Distributionally Robust Optimization: A Review

1. Introduction.

• DRO addresses ambiguous parameter distributions by optimizing against the worst-case probability distribution in an ambiguity set, extending earlier sensitivity analyses that lacked systematic hedging.The review traces this approach to Scarf’s worst-case newsvendor model with known mean and variance.Source: §1 · pp. 1–2

1.1. A General DRO Model.

• DRO minimizes worst-case uncertainty over an ambiguity set, reducing to stochastic optimization when the true distribution is known and to robust optimization when only support information remains.Choosing the ambiguity set can therefore place DRO between stochastic and robust optimization and make it less conservative than robust optimization.Source: §1.1 · pp. 3–4

1.2. Motivation and Contributions.

• The review unifies operations research and machine learning perspectives, connecting DRO with game theory, risk aversion, chance constraints, robust optimization, and function regularization.It also surveys solution techniques, ambiguity-set constructions and calibration, and uncertainty functionals.Source: §1.2 · pp. 4–5

1.3. Organization of this Paper.

• The section establishes notation and definitions used throughout the paper to keep the presentation self-contained.These preliminaries are introduced before the review’s modeling, theoretical, computational, and application discussions.Source: §2 · p. 5

Measure and probability notation

• The paper distinguishes measurable spaces, probability measures, induced probability distributions, random vectors, integrability, and distributional equivalence using explicit notation.A random vector is written with a tilde, while its realization omits the tilde; induced distributions are distinguished from probability measures.Source: §2 · pp. 5–6

Finite-space representation

• For finite spaces, probability measures and measurable functions are represented as vectors of atom probabilities and corresponding outcomes.With M atoms, q=[q1,…,qM]^T stores event probabilities and z=[z1,…,zM]^T stores function outcomes.Source: §2 · p. 6

Algebraic and convex analysis notation

• The notation defines dual spaces, adjoint mappings, matrix inner products, Hadamard products, conjugates, biconjugates, characteristic functions, and support functions.These constructions provide the algebraic and convex-analytic language for later optimization formulations.Source: §2 · p. 6

Norm and conic notation

• The section introduces essentially bounded function spaces, ℓp norms, simplices, positive parts, cone orders, dual and polar cones, and conic representability.It also defines positive semidefinite ordering, convex hulls, and interiors for subsequent conic and convex formulations.Source: §2 · p. 6

Statistical learning terminology

• For data-driven DRO, out-of-sample performance is the expected test cost under an unknown true distribution, assessed through finite-sample guarantees and asymptotic consistency.A finite-sample certificate provides a (1−α) confidence guarantee, while asymptotic consistency concerns convergence of the certificate and solution to the true problem.Source: §2 · p. 7

3.1. Relationship with Game Theory.

• Under suitable convexity, compactness, semicontinuity, and quasiconvexity conditions, DRO’s minimax problem admits an equivalent maximin formulation with an equilibrium saddle point.The decision maker minimizes worst-case expected cost while nature selects a distribution maximizing the minimum expected cost; exchanging inf and sup preserves the optimal value.Source: §3.1 · pp. 7–8

3.2. Relationship between DRO and RO.

• For finite sample spaces, DRO can be written as robust optimization over an uncertainty set of probability vectors, while retaining statistical modeling tools and allowing continuous sample spaces.Writing f(x, p) := p⊤h(x) yields inf_x∈X sup_p∈P f(x, p), so robust-optimization techniques can specify P.Source: §3.2 · p. 8

3.3.1. Relationship between DRO and Coherent and Law Invariant Risk Measures.

• Under mild conditions, worst-case expectations over convex compact ambiguity sets are equivalent to coherent risk measures, and law invariance transfers between the ambiguity set and the induced risk measure.Conversely, every real-valued coherent risk measure can be represented using a convex compact set of probability measures, making the ambiguity set a consequence of the selected risk preference.Source: §3.3.1 · pp. 10–12

3.3.2. Relationship with Chance-Constrained Optimization.

• DRO connects robust constraints to chance constraints by replacing unknown true distributions with ambiguity sets whose feasible regions conservatively approximate probabilistic feasibility.The robust implication requires feasibility for every ξ in Uϵ to imply probability at least 1−ϵ under P_true; DRO handles distributional ambiguity through scenario and ambiguity-set approaches.Source: §3.3.2 · pp. 12–13
• DRO ambiguity sets provide safe approximations to chance constraints and correspond one-to-one with ambiguity sets satisfying the associated probabilistic guarantee.The guarantee holds with probability at least 1 − α under the sample probability measure.Source: §3.3.2 · p. 14

3.4. Relationship with Function Regularization.

• Optimal-transport DRO connects regularization to protection against data disturbances in SVM, regularized logistic regression, and LASSO.This interpretation is reported in Shafieezadeh-Abadeh et al., Blanchet et al. [50], and Gao and Kleywegt [110].Source: §3.4.1 · p. 14

