Previous: lmstr Up: ../minpack.html First: chkder
Page 1
Documentation for MINPACK subroutine LMSTR1
Double precision version
Argonne National Laboratory
Burton S. Garbow, Kenneth E. Hillstrom, Jorge J. More
March 1980
1. Purpose.
The purpose of LMSTR1 is to minimize the sum of the squares of M
nonlinear functions in N variables by a modification of the
Levenberg-Marquardt algorithm which uses minimal storage. This
is done by using the more general least-squares solver LMSTR.
The user must provide a subroutine which calculates the func-
tions and the rows of the Jacobian.
2. Subroutine and type statements.
SUBROUTINE LMSTR1(FCN,M,N,X,FVEC,FJAC,LDFJAC,TOL,
* INFO,IPVT,WA,LWA)
INTEGER M,N,LDFJAC,INFO,LWA
INTEGER IPVT(N)
DOUBLE PRECISION TOL
DOUBLE PRECISION X(N),FVEC(M),FJAC(LDFJAC,N),WA(LWA)
EXTERNAL FCN
3. Parameters.
Parameters designated as input parameters must be specified on
entry to LMSTR1 and are not changed on exit, while parameters
designated as output parameters need not be specified on entry
and are set to appropriate values on exit from LMSTR1.
FCN is the name of the user-supplied subroutine which calculates
the functions and the rows of the Jacobian. FCN must be
declared in an EXTERNAL statement in the user calling program,
and should be written as follows.
SUBROUTINE FCN(M,N,X,FVEC,FJROW,IFLAG)
INTEGER M,N,IFLAG
DOUBLE PRECISION X(N),FVEC(M),FJROW(N)
----------
IF IFLAG = 1 CALCULATE THE FUNCTIONS AT X AND
RETURN THIS VECTOR IN FVEC.
IF IFLAG = I CALCULATE THE (I-1)-ST ROW OF THE
JACOBIAN AT X AND RETURN THIS VECTOR IN FJROW.
----------
RETURN
Page 2
END
The value of IFLAG should not be changed by FCN unless the
user wants to terminate execution of LMSTR1. In this case set
IFLAG to a negative integer.
M is a positive integer input variable set to the number of
functions.
N is a positive integer input variable set to the number of
variables. N must not exceed M.
X is an array of length N. On input X must contain an initial
estimate of the solution vector. On output X contains the
final estimate of the solution vector.
FVEC is an output array of length M which contains the functions
evaluated at the output X.
FJAC is an output N by N array. The upper triangle of FJAC con-
tains an upper triangular matrix R such that
T T T
P *(JAC *JAC)*P = R *R,
where P is a permutation matrix and JAC is the final calcu-
lated Jacobian. Column j of P is column IPVT(j) (see below)
of the identity matrix. The lower triangular part of FJAC
contains information generated during the computation of R.
LDFJAC is a positive integer input variable not less than N
which specifies the leading dimension of the array FJAC.
TOL is a nonnegative input variable. Termination occurs when
the algorithm estimates either that the relative error in the
sum of squares is at most TOL or that the relative error
between X and the solution is at most TOL. Section 4 contains
more details about TOL.
INFO is an integer output variable. If the user has terminated
execution, INFO is set to the (negative) value of IFLAG. See
description of FCN. Otherwise, INFO is set as follows.
INFO = 0 Improper input parameters.
INFO = 1 Algorithm estimates that the relative error in the
sum of squares is at most TOL.
INFO = 2 Algorithm estimates that the relative error between
X and the solution is at most TOL.
INFO = 3 Conditions for INFO = 1 and INFO = 2 both hold.
INFO = 4 FVEC is orthogonal to the columns of the Jacobian to
Page 3
machine precision.
INFO = 5 Number of calls to FCN with IFLAG = 1 has reached
100*(N+1).
INFO = 6 TOL is too small. No further reduction in the sum
of squares is possible.
INFO = 7 TOL is too small. No further improvement in the
approximate solution X is possible.
Sections 4 and 5 contain more details about INFO.
IPVT is an integer output array of length N. IPVT defines a
permutation matrix P such that JAC*P = Q*R, where JAC is the
final calculated Jacobian, Q is orthogonal (not stored), and R
is upper triangular. Column j of P is column IPVT(j) of the
identity matrix.
WA is a work array of length LWA.
LWA is a positive integer input variable not less than 5*N+M.
4. Successful completion.
The accuracy of LMSTR1 is controlled by the convergence parame-
ter TOL. This parameter is used in tests which make three types
of comparisons between the approximation X and a solution XSOL.
LMSTR1 terminates when any of the tests is satisfied. If TOL is
less than the machine precision (as defined by the MINPACK func-
tion DPMPAR(1)), then LMSTR1 only attempts to satisfy the test
defined by the machine precision. Further progress is not usu-
ally possible. Unless high precision solutions are required,
the recommended value for TOL is the square root of the machine
precision.
