Multi-Channel Factor Analysis for Temporally and Spatially Correlated Time Series.

In Multi-Channel Factor Analysis (MFA), the spatial covariance of a multi-channel observation is decomposed into the covariances of latent signal, interference, and noise components. In proposed applications, the observations also have temporal correlations which may be of independent interest or may influence the spatial covariance estimates. An extension to MFA is proposed where the common and unique factor series are synthesized using LTI filters with unknown transfer functions. A novel block majorization-minimization procedure for semi-parametric estimation of both spatial and temporal correlations is summarized. Experiments show that the resulting technique for spatio-temporal correlation analysis of a multi- channel observation series improves on MFA when the factor series are time-dependent.