Comparing Data Structures via Persistent Homology

This direction is merged into another direction: persistent homology. Its last version stays readable and citable here.

Version 1 of 137 papers38 conceptsfrozen 2026-10-04

Comparing complex datasets—whether they are brain networks, molecular structures, or neural representations—is a fundamental challenge in data science. Traditional metrics often rely on point-wise comparisons or rigid statistical assumptions that fail to capture the underlying "shape" of the data. A growing body of research suggests that Topological Data Analysis (TDA), specifically the use of distances between persistence diagrams, provides a robust, coordinate-free framework for quantifying structural similarity.

The Intuition: Measuring the Cost of Structural Transformation

At the heart of this approach is the persistence diagram, a 2D summary that tracks the "birth" and "death" of topological features (like loops, voids, or connected components) as one looks at the data across different scales. To compare two datasets, we compare their diagrams.

The most intuitive way to think about this is through Optimal Transport. Imagine one persistence diagram as a set of holes in the ground and the other as piles of dirt. The distance between them is the minimum "work" required to move the dirt into the holes.

  • Wasserstein Distance: This metric calculates the optimal matching between points in two diagrams. It is used by Moo K. Chung to cluster dynamic brain networks and by Roberto Fontana to identify non-isomorphic experimental designs.
  • Bottleneck Distance: A more conservative relative of the Wasserstein distance, it focuses only on the single largest discrepancy between features. Stefan Groha and Bastian Rieck use this to find structural outliers in cancer genomic data.

Bridging Topology and Deep Learning

One of the most powerful applications of these distances is in making topology differentiable. By treating the distance between persistence diagrams as a loss function, researchers can force neural networks to respect global structural constraints.

In Topology-Preserving Deep Image Segmentation, Chao Chen and colleagues use a topological loss to ensure that segmented road networks or neurons don't have unrealistic breaks. Similarly, Fan Wang introduced TopoGAN, which uses optimal transport on persistence diagrams to ensure synthetic images share the same "loopy-ness" as real-world data.

Beyond Simple Loops: Sensitivity and Hardness

While $L^\infty$-style metrics like the bottleneck distance are common, they can be insensitive to many small changes. To address this, Håvard Bakke Bjerkevik and Magnus Bakke Botnan have explored the $p$-presentation distance. This $L^p$-style metric aggregates all structural differences rather than just the maximum one, providing a more nuanced view of total variation. However, their work in Computing p-presentation distances is hard demonstrates that this increased sensitivity comes with a high computational cost, proving the problem to be NP-hard.

Applications Across Domains

The versatility of topological distances allows them to serve as a "universal language" for structural comparison:

  • Molecular Chemistry: Kelin Xia uses cohomology-based distances to quantify the similarity of molecular "voids" in materials like perovskites.
  • Model Evaluation: Serguei Barannikov and Ilya Trofimov developed Representation Topology Divergence (RTD) to compare how different neural networks "see" the same data manifold, finding that topological divergence correlates strongly with model disagreement.
  • Network Science: Weiyu Huang proved that persistent homology can provide efficient lower bounds for the Gromov-Hausdorff distance, making it possible to compare massive hypergraphs that would otherwise be computationally inaccessible.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_comparing_data_structures_via_persistent_97e4f37b,
  title = {Comparing Data Structures via Persistent Homology},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/comparing-data-structures-via-persistent-homology/rdr_97e4f37bf06c4bd0905463fc16f9bbf2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_comparing_data_structures_via_persistent_97e4f37b_v1,
  title = {Comparing Data Structures via Persistent Homology},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/comparing-data-structures-via-persistent-homology/rdr_97e4f37bf06c4bd0905463fc16f9bbf2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}