Belief Identification in Populations

We study the identification of belief distributions in a population of Bayesian agents from anonymous aggregate belief data.

While a single Bayesian agent's full belief can be recovered from beliefs over a suitable collection of binary events, this principle need not extend to populations: event-by-event distributions of beliefs may fail to identify the underlying distribution of priors.

We study when this failure is generic and when it is exceptional.

Identification is governed by the graph-theoretic structure induced by the observed family of events on the state space.

Among $n$-agent distributions, identification is generic if the induced graph is nonseparable, while non-identification is generic if the graph is separable.

The results establish both limits and design principles for recovering belief heterogeneity from aggregate belief data.