Bayesian Economist ... Bayesian Agents I: An Alternative Approach to Optimal Learning

Abstract

We study the framework of optimal decision making under uncertainty where the agents do not know the full structure of the model and try to learn it optimally. We generalize the results on Bayesian learning based on the martingale convergence theorem to the sequential framework instead of the repeated framework for which results are currently available. We also show that the variability introduced by the sequential framework is sufficient under very mild identifiability conditions to circumvent the incomplete learning results that characterize the literature. We then question the type of convergence so achieved, and give an alternative Bayesian approach whereby we let the economist himself be a Bayesian with a prior on the priors that his agents may have. We prove that such an economist cannot justify endowing all his agents with the same (much less the true) prior on the basis that the model has been running long enough that we can almost surely approximate any agent's beliefs by any other's. We then examine a possibly weaker justification based on the convergence of the economist's measure on beliefs, and fully characterize it by the Harris ergodicity of the relevant Markov kernel. By means of very simple examples, we then show that learning, partial learning, and non-learning may all occur under the weak conditions that we impose. For complicated models where the Harris ergodicity of the Markov kernel in question can neither be proved nor disproved, the mathematical/statistical test of Domowitz and El-Gamal (1989) can be utilized.