Bilinear precoding for FDD massive MIMO system with imperfect covariance matrices

This work deals with a massive multiple-input-multiple-output (MIMO) system operating in frequency division duplex (FDD) mode. We propose a linear precoding technique for the downlink (DL) transmission, where we exploit the second-order information of the DL channel gained from an extrapolation of the uplink (UL) channel covariance. Contrary to the instantaneous channel state information (CSI), which is not reciprocal, the channel covariance matrix depends on the so-called angular scattering function, which is invariant over the UL-DL frequency gap and can be therefore extracted from the UL training signals. An extrapolation of the estimated UL channel covariance matrices is then performed to obtain the DL covariance matrices. We also propose to make use of the feedback symbols to update these estimates of the DL covariance matrices based on the observation of the actual channel. Simulations show a satisfactory performance in terms of the spectral efficiency and an approach is suggested to handle the mismatch of the estimated covariance matrices especially in the high transmit power region.