A Deep-Learning Based Framework for Joint Downlink Precoding and CSI Sparsification

Optimal pilot design to acquire channel state information (CSI) is of critical importance for FDD downlink massive MIMO systems, and is still an open problem. To tackle this issue, in this paper we propose a two-stage precoding approach based on reduced CSI (rCSI-TSP) design framework and an efficient algorithm, whose core is to obtain an optimal precoder while also sparsifying physical CSI (pCSI), so as to save on CSI estimation. The advantages of the rCSI-TSP framework are three-fold. First, the framework enables to simultaneously extract and exploit statistical and instantaneous CSI. Second, it guarantees the most needed rCSI can be obtained and thus avoids performance loss due to heuristic pilot design. Third, we tailor an efficient online deep-learning based method for the TSP framework, which paves the way for practical applications. As an example, we apply the framework to the multi-user symbol-level precoding (SLP) and verify performance improvements.