The purpose of this paper is to extend the recent formulation of locally optimum Bayes (LOB) detection in nonadditive non-Gaussian noise with independent sampling to the case where dependence between the noise samples is modeled via an ergodic first-order Markov discrete-time process. Moreover, unlike previous related work, numerical results are provided which are based on an empirically derived second-order transition density, the marginal PDF of which is Middleton's (1993) Class A noise, an experimentally verifiable and widely applicable non-Gaussian model. Canonical, in signal waveform and noise statistics, asymptotically and LOB detectors in both coherent and incoherent modes are derived and their statistics are computed under both "signal present" and "signal absent" hypotheses. Performance measures are thus obtained together with the correlation gain G/sup (M.P.)/, which is used for systems comparison. Explicit forms for the transition density and and the nonlinearities involved, as well as numerical values of the noise indices, are calculated from a generalized observation model containing multiplicative and additive Markov noise components. It is shown that significant performance gains over the case with independent sampling can be achieved, depending upon the degree of correlation between the noise samples.

The locally optimum Bayes detection (LOBD) framework for nonadditive non-Gaussian noise is applied to a particular additive and signal-dependent noise model where the noise components are correlated. Under the proposed observation model, the performance of the LOB detectors is compared with that of well-known conventional receivers.

The locally optimum Bayes theory of signal detection in additive non-Gaussian noise/interference is extended to independent observations of the received data without the additive noise restriction. The methodology employed parallels very closely the original development of threshold detection theory and utilizes the mathematical machinery of asymptotic decision theory, especially, the concepts of contiguity and locally asymptotically normal (LAN) log-likelihood ratio, which are needed in the determination of the detector structure in both coherent and incoherent modes and its statistics under both hypotheses. Under the present framework, the canonical (in signal waveform and noise statistics) optimum detection algorithms retain their asymptotically optimum character. An example is provided in order to demonstrate the applicability of the theory to a specific noise environment, where explicit forms of the non-linearities involved and numerical values of the new noise indices are obtained. Moreover, a significant improvement in performance (0 (24-27) dB) over that of optimum detectors in independent, additive Gaussian and non-Gaussian noise is noted.>