Why do we need this method? What’s wrong with just generating observations from the distribution itself?