On the Performance of ML Channel Estimation for OQPSK in Frequency Non-Selective Fading

The maximum likelihood estimator for a constant channel gain based on offset QPSK (OQPSK) training symbols is derived and analyzed. The estimator is unbiased and only asymptotically efficient. The estimator error variance is a function of the variance of an inphase/quadrature (I/Q) interference term that creates a floor for the estimator error variance. The I/Q interference term is not present in estimators based on QPSK and constitutes the primary difference between the estimators for OQPSK and QPSK. Examples using the non-return to zero (NRZ) and square-root raised cosine (SRRC) pulse shapes demonstrate that for a fixed training sequence length, the performance floor for the estimator error variance is lower (better) with full response pulse shapes (NRZ) than with partial response pulse shapes (SRRC). The criterion for training sequences that minimize the undesired performance floor is also set forth.