On the Minimal Error of Empirical Risk Minimization