Decision-Dependent DRO

• Decision-dependent distributionally robust optimization (DD-DRO) is an optimization framework where the ambiguity set of probability distributions is a function of the decision.
• It employs diverse constructions such as moment, Wasserstein, and phi-divergence sets to model how choices influence uncertainty and ensure tractable reformulations.
• DD-DRO has practical applications in portfolio optimization, dynamic pricing, and facility location, offering finite-sample guarantees and adaptive robust solutions.

Searching arXiv for recent and foundational papers on decision-dependent distributionally robust optimization. Decision-dependent distributionally robust optimization (DD-DRO) studies optimization problems in which the ambiguity set of probability distributions depends on the decision itself. In its basic form, one solves

$$\min_{x\in X}\;\sup_{P\in\mathcal P(x)}\; \mathbb E_P[h(x,\xi)],$$

where $x$ is the decision, $h(x,\xi)$ is the random cost or recourse value, and $\mathcal P(x)$ is a decision-dependent ambiguity set. This formulation is designed for situations with endogenous uncertainty: the decision affects the uncertainty description, the nominal distribution, the admissible deviation from that nominal model, or the information that becomes available before recourse. The framework includes two-stage decision-dependent distributionally robust stochastic programming as a special case, and its tractability depends on how $\mathcal P(x)$ is constructed and how the inner worst-case expectation is dualized or approximated (Luo et al., 2018).

1. Core formulation and conceptual scope

The canonical DD-DRO model uses a decision set $X\subseteq\mathbb R^n$, an uncertain vector $\xi$ on $(\Xi,\mathcal F)$, and a cost function $h(x,\xi)$. Nature selects the worst-case distribution $P$ within the decision-dependent ambiguity set $x$0, so the inner supremum explicitly models an adversary that reacts to the chosen decision. In this sense, DD-DRO differs from classical DRO not by replacing the min–max structure, but by making the feasible distributions endogenous to $x$1 (Luo et al., 2018).

This endogenous dependence takes several forms in the literature. The ambiguity-set radius may vary with the decision; moment targets, covariance bounds, or cumulative-distribution tolerances may be functions of the decision; the nominal distribution itself may be rebuilt at each decision; or only part of the uncertainty may be revealed after a first-stage investment. A recurring interpretation is that DD-DRO captures settings in which the decision affects either the uncertainty-generating mechanism or the decision maker’s observational access to it.

Two-stage variants make this especially explicit. In decision-dependent information discovery, first-stage binary measurement decisions $x$2 determine which components of $x$3 will be observed before recourse, while the ambiguity set remains distributionally robust. The resulting model can be written as a min–max–min–max problem, because one chooses first-stage actions, nature chooses a latent scenario, the decision maker selects a recourse action after partial revelation, and nature then chooses a compatible realization within the revealed-information class (Jin et al., 2024).

2. Ambiguity-set constructions

A foundational contribution of the finite-support DD-DRO framework is the identification of five decision-dependent ambiguity-set families. Each family preserves the basic DRO logic while relocating one or more uncertainty descriptors from fixed parameters to functions of the decision (Luo et al., 2018).

Family	Definition pattern	Decision-dependent quantities
Simple measure- and moment-inequality sets	$x$4, $x$5	Measure bounds, moment intervals
Mean-covariance ellipsoid sets	Delage–Ye type bounds on $x$6 and $x$7	$x$8
Wasserstein-ball sets	$x$9	Radius $h(x,\xi)$0, possibly $h(x,\xi)$1
$h(x,\xi)$2-divergence sets	$h(x,\xi)$3	Divergence budget $h(x,\xi)$4
Kolmogorov–Smirnov sets	$h(x,\xi)$5	K-S tolerance $h(x,\xi)$6

These five families already cover several distinct notions of ambiguity: bounds on moments, transportation-based neighborhoods, likelihood-ratio-type neighborhoods, and empirical-distribution tolerances. In the finite-support K-S form, for example, the cumulative-probability deviations are constrained by a decision-dependent scalar $h(x,\xi)$7, while in Wasserstein DD-DRO the radius $h(x,\xi)$8 governs how far the adversary may move mass away from a reference distribution.

Subsequent work has specialized and expanded these constructions. One Wasserstein-based approach defines a scalar random variable $h(x,\xi)$9, forms its empirical distribution $\mathcal P(x)$0, and centers the ambiguity set at $\mathcal P(x)$1 with radius $\mathcal P(x)$2, where $\mathcal P(x)$3 is the $\mathcal P(x)$4-Lipschitz constant of $\mathcal P(x)$5. In that construction, both the center and the radius depend on the decision, and the ambiguity set lives on the scalar loss space rather than directly on the original uncertainty space (Fonseca et al., 2023).

