Algorithm Improvements for Fast Lomb Scargle

The fast Lomb-Scargle method implemented in astropy.stats.LombScargle is based on a what is effectively fast nonuniform Fourier transform approximation proposed by Press & Rybicki (1989); the utility code is here. The result is an approximation good to roughly one part in 10^4-10^5 or so.

Since 1989, there has been a lot of work on more efficient and effective versions of this flavor of operation; for example the NFFT, with the basic algorithm outlined here. The NFFT package is GPL-licensed, and so can’t be used with astropy, but it would be fairly straightforward to replace astropy’s current trig_sum utility with a Python version of the basic algorithm outlined there. The result would be a much faster and more robust periodogram.

This could be a really fun contribution for someone who wants to dig into fast numerical algorithms…

[jakevdp]

I just released a related package: http://github.com/jakevdp/nfft

I think the best thing would be to basically replace the trig_sum utility with the nfft_adjoint code in that package (and update references in docs to reflect the algorithm change). I haven’t benchmarked it, but I believe it should be several times faster, and it will definitely lead to more accurate results – we can also tighten the tolerance in unit tests of method='fast'.

[bsipocz]

Yes, please remove milestones from non critical issues.