Scilab Home page | Wiki | Bug tracker | Forge | Mailing list archives | ATOMS | File exchange
Please login or create an account
Change language to: Français - Português - Русский - 日本語
Scilab Help >> ARnoldi PACKage (ARPACK binding) > eigs

eigs

calculates largest eigenvalues and eigenvectors of matrices

Syntax

d = eigs(A [,B [,k [,sigma [,opts]]]])
[d, v] = eigs(A [,B [,k [,sigma [,opts]]]])

d = eigs(Af, n [,B [,k [,sigma [,opts]]]])
[d, v] = eigs(Af, n [,B [,k [,sigma [,opts]]]])

Arguments

A

a full or sparse, real or complex, symmetric or non-symmetric square matrix

Af

a function

n

a scalar, defined only if A is a function

B

a sparse, real or complex, square matrix with same dimensions as A

k

an integer, number of eigenvalues to be computed

sigma

a real scalar or a string of length 2

opts

a structure

d

a real or complex eigenvalues vector or diagonal matrix (eigenvalues along the diagonal)

v

real or complex eigenvector matrix

Description

The purpose of the eigs function is to compute the largest eigenvalues of sparse, large matrices.

d = eigs(A) or d = eigs(Af, n)

solves the eigenvalue problem A * v = lambda * v. This calling returns a vector d containing the six largest magnitude eigenvalues. A is either a square matrix, which can be symmetric or non-symmetric, real or complex, full or sparse.

A should be represented by a function Af. In this instance, a scalar n designating the length of the vector argument, must be defined. It must have the following header :

function y=A(x)

This function Af must return one of the four following expressions :

  • A * x

    if sigma is not given or is a string other than 'SM'.

  • A \ x

    if sigma is 0 or 'SM'.

  • (A - sigma * I) \ x

    for the standard eigenvalue problem, where I is the identity matrix.

  • (A - sigma * B) \ x

    for the generalized eigenvalue problem.

[d, v] = eigs(A) or [d, v] = eigs(Af, n)

returns a diagonal matrix d containing the six largest magnitude eigenvalues on the diagonal. v is a n by six matrix whose columns are the six eigenvectors corresponding to the returned eigenvalues.

d = eigs(A, B)

solves the generalized eigenvalue problem A * v = lambda * B * v with positive, definite matrix B.

  • if B is not specified, B = [] is used.

  • if B is specified, B must be the same size as A.

d = eigs(A, B, k)

returns in vector d the k eigenvalues. If k is not specified, k = min(n, 6), where n is the row number of A.

d = eigs(A, B, k, sigma)

returns in vector d the k eigenvalues determined by sigma. sigma can be either a real or complex including 0 scalar or string. If sigma is a string of length 2, it takes one of the following values :

  • 'LM' compute the k largest in magnitude eigenvalues (by default).

  • 'SM' compute the k smallest in magnitude eigenvalues (same as sigma = 0).

  • 'LA' compute the k Largest Algebraic eigenvalues, only for real symmetric problems.

  • 'SA' compute the k Smallest Algebraic eigenvalues, only for real symmetric problems.

  • 'BE' compute k eigenvalues, half from each end of the spectrum, only for real symmetric problems.

  • 'LR' compute the k eigenvalues of Largest Real part, only for real non-symmetric or complex problems.

  • 'SR' compute the k eigenvalues of Smallest Real part, only for real non-symmetric or complex problems.

  • 'LI' compute the k eigenvalues of Largest Imaginary part, only for real non-symmetric or complex problems.

  • 'SI' compute the k eigenvalues of Smallest Imaginary part, only for real non-symmetric or complex problems.

d = eigs(A, B, k, sigma, opts)

If the opts structure is specified, different options can be used to compute the k eigenvalues :

  • tol

    required convergence tolerance. By default, tol = %eps.

  • maxiter

    maximum number of iterations. By default, maxiter = 300.

  • ncv

    number of Lanzcos basis vectors to use. For real non-symmetric problems, the ncv value must be greater or equal than 2 * k + 1 and, by default, ncv = min(max(2 * k + 1, 20), nA). For real symmetric or complex problems, ncv must be greater or equal 2 * k and, by default, ncv = min(max(2 * k, 20), nA) with nA = size(A, 2).

  • resid

    starting vector whose contains the initial residual vector, possibly from a previous run. By default, resid is a random initial vector.

  • cholB

    if chol(B) is passed rather than B. By default, cholB is %f.

  • isreal

    if Af is given, isreal can be defined. By default, isreal is %t. This argument must not be indicated if A is a matrix.

  • issym

    if Af is given, issym can be defined. By default, issym is %f. This argument must not be indicated if A is a matrix.

References

This function is based on the ARPACK package written by R. Lehoucq, K. Maschhoff, D. Sorensen, and C. Yang.

  • DSAUPD and DSEUPD routines for real symmetric problems,

  • DNAUPD and DNEUPD routines for real non-symmetric problems.

  • ZNAUPD and ZNEUPD routines for complex problems.

