Joint Source-Channel Coding with Learnable Codebook Quantization for CSI Feedback

Abstract

In this paper, we propose a joint source-channel coding (JSCC) framework with a learnable codebook quantizer (LCQ), termed JSCC-LCQ, for channel state information (CSI) feedback in frequency division duplex orthogonal frequency division multiplexing (FDD-OFDM) systems. Different from conventional codeword lookup or index-oriented discrete representation methods, the proposed framework generates feedback bitstreams through selector-guided codeword selection, position-wise codebook learning, and codeword-guided aggregation within an end-to-end trainable architecture. Specifically, for each latent position, the learnable codebook quantizer forms a bit-level representation through soft codeword aggregation and hard binarization, while preserving gradient propagation via differentiable training. Simulation results under clustered delay line channels with imperfect CSI (CDL-impCSI) and perfect CSI (CDL-pCSI) demonstrate that the proposed method achieves stable and competitive CSI reconstruction performance across a wide SNR range, with a 9.21% relative SGCS improvement over the JSCC baseline with uniform quantizer (JSCC-UQ) under representative limited-feedback settings. Furthermore, analyses of codebook entropy, codebook size, and computational complexity demonstrate that JSCC-LCQ sustains stable codeword utilization while incurring negligible inference-time overhead.