C ALGORITHM 611, COLLECTED ALGORITHMS FROM ACM. C ALGORITHM APPEARED IN ACM-TRANS. MATH. SOFTWARE, VOL.9, NO. 4, C DEC., 1983, P. 503-524. integer function imdcon(k) c integer k c c *** return integer machine-dependent constants *** c c *** k = 1 means return standard output unit number. *** c *** k = 2 means return alternate output unit number. *** c *** k = 3 means return input unit number. *** c (note -- k = 2, 3 are used only by test programs.) c c +++ port version follows... c external i1mach c integer i1mach c integer mdperm(3) c data mdperm(1)/2/, mdperm(2)/4/, mdperm(3)/1/ c imdcon = i1mach(mdperm(k)) c +++ end of port version +++ c c +++ non-port version follows... integer mdcon(3) data mdcon(1)/6/, mdcon(2)/8/, mdcon(3)/5/ imdcon = mdcon(k) c +++ end of non-port version +++ c 999 return c *** last card of imdcon follows *** end double precision function rmdcon(k) c c *** return machine dependent constants used by nl2sol *** c c +++ comments below contain data statements for various machines. +++ c +++ to convert to another machine, place a c in column 1 of the +++ c +++ data statement line(s) that correspond to the current machine +++ c +++ and remove the c from column 1 of the data statement line(s) +++ c +++ that correspond to the new machine. +++ c integer k c c *** the constant returned depends on k... c c *** k = 1... smallest pos. eta such that -eta exists. c *** k = 2... square root of eta. c *** k = 3... unit roundoff = smallest pos. no. machep such c *** that 1 + machep .gt. 1 .and. 1 - machep .lt. 1. c *** k = 4... square root of machep. c *** k = 5... square root of big (see k = 6). c *** k = 6... largest machine no. big such that -big exists. c double precision big, eta, machep integer bigi(4), etai(4), machei(4) c/+ double precision dsqrt c/ equivalence (big,bigi(1)), (eta,etai(1)), (machep,machei(1)) c c +++ ibm 360, ibm 370, or xerox +++ c c data big/z7fffffffffffffff/, eta/z0010000000000000/, c 1 machep/z3410000000000000/ c c +++ data general +++ c c data big/0.7237005577d+76/, eta/0.5397605347d-78/, c 1 machep/2.22044605d-16/ c c +++ dec 11 +++ c c data big/1.7d+38/, eta/2.938735878d-39/, machep/2.775557562d-17/ c c +++ hp3000 +++ c c data big/1.157920892d+77/, eta/8.636168556d-78/, c 1 machep/5.551115124d-17/ c c +++ honeywell +++ c c data big/1.69d+38/, eta/5.9d-39/, machep/2.1680435d-19/ c c +++ dec10 +++ c c data big/"377777100000000000000000/, c 1 eta/"002400400000000000000000/, c 2 machep/"104400000000000000000000/ c c +++ burroughs +++ c c data big/o0777777777777777,o7777777777777777/, c 1 eta/o1771000000000000,o7770000000000000/, c 2 machep/o1451000000000000,o0000000000000000/ c c +++ control data +++ c c data big/37767777777777777777b,37167777777777777777b/, c 1 eta/00014000000000000000b,00000000000000000000b/, c 2 machep/15614000000000000000b,15010000000000000000b/ c c +++ prime +++ c c data big/1.0d+9786/, eta/1.0d-9860/, machep/1.4210855d-14/ c c +++ univac +++ c c data big/8.988d+307/, eta/1.2d-308/, machep/1.734723476d-18/ c c +++ vax +++ c data big/1.7d+38/, eta/2.939d-39/, machep/1.3877788d-17/ c c +++ cray 1 +++ c c data bigi(1)/577767777777777777777b/, c 1 bigi(2)/000007777777777777776b/, c 2 etai(1)/200004000000000000000b/, c 3 etai(2)/000000000000000000000b/, c 4 machei(1)/377224000000000000000b/, c 5 machei(2)/000000000000000000000b/ c c +++ port library -- requires more than just a data statement... +++ c c external d1mach c double precision d1mach, zero c data big/0.d+0/, eta/0.d+0/, machep/0.d+0/, zero/0.d+0/ c if (big .gt. zero) go to 1 c big = d1mach(2) c eta = d1mach(1) c machep = d1mach(4) c1 continue c c +++ end of port +++ c c------------------------------- body -------------------------------- c go to (10, 20, 30, 40, 50, 60), k c 10 rmdcon = eta go to 999 c 20 rmdcon = dsqrt(256.d+0*eta)/16.d+0 go to 999 c 30 rmdcon = machep go to 999 c 40 rmdcon = dsqrt(machep) go to 999 c 50 rmdcon = dsqrt(big/256.d+0)*16.d+0 go to 999 c 60 rmdcon = big c 999 return c *** last card of rmdcon follows *** end 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 iv 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 liv 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 lv 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 iv 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 data nextv/47/, nfcall/6/, nfgcal/7/, g/28/, toobig/2/, vneed/4/ c/7 c 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 smsno(n, d, x, calcf, iv, liv, lv, v, 1 uiparm, urparm, ufparm) c c *** minimize general unconstrained objective function using c *** finite-difference gradients and secant hessian approximations. c integer n, liv, lv integer iv(liv), uiparm(1) double precision d(n), x(n), v(lv), urparm(1) c dimension v(77 + n*(n+17)/2), uiparm(*), urparm(*) external calcf, ufparm c c *** purpose *** c c this routine interacts with subroutine snoit 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 *** parameters *** c c the parameters for smsno are the same as those for sumsl c (which see), except that calcg is omitted. instead of calling c calcg to obtain the gradient of the objective function at x, c smsno calls sgrad2, which computes an approximation to the c gradient by finite (forward and central) differences using the c method of ref. 1. the following input component is of interest c in this regard (and is not described in sumsl). c c v(eta0)..... v(42) is an estimated bound on the relative error in the c objective function value computed by calcf... c (true value) = (computed value) * (1 + e), c where abs(e) .le. v(eta0). default = machep * 10**3, c where machep is the unit roundoff. c c the output values iv(nfcall) and iv(ngcall) have different c meanings for smsno than for sumsl... c c iv(nfcall)... iv(6) is the number of calls so far made on calcf (i.e., c function evaluations) excluding those made only for c computing gradients. the input value iv(mxfcal) is a c limit on iv(nfcall). c iv(ngcall)... iv(30) is the number of function evaluations made only c for computing gradients. the total number of function c evaluations is thus iv(nfcall) + iv(ngcall). c c *** reference *** c c 1. stewart, g.w. (1967), a modification of davidon*s minimization c method to accept difference approximations of derivatives, c j. assoc. comput. mach. 14, pp. 72-83. c. c *** general *** c c coded by david m. gay (winter 1980). revised sept. 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 snoit c c snoit.... oversees computation of finite-difference gradient and c calls sumit to carry out sumsl algorithm. c integer nf double precision fx c c *** subscripts for iv *** c integer nfcall, toobig c c/6 data nfcall/6/, toobig/2/ c/7 c parameter (nfcall=6, toobig=2) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c 10 call snoit(d, fx, iv, liv, lv, n, v, x) if (iv(1) .gt. 2) go to 999 c c *** compute function *** c nf = iv(nfcall) call calcf(n, x, nf, fx, uiparm, urparm, ufparm) if (nf .le. 0) iv(toobig) = 1 go to 10 c c 999 return c *** last card of smsno 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 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 data 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/7 c parameter (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/ c c *** v subscript values *** c c/6 data 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/7 c parameter (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/ c c/6 data half/0.5d+0/, negone/-1.d+0/, one/1.d+0/, onep2/1.2d+0/, 1 zero/0.d+0/ c/7 c parameter (half=0.5d+0, negone=-1.d+0, one=1.d+0, onep2=1.2d+0, c 1 zero=0.d+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c 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 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 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 snoit(d, fx, iv, liv, lv, n, v, x) c c *** iteration driver for smsno... c *** minimize general unconstrained objective function using c *** finite-difference gradients and secant hessian approximations. c integer liv, lv, n integer iv(liv) double precision d(n), fx, x(n), v(lv) c dimension v(77 + n*(n+17)/2) 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 fx = f(x) computed by the caller. (often c the x* found is a local minimizer rather than a global one.) c c *** parameters *** c c the parameters for snoit are the same as those for sumsl c (which see), except that calcf, calcg, uiparm, urparm, and ufparm c are omitted, and a parameter fx for the objective function c value at x is added. instead of calling calcg to obtain the c gradient of the objective function at x, snoit calls sgrad2, c which computes an approximation to the gradient by finite c (forward and central) differences using the method of ref. 1. c the following input component is of interest in this regard c (and is not described in sumsl). c c v(eta0)..... v(42) is an estimated bound on the relative error in the c objective function value computed by calcf... c (true value) = (computed value) * (1 + e), c where abs(e) .le. v(eta0). default = machep * 10**3, c where machep is the unit roundoff. c c the output values iv(nfcall) and iv(ngcall) have different c meanings for smsno than for sumsl... c c iv(nfcall)... iv(6) is the number of calls so far made on calcf (i.e., c function evaluations) excluding those made only for c computing gradients. the input value iv(mxfcal) is a c limit on iv(nfcall). c iv(ngcall)... iv(30) is the number of function evaluations made only c for computing gradients. the total number of function c evaluations is thus iv(nfcall) + iv(ngcall). c c *** references *** c c 1. stewart, g.w. (1967), a modification of davidon*s minimization c method to accept difference approximations of derivatives, c j. assoc. comput. mach. 14, pp. 72-83. c. c *** general *** c c coded by david m. gay (august 1982). c c---------------------------- declarations --------------------------- c external deflt, dotprd, sgrad2, sumit, vscopy double precision dotprd c c deflt.... supplies default parameter values. c dotprd... returns inner product of two vectors. c sgrad2... computes finite-difference gradient approximation. c sumit.... reverse-communication routine that does sumsl algorithm. c vscopy... sets all elements of a vector to a scalar. c integer alpha, g1, i, iv1, j, k, w double precision one, zero c c *** subscripts for iv *** c integer dtype, eta0, f, g, lmat, nextv, nfcall, nfgcal, ngcall, 1 niter, sgirc, toobig, vneed c c/6 data dtype/16/, eta0/42/, f/10/, g/28/, lmat/42/, nextv/47/, 1 nfcall/6/, nfgcal/7/, ngcall/30/, niter/31/, sgirc/57/, 2 toobig/2/, vneed/4/ c/7 c parameter (dtype=16, eta0=42, f=10, g=28, lmat=42, nextv=47, c 1 nfcall=6, nfgcal=7, ngcall=30, niter=31, sgirc=57, c 2 toobig=2, vneed=4) c/ c/6 data one/1.d+0/, zero/0.d+0/ c/7 c parameter (one=1.d+0, zero=0.d+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c iv1 = iv(1) if (iv1 .eq. 1) go to 10 if (iv1 .eq. 2) go to 50 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) + 2*n + 6 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, fx, v(g1), iv, liv, lv, n, v, x) if (iv(1) - 2) 999, 30, 70 c c *** compute gradient *** c 30 if (iv(niter) .eq. 0) call vscopy(n, v(g1), zero) j = iv(lmat) k = g1 - n do 40 i = 1, n v(k) = dotprd(i, v(j), v(j)) k = k + 1 j = j + i 40 continue c *** undo increment of iv(ngcall) done by sumit *** iv(ngcall) = iv(ngcall) - 1 c *** store return code from sgrad2 in iv(sgirc) *** iv(sgirc) = 0 c *** x may have been restored, so copy back fx... *** fx = v(f) go to 60 c c *** gradient loop *** c 50 if (iv(toobig) .eq. 0) go to 60 iv(nfgcal) = 0 go to 10 c 60 g1 = iv(g) alpha = g1 - n w = alpha - 6 call sgrad2(v(alpha), d, v(eta0), fx, v(g1), iv(sgirc), n, v(w),x) if (iv(sgirc) .eq. 0) go to 10 iv(ngcall) = iv(ngcall) + 1 go to 999 c 70 if (iv(1) .ne. 14) go to 999 c c *** storage allocation *** c iv(g) = iv(nextv) + n + 6 iv(nextv) = iv(g) + n if (iv1 .ne. 13) go to 10 c 999 return c *** last card of snoit 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 data half/0.5d+0/, one/1.d+0/, two/2.d+0/, zero/0.d+0/ c/7 c parameter (half=0.5d+0, one=1.d+0, two=2.d+0, zero=0.d+0) c/ c c/6 data 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/7 c parameter (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/ 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 double precision x(n), l(1), y(n) c dimension l(n*(n+1)/2) integer i, ij, i0, j double precision yi, zero c/6 data zero/0.d+0/ c/7 c 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 double precision beta(n), gamma(n), l(1), lambda(n), lplus(1), 1 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 data one/1.d+0/, zero/0.d+0/ c/7 c 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 double precision x(n), l(1), y(n) c dimension l(n*(n+1)/2) integer i, ii, ij, i0, j, np1 double precision t, zero c/6 data zero/0.d+0/ c/7 c 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 sgrad2 (alpha, d, eta0, fx, g, irc, n, w, x) c c *** compute finite difference gradient by stweart*s scheme *** c c *** parameters *** c integer irc, n double precision alpha(n), d(n), eta0, fx, g(n), w(6), x(n) c c....................................................................... c c *** purpose *** c c this subroutine uses an embellished form of the finite-differ- c ence scheme proposed by stewart (ref. 1) to approximate the c gradient of the function f(x), whose values are supplied by c reverse communication. c c *** parameter description *** c c alpha in (approximate) diagonal elements of the hessian of f(x). c d in scale vector such that d(i)*x(i), i = 1,...,n, are in c comparable units. c eta0 in estimated bound on relative error in the function value... c (true value) = (computed value)*(1+e), where c abs(e) .le. eta0. c fx i/o on input, fx must be the computed value of f(x). on c output with irc = 0, fx has been restored to its original c value, the one it had when sgrad2 was last called with c irc = 0. c g i/o on input with irc = 0, g should contain an approximation c to the gradient of f near x, e.g., the gradient at the c previous iterate. when sgrad2 returns with irc = 0, g is c the desired finite-difference approximation to the c gradient at x. c irc i/o input/return code... before the very first call on sgrad2, c the caller must set irc to 0. whenever sgrad2 returns a c nonzero value for irc, it has perturbed some component of c x... the caller should evaluate f(x) and call sgrad2 c again with fx = f(x). c n in the number of variables (components of x) on which f c depends. c x i/o on input with irc = 0, x is the point at which the c gradient of f is desired. on output with irc nonzero, x c is the point at which f should be evaluated. on output c with irc = 0, x has been restored to its original value c (the one it had when sgrad2 was last called with irc = 0) c and g contains the desired gradient approximation. c w i/o work vector of length 6 in which sgrad2 saves certain c quantities while the caller is evaluating f(x) at a c perturbed x. c c *** application and usage restrictions *** c c this routine is intended for use with quasi-newton routines c for unconstrained minimization (in which case alpha comes from c the diagonal of the quasi-newton hessian approximation). c c *** algorithm notes *** c c this code departs from the scheme proposed by stewart (ref. 1) c in its guarding against overly large or small step sizes and its c handling of special cases (such as zero components of alpha or g). c c *** references *** c c 1. stewart, g.w. (1967), a modification of davidon*s minimization c method to accept difference approximations of derivatives, c j. assoc. comput. mach. 14, pp. 72-83. c c *** history *** c c designed and coded by david m. gay (summer 1977/summer 1980). c c *** general *** c c this routine was prepared in connection with work supported by c the national science foundation under grants mcs76-00324 and c mcs-7906671. c c....................................................................... c c ***** external function ***** c external rmdcon double precision rmdcon c rmdcon... returns machine-dependent constants. c c ***** intrinsic functions ***** c/+ integer iabs double precision dabs, dmax1, dsqrt c/ c ***** local variables ***** c integer fh, fx0, hsave, i, xisave double precision aai, afx, afxeta, agi, alphai, axi, axibar, 1 discon, eta, gi, h, hmin double precision c2000, four, hmax0, hmin0, h0, machep, one, p002, 1 three, two, zero c c/6 data c2000/2.0d+3/, four/4.0d+0/, hmax0/0.02d+0/, hmin0/5.0d+1/, 1 one/1.0d+0/, p002/0.002d+0/, three/3.0d+0/, 2 two/2.0d+0/, zero/0.0d+0/ c/7 c parameter (c2000=2.0d+3, four=4.0d+0, hmax0=0.02d+0, hmin0=5.0d+1, c 1 one=1.0d+0, p002=0.002d+0, three=3.0d+0, c 2 two=2.0d+0, zero=0.0d+0) c/ c/6 data fh/3/, fx0/4/, hsave/5/, xisave/6/ c/7 c parameter (fh=3, fx0=4, hsave=5, xisave=6) c/ c c--------------------------------- body ------------------------------ c if (irc) 140, 100, 210 c c *** fresh start -- get machine-dependent constants *** c c store machep in w(1) and h0 in w(2), where machep is the unit c roundoff (the smallest positive number such that c 1 + machep .gt. 1 and 1 - machep .lt. 1), and h0 is the c square-root of machep. c 100 w(1) = rmdcon(3) w(2) = dsqrt(w(1)) c w(fx0) = fx c c *** increment i and start computing g(i) *** c 110 i = iabs(irc) + 1 if (i .gt. n) go to 300 irc = i afx = dabs(w(fx0)) machep = w(1) h0 = w(2) hmin = hmin0 * machep w(xisave) = x(i) axi = dabs(x(i)) axibar = dmax1(axi, one/d(i)) gi = g(i) agi = dabs(gi) eta = dabs(eta0) if (afx .gt. zero) eta = dmax1(eta, agi*axi*machep/afx) alphai = alpha(i) if (alphai .eq. zero) go to 170 if (gi .eq. zero .or. fx .eq. zero) go to 180 afxeta = afx*eta aai = dabs(alphai) c c *** compute h = stewart*s forward-difference step size. c if (gi**2 .le. afxeta*aai) go to 120 h = two*dsqrt(afxeta/aai) h = h*(one - aai*h/(three*aai*h + four*agi)) go to 130 120 h = two*(afxeta*agi/(aai**2))**(one/three) h = h*(one - two*agi/(three*aai*h + four*agi)) c c *** ensure that h is not insignificantly small *** c 130 h = dmax1(h, hmin*axibar) c c *** use forward difference if bound on truncation error is at c *** most 10**-3. c if (aai*h .le. p002*agi) go to 160 c c *** compute h = stewart*s step for central difference. c discon = c2000*afxeta h = discon/(agi + dsqrt(gi**2 + aai*discon)) c c *** ensure that h is neither too small nor too big *** c h = dmax1(h, hmin*axibar) if (h .ge. hmax0*axibar) h = axibar * h0**(two/three) c c *** compute central difference *** c irc = -i go to 200 c 140 h = -w(hsave) i = iabs(irc) if (h .gt. zero) go to 150 w(fh) = fx go to 200 c 150 g(i) = (w(fh) - fx) / (two * h) x(i) = w(xisave) go to 110 c c *** compute forward differences in various cases *** c 160 if (h .ge. hmax0*axibar) h = h0 * axibar if (alphai*gi .lt. zero) h = -h go to 200 170 h = axibar go to 200 180 h = h0 * axibar c 200 x(i) = w(xisave) + h w(hsave) = h go to 999 c c *** compute actual forward difference *** c 210 g(irc) = (fx - w(fx0)) / w(hsave) x(irc) = w(xisave) go to 110 c c *** restore fx and indicate that g has been computed *** c 300 fx = w(fx0) irc = 0 c 999 return c *** last card of sgrad2 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 double precision l(1), 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 data eps/0.1d+0/, one/1.d+0/ c/7 c 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 subroutine assst(iv, liv, lv, v) c c *** assess candidate step (***sol version 2.3) *** c integer liv, lv integer iv(liv) double precision v(lv) c c *** purpose *** c c this subroutine is called by an unconstrained minimization c routine to assess the next candidate step. it may recommend one c of several courses of action, such as accepting the step, recom- c puting it using the same or a new quadratic model, or halting due c to convergence or false convergence. see the return code listing c below. c c-------------------------- parameter usage -------------------------- c c iv (i/o) integer parameter and scratch vector -- see description c below of iv values referenced. c liv (in) length of iv array. c lv (in) length of v array. c v (i/o) real parameter and scratch vector -- see description c below of v values referenced. c c *** iv values referenced *** c c iv(irc) (i/o) on input for the first step tried in a new iteration, c iv(irc) should be set to 3 or 4 (the value to which it is c set when step is definitely to be accepted). on input c after step has been recomputed, iv(irc) should be c unchanged since the previous return of assst. c on output, iv(irc) is a return code having one of the c following values... c 1 = switch models or try smaller step. c 2 = switch models or accept step. c 3 = accept step and determine v(radfac) by gradient c tests. c 4 = accept step, v(radfac) has been determined. c 5 = recompute step (using the same model). c 6 = recompute step with radius = v(lmaxs) but do not c evaulate the objective function. c 7 = x-convergence (see v(xctol)). c 8 = relative function convergence (see v(rfctol)). c 9 = both x- and relative function convergence. c 10 = absolute function convergence (see v(afctol)). c 11 = singular convergence (see v(lmaxs)). c 12 = false convergence (see v(xftol)). c 13 = iv(irc) was out of range on input. c return code i has precdence over i+1 for i = 9, 10, 11. c iv(mlstgd) (i/o) saved value of iv(model). c iv(model) (i/o) on input, iv(model) should be an integer identifying c the current quadratic model of the objective function. c if a previous step yielded a better function reduction, c then iv(model) will be set to iv(mlstgd) on output. c iv(nfcall) (in) invocation count for the objective function. c iv(nfgcal) (i/o) value of iv(nfcall) at step that gave the biggest c function reduction this iteration. iv(nfgcal) remains c unchanged until a function reduction is obtained. c iv(radinc) (i/o) the number of radius increases (or minus the number c of decreases) so far this iteration. c iv(restor) (out) set to 1 if v(f) has been restored and x should be c restored to its initial value, to 2 if x should be saved, c to 3 if x should be restored from the saved value, and to c 0 otherwise. c iv(stage) (i/o) count of the number of models tried so far in the c current iteration. c iv(stglim) (in) maximum number of models to consider. c iv(switch) (out) set to 0 unless a new model is being tried and it c gives a smaller function value than the previous model, c in which case assst sets iv(switch) = 1. c iv(toobig) (in) is nonzero if step was too big (e.g. if it caused c overflow). c iv(xirc) (i/o) value that iv(irc) would have in the absence of c convergence, false convergence, and oversized steps. c c *** v values referenced *** c c v(afctol) (in) absolute function convergence tolerance. if the c absolute value of the current function value v(f) is less c than v(afctol), then assst returns with iv(irc) = 10. c v(decfac) (in) factor by which to decrease radius when iv(toobig) is c nonzero. c v(dstnrm) (in) the 2-norm of d*step. c v(dstsav) (i/o) value of v(dstnrm) on saved step. c v(dst0) (in) the 2-norm of d times the newton step (when defined, c i.e., for v(nreduc) .ge. 0). c v(f) (i/o) on both input and output, v(f) is the objective func- c tion value at x. if x is restored to a previous value, c then v(f) is restored to the corresponding value. c v(fdif) (out) the function reduction v(f0) - v(f) (for the output c value of v(f) if an earlier step gave a bigger function c decrease, and for the input value of v(f) otherwise). c v(flstgd) (i/o) saved value of v(f). c v(f0) (in) objective function value at start of iteration. c v(gtslst) (i/o) value of v(gtstep) on saved step. c v(gtstep) (in) inner product between step and gradient. c v(incfac) (in) minimum factor by which to increase radius. c v(lmaxs) (in) maximum reasonable step size (and initial step bound). c if the actual function decrease is no more than twice c what was predicted, if a return with iv(irc) = 7, 8, 9, c or 10 does not occur, if v(dstnrm) .gt. v(lmaxs), and if c v(preduc) .le. v(sctol) * abs(v(f0)), then assst re- c turns with iv(irc) = 11. if so doing appears worthwhile, c then assst repeats this test with v(preduc) computed for c a step of length v(lmaxs) (by a return with iv(irc) = 6). c v(nreduc) (i/o) function reduction predicted by quadratic model for c newton step. if assst is called with iv(irc) = 6, i.e., c if v(preduc) has been computed with radius = v(lmaxs) for c use in the singular convervence test, then v(nreduc) is c set to -v(preduc) before the latter is restored. c v(plstgd) (i/o) value of v(preduc) on saved step. c v(preduc) (i/o) function reduction predicted by quadratic model for c current step. c v(radfac) (out) factor to be used in determining the new radius, c which should be v(radfac)*dst, where dst is either the c output value of v(dstnrm) or the 2-norm of c diag(newd)*step for the output value of step and the c updated version, newd, of the scale vector d. for c iv(irc) = 3, v(radfac) = 1.0 is returned. c v(rdfcmn) (in) minimum value for v(radfac) in terms of the input c value of v(dstnrm) -- suggested value = 0.1. c v(rdfcmx) (in) maximum value for v(radfac) -- suggested value = 4.0. c v(reldx) (in) scaled relative change in x caused by step, computed c (e.g.) by function reldst as c max (d(i)*abs(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 v(rfctol) (in) relative function convergence tolerance. if the c actual function reduction is at most twice what was pre- c dicted and v(nreduc) .le. v(rfctol)*abs(v(f0)), then c assst returns with iv(irc) = 8 or 9. c v(stppar) (in) marquardt parameter -- 0 means full newton step. c v(tuner1) (in) tuning constant used to decide if the function c reduction was much less than expected. suggested c value = 0.1. c v(tuner2) (in) tuning constant used to decide if the function c reduction was large enough to accept step. suggested c value = 10**-4. c v(tuner3) (in) tuning constant used to decide if the radius c should be increased. suggested value = 0.75. c v(xctol) (in) x-convergence criterion. if step is a newton step c (v(stppar) = 0) having v(reldx) .le. v(xctol) and giving c at most twice the predicted function decrease, then c assst returns iv(irc) = 7 or 9. c v(xftol) (in) false convergence tolerance. if step gave no or only c a small function decrease and v(reldx) .le. v(xftol), c then assst returns with iv(irc) = 12. c c------------------------------- notes ------------------------------- c c *** application and usage restrictions *** c c this routine is called as part of the nl2sol (nonlinear c least-squares) package. it may be used in any unconstrained c minimization solver that uses dogleg, goldfeld-quandt-trotter, c or levenberg-marquardt steps. c c *** algorithm notes *** c c see (1) for further discussion of the assessing and model c switching strategies. while nl2sol considers only two models, c assst is designed to handle any number of models. c c *** usage notes *** c c on the first call of an iteration, only the i/o variables c step, x, iv(irc), iv(model), v(f), v(dstnrm), v(gtstep), and c v(preduc) need have been initialized. between calls, no i/o c values execpt step, x, iv(model), v(f) and the stopping toler- c ances should be changed. c after a return for convergence or false convergence, one can c change the stopping tolerances and call assst again, in which c case the stopping tests will be repeated. c c *** references *** c c (1) dennis, j.e., jr., gay, d.m., and welsch, r.e. (1981), c an adaptive nonlinear least-squares algorithm, c acm trans. math. software, vol. 7, no. 3. c c (2) powell, m.j.d. (1970) a fortran subroutine for solving c systems of nonlinear algebraic equations, in numerical c methods for nonlinear algebraic equations, edited by c p. rabinowitz, gordon and breach, london. c c *** history *** c c john dennis designed much of this routine, starting with c ideas in (2). roy welsch suggested the model switching strategy. c david gay and stephen peters cast this subroutine into a more c portable form (winter 1977), and david gay cast it into its c present form (fall 1978). c c *** general *** c c this subroutine was written in connection with research c supported by the national science foundation under grants c mcs-7600324, dcr75-10143, 76-14311dss, mcs76-11989, and c mcs-7906671. c c------------------------ external quantities ------------------------ c c *** no external functions and subroutines *** c c *** intrinsic functions *** c/+ double precision dabs, dmax1 c/ c *** no common blocks *** c c-------------------------- local variables -------------------------- c logical goodx integer i, nfc double precision emax, emaxs, gts, rfac1, xmax double precision half, one, onep2, two, zero c c *** subscripts for iv and v *** c integer afctol, decfac, dstnrm, dstsav, dst0, f, fdif, flstgd, f0, 1 gtslst, gtstep, incfac, irc, lmaxs, mlstgd, model, nfcall, 2 nfgcal, nreduc, plstgd, preduc, radfac, radinc, rdfcmn, 3 rdfcmx, reldx, restor, rfctol, sctol, stage, stglim, 4 stppar, switch, toobig, tuner1, tuner2, tuner3, xctol, 5 xftol, xirc c c *** data initializations *** c c/6 data half/0.5d+0/, one/1.d+0/, onep2/1.2d+0/, two/2.d+0/, 1 zero/0.d+0/ c/7 c parameter (half=0.5d+0, one=1.d+0, onep2=1.2d+0, two=2.d+0, c 1 zero=0.d+0) c/ c c/6 data irc/29/, mlstgd/32/, model/5/, nfcall/6/, nfgcal/7/, 1 radinc/8/, restor/9/, stage/10/, stglim/11/, switch/12/, 2 toobig/2/, xirc/13/ c/7 c parameter (irc=29, mlstgd=32, model=5, nfcall=6, nfgcal=7, c 1 radinc=8, restor=9, stage=10, stglim=11, switch=12, c 2 toobig=2, xirc=13) c/ c/6 data afctol/31/, decfac/22/, dstnrm/2/, dst0/3/, dstsav/18/, 1 f/10/, fdif/11/, flstgd/12/, f0/13/, gtslst/14/, gtstep/4/, 2 incfac/23/, lmaxs/36/, nreduc/6/, plstgd/15/, preduc/7/, 3 radfac/16/, rdfcmn/24/, rdfcmx/25/, reldx/17/, rfctol/32/, 4 sctol/37/, stppar/5/, tuner1/26/, tuner2/27/, tuner3/28/, 5 xctol/33/, xftol/34/ c/7 c parameter (afctol=31, decfac=22, dstnrm=2, dst0=3, dstsav=18, c 1 f=10, fdif=11, flstgd=12, f0=13, gtslst=14, gtstep=4, c 2 incfac=23, lmaxs=36, nreduc=6, plstgd=15, preduc=7, c 3 radfac=16, rdfcmn=24, rdfcmx=25, reldx=17, rfctol=32, c 4 sctol=37, stppar=5, tuner1=26, tuner2=27, tuner3=28, c 5 xctol=33, xftol=34) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c nfc = iv(nfcall) iv(switch) = 0 iv(restor) = 0 rfac1 = one goodx = .true. i = iv(irc) if (i .ge. 1 .and. i .le. 12) 1 go to (20,30,10,10,40,280,220,220,220,220,220,170), i iv(irc) = 13 go to 999 c c *** initialize for new iteration *** c 10 iv(stage) = 1 iv(radinc) = 0 v(flstgd) = v(f0) if (iv(toobig) .eq. 0) go to 110 iv(stage) = -1 iv(xirc) = i go to 60 c c *** step was recomputed with new model or smaller radius *** c *** first decide which *** c 20 if (iv(model) .ne. iv(mlstgd)) go to 30 c *** old model retained, smaller radius tried *** c *** do not consider any more new models this iteration *** iv(stage) = iv(stglim) iv(radinc) = -1 go to 110 c c *** a new model is being tried. decide whether to keep it. *** c 30 iv(stage) = iv(stage) + 1 c c *** now we add the possibiltiy that step was recomputed with *** c *** the same model, perhaps because of an oversized step. *** c 40 if (iv(stage) .gt. 0) go to 50 c c *** step was recomputed because it was too big. *** c if (iv(toobig) .ne. 0) go to 60 c c *** restore iv(stage) and pick up where we left off. *** c iv(stage) = -iv(stage) i = iv(xirc) go to (20, 30, 110, 110, 70), i c 50 if (iv(toobig) .eq. 0) go to 70 c c *** handle oversize step *** c if (iv(radinc) .gt. 0) go to 80 iv(stage) = -iv(stage) iv(xirc) = iv(irc) c 60 v(radfac) = v(decfac) iv(radinc) = iv(radinc) - 1 iv(irc) = 5 iv(restor) = 1 go to 999 c 70 if (v(f) .lt. v(flstgd)) go to 110 c c *** the new step is a loser. restore old model. *** c if (iv(model) .eq. iv(mlstgd)) go to 80 iv(model) = iv(mlstgd) iv(switch) = 1 c c *** restore step, etc. only if a previous step decreased v(f). c 80 if (v(flstgd) .ge. v(f0)) go to 110 iv(restor) = 1 v(f) = v(flstgd) v(preduc) = v(plstgd) v(gtstep) = v(gtslst) if (iv(switch) .eq. 0) rfac1 = v(dstnrm) / v(dstsav) v(dstnrm) = v(dstsav) nfc = iv(nfgcal) goodx = .false. c 110 v(fdif) = v(f0) - v(f) if (v(fdif) .gt. v(tuner2) * v(preduc)) go to 140 if(iv(radinc).gt.0) go to 140 c c *** no (or only a trivial) function decrease c *** -- so try new model or smaller radius c if (v(f) .lt. v(f0)) go to 120 iv(mlstgd) = iv(model) v(flstgd) = v(f) v(f) = v(f0) iv(restor) = 1 go to 130 120 iv(nfgcal) = nfc 130 iv(irc) = 1 if (iv(stage) .lt. iv(stglim)) go to 160 iv(irc) = 5 iv(radinc) = iv(radinc) - 1 go to 160 c c *** nontrivial function decrease achieved *** c 140 iv(nfgcal) = nfc rfac1 = one v(dstsav) = v(dstnrm) if (v(fdif) .gt. v(preduc)*v(tuner1)) go to 190 c c *** decrease was much less than predicted -- either change models c *** or accept step with decreased radius. c if (iv(stage) .ge. iv(stglim)) go to 150 c *** consider switching models *** iv(irc) = 2 go to 160 c c *** accept step with decreased radius *** c 150 iv(irc) = 4 c c *** set v(radfac) to fletcher*s decrease factor *** c 160 iv(xirc) = iv(irc) emax = v(gtstep) + v(fdif) v(radfac) = half * rfac1 if (emax .lt. v(gtstep)) v(radfac) = rfac1 * dmax1(v(rdfcmn), 1 half * v(gtstep)/emax) c c *** do false convergence test *** c 170 if (v(reldx) .le. v(xftol)) go to 180 iv(irc) = iv(xirc) if (v(f) .lt. v(f0)) go to 200 go to 230 c 180 iv(irc) = 12 go to 240 c c *** handle good function decrease *** c 190 if (v(fdif) .lt. (-v(tuner3) * v(gtstep))) go to 210 c c *** increasing radius looks worthwhile. see if we just c *** recomputed step with a decreased radius or restored step c *** after recomputing it with a larger radius. c if (iv(radinc) .lt. 0) go to 210 if (iv(restor) .eq. 1) go to 210 c c *** we did not. try a longer step unless this was a newton c *** step. c v(radfac) = v(rdfcmx) gts = v(gtstep) if (v(fdif) .lt. (half/v(radfac) - one) * gts) 1 v(radfac) = dmax1(v(incfac), half*gts/(gts + v(fdif))) iv(irc) = 4 if (v(stppar) .eq. zero) go to 230 if (v(dst0) .ge. zero .and. (v(dst0) .lt. two*v(dstnrm) 1 .or. v(nreduc) .lt. onep2*v(fdif))) go to 230 c *** step was not a newton step. recompute it with c *** a larger radius. iv(irc) = 5 iv(radinc) = iv(radinc) + 1 c c *** save values corresponding to good step *** c 200 v(flstgd) = v(f) iv(mlstgd) = iv(model) if (iv(restor) .ne. 1) iv(restor) = 2 v(dstsav) = v(dstnrm) iv(nfgcal) = nfc v(plstgd) = v(preduc) v(gtslst) = v(gtstep) go to 230 c c *** accept step with radius unchanged *** c 210 v(radfac) = one iv(irc) = 3 go to 230 c c *** come here for a restart after convergence *** c 220 iv(irc) = iv(xirc) if (v(dstsav) .ge. zero) go to 240 iv(irc) = 12 go to 240 c c *** perform convergence tests *** c 230 iv(xirc) = iv(irc) 240 if (iv(restor) .eq. 1 .and. v(flstgd) .lt. v(f0)) iv(restor) = 3 if (dabs(v(f)) .lt. v(afctol)) iv(irc) = 10 if (half * v(fdif) .gt. v(preduc)) go to 999 emax = v(rfctol) * dabs(v(f0)) emaxs = v(sctol) * dabs(v(f0)) if (v(dstnrm) .gt. v(lmaxs) .and. v(preduc) .le. emaxs) 1 iv(irc) = 11 if (v(dst0) .lt. zero) go to 250 i = 0 if ((v(nreduc) .gt. zero .and. v(nreduc) .le. emax) .or. 1 (v(nreduc) .eq. zero. and. v(preduc) .eq. zero)) i = 2 if (v(stppar) .eq. zero .and. v(reldx) .le. v(xctol) 1 .and. goodx) i = i + 1 if (i .gt. 0) iv(irc) = i + 6 c c *** consider recomputing step of length v(lmaxs) for singular c *** convergence test. c 250 if (iv(irc) .gt. 5 .and. iv(irc) .ne. 12) go to 999 if (v(dstnrm) .gt. v(lmaxs)) go to 260 if (v(preduc) .ge. emaxs) go to 999 if (v(dst0) .le. zero) go to 270 if (half * v(dst0) .le. v(lmaxs)) go to 999 go to 270 260 if (half * v(dstnrm) .le. v(lmaxs)) go to 999 xmax = v(lmaxs) / v(dstnrm) if (xmax * (two - xmax) * v(preduc) .ge. emaxs) go to 999 270 if (v(nreduc) .lt. zero) go to 290 c c *** recompute v(preduc) for use in singular convergence test *** c v(gtslst) = v(gtstep) v(dstsav) = v(dstnrm) if (iv(irc) .eq. 12) v(dstsav) = -v(dstsav) v(plstgd) = v(preduc) i = iv(restor) iv(restor) = 2 if (i .eq. 3) iv(restor) = 0 iv(irc) = 6 go to 999 c c *** perform singular convergence test with recomputed v(preduc) *** c 280 v(gtstep) = v(gtslst) v(dstnrm) = dabs(v(dstsav)) iv(irc) = iv(xirc) if (v(dstsav) .le. zero) iv(irc) = 12 v(nreduc) = -v(preduc) v(preduc) = v(plstgd) iv(restor) = 3 290 if (-v(nreduc) .le. v(rfctol) * dabs(v(f0))) iv(irc) = 11 c 999 return c c *** last card of assst follows *** end subroutine deflt(alg, iv, liv, lv, v) c c *** supply ***sol (version 2.3) default values to iv and v *** c c *** alg = 1 means regression constants. c *** alg = 2 means general unconstrained optimization constants. c integer liv, lv integer alg, iv(liv) double precision v(lv) c external imdcon, vdflt integer imdcon c imdcon... returns machine-dependent integer constants. c vdflt.... provides default values to v. c integer miv, mv integer miniv(2), minv(2) c c *** subscripts for iv *** c integer algsav, covprt, covreq, dtype, hc, ierr, inith, inits, 1 ipivot, ivneed, lastiv, lastv, lmat, mxfcal, mxiter, 2 nfcov, ngcov, nvdflt, outlev, parprt, parsav, perm, 3 prunit, qrtyp, rdreq, rmat, solprt, statpr, vneed, 4 vsave, x0prt c c *** iv subscript values *** c c/6 data algsav/51/, covprt/14/, covreq/15/, dtype/16/, hc/71/, 1 ierr/75/, inith/25/, inits/25/, ipivot/76/, ivneed/3/, 2 lastiv/44/, lastv/45/, lmat/42/, mxfcal/17/, mxiter/18/, 3 nfcov/52/, ngcov/53/, nvdflt/50/, outlev/19/, parprt/20/, 4 parsav/49/, perm/58/, prunit/21/, qrtyp/80/, rdreq/57/, 5 rmat/78/, solprt/22/, statpr/23/, vneed/4/, vsave/60/, 6 x0prt/24/ c/7 c parameter (algsav=51, covprt=14, covreq=15, dtype=16, hc=71, c 1 ierr=75, inith=25, inits=25, ipivot=76, ivneed=3, c 2 lastiv=44, lastv=45, lmat=42, mxfcal=17, mxiter=18, c 3 nfcov=52, ngcov=53, nvdflt=50, outlev=19, parprt=20, c 4 parsav=49, perm=58, prunit=21, qrtyp=80, rdreq=57, c 5 rmat=78, solprt=22, statpr=23, vneed=4, vsave=60, c 6 x0prt=24) c/ data miniv(1)/80/, miniv(2)/59/, minv(1)/98/, minv(2)/71/ c c------------------------------- body -------------------------------- c if (alg .lt. 1 .or. alg .gt. 2) go to 40 miv = miniv(alg) if (liv .lt. miv) go to 20 mv = minv(alg) if (lv .lt. mv) go to 30 call vdflt(alg, lv, v) iv(1) = 12 iv(algsav) = alg iv(ivneed) = 0 iv(lastiv) = miv iv(lastv) = mv iv(lmat) = mv + 1 iv(mxfcal) = 200 iv(mxiter) = 150 iv(outlev) = 1 iv(parprt) = 1 iv(perm) = miv + 1 iv(prunit) = imdcon(1) iv(solprt) = 1 iv(statpr) = 1 iv(vneed) = 0 iv(x0prt) = 1 c if (alg .ge. 2) go to 10 c c *** regression values c iv(covprt) = 3 iv(covreq) = 1 iv(dtype) = 1 iv(hc) = 0 iv(ierr) = 0 iv(inits) = 0 iv(ipivot) = 0 iv(nvdflt) = 32 iv(parsav) = 67 iv(qrtyp) = 1 iv(rdreq) = 3 iv(rmat) = 0 iv(vsave) = 58 go to 999 c c *** general optimization values c 10 iv(dtype) = 0 iv(inith) = 1 iv(nfcov) = 0 iv(ngcov) = 0 iv(nvdflt) = 25 iv(parsav) = 47 go to 999 c 20 iv(1) = 15 go to 999 c 30 iv(1) = 16 go to 999 c 40 iv(1) = 67 c 999 return c *** last card of deflt follows *** end double precision function dotprd(p, x, y) c c *** return the inner product of the p-vectors x and y. *** c integer p double precision x(p), y(p) c integer i double precision one, sqteta, t, zero c/+ double precision dmax1, dabs c/ external rmdcon double precision rmdcon c c *** rmdcon(2) returns a machine-dependent constant, sqteta, which c *** is slightly larger than the smallest positive number that c *** can be squared without underflowing. c c/6 data one/1.d+0/, sqteta/0.d+0/, zero/0.d+0/ c/7 c parameter (one=1.d+0, zero=0.d+0) c data sqteta/0.d+0/ c/ c dotprd = zero if (p .le. 0) go to 999 if (sqteta .eq. zero) sqteta = rmdcon(2) do 20 i = 1, p t = dmax1(dabs(x(i)), dabs(y(i))) if (t .gt. one) go to 10 if (t .lt. sqteta) go to 20 t = (x(i)/sqteta)*y(i) if (dabs(t) .lt. sqteta) go to 20 10 dotprd = dotprd + x(i)*y(i) 20 continue c 999 return c *** last card of dotprd follows *** end subroutine itsum(d, g, iv, liv, lv, p, v, x) c c *** print iteration summary for ***sol (version 2.3) *** c c *** parameter declarations *** c integer liv, lv, p integer iv(liv) double precision d(p), g(p), v(lv), x(p) c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** local variables *** c integer alg, i, iv1, m, nf, ng, ol, pu c/6 real model1(6), model2(6) c/7 c character*4 model1(6), model2(6) c/ double precision nreldf, oldf, preldf, reldf, zero c c *** intrinsic functions *** c/+ integer iabs double precision dabs, dmax1 c/ c *** no external functions or subroutines *** c c *** subscripts for iv and v *** c integer algsav, dstnrm, f, fdif, f0, needhd, nfcall, nfcov, ngcov, 1 ngcall, niter, nreduc, outlev, preduc, prntit, prunit, 2 reldx, solprt, statpr, stppar, sused, x0prt c c *** iv subscript values *** c c/6 data algsav/51/, needhd/36/, nfcall/6/, nfcov/52/, ngcall/30/, 1 ngcov/53/, niter/31/, outlev/19/, prntit/39/, prunit/21/, 2 solprt/22/, statpr/23/, sused/64/, x0prt/24/ c/7 c parameter (algsav=51, needhd=36, nfcall=6, nfcov=52, ngcall=30, c 1 ngcov=53, niter=31, outlev=19, prntit=39, prunit=21, c 2 solprt=22, statpr=23, sused=64, x0prt=24) c/ c c *** v subscript values *** c c/6 data dstnrm/2/, f/10/, f0/13/, fdif/11/, nreduc/6/, preduc/7/, 1 reldx/17/, stppar/5/ c/7 c parameter (dstnrm=2, f=10, f0=13, fdif=11, nreduc=6, preduc=7, c 1 reldx=17, stppar=5) c/ c c/6 data zero/0.d+0/ c/7 c parameter (zero=0.d+0) c/ c/6 data model1(1)/4h /, model1(2)/4h /, model1(3)/4h /, 1 model1(4)/4h /, model1(5)/4h g /, model1(6)/4h s /, 2 model2(1)/4h g /, model2(2)/4h s /, model2(3)/4hg-s /, 3 model2(4)/4hs-g /, model2(5)/4h-s-g/, model2(6)/4h-g-s/ c/7 c data model1/' ',' ',' ',' ',' g ',' s '/, c 1 model2/' g ',' s ','g-s ','s-g ','-s-g','-g-s'/ c/ c c------------------------------- body -------------------------------- c pu = iv(prunit) if (pu .eq. 0) go to 999 iv1 = iv(1) if (iv1 .gt. 62) iv1 = iv1 - 51 ol = iv(outlev) alg = iv(algsav) if (iv1 .lt. 2 .or. iv1 .gt. 15) go to 370 if (iv1 .ge. 12) go to 120 if (iv1 .eq. 2 .and. iv(niter) .eq. 0) go to 390 if (ol .eq. 0) go to 120 if (iv1 .ge. 10 .and. iv(prntit) .eq. 0) go to 120 if (iv1 .gt. 2) go to 10 iv(prntit) = iv(prntit) + 1 if (iv(prntit) .lt. iabs(ol)) go to 999 10 nf = iv(nfcall) - iabs(iv(nfcov)) iv(prntit) = 0 reldf = zero preldf = zero oldf = dmax1(dabs(v(f0)), dabs(v(f))) if (oldf .le. zero) go to 20 reldf = v(fdif) / oldf preldf = v(preduc) / oldf 20 if (ol .gt. 0) go to 60 c c *** print short summary line *** c if (iv(needhd) .eq. 1 .and. alg .eq. 1) write(pu,30) 30 format(/10h it nf,6x,1hf,7x,5hreldf,3x,6hpreldf,3x,5hreldx, 1 2x,13hmodel stppar) if (iv(needhd) .eq. 1 .and. alg .eq. 2) write(pu,40) 40 format(/11h it nf,7x,1hf,8x,5hreldf,4x,6hpreldf,4x,5hreldx, 1 3x,6hstppar) iv(needhd) = 0 if (alg .eq. 2) go to 50 m = iv(sused) write(pu,100) iv(niter), nf, v(f), reldf, preldf, v(reldx), 1 model1(m), model2(m), v(stppar) go to 120 c 50 write(pu,110) iv(niter), nf, v(f), reldf, preldf, v(reldx), 1 v(stppar) go to 120 c c *** print long summary line *** c 60 if (iv(needhd) .eq. 1 .and. alg .eq. 1) write(pu,70) 70 format(/11h it nf,6x,1hf,7x,5hreldf,3x,6hpreldf,3x,5hreldx, 1 2x,13hmodel stppar,2x,6hd*step,2x,7hnpreldf) if (iv(needhd) .eq. 1 .and. alg .eq. 2) write(pu,80) 80 format(/11h it nf,7x,1hf,8x,5hreldf,4x,6hpreldf,4x,5hreldx, 1 3x,6hstppar,3x,6hd*step,3x,7hnpreldf) iv(needhd) = 0 nreldf = zero if (oldf .gt. zero) nreldf = v(nreduc) / oldf if (alg .eq. 2) go to 90 m = iv(sused) write(pu,100) iv(niter), nf, v(f), reldf, preldf, v(reldx), 1 model1(m), model2(m), v(stppar), v(dstnrm), nreldf go to 120 c 90 write(pu,110) iv(niter), nf, v(f), reldf, preldf, 1 v(reldx), v(stppar), v(dstnrm), nreldf 100 format(i6,i5,d10.3,2d9.2,d8.1,a3,a4,2d8.1,d9.2) 110 format(i6,i5,d11.3,2d10.2,3d9.1,d10.2) c 120 if (iv(statpr) .lt. 0) go to 430 go to (999, 999, 130, 150, 170, 190, 210, 230, 250, 270, 290, 310, 1 330, 350, 520), iv1 c 130 write(pu,140) 140 format(/26h ***** x-convergence *****) go to 430 c 150 write(pu,160) 160 format(/42h ***** relative function convergence *****) go to 430 c 170 write(pu,180) 180 format(/49h ***** x- and relative function convergence *****) go to 430 c 190 write(pu,200) 200 format(/42h ***** absolute function convergence *****) go to 430 c 210 write(pu,220) 220 format(/33h ***** singular convergence *****) go to 430 c 230 write(pu,240) 240 format(/30h ***** false convergence *****) go to 430 c 250 write(pu,260) 260 format(/38h ***** function evaluation limit *****) go to 430 c 270 write(pu,280) 280 format(/28h ***** iteration limit *****) go to 430 c 290 write(pu,300) 300 format(/18h ***** stopx *****) go to 430 c 310 write(pu,320) 320 format(/44h ***** initial f(x) cannot be computed *****) c go to 390 c 330 write(pu,340) 340 format(/37h ***** bad parameters to assess *****) go to 999 c 350 write(pu,360) 360 format(/43h ***** gradient could not be computed *****) if (iv(niter) .gt. 0) go to 480 go to 390 c 370 write(pu,380) iv(1) 380 format(/14h ***** iv(1) =,i5,6h *****) go to 999 c c *** initial call on itsum *** c 390 if (iv(x0prt) .ne. 0) write(pu,400) (i, x(i), d(i), i = 1, p) 400 format(/23h i initial x(i),8x,4hd(i)//(1x,i5,d17.6,d14.3)) c *** the following are to avoid undefined variables when the c *** function evaluation limit is 1... v(dstnrm) = zero v(fdif) = zero v(nreduc) = zero v(preduc) = zero v(reldx) = zero if (iv1 .ge. 12) go to 999 iv(needhd) = 0 iv(prntit) = 0 if (ol .eq. 0) go to 999 if (ol .lt. 0 .and. alg .eq. 1) write(pu,30) if (ol .lt. 0 .and. alg .eq. 2) write(pu,40) if (ol .gt. 0 .and. alg .eq. 1) write(pu,70) if (ol .gt. 0 .and. alg .eq. 2) write(pu,80) if (alg .eq. 1) write(pu,410) v(f) if (alg .eq. 2) write(pu,420) v(f) 410 format(/11h 0 1,d10.3) c365 format(/11h 0 1,e11.3) 420 format(/11h 0 1,d11.3) go to 999 c c *** print various information requested on solution *** c 430 iv(needhd) = 1 if (iv(statpr) .eq. 0) go to 480 oldf = dmax1(dabs(v(f0)), dabs(v(f))) preldf = zero nreldf = zero if (oldf .le. zero) go to 440 preldf = v(preduc) / oldf nreldf = v(nreduc) / oldf 440 nf = iv(nfcall) - iv(nfcov) ng = iv(ngcall) - iv(ngcov) write(pu,450) v(f), v(reldx), nf, ng, preldf, nreldf 450 format(/9h function,d17.6,8h reldx,d17.3/12h func. evals, 1 i8,9x,11hgrad. evals,i8/7h preldf,d16.3,6x,7hnpreldf,d15.3) c if (iv(nfcov) .gt. 0) write(pu,460) iv(nfcov) 460 format(/1x,i4,50h extra func. evals for covariance and diagnost 1ics.) if (iv(ngcov) .gt. 0) write(pu,470) iv(ngcov) 470 format(1x,i4,50h extra grad. evals for covariance and diagnosti 1cs.) c 480 if (iv(solprt) .eq. 0) go to 999 iv(needhd) = 1 write(pu,490) 490 format(/22h i final x(i),8x,4hd(i),10x,4hg(i)/) do 500 i = 1, p write(pu,510) i, x(i), d(i), g(i) 500 continue 510 format(1x,i5,d16.6,2d14.3) go to 999 c 520 write(pu,530) 530 format(/24h inconsistent dimensions) 999 return c *** last card of itsum follows *** end subroutine litvmu(n, x, l, y) c c *** solve (l**t)*x = 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 double precision x(n), l(1), y(n) integer i, ii, ij, im1, i0, j, np1 double precision xi, zero c/6 data zero/0.d+0/ c/7 c parameter (zero=0.d+0) c/ c do 10 i = 1, n 10 x(i) = y(i) np1 = n + 1 i0 = n*(n+1)/2 do 30 ii = 1, n i = np1 - ii xi = x(i)/l(i0) x(i) = xi if (i .le. 1) go to 999 i0 = i0 - i if (xi .eq. zero) go to 30 im1 = i - 1 do 20 j = 1, im1 ij = i0 + j x(j) = x(j) - xi*l(ij) 20 continue 30 continue 999 return c *** last card of litvmu follows *** end subroutine livmul(n, x, l, y) c c *** solve l*x = 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 double precision x(n), l(1), y(n) external dotprd double precision dotprd integer i, j, k double precision t, zero c/6 data zero/0.d+0/ c/7 c parameter (zero=0.d+0) c/ c do 10 k = 1, n if (y(k) .ne. zero) go to 20 x(k) = zero 10 continue go to 999 20 j = k*(k+1)/2 x(k) = y(k) / l(j) if (k .ge. n) go to 999 k = k + 1 do 30 i = k, n t = dotprd(i-1, l(j+1), x) j = j + i x(i) = (y(i) - t)/l(j) 30 continue 999 return c *** last card of livmul follows *** end subroutine parck(alg, d, iv, liv, lv, n, v) c c *** check ***sol (version 2.3) parameters, print changed values *** c c *** alg = 1 for regression, alg = 2 for general unconstrained opt. c integer alg, liv, lv, n integer iv(liv) double precision d(n), v(lv) c external rmdcon, vcopy, vdflt double precision rmdcon c rmdcon -- returns machine-dependent constants. c vcopy -- copies one vector to another. c vdflt -- supplies default parameter values to v alone. c/+ integer max0 c/ c c *** local variables *** c integer i, ii, iv1, j, k, l, m, miv1, miv2, ndfalt, parsv1, pu integer ijmp, jlim(2), miniv(2), ndflt(2) c/6 integer varnm(2), sh(2) real cngd(3), dflt(3), vn(2,34), which(3) c/7 c character*1 varnm(2), sh(2) c character*4 cngd(3), dflt(3), vn(2,34), which(3) c/ double precision big, machep, tiny, vk, vm(34), vx(34), zero c c *** iv and v subscripts *** c integer algsav, dinit, dtype, dtype0, epslon, inits, ivneed, 1 lastiv, lastv, lmat, nextiv, nextv, nvdflt, oldn, 2 parprt, parsav, perm, prunit, vneed c c c/6 data algsav/51/, dinit/38/, dtype/16/, dtype0/54/, epslon/19/, 1 inits/25/, ivneed/3/, lastiv/44/, lastv/45/, lmat/42/, 2 nextiv/46/, nextv/47/, nvdflt/50/, oldn/38/, parprt/20/, 3 parsav/49/, perm/58/, prunit/21/, vneed/4/ c/7 c parameter (algsav=51, dinit=38, dtype=16, dtype0=54, epslon=19, c 1 inits=25, ivneed=3, lastiv=44, lastv=45, lmat=42, c 2 nextiv=46, nextv=47, nvdflt=50, oldn=38, parprt=20, c 3 parsav=49, perm=58, prunit=21, vneed=4) c save big, machep, tiny c/ c data big/0.d+0/, machep/-1.d+0/, tiny/1.d+0/, zero/0.d+0/ c/6 data vn(1,1),vn(2,1)/4hepsl,4hon../ data vn(1,2),vn(2,2)/4hphmn,4hfc../ data vn(1,3),vn(2,3)/4hphmx,4hfc../ data vn(1,4),vn(2,4)/4hdecf,4hac../ data vn(1,5),vn(2,5)/4hincf,4hac../ data vn(1,6),vn(2,6)/4hrdfc,4hmn../ data vn(1,7),vn(2,7)/4hrdfc,4hmx../ data vn(1,8),vn(2,8)/4htune,4hr1../ data vn(1,9),vn(2,9)/4htune,4hr2../ data vn(1,10),vn(2,10)/4htune,4hr3../ data vn(1,11),vn(2,11)/4htune,4hr4../ data vn(1,12),vn(2,12)/4htune,4hr5../ data vn(1,13),vn(2,13)/4hafct,4hol../ data vn(1,14),vn(2,14)/4hrfct,4hol../ data vn(1,15),vn(2,15)/4hxcto,4hl.../ data vn(1,16),vn(2,16)/4hxfto,4hl.../ data vn(1,17),vn(2,17)/4hlmax,4h0.../ data vn(1,18),vn(2,18)/4hlmax,4hs.../ data vn(1,19),vn(2,19)/4hscto,4hl.../ data vn(1,20),vn(2,20)/4hdini,4ht.../ data vn(1,21),vn(2,21)/4hdtin,4hit../ data vn(1,22),vn(2,22)/4hd0in,4hit../ data vn(1,23),vn(2,23)/4hdfac,4h..../ data vn(1,24),vn(2,24)/4hdltf,4hdc../ data vn(1,25),vn(2,25)/4hdltf,4hdj../ data vn(1,26),vn(2,26)/4hdelt,4ha0../ data vn(1,27),vn(2,27)/4hfuzz,4h..../ data vn(1,28),vn(2,28)/4hrlim,4hit../ data vn(1,29),vn(2,29)/4hcosm,4hin../ data vn(1,30),vn(2,30)/4hhube,4hrc../ data vn(1,31),vn(2,31)/4hrspt,4hol../ data vn(1,32),vn(2,32)/4hsigm,4hin../ data vn(1,33),vn(2,33)/4heta0,4h..../ data vn(1,34),vn(2,34)/4hbias,4h..../ c/7 c data vn(1,1),vn(2,1)/'epsl','on..'/ c data vn(1,2),vn(2,2)/'phmn','fc..'/ c data vn(1,3),vn(2,3)/'phmx','fc..'/ c data vn(1,4),vn(2,4)/'decf','ac..'/ c data vn(1,5),vn(2,5)/'incf','ac..'/ c data vn(1,6),vn(2,6)/'rdfc','mn..'/ c data vn(1,7),vn(2,7)/'rdfc','mx..'/ c data vn(1,8),vn(2,8)/'tune','r1..'/ c data vn(1,9),vn(2,9)/'tune','r2..'/ c data vn(1,10),vn(2,10)/'tune','r3..'/ c data vn(1,11),vn(2,11)/'tune','r4..'/ c data vn(1,12),vn(2,12)/'tune','r5..'/ c data vn(1,13),vn(2,13)/'afct','ol..'/ c data vn(1,14),vn(2,14)/'rfct','ol..'/ c data vn(1,15),vn(2,15)/'xcto','l...'/ c data vn(1,16),vn(2,16)/'xfto','l...'/ c data vn(1,17),vn(2,17)/'lmax','0...'/ c data vn(1,18),vn(2,18)/'lmax','s...'/ c data vn(1,19),vn(2,19)/'scto','l...'/ c data vn(1,20),vn(2,20)/'dini','t...'/ c data vn(1,21),vn(2,21)/'dtin','it..'/ c data vn(1,22),vn(2,22)/'d0in','it..'/ c data vn(1,23),vn(2,23)/'dfac','....'/ c data vn(1,24),vn(2,24)/'dltf','dc..'/ c data vn(1,25),vn(2,25)/'dltf','dj..'/ c data vn(1,26),vn(2,26)/'delt','a0..'/ c data vn(1,27),vn(2,27)/'fuzz','....'/ c data vn(1,28),vn(2,28)/'rlim','it..'/ c data vn(1,29),vn(2,29)/'cosm','in..'/ c data vn(1,30),vn(2,30)/'hube','rc..'/ c data vn(1,31),vn(2,31)/'rspt','ol..'/ c data vn(1,32),vn(2,32)/'sigm','in..'/ c data vn(1,33),vn(2,33)/'eta0','....'/ c data vn(1,34),vn(2,34)/'bias','....'/ c/ c data vm(1)/1.0d-3/, vm(2)/-0.99d+0/, vm(3)/1.0d-3/, vm(4)/1.0d-2/, 1 vm(5)/1.2d+0/, vm(6)/1.d-2/, vm(7)/1.2d+0/, vm(8)/0.d+0/, 2 vm(9)/0.d+0/, vm(10)/1.d-3/, vm(11)/-1.d+0/, vm(13)/0.d+0/, 3 vm(15)/0.d+0/, vm(16)/0.d+0/, vm(19)/0.d+0/, vm(20)/-10.d+0/, 4 vm(21)/0.d+0/, vm(22)/0.d+0/, vm(23)/0.d+0/, vm(27)/1.01d+0/, 5 vm(28)/1.d+10/, vm(30)/0.d+0/, vm(31)/0.d+0/, vm(32)/0.d+0/, 6 vm(34)/0.d+0/ data vx(1)/0.9d+0/, vx(2)/-1.d-3/, vx(3)/1.d+1/, vx(4)/0.8d+0/, 1 vx(5)/1.d+2/, vx(6)/0.8d+0/, vx(7)/1.d+2/, vx(8)/0.5d+0/, 2 vx(9)/0.5d+0/, vx(10)/1.d+0/, vx(11)/1.d+0/, vx(14)/0.1d+0/, 3 vx(15)/1.d+0/, vx(16)/1.d+0/, vx(19)/1.d+0/, vx(23)/1.d+0/, 4 vx(24)/1.d+0/, vx(25)/1.d+0/, vx(26)/1.d+0/, vx(27)/1.d+10/, 5 vx(29)/1.d+0/, vx(31)/1.d+0/, vx(32)/1.d+0/, vx(33)/1.d+0/, 6 vx(34)/1.d+0/ c c/6 data varnm(1)/1hp/, varnm(2)/1hn/, sh(1)/1hs/, sh(2)/1hh/ data cngd(1),cngd(2),cngd(3)/4h---c,4hhang,4hed v/, 1 dflt(1),dflt(2),dflt(3)/4hnond,4hefau,4hlt v/ c/7 c data varnm(1)/'p'/, varnm(2)/'n'/, sh(1)/'s'/, sh(2)/'h'/ c data cngd(1),cngd(2),cngd(3)/'---c','hang','ed v'/, c 1 dflt(1),dflt(2),dflt(3)/'nond','efau','lt v'/ c/ data ijmp/33/, jlim(1)/0/, jlim(2)/24/, ndflt(1)/32/, ndflt(2)/25/ data miniv(1)/80/, miniv(2)/59/ c c............................... body ................................ c pu = 0 if (prunit .le. liv) pu = iv(prunit) if (alg .lt. 1 .or. alg .gt. 2) go to 340 if (iv(1) .eq. 0) call deflt(alg, iv, liv, lv, v) iv1 = iv(1) if (iv1 .ne. 13 .and. iv1 .ne. 12) go to 10 miv1 = miniv(alg) if (perm .le. liv) miv1 = max0(miv1, iv(perm) - 1) if (ivneed .le. liv) miv2 = miv1 + max0(iv(ivneed), 0) if (lastiv .le. liv) iv(lastiv) = miv2 if (liv .lt. miv1) go to 300 iv(ivneed) = 0 iv(lastv) = max0(iv(vneed), 0) + iv(lmat) - 1 iv(vneed) = 0 if (liv .lt. miv2) go to 300 if (lv .lt. iv(lastv)) go to 320 10 if (alg .eq. iv(algsav)) go to 30 if (pu .ne. 0) write(pu,20) alg, iv(algsav) 20 format(/39h the first parameter to deflt should be,i3, 1 12h rather than,i3) iv(1) = 82 go to 999 30 if (iv1 .lt. 12 .or. iv1 .gt. 14) go to 60 if (n .ge. 1) go to 50 iv(1) = 81 if (pu .eq. 0) go to 999 write(pu,40) varnm(alg), n 40 format(/8h /// bad,a1,2h =,i5) go to 999 50 if (iv1 .ne. 14) iv(nextiv) = iv(perm) if (iv1 .ne. 14) iv(nextv) = iv(lmat) if (iv1 .eq. 13) go to 999 k = iv(parsav) - epslon call vdflt(alg, lv-k, v(k+1)) iv(dtype0) = 2 - alg iv(oldn) = n which(1) = dflt(1) which(2) = dflt(2) which(3) = dflt(3) go to 110 60 if (n .eq. iv(oldn)) go to 80 iv(1) = 17 if (pu .eq. 0) go to 999 write(pu,70) varnm(alg), iv(oldn), n 70 format(/5h /// ,1a1,14h changed from ,i5,4h to ,i5) go to 999 c 80 if (iv1 .le. 11 .and. iv1 .ge. 1) go to 100 iv(1) = 80 if (pu .ne. 0) write(pu,90) iv1 90 format(/13h /// iv(1) =,i5,28h should be between 0 and 14.) go to 999 c 100 which(1) = cngd(1) which(2) = cngd(2) which(3) = cngd(3) c 110 if (iv1 .eq. 14) iv1 = 12 if (big .gt. tiny) go to 120 tiny = rmdcon(1) machep = rmdcon(3) big = rmdcon(6) vm(12) = machep vx(12) = big vx(13) = big vm(14) = machep vm(17) = tiny vx(17) = big vm(18) = tiny vx(18) = big vx(20) = big vx(21) = big vx(22) = big vm(24) = machep vm(25) = machep vm(26) = machep vx(28) = rmdcon(5) vm(29) = machep vx(30) = big vm(33) = machep 120 m = 0 i = 1 j = jlim(alg) k = epslon ndfalt = ndflt(alg) do 150 l = 1, ndfalt vk = v(k) if (vk .ge. vm(i) .and. vk .le. vx(i)) go to 140 m = k if (pu .ne. 0) write(pu,130) vn(1,i), vn(2,i), k, vk, 1 vm(i), vx(i) 130 format(/6h /// ,2a4,5h.. v(,i2,3h) =,d11.3,7h should, 1 11h be between,d11.3,4h and,d11.3) 140 k = k + 1 i = i + 1 if (i .eq. j) i = ijmp 150 continue c if (iv(nvdflt) .eq. ndfalt) go to 170 iv(1) = 51 if (pu .eq. 0) go to 999 write(pu,160) iv(nvdflt), ndfalt 160 format(/13h iv(nvdflt) =,i5,13h rather than ,i5) go to 999 170 if ((iv(dtype) .gt. 0 .or. v(dinit) .gt. zero) .and. iv1 .eq. 12) 1 go to 200 do 190 i = 1, n if (d(i) .gt. zero) go to 190 m = 18 if (pu .ne. 0) write(pu,180) i, d(i) 180 format(/8h /// d(,i3,3h) =,d11.3,19h should be positive) 190 continue 200 if (m .eq. 0) go to 210 iv(1) = m go to 999 c 210 if (pu .eq. 0 .or. iv(parprt) .eq. 0) go to 999 if (iv1 .ne. 12 .or. iv(inits) .eq. alg-1) go to 230 m = 1 write(pu,220) sh(alg), iv(inits) 220 format(/22h nondefault values..../5h init,a1,14h..... iv(25) =, 1 i3) 230 if (iv(dtype) .eq. iv(dtype0)) go to 250 if (m .eq. 0) write(pu,260) which m = 1 write(pu,240) iv(dtype) 240 format(20h dtype..... iv(16) =,i3) 250 i = 1 j = jlim(alg) k = epslon l = iv(parsav) ndfalt = ndflt(alg) do 290 ii = 1, ndfalt if (v(k) .eq. v(l)) go to 280 if (m .eq. 0) write(pu,260) which 260 format(/1h ,3a4,9halues..../) m = 1 write(pu,270) vn(1,i), vn(2,i), k, v(k) 270 format(1x,2a4,5h.. v(,i2,3h) =,d15.7) 280 k = k + 1 l = l + 1 i = i + 1 if (i .eq. j) i = ijmp 290 continue c iv(dtype0) = iv(dtype) parsv1 = iv(parsav) call vcopy(iv(nvdflt), v(parsv1), v(epslon)) go to 999 c 300 iv(1) = 15 if (pu .eq. 0) go to 999 write(pu,310) liv, miv2 310 format(/10h /// liv =,i5,17h must be at least,i5) if (liv .lt. miv1) go to 999 if (lv .lt. iv(lastv)) go to 320 go to 999 c 320 iv(1) = 16 if (pu .eq. 0) go to 999 write(pu,330) lv, iv(lastv) 330 format(/9h /// lv =,i5,17h must be at least,i5) go to 999 c 340 iv(1) = 67 if (pu .eq. 0) go to 999 write(pu,350) alg 350 format(/10h /// alg =,i5,15h must be 1 or 2) c 999 return c *** last card of parck follows *** end double precision function reldst(p, d, x, x0) c c *** compute and return relative difference between x and x0 *** c *** nl2sol version 2.2 *** c integer p double precision d(p), x(p), x0(p) c/+ double precision dabs c/ integer i double precision emax, t, xmax, zero c/6 data zero/0.d+0/ c/7 c parameter (zero=0.d+0) c/ c emax = zero xmax = zero do 10 i = 1, p t = dabs(d(i) * (x(i) - x0(i))) if (emax .lt. t) emax = t t = d(i) * (dabs(x(i)) + dabs(x0(i))) if (xmax .lt. t) xmax = t 10 continue reldst = zero if (xmax .gt. zero) reldst = emax / xmax 999 return c *** last card of reldst follows *** end logical function stopx(idummy) c *****parameters... integer idummy c c .................................................................. c c *****purpose... c this function may serve as the stopx (asynchronous interruption) c function for the nl2sol (nonlinear least-squares) package at c those installations which do not wish to implement a c dynamic stopx. c c *****algorithm notes... c at installations where the nl2sol system is used c interactively, this dummy stopx should be replaced by a c function that returns .true. if and only if the interrupt c (break) key has been pressed since the last call on stopx. c c .................................................................. c stopx = .false. return end subroutine vaxpy(p, w, a, x, y) c c *** set w = a*x + y -- w, x, y = p-vectors, a = scalar *** c integer p double precision a, w(p), x(p), y(p) c integer i c do 10 i = 1, p 10 w(i) = a*x(i) + y(i) return end subroutine vcopy(p, y, x) c c *** set y = x, where x and y are p-vectors *** c integer p double precision x(p), y(p) c integer i c do 10 i = 1, p 10 y(i) = x(i) return end subroutine vdflt(alg, lv, v) c c *** supply ***sol (version 2.3) default values to v *** c c *** alg = 1 means regression constants. c *** alg = 2 means general unconstrained optimization constants. c integer alg, lv double precision v(lv) c/+ double precision dmax1 c/ external rmdcon double precision rmdcon c rmdcon... returns machine-dependent constants c double precision machep, mepcrt, one, sqteps, three c c *** subscripts for v *** c integer afctol, bias, cosmin, decfac, delta0, dfac, dinit, dltfdc, 1 dltfdj, dtinit, d0init, epslon, eta0, fuzz, huberc, 2 incfac, lmax0, lmaxs, phmnfc, phmxfc, rdfcmn, rdfcmx, 3 rfctol, rlimit, rsptol, sctol, sigmin, tuner1, tuner2, 4 tuner3, tuner4, tuner5, xctol, xftol c c/6 data one/1.d+0/, three/3.d+0/ c/7 c parameter (one=1.d+0, three=3.d+0) c/ c c *** v subscript values *** c c/6 data afctol/31/, bias/43/, cosmin/47/, decfac/22/, delta0/44/, 1 dfac/41/, dinit/38/, dltfdc/42/, dltfdj/43/, dtinit/39/, 2 d0init/40/, epslon/19/, eta0/42/, fuzz/45/, huberc/48/, 3 incfac/23/, lmax0/35/, lmaxs/36/, phmnfc/20/, phmxfc/21/, 4 rdfcmn/24/, rdfcmx/25/, rfctol/32/, rlimit/46/, rsptol/49/, 5 sctol/37/, sigmin/50/, tuner1/26/, tuner2/27/, tuner3/28/, 6 tuner4/29/, tuner5/30/, xctol/33/, xftol/34/ c/7 c parameter (afctol=31, bias=43, cosmin=47, decfac=22, delta0=44, c 1 dfac=41, dinit=38, dltfdc=42, dltfdj=43, dtinit=39, c 2 d0init=40, epslon=19, eta0=42, fuzz=45, huberc=48, c 3 incfac=23, lmax0=35, lmaxs=36, phmnfc=20, phmxfc=21, c 4 rdfcmn=24, rdfcmx=25, rfctol=32, rlimit=46, rsptol=49, c 5 sctol=37, sigmin=50, tuner1=26, tuner2=27, tuner3=28, c 6 tuner4=29, tuner5=30, xctol=33, xftol=34) c/ c c------------------------------- body -------------------------------- c machep = rmdcon(3) v(afctol) = 1.d-20 if (machep .gt. 1.d-10) v(afctol) = machep**2 v(decfac) = 0.5d+0 sqteps = rmdcon(4) v(dfac) = 0.6d+0 v(delta0) = sqteps v(dtinit) = 1.d-6 mepcrt = machep ** (one/three) v(d0init) = 1.d+0 v(epslon) = 0.1d+0 v(incfac) = 2.d+0 v(lmax0) = 1.d+0 v(lmaxs) = 1.d+0 v(phmnfc) = -0.1d+0 v(phmxfc) = 0.1d+0 v(rdfcmn) = 0.1d+0 v(rdfcmx) = 4.d+0 v(rfctol) = dmax1(1.d-10, mepcrt**2) v(sctol) = v(rfctol) v(tuner1) = 0.1d+0 v(tuner2) = 1.d-4 v(tuner3) = 0.75d+0 v(tuner4) = 0.5d+0 v(tuner5) = 0.75d+0 v(xctol) = sqteps v(xftol) = 1.d+2 * machep c if (alg .ge. 2) go to 10 c c *** regression values c v(cosmin) = dmax1(1.d-6, 1.d+2 * machep) v(dinit) = 0.d+0 v(dltfdc) = mepcrt v(dltfdj) = sqteps v(fuzz) = 1.5d+0 v(huberc) = 0.7d+0 v(rlimit) = rmdcon(5) v(rsptol) = 1.d-3 v(sigmin) = 1.d-4 go to 999 c c *** general optimization values c 10 v(bias) = 0.8d+0 v(dinit) = -1.0d+0 v(eta0) = 1.0d+3 * machep c 999 return c *** last card of vdflt follows *** end subroutine vscopy(p, y, s) c c *** set p-vector y to scalar s *** c integer p double precision s, y(p) c integer i c do 10 i = 1, p 10 y(i) = s return end double precision function v2norm(p, x) c c *** return the 2-norm of the p-vector x, taking *** c *** care to avoid the most likely underflows. *** c integer p double precision x(p) c integer i, j double precision one, r, scale, sqteta, t, xi, zero c/+ double precision dabs, dsqrt c/ external rmdcon double precision rmdcon c c/6 data one/1.d+0/, zero/0.d+0/ c/7 c parameter (one=1.d+0, zero=0.d+0) c save sqteta c/ data sqteta/0.d+0/ c if (p .gt. 0) go to 10 v2norm = zero go to 999 10 do 20 i = 1, p if (x(i) .ne. zero) go to 30 20 continue v2norm = zero go to 999 c 30 scale = dabs(x(i)) if (i .lt. p) go to 40 v2norm = scale go to 999 40 t = one if (sqteta .eq. zero) sqteta = rmdcon(2) c c *** sqteta is (slightly larger than) the square root of the c *** smallest positive floating point number on the machine. c *** the tests involving sqteta are done to prevent underflows. c j = i + 1 do 60 i = j, p xi = dabs(x(i)) if (xi .gt. scale) go to 50 r = xi / scale if (r .gt. sqteta) t = t + r*r go to 60 50 r = scale / xi if (r .le. sqteta) r = zero t = one + t * r*r scale = xi 60 continue c v2norm = scale * dsqrt(t) 999 return c *** last card of v2norm follows *** end subroutine humsl(n, d, x, calcf, calcgh, iv, liv, lv, v, 1 uiparm, urparm, ufparm) c c *** minimize general unconstrained objective function using *** c *** (analytic) gradient and hessian provided by the caller. *** c integer liv, lv, n integer iv(liv), uiparm(1) double precision d(n), x(n), v(lv), urparm(1) c dimension v(78 + n*(n+12)), uiparm(*), urparm(*) external calcf, calcgh, ufparm c c------------------------------ discussion --------------------------- c c this routine is like sumsl, except that the subroutine para- c meter calcg of sumsl (which computes the gradient of the objec- c tive function) is replaced by the subroutine parameter calcgh, c which computes both the gradient and (lower triangle of the) c hessian of the objective function. the calling sequence is... c call calcgh(n, x, nf, g, h, uiparm, urparm, ufparm) c parameters n, x, nf, g, uiparm, urparm, and ufparm are the same c as for sumsl, while h is an array of length n*(n+1)/2 in which c calcgh must store the lower triangle of the hessian at x. start- c ing at h(1), calcgh must store the hessian entries in the order c (1,1), (2,1), (2,2), (3,1), (3,2), (3,3), ... c the value printed (by itsum) in the column labelled stppar c is the levenberg-marquardt used in computing the current step. c zero means a full newton step. if the special case described in c ref. 1 is detected, then stppar is negated. the value printed c in the column labelled npreldf is zero if the current hessian c is not positive definite. c it sometimes proves worthwhile to let d be determined from the c diagonal of the hessian matrix by setting iv(dtype) = 1 and c v(dinit) = 0. the following iv and v components are relevant... c c iv(dtol)..... iv(59) gives the starting subscript in v of the dtol c array used when d is updated. (iv(dtol) can be c initialized by calling humsl with iv(1) = 13.) c iv(dtype).... iv(16) tells how the scale vector d should be chosen. c iv(dtype) .le. 0 means that d should not be updated, and c iv(dtype) .ge. 1 means that d should be updated as c described below with v(dfac). default = 0. c v(dfac)..... v(41) and the dtol and d0 arrays (see v(dtinit) and c v(d0init)) are used in updating the scale vector d when c iv(dtype) .gt. 0. (d is initialized according to c v(dinit), described in sumsl.) let c d1(i) = max(sqrt(abs(h(i,i))), v(dfac)*d(i)), c where h(i,i) is the i-th diagonal element of the current c hessian. if iv(dtype) = 1, then d(i) is set to d1(i) c unless d1(i) .lt. dtol(i), in which case d(i) is set to c max(d0(i), dtol(i)). c if iv(dtype) .ge. 2, then d is updated during the first c iteration as for iv(dtype) = 1 (after any initialization c due to v(dinit)) and is left unchanged thereafter. c default = 0.6. c v(dtinit)... v(39), if positive, is the value to which all components c of the dtol array (see v(dfac)) are initialized. if c v(dtinit) = 0, then it is assumed that the caller has c stored dtol in v starting at v(iv(dtol)). c default = 10**-6. c v(d0init)... v(40), if positive, is the value to which all components c of the d0 vector (see v(dfac)) are initialized. if c v(dfac) = 0, then it is assumed that the caller has c stored d0 in v starting at v(iv(dtol)+n). default = 1.0. c c *** reference *** c c 1. gay, d.m. (1981), computing optimal locally constrained steps, c siam j. sci. statist. comput. 2, pp. 186-197. c. c *** general *** c c coded by david m. gay (winter 1980). 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---------------------------- declarations --------------------------- c external deflt, humit c c deflt... provides default input values for iv and v. c humit... reverse-communication routine that does humsl algorithm. c integer g1, h1, iv1, lh, nf double precision f c c *** subscripts for iv *** c integer g, h, nextv, nfcall, nfgcal, toobig, vneed c c/6 data nextv/47/, nfcall/6/, nfgcal/7/, g/28/, h/56/, toobig/2/, 1 vneed/4/ c/7 c parameter (nextv=47, nfcall=6, nfgcal=7, g=28, h=56, toobig=2, c 1 vneed=4) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c lh = n * (n + 1) / 2 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+3)/2 iv1 = iv(1) if (iv1 .eq. 14) go to 10 if (iv1 .gt. 2 .and. iv1 .lt. 12) go to 10 g1 = 1 h1 = 1 if (iv1 .eq. 12) iv(1) = 13 go to 20 c 10 g1 = iv(g) h1 = iv(h) c 20 call humit(d, f, v(g1), v(h1), iv, lh, 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 calcgh(n, x, iv(nfgcal), v(g1), v(h1), uiparm, urparm, 1 ufparm) go to 20 c 50 if (iv(1) .ne. 14) go to 999 c c *** storage allocation c iv(g) = iv(nextv) iv(h) = iv(g) + n iv(nextv) = iv(h) + n*(n+1)/2 if (iv1 .ne. 13) go to 10 c 999 return c *** last card of humsl follows *** end subroutine humit(d, fx, g, h, iv, lh, liv, lv, n, v, x) c c *** carry out humsl (unconstrained minimization) iterations, using c *** hessian matrix provided by the caller. c c *** parameter declarations *** c integer lh, liv, lv, n integer iv(liv) double precision d(n), fx, g(n), h(lh), v(lv), x(n) c c-------------------------- parameter usage -------------------------- c c d.... scale vector. c fx... function value. c g.... gradient vector. c h.... lower triangle of the hessian, stored rowwise. c iv... integer value array. c lh... length of h = p*(p+1)/2. c liv.. length of iv (at least 60). c lv... length of v (at least 78 + n*(n+21)/2). c n.... number of variables (components in x and g). c v.... floating-point value array. c x.... parameter vector. c c *** discussion *** c c parameters iv, n, v, and x are the same as the corresponding c ones to humsl (which see), except that v can be shorter (since c the part of v that humsl uses for storing g and h is not needed). c moreover, compared with humsl, 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 humsl, is not referenced by humit or the c subroutines it calls. c c iv(1) = 1 means the caller should set fx to f(x), the function value c at x, and call humit again, having changed none of the c other parameters. an exception occurs if f(x) cannot be c computed (e.g. if overflow would occur), which may happen c because of an oversized step. in this case the caller c should set iv(toobig) = iv(2) to 1, which will cause c humit to ignore fx and try a smaller step. the para- c meter nf that humsl passes to calcf (for possible use by c calcgh) is a copy of iv(nfcall) = iv(6). c iv(1) = 2 means the caller should set g to g(x), the gradient of f at c x, and h to the lower triangle of h(x), the hessian of f c at x, and call humit again, having changed none of the c other parameters except perhaps the scale vector d. c the parameter nf that humsl passes to calcg is c iv(nfgcal) = iv(7). if g(x) and h(x) cannot be evaluated, c then the caller may set iv(nfgcal) to 0, in which case c humit will return with iv(1) = 65. c note -- humit overwrites h with the lower triangle c of diag(d)**-1 * h(x) * diag(d)**-1. c. c *** general *** c c coded by david m. gay (winter 1980). 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 and humsl for references.) c c+++++++++++++++++++++++++++ declarations ++++++++++++++++++++++++++++ c c *** local variables *** c integer dg1, dummy, i, j, k, l, lstgst, nn1o2, step1, 1 temp1, w1, x01 double precision t c c *** constants *** c double precision one, onep2, zero c c *** no intrinsic functions *** c c *** external functions and subroutines *** c external assst, deflt, dotprd, dupdu, gqtst, itsum, parck, 1 reldst, slvmul, stopx, vaxpy, vcopy, vscopy, v2norm logical stopx double precision dotprd, reldst, v2norm c c assst.... assesses candidate step. c deflt.... provides default iv and v input values. c dotprd... returns inner product of two vectors. c dupdu.... updates scale vector d. c gqtst.... computes optimally locally constrained step. c itsum.... prints iteration summary and info on initial and final x. c parck.... checks validity of input iv and v values. c reldst... computes v(reldx) = relative step size. c slvmul... multiplies symmetric matrix times vector, given the lower c triangle of the matrix. 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 v2norm... returns the 2-norm of a vector. c c *** subscripts for iv and v *** c integer cnvcod, dg, dgnorm, dinit, dstnrm, dtinit, dtol, 1 dtype, d0init, f, f0, fdif, gtstep, incfac, irc, kagqt, 2 lmat, lmax0, lmaxs, mode, model, mxfcal, mxiter, nextv, 3 nfcall, nfgcal, ngcall, niter, preduc, radfac, radinc, 4 radius, rad0, reldx, restor, step, stglim, stlstg, stppar, 5 toobig, tuner4, tuner5, vneed, w, xirc, x0 c c *** iv subscript values *** c c/6 data cnvcod/55/, dg/37/, dtol/59/, dtype/16/, irc/29/, kagqt/33/, 1 lmat/42/, mode/35/, model/5/, mxfcal/17/, mxiter/18/, 2 nextv/47/, nfcall/6/, nfgcal/7/, ngcall/30/, niter/31/, 3 radinc/8/, restor/9/, step/40/, stglim/11/, stlstg/41/, 4 toobig/2/, vneed/4/, w/34/, xirc/13/, x0/43/ c/7 c parameter (cnvcod=55, dg=37, dtol=59, dtype=16, irc=29, kagqt=33, c 1 lmat=42, mode=35, model=5, mxfcal=17, mxiter=18, c 2 nextv=47, nfcall=6, nfgcal=7, ngcall=30, niter=31, c 3 radinc=8, restor=9, step=40, stglim=11, stlstg=41, c 4 toobig=2, vneed=4, w=34, xirc=13, x0=43) c/ c c *** v subscript values *** c c/6 data dgnorm/1/, dinit/38/, dstnrm/2/, dtinit/39/, d0init/40/, 1 f/10/, f0/13/, fdif/11/, gtstep/4/, incfac/23/, lmax0/35/, 2 lmaxs/36/, preduc/7/, radfac/16/, radius/8/, rad0/9/, 3 reldx/17/, stppar/5/, tuner4/29/, tuner5/30/ c/7 c parameter (dgnorm=1, dinit=38, dstnrm=2, dtinit=39, d0init=40, c 1 f=10, f0=13, fdif=11, gtstep=4, incfac=23, lmax0=35, c 2 lmaxs=36, preduc=7, radfac=16, radius=8, rad0=9, c 3 reldx=17, stppar=5, tuner4=29, tuner5=30) c/ c c/6 data one/1.d+0/, onep2/1.2d+0/, zero/0.d+0/ c/7 c parameter (one=1.d+0, onep2=1.2d+0, zero=0.d+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c i = iv(1) if (i .eq. 1) go to 30 if (i .eq. 2) go to 40 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+21)/2 + 7 call parck(2, d, iv, liv, lv, n, v) i = iv(1) - 2 if (i .gt. 12) go to 999 nn1o2 = n * (n + 1) / 2 if (lh .ge. nn1o2) go to (210,210,210,210,210,210,160,120,160, 1 10,10,20), i iv(1) = 66 go to 350 c c *** storage allocation *** c 10 iv(dtol) = iv(lmat) + nn1o2 iv(x0) = iv(dtol) + 2*n iv(step) = iv(x0) + n iv(stlstg) = iv(step) + n iv(dg) = iv(stlstg) + n iv(w) = iv(dg) + n iv(nextv) = iv(w) + 4*n + 7 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 v(stppar) = zero if (v(dinit) .ge. zero) call vscopy(n, d, v(dinit)) k = iv(dtol) if (v(dtinit) .gt. zero) call vscopy(n, v(k), v(dtinit)) k = k + n if (v(d0init) .gt. zero) call vscopy(n, v(k), v(d0init)) iv(1) = 1 go to 999 c 30 v(f) = fx if (iv(mode) .ge. 0) go to 210 iv(1) = 2 if (iv(toobig) .eq. 0) go to 999 iv(1) = 63 go to 350 c c *** make sure gradient could be computed *** c 40 if (iv(nfgcal) .ne. 0) go to 50 iv(1) = 65 go to 350 c c *** update the scale vector d *** c 50 dg1 = iv(dg) if (iv(dtype) .le. 0) go to 70 k = dg1 j = 0 do 60 i = 1, n j = j + i v(k) = h(j) k = k + 1 60 continue call dupdu(d, v(dg1), iv, liv, lv, n, v) c c *** compute scaled gradient and its norm *** c 70 dg1 = iv(dg) k = dg1 do 80 i = 1, n v(k) = g(i) / d(i) k = k + 1 80 continue v(dgnorm) = v2norm(n, v(dg1)) c c *** compute scaled hessian *** c k = 1 do 100 i = 1, n t = one / d(i) do 90 j = 1, i h(k) = t * h(k) / d(j) k = k + 1 90 continue 100 continue c if (iv(cnvcod) .ne. 0) go to 340 if (iv(mode) .eq. 0) go to 300 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 110 call itsum(d, g, iv, liv, lv, n, v, x) 120 k = iv(niter) if (k .lt. iv(mxiter)) go to 130 iv(1) = 10 go to 350 c 130 iv(niter) = k + 1 c c *** initialize for start of next iteration *** c dg1 = iv(dg) x01 = iv(x0) v(f0) = v(f) iv(irc) = 4 iv(kagqt) = -1 c c *** copy x to x0 *** c call vcopy(n, v(x01), x) c c *** update radius *** c if (k .eq. 0) go to 150 step1 = iv(step) k = step1 do 140 i = 1, n v(k) = d(i) * v(k) k = k + 1 140 continue v(radius) = v(radfac) * v2norm(n, v(step1)) c c *** check stopx and function evaluation limit *** c 150 if (.not. stopx(dummy)) go to 170 iv(1) = 11 go to 180 c c *** come here when restarting after func. eval. limit or stopx. c 160 if (v(f) .ge. v(f0)) go to 170 v(radfac) = one k = iv(niter) go to 130 c 170 if (iv(nfcall) .lt. iv(mxfcal)) go to 190 iv(1) = 9 180 if (v(f) .ge. v(f0)) go to 350 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 290 c c. . . . . . . . . . . . . compute candidate step . . . . . . . . . . c 190 step1 = iv(step) dg1 = iv(dg) l = iv(lmat) w1 = iv(w) call gqtst(d, v(dg1), h, iv(kagqt), v(l), n, v(step1), v, v(w1)) if (iv(irc) .eq. 6) go to 210 c c *** check whether evaluating f(x0 + step) looks worthwhile *** c if (v(dstnrm) .le. zero) go to 210 if (iv(irc) .ne. 5) go to 200 if (v(radfac) .le. one) go to 200 if (v(preduc) .le. onep2 * v(fdif)) go to 210 c c *** compute f(x0 + step) *** c 200 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 210 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 220 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 220 k = iv(irc) go to (230,260,260,260,230,240,250,250,250,250,250,250,330,300), k c c *** recompute step with new radius *** c 230 v(radius) = v(radfac) * v(dstnrm) go to 150 c c *** compute step of length v(lmaxs) for singular convergence test. c 240 v(radius) = v(lmaxs) go to 190 c c *** convergence or false convergence *** c 250 iv(cnvcod) = k - 4 if (v(f) .ge. v(f0)) go to 340 if (iv(xirc) .eq. 14) go to 340 iv(xirc) = 14 c c. . . . . . . . . . . . process acceptable step . . . . . . . . . . . c 260 if (iv(irc) .ne. 3) go to 290 temp1 = lstgst c c *** prepare for gradient tests *** c *** set temp1 = hessian * step + g(x0) c *** = diag(d) * (h * step + g(x0)) c c use x0 vector as temporary. k = x01 do 270 i = 1, n v(k) = d(i) * v(step1) k = k + 1 step1 = step1 + 1 270 continue call slvmul(n, v(temp1), h, v(x01)) do 280 i = 1, n v(temp1) = d(i) * v(temp1) + g(i) temp1 = temp1 + 1 280 continue c c *** compute gradient and hessian *** c 290 iv(ngcall) = iv(ngcall) + 1 iv(1) = 2 go to 999 c 300 iv(1) = 2 if (iv(irc) .ne. 3) go to 110 c c *** set v(radfac) by gradient tests *** c temp1 = iv(stlstg) step1 = iv(step) c c *** set temp1 = diag(d)**-1 * (hessian*step + (g(x0)-g(x))) *** c k = temp1 do 310 i = 1, n v(k) = (v(k) - g(i)) / d(i) k = k + 1 310 continue c c *** do gradient tests *** c if (v2norm(n, v(temp1)) .le. v(dgnorm) * v(tuner4)) go to 320 if (dotprd(n, g, v(step1)) 1 .ge. v(gtstep) * v(tuner5)) go to 110 320 v(radfac) = v(incfac) go to 110 c c. . . . . . . . . . . . . . misc. details . . . . . . . . . . . . . . c c *** bad parameters to assess *** c 330 iv(1) = 64 go to 350 c c *** print summary of final iteration and other requested items *** c 340 iv(1) = iv(cnvcod) iv(cnvcod) = 0 350 call itsum(d, g, iv, liv, lv, n, v, x) c 999 return c c *** last card of humit follows *** end subroutine dupdu(d, hdiag, iv, liv, lv, n, v) c c *** update scale vector d for humsl *** c c *** parameter declarations *** c integer liv, lv, n integer iv(liv) double precision d(n), hdiag(n), v(lv) c c *** local variables *** c integer dtoli, d0i, i double precision t, vdfac c c *** intrinsic functions *** c/+ double precision dabs, dmax1, dsqrt c/ c *** subscripts for iv and v *** c integer dfac, dtol, dtype, niter c/6 data dfac/41/, dtol/59/, dtype/16/, niter/31/ c/7 c parameter (dfac=41, dtol=59, dtype=16, niter=31) c/ c c------------------------------- body -------------------------------- c i = iv(dtype) if (i .eq. 1) go to 10 if (iv(niter) .gt. 0) go to 999 c 10 dtoli = iv(dtol) d0i = dtoli + n vdfac = v(dfac) do 20 i = 1, n t = dmax1(dsqrt(dabs(hdiag(i))), vdfac*d(i)) if (t .lt. v(dtoli)) t = dmax1(v(dtoli), v(d0i)) d(i) = t dtoli = dtoli + 1 d0i = d0i + 1 20 continue c 999 return c *** last card of dupdu follows *** end subroutine gqtst(d, dig, dihdi, ka, l, p, step, v, w) c c *** compute goldfeld-quandt-trotter step by more-hebden technique *** c *** (nl2sol version 2.2), modified a la more and sorensen *** c c *** parameter declarations *** c integer ka, p double precision d(p), dig(p), dihdi(1), l(1), v(21), step(p), 1 w(1) c dimension dihdi(p*(p+1)/2), l(p*(p+1)/2), w(4*p+7) c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** purpose *** c c given the (compactly stored) lower triangle of a scaled c hessian (approximation) and a nonzero scaled gradient vector, c this subroutine computes a goldfeld-quandt-trotter step of c approximate length v(radius) by the more-hebden technique. in c other words, step is computed to (approximately) minimize c psi(step) = (g**t)*step + 0.5*(step**t)*h*step such that the c 2-norm of d*step is at most (approximately) v(radius), where