Norm- and Moment-Constrained Ambiguity Sets

• Norm- and moment-constrained ambiguity sets are frameworks in distributionally robust optimization that restrict feasible distributions via moment bounds and norm metrics.
• They enable tractable reformulations into MILP, SOCP, and SDP models, balancing conservatism with computational efficiency in applications like chance-constrained and multistage programming.
• Practical studies show these sets effectively manage uncertainty, offering reliable performance in robust control and risk minimization while addressing tradeoffs between tightness and scalability.

Norm- and moment-constrained ambiguity sets are foundational structures in distributionally robust optimization (DRO) that capture uncertainty about underlying probability distributions through constraints on norms, moments, or deviations in functional geometry. These sets formally encode allowed distributions by restricting either their moments (means, (co)variances), or their proximity to a reference measure as determined by a norm, such as Wasserstein distance or @@@@1@@@@ (RKHS) norm. The resulting DRO models facilitate tractable reformulations of worst-case expected values, regret, or constraint satisfaction under distributional ambiguity, especially in settings such as chance-constrained programming, multistage robust optimization, and robust control, where only incomplete or empirical distributional information is available.

1. Formal Definitions of Ambiguity Sets

Moment-based ambiguity sets restrict feasible probability distributions via bounds on their mean vector and covariance matrix, typically around empirical estimates. For example, a standard construction is: $$\mathcal{P}_{mom} = \left\{ f\,\middle|\, (E_f[\widetilde{P}_{PV}] - \mu)^\top \Sigma^{-1} (E_f[\widetilde{P}_{PV}] - \mu) \le \gamma_1,~~ E_f[(\widetilde{P}_{PV} - \mu)(\widetilde{P}_{PV} - \mu)^\top] \preceq \gamma_2 \Sigma \right\}$$ where $\mu$ and $\Sigma$ are empirical mean/covariance and $\gamma_1, \gamma_2$ parameterize the permissible deviation (Zhang et al., 2021). Extensions support decision-dependent bounds in multistage settings, leading to sets defined by per-scenario affine moment conditions and probability bounds (Yu et al., 2020), or ellipsoidal (Mahalanobis) constraints on the mean coupled with positive semidefinite cone inclusions for the covariance.

Norm-constrained ambiguity sets typically bound the Wasserstein, total variation, or RKHS (via MMD) distance between candidate distributions and an empirical reference. The Wasserstein ball construction is: $$\mathcal{P}_{Wass} = \left\{ f\,\mid\, W_p(f, P_0) \le \delta \right\}$$ where $P_0$ is the empirical measure and $\delta > 0$ is the ball radius (Zhang et al., 2021). Similarly, a kernel mean embedding defines the ambiguity set as an MMD-ball: $$A(\delta) = \left\{ P \,|\, \|\mu_P - \mu_{\hat{P}}\|_{\mathcal{H}} \le \delta \right\}$$ where $\mathcal{H}$ is the RKHS, and $\mu_P$ is the kernel mean embedding of $P$ (Zhu et al., 2020).

2. Applications in Chance-Constrained and Multistage Programs

Norm- and moment-constrained ambiguity sets play a critical role in distributionally robust chance-constrained (DRCC) programming, often formulated as: $$\inf_{f \in \mathcal{P}} \mathbb{P}_f \left[ \sum_\ell P_\ell u_\ell - \widetilde{P}_{PV} \ge 0 \right] \ge 1 - \alpha$$ where $\mathcal{P}$ may be either a moment-based or norm-constrained set (Zhang et al., 2021).

In multistage settings with endogenous uncertainty, ambiguity sets are constructed to be decision-dependent, allowing moment conditions and support to adjust based on prior-stage decisions. The general forms include:

• Type 1: Separate bounds on individual moments and scenario probabilities.
• Type 2: Exact moment-matching (mean and covariance).
• Type 3: Delage–Ye ellipsoidal sets combining Mahalanobis-norm bounds on the mean and PSD-cone constraints on covariance (Yu et al., 2020).

These structures provide rigorous means for propagating uncertainty and maintaining performance guarantees under dynamically-evolving information.

3. Tractable Reformulations: MILP, SOCP, and SDP

DRO models with these ambiguity sets are reformulated into tractable optimization problems using duality and convexification.

