FDD Massive MIMO Based on Efficient Downlink Channel Reconstruction

This paper focuses on frequency division duplex (FDD) massive multiple-input multiple-output (MIMO) systems and proposes a transceiver design that fully exploits the downlink spatial multiplexing gain with only a small amount of overhead. The bottleneck lies in the acquisition of downlink channel state information (CSI), which occurs when large scale antenna array is employed in FDD transmission systems. Fortunately, the spatial reciprocity between uplink and downlink inspires us to reconstruct the downlink channel based on the frequency-independent parameters (downtilts, azimuths and delays) that can be derived in the uplink. We first extract these parameters through an enhanced Newtonized orthogonal matching pursuit (e-NOMP) algorithm which is proposed in this paper to fit the massive MIMO orthogonal frequency division multiplexing (OFDM) system. After formulating the requirement to achieve an acceptable estimation error rate, we propose a low-cost downlink training scheme to estimate the downlink gains of each user channel. This scheme saves the training time resource by introducing a predefined spatial angle grid which corresponds to a beam set and by minimizing the number of selected beams which is equal to the number of OFDM symbols used for downlink training. Having obtained the reconstructed multiuser channel, the BS can maximize the spatial multiplexing gain by serving all the users simultaneously without causing severe interference. Numerical results verify the precision of the e-NOMP algorithm, and demonstrate that sum-rate performance of the reconstructionbased transceiver design approximates that of using perfect CSI