📝 SummarXiv

Abstract

Holographic multiple-input multiple-output (HMIMO) has emerged as a potential technique for enhancing spectral efficiency (SE) while maintaining low hardware costs and power consumption. Conventional alternating optimization algorithms have been widely employed to design HMIMO beamforming, but their high computational complexity hinders real-time deployment. Graph neural networks (GNNs) provide a learning-based alternative with low inference time. However, in HMIMO systems, the fixed phase pattern, which captures the distribution of reference-wave phases across the holographic surface disrupts the permutation equivariance (PE) property, which is essential for GNN design. To address this issue, we reformulate the beamforming problem into a PE-compliant equivalent form and propose a novel network architecture consisting of a gradient-based GNN (GGNN) followed by two projection modules. Simulation results demonstrate that the proposed method achieves higher SE with significantly reduced inference time than alternating optimization methods and exhibits superior generalizability compared to other learning-based baselines.