1 ## Copyright (C) 2004-2012 Piotr Krzyzanowski
3 ## This file is part of Octave.
5 ## Octave is free software; you can redistribute it and/or modify it
6 ## under the terms of the GNU General Public License as published by
7 ## the Free Software Foundation; either version 3 of the License, or (at
8 ## your option) any later version.
10 ## Octave is distributed in the hope that it will be useful, but
11 ## WITHOUT ANY WARRANTY; without even the implied warranty of
12 ## MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 ## General Public License for more details.
15 ## You should have received a copy of the GNU General Public License
16 ## along with Octave; see the file COPYING. If not, see
17 ## <http://www.gnu.org/licenses/>.
20 ## @deftypefn {Function File} {@var{x} =} pcg (@var{A}, @var{b}, @var{tol}, @var{maxit}, @var{m1}, @var{m2}, @var{x0}, @dots{})
21 ## @deftypefnx {Function File} {[@var{x}, @var{flag}, @var{relres}, @var{iter}, @var{resvec}, @var{eigest}] =} pcg (@dots{})
23 ## Solve the linear system of equations @code{@var{A} * @var{x} = @var{b}}
24 ## by means of the Preconditioned Conjugate Gradient iterative
25 ## method. The input arguments are
29 ## @var{A} can be either a square (preferably sparse) matrix or a
30 ## function handle, inline function or string containing the name
31 ## of a function which computes @code{@var{A} * @var{x}}. In principle
32 ## @var{A} should be symmetric and positive definite; if @code{pcg}
33 ## finds @var{A} to not be positive definite, you will get a warning
34 ## message and the @var{flag} output parameter will be set.
37 ## @var{b} is the right hand side vector.
40 ## @var{tol} is the required relative tolerance for the residual error,
41 ## @code{@var{b} - @var{A} * @var{x}}. The iteration stops if
42 ## @code{norm (@var{b} - @var{A} * @var{x}) <=
43 ## @var{tol} * norm (@var{b} - @var{A} * @var{x0})}.
44 ## If @var{tol} is empty or is omitted, the function sets
45 ## @code{@var{tol} = 1e-6} by default.
48 ## @var{maxit} is the maximum allowable number of iterations; if
49 ## @code{[]} is supplied for @code{maxit}, or @code{pcg} has less
50 ## arguments, a default value equal to 20 is used.
53 ## @var{m} = @var{m1} * @var{m2} is the (left) preconditioning matrix, so that
54 ## the iteration is (theoretically) equivalent to solving by @code{pcg}
56 ## @var{x} = @var{m} \ @var{b}}, with @code{@var{P} = @var{m} \ @var{A}}.
57 ## Note that a proper choice of the preconditioner may dramatically
58 ## improve the overall performance of the method. Instead of matrices
59 ## @var{m1} and @var{m2}, the user may pass two functions which return
60 ## the results of applying the inverse of @var{m1} and @var{m2} to
61 ## a vector (usually this is the preferred way of using the preconditioner).
62 ## If @code{[]} is supplied for @var{m1}, or @var{m1} is omitted, no
63 ## preconditioning is applied. If @var{m2} is omitted, @var{m} = @var{m1}
64 ## will be used as preconditioner.
67 ## @var{x0} is the initial guess. If @var{x0} is empty or omitted, the
68 ## function sets @var{x0} to a zero vector by default.
71 ## The arguments which follow @var{x0} are treated as parameters, and
72 ## passed in a proper way to any of the functions (@var{A} or @var{m})
73 ## which are passed to @code{pcg}. See the examples below for further
74 ## details. The output arguments are
78 ## @var{x} is the computed approximation to the solution of
79 ## @code{@var{A} * @var{x} = @var{b}}.
82 ## @var{flag} reports on the convergence. @code{@var{flag} = 0} means
83 ## the solution converged and the tolerance criterion given by @var{tol}
84 ## is satisfied. @code{@var{flag} = 1} means that the @var{maxit} limit
85 ## for the iteration count was reached. @code{@var{flag} = 3} reports that
86 ## the (preconditioned) matrix was found not positive definite.
89 ## @var{relres} is the ratio of the final residual to its initial value,
90 ## measured in the Euclidean norm.
93 ## @var{iter} is the actual number of iterations performed.
96 ## @var{resvec} describes the convergence history of the method.
