subroutine sumsl(n, d, x, calcf, calcg, iv, liv, lv, v, 1 uiparm, urparm, ufparm) c c *** minimize general unconstrained objective function using *** c *** analytic gradient and hessian approx. from secant update *** c integer n, liv, lv integer iv(liv), uiparm(1) double precision d(n), x(n), v(lv), urparm(1) c dimension v(71 + n*(n+15)/2), uiparm(*), urparm(*) external calcf, calcg, ufparm c c *** purpose *** c c this routine interacts with subroutine sumit in an attempt c to find an n-vector x* that minimizes the (unconstrained) c objective function computed by calcf. (often the x* found is c a local minimizer rather than a global one.) c c-------------------------- parameter usage -------------------------- c c n........ (input) the number of variables on which f depends, i.e., c the number of components in x. c d........ (input/output) a scale vector such that d(i)*x(i), c i = 1,2,...,n, are all in comparable units. c d can strongly affect the behavior of sumsl. c finding the best choice of d is generally a trial- c and-error process. choosing d so that d(i)*x(i) c has about the same value for all i often works well. c the defaults provided by subroutine deflt (see i c below) require the caller to supply d. c x........ (input/output) before (initially) calling sumsl, the call- c er should set x to an initial guess at x*. when c sumsl returns, x contains the best point so far c found, i.e., the one that gives the least value so c far seen for f(x). c calcf.... (input) a subroutine that, given x, computes f(x). calcf c must be declared external in the calling program. c it is invoked by c call calcf(n, x, nf, f, uiparm, urparm, ufparm) c when calcf is called, nf is the invocation c count for calcf. nf is included for possible use c with calcg. if x is out of bounds (e.g., if it c would cause overflow in computing f(x)), then calcf c should set nf to 0. this will cause a shorter step c to be attempted. (if x is in bounds, then calcf c should not change nf.) the other parameters are as c described above and below. calcf should not change c n, p, or x. c calcg.... (input) a subroutine that, given x, computes g(x), the gra- c dient of f at x. calcg must be declared external in c the calling program. it is invoked by c call calcg(n, x, nf, g, uiparm, urparm, ufaprm) c when calcg is called, nf is the invocation c count for calcf at the time f(x) was evaluated. the c x passed to calcg is usually the one passed to calcf c on either its most recent invocation or the one c prior to it. if calcf saves intermediate results c for use by calcg, then it is possible to tell from c nf whether they are valid for the current x (or c which copy is valid if two copies are kept). if g c cannot be computed at x, then calcg should set nf to c 0. in this case, sumsl will return with iv(1) = 65. c (if g can be computed at x, then calcg should not c changed nf.) the other parameters to calcg are as c described above and below. calcg should not change c n or x. c iv....... (input/output) an integer value array of length liv (see c below) that helps control the sumsl algorithm and c that is used to store various intermediate quanti- c ties. of particular interest are the initialization/ c return code iv(1) and the entries in iv that control c printing and limit the number of iterations and func- c tion evaluations. see the section on iv input c values below. c liv...... (input) length of iv array. must be at least 60. if li c is too small, then sumsl returns with iv(1) = 15. c when sumsl returns, the smallest allowed value of c liv is stored in iv(lastiv) -- see the section on c iv output values below. (this is intended for use c with extensions of sumsl that handle constraints.) c lv....... (input) length of v array. must be at least 71+n*(n+15)/2. c (at least 77+n*(n+17)/2 for smsno, at least c 78+n*(n+12) for humsl). if lv is too small, then c sumsl returns with iv(1) = 16. when sumsl returns, c the smallest allowed value of lv is stored in c iv(lastv) -- see the section on iv output values c below. c v........ (input/output) a floating-point value array of length l c (see below) that helps control the sumsl algorithm c and that is used to store various intermediate c quantities. of particular interest are the entries c in v that limit the length of the first step c attempted (lmax0) and specify convergence tolerances c (afctol, lmaxs, rfctol, sctol, xctol, xftol). c uiparm... (input) user integer parameter array passed without change c to calcf and calcg. c urparm... (input) user floating-point parameter array passed without c change to calcf and calcg. c ufparm... (input) user external subroutine or function passed without c change to calcf and calcg. c c *** iv input values (from subroutine deflt) *** c c iv(1)... on input, iv(1) should have a value between 0 and 14...... c 0 and 12 mean this is a fresh start. 0 means that c deflt(2, iv, liv, lv, v) c is to be called to provide all default values to iv and c v. 12 (the value that deflt assigns to iv(1)) means the c caller has already called deflt and has possibly changed c some iv and/or v entries to non-default values. c 13 means deflt has been called and that sumsl (and c sumit) should only do their storage allocation. that is, c they should set the output components of iv that tell c where various subarrays arrays of v begin, such as iv(g) c (and, for humsl and humit only, iv(dtol)), and return. c 14 means that a storage has been allocated (by a call c with iv(1) = 13) and that the algorithm should be c started. when called with iv(1) = 13, sumsl returns c iv(1) = 14 unless liv or lv is too small (or n is not c positive). default = 12. c iv(inith).... iv(25) tells whether the hessian approximation h should c be initialized. 1 (the default) means sumit should c initialize h to the diagonal matrix whose i-th diagonal c element is d(i)**2. 0 means the caller has supplied a c cholesky factor l of the initial hessian approximation c h = l*(l**t) in v, starting at v(iv(lmat)) = v(iv(42)) c (and stored compactly by rows). note that iv(lmat) may c be initialized by calling sumsl with iv(1) = 13 (see c the iv(1) discussion above). default = 1. c iv(mxfcal)... iv(17) gives the maximum number of function evaluations c (calls on calcf) allowed. if this number does not suf- c fice, then sumsl returns with iv(1) = 9. default = 200. c iv(mxiter)... iv(18) gives the maximum number of iterations allowed. c it also indirectly limits the number of gradient evalua- c tions (calls on calcg) to iv(mxiter) + 1. if iv(mxiter) c iterations do not suffice, then sumsl returns with c iv(1) = 10. default = 150. c iv(outlev)... iv(19) controls the number and length of iteration sum- c mary lines printed (by itsum). iv(outlev) = 0 means do c not print any summary lines. otherwise, print a summary c line after each abs(iv(outlev)) iterations. if iv(outlev) c is positive, then summary lines of length 78 (plus carri- c age control) are printed, including the following... the c iteration and function evaluation counts, f = the current c function value, relative difference in function values c achieved by the latest step (i.e., reldf = (f0-v(f))/f01, c where f01 is the maximum of abs(v(f)) and abs(v(f0)) and c v(f0) is the function value from the previous itera- c tion), the relative function reduction predicted for the c step just taken (i.e., preldf = v(preduc) / f01, where c v(preduc) is described below), the scaled relative change c in x (see v(reldx) below), the step parameter for the c step just taken (stppar = 0 means a full newton step, c between 0 and 1 means a relaxed newton step, between 1 c and 2 means a double dogleg step, greater than 2 means c a scaled down cauchy step -- see subroutine dbldog), the c 2-norm of the scale vector d times the step just taken c (see v(dstnrm) below), and npreldf, i.e., c v(nreduc)/f01, where v(nreduc) is described below -- if c npreldf is positive, then it is the relative function c reduction predicted for a newton step (one with c stppar = 0). if npreldf is negative, then it is the c negative of the relative function reduction predicted c for a step computed with step bound v(lmaxs) for use in c testing for singular convergence. c if iv(outlev) is negative, then lines of length 50 c are printed, including only the first 6 items listed c above (through reldx). c default = 1. c iv(parprt)... iv(20) = 1 means print any nondefault v values on a c fresh start or any changed v values on a restart. c iv(parprt) = 0 means skip this printing. default = 1. c iv(prunit)... iv(21) is the output unit number on which all printing c is done. iv(prunit) = 0 means suppress all printing. c default = standard output unit (unit 6 on most systems). c iv(solprt)... iv(22) = 1 means print out the value of x returned (as c well as the gradient and the scale vector d). c iv(solprt) = 0 means skip this printing. default = 1. c iv(statpr)... iv(23) = 1 means print summary statistics upon return- c ing. these consist of the function value, the scaled c relative change in x caused by the most recent step (see c v(reldx) below), the number of function and gradient c evaluations (calls on calcf and calcg), and the relative c function reductions predicted for the last step taken and c for a newton step (or perhaps a step bounded by v(lmaxs) c -- see the descriptions of preldf and npreldf under c iv(outlev) above). c iv(statpr) = 0 means skip this printing. c iv(statpr) = -1 means skip this printing as well as that c of the one-line termination reason message. default = 1. c iv(x0prt).... iv(24) = 1 means print the initial x and scale vector d c (on a fresh start only). iv(x0prt) = 0 means skip this c printing. default = 1. c c *** (selected) iv output values *** c c iv(1)........ on output, iv(1) is a return code.... c 3 = x-convergence. the scaled relative difference (see c v(reldx)) between the current parameter vector x and c a locally optimal parameter vector is very likely at c most v(xctol). c 4 = relative function convergence. the relative differ- c ence between the current function value and its lo- c cally optimal value is very likely at most v(rfctol). c 5 = both x- and relative function convergence (i.e., the c conditions for iv(1) = 3 and iv(1) = 4 both hold). c 6 = absolute function convergence. the current function c value is at most v(afctol) in absolute value. c 7 = singular convergence. the hessian near the current c iterate appears to be singular or nearly so, and a c step of length at most v(lmaxs) is unlikely to yield c a relative function decrease of more than v(sctol). c 8 = false convergence. the iterates appear to be converg- c ing to a noncritical point. this may mean that the c convergence tolerances (v(afctol), v(rfctol), c v(xctol)) are too small for the accuracy to which c the function and gradient are being computed, that c there is an error in computing the gradient, or that c the function or gradient is discontinuous near x. c 9 = function evaluation limit reached without other con- c vergence (see iv(mxfcal)). c 10 = iteration limit reached without other convergence c (see iv(mxiter)). c 11 = stopx returned .true. (external interrupt). see the c usage notes below. c 14 = storage has been allocated (after a call with c iv(1) = 13). c 17 = restart attempted with n changed. c 18 = d has a negative component and iv(dtype) .le. 0. c 19...43 = v(iv(1)) is out of range. c 63 = f(x) cannot be computed at the initial x. c 64 = bad parameters passed to assess (which should not c occur). c 65 = the gradient could not be computed at x (see calcg c above). c 67 = bad first parameter to deflt. c 80 = iv(1) was out of range. c 81 = n is not positive. c iv(g)........ iv(28) is the starting subscript in v of the current c gradient vector (the one corresponding to x). c iv(lastiv)... iv(44) is the least acceptable value of liv. (it is c only set if liv is at least 44.) c iv(lastv).... iv(45) is the least acceptable value of lv. (it is c only set if liv is large enough, at least iv(lastiv).) c iv(nfcall)... iv(6) is the number of calls so far made on calcf (i.e., c function evaluations). c iv(ngcall)... iv(30) is the number of gradient evaluations (calls on c calcg). c iv(niter).... iv(31) is the number of iterations performed. c c *** (selected) v input values (from subroutine deflt) *** c c v(bias)..... v(43) is the bias parameter used in subroutine dbldog -- c see that subroutine for details. default = 0.8. c v(afctol)... v(31) is the absolute function convergence tolerance. c if sumsl finds a point where the function value is less c than v(afctol) in absolute value, and if sumsl does not c return with iv(1) = 3, 4, or 5, then it returns with c iv(1) = 6. this test can be turned off by setting c v(afctol) to zero. default = max(10**-20, machep**2), c where machep is the unit roundoff. c v(dinit).... v(38), if nonnegative, is the value to which the scale c vector d is initialized. default = -1. c v(lmax0).... v(35) gives the maximum 2-norm allowed for d times the c very first step that sumsl attempts. this parameter can c markedly affect the performance of sumsl. c v(lmaxs).... v(36) is used in testing for singular convergence -- if c the function reduction predicted for a step of length c bounded by v(lmaxs) is at most v(sctol) * abs(f0), where c f0 is the function value at the start of the current c iteration, and if sumsl does not return with iv(1) = 3, c 4, 5, or 6, then it returns with iv(1) = 7. default = 1. c v(rfctol)... v(32) is the relative function convergence tolerance. c if the current model predicts a maximum possible function c reduction (see v(nreduc)) of at most v(rfctol)*abs(f0) c at the start of the current iteration, where f0 is the c then current function value, and if the last step attempt- c ed achieved no more than twice the predicted function c decrease, then sumsl returns with iv(1) = 4 (or 5). c default = max(10**-10, machep**(2/3)), where machep is c the unit roundoff. c v(sctol).... v(37) is the singular convergence tolerance -- see the c description of v(lmaxs) above. c v(tuner1)... v(26) helps decide when to check for false convergence. c this is done if the actual function decrease from the c current step is no more than v(tuner1) times its predict- c ed value. default = 0.1. c v(xctol).... v(33) is the x-convergence tolerance. if a newton step c (see v(nreduc)) is tried that has v(reldx) .le. v(xctol) c and if this step yields at most twice the predicted func- c tion decrease, then sumsl returns with iv(1) = 3 (or 5). c (see the description of v(reldx) below.) c default = machep**0.5, where machep is the unit roundoff. c v(xftol).... v(34) is the false convergence tolerance. if a step is c tried that gives no more than v(tuner1) times the predict- c ed function decrease and that has v(reldx) .le. v(xftol), c and if sumsl does not return with iv(1) = 3, 4, 5, 6, or c 7, then it returns with iv(1) = 8. (see the description c of v(reldx) below.) default = 100*machep, where c machep is the unit roundoff. c v(*)........ deflt supplies to v a number of tuning constants, with c which it should ordinarily be unnecessary to tinker. see c section 17 of version 2.2 of the nl2sol usage summary c (i.e., the appendix to ref. 1) for details on v(i), c i = decfac, incfac, phmnfc, phmxfc, rdfcmn, rdfcmx, c tuner2, tuner3, tuner4, tuner5. c c *** (selected) v output values *** c c v(dgnorm)... v(1) is the 2-norm of (diag(d)**-1)*g, where g is the c most recently computed gradient. c v(dstnrm)... v(2) is the 2-norm of diag(d)*step, where step is the c current step. c v(f)........ v(10) is the current function value. c v(f0)....... v(13) is the function value at the start of the current c iteration. c v(nreduc)... v(6), if positive, is the maximum function reduction c possible according to the current model, i.e., the func- c tion reduction predicted for a newton step (i.e., c step = -h**-1 * g, where g is the current gradient and c h is the current hessian approximation). c if v(nreduc) is negative, then it is the negative of c the function reduction predicted for a step computed with c a step bound of v(lmaxs) for use in testing for singular c convergence. c v(preduc)... v(7) is the function reduction predicted (by the current c quadratic model) for the current step. this (divided by c v(f0)) is used in testing for relative function c convergence. c v(reldx).... v(17) is the scaled relative change in x caused by the c current step, computed as c max(abs(d(i)*(x(i)-x0(i)), 1 .le. i .le. p) / c max(d(i)*(abs(x(i))+abs(x0(i))), 1 .le. i .le. p), c where x = x0 + step. c c------------------------------- notes ------------------------------- c c *** algorithm notes *** c c this routine uses a hessian approximation computed from the c bfgs update (see ref 3). only a cholesky factor of the hessian c approximation is stored, and this is updated using ideas from c ref. 4. steps are computed by the double dogleg scheme described c in ref. 2. the steps are assessed as in ref. 1. c c *** usage notes *** c c after a return with iv(1) .le. 11, it is possible to restart, c i.e., to change some of the iv and v input values described above c and continue the algorithm from the point where it was interrupt- c ed. iv(1) should not be changed, nor should any entries of i c and v other than the input values (those supplied by deflt). c those who do not wish to write a calcg which computes the c gradient analytically should call smsno rather than sumsl. c smsno uses finite differences to compute an approximate gradient. c those who would prefer to provide f and g (the function and c gradient) by reverse communication rather than by writing subrou- c tines calcf and calcg may call on sumit directly. see the com- c ments at the beginning of sumit. c those who use sumsl interactively may wish to supply their c own stopx function, which should return .true. if the break key c has been pressed since stopx was last invoked. this makes it c possible to externally interrupt sumsl (which will return with c iv(1) = 11 if stopx returns .true.). c storage for g is allocated at the end of v. thus the caller c may make v longer than specified above and may