19
$\begingroup$

Is there a multivariate alternative to two-sample Kolmogorov–Smirnov test? What I mean is a test that can be used to check whenever two underlying multidimensional distributions differ.

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ I am facing the same question. Have you learned anything useful and made any conclusions? I do see the answer, but beyond it I wonder about your experience with the suggestions and what worked best. $\endgroup$
    – Richard Hardy
    Commented Mar 31, 2021 at 14:06
  • $\begingroup$ @RichardHardy nope. $\endgroup$
    – Tim
    Commented Mar 31, 2021 at 14:22

3 Answers 3

Reset to default
20
$\begingroup$

A 2004 article On a new multivariate two-sample test by Baringhaus and Franz maybe helpful, they provided a brief literature review on the two-sample multivariate GoF tests and then a R package cramer. As the package name suggested their method is related to Cramer's test, a predecessor of Cramer-von Mises.

For one-sample problem Justel et al. developed a generalization of Kolmogorov-Smirnov test. In general it seems the difficulty in multivariate case rooted from extending the definition of EDF (empirical distribution function), so methods based on other measures are worth exploring, e.g. multivariate tests based on ECF (empirical characteristic function) by Fan.

Improve this answer
$\endgroup$
4
  • 3
    $\begingroup$ Please spell out abbrevs EDF, ECF $\endgroup$
    – kjetil b halvorsen ♦
    Commented Apr 13, 2017 at 22:24
  • 2
    $\begingroup$ @kjetilbhalvorsen: fixed, though it's quite clear what they are once click the links provided. $\endgroup$
    – Francis
    Commented Apr 13, 2017 at 23:00
  • $\begingroup$ Is there a python implementation of this? $\endgroup$
    – Carlos Mougan
    Commented May 4, 2022 at 21:15
  • 1
    $\begingroup$ @CarlosMougan: Not that I know of. But you can find the source code in the package page I linked (or here), which doesn't seem to be too difficult to convert to python. $\endgroup$
    – Francis
    Commented Jun 15, 2022 at 10:02
0
$\begingroup$

Yes, the two sample KS test was extended to the multivariate case in 2021. https://www.sciencedirect.com/science/article/pii/S016771522100050X

Improve this answer
$\endgroup$
1
  • 2
    $\begingroup$ While the link maybe helpful in answering OP's query, in the long run, link only answers prove useless once the link dies. So, it would be appreciable if you are able to jot down the crux of what has been presented there. $\endgroup$
    – User1865345
    Commented Jan 2, 2023 at 17:06
0
$\begingroup$

For a practical application of the multivariate KS distance in high dimensions, see my article New Python Library to Evaluate AI-generated Data and Compare Models. I actually created a Python library to do the computations.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.