Combining Scheme Design for the User Multi-Antenna Cell-Free Massive MIMO System

Federated learning has emerged as a research hotspot in communication scenarios due to its potential to enable efficient, privacy-preserving model training over resource-constrained and unreliable networks. In this paper, we investigate a federated learning task in user multi-antenna Cell-Free massive MIMO (CF mMIMO) system. Through theoretical analysis, we point out that the design based on the maximum ratio combining (MRC) scheme, as discussed in existing works, suffers from significant performance degradation in scenarios with multiple antennas and multi-stream transmission, due to the lack of effective suppression of data stream interference across different antennas of the same user. Therefore, without assuming a predefined combining scheme, we perform a convergence analysis of federated learning algorithm under the CF mMIMO system and conclude that the optimal combining scheme corresponds to the minimum mean square error (MMSE) detector. Through the simulation results of the federated learning task under the investigated system, we confirm the validity of the convergence analysis, demonstrating that optimizing the MMSE can improve the convergence of federated learning.