In an environment where agents face similar but distinct sets of problems, imitating others' successful behavior consists of inferring patterns, or statistical regularities, that can help in dealing with one's own unique problems. The extent to which successful behavior is copiable is determined by the complexity of the environment: in simple environments, imitation and learning wash out all initial competitive differences while in complex environments sustained differences in performance persist. Learning is modeled as nonparametric inference based on no prior, structural knowledge of the environment. Complexity is a barrier to imitation because it creates a statistical identification problem that hampers agents' inference about the behavioral rules followed by others. The model provides some basis to common informal arguments appealing to causal ambiguity, the value of focused business strategies, and inflexibility in combining different behavioral rules.

This is an old working paper which was not widely circulated or submitted for publication. The tools needed to model the issues raised in that paper either did not exist or were unknown to me. Much of my subsequent work consisted of attempts to fill the gaps (this includes works on completeness of Lp spaces with finitely additive probabilities, VC-theory and learning on discrete spaces and so on). A major revision of this paper should be available soon.

Original 1999 version to be revised extensively soon.