Supplementary Materials WASP: Scalable Bayes via barycenters of subset posteriors
Sanvesh Srivastava, Volkan Cevher, Quoc Tran-Dinh, David B. Dunson · 2000
, which is verysparse when Nis large. Due to the simplex constraints, problem (26) always admits an optimal solution.Although (26) is a linear program, but it is large-scale when Nis large. By exploiting the sparsity ofthis problem, one can solve it efficiently by using off-the-shelf centralized LP solvers such as CPLEX orGurobi. Alternatively, we can also exploit specific structure of (26) to develop appropriate decompositionmethods that can be scaled naturally to sufficiently large dimension and can be implemented in a parallel ordistribution fashion.