Online Recovery of Temporally Correlated Sparse Signals Using Multiple Measurement Vectors

This work addresses the problem of sequential recovery of temporally correlated sparse vectors with common support from noisy under-determined linear measurements. The Kalman sparse Bayesian learning (SBL) algorithm is an efficient tool for solving the problem when the temporal correlation is modeled using a first order autoregressive model. However, this method processes the input data in a batch mode, which results in high latency. We propose two online SBL algorithms which operate on the observations in a serial fashion. They are sequential expectation maximization (EM) schemes, implemented using fixed lag smoothing and sawtooth lag smoothing. The online algorithms require significantly lower computational and memory resources compared to their offline counterparts. Also, estimates of the sparse vectors become available after a fixed delay from the time observations arrive. Using Monte Carlo simulations, we illustrate that the mean square error and support recovery performance of the proposed algorithms is very close to the offline Kalman SBL algorithm.