97 ## @code{@var{resvec} (i,1)} is the Euclidean norm of the residual, and
98 ## @code{@var{resvec} (i,2)} is the preconditioned residual norm,
99 ## after the (@var{i}-1)-th iteration, @code{@var{i} =
100 ## 1, 2, @dots{}, @var{iter}+1}. The preconditioned residual norm
102 ## @code{norm (@var{r}) ^ 2 = @var{r}' * (@var{m} \ @var{r})} where
103 ## @code{@var{r} = @var{b} - @var{A} * @var{x}}, see also the
104 ## description of @var{m}. If @var{eigest} is not required, only
105 ## @code{@var{resvec} (:,1)} is returned.
108 ## @var{eigest} returns the estimate for the smallest @code{@var{eigest}
109 ## (1)} and largest @code{@var{eigest} (2)} eigenvalues of the
110 ## preconditioned matrix @code{@var{P} = @var{m} \ @var{A}}. In
111 ## particular, if no preconditioning is used, the estimates for the
112 ## extreme eigenvalues of @var{A} are returned. @code{@var{eigest} (1)}
113 ## is an overestimate and @code{@var{eigest} (2)} is an underestimate,
114 ## so that @code{@var{eigest} (2) / @var{eigest} (1)} is a lower bound
115 ## for @code{cond (@var{P}, 2)}, which nevertheless in the limit should
116 ## theoretically be equal to the actual value of the condition number.
117 ## The method which computes @var{eigest} works only for symmetric positive
118 ## definite @var{A} and @var{m}, and the user is responsible for
119 ## verifying this assumption.
122 ## Let us consider a trivial problem with a diagonal matrix (we exploit the
128 ## A = diag (sparse (1:n));
130 ## [l, u, p, q] = luinc (A, 1.e-3);
134 ## @sc{Example 1:} Simplest use of @code{pcg}
140 ## @sc{Example 2:} @code{pcg} with a function which computes
141 ## @code{@var{A} * @var{x}}
145 ## function y = apply_a (x)
149 ## x = pcg ("apply_a", b)
153 ## @sc{Example 3:} @code{pcg} with a preconditioner: @var{l} * @var{u}
156 ## x = pcg (A, b, 1.e-6, 500, l*u)
159 ## @sc{Example 4:} @code{pcg} with a preconditioner: @var{l} * @var{u}.
160 ## Faster than @sc{Example 3} since lower and upper triangular matrices
161 ## are easier to invert
164 ## x = pcg (A, b, 1.e-6, 500, l, u)
167 ## @sc{Example 5:} Preconditioned iteration, with full diagnostics. The
168 ## preconditioner (quite strange, because even the original matrix
169 ## @var{A} is trivial) is defined as a function
173 ## function y = apply_m (x)
174 ## k = floor (length (x) - 2);
176 ## y(1:k) = x(1:k) ./ [1:k]';
179 ## [x, flag, relres, iter, resvec, eigest] = ...
180 ## pcg (A, b, [], [], "apply_m");
181 ## semilogy (1:iter+1, resvec);
185 ## @sc{Example 6:} Finally, a preconditioner which depends on a
186 ## parameter @var{k}.
190 ## function y = apply_M (x, varargin)
191 ## K = varargin@{1@};
193 ## y(1:K) = x(1:K) ./ [1:K]';
196 ## [x, flag, relres, iter, resvec, eigest] = ...
197 ## pcg (A, b, [], [], "apply_m", [], [], 3)
205 ## C.T. Kelley, @cite{Iterative Methods for Linear and Nonlinear Equations},
206 ## SIAM, 1995. (the base PCG algorithm)
209 ## Y. Saad, @cite{Iterative Methods for Sparse Linear Systems}, PWS 1996.
210 ## (condition number estimate from PCG) Revised version of this book is
211 ## available online at @url{http://www-users.cs.umn.edu/~saad/books.html}
214 ## @seealso{sparse, pcr}
217 ## Author: Piotr Krzyzanowski <piotr.krzyzanowski@mimuw.edu.pl>
218 ## Modified by: Vittoria Rezzonico <vittoria.rezzonico@epfl.ch>
219 ## - Add the ability to provide the pre-conditioner as two separate matrices
221 function [x, flag, relres, iter, resvec, eigest] = pcg (A, b, tol, maxit, m1, m2, x0, varargin)
225 if (nargin < 7 || isempty (x0))
226 x = zeros (size (b));
231 if (nargin < 5 || isempty (m1))
237 if (nargin < 6 || isempty (m2))
243 if (nargin < 4 || isempty (maxit))
244 maxit = min (size (b, 1), 20);
249 if (nargin < 3 || isempty (tol))
253 preconditioned_residual_out = false;
255 T = zeros (maxit, maxit);
256 preconditioned_residual_out = true;
259 ## Assume A is positive definite.
