A From-Scratch Approach to Deep Unfolding-Aided Bilinear Gaussian Belief Propagation for Correlated Large MIMO

This paper proposes a trainable bilinear Gaussian belief propagation (BiGaBP) aided by deep unfolding built from scratch. In Massive MIMO scenarios, joint channel and data estimation (JCDE) based on belief propagation (BP) has been achieved with low computational complexity through bilinear generalized approximate message passing (BiGAMP) and bilinear Gaussian belief propagation (BiGaBP). However, although BiGaBP has a lower dependence on the large-system limit approximation compared to BiGAMP, both message-passing algorithms are still derived based on the large system assuming independent and identically distributed (i.i.d.) observations with mean zero, resulting in significant performance degradation in the presence of spatial fading correlation. To tackle this issue, we introduce the parameters in BiGaBP for belief design, which were previously set manually in conventional method, as trainable parameters and improve performance by conducting data-driven tuning under spatially correlated fading. We conduct computer simulations from scratch based on a theoretically derived iterative error backpropagation algorithm from the signal flow graph, demonstrating the effectiveness of the proposed method in terms of bit error rate (BER) and achieving transparency in deep unfolding.