Exjobbspresentation: Channel Estimation for Extra-Large (XL) MIMO: Joint Visibility-Region and Polar-Domain Group Sparse Estimation with Coarse-to-Fine Refinement

Exjobbet genomfördes på Tieto Evry med Bengt Halllinger som handledare och Juan Vidal Alegría som akademisk handledare. Ove Edfors är examinator.

XL-MIMO systems place hundreds of antenna elements across a physically large aperture, enabling high spatial resolution for 6G deployments. At sub-6 GHz carrier frequencies and user distances below roughly 30~m, the aperture is large enough that two propagation effects dominate channel structure: the wavefront arriving from a near-field source is spherical rather than planar, so both range and angle determine the phase at each element; and the channel is spatially non-stationary, meaning that different groups of antennas receive meaningfully different signal levels from the same user. Standard channel estimation methods, least squares, scalar LMMSE, and DFT beamspace compressed sensing, were designed for neither effect.

This thesis develops a practical uplink channel estimation framework that treats both effects as a structure to exploit. Three components are combined: a non-oracle visibility region detector that identifies each user's active antenna subset from per-antenna pilot energy; a polar-domain dictionary parameterised by range, azimuth, and elevation, with steering vectors weighted by the amplitude term $1/r^{\gamma}$ to reflect the spherical-wave fall-off; and a joint recovery step using simultaneous orthogonal matching pursuit that estimates a common polar-domain support across nearby users who share the same dominant scatterers. A coarse-to-fine two-pass variant reduces off-grid mismatch without proportionally increasing computation.

Monte Carlo results averaged over 5000 independent realisations show consistent improvement over least squares, oracle VR-Wiener, beamspace compressed LS, and per-user polar OMP, in both normalised mean-square error and uplink sum-rate under maximum ratio combining. Notably, a group-SOMP beamspace baseline, which applies joint processing in the DFT domain, performs below independent per-user polar OMP, confirming that a well-matched physical model is a prerequisite for joint recovery to be effective. The proposed framework requires no second order channel statistics, no oracle visibility region knowledge, and no additional pilot overhead, yet achieves near-oracle estimation accuracy at moderate-to-high SNR, making it directly applicable to practical 6G XL-MIMO deployments without a separate channel statistics acquisition  phase.