EcoPrecise: Lightweight Wi-Fi CSI Feedback via Product Quantization VA

This paper proposes EcoPrecise, a lightweight deep learning framework based on Vector Quantized Variational Autoencoder (VQ-VAE) to address the efficiency-accuracy tradeoff in high dimensional Channel State Information (CSI) feedback for Wi-Fi systems. Unlike traditional Givens rotation quantization per subcarrier, EcoPrecise directly compresses the right singular matrix V via end-to-end learning through three key innovations. A lightweight architecture integrating depthwise separable convolutions and Squeeze&Excitation (SE) modules for deployment on edge devices; Product quantization with orthogonal regularization enabling flexible resource allocation; And dynamic exponential moving average (EMA) based codebook adaptation leveraging channel temporal correlation for noiserobust online fine-tuning. Experiments show that the proposed method can reduce the feedback overhead by about 80% compared with traditional methods while maintaining high reconstruction fidelity and robustness in indoor wireless environments. This work establishes a new paradigm for efficient CSI representation learning in Wi-Fi systems.