allow calcg to use c elements of g beyond the first n as scratch storage. c c *** portability notes *** c c the sumsl distribution tape contains both single- and double- c precision versions of the sumsl source code, so it should be un- c necessary to change precisions. c only the functions imdcon and rmdcon contain machine-dependent c constants. to change from one machine to another, it should c suffice to change the (few) relevant lines in these functions. c intrinsic functions are explicitly declared. on certain com- c puters (e.g. univac), it may be necessary to comment out these c declarations. so that this may be done automatically by a simple c program, such declarations are preceded by a comment having c/+ c in columns 1-3 and blanks in columns 4-72 and are followed by c a comment having c/ in columns 1 and 2 and blanks in columns 3-72. c the sumsl source code is expressed in 1966 ansi standard c fortran. it may be converted to fortran 77 by commenting out all c lines that fall between a line having c/6 in columns 1-3 and a c line having c/7 in columns 1-3 and by removing (i.e., replacing c by a blank) the c in column 1 of the lines that follow the c/7 c line and precede a line having c/ in columns 1-2 and blanks in c columns 3-72. these changes convert some data statements into c parameter statements, convert some variables from real to c character*4, and make the data statements that initialize these c variables use character strings delimited by primes instead c of hollerith constants. (such variables and data statements c appear only in modules itsum and parck. parameter statements c appear nearly everywhere.) these changes also add save state- c ments for variables given machine-dependent constants by rmdcon. c c *** references *** c c 1. dennis, j.e., gay, d.m., and welsch, r.e. (1981), algorithm 573 -- c an adaptive nonlinear least-squares algorithm, acm trans. c math. software 7, pp. 369-383. c c 2. dennis, j.e., and mei, h.h.w. (1979), two new unconstrained opti- c mization algorithms which use function and gradient c values, j. optim. theory applic. 28, pp. 453-482. c c 3. dennis, j.e., and more, j.j. (1977), quasi-newton methods, motiva- c tion and theory, siam rev. 19, pp. 46-89. c c 4. goldfarb, d. (1976), factorized variable metric methods for uncon- c strained optimization, math. comput. 30, pp. 796-811. c c *** general *** c c coded by david m. gay (winter 1980). revised summer 1982. c this subroutine was written in connection with research c supported in part by the national science foundation under c grants mcs-7600324, dcr75-10143, 76-14311dss, mcs76-11989, c and mcs-7906671. c. c c---------------------------- declarations --------------------------- c external deflt, sumit c c deflt... supplies default iv and v input components. c sumit... reverse-communication routine that carries out sumsl algo- c rithm. c integer g1, iv1, nf double precision f c c *** subscripts for iv *** c integer nextv, nfcall, nfgcal, g, toobig, vneed c c/6 c data nextv/47/, nfcall/6/, nfgcal/7/, g/28/, toobig/2/, vneed/4/ c/7 parameter (nextv=47, nfcall=6, nfgcal=7, g=28, toobig=2, vneed=4) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c if (iv(1) .eq. 0) call deflt(2, iv, liv, lv, v) iv1 = iv(1) if (iv1 .eq. 12 .or. iv1 .eq. 13) iv(vneed) = iv(vneed) + n if (iv1 .eq. 14) go to 10 if (iv1 .gt. 2 .and. iv1 .lt. 12) go to 10 g1 = 1 if (iv1 .eq. 12) iv(1) = 13 go to 20 c 10 g1 = iv(g) c 20 call sumit(d, f, v(g1), iv, liv, lv, n, v, x) if (iv(1) - 2) 30, 40, 50 c 30 nf = iv(nfcall) call calcf(n, x, nf, f, uiparm, urparm, ufparm) if (nf .le. 0) iv(toobig) = 1 go to 20 c 40 call calcg(n, x, iv(nfgcal), v(g1), uiparm, urparm, ufparm) go to 20 c 50 if (iv(1) .ne. 14) go to 999 c c *** storage allocation c iv(g) = iv(nextv) iv(nextv) = iv(g) + n if (iv1 .ne. 13) go to 10 c 999 return c *** last card of sumsl follows *** end subroutine sumit(d, fx, g, iv, liv, lv, n, v, x) c c *** carry out sumsl (unconstrained minimization) iterations, using c *** double-dogleg/bfgs steps. c c *** parameter declarations *** c integer liv, lv, n integer iv(liv) double precision d(n), fx, g(n), v(lv), x(n) c c-------------------------- parameter usage -------------------------- c c d.... scale vector. c fx... function value. c g.... gradient vector. c iv... integer value array. c liv.. length of iv (at least 60). c lv... length of v (at least 71 + n*(n+13)/2). c n.... number of variables (components in x and g). c v.... floating-point value array. c x.... vector of parameters to be optimized. c c *** discussion *** c c parameters iv, n, v, and x are the same as the corresponding c ones to sumsl (which see), except that v can be shorter (since c the part of v that sumsl uses for storing g is not needed). c moreover, compared with sumsl, iv(1) may have the two additional c output values 1 and 2, which are explained below, as is the use c of iv(toobig) and iv(nfgcal). the value iv(g), which is an c output value from sumsl (and smsno), is not referenced by c sumit or the subroutines it calls. c fx and g need not have been initialized when sumit is called c with iv(1) = 12, 13, or 14. c c iv(1) = 1 means the caller should set fx to f(x), the function value c at x, and call sumit again, having changed none of the c other parameters. an exception occurs if f(x) cannot be c (e.g. if overflow would occur), which may happen because c of an oversized step. in this case the caller should set c iv(toobig) = iv(2) to 1, which will cause sumit to ig- c nore fx and try a smaller step. the parameter nf that c sumsl passes to calcf (for possible use by calcg) is a c copy of iv(nfcall) = iv(6). c iv(1) = 2 means the caller should set g to g(x), the gradient vector c of f at x, and call sumit again, having changed none of c the other parameters except possibly the scale vector d c when iv(dtype) = 0. the parameter nf that sumsl passes c to calcg is iv(nfgcal) = iv(7). if g(x) cannot be c evaluated, then the caller may set iv(nfgcal) to 0, in c which case sumit will return with iv(1) = 65. c. c *** general *** c c coded by david m. gay (december 1979). revised sept. 1982. c this subroutine was written in connection with research supported c in part by the national science foundation under grants c mcs-7600324 and mcs-7906671. c c (see sumsl for references.) c c+++++++++++++++++++++++++++ declarations ++++++++++++++++++++++++++++ c c *** local variables *** c integer dg1, dummy, g01, i, k, l, lstgst, nwtst1, step1, 1 temp1, w, x01, z double precision t c c *** constants *** c double precision half, negone, one, onep2, zero c c *** no intrinsic functions *** c c *** external functions and subroutines *** c external assst, dbdog, deflt, dotprd, itsum, litvmu, livmul, 1 ltvmul, lupdat, lvmul, parck, reldst, stopx, vaxpy, 2 vcopy, vscopy, vvmulp, v2norm, wzbfgs logical stopx double precision dotprd, reldst, v2norm c c assst.... assesses candidate step. c dbdog.... computes double-dogleg (candidate) step. c deflt.... supplies default iv and v input components. c dotprd... returns inner product of two vectors. c itsum.... prints iteration summary and info on initial and final x. c litvmu... multiplies inverse transpose of lower triangle times vector. c livmul... multiplies inverse of lower triangle times vector. c ltvmul... multiplies transpose of lower triangle times vector. c lupdt.... updates cholesky factor of hessian approximation. c lvmul.... multiplies lower triangle times vector. c parck.... checks validity of input iv and v values. c reldst... computes v(reldx) = relative step size. c stopx.... returns .true. if the break key has been pressed. c vaxpy.... computes scalar times one vector plus another. c vcopy.... copies one vector to another. c vscopy... sets all elements of a vector to a scalar. c vvmulp... multiplies vector by vector raised to power (componentwise). c v2norm... returns the 2-norm of a vector. c wzbfgs... computes w and z for lupdat corresponding to bfgs update. c c *** subscripts for iv and v *** c integer afctol integer cnvcod, dg, dgnorm, dinit, dstnrm, dst0, f, f0, fdif, 1 gthg, gtstep, g0, incfac, inith, irc, kagqt, lmat, lmax0, 2 lmaxs, mode, model, mxfcal, mxiter, nextv, nfcall, nfgcal, 3 ngcall, niter, nreduc, nwtstp, preduc, radfac, radinc, 4 radius, rad0, reldx, restor, step, stglim, stlstg, toobig, 5 tuner4, tuner5, vneed, xirc, x0 c c *** iv subscript values *** c c/6 c data cnvcod/55/, dg/37/, g0/48/, inith/25/, irc/29/, kagqt/33/, c 1 mode/35/, model/5/, mxfcal/17/, mxiter/18/, nfcall/6/, c 2 nfgcal/7/, ngcall/30/, niter/31/, nwtstp/34/, radinc/8/, c 3 restor/9/, step/40/, stglim/11/, stlstg/41/, toobig/2/, c 4 vneed/4/, xirc/13/, x0/43/ c/7 parameter (cnvcod=55, dg=37, g0=48, inith=25, irc=29, kagqt=33, 1 mode=35, model=5, mxfcal=17, mxiter=18, nfcall=6, 2 nfgcal=7, ngcall=30, niter=31, nwtstp=34, radinc=8, 3 restor=9, step=40, stglim=11, stlstg=41, toobig=2, 4 vneed=4, xirc=13, x0=43) c/ c c *** v subscript values *** c c/6 c data afctol/31/ c data dgnorm/1/, dinit/38/, dstnrm/2/, dst0/3/, f/10/, f0/13/, c 1 fdif/11/, gthg/44/, gtstep/4/, incfac/23/, lmat/42/, c 2 lmax0/35/, lmaxs/36/, nextv/47/, nreduc/6/, preduc/7/, c 3 radfac/16/, radius/8/, rad0/9/, reldx/17/, tuner4/29/, c 4 tuner5/30/ c/7 parameter (afctol=31) parameter (dgnorm=1, dinit=38, dstnrm=2, dst0=3, f=10, f0=13, 1 fdif=11, gthg=44, gtstep=4, incfac=23, lmat=42, 2 lmax0=35, lmaxs=36, nextv=47, nreduc=6, preduc=7, 3 radfac=16, radius=8, rad0=9, reldx=17, tuner4=29, 4 tuner5=30) c/ c c/6 c data half/0.5d+0/, negone/-1.d+0/, one/1.d+0/, onep2/1.2d+0/, c 1 zero/0.d+0/ c/7 parameter (half=0.5d+0, negone=-1.d+0, one=1.d+0, onep2=1.2d+0, 1 zero=0.d+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c C Following SAVE statement inserted. save l i = iv(1) if (i .eq. 1) go to 50 if (i .eq. 2) go to 60 c c *** check validity of iv and v input values *** c if (iv(1) .eq. 0) call deflt(2, iv, liv, lv, v) if (iv(1) .eq. 12 .or. iv(1) .eq. 13) 1 iv(vneed) = iv(vneed) + n*(n+13)/2 call parck(2, d, iv, liv, lv, n, v) i = iv(1) - 2 if (i .gt. 12) go to 999 go to (180, 180, 180, 180, 180, 180, 120, 90, 120, 10, 10, 20), i c c *** storage allocation *** c 10 l = iv(lmat) iv(x0) = l + n*(n+1)/2 iv(step) = iv(x0) + n iv(stlstg) = iv(step) + n iv(g0) = iv(stlstg) + n iv(nwtstp) = iv(g0) + n iv(dg) = iv(nwtstp) + n iv(nextv) = iv(dg) + n if (iv(1) .ne. 13) go to 20 iv(1) = 14 go to 999 c c *** initialization *** c 20 iv(niter) = 0 iv(nfcall) = 1 iv(ngcall) = 1 iv(nfgcal) = 1 iv(mode) = -1 iv(model) = 1 iv(stglim) = 1 iv(toobig) = 0 iv(cnvcod) = 0 iv(radinc) = 0 v(rad0) = zero if (v(dinit) .ge. zero) call vscopy(n, d, v(dinit)) if (iv(inith) .ne. 1) go to 40 c c *** set the initial hessian approximation to diag(d)**-2 *** c l = iv(lmat) call vscopy(n*(n+1)/2, v(l), zero) k = l - 1 do 30 i = 1, n k = k + i t = d(i) if (t .le. zero) t = one v(k) = t 30 continue c c *** compute initial function value *** c 40 iv(1) = 1 go to 999 c 50 v(f) = fx if (iv(mode) .ge. 0) go to 180 iv(1) = 2 if (iv(toobig) .eq. 0) go to 999 iv(1) = 63 go to 300 c c *** make sure gradient could be computed *** c 60 if (iv(nfgcal) .ne. 0) go to 70 iv(1) = 65 go to 300 c 70 dg1 = iv(dg) call vvmulp(n, v(dg1), g, d, -1) v(dgnorm) = v2norm(n, v(dg1)) c c *** test norm of gradient *** c if (v(dgnorm) .gt. v(afctol)) go to 75 iv(irc) = 10 iv(cnvcod) = iv(irc) - 4 c 75 if (iv(cnvcod) .ne. 0) go to 290 if (iv(mode) .eq. 0) go to 250 c c *** allow first step to have scaled 2-norm at most v(lmax0) *** c v(radius) = v(lmax0) c iv(mode) = 0 c c c----------------------------- main loop ----------------------------- c c c *** print iteration summary, check iteration limit *** c 80 call itsum(d, g, iv, liv, lv, n, v, x) 90 k = iv(niter) if (k .lt. iv(mxiter)) go to 100 iv(1) = 10 go to 300 c c *** update radius *** c 100 iv(niter) = k + 1 if(k.gt.0)v(radius) = v(radfac) * v(dstnrm) c c *** initialize for start of next iteration *** c g01 = iv(g0) x01 = iv(x0) v(f0) = v(f) iv(irc) = 4 iv(kagqt) = -1 c c *** copy x to x0, g to g0 *** c call vcopy(n, v(x01), x) call vcopy(n, v(g01), g) c c *** check stopx and function evaluation