About AMP-MMV

AMP-MMV is a Bayesian message passing algorithm that solves the multiple measurement vector (MMV) problem, in which a matrix of noisy measurements, Y, is acquired from a sparse signal matrix, X, through the linear measurement process Y = AX + E, where A is a measurement matrix (typically with more columns than rows), and E is corrupting noise. Each column in X can represent the value of some time-varying signal at a particular instant in time, or can represent a time-invariant signal acquired in one of multiple channels. The unique feature of the MMV problem is the assumption that each column of X shares the same support, i.e., X is row-sparse.

Benefits of AMP-MMV:
• AMP-MMV works really quickly, especially in high-dimensions!
• Designed for cases with significant column correlations in X
• Model parameters learned automatically from data
• Performs near theoretical bounds for many problems
• Support for implicit matrix operators, e.g., FFTs

AMP-MMV has been implemented in MATLAB, and has been shown to work extremely quickly, requiring only simple matrix-vector products to perform its computations. Click here to try it out.

AMP-MMV: An algorithm for efficient high-dimensional inference in the multiple measurement vector (MMV) problem

Authors

• Justin Ziniel
• Philip Schniter

See Also:

• Nuit Blanche
• Rice CS Resources
• T-MSBL Homepage
• IPS Lab