We compare the effect of blind and nonblind channel estimates on the performance of Global System for Mobile communications (GSM) receivers. More precisely, we investigate whether two blind approaches, based on higher order statistics (HOS), can compete with two conventional methods, exploiting training sequences. For blind and nonblind estimates of six fast and slowly fading mobile radio channels, we give simulated bit error rates (BERs), after Viterbi detection, in terms of the signal-to-noise ratio (SNR). We also study the influence of cochannel interferers at different values of the signal-to-interference ratio (SIR). Averaged over the six channel examples, we demonstrate that the blind channel estimation algorithm eigenvector approach to blind identification (EVI) leads to an SNR loss of 1.2-1.3 dB only, while it saves the 22% overhead in GSM data rate caused by the transmission of training sequences. Since just 142 samples are used for blind channel estimation, we consider this performance outstanding for an approach based on HOS.

We investigate the applicability of two algorithms for the blind identification of mixed-phase linear time-invariant FIR systems to the estimation of mobile radio channels on GSM conditions. One approach is based on second order cyclostationary statistics (SOCS), whereas the other exploits higher order stationary statistics (HOSS) of the received signal. While the former class of algorithms suffers from "singular" systems which can not be identified, the latter class is said to require an excessive number of samples of the received signal to achieve comparable performance levels. The purpose of this paper is two-fold: first, we demonstrate that "singular" systems represent a severe limitation to SOCS-based methods when it comes to the estimation of time-variant mobile radio channels from a small number of received samples. Secondly, we reveal that the approach relying on 4th order statistics yields a superior estimation performance: at a signal-to-noise-ratio of 10 dB, all channel examples can be identified from 142 samples of a GSM burst within a normalized mean square error bound of 7 per cent.

The performance of many communication systems could be improved if the transmission channel was estimated blindly, i.e. without training sequences. As an example, we investigate whether, on GSM conditions, the blind channel estimation method EVI (eigen vector approach to blind identification) can compete with the non-blind least squares scheme based on the cross-correlation. For Gaussian stationary uncorrelated scattering channels, we give simulated bit error rates (BER) after Viterbi detection in terms of the mean signal-to-noise ratio (S~N~R~) for blind, non-blind, and ideal channel estimation. Averaged over three COST-207 propagation environments, EVI leads to an S~N~R~ loss of 1.1 dB only, which is quite remarkable for an approach based on higher order statistics, as just 142 samples can be used for blind channel estimation.