Chance-Constrained Control

• Chance-constrained control is a framework that enforces probabilistic constraints on stochastic system dynamics to balance performance with risk management.
• Deterministic reformulation methods, such as analytic quantile tightening and risk allocation, convert probability constraints into tractable optimization problems.
• Numerical algorithms and data-driven techniques, including convex programming and deep reinforcement learning, enable real-time robust control in safety-critical applications.

Chance-constrained control is a paradigm in optimal and model predictive control that enforces hard constraints on the probability of undesirable events in stochastic dynamical systems. Rather than requiring constraints to hold deterministically, chance-constrained formulations allow them to be violated with user-specified probability levels, yielding a principled balance between performance and risk. This approach has become central to safety-critical applications where uncertainty affects both system dynamics and operational specifications.

1. Mathematical Foundations and Formal Problem Definition

Chance-constrained control problems consider discrete- or continuous-time stochastic dynamics of the form

$$x_{t+1} = f(x_t, u_t, \omega_t), \qquad x_0 \;\; \text{given},$$

where $x_t$ is the state, $u_t$ the control input, and $\omega_t$ is a disturbance process with known or partially known probability law. The objective is to minimize an expected finite- or infinite-horizon cost, typically additive in state and control, subject to chance constraints on, e.g., state, input, or output: $$P\left\{ g(x_k, u_k) \leq 0 \right\} \geq 1-\epsilon$$ for risk parameter $\epsilon\in(0,1)$. Frequently, the constraints are expressed as requirements that certain sets (e.g., "target reach," "safe operation," "input bounds") are satisfied with high probability, or that the closed-loop trajectory remains within a prescribed region with probability at least $1-\alpha$ (Jasour et al., 2016, Kordabad et al., 2024, Patil et al., 19 Nov 2025).

Formulations extend beyond individual stage-wise constraints to joint, mission-wide, or temporal logic predicates (Wang et al., 2022, Kordabad et al., 2024), as well as include binary effectors or hybrid system elements.

2. Deterministic Reformulation Techniques

Chance constraints are probability inequalities, nonconvex in control parameters. Deterministic reformulation is essential for efficient solution:

• Analytic Quantile Tightening:

For affine or Gaussian settings, the chance constraint reduces to a deterministic inequality involving the mean and variance. For a scalar affine function, $a^T x + b^T u$,

$$P[a^T x + b^T u \leq h] \geq 1-\epsilon \implies a^T E[x] + b^T u \leq h - \Phi^{-1}(1-\epsilon) \sqrt{a^T \text{Cov}(x) a}$$

where $\Phi^{-1}$ is the standard normal quantile (Svensen et al., 2020, Oguri, 2024, Schildbach et al., 2014).

• Boole’s Inequality and Risk Allocation:

Joint chance constraints are intractable, so risk is allocated to individual constraints using Boole’s inequality, introducing auxiliary risk variables subject to $\sum \delta_j \leq \alpha$ (Kato et al., 26 Jan 2026). Deterministic surrogates are derived using Chebyshev/Cantelli for known mean/variance (non-Gaussian), Hoeffding for bounded support, or inverse CDFs for known tails.

• Samples, Kernel Density, and Scenario Methods:

When only samples are available (unknown or nonparametric disturbances), sample-based deterministic reformulations use empirical statistics (Priore et al., 2023) or biased kernel density estimators (KDEs) to upper bound chance constraints for specified violation probabilities (Keil et al., 2020, Kiel et al., 2020). Such methods transform the probability bound into a smooth nonlinear constraint that can be incorporated in an optimization framework.

• Distributional Robustness:

Under unknown/distorted distributions, distributionally robust optimization is employed: empirical samples are used to define an ambiguity set (e.g., Wasserstein ball), and the design seeks to enforce the chance constraint for all distributions in the ambiguity set, guaranteeing high-confidence satisfaction (Kordabad et al., 2024).

• Measure and Moment-SOS Methods:

For nonlinear polynomial dynamics and cost, the chance-constrained optimization is lifted to an infinite-dimensional linear program over measures, then relaxed to a sequence of tractable semidefinite programs using moments and sum-of-squares polynomials, with proven convergence (Jasour et al., 2016).

3. Numerical Solution Algorithms and Complexity

• Convex Programming (SDP, SOCP, QCQP):
  • Linear systems with affine-Gaussian uncertainty lead to SDPs or SOCPs via quantile-based tightness (Oguri, 2024, Schildbach et al., 2014, Kato et al., 26 Jan 2026).
  • Problems with non-Gaussian uncertainty but known moments, or hybrid risk allocation, are managed via convex risk relaxations plus regularization to ensure strict convexity (Kato et al., 26 Jan 2026).
• Sampling-based Methods:

When explicit probability distributions are unknown, Monte Carlo simulation or scenario optimization combined with kernel smoothing (e.g., Epanechnikov, Split–Bernstein kernels) is applied to reformulate chance constraints as sample averages subject to bias bounds. Warm-start, incremental sampling, and kernel-switching reduce computational burden (Kiel et al., 2020, Keil et al., 2020).

• Successive Convexification, Difference-of-Convex Programming:

Sample-mean concentration-based constraints and norm-based safety constraints are often "difference-of-convex" (DC) in parameters, and are solved using the convex-concave procedure, iteratively linearizing nonconvexities (Pilipovsky et al., 2023, Priore et al., 2023).

• Policy Classes and Bellman Approaches:

Dynamic programming is feasible for mission-wide or joint-chance constraints via state augmentation, representing the “probability of safety” as a functional state; this is exact but only computationally tractable for low dimensional problems (Wang et al., 2022).

