Develop channel-prior learning for multisource AMP

Develop message-passing algorithms or expectation-maximization integrations with multisource approximate message passing to learn the channel priors required for activity detection and localization in frequency-selective cell-free random-access networks.

Background

The frequency-domain receiver uses multisource approximate message passing (AMP), whose denoising function depends on assumed channel statistics. The paper adopts simplified zero-mean Gaussian channel models with position-dependent large-scale fading coefficients because the detailed spatial covariance structure is generally unavailable to the receiver.

The authors note that expectation-maximization (EM)-AMP and related message-passing methods for learning channel priors have been studied in other settings, but that corresponding algorithms have not been developed for the multisource AMP framework used here. Such methods could reduce the mismatch between assumed and actual channel statistics and potentially improve activity detection and localization in realistic spatially consistent, frequency-selective channels.