Machine-Learning Mathematical Structures

Yang‐Hui He

TL;DR

This paper explores a "top-down" approach to mathematics, using standard machine learning (ML) tools to uncover patterns in vast datasets across geometry, algebra, and number theory. By treating mathematical structures as labeled data, the author demonstrates that AI can predict complex invariants—like Hodge numbers or group properties—with surprising accuracy, offering a new "bag of tricks" for conjecture formulation and accelerated computation.

What Problem Does This Paper Address?

Traditionally, mathematical discovery follows two paths: "bottom-up" formal logic (axioms leading to proofs) and "top-down" intuition (experimental observation leading to conjectures). While automated theorem provers (like Lean or Coq) have made strides in the former, the latter has remained a human stronghold.

However, modern mathematics has generated massive databases of objects (Calabi-Yau manifolds, graphs, elliptic curves) that are too complex for manual inspection. Traditional algorithms for computing their properties, such as Gröbner bases, often suffer from double-exponential complexity. This paper investigates whether ML can bridge this gap by acting as a "digital intuition," identifying patterns in these datasets that might lead to new conjectures or provide near-instant approximations for computationally expensive problems.

Methods

The author purposefully avoids "over-engineering" the AI. Instead of developing bespoke architectures, the paper uses a standard "bag of tricks" from supervised learning to see how well mathematical structure responds to general-purpose tools:

  • Neural Networks (MLP): Simple feed-forward architectures with a few layers.
  • Support Vector Machines (SVM): Used to find hyperplanes that separate different mathematical categories.
  • Naive Bayesian Classifiers: Used for statistical pattern matching, particularly in number theory.
  • Decision Trees & Clustering: For categorizing labeled data into tree-like structures.

The primary metric for success is Accuracy, reported as a pair: $(p, \phi)$, where $p$ is naive precision and $\phi$ is the Matthews’ correlation coefficient (a measure of prediction quality).

Datasets

The paper synthesizes results from several major mathematical databases:

  • CICY (Complete Intersection Calabi-Yau): Matrices representing manifold configurations.
  • Kreuzer-Skarke Dataset: A massive collection of nearly 500 million reflexive polytopes.
  • GAP (Groups, Algorithms, Programming): Used to generate Cayley tables for finite groups.
  • Wolfram Graph Database: Approximately 8,000 connected, simple, undirected graphs.
  • LMFDB (L-functions and Modular Forms Database): Specifically for elliptic curve properties and coefficients of L-functions.

Key Insights and Ideas

  • The Hierarchy of Difficulty: The most striking insight is that different branches of math have varying levels of "learnability." The author proposes a hierarchy: Numerical Analysis < Algebraic Geometry < Representation Theory < Combinatorics < Analytic Number Theory.
  • Representation is Everything: ML struggles to predict the next prime number from a simple list. However, when elliptic curves are represented by their L-function coefficients, ML can predict their rank and torsion with near-perfect accuracy.
  • The "Simple Group" Hyperplane: The paper suggests a "proto-conjecture": there may exist a high-dimensional hyperplane that cleanly separates simple finite groups from non-simple ones based solely on their Cayley tables.
  • Top-Down vs. Bottom-Up: ML provides the "semantics" (meaning/patterns) while formal logic provides the "syntax" (rules/proofs). The author argues that machine-aided conjectures will soon go hand-in-hand with automated proofs.

Key Results

  • Algebraic Geometry: Predicting Hodge numbers for CICY manifolds reached >99% accuracy using convolutional neural networks.
  • Representation Theory: Distinguishing simple groups from non-simple groups achieved ~98% accuracy using an SVM with a Gaussian kernel.
  • Number Theory: While predicting the Liouville $\lambda$-function was no better than a random guess (0.50), predicting the rank and torsion of elliptic curves achieved 98% to 100% accuracy using Naive Bayes.
  • Combinatorics: Graph properties like planarity and genus were learnable with ~80% accuracy, while testing if a graph is acyclic reached 95%.

Limitations and Future Work

  • Interpolation vs. Extrapolation: A major risk is that ML might only "interpolate" within the known dataset. Future work must ensure models can "extrapolate" to larger, unseen mathematical structures (e.g., training on small graphs and validating on much larger ones).
  • The Black Box Problem: While an NN might give the correct answer 100% of the time, it doesn't provide a "proof." The field of Interpretable ML is needed to extract human-readable formulae from these trained models.
  • The Skewes Number Caveat: The author warns that patterns in math can hold for incredibly long sequences before failing (like the famous Skewes number), meaning data-mined conjectures must still be treated with caution until formally proven.

Summary updated

Abstract

We review, for a general audience, a variety of recent experiments on extracting structure from machine-learning mathematical data that have been compiled over the years. Focusing on supervised machine-learning on labeled data from different fields ranging from geometry to representation theory, from combinatorics to number theory, we present a comparative study of the accuracies on different problems. The paradigm should be useful for conjecture formulation, finding more efficient methods of computation, as well as probing into certain hierarchy of structures in mathematics.
BibTeX
@article{he2023machine,
  author = {Yang‐Hui He},
  title = {{Machine-Learning Mathematical Structures}},
  journal = {Int. J. Data Sci. Math. sci},
  year = {2023},
  doi = {10.1142/s2810939222500010},
}
Figure