The definition of a linear program is following:
Find a vector such that: , subject to and .
Generally, is assumed to be a fixed constant. However is it possible to construct a program where values of are part of the optimization? Could I included b in the optimization by changing to . If so, would I also be able to place constraints upon like and ? Finally, would such a program be possible to solve efficiently?
I am trying to solve the linear program for Wasserstein Distance between two discrete distributions. In the standard case, b represents the marginals for each datapoint. I know the marginals for the target distribution but the marginals from my source distribution are unknown. I am wondering if there is an efficient way to optimize the marginals for my source distribution such that the Wasserstein distance is minimized.