Axiomatic Characterization of Generalized NEO-EU Preferences and the Falsifiability of Ambiguity Theories

We introduce a new weakening of the independence axiom, co-extreme independence, which we show characterizes a class that encompasses popular ambiguity models including Hurwicz preferences, ϵ-contamination preferences, and NEOEU preferences. Importantly, our representation unifies these as forms of simple ambiguity models as we show that the Vapnik-Chervonenkis (VC) dimension of these models is either unity or increases linearly with the number of states. In contrast, standard ambiguity models characterized in the Anscombe-Aumann framework are complex ambiguity models that have a VC dimension that is either infinite or increases exponentially in the number of states. We show that simple ambiguity preferences are efficient to compute by introducing a linear program for finding the parameters of a NEO-EU representation given an arbitrary dataset of binary choices between acts. In contrast to recent work showing that ambiguity theories have large VC dimension compared to expected utility theory (EU), simple ambiguity preferences provide a counter-example to the conjectures that axiomatic models of ambiguity attitudes are substantially more difficult to falsify or are substantially more complex than EU.