Paper: arXiv 2611.04021

Authors: Peng Liu, Steven Vanduffel, Carole Bernard

Abstract

We study the worst-case values of distortion riskmetrics over Bregman–Wasserstein balls around a reference distribution. Unlike the squared $2$-Wasserstein distance, Bregman–Wasserstein divergences can penalize upward and downward deviations differently. Our main result applies to generalized distortion functions $g:[0,1]\to\R$ of bounded variation, with $g(0)=0$ and suitable endpoint regularity, which need not be monotone, normalized, or upper semicontinuous. For strictly convex generators satisfying a two-sided $p$-growth condition, we obtain a one-dimensional dual representation of the worst-case value. When a positive Lagrange multiplier makes the divergence constraint binding, we identify a unique quantile function through a least concave majorant and characterize when it attains the supremum. For absolutely continuous generalized distortion functions, we characterize the solution through an isotonic projection; we also study more general generators and shape constraints, including unimodality. Finally, we extend the analysis to intersections of Bregman–Wasserstein balls, describing models acceptable to several parties involved in decision making, and illustrate the results numerically.

Complexity vs Empirical Score

  • Math Complexity: 9.0/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: The paper presents highly advanced mathematical derivations for robust riskmetrics, focusing on theoretical existence and characterization. While it includes a numerical illustration, the empirical validation is limited, lacking extensive backtesting or real-world data application beyond a single case study.

Research Flowchart

  flowchart TD
    A[Research Goal: Worst-case distortion riskmetrics over Bregman-Wasserstein balls] --> B{Methodology: Dual Representation & Isotonic Projection};
    B --> C[Inputs: Reference distribution, Generalized distortion functions, Bregman-Wasserstein generators];
    C --> D{Computational Process: Least concave majorant, Isotonic projection, Numerical illustration};
    D --> E[Outcome 1: One-dimensional dual representation for worst-case value];
    D --> F[Outcome 2: Characterization of unique quantile function & solution for absolutely continuous functions];
    D --> G[Outcome 3: Extension to intersections of Bregman-Wasserstein balls];