c g is the gradient, h is the hessian, and d is a diagonal c scale matrix whose diagonal is stored in the parameter d. c (gqtst assumes dig = d**-1 * g and dihdi = d**-1 * h * d**-1.) c c *** parameter description *** c c d (in) = the scale vector, i.e. the diagonal of the scale c matrix d mentioned above under purpose. c dig (in) = the scaled gradient vector, d**-1 * g. if g = 0, then c step = 0 and v(stppar) = 0 are returned. c dihdi (in) = lower triangle of the scaled hessian (approximation), c i.e., d**-1 * h * d**-1, stored compactly by rows., i.e., c in the order (1,1), (2,1), (2,2), (3,1), (3,2), etc. c ka (i/o) = the number of hebden iterations (so far) taken to deter- c mine step. ka .lt. 0 on input means this is the first c attempt to determine step (for the present dig and dihdi) c -- ka is initialized to 0 in this case. output with c ka = 0 (or v(stppar) = 0) means step = -(h**-1)*g. c l (i/o) = workspace of length p*(p+1)/2 for cholesky factors. c p (in) = number of parameters -- the hessian is a p x p matrix. c step (i/o) = the step computed. c v (i/o) contains various constants and variables described below. c w (i/o) = workspace of length 4*p + 6. c c *** entries in v *** c c v(dgnorm) (i/o) = 2-norm of (d**-1)*g. c v(dstnrm) (output) = 2-norm of d*step. c v(dst0) (i/o) = 2-norm of d*(h**-1)*g (for pos. def. h only), or c overestimate of smallest eigenvalue of (d**-1)*h*(d**-1). c v(epslon) (in) = max. rel. error allowed for psi(step). for the c step returned, psi(step) will exceed its optimal value c by less than -v(epslon)*psi(step). suggested value = 0.1. c v(gtstep) (out) = inner product between g and step. c v(nreduc) (out) = psi(-(h**-1)*g) = psi(newton step) (for pos. def. c h only -- v(nreduc) is set to zero otherwise). c v(phmnfc) (in) = tol. (together with v(phmxfc)) for accepting step c (more*s sigma). the error v(dstnrm) - v(radius) must lie c between v(phmnfc)*v(radius) and v(phmxfc)*v(radius). c v(phmxfc) (in) (see v(phmnfc).) c suggested values -- v(phmnfc) = -0.25, v(phmxfc) = 0.5. c v(preduc) (out) = psi(step) = predicted obj. func. reduction for step. c v(radius) (in) = radius of current (scaled) trust region. c v(rad0) (i/o) = value of v(radius) from previous call. c v(stppar) (i/o) is normally the marquardt parameter, i.e. the alpha c described below under algorithm notes. if h + alpha*d**2 c (see algorithm notes) is (nearly) singular, however, c then v(stppar) = -alpha. c c *** usage notes *** c c if it is desired to recompute step using a different value of c v(radius), then this routine may be restarted by calling it c with all parameters unchanged except v(radius). (this explains c why step and w are listed as i/o). on an initial call (one with c ka .lt. 0), step and w need not be initialized and only compo- c nents v(epslon), v(stppar), v(phmnfc), v(phmxfc), v(radius), and c v(rad0) of v must be initialized. c c *** algorithm notes *** c c the desired g-q-t step (ref. 2, 3, 4, 6) satisfies c (h + alpha*d**2)*step = -g for some nonnegative alpha such that c h + alpha*d**2 is positive semidefinite. alpha and step are c computed by a scheme analogous to the one described in ref. 5. c estimates of the smallest and largest eigenvalues of the hessian c are obtained from the gerschgorin circle theorem enhanced by a c simple form of the scaling described in ref. 7. cases in which c h + alpha*d**2 is nearly (or exactly) singular are handled by c the technique discussed in ref. 2. in these cases, a step of c (exact) length v(radius) is returned for which psi(step) exceeds c its optimal value by less than -v(epslon)*psi(step). the test c suggested in ref. 6 for detecting the special case is performed c once two matrix factorizations have been done -- doing so sooner c seems to degrade the performance of optimization routines that c call this routine. c c *** functions and subroutines called *** c c dotprd - returns inner product of two vectors. c litvmu - applies inverse-transpose of compact lower triang. matrix. c livmul - applies inverse of compact lower triang. matrix. c lsqrt - finds cholesky factor (of compactly stored lower triang.). c lsvmin - returns approx. to min. sing. value of lower triang. matrix. c rmdcon - returns machine-dependent constants. c v2norm - returns 2-norm of a vector. c c *** references *** c c 1. dennis, j.e., gay, d.m., and welsch, r.e. (1981), an adaptive c nonlinear least-squares algorithm, acm trans. math. c software, vol. 7, no. 3. c 2. gay, d.m. (1981), computing optimal locally constrained steps, c siam j. sci. statist. computing, vol. 2, no. 2, pp. c 186-197. c 3. goldfeld, s.m., quandt, r.e., and trotter, h.f. (1966), c maximization by quadratic hill-climbing, econometrica 34, c pp. 541-551. c 4. hebden, m.d. (1973), an algorithm for minimization using exact c second derivatives, report t.p. 515, theoretical physics c div., a.e.r.e. harwell, oxon., england. c 5. more, j.j. (1978), the levenberg-marquardt algorithm, implemen- c tation and theory, pp.105-116 of springer lecture notes c in mathematics no. 630, edited by g.a. watson, springer- c verlag, berlin and new york. c 6. more, j.j., and sorensen, d.c. (1981), computing a trust region c step, technical report anl-81-83, argonne national lab. c 7. varga, r.s. (1965), minimal gerschgorin sets, pacific j. math. 15, c pp. 719-729. c c *** general *** c c coded by david m. gay. c this subroutine was written in connection with research c supported by the national science foundation under grants c mcs-7600324, dcr75-10143, 76-14311dss, mcs76-11989, and c mcs-7906671. c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** local variables *** c logical restrt integer dggdmx, diag, diag0, dstsav, emax, emin, i, im1, inc, irc, 1 j, k, kalim, kamin, k1, lk0, phipin, q, q0, uk0, x double precision alphak, aki, akk, delta, dst, eps, gtsta, lk, 1 oldphi, phi, phimax, phimin, psifac, rad, radsq, 2 root, si, sk, sw, t, twopsi, t1, t2, uk, wi c c *** constants *** double precision big, dgxfac, epsfac, four, half, kappa, negone, 1 one, p001, six, three, two, zero c c *** intrinsic functions *** c/+ double precision dabs, dmax1, dmin1, dsqrt c/ c *** external functions and subroutines *** c external dotprd, litvmu, livmul, lsqrt, lsvmin, rmdcon, v2norm double precision dotprd, lsvmin, rmdcon, v2norm c c *** subscripts for v *** c integer dgnorm, dstnrm, dst0, epslon, gtstep, stppar, nreduc, 1 phmnfc, phmxfc, preduc, radius, rad0 c/6 data dgnorm/1/, dstnrm/2/, dst0/3/, epslon/19/, gtstep/4/, 1 nreduc/6/, phmnfc/20/, phmxfc/21/, preduc/7/, radius/8/, 2 rad0/9/, stppar/5/ c/7 c parameter (dgnorm=1, dstnrm=2, dst0=3, epslon=19, gtstep=4, c 1 nreduc=6, phmnfc=20, phmxfc=21, preduc=7, radius=8, c 2 rad0=9, stppar=5) c/ c c/6 data epsfac/50.0d+0/, four/4.0d+0/, half/0.5d+0/, 1 kappa/2.0d+0/, negone/-1.0d+0/, one/1.0d+0/, p001/1.0d-3/, 2 six/6.0d+0/, three/3.0d+0/, two/2.0d+0/, zero/0.0d+0/ c/7 c parameter (epsfac=50.0d+0, four=4.0d+0, half=0.5d+0, c 1 kappa=2.0d+0, negone=-1.0d+0, one=1.0d+0, p001=1.0d-3, c 2 six=6.0d+0, three=3.0d+0, two=2.0d+0, zero=0.0d+0) c save dgxfac c/ data big/0.d+0/, dgxfac/0.d+0/ c c *** body *** c c *** store largest abs. entry in (d**-1)*h*(d**-1) at w(dggdmx). dggdmx = p + 1 c *** store gerschgorin over- and underestimates of the largest c *** and smallest eigenvalues of (d**-1)*h*(d**-1) at w(emax) c *** and w(emin) respectively. emax = dggdmx + 1 emin = emax + 1 c *** for use in recomputing step, the final values of lk, uk, dst, c *** and the inverse derivative of more*s phi at 0 (for pos. def. c *** h) are stored in w(lk0), w(uk0), w(dstsav), and w(phipin) c *** respectively. lk0 = emin + 1 phipin = lk0 + 1 uk0 = phipin + 1 dstsav = uk0 + 1 c *** store diag of (d**-1)*h*(d**-1) in w(diag),...,w(diag0+p). diag0 = dstsav diag = diag0 + 1 c *** store -d*step in w(q),...,w(q0+p). q0 = diag0 + p q = q0 + 1 c *** allocate storage for scratch vector x *** x = q + p rad = v(radius) radsq = rad**2 c *** phitol = max. error allowed in dst = v(dstnrm) = 2-norm of c *** d*step. phimax = v(phmxfc) * rad phimin = v(phmnfc) * rad psifac = two * v(epslon) / (three * (four * (v(phmnfc) + one) * 1 (kappa + one) + kappa + two) * rad**2) c *** oldphi is used to detect limits of numerical accuracy. if c *** we recompute step and it does not change, then we accept it. oldphi = zero eps = v(epslon) irc = 0 restrt = .false. kalim = ka + 50 c c *** start or restart, depending on ka *** c if (ka .ge. 0) go to 290 c c *** fresh start *** c k = 0 uk = negone ka = 0 kalim = 50 v(dgnorm) = v2norm(p, dig) v(nreduc) = zero v(dst0) = zero kamin = 3 if (v(dgnorm) .eq. zero) kamin = 0 c c *** store diag(dihdi) in w(diag0+1),...,w(diag0+p) *** c j = 0 do 10 i = 1, p j = j + i k1 = diag0 + i w(k1) = dihdi(j) 10 continue c c *** determine w(dggdmx), the largest element of dihdi *** c t1 = zero j = p * (p + 1) / 2 do 20 i = 1, j t = dabs(dihdi(i)) if (t1 .lt. t) t1 = t 20 continue w(dggdmx) = t1 c c *** try alpha = 0 *** c 30 call lsqrt(1, p, l, dihdi, irc) if (irc .eq. 0) go to 50 c *** indef. h -- underestimate smallest eigenvalue, use this c *** estimate to initialize lower bound lk on alpha. j = irc*(irc+1)/2 t = l(j) l(j) = one do 40 i = 1, irc 40 w(i) = zero w(irc) = one call litvmu(irc, w, l, w) t1 = v2norm(irc, w) lk = -t / t1 / t1 v(dst0) = -lk if (restrt) go to 210 go to 70 c c *** positive definite h -- compute unmodified newton step. *** 50 lk = zero t = lsvmin(p, l, w(q), w(q)) if (t .ge. one) go to 60 if (big .le. zero) big = rmdcon(6) if (v(dgnorm) .ge. t*t*big) go to 70 60 call livmul(p, w(q), l, dig) gtsta = dotprd(p, w(q), w(q)) v(nreduc) = half * gtsta call litvmu(p, w(q), l, w(q)) dst = v2norm(p, w(q)) v(dst0) = dst phi = dst - rad if (phi .le. phimax) go to 260 if (restrt) go to 210 c c *** prepare to compute gerschgorin estimates of largest (and c *** smallest) eigenvalues. *** c 70 k = 0 do 100 i = 1, p wi = zero if (i .eq. 1) go to 90 im1 = i - 1 do 80 j = 1, im1 k = k + 1 t = dabs(dihdi(k)) wi = wi + t w(j) = w(j) + t 80 continue 90 w(i) = wi k = k + 1 100 continue c c *** (under-)estimate smallest eigenvalue of (d**-1)*h*(d**-1) *** c k = 1 t1 = w(diag) - w(1) if (p .le. 1) go to 120 do 110 i = 2, p j = diag0 + i t = w(j) - w(i) if (t .ge. t1) go to 110 t1 = t k = i 110 continue c 120 sk = w(k) j = diag0 + k akk = w(j) k1 = k*(k-1)/2 + 1 inc = 1 t = zero do 150 i = 1, p if (i .eq. k) go to 130 aki = dabs(dihdi(k1)) si = w(i) j = diag0 + i t1 = half * (akk - w(j) + si - aki) t1 = t1 + dsqrt(t1*t1 + sk*aki) if (t .lt. t1) t = t1 if (i .lt. k) go to 140 130 inc = i 140 k1 = k1 + inc 150 continue c w(emin) = akk - t uk = v(dgnorm)/rad - w(emin) if (v(dgnorm) .eq. zero) uk = uk + p001 + p001*uk if (uk .le. zero) uk = p001 c c *** compute gerschgorin (over-)estimate of largest eigenvalue *** c k = 1 t1 = w(diag) + w(1) if (p .le. 1) go to 170 do 160 i = 2, p j = diag0 + i t = w(j) + w(i) if (t .le. t1) go to 160 t1 = t k = i 160 continue c 170 sk = w(k) j = diag0 + k akk = w(j) k1 = k*(k-1)/2 + 1 inc = 1 t = zero do 200 i = 1, p if (i .eq. k) go to 180 aki = dabs(dihdi(k1)) si = w(i) j = diag0 + i t1 = half * (w(j) + si - aki - akk) t1 = t1 + dsqrt(t1*t1 + sk*aki) if (t .lt. t1) t = t1 if (i .lt. k) go to 190 180 inc = i 190 k1 = k1 + inc 200 continue c w(emax) = akk + t lk = dmax1(lk, v(dgnorm)/rad - w(emax)) c c *** alphak = current value of alpha (see alg. notes above). we c *** use more*s scheme for initializing it. alphak = dabs(v(stppar)) * v(rad0)/rad c if (irc .ne. 0) go to 210 c c *** compute l0 for positive definite h *** c call livmul(p, w, l, w(q)) t = v2norm(p, w) w(phipin) = dst / t / t lk = dmax1(lk, phi*w(phipin)) c c *** safeguard alphak and add alphak*i to (d**-1)*h*(d**-1) *** c 210 ka = ka + 1 if (-v(dst0) .ge. alphak .or. alphak .lt. lk .or. alphak .ge. uk) 1 alphak = uk * dmax1(p001, dsqrt(lk/uk)) if (alphak .le. zero) alphak = half * uk if (alphak .le. zero) alphak = uk k = 0 do 220 i = 1, p k = k + i j = diag0 + i dihdi(k) = w(j) + alphak 220 continue c c *** try computing cholesky decomposition *** c call lsqrt(1, p, l, dihdi, irc) if (irc .eq. 0) go to 240 c c *** (d**-1)*h*(d**-1) + alphak*i is indefinite -- overestimate c *** smallest eigenvalue for use in updating lk *** c j = (irc*(irc+1))/2 t = l(j) l(j) = one do 230 i = 1, irc 230 w(i) = zero w(irc) = one call litvmu(irc, w, l, w) t1 = v2norm(irc, w) lk = alphak - t/t1/t1 v(dst0) = -lk go to 210 c c *** alphak makes (d**-1)*h*(d**-1) positive definite. c *** compute q = -d*step, check for convergence. *** c 240 call livmul(p, w(q), l, dig) gtsta = dotprd(p, w(q), w(q)) call litvmu(p, w(q), l, w(q)) dst = v2norm(p, w(q)) phi = dst - rad if (phi .le. phimax .and. phi .ge. phimin) go to 270 if (phi .eq. oldphi) go to 270 oldphi = phi if (phi .lt. zero) go to 330 c c *** unacceptable alphak -- update lk, uk, alphak *** c 250 if (ka .ge. kalim) go to 270 c *** the following dmin1 is necessary because of restarts *** if (phi .lt. zero) uk = dmin1(uk, alphak) c *** kamin = 0 only iff the gradient vanishes *** if (kamin .eq. 0) go to 210 call livmul(p, w, l, w(q)) t1 = v2norm(p, w) alphak = alphak + (phi/t1) * (dst/t1) * (dst/rad) lk = dmax1(lk, alphak) go to 210 c c *** acceptable step on first try *** c 260 alphak = zero c c *** successful step in general. compute step = -(d**-1)*q *** c 270 do 280 i = 1, p j = q0 + i step(i) = -w(j)/d(i) 280 continue v(gtstep) = -gtsta v(preduc) = half * (dabs(alphak)*dst*dst + gtsta) go to 410 c c c *** restart with new radius *** c 290 if (v(dst0) .le. zero .or. v(dst0) - rad .gt. phimax) go to 310 c c *** prepare to return newton step *** c restrt = .true. ka = ka + 1 k = 0 do 300 i = 1, p k = k + i j = diag0 + i dihdi(k) = w(j) 300 continue uk = negone go to 30 c 310 kamin = ka + 3 if (v(dgnorm) .eq. zero) kamin = 0 if (ka .eq. 0) go to 50 c dst = w(dstsav) alphak = dabs(v(stppar)) phi = dst - rad t = v(dgnorm)/rad uk = t - w(emin) if (v(dgnorm) .eq. zero) uk = uk + p001 + p001*uk if (uk .le. zero) uk = p001 if (rad .gt. v(rad0)) go to 320 c c *** smaller radius *** lk = zero if (alphak .gt. zero) lk = w(lk0) lk = dmax1(lk, t - w(emax)) if (v(dst0) .gt. zero) lk = dmax1(lk, (v(dst0)-rad)*w(phipin)) go to 250 c c *** bigger radius *** 320 if (alphak .gt. zero) uk = dmin1(uk, w(uk0)) lk = dmax1(zero, -v(dst0), t - w(emax)) if (v(dst0) .gt. zero) lk = dmax1(lk, (v(dst0)-rad)*w(phipin)) go to 250 c c *** decide whether to check for special case... in practice (from c *** the standpoint of the calling optimization code) it seems best c *** not to check until a few iterations have failed -- hence the c *** test on kamin below. c 330 delta = alphak + dmin1(zero, v(dst0)) twopsi = alphak*dst*dst + gtsta if (ka .ge. kamin) go to 340 c *** if the test in ref. 2 is satisfied, fall through to handle c *** the special case (as soon as the more-sorensen test detects c *** it). if (delta .ge. psifac*twopsi) go to 370 c c *** check for the special case of h + alpha*d**2 (nearly) c *** singular. use one step of inverse power method with start c *** from lsvmin to obtain approximate eigenvector corresponding c *** to smallest eigenvalue of (d**-1)*h*(d**-1). lsvmin returns c *** x and w with l*w = x. c 340 t = lsvmin(p, l, w(x), w) c c *** normalize w *** do 350 i = 1, p 350 w(i) = t*w(i) c *** complete current inv. power iter. -- replace w by (l**-t)*w. call litvmu(p, w, l, w) t2 = one/v2norm(p, w) do 360 i = 1, p 360 w(i) = t2*w(i) t = t2 * t c c *** now w is the desired approximate (unit) eigenvector and c *** t*x = ((d**-1)*h*(d**-1) + alphak*i)*w. c sw = dotprd(p, w(q), w) t1 = (rad + dst) * (rad - dst) root = dsqrt(sw*sw + t1) if (sw .lt. zero) root = -root si = t1 / (sw + root) c c *** the actual test for the special case... c if ((t2*si)**2 .le. eps*(dst**2 + alphak*radsq)) go to 380 c c *** update upper bound on smallest eigenvalue (when not positive) c *** (as recommended by more and sorensen) and continue... c if (v(dst0) .le. zero) v(dst0) = dmin1(v(dst0), t2**2 - alphak) lk = dmax1(lk, -v(dst0)) c c *** check whether we can hope to detect the special case in c *** the available arithmetic. accept step as it is if not. c c *** if not yet available, obtain machine dependent value dgxfac. 370 if (dgxfac .eq. zero) dgxfac = epsfac * rmdcon(3) c if (delta .gt. dgxfac*w(dggdmx)) go to 250 go to 270 c c *** special case detected... negate alphak to indicate special case c 380 alphak = -alphak v(preduc) = half * twopsi c c *** accept current step if adding si*w would lead to a c *** further relative reduction in psi of less than v(epslon)/3. c t1 = zero t = si*(alphak*sw - half*si*(alphak + t*dotprd(p,w(x),w))) if (t .lt. eps*twopsi/six) go to 390 v(preduc) = v(preduc) + t dst = rad t1 = -si 390 do 400 i = 1, p j = q0 + i w(j) = t1*w(i) - w(j) step(i) = w(j) / d(i) 400 continue v(gtstep) = dotprd(p, dig, w(q)) c c *** save values for use in a possible restart *** c 410 v(dstnrm) = dst v(stppar) = alphak w(lk0) = lk w(uk0) = uk v(rad0) = rad w(dstsav) = dst c c *** restore diagonal of dihdi *** c j = 0 do 420 i = 1, p j = j + i k = diag0 + i dihdi(j) = w(k) 420 continue c 999 return c c *** last card of gqtst follows *** end subroutine lsqrt(n1, n, l, a, irc) c c *** compute rows n1 through n of the cholesky factor l of c *** a = l*(l**t), where l and the lower triangle of a are both c *** stored compactly by rows (and may occupy the same storage). c *** irc = 0 means all went well. irc = j means the leading c *** principal j x j submatrix of a is not positive definite -- c *** and l(j*(j+1)/2) contains the (nonpos.) reduced j-th diagonal. c c *** parameters *** c integer n1, n, irc double precision l(1), a(1) c dimension l(n*(n+1)/2), a(n*(n+1)/2) c c *** local variables *** c integer i, ij, ik, im1, i0, j, jk, jm1, j0, k double precision t, td, zero c c *** intrinsic functions *** c/+ double precision dsqrt c/ c/6 data zero/0.d+0/ c/7 c parameter (zero=0.d+0) c/ c c *** body *** c i0 = n1 * (n1 - 1) / 2 do 50 i = n1, n td = zero if (i .eq. 1) go to 40 j0 = 0 im1 = i - 1 do 30 j = 1, im1 t = zero if (j .eq. 1) go to 20 jm1 = j - 1 do 10 k = 1, jm1 ik = i0 + k jk = j0 + k t = t + l(ik)*l(jk) 10 continue 20 ij = i0 + j j0 = j0 + j t = (a(ij) - t) / l(j0) l(ij) = t td = td + t*t 30 continue 40 i0 = i0 + i t = a(i0) - td if (t .le. zero) go to 60 l(i0) = dsqrt(t) 50 continue c irc = 0 go to 999 c 60 l(i0) = t irc = i c 999 return c c *** last card of lsqrt *** end double precision function lsvmin(p, l, x, y) c c *** estimate smallest sing. value of packed lower triang. matrix l c c *** parameter declarations *** c integer p double precision l(1), x(p), y(p) c dimension l(p*(p+1)/2) c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** purpose *** c c this function returns a good over-estimate of the smallest c singular value of the packed lower triangular matrix l. c c *** parameter description *** c c p (in) = the order of l. l is a p x p lower triangular matrix. c l (in) = array holding the elements of l in row order, i.e. c l(1,1), l(2,1), l(2,2), l(3,1), l(3,2), l(3,3), etc. c x (out) if lsvmin returns a positive value, then x is a normalized c approximate left singular vector corresponding to the c smallest singular value. this approximation may be very c crude. if lsvmin returns zero, then some components of x c are zero and the rest retain their input values. c y (out) if lsvmin returns a positive value, then y = (l**-1)*x is an c unnormalized approximate right singular vector correspond- c ing to the smallest singular value. this approximation c may be crude. if lsvmin returns zero, then y retains its c input value. the caller may pass the same vector for x c and y (nonstandard fortran usage), in which case y over- c writes x (for nonzero lsvmin returns). c c *** algorithm notes *** c c the algorithm is based on (1), with the additional provision that c lsvmin = 0 is returned if the smallest diagonal element of l c (in magnitude) is not more than the unit roundoff times the c largest. the algorithm uses a random number generator proposed c in (4), which passes the spectral test with flying colors -- see c (2) and (3). c c *** subroutines and functions called *** c c v2norm - function, returns the 2-norm of a vector. c c *** references *** c c (1) cline, a., moler, c., stewart, g., and wilkinson, j.h.(1977), c an estimate for the condition number of a matrix, report c tm-310, applied math. div., argonne national laboratory. c c (2) hoaglin, d.c. (1976), theoretical properties of congruential c random-number generators -- an empirical view, c memorandum ns-340, dept. of statistics, harvard univ. c c (3) knuth, d.e. (1969), the art of computer programming, vol. 2 c (seminumerical algorithms), addison-wesley, reading, mass. c c (4) smith, c.s. (1971), multiplicative pseudo-random number c generators with prime modulus, j. assoc. comput. mach. 18, c pp. 586-593. c c *** history *** c c designed and coded by david m. gay (winter 1977/summer 1978). c c *** general *** c c this subroutine was written in connection with research c supported by the national science foundation under grants c mcs-7600324, dcr75-10143, 76-14311dss, and mcs76-11989. c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** local variables *** c integer i, ii, ix, j, ji, jj, jjj, jm1, j0, pm1 double precision b, sminus, splus, t, xminus, xplus c c *** constants *** c double precision half, one, r9973, zero c c *** intrinsic functions *** c/+ integer mod real float double precision dabs c/ c *** external functions and subroutines *** c external dotprd, v2norm, vaxpy double precision dotprd, v2norm c c/6 data half/0.5d+0/, one/1.d+0/, r9973/9973.d+0/, zero/0.d+0/ c/7 c parameter (half=0.5d+0, one=1.d+0, r9973=9973.d+0, zero=0.d+0) c/ c c *** body *** c ix = 2 pm1 = p - 1 c c *** first check whether to return lsvmin = 0 and initialize x *** c ii = 0 j0 = p*pm1/2 jj = j0 + p if (l(jj) .eq. zero) go to 110 ix = mod(3432*ix, 9973) b = half*(one + float(ix)/r9973) xplus = b / l(jj) x(p) = xplus if (p .le. 1) go to 60 do 10 i = 1, pm1 ii = ii + i if (l(ii) .eq. zero) go to 110 ji = j0 + i x(i) = xplus * l(ji) 10 continue c c *** solve (l**t)*x = b, where the components of b have randomly c *** chosen magnitudes in (.5,1) with signs chosen to make x large. c c do j = p-1 to 1 by -1... do 50 jjj = 1, pm1 j = p - jjj c *** determine x(j) in this iteration. note for i = 1,2,...,j c *** that x(i) holds the current partial sum for row i. ix = mod(3432*ix, 9973) b = half*(one + float(ix)/r9973) xplus = (b - x(j)) xminus = (-b - x(j)) splus = dabs(xplus) sminus = dabs(xminus) jm1 = j - 1 j0 = j*jm1/2 jj = j0 + j xplus = xplus/l(jj) xminus = xminus/l(jj) if (jm1 .eq. 0) go to 30 do 20 i = 1, jm1 ji = j0 + i splus = splus + dabs(x(i) + l(ji)*xplus) sminus = sminus + dabs(x(i) + l(ji)*xminus) 20 continue 30 if (sminus .gt. splus) xplus = xminus x(j) = xplus c *** update partial sums *** if (jm1 .gt. 0) call vaxpy(jm1, x, xplus, l(j0+1), x) 50 continue c c *** normalize x *** c 60 t = one/v2norm(p, x) do 70 i = 1, p 70 x(i) = t*x(i) c c *** solve l*y = x and return lsvmin = 1/twonorm(y) *** c do 100 j = 1, p jm1 = j - 1 j0 = j*jm1/2 jj = j0 + j t = zero if (jm1 .gt. 0) t = dotprd(jm1, l(j0+1), y) y(j) = (x(j) - t) / l(jj) 100 continue c lsvmin = one/v2norm(p, y) go to 999 c 110 lsvmin = zero 999 return c *** last card of lsvmin follows *** end subroutine slvmul(p, y, s, x) c c *** set y = s * x, s = p x p symmetric matrix. *** c *** lower triangle of s stored rowwise. *** c c *** parameter declarations *** c integer p double precision s(1), x(p), y(p) c dimension s(p*(p+1)/2) c c *** local variables *** c integer i, im1, j, k double precision xi c c *** no intrinsic functions *** c c *** external function *** c external dotprd double precision dotprd c c----------------------------------------------------------------------- c j = 1 do 10 i = 1, p y(i) = dotprd(i, s(j), x) j = j + i 10 continue c if (p .le. 1) go to 999 j = 1 do 40 i = 2, p xi = x(i) im1 = i - 1 j = j + 1 do 30 k = 1, im1 y(k) = y(k) + s(j)*xi j = j + 1 30 continue 40 continue c 999 return c *** last card of slvmul follows *** end c *** simple test program for smsno, sumsl, and humsl *** c c in these examples, n = 4 and f(x) = (1.0 + 0.5*(x1**t)*a*x1)**0.5, c where x1(i) = d1(i)*x(i) - i, **t denotes transpose, and a is a c matrix having fives on the main diagonal and ones everywhere else. c the scale vector d1 is passed to qdrtf, the subroutine that c evaluates f, as part of urparm. specifically, the matrix urp c declared below is passed for ufparm, and d1 is urp(*,1), the first c column of urp. this main program repeatedly minimizes f, starting c from x = 0, by calling smsno, sumsl, and humsl. we actually c use two different objective functions, since we change d1 after c the first call on sumsl. all runs but the last use d = d1. c c f(x) is minimized at x1 = 0 (a vector of zeros), i.e., at c x(i) = i/d1(i). c external deflt, humsl, imdcon, qdrtf, qdrtg, qdrtgh, smsno, 1 sumsl integer imdcon c c imdcon - supplies nout, the output unit number. c qdrtf - passed for calcf to smsno, sumsl, and humsl. c qdrtg - passed for calcg to sumsl. c qdrtgh - passed for calcgh to humsl. c integer i, iv(60), liv, lv, nout, uip(1) double precision d(4), urp(4,3), v(150), x(4) c c the length of v is dictated by humsl. c data liv/60/, lv/150/ c c initialize nout, d, d1, and x. c nout = imdcon(1) c do 10 i = 1, 4 d(i) = 1.d+0 urp(i,1) = d(i) x(i) = 0.d+0 10 continue write(nout,20) 20 format(/16h smsno on qdrtf ) c c before this first call, we set iv(1) to 0 so that all input c components of iv and v will be given default values. before c subsequent calls, we set iv(1) to 12 so that the old input values c of iv and v are used. c c qdrtf does not make use of ufparm. in calling smsno, we c arbitrarily pass qdrtf for ufparm to satifsy the calling sequence. c iv(1) = 0 call smsno(4, d, x, qdrtf, iv, liv, lv, v, uip, urp, qdrtf) c c we reinitialize x and minimize f again, this time using sumsl. c qdrtg, the subroutine passed for calcg, assumes that ufparm is qdrtf. c do 30 i = 1, 4 30 x(i) = 0.d+0 write(nout,40) 40 format(/16h sumsl on qdrtf ) c iv(1) = 12 call sumsl(4, d, x, qdrtf, qdrtg, iv, liv, lv, v, uip,urp,qdrtf) c c now we modify f by using a different choice of d1. we still use c d = d1, so the performance of sumsl should stay the same -- only d c and the final x and gradient should be affected. c do 50 i = 1, 4 x(i) = 0.d+0 d(i) = 1.d2 ** i urp(i,1) = d(i) 50 continue write(nout,40) c iv(1) = 12 call sumsl(4, d, x, qdrtf, qdrtg, iv, liv, lv, v, uip,urp,qdrtf) c