1: \begin{abstract}

2: We investigate the problem of recovering a structured sparse signal

3: from a linear observation model with an uncertain dynamic grid in

4: the sensing matrix. The state-of-the-art expectation maximization

5: based compressed sensing (EM-CS) methods, such as turbo compressed

6: sensing (Turbo-CS) and turbo variational Bayesian inference (Turbo-VBI),

7: have a relatively slow convergence speed due to the double-loop iterations

8: between the E-step and M-step. Moreover, each inner iteration in the

9: E-step involves a high-dimensional matrix inverse in general, which

10: is unacceptable for problems with large signal dimensions or real-time

11: calculation requirements. Although there are some attempts to avoid

12: the high-dimensional matrix inverse by majorization minimization,

13: the convergence speed and accuracy are often sacrificed. To better

14: address this problem, we propose an alternating estimation framework

15: based on a novel subspace constrained VBI (SC-VBI) method, in which

16: the high-dimensional matrix inverse is replaced by a low-dimensional

17: subspace constrained matrix inverse (with the dimension equal to the

18: sparsity level). We further prove the convergence of the SC-VBI to

19: a stationary solution of the Kullback-Leibler divergence minimization

20: problem. Simulations demonstrate that the proposed SC-VBI algorithm

21: can achieve a much better tradeoff between complexity per iteration,

22: convergence speed, and performance compared to the state-of-the-art

23: algorithms.

24: \end{abstract}