Observational learning occurs when privately-informed individuals sequentially choose among finitely many actions, after seeing predecessors' choices. We summarize the general theory of this paradigm: Belief convergence forces action convergence, specifically, copycat "herds" arise. Also, beliefs converge to a point mass on the truth exactly when the private information is not uniformly bounded.This subsumes two key findings of the original herding literature: With multinomial signals, cascades occur, where individuals rationally ignore their private signals, and incorrect herds start with positive probability. The framework is flexible -some individuals may be committed to an action, or individuals may have divergent cardinal or even ordinal preferences.