Abstract:We report on two machine learning experiments in search of statistical relationships between Dirichlet coefficients and root numbers or analytic ranks of certain low-degree $L$-functions. The first experiment is to construct interpretable models based on murmurations, a recently discovered correlation between Dirichlet coefficients and root numbers. We show experimentally that these models achieve high accuracy by learning a combination of Mestre-Nagao type heuristics and murmurations, noting that the relative importance of these features varies with degree. The second experiment is to search for a low-complexity statistic of Dirichlet coefficients that can be used to predict root numbers in polynomial time. We give experimental evidence and provide heuristics that suggest this can not be done with standard machine learning techniques.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is how to use neural networks to predict the root numbers of L - functions. Specifically, the author explores this problem through two experiments: 1. **Constructing interpretable models**: Based on the recently discovered "murmurations" phenomenon (i.e., the correlation between Dirichlet coefficients and root numbers), the author attempts to construct interpretable machine - learning models. These models combine Mestre - Nagao - type heuristic algorithms and murmurations features, showing that the importance of these features varies in L - functions of different degrees. 2. **Finding low - complexity statistics**: The author attempts to find a low - complexity Dirichlet coefficient statistic that can predict the root number in polynomial time. However, the experimental results show that it is difficult to achieve this using standard machine - learning techniques. In addition, the paper also explores the following related issues: - Whether deep - learning techniques such as convolutional neural networks (CNN) can be used to classify elliptic curves more efficiently. - Whether the root number can be predicted using the Frobenius trace without factoring the discriminant. - Sarnak's problem of predicting the root number in polynomial time is explored, and its relationship with the Mobius function is analyzed. ### Formula summary 1. **Euler product form of L - function**: \[ L(E, s)=\prod_{p \in P} L_p(E, s)^{-1} \] where \[ L_p(E, s)=\begin{cases} 1 - a_p(E) p^{-s}+p^{1 - 2s}&\text{if }p\text{ is good},\\ 1 - a_p(E) p^{-s}&\text{if }p\text{ is bad}. \end{cases} \] 2. **Definition of Frobenius trace**: \[ a_p(E)=p + 1-\#E(\mathbb{F}_p) \] 3. **Definition of root number**: \[ w_E=(-1)^{r_E} \] where $r_E$ is the rank of the elliptic curve $E$. 4. **Local zeta function**: \[ Z(C/\mathbb{F}_p, T)=\exp\left(\sum_{k = 1}^{\infty}\frac{\#C(\mathbb{F}_{p^k}) T^k}{k}\right) \] 5. **Convolution layer output**: \[ C(x)=\left(\sum_{k=-\lceil K/2\rceil}^{\lceil K/2 - 1\rceil} x_{S(i - 1)+1 - k} h_{j,k}\right)_{1\leq i\leq n', 1\leq j\leq F} \] where $n'=\frac{n - K + 2P}{S}+1$. Through these experiments and analyses, the author aims to reveal the potential of machine learning in number theory research and explore its limitations in predicting root numbers.

Predicting root numbers with neural networks

Predicting the cardinality and maximum degree of a reduced Gröbner basis

Learning from higher-order statistics, efficiently: hypothesis tests, random features, and neural networks

Hilbert series, machine learning, and applications to physics

Machine-Learning Kronecker Coefficients

An Interpretation of Long Short-Term Memory Recurrent Neural Network for Approximating Roots of Polynomials

Machine Learning Clifford invariants of ADE Coxeter elements

Root statistics of random polynomials with bounded Mahler measure

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

Low-rank lottery tickets: finding efficient low-rank neural networks via matrix differential equations

Surprises in high-dimensional ridgeless least squares interpolation

Exponential Expressivity of ReLU$^k$ Neural Networks on Gevrey Classes with Point Singularities

Using neural ordinary differential equations to predict complex ecological dynamics from population density data

Predicting interacting Green's functions with neural networks

LLMs learn governing principles of dynamical systems, revealing an in-context neural scaling law

Neural Networks Learn Statistics of Increasing Complexity

On the numerical rank of radial basis function kernels in high dimension

New practical advances in polynomial root clustering

Predicting Higher-order Dynamics Without Network Topology by Ridge Regression

Random matrices, the Cohen-Lenstra heuristics, and roots of unity