ML and EM algorithm for non-data-aided SNR estimation of linearly modulated signals

The recently published Cramer-Rao lower bound for non-data-aided (NDA) estimation of the signal-to-noise ratio (SNR) reveals a considerable gap, when compared to the jitter performance of NDA algorithms available from the open literature. The maximum-likelihood (ML) solution derived in this paper closes this gap. However, the latter provides a set of two nonlinear vector equations, which might be simplified only for modulation schemes with constant envelope like M-ary PSK. For signals with nonconstant envelope, like 16-QAM as most prominent example in this respect, a much less complex approach based on the expectation-maximization (EM) principle is developed in this paper. In the medium SNR range, this bridges part of the performance gap mentioned previously. Over the full SNR range, we propose a hybrid algorithm, where the EM estimate is replaced by a moment-based method as soon as the true SNR drops below a predefined threshold.