Probability-signature dynamics explain Fourier circuit emergence in neural networks

This research demonstrates how gradient-based training selects Fourier-structured representations during modular addition tasks. Using probability signatures, the study shows that gradient interactions act as cyclic shift operators, explaining how networks achieve exact generalization and frequency matching.