limit *** c C AL 4/30/95 dummy=iv(nfcall) 110 if (.not. stopx(dummy)) go to 130 iv(1) = 11 go to 140 c c *** come here when restarting after func. eval. limit or stopx. c 120 if (v(f) .ge. v(f0)) go to 130 v(radfac) = one k = iv(niter) go to 100 c 130 if (iv(nfcall) .lt. iv(mxfcal)) go to 150 iv(1) = 9 140 if (v(f) .ge. v(f0)) go to 300 c c *** in case of stopx or function evaluation limit with c *** improved v(f), evaluate the gradient at x. c iv(cnvcod) = iv(1) go to 240 c c. . . . . . . . . . . . . compute candidate step . . . . . . . . . . c 150 step1 = iv(step) dg1 = iv(dg) nwtst1 = iv(nwtstp) if (iv(kagqt) .ge. 0) go to 160 l = iv(lmat) call livmul(n, v(nwtst1), v(l), g) v(nreduc) = half * dotprd(n, v(nwtst1), v(nwtst1)) call litvmu(n, v(nwtst1), v(l), v(nwtst1)) call vvmulp(n, v(step1), v(nwtst1), d, 1) v(dst0) = v2norm(n, v(step1)) call vvmulp(n, v(dg1), v(dg1), d, -1) call ltvmul(n, v(step1), v(l), v(dg1)) v(gthg) = v2norm(n, v(step1)) iv(kagqt) = 0 160 call dbdog(v(dg1), lv, n, v(nwtst1), v(step1), v) if (iv(irc) .eq. 6) go to 180 c c *** check whether evaluating f(x0 + step) looks worthwhile *** c if (v(dstnrm) .le. zero) go to 180 if (iv(irc) .ne. 5) go to 170 if (v(radfac) .le. one) go to 170 if (v(preduc) .le. onep2 * v(fdif)) go to 180 c c *** compute f(x0 + step) *** c 170 x01 = iv(x0) step1 = iv(step) call vaxpy(n, x, one, v(step1), v(x01)) iv(nfcall) = iv(nfcall) + 1 iv(1) = 1 iv(toobig) = 0 go to 999 c c. . . . . . . . . . . . . assess candidate step . . . . . . . . . . . c 180 x01 = iv(x0) v(reldx) = reldst(n, d, x, v(x01)) call assst(iv, liv, lv, v) step1 = iv(step) lstgst = iv(stlstg) if (iv(restor) .eq. 1) call vcopy(n, x, v(x01)) if (iv(restor) .eq. 2) call vcopy(n, v(lstgst), v(step1)) if (iv(restor) .ne. 3) go to 190 call vcopy(n, v(step1), v(lstgst)) call vaxpy(n, x, one, v(step1), v(x01)) v(reldx) = reldst(n, d, x, v(x01)) c 190 k = iv(irc) go to (200,230,230,230,200,210,220,220,220,220,220,220,280,250), k c c *** recompute step with changed radius *** c 200 v(radius) = v(radfac) * v(dstnrm) go to 110 c c *** compute step of length v(lmaxs) for singular convergence test. c 210 v(radius) = v(lmaxs) go to 150 c c *** convergence or false convergence *** c 220 iv(cnvcod) = k - 4 if (v(f) .ge. v(f0)) go to 290 if (iv(xirc) .eq. 14) go to 290 iv(xirc) = 14 c c. . . . . . . . . . . . process acceptable step . . . . . . . . . . . c 230 if (iv(irc) .ne. 3) go to 240 step1 = iv(step) temp1 = iv(stlstg) c c *** set temp1 = hessian * step for use in gradient tests *** c l = iv(lmat) call ltvmul(n, v(temp1), v(l), v(step1)) call lvmul(n, v(temp1), v(l), v(temp1)) c c *** compute gradient *** c 240 iv(ngcall) = iv(ngcall) + 1 iv(1) = 2 go to 999 c c *** initializations -- g0 = g - g0, etc. *** c 250 g01 = iv(g0) call vaxpy(n, v(g01), negone, v(g01), g) step1 = iv(step) temp1 = iv(stlstg) if (iv(irc) .ne. 3) go to 270 c c *** set v(radfac) by gradient tests *** c c *** set temp1 = diag(d)**-1 * (hessian*step + (g(x0)-g(x))) *** c call vaxpy(n, v(temp1), negone, v(g01), v(temp1)) call vvmulp(n, v(temp1), v(temp1), d, -1) c c *** do gradient tests *** c if (v2norm(n, v(temp1)) .le. v(dgnorm) * v(tuner4)) 1 go to 260 if (dotprd(n, g, v(step1)) 1 .ge. v(gtstep) * v(tuner5)) go to 270 260 v(radfac) = v(incfac) c c *** update h, loop *** c 270 w = iv(nwtstp) z = iv(x0) l = iv(lmat) call wzbfgs(v(l), n, v(step1), v(w), v(g01), v(z)) c c ** use the n-vectors starting at v(step1) and v(g01) for scratch.. call lupdat(v(temp1), v(step1), v(l), v(g01), v(l), n, v(w), v(z)) iv(1) = 2 go to 80 c c. . . . . . . . . . . . . . misc. details . . . . . . . . . . . . . . c c *** bad parameters to assess *** c 280 iv(1) = 64 go to 300 c c *** print summary of final iteration and other requested items *** c 290 iv(1) = iv(cnvcod) iv(cnvcod) = 0 300 call itsum(d, g, iv, liv, lv, n, v, x) c 999 return c c *** last line of sumit follows *** end subroutine dbdog(dig, lv, n, nwtstp, step, v) c c *** compute double dogleg step *** c c *** parameter declarations *** c integer lv, n double precision dig(n), nwtstp(n), step(n), v(lv) c c *** purpose *** c c this subroutine computes a candidate step (for use in an uncon- c strained minimization code) by the double dogleg algorithm of c dennis and mei (ref. 1), which is a variation on powell*s dogleg c scheme (ref. 2, p. 95). c c-------------------------- parameter usage -------------------------- c c dig (input) diag(d)**-2 * g -- see algorithm notes. c g (input) the current gradient vector. c lv (input) length of v. c n (input) number of components in dig, g, nwtstp, and step. c nwtstp (input) negative newton step -- see algorithm notes. c step (output) the computed step. c v (i/o) values array, the following components of which are c used here... c v(bias) (input) bias for relaxed newton step, which is v(bias) of c the way from the full newton to the fully relaxed newton c step. recommended value = 0.8 . c v(dgnorm) (input) 2-norm of diag(d)**-1 * g -- see algorithm notes. c v(dstnrm) (output) 2-norm of diag(d) * step, which is v(radius) c unless v(stppar) = 0 -- see algorithm notes. c v(dst0) (input) 2-norm of diag(d) * nwtstp -- see algorithm notes. c v(grdfac) (output) the coefficient of dig in the step returned -- c step(i) = v(grdfac)*dig(i) + v(nwtfac)*nwtstp(i). c v(gthg) (input) square-root of (dig**t) * (hessian) * dig -- see c algorithm notes. c v(gtstep) (output) inner product between g and step. c v(nreduc) (output) function reduction predicted for the full newton c step. c v(nwtfac) (output) the coefficient of nwtstp in the step returned -- c see v(grdfac) above. c v(preduc) (output) function reduction predicted for the step returned. c v(radius) (input) the trust region radius. d times the step returned c has 2-norm v(radius) unless v(stppar) = 0. c v(stppar) (output) code telling how step was computed... 