c using the last choice of d and d1, we now use humsl to minimize f. c like qdrtg, qdrtgh assumes that ufparm is qdrtf. c do 60 i = 1, 4 60 x(i) = 0.d+0 write(nout,70) 70 format(/16h humsl on qdrtf ) c iv(1) = 12 call humsl(4, d, x, qdrtf, qdrtgh, iv, liv, lv, v, uip,urp,qdrtf) c c we repeat the last run with iv(dtype) = 1 and v(dinit) = 0.0, so c that humsl will determine d from the diagonal of the hessian. c this run also demonstrates the use of subroutine deflt and the c passing of nondefault parameters. (since the iv and v input c components still have their default values at his point, it is not c really necessary to call deflt. it is necessary to reset iv(1) to c 12, however, and deflt does this for us.) c write(nout,80) 80 format(/18h humsl updating d ) c do 90 i = 1, 4 90 x(i) = 0.d+0 c call deflt(2, iv, liv, lv, v) iv(16) = 1 v(38) = 0.d+0 call humsl(4, d, x, qdrtf, qdrtgh, iv, liv, lv, v, uip,urp,qdrtf) c 999 stop end c c*********************************************************************** c c q d r t f c c*********************************************************************** c subroutine qdrtf(n, x, nf, f, uip, urp, ufp) c c this routine evaluates the objective function f(x) described in the c main program above. it stores in urp(*,2) and urp(*,3) some c information useful in evaluating the gradient and hessian of f, and c it stores nf in uip(1) to identify the x corresponding to this c information. f(x) has the form f(x) = phi(q(x)), where q is a c quadratic form and phi(y) = y**0.5. the gradient of f is c g(x) = phiprm(q(x))*gq(x), where phiprm is the derivative of phi c and gq is the gradient of q. this routine stores phiprm(q(x)) in c urp(1,3) and gq(x) in urp(*,2). the hessian of f is c h(x) = phi2prm(q(x))*gq(x)*gq(x)**t + phiprm(q(x))*hq(x), where c phi2prm is the second derivative of phi, **t denotes transpose, c and hq is the hessian of q. this routine stores phi2prm(q(x)) in c urp(2,3). the subroutines qdrtg and qdrtgh given below would work c without change on any other choice of phi. qdrtg would also work c with any other differentiable function q. qdrtgh, on the other c hand, assumes that hq(x) is the matrix a described in the main c program above. c integer n, nf, uip(1) double precision x(n), f, urp(n,3) external ufp c/+ real float double precision dsqrt c/ integer i double precision dn, f2, t, t1 c c uip(1) = nf dn = n t = 0.d+0 do 10 i = 1, n urp(i,2) = urp(i,1)*x(i) - float(i) t = t + urp(i,2) 10 continue f2 = 0.d+0 do 20 i = 1, n t1 = dn*urp(i,2) + t f2 = f2 + t1*urp(i,2) urp(i,2) = urp(i,1) * t1 20 continue f2 = 1.d+0 + 0.5d+0 * f2 f = dsqrt(f2) urp(1,3) = 0.5d+0 / f urp(2,3) = -0.5d+0 / (f * f2) 999 return end c c*********************************************************************** c c q d r t g c c*********************************************************************** c subroutine qdrtg(n, x, nf, g, uip, urp, qdrtf) c c this routine evaluates the gradient of the objective function f(x) c described in the main program above. see the comments there and in c subroutine qdrtf above. c integer n, nf, uip(1) double precision x(n), g(n), urp(n,3) external qdrtf c integer i double precision f c if (nf .ne. uip(1)) call qdrtf(n, x, nf, f, uip, urp, qdrtf) do 10 i = 1, n 10 g(i) = urp(1,3) * urp(i,2) 999 return end c c*********************************************************************** c c q d r t g h c c*********************************************************************** c subroutine qdrtgh(n, x, nf, g, h, uip, urp, qdrtf) c c this routine evaluates the gradient and hessian of the objective c function f(x) described in the main program above. see the comments c there and in subroutine qdrtf above. note that the h returned is c the lower triangle of the hessian, stored row-wise. c integer n, nf, uip(1) double precision x(n), g(n), h(1), urp(n,3) c dimension h(n*(n+1)/2) external qdrtf c integer i, j, k double precision dn, f, t1, t2 c if (nf .ne. uip(1)) call qdrtf(n, x, nf, f, uip, urp, qdrtf) k = 0 dn = n do 20 i = 1, n g(i) = urp(1,3) * urp(i,2) t1 = urp(1,3) * urp(i,1) t2 = urp(i,2) * urp(2,3) do 10 j = 1, i k = k + 1 h(k) = t2*urp(j,2) + t1*urp(j,1) 10 continue h(k) = h(k) + dn*urp(i,1)*t1 20 continue 999 return end integer function imdcon(k) c integer k c c *** return integer machine-dependent constants *** c c *** k = 1 means return standard output unit number. *** c *** k = 2 means return alternate output unit number. *** c *** k = 3 means return input unit number. *** c (note -- k = 2, 3 are used only by test programs.) c c +++ port version follows... c external i1mach c integer i1mach c integer mdperm(3) c data mdperm(1)/2/, mdperm(2)/4/, mdperm(3)/1/ c imdcon = i1mach(mdperm(k)) c +++ end of port version +++ c c +++ non-port version follows... integer mdcon(3) data mdcon(1)/6/, mdcon(2)/8/, mdcon(3)/5/ imdcon = mdcon(k) c +++ end of non-port version +++ c 999 return c *** last card of imdcon follows *** end real function rmdcon(k) c c *** return machine dependent constants used by nl2sol *** c c +++ comments below contain data statements for various machines. +++ c +++ to convert to another machine, place a c in column 1 of the +++ c +++ data statement line(s) that correspond to the current machine +++ c +++ and remove the c from column 1 of the data statement line(s) +++ c +++ that correspond to the new machine. +++ c integer k c c *** the constant returned depends on k... c c *** k = 1... smallest pos. eta such that -eta exists. c *** k = 2... square root of eta. c *** k = 3... unit roundoff = smallest pos. no. machep such c *** that 1 + machep .gt. 1 .and. 1 - machep .lt. 1. c *** k = 4... square root of machep. c *** k = 5... square root of big (see k = 6). c *** k = 6... largest machine no. big such that -big exists. c real big, eta, machep c/+ real sqrt c/ c c +++ ibm 360, ibm 370, or xerox +++ c c data big/z7fffffff/, eta/z00100000/, machep/z3c100000/ c c +++ data general +++ c c data big/0.7237e+76/, eta/0.5398e-78/, machep/0.9537e-06/ c c +++ dec 11 +++ c c data big/1.7e+38/, eta/2.9388e-39/, machep/1.1921e-07/ c c +++ hp3000 +++ c c data big/1.1579e+77/, eta/8.6362e-78/, machep/2.3842e-07/ c c +++ honeywell +++ c c data big/o376777000000/, eta/o404400400000/, c 1 machep/o716400000000/ c c +++ dec10 +++ c c data big/"377777777777/, eta/"000400000021/, c 1 machep/"147400000000/ c c +++ burroughs +++ c c data big/o0777777777777777/, eta/o1771000000000000/, c 1 machep/o1301000000000000/ c c +++ control data +++ c c data big/37754000000000000000b/, eta/00024000000000000000b/, c 1 machep/16414000000000000000b/ c c +++ prime +++ c c data big/1.7e+38/, eta/1.47e-39/, machep/2.38419e-7/ c c +++ univac +++ c c data big/1.69e+38/, eta/5.9e-39/, machep/1.4901162e-8/ c c +++ vax +++ c c data big/1.7e+38/, eta/2.939e-39/, machep/5.9604645e-08/ c c +++ cray 1 +++ c data big/577767777777777777776b/, eta/200004000000000000000b/, 1 machep/377224000000000000000b/ c c +++ port library -- requires more than just a data statement... +++ c c external r1mach c real r1mach, zero c data big/0.e+0/, eta/0.e+0/, machep/0.e+0/, zero/0.e+0/ c if (big .gt. zero) go to 1 c big = r1mach(2) c eta = r1mach(1) c machep = r1mach(4) c1 continue c c +++ end of port +++ c c------------------------------- body -------------------------------- c go to (10, 20, 30, 40, 50, 60), k c 10 rmdcon = eta go to 999 c 20 rmdcon = sqrt(256.e+0*eta)/16.e+0 go to 999 c 30 rmdcon = machep go to 999 c 40 rmdcon = sqrt(machep) go to 999 c 50 rmdcon = sqrt(big/256.e+0)*16.e+0 go to 999 c 60 rmdcon = big c 999 return c *** last card of rmdcon follows *** end 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) real 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 iv 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 liv 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 lv 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 iv 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 real f c c *** subscripts for iv *** c integer nextv, nfcall, nfgcal, g, toobig, vneed c c/6 data nextv/47/, nfcall/6/, nfgcal/7/, g/28/, toobig/2/, vneed/4/ c/7 c 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 smsno(n, d, x, calcf, iv, liv, lv, v, 1 uiparm, urparm, ufparm) c c *** minimize general unconstrained objective function using c *** finite-difference gradients and secant hessian approximations. c integer n, liv, lv integer iv(liv), uiparm(1) real d(n), x(n), v(lv), urparm(1) c dimension v(77 + n*(n+17)/2), uiparm(*), urparm(*) external calcf, ufparm c c *** purpose *** c c this routine interacts with subroutine snoit 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 *** parameters *** c c the parameters for smsno are the same as those for sumsl c (which see), except that calcg is omitted. instead of calling c calcg to obtain the gradient of the objective function at x, c smsno calls sgrad2, which computes an approximation to the c gradient by finite (forward and central) differences using the c method of ref. 1. the following input component is of interest c in this regard (and is not described in sumsl). c c v(eta0)..... v(42) is an estimated bound on the relative error in the c objective function value computed by calcf... c (true value) = (computed value) * (1 + e), c where abs(e) .le. v(eta0). default = machep * 10**3, c where machep is the unit roundoff. c c the output values iv(nfcall) and iv(ngcall) have different c meanings for smsno than for sumsl... c c iv(nfcall)... iv(6) is the number of calls so far made on calcf (i.e., c function evaluations) excluding those made only for c computing gradients. the input value iv(mxfcal) is a c limit on iv(nfcall). c iv(ngcall)... iv(30) is the number of function evaluations made only c for computing gradients. the total number of function c evaluations is thus iv(nfcall) + iv(ngcall). c c *** reference *** c c 1. stewart, g.w. (1967), a modification of davidon*s minimization c method to accept difference approximations of derivatives, c j. assoc. comput. mach. 14, pp. 72-83. c. c *** general *** c c coded by david m. gay (winter 1980). revised sept. 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 snoit c c snoit.... oversees computation of finite-difference gradient and c calls sumit to carry out sumsl algorithm. c integer nf real fx c c *** subscripts for iv *** c integer nfcall, toobig c c/6 data nfcall/6/, toobig/2/ c/7 c parameter (nfcall=6, toobig=2) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c 10 call snoit(d, fx, iv, liv, lv, n, v, x) if (iv(1) .gt. 2) go to 999 c c *** compute function *** c nf = iv(nfcall) call calcf(n, x, nf, fx, uiparm, urparm, ufparm) if (nf .le. 0) iv(toobig) = 1 go to 10 c c 999 return c *** last card of smsno 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) real 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 real t c c *** constants *** c real 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 real 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 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 data 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/7 c parameter (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/ c c *** v subscript values *** c c/6 data 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/7 c parameter (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/ c c/6 data half/0.5e+0/, negone/-1.e+0/, one/1.e+0/, onep2/1.2e+0/, 1 zero/0.e+0/ c/7 c parameter (half=0.5e+0, negone=-1.e+0, one=1.e+0, onep2=1.2e+0, c 1 zero=0.e+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c 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 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 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 snoit(d, fx, iv, liv, lv, n, v, x) c c *** iteration driver for smsno... c *** minimize general unconstrained objective function using c *** finite-difference gradients and secant hessian approximations. c integer liv, lv, n integer iv(liv) real d(n), fx, x(n), v(lv) c dimension v(77 + n*(n+17)/2) 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 fx = f(x) computed by the caller. (often c the x* found is a local minimizer rather than a global one.) c c *** parameters *** c c the parameters for snoit are the same as those for sumsl c (which see), except that calcf, calcg, uiparm, urparm, and ufparm c are omitted, and a parameter fx for the objective function c value at x is added. instead of calling calcg to obtain the c gradient of the objective function at x, snoit calls sgrad2, c which computes an approximation to the gradient by finite c (forward and central) differences using the method of ref. 1. c the following input component is of interest in this regard c (and is not described in sumsl). c c v(eta0)..... v(42) is an estimated bound on the relative error in the c objective function value computed by calcf... c (true value) = (computed value) * (1 + e), c where abs(e) .le. v(eta0). default = machep * 10**3, c where machep is the unit roundoff. c c the output values iv(nfcall) and iv(ngcall) have different c meanings for smsno than for sumsl... c c iv(nfcall)... iv(6) is the number of calls so far made on calcf (i.e., c function evaluations) excluding those made only for c computing gradients. the input value iv(mxfcal) is a c limit on iv(nfcall). c iv(ngcall)... iv(30) is the number of function evaluations made only c for computing gradients. the total number of function c evaluations is thus iv(nfcall) + iv(ngcall). c c *** references *** c c 1. stewart, g.w. (1967), a modification of davidon*s minimization c method to accept difference approximations of derivatives, c j. assoc. comput. mach. 14, pp. 72-83. c. c *** general *** c c coded by david m. gay (august 1982). c c---------------------------- declarations --------------------------- c external deflt, dotprd, sgrad2, sumit, vscopy real dotprd c c deflt.... supplies default parameter values. c dotprd... returns inner product of two vectors. c sgrad2... computes finite-difference gradient approximation. c sumit.... reverse-communication routine that does sumsl algorithm. c vscopy... sets all elements of a vector to a scalar. c integer alpha, g1, i, iv1, j, k, w real one, zero c c *** subscripts for iv *** c integer dtype, eta0, f, g, lmat, nextv, nfcall, nfgcal, ngcall, 1 niter, sgirc, toobig, vneed c c/6 data dtype/16/, eta0/42/, f/10/, g/28/, lmat/42/, nextv/47/, 1 nfcall/6/, nfgcal/7/, ngcall/30/, niter/31/, sgirc/57/, 2 toobig/2/, vneed/4/ c/7 c parameter (dtype=16, eta0=42, f=10, g=28, lmat=42, nextv=47, c 1 nfcall=6, nfgcal=7, ngcall=30, niter=31, sgirc=57, c 2 toobig=2, vneed=4) c/ c/6 data one/1.e+0/, zero/0.e+0/ c/7 c parameter (one=1.e+0, zero=0.e+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c iv1 = iv(1) if (iv1 .eq. 1) go to 10 if (iv1 .eq. 2) go to 50 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) + 2*n + 6 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, fx, v(g1), iv, liv, lv, n, v, x) if (iv(1) - 2) 999, 30, 70 c c *** compute gradient *** c 30 if (iv(niter) .eq. 0) call vscopy(n, v(g1), zero) j = iv(lmat) k = g1 - n do 40 i = 1, n v(k) = dotprd(i, v(j), v(j)) k = k + 1 j = j + i 40 continue c *** undo increment of iv(ngcall) done by sumit *** iv(ngcall) = iv(ngcall) - 1 c *** store return code from sgrad2 in iv(sgirc) *** iv(sgirc) = 0 c *** x may have been restored, so copy back fx... *** fx = v(f) go to 60 c c *** gradient loop *** c 50 if (iv(toobig) .eq. 0) go to 60 iv(nfgcal) = 0 go to 10 c 60 g1 = iv(g) alpha = g1 - n w = alpha - 6 call sgrad2(v(alpha), d, v(eta0), fx, v(g1), iv(sgirc), n, v(w),x) if (iv(sgirc) .eq. 0) go to 10 iv(ngcall) = iv(ngcall) + 1 go to 999 c 70 if (iv(1) .ne. 14) go to 999 c c *** storage allocation *** c iv(g) = iv(nextv) + n + 6 iv(nextv) = iv(g) + n if (iv1 .ne. 13) go to 10 c 999 return c *** last card of snoit follows *** end subroutine dbdog(dig, lv, n, nwtstp, step, v) c c *** compute double dogleg step *** c c *** parameter declarations *** c integer lv, n real 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 real dotprd, v2norm c c dotprd... returns inner product of two vectors. c v2norm... returns 2-norm of a vector. c c *** intrinsic functions *** c/+ real sqrt c/ c-------------------------- local variables -------------------------- c integer i real cfact, cnorm, ctrnwt, ghinvg, femnsq, gnorm, 1 nwtnrm, relax, rlambd, t, t1, t2 real 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 data half/0.5e+0/, one/1.e+0/, two/2.e+0/, zero/0.e+0/ c/7 c parameter (half=0.5e+0, one=1.e+0, two=2.e+0, zero=0.e+0) c/ c c/6 data 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/7 c parameter (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/ 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 + sqrt(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 real x(n), l(1), y(n) c dimension l(n*(n+1)/2) integer i, ij, i0, j real yi, zero c/6 data zero/0.e+0/ c/7 c parameter (zero=0.e+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 real beta(n), gamma(n), l(1), lambda(n), lplus(1), 1 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/+ real sqrt c/ c-------------------------- local variables -------------------------- c integer i, ij, j, jj, jp1, k, nm1, np1 real a, b, bj, eta, gj, lj, lij, ljj, nu, s, theta, 1 wj, zj real one, zero c c *** data initializations *** c c/6 data one/1.e+0/, zero/0.e+0/ c/7 c parameter (one=1.e+0, zero=0.e+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 = sqrt(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 real x(n), l(1), y(n) c dimension l(n*(n+1)/2) integer i, ii, ij, i0, j, np1 real t, zero c/6 data zero/0.e+0/ c/7 c parameter (zero=0.e+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 sgrad2 (alpha, d, eta0, fx, g, irc, n, w, x) c c *** compute finite difference gradient by stweart*s scheme *** c c *** parameters *** c integer irc, n real alpha(n), d(n), eta0, fx, g(n), w(6), x(n) c c....................................................................... c c *** purpose *** c c this subroutine uses an embellished form of the finite-differ- c ence scheme proposed by stewart (ref. 1) to approximate the c gradient of the function f(x), whose values are supplied by c reverse communication. c c *** parameter description *** c c alpha in (approximate) diagonal elements of the hessian of f(x). c d in scale vector such that d(i)*x(i), i = 1,...,n, are in c comparable units. c eta0 in estimated bound on relative error in the function value... c (true value) = (computed value)*(1+e), where c abs(e) .le. eta0. c fx i/o on input, fx must be the computed value of f(x). on c output with irc = 0, fx has been restored to its original c value, the one it had when sgrad2 was last called with c irc = 0. c g i/o on input with irc = 0, g should contain an approximation c to the gradient of f near x, e.g., the gradient at the c previous iterate. when sgrad2 returns with irc = 0, g is c the desired finite-difference approximation to the c gradient at x. c irc i/o input/return code... before the very first call on sgrad2, c the caller must set irc to 0. whenever sgrad2 returns a c nonzero value for irc, it has perturbed some component of c x... the caller should evaluate f(x) and call sgrad2 c again with fx = f(x). c n in the number of variables (components of x) on which f c depends. c x i/o on input with irc = 0, x is the point at which the c gradient of f is desired. on output with irc nonzero, x c is the point at which f should be evaluated. on output c with irc = 0, x has been restored to its original value c (the one it had when sgrad2 was last called with irc = 0) c and g contains the desired gradient approximation. c w i/o work vector of length 6 in which sgrad2 saves certain c quantities while the caller is evaluating f(x) at a c perturbed x. c c *** application and usage restrictions *** c c this routine is intended for use with quasi-newton routines c for unconstrained minimization (in which case alpha comes from c the diagonal of the quasi-newton hessian approximation). c c *** algorithm notes *** c c this code departs from the scheme proposed by stewart (ref. 1) c in its guarding against overly large or small step sizes and its c handling of special cases (such as zero components of alpha or g). c c *** references *** c c 1. stewart, g.w. (1967), a modification of davidon*s minimization c method to accept difference approximations of derivatives, c j. assoc. comput. mach. 14, pp. 72-83. c c *** history *** c c designed and coded by david m. gay (summer 1977/summer 1980). c c *** general *** c c this routine was prepared in connection with work supported by c the national science foundation under grants mcs76-00324 and c mcs-7906671. c c....................................................................... c c ***** external function ***** c external rmdcon real rmdcon c rmdcon... returns machine-dependent constants. c c ***** intrinsic functions ***** c/+ integer iabs real abs, amax1, sqrt c/ c ***** local variables ***** c integer fh, fx0, hsave, i, xisave real aai, afx, afxeta, agi, alphai, axi, axibar, 1 discon, eta, gi, h, hmin real c2000, four, hmax0, hmin0, h0, machep, one, p002, 1 three, two, zero c c/6 data c2000/2.0e+3/, four/4.0e+0/, hmax0/0.02e+0/, hmin0/5.0e+1/, 1 one/1.0e+0/, p002/0.002e+0/, three/3.0e+0/, 2 two/2.0e+0/, zero/0.0e+0/ c/7 c parameter (c2000=2.0e+3, four=4.0e+0, hmax0=0.02e+0, hmin0=5.0e+1, c 1 one=1.0e+0, p002=0.002e+0, three=3.0e+0, c 2 two=2.0e+0, zero=0.0e+0) c/ c/6 data fh/3/, fx0/4/, hsave/5/, xisave/6/ c/7 c parameter (fh=3, fx0=4, hsave=5, xisave=6) c/ c c--------------------------------- body ------------------------------ c if (irc) 140, 100, 210 c c *** fresh start -- get machine-dependent constants *** c c store machep in w(1) and h0 in w(2), where machep is the unit c roundoff (the smallest positive number such that c 1 + machep .gt. 1 and 1 - machep .lt. 1), and h0 is the c square-root of machep. c 100 w(1) = rmdcon(3) w(2) = sqrt(w(1)) c w(fx0) = fx c c *** increment i and start computing g(i) *** c 110 i = iabs(irc) + 1 if (i .gt. n) go to 300 irc = i afx = abs(w(fx0)) machep = w(1) h0 = w(2) hmin = hmin0 * machep w(xisave) = x(i) axi = abs(x(i)) axibar = amax1(axi, one/d(i)) gi = g(i) agi = abs(gi) eta = abs(eta0) if (afx .gt. zero) eta = amax1(eta, agi*axi*machep/afx) alphai = alpha(i) if (alphai .eq. zero) go to 170 if (gi .eq. zero .or. fx .eq. zero) go to 180 afxeta = afx*eta aai = abs(alphai) c c *** compute h = stewart*s forward-difference step size. c if (gi**2 .le. afxeta*aai) go to 120 h = two* sqrt(afxeta/aai) h = h*(one - aai*h/(three*aai*h + four*agi)) go to 130 120 h = two*(afxeta*agi/(aai**2))**(one/three) h = h*(one - two*agi/(three*aai*h + four*agi)) c c *** ensure that h is not insignificantly small *** c 130 h = amax1(h, hmin*axibar) c c *** use forward difference if bound on truncation error is at c *** most 10**-3. c if (aai*h .le. p002*agi) go to 160 c c *** compute h = stewart*s step for central difference. c discon = c2000*afxeta h = discon/(agi + sqrt(gi**2 + aai*discon)) c c *** ensure that h is neither too small nor too big *** c h = amax1(h, hmin*axibar) if (h .ge. hmax0*axibar) h = axibar * h0**(two/three) c c *** compute central difference *** c irc = -i go to 200 c 140 h = -w(hsave) i = iabs(irc) if (h .gt. zero) go to 150 w(fh) = fx go to 200 c 150 g(i) = (w(fh) - fx) / (two * h) x(i) = w(xisave) go to 110 c c *** compute forward differences in various cases *** c 160 if (h .ge. hmax0*axibar) h = h0 * axibar if (alphai*gi .lt. zero) h = -h go to 200 170 h = axibar go to 200 180 h = h0 * axibar c 200 x(i) = w(xisave) + h w(hsave) = h go to 999 c c *** compute actual forward difference *** c 210 g(irc) = (fx - w(fx0)) / w(hsave) x(irc) = w(xisave) go to 110 c c *** restore fx and indicate that g has been computed *** c 300 fx = w(fx0) irc = 0 c 999 return c *** last card of sgrad2 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 real 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 real l(1), 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 real 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/+ real sqrt c/ c-------------------------- local variables -------------------------- c integer i real cs, cy, eps, epsrt, one, shs, ys, theta c c *** data initializations *** c c/6 data eps/0.1e+0/, one/1.e+0/ c/7 c parameter (eps=0.1e+0, one=1.e+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 = sqrt(eps) cy = theta / (shs * epsrt) cs = (one + (theta-one)/epsrt) / shs go to 20 10 cy = one / ( sqrt(ys) * sqrt(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 subroutine assst(iv, liv, lv, v) c c *** assess candidate step (***sol version 2.3) *** c integer liv, lv integer iv(liv) real v(lv) c c *** purpose *** c c this subroutine is called by an unconstrained minimization c routine to assess the next candidate step. it may recommend one c of several courses of action, such as accepting the step, recom- c puting it using the same or a new quadratic model, or halting due c to convergence or false convergence. see the return code listing c below. c c-------------------------- parameter usage -------------------------- c c iv (i/o) integer parameter and scratch vector -- see description c below of iv values referenced. c liv (in) length of iv array. c lv (in) length of v array. c v (i/o) real parameter and scratch vector -- see description c below of v values referenced. c c *** iv values referenced *** c c iv(irc) (i/o) on input for the first step tried in a new iteration, c iv(irc) should be set to 3 or 4 (the value to which it is c set when step is definitely to be accepted). on input c after step has been recomputed, iv(irc) should be c unchanged since the previous return of assst. c on output, iv(irc) is a return code having one of the c following values... c 1 = switch models or try smaller step. c 2 = switch models or accept step. c 3 = accept step and determine v(radfac) by gradient c tests. c 4 = accept step, v(radfac) has been determined. c 5 = recompute step (using the same model). c 6 = recompute step with radius = v(lmaxs) but do not c evaulate the objective function. c 7 = x-convergence (see v(xctol)). c 8 = relative function convergence (see v(rfctol)). c 9 = both x- and relative function convergence. c 10 = absolute function convergence (see v(afctol)). c 11 = singular convergence (see v(lmaxs)). c 12 = false convergence (see v(xftol)). c 13 = iv(irc) was out of range on input. c return code i has precdence over i+1 for i = 9, 10, 11. c iv(mlstgd) (i/o) saved value of