Algebraic multilevel Krylov methods

Yogi A. Erlangga, Reinhard Nabben

Research output: Contribution to journalArticle

13 Citations (Scopus)

Abstract

In [Erlangga and Nabben, SIAM J. Sci. Comput., 30 (2008), pp. 1572-1595], we developed a new type of multilevel method, called the multilevel Krylov (MK) method, to solve linear systems of equations. The basic idea of this type of method is to shift small eigenvalues that are responsible for slow convergence of Krylov methods to an a priori fixed constant. This shifting of the eigenvalues is similar to projection-type methods and is achieved via the solution of subspace or coarse level systems. Numerical results show that MK works very well for the two-dimensional (2D) Poisson and convection-diffusion equation, i.e., the convergence can be made almost independent of the grid size h and the physical parameter involved. In a follow-up paper we have used MK in the context of the preconditioned Helmholtz equation. For this problem, we show that the convergence can be made only mildly dependent on the wavenumber. Even though the coarse level system of MK is algebraically related to that of multigrid, the construction of interlevel transfer operators does not require a smoothness condition; MK requires only that the interlevel transfer matrices are full rank. We will highlight in detail the differences between multigrid methods on one side and MK methods on the other side. This means that many multigrid transfer operators are more than sufficient for MK and henceforth can potentially be used in the MK framework. Motivated by the evolution of geometric to algebraic multigrid methods (AMG), in this paper we bring the algebraic way of choosing the coarse level system and the transfer operators into the MK context. We evaluate two techniques common in AMG: the interpolation-based technique, in particular that of Ruge and Sẗuben, and the aggregation-based technique. The resulting MK method is thus called the algebraic multilevel Krylov method (AMK). Both techniques are tested for various matrices arising from discretization of some 2D diffusion and convection-diffusion problems, as well as for several matrices taken from the matrix-market collections. The numerical results show that AMK works as well as the geometric MK methods. For the convection-diffusion equations, AMK leads to better convergence rates than Ruge-Sẗuben's AMG with lexicographic Gauss-Seidel relaxation.

Original languageEnglish
Pages (from-to)3417-3437
Number of pages21
JournalSIAM Journal on Scientific Computing
Volume31
Issue number5
DOIs
Publication statusPublished - 2009
Externally publishedYes

Fingerprint

Krylov Methods
Multilevel Methods
Algebraic multigrid Method
Transfer Operator
Helmholtz equation
Convection-diffusion Equation
Linear systems
Interpolation
Agglomeration
Numerical Results
Gauss-Seidel
Convection-diffusion Problems
Smallest Eigenvalue
Linear system of equations
Multigrid Method
Transfer Matrix
Helmholtz Equation
Convergence Rate
Convection
Smoothness

Keywords

  • Algebraic multigrid
  • Convectiondiffusion equation
  • Diffusion equation
  • GMRES
  • Multilevel Krylov

ASJC Scopus subject areas

  • Applied Mathematics
  • Computational Mathematics

Cite this

Algebraic multilevel Krylov methods. / Erlangga, Yogi A.; Nabben, Reinhard.

In: SIAM Journal on Scientific Computing, Vol. 31, No. 5, 2009, p. 3417-3437.

Research output: Contribution to journalArticle

Erlangga, Yogi A. ; Nabben, Reinhard. / Algebraic multilevel Krylov methods. In: SIAM Journal on Scientific Computing. 2009 ; Vol. 31, No. 5. pp. 3417-3437.
@article{ee31467f328d4ed39ba635c2de59b24e,
title = "Algebraic multilevel Krylov methods",
abstract = "In [Erlangga and Nabben, SIAM J. Sci. Comput., 30 (2008), pp. 1572-1595], we developed a new type of multilevel method, called the multilevel Krylov (MK) method, to solve linear systems of equations. The basic idea of this type of method is to shift small eigenvalues that are responsible for slow convergence of Krylov methods to an a priori fixed constant. This shifting of the eigenvalues is similar to projection-type methods and is achieved via the solution of subspace or coarse level systems. Numerical results show that MK works very well for the two-dimensional (2D) Poisson and convection-diffusion equation, i.e., the convergence can be made almost independent of the grid size h and the physical parameter involved. In a follow-up paper we have used MK in the context of the preconditioned Helmholtz equation. For this problem, we show that the convergence can be made only mildly dependent on the wavenumber. Even though the coarse level system of MK is algebraically related to that of multigrid, the construction of interlevel transfer operators does not require a smoothness condition; MK requires only that the interlevel transfer matrices are full rank. We will highlight in detail the differences between multigrid methods on one side and MK methods on the other side. This means that many multigrid transfer operators are more than sufficient for MK and henceforth can potentially be used in the MK framework. Motivated by the evolution of geometric to algebraic multigrid methods (AMG), in this paper we bring the algebraic way of choosing the coarse level system and the transfer operators into the MK context. We evaluate two techniques common in AMG: the interpolation-based technique, in particular that of Ruge and Sẗuben, and the aggregation-based technique. The resulting MK method is thus called the algebraic multilevel Krylov method (AMK). Both techniques are tested for various matrices arising from discretization of some 2D diffusion and convection-diffusion problems, as well as for several matrices taken from the matrix-market collections. The numerical results show that AMK works as well as the geometric MK methods. For the convection-diffusion equations, AMK leads to better convergence rates than Ruge-Sẗuben's AMG with lexicographic Gauss-Seidel relaxation.",
keywords = "Algebraic multigrid, Convectiondiffusion equation, Diffusion equation, GMRES, Multilevel Krylov",
author = "Erlangga, {Yogi A.} and Reinhard Nabben",
year = "2009",
doi = "10.1137/080731657",
language = "English",
volume = "31",
pages = "3417--3437",
journal = "SIAM Journal of Scientific Computing",
issn = "1064-8275",
publisher = "Society for Industrial and Applied Mathematics Publications",
number = "5",

