Reducing CSIR Sounding Overhead in MIMO-OFDM System via Bayesian End-to-End Learning With Delay-Domain Sparse Precoding

This paper introduces a novel precoder design aimed at reducing pilot overhead for estimating channel state information at the receiver (CSIR) in multiple-input multiple-output orthogonal frequency division multiplexing (MIMOOFDM) applications utilizing high-order modulation. We propose an innovative demodulation reference signal scheme that achieves up to an 8x reduction in overhead by implementing a delay-domain sparsity constraint on the precoder. Furthermore, we present a deep neural network (DNN)-based end-to-end architecture that integrates a propagation channel estimation module, a precoder design module, and an effective channel estimation module. Additionally, we propose a Bayesian modelassisted training framework that incorporates domain knowledge, resulting in an interpretable datapath design. Simulation results demonstrate that our proposed solution significantly outperforms various baseline schemes while exhibiting substantially lower computational complexity.