Robust expected utility maximization with medial limits

Daniel Bartl, Patrick Cheridito, Michael KupperJournal of Mathematical Analysis and Applications 2018 · DOI · arXiv

Introduction

In classical mathematical finance, portfolio optimization typically assumes a known probability distribution for asset prices. However, "robust finance" acknowledges that the true model is often unknown, leading to the problem of robust expected utility maximization. This involves finding a trading strategy that maximizes the worst-case expected utility across a set of possible models $\mathcal{P}$.

Historically, solving this problem has relied on two restrictive "crutches": model domination (the assumption that all models agree on which events are impossible) and time-consistency (the assumption that the model set is stable enough to allow for step-by-step dynamic programming). This paper presents a significant theoretical breakthrough by removing these assumptions, allowing for "nondominated" model uncertainty and models that do not fit the neat structure required for backward induction.

Overcoming the Limitations of Dynamic Programming

The traditional method for solving multi-period optimization problems is dynamic programming. This approach requires the set of models $\mathcal{P}$ to be time-consistent, allowing the problem to be solved backward from the terminal date. In many realistic scenarios—particularly those involving Knightian uncertainty—this structure is absent.

The authors bypass the need for backward induction by adopting a "global" functional analytic approach. Instead of a step-by-step decomposition, they treat the optimization as a static problem over the entire time horizon. This shift allows for the inclusion of highly irregular model sets that do not support the standard recursive structure.

Technical Innovations: Medial Limits and Capacitability

To handle the lack of a dominant measure and the potential irregularity of the model set, the paper introduces several sophisticated mathematical tools:

Medial Limits

A central challenge in nondominated models is defining a consistent expectation when sequences of functions do not converge in the traditional sense. The authors employ medial limits, mathematical objects that sit between a lim inf and a lim sup. Medial limits provide a linear and monotonic way to "select" a limit, effectively acting as the "secret sauce" that maintains the properties of an expectation in a robust environment. While these require set-theoretic axioms beyond standard Zermelo-Fraenkel (such as Martin's Axiom), they provide the necessary regularity for the robust setting.

Choquet’s Capacitability Theorem

The authors derive a functional version of Choquet’s capacitability theorem. This allows them to extend results from "nice" continuous functions to more complex Borel measurable functions by approximating them from within. This technical bridge is essential for proving the existence of optimal strategies for general random endowments.

Convex Duality

The paper establishes a dual representation of the utility problem. Rather than directly maximizing utility over a space of trading strategies, the authors show that the problem can be solved by minimizing a specific functional (involving "v-divergence") over a set of dual martingale measures. This duality often simplifies the analysis and provides a clearer characterization of the optimal solution.

Key Results and Applications

The primary contribution of the paper is an existence theorem: under a general "No-Arbitrage" (NA) condition—where every model in the set is dominated by another model that does not allow for "free lunches"—and the assumption of a medial limit, an optimal trading strategy is guaranteed to exist for any Borel measurable random endowment.

Furthermore, the framework is shown to be highly versatile, accommodating complex modern financial constraints such as:

  • Moment Constraints: Scenarios where only the mean or variance of prices is known.
  • Wasserstein Constraints/Penalties: Situations where the "true" model is assumed to be within a certain optimal transport distance from a reference model.

Limitations and Future Directions

While the paper provides a robust theoretical foundation, it introduces certain complexities. The reliance on medial limits adds a layer of set-theoretic complexity, as their existence is not guaranteed by standard ZF axioms. Additionally, the current results are derived for discrete-time trading.

Future research remains to be done in extending this "medial limit" approach to continuous-time models. Furthermore, as a purely theoretical work, translating these dual representations into practical computational algorithms—perhaps through deep learning or high-dimensional numerical optimization—is a significant open area for the field.

Summary updated

BibTeX
@article{bartl2018robust,
  author = {Daniel Bartl and Patrick Cheridito and Michael Kupper},
  title = {{Robust expected utility maximization with medial limits}},
  journal = {Journal of Mathematical Analysis and Applications},
  year = {2018},
  doi = {10.1016/j.jmaa.2018.11.012},
}
Figure