cpkrylov

A Matlab package for the solution of regularized saddle-point systems by constraint-preconditioned Krylov solvers

Authors

Daniela di Serafino, University of Campania "Luigi Vanvitelli", Caserta, Italy, [email protected].
Dominique Orban, GERAD and Polytechnique Montréal, QC, Canada, [email protected].

Last update

Version 1.0 - October 8, 2019.

Description

cpkrylov is a Matlab package implementing constraint-preconditioned variants of Krylov solvers for the solution of regularized saddle-point linear systems. The saddle-point matrix is assumed to be nonsingular; its leading block may be nonsymmetric, and its trailing block must be nonzero and symmetric. In particular, cpkrylov implements constraint-preconditioned variants of the CG, CG-Lanczos, MINRES, SYMMLQ, GMRES(l) and DQGMRES methods. Details on the solvers and the related constraint preconditioners are provided in

Daniela di Serafino and Dominique Orban, Constraint-Preconditioned Krylov Solvers for Regularized Saddle-Point Systems, SIAM Journal on Scientific Computing, 43 (2), 2021, pp. A1001-A1026.
Preprint available as Cahier du GERAD G-2019-72, GERAD, Montréal, QC, Canada (revised in July 2020), and from http://www.optimization-online.org/DB_HTML/2019/10/7411.html and https://arxiv.org/abs/1910.02552.

The leading block of the saddle-point matrix may be a matrix or a linear operator, but the other blocks must be explicit matrices. The constraint preconditioner P is implemented as a linear operator M such that M*z returns the solution of Pv=z. The operator M also implements iterative refinement and residual update, as suggested (for the case where the trailing block of the saddle-point matrix is zero) in

Nicholas I. M. Gould, Mary E. Hribar, and Jorge Nocedal, On the solution of equality constrained quadratic programming problems arising in optimization, SIAM Journal on Scientific Computing, 23(4), pp. 1376-1395, 2001.

The linear operators are defined using the Spot toolbox by Ewout van den Berg and Michael P. Friedlander. See https://github.com/mpf/spot.

Installation

cpkrylov runs under MATLAB (it has been tested under MATLAB 2018b) and requires the Spot toolbox (see https://github.com/mpf/spot). In order to use Spot, the master branch of the Spot repository must be added to the Matlab path, e.g., with the Matlab command

    addpath('path/to/spot/master')

The user must also add to the MATLAB path the root cpkrylov directory and its subdirectories kernels, ops and utils. This can be done by running cpk_path_setup.m.

Contents of the package

• reg_cpkrylov.m:
  main driver, which performs pre-processing operations, calls the requested solver, performs post-processing operations, and returns solutions and statistics to the user;
• kernels/cpcg.m:
  function implementing the constraint-preconditioned CG method;
• kernels/cpcglanczos.m:
  function implementing the constraint-preconditioned Lanczos version of CG;
• kernels/cpminres.m:
  function implementing the constraint-preconditioned MINRES method;
• kernels/cpsymmlq.m:
  function implementing the constraint-preconditioned SYMMLQ method;
• kernels/cpdqgmres.m:
  function implementing the constraint-preconditioned DQGMRES method;
• kernels/cpgmres.m:
  function implementing the constraint-preconditioned GMRES(l) method;
• ops/LDL2.m:
  operator representing the LDL factorization of a symmetric indefinite matrix with optional iterative refinement (needed to apply the constraint preconditioner);
• utils/SymGivens:
  function implementing a symmetric Given rotation (by M. A. Saunders and S.-C. Choi), called by cpdqgmres and cpgmres.

See the documentation inside each file for further details.

Examples of use of cpkrylov

Examples may be found in the examples directory.

• cpk_exprog1.m:
  example program 1, which runs cpminres (or cpgc, or cpcglanczos, or cpdqgmres) on the symmetric saddle-point linear system stored in cvxqp1_m_2x2_symm_iter10.mat;
• cpk_exprog2.m:
  example program 2, which runs cpgmres (or cpdqgmres) on the nonsymmetric saddle-point linear system stored in cvxqp2_s_3x3_nonsymm_perm_iter10.mat;
• cvxqp1_m_2x2_symm_iter10.mat, cvxqp2_s_3x3_nonsymm_perm_iter10.mat:
  MATLAB mat-files containing the data for the example programs.

License