Reconfigurable holographic surfaces (RHS) have emerged as a promising enablers for Integrated Sensing and Communication (ISAC) owing to their ultra-dense electromagnetic control capability. In this work, we consider near-field multi-user localization as a key ISAC function. However, hardware-induced phase-shift constraints and complex-plane coupling effects severely degrade the performance of classical subspace-based estimators such as multiple signal classification (MUSIC). In this work, we analytically characterize the impact of practical phase-shift constraints on the MUSIC spectrum and reveal the resulting pseudo-peak and angular bias phenomena. To address these challenges, we propose a hierarchical sensing framework that integrates subspace estimation with sparsity-aware recovery. Specifically, we introduce a regularized orthogonal matching pursuit (ROMP)-assisted multi-sector sensing mechanism that exploits directional sparsity and sector-level structural partitioning of RHS, enabling robust multi-user angle-of-arrival (AoA) estimation under limited-resolution hardware. Furthermore, to alleviate near-field distance estimation bias, we design a nonlinear fusion estimator that aggregates angle observations from multiple local far-field sectors using confidence-weighted integration based on sector-specific priors. A confidence-aware sector selection scheme is also developed to suppress unreliable observations without relying on data-driven learning models. Extensive Monte Carlo simulations demonstrate that our method achieves over 80% reduction in root mean square error (RMSE) localization compared to conventional MUSIC and least-squares schemes, and maintains meter-level accuracy even under severe hardware distortions and sector limitations.

Reconfigurable holographic surfaces (RHS) are expected to play a key role in future mobile networks. However, the substantial increase in antenna aperture and operating frequency brings new challenges for near-field communication. We propose a near-field multi-user 3D hierarchical beam training scheme tailored for RHS-based multi-input multi-output (MIMO) systems, supporting both near-field and far-field user deployment, while considering hardware constraints. Since the hierarchical beam training scheme involves activating varying numbers of transmitting elements at each search layer, and RHS elements are densely packed, significant mutual coupling effects may arise. To mitigate this, we propose two element activation strategies: centered activation and sparse activation based on different RHS element positioning patterns within the hierarchical beam training framework. Furthermore, we design a practical beam training approach tailored to a hybrid digital–holographic architecture, optimized through an alternating algorithm that accounts for both binary and coupled amplitude-phase hardware constraints on RHS meta-elements. Simulation results demonstrate strong robustness under various hardware and channel state information (CSI) imperfections, achieving performance close to that of fully digital systems. Finally, we further analyse the asymptotic orthogonality of near-field beam focusing vectors under different RHS surface geometries. The results show that rectangular surfaces offer superior beam orthogonality for beams steered in the same direction but located at different distances.