3.4.1. DRO and Regularization.

• For linear regression, logistic regression, and SVM, optimal-transport DRO can be rewritten as empirical expected cost plus a regularization term.The corresponding losses are square loss, logistic loss, and hinge loss, respectively.Source: §3.4.1 · p. 15
• The transportation-cost shape determines the regularization type, while ambiguity-set size determines the regularization parameter in the equivalent DRO reformulation.This equivalence supports choosing a regularization parameter through the DRO formulation; related results are also noted in Shafieezadeh-Abadeh et al. and Gao and Kleywegt [110].Source: §3.4.1 · p. 15
• φ-divergence DRO yields an expected-cost objective plus a regularization term accounting for the empirical standard deviation of cost.This connection is associated with Duchi et al. [92] and Namkoong and Duchi.Source: §3.4.1 · pp. 14, 16

4. General Solution Techniques to Solve DRO Models.

• DRO reformulations are semi-infinite programs, so solvable reformulations or approximations require assumptions on the ambiguity set and functions h(x, ˜ξ) and g(x, ˜ξ).Source: §4 · p. 16

4.1. Cutting-Surface Method.

• The cutting-surface method iteratively adds probability distributions to finite relaxations until optimality criteria are met, requiring solvable master problems and an ϵ-optimal distribution-separation procedure.Under the theorem’s oracle assumptions, finitely many iterations yield an ϵ-feasible solution whose objective value lower bounds the optimum; the separation subproblem may be nonconvex.Source: §4.1 · p. 17

4.2. Dual Method.

• Dual methods use strong duality to replace the distributional supremum with a dual infimum and are especially useful when the ambiguity set depends on the decision x.Existing DRO work exploits linear, Lagrangian, convex-analytic, and conic duality; strong duality can require a Slater-type condition, such as an interior probability measure.Source: §4.2 · p. 18
• Other SIP techniques, including penalty, smoothing, projection, and primal methods, are less common in DRO, although specialized primal-dual algorithms have been applied to discretized, moment-based, Wasserstein, and divergence-based models.Liu et al. report quantitative convergence for discretized DRO, while related primal methods address χ2-distance, general φ-divergences, moment constraints, and Wasserstein ambiguity sets.Source: §4.2 · p. 19

• Ambiguity sets should be as small as possible while containing the unknown true distribution with certainty or high confidence, reducing conservatism while improving robustness.This motivates separate questions about the set’s shape and size; Section 5 focuses on shape, while Section 6 addresses size.Source: §5 · p. 19
• Researchers commonly choose ambiguity-set shapes using modeler judgment, then determine their size from data, although the two choices may be addressed simultaneously.Section 5.4 reviews approaches that jointly address shape and size.Source: §5 · pp. 19–20
• Ambiguity-set design balances tractable formulations, such as LP, SOCP, or SDP, against expressiveness from descriptive statistics and structural properties of the unknown distribution.Source: §5 · p. 20
• The review groups ambiguity sets into discrepancy-based, moment-based, shape-preserving, and kernel-based families, while noting that these categories can overlap.Moment-based sets constrain distributional moments; discrepancy-based sets remain close to a nominal distribution, shape-preserving sets retain structural properties, and kernel-based sets use kernels near nominal parameters.Source: §5 · pp. 20–21
• Named examples include Markov, Chebyshev, Gauss, median-absolute deviation, Huber, Hoeffding, Bernstein, Choquet, and mixture ambiguity sets.The paper reviews these four groups in Sections 5.1–5.4 and general ambiguity-set treatments in Section 5.5.Source: §5 · pp. 20–21

5.1. Discrepancy-Based Ambiguity Sets.

• The ambiguity set contains distributions within discrepancy ϵ of nominal measure P0, where the discrepancy determines its shape and ϵ represents robustness or belief in P0.The review distinguishes optimal transport, φ-divergence, total variation, goodness-of-fit, Prohorov, ℓp-norm, ζ-structure, Levy, and contamination discrepancies, which need not all be metrics.Source: §5.1 · p. 21

5.1.1. Optimal Transport Discrepancy.

• Optimal transport DRO admits strong duality under mild semicontinuity and integrability conditions, including general Polish spaces and lower semicontinuous transportation costs.The dual uses a nonnegative multiplier and an explicit worst-case transformation of the loss; the robustness radius enters a univariate reformulation linearly. These results also connect DRO with regularized logistic regression and LASSO [50] [48] [47] [112] [274].Source: §5.1.1 · p. 23
• Wasserstein DRO yields strong duality, finite-dimensional convex reformulations, extremal distributions, and consistency guarantees for data-driven problems under norms on Rn or metrics on Polish spaces [110].Mohajerin Esfahani and Kuhn apply these results to portfolio optimization and uncertainty quantification, while Gao and Kleywegt [110] show data-driven DRO can be approximated by robust optimization.Source: §5.1.1 · pp. 23–24

