Circular Statistics: astropy vs scipy #13749

Activity

1. pllim

   added

   stats

   needs-discussion

   on Sep 23, 2022
2. dschmitz89 commented on Mar 11, 2023

   

   dschmitz89

   on Mar 11, 2023

   Last edited by dschmitz89

   More actions

   Hi astropy people,

   I recently worked on circular/directional stats in scipy. Indeed there is some duplication, it would be great to collaborate/coordinate a little bit.

   Before talking about APIs, I wanted to point out recent additions to scipy that could be helpful for astronomy:

   • we implemented a numerical fitting method for the von Mises distribution (see here) that gives a quick and accurate result also for the concentration parameter $\kappa$. astropy currently uses an approximation for the concentration parameter. You could consider using the scipy method instead (after a deprecation cycle maybe). Be aware that this is only available in scipy's main branch and not released yet.
   • directional_stats is a generalization of circmean and circvar that works in arbitrary dimensions. I would expect that especially $n=3$ must be very interesting for astropy.
   • the new distribution uniform_direction enables to sample uniformly from the surface of a hyphersphere
   • I am currently working on implementing the Von Mises-Fisher distribution, a generalization of the Von Mises distribution (see here).

   One function from astropy that would be useful for scipy would be the Rayleigh test for uniformity on the hypersphere. Can your implementation be generalized to $n$ dimensions maybe?

   Regarding the duplication of circmean, circvar and circstd: could astropy just use the scipy methods under the hood for some time and then maybe deprecate them in the long run? From what I see, scipy's implementation lets the user choose axis and a nan policy (not available in astropy) but does not allow weights yet (available in astropy). Those could be added if we decide to go this route.

   CC @mdhaber
3. pllim commented on Mar 13, 2023

   

   pllim

   on Mar 13, 2023

   MemberAuthor

   More actions

   Thank you for the update! Looks like a bulk of these were implemented here by @mirca back in #4472 . Our current stats subpackage maintainer is @larrybradley . Personally, I am very interested in re-using scipy but the problem is getting devs time to work on this campaign that would probably drag on for months if not years.

   For a feature that is not in any released scipy, we would be hesitant to use. It is much simpler for us to switch over if the feature is already in our minversion of scipy, which we could consider bumping as long as it adheres to APE 18.

   astropy/setup.cfg

   Line 77 in 386b52e

   scipy>=1.5
4. mhvk commented on Mar 13, 2023

   

   mhvk

   on Mar 13, 2023

   Contributor

   More actions

   Bit from the sidelines, but yes to using scipy where possible - these routines are not astronomy specific! But it will be a bit of a slow trajectory, obviously.
5. mdhaber commented on Mar 13, 2023

   

   mdhaber

   on Mar 13, 2023

   Last edited by mdhaber

   More actions

   Just thought I'd note that the SciPy and Astropy definitions of circular variance were different in the past. SciPy's was corrected in 1.9.0 by gh-5747scipy/scipy#5747, so they are consistent now. (The documentation of Astropy's circvar mentions the discrepancy; that could be updated.)

   👍React with 👍1Reacted by Larry Bradley
6. larrybradley commented on Mar 13, 2023

   

   larrybradley

   on Mar 13, 2023

   Member

   More actions

   Thanks, @mdhaber. I'll update the astropy circvar docs.

   The scipy PR that made them consistent is scipy/scipy#5747

   👍React with 👍1Reacted by P. L. Lim
7. mdhaber commented on Dec 16, 2025

   

   mdhaber

   on Dec 16, 2025 · Hidden as outdated

   Last edited by mdhaber

   show commentMore actions
8. pllim commented on Dec 17, 2025

   

   pllim

   on Dec 17, 2025 · Hidden as outdated

   Author

   show commentMore actions
9. mdhaber commented on Aug 23, 2026

   

   mdhaber

   on Aug 23, 2026

   Last edited by mdhaber

   More actions

   Thanks @pllim. For future participants, my interest in #13749 (comment) has expired, but I thought I'd median that SciPy's circular median PR (scipy/scipy#24166) has merged.

   👍React with 👍1Reacted by P. L. Lim