Semi-Blind Channel Estimation for RIS-Assisted mmWave Systems Based on Diffusion Models

Abstract

Channel estimation in reconfigurable intelligent surface (RIS)-assisted millimeter-wave (mmWave) systems is challenging due to the high-dimensional nature of cascaded channels. Existing semi-blind schemes mitigate pilot overhead by exploiting unknown data symbols. Nevertheless, these schemes often fail to fully leverage channel priors. To address these limitations, we propose a diffusion model (DM)-based semi-blind posterior sampling (DMSBPS) algorithm for RIS-assisted mmWave channel estimation. The proposed algorithm utilizes a pre-trained DM as a learned generative prior and approximates the unknown data symbols as Gaussian random variables. A novel data likelihood score is then derived and combined with the pilot likelihood score and channel prior to jointly guide posterior sampling. Simulation results demonstrate that DMSBPS outperforms state-of-the-art pilot-based and semi-blind methods under limited pilot resources.