A question about your code compared with the paper

[LJSthu]

In paper's appendix, after introducing variable v, the original formulation became projection vector v onto the possibility simplex. I think the constraints left are only sum(p)=1 and p >= 0, and I wonder why your codes still have the 'distance' constraint?

Looking forward to your reply! THANKS!

[jondeaton]

The additional distance constraint comes from placing a limit on the distributional shift as described in this paper "Learning Models with Uniform Performancevia Distributionally Robust Optimization". The idea is to limit the divergence from the uniform empirical distribution over samples to be no larger than rho.