A deterministic constraint demands feasible, always. A chance constraint relaxes this to feasible, with at least probability 1 − α. Formally, given a random quantity X and a threshold b, the constraint P(X ≤ b) ≥ 1 − α says the requirement X ≤ b is allowed to fail in up to a fraction α of scenarios.
The parameter α is the risk tolerance, picked explicitly by the modeller (typically 1–5%). It is the single most consequential number in a chance-constrained programme: shrink α and the model converges on worst-case robust optimisation; enlarge it and the model converges on the expected-value solution. The choice of α is a policy question dressed as a parameter.
Chance constraints exist because most hard requirements in practice are not actually hard. A service-level target of "cover 100% of emergency demand always" forces the model to size against the single worst historical day; a target of "cover emergency demand with probability 0.95 on any given day" produces a working plan that is honest about the failure rate it accepts.
Setup: A hospital is sizing the number of operating-theatre slots it reserves for emergency admissions each day. Reserve too many, and elective procedures get pushed out, destroying revenue. Reserve too few, and the hospital turns emergencies away, breaching its service-level promise.
The three options: The worst-case reservation covers the single worst historical day and blocks a large share of elective capacity. The mean-demand reservation covers roughly half of days and breaches the service promise on the other half. The chance-constrained reservation picks a middle number, tied to a stated failure rate.
The model: Choose the reserved slot count s so that P(emergencies > s) ≤ 0.05. The binding quantity is the 95th-percentile daily demand, not the mean and not the maximum. The reserved capacity is larger than the mean but much smaller than the worst case, and the slippage the model accepts is explicit.
What α controls: Tighten α to 0.01 and the model protects against the 99th percentile, pushing the reservation closer to worst case. Loosen to 0.10 and the reservation drops toward the 90th percentile. The dial is continuous and the policy implication is direct.
A chance constraint at 5% risk is much tighter than a constraint on the mean. It asks about the 95th-percentile scenario, not the typical one. Modellers who mentally substitute "so the constraint holds on average" under-dimension their reserves every time.
Under simple distributional assumptions (Gaussian demand, independent noise, discrete scenarios), a chance constraint linearises to a deterministic quantile constraint and stays inside a Mixed Integer Linear Programming (MILP) solver. Moving to Mixed Integer Nonlinear Programming (MINLP) is a last resort, not the default.
A chance constraint assumes the distribution is known. A distributionally robust constraint assumes only that the distribution lies in a set and asks the constraint to hold for the worst distribution in that set. The two are siblings, not synonyms.
A chance constraint lets rare scenarios violate the requirement in exchange for a model that is achievable in the common case, with the risk tolerance α as the single most consequential number the modeller picks.