Improved GPR-Based CSI Acquisition via Spatial-Correlation Kernel

TL;DR

This paper introduces a novel Gaussian process regression framework with a spatial-correlation kernel for improved wireless channel estimation, reducing pilot overhead and computational complexity while enhancing accuracy.

Contribution

It proposes a new SC kernel for GPR that explicitly models second-order channel statistics and derives an optimal estimator without assuming Gaussian channels.

Findings

Achieves up to 50% pilot overhead reduction.

Outperforms benchmark estimators in MSE and spectral efficiency.

Maintains lower computational complexity than MMSE.

Abstract

Accurate channel estimation with low pilot overhead and computational complexity is key to efficiently utilizing multi-antenna wireless systems. Motivated by the evolution from purely statistical descriptions toward physics- and geometry-aware propagation models, this work focuses on incorporating channel information into a Gaussian process regression (GPR) framework for improving the channel estimation accuracy. In this work, we propose a GPR-based channel estimation framework along with a novel Spatial-correlation (SC) kernel that explicitly captures the channel's second-order statistics. We derive a closed-form expression of the proposed SC-based GPR estimator and prove that its posterior mean is optimal in terms of minimum mean-square error (MMSE) under the same second-order statistics, without requiring the underlying channel distribution to be Gaussian. Our analysis reveals that, with up to 50% pilot overhead reduction, the proposed method achieves the lowest normalized mean-square error, the highest empirical 95% credible-interval coverage, and superior preservation of spectral efficiency compared to benchmark estimators, while maintaining lower computational complexity than the conventional MMSE estimator.