3D Near-Field Beam Training for Uniform Planar Arrays: Low-Overhead Approaches Leveraging Beam Diverging
Abstract
Large-scale millimeter-wave arrays improve link budget and spatial resolution, but their narrow beams make beam alignment increasingly costly. The challenge is more severe for large uniform planar arrays (UPAs) operating in the near-field (NF), where the channel depends jointly on azimuth, elevation, and distance. Existing far-field steering or NF focusing codebooks therefore require either dense 3D search or additional hardware support, and their decisions are sensitive to location mismatch. This paper develops a low-overhead NF beam-training framework for UPAs based on beam-diverging codewords. A virtual focal point behind the array induces a controllable 3D high-power region in front of the array, which enables coarse localization using phase-only analog beams. We first characterize the resulting UPA beam-diverging geometry and construct a hierarchical codebook that recursively partitions the 3D service region. We then design a finite NF refinement codebook for the localized region and propose three training protocols: a single-radio frequency (RF)-chain two-phase method, a three-RF-chain uniform linear array (ULA)-assisted extension, and a five-RF-chain multipath extension. Numerical results show that the proposed methods achieve small SNR loss with substantially lower pilot overhead than focusing- or DFT-based baselines.