Robust Non-Bayesian Learning

This paper investigates the robustness of three of the core insights of the DeGroot model of boundedly rational updating in social networks. Two updating systems are ε-close if their supremum norm distance is equal to ε. An ε-perturbation of a DeGroot (weighted average) updating system is any continuous updating system that is ε-close. A property of long run opinions is robust if it is retained in the limit for all ε-perturbations as ε converges to zero. We show consensus is a robust outcome but influence and naive learning are not