Purpose
To solve the system of linear equations Ax = b, with A symmetric, positive definite, or, in the implicit form, f(A, x) = b, where y = f(A, x) is a symmetric positive definite linear mapping from x to y, using the conjugate gradient (CG) algorithm without preconditioning.Specification
SUBROUTINE MB02WD( FORM, F, N, IPAR, LIPAR, DPAR, LDPAR, ITMAX, $ A, LDA, B, INCB, X, INCX, TOL, DWORK, LDWORK, $ IWARN, INFO ) C .. Scalar Arguments .. CHARACTER FORM INTEGER INCB, INCX, INFO, ITMAX, IWARN, LDA, LDPAR, $ LDWORK, LIPAR, N DOUBLE PRECISION TOL C .. Array Arguments .. DOUBLE PRECISION A(LDA,*), B(*), DPAR(*), DWORK(*), X(*) INTEGER IPAR(*)Arguments
Mode Parameters
FORM CHARACTER*1 Specifies the form of the system of equations, as follows: = 'U' : Ax = b, the upper triagular part of A is used; = 'L' : Ax = b, the lower triagular part of A is used; = 'F' : the implicit, function form, f(A, x) = b.Function Parameters
F EXTERNAL If FORM = 'F', then F is a subroutine which calculates the value of f(A, x), for given A and x. If FORM <> 'F', then F is not called. F must have the following interface: SUBROUTINE F( N, IPAR, LIPAR, DPAR, LDPAR, A, LDA, X, $ INCX, DWORK, LDWORK, INFO ) where N (input) INTEGER The dimension of the vector x. N >= 0. IPAR (input) INTEGER array, dimension (LIPAR) The integer parameters describing the structure of the matrix A. LIPAR (input) INTEGER The length of the array IPAR. LIPAR >= 0. DPAR (input) DOUBLE PRECISION array, dimension (LDPAR) The real parameters needed for solving the problem. LDPAR (input) INTEGER The length of the array DPAR. LDPAR >= 0. A (input) DOUBLE PRECISION array, dimension (LDA, NC), where NC is the number of columns. The leading NR-by-NC part of this array must contain the (compressed) representation of the matrix A, where NR is the number of rows of A (function of IPAR entries). LDA (input) INTEGER The leading dimension of the array A. LDA >= MAX(1,NR). X (input/output) DOUBLE PRECISION array, dimension (1+(N-1)*INCX) On entry, this incremented array must contain the vector x. On exit, this incremented array contains the value of the function f, y = f(A, x). INCX (input) INTEGER The increment for the elements of X. INCX > 0. DWORK DOUBLE PRECISION array, dimension (LDWORK) The workspace array for subroutine F. LDWORK (input) INTEGER The size of the array DWORK (as large as needed in the subroutine F). INFO INTEGER Error indicator, set to a negative value if an input scalar argument is erroneous, and to positive values for other possible errors in the subroutine F. The LAPACK Library routine XERBLA should be used in conjunction with negative INFO. INFO must be zero if the subroutine finished successfully. Parameters marked with "(input)" must not be changed.Input/Output Parameters
N (input) INTEGER The dimension of the vector x. N >= 0. If FORM = 'U' or FORM = 'L', N is also the number of rows and columns of the matrix A. IPAR (input) INTEGER array, dimension (LIPAR) If FORM = 'F', the integer parameters describing the structure of the matrix A. This parameter is ignored if FORM = 'U' or FORM = 'L'. LIPAR (input) INTEGER The length of the array IPAR. LIPAR >= 0. DPAR (input) DOUBLE PRECISION array, dimension (LDPAR) If FORM = 'F', the real parameters needed for solving the problem. This parameter is ignored if FORM = 'U' or FORM = 'L'. LDPAR (input) INTEGER The length of the array DPAR. LDPAR >= 0. ITMAX (input) INTEGER The maximal number of iterations to do. ITMAX >= 0. A (input) DOUBLE PRECISION array, dimension (LDA, NC), if FORM = 'F', dimension (LDA, N), otherwise. If FORM = 'F', the leading NR-by-NC part of this array