A Decomposed Learning Framework for Hybrid Precoding

Abstract

Learning-based hybrid precoding has emerged as a promising solution for low-latency multi-user transmission in large-scale antenna systems. However, its practical deployment is often hindered by high training overhead, and poor generalization across varying system sizes. These limitations fundamentally stem from the fact that most existing approaches learn coupled mappings defined over heterogeneous spaces under distinct constraints with a single neural network. To address this issue, we propose a learning framework that decomposes hybrid precoding into three structurally distinct sub-tasks: analog precoding, digital beamforming, and power allocation. This decomposition is grounded in a structural condition analysis, which reveals that the analog and digital precoders – the latter encompassing beamforming and power allocation – can be designed in partially separable forms. Based on these insights, we develop dedicated neural modules for each sub-task, explicitly leveraging their respective structural properties, and jointly train the modules in an end-to-end manner. Simulation results demonstrate that the proposed learning framework achieves spectral efficiency comparable to optimization-based benchmarks, while requiring significantly fewer floating-point operations for inference in large-scale settings (e.g., 1024 antennas). The proposed design also exhibits favorable generalizability across a range of system sizes, and can be readily adapted to alternative objectives such as energy efficiency maximization.