Performance-Guaranteed CSI Feedback via Model-Agnostic Incremental Residual Compression Framework

This paper proposes a framework that utilizes multiple deep autoencoders to compress channel state information (CSI) in massive multiple-input multiple-output (MIMO) systems. A key limitation of existing methods is their focus on overall average performance, which fails to guarantee any error thresholds for individual instances on a designated compression rate (CR). To address this issue, we propose an iterative residual-compression method that integrates weight binarization, with our error control mechanism enabling its effective application and thereby reducing the computational burden on user equipment (UE). Experimental results show that our framework improves worst-case error levels by up to 6.98 dB compared to baseline models, with further improvements achievable through the use of additional residual stages. Moreover, the framework is designed to adaptively generate codes that meet specific target error levels (e.g., achieving -10 dB), rather than merely adjusting the CR. Finally, numerical evaluations indicate that our method reduces the average code length by up to 4.3 times relative to that of backbone models, highlighting its efficiency.