EM Algorithm

Grung and Manne [9] studied the EM-like algorithm for PCA in the case of missing values. In the E-step, $\mathbf{A}$ is fixed and $\mathbf{S}$ is solved as a least squares problem. In the M-step, $\mathbf{S}$ is fixed and $\mathbf{A}$ is solved again as a least squares problem. Computations are a lot heavier than in the fully observed case, but still, experiments in [9] showed a faster convergence compared to the iterative imputation algorithm.