Code for Limits of spectral learning under noise

This repository contains the Julia code used to generate the numerical results for the manuscript Limits of spectral learning under noise. The code implements sparse spectral regression with centered and whitened design matrices, quantifies how additive label noise affects learned spectral coefficients, and computes metrics such as coefficient overlap, normalized coefficient distance, spectral entropy, reconstruction error, and test RMSE. The experiments cover one- and two-dimensional target functions and compare several orthonormal spectral bases, including Fourier, Legendre, Bessel, Jacobi, Chebyshev, and Haar bases.