Spectral Efficiency Enhancement for Spatially Non-Stationary XL-RIS-Aided Systems With Dual-Side VR Estimation

Abstract

This letter investigates spatial non-stationarities at both the reconfigurable intelligent surface (RIS) and base station, where only partial arrays, referred to as the visibility region (VR), can receive user signals. To improve spectral efficiency (SE) in frequency division duplexing systems, we embed VR estimation into the uplink training phase, which reduces downlink training overhead by merely activating VR elements. Specifically, VR estimation is separately accomplished via energy detection at the base station and RIS. A deterministic approximation of the average ergodic SE is then derived using large-system analysis, whereas its complexity remains prohibitive due to exhaustive enumeration of all possible VR detection results. To address this, we further approximate the average SE by collecting only dominant detection results, thus reducing complexity significantly. Moreover, we leverage historical pilot energy over multiple coherence times to enhance VR detection accuracy. By maximizing the low-complexity approximation of average SE, we alternately optimize detection thresholds, pilot length for VR training, and the downlink RIS phase-shift matrix. Numerical results validate that the proposed scheme improves VR detection accuracy and can closely approach the SE performance of perfect VR knowledge.