260 matrix_positive_definite = true;
262 p = zeros (size (b));
268 ## A should be a function.
269 r = b - feval (A, x, varargin{:});
272 resvec(1,1) = norm (r);
276 while (resvec (iter-1,1) > tol * resvec (1,1) && iter < maxit)
281 y = feval (m1, r, varargin{:});
290 z = feval (m2, y, varargin{:});
296 resvec (iter-1,2) = sqrt (tau);
304 ## A should be a function.
305 w = feval (A, p, varargin{:});
307 ## Needed only for eigest.
309 alpha = tau / (p'*w);
312 matrix_positive_definite = false;
316 if (nargout > 5 && iter > 2)
317 T(iter-1:iter, iter-1:iter) = T(iter-1:iter, iter-1:iter) + ...
318 [1 sqrt(beta); sqrt(beta) beta]./oldalpha;
319 ## EVS = eig(T(2:iter-1,2:iter-1));
320 ## fprintf(stderr,"PCG condest: %g (iteration: %d)\n", max(EVS)/min(EVS),iter);
322 resvec (iter,1) = norm (r);
327 if (matrix_positive_definite)
329 T = T(2:iter-2,2:iter-2);
331 eigest = [min(l), max(l)];
332 ## fprintf (stderr, "pcg condest: %g\n", eigest(2)/eigest(1));
335 warning ("pcg: eigenvalue estimate failed: iteration converged too fast");
341 ## Apply the preconditioner once more and finish with the precond
347 y = feval (m1, r, varargin{:});
356 z = feval (m2, y, varargin{:});
362 resvec (iter-1,2) = sqrt (r' * z);
364 resvec = resvec(:,1);
368 relres = resvec (iter-1,1) ./ resvec(1,1);
370 if (iter >= maxit - 2)
373 warning ("pcg: maximum number of iterations (%d) reached\n", iter);
374 warning ("the initial residual norm was reduced %g times.\n", ...
378 fprintf (stderr, "pcg: converged in %d iterations. ", iter);
379 fprintf (stderr, "the initial residual norm was reduced %g times.\n",...
383 if (! matrix_positive_definite)
386 warning ("pcg: matrix not positive definite?\n");
393 %! # Simplest usage of pcg (see also 'help pcg')
396 %! A = diag ([1:N]); b = rand (N, 1); y = A \ b; #y is the true solution
398 %! printf('The solution relative error is %g\n', norm (x - y) / norm (y));
400 %! # You shouldn't be afraid if pcg issues some warning messages in this
401 %! # example: watch out in the second example, why it takes N iterations
402 %! # of pcg to converge to (a very accurate, by the way) solution
405 %! # Full output from pcg, except for the eigenvalue estimates
406 %! # We use this output to plot the convergence history
409 %! A = diag ([1:N]); b = rand (N, 1); X = A \ b; #X is the true solution
410 %! [x, flag, relres, iter, resvec] = pcg (A, b);
411 %! printf('The solution relative error is %g\n', norm (x - X) / norm (X));
412 %! title('Convergence history'); xlabel('Iteration'); ylabel('log(||b-Ax||/||b||)');
413 %! semilogy([0:iter], resvec / resvec(1),'o-g');
414 %! legend('relative residual');
417 %! # Full output from pcg, including the eigenvalue estimates
418 %! # Hilbert matrix is extremely ill conditioned, so pcg WILL have problems
421 %! A = hilb (N); b = rand (N, 1); X = A \ b; #X is the true solution
422 %! [x, flag, relres, iter, resvec, eigest] = pcg (A, b, [], 200);
423 %! printf('The solution relative error is %g\n', norm (x - X) / norm (X));
424 %! printf('Condition number estimate is %g\n', eigest(2) / eigest (1));
425 %! printf('Actual condition number is %g\n', cond (A));
426 %! title('Convergence history'); xlabel('Iteration'); ylabel('log(||b-Ax||)');
427 %! semilogy([0:iter], resvec,['o-g';'+-r']);
428 %! legend('absolute residual','absolute preconditioned residual');
431 %! # Full output from pcg, including the eigenvalue estimates
432 %! # We use the 1-D Laplacian matrix for A, and cond(A) = O(N^2)
433 %! # and that's the reasone we need some preconditioner; here we take
434 %! # a very simple and not powerful Jacobi preconditioner,
435 %! # which is the diagonal of A
439 %! for i=1 : N - 1 # form 1-D Laplacian matrix
440 %! A (i:i+1, i:i+1) = [2 -1; -1 2];
442 %! b = rand (N, 1); X = A \ b; #X is the true solution
444 %! printf('System condition number is %g\n', cond (A));
445 %! # No preconditioner: the convergence is very slow!