The tests assume that the functions and the Jacobian are coded
consistently, and that the functions are reasonably well
behaved. If these conditions are not satisfied, then LMSTR1 may
incorrectly indicate convergence. The coding of the Jacobian
can be checked by the MINPACK subroutine CHKDER. If the Jaco-
bian is coded correctly, then the validity of the answer can be
checked, for example, by rerunning LMSTR1 with a tighter toler-
ance.
First convergence test. If ENORM(Z) denotes the Euclidean norm
of a vector Z, then this test attempts to guarantee that
ENORM(FVEC) .LE. (1+TOL)*ENORM(FVECS),
where FVECS denotes the functions evaluated at XSOL. If this
condition is satisfied with TOL = 10**(-K), then the final
residual norm ENORM(FVEC) has K significant decimal digits and
Page 4
INFO is set to 1 (or to 3 if the second test is also satis-
fied).
Second convergence test. If D is a diagonal matrix (implicitly
generated by LMSTR1) whose entries contain scale factors for
the variables, then this test attempts to guarantee that
ENORM(D*(X-XSOL)) .LE. TOL*ENORM(D*XSOL).
If this condition is satisfied with TOL = 10**(-K), then the
larger components of D*X have K significant decimal digits and
INFO is set to 2 (or to 3 if the first test is also satis-
fied). There is a danger that the smaller components of D*X
may have large relative errors, but the choice of D is such
that the accuracy of the components of X is usually related to
their sensitivity.
Third convergence test. This test is satisfied when FVEC is
orthogonal to the columns of the Jacobian to machine preci-
sion. There is no clear relationship between this test and
the accuracy of LMSTR1, and furthermore, the test is equally
well satisfied at other critical points, namely maximizers and
saddle points. Therefore, termination caused by this test
(INFO = 4) should be examined carefully.
5. Unsuccessful completion.
Unsuccessful termination of LMSTR1 can be due to improper input
parameters, arithmetic interrupts, or an excessive number of
function evaluations.
Improper input parameters. INFO is set to 0 if N .LE. 0, or
M .LT. N, or LDFJAC .LT. N, or TOL .LT. 0.D0, or
LWA .LT. 5*N+M.
Arithmetic interrupts. If these interrupts occur in the FCN
subroutine during an early stage of the computation, they may
be caused by an unacceptable choice of X by LMSTR1. In this
case, it may be possible to remedy the situation by not evalu-
ating the functions here, but instead setting the components
of FVEC to numbers that exceed those in the initial FVEC,
thereby indirectly reducing the step length. The step length
can be more directly controlled by using instead LMSTR, which
includes in its calling sequence the step-length- governing
parameter FACTOR.
Excessive number of function evaluations. If the number of
calls to FCN with IFLAG = 1 reaches 100*(N+1), then this indi-
cates that the routine is converging very slowly as measured
by the progress of FVEC, and INFO is set to 5. In this case,
it may be helpful to restart LMSTR1, thereby forcing it to
disregard old (and possibly harmful) information.
Page 5
6. Characteristics of the algorithm.
LMSTR1 is a modification of the Levenberg-Marquardt algorithm.
Two of its main characteristics involve the proper use of
implicitly scaled variables and an optimal choice for the cor-
rection. The use of implicitly scaled variables achieves scale
invariance of LMSTR1 and limits the size of the correction in
any direction where the functions are changing rapidly. The
optimal choice of the correction guarantees (under reasonable
conditions) global convergence from starting points far from the
solution and a fast rate of convergence for problems with small
residuals.
Timing. The time required by LMSTR1 to solve a given problem
depends on M and N, the behavior of the functions, the accu-
racy requested, and the starting point. The number of arith-
metic operations needed by LMSTR1 is about N**3 to process
each evaluation of the functions (call to FCN with IFLAG = 1)
and 1.5*(N**2) to process each row of the Jacobian (call to
FCN with IFLAG .GE. 2). Unless FCN can be evaluated quickly,
the timing of LMSTR1 will be strongly influenced by the time
spent in FCN.
Storage. LMSTR1 requires N**2 + 2*M + 6*N double precision sto-
rage locations and N integer storage locations, in addition to
the storage required by the program. There are no internally
declared storage arrays.
7. Subprograms required.
USER-supplied ...... FCN
MINPACK-supplied ... DPMPAR,ENORM,LMSTR,LMPAR,QRFAC,QRSOLV,
RWUPDT
FORTRAN-supplied ... DABS,DMAX1,DMIN1,DSQRT,MOD
8. References.
Jorge J. More, The Levenberg-Marquardt Algorithm, Implementation
and Theory. Numerical Analysis, G. A. Watson, editor.
