Decision-Dependent Ambiguity Sets

• Decision-dependent ambiguity sets are families of probability measures that adjust based on decision variables, extending traditional DRO models.
• They leverage techniques such as Wasserstein balls, moment constraints, and φ-divergence metrics to capture endogenous uncertainty with finite-sample guarantees.
• Applications include dynamic pricing, robust control, and facility location, supported by scalable algorithms and theoretical performance bounds.

Decision-dependent ambiguity sets are families of probability measures used in Distributionally Robust Optimization (DRO) in which the parameters defining the ambiguity set are functions of the optimization decision variables. This modeling paradigm captures situations where the underlying data-generating process is endogenous to the decisions themselves—a setting termed decision-dependent uncertainty (DDU). The introduction of such ambiguity sets gives rise to decision-dependent distributionally robust optimization (DD-DRO) models, which generalize classical DRO by allowing for endogenous changes in both the type and extent of distributional ambiguity as a function of chosen actions. These constructions have recently achieved significant advances in domains such as dynamic pricing, contextual optimization, resilient network planning, robust control, sequential facility location, and Markov decision processes, supported by theoretical guarantees and scalable algorithms.

1. Formal Frameworks for Decision-Dependent Ambiguity Sets

Decision-dependent ambiguity sets can take a range of forms, depending on the application domain and the nature of the uncertainty linked to decisions. A generic DD-DRO problem is

$$\min_{x \in X} \ f(x) + \sup_{P \in \mathcal{A}(x)} \mathbb{E}_{P}[h(x,\xi)],$$

where $x$ denotes the decision, $\xi$ is the uncertain parameter, and $\mathcal{A}(x)$ is an ambiguity set whose structure (support, center, radius, or all) depends explicitly on $x$. Representative families include:

Wasserstein Balls with Decision-Dependent Centers and Radii

$$\mathcal{A}(x) = \left\{ P \in \mathcal{M}(\Xi) : W_p(P, \hat{P}_N(x)) \leq r_N(x) \right\}$$

where the nominal distribution $\hat{P}_N(x)$ (center) is constructed via empirical interpolation or regression and $r_N(x)$ reflects both data coverage and statistical error in a decision-dependent manner (Qu et al., 9 Aug 2025, Fonseca et al., 2023).

Moment-Based and φ-Divergence-Based Sets

Moment set parameters (means, variances, bounds) and divergence radii are determined affinely or nonlinearly as functions of $x$ (Luo et al., 2018, Yu et al., 2020, Yu et al., 2024).

Decision-Dependent Weighted Norm Balls

In robust MDPs, the ambiguity set for the transition probabilities from a given state-action pair $(s,a)$ is an $x$0 ball with a data-driven, value-function-dependent weight vector $x$1:

$x$2

where $x$3 reflects local risk sensitivity induced by $x$4 (Russel et al., 2019).

KL and φ-Divergence with Endogenous Radius

Ambiguity budgets grow with decision features, e.g., for the process noise law in LQG control,

$x$5

with $x$6 (Fochesato et al., 13 May 2025).

Scenario- and Mode-wise Decision-Dependence

In settings with scenario-based or multimodal uncertainty, ambiguity sets are defined for each scenario or mode, with support and moment/domain parameters indexed by decisions (Li et al., 2022, Yu et al., 2024).

A comprehensive taxonomy and dual reformulations for these constructions are compiled in (Luo et al., 2018).

2. Construction and Data-Driven Learning of Ambiguity Sets

The robust modeling of DDU requires careful construction of $x$7 that faithfully tracks the decision-induced change in the law of $x$8.

• Empirical Interpolation: Discrete empirical laws are constructed at historic decision points and combined via Lipschitz-continuous interpolation (nearest neighbor, barycentric, convex) to yield a nominal measure for any $x$9 (Qu et al., 9 Aug 2025).
• Regression-Based Residual Recycling: Contextual regression models estimate the conditional mean $\xi$0; residuals are recycled to construct empirical distributions at new $\xi$1 pairs, around which ambiguity balls are centered (Zhu et al., 2024).
• Reference Mode Probability and Conditional Law: In multimodal settings, a decision-dependent nominal probability vector for the mode is matched with scenario-dependent inner distributions, both indexed by $\xi$2 (Yu et al., 2024).
• Parametric or Learned Moment Mappings: Empirical means and covariances estimated from observed data (or based on risk models) are made affine or nonlinear functions of $\xi$3 in multistage or facility location problems (Yu et al., 2020).

Data-driven calibration of ambiguity set radii is critical. Cross-validation is employed to select radii that optimize out-of-sample performance, and statistical measure-concentration results are used to provide finite-sample coverage guarantees (Qu et al., 9 Aug 2025, Zhu et al., 2024).

