Redefining Coherent Risk Measures: From Gauge Optimization to Regularization

Abstract: It is well understood that each coherent risk measure can be represented as the expectation with respect to the worst-case reweighted density function, chosen from an abstract risk envelope. This paper introduces an equivalent but more explicit definition of the risk envelope that uses gauge sets (i.e., a type of convex sets widely utilized in convex analysis and gauge optimization) to provide a generalized measure of distance between any reweighting function and the nominal one. Using the primal gauge set reweighting problem, we provide a unified framework for various existing methods in optimization under uncertainty, including risk-neutral/risk-averse stochastic programming, robust optimization, and distributionally robust optimization with moment-based and distance-based ambiguity sets. On the other hand, the associated dual problem offers an intuitive interpretation from the regularization perspective. This approach not only simplifies the derivation of classic results but also provides a versatile framework for robustness design via manipulations of the gauge sets (e.g., intersection, union, summation, convex combination, and function basis enforcement). To demonstrate this flexibility, we present approaches for customizing robustness to specific managerial needs, including methods for selecting flexible tail behaviors, addressing spatial distributional ambiguities, combining multiple robustness metrics, and achieving heterogeneous distributional robustness. We also discuss general reformulation techniques and computational approaches for this unified framework.