Improving Channel Spatial Extrapolation Using Smooth and Low-Rank Graph Signal Features

This paper presents a graph signal-based approach for channel spatial extrapolation to tackle the increasing challenge of obtaining precise channel estimates, especially in large-scale data contexts. The objective of this method is to infer channel characteristics for the entire scenario by extrapolating locally observed channel data. Graph signals are constructed based on spatial locations, and smooth, low-rank spatio-temporal features are employed to capture spatial consistency principles and scatterer distributions. This strategy transforms the channel spatial extrapolation issue into a graph signal sampling recovery model. The proposed solution leverages the Alternating Direction Method of Multipliers (ADMM) algorithm. Furthermore, to alleviate computational overhead, we introduce an ADMM-Net framework, which integrates model-driven neural networks, thus reducing computation time without compromising accuracy. Simulation results demonstrate a significant decrease in extrapolation errors for the optimized model compared to traditional interpolation methods, particularly in scenarios with limited channel information.