Learning 3-manifold triangulations

Francesco Costantino, Yang‐Hui He, Elli Heyes, Edward Hirst

This paper explores the intersection of low-dimensional topology and artificial intelligence, demonstrating how machine learning and network science can help us navigate the complex "space of shapes." By representing 3-manifold triangulations as compact strings and analyzing the graphs formed by their transformations, the authors uncover hidden relationships between a manifold's geometry and its combinatorial representation.

TL;DR

The researchers used neural networks to classify 3-manifolds based on compact "isomorphism signatures" (IsoSigs) and applied network science to analyze "Pachner graphs"—the networks of all possible triangulations of a space. Their standout finding is a conjectural link between the growth of these graphs and the "systole" (the shortest loop) of hyperbolic manifolds.

What Problem Does This Paper Address?

In topology, a 3-manifold (a 3D space) can be represented by a "triangulation"—essentially a collection of tetrahedra glued together at their faces. However, a single manifold can be triangulated in infinitely many ways. Determining if two different-looking triangulations actually represent the same space is a notoriously difficult problem; while algorithms exist, their theoretical complexity is "huge" (involving towers of powers).

The authors seek to bridge this gap using statistical methods. Can machine learning "learn" the language of triangulations to identify manifolds quickly? And can we use network science to understand the structure of the Pachner Graph—a massive network where each node is a triangulation and each edge is a "Pachner move" (a local rule for changing the triangulation without changing the manifold)?

Methods

The authors employed a multi-pronged approach:

  1. Data Generation: Using the software libraries Regina and SnapPy, they generated thousands of unique triangulations for specific manifolds and knots.
  2. Representation: They used IsoSigs (Isomorphism Signatures), which are highly compressed strings that uniquely identify a triangulation. These were converted into one-hot encoded vectors for machine learning.
  3. Machine Learning: They trained standard Feed-forward Neural Networks (NNs) and Transformers (specifically BERT) to perform binary classification (e.g., "Does this IsoSig belong to Manifold A or Manifold B?").
  4. Network Analysis: They treated the Pachner moves as a graph and analyzed it using metrics like eigenvector centrality, the Wiener index (average path length), and cycle basis analysis.
  5. Interpretability: They used Gradient Saliency to "peek under the hood" of the neural networks, identifying which parts of the IsoSig string were most important for the AI's decision-making.

Datasets

  • Focus Manifolds: Eight specific 3-manifolds chosen for their diverse properties, including the 3-sphere ($S^3$), the 3-torus ($T^3$), and the "Weeks manifold" (the smallest volume hyperbolic manifold).
  • Knot Complements: A selection of fundamental knots (Unknot, Trefoil, Figure Eight, etc.).
  • The Census: A subset of the SnapPy Orientable Cusped Census, consisting of 4,815 hyperbolic manifolds with fewer than 8 initial tetrahedra.

Key Insights and Ideas

  • The "Language" of Topology: Neural networks were surprisingly good at distinguishing manifolds based on IsoSigs. More importantly, saliency analysis showed the NNs were actually "reading" the structure of the IsoSig correctly, focusing on the permutation sequences that dictate how vertices are matched during gluing.
  • Hyperbolic Explosion: The Pachner graphs of hyperbolic manifolds grow significantly faster than those of non-hyperbolic manifolds. This suggests that hyperbolic spaces have a much more "complex" or "dense" landscape of possible triangulations.
  • The Systole Conjecture: The authors propose Conjecture 1: the growth rate of a hyperbolic manifold's Pachner graph is inversely related to its systole (the length of its shortest closed geodesic). Essentially, the "tighter" the manifold is pinched, the more nodes its Pachner graph contains at a given depth.
  • Transformer Limitations: Interestingly, standard Transformer models (like BERT) performed worse than simple NNs. This suggests that the "grammar" of topological IsoSigs is fundamentally different from natural human language, making off-the-shelf NLP transfer learning less effective.

Key Results

  • High Accuracy: NNs achieved near-perfect accuracy (>0.99) in many cases. Even for "hard" pairs like the lens spaces $L(7,1)$ and $L(7,2)$ (which are homotopy equivalent but not homeomorphic), the NN achieved an accuracy of ~0.74.
  • Dehn Surgery Success: When distinguishing the Unknot from the Trefoil, accuracy jumped from ~0.63 (knot complements) to over 0.97 after performing Dehn surgery, showing that the resulting closed manifolds are much easier for the AI to differentiate.
  • Graph Centrality: In most cases, the "minimal" triangulation (the one with the fewest tetrahedra) was not the most central node in the Pachner graph. For $L(7,2)$, the most central triangulation was actually two moves away from the minimal one.

Limitations and Future Work

  • Computational Ceiling: Generating Pachner graphs is memory-intensive. Even with 200GB of RAM, the authors could only explore graphs to a depth of 5 before hitting limits.
  • Architecture Optimization: Since Transformers struggled, future work could involve designing custom architectures specifically for the "syntax" of IsoSigs.
  • Graph Neural Networks (GNNs): The authors suggest that instead of just looking at strings, future models should operate directly on the Pachner graphs themselves to learn the "navigation rules" of the triangulation space.
  • Proving the Conjecture: The observed relationship between graph size and the systole remains a conjecture based on empirical data; a formal mathematical proof is the next big hurdle.

Summary updated

Abstract

Abstract Real 3-manifold triangulations can be uniquely represented by isomorphism signatures. Databases of these isomorphism signatures are generated for a variety of 3-manifolds and knot complements, using SnapPy and Regina, then these language-like inputs are used to train various machine learning architectures to differentiate the manifolds, as well as their Dehn surgeries, via their triangulations. Gradient saliency analysis then extracts key parts of this language-like encoding scheme from the trained models. The isomorphism signature databases are taken from the 3-manifolds’ Pachner graphs, which are also generated in bulk for some selected manifolds of focus and for the subset of the SnapPy orientable cusped census with <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mrow> <mml:mo>&lt;</mml:mo> <mml:mn>8</mml:mn> </mml:mrow> </mml:math> initial tetrahedra. These Pachner graphs are further analysed through the lens of network science to identify new structure in the triangulation representation; in particular for the hyperbolic case, a relation between the length of the shortest geodesic (systole) and the size of the Pachner graph’s ball is observed.
BibTeX
@article{costantino2025learning,
  author = {Francesco Costantino and Yang‐Hui He and Elli Heyes and Edward Hirst},
  title = {{Learning 3-manifold triangulations}},
  journal = {Journal of Physics A Mathematical and Theoretical},
  year = {2025},
  doi = {10.1088/1751-8121/adb5de},
}

Loading related content...

Figure