Fully Decentralized Massive MIMO Detection with Gradient-Based MCMC

Abstract

To cope with the increasing computational complexity and interconnection bandwidth of centralized baseband processing in massive multiple-input multiple-output (MIMO), decentralized baseband processing (DBP) has been proposed as an alternative. Under this architecture, existing partially decentralized MIMO detection schemes require frequent interchange of knowledge on estimated data, incurring high interconnection bandwidth and non-negligible central unit (CU) processing. In this paper, we propose a fully decentralized gradient-guided Markov chain Monte Carlo (MCMC) detector that completely offloads sampling to distributed units (DUs), which are coordinated via communication-efficient message passing with compact statistics. The proposed design minimizes CU-side workload and reduces interconnection overhead while maintaining manageable detection complexity. Extensive simulation results demonstrate that the proposed detector achieves enhanced bit error rate (BER) performance within a few message-passing rounds and exhibits superior BER-complexity-interconnection trade-off over existing decentralized detectors.