This paper characterises completely the circumstances in which maximum likelihood estimation of the factor model is feasible when the sample covariance matrix is rank deficient. This situation will arise when the number of variables exceeds the number of observations.