Compressive RF training for massive MIMO with channel support side information

Hybrid beamforming (BF) is a promising solution for massive MIMO with limited RF chains. To reduce the amount of pilot overhead for channel estimation, compressed sensing techniques that exploit channel sparsity have been proposed. One key issue is how to design the RF (analog) training vectors to achieve higher BF gain with fewer pilots. Specifically, narrow-beam RF training requires large pilot overhead for finding strongest paths, and random RF training suffers from low BF gain. We propose to use a mixture of narrow-beam and random RF training vectors, and optimize the fraction of the two sets of RF training vectors based on channel support side information (CSSI) at the BS. We show that this optimized fraction exhibits a phase transition: When the CSSI accuracy exceeds a certain threshold, the maximum number of narrow-beam RF training vectors should be used to focus beams in all directions indicated by the CSSI. Otherwise, only random RF training vectors should be used to explore the unknown channel support. Moreover, we derive closed-form bounds on the channel estimation error. Both the analysis and simulations show that the proposed method can achieve substantial gains over various baseline methods.