Example for real symmetric problems

clear opts
A            = diag(10*ones(10,1));
A(1:$-1,2:$) = A(1:$-1,2:$) + diag(6*ones(9,1));
A(2:$,1:$-1) = A(2:$,1:$-1) + diag(6*ones(9,1));

B = eye(10,10);
k = 8;
sigma = 'SM';
opts.cholB = %t;

d = eigs(A)
[d, v] = eigs(A)

d = eigs(A, B, k, sigma)
[d, v] = eigs(A, B, k, sigma)

d = eigs(A, B, k, sigma, opts)
[d, v] = eigs(A, B, k, sigma, opts)

// With sparses
AS = sparse(A);
BS = sparse(B);

d = eigs(AS)
[d, v] = eigs(AS)

d = eigs(AS, BS, k, sigma)
[d, v] = eigs(AS, BS, k, sigma)

d = eigs(AS, BS, k, sigma, opts)
[d, v] = eigs(AS, BS, k, sigma, opts)

// With function
clear opts
function y=fn(x)
   y = A * x;
endfunction

opts.isreal = %t;
opts.issym = %t;

d = eigs(fn, 10, [], k, 'LM', opts)

function y=fn(x)
   y = A \ x;
endfunction

d = eigs(fn, 10, [], k, 'SM', opts)

function y=fn(x)
   y = (A - 4 * eye(10,10)) \ x;
endfunction

d = eigs(fn, 10, [], k, 4, opts)

Example for real non-symmetric problems

clear opts
   A            = diag(10*ones(10,1));
   A(1:$-1,2:$) = A(1:$-1,2:$) + diag(6*ones(9,1));
   A(2:$,1:$-1) = A(2:$,1:$-1) + diag(-6*ones(9,1));

   B = eye(10,10);
   k = 8;
   sigma = 'SM';
   opts.cholB = %t;
   
   d = eigs(A)
[d, v] = eigs(A)
   
   d = eigs(A, B, k, sigma)
   [d, v] = eigs(A, B, k, sigma) 

   d = eigs(A, B, k, sigma, opts)
   [d, v] = eigs(A, B, k, sigma, opts)

// With sparses
   AS = sparse(A);
   BS = sparse(B);

d = eigs(AS)
[d, v] = eigs(AS)
   d = eigs(AS, BS, k, sigma)
   [d, v] = eigs(AS, BS, k, sigma)

   d = eigs(AS, BS, k, sigma, opts)
   [d, v] = eigs(AS, BS, k, sigma, opts)
   
   // With function
clear opts
function y=fn(x)
  y = A * x;
endfunction

opts.isreal = %t;
opts.issym = %f;

d = eigs(fn, 10, [], k, 'LM', opts)

function y=fn(x)
  y = A \ x;
endfunction

d = eigs(fn, 10, [], k, 'SM', opts)

function y=fn(x)
  y = (A - 4 * eye(10,10)) \ x;
endfunction

d = eigs(fn, 10, [], k, 4, opts)

Example for complex problems

clear opts
   A            = diag(10*ones(10,1) + %i * ones(10,1));
   A(1:$-1,2:$) = A(1:$-1,2:$) + diag(6*ones(9,1));
   A(2:$,1:$-1) = A(2:$,1:$-1) + diag(-6*ones(9,1));

   B = eye(10,10);
   k = 8;
   sigma = 'LM';
   opts.cholB = %t;
   
   d = eigs(A)
[d, v] = eigs(A)

   d = eigs(A, B, k, sigma)
   [d, v] = eigs(A, B, k, sigma)
   d = eigs(A, B, k, sigma, opts)
   [d, v] = eigs(A, B, k, sigma, opts)
   
   // With sparses
   AS = sparse(A);
   BS = sparse(B);
   
   d = eigs(AS)
[d, v] = eigs(AS)

   d = eigs(AS, BS, k, sigma)
   [d, v] = eigs(AS, BS, k, sigma)

   d = eigs(AS, BS, k, sigma, opts)
   [d, v] = eigs(AS, BS, k, sigma, opts)
   
   // With function
clear opts
function y=fn(x)
  y = A * x;
endfunction

opts.isreal = %f;
opts.issym = %f;

d = eigs(fn, 10, [], k, 'LM', opts)

function y=fn(x)
  y = A \ x;
endfunction

d = eigs(fn, 10, [], k, 'SM', opts)

function y=fn(x)
  y = (A - 4 * eye(10,10)) \ x;
endfunction

d = eigs(fn, 10, [], k, 4, opts)

See Also

  • spec — eigenvalues of matrices and pencils

History

VersionDescription
5.4.0 Function introduced. Deprecates dnaupd, dneupd, dsaupd, dseupd, znaupd and zneupd.

Comments

Author : Francois Granade posted the 17/06/2015 14:03
The help says "See Also: spec" and you may wonder what are the differences.

As far as I know, Eigs is an interface to the ARnoldi PACKage module. Eigs is good at computation on sparse matrices, whereas spec is more targeted at full matrices. ARPACK's cleverness is to choose the right algorithme based on the characteristics of the sparse matrix (symmetry, among others).

Additional opinions welcome...
Reply to this comment
Please login to comment this page


Add a comment:
Please login to comment this page.

Report an issue
<< dseupd ARnoldi PACKage (ARPACK binding) znaupd >>

Scilab Enterprises
Copyright (c) 2011-2015 (Scilab Enterprises)
Copyright (c) 1989-2012 (INRIA)
Copyright (c) 1989-2007 (ENPC)
with contributors
Last updated:
Wed Jun 15 08:27:37 CEST 2016