Path-Level Radio Map-Aided Fast and Robust Channel Estimation for Pilot-Starved MIMO-OFDM Systems

TL;DR

This paper introduces CHARM, a novel framework for fast, robust channel estimation in MIMO-OFDM systems that leverages path-level radio maps to significantly reduce computational complexity and improve accuracy under pilot-starved conditions.

Contribution

The paper proposes a new ADPS prior extraction method from radio maps, enabling one-dimensional angle-of-departure search and a trust-region constraint to enhance robustness against dictionary mismatch.

Findings

Achieves accuracy comparable to 3D joint OMP with 34.8x speedup at T ≤ 4.

Trust-region variant degrades only 3.7 dB under severe mismatch, versus 8.2 dB without it.

Significantly reduces computational cost in pilot-starved MIMO-OFDM channel estimation.

Abstract

Accurate channel estimation in massive multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems is challenging when the number of pilot symbols is much smaller than the number of transmit antennas. Conventional compressed sensing methods perform a three-dimensional search over the angle-of-arrival, angle-of-departure, and delay domains, which incurs high computational cost. In this paper, we propose CHARM (channel estimation with angular-delay radio map), a framework that extracts an angular-delay power spectrum (ADPS) prior from path-level radio maps. The ADPS identifies the joint angle-of-arrival and delay support of the dominant multipath components offline, reducing the online estimation to a one-dimensional angle-of-departure search per path. A trust-region constraint is further introduced to prevent sub-grid refinement from diverging under dictionary mismatch. Simulation results show that CHARM achieves accuracy comparable to three-dimensional joint orthogonal matching pursuit (OMP) with $34.8\times$ speedup at pilot length $T \leq 4$, and that the trust-region variant degrades by only 3.7~dB under severe dictionary mismatch of 0.2~rad standard deviation, compared with 8.2~dB without the constraint.