Covariance estimation in decomposable Gaussian graphical models
IEEE Transactions on Signal Processing
| Hi-index | 35.69 |
In a series of two papers, a new class of parametric models for two-dimensional multivariate (matrix-valued, space-time) adaptive processing is introduced. This class is based on the maximum-entropy extension and/or completion of partially specified matrix-valued Hermitian covariance matrices in both the space and time dimensions. The first paper considered the more restricted class of Hermitian Toeplitz-block covariance matrices that model stationary clutter. This second paper deals with the more general class of Hermitian-block covariance matrices that model nonstationary clutter. For our recently proposed 2-D time-varying autoregressive (TVAR) model, we derive optimal and computationally practical suboptimal methods for calculating such parametric models. The maximum-likelihood covariance matrix estimate for the 2-D TVAR model is also derived. The efficacy of the introduced models is illustrated by signal-to-interference-plus-noise ratio (SINR) degradation results obtained when applying the covariance matrix models to space-time adaptive processing filter design, compared with the true clutter covariance matrix provided by the DARPA KASSPER dataset.