}

TY - JOUR

T1 - Algebraic multilevel Krylov methods

AU - Erlangga, Yogi A.

AU - Nabben, Reinhard

PY - 2009

Y1 - 2009

N2 - In [Erlangga and Nabben, SIAM J. Sci. Comput., 30 (2008), pp. 1572-1595], we developed a new type of multilevel method, called the multilevel Krylov (MK) method, to solve linear systems of equations. The basic idea of this type of method is to shift small eigenvalues that are responsible for slow convergence of Krylov methods to an a priori fixed constant. This shifting of the eigenvalues is similar to projection-type methods and is achieved via the solution of subspace or coarse level systems. Numerical results show that MK works very well for the two-dimensional (2D) Poisson and convection-diffusion equation, i.e., the convergence can be made almost independent of the grid size h and the physical parameter involved. In a follow-up paper we have used MK in the context of the preconditioned Helmholtz equation. For this problem, we show that the convergence can be made only mildly dependent on the wavenumber. Even though the coarse level system of MK is algebraically related to that of multigrid, the construction of interlevel transfer operators does not require a smoothness condition; MK requires only that the interlevel transfer matrices are full rank. We will highlight in detail the differences between multigrid methods on one side and MK methods on the other side. This means that many multigrid transfer operators are more than sufficient for MK and henceforth can potentially be used in the MK framework. Motivated by the evolution of geometric to algebraic multigrid methods (AMG), in this paper we bring the algebraic way of choosing the coarse level system and the transfer operators into the MK context. We evaluate two techniques common in AMG: the interpolation-based technique, in particular that of Ruge and Sẗuben, and the aggregation-based technique. The resulting MK method is thus called the algebraic multilevel Krylov method (AMK). Both techniques are tested for various matrices arising from discretization of some 2D diffusion and convection-diffusion problems, as well as for several matrices taken from the matrix-market collections. The numerical results show that AMK works as well as the geometric MK methods. For the convection-diffusion equations, AMK leads to better convergence rates than Ruge-Sẗuben's AMG with lexicographic Gauss-Seidel relaxation.

AB - In [Erlangga and Nabben, SIAM J. Sci. Comput., 30 (2008), pp. 1572-1595], we developed a new type of multilevel method, called the multilevel Krylov (MK) method, to solve linear systems of equations. The basic idea of this type of method is to shift small eigenvalues that are responsible for slow convergence of Krylov methods to an a priori fixed constant. This shifting of the eigenvalues is similar to projection-type methods and is achieved via the solution of subspace or coarse level systems. Numerical results show that MK works very well for the two-dimensional (2D) Poisson and convection-diffusion equation, i.e., the convergence can be made almost independent of the grid size h and the physical parameter involved. In a follow-up paper we have used MK in the context of the preconditioned Helmholtz equation. For this problem, we show that the convergence can be made only mildly dependent on the wavenumber. Even though the coarse level system of MK is algebraically related to that of multigrid, the construction of interlevel transfer operators does not require a smoothness condition; MK requires only that the interlevel transfer matrices are full rank. We will highlight in detail the differences between multigrid methods on one side and MK methods on the other side. This means that many multigrid transfer operators are more than sufficient for MK and henceforth can potentially be used in the MK framework. Motivated by the evolution of geometric to algebraic multigrid methods (AMG), in this paper we bring the algebraic way of choosing the coarse level system and the transfer operators into the MK context. We evaluate two techniques common in AMG: the interpolation-based technique, in particular that of Ruge and Sẗuben, and the aggregation-based technique. The resulting MK method is thus called the algebraic multilevel Krylov method (AMK). Both techniques are tested for various matrices arising from discretization of some 2D diffusion and convection-diffusion problems, as well as for several matrices taken from the matrix-market collections. The numerical results show that AMK works as well as the geometric MK methods. For the convection-diffusion equations, AMK leads to better convergence rates than Ruge-Sẗuben's AMG with lexicographic Gauss-Seidel relaxation.

KW - Algebraic multigrid

KW - Convectiondiffusion equation

KW - Diffusion equation

KW - GMRES

KW - Multilevel Krylov

UR - http://www.scopus.com/inward/record.url?scp=73349110451&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=73349110451&partnerID=8YFLogxK

U2 - 10.1137/080731657

DO - 10.1137/080731657

M3 - Article

VL - 31

SP - 3417

EP - 3437

JO - SIAM Journal of Scientific Computing

JF - SIAM Journal of Scientific Computing

SN - 1064-8275

IS - 5

ER -