Assortative Matching, Reputation, and the Beatles Break-Up

Consider Becker's (1973) classic static matching model, with output a stochastic function of unobserved types. Assume symmetric incomplete information about types, and thus commonly observed Bayesian posteriors. Matching is then assortative in these `reputations' if expected output is supermodular in types. We instead consider a standard dynamic version of this world, and discover a robust failure of Becker's global result. We show that as the production outcomes grow, assortative matching is neither efficient nor an equilibrium for high enough discount factors. Specifically, assortative matching fails around the highest reputation agents for `low-skill concealing' technologies. Our theory implies the dynamic result that high-skill matches (like the Beatles) eventually break~up. Our results owe especially to two findings: (a) value convexity due to learning undermines match supermodularity; and (b) for a fixed policy in optimal learning, the second derivative of the value function explodes geometrically at extremes.