On Learning-Curve Monotonicity for Maximum Likelihood Estimators