0 means a c full newton step. between 0 and 1 means v(stppar) of the c way from the newton to the relaxed newton step. between c 1 and 2 means a true double dogleg step, v(stppar) - 1 of c the way from the relaxed newton to the cauchy step. c greater than 2 means 1 / (v(stppar) - 1) times the cauchy c step. c c------------------------------- notes ------------------------------- c c *** algorithm notes *** c c let g and h be the current gradient and hessian approxima- c tion respectively and let d be the current scale vector. this c routine assumes dig = diag(d)**-2 * g and nwtstp = h**-1 * g. c the step computed is the same one would get by replacing g and h c by diag(d)**-1 * g and diag(d)**-1 * h * diag(d)**-1, c computing step, and translating step back to the original c variables, i.e., premultiplying it by diag(d)**-1. c c *** references *** c c 1. dennis, j.e., and mei, h.h.w. (1979), two new unconstrained opti- c mization algorithms which use function and gradient c values, j. optim. theory applic. 28, pp. 453-482. c 2. powell, m.j.d. (1970), a hybrid method for non-linear equations, c in numerical methods for non-linear equations, edited by c p. rabinowitz, gordon and breach, london. c c *** general *** c c coded by david m. gay. c this subroutine was written in connection with research supported c by the national science foundation under grants mcs-7600324 and c mcs-7906671. c c------------------------ external quantities ------------------------ c c *** functions and subroutines called *** c external dotprd, v2norm double precision dotprd, v2norm c c dotprd... returns inner product of two vectors. c v2norm... returns 2-norm of a vector. c c *** intrinsic functions *** c/+ double precision dsqrt c/ c-------------------------- local variables -------------------------- c integer i double precision cfact, cnorm, ctrnwt, ghinvg, femnsq, gnorm, 1 nwtnrm, relax, rlambd, t, t1, t2 double precision half, one, two, zero c c *** v subscripts *** c integer bias, dgnorm, dstnrm, dst0, grdfac, gthg, gtstep, 1 nreduc, nwtfac, preduc, radius, stppar c c *** data initializations *** c c/6 c data half/0.5d+0/, one/1.d+0/, two/2.d+0/, zero/0.d+0/ c/7 parameter (half=0.5d+0, one=1.d+0, two=2.d+0, zero=0.d+0) c/ c c/6 c data bias/43/, dgnorm/1/, dstnrm/2/, dst0/3/, grdfac/45/, c 1 gthg/44/, gtstep/4/, nreduc/6/, nwtfac/46/, preduc/7/, c 2 radius/8/, stppar/5/ c/7 parameter (bias=43, dgnorm=1, dstnrm=2, dst0=3, grdfac=45, 1 gthg=44, gtstep=4, nreduc=6, nwtfac=46, preduc=7, 2 radius=8, stppar=5) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c nwtnrm = v(dst0) rlambd = one if (nwtnrm .gt. zero) rlambd = v(radius) / nwtnrm gnorm = v(dgnorm) ghinvg = two * v(nreduc) v(grdfac) = zero v(nwtfac) = zero if (rlambd .lt. one) go to 30 c c *** the newton step is inside the trust region *** c v(stppar) = zero v(dstnrm) = nwtnrm v(gtstep) = -ghinvg v(preduc) = v(nreduc) v(nwtfac) = -one do 20 i = 1, n 20 step(i) = -nwtstp(i) go to 999 c 30 v(dstnrm) = v(radius) cfact = (gnorm / v(gthg))**2 c *** cauchy step = -cfact * g. cnorm = gnorm * cfact relax = one - v(bias) * (one - gnorm*cnorm/ghinvg) if (rlambd .lt. relax) go to 50 c c *** step is between relaxed newton and full newton steps *** c v(stppar) = one - (rlambd - relax) / (one - relax) t = -rlambd v(gtstep) = t * ghinvg v(preduc) = rlambd * (one - half*rlambd) * ghinvg v(nwtfac) = t do 40 i = 1, n 40 step(i) = t * nwtstp(i) go to 999 c 50 if (cnorm .lt. v(radius)) go to 70 c c *** the cauchy step lies outside the trust region -- c *** step = scaled cauchy step *** c t = -v(radius) / gnorm v(grdfac) = t v(stppar) = one + cnorm / v(radius) v(gtstep) = -v(radius) * gnorm v(preduc) = v(radius)*(gnorm - half*v(radius)*(v(gthg)/gnorm)**2) do 60 i = 1, n 60 step(i) = t * dig(i) go to 999 c c *** compute dogleg step between cauchy and relaxed newton *** c *** femur = relaxed newton step minus cauchy step *** c 70 ctrnwt = cfact * relax * ghinvg / gnorm c *** ctrnwt = inner prod. of cauchy and relaxed newton steps, c *** scaled by gnorm**-1. t1 = ctrnwt - gnorm*cfact**2 c *** t1 = inner prod. of femur and cauchy step, scaled by c *** gnorm**-1. t2 = v(radius)*(v(radius)/gnorm) - gnorm*cfact**2 t = relax * nwtnrm femnsq = (t/gnorm)*t - ctrnwt - t1 c *** femnsq = square of 2-norm of femur, scaled by gnorm**-1. t = t2 / (t1 + dsqrt(t1**2 + femnsq*t2)) c *** dogleg step = cauchy step + t * femur. t1 = (t - one) * cfact v(grdfac) = t1 t2 = -t * relax v(nwtfac) = t2 v(stppar) = two - t v(gtstep) = t1*gnorm**2 + t2*ghinvg v(preduc) = -t1*gnorm * ((t2 + one)*gnorm) 1 - t2 * (one + half*t2)*ghinvg 2 - half * (v(gthg)*t1)**2 do 80 i = 1, n 80 step(i) = t1*dig(i) + t2*nwtstp(i) c 999 return c *** last line of dbdog follows *** end subroutine ltvmul(n, x, l, y) c c *** compute x = (l**t)*y, where l is an n x n lower c *** triangular matrix stored compactly by rows. x and y may c *** occupy the same storage. *** c integer n cal double precision x(n), l(1), y(n) double precision x(n), l(n*(n+1)/2), y(n) c dimension l(n*(n+1)/2) integer i, ij, i0, j double precision yi, zero c/6 c data zero/0.d+0/ c/7 parameter (zero=0.d+0) c/ c i0 = 0 do 20 i = 1, n yi = y(i) x(i) = zero do 10 j = 1, i ij = i0 + j x(j) = x(j) + yi*l(ij) 10 continue i0 = i0 + i 20 continue 999 return c *** last card of ltvmul follows *** end subroutine lupdat(beta, gamma, l, lambda, lplus, n, w, z) c c *** compute lplus = secant update of l *** c c *** parameter declarations *** c integer n cal double precision beta(n), gamma(n), l(1), lambda(n), lplus(1), double precision beta(n), gamma(n), l(n*(n+1)/2), lambda(n), 1 lplus(n*(n+1)/2),w(n), z(n) c dimension l(n*(n+1)/2), lplus(n*(n+1)/2) c c-------------------------- parameter usage -------------------------- c c beta = scratch vector. c gamma = scratch vector. c l (input) lower triangular matrix, stored rowwise. c lambda = scratch vector. c lplus (output) lower triangular matrix, stored rowwise, which may c occupy the same storage as l. c n (input) length of vector parameters and order of matrices. c w (input, destroyed on output) right singular vector of rank 1 c correction to l. c z (input, destroyed on output) left singular vector of rank 1 c correction to l. c c------------------------------- notes ------------------------------- c c *** application and usage restrictions *** c c this routine updates the cholesky factor l of a symmetric c positive definite matrix to which a secant update is being c applied -- it computes a cholesky factor lplus of c l * (i + z*w**t) * (i + w*z**t) * l**t. it is assumed that w c and z have been chosen so that the updated matrix is strictly c positive definite. c c *** algorithm notes *** c c this code uses recurrence 3 of ref. 1 (with d(j) = 1 for all j) c to compute lplus of the form l * (i + z*w**t) * q, where q c is an orthogonal matrix that makes the result lower triangular. c lplus may have some negative diagonal elements. c c *** references *** c c 1. goldfarb, d. (1976), factorized variable metric methods for uncon- c strained optimization, math. comput. 