Information theoretic comparisons of training based channel estimation and semi-blind estimation in fading channels with memory

In this paper, we analyze and compare the semi-blind estimation scheme with the conventional training based scheme in linear block fading channels with memory. Our results show that the performance of semi-blind estimation in terms of the channel mutual information and the Cramer-Rao bound is superior to the conventional training based scheme, but the performance gain diminishes as the signal-to-noise ratio (SNR) increases. We have derived the optimal lengths of training sequence for both estimation schemes in terms of the mutual information of the channel. Provided that the length of the training sequence used for channel estimation is optimized, we found that by using either channel estimation scheme, the mutual information of the channel drops sharply with the memory of the channel at low SNR. On the contrary, the mutual information drops only mildly with the memory of the channel at high SNR.