5.1.1.1. Discrete Problems.

• Wasserstein ambiguity sets support tractable discrete DRO formulations, including semi-infinite, copositive, semidefinite-approximated, and linear programs for stochastic and mixed-integer models.Hanasusanto and Kuhn obtain copositive formulations for 2-Wasserstein models and linear programs for certain 1-Wasserstein models, while Xu and Burer derive copositive reformulations with semidefinite approximations.Source: §5.1.1 · p. 24, §5.1.1.1 · p. 27

5.1.1.2. Chance Constraints.

• Distributionally robust chance constraints over Wasserstein sets admit exact mixed-integer conic reformulations in broad affine settings, while CVaR and Bonferroni approximations can be inferior to the optimal reformulation [76].For ℓ1-norm or ℓ∞-norm transportation costs, the reformulation becomes a mixed-integer linear program; random-technology constraints receive a relaxation that is exact for binary or bounded integer decisions.Source: §5.1.1.2 · p. 27

5.1.1.3. Statistical Learning.

• In statistical learning, Wasserstein DRO has been developed for covariance estimation, support vector machines, domain adaptation, hypothesis testing, and regularization.The transportation-cost shape determines the regularizer, while ambiguity-set size determines its parameter; for affine losses, worst-case risk measures can equal nominal risk plus a regularization term.Source: §5.1.1.3 · pp. 27–28
• Shafieezadeh-Abadeh et al. [274] establish Wasserstein-DRO generalization bounds for regression and classification under dispersion or separability conditions, avoiding hypothesis-space complexity measures.Bounding the decision space improves prior out-of-sample confidence bounds, and the framework extends to nonlinear hypotheses through lifted RKHS formulations.Source: §5.1.1.3 · p. 29
• Wasserstein-based DRO in statistical learning often yields tractable convex reformulations and can be interpreted as regularization of empirical risk minimization.These results cover logistic regression, inverse optimization, regression, and broader cost functions, with regularization penalties linked to Wasserstein geometry.Source: §5.1.1.3 · pp. 29–30
• Wasserstein-DRO supports specialized learning methods including group-square-root LASSO, semi-supervised learning, and outlier detection, with unlabeled data restricting worst-case support.Chen and Paschalidis report higher performance than M-estimation for abnormal radiation exposure detection, while Blanchet and Kang select robustness by cross-validation.Source: §5.1.1.3 · pp. 29–30
• Metric-learning and robust transportation-cost design focus Wasserstein ambiguity on performance-relevant regions, improving generalization and reducing variability in regularized estimators.Robustness levels are selected by cross-validation; robust cost design adds an additional robustification layer beyond metric learning.Source: §5.1.1.4 · p. 30

5.1.1.5. Multistage Setting.

• Multistage stochastic programs use Wasserstein-based nested distance to vary scenario probabilities on fixed trees, enabling applications to production control and contingent-claim acceptability pricing.Glanzer et al. show data-built scenario trees converge in nested distance to the true model in probability at an exponential rate.Source: §5.1.1.5 · p. 30
• φ-divergences provide another information-theoretic framework for modeling distributional ambiguity by measuring discrepancy between probability measures.The review also tabulates example φ-divergences, conjugate functions, and corresponding DRO formulations.Source: §5.1.2 · p. 31, Table 1 · p. 31

5.1.2. φ-Divergences.

• φ-divergence ambiguity sets are parameterized around a nominal measure P0 by ϵ, and robust counterparts are tractable for most commonly used choices in finite-dimensional settings.Table 1 lists common φ-divergence functions and their conjugates; φ(1)=0 is without loss of generality.Source: §5.1.2 · pp. 31–32
• Theorem 5.3 characterizes worst-case expected costs over φ-divergence ambiguity sets through a Lagrangian-dual formulation.The result is obtained by taking the Lagrangian dual of the worst-case expectation problem.Source: §5.1.2 · p. 31
• φ-divergence DRO chance constraints can be transformed into nominal-distribution chance constraints with adjusted risk levels, while ambiguous probability minimization retains the complexity of its nominal counterpart.For general divergences, the adjusted risk level can be found by bisection; closed forms are available for variation distance, Kullback-Leibler divergence, and χ2-distance.Source: §5.1.2 · pp. 32–33

5.1.2.1. Statistical Learning.

• φ-divergence DRO in statistical learning can produce classifiers optimal for the training distribution rather than robust over the ambiguity set, motivating ambiguity sets that preserve conditional structure given latent variables.Hu et al. attribute this pessimism to the classification losses and over-conservatism of φ-divergence ambiguity sets.Source: §5.1.2.1 · p. 33

5.1.2.2. Specific φ-Divergences.

• Specific φ-divergence models span portfolio, extreme-value, Markov-decision, inventory, dynamic-programming, multistage, and chance-constrained applications, with tractability or statistical guarantees established for several choices.Examples include Burg entropy, χ2-distance, modified χ2-distance, Kullback-Leibler divergence, Rényi divergences, and variation distance.Source: §5.1.2.2 · pp. 33–34

