Answers are grounded in this paper and cite their sources.
Reliable channel state information (CSI) acquisition in wideband beamspace massive multiple-input multiple-output (MIMO) systems is hampered by the beam-squint effect and the structural mismatch of conventional deep learning (DL) estimators. To address these challenges, we propose a physics-informed conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP) scheme, termed initial-estimate-enhanced Pix2pix (IE-Pix2pix), which integrates model-driven initialization with structure-preserving adversarial refinement. Specifically, a linear signal preprocessing (LSP) operator generates a physics-informed initial estimate as the conditional input, and a trajectory-aware loss is incorporated to preserve the frequency-dependent sparsity of the beamspace channel. Simulation results show that the proposed scheme attains an NMSE of −23.5 dB at 10 dB SNR, about 2.5 dB lower than the strongest DL baseline, and reaches −31.2 dB at 30 dB, about 2.2 dB lower than the strongest baseline while clearly avoiding the error floor of conventional compressive-sensing estimators. Moreover, the achievable spectral efficiency at 30 dB reaches about 166 bps/Hz, approximately 12 bps/Hz higher than the conditional GAN baseline and 98 bps/Hz higher than OMP.
No takes yet. Share an insight, caveat, or question.
Nie et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: