FDD Multiuser Massive MIMO Systems with Adaptive Channel Covariance Feedback

Abstract

Multiple works that address training and pre-coding in FDD massive MIMO systems are based on the assumption that the base station (BS) has a priori knowledge of the users' downlink (DL) channel covariance matrices. It is already known that columns of DFT (resp. Kronecker product of DFT) matrices approximate covariance eigen-vectors of large uniform linear (resp. rectangular) arrays in the large system limit. Therefore, DFT-based codebooks can be used for covariance eigenvector feedback from the user to the BS in FDD mode for the aforementioned arrays. It is not known, however, how much feedback is required, i.e., how many eigenvectors need to be fed back to the BS for a given DL signal-to-noise-ratio (SNR). The main paper contribution is discussing how the number of fed back covariance eigenvectors should be adapted to the DL SNR to achieve increasing sum rates with increasing SNR, and further discussing the feedback implications on training design. Numerical results scheme show only negligible sum rate losses compared to the case with perfect covariance knowledge at the BS.