For moment-based sets, DRCC constraints admit reformulations as single linear inequalities: $$\sum_\ell P_\ell u_\ell \ge \theta + \Omega(\alpha, \gamma_1, \gamma_2) \cdot \sigma$$ where $\theta$ and $\sigma$ are derived from empirical moments, and $\Omega$ is a safety factor depending on conservatism parameters (Zhang et al., 2021). In adjustable-risk variants, the nonlinear dependence on $\alpha$ necessitates SOCP or SDP reformulations.

Norm-constrained (Wasserstein) sets yield MILP reformulations via CVaR duality, with variants:

• MILP1: Full primal MILP with $O(N)$ binaries (N = sample size).
• MILP2: Compact dual formulation using only $O(\alpha N)$ binaries (Zhang et al., 2021).

In control contexts, moment-constrained ambiguity sets admit convex program reformulations. For finite-horizon linear-quadratic regret-optimal control, the minimax DRO problem becomes the regularized objective: $$\min_{K}~ \operatorname{Tr}[\Sigma_0 C(K)] + r_1 \|C(K)\|_\infty + r_2 \|C(K)\|_q$$ where $C(K)$ encodes control policy sensitivity, and regularization terms enforce robustness to moment uncertainty. Standard SDP techniques and scalable dual projected subgradient methods enable computation for large-scale systems (Taha et al., 11 Dec 2025).

4. Kernel Mean Embedding and Generalized Moment Problems

Kernel mean embedding via MMD defines ambiguity sets in infinite or high-dimensional moment spaces, controlling simultaneously all empirical moments up to a chosen kernel order. This yields the generalized moment program: $$\sup_{P: \|\mu_P - \mu_{\hat{P}}\|_{\mathcal{H}} \le \delta} \mathbb{E}_P[f(X)]$$ The dual is a semi-infinite program (SIP) imposing conic constraints in the RKHS. Finite-dimensional tractable approximations using atomic measures and Gram matrices transform the problem into a convex quadratic program or SOCP in the atom weights (Zhu et al., 2020).

This approach achieves monotonic lower bounds converging to the true worst-case, with explicit convergence and error guarantees linked to sampling density (fill distance).

5. Tradeoffs: Conservatism, Tightness, and Computational Complexity

Key tradeoffs arise between tractability and conservatism:

• Moment-based sets (e.g., $\mathcal{P}_{mom}$): yield very tractable MILPs (or SOCPs for adjustable risk), but can be over-conservative if bounds are loose or the true distribution is markedly non-Gaussian (Zhang et al., 2021).
• Wasserstein/norm-constrained sets: provide tighter uncertainty description, require moderate sample sizes ($N \ge 100$), and enable MILPs whose complexity is linear or sublinear in sample number and risk threshold; adjustable-risk variants remain tractable and avoid big-M artifacts (Zhang et al., 2021).
• Kernel mean sets: control higher-order deviations and are computationally feasible for moderate N via convex QP/SOCP, offering simultaneous moment robustness (Zhu et al., 2020).

For multistage programs, simpler moment-bound sets provide practical solutions as MILPs, but ellipsoidal norm constraints (Type 3) yield lower conservatism at the cost of MISDP complexity (Yu et al., 2020). In regret-optimal control, spectral and Schatten-norm regularizations achieve robust designs, but interior-point solvers for the resultant SDP may scale poorly, motivating projected subgradient methods (Taha et al., 11 Dec 2025).

6. Practical Implications and Computational Studies

Computational studies demonstrate:

• For building load control, moment-based DRCC delivers zero out-of-sample violations with minimal samples but incurs higher operating penalties; Wasserstein DRCC achieves superior cost-risk balance as sample sizes grow (Zhang et al., 2021).
• Multistage DRO with decision-dependent moment sets (Type 1, 2) converges efficiently in SDDiP algorithms, while MISDP approximations for ellipsoidal sets trade off tightness for solution speed (Yu et al., 2020).
• For kernel-mean ambiguity sets, worst-case risk quantification can be implemented efficiently using off-the-shelf convex solvers, with rigorous monotonicity and convergence properties, and empirical scenarios demonstrate near-optimal cost-risk calibration (Zhu et al., 2020).
• Distributionally robust regret-optimal control under Schatten-norm ball ambiguity sets demonstrates scalable, accurate policy computation via dual subgradient ascent, outperforming conventional SDP solvers in high dimension (Taha et al., 11 Dec 2025).

These results underline the centrality of norm- and moment-constrained ambiguity sets in rigorous, tractable DRO formulations across optimization, control, and decision sciences.