Contextual and data-driven formulations generalize this further. In a residuals-based model, the uncertain parameter $\mathcal P(x)$6 depends on covariates $\mathcal P(x)$7 and decisions $\mathcal P(x)$8, a regression model $\mathcal P(x)$9 is fitted, empirical residuals $\mathcal P(x)$0 are formed, and the nominal empirical distribution becomes

$\mathcal P(x)$1

Around this nominal model one may build a Wasserstein ball, a sample-robust ambiguity set that perturbs atoms, or a same-support/$\mathcal P(x)$2-divergence set that perturbs probabilities; in each case, the radius may depend on $\mathcal P(x)$3 or only on $\mathcal P(x)$4 (Zhu et al., 2024).

Another data-driven approach starts from offline observations $\mathcal P(x)$5 generated under decision-dependent distributions $\mathcal P(x)$6. It constructs empirical measures $\mathcal P(x)$7 at distinct decision points, interpolates them via Lipschitz weights $\mathcal P(x)$8, and defines a decision-dependent nominal measure

$\mathcal P(x)$9

The ambiguity set is then a Wasserstein ball around $X\subseteq\mathbb R^n$0 with radius $X\subseteq\mathbb R^n$1 determined by interpolation error and sample size (Qu et al., 9 Aug 2025).

Multimodal DD-DRO adds a second layer of dependence by modeling the true distribution as a mixture $X\subseteq\mathbb R^n$2, where both the nominal mode probabilities $X\subseteq\mathbb R^n$3 and the nominal within-mode distributions $X\subseteq\mathbb R^n$4 depend on the first-stage decision $X\subseteq\mathbb R^n$5. Mode-probability ambiguity is captured by a $X\subseteq\mathbb R^n$6-divergence set $X\subseteq\mathbb R^n$7, while each conditional distribution $X\subseteq\mathbb R^n$8 lies in a moment-based or Wasserstein-based ambiguity set $X\subseteq\mathbb R^n$9 (Yu et al., 2024).

3. Reformulations and tractability

For finite-support uncertainty $\xi$0 that is fixed independently of the decision, each distribution $\xi$1 can be identified with a probability vector $\xi$2, $\xi$3, $\xi$4. Under that assumption, the inner maximization admits explicit dualizations whose form depends on the ambiguity set (Luo et al., 2018).

Simple measure and moment bounds yield a linear-program dual. Mean-covariance ambiguity sets produce a conic dual and an SOCP reformulation. Wasserstein ambiguity sets admit an LP-duality reformulation with dual variables coupled to the decision through the radius and the nominal probabilities. $\xi$5-divergence ambiguity leads either to a semi-infinite saddle form or, via KKT stationarity, to a finite nonconvex program. K-S ambiguity sets again reduce to an LP-type dual. Collectively, these reformulations end up as either finite LPs, SDPs, or SOCPs with explicit $\xi$6–dual coupling, or as small nonconvex problems in the $\xi$7-divergence case (Luo et al., 2018).

The practical tractability statement is correspondingly nuanced. If the reformulation is convex in $\xi$8 and $\xi$9 is convex in $(\Xi,\mathcal F)$0, then the DD-DRO problem becomes a convex conic program solvable by off-the-shelf solvers such as MOSEK and CPLEX. If the reformulation is nonconvex, as in explicit KKT forms for $(\Xi,\mathcal F)$1-divergence or when $(\Xi,\mathcal F)$2 is nonconvex in $(\Xi,\mathcal F)$3, one can use global optimization solvers such as BARON or ANTIGONE. For continuous-support Wasserstein or K-S sets, the reformulation becomes a semi-infinite program, and a cutting-surface or exchange algorithm alternates between a master relaxation and a separation problem; under compactness and continuity, this method terminates finitely within $(\Xi,\mathcal F)$4-tolerance (Luo et al., 2018).

A distinct tractability route appears in the decision-dependent Wasserstein formulation on scalar losses. Using Kantorovich duality, the worst-case expectation over a ball $(\Xi,\mathcal F)$5 can be reduced to a finite-dimensional minimax. Two cases receive closed forms. If $(\Xi,\mathcal F)$6 and $(\Xi,\mathcal F)$7 is an interval $(\Xi,\mathcal F)$8, then the DD-DRO objective becomes

$(\Xi,\mathcal F)$9

where $h(x,\xi)$0 and $h(x,\xi)$1. If $h(x,\xi)$2 and $h(x,\xi)$3 is unbounded above, then the objective becomes

$h(x,\xi)$4

Under joint convexity of $h(x,\xi)$5 in $h(x,\xi)$6 and suitable convexity of $h(x,\xi)$7, these are convex reformulations; in particular, when $h(x,\xi)$8 and $h(x,\xi)$9 or $P$0, the DD-DRO problem coincides exactly with standard DRO with a constant-center ball (Fonseca et al., 2023).

