This paper investigates Mirrlees' model of optimal income taxation. It provides a concrete example of utility and density functions for which the solution to the usual (first-order) model is not implementable, i.e. an example where the first-order approach does not work. Adding second-order conditions leads to an extended model and to implementable solutions. If these conditions are binding one gets a kink in the optimal net-income schedule and bunching of individuals occurs. The properties of an optimal nonlinear income tax are reexamined within the extended model.