v0.22.0

Added

• Shift estimator choice: Warping and RawWarping accept shift_estimator ("xcorr", the new default, or "peaks") and min_gain, the relative gain in correlation with the reference that a shift (or a DTW result) must bring before it is applied.
• Peak-supported warps: Warping and RawWarping accept min_support (default 0.6), the share of matched peak pairs that a non-rigid warp must bring within a third of the maximum shift of each other, and n_peaks (default 200), the number of most prominent peaks used for matching.
• Two-step alignment: Warping and RawWarping accept prealign (default True): every method first applies a global shift and then aligns to the reference of the shifted spectra, which keeps the reference sharp when spectra come from different calibrations; the global shift of both steps stays within the maximum shift. RawWarping runs both steps at full resolution before binning.
• Optional renormalization: Warping and RawWarping accept renormalize (default False), which divides each aligned spectrum by its total intensity so that it sums to 1.

Changed

• Alignment applied only when justified: the default "xcorr" shift is chosen by cross-correlation with the reference; a shift or a DTW result is applied only if it raises the spectrum's correlation with the reference by min_gain, and non-rigid warps, fitted after that shift on the matched peaks (prealign), are kept only if the matched peaks support them (min_support). Non-rigid output preserves the total intensity (apart from what is moved past the spectrum edges); "dtw" does not. Spectra of several species are best aligned one species at a time.
• Tighter default search: Warping(max_shift=...) defaults to 3 bins (about 9 Da with the default 3 Da bins) and RawWarping(max_shift_da=...) to 10 Da, instead of 50.
• Output dtype: Warping.transform returns the floating dtype of its input for every method.
• Migration: aligned spectra differ from earlier releases. shift_estimator="peaks" with the former max_shift is the closest setting, but the earlier estimates were biased by noise peaks and cannot be reproduced exactly.
• Faster, leaner Warping.transform: spectra are aligned in chunks, so large batches run much faster in parallel and use far less memory.
• Lazy DTW dependency: tslearn is imported only when DTW alignment runs.

Fixed

• Peak matching: sample and reference peaks are paired only when mutually nearest and within the maximum shift (also in align_peaks). Previously every sample peak was paired with its nearest reference peak, so noise peaks biased the shift and peaks beyond the last reference peak produced very large warps.
• LOWESS edges and robustness: the "lowess" warp keeps the shift constant beyond the outermost matched peaks instead of collapsing those regions onto a single position, and its robustness iterations no longer break down when the peak offsets are fitted exactly (which could move part of a spectrum by hundreds of bins).
• No silent reordering: a non-monotonic warp is no longer sorted and averaged silently; the global shift is used for that spectrum and the aggregated warning reports it.
• Piecewise precision: each "piecewise" segment shift is the mean of the peak offsets close to the segment median, instead of the median itself, which on binned spectra could only take whole-bin values.