CsiMamba: Efficient Channel State Information Feedback for Frequency‐Division Duplex Massive Multiple‐Input Multiple‐Output Systems

Abstract

In frequency‐division duplex massive multiple‐input multiple‐output systems, channel state information (CSI) feedback is essential for downlink transmission but incurs substantial overhead. To improve CSI feedback efficiency while enhancing downlink transmission reliability, we propose CsiMamba, a Mamba‐based CSI feedback model. CsiMamba employs state‐space‐model–based sequence modeling with convolutional and lightweight MB blocks to capture global and local CSI characteristics with low complexity. Experiments using 5G new radio clustered delay line‐C channels demonstrate that CsiMamba achieves superior bit‐error‐rate and block‐error‐rate performance while reducing the parameter count by 14.3% and computational complexity by 36.9% compared with CsiFormer. The performance advantage becomes more pronounced under quantized feedback and remains comparable to that of the traditional CSI feedback scheme. These results indicate that CsiMamba provides an efficient CSI feedback solution that improves practical downlink communication reliability.