GIST: Location-Dependent Uplink Path Structure Transfer for FDD Downlink Channel Estimation

Abstract

A key challenge of downlink (DL) channel estimation in frequency division duplex (FDD) massive multiple-input multiple-output (MIMO) systems is the pilot overhead, which grows with the numbers of antennas and subcarriers. However, a prior on path delays and directions shared by the uplink (UL) and DL channels can reduce this overhead. A model trained only on UL channels measured at a base station (BS) can predict the path structure (delays, directions and power gains) from the location of a user equipment (UE) and thereby assist DL channel estimation. In this paper, we propose Gaussian-Informed Structure Transfer (GIST), in which each Gaussian primitive models one propagation path, whose delay, directions and power gain are predicted from the UE location relative to the primitive center. The complex gains of the predicted paths are estimated from DL pilots by linear minimum mean square error estimation with this prior. Experiments on a ray-traced and a measured dataset show that GIST outperforms all baselines, reaching a lower DL estimation error with fewer pilot subcarriers. Further experiments demonstrate its robustness to UE location errors, its low storage and computational requirements.