florisvb / PyNumDiff / 34169973048

94%

push

github

Drop splinediff's num_iterations, which was doing nothing

Refitting a spline to its own output holds the residual budget s*N*sigma^2 fixed while the input gets cleaner, so each pass satisfies the same budget with fewer knots and can only smooth harder. Measured, it does: on a sine at s=1 the derivative RMSE runs 0.80, 2.55, 3.93, 4.26 for one through five passes, and peak-to-peak falls from 1.80 to 1.10 against a truth of 1.95, the opposite of the peak preservation the parameter was there to provide. Iterating only ever rescues an s set too small, and simply raising s beats every iterated variant. The search agrees: across 36 sim/step/noise combinations it picked one pass 33 times, and the three exceptions all sat at the s floor and scored identically with iteration forced off. Also tidies a docstring in basis_fit.

1 of 1 new or added line in 1 file covered. (100.0%)

800 of 848 relevant lines covered (94.34%)

0.94 hits per line