iv(model). c iv(model) (i/o) on input, iv(model) should be an integer identifying c the current quadratic model of the objective function. c if a previous step yielded a better function reduction, c then iv(model) will be set to iv(mlstgd) on output. c iv(nfcall) (in) invocation count for the objective function. c iv(nfgcal) (i/o) value of iv(nfcall) at step that gave the biggest c function reduction this iteration. iv(nfgcal) remains c unchanged until a function reduction is obtained. c iv(radinc) (i/o) the number of radius increases (or minus the number c of decreases) so far this iteration. c iv(restor) (out) set to 1 if v(f) has been restored and x should be c restored to its initial value, to 2 if x should be saved, c to 3 if x should be restored from the saved value, and to c 0 otherwise. c iv(stage) (i/o) count of the number of models tried so far in the c current iteration. c iv(stglim) (in) maximum number of models to consider. c iv(switch) (out) set to 0 unless a new model is being tried and it c gives a smaller function value than the previous model, c in which case assst sets iv(switch) = 1. c iv(toobig) (in) is nonzero if step was too big (e.g. if it caused c overflow). c iv(xirc) (i/o) value that iv(irc) would have in the absence of c convergence, false convergence, and oversized steps. c c *** v values referenced *** c c v(afctol) (in) absolute function convergence tolerance. if the c absolute value of the current function value v(f) is less c than v(afctol), then assst returns with iv(irc) = 10. c v(decfac) (in) factor by which to decrease radius when iv(toobig) is c nonzero. c v(dstnrm) (in) the 2-norm of d*step. c v(dstsav) (i/o) value of v(dstnrm) on saved step. c v(dst0) (in) the 2-norm of d times the newton step (when defined, c i.e., for v(nreduc) .ge. 0). c v(f) (i/o) on both input and output, v(f) is the objective func- c tion value at x. if x is restored to a previous value, c then v(f) is restored to the corresponding value. c v(fdif) (out) the function reduction v(f0) - v(f) (for the output c value of v(f) if an earlier step gave a bigger function c decrease, and for the input value of v(f) otherwise). c v(flstgd) (i/o) saved value of v(f). c v(f0) (in) objective function value at start of iteration. c v(gtslst) (i/o) value of v(gtstep) on saved step. c v(gtstep) (in) inner product between step and gradient. c v(incfac) (in) minimum factor by which to increase radius. c v(lmaxs) (in) maximum reasonable step size (and initial step bound). c if the actual function decrease is no more than twice c what was predicted, if a return with iv(irc) = 7, 8, 9, c or 10 does not occur, if v(dstnrm) .gt. v(lmaxs), and if c v(preduc) .le. v(sctol) * abs(v(f0)), then assst re- c turns with iv(irc) = 11. if so doing appears worthwhile, c then assst repeats this test with v(preduc) computed for c a step of length v(lmaxs) (by a return with iv(irc) = 6). c v(nreduc) (i/o) function reduction predicted by quadratic model for c newton step. if assst is called with iv(irc) = 6, i.e., c if v(preduc) has been computed with radius = v(lmaxs) for c use in the singular convervence test, then v(nreduc) is c set to -v(preduc) before the latter is restored. c v(plstgd) (i/o) value of v(preduc) on saved step. c v(preduc) (i/o) function reduction predicted by quadratic model for c current step. c v(radfac) (out) factor to be used in determining the new radius, c which should be v(radfac)*dst, where dst is either the c output value of v(dstnrm) or the 2-norm of c diag(newd)*step for the output value of step and the c updated version, newd, of the scale vector d. for c iv(irc) = 3, v(radfac) = 1.0 is returned. c v(rdfcmn) (in) minimum value for v(radfac) in terms of the input c value of v(dstnrm) -- suggested value = 0.1. c v(rdfcmx) (in) maximum value for v(radfac) -- suggested value = 4.0. c v(reldx) (in) scaled relative change in x caused by step, computed c (e.g.) by function reldst as c max (d(i)*abs(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 v(rfctol) (in) relative function convergence tolerance. if the c actual function reduction is at most twice what was pre- c dicted and v(nreduc) .le. v(rfctol)*abs(v(f0)), then c assst returns with iv(irc) = 8 or 9. c v(stppar) (in) marquardt parameter -- 0 means full newton step. c v(tuner1) (in) tuning constant used to decide if the function c reduction was much less than expected. suggested c value = 0.1. c v(tuner2) (in) tuning constant used to decide if the function c reduction was large enough to accept step. suggested c value = 10**-4. c v(tuner3) (in) tuning constant used to decide if the radius c should be increased. suggested value = 0.75. c v(xctol) (in) x-convergence criterion. if step is a newton step c (v(stppar) = 0) having v(reldx) .le. v(xctol) and giving c at most twice the predicted function decrease, then c assst returns iv(irc) = 7 or 9. c v(xftol) (in) false convergence tolerance. if step gave no or only c a small function decrease and v(reldx) .le. v(xftol), c then assst returns with iv(irc) = 12. c c------------------------------- notes ------------------------------- c c *** application and usage restrictions *** c c this routine is called as part of the nl2sol (nonlinear c least-squares) package. it may be used in any unconstrained c minimization solver that uses dogleg, goldfeld-quandt-trotter, c or levenberg-marquardt steps. c c *** algorithm notes *** c c see (1) for further discussion of the assessing and model c switching strategies. while nl2sol considers only two models, c assst is designed to handle any number of models. c c *** usage notes *** c c on the first call of an iteration, only the i/o variables c step, x, iv(irc), iv(model), v(f), v(dstnrm), v(gtstep), and c v(preduc) need have been initialized. between calls, no i/o c values execpt step, x, iv(model), v(f) and the stopping toler- c ances should be changed. c after a return for convergence or false convergence, one can c change the stopping tolerances and call assst again, in which c case the stopping tests will be repeated. c c *** references *** c c (1) dennis, j.e., jr., gay, d.m., and welsch, r.e. (1981), c an adaptive nonlinear least-squares algorithm, c acm trans. math. software, vol. 7, no. 3. c c (2) powell, m.j.d. (1970) a fortran subroutine for solving c systems of nonlinear algebraic equations, in numerical c methods for nonlinear algebraic equations, edited by c p. rabinowitz, gordon and breach, london. c c *** history *** c c john dennis designed much of this routine, starting with c ideas in (2). roy welsch suggested the model switching strategy. c david gay and stephen peters cast this subroutine into a more c portable form (winter 1977), and david gay cast it into its c present form (fall 1978). c c *** general *** c c this subroutine was written in connection with research c supported by the national science foundation under grants c mcs-7600324, dcr75-10143, 76-14311dss, mcs76-11989, and c mcs-7906671. c c------------------------ external quantities ------------------------ c c *** no external functions and subroutines *** c c *** intrinsic functions *** c/+ real abs, amax1 c/ c *** no common blocks *** c c-------------------------- local variables -------------------------- c logical goodx integer i, nfc real emax, emaxs, gts, rfac1, xmax real half, one, onep2, two, zero c c *** subscripts for iv and v *** c integer afctol, decfac, dstnrm, dstsav, dst0, f, fdif, flstgd, f0, 1 gtslst, gtstep, incfac, irc, lmaxs, mlstgd, model, nfcall, 2 nfgcal, nreduc, plstgd, preduc, radfac, radinc, rdfcmn, 3 rdfcmx, reldx, restor, rfctol, sctol, stage, stglim, 4 stppar, switch, toobig, tuner1, tuner2, tuner3, xctol, 5 xftol, xirc c c *** data initializations *** c c/6 data half/0.5e+0/, one/1.e+0/, onep2/1.2e+0/, two/2.e+0/, 1 zero/0.e+0/ c/7 c parameter (half=0.5e+0, one=1.e+0, onep2=1.2e+0, two=2.e+0, c 1 zero=0.e+0) c/ c c/6 data irc/29/, mlstgd/32/, model/5/, nfcall/6/, nfgcal/7/, 1 radinc/8/, restor/9/, stage/10/, stglim/11/, switch/12/, 2 toobig/2/, xirc/13/ c/7 c parameter (irc=29, mlstgd=32, model=5, nfcall=6, nfgcal=7, c 1 radinc=8, restor=9, stage=10, stglim=11, switch=12, c 2 toobig=2, xirc=13) c/ c/6 data afctol/31/, decfac/22/, dstnrm/2/, dst0/3/, dstsav/18/, 1 f/10/, fdif/11/, flstgd/12/, f0/13/, gtslst/14/, gtstep/4/, 2 incfac/23/, lmaxs/36/, nreduc/6/, plstgd/15/, preduc/7/, 3 radfac/16/, rdfcmn/24/, rdfcmx/25/, reldx/17/, rfctol/32/, 4 sctol/37/, stppar/5/, tuner1/26/, tuner2/27/, tuner3/28/, 5 xctol/33/, xftol/34/ c/7 c parameter (afctol=31, decfac=22, dstnrm=2, dst0=3, dstsav=18, c 1 f=10, fdif=11, flstgd=12, f0=13, gtslst=14, gtstep=4, c 2 incfac=23, lmaxs=36, nreduc=6, plstgd=15, preduc=7, c 3 radfac=16, rdfcmn=24, rdfcmx=25, reldx=17, rfctol=32, c 4 sctol=37, stppar=5, tuner1=26, tuner2=27, tuner3=28, c 5 xctol=33, xftol=34) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c nfc = iv(nfcall) iv(switch) = 0 iv(restor) = 0 rfac1 = one goodx = .true. i = iv(irc) if (i .ge. 1 .and. i .le. 12) 1 go to (20,30,10,10,40,280,220,220,220,220,220,170), i iv(irc) = 13 go to 999 c c *** initialize for new iteration *** c 10 iv(stage) = 1 iv(radinc) = 0 v(flstgd) = v(f0) if (iv(toobig) .eq. 0) go to 110 iv(stage) = -1 iv(xirc) = i go to 60 c c *** step was recomputed with new model or smaller radius *** c *** first decide which *** c 20 if (iv(model) .ne. iv(mlstgd)) go to 30 c *** old model retained, smaller radius tried *** c *** do not consider any more new models this iteration *** iv(stage) = iv(stglim) iv(radinc) = -1 go to 110 c c *** a new model is being tried. decide whether to keep it. *** c 30 iv(stage) = iv(stage) + 1 c c *** now we add the possibiltiy that step was recomputed with *** c *** the same model, perhaps because of an oversized step. *** c 40 if (iv(stage) .gt. 0) go to 50 c c *** step was recomputed because it was too big. *** c if (iv(toobig) .ne. 0) go to 60 c c *** restore iv(stage) and pick up where we left off. *** c iv(stage) = -iv(stage) i = iv(xirc) go to (20, 30, 110, 110, 70), i c 50 if (iv(toobig) .eq. 0) go to 70 c c *** handle oversize step *** c if (iv(radinc) .gt. 0) go to 80 iv(stage) = -iv(stage) iv(xirc) = iv(irc) c 60 v(radfac) = v(decfac) iv(radinc) = iv(radinc) - 1 iv(irc) = 5 iv(restor) = 1 go to 999 c 70 if (v(f) .lt. v(flstgd)) go to 110 c c *** the new step is a loser. restore old model. *** c if (iv(model) .eq. iv(mlstgd)) go to 80 iv(model) = iv(mlstgd) iv(switch) = 1 c c *** restore step, etc. only if a previous step decreased v(f). c 80 if (v(flstgd) .ge. v(f0)) go to 110 iv(restor) = 1 v(f) = v(flstgd) v(preduc) = v(plstgd) v(gtstep) = v(gtslst) if (iv(switch) .eq. 0) rfac1 = v(dstnrm) / v(dstsav) v(dstnrm) = v(dstsav) nfc = iv(nfgcal) goodx = .false. c 110 v(fdif) = v(f0) - v(f) if (v(fdif) .gt. v(tuner2) * v(preduc)) go to 140 if(iv(radinc).gt.0) go to 140 c c *** no (or only a trivial) function decrease c *** -- so try new model or smaller radius c if (v(f) .lt. v(f0)) go to 120 iv(mlstgd) = iv(model) v(flstgd) = v(f) v(f) = v(f0) iv(restor) = 1 go to 130 120 iv(nfgcal) = nfc 130 iv(irc) = 1 if (iv(stage) .lt. iv(stglim)) go to 160 iv(irc) = 5 iv(radinc) = iv(radinc) - 1 go to 160 c c *** nontrivial function decrease achieved *** c 140 iv(nfgcal) = nfc rfac1 = one v(dstsav) = v(dstnrm) if (v(fdif) .gt. v(preduc)*v(tuner1)) go to 190 c c *** decrease was much less than predicted -- either change models c *** or accept step with decreased radius. c if (iv(stage) .ge. iv(stglim)) go to 150 c *** consider switching models *** iv(irc) = 2 go to 160 c c *** accept step with decreased radius *** c 150 iv(irc) = 4 c c *** set v(radfac) to fletcher*s decrease factor *** c 160 iv(xirc) = iv(irc) emax = v(gtstep) + v(fdif) v(radfac) = half * rfac1 if (emax .lt. v(gtstep)) v(radfac) = rfac1 * amax1(v(rdfcmn), 1 half * v(gtstep)/emax) c c *** do false convergence test *** c 170 if (v(reldx) .le. v(xftol)) go to 180 iv(irc) = iv(xirc) if (v(f) .lt. v(f0)) go to 200 go to 230 c 180 iv(irc) = 12 go to 240 c c *** handle good function decrease *** c 190 if (v(fdif) .lt. (-v(tuner3) * v(gtstep))) go to 210 c c *** increasing radius looks worthwhile. see if we just c *** recomputed step with a decreased radius or restored step c *** after recomputing it with a larger radius. c if (iv(radinc) .lt. 0) go to 210 if (iv(restor) .eq. 1) go to 210 c c *** we did not. try a longer step unless this was a newton c *** step. c v(radfac) = v(rdfcmx) gts = v(gtstep) if (v(fdif) .lt. (half/v(radfac) - one) * gts) 1 v(radfac) = amax1(v(incfac), half*gts/(gts + v(fdif))) iv(irc) = 4 if (v(stppar) .eq. zero) go to 230 if (v(dst0) .ge. zero .and. (v(dst0) .lt. two*v(dstnrm) 1 .or. v(nreduc) .lt. onep2*v(fdif))) go to 230 c *** step was not a newton step. recompute it with c *** a larger radius. iv(irc) = 5 iv(radinc) = iv(radinc) + 1 c c *** save values corresponding to good step *** c 200 v(flstgd) = v(f) iv(mlstgd) = iv(model) if (iv(restor) .ne. 1) iv(restor) = 2 v(dstsav) = v(dstnrm) iv(nfgcal) = nfc v(plstgd) = v(preduc) v(gtslst) = v(gtstep) go to 230 c c *** accept step with radius unchanged *** c 210 v(radfac) = one iv(irc) = 3 go to 230 c c *** come here for a restart after convergence *** c 220 iv(irc) = iv(xirc) if (v(dstsav) .ge. zero) go to 240 iv(irc) = 12 go to 240 c c *** perform convergence tests *** c 230 iv(xirc) = iv(irc) 240 if (iv(restor) .eq. 1 .and. v(flstgd) .lt. v(f0)) iv(restor) = 3 if ( abs(v(f)) .lt. v(afctol)) iv(irc) = 10 if (half * v(fdif) .gt. v(preduc)) go to 999 emax = v(rfctol) * abs(v(f0)) emaxs = v(sctol) * abs(v(f0)) if (v(dstnrm) .gt. v(lmaxs) .and. v(preduc) .le. emaxs) 1 iv(irc) = 11 if (v(dst0) .lt. zero) go to 250 i = 0 if ((v(nreduc) .gt. zero .and. v(nreduc) .le. emax) .or. 1 (v(nreduc) .eq. zero. and. v(preduc) .eq. zero)) i = 2 if (v(stppar) .eq. zero .and. v(reldx) .le. v(xctol) 1 .and. goodx) i = i + 1 if (i .gt. 0) iv(irc) = i + 6 c c *** consider recomputing step of length v(lmaxs) for singular c *** convergence test. c 250 if (iv(irc) .gt. 5 .and. iv(irc) .ne. 12) go to 999 if (v(dstnrm) .gt. v(lmaxs)) go to 260 if (v(preduc) .ge. emaxs) go to 999 if (v(dst0) .le. zero) go to 270 if (half * v(dst0) .le. v(lmaxs)) go to 999 go to 270 260 if (half * v(dstnrm) .le. v(lmaxs)) go to 999 xmax = v(lmaxs) / v(dstnrm) if (xmax * (two - xmax) * v(preduc) .ge. emaxs) go to 999 270 if (v(nreduc) .lt. zero) go to 290 c c *** recompute v(preduc) for use in singular convergence test *** c v(gtslst) = v(gtstep) v(dstsav) = v(dstnrm) if (iv(irc) .eq. 12) v(dstsav) = -v(dstsav) v(plstgd) = v(preduc) i = iv(restor) iv(restor) = 2 if (i .eq. 3) iv(restor) = 0 iv(irc) = 6 go to 999 c c *** perform singular convergence test with recomputed v(preduc) *** c 280 v(gtstep) = v(gtslst) v(dstnrm) = abs(v(dstsav)) iv(irc) = iv(xirc) if (v(dstsav) .le. zero) iv(irc) = 12 v(nreduc) = -v(preduc) v(preduc) = v(plstgd) iv(restor) = 3 290 if (-v(nreduc) .le. v(rfctol) * abs(v(f0))) iv(irc) = 11 c 999 return c c *** last card of assst follows *** end subroutine deflt(alg, iv, liv, lv, v) c c *** supply ***sol (version 2.3) default values to iv and v *** c c *** alg = 1 means regression constants. c *** alg = 2 means general unconstrained optimization constants. c integer liv, lv integer alg, iv(liv) real v(lv) c external imdcon, vdflt integer imdcon c imdcon... returns machine-dependent integer constants. c vdflt.... provides default values to v. c integer miv, mv integer miniv(2), minv(2) c c *** subscripts for iv *** c integer algsav, covprt, covreq, dtype, hc, ierr, inith, inits, 1 ipivot, ivneed, lastiv, lastv, lmat, mxfcal, mxiter, 2 nfcov, ngcov, nvdflt, outlev, parprt, parsav, perm, 3 prunit, qrtyp, rdreq, rmat, solprt, statpr, vneed, 4 vsave, x0prt c c *** iv subscript values *** c c/6 data algsav/51/, covprt/14/, covreq/15/, dtype/16/, hc/71/, 1 ierr/75/, inith/25/, inits/25/, ipivot/76/, ivneed/3/, 2 lastiv/44/, lastv/45/, lmat/42/, mxfcal/17/, mxiter/18/, 3 nfcov/52/, ngcov/53/, nvdflt/50/, outlev/19/, parprt/20/, 4 parsav/49/, perm/58/, prunit/21/, qrtyp/80/, rdreq/57/, 5 rmat/78/, solprt/22/, statpr/23/, vneed/4/, vsave/60/, 6 x0prt/24/ c/7 c parameter (algsav=51, covprt=14, covreq=15, dtype=16, hc=71, c 1 ierr=75, inith=25, inits=25, ipivot=76, ivneed=3, c 2 lastiv=44, lastv=45, lmat=42, mxfcal=17, mxiter=18, c 3 nfcov=52, ngcov=53, nvdflt=50, outlev=19, parprt=20, c 4 parsav=49, perm=58, prunit=21, qrtyp=80, rdreq=57, c 5 rmat=78, solprt=22, statpr=23, vneed=4, vsave=60, c 6 x0prt=24) c/ data miniv(1)/80/, miniv(2)/59/, minv(1)/98/, minv(2)/71/ c c------------------------------- body -------------------------------- c if (alg .lt. 1 .or. alg .gt. 2) go to 40 miv = miniv(alg) if (liv .lt. miv) go to 20 mv = minv(alg) if (lv .lt. mv) go to 30 call vdflt(alg, lv, v) iv(1) = 12 iv(algsav) = alg iv(ivneed) = 0 iv(lastiv) = miv iv(lastv) = mv iv(lmat) = mv + 1 iv(mxfcal) = 200 iv(mxiter) = 150 iv(outlev) = 1 iv(parprt) = 1 iv(perm) = miv + 1 iv(prunit) = imdcon(1) iv(solprt) = 1 iv(statpr) = 1 iv(vneed) = 0 iv(x0prt) = 1 c if (alg .ge. 2) go to 10 c c *** regression values c iv(covprt) = 3 iv(covreq) = 1 iv(dtype) = 1 iv(hc) = 0 iv(ierr) = 0 iv(inits) = 0 iv(ipivot) = 0 iv(nvdflt) = 32 iv(parsav) = 67 iv(qrtyp) = 1 iv(rdreq) = 3 iv(rmat) = 0 iv(vsave) = 58 go to 999 c c *** general optimization values c 10 iv(dtype) = 0 iv(inith) = 1 iv(nfcov) = 0 iv(ngcov) = 0 iv(nvdflt) = 25 iv(parsav) = 47 go to 999 c 20 iv(1) = 15 go to 999 c 30 iv(1) = 16 go to 999 c 40 iv(1) = 67 c 999 return c *** last card of deflt follows *** end real function dotprd(p, x, y) c c *** return the inner product of the p-vectors x and y. *** c integer p real x(p), y(p) c integer i real one, sqteta, t, zero c/+ real amax1, abs c/ external rmdcon real rmdcon c c *** rmdcon(2) returns a machine-dependent constant, sqteta, which c *** is slightly larger than the smallest positive number that c *** can be squared without underflowing. c c/6 data one/1.e+0/, sqteta/0.e+0/, zero/0.e+0/ c/7 c parameter (one=1.e+0, zero=0.e+0) c data sqteta/0.e+0/ c/ c dotprd = zero if (p .le. 0) go to 999 if (sqteta .eq. zero) sqteta = rmdcon(2) do 20 i = 1, p t = amax1( abs(x(i)), abs(y(i))) if (t .gt. one) go to 10 if (t .lt. sqteta) go to 20 t = (x(i)/sqteta)*y(i) if ( abs(t) .lt. sqteta) go to 20 10 dotprd = dotprd + x(i)*y(i) 20 continue c 999 return c *** last card of dotprd follows *** end subroutine itsum(d, g, iv, liv, lv, p, v, x) c c *** print iteration summary for ***sol (version 2.3) *** c c *** parameter declarations *** c integer liv, lv, p integer iv(liv) real d(p), g(p), v(lv), x(p) c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** local variables *** c integer alg, i, iv1, m, nf, ng, ol, pu c/6 real model1(6), model2(6) c/7 c character*4 model1(6), model2(6) c/ real nreldf, oldf, preldf, reldf, zero c c *** intrinsic functions *** c/+ integer iabs real abs, amax1 c/ c *** no external functions or subroutines *** c c *** subscripts for iv and v *** c integer algsav, dstnrm, f, fdif, f0, needhd, nfcall, nfcov, ngcov, 1 ngcall, niter, nreduc, outlev, preduc, prntit, prunit, 2 reldx, solprt, statpr, stppar, sused, x0prt c c *** iv subscript values *** c c/6 data algsav/51/, needhd/36/, nfcall/6/, nfcov/52/, ngcall/30/, 1 ngcov/53/, niter/31/, outlev/19/, prntit/39/, prunit/21/, 2 solprt/22/, statpr/23/, sused/64/, x0prt/24/ c/7 c parameter (algsav=51, needhd=36, nfcall=6, nfcov=52, ngcall=30, c 1 ngcov=53, niter=31, outlev=19, prntit=39, prunit=21, c 2 solprt=22, statpr=23, sused=64, x0prt=24) c/ c c *** v subscript values *** c c/6 data dstnrm/2/, f/10/, f0/13/, fdif/11/, nreduc/6/, preduc/7/, 1 reldx/17/, stppar/5/ c/7 c parameter (dstnrm=2, f=10, f0=13, fdif=11, nreduc=6, preduc=7, c 1 reldx=17, stppar=5) c/ c c/6 data zero/0.e+0/ c/7 c parameter (zero=0.e+0) c/ c/6 data model1(1)/4h /, model1(2)/4h /, model1(3)/4h /, 1 model1(4)/4h /, model1(5)/4h g /, model1(6)/4h s /, 2 model2(1)/4h g /, model2(2)/4h s /, model2(3)/4hg-s /, 3 model2(4)/4hs-g /, model2(5)/4h-s-g/, model2(6)/4h-g-s/ c/7 c data model1/' ',' ',' ',' ',' g ',' s '/, c 1 model2/' g ',' s ','g-s ','s-g ','-s-g','-g-s'/ c/ c c------------------------------- body -------------------------------- c pu = iv(prunit) if (pu .eq. 0) go to 999 iv1 = iv(1) if (iv1 .gt. 62) iv1 = iv1 - 51 ol = iv(outlev) alg = iv(algsav) if (iv1 .lt. 2 .or. iv1 .gt. 15) go to 370 if (iv1 .ge. 12) go to 120 if (iv1 .eq. 2 .and. iv(niter) .eq. 0) go to 390 if (ol .eq. 0) go to 120 if (iv1 .ge. 10 .and. iv(prntit) .eq. 0) go to 120 if (iv1 .gt. 2) go to 10 iv(prntit) = iv(prntit) + 1 if (iv(prntit) .lt. iabs(ol)) go to 999 10 nf = iv(nfcall) - iabs(iv(nfcov)) iv(prntit) = 0 reldf = zero preldf = zero oldf = amax1( abs(v(f0)), abs(v(f))) if (oldf .le. zero) go to 20 reldf = v(fdif) / oldf preldf = v(preduc) / oldf 20 if (ol .gt. 0) go to 60 c c *** print short summary line *** c if (iv(needhd) .eq. 1 .and. alg .eq. 1) write(pu,30) 30 format(/10h it nf,6x,1hf,7x,5hreldf,3x,6hpreldf,3x,5hreldx, 1 2x,13hmodel stppar) if (iv(needhd) .eq. 1 .and. alg .eq. 2) write(pu,40) 40 format(/11h it nf,7x,1hf,8x,5hreldf,4x,6hpreldf,4x,5hreldx, 1 3x,6hstppar) iv(needhd) = 0 if (alg .eq. 2) go to 50 m = iv(sused) write(pu,100) iv(niter), nf, v(f), reldf, preldf, v(reldx), 1 model1(m), model2(m), v(stppar) go to 120 c 50 write(pu,110) iv(niter), nf, v(f), reldf, preldf, v(reldx), 1 v(stppar) go to 120 c c *** print long summary line *** c 60 if (iv(needhd) .eq. 1 .and. alg .eq. 1) write(pu,70) 70 format(/11h it nf,6x,1hf,7x,5hreldf,3x,6hpreldf,3x,5hreldx, 1 2x,13hmodel stppar,2x,6hd*step,2x,7hnpreldf) if (iv(needhd) .eq. 1 .and. alg .eq. 2) write(pu,80) 80 format(/11h it nf,7x,1hf,8x,5hreldf,4x,6hpreldf,4x,5hreldx, 1 3x,6hstppar,3x,6hd*step,3x,7hnpreldf) iv(needhd) = 0 nreldf = zero if (oldf .gt. zero) nreldf = v(nreduc) / oldf if (alg .eq. 2) go to 90 m = iv(sused) write(pu,100) iv(niter), nf, v(f), reldf, preldf, v(reldx), 1 model1(m), model2(m), v(stppar), v(dstnrm), nreldf go to 120 c 90 write(pu,110) iv(niter), nf, v(f), reldf, preldf, 1 v(reldx), v(stppar), v(dstnrm), nreldf 100 format(i6,i5,e10.3,2e9.2,e8.1,a3,a4,2e8.1,e9.2) 110 format(i6,i5,e11.3,2e10.2,3e9.1,e10.2) c 120 if (iv(statpr) .lt. 0) go to 430 go to (999, 999, 130, 150, 170, 190, 210, 230, 250, 270, 290, 310, 1 330, 350, 520), iv1 c 130 write(pu,140) 140 format(/26h ***** x-convergence *****) go to 430 c 150 write(pu,160) 160 format(/42h ***** relative function convergence *****) go to 430 c 170 write(pu,180) 180 format(/49h ***** x- and relative function convergence *****) go to 430 c 190 write(pu,200) 200 format(/42h ***** absolute function convergence *****) go to 430 c 210 write(pu,220) 220 format(/33h ***** singular convergence *****) go to 430 c 230 write(pu,240) 240 format(/30h ***** false convergence *****) go to 430 c 250 write(pu,260) 260 format(/38h ***** function evaluation limit *****) go to 430 c 270 write(pu,280) 280 format(/28h ***** iteration limit *****) go to 430 c 290 write(pu,300) 300 format(/18h ***** stopx *****) go to 430 c 310 write(pu,320) 320 format(/44h ***** initial f(x) cannot be computed *****) c go to 390 c 330 write(pu,340) 340 format(/37h ***** bad parameters to assess *****) go to 999 c 350 write(pu,360) 360 format(/43h ***** gradient could not be computed *****) if (iv(niter) .gt. 0) go to 480 go to 390 c 370 write(pu,380) iv(1) 380 format(/14h ***** iv(1) =,i5,6h *****) go to 999 c c *** initial call on itsum *** c 390 if (iv(x0prt) .ne. 0) write(pu,400) (i, x(i), d(i), i = 1, p) 400 format(/23h i initial x(i),8x,4hd(i)//(1x,i5,e17.6,e14.3)) c *** the following are to avoid undefined variables when the c *** function evaluation limit is 1... v(dstnrm) = zero v(fdif) = zero v(nreduc) = zero v(preduc) = zero v(reldx) = zero if (iv1 .ge. 12) go to 999 iv(needhd) = 0 iv(prntit) = 0 if (ol .eq. 0) go to 999 if (ol .lt. 0 .and. alg .eq. 1) write(pu,30) if (ol .lt. 0 .and. alg .eq. 2) write(pu,40) if (ol .gt. 0 .and. alg .eq. 1) write(pu,70) if (ol .gt. 0 .and. alg .eq. 2) write(pu,80) if (alg .eq. 1) write(pu,410) v(f) if (alg .eq. 2) write(pu,420) v(f) 410 format(/11h 0 1,e10.3) c365 format(/11h 0 1,e11.3) 420 format(/11h 0 1,d11.3) go to 999 c c *** print various information requested on solution *** c 430 iv(needhd) = 1 if (iv(statpr) .eq. 0) go to 480 oldf = amax1( abs(v(f0)), abs(v(f))) preldf = zero nreldf = zero if (oldf .le. zero) go to 440 preldf = v(preduc) / oldf nreldf = v(nreduc) / oldf 440 nf = iv(nfcall) - iv(nfcov) ng = iv(ngcall) - iv(ngcov) write(pu,450) v(f), v(reldx), nf, ng, preldf, nreldf 450 format(/9h function,e17.6,8h reldx,e17.3/12h func. evals, 1 i8,9x,11hgrad. evals,i8/7h preldf,e16.3,6x,7hnpreldf,e15.3) c if (iv(nfcov) .gt. 0) write(pu,460) iv(nfcov) 460 format(/1x,i4,50h extra func. evals for covariance and diagnost 1ics.) if (iv(ngcov) .gt. 0) write(pu,470) iv(ngcov) 470 format(1x,i4,50h extra grad. evals for covariance and diagnosti 1cs.) c 480 if (iv(solprt) .eq. 0) go to 999 iv(needhd) = 1 write(pu,490) 490 format(/22h i final x(i),8x,4hd(i),10x,4hg(i)/) do 500 i = 1, p write(pu,510) i, x(i), d(i), g(i) 500 continue 510 format(1x,i5,e16.6,2e14.3) go to 999 c 520 write(pu,530) 530 format(/24h inconsistent dimensions) 999 return c *** last card of itsum follows *** end subroutine litvmu(n, x, l, y) c c *** solve (l**t)*x = 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 real x(n), l(1), y(n) integer i, ii, ij, im1, i0, j, np1 real xi, zero c/6 data zero/0.e+0/ c/7 c parameter (zero=0.e+0) c/ c do 10 i = 1, n 10 x(i) = y(i) np1 = n + 1 i0 = n*(n+1)/2 do 30 ii = 1, n i = np1 - ii xi = x(i)/l(i0) x(i) = xi if (i .le. 1) go to 999 i0 = i0 - i if (xi .eq. zero) go to 30 im1 = i - 1 do 20 j = 1, im1 ij = i0 + j x(j) = x(j) - xi*l(ij) 20 continue 30 continue 999 return c *** last card of litvmu follows *** end subroutine livmul(n, x, l, y) c c *** solve l*x = 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 real x(n), l(1), y(n) external dotprd real dotprd integer i, j, k real t, zero c/6 data zero/0.e+0/ c/7 c parameter (zero=0.e+0) c/ c do 10 k = 1, n if (y(k) .ne. zero) go to 20 x(k) = zero 10 continue go to 999 20 j = k*(k+1)/2 x(k) = y(k) / l(j) if (k .ge. n) go to 999 k = k + 1 do 30 i = k, n t = dotprd(i-1, l(j+1), x) j = j + i x(i) = (y(i) - t)/l(j) 30 continue 999 return c *** last card of livmul follows *** end subroutine parck(alg, d, iv, liv, lv, n, v) c c *** check ***sol (version 2.3) parameters, print changed values *** c c *** alg = 1 for regression, alg = 2 for general unconstrained opt. c integer alg, liv, lv, n integer iv(liv) real d(n), v(lv) c external rmdcon, vcopy, vdflt real rmdcon c rmdcon -- returns machine-dependent constants. c vcopy -- copies one vector to another. c vdflt -- supplies default parameter values to v alone. c/+ integer max0 c/ c c *** local variables *** c integer i, ii, iv1, j, k, l, m, miv1, miv2, ndfalt, parsv1, pu integer ijmp, jlim(2), miniv(2), ndflt(2) c/6 integer varnm(2), sh(2) real cngd(3), dflt(3), vn(2,34), which(3) c/7 c character*1 varnm(2), sh(2) c character*4 cngd(3), dflt(3), vn(2,34), which(3) c/ real big, machep, tiny, vk, vm(34), vx(34), zero c c *** iv and v subscripts *** c integer algsav, dinit, dtype, dtype0, epslon, inits, ivneed, 1 lastiv, lastv, lmat, nextiv, nextv, nvdflt, oldn, 2 parprt, parsav, perm, prunit, vneed c c c/6 data algsav/51/, dinit/38/, dtype/16/, dtype0/54/, epslon/19/, 1 inits/25/, ivneed/3/, lastiv/44/, lastv/45/, lmat/42/, 2 nextiv/46/, nextv/47/, nvdflt/50/, oldn/38/, parprt/20/, 3 parsav/49/, perm/58/, prunit/21/, vneed/4/ c/7 c parameter (algsav=51, dinit=38, dtype=16, dtype0=54, epslon=19, c 1 inits=25, ivneed=3, lastiv=44, lastv=45, lmat=42, c 2 nextiv=46, nextv=47, nvdflt=50, oldn=38, parprt=20, c 3 parsav=49, perm=58, prunit=21, vneed=4) c save big, machep, tiny c/ c data big/0.e+0/, machep/-1.e+0/, tiny/1.e+0/, zero/0.e+0/ c/6 data vn(1,1),vn(2,1)/4hepsl,4hon../ data vn(1,2),vn(2,2)/4hphmn,4hfc../ data vn(1,3),vn(2,3)/4hphmx,4hfc../ data vn(1,4),vn(2,4)/4hdecf,4hac../ data vn(1,5),vn(2,5)/4hincf,4hac../ data vn(1,6),vn(2,6)/4hrdfc,4hmn../ data vn(1,7),vn(2,7)/4hrdfc,4hmx../ data vn(1,8),vn(2,8)/4htune,4hr1../ data vn(1,9),vn(2,9)/4htune,4hr2../ data vn(1,10),vn(2,10)/4htune,4hr3../ data vn(1,11),vn(2,11)/4htune,4hr4../ data vn(1,12),vn(2,12)/4htune,4hr5../ data vn(1,13),vn(2,13)/4hafct,4hol../ data vn(1,14),vn(2,14)/4hrfct,4hol../ data vn(1,15),vn(2,15)/4hxcto,4hl.../ data vn(1,16),vn(2,16)/4hxfto,4hl.../ data vn(1,17),vn(2,17)/4hlmax,4h0.../ data vn(1,18),vn(2,18)/4hlmax,4hs.../ data vn(1,19),vn(2,19)/4hscto,4hl.../ data vn(1,20),vn(2,20)/4hdini,4ht.../ data vn(1,21),vn(2,21)/4hdtin,4hit../ data vn(1,22),vn(2,22)/4hd0in,4hit../ data vn(1,23),vn(2,23)/4hdfac,4h..../ data vn(1,24),vn(2,24)/4hdltf,4hdc../ data vn(1,25),vn(2,25)/4hdltf,4hdj../ data vn(1,26),vn(2,26)/4hdelt,4ha0../ data vn(1,27),vn(2,27)/4hfuzz,4h..../ data vn(1,28),vn(2,28)/4hrlim,4hit../ data vn(1,29),vn(2,29)/4hcosm,4hin../ data vn(1,30),vn(2,30)/4hhube,4hrc../ data vn(1,31),vn(2,31)/4hrspt,4hol../ data vn(1,32),vn(2,32)/4hsigm,4hin../ data vn(1,33),vn(2,33)/4heta0,4h..../ data vn(1,34),vn(2,34)/4hbias,4h..../ c/7 c data vn(1,1),vn(2,1)/'epsl','on..'