CSI Compression for Massive MIMO: Model-Based or Data-Driven?

With increasing numbers of transmit antennas and the use of wider spectra, channel state information (CSI) feedback for massive MIMO systems can incur a large overhead in uplink capacity. For a 32-antenna system operating on 13 frequency subbands, uncompressed channel state feedback would consume about two Mb/s of uplink capacity for each receive antenna, and so compression of the feed-back is essential in practice. Communication theory provides foundational models for the compression of large arrays of numbers and it has been applied to reduce this overhead significantly. However, simplifying assumptions typically used to make these models tractable leave significant room for improvement. With the maturity of machine learning, alternative data-driven approaches have received much attention over the past few years. These approaches have a higher complexity and computational footprint, but demonstrate improved accuracy compared to legacy models and promise more robust customization and interoperability. In this article, we give an overview of both the legacy model-driven and the emerging data-driven approaches for CSI compression, provide macroscopic comparisons of the performance of the two approaches based on both efficiency and flexibility, and discuss the maturity of this technology for large-scale deployment in future 5G-Advanced and 6G systems.