1: \begin{abstract}

2: Approximate Message Passing (AMP) has been shown

3: to be a superior method for inference problems, such

4: as the recovery of signals from

5: sets of noisy, lower-dimensionality measurements, both

6: in terms of reconstruction accuracy and in computational

7: efficiency. However, AMP suffers from serious convergence

8: issues in contexts that do not exactly match its assumptions.

9: We propose a new approach to stabilizing AMP in these

10: contexts by applying AMP updates to individual coefficients 

11: rather than in parallel. Our results show that this change

12: to the AMP iteration can provide theoretically expected, but

13: hitherto unobtainable, performance for problems on which the

14: standard AMP iteration diverges. Additionally, we find that

15: the computational costs of this \emph{swept} coefficient update

16: scheme is not unduly burdensome, allowing it to be applied 

17: efficiently to signals of large dimensionality.

18: \end{abstract}

19: