Unbounded expected utility and continuity (Q761228): Difference between revisions

A decision-maker often does not know what the outcome of a particular action will be; instead, he associates with each possible action a (subjective) probability distribution of outcomes. His problem is to choose a preferred probability distribution subject to whatever constraints he is facing. The decision-maker's preference structure can be described by a measurable space of outcomes, a set of probability measures on the set of outcomes, and a binary relation (preference relation) on those probability measures. Subject to certain basic assumptions, this paper gives necessary and sufficient conditions for a preference structure to conform to the expected-utility hypothesis: there exists a real function (utility function) defined on the set of outcomes, integrable with respect to all members of the set of probability measures where the preference relation is defined, such that one probability measure is preferred to another if and only if it yields a higher expected value of the utility function. Nothing is assumed or concluded about boundedness of the utility function. Most previous axiomatic treatments of expected utility have required the utility function to be bounded because integrability implies boundedness when the set of probability measures is large enough. However, since applied work often uses unbounded functional forms, it is desirable to have an axiomatic theory available which allows for unboundedness by restricting the set of probability measures. The difficulty is to impose axioms which are formulated entirely in terms of the preference relation and the set of probability measures without direct reference to the utility function. The issue is closely related to the St. Petersburg paradox. Another advantage of this paper over previous contributions is that the set of probability measures is not assumed to be convex. The mathematical tools used in the paper are topological in nature and have to do with weak convergence of probability measures.