Low Complexity Detection and Precoding for Massive MIMO Systems: Algorithm, Architecture, and Applicatio

Massive (or large-scale) MIMO is an emerging technology to improve the spectral efficiency of existing (small-scale) MIMO wireless communication systems. The main idea is to equip the base station (BS) with hundreds of antennas that serves a small number of users (in the orders of tens) simultaneously in the same frequency band. In such a system, the data detection and precoding are among the most challenging tasks in terms of computational complexity and performance. Although theoretical results show that simple detection and precoding algorithms are able to achieve optimal error rate performance when the number of BS antennas approaches infinity, the systems with realistic antenna configurations have to resort to computationally expensive algorithms to achieve near-optimal performance.;In this research, we show that by utilizing the special property of massive MIMO systems, approximate linear detection and precoding can deliver near-optimal error rate performance with low complexity. We first propose approximate methods relying on Neumann series. This approach requires lower computational complexity than that of an exact inversion while delivering near-optimal results when there is a large ratio between BS and user antennas. We then develop a novel reconfigurable VLSI architecture to perform both the necessary Gram matrix computation and Neumann series based matrix inversion. The Neumann series approach, however, suffers from a considerable error-rate performance loss if the ratio of BS to user antennas is not large enough. To improve the performance, we investigate the conjugate gradient (CG) method (without explicitly computing matrix inversion) and conjugate gradient least square (CGLS) method (without explicitly computing Gram matrix and matrix inversion). Although CG and CGLS for precoding are rather straightforward, the necessary signal-to-interference-and-noise-ratio (SINR) for soft-output detection is not computed by CG and CGLS. To solve this problem, we propose an exact and an approximate method to compute the SINR within CG and CGLS algorithm with low complexity. We show that compared to exact and Neumann series based linear methods, CG based detection and precoding method is suitable for systems with small to medium number of users, while CGLS is suitable for systems with large number of users. A novel reconfigurable VLSI architecture is then proposed to support the both CG and CGLS.