• Hamilton-Jacobi-Bellman (HJB) and Path-Integral Methods:

For continuous-time SOC, the chance constraint becomes an exit-time probability controlled via dualization (Lagrange multiplier, risk tuning in the boundary cost). Solving the resulting HJB PDE (linearizable via Cole-Hopf in matched noise–control cases) produces optimal feedback, which may be approximated via finite differences or path-integral (Monte Carlo importance sampling) (Patil et al., 2022, Patil et al., 19 Nov 2025).

• Deep Reinforcement Learning (RL):

Lexicographic Deep Q-Networks separately train primary cost and constraint critics, enabling RL control with explicit state-level chance constraints enforced by lex-minimization, requiring no penalty weight tuning or retraining when risk thresholds change (Giuseppi et al., 2020).

4. Theoretical Guarantees

• Feasibility, Recursive Feasibility, and Safety:
  • Hierarchies of convex relaxations converge to the optimal chance-constrained value as the order increases (moment-SOS, (Jasour et al., 2016)).
  • Risk allocation methods guarantee satisfaction of the original joint chance constraints if marginal constraint regularity and convexity conditions are met (Kato et al., 26 Jan 2026).
  • Under GMM uncertainty, robust tightening ensures recursive feasibility (feasibility at one MPC step implies feasibility at all subsequent steps) (Ren et al., 2024).
  • Distributionally robust formulations deliver finite-sample guarantees with user-level confidence by calibrating Wasserstein ball radii (Kordabad et al., 2024).
  • Sample-statistic reformulations may provide almost-sure chance constraint satisfaction at the expense of moderate conservatism (Priore et al., 2023).
• Performance vs. Conservatism:

Analytic relaxations and sample-statistic methods trade off computational tractability with conservatism. Explicit risk allocation and convex surrogates reduce over-conservatism observed in uniform allocation or deterministic bounds. Data-driven kernel methods (RKHS embedding) provide convergence guarantees as dataset size grows, with performance close to scenario-based methods (Thorpe et al., 2022).

5. Application Domains and Case Studies

Chance-constrained control methodologies have been demonstrated across diverse domains:

• Spacecraft Autonomy: Convex chance-constrained SOCPs have been deployed for robust rendezvous, station-keeping under navigation and actuation uncertainty, and safety-aware path planning, achieving validated empirical risk bounds (Oguri, 2024).
• Powertrain and Energy Systems: Exactly-linearizable SCMPC controllers with risk allocation have been applied for uncertain powertrain control, with joint input and state risk optimized using learned models (Kato et al., 26 Jan 2026).
• Autonomous Vehicles and Obstacle Avoidance: GMM-based chance-constrained MPC allows for safe navigation amid multi-modal obstacle prediction; contingency planning reduces conservatism by branching on scenario modes (Ren et al., 2024).
• Safety via Barrier and Lyapunov Functions: Convex robust QP synthesis enforces probabilistic control barrier (CBF) and Lyapunov (CLF) conditions under state estimation noise, with empirical guarantees on forward invariance and stability (Wang et al., 2020).

6. Extensions and Open Challenges

• Non-Gaussian, Nonlinear, and Nonparametric Uncertainty: Recent work has extended chance-constrained formulations to systems with non-Gaussian, unknown, or multi-modal uncertainties, including learning-based and data-driven control strategies (e.g., RKHS embedding, Wasserstein DRO) (Thorpe et al., 2022, Kordabad et al., 2024).
• Temporal Logic and Logic-Based Safety Specifications: Distributionally robust approaches for STL-predicate chance constraints have been developed when atomic predicates are Lipschitz and disturbances admit concentration of measure (Kordabad et al., 2024).
• Scalability and Real-Time Implementation: Convexification, warm-start methods, and sample-efficient RL enable practical real-time chance-constrained control for high-dimensional and mission-critical systems (Kiel et al., 2020, Giuseppi et al., 2020).
• Theoretical Gaps: For general nonlinear, nonconvex systems, global guarantees remain limited; moment-SOS relaxations and dual/measure-theoretic methods address select cases, but scalability and practical computation for high degree-of-freedom problems are still open challenges (Jasour et al., 2016).

7. Comparative Summary of Solution Approaches

System Setting/Uncertainty	Core Reformulation	Optimization Class	Notable References
Linear/Gaussian	Quantile-based tightening	SDP, SOCP	(Oguri, 2024, Schildbach et al., 2014)
General random samples	Biased-KDE overapproximation	NLP, CCP (DC)	(Keil et al., 2020, Kiel et al., 2020)
Nonlinear polynomial	Measure/moment-SOS hierarchy	SDP sequence	(Jasour et al., 2016)
Unknown/Distributionally robust	Wasserstein ambiguity, DRO via duality	Finite-dimensional convex	(Kordabad et al., 2024)
GMM/multi-modal	SOCP per mode, robust risk allocation	SOCP, Mixed-integer	(Ren et al., 2024)
RL/Black-box dynamics	Lexicographic DQN critics	DNN/Policy iteration	(Giuseppi et al., 2020)
Data-driven (RKHS embedding)	Empirical conditional mean/LP	Linear program	(Thorpe et al., 2022)
Mission-wide chance constraint	State augmentation, backward DP	DP/recursive Bellman	(Wang et al., 2022)

Chance-constrained control is now a foundational framework for synthesizing risk-aware controllers in complex stochastic environments. Deployed methodologies span convex optimization, measure-theoretic relaxations, data-driven and distributionally robust optimization, and deep RL, each offering trade-offs between conservatism, computational scalability, and applicability to nonlinear and/or data-driven contexts.