This body of reformulation results directly contradicts the common assumption that decision dependence necessarily destroys tractability. The literature instead shows a spectrum: some DD-DRO models remain LP-, SOCP-, or SDP-representable, some reduce to semi-infinite programs addressable by exchange methods, and others require global optimization because the source of difficulty lies in nonconvexity rather than in decision dependence per se.

4. Statistical learning and data-driven calibration

A central issue in DD-DRO is that the true decision-dependent distribution is usually unobservable. Recent work therefore builds ambiguity sets from samples collected under varying decisions and then derives finite-sample or asymptotic guarantees for the resulting robust solutions.

In the residuals-based contextual framework, the theoretical analysis rests on a light-tail condition for the residual $P$1, regression-error control bounds for $P$2, lower semi-continuity of the cost, Lipschitz continuity in $P$3, and vanishing ambiguity radii. Under these assumptions, the Wasserstein radius can be chosen as

$P$4

which yields a finite-sample certificate; the robust value converges to the true value; solution sets converge in probability; and if $P$5, then both the value error and the out-of-sample suboptimality are $P$6. The same framework also gives a finite-sample solution guarantee of exponential form for the distance from the robust optimizer to the true solution set (Zhu et al., 2024).

That framework includes explicit data-driven calibration. An $P$7-fold cross-validation scheme trains the regression model on $P$8, constructs the ambiguity set from residuals, solves the corresponding ER-D$P$9RO problem, and evaluates the resulting decision on the validation fold. Radius forms may be decision-dependent, such as $x$00, and grid search selects the best parameters (Zhu et al., 2024).

Interpolation-based DD-DRO offers a different statistical route. Under compactness of $x$01 and $x$02, Lipschitz continuity of $x$03, Lipschitz continuity of the true map $x$04 in Wasserstein distance, and Lipschitz interpolation weights, the radius

$x$05

ensures high-probability coverage: $x$06 If $x$07 solves DD-DRO, $x$08 is its robust objective value, $x$09 is the true performance at $x$10, and $x$11 is the true optimum, then with probability $x$12,

$x$13

This provides a non-asymptotic out-of-sample guarantee and an optimality-gap bound tied directly to the ambiguity radius (Qu et al., 9 Aug 2025).

A plausible implication of these results is that DD-DRO has moved from a purely structural generalization of classical DRO to a statistically calibrated framework in which the ambiguity set itself can be learned from regression residuals, offline decision-response samples, or interpolated empirical measures, while still preserving formal coverage and performance guarantees.

5. Two-stage adaptivity, information revelation, and multimodality

Two-stage DD-DRO models extend the endogenous-uncertainty perspective from static ambiguity descriptions to adaptive information structures. In decision-dependent information discovery, first-stage measurement decisions $x$14 determine which components of $x$15 become observable before recourse. Starting from a moment-based ambiguity set

$x$16

strong duality converts the problem into an equivalent min–max–min–max robust counterpart. Because the exact problem optimizes over all measurable recourse rules, a $x$17-adaptability approximation is introduced: one selects $x$18 candidate recourse actions here-and-now and implements the best feasible one after the chosen observations are revealed (Jin et al., 2024).

The algorithmic consequence is a nested decomposition. The outer problem minimizes over the binary information-discovery decisions $x$19 using feasibility cuts and integer optimality cuts; the evaluation of a candidate $x$20 is performed by a branch-and-cut method that solves the inner min–max–min problem exactly. The outer scheme terminates finitely because there are finitely many binary $x$21, and the evaluation branch-and-cut also converges finitely because each node’s LP relaxation is strengthened by valid Benders-type cuts and the search tree is finite. Numerically, the best-box problem shows up to $x$22 improvement in worst-case expected return when moving from static $x$23 to adaptive $x$24, and a purely robust solution can be up to $x$25 suboptimal under the worst-case distribution. In the R&D project portfolio problem, the decomposition-plus-cuts algorithm solves many more instances to optimality than direct MINLO, with optimality gaps $x$26 on moderate sizes, while adaptive solutions with $x$27 improve worst-case return by up to $x$28 over static and by roughly $x$29 over purely robust solutions (Jin et al., 2024).

