Bayesian End-to-End Learning for FDD-Massive MIMO Physical-Layer Design

This paper proposes a variational Bayesian learning framework for the end-to-end design of frequency-division duplex (FDD) massive Multiple Input Multiple Output (MIMO) communication systems. By factorizing the posterior distribution of involved variables, we effectively model various components of the system, including MIMO precoding and channel state information (CSI) at the transmitter (CSIT) recovery at the transmitter, along with channel estimation (CE), CSIT compression, and MIMO detection at the receiver. This factorization enhances adaptability and incorporates estimation errors during training, as each block outputs a probability distribution rather than deterministic values. Furthermore, we exploit the inherent sparsity of the MIMO channel using a two-layer hierarchical prior model to achieve efficient estimation. To address the intractability of the original problem, we propose a low-complexity unsupervised variational Bayesian training algorithm that does not require true channel realizations. Extensive simulations demonstrate that the proposed framework significantly outperforms existing methods while minimizing signaling overhead.