Sparse Channel Reconstruction: A Generative-Adversarial-Network-Based Approach

Wireless channels typically exhibit a sparse structure, with sparsity regarded as an inherent characteristics. A deeper understanding and accurate representation of channel sparsity contribute to a more comprehensive revelation of channel features and structure, thereby enhancing the performance of communication systems. However, current research on channel sparsity remains limited to the measurement, and traditional channel models have not sufficiently incorporated this critical characteristic. In this paper, considering channel deep sparse characteristics, a generative adversarial network (GAN)-based channel sparsity-tunable framework is proposed to reconstruct channel samples with various levels of sparsity. Specifically, continuous sparsity factors and a sparsity measurement module are introduced to manipulate and predict the sparsity of the reconstructed channel in the framework design. Channel power and delay are employed for model training to better consider the channel behavior of sparsity. Furthermore, the mutual information between input sparsity factors and reconstructed channels is enhanced to ensure that the sparsity factor has an actual impact on the level of reconstructed channel sparsity, thereby enabling the flexible and continuous reconstruction of channels with varying levels of sparsity. Simulation results prove that the proposed framework can effectively and accurately reconstruct channels with different levels of sparsity by changing sparsity factor, as well as providing good performance in measuring the sparsity of wireless channel. Finally, sparse manipulation performance is further validated using channel capacity simulation.