Stability Guarantees for Topological Machine Learning

359 concepts228 papers1 proposal

Imagine you have a point cloud sampled from the surface of a circle. If you add a small amount of Gaussian noise to the data, the exact coordinates of every single point will shift. Traditional geometric models that rely on precise coordinate distances might struggle, seeing a completely different set of local neighborhoods.

However, the global shape of the data—the fact that it forms a single connected loop with a hole in the center—remains completely unchanged.

This resilience is the core promise of Topological Data Analysis (TDA). By tracking structural features like connected components, loops, and voids through a process called persistent homology, TDA allows us to capture the fundamental "shape" of data. But to use these topological features reliably in machine learning, we need a strict mathematical guarantee: if a tiny perturbation or outlier enters our dataset, our topological features must not change erratically.

This guarantee is known as topological stability. It ensures that the mapping from raw, noisy data to topological feature space is continuous and bounded, providing a robust inductive bias for downstream machine learning models.


The Intuition: Global Invariants vs. Local Noise

Why is topology inherently robust to noise? As noted by Felix Hensel et al., topological features focus on global structural invariants rather than local data fluctuations. When individual data points are corrupted, the macro-level connectivity of the space typically survives.

This property is incredibly valuable in real-world applications where data is inherently messy:

  • Time-Series Forecasting: In complex networks like traffic or blockchain transactions, Yuzhou Chen et al. demonstrate that tracking the evolving shape of a network via "zigzag persistence" provides high robustness because the global structural consistency outlasts localized data corruption.
  • Chaotic Systems: When inferring parameters for chaotic systems, Julius Von Rohrscheidt et al. show that topological descriptors provide a more stable inductive bias than standard geometric distances. Even when a tiny parameter change causes a massive, non-continuous shift in raw spatial coordinates, the underlying connectivity and holes remain stable.
  • Density Estimation: As Steve Oudot demonstrates, stability ensures that small variations in a density function (such as sampling noise) result in only minor changes to the resulting persistence diagram, allowing for reliable clustering from finite, noisy samples.
  • Structural Skeletons: Similarly, Jun Aishima shows that focusing on the topological skeleton (the medial axis) of an electron-density map provides greater stability against intensity fluctuations than tracking local intensity peaks.

Formalizing Stability: Bounding the Perturbations

To prove that a topological representation is stable, researchers establish mathematical bounds showing that the distance between two topological summaries is strictly controlled by the distance between the input data representations.

Input Space Perturbation (e.g., Hausdorff / Prohorov Distance)
               │
               ▼ (Bounded Mapping)
Topological Feature Space Change (e.g., Bottleneck / Wasserstein Distance)

A foundational way to express this is by bounding the bottleneck distance or Wasserstein distance between topological outputs by the Hausdorff distance of the inputs. For example, Vitaliy Kurlin et al. and Vitaliy Kurlin et al. use this framework to prove that hierarchical clustering descriptors (such as the mergegram) remain robust under noise, ensuring that small spatial perturbations in point sets cannot cause drastic jumps in the resulting representation.

However, Hausdorff distance can be sensitive to extreme outliers. To handle this, Andrew J. Blumberg et al. treat datasets as probability measures and use the Prohorov distance to establish stability. They prove that 2-parameter persistent homology can achieve a 1-Lipschitz stability bound, meaning that introducing a small percentage of outliers causes a strictly bounded, proportional change in the topological descriptor.


Bridging Topology and Machine Learning

The raw output of persistent homology is a persistence diagram—a collection of points in a 2D plane representing the "birth" and "death" times of topological features. Because persistence diagrams have varying numbers of points, they cannot be directly fed into standard machine learning models like Support Vector Machines (SVMs) or Neural Networks.

The core challenge of topological machine learning is transforming these diagrams into fixed-length vectors or kernels without destroying the stability guarantee.

Stable Vectorizations

If we vectorize a persistence diagram carelessly, minor noise could cause massive jumps in the vector coordinates. Researchers prevent this using clever mathematical design:

  • Persistence Duration Weighting: Zhichao Wang et al. show that weighting topological features by their persistence (death minus birth) creates a vector representation stable under the 1-Wasserstein distance. Because short-lived features (topological noise) have a weight near zero, they are naturally dampened.
  • Tropical Geometry: Anthea Monod et al. utilize max-plus semiring operations to achieve Lipschitz stability. These min-max operations naturally smooth out sensitivity to exact birth and death times, making the vectorized features highly resilient to noise.
  • Neural Network Constraints: In end-to-end deep learning, the authors of Deep Learning with Topological Signatures explicitly constrain the neural network layer's transformation to be stable under the 1-Wasserstein distance, ensuring that small variations in input data yield proportionally small changes in learned representations.

Stable Kernels

Instead of vectorizing, we can define a kernel function that measures the similarity between two persistence diagrams.

Jan Reininghaus et al. designed a multi-scale kernel by treating persistence diagrams as configurations of delta forces and modeling their heat diffusion. This formulation is provably stable under the 1-Wasserstein distance.

For more complex data, René Corbet et al. extended stable kernel formulations to multi-parameter persistent homology, while Tam Le et al. developed the Persistence Fisher Kernel using Fisher information geometry to guarantee both stability and positive definiteness on noisy experimental data.


Stability in Graphs and Relational Deep Learning

As machine learning moves beyond Euclidean data toward graphs and complex relational structures, topological stability has become a vital design principle for Graph Neural Networks (GNNs) and Topological Deep Learning (TDL).

Standard GNNs are notoriously sensitive to structural perturbations—adding or removing a few edges can drastically alter their embeddings. Tolga Birdal argues that shifting from graphs to higher-order topological structures like simplicial or cell complexes inherently enhances model robustness, creating representations that remain stable despite small deformations in the underlying graph.

To enforce this stability mathematically, researchers are embedding topological invariants directly into graph architectures:

  • Functorial Stability: Justin Curry et al. formalize topological attributes as functors from a graph category to a category of persistence modules. This mathematical framework guarantees that small amounts of input noise do not cause disproportionate changes in the resulting graph representation.
  • Stable Slicing: For multi-parameter graph data, Cheng Xin et al. introduce GRIL (Generalized Rank Invariant Landscape), proving it preserves 1-Lipschitz stability. Similarly, Yuzhou Chen et al. show that slicing multidimensional persistence into structured summaries creates a stable fingerprint that protects downstream classifiers from graph attribute noise.
  • Continuous Dynamics: Moving away from discrete message-passing steps, Aref Einizade et al. introduce continuous-time simplicial neural networks governed by partial differential equations. By modeling information flow as continuous heat diffusion on a complex, they provide mathematical guarantees of robustness against structural noise.

By anchoring machine learning models to these provably stable topological foundations, researchers can build systems that ignore local, superficial noise and focus entirely on the true, underlying shape of data.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 233 papers
  2. Version 1 · 2026-10-04 · 327 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_topological_stability_b3a9a83a,
  title = {topological stability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-stability/rdr_b3a9a83a1e2e4a1e8159518469093cbb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_topological_stability_b3a9a83a_v2,
  title = {topological stability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-stability/rdr_b3a9a83a1e2e4a1e8159518469093cbb/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake