
Predicting L-Function Properties from Trace-Index Graphs using Graph Neural Networks
Can GNNs predict arithmetic properties of modular forms? We construct trace-index graphs from Fourier coefficient data — 1000 nodes per modular form, connected by sequential, primality, and kNN edges — and show ChebConv achieves R²=0.625 for L-function zero prediction on 46,347 newforms from the LMFDB. Spectral filters outperform GCN by +14.61pp in macro F1 on rank classification.