5.1.3. Total Variation Distance.

• Total variation distance is a φ-divergence, half the ℓ1-norm, and an optimal-transport discrepancy.Its φ function is φ(t) = 1/2|t−1|.Source: §5.1.3 · p. 35
• Total variation ambiguity sets are centered on a nominal probability measure P0, with ϵ controlling their size.The section presents a duality result for worst-case expectations over these sets.Source: §5.1.3 · p. 35
• One-sided variation distances are φ-divergences but, unlike total variation distance, are not probability metrics.Their φ functions are φ(t) = 1/2(t−1)+ and φ(t) = 1/2(1−t)+, respectively.Source: §5.1.3 · p. 35
• Jiang and Guan study two-stage stochastic programs, while Rahimian et al. analyze effective scenarios and newsvendor solutions under total variation ambiguity.Rahimian et al. provide conditions for effective and ineffective scenarios; Rahimian et al. derive a closed-form newsvendor solution and maximal effective demand subsets.Source: §5.1.3 · p. 36
• Postek et al. review tractable risk-constraint reformulations, while Bertsimas et al. study goodness-of-fit tests for constructing ambiguity sets.Applications include portfolio optimization and antenna array design.Source: §5.1.4 · p. 36

• The analysis considers independent components with continuous support for ˜ξ, asynchronous samples from continuous marginal distributions, and joint-distribution samples from Ptrue.Source: §5.1.4 · p. 36
• It examines χ2, G, Kolmogrov-Smirnov, Kuiper, Cramer-von Mises, Watson, and Anderson-Darling goodness-of-fit tests.Source: §5.1.4 · p. 36
• The corresponding statistical hypothesis tests are used to characterize the geometric shape of the resulting ambiguity sets.Source: §5.1.4 · p. 36

5.1.5. Prohorov Metric.

• The Prohorov metric compares probability measures through enlarged measurable sets and takes values in [0, 1].Its ambiguity set contains distributions within radius ϵ of nominal measure P0.Source: §5.1.5 · p. 36
• For univariate distributions, the Prohorov metric specializes to the Lévy metric, which bounds distribution-function deviations over shifted thresholds.The radius ϵ again controls the resulting distributional ambiguity around P0.Source: §5.1.5 · p. 36
• Under Prohorov ambiguity, Erdo˘gan and Iyengar analyze parameterized convex constraints and bound samples needed for high-probability scenario approximation of ambiguous chance constraints.Their approach extends prior work.Source: §5.1.5 · p. 37
• Other Prohorov-related DRO studies address radially symmetric chance constraints, polynomial optimization with SDP relaxations, and multistage stochastic programs using expectation–CVaR objectives [158].These works include consistency, asymptotic, convergence, and application analyses for water networks and hydrothermal scheduling.Source: §5.1.5 · p. 37

5.1.7. ζ-Structure Metrics.

• ζ-structure metrics encompass total variation, Kantorovich, Fortet-Mourier, and uniform (Kolmogorov) metrics, with Fortet-Mourier reducing to Kantorovich when q = 1.They use Lipschitz-type function restrictions and can define ambiguity sets around nominal measure P0 with radius ϵ.Source: §5.1.7 · pp. 37–38
• Lemma 5.7 states that bounded support with diameter θ yields explicit relationships among ζ-structure metrics.The lemma assumes θ := sup{d(ξ1, ξ2) : ξ1, ξ2 ∈Ω} is finite.Source: §5.1.7 · p. 38
• Zhao and Guan construct and solve ζ-structure ambiguity sets for two-stage stochastic programs, analyze convergence, and test them on newsvendor and facility-location problems.Their framework covers discrete and continuous true distributions.Source: §5.1.7 · p. 38
• Pichler and Xu study how varying a ζ-structure ambiguity set changes optimal values and solutions in expectation-based two-stage recourse models.They illustrate the effects in a two-stage stochastic program with recourse.Source: §5.1.7 · p. 38

5.1.8. Contamination

• Contamination neighborhoods model distributions as P = (1 − ϵ)P0 + ϵQ, with Q ⊆ M(Ξ, F) and ϵ ∈ [0, 1].The nominal distribution P0 is perturbed by an ϵ-weighted distribution Q.Source: §5.1.8 · p. 39
• Contamination neighborhoods are used in robust statistics, economics, statistical learning, and hypothesis testing, including work by Huber, Huber and Ronchetti, Nishimura and Ozaki, Duchi et al., and Huber.Bose and Daripa apply a maximin expected utility model of Gilboa and Schmeidler to ambiguity aversion in mechanism design.Source: §5.1.8 · p. 39

5.1.9. General Discrepancy-Based Ambiguity Sets.

• General discrepancy-based models cover risk measures ranging from expectation and variance to VaR, CVaR, entropic VaR, coherent, and spectral risk measures.Postek et al. survey these combinations for discrete distributions and derive tractable reformulations using conjugate duality and ambiguity-set support functions.Source: §5.1.9 · p. 39
• The surveyed discrepancy-based ambiguity sets include φ-divergences, Kolmogorov-Smirnov, Wasserstein, Anderson-Darling, Cramer-von Mises, Watson, and Kuiper models.Postek et al. exemplify the reformulations in portfolio optimization and antenna array design.Source: §5.1.9 · p. 39
• Shapiro connects DRO models based on Wasserstein metrics, φ-divergences, and total variation distance to law-invariant risk measures.The work also proposes an SAA method for the corresponding dual problems and establishes statistical properties of optimal solutions and values comparable to risk-neutral stochastic programs, as in Shapiro et al. and Shapiro.Source: §5.1.9 · p. 39
• Moment-based ambiguity sets contain all probability distributions whose moments satisfy specified properties and are categorized into overlapping subgroups.The paper introduces this as a common modeling approach for ambiguity sets.Source: §5.2 · p. 39

5.2.1. Chebyshev.

• Chebyshev ambiguity sets recover Scarf’s closed-form newsvendor solution, whose worst-case demand distribution has two-point support, and motivate extensions with recourse, fixed ordering costs, random output, and scarce resources.Source: §5.2.1 · pp. 39–40
• Portfolio DRO models with uncertain means and covariances admit SOCP or SDP reformulations under polytopic, interval, ellipsoidal, or componentwise uncertainty, including worst-case VaR and CVaR problems [189].Goldfarb and Iyengar’s factor-model portfolio formulations, including variance, expected-return, and VaR-constrained classes, are reformulated as SOCPs; dependent mean-covariance uncertainty yields an SDP upper bound on worst-case VaR.Source: §5.2.1 · pp. 40–41

5.2.2. Delage and Ye.

• Delage–Ye ambiguity sets constrain the mean and second central-moment matrix, support data-driven confidence regions, and provide probabilistic out-of-sample upper-bound guarantees; for ellipsoidal supports, the reformulation reduces to finite-size SDPs.Under convexity, compactness, oracle, and piecewise convex-concave assumptions, the reformulation is polynomial-time solvable. PCA-based approximation yields a relaxation with theoretical bounds on the gap.Source: §5.2.2 · pp. 42–43

5.2.2.1. Discrete Problems.

• Discrete DRO models produce tractable formulations including SDP upper bounds for mixed-integer problems, mixed-integer SOCPs for two-stage integer recourse, and binary SOC constraints for chance constraints [211, 323, 340].Project-crashing DRO instead exploits first-two-moment structure through a convex-concave saddle-point formulation solvable in polynomial time.Source: §5.2.2.1 · pp. 43–44

5.2.2.2. Risk and Chance Constraints.

• Worst-case chance constraints are SOC-representable under mean-covariance information, while worst-case CVaR gives exact SDP reformulations for several individual and joint constraint classes.For joint chance constraints in a hydro-power reservoir-control problem, the CVaR approximation outperforms the Bonferroni approximation and Chen et al.’s method.Source: §5.2.2.2 · p. 45

5.2.2.3. Statistical Learning.

• In statistical learning, worst-case misclassification probability with moment ambiguity becomes an SOCP for linear hypotheses, including regularization under covariance norm-ball uncertainty, and remains an SOCP after kernelizing nonlinear hypotheses.Source: §5.2.2.3 · p. 46

5.2.2.4. Multistage Setting.

• Multistage moment-based DRO studies characterize time consistency through rectangularity or martingale demand processes, with closed-form optimal policies and two-point worst-case distributions in the latter setting [326] [325].For martingale demand, one worst-case support point is zero, and at every stage demand is either at least inventory or zero, making all future demands zero in the latter case [325].Source: §5.2.2.4 · p. 46

5.2.3. Generalized Moment and Measure Inequalities.

• Generalized moment-and-measure ambiguity sets bound probability measures and measurable functionals, encompassing moments, deviations, covariances, higher moments, Huber loss, unimodality, and symmetry.They support dual reformulations and convergence as increasingly many reference-matched moments are imposed, while also accommodating reference-measure risk models and decision-dependent ambiguity sets [291, 238, 33, 202, 265].Source: §5.2.3 · pp. 46–47
• Generalized moment models support discrete stochastic and integer programs, risk and chance constraints, and statistical applications through mixed-integer, semidefinite, conic, or approximation-based reformulations.Examples include exact newsvendor approximations, decomposition algorithms with finite convergence, semidefinite bounds for polynomial moment information, reward-risk portfolio models, and conic regression.Source: §5.2.3 · p. 48, §5.2.3.1 · p. 48, §5.2.3.2 · p. 49

