Near-Field Beam Training for Extremely Large-Scale IRS

In this paper, we investigate codebook-based near-field beam training for extremely large-scale intelligent reflecting surface (XL-IRS). Compared with the conventional far-field beam training method that only searches for the best beam direction, the near-field beam training is more challenging since it requires a beam search over both the angular and distance domains due to the spherical wavefront propagation model. To reduce the near-field beam-training overhead of two-dimensional exhaustive search, we propose a novel two-layer codebook-based near-field beam training scheme that decomposes the two-dimensional search into two sequential phases. Specifically, the layer-l codebook designed based on the omnidirectivity of random-phase beam pattern is firstly employed to estimate the user distance. Then, given the estimated user distance of the layer-1, a customized layer-2 codebook is employed to scan the candidate locations of the user. Numerical results demonstrate that the proposed scheme can achieve more accurate estimation of the user distance and angle, as well as higher data rate with smaller training overhead, compared with benchmarks.