Implicit Neural Representation of Beamforming for Continuous Aperture Array (CAPA) System

Abstract: In this paper, a learning-based approach for optimizing downlink beamforming in continuous aperture array (CAPA) systems is proposed, where a MIMO scenario that both the base station (BS) and the user are equipped with CAPA is considered. As the beamforming in the CAPA system is a function that maps a coordinate on the aperture to the beamforming weight at the coordinate, a DNN called BeaINR is proposed to parameterize this function, which is called implicit neural representation (INR). We further find that the optimal beamforming function lies in the subspace of channel function, i.e., it can be expressed as a weighted integral of channel function. Based on this finding, we propose another DNN called CoefINR to learn the weighting coefficient with INR, which has lower complexity than learning the beamforming function with BeaINR. Simulation results show that the proposed INR-based methods outperform numerical baselines in both spectral efficiency (SE) and inference time, with CoefINR offering additional training efficiency.