Sparse Recovery Experiments with Sparse Matrices

From Sparse

(Redirected from Sparse Recovery Experiments)
Jump to: navigation, search

Contents

Intro

This page contains the Matlab/C toolkit we used to get experimental results for sparse recovery using sparse matrices, reported in BGIKS08, BIR08 and BI08. The code provides the following functionality: (i) given an n-dimensional vector x, compute its m-dimensional sketch Ax, where A is an mxn sketch matrix, m << n, and (ii) given a sketch Ax, recover a k-sparse approximation x* of x. In addition, the code provides subroutines for generating test vectors x and sparse matrices A.

The toolkit supports the following recovery algorithms: Sparse Matching Pursuit (SMP), Countmin and L1 minimization (although one needs to download either l1-magic or GPSR to use the last option).

Code

The archive can be found here [1] . See the "Readme" file for instructions and descriptions of files and key functions.

Benchmark

The running times of the different recovery algorithms vs. signal length (n) are depicted below. The signal sparsity was set to k = 0.002n, the sketch size was set to m=0.1n. All methods used sparse 0-1 matrices A. Unless specified otherwise, the number d of ones per column was set to 10.

Sparse Recovery Experiments

Exact recovery plots show the empirical sketch sizes necessary to recover a sparse vector exactly. The plots are created by dividing the plot space into datapoints. For each datapoint, a matrix is randomly generated, and a number of experiments are carried out with this matrix. In each experiment, a sparse signal is generated, measured, and recovered. If the recovered vector is identical to the original vector, the experiment is deemed successful. The fraction of experiments which are successful for a datapoint is indicative of the probability that one can recover sparse vector exactly from a sketch with the corresponding parameters. The resulting data is used to create a contour plot, which shows the level curves.

The plot titles show some information about the plot: the type of measurement matrices, the type of test signals, and the method/algorithm for recovery. A description of the "codenames" for these is given below. The title also shows the resolution of the plot: the number of datapoints per each axis, and the number of experiments (trials) per datapoint.

Signal types:

  • plus_minus_one_peaks: truly sparse signal; the non-zero components are chosen randomly to be either +1 or -1.

Matrices:

  • countmin<d>: random Count-Min matrix (the kind of matrices used by the Count-Min algorithmCM04). It is a binary sparse matrix, which (with the right parameters) is with high probability an expander of degree d. The m rows are divided into d "row sections", and for each column there is one 1 placed randomly inside each section of each column)

Methods:

  • smp(<it>,<r>): SMPBIR04 recovery algorithm; it is the number of iterations, and r is a sparsity multiplier (SMP recovery sparsity is set to Kr, where K is the real sparsity)
  • countmin: CountMinCM04 recovery algorithm for general signals (using median)
  • gpsr: L1 minimization using GPSRFNW07, with a small tau value.
  • lp: L1 minimization using l1-MagicCR05.

SMP

Countmin


GPSR

l1-Magic

References

  • ^ CM04 G. Cormode and S. Muthukrishnan. Improved data stream summaries: The count-min sketch and its applications, FSTTCS, 2004.
  • ^ CR05 E.J. Candes and J. Romberg. l1-MAGIC: Recovery of Sparse Signals via Convex Programming, 2005.
  • ^ FNW07 M.A.T. Figueiredo, R.D. Nowak, and S.J. Wright. Gradient projection for sparse reconstruction: Application to compressed sensing and other inverse problems. IEEE Journal of Selected Topics in Signal Processing, 1, 2007.
Personal tools