must contain the (compressed) representation of the matrix A, where NR and NC are the number of rows and columns, respectively, of the matrix A. The array A is not referenced by this routine itself, except in the calls to the routine F. If FORM <> 'F', the leading N-by-N part of this array must contain the matrix A, assumed to be symmetric; only the triangular part specified by FORM is referenced. LDA (input) INTEGER The leading dimension of array A. LDA >= MAX(1,NR), if FORM = 'F'; LDA >= MAX(1,N), if FORM = 'U' or FORM = 'L'. B (input) DOUBLE PRECISION array, dimension (1+(N-1)*INCB) The incremented vector b. INCB (input) INTEGER The increment for the elements of B. INCB > 0. X (input/output) DOUBLE PRECISION array, dimension (1+(N-1)*INCX) On entry, this incremented array must contain an initial approximation of the solution. If an approximation is not known, setting all elements of x to zero is recommended. On exit, this incremented array contains the computed solution x of the system of linear equations. INCX (input) INTEGER The increment for the elements of X. INCX > 0.Tolerances
TOL DOUBLE PRECISION If TOL > 0, absolute tolerance for the iterative process. The algorithm will stop if || Ax - b ||_2 <= TOL. Since it is advisable to use a relative tolerance, say TOLER, TOL should be chosen as TOLER*|| b ||_2. If TOL <= 0, a default relative tolerance, TOLDEF = N*EPS*|| b ||_2, is used, where EPS is the machine precision (see LAPACK Library routine DLAMCH).Workspace
DWORK DOUBLE PRECISION array, dimension (LDWORK) On exit, if INFO = 0, DWORK(1) returns the number of iterations performed and DWORK(2) returns the remaining residual, || Ax - b ||_2. LDWORK INTEGER The length of the array DWORK. LDWORK >= MAX(2,3*N + DWORK(F)), if FORM = 'F', where DWORK(F) is the workspace needed by F; LDWORK >= MAX(2,3*N), if FORM = 'U' or FORM = 'L'.Warning Indicator
IWARN INTEGER = 0: no warning; = 1: the algorithm finished after ITMAX > 0 iterations, without achieving the desired precision TOL; = 2: ITMAX is zero; in this case, DWORK(2) is not set.Error Indicator
INFO INTEGER = 0: successful exit; < 0: if INFO = -i, the i-th argument had an illegal value; > 0: if INFO = i, then F returned with INFO = i.Method
The following CG iteration is used for solving Ax = b: Start: q(0) = r(0) = Ax - b < q(k), r(k) > ALPHA(k) = - ---------------- < q(k), Aq(k) > x(k+1) = x(k) - ALPHA(k) * q(k) r(k+1) = r(k) - ALPHA(k) * Aq(k) < r(k+1), r(k+1) > BETA(k) = -------------------- < r(k) , r(k) > q(k+1) = r(k+1) + BETA(k) * q(k) where <.,.> denotes the scalar product.References
[1] Golub, G.H. and van Loan, C.F. Matrix Computations. Third Edition. M. D. Johns Hopkins University Press, Baltimore, pp. 520-528, 1996. [2] Luenberger, G. Introduction to Linear and Nonlinear Programming. Addison-Wesley, Reading, MA, p.187, York, 1973.Numerical Aspects
Since the residuals are orthogonal in the scalar product <x, y> = y'Ax, the algorithm is theoretically finite. But rounding errors cause a loss of orthogonality, so a finite termination cannot be guaranteed. However, one can prove [2] that || x-x_k ||_A := sqrt( (x-x_k)' * A * (x-x_k) ) sqrt( kappa_2(A) ) - 1 <= 2 || x-x_0 ||_A * ------------------------ , sqrt( kappa_2(A) ) + 1 where kappa_2 is the condition number. The approximate number of floating point operations is (k*(N**2 + 15*N) + N**2 + 3*N)/2, if FORM <> 'F', k*(f + 7*N) + f, if FORM = 'F', where k is the number of CG iterations performed, and f is the number of floating point operations required by the subroutine F.Further Comments
NoneExample
Program Text
NoneProgram Data
NoneProgram Results
None