Decentralized regression with asynchronous sub-Nyquist sampling

When capturing data on a sensor field to uncover its latent structure, there are often nuisance parameters in the observation model that turn even linear regression problems into non-convex optimizations. One common case is the lack of common timing source in ADCs, therefore samplings are done with time offsets. Motivated by the desire of estimating jointly the sensor field and nuisance parameters in a wide area deployment, this paper derives a new decentralized algorithm that combines alternating optimization and gossip-based learning. The proposed algorithm is shown to converge to the neighborhood of a local minimum, both analytically and empirically.