30, pp. 796-811. c c *** general *** c c coded by david m. gay (fall 1979). c this subroutine was written in connection with research supported c by the national science foundation under grants mcs-7600324 and c mcs-7906671. c c------------------------ external quantities ------------------------ c c *** intrinsic functions *** c/+ double precision dsqrt c/ c-------------------------- local variables -------------------------- c integer i, ij, j, jj, jp1, k, nm1, np1 double precision a, b, bj, eta, gj, lj, lij, ljj, nu, s, theta, 1 wj, zj double precision one, zero c c *** data initializations *** c c/6 c data one/1.d+0/, zero/0.d+0/ c/7 parameter (one=1.d+0, zero=0.d+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c nu = one eta = zero if (n .le. 1) go to 30 nm1 = n - 1 c c *** temporarily store s(j) = sum over k = j+1 to n of w(k)**2 in c *** lambda(j). c s = zero do 10 i = 1, nm1 j = n - i s = s + w(j+1)**2 lambda(j) = s 10 continue c c *** compute lambda, gamma, and beta by goldfarb*s recurrence 3. c do 20 j = 1, nm1 wj = w(j) a = nu*z(j) - eta*wj theta = one + a*wj s = a*lambda(j) lj = dsqrt(theta**2 + a*s) if (theta .gt. zero) lj = -lj lambda(j) = lj b = theta*wj + s gamma(j) = b * nu / lj beta(j) = (a - b*eta) / lj nu = -nu / lj eta = -(eta + (a**2)/(theta - lj)) / lj 20 continue 30 lambda(n) = one + (nu*z(n) - eta*w(n))*w(n) c c *** update l, gradually overwriting w and z with l*w and l*z. c np1 = n + 1 jj = n * (n + 1) / 2 do 60 k = 1, n j = np1 - k lj = lambda(j) ljj = l(jj) lplus(jj) = lj * ljj wj = w(j) w(j) = ljj * wj zj = z(j) z(j) = ljj * zj if (k .eq. 1) go to 50 bj = beta(j) gj = gamma(j) ij = jj + j jp1 = j + 1 do 40 i = jp1, n lij = l(ij) lplus(ij) = lj*lij + bj*w(i) + gj*z(i) w(i) = w(i) + lij*wj z(i) = z(i) + lij*zj ij = ij + i 40 continue 50 jj = jj - j 60 continue c 999 return c *** last card of lupdat follows *** end subroutine lvmul(n, x, l, y) c c *** compute x = l*y, where l is an n x n lower triangular c *** matrix stored compactly by rows. x and y may occupy the same c *** storage. *** c integer n cal double precision x(n), l(1), y(n) double precision x(n), l(n*(n+1)/2), y(n) c dimension l(n*(n+1)/2) integer i, ii, ij, i0, j, np1 double precision t, zero c/6 c data zero/0.d+0/ c/7 parameter (zero=0.d+0) c/ c np1 = n + 1 i0 = n*(n+1)/2 do 20 ii = 1, n i = np1 - ii i0 = i0 - i t = zero do 10 j = 1, i ij = i0 + j t = t + l(ij)*y(j) 10 continue x(i) = t 20 continue 999 return c *** last card of lvmul follows *** end subroutine vvmulp(n, x, y, z, k) c c *** set x(i) = y(i) * z(i)**k, 1 .le. i .le. n (for k = 1 or -1) *** c integer n, k double precision x(n), y(n), z(n) integer i c if (k .ge. 0) go to 20 do 10 i = 1, n 10 x(i) = y(i) / z(i) go to 999 c 20 do 30 i = 1, n 30 x(i) = y(i) * z(i) 999 return c *** last card of vvmulp follows *** end subroutine wzbfgs (l, n, s, w, y, z) c c *** compute y and z for lupdat corresponding to bfgs update. c integer n cal double precision l(1), s(n), w(n), y(n), z(n) double precision l(n*(n+1)/2), s(n), w(n), y(n), z(n) c dimension l(n*(n+1)/2) c c-------------------------- parameter usage -------------------------- c c l (i/o) cholesky factor of hessian, a lower triang. matrix stored c compactly by rows. c n (input) order of l and length of s, w, y, z. c s (input) the step just taken. c w (output) right singular vector of rank 1 correction to l. c y (input) change in gradients corresponding to s. c z (output) left singular vector of rank 1 correction to l. c c------------------------------- notes ------------------------------- c c *** algorithm notes *** c c when s is computed in certain ways, e.g. by gqtstp or c dbldog, it is possible to save n**2/2 operations since (l**t)*s c or l*(l**t)*s is then known. c if the bfgs update to l*(l**t) would reduce its determinant to c less than eps times its old value, then this routine in effect c replaces y by theta*y + (1 - theta)*l*(l**t)*s, where theta c (between 0 and 1) is chosen to make the reduction factor = eps. c c *** general *** c c coded by david m. gay (fall 1979). c this subroutine was written in connection with research supported c by the national science foundation under grants mcs-7600324 and c mcs-7906671. c c------------------------ external quantities ------------------------ c c *** functions and subroutines called *** c external dotprd, livmul, ltvmul double precision dotprd c dotprd returns inner product of two vectors. c livmul multiplies l**-1 times a vector. c ltvmul multiplies l**t times a vector. c c *** intrinsic functions *** c/+ double precision dsqrt c/ c-------------------------- local variables -------------------------- c integer i double precision cs, cy, eps, epsrt, one, shs, ys, theta c c *** data initializations *** c c/6 c data eps/0.1d+0/, one/1.d+0/ c/7 parameter (eps=0.1d+0, one=1.d+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c call ltvmul(n, w, l, s) shs = dotprd(n, w, w) ys = dotprd(n, y, s) if (ys .ge. eps*shs) go to 10 theta = (one - eps) * shs / (shs - ys) epsrt = dsqrt(eps) cy = theta / (shs * epsrt) cs = (one + (theta-one)/epsrt) / shs go to 20 10 cy = one / (dsqrt(ys) * dsqrt(shs)) cs = one / shs 20 call livmul(n, z, l, y) do 30 i = 1, n 30 z(i) = cy * z(i) - cs * w(i) c 999 return c *** last card of wzbfgs follows *** end