/ c data vn(1,2),vn(2,2)/'phmn','fc..'/ c data vn(1,3),vn(2,3)/'phmx','fc..'/ c data vn(1,4),vn(2,4)/'decf','ac..'/ c data vn(1,5),vn(2,5)/'incf','ac..'/ c data vn(1,6),vn(2,6)/'rdfc','mn..'/ c data vn(1,7),vn(2,7)/'rdfc','mx..'/ c data vn(1,8),vn(2,8)/'tune','r1..'/ c data vn(1,9),vn(2,9)/'tune','r2..'/ c data vn(1,10),vn(2,10)/'tune','r3..'/ c data vn(1,11),vn(2,11)/'tune','r4..'/ c data vn(1,12),vn(2,12)/'tune','r5..'/ c data vn(1,13),vn(2,13)/'afct','ol..'/ c data vn(1,14),vn(2,14)/'rfct','ol..'/ c data vn(1,15),vn(2,15)/'xcto','l...'/ c data vn(1,16),vn(2,16)/'xfto','l...'/ c data vn(1,17),vn(2,17)/'lmax','0...'/ c data vn(1,18),vn(2,18)/'lmax','s...'/ c data vn(1,19),vn(2,19)/'scto','l...'/ c data vn(1,20),vn(2,20)/'dini','t...'/ c data vn(1,21),vn(2,21)/'dtin','it..'/ c data vn(1,22),vn(2,22)/'d0in','it..'/ c data vn(1,23),vn(2,23)/'dfac','....'/ c data vn(1,24),vn(2,24)/'dltf','dc..'/ c data vn(1,25),vn(2,25)/'dltf','dj..'/ c data vn(1,26),vn(2,26)/'delt','a0..'/ c data vn(1,27),vn(2,27)/'fuzz','....'/ c data vn(1,28),vn(2,28)/'rlim','it..'/ c data vn(1,29),vn(2,29)/'cosm','in..'/ c data vn(1,30),vn(2,30)/'hube','rc..'/ c data vn(1,31),vn(2,31)/'rspt','ol..'/ c data vn(1,32),vn(2,32)/'sigm','in..'/ c data vn(1,33),vn(2,33)/'eta0','....'/ c data vn(1,34),vn(2,34)/'bias','....'/ c/ c data vm(1)/1.0e-3/, vm(2)/-0.99e+0/, vm(3)/1.0e-3/, vm(4)/1.0e-2/, 1 vm(5)/1.2e+0/, vm(6)/1.e-2/, vm(7)/1.2e+0/, vm(8)/0.e+0/, 2 vm(9)/0.e+0/, vm(10)/1.e-3/, vm(11)/-1.e+0/, vm(13)/0.e+0/, 3 vm(15)/0.e+0/, vm(16)/0.e+0/, vm(19)/0.e+0/, vm(20)/-10.e+0/, 4 vm(21)/0.e+0/, vm(22)/0.e+0/, vm(23)/0.e+0/, vm(27)/1.01e+0/, 5 vm(28)/1.e+10/, vm(30)/0.e+0/, vm(31)/0.e+0/, vm(32)/0.e+0/, 6 vm(34)/0.e+0/ data vx(1)/0.9e+0/, vx(2)/-1.e-3/, vx(3)/1.e+1/, vx(4)/0.8e+0/, 1 vx(5)/1.e+2/, vx(6)/0.8e+0/, vx(7)/1.e+2/, vx(8)/0.5e+0/, 2 vx(9)/0.5e+0/, vx(10)/1.e+0/, vx(11)/1.e+0/, vx(14)/0.1e+0/, 3 vx(15)/1.e+0/, vx(16)/1.e+0/, vx(19)/1.e+0/, vx(23)/1.e+0/, 4 vx(24)/1.e+0/, vx(25)/1.e+0/, vx(26)/1.e+0/, vx(27)/1.e+10/, 5 vx(29)/1.e+0/, vx(31)/1.e+0/, vx(32)/1.e+0/, vx(33)/1.e+0/, 6 vx(34)/1.e+0/ c c/6 data varnm(1)/1hp/, varnm(2)/1hn/, sh(1)/1hs/, sh(2)/1hh/ data cngd(1),cngd(2),cngd(3)/4h---c,4hhang,4hed v/, 1 dflt(1),dflt(2),dflt(3)/4hnond,4hefau,4hlt v/ c/7 c data varnm(1)/'p'/, varnm(2)/'n'/, sh(1)/'s'/, sh(2)/'h'/ c data cngd(1),cngd(2),cngd(3)/'---c','hang','ed v'/, c 1 dflt(1),dflt(2),dflt(3)/'nond','efau','lt v'/ c/ data ijmp/33/, jlim(1)/0/, jlim(2)/24/, ndflt(1)/32/, ndflt(2)/25/ data miniv(1)/80/, miniv(2)/59/ c c............................... body ................................ c pu = 0 if (prunit .le. liv) pu = iv(prunit) if (alg .lt. 1 .or. alg .gt. 2) go to 340 if (iv(1) .eq. 0) call deflt(alg, iv, liv, lv, v) iv1 = iv(1) if (iv1 .ne. 13 .and. iv1 .ne. 12) go to 10 miv1 = miniv(alg) if (perm .le. liv) miv1 = max0(miv1, iv(perm) - 1) if (ivneed .le. liv) miv2 = miv1 + max0(iv(ivneed), 0) if (lastiv .le. liv) iv(lastiv) = miv2 if (liv .lt. miv1) go to 300 iv(ivneed) = 0 iv(lastv) = max0(iv(vneed), 0) + iv(lmat) - 1 iv(vneed) = 0 if (liv .lt. miv2) go to 300 if (lv .lt. iv(lastv)) go to 320 10 if (alg .eq. iv(algsav)) go to 30 if (pu .ne. 0) write(pu,20) alg, iv(algsav) 20 format(/39h the first parameter to deflt should be,i3, 1 12h rather than,i3) iv(1) = 82 go to 999 30 if (iv1 .lt. 12 .or. iv1 .gt. 14) go to 60 if (n .ge. 1) go to 50 iv(1) = 81 if (pu .eq. 0) go to 999 write(pu,40) varnm(alg), n 40 format(/8h /// bad,a1,2h =,i5) go to 999 50 if (iv1 .ne. 14) iv(nextiv) = iv(perm) if (iv1 .ne. 14) iv(nextv) = iv(lmat) if (iv1 .eq. 13) go to 999 k = iv(parsav) - epslon call vdflt(alg, lv-k, v(k+1)) iv(dtype0) = 2 - alg iv(oldn) = n which(1) = dflt(1) which(2) = dflt(2) which(3) = dflt(3) go to 110 60 if (n .eq. iv(oldn)) go to 80 iv(1) = 17 if (pu .eq. 0) go to 999 write(pu,70) varnm(alg), iv(oldn), n 70 format(/5h /// ,1a1,14h changed from ,i5,4h to ,i5) go to 999 c 80 if (iv1 .le. 11 .and. iv1 .ge. 1) go to 100 iv(1) = 80 if (pu .ne. 0) write(pu,90) iv1 90 format(/13h /// iv(1) =,i5,28h should be between 0 and 14.) go to 999 c 100 which(1) = cngd(1) which(2) = cngd(2) which(3) = cngd(3) c 110 if (iv1 .eq. 14) iv1 = 12 if (big .gt. tiny) go to 120 tiny = rmdcon(1) machep = rmdcon(3) big = rmdcon(6) vm(12) = machep vx(12) = big vx(13) = big vm(14) = machep vm(17) = tiny vx(17) = big vm(18) = tiny vx(18) = big vx(20) = big vx(21) = big vx(22) = big vm(24) = machep vm(25) = machep vm(26) = machep vx(28) = rmdcon(5) vm(29) = machep vx(30) = big vm(33) = machep 120 m = 0 i = 1 j = jlim(alg) k = epslon ndfalt = ndflt(alg) do 150 l = 1, ndfalt vk = v(k) if (vk .ge. vm(i) .and. vk .le. vx(i)) go to 140 m = k if (pu .ne. 0) write(pu,130) vn(1,i), vn(2,i), k, vk, 1 vm(i), vx(i) 130 format(/6h /// ,2a4,5h.. v(,i2,3h) =,e11.3,7h should, 1 11h be between,e11.3,4h and,d11.3) 140 k = k + 1 i = i + 1 if (i .eq. j) i = ijmp 150 continue c if (iv(nvdflt) .eq. ndfalt) go to 170 iv(1) = 51 if (pu .eq. 0) go to 999 write(pu,160) iv(nvdflt), ndfalt 160 format(/13h iv(nvdflt) =,i5,13h rather than ,i5) go to 999 170 if ((iv(dtype) .gt. 0 .or. v(dinit) .gt. zero) .and. iv1 .eq. 12) 1 go to 200 do 190 i = 1, n if (d(i) .gt. zero) go to 190 m = 18 if (pu .ne. 0) write(pu,180) i, d(i) 180 format(/8h /// d(,i3,3h) =,e11.3,19h should be positive) 190 continue 200 if (m .eq. 0) go to 210 iv(1) = m go to 999 c 210 if (pu .eq. 0 .or. iv(parprt) .eq. 0) go to 999 if (iv1 .ne. 12 .or. iv(inits) .eq. alg-1) go to 230 m = 1 write(pu,220) sh(alg), iv(inits) 220 format(/22h nondefault values..../5h init,a1,14h..... iv(25) =, 1 i3) 230 if (iv(dtype) .eq. iv(dtype0)) go to 250 if (m .eq. 0) write(pu,260) which m = 1 write(pu,240) iv(dtype) 240 format(20h dtype..... iv(16) =,i3) 250 i = 1 j = jlim(alg) k = epslon l = iv(parsav) ndfalt = ndflt(alg) do 290 ii = 1, ndfalt if (v(k) .eq. v(l)) go to 280 if (m .eq. 0) write(pu,260) which 260 format(/1h ,3a4,9halues..../) m = 1 write(pu,270) vn(1,i), vn(2,i), k, v(k) 270 format(1x,2a4,5h.. v(,i2,3h) =,e15.7) 280 k = k + 1 l = l + 1 i = i + 1 if (i .eq. j) i = ijmp 290 continue c iv(dtype0) = iv(dtype) parsv1 = iv(parsav) call vcopy(iv(nvdflt), v(parsv1), v(epslon)) go to 999 c 300 iv(1) = 15 if (pu .eq. 0) go to 999 write(pu,310) liv, miv2 310 format(/10h /// liv =,i5,17h must be at least,i5) if (liv .lt. miv1) go to 999 if (lv .lt. iv(lastv)) go to 320 go to 999 c 320 iv(1) = 16 if (pu .eq. 0) go to 999 write(pu,330) lv, iv(lastv) 330 format(/9h /// lv =,i5,17h must be at least,i5) go to 999 c 340 iv(1) = 67 if (pu .eq. 0) go to 999 write(pu,350) alg 350 format(/10h /// alg =,i5,15h must be 1 or 2) c 999 return c *** last card of parck follows *** end real function reldst(p, d, x, x0) c c *** compute and return relative difference between x and x0 *** c *** nl2sol version 2.2 *** c integer p real d(p), x(p), x0(p) c/+ real abs c/ integer i real emax, t, xmax, zero c/6 data zero/0.e+0/ c/7 c parameter (zero=0.e+0) c/ c emax = zero xmax = zero do 10 i = 1, p t = abs(d(i) * (x(i) - x0(i))) if (emax .lt. t) emax = t t = d(i) * ( abs(x(i)) + abs(x0(i))) if (xmax .lt. t) xmax = t 10 continue reldst = zero if (xmax .gt. zero) reldst = emax / xmax 999 return c *** last card of reldst follows *** end logical function stopx(idummy) c *****parameters... integer idummy c c .................................................................. c c *****purpose... c this function may serve as the stopx (asynchronous interruption) c function for the nl2sol (nonlinear least-squares) package at c those installations which do not wish to implement a c dynamic stopx. c c *****algorithm notes... c at installations where the nl2sol system is used c interactively, this dummy stopx should be replaced by a c function that returns .true. if and only if the interrupt c (break) key has been pressed since the last call on stopx. c c .................................................................. c stopx = .false. return end subroutine vaxpy(p, w, a, x, y) c c *** set w = a*x + y -- w, x, y = p-vectors, a = scalar *** c integer p real a, w(p), x(p), y(p) c integer i c do 10 i = 1, p 10 w(i) = a*x(i) + y(i) return end subroutine vcopy(p, y, x) c c *** set y = x, where x and y are p-vectors *** c integer p real x(p), y(p) c integer i c do 10 i = 1, p 10 y(i) = x(i) return end subroutine vdflt(alg, lv, v) c c *** supply ***sol (version 2.3) default values to v *** c c *** alg = 1 means regression constants. c *** alg = 2 means general unconstrained optimization constants. c integer alg, lv real v(lv) c/+ real amax1 c/ external rmdcon real rmdcon c rmdcon... returns machine-dependent constants c real machep, mepcrt, one, sqteps, three c c *** subscripts for v *** c integer afctol, bias, cosmin, decfac, delta0, dfac, dinit, dltfdc, 1 dltfdj, dtinit, d0init, epslon, eta0, fuzz, huberc, 2 incfac, lmax0, lmaxs, phmnfc, phmxfc, rdfcmn, rdfcmx, 3 rfctol, rlimit, rsptol, sctol, sigmin, tuner1, tuner2, 4 tuner3, tuner4, tuner5, xctol, xftol c c/6 data one/1.e+0/, three/3.e+0/ c/7 c parameter (one=1.e+0, three=3.e+0) c/ c c *** v subscript values *** c c/6 data afctol/31/, bias/43/, cosmin/47/, decfac/22/, delta0/44/, 1 dfac/41/, dinit/38/, dltfdc/42/, dltfdj/43/, dtinit/39/, 2 d0init/40/, epslon/19/, eta0/42/, fuzz/45/, huberc/48/, 3 incfac/23/, lmax0/35/, lmaxs/36/, phmnfc/20/, phmxfc/21/, 4 rdfcmn/24/, rdfcmx/25/, rfctol/32/, rlimit/46/, rsptol/49/, 5 sctol/37/, sigmin/50/, tuner1/26/, tuner2/27/, tuner3/28/, 6 tuner4/29/, tuner5/30/, xctol/33/, xftol/34/ c/7 c parameter (afctol=31, bias=43, cosmin=47, decfac=22, delta0=44, c 1 dfac=41, dinit=38, dltfdc=42, dltfdj=43, dtinit=39, c 2 d0init=40, epslon=19, eta0=42, fuzz=45, huberc=48, c 3 incfac=23, lmax0=35, lmaxs=36, phmnfc=20, phmxfc=21, c 4 rdfcmn=24, rdfcmx=25, rfctol=32, rlimit=46, rsptol=49, c 5 sctol=37, sigmin=50, tuner1=26, tuner2=27, tuner3=28, c 6 tuner4=29, tuner5=30, xctol=33, xftol=34) c/ c c------------------------------- body -------------------------------- c machep = rmdcon(3) v(afctol) = 1.e-20 if (machep .gt. 1.e-10) v(afctol) = machep**2 v(decfac) = 0.5e+0 sqteps = rmdcon(4) v(dfac) = 0.6e+0 v(delta0) = sqteps v(dtinit) = 1.e-6 mepcrt = machep ** (one/three) v(d0init) = 1.e+0 v(epslon) = 0.1e+0 v(incfac) = 2.e+0 v(lmax0) = 1.e+0 v(lmaxs) = 1.e+0 v(phmnfc) = -0.1e+0 v(phmxfc) = 0.1e+0 v(rdfcmn) = 0.1e+0 v(rdfcmx) = 4.e+0 v(rfctol) = amax1(1.e-10, mepcrt**2) v(sctol) = v(rfctol) v(tuner1) = 0.1e+0 v(tuner2) = 1.e-4 v(tuner3) = 0.75e+0 v(tuner4) = 0.5e+0 v(tuner5) = 0.75e+0 v(xctol) = sqteps v(xftol) = 1.e+2 * machep c if (alg .ge. 2) go to 10 c c *** regression values c v(cosmin) = amax1(1.e-6, 1.e+2 * machep) v(dinit) = 0.e+0 v(dltfdc) = mepcrt v(dltfdj) = sqteps v(fuzz) = 1.5e+0 v(huberc) = 0.7e+0 v(rlimit) = rmdcon(5) v(rsptol) = 1.e-3 v(sigmin) = 1.e-4 go to 999 c c *** general optimization values c 10 v(bias) = 0.8e+0 v(dinit) = -1.0e+0 v(eta0) = 1.0e+3 * machep c 999 return c *** last card of vdflt follows *** end subroutine vscopy(p, y, s) c c *** set p-vector y to scalar s *** c integer p real s, y(p) c integer i c do 10 i = 1, p 10 y(i) = s return end real function v2norm(p, x) c c *** return the 2-norm of the p-vector x, taking *** c *** care to avoid the most likely underflows. *** c integer p real x(p) c integer i, j real one, r, scale, sqteta, t, xi, zero c/+ real abs, sqrt c/ external rmdcon real rmdcon c c/6 data one/1.e+0/, zero/0.e+0/ c/7 c parameter (one=1.e+0, zero=0.e+0) c save sqteta c/ data sqteta/0.e+0/ c if (p .gt. 0) go to 10 v2norm = zero go to 999 10 do 20 i = 1, p if (x(i) .ne. zero) go to 30 20 continue v2norm = zero go to 999 c 30 scale = abs(x(i)) if (i .lt. p) go to 40 v2norm = scale go to 999 40 t = one if (sqteta .eq. zero) sqteta = rmdcon(2) c c *** sqteta is (slightly larger than) the square root of the c *** smallest positive floating point number on the machine. c *** the tests involving sqteta are done to prevent underflows. c j = i + 1 do 60 i = j, p xi = abs(x(i)) if (xi .gt. scale) go to 50 r = xi / scale if (r .gt. sqteta) t = t + r*r go to 60 50 r = scale / xi if (r .le. sqteta) r = zero t = one + t * r*r scale = xi 60 continue c v2norm = scale * sqrt(t) 999 return c *** last card of v2norm follows *** end subroutine humsl(n, d, x, calcf, calcgh, iv, liv, lv, v, 1 uiparm, urparm, ufparm) c c *** minimize general unconstrained objective function using *** c *** (analytic) gradient and hessian provided by the caller. *** c integer liv, lv, n integer iv(liv), uiparm(1) real d(n), x(n), v(lv), urparm(1) c dimension v(78 + n*(n+12)), uiparm(*), urparm(*) external calcf, calcgh, ufparm c c------------------------------ discussion --------------------------- c c this routine is like sumsl, except that the subroutine para- c meter calcg of sumsl (which computes the gradient of the objec- c tive function) is replaced by the subroutine parameter calcgh, c which computes both the gradient and (lower triangle of the) c hessian of the objective function. the calling sequence is... c call calcgh(n, x, nf, g, h, uiparm, urparm, ufparm) c parameters n, x, nf, g, uiparm, urparm, and ufparm are the same c as for sumsl, while h is an array of length n*(n+1)/2 in which c calcgh must store the lower triangle of the hessian at x. start- c ing at h(1), calcgh must store the hessian entries in the order c (1,1), (2,1), (2,2), (3,1), (3,2), (3,3), ... c the value printed (by itsum) in the column labelled stppar c is the levenberg-marquardt used in computing the current step. c zero means a full newton step. if the special case described in c ref. 1 is detected, then stppar is negated. the value printed c in the column labelled npreldf is zero if the current hessian c is not positive definite. c it sometimes proves worthwhile to let d be determined from the c diagonal of the hessian matrix by setting iv(dtype) = 1 and c v(dinit) = 0. the following iv and v components are relevant... c c iv(dtol)..... iv(59) gives the starting subscript in v of the dtol c array used when d is updated. (iv(dtol) can be c initialized by calling humsl with iv(1) = 13.) c iv(dtype).... iv(16) tells how the scale vector d should be chosen. c iv(dtype) .le. 0 means that d should not be updated, and c iv(dtype) .ge. 1 means that d should be updated as c described below with v(dfac). default = 0. c v(dfac)..... v(41) and the dtol and d0 arrays (see v(dtinit) and c v(d0init)) are used in updating the scale vector d when c iv(dtype) .gt. 0. (d is initialized according to c v(dinit), described in sumsl.) let c d1(i) = max(sqrt(abs(h(i,i))), v(dfac)*d(i)), c where h(i,i) is the i-th diagonal element of the current c hessian. if iv(dtype) = 1, then d(i) is set to d1(i) c unless d1(i) .lt. dtol(i), in which case d(i) is set to c max(d0(i), dtol(i)). c if iv(dtype) .ge. 2, then d is updated during the first c iteration as for iv(dtype) = 1 (after any initialization c due to v(dinit)) and is left unchanged thereafter. c default = 0.6. c v(dtinit)... v(39), if positive, is the value to which all components c of the dtol array (see v(dfac)) are initialized. if c v(dtinit) = 0, then it is assumed that the caller has c stored dtol in v starting at v(iv(dtol)). c default = 10**-6. c v(d0init)... v(40), if positive, is the value to which all components c of the d0 vector (see v(dfac)) are initialized. if c v(dfac) = 0, then it is assumed that the caller has c stored d0 in v starting at v(iv(dtol)+n). default = 1.0. c c *** reference *** c c 1. gay, d.m. (1981), computing optimal locally constrained steps, c siam j. sci. statist. comput. 2, pp. 186-197. c. c *** general *** c c coded by david m. gay (winter 1980). 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---------------------------- declarations --------------------------- c external deflt, humit c c deflt... provides default input values for iv and v. c humit... reverse-communication routine that does humsl algorithm. c integer g1, h1, iv1, lh, nf real f c c *** subscripts for iv *** c integer g, h, nextv, nfcall, nfgcal, toobig, vneed c c/6 data nextv/47/, nfcall/6/, nfgcal/7/, g/28/, h/56/, toobig/2/, 1 vneed/4/ c/7 c parameter (nextv=47, nfcall=6, nfgcal=7, g=28, h=56, toobig=2, c 1 vneed=4) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c lh = n * (n + 1) / 2 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+3)/2 iv1 = iv(1) if (iv1 .eq. 14) go to 10 if (iv1 .gt. 2 .and. iv1 .lt. 12) go to 10 g1 = 1 h1 = 1 if (iv1 .eq. 12) iv(1) = 13 go to 20 c 10 g1 = iv(g) h1 = iv(h) c 20 call humit(d, f, v(g1), v(h1), iv, lh, 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 calcgh(n, x, iv(nfgcal), v(g1), v(h1), uiparm, urparm, 1 ufparm) go to 20 c 50 if (iv(1) .ne. 14) go to 999 c c *** storage allocation c iv(g) = iv(nextv) iv(h) = iv(g) + n iv(nextv) = iv(h) + n*(n+1)/2 if (iv1 .ne. 13) go to 10 c 999 return c *** last card of humsl follows *** end subroutine humit(d, fx, g, h, iv, lh, liv, lv, n, v, x) c c *** carry out humsl (unconstrained minimization) iterations, using c *** hessian matrix provided by the caller. c c *** parameter declarations *** c integer lh, liv, lv, n integer iv(liv) real d(n), fx, g(n), h(lh), v(lv), x(n) c c-------------------------- parameter usage -------------------------- c c d.... scale vector. c fx... function value. c g.... gradient vector. c h.... lower triangle of the hessian, stored rowwise. c iv... integer value array. c lh... length of h = p*(p+1)/2. c liv.. length of iv (at least 60). c lv... length of v (at least 78 + n*(n+21)/2). c n.... number of variables (components in x and g). c v.... floating-point value array. c x.... parameter vector. c c *** discussion *** c c parameters iv, n, v, and x are the same as the corresponding c ones to humsl (which see), except that v can be shorter (since c the part of v that humsl uses for storing g and h is not needed). c moreover, compared with humsl, 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 humsl, is not referenced by humit or the c subroutines it calls. c c iv(1) = 1 means the caller should set fx to f(x), the function value c at x, and call humit again, having changed none of the c other parameters. an exception occurs if f(x) cannot be c computed (e.g. if overflow would occur), which may happen c because of an oversized step. in this case the caller c should set iv(toobig) = iv(2) to 1, which will cause c humit to ignore fx and try a smaller step. the para- c meter nf that humsl passes to calcf (for possible use by c calcgh) is a copy of iv(nfcall) = iv(6). c iv(1) = 2 means the caller should set g to g(x), the gradient of f at c x, and h to the lower triangle of h(x), the hessian of f c at x, and call humit again, having changed none of the c other parameters except perhaps the scale vector d. c the parameter nf that humsl passes to calcg is c iv(nfgcal) = iv(7). if g(x) and h(x) cannot be evaluated, c then the caller may set iv(nfgcal) to 0, in which case c humit will return with iv(1) = 65. c note -- humit overwrites h with the lower triangle c of diag(d)**-1 * h(x) * diag(d)**-1. c. c *** general *** c c coded by david m. gay (winter 1980). 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 and humsl for references.) c c+++++++++++++++++++++++++++ declarations ++++++++++++++++++++++++++++ c c *** local variables *** c integer dg1, dummy, i, j, k, l, lstgst, nn1o2, step1, 1 temp1, w1, x01 real t c c *** constants *** c real one, onep2, zero c c *** no intrinsic functions *** c c *** external functions and subroutines *** c external assst, deflt, dotprd, dupdu, gqtst, itsum, parck, 1 reldst, slvmul, stopx, vaxpy, vcopy, vscopy, v2norm logical stopx real dotprd, reldst, v2norm c c assst.... assesses candidate step. c deflt.... provides default iv and v input values. c dotprd... returns inner product of two vectors. c dupdu.... updates scale vector d. c gqtst.... computes optimally locally constrained step. c itsum.... prints iteration summary and info on initial and final x. c parck.... checks validity of input iv and v values. c reldst... computes v(reldx) = relative step size. c slvmul... multiplies symmetric matrix times vector, given the lower c triangle of the matrix. 