Abstract
To alleviate the downlink training and uplink feedback overhead in frequency-division duplexing massive multiple-input multiple-output systems, a two-stage precoder is proposed for hybrid analog and digital precoding. The analog beamformer and digital precoder are jointly designed by leveraging the signal-to-leakage-plus-noise ratio metric. The design of the analog beamformer is based only on the long-term channel statistics information, exploiting both the channel mean and low rank covariance statistics, whereas the digital precoder utilizes the knowledge of an effective channel with significantly reduced dimensionality. Consequently, we can considerably reduce the signalling and feedback overhead to the dimensions of the resultant effective channel. For comparison purposes, the cases of full channel state information at the transmitter (CSIT) and statistical CSIT are also investigated. It is shown that the digital precoder design problem reduces to the generalized Rayleigh quotient problem, while the analog beamformer design problem reduces to the quotient trace problem. These dimensionality reduction problems are solved via the generalized eigenvalue decomposition method.