Dual and primal loss don't align for small blur values

[ericdunipace]

Hello,

First, let me say thank you for such a great program. I have found GeomLoss to be very helpful in my research.

Second, I noticed a discrepancy between the dual and primal losses for small values of the blur parameter when I was attempting to use the solutions to GeomLoss to perform some barycentric projections. I've attached some sample code that should replicate the observation. I also tried some smaller scaling values and get the same problem.

Thanks for any help you can provide!

geomloss_transport_matrix.py.zip

software

MacOSX 11.6
GeomLoss version 0.2.4
Python 3.8

[ericdunipace]

As an update, I tried this again with GeomLoss v 0.2.5 and get the same values as the python file above