1: \begin{abstract}

2:   We propose a Bayesian \gls{EM} algorithm for reconstructing

3:   Markov-tree sparse signals via belief propagation.  The measurements

4:   follow an underdetermined linear model where the

5:   regression-coefficient vector is the sum of an unknown approximately

6:   sparse signal and a zero-mean white Gaussian noise with an unknown

7:   variance.  The signal is composed of large- and small-magnitude

8:   components identified by binary state variables whose probabilistic

9:   dependence structure is described by a Markov tree. Gaussian priors

10:   are assigned to the signal coefficients given their state variables

11:   and the Jeffreys' noninformative prior is assigned to the noise

12:   variance.  Our signal reconstruction scheme is based on an \gls{EM}

13:   iteration that aims at maximizing the posterior distribution of the

14:   signal and its state variables given the noise variance.   We

15:   construct the missing data for the \gls{EM} iteration so that the

16:   complete-data posterior distribution corresponds to a \gls{HMT}

17:   probabilistic graphical model that contains no loops and implement

18:   its \gls{M} step via a max-product algorithm.  This \gls{EM}

19:   algorithm estimates the vector of state variables \emph{as well as}

20:   solves iteratively a linear system of equations to obtain the

21:   corresponding signal estimate.  We select the noise variance so that

22:   the corresponding estimated signal and state variables obtained upon

23:   convergence of the \gls{EM} iteration have the largest marginal

24:   posterior distribution.  We compare the proposed and existing

25:   state-of-the-art reconstruction methods via signal and image

26:   reconstruction experiments.

27: %

28: %

29: %

30: \end{abstract}

31: