Referential Approximate Message Passing for Quantized Large MIMO Detection

This paper proposes a look-up table (LUT)-based multi-user detection (MUD) scheme imitating approximate message passing (AMP) for uplink signal detection in large multi-user multi-input multi-output (MU-MIMO) systems. When AMP is implemented with double-precision arithmetic, the increase in power consumption and memory occupation becomes a major obstacle to practical application as the system scale expands. To tackle this issue, we design a novel referential AMP detector composed by hierarchically cascading many small LUTs, where only informative integer-valued messages are exchanged on a factor graph. The quantization thresholds of LUTs are sequentially computed by tracking the discrete distribution of the LUT outputs at each layer to minimize the performance degradation owing to quantization errors. Based on the distribution at each layer, the thresholds are optimized by clustering with the Lloyd-Max algorithm using the initial values given by the k-means++ method. Simulation results demonstrate the efficacy of the proposed method in terms of the bit error rate (BER) performance and memory usage. Notably, we show that the referential AMP is robust to changes in communication environments, and then clarify the reason in terms of the algorithmic structure.