Blind subspace-based channel estimation using the EM algorithm

We propose an application of the Expectation-Maximization (EM) algorithm to the problem of blind estimation of single-input multiple-output (SIMO), finite-impulse-response (FIR) channels. We first assume Gaussian input to formulate an EM-based estimation of the signal subspace of the output covariance matrix. This Gaussian assumption allows us to utilize knowledge from EM-based probabilistic principle component analysis (P-PCA). Next, we show that the equilibrium point of the EM iteration equations is reached without the Gaussian assumption, which suggests usage of non-Gaussian communication input signals. The estimated signal subspace is then utilized to identify the channels. In principle, the proposed method yields the same channel estimates as the widely-known subspace method, but is computationally more efficient. In addition, unlike typical EM applications, the proposed scheme is free from cumbersome parameter initialization issue, which greatly increases flexibility of the proposed scheme.