Kyle Beauchamp

OK, I also just added a `svd-ew-kab` method to my pull request, which is just the `svd-ew` style calculation using `np.linalg.pinv` instead of the custom `pseudoinverse`. From the docstrings, it...

Here's a speed benchmark for the method `svd-ew-kab`: ``` name K N time 0 25x100 exponentials 25 2500 0.00436 1 100x100 exponentials 100 10000 0.0322 2 250x250 exponentials 250 62500...

For the record, here are some representative timings for the various linear algebra operations on a 5000x5000 matrix. As expected, eigenvalues are cheap, SVD / pinv are expensive. ``` In...

Here are the covariance timings with the `inverse` method. Note that is with the fix to #106, as otherwise these systems would blow up from memory usage. ``` name K...

Conclusion: the more general `svd-ew-kab` method is only 2X slower than the method that relies on the numerical inverse (and unique states), so there's not that much lost by defaulting...

Also, with the inverse method, these systems _all_ give warnings about singularity, despite having unique states. ``` Warning: W'W appears to be singular, yet 'inverse' method of uncertainty estimation requires...

It's actually discussed in that appendix, equation D5.

Yes, kab is just the svd-ew + `np.linalg.pinv`. Inverse and svd-ew should have the same memory footprint--both operate entirely on KxK matrices. I can't comment on the accuracy / singularity...

What do you mean by "does anyone know how to support pymbar3"?

So I think Michael has some familiarity with many of the changes. The changes he is less familiar with mostly live here: https://github.com/choderalab/pymbar/blob/master/pymbar/mbar_solvers.py. IMHO, these changes actually bring about substantial...