Abstract
In modern Multiple-Input Multiple-Output (MIMO) wireless communication systems, the overhead associated with Channel State Information (CSI) feedback presents a significant bottleneck. This paper proposes a novel framework that treats CSI compression as a Wyner-Ziv coding problem—compression with side information—and addresses it using generative diffusion models. By pairing a trainable codebook-based encoder at the User Equipment (UE) with a conditional diffusion decoder at the Base Station (BS), the framework achieves a twofold improvement in compression efficiency over current state-of-the-art methods. The approach is particularly effective in complex scenarios where the Base Station must predict future Downlink CSI using compressed feedback and available Uplink side information.
Introduction
The evolution toward massive MIMO in 5G and 6G networks has dramatically increased the dimensionality of wireless channels. To optimize beamforming and data transmission, the Base Station requires accurate CSI, which must be measured and fed back by the User Equipment. However, as antenna arrays grow, the volume of this feedback data threatens to saturate the available bandwidth.
While traditional neural network-based lossy compression (e.g., Autoencoders) has provided some relief, these models often struggle with practical, high-complexity tasks. Specifically, when the objective is not just reconstruction but the prediction of future CSI using side information (such as Uplink CSI already known to the BS), standard neural compression models hit a performance ceiling. This paper introduces a diffusion-based approach designed to break through these saturation points and provide high-fidelity CSI recovery at significantly lower bitrates.
Methodology
The proposed framework reframes CSI feedback as a Wyner-Ziv coding task, utilizing a two-stage process involving discrete vector quantization and generative denoising.
Fixed-Rate Encoding with a Trainable Codebook
At the UE, raw CSI data is processed through a neural network to produce continuous latent vectors. To ensure a fixed-rate transmission suitable for standard communication protocols, these vectors undergo vector quantization. They are mapped to the nearest entries in a trainable codebook, resulting in a highly compressed, discrete codeword. This codeword represents the minimal necessary information required for the BS to reconstruct the channel state.
Conditional Diffusion Decoding
The Base Station employs a backward diffusion process to reconstruct the CSI. Unlike standard decoders that perform a single-shot mapping, the diffusion decoder utilizes a U-Net architecture to iteratively denoise a signal, starting from pure Gaussian noise. This process is conditioned on two critical inputs:
- The received codeword: The compressed representation sent by the UE.
- Side information: Available Uplink CSI, which provides context about the current environment.
By treating CSI reconstruction as a conditional generation task, the model can produce high-fidelity estimates that align with the physical constraints of the wireless channel.
Experimental Setup
The framework was evaluated across two distinct tasks using simulated and standard datasets:
- Future Channel Prediction: Using the NVIDIA Sionna simulator, the authors modeled a vehicular scenario (UE moving at 5 m/s) with 32 BS antennas. The 3GPP CDL-C channel profile was used to generate dynamic MIMO CSI data, testing the model's ability to predict future states based on past side information.
- Channel Reconstruction: The model was benchmarked on the COST2100 outdoor dataset, a standard in the field containing 120,000 samples. This task focused on pure compression and decompression without side information to compare against existing literature.
Key Results and Discussion
Breaking the Performance Ceiling
In the future channel prediction task, traditional baselines like CsiNet and CRNet exhibited clear performance saturation. Regardless of the number of bits allocated, these models struggled to improve beyond a -6 dB Normalized Mean Squared Error (NMSE). In contrast, the diffusion-based model scaled effectively; increasing the bit budget from 64 to 128 bits resulted in a gain of over 1 dB, demonstrating its superior representational capacity.
Efficiency Gains
On the COST2100 dataset, the proposed method significantly outperformed state-of-the-art algorithms including CsiConformer and CsiNet+. To achieve a target distortion of -7 dB NMSE, the diffusion-based approach required fewer than 150 bits. In comparison, all baseline methods required more than 400 bits to reach the same quality. This represents a greater than 2x improvement in data rate efficiency, allowing for more precise channel control with less overhead.
Conclusion
This research demonstrates that generative diffusion models, while typically associated with image generation, are exceptionally well-suited for high-precision industrial tasks like MIMO CSI compression. By leveraging the Wyner-Ziv framework and conditional denoising, the proposed model overcomes the limitations of traditional neural compression.
While the iterative nature of diffusion models introduces computational complexity, future work focusing on fast sampling techniques (such as rectified flow) and lightweight architectures will be essential for real-time deployment in 6G systems. The current results provide a strong foundation for a new class of generative communication protocols.
Summary updated