Ambiguity and robustness

Expected utility theory assumes the decision-maker has (or acts as if they have) one precise probability distribution over states. Two classic experiments show real people systematically violate this.

The Ellsberg paradox (1961). Urn A has 50 red and 50 black balls; Urn B has 100 red-or-black balls in unknown proportion. Most people are indifferent between betting on red vs. black within either urn (so implied probabilities are 50/50 for both), yet strictly prefer betting on Urn A (known odds) over Urn B (unknown odds) for either colour. No single probability distribution over Urn B's composition can rationalise this — it is not that people misjudge probabilities, it is that they display ambiguity aversion: a preference for known risk over unknown ("Knightian") uncertainty, which pure expected utility cannot represent at all.

The Allais paradox (1953). Offered a choice between (A) a certain $1M and (B) a lottery with 10% chance of $5M, 89% of $1M, 1% of nothing, most people choose A. But offered a choice between (C) 11% chance of $1M vs. 89% of nothing, and (D) 10% chance of $5M vs. 90% of nothing, most people choose D. This violates the vNM independence axiom: A and B, and C and D, differ only in a common consequence (a 89%-vs-90%-of-nothing tail), which should not flip the ranking under independence, yet it does — attributed to people overweighting certainty itself (the "certainty effect"), formalised later in prospect theory and rank-dependent utility. The distortion function $g$ in ADE captures exactly this kind of systematic probability overweighting/underweighting, giving a normative (not merely descriptive) tool that still respects monotonicity and avoids other paradoxes.

Formal responses.

• Gilboa–Schmeidler maxmin expected utility (1989): instead of one prior, admit a set of priors $\mathcal C$ consistent with the available evidence, and evaluate each action by its worst-case expected utility, $V(a)=\min_{p\in\mathcal C}E_p[u(X_a)]$ — a decision rule that is itself axiomatically founded (relaxing only the independence axiom to a weaker "certainty independence") and that rationalises Ellsberg-type behaviour.
• ε-contamination: a tractable special case where $\mathcal C = \{(1-\varepsilon)p+\varepsilon q : q \text{ arbitrary}\}$ for a reference $p$ and "contamination" size $\varepsilon$. The worst case over this whole set collapses to a simple formula: $(1-\varepsilon)E_p[u(X_a)]+\varepsilon\min_s u(X_a(s))$ — exactly ADE's ambiguity term. $\varepsilon=0$ is full Bayesian trust in $p$; $\varepsilon=1$ is pure Wald maximin, ignoring probabilities altogether.
• Multiplier preferences (Hansen–Sargent robust control): penalise deviation from a reference model by a relative-entropy cost, yielding a "robust" certainty-equivalent that is more tractable in dynamic/control settings but conceptually parallel.
• Info-gap decision theory (Ben-Haim): sidesteps probability altogether, asking "how much can my model be wrong and still deliver an acceptable outcome?" — a purely non-probabilistic robustness measure useful when even a set of priors is hard to justify (deeply novel risks, e.g. an entirely new peril).

ADE's stance. ADE does not claim ambiguity aversion is always "correct" — reasonable decision-makers can be purely Bayesian ($\varepsilon=0$). It offers $\varepsilon$ as a disclosed dial so that the degree of robustness sought is explicit and auditable, and the recommendation under $\varepsilon=0$ vs. $\varepsilon>0$ can be compared side by side, exposing exactly how much the choice is driven by ambiguity aversion versus by expected value.