Adaptive Beam Pairing and Local Interpolation for Robust Analytical Beam Training in RIS-Assisted Wideband THz Systems

Wideband terahertz (THz) communication systems offer ultra-high data rates and low latency for future sixth-generation (6G) networks, but they face severe path loss and beam split effects that degrade direction estimation in reconfigurable intelligent surface (RIS)-assisted deployments. Accurate beam training in these systems is particularly challenging under coarse codebooks and low signal-to-noise ratio (SNR) conditions. To overcome these limitations, we propose a robust analytical beam training framework, termed WBPC-AI, which integrates two key enhancements. First, a Weighted Beam Pair Combination (WBPC) mechanism aggregates information from the top k beam pairs using SNR-aware weighting, improving the stability of the estimation under noise and quantization. Second, an Adaptive Interpolation (AI) step refines direction estimates to sub-codebook resolution by exploiting parametric power asymmetries between neighboring beams. The proposed method avoids the drawbacks of hard decision-based schemes while maintaining low complexity. Simulation results demonstrate that WBPC-AI outperforms conventional analytical, hierarchical, and narrow-beam exhaustive search baselines in both angle estimation accuracy and achievable rate, particularly in low-SNR and large-RIS scenarios. Moreover, WBPC-AI exhibits strong scalability across RIS configurations and robustness under quantized phase control, validating its suitability for real-time RIS-assisted wideband THz systems.