5.2.3.2. Risk and Chance Constraints.

• A Bernstein approximation provides a conservative, efficiently solvable convex program for ambiguous joint chance constraints under independence, finite moment-generating functions, convexity, and a convex ambiguity set.Every feasible solution of the approximation is feasible for the original chance-constrained problem; related conic representations apply to pessimistic and optimistic models under additional structural conditions.Source: §5.2.3.2 · pp. 49–50

5.2.3.3. Statistical Learning.

• In statistical learning, DRO graphical models preserve probabilistic graphical-model consistency while allowing customized loss metrics through first-order feature-moment matching.Fathony et al. [104] optimize worst-case conditional prediction loss over distributions matching empirical feature moments.Source: §5.2.3.3 · p. 50

5.2.4. Moment Matrix Inequalities.

• Moment-matrix inequalities generalize scalar moment constraints and, under a Slater-type condition, yield strong-duality-based conic reformulations.Xu et al. develop exchange and dual cutting-plane schemes with convergence of optimal values and solutions under compact-support or boundedness conditions.Source: §5.2.5 · p. 52, §5.2.4 · p. 52

5.2.5. Cross-Moment or Nested Moment.

• Conic-representable ambiguity sets with auxiliary random vectors unify support, moment, divergence, robust-statistics, and higher-order information, but exclude infinitely many restrictions such as symmetry or independence.Wiesemann et al. show that many existing ambiguity sets arise as projections of their lifted framework, while strict nesting supports tractable formulations and violations motivate conservative outer approximations.Source: §5.2.5 · pp. 53–54

5.2.5.2. Multistage Setting..

• Lifted and nested-moment formulations extend to adaptive and multistage DRO, where recourse can be approximated or transformed into robust optimization models with tractable conic structure.Bertsimas et al. use generalized linear decision rules and lifting to obtain second-order-conic robust formulations, while related work covers nested uncertainty and marginal-information models.Source: §5.2.5 · p. 55, §5.2.5.2 · p. 56, §5.2.6.1 · p. 56

5.2.6.2. Risk and Chance Constraints.

• Moment and marginal information supports tractable DRO reformulations for CVaR, chance constraints, and combinatorial optimization, including LP, mixed-integer LP, mixed-integer SOCP, and mixed-binary formulations.Dhara et al. [89] and Doan et al. derive convex or linear reformulations for worst-case CVaR; Natarajan et al. [212] obtain polynomial-sized mixed-integer formulations; Zhang et al. use a three-point worst-case distribution to derive a closed-form chance constraint and mixed-binary program.Source: §5.2.6.2 · pp. 56–57

5.2.6.3. Statistical Learning.

• Matching empirical covariate marginals while constraining response–covariate cross-moments makes DRO equivalent to regularized maximum likelihood for several generalized linear-model losses.Farnia and Tse [103] recover regularized maximum likelihood for linear regression with quadratic loss, logistic regression with logarithmic loss, and SVM with 0-1 loss.Source: §5.2.6.3 · p. 57
• DRO classification methods based on nominal pairwise marginals achieve 2-approximation upper bounds for worst-case loss or misclassification probability.Eban et al. derive the bound for a classifier using zero-one hinge loss, while Razaviyayn et al. obtain it for a randomized classifier after changing the infimum-supremum order and smoothing the objective.Source: §5.2.6.3 · pp. 57–58

5.2.7. Mixture Distribution.

• Mixture ambiguity sets model distributions as convex combinations of parameterized component distributions, and polynomial SDP approximations can converge to ambiguous chance-constraint feasibility sets.Lasserre and Weisser [181] establish convergence as the degree of the approximating polynomials increases.Source: §5.2.7.1 · p. 58

5.2.7.1. Risk and Chance Constraints.

• Mixture-based risk models produce LP, SOCP, and SDP reformulations for portfolio and newsvendor problems, while shape-preserving models require shared structural properties across ambiguity-set distributions.Kapsos et al. use box or ellipsoidal uncertainty for mixture probabilities; Hanasusanto et al. obtain exact and conservative SDP formulations, including under χ2-distance ambiguity in mixture weights.Source: §5.2.7.1 · pp. 58–59, §5.3 · p. 59

Shape-preserving ambiguity sets

• Popescu obtains generalized Chebyshev bounds for structurally constrained distributions using Choquet representations and semidefinite programs.The framework covers symmetric, unimodal, multimodal, and convex or concave density classes, with worst-case measures represented by at most m + 1 extremal measures when attained.Source: §5.3 · p. 59
• Van Parys et al. extend these results to α-unimodal multivariate distributions on polytopic supports, obtaining generalized Gauss bounds through semidefinite representations.Their results include α-multimodal distributions and recover classical generalized Chebyshev and Gauss bounds as special cases.Source: §5.3 · p. 59

