A Dynamic Vector Quantization Framework for CSI Feedback in Massive MIMO Systems

Abstract

Obtaining accurate downlink channel state information (CSI) is critical for precoding in massive multi-input multi-output (MIMO) systems, yet it imposes a significant feedback overhead on the uplink bandwidth in frequency division duplex (FDD) systems. While deep learning (DL) based CSI feedback has demonstrated remarkable performance, conventional vector quantization (VQ) schemes remain structurally rigid and have limited ability to adapt their quantization space to the training-data distribution. To address this challenge, this paper proposes a dynamic vector quantization (DVQ) framework for CSI feedback. Specifically, to eliminate feature redundancy, the framework introduces an orthogonal transformation that concentrates feature energy for effective dimensionality reduction. A data-driven initialization is developed to mitigate codebook misalignment, and a geometry-aware binary labeling strategy is introduced to reorder codebook indices to improve robustness against bit-flip errors. Experimental results demonstrate that compared with commonly used Transformer-based vector quantization (TransVQ) models, DVQ could significantly reduce the air-interface overhead by more than 33.3% while enhancing the reconstruction accuracy. It provides a flexible and bit-efficient solution for high-capacity CSI feedback in future 6G wireless networks.