Blind multiple-input multiple-output system identification with applications in multiuser communications

Multiple-Input Multiple-Output (MIMO) system models appear in many applications, such as telecommunications, earthquake engineering, speech processing and biomedical engineering. Therefore, MIMO system identification has attracted a lot of interest in recent years. The goal of blind identification of a ( P x M) MIMO system is to identify an unknown system driven by P unobservable inputs, based on the M system outputs, and subsequently use the system estimate to recover the input signals (sources).; In this thesis we propose a frequency domain approach for blind MIMO system identification, that uses second order spectra correlation of the system output. We first consider the case of stationary or cyclostationary, non-white inputs with known correlations. We propose a novel criterion that involves correlations of discrete Fourier transforms (DFT) of the system output, minimization of which yields the system impulse response within a scalar ambiguity. A strength of the proposed approach is insensitivity to channel length mismatch. The performance of the proposed method has been demonstrated via extensive simulations and also real measurements obtained using our wireless communications testbed.; Next we consider the case where the input statistics are unknown, a problem of great interest in separation of competing speakers in speech processing applications. We show that under certain conditions, second order spectra correlations of the system output are sufficient for the unique identification of a P x P system. Although identifiability results for the same problem existed for the 2 x 2 case, this is the first result for the general P x P MIMO case.; The case of white stationary inputs cannot be identified unless the higher-order statistics (HOS) are used. HOS have been criticized for high complexity and the need for long data in order to maintain small variance of the system estimates. We propose a low rank approach for reducing the variance of the HOS estimates for a fixed data length. We demonstrate the usefulness of the approach in Single-Input Single-Output (SISO) blind system identification.