On estimation with missing data

The maximum likelihood estimation of the mean and covariance of a normal random vector X, based on a sequence of independently and identically distributed observations {X1, X2,.., XN} with possibly missing components in Xi for some i is examined. It is shown that a systematic procedure can be obtained via the general theory of the E-M (Expectation and Maximization) algorithm (Ref. 3). A numerical example is discussed.