Canonical structural ambiguity sets

• Hanasusanto et al. unify moment and structural ambiguity sets, enabling conic formulations for distributionally robust uncertainty quantification and chance constraints.Their conservative chance-constraint approximation alternates worst-case probability evaluation for a fixed decision with decision improvement, while recovering generalized Chebyshev and Gauss inequalities.Source: §5.3 · pp. 59–60

Risk-constrained stochastic programs

• Li et al. formulate chance- and CVaR-constrained programs over α-unimodal distributions as infinite SOC systems and propose separation, conservative, and relaxation approximations.The safe region is bi-affine in uncertainty and decisions, and the approximations replace the infinite constraints with finitely many constraints.Source: §5.3 · p. 60

Functionally robust decisions

• Hu et al. hedge misspecified price-demand functions by requiring decreasing convex or concave demand, reducing the minimax newsvendor model to a univariate pricing problem.A two-sided cutting-surface algorithm generates function cuts to shrink admissible functions, while ambiguity-set size exposes the pricing and ordering risk-reward trade-off.Source: §5.3 · p. 60

Tail-shape ambiguity

• Lam and Mottet compute worst-case bounds under tail convexity, showing that extremal convex-tail behavior can be either extremely light-tailed or extremely heavy-tailed.The approach addresses the difficulty of evaluating extremal performance when data are insufficient.Source: §5.3 · p. 60

• Kernel-weighted prescriptions learn conditional decisions from covariates and uncertain observations, extending conditional stochastic optimization beyond direct SAA.Bertsimas and Kallus [32] use locally learned weights motivated by nearest-neighbors, Nadaraya-Watson, local linear regression, trees, and random forests; under mild conditions, the resulting problem is polynomially solvable, asymptotically optimal, and consistent, whereas direct SAA can be suboptimal and inconsistent.Source: §5.4 · p. 61
• Kernel density estimation combined with principal component analysis constructs uncertainty sets by capturing correlations among uncertain parameters and distributional information along uncorrelated principal components.Source: §5.4 · p. 62
• SVC with a piecewise-linear weighted generalized intersection kernel produces a positive-definite kernel matrix and a polyhedral uncertainty set whose robust counterpart has deterministic-problem complexity.Common polynomial, radial-basis, and sigmoid kernels can yield intractable robust counterparts, motivating the piecewise-linear construction proposed by Shang et al. [282].Source: §5.4 · p. 64
• The SVC regularization parameter γ bounds the outlier fraction, makes robust feasibility imply feasibility for a SAA-based chance constraint, and determines the asymptotic outlier fraction.As the number of data points increases, the outlier fraction converges to γ with probability one.Source: §5.4 · p. 64

Unified ambiguity sets

• Chen et al. propose a scenario-wise ambiguity set that unifies generalized moments, mixture distributions, Wasserstein metrics, φ-divergences, and k-means clustering, with second-order conic representations and affine recourse approximations.The framework introduces a discrete scenario variable and supports adaptive robust optimization models unifying stochastic programming and distributionally robust optimization with recourse.Source: §5.5 · pp. 64–65

Decision-dependent ambiguity

• Luo and Mehrotra [199] allow ambiguity sets to depend on decisions and derive duality-based equivalent reformulations for moment-, discrepancy-, Wasserstein-, φ-divergence-, and Kolmogorov-Smirnov-based sets.Their considered sets include measure and moment inequalities, bounded moments, and 1-Wasserstein distance with the ℓ1-norm.Source: §5.5 · p. 65

Algorithms and convergence

• Pflug and Wozabal use an exchange method to generate finite inner approximations when ambiguity affects both objectives and constraints, proving finite convergence for compact convex sets and jointly continuous risk measures.The convergence conditions require compactness and convexity of the ambiguity set and joint continuity in x and P.Source: §5.5 · p. 65
• Bansal and Mehrotra develop decomposition algorithms for two-stage distributionally robust disjunctive programs that finitely converge when the distribution-separation subproblem is finitely solvable, including for PMM and Wasserstein or total-variation sets.The methods combine Balas’ linear-programming equivalent or sequential convexification within an L-shaped framework and generalize the distributionally robust integer L-shaped algorithm of Bansal et al..Source: §5.5 · pp. 65–66

Chance-constrained and integer models

• Wang et al. formulate distributionally robust chance-constrained bin packing as binary bilinear knapsack constraints and solve it with lifted-cover inequalities, probability cuts, and branch-and-cut.They establish finite convergence when the distribution-separation subproblem is finitely solvable and apply the approach to operating-room scheduling.Source: §5.5 · p. 66

Extensions and applications

• Other studies bound policy randomization value, approximate entropic-VaR binary programs, analyze ambiguity-set convergence, and extend methods to mixed-integer, multistage, and stochastic dynamic programs.Delage and Saif show the randomization bound is tight for expected risk with affine costs; Guo et al. study convergence under Kolmogorov and total-variation discrepancies; Long and Qi use sequential approximation; Bansal and Zhang derive convex equivalents; and Shapiro et al. modify SDDP.Source: §5.5 · pp. 65–66

