SLRA: Software for weighted structured low-rank approximation

Overview

SLRA is a C software with interface to MATLAB, Octave, and R for solving low-rank approximation problems

 text{minimize} quad text{over}  widehat p quad |p - widehat p|_w quad text{subject to} quad text{rank}({cal S}(widehat p)) leq r,

with the following features:

 |p|_w = textstylesqrt{sum_{i=1}^{n_p} w_i p_i^2},
  • fixed elements in the approximating matrix,

  • missing elements in the data matrix,

  • linear constraints on an approximating matrix's left kernel basis.

For an mtimes n matrix {cal S}(widehat p), with m < n, the computational complexity of the cost function and derivative evaluation is O(m^2n), so that the package is suitable for applications with ngg m. Problems with n_p leq n(m - r) can not be solved by the method, used in the software.

Usage

In MATLAB/Octave, once the structure and approximation criterion are specified in a variable s (see examples and the manual), the problem is solved by a call of the slra function:

[ph, info] = slra(p, s, r, opt);
  • The optional parameter opt is used to specify an initial approximation and options for the optimization method.

  • The output parameter info contains information about the convergence of the optimization method.

System identification

A major application area of mosaic Hankel structured low-rank approximation methods is system identification. The slra package includes a wrapper function ident for identification of linear time-invariant systems. For details see the ident package.

Download and installation

Latest version

Installation

  • For MATLAB/Octave users: unpack the archive and use addpath in MATLAB/Octave

  • For R users (Linux): unpack the archive and run make R in console

For advanced users/developers

  • The instructions how to compile from source are contained here

Citing

@Article{slra-software,
  author = 	 {I. Markovsky and K. Usevich},
  title = 	 {Software for weighted structured low-rank approximation},
  journal =      {J. Comput. Appl. Math.},
  volume =       {256},
  pages =        {278--292},
  year = 	 {2014},
}