3. Theoretical Guarantees and Finite-Sample Performance

A hallmark of DD ambiguity set models is the provision of tight non-asymptotic performance certificates:

• Coverage Theorems: For Wasserstein/interpolated sets, if $\xi$4 is chosen appropriately, then with high probability, the true $\xi$5 lies in $\xi$6 for all $\xi$7, uniformly over the decision domain (Qu et al., 9 Aug 2025):

$\xi$8

• Optimality Gap Bounds: The realized out-of-sample cost $\xi$9 and the robust value $\mathcal{A}(x)$0 satisfy

$\mathcal{A}(x)$1

with $\mathcal{A}(x)$2 the Lipschitz constant of the cost function (Qu et al., 9 Aug 2025).

• Consistency and Rates: Under vanishing radii and consistent regression, DRO optima converge to the true optimum at rate $\mathcal{A}(x)$3 or faster under exponential error concentration (Zhu et al., 2024).
• Finite-Time Bounds for MDPs and MIPs: Weighted ambiguity balls in MDPs yield confidence intervals on the true transition law, with explicit sampling metrics (Russel et al., 2019); SDDiP-based MIP solvers for multistage problems provide UB/LB gaps on the robust objective (Yu et al., 2020).

4. Reformulations and Computational Methods

Decision-dependent ambiguity sets retain many of the tractability and duality properties of their parameter-agnostic counterparts while introducing new convex and nonconvex interactions in the optimization problem:

• Kantorovich Duality: Wasserstein-based DD-DRO admits finite-dimensional convex or conic reformulations when the nominal and radius are data/interpolation-driven (Qu et al., 9 Aug 2025, Fonseca et al., 2023).
• MILPs and MISDPs: Moment-based sets lead to tractable MILPs for box and exact-moment constraints, and MISDPs or conic programs for ellipsoidal and higher-order moment specifications (Yu et al., 2020, Luo et al., 2018).
• Dynamic Programming and Coordinate Descent: Minimax LQG with decision-dependent KL balls is addressed by a Lagrangian relaxation and backward Riccati-type recursions with coordinate gradient-descent outer loops (Fochesato et al., 13 May 2025).
• Column-and-Constraint Generation: Scenario-wise sets in resilient network planning are handled by iterative master–subproblem decompositions, with McCormick linearization for bilinear terms (Li et al., 2022).
• Bayesian and Frequentist Calibration: Sampling strategies yield objective-calibrated ambiguity radii and weights for robust MDPs, optimizing conservative Bellman updates (Russel et al., 2019).

5. Representative Applications

Recent work has demonstrated DD ambiguity set models across varied domains:

• Dynamic Pricing with Nonstationary Demand: Data-driven Wasserstein balls interpolate historical price–demand pairs to learn decision-dependent demand laws, yielding pricing strategies with quantifiable revenue guarantees (Qu et al., 9 Aug 2025).
• Contextual Pricing and Logistics: Regression-driven ambiguity construction yields robust pricing and stock/shipment strategies in two-stage models, with demonstrable outperformance over traditional ERM/SAA and static-DRO (Zhu et al., 2024).
• Distribution Network Resilience: Scenario-wise moment-constrained ambiguity sets encode how line hardening decisions shift fragility curve parameters under extreme weather, with performance superior to stochastic and robust optimization (Li et al., 2022).
• Robust Stochastic Control: LQG controllers using KL-ball ambiguity sets with decision-dependent radius retain affine optimal policy structure and solution via closed-form Riccati recursions plus dual optimization (Fochesato et al., 13 May 2025).
• Facility Location under Endogenous Demand: Multistage MILPs and MISDPs with decision-dependent moment sets address endogenous spatial demand induced by facility placement (Yu et al., 2020, Yu et al., 2024).
• Robust MDPs with Value-Sensitive Ambiguity: Weighted $\mathcal{A}(x)$4 and $\mathcal{A}(x)$5 ambiguity sets centered at nominal transitions and weighted by future value function sensitivities improve worst-case policy guarantees and empirical performance (Russel et al., 2019).
• Ambiguity Perception in Decision Theory: Models where decision makers endogenously choose their ambiguity sets, trading off ambiguity reduction and associated costs, have been shown to explain paradoxes not captured by standard frameworks (Akita et al., 5 Sep 2025).

6. Modeling Scope, Directions, and Guidelines

Decision-dependent ambiguity sets extend DRO to encompass a wide spectrum of endogenous uncertainty phenomena: price-influenced demand, investment-tuned reliability, scenario weighting, and sequential decision-driven stochasticity. While leading to more faithful risk modeling and enhanced out-of-sample performance, the construction and calibration of these sets entail increased modeling and computational complexity. Practitioners are advised to:

• Utilize MILP/MISOCP formulations for piecewise affine/ellipsoidal moment sets when computational resources permit;
• Deploy data-driven or cross-validated methods for ambiguity radius selection to avoid excessive conservatism;
• Prefer Wasserstein/φ-divergence sets when seeking asymptotic calibration and tractable duality;
• Recognize that modeling multimodality and decision-dependence leads to strictly tighter sets and improved guarantees under finite samples or model mis-specification (Yu et al., 2024).

Advancements in cut generation for MISDP, state aggregation for long horizons, and statistical learning of scenario-dependent ambiguity structures remain active research frontiers in the field (Qu et al., 9 Aug 2025, Yu et al., 2020, Zhu et al., 2024, Li et al., 2022).