This paper generalizes the result of Mei and Montanari to show that (under some assumptions) as long as the population risk has nice behavior, the empirical risk also have nice properties. The main novelty is that it removed the assumption that saddle points need to be hyperbolic - this is important for the applications.