Suppression of Self-Noise Feedback in GAMP for Highly Correlated Large MIMO Detection

This paper deals with uplink multi-user detection (MUD) via generalized approximate message passing (GAMP) in highly correlated large multi-user multi-input multi-output (MUMIMO) systems. The most vital mechanism of GAMP is Onsager correction for decoupling the self-noise feedback of beliefs across iterations; this makes it possible to exchange extrinsic values. First, we show that by introducing the belief scaling method proposed in the context of Gaussian belief propagation (GaBP) for adjusting the convergence speed into GAMP, the Onsager correction works properly even under spatial fading correlation, which significantly improves detection capability. Surprisingly, the performance is much better compared to that of the GaBP with belief scaling, and even asymptotically approaches that of computationally expensive expectation propagation (EP) detectors. Based on the results, we clarify why such a dramatic performance improvement is possible only for GAMP in terms of the suppression mechanism of self-noise feedback.