Multimodal DD-DRO addresses a different structural extension. The ambiguity set becomes

$x$30

where $x$31 is a $x$32-divergence neighborhood of the decision-dependent nominal mode probabilities and each $x$33 is a decision-dependent moment-based or Wasserstein-based ambiguity set for mode $x$34. For general $x$35, dualization produces a master problem involving the convex conjugate $x$36; for variation distance, the resulting reformulation is linear; for $x$37-distance, it becomes second-order conic (Yu et al., 2024).

This multimodal formulation is not merely descriptive. Under variation distance and moment-based mode-wise ambiguity, the multimodal ambiguity set is contained in an aggregated single-modal ambiguity set, implying $x$38; a similar containment holds for Wasserstein-based mode-wise ambiguity. Computationally, in an uncapacitated facility-location study with three modes, omission of multimodality and decision-dependent uncertainties led to worse in-sample and out-of-sample performance under various settings. The multimodal decision-dependent model yielded stable out-of-sample costs of approximately $x$39 across robustness levels in the moment-based setting, whereas the single-modal variant degraded as the robustness parameter grew (Yu et al., 2024).

6. Applications, comparative behavior, and open directions

DD-DRO has been instantiated in portfolio optimization, dynamic pricing, shipment planning with pricing, facility location, best-box selection, and R&D project portfolio optimization. These examples are heterogeneous, but they share a common modeling feature: the decision changes either the loss distribution, the nominal model around which robustness is built, the ambiguity radius, the mode structure, or the available information.

In portfolio optimization, the scalar-loss Wasserstein formulation was compared with standard DRO in a mean-risk portfolio problem with a Rockafellar–Uryasev CVaR reformulation. Both methods showed similar portfolio-focused out-of-sample performance and high reliability for large $x$40, but the decision-dependent formulation was far easier to solve, and in comprehensive out-of-sample performance its estimated $x$41 tracked the true VaR much more closely; as sample size increased from $x$42 to $x$43, both methods converged, while the decision-dependent method exhibited better small-sample robustness (Fonseca et al., 2023).

In dynamic pricing with nonstationary demand, the interpolation-based DD-DRO framework specialized the semi-infinite dual into finitely many convex constraints. For Gaussian demand with time-varying mean, $x$44, $x$45, and $x$46, the out-of-sample revenue $x$47 lay within the predicted band $x$48, and as $x$49 grew or $x$50 varied appropriately, $x$51 with small optimality gap. The paper interprets the resulting pricing policies as having guaranteed expected revenue (Qu et al., 9 Aug 2025).

In the residuals-based shipment-planning-and-pricing study, ER-D$x$52RO-W and ER-D$x$53RO-SR reduced out-of-sample cost by up to $x$54–$x$55 relative to ER-DD-SAA when $x$56 was between $x$57 and $x$58, decision-dependent ER-D$x$59RO outperformed decision-independent ER-DRO by several percent, and cross-validation with a radius depending on $x$60 achieved up to $x$61 further gain over a radius depending only on $x$62. Among regressors, the reported ranking was Ridge $x$63 OLS $x$64 Lasso (Zhu et al., 2024).

Several comparative lessons recur across these applications. First, DD-DRO is not a single ambiguity-set recipe; it is a modeling principle that can be instantiated with moments, Wasserstein balls, $x$65-divergence, K-S tolerances, sample-robust perturbations, regression-residual constructions, interpolated empirical measures, and multimodal mixtures. Second, decision dependence does not uniformly enlarge or shrink conservatism; it changes which distributions are deemed plausible at each decision, and in multimodal settings the resulting ambiguity set can be strictly smaller than a single-modal aggregate. Third, DD-DRO is not uniformly harder than classical DRO: some formulations remain LP-, SOCP-, or SDP-representable, and some even collapse to regularized empirical-risk forms or coincide with standard DRO in specific linear cases (Luo et al., 2018).

Open directions stated in the recent literature include multi-stage DRO with decision-dependent information and efficient approximations; handling very large $x$66 in $x$67-adaptability; learning ambiguity-set parameters dynamically from partial observations; nonparametric or machine-learning regression models such as random forests and neural nets; data-efficient radius calibration in the very small-$x$68 regime; and specialized branch-and-cut, Benders, SDDP, proximal, or spatial line-search methods for large-scale DD-DRO (Jin et al., 2024). These questions indicate that DD-DRO has become both a structural extension of DRO for endogenous uncertainty and a computational-statistical research program focused on learning, calibrating, and solving ambiguity sets that vary with the decision itself.