On the convergence of Bayesian posterior processes in linear economic models. Counting equations and unknowns (Q1177287)

This paper focuses on the question whether there will be a complete learning of the true parameter vector of a linear regression process. A technique of counting equations and unknowns is proposed which allows to decide when the posterior distributions of the parameters converge to their true values. This technique is illustrated for i.i.d. regressors and lagged dependent regressors. Section 2 shows that joint restrictions on both the distributional form of the priors and errors as in Bayesian statistics are not required for the convergence of the posterior process (theorem 2.2). For the case of multiple regressors, the technique requires much weaker conditions than others used in literature. In case of one regressor, the theorem yields the convergence of the Bayesian posterior to any given sample path with at least two limit points. Section 3 provides a number of examples that show how to use the theorem. In section 4, the problem of the optimal control of a linear regression process is discussed. Some of the regressors are obtained via an infinite-horizon optimization process, facing an agent in an economy (say, a monopolist). The limit points of the optimal action process solve certain one-period problems. For a nonlinear policy function it is shown for the static case that the true regression parameters may be learned completely. Finally, these results are illustrated with an example of a monopolist with unknown demand curve.