Learning in an Uncertain World: MIMO Covariance Matrix Optimization With Imperfect Feedback