Lecture Notes in Mathematics 630, Springer-Verlag, 1977.
9. Example.
The problem is to determine the values of x(1), x(2), and x(3)
which provide the best fit (in the least squares sense) of
x(1) + u(i)/(v(i)*x(2) + w(i)*x(3)), i = 1, 15
Page 6
to the data
y = (0.14,0.18,0.22,0.25,0.29,0.32,0.35,0.39,
0.37,0.58,0.73,0.96,1.34,2.10,4.39),
where u(i) = i, v(i) = 16 - i, and w(i) = min(u(i),v(i)). The
i-th component of FVEC is thus defined by
y(i) - (x(1) + u(i)/(v(i)*x(2) + w(i)*x(3))).
C **********
C
C DRIVER FOR LMSTR1 EXAMPLE.
C DOUBLE PRECISION VERSION
C
C **********
INTEGER J,M,N,LDFJAC,INFO,LWA,NWRITE
INTEGER IPVT(3)
DOUBLE PRECISION TOL,FNORM
DOUBLE PRECISION X(3),FVEC(15),FJAC(3,3),WA(30)
DOUBLE PRECISION ENORM,DPMPAR
EXTERNAL FCN
C
C LOGICAL OUTPUT UNIT IS ASSUMED TO BE NUMBER 6.
C
DATA NWRITE /6/
C
M = 15
N = 3
C
C THE FOLLOWING STARTING VALUES PROVIDE A ROUGH FIT.
C
X(1) = 1.D0
X(2) = 1.D0
X(3) = 1.D0
C
LDFJAC = 3
LWA = 30
C
C SET TOL TO THE SQUARE ROOT OF THE MACHINE PRECISION.
C UNLESS HIGH PRECISION SOLUTIONS ARE REQUIRED,
C THIS IS THE RECOMMENDED SETTING.
C
TOL = DSQRT(DPMPAR(1))
C
CALL LMSTR1(FCN,M,N,X,FVEC,FJAC,LDFJAC,TOL,
* INFO,IPVT,WA,LWA)
FNORM = ENORM(M,FVEC)
WRITE (NWRITE,1000) FNORM,INFO,(X(J),J=1,N)
STOP
1000 FORMAT (5X,31H FINAL L2 NORM OF THE RESIDUALS,D15.7 //
* 5X,15H EXIT PARAMETER,16X,I10 //
* 5X,27H FINAL APPROXIMATE SOLUTION // 5X,3D15.7)
C
Page 7
C LAST CARD OF DRIVER FOR LMSTR1 EXAMPLE.
C
END
SUBROUTINE FCN(M,N,X,FVEC,FJROW,IFLAG)
INTEGER M,N,IFLAG
DOUBLE PRECISION X(N),FVEC(M),FJROW(N)
C
C SUBROUTINE FCN FOR LMSTR1 EXAMPLE.
C
INTEGER I
DOUBLE PRECISION TMP1,TMP2,TMP3,TMP4
DOUBLE PRECISION Y(15)
DATA Y(1),Y(2),Y(3),Y(4),Y(5),Y(6),Y(7),Y(8),
* Y(9),Y(10),Y(11),Y(12),Y(13),Y(14),Y(15)
* /1.4D-1,1.8D-1,2.2D-1,2.5D-1,2.9D-1,3.2D-1,3.5D-1,3.9D-1,
* 3.7D-1,5.8D-1,7.3D-1,9.6D-1,1.34D0,2.1D0,4.39D0/
C
IF (IFLAG .GE. 2) GO TO 20
DO 10 I = 1, 15
TMP1 = I
TMP2 = 16 - I
TMP3 = TMP1
IF (I .GT. 8) TMP3 = TMP2
FVEC(I) = Y(I) - (X(1) + TMP1/(X(2)*TMP2 + X(3)*TMP3))
10 CONTINUE
GO TO 40
20 CONTINUE
I = IFLAG - 1
TMP1 = I
TMP2 = 16 - I
TMP3 = TMP1
IF (I .GT. 8) TMP3 = TMP2
TMP4 = (X(2)*TMP2 + X(3)*TMP3)**2
FJROW(1) = -1.D0
FJROW(2) = TMP1*TMP2/TMP4
FJROW(3) = TMP1*TMP3/TMP4
30 CONTINUE
40 CONTINUE
RETURN
C
C LAST CARD OF SUBROUTINE FCN.
C
END
Results obtained with different compilers or machines
may be slightly different.
FINAL L2 NORM OF THE RESIDUALS 0.9063596D-01
EXIT PARAMETER 1
FINAL APPROXIMATE SOLUTION
0.8241058D-01 0.1133037D+01 0.2343695D+01