Optimal relay functionality for SNR maximization in memoryless relay networks

We explore the SNR-optimal relay functionality in a mentoryless relay network, i.e. a network where, during each channel use, the signal transmitted by a relay depends only on the last received symbol. at that relay. We develop a generalized notion of SNR for the class of memoryless relay functions. The solution to the generalized SNR optimization problem leads to the novel concept of minimum mean squared uncorrelated error (MMSUE) estimation. For the elemental case of a single relay, we show that MMSUE estimate is a scaled version of the MMSE estimate. This scheme, that we call estimate and forward (EF), performs better than the best of amplify and forward (AF) and demodulate and forward (DF) in both parallel and serial relay networks. We determine that AF is near-optimal at low transmit power in a parallel network, while DF is near-optimal at high transmit power in a serial network. For hybrid networks that contain both serial and parallel elements, the advantage of EF over the best of AF and DF is found to be significant. Error probabilities are provided to substantiate the performance gain obtained through SNR optimality. We also show that, for Gaussian inputs, AF, DF and EF are identical.