Mutual Information with Grassberger/Shuffle Not Matching KNN

[cing]

There appears to be some discrepancy when using the standard CLI mutual information calculation with the grassberger estimator in the latest dev version of mdentropy. When you do >0 shuffle, there appears to be only non-zero mutual information on the diagonal.

Issue identified here: 63441b1

@cxhernandez Do you need anything from me to reproduce this problem?

[cxhernandez]

So, I couldn't exactly recreate this issue with Fs Peptide. There's definitely a trend that shuffling drastically reduces signal for the Grassberger estimator, but for Fs peptide, at least, it doesn't completely eliminate it:

MutInf Benchmark.ipynb.zip

I'm going to be chatting with a few people in the EE dept tomorrow about entropy estimators, and I'll be sure to bring this up to them.