FedRep with PA-DNN recorded the highest reported weighted sum-rate, or WSR, in a simulated three-cell scenario, at 1.34 bps/Hz. The scenario distributed two, three and four users across the three cells. Traditional federated learning plateaued at 0.9 bps/Hz, EM-based personalized federated learning reached 1.29 bps/Hz, and pFedMe converged after 200 epochs.
The study's objective was to maximize WSR, reported in bits per second per hertz, under power constraints by jointly optimizing active beamforming at base stations and users' fluid-antenna port selection. In practical terms, the system chooses both how base stations shape their transmissions and which port each user's antenna uses.
A shared model with local choices
The proposed PA-DNN uses channel state information, or CSI, together with positional encoding for the antenna ports. It has two task-specific branches: one produces beamforming vectors and the other produces port selections.
Port selection is a discrete choice. The network uses a hard one-hot selection in its output calculation and a continuous Gumbel-Softmax relaxation while calculating the training update.
FedRep splits the learned parameters into shared and local portions. Beamforming-related parameters are shared among base stations, while port-selection parameters stay local for cell-specific adaptation. The two portions are updated through alternating block-coordinate descent, a routine that takes turns updating each group.
The simulated setup
The simulations used a 6 GHz carrier, a wavelength of 0.05 metres, eight transmit antennas at each base station and a maximum transmit power of 30 dBm. Each user had 25 candidate ports arranged in a square measuring 0.02 metres by 0.02 metres. The channel model included 30 multipath components, free-space path loss and Rayleigh fading.
PA-DNN and FedRep training used Adam, batches of 64 and 10,000 generated samples. The learning rate was reduced by a factor of 0.9975, and FedRep performed two local updates for every global epoch. Training stopped at a predefined round count or when convergence was reached.
Scores shifted with the scenario
In a separate seven-user comparison, PA-DNN was tested against fixed antenna, Portrandom and Maxnorm baselines. It had the highest reported WSR at 1.57 bps/Hz.
The reported training curves also changed rank over time. Maxnorm was above Portrandom during the first 50 epochs but fell behind after 100 epochs, while PA-DNN fully converged at 150 epochs.
Results varied by base station. For BS1, FedRep reached 0.42 bps/Hz, with EM-based PFL close behind at 0.41 bps/Hz. For BS2, FedRep reached 0.61 bps/Hz, compared with 0.52 bps/Hz for pFedMe and 0.44 bps/Hz for Traditional FL; EM-based PFL ranked second overall in that comparison.
BS3 showed a different balance between speed and final score. FedRep reached 0.31 bps/Hz by epoch 100, but pFedMe later overtook it around epoch 170 with 0.32 bps/Hz.
What the authors conclude
The authors interpret the simulation comparisons as support for explicit global and local parameter decoupling and position-aware learning, particularly in heterogeneous network conditions. They conclude that FedRep with PA-DNN jointly optimizes beamforming and port selection.
The paper is a preprint labeled arXiv:2608.25514v1 and dated 26 August 2026.
Paper data and sources
Original title: Joint Beamforming Design and Port Selection in Fluid Antenna-Assisted Multi-Cell Networks: A Personalized Federated Learning Approach
Authors: Liwen Gao, Li Zheng, Xing Hao et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text