Persistent Homology & Topological Data Analysis

Why It Matters for Modern Models

Persistent homology summarizes the shape of a point cloud, its clusters, loops and voids, at every scale at once rather than at a single chosen threshold (Carlsson, 2009).
Its output can be read directly: a diagram with three long-lived points in H₀ and one in H₁ says that the data form three clusters and contain one loop.
Topological terms can be added to training losses: Hu et al. (2019) used a differentiable loss that makes a segmentation's numbers of components and loops (its Betti numbers) match the ground truth, reducing broken connections in thin structures, and Chen et al. (2019) used persistence to penalize classifiers with topologically complex decision boundaries.

Questions to Think About

Why do features that persist over a long range of scales reflect the data rather than noise, and what exactly does the stability theorem guarantee (Cohen-Steiner, Edelsbrunner & Harer, 2007)?
Can topological summaries of a trained network, such as the persistence of its weight graph, track how well it learns and generalizes, and how strong is the evidence so far (Rieck et al., 2019)?

Core Math

Key Equation

\text{persistence} = \text{death} - \text{birth}

Persistent homology describes the shape of data, such as a cloud of points, at all scales at once.

Filtration. Join points that lie within distance $\epsilon$ of each other, and fill in triangles and higher-dimensional simplices, to get a simplicial complex $K_\epsilon$ . As $\epsilon$ grows the complexes are nested:

\emptyset \subseteq K_{\epsilon_0} \subseteq K_{\epsilon_1} \subseteq \cdots \subseteq K_{\epsilon_n}

Homology. For each complex, the homology group $H_k(K_\epsilon)$ counts $k$ -dimensional holes:

$H_0$ : connected components
$H_1$ : loops
$H_2$ : enclosed voids

Persistence diagram. A feature appears at some scale (its birth) and can disappear when components merge or a hole is filled (its death). A feature that never disappears has death at infinity. Plotting the (birth, death) pairs gives the persistence diagram, and each feature's lifetime is

\text{persistence} = \text{death} - \text{birth}

Stability. A small change to the input moves the persistence diagram by at most a comparably small amount (Cohen-Steiner, Edelsbrunner & Harer, 2007), so features that persist much longer than the size of the noise cannot have been created by the noise alone.