Differentiable Optimization of Persistent Homology

This direction is merged into another direction: topological computation scalability. Its last version stays readable and citable here.

Version 1 of 111 papers23 conceptsfrozen 2026-10-04

In traditional machine learning, we often define "similarity" through local geometry—minimizing the distance between two points or maximizing the margin of a classifier. However, geometry alone often fails to capture the global structure of data, such as the connectivity of a graph or the presence of loops in a biological structure.

To capture these global features, researchers use Persistent Homology (PH), a tool from Topological Data Analysis (TDA) that identifies "holes" and "voids" across different scales. While PH was originally a descriptive tool, a new research direction focuses on turning it into an optimization objective. By making topological features differentiable, we can "sculpt" the shape of data directly.

The Challenge of Optimizing Shape

The core difficulty in optimizing for topology is that topological features are discrete. A hole either exists or it doesn't. To bridge this gap, Mathieu Carrière, Frédéric Chazal, and their colleagues established a theoretical framework in Optimizing persistent homology based functions. They proved that persistence maps are "tame" (specifically, semi-algebraic), which allows for the use of stochastic subgradient descent.

This means we can define a loss function $\mathcal{L}_{topo}$ that penalizes the absence of a specific topological feature. For example, if we know a dataset should represent a circle, we can penalize the optimization process until a one-dimensional "hole" appears in the persistence diagram.

The Instability Problem: Why Points Fly to Infinity

A recurring insight across this literature is that optimizing for topology in isolation is dangerously unstable. In Optimizing persistent homology based functions, the authors observe that if you tell an algorithm to "maximize the size of a hole," the most efficient way for the algorithm to satisfy that requirement is to push the data points toward infinity.

Because topological features like "connectedness" or "voids" are scale-invariant, a purely topological loss function lacks a sense of physical scale. Without constraints, the gradient descent process will expand the point cloud indefinitely to make the topological features more "persistent."

The Solution: Geometric Regularization

To create meaningful structures, researchers have found that topological losses must be anchored by geometric constraints. This creates a "tug-of-war" between the desired global shape and the local spatial integrity.

  1. Spatial Penalties: In medical imaging, Jinah Park demonstrates in LV-Net that reconstructing the brain's lateral ventricles requires a multi-objective loss. By combining topological integrity with Laplacian regularization (which ensures smoothness) and anatomical templates, the model prevents the shape from deforming into unrealistic configurations.
  2. Avoiding Local Minima: Interestingly, topological penalties can actually help optimization. Amina Shabbeer and Cagri Ozcaglar show in Crossing Minimization within Graph Embeddings that incorporating topological penalties into non-convex objectives acts as a regularizer. It pushes the system out of local minima toward configurations that are both spatially accurate and visually clear.

Scaling Topology with Minibatches

Even with regularization, calculating persistent homology on a full dataset is computationally expensive and can lead to "inward distortion," where the gradient influence is too aggressive.

In Topological Node2vec, Yusuke Imoto, Yasuaki Hiraoka, and their team propose a more stable alternative: topological regularization on minibatches. By applying the topological loss to small, random subsets of the data, the gradient influence is spread out over time. This prevents any single topological feature from dominating the geometry, ensuring that the final embedding preserves the global loops and voids of the original graph without collapsing the local structure.

Summary of the Intuition

Optimizing persistent homology works by treating the "birth" and "death" of data features as differentiable events. However, because topology doesn't care about size, these functions are naturally divergent. The "secret sauce" of this research direction is the integration of geometric anchors—spatial constraints, anatomical templates, or minibatch sampling—that force the abstract topological requirements to manifest within a bounded, physically meaningful space.

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_differentiable_optimization_of_persisten_d2d62ec2,
  title = {Differentiable Optimization of Persistent Homology},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/differentiable-optimization-of-persistent-homology/rdr_d2d62ec2600a4e9da7d7af3872e5028b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_differentiable_optimization_of_persisten_d2d62ec2_v1,
  title = {Differentiable Optimization of Persistent Homology},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/differentiable-optimization-of-persistent-homology/rdr_d2d62ec2600a4e9da7d7af3872e5028b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}