Design and implementation of multiuser MIMO precoding algorithms

Abstract

The demand for high-speed communications required by cutting-edge applications has put a strain on the already saturated wireless spectrum. The incorporation of antenna arrays at both ends of the communication link has provided improved spectral efficiency and link reliability to the inherently complex wireless environment, thus allowing for the thriving of high data-rate applications without the cost of extra bandwidth consumption. As a consequence to this, multiple-input multiple-output (MIMO) systems have become the key technology for wideband communication standards both in single-user and multi-user setups. The main difficulty in single-user MIMO systems stems from the signal detection stage at the receiver, whereas multi-user downlink systems struggle with the challenge of enabling non-cooperative signal acquisition at the user terminals. In this respect, precoding techniques perform a pre-equalization stage at the base station so that the signal at each receiver can be interpreted independently and without the knowledge of the overall channel state. Vector precoding (VP) has been recently proposed for non-cooperative signal acquisition in the multi-user broadcast channel. The performance advantage with respect to the more straightforward linear precoding algorithms is the result of an added perturbation vector which enhances the properties of the precoded signal. Nevertheless, the computation of the perturbation signal entails a search for the closest point in an in nite lattice, which is known to be in the class of non-deterministic polynomial-time hard (NP-hard) problems. This thesis addresses the difficulties that stem from the perturbation process in VP systems from both theoretical and practical perspectives. On one hand, the asymptotic performance of VP is analyzed assuming optimal decoding. Since the perturbation process hinders the analytical assessment of the VP performance, lower and upper bounds on the expected data rate are reviewed and proposed. Based on these bounds, VP is compared to linear precoding with respect to the performance after a weighted sum rate optimization, the power resulting from a quality of service (QoS) formulation, and the performance when balancing the user rates. On the other hand, the intricacies of performing an efficient computation of the perturbation vector are analyzed. This study is focused on tree-search techniques that, by means of an strategic node pruning policy, reduce the complexity derived from an exhaustive search and yield a close-to-optimum performance. To that respect, three tree-search algorithms are proposed. The xed-sphere encoder (FSE) features a constant data path and a non-iterative architecture that enable the parallel processing of the set of vector hypotheses and thus, allow for high-data processing rates. The sequential best-node expansion (SBE) algorithm applies a distance control policy to reduce the amount of metric computations performed during the tree traversal. Finally, the low-complexity SBE (LC-SBE) aims at reducing the complexity and latency of the aforementioned algorithm by combining an approximate distance computation model and a novel approach of variable run-time constraints. Furthermore, the hardware implementation of non-recursive tree-search algorithms for the precoding scenario is also addressed in this thesis. More specifically, the hardware architecture design and resource occupation of the FSE and K-Best xed-complexity treesearch techniques are presented. The determination of the ordered sequence of complexvalued nodes, also known as the Schnorr-Euchner enumeration, is required in order to select the nodes to be evaluated during the tree traversal. With the aim of minimizing the hardware resource demand of such a computationally-expensive task, a novel non-sequential and lowcomplexity enumeration algorithm is presented, which enables the independent selection of the nodes within the ordered sequence. The incorporation of the proposed enumeration technique along with a fully-pipelined architecture of the FSE and K-Best approaches, allow for data processing throughputs of up to 5 Gbps in a 4x4 antenna setup.