447 %! [x, flag, relres, iter, resvec, eigest] = pcg (A, b, [], maxit);
448 %! printf('System condition number estimate is %g\n', eigest(2) / eigest(1));
449 %! title('Convergence history'); xlabel('Iteration'); ylabel('log(||b-Ax||)');
450 %! semilogy([0:iter], resvec(:,1), 'o-g');
451 %! legend('NO preconditioning: absolute residual');
454 %! # Test Jacobi preconditioner: it will not help much!!!
456 %! M = diag (diag (A)); # Jacobi preconditioner
457 %! [x, flag, relres, iter, resvec, eigest] = pcg (A, b, [], maxit, M);
458 %! printf('JACOBI preconditioned system condition number estimate is %g\n', eigest(2) / eigest(1));
460 %! semilogy([0:iter], resvec(:,1), 'o-r');
461 %! legend('NO preconditioning: absolute residual', ...
462 %! 'JACOBI preconditioner: absolute residual');
465 %! # Test nonoverlapping block Jacobi preconditioner: it will help much!
467 %! M = zeros (N, N); k = 4;
468 %! for i = 1 : k : N # form 1-D Laplacian matrix
469 %! M (i:i+k-1, i:i+k-1) = A (i:i+k-1, i:i+k-1);
471 %! [x, flag, relres, iter, resvec, eigest] = pcg (A, b, [], maxit, M);
472 %! printf('BLOCK JACOBI preconditioned system condition number estimate is %g\n', eigest(2) / eigest(1));
473 %! semilogy ([0:iter], resvec(:,1),'o-b');
474 %! legend('NO preconditioning: absolute residual', ...
475 %! 'JACOBI preconditioner: absolute residual', ...
476 %! 'BLOCK JACOBI preconditioner: absolute residual');
480 %! #solve small diagonal system
483 %! A = diag ([1:N]); b = rand (N, 1); X = A \ b; #X is the true solution
484 %! [x, flag] = pcg (A, b, [], N+1);
485 %! assert(norm (x - X) / norm (X), 0, 1e-10);
490 %! #solve small indefinite diagonal system
491 %! #despite A is indefinite, the iteration continues and converges
492 %! #indefiniteness of A is detected
495 %! A = diag([1:N] .* (-ones(1, N) .^ 2)); b = rand (N, 1); X = A \ b; #X is the true solution
496 %! [x, flag] = pcg (A, b, [], N+1);
497 %! assert(norm (x - X) / norm (X), 0, 1e-10);
502 %! #solve tridiagonal system, do not converge in default 20 iterations
506 %! for i = 1 : N - 1 # form 1-D Laplacian matrix
507 %! A (i:i+1, i:i+1) = [2 -1; -1 2];
509 %! b = ones (N, 1); X = A \ b; #X is the true solution
510 %! [x, flag, relres, iter, resvec, eigest] = pcg (A, b, 1e-12);
512 %! assert(relres > 1.0);
513 %! assert(iter, 20); #should perform max allowable default number of iterations
517 %! #solve tridiagonal system with 'prefect' preconditioner
518 %! #converges in one iteration, so the eigest does not work
519 %! #and issues a warning
523 %! for i = 1 : N - 1 # form 1-D Laplacian matrix
524 %! A (i:i+1, i:i+1) = [2 -1; -1 2];
526 %! b = ones (N, 1); X = A \ b; #X is the true solution
527 %! [x, flag, relres, iter, resvec, eigest] = pcg (A, b, [], [], A, [], b);
528 %! assert(norm (x - X) / norm (X), 0, 1e-6);
530 %! assert(iter, 1); #should converge in one iteration
531 %! assert(isnan (eigest), isnan ([NaN, NaN]));