Random Distributionally Robust Optimization under Phi-divergence

This paper introduces a novel framework, Random Distributionally Robust Optimization (RDRO), which extends classical Distributionally Robust Optimization (DRO) by allowing the decision variable to be a random variable. We formulate the RDRO problem using a bivariate utility function and $φ$-divergence ambiguity sets, enabling a more flexible and realistic treatment of uncertainty. The RDRO framework encompasses a broad range of robust decision-making applications, including portfolio optimization, healthcare resource allocation, and reliable facility location. By optimal transport theory and convex analysis, we characterize key structural properties of the RDRO problem. Our main theoretical contributions include establishing the existence and uniqueness of optimal randomized decisions and proving a duality theorem that links the constrained RDRO formulation to its penalized counterpart. We further propose an efficient numerical scheme that combines the scaling algorithm for unbalanced optimal transport with projected gradient descent, and demonstrate its effectiveness through numerical experiments.