Distributionally robust optimization with polynomial densities: theory, models and algorithms (Q2189441)

The authors present first steps towards using SOS polynomial densities in distributionally robust optimization for problems that display a polynomial dependence on the uncertain parameters. The proposed framework is tractable for SOS density functions of any fixed degree. This approach shows flexibility in modelling in sense of encoding various salient features of the unknown distribution of the uncertain parameters through linear constraints and/or linear matrix inequalities. In the limit, as the degree of the SOS density functions tends to infinity, one recovers the usual robust counterpart or generalized moment problem. On the other hand, the approach is not applicable to objective or constraint functions that display a general (decision-dependent) piecewise polynomial dependence on the uncertain parameters as it is the case for the recourse functions of linear two-stage stochastic programs. Moreover, this approach often reduces distributionally robust optimization problems to generalized eigenvalue problems or even semidefinite programs of large sizes that are poorly conditioned.