6.1. Choice of the Nominal Parameters.

• Nominal distributions are typically estimated by empirical distributions for discrete data, kernel density estimation for continuous data, and empirical mean/covariance estimates for moment-based sets.These choices provide the inputs around which discrepancy- or moment-based ambiguity sets are constructed.Source: §6.1 · p. 67

6.2. Choice of Robustness Parameters.

• Robustness parameters control ambiguity-set conservatism and are generally chosen inversely with sample size, using convergence arguments and probabilistic guarantees under i.i.d. sampling.Such guarantees can make the ambiguity set a confidence set for the true distribution and the DRO objective an upper bound on out-of-sample performance with high probability.Source: §6.2 · p. 67, §6.2.1 · p. 67

6.2.1.1. Optimal Transport Discrepancy.

• Wasserstein ambiguity-set radii can be theoretically conservative because probabilistic bounds rely on an exogenous constant C, whereas Blanchet et al. [50] use an RWP quantile for optimization-aware calibration.For confidence level α, choosing ε as the (1−α)-quantile of RWP at the true-distribution optimal solution makes that solution belong to the DRO optimal-solution set with (1−α) confidence.Source: §6.2.1.1 · pp. 68–69

6.2.1.2. Goodness-of-Fit Test / 6.2.1.3. φ-Divergences.

• Goodness-of-fit tests and φ-divergence asymptotics calibrate ambiguity levels as confidence thresholds, yielding approximate or asymptotically exact confidence coverage under stated regularity conditions.For φ-divergences, Duchi et al. [92] provide an upper confidence interval with asymptotic exact coverage, while variation distance requires a Hellinger-based workaround because its φ-function is not twice differentiable at 1.Source: §6.2.1.3 · pp. 69–70

6.2.1.5. ζ-Structure.

• For ζ-structure metrics, discrete-support robustness levels depend on support diameter, sample size, and confidence level, whereas continuous-case prescriptions can remain practically conservative.Theorem 6.3 supplies high-probability coverage for Kantorovich and bounded-Lipschitz metrics.Source: §6.2.1.5 · p. 71

6.2.1.6. Chebyshev / 6.2.1.7. Delage and Ye / 6.2.2. Non-Data-Driven DROs.

• Chebyshev and Delage–Ye calibrations construct uncertainty sets from regression confidence regions or estimated moments, while non-data-driven methods balance optimism–pessimism or nominal–worst-case regret when i.i.d. calibration is unsuitable.Goldfarb and Iyengar use least-squares estimates, projected confidence regions, and bootstrap intervals; Rahimian et al. define the smallest robustness levels achieving the respective balances.Source: §6.2.1.6 · p. 72, §6.2.1.7 · p. 72, §6.2.2 · p. 73

7.1. Regret Function.

• DRO regret functions quantify expected additional loss or disappointment relative to possible future outcomes, with applications in combinatorial optimization, multicriteria decision-making, and newsvendor models.Perakis and Roels derive closed-form single-item newsvendor solutions under support, moment, median, symmetry, and unimodality information.Source: §7.1 · pp. 73–74

7.2. Risk Measure.

• DRO models explicitly incorporate risk measures, including spectral, distortion, variance, mean absolute-deviation, optimized certainty equivalent, CVaR, and other risk measures.Risk measures quantify uncertainty, typically impose aversion relative to expectation, and induce preference orderings over random variables.Source: §7.2 · p. 74

7.3. Utility Function.

• Expected utility compares random variables through expected utility, while stochastic dominance addresses ambiguity in the decision maker’s utility function or an equivalent target distribution.A bounded utility can be normalized between zero and one and interpreted as a cumulative distribution function.Source: §7.3 · p. 74
• Hu and Mehrotra formulate utility ambiguity with concave, piecewise-linear bounds and solve discretized approximations, while Hu et al. use SAA for mixed-integer LPs without shape restrictions.Applications include portfolio optimization and streaming bandwidth allocation; convergence properties are analyzed for both approximation approaches.Source: §7.3 · pp. 74–75
• DRO extends utility-based comparison to random vectors through multivariate utilities and weighted scalarization, including stochastic weights and reformulations for polyhedral or conic-representable weight sets.Hu and Mehrotra illustrate stochastic weighting in disaster planning and agriculture revenue management, including dependent weights and random parameters.Source: §7.3 · pp. 75–76

8. Modeling Toolboxes.

• ROME, ROC, XProg, and AROMA provide algebraic modeling tools for DRO, supporting ambiguity-set declarations, reformulations, decision rules, and practical applications in inventory, projects, portfolios, and other settings.ROME targets conic-representable support and mean sets with known covariance and directional-deviation bounds; ROC emphasizes adaptive DRO scalability.Source: §8 · p. 76