Can a perfect training fit predict well?
Compare one alias block with nine while keeping preference fixed. Both fit the orange points; watch the curve between them and its test error.
Check other assumptions
Try other targets, noise levels, or favored frequency bands after comparing feature count and preference.
The preference applies to frequencies −3 through +3.
Exact model and numerical audit
Frequencies in one residue class modulo n are exact aliases on the regular grid. With feature masses lambda_k and sampled coefficient c, the weighted norm is sum |a_k|² / lambda_k, subject to sum a_k = c. Its minimum assigns a_k = c lambda_k / sum(lambda_aliases). The implementation uses the equivalent complex Fourier notation from the JSAIT treatment and always returns a real function.
n=49, d=441, s=7, favored |k|<=3 low-frequency mass=96.000000 max |f_hat(x_i)-y_i|=< 1e-12 train MSE=< 1e-12
