Blind MIMO-AR system identification and source separation with finite-alphabet

In this paper, a new method for system identification and blind source separation in a multiple-input multiple-output (MIMO) system is proposed. The MIMO channel is modeled by a multi-dimensional autoregressive (AR) system. The transmitted signals are assumed to take values from a finite alphabet, modeled by the Gaussian mixture model (GMM) with infinitesimal variances. The expectation-maximization (EM) algorithm for estimation of the MIMO-AR model parameters is derived. The performance of the proposed algorithm in terms of probability of error in signal detection and root mean squared error (RMSE) of the system parameters and system transfer function estimates is evaluated via simulations. It is shown that the obtained probability of error is very close to the probability of error of the optimal algorithm which assumes known channel state information.