Ranks of elliptic curves and deep neural networks (Q6175154)

Determine the (algebraic) rank of elliptic curves is an important and difficult problem. The BSD conjecture says that the (algebraic) rank of an elliptic curve over \(\mathbb{Q}\) equals to its analytic rank. Although this conjecture has been partially proved in many case, see [\textit{B. Gross} and \textit{D. Zagier}, C. R. Acad. Sci., Paris, Sér. I 297, 85--87 (1983; Zbl 0538.14023)], the general conjecture is still open. However, this paper uses deep convolutional neural network (CNNs) to predict the rank of an elliptic curve with a sequence of normalized Frobenius traces \(a_p\) for prime \(p\) in a certain range as input, and compare this method with eight simple neural network models of the Mestre-Nagao sums based on the elliptic curves in two datasets: the LMFDB and the custom dataset. The conclusion is that the CNNs outperform the Mestre-Nagao sums on the LMFDB dataset, and has a comparable performance as the Metre-Nagao sum on the custom dataset. The accuracy index is the Mathews correlation coefficient (MCC), which is generally regarded as a balanced measure of the quality of the classification even if the classes are of very different sizes [\textit{A. Boughorbel et al}, Ann. Math. (2) 141, No. 3, 443--551 (1995; Zbl 0823.11029)]. To me, although the prediction of the rank by the method is not \(100\%\) accurate, we may use this method to find high-rank elliptic curves. This also provides an example of application of machine learning in number theory. It will be interesting if we can find out why the CNN model outperforms the Mestre-Nagao sum-based model, and provide a rigorous proof for such statement.