Sparsifying dictionary learning for beamspace channel representation and estimation in millimeter-wave massive MIMO

Abstract

Millimeter-wave (mmWave) massive multiple-input-multiple-output (mMIMO) is reported as a key enabler in fifth-generation communication and beyond. It is customary to use a lens antenna array to transform a mmWave mMIMO channel into a beamspace where the channel exhibits sparsity. This beamspace transformation is equivalent to performing a Fourier transformation of the channel. Still, a Fourier transformation is not necessarily optimal for many reasons. For example, it can cause a power leakage problem. Accordingly, this paper proposes using a learned sparsifying dictionary as the transformation operator leading to another beamspace for channel representation. Since a dictionary is obtained by training over actual channel measurements in an end-to-end manner, this transformation is shown to yield two immediate advantages. First is enhancing channel sparsity, thereby leading to more efficient pilot reduction. Second is improving the channel representation quality, thus reducing the underlying power leakage phenomenon. Consequently, this allows for improved channel estimation and facilitates beam selection in mmWave mMIMO. In addition, a learned dictionary is used as the channel estimation operator for the same reasons. Extensive simulations under various operating scenarios and environments validate the added benefits of using learned dictionaries in improving the channel estimation quality and beam selectivity, thus improving spectral efficiency.