EM-BG-AMP

Version History

Version 5.0 (11-23-15)

Made compatible with noise-variance learning implemented inside the EstimOut operators, allowing us to remove the EM loop.

Version 4.3 (6-14-13)

Changed to default option optEM.heavy_tailed = true and added the option optEM.robust_mat to handle cases when the sensing matrix gives issues.

Version 4.1 & 4.2 (6-05-13)

Various bugs and alterations were made to code:

Cleaned up GAMP warm starting to account for adaptive step size.
Various outputs (EMstate, param, xvar) now all collected in structure EMfin
Now user can alter GAMP and EM options without knowing all other defauls. For instance, if one wants to decrease the EM tolerance, they simply need to do EMopt.EMtol = 1e-5, and then call xhat = EMGMAMP(y,A,EMopt).
Switched the order of GAMPopt and EMopt as input parameters.
Decreased GAMP iterations to 4 and increased EMiterations to 60.

Version 4.0 (1-15-13)

Added the EMGMAMP Model order selection (EM-GM-AMP-MOS), which automatically selects the number of GM components L. To call simply use xhat = EMGMAMPMOS(y,A).

Version 3.0 (12-01-12)

Added option for estHist output. This structure contains per-iteration GAMP quantities such as the step size and the cost function.

Version 2.0 (4-18-12)

Updated for the MMV case. Now code is included in the GAMP sourceforge package (See download section).

plot_GM - Plots the estimated pdf for real or imaginary case.
EMBGAMPdemo - Updated to show plot_GM's usage.

Version 1.3 (12-03-11)

Fourth release of the EM-BG-AMP algorithm.

EMBGAMP - Allows toggling the learning of each parameter. Also initialized active variance (phi) with unnormalized matrices
EMOpt- Default initializations for learning of each parameter included.

Version 1.2 (11-21-11)

Third release of the EM-BG-AMP algorithm. Make sure to download new GAMP package as well. Contains the following updates

EMBGAMP - Changes made to noise variance parameter update equation. Better estimates for lower SNR.

Version 1.1 (11-2-11)

Second release of the EM-BG-AMP algorithm. Contains the following updates

generate_Amat.m - Now able to generate DFT operators.
EMBGAMP.m - Now handles either explicit A matrices or GAMP object operators.
EMBGAMPdemo.m - One example now uses oversampled DFT operator.

Version 1.0 (10-20-11)

First release of the EM-BG-AMP algorithm. Supports real and complex signal types for a single measurement vector (SMV). Includes the following features:

generate_Amat.m - Generates a random M by N matrix of specified type.
EMOpt.m - Initilizes optional EM inputs.
EMBGAMP.m - This is the main function which performs sparse signal estimation.
EMBGAMPdemo.m - This is an example demo file which demonstrates how to use EMBGAMP and interpret the results.