Merge Dcorr/Energy/MMD/Hsic into scipy

[sampan501]

No description provided.

[bstraus1]

Can I be assigned to this issue?

[bstraus1]

I've added the distance_correlation function (and its helpers) to my fork of SciPy in the ben-master branch.

The distance_correlation function is modeled off of the multiscale_graphcorr function which is already merged into SciPy and the helpers are from the hyppo repo (primarily the dcorr function).

I've also added a simple test (more to see that the code runs and returns an expected result rather than for accuracy). The function passes the test. This link goes directly to the distance_correlation function which relies on multiple other functions in the _stats.pyx file.

See attached screenshot of the test passed. Tests are in the same branch here. @sampan501

Suggested future steps include:

• Add more tests to assess the accuracy of the distance_correlation function to ensure that it is the same as Sambit's implementation.
• Add a doc_string to the distance_correlation function. Currently the docstring matches the multiscale_graphcorr docstring since that is what the distance_correlation function was modeled off of. The docstring should be changed to refer specifically to distance_correlation.
• Merge these changes into SciPy

[sampan501]

None of the functions that you added have been Cythonized. All the functions added use def which is calling pure Python

[bstraus1]

Then I don’t really understand why cythonizing is necessary if pure python works. @sampan501

[sampan501]

Then I don’t really understand why cythonizing is necessary if pure python works.

Pure python is slow. With large datasets, it will long time to run.

[bstraus1]

Oh I see. I wasn't aware of that being an essential part for any reason beyond making the code run-able in SciPy.

[bstraus1]

Updates:

1. _dcorr has been cythonized
2. _center_distance_matrix has been changed to work for biased and unbiased data and will now work for mgc (multi-scale graph correlation) and dcorr (distance correlation).
3. Bug Fix: The distance_correlationfunction now uses _dcorr in the pvalue computation function _perm_test to compute the test statistic rather than _mgc_stat
4. Documentation: Added more comments throughout the distance_correlation function call tree

@sampan501

[bstraus1]

Documentation for the distance_correlation function still needs to be changed (it is currently just copied from multiscale_graphcorr.

[kfangster]

Can I be assigned this issue?

[sampan501]

https://github.com/kfangster/scipy