Toward the Fundamental Limits of Imitation Learning