Chance Constraints

Definition

A chance constraint requires a safety (or feasibility) constraint to hold not deterministically but
with at least a prescribed probability, thereby tolerating a small bound on the probability of
violation. Given a state/decision $x$, an uncertain parameter $\delta$ distributed as $p_{\ast }$, and a
constraint function $C(x,\delta )$ with $C\le 0$ denoting satisfaction, the constraint is enforced only
in the probabilistic sense ${\mathbb{P}}_{\delta \sim p_{\ast }}(C(x,\delta )\le 0)\ge 1-ϵ$
(ren2022chance). This relaxes the conservatism of robust (worst-case)
formulations, which must hold for every realization, while still bounding the safety risk
(dixit2023risk). The formulation is regime-agnostic — it constrains the
planning/control optimization, not the dynamics — so it transfers to a free-flying base unchanged,
provided the propagated uncertainty $\delta$ reflects the actuated-base model. A key limitation: a chance
constraint scores only the boolean event “violated / not violated” and is blind to the magnitude of a
violation in the tail, which motivates the coherent-risk-measure refinements below
(majumdar2017how).

Key Equations

Symbols per notation.md. The risk-layer confidence is $\alpha$; the per-constraint
violation tolerance $ϵ\in (0,\frac{1}{2})$ (confidence $=1-ϵ$, so $\alpha =1-ϵ$ and the
risk-operator subscript $1-\alpha =ϵ$) is not yet in notation.md — used here as in
ren2022chance, dixit2023risk. The scalar
quantile multiplier $\Gamma ={\Psi }^{-1}(1-ϵ)$ below is the source’s notation
(ren2022chance); it is distinct from the repo’s coordinate-transform
matrix $\mathbf{\Gamma }$ in notation.md.

The single chance constraint:

${\mathbb{P}}_{\delta \sim p_{\ast }}(C(x,\delta )\le 0)\ge 1-ϵ.$

For a linear constraint $C={\delta }^{\top }\tilde{x}$ with Gaussian $\delta \sim \mathcal{N}(\mu ,\Sigma )$, it has
the exact deterministic second-order-cone reformulation
(ren2022chance, Lemma 1; Calafiore–El Ghaoui):

\qquad \Gamma=\Psi^{-1}(1-\epsilon),$$ where $\Psi^{-1}$ is the standard-normal inverse CDF. Replacing the probability operator with a **coherent risk measure** $\rho_{1-\alpha}$ generalizes the chance constraint to a *risk-sensitive safety constraint* ([dixit2023risk](../sources/dixit2023risk.md)): $$\rho_{1-\alpha}\big(C(x,\delta)\big)\ \le\ \epsilon_l .$$ A [conditional_value_at_risk](conditional_value_at_risk.md) (CVaR) constraint is a convex *inner* approximation: $\mathrm{CVaR}_{\epsilon}(C)\le 0 \Rightarrow \mathbb{P}(C\le 0)\ge 1-\epsilon$ ([ren2022chance](../sources/ren2022chance.md), [majumdar2017how](../sources/majumdar2017how.md)). ## Source Support - [ren2022chance](../sources/ren2022chance.md) — primary: states the chance constraint, its exact SOC reformulation for (mixtures of) Gaussians, and its CVaR inner approximation; introduces *risk allocation* $\sum_k\pi_k\epsilon_k=\epsilon$ across GMM modes. - [dixit2023risk](../sources/dixit2023risk.md) — generalizes the probability operator to a coherent risk measure $\rho_{1-\alpha}$ (risk-sensitive safety constraints) inside a receding-horizon MPC. - [majumdar2017how](../sources/majumdar2017how.md) — frames chance-constrained programming as one risk metric among many; argues it is *tail-blind* (boolean), motivating coherent/distortion risk metrics. - [akella2024risk](../sources/akella2024risk.md) — survey placing chance constraints in the worst-case / risk-neutral / risk-aware taxonomy; tail risk measures (CVaR, EVaR) give convex inner approximations. - [zheng2024informed](../sources/zheng2024informed.md) — uses a per-step chance constraint $\mathbb{P}(\text{collision})\le\delta$ as the safety condition in belief-space sampling-based planning. - naumann2020probabilistic — broader probabilistic-planning context (MDP/POMDP, probabilistic occupancy); supplies the surrounding uncertainty-modelling vocabulary. - vasquezgomez2017view — inspection/NBV planning that selects views by *expected utility* and a sampled probability of a collision-free trajectory; a probabilistic-feasibility analogue rather than an explicit chance constraint. ## Related Topics - [coherent_risk_measures](coherent_risk_measures.md) — the axiomatic class ($\rho_{1-\alpha}$) that generalizes the chance-constraint probability operator while keeping convexity. - [conditional_value_at_risk](conditional_value_at_risk.md) — the specific coherent measure giving a convex inner approximation of a chance constraint; also penalizes violation *severity*, unlike the chance constraint itself. - motion_planning_under_uncertainty — the host problem: chance constraints are the safety conditions a stochastic planner must satisfy. - [covariance_propagation](covariance_propagation.md) — supplies the state distribution $(\mu_k,\Sigma_k)$ that the deterministic SOC reformulation needs at each step. - responsibility_sensitive_safety — a deterministic, rule-based safety envelope; contrast with the probabilistic guarantee a chance constraint provides. - [gaussian_mixture_model](gaussian_mixture_model.md) — the multimodal uncertainty model under which [ren2022chance](../sources/ren2022chance.md) decomposes the chance constraint per mode with risk allocation. - [model_predictive_control](model_predictive_control.md) / [risk_aware_mpc](risk_aware_mpc.md) — the receding-horizon setting where chance / risk-sensitive constraints are most commonly imposed. ## Open Questions - All cited sources assume a free-floating / ground / automotive / aerial setting; none derive the constraint for a free-*flying* space manipulator. Does the SOC reformulation stay tractable when $\delta$ carries the **dynamic coupling** between the actuated base and the redundant arm, or does the coupling inflate $\Sigma$ enough to make the constraint dominate the optimization? - The exact SOC reformulation requires (mixtures of) Gaussian moments; our inspection uncertainty (estimation + thruster/contact disturbance) may be non-Gaussian. When do the CVaR / coherent-risk inner approximations stay tight rather than over-conservative ([majumdar2017how](../sources/majumdar2017how.md), [dixit2023risk](../sources/dixit2023risk.md))? - Risk allocation ($\sum_k\pi_k\epsilon_k=\epsilon$, [ren2022chance](../sources/ren2022chance.md)) splits a single budget across modes; how should a per-step budget be allocated across an *inspection trajectory* (many viewpoints, many keep-out surfaces) without compounding into excessive conservatism?