Functions

IRLS( MeasurementMatrix , Measurement ; x...)

This function reconstructs a signal using equality constrained Lp minimization, where 0 < p <= 1. It accepts these options:

  • verbose = false - Print iteration and convergence information
  • maxiter = 1000 - The maximum number of iterations before giving up.
  • p = .5 - The p in Lp
  • theshold = 1e-5 - Threshold for convergence, the smaller the number the more the algorithm onverges.
  • eps = x->1/x^3 - A function that converges to 0 as x->Inf. The faster eps approaches 0 the faster the algorithm converges, however it also becomes more likely to fail. 1/x^3 seems to be something of a conservative setting, resulting in good convergence at the cost of taking slightly more time to run.
  • debug = false - Returns additional output.
UIRLS( MeasurementMatrix , Measurement ; x...)

This function reconstructs a signal using unconstrained Lp minimization, where 0 < p <= 1. This algorithm can generally reconstruct noisy signals better then the IRLS function. It accepts the same options as IRLS, with one addition:

  • lambda = 1 - This value is the tradeoff between finding a sparser solution and a solution that is closer to the measured output. It is in essence a measure of how much noise is expected in the measurement. Smaller lambdas would tend to mean there is less noise, so the solution approaches IRLS as lambda goes to 0. lambda>0. See [1]

These functions returns a vector of the reconstructed signal.

ZAP(Y, A ; x...)

Implementation of the algorith presented in [2]. This identifies a sparse solution to Multiple Measurement Vector, using a Zero-point Attracting Projection. The algorithm accepts a measured output matrix Y and a sampling/measurement matrix A, and returns a joint sparse input signal. Additional Options:

  • verbose = false - Print iteration and convergence information.
  • maxIter = 5000 - maximum number of iterations before giving up.
  • threshold = 1e-6 - threshold for convergence.
  • kappa = .1 - See [2] for full description.
  • neu = .1 - See [2] for full description.
  • Q = 11 - See [2] for full description.
  • alpha = 1 - See [2] for full description.
nGMCA( Y , r ; x...)

An implementation of the algorithm presented in [3]. This takes a matrix and performs blind source separation with the added restrictions that the source signals and mixing matrix be as sparse (L1 sparsity) as possible, and non-negative. In addition to the starting matrix, the number of components must be specified. Additional Options:

  • verbose = false - Print iteration and convergence information.
  • maxIter = 5000 - maximum number of iterations before giving up.
  • threshold = 1e-6 - threshold for convergence.
  • phaseRatio=0.30 - there are 2 stages of convergence, this is the ratio of time spent on either stage. This should probably not be edited except when algorithm isn't converging properly.
  • kickstart=true - this runs an initial fitting step without using the algorithm, greatly improving the speed of intial convergence.

It returns two matricies, that when multiplied together result in the closest to the input matrix as possible.

  • return {A,S}

Where Y=A*S

GI( X )

This is the absolute gini index, which is an excelent measure of sparsity, see [4]. This function accepts any vector or matrix of numbers and returns a value from 0 to 1, with 1 being as sparse as possible.

Coherence( X )

This is an implementation of the standard coherence measure used with the RIP, the definition can be seen in [5]. X must be a matrix.