comparing PH against standard machine learning baselines and specialized architectures like PointNet, the paper seeks to define the "boundaries of effectiveness" for topological methods.

Methods

The researchers utilize the standard PH pipeline—turning point clouds into simplicial complexes (like the Vietoris-Rips or Alpha complex) and tracking how topological features persist across different scales. However, they introduce several specific approaches for different tasks:

  • Tubular Filtrations: To tackle convexity, the authors introduce a novel "tubular filtration." They define a function based on the distance of points from a given line. If a shape is concave, this filtration will "see" multiple disconnected components that eventually merge.
  • Theoretical Proofs: They provide a new theoretical guarantee (Theorem 1) proving that a shape is convex if and only if every possible tubular filtration results in exactly one persistent component in the persistence diagram.
  • Signature Comparison: They test various ways to "vectorize" the topological data, including Persistence Images (PI), Persistence Landscapes (PL), and simple sorted lists of lifespans, feeding these into an SVM.

Datasets

The study uses a mix of synthetic benchmarks and real-world data:

  • Synthetic Holes: 2D and 3D point clouds sampled from shapes with 0, 1, 2, 4, or 9 holes.
  • Constant Curvature Disks: Point clouds sampled from surfaces with constant negative (hyperbolic), zero (Euclidean), and positive (spherical) curvature.
  • Convexity Benchmark: A balanced set of "regular" shapes (triangles, squares) and "random" shapes generated via convex hulls and alpha shapes.
  • FLAVIA Dataset: A real-world dataset of 1,907 plant leaf images used to test the PH pipeline’s ability to predict a continuous measure of convexity in nature.

Key Insights and Ideas

  • Short Bars are Signal, Not Noise: A common "rule of thumb" in TDA is that long persistence intervals represent the true signal (topology) while short intervals are noise. This paper flips that script: for tasks like curvature detection, the "short bars" actually contain the essential geometric information.
  • The Low-Data Hero: One of the most striking insights is PH's data efficiency. While deep learning models like PointNet require large amounts of training data to learn shape features, PH achieves high accuracy with very few samples because the "knowledge" of shape is already baked into the mathematical framework.
  • Out-of-Distribution Robustness: PH proved much more resilient than PointNet when faced with "out-of-distribution" data. For example, when trained on simple regular polygons but tested on complex random shapes, PH maintained its performance while neural networks struggled to generalize.
  • Invariance by Design: Because PH is built on relative distances, it is naturally invariant to rotation and translation—properties that standard neural networks often have to "learn" through expensive data augmentation.

Key Results

  • Hole Detection: PH achieved nearly 100% accuracy across all transformations (rotation, shear, stretch). In contrast, PointNet’s accuracy plummeted to ~20% under simple translations or rotations.
  • Curvature Estimation: PH (specifically using 0-dimensional lifespans) achieved a Mean Squared Error (MSE) of 0.06, vastly outperforming PointNet, which had an MSE of 578.28. This suggests PointNet struggles to capture the subtle metric properties required to distinguish curvature.
  • Convexity: On the FLAVIA leaf dataset, the PH pipeline effectively predicted convexity measures with a low MSE of 0.00065, demonstrating that the theoretical "tubular filtration" works on noisy, real-world biological data.

Limitations and Future Work

The authors acknowledge a few hurdles:

  • Computational Complexity: While PH is data-efficient, calculating filtrations and tuning persistence signatures (like PIs and PLs) can be computationally expensive compared to a forward pass of a trained neural network.
  • High-Dimensional Scaling: The tubular filtration method for convexity requires "sweeping" lines in many directions. As the dimension of the data increases, the number of lines needed to guarantee detection grows, potentially leading to a "curse of dimensionality."
  • Non-Constant Curvature: The current study focused on surfaces with constant curvature. Future work is needed to see how well PH handles objects with varying local geometry (e.g., a "bumpy" sphere).

The paper concludes that PH should be viewed not just as a niche tool for topology, but as a robust, general-purpose framework for geometric feature extraction in any shape-related machine learning task.

Summary updated

Abstract

Persistent homology (PH) is one of the most popular methods in Topological Data Analysis. Even though PH has been used in many different types of applications, the reasons behind its success remain elusive; in particular, it is not known for which classes of problems it is most effective, or to what extent it can detect geometric or topological features. The goal of this work is to identify some types of problems where PH performs well or even better than other methods in data analysis. We consider three fundamental shape analysis tasks: the detection of the number of holes, curvature and convexity from 2D and 3D point clouds sampled from shapes. Experiments demonstrate that PH is successful in these tasks, outperforming several baselines, including PointNet, an architecture inspired precisely by the properties of point clouds. In addition, we observe that PH remains effective for limited computational resources and limited training data, as well as out-of-distribution test data, including various data transformations and noise. For convexity detection, we provide a theoretical guarantee that PH is effective for this task in $\mathbb{R}^d$, and demonstrate the detection of a convexity measure on the FLAVIA dataset of plant leaf images. Due to the crucial role of shape classification in understanding mathematical and physical structures and objects, and in many applications, the findings of this work will provide some knowledge about the types of problems that are appropriate for PH, so that it can --- to borrow the words from Wigner 1960 --- ``remain valid in future research, and extend, to our pleasure", but to our lesser bafflement, to a variety of applications.
BibTeX
@inproceedings{turkes2022effectiveness,
  author = {Renata Turkes and Guido Montufar and Nina Otter},
  title = {{On the Effectiveness of Persistent Homology}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
}
Figure