Machine Learning meets Number Theory: The Data Science of Birch-Swinnerton-Dyer

Yang-Hui He 何楊輝 何杨辉, Laura Alessandretti, Andrea Baronchelli, Yang-Hui HearXiv 2019 · arXiv

Abstract

The Birch-Swinnerton-Dyer (BSD) conjecture remains one of the most profound unsolved problems in mathematics, forming one of the seven Millennium Prize Problems. This study applies modern data science, machine learning, and topological data analysis (TDA) to a massive dataset of over 2.5 million elliptic curves. By treating number-theoretic properties as features in a high-dimensional dataset, we uncover hidden statistical patterns and demonstrate that while individual arithmetic quantities are difficult to predict from raw equations, they exhibit strong inter-correlations that gradient-boosted models can successfully exploit. Our findings suggest a hierarchy of mathematical difficulty for artificial intelligence and propose a "stochastic laboratory" approach to generating new number-theoretic conjectures.

1. Introduction

For decades, number theory has progressed through a cycle of empirical observation followed by rigorous proof. The Birch-Swinnerton-Dyer (BSD) conjecture itself originated from such observations made by Bryan Birch and Peter Swinnerton-Dyer in the 1960s using early computational power. The conjecture relates the rank of an elliptic curve—a measure of the size of its set of rational points—to the behavior of its associated $L$-function at a specific point.

As the complexity and volume of mathematical data grow, the challenge lies in scaling human intuition to "Big Data" levels. This paper investigates whether modern machine learning (ML) and topological data analysis can identify patterns in the arithmetic of elliptic curves that have eluded traditional analysis, specifically focusing on predicting properties like rank and the Tate-Shafarevich group.

2. Methodology

To explore the landscape of the BSD conjecture, we employed a multi-disciplinary data science toolkit:

  • Statistical Distribution Fitting: We evaluated 85 continuous distributions against the dataset. Using the Akaike Information Criterion (AIC), we identified the optimal statistical models to describe the "Right Hand Side" of the BSD equation.
  • Topological Data Analysis (TDA): Using the Eirene package for persistent homology, we visualized high-dimensional point clouds of elliptic curve data. By calculating Betti numbers and generating persistence barcodes, we sought to identify topological features—such as clusters or voids—that correlate with specific arithmetic properties like curve rank.
  • Machine Learning Models: We utilized Gradient Boosted Trees (XGBoost) for both classification (e.g., predicting rank and torsion group size) and regression (e.g., predicting the real period and regulator). These models were benchmarked against linear regression and dummy baselines to quantify the non-linear predictive power of the features.

3. Datasets and Features

The study utilizes the Cremona database, a comprehensive catalog containing approximately 2,483,649 isomorphism classes of elliptic curves over $\mathbb{Q}$. The dataset includes curves with conductors up to 400,000. For each curve, we extracted the Weierstrass coefficients ($a_1, a_2, a_3, a_4, a_6$) as raw inputs, alongside pre-computed BSD quantities including the rank, conductor, regulator, Tamagawa numbers, and the order of the torsion subgroup.

4. Key Insights and Observations

4.1 The "Cross" Symmetry

Analysis of the Weierstrass coefficients $a_4$ and $a_6$ (visualized on a signed log scale) revealed a striking, approximate cross-like symmetry. This geometric pattern is unexpected because the transformations generating this symmetry do not necessarily preserve the curve's rank, yet the structural regularity persists across the dataset.

4.2 Statistical Distributions

Our fitting procedures discovered that the distribution of the "Strong BSD" ratio is best described by a Beta distribution. This provides a new statistical lens for viewing the conjecture's global behavior across a large population of curves, moving beyond individual case studies.

4.3 The Hierarchy of Mathematical Difficulty

Our results suggest that number theory presents a unique challenge for AI compared to other fields like algebraic geometry. While previous research indicated that ML could easily predict geometric properties (such as cohomology), predicting number-theoretic properties from raw coefficients alone proved significantly more difficult, suggesting these properties are more deeply "encoded."

5. Results and Discussion

5.1 Prediction Performance

Predicting arithmetic properties from raw Weierstrass coefficients alone proved to be a "hard" problem. XGBoost achieved an F1 score of approximately 0.50 for rank and torsion—only marginally better than a baseline model. This confirms that the mapping from a curve's equation to its arithmetic is highly non-trivial.

However, the "mixed" prediction models—where BSD quantities were used to predict one another—showed remarkable success. XGBoost achieved an F1 score of 0.90 for rank and 0.91 for torsion. Furthermore, the model predicted the Tamagawa number and the regulator with a Normalized Median Absolute Error (NMAE) of 0.012 and 0.014, respectively. This demonstrates that while the raw coefficients are opaque, the BSD variables themselves are deeply inter-linked in ways the ML model can quantify.

5.2 Topological Signatures

TDA barcodes successfully distinguished between different ranks within a 6-dimensional point cloud. This provides visual and mathematical evidence that the "shape" of the data in the parameter space changes fundamentally as the rank of the elliptic curve increases.

6. Conclusion and Future Work

This research demonstrates that while machine learning cannot yet replace the rigor of a mathematical proof, it serves as a powerful "stochastic laboratory." The ability of ML to identify inter-correlations between BSD variables suggests that there are still undiscovered relationships within the conjecture's framework.

Future work will focus on expanding the dataset to include curves with larger conductors and higher ranks to determine if these patterns hold at the limits of current mathematical knowledge. Ultimately, we hope these data-driven insights provide the intuition necessary for number theorists to formulate new, provable conjectures in the study of elliptic curves.

Summary updated

Abstract

Empirical analysis is often the first step towards the birth of a conjecture. This is the case of the Birch-Swinnerton-Dyer (BSD) Conjecture describing the rational points on an elliptic curve, one of the most celebrated unsolved problems in mathematics. Here we extend the original empirical approach, to the analysis of the Cremona database of quantities relevant to BSD, inspecting more than 2.5 million elliptic curves by means of the latest techniques in data science, machine-learning and topological data analysis. Key quantities such as rank, Weierstrass coefficients, period, conductor, Tamagawa number, regulator and order of the Tate-Shafarevich group give rise to a high-dimensional point-cloud whose statistical properties we investigate. We reveal patterns and distributions in the rank versus Weierstrass coefficients, as well as the Beta distribution of the BSD ratio of the quantities. Via gradient boosted trees, machine learning is applied in finding inter-correlation amongst the various quantities. We anticipate that our approach will spark further research on the statistical properties of large datasets in Number Theory and more in general in pure Mathematics.
BibTeX
@misc{2019machine,
  author = {Yang-Hui He 何楊輝 何杨辉 and Laura Alessandretti and Andrea Baronchelli and Yang-Hui He},
  title = {{Machine Learning meets Number Theory: The Data Science of Birch-Swinnerton-Dyer}},
  year = {2019},
  eprint = {1911.02008},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.02008},
}
Figure