Compressed CSI Feedback With Learned Measurement Matrix for mmWave Massive MIMO.

A major challenge to implement compressed sensing method for channel state information (CSI) acquisition lies in the design of a well-performed measurement matrix to subsample sparse channel vectors. The widely adopted randomized measurement matrices drawn from Gaussian or Bernoulli distribution may not be optimal. To tackle this problem, we propose a fully data-driven approach to learn a measurement matrix for beamspace channel compression, and this method trains a mathematically interpretable autoencoder constructed according to the iterative solution of sparse recovery. The learned measurement matrix can achieve near perfect CSI recovery with fewer required measurements, thus the feedback overhead can be substantially reduced.