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 v2norm... returns the 2-norm of a vector. c c *** subscripts for iv and v *** c integer cnvcod, dg, dgnorm, dinit, dstnrm, dtinit, dtol, 1 dtype, d0init, f, f0, fdif, gtstep, incfac, irc, kagqt, 2 lmat, lmax0, lmaxs, mode, model, mxfcal, mxiter, nextv, 3 nfcall, nfgcal, ngcall, niter, preduc, radfac, radinc, 4 radius, rad0, reldx, restor, step, stglim, stlstg, stppar, 5 toobig, tuner4, tuner5, vneed, w, xirc, x0 c c *** iv subscript values *** c c/6 data cnvcod/55/, dg/37/, dtol/59/, dtype/16/, irc/29/, kagqt/33/, 1 lmat/42/, mode/35/, model/5/, mxfcal/17/, mxiter/18/, 2 nextv/47/, nfcall/6/, nfgcal/7/, ngcall/30/, niter/31/, 3 radinc/8/, restor/9/, step/40/, stglim/11/, stlstg/41/, 4 toobig/2/, vneed/4/, w/34/, xirc/13/, x0/43/ c/7 c parameter (cnvcod=55, dg=37, dtol=59, dtype=16, irc=29, kagqt=33, c 1 lmat=42, mode=35, model=5, mxfcal=17, mxiter=18, c 2 nextv=47, nfcall=6, nfgcal=7, ngcall=30, niter=31, c 3 radinc=8, restor=9, step=40, stglim=11, stlstg=41, c 4 toobig=2, vneed=4, w=34, xirc=13, x0=43) c/ c c *** v subscript values *** c c/6 data dgnorm/1/, dinit/38/, dstnrm/2/, dtinit/39/, d0init/40/, 1 f/10/, f0/13/, fdif/11/, gtstep/4/, incfac/23/, lmax0/35/, 2 lmaxs/36/, preduc/7/, radfac/16/, radius/8/, rad0/9/, 3 reldx/17/, stppar/5/, tuner4/29/, tuner5/30/ c/7 c parameter (dgnorm=1, dinit=38, dstnrm=2, dtinit=39, d0init=40, c 1 f=10, f0=13, fdif=11, gtstep=4, incfac=23, lmax0=35, c 2 lmaxs=36, preduc=7, radfac=16, radius=8, rad0=9, c 3 reldx=17, stppar=5, tuner4=29, tuner5=30) c/ c c/6 data one/1.e+0/, onep2/1.2e+0/, zero/0.e+0/ c/7 c parameter (one=1.e+0, onep2=1.2e+0, zero=0.e+0) c/ c c+++++++++++++++++++++++++++++++ body ++++++++++++++++++++++++++++++++ c i = iv(1) if (i .eq. 1) go to 30 if (i .eq. 2) go to 40 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+21)/2 + 7 call parck(2, d, iv, liv, lv, n, v) i = iv(1) - 2 if (i .gt. 12) go to 999 nn1o2 = n * (n + 1) / 2 if (lh .ge. nn1o2) go to (210,210,210,210,210,210,160,120,160, 1 10,10,20), i iv(1) = 66 go to 350 c c *** storage allocation *** c 10 iv(dtol) = iv(lmat) + nn1o2 iv(x0) = iv(dtol) + 2*n iv(step) = iv(x0) + n iv(stlstg) = iv(step) + n iv(dg) = iv(stlstg) + n iv(w) = iv(dg) + n iv(nextv) = iv(w) + 4*n + 7 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 v(stppar) = zero if (v(dinit) .ge. zero) call vscopy(n, d, v(dinit)) k = iv(dtol) if (v(dtinit) .gt. zero) call vscopy(n, v(k), v(dtinit)) k = k + n if (v(d0init) .gt. zero) call vscopy(n, v(k), v(d0init)) iv(1) = 1 go to 999 c 30 v(f) = fx if (iv(mode) .ge. 0) go to 210 iv(1) = 2 if (iv(toobig) .eq. 0) go to 999 iv(1) = 63 go to 350 c c *** make sure gradient could be computed *** c 40 if (iv(nfgcal) .ne. 0) go to 50 iv(1) = 65 go to 350 c c *** update the scale vector d *** c 50 dg1 = iv(dg) if (iv(dtype) .le. 0) go to 70 k = dg1 j = 0 do 60 i = 1, n j = j + i v(k) = h(j) k = k + 1 60 continue call dupdu(d, v(dg1), iv, liv, lv, n, v) c c *** compute scaled gradient and its norm *** c 70 dg1 = iv(dg) k = dg1 do 80 i = 1, n v(k) = g(i) / d(i) k = k + 1 80 continue v(dgnorm) = v2norm(n, v(dg1)) c c *** compute scaled hessian *** c k = 1 do 100 i = 1, n t = one / d(i) do 90 j = 1, i h(k) = t * h(k) / d(j) k = k + 1 90 continue 100 continue c if (iv(cnvcod) .ne. 0) go to 340 if (iv(mode) .eq. 0) go to 300 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 110 call itsum(d, g, iv, liv, lv, n, v, x) 120 k = iv(niter) if (k .lt. iv(mxiter)) go to 130 iv(1) = 10 go to 350 c 130 iv(niter) = k + 1 c c *** initialize for start of next iteration *** c dg1 = iv(dg) x01 = iv(x0) v(f0) = v(f) iv(irc) = 4 iv(kagqt) = -1 c c *** copy x to x0 *** c call vcopy(n, v(x01), x) c c *** update radius *** c if (k .eq. 0) go to 150 step1 = iv(step) k = step1 do 140 i = 1, n v(k) = d(i) * v(k) k = k + 1 140 continue v(radius) = v(radfac) * v2norm(n, v(step1)) c c *** check stopx and function evaluation limit *** c 150 if (.not. stopx(dummy)) go to 170 iv(1) = 11 go to 180 c c *** come here when restarting after func. eval. limit or stopx. c 160 if (v(f) .ge. v(f0)) go to 170 v(radfac) = one k = iv(niter) go to 130 c 170 if (iv(nfcall) .lt. iv(mxfcal)) go to 190 iv(1) = 9 180 if (v(f) .ge. v(f0)) go to 350 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 290 c c. . . . . . . . . . . . . compute candidate step . . . . . . . . . . c 190 step1 = iv(step) dg1 = iv(dg) l = iv(lmat) w1 = iv(w) call gqtst(d, v(dg1), h, iv(kagqt), v(l), n, v(step1), v, v(w1)) if (iv(irc) .eq. 6) go to 210 c c *** check whether evaluating f(x0 + step) looks worthwhile *** c if (v(dstnrm) .le. zero) go to 210 if (iv(irc) .ne. 5) go to 200 if (v(radfac) .le. one) go to 200 if (v(preduc) .le. onep2 * v(fdif)) go to 210 c c *** compute f(x0 + step) *** c 200 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 210 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 220 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 220 k = iv(irc) go to (230,260,260,260,230,240,250,250,250,250,250,250,330,300), k c c *** recompute step with new radius *** c 230 v(radius) = v(radfac) * v(dstnrm) go to 150 c c *** compute step of length v(lmaxs) for singular convergence test. c 240 v(radius) = v(lmaxs) go to 190 c c *** convergence or false convergence *** c 250 iv(cnvcod) = k - 4 if (v(f) .ge. v(f0)) go to 340 if (iv(xirc) .eq. 14) go to 340 iv(xirc) = 14 c c. . . . . . . . . . . . process acceptable step . . . . . . . . . . . c 260 if (iv(irc) .ne. 3) go to 290 temp1 = lstgst c c *** prepare for gradient tests *** c *** set temp1 = hessian * step + g(x0) c *** = diag(d) * (h * step + g(x0)) c c use x0 vector as temporary. k = x01 do 270 i = 1, n v(k) = d(i) * v(step1) k = k + 1 step1 = step1 + 1 270 continue call slvmul(n, v(temp1), h, v(x01)) do 280 i = 1, n v(temp1) = d(i) * v(temp1) + g(i) temp1 = temp1 + 1 280 continue c c *** compute gradient and hessian *** c 290 iv(ngcall) = iv(ngcall) + 1 iv(1) = 2 go to 999 c 300 iv(1) = 2 if (iv(irc) .ne. 3) go to 110 c c *** set v(radfac) by gradient tests *** c temp1 = iv(stlstg) step1 = iv(step) c c *** set temp1 = diag(d)**-1 * (hessian*step + (g(x0)-g(x))) *** c k = temp1 do 310 i = 1, n v(k) = (v(k) - g(i)) / d(i) k = k + 1 310 continue c c *** do gradient tests *** c if (v2norm(n, v(temp1)) .le. v(dgnorm) * v(tuner4)) go to 320 if (dotprd(n, g, v(step1)) 1 .ge. v(gtstep) * v(tuner5)) go to 110 320 v(radfac) = v(incfac) go to 110 c c. . . . . . . . . . . . . . misc. details . . . . . . . . . . . . . . c c *** bad parameters to assess *** c 330 iv(1) = 64 go to 350 c c *** print summary of final iteration and other requested items *** c 340 iv(1) = iv(cnvcod) iv(cnvcod) = 0 350 call itsum(d, g, iv, liv, lv, n, v, x) c 999 return c c *** last card of humit follows *** end subroutine dupdu(d, hdiag, iv, liv, lv, n, v) c c *** update scale vector d for humsl *** c c *** parameter declarations *** c integer liv, lv, n integer iv(liv) real d(n), hdiag(n), v(lv) c c *** local variables *** c integer dtoli, d0i, i real t, vdfac c c *** intrinsic functions *** c/+ real abs, amax1, sqrt c/ c *** subscripts for iv and v *** c integer dfac, dtol, dtype, niter c/6 data dfac/41/, dtol/59/, dtype/16/, niter/31/ c/7 c parameter (dfac=41, dtol=59, dtype=16, niter=31) c/ c c------------------------------- body -------------------------------- c i = iv(dtype) if (i .eq. 1) go to 10 if (iv(niter) .gt. 0) go to 999 c 10 dtoli = iv(dtol) d0i = dtoli + n vdfac = v(dfac) do 20 i = 1, n t = amax1( sqrt( abs(hdiag(i))), vdfac*d(i)) if (t .lt. v(dtoli)) t = amax1(v(dtoli), v(d0i)) d(i) = t dtoli = dtoli + 1 d0i = d0i + 1 20 continue c 999 return c *** last card of dupdu follows *** end subroutine gqtst(d, dig, dihdi, ka, l, p, step, v, w) c c *** compute goldfeld-quandt-trotter step by more-hebden technique *** c *** (nl2sol version 2.2), modified a la more and sorensen *** c c *** parameter declarations *** c integer ka, p real d(p), dig(p), dihdi(1), l(1), v(21), step(p), 1 w(1) c dimension dihdi(p*(p+1)/2), l(p*(p+1)/2), w(4*p+7) c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** purpose *** c c given the (compactly stored) lower triangle of a scaled c hessian (approximation) and a nonzero scaled gradient vector, c this subroutine computes a goldfeld-quandt-trotter step of c approximate length v(radius) by the more-hebden technique. in c other words, step is computed to (approximately) minimize c psi(step) = (g**t)*step + 0.5*(step**t)*h*step such that the c 2-norm of d*step is at most (approximately) v(radius), where c g is the gradient, h is the hessian, and d is a diagonal c scale matrix whose diagonal is stored in the parameter d. c (gqtst assumes dig = d**-1 * g and dihdi = d**-1 * h * d**-1.) c c *** parameter description *** c c d (in) = the scale vector, i.e. the diagonal of the scale c matrix d mentioned above under purpose. c dig (in) = the scaled gradient vector, d**-1 * g. if g = 0, then c step = 0 and v(stppar) = 0 are returned. c dihdi (in) = lower triangle of the scaled hessian (approximation), c i.e., d**-1 * h * d**-1, stored compactly by rows., i.e., c in the order (1,1), (2,1), (2,2), (3,1), (3,2), etc. c ka (i/o) = the number of hebden iterations (so far) taken to deter- c mine step. ka .lt. 0 on input means this is the first c attempt to determine step (for the present dig and dihdi) c -- ka is initialized to 0 in this case. output with c ka = 0 (or v(stppar) = 0) means step = -(h**-1)*g. c l (i/o) = workspace of length p*(p+1)/2 for cholesky factors. c p (in) = number of parameters -- the hessian is a p x p matrix. c step (i/o) = the step computed. c v (i/o) contains various constants and variables described below. c w (i/o) = workspace of length 4*p + 6. c c *** entries in v *** c c v(dgnorm) (i/o) = 2-norm of (d**-1)*g. c v(dstnrm) (output) = 2-norm of d*step. c v(dst0) (i/o) = 2-norm of d*(h**-1)*g (for pos. def. h only), or c overestimate of smallest eigenvalue of (d**-1)*h*(d**-1). c v(epslon) (in) = max. rel. error allowed for psi(step). for the c step returned, psi(step) will exceed its optimal value c by less than -v(epslon)*psi(step). suggested value = 0.1. c v(gtstep) (out) = inner product between g and step. c v(nreduc) (out) = psi(-(h**-1)*g) = psi(newton step) (for pos. def. c h only -- v(nreduc) is set to zero otherwise). c v(phmnfc) (in) = tol. (together with v(phmxfc)) for accepting step c (more*s sigma). the error v(dstnrm) - v(radius) must lie c between v(phmnfc)*v(radius) and v(phmxfc)*v(radius). c v(phmxfc) (in) (see v(phmnfc).) c suggested values -- v(phmnfc) = -0.25, v(phmxfc) = 0.5. c v(preduc) (out) = psi(step) = predicted obj. func. reduction for step. c v(radius) (in) = radius of current (scaled) trust region. c v(rad0) (i/o) = value of v(radius) from previous call. c v(stppar) (i/o) is normally the marquardt parameter, i.e. the alpha c described below under algorithm notes. if h + alpha*d**2 c (see algorithm notes) is (nearly) singular, however, c then v(stppar) = -alpha. c c *** usage notes *** c c if it is desired to recompute step using a different value of c v(radius), then this routine may be restarted by calling it c with all parameters unchanged except v(radius). (this explains c why step and w are listed as i/o). on an initial call (one with c ka .lt. 0), step and w need not be initialized and only compo- c nents v(epslon), v(stppar), v(phmnfc), v(phmxfc), v(radius), and c v(rad0) of v must be initialized. c c *** algorithm notes *** c c the desired g-q-t step (ref. 2, 3, 4, 6) satisfies c (h + alpha*d**2)*step = -g for some nonnegative alpha such that c h + alpha*d**2 is positive semidefinite. alpha and step are c computed by a scheme analogous to the one described in ref. 5. c estimates of the smallest and largest eigenvalues of the hessian c are obtained from the gerschgorin circle theorem enhanced by a c simple form of the scaling described in ref. 7. cases in which c h + alpha*d**2 is nearly (or exactly) singular are handled by c the technique discussed in ref. 2. in these cases, a step of c (exact) length v(radius) is returned for which psi(step) exceeds c its optimal value by less than -v(epslon)*psi(step). the test c suggested in ref. 6 for detecting the special case is performed c once two matrix factorizations have been done -- doing so sooner c seems to degrade the performance of optimization routines that c call this routine. c c *** functions and subroutines called *** c c dotprd - returns inner product of two vectors. c litvmu - applies inverse-transpose of compact lower triang. matrix. c livmul - applies inverse of compact lower triang. matrix. c lsqrt - finds cholesky factor (of compactly stored lower triang.). c lsvmin - returns approx. to min. sing. value of lower triang. matrix. c rmdcon - returns machine-dependent constants. c v2norm - returns 2-norm of a vector. c c *** references *** c c 1. dennis, j.e., gay, d.m., and welsch, r.e. (1981), an adaptive c nonlinear least-squares algorithm, acm trans. math. c software, vol. 7, no. 3. c 2. gay, d.m. (1981), computing optimal locally constrained steps, c siam j. sci. statist. computing, vol. 2, no. 2, pp. c 186-197. c 3. goldfeld, s.m., quandt, r.e., and trotter, h.f. (1966), c maximization by quadratic hill-climbing, econometrica 34, c pp. 541-551. c 4. hebden, m.d. (1973), an algorithm for minimization using exact c second derivatives, report t.p. 515, theoretical physics c div., a.e.r.e. harwell, oxon., england. c 5. more, j.j. (1978), the levenberg-marquardt algorithm, implemen- c tation and theory, pp.105-116 of springer lecture notes c in mathematics no. 630, edited by g.a. watson, springer- c verlag, berlin and new york. c 6. more, j.j., and sorensen, d.c. (1981), computing a trust region c step, technical report anl-81-83, argonne national lab. c 7. varga, r.s. (1965), minimal gerschgorin sets, pacific j. math. 15, c pp. 719-729. c c *** general *** c c coded by david m. gay. c this subroutine was written in connection with research c supported by the national science foundation under grants c mcs-7600324, dcr75-10143, 76-14311dss, mcs76-11989, and c mcs-7906671. c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** local variables *** c logical restrt integer dggdmx, diag, diag0, dstsav, emax, emin, i, im1, inc, irc, 1 j, k, kalim, kamin, k1, lk0, phipin, q, q0, uk0, x real alphak, aki, akk, delta, dst, eps, gtsta, lk, 1 oldphi, phi, phimax, phimin, psifac, rad, radsq, 2 root, si, sk, sw, t, twopsi, t1, t2, uk, wi c c *** constants *** real big, dgxfac, epsfac, four, half, kappa, negone, 1 one, p001, six, three, two, zero c c *** intrinsic functions *** c/+ real abs, amax1, amin1, sqrt c/ c *** external functions and subroutines *** c external dotprd, litvmu, livmul, lsqrt, lsvmin, rmdcon, v2norm real dotprd, lsvmin, rmdcon, v2norm c c *** subscripts for v *** c integer dgnorm, dstnrm, dst0, epslon, gtstep, stppar, nreduc, 1 phmnfc, phmxfc, preduc, radius, rad0 c/6 data dgnorm/1/, dstnrm/2/, dst0/3/, epslon/19/, gtstep/4/, 1 nreduc/6/, phmnfc/20/, phmxfc/21/, preduc/7/, radius/8/, 2 rad0/9/, stppar/5/ c/7 c parameter (dgnorm=1, dstnrm=2, dst0=3, epslon=19, gtstep=4, c 1 nreduc=6, phmnfc=20, phmxfc=21, preduc=7, radius=8, c 2 rad0=9, stppar=5) c/ c c/6 data epsfac/50.0e+0/, four/4.0e+0/, half/0.5e+0/, 1 kappa/2.0e+0/, negone/-1.0e+0/, one/1.0e+0/, p001/1.0e-3/, 2 six/6.0e+0/, three/3.0e+0/, two/2.0e+0/, zero/0.0e+0/ c/7 c parameter (epsfac=50.0e+0, four=4.0e+0, half=0.5e+0, c 1 kappa=2.0e+0, negone=-1.0e+0, one=1.0e+0, p001=1.0e-3, c 2 six=6.0e+0, three=3.0e+0, two=2.0e+0, zero=0.0e+0) c save dgxfac c/ data big/0.e+0/, dgxfac/0.e+0/ c c *** body *** c c *** store largest abs. entry in (d**-1)*h*(d**-1) at w(dggdmx). dggdmx = p + 1 c *** store gerschgorin over- and underestimates of the largest c *** and smallest eigenvalues of (d**-1)*h*(d**-1) at w(emax) c *** and w(emin) respectively. emax = dggdmx + 1 emin = emax + 1 c *** for use in recomputing step, the final values of lk, uk, dst, c *** and the inverse derivative of more*s phi at 0 (for pos. def. c *** h) are stored in w(lk0), w(uk0), w(dstsav), and w(phipin) c *** respectively. lk0 = emin + 1 phipin = lk0 + 1 uk0 = phipin + 1 dstsav = uk0 + 1 c *** store diag of (d**-1)*h*(d**-1) in w(diag),...,w(diag0+p). diag0 = dstsav diag = diag0 + 1 c *** store -d*step in w(q),...,w(q0+p). q0 = diag0 + p q = q0 + 1 c *** allocate storage for scratch vector x *** x = q + p rad = v(radius) radsq = rad**2 c *** phitol = max. error allowed in dst = v(dstnrm) = 2-norm of c *** d*step. phimax = v(phmxfc) * rad phimin = v(phmnfc) * rad psifac = two * v(epslon) / (three * (four * (v(phmnfc) + one) * 1 (kappa + one) + kappa + two) * rad**2) c *** oldphi is used to detect limits of numerical accuracy. if c *** we recompute step and it does not change, then we accept it. oldphi = zero eps = v(epslon) irc = 0 restrt = .false. kalim = ka + 50 c c *** start or restart, depending on ka *** c if (ka .ge. 0) go to 290 c c *** fresh start *** c k = 0 uk = negone ka = 0 kalim = 50 v(dgnorm) = v2norm(p, dig) v(nreduc) = zero v(dst0) = zero kamin = 3 if (v(dgnorm) .eq. zero) kamin = 0 c c *** store diag(dihdi) in w(diag0+1),...,w(diag0+p) *** c j = 0 do 10 i = 1, p j = j + i k1 = diag0 + i w(k1) = dihdi(j) 10 continue c c *** determine w(dggdmx), the largest element of dihdi *** c t1 = zero j = p * (p + 1) / 2 do 20 i = 1, j t = abs(dihdi(i)) if (t1 .lt. t) t1 = t 20 continue w(dggdmx) = t1 c c *** try alpha = 0 *** c 30 call lsqrt(1, p, l, dihdi, irc) if (irc .eq. 0) go to 50 c *** indef. h -- underestimate smallest eigenvalue, use this c *** estimate to initialize lower bound lk on alpha. j = irc*(irc+1)/2 t = l(j) l(j) = one do 40 i = 1, irc 40 w(i) = zero w(irc) = one call litvmu(irc, w, l, w) t1 = v2norm(irc, w) lk = -t / t1 / t1 v(dst0) = -lk if (restrt) go to 210 go to 70 c c *** positive definite h -- compute unmodified newton step. *** 50 lk = zero t = lsvmin(p, l, w(q), w(q)) if (t .ge. one) go to 60 if (big .le. zero) big = rmdcon(6) if (v(dgnorm) .ge. t*t*big) go to 70 60 call livmul(p, w(q), l, dig) gtsta = dotprd(p, w(q), w(q)) v(nreduc) = half * gtsta call litvmu(p, w(q), l, w(q)) dst = v2norm(p, w(q)) v(dst0) = dst phi = dst - rad if (phi .le. phimax) go to 260 if (restrt) go to 210 c c *** prepare to compute gerschgorin estimates of largest (and c *** smallest) eigenvalues. *** c 70 k = 0 do 100 i = 1, p wi = zero if (i .eq. 1) go to 90 im1 = i - 1 do 80 j = 1, im1 k = k + 1 t = abs(dihdi(k)) wi = wi + t w(j) = w(j) + t 80 continue 90 w(i) = wi k = k + 1 100 continue c c *** (under-)estimate smallest eigenvalue of (d**-1)*h*(d**-1) *** c k = 1 t1 = w(diag) - w(1) if (p .le. 1) go to 120 do 110 i = 2, p j = diag0 + i t = w(j) - w(i) if (t .ge. t1) go to 110 t1 = t k = i 110 continue c 120 sk = w(k) j = diag0 + k akk = w(j) k1 = k*(k-1)/2 + 1 inc = 1 t = zero do 150 i = 1, p if (i .eq. k) go to 130 aki = abs(dihdi(k1)) si = w(i) j = diag0 + i t1 = half * (akk - w(j) + si - aki) t1 = t1 + sqrt(t1*t1 + sk*aki) if (t .lt. t1) t = t1 if (i .lt. k) go to 140 130 inc = i 140 k1 = k1 + inc 150 continue c w(emin) = akk - t uk = v(dgnorm)/rad - w(emin) if (v(dgnorm) .eq. zero) uk = uk + p001 + p001*uk if (uk .le. zero) uk = p001 c c *** compute gerschgorin (over-)estimate of largest eigenvalue *** c k = 1 t1 = w(diag) + w(1) if (p .le. 1) go to 170 do 160 i = 2, p j = diag0 + i t = w(j) + w(i) if (t .le. t1) go to 160 t1 = t k = i 160 continue c 170 sk = w(k) j = diag0 + k akk = w(j) k1 = k*(k-1)/2 + 1 inc = 1 t = zero do 200 i = 1, p if (i .eq. k) go to 180 aki = abs(dihdi(k1)) si = w(i) j = diag0 + i t1 = half * (w(j) + si - aki - akk) t1 = t1 + sqrt(t1*t1 + sk*aki) if (t .lt. t1) t = t1 if (i .lt. k) go to 190 180 inc = i 190 k1 = k1 + inc 200 continue c w(emax) = akk + t lk = amax1(lk, v(dgnorm)/rad - w(emax)) c c *** alphak = current value of alpha (see alg. notes above). we c *** use more*s scheme for initializing it. alphak = abs(v(stppar)) * v(rad0)/rad c if (irc .ne. 0) go to 210 c c *** compute l0 for positive definite h *** c call livmul(p, w, l, w(q)) t = v2norm(p, w) w(phipin) = dst / t / t lk = amax1(lk, phi*w(phipin)) c c *** safeguard alphak and add alphak*i to (d**-1)*h*(d**-1) *** c 210 ka = ka + 1 if (-v(dst0) .ge. alphak .or. alphak .lt. lk .or. alphak .ge. uk) 1 alphak = uk * amax1(p001, sqrt(lk/uk)) if (alphak .le. zero) alphak = half * uk if (alphak .le. zero) alphak = uk k = 0 do 220 i = 1, p k = k + i j = diag0 + i dihdi(k) = w(j) + alphak 220 continue c c *** try computing cholesky decomposition *** c call lsqrt(1, p, l, dihdi, irc) if (irc .eq. 0) go to 240 c c *** (d**-1)*h*(d**-1) + alphak*i is indefinite -- overestimate c *** smallest eigenvalue for use in updating lk *** c j = (irc*(irc+1))/2 t = l(j) l(j) = one do 230 i = 1, irc 230 w(i) = zero w(irc) = one call litvmu(irc, w, l, w) t1 = v2norm(irc, w) lk = alphak - t/t1/t1 v(dst0) = -lk go to 210 c c *** alphak makes (d**-1)*h*(d**-1) positive definite. c *** compute q = -d*step, check for convergence. *** c 240 call livmul(p, w(q), l, dig) gtsta = dotprd(p, w(q), w(q)) call litvmu(p, w(q), l, w(q)) dst = v2norm(p, w(q)) phi = dst - rad if (phi .le. phimax .and. phi .ge. phimin) go to 270 if (phi .eq. oldphi) go to 270 oldphi = phi if (phi .lt. zero) go to 330 c c *** unacceptable alphak -- update lk, uk, alphak *** c 250 if (ka .ge. kalim) go to 270 c *** the following amin1 is necessary because of restarts *** if (phi .lt. zero) uk = amin1(uk, alphak) c *** kamin = 0 only iff the gradient vanishes *** if (kamin .eq. 0) go to 210 call livmul(p, w, l, w(q)) t1 = v2norm(p, w) alphak = alphak + (phi/t1) * (dst/t1) * (dst/rad) lk = amax1(lk, alphak) go to 210 c c *** acceptable step on first try *** c 260 alphak = zero c c *** successful step in general. compute step = -(d**-1)*q *** c 270 do 280 i = 1, p j = q0 + i step(i) = -w(j)/d(i) 280 continue v(gtstep) = -gtsta v(preduc) = half * ( abs(alphak)*dst*dst + gtsta) go to 410 c c c *** restart with new radius *** c 290 if (v(dst0) .le. zero .or. v(dst0) - rad .gt. phimax) go to 310 c c *** prepare to return newton step *** c restrt = .true. ka = ka + 1 k = 0 do 300 i = 1, p k = k + i j = diag0 + i dihdi(k) = w(j) 300 continue uk = negone go to 30 c 310 kamin = ka + 3 if (v(dgnorm) .eq. zero) kamin = 0 if (ka .eq. 0) go to 50 c dst = w(dstsav) alphak = abs(v(stppar)) phi = dst - rad t = v(dgnorm)/rad uk = t - w(emin) if (v(dgnorm) .eq. zero) uk = uk + p001 + p001*uk if (uk .le. zero) uk = p001 if (rad .gt. v(rad0)) go to 320 c c *** smaller radius *** lk = zero if (alphak .gt. zero) lk = w(lk0) lk = amax1(lk, t - w(emax)) if (v(dst0) .gt. zero) lk = amax1(lk, (v(dst0)-rad)*w(phipin)) go to 250 c c *** bigger radius *** 320 if (alphak .gt. zero) uk = amin1(uk, w(uk0)) lk = amax1(zero, -v(dst0), t - w(emax)) if (v(dst0) .gt. zero) lk = amax1(lk, (v(dst0)-rad)*w(phipin)) go to 250 c c *** decide whether to check for special case... in practice (from c *** the standpoint of the calling optimization code) it seems best c *** not to check until a few iterations have failed -- hence the c *** test on kamin below. c 330 delta = alphak + amin1(zero, v(dst0)) twopsi = alphak*dst*dst + gtsta if (ka .ge. kamin) go to 340 c *** if the test in ref. 2 is satisfied, fall through to handle c *** the special case (as soon as the more-sorensen test detects c *** it). if (delta .ge. psifac*twopsi) go to 370 c c *** check for the special case of h + alpha*d**2 (nearly) c *** singular. use one step of inverse power method with start c *** from lsvmin to obtain approximate eigenvector corresponding c *** to smallest eigenvalue of (d**-1)*h*(d**-1). lsvmin returns c *** x and w with l*w = x. c 340 t = lsvmin(p, l, w(x), w) c c *** normalize w *** do 350 i = 1, p 350 w(i) = t*w(i) c *** complete current inv. power iter. -- replace w by (l**-t)*w. call litvmu(p, w, l, w) t2 = one/v2norm(p, w) do 360 i = 1, p 360 w(i) = t2*w(i) t = t2 * t c c *** now w is the desired approximate (unit) eigenvector and c *** t*x = ((d**-1)*h*(d**-1) + alphak*i)*w. c sw = dotprd(p, w(q), w) t1 = (rad + dst) * (rad - dst) root = sqrt(sw*sw + t1) if (sw .lt. zero) root = -root si = t1 / (sw + root) c c *** the actual test for the special case... c if ((t2*si)**2 .le. eps*(dst**2 + alphak*radsq)) go to 380 c c *** update upper bound on smallest eigenvalue (when not positive) c *** (as recommended by more and sorensen) and continue... c if (v(dst0) .le. zero) v(dst0) = amin1(v(dst0), t2**2 - alphak) lk = amax1(lk, -v(dst0)) c c *** check whether we can hope to detect the special case in c *** the available arithmetic. accept step as it is if not. c c *** if not yet available, obtain machine dependent value dgxfac. 370 if (dgxfac .eq. zero) dgxfac = epsfac * rmdcon(3) c if (delta .gt. dgxfac*w(dggdmx)) go to 250 go to 270 c c *** special case detected... negate alphak to indicate special case c 380 alphak = -alphak v(preduc) = half * twopsi c c *** accept current step if adding si*w would lead to a c *** further relative reduction in psi of less than v(epslon)/3. c t1 = zero t = si*(alphak*sw - half*si*(alphak + t*dotprd(p,w(x),w))) if (t .lt. eps*twopsi/six) go to 390 v(preduc) = v(preduc) + t dst = rad t1 = -si 390 do 400 i = 1, p j = q0 + i w(j) = t1*w(i) - w(j) step(i) = w(j) / d(i) 400 continue v(gtstep) = dotprd(p, dig, w(q)) c c *** save values for use in a possible restart *** c 410 v(dstnrm) = dst v(stppar) = alphak w(lk0) = lk w(uk0) = uk v(rad0) = rad w(dstsav) = dst c c *** restore diagonal of dihdi *** c j = 0 do 420 i = 1, p j = j + i k = diag0 + i dihdi(j) = w(k) 420 continue c 999 return c c *** last card of gqtst follows *** end subroutine lsqrt(n1, n, l, a, irc) c c *** compute rows n1 through n of the cholesky factor l of c *** a = l*(l**t), where l and the lower triangle of a are both c *** stored compactly by rows (and may occupy the same storage). c *** irc = 0 means all went well. irc = j means the leading c *** principal j x j submatrix of a is not positive definite -- c *** and l(j*(j+1)/2) contains the (nonpos.) reduced j-th diagonal. c c *** parameters *** c integer n1, n, irc real l(1), a(1) c dimension l(n*(n+1)/2), a(n*(n+1)/2) c c *** local variables *** c integer i, ij, ik, im1, i0, j, jk, jm1, j0, k real t, td, zero c c *** intrinsic functions *** c/+ real sqrt c/ c/6 data zero/0.e+0/ c/7 c parameter (zero=0.e+0) c/ c c *** body *** c i0 = n1 * (n1 - 1) / 2 do 50 i = n1, n td = zero if (i .eq. 1) go to 40 j0 = 0 im1 = i - 1 do 30 j = 1, im1 t = zero if (j .eq. 1) go to 20 jm1 = j - 1 do 10 k = 1, jm1 ik = i0 + k jk = j0 + k t = t + l(ik)*l(jk) 10 continue 20 ij = i0 + j j0 = j0 + j t = (a(ij) - t) / l(j0) l(ij) = t td = td + t*t 30 continue 40 i0 = i0 + i t = a(i0) - td if (t .le. zero) go to 60 l(i0) = sqrt(t) 50 continue c irc = 0 go to 999 c 60 l(i0) = t irc = i c 999 return c c *** last card of lsqrt *** end real function lsvmin(p, l, x, y) c c *** estimate smallest sing. value of packed lower triang. matrix l c c *** parameter declarations *** c integer p real l(1), x(p), y(p) c dimension l(p*(p+1)/2) c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** purpose *** c c this function returns a good over-estimate of the smallest c singular value of the packed lower triangular matrix l. c c *** parameter description *** c c p (in) = the order of l. l is a p x p lower triangular matrix. c l (in) = array holding the elements of l in row order, i.e. c l(1,1), l(2,1), l(2,2), l(3,1), l(3,2), l(3,3), etc. c x (out) if lsvmin returns a positive value, then x is a normalized c approximate left singular vector corresponding to the c smallest singular value. this approximation may be very c crude. if lsvmin returns zero, then some components of x c are zero and the rest retain their input values. c y (out) if lsvmin returns a positive value, then y = (l**-1)*x is an c unnormalized approximate right singular vector correspond- c ing to the smallest singular value. this approximation c may be crude. if lsvmin returns zero, then y retains its c input value. the caller may pass the same vector for x c and y (nonstandard fortran usage), in which case y over- c writes x (for nonzero lsvmin returns). c c *** algorithm notes *** c c the algorithm is based on (1), with the additional provision that c lsvmin = 0 is returned if the smallest diagonal element of l c (in magnitude) is not more than the unit roundoff times the c largest. the algorithm uses a random number generator proposed c in (4), which passes the spectral test with flying colors -- see c (2) and (3). c c *** subroutines and functions called *** c c v2norm - function, returns the 2-norm of a vector. c c *** references *** c c (1) cline, a., moler, c., stewart, g., and wilkinson, j.h.(1977), c an estimate for the condition number of a matrix, report c tm-310, applied math. div., argonne national laboratory. c c (2) hoaglin, d.c. (1976), theoretical properties of congruential c random-number generators -- an empirical view, c memorandum ns-340, dept. of statistics, harvard univ. c c (3) knuth, d.e. (1969), the art of computer programming, vol. 2 c (seminumerical algorithms), addison-wesley, reading, mass. c c (4) smith, c.s. (1971), multiplicative pseudo-random number c generators with prime modulus, j. assoc. comput. mach. 18, c pp. 586-593. c c *** history *** c c designed and coded by david m. gay (winter 1977/summer 1978). c c *** general *** c c this subroutine was written in connection with research c supported by the national science foundation under grants c mcs-7600324, dcr75-10143, 76-14311dss, and mcs76-11989. c c+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ c c *** local variables *** c integer i, ii, ix, j, ji, jj, jjj, jm1, j0, pm1 real b, sminus, splus, t, xminus, xplus c c *** constants *** c real half, one, r9973, zero c c *** intrinsic functions *** c/+ integer mod real abs, float c/ c *** external functions and subroutines *** c external dotprd, v2norm, vaxpy real dotprd, v2norm c c/6 data half/0.5e+0/, one/1.e+0/, r9973/9973.e+0/, zero/0.e+0/ c/7 c parameter (half=0.5e+0, one=1.e+0, r9973=9973.e+0, zero=0.e+0) c/ c c *** body *** c ix = 2 pm1 = p - 1 c c *** first check whether to return lsvmin = 0 and initialize x *** c ii = 0 j0 = p*pm1/2 jj = j0 + p if (l(jj) .eq. zero) go to 110 ix = mod(3432*ix, 9973) b = half*(one + float(ix)/r9973) xplus = b / l(jj) x(p) = xplus if (p .le. 1) go to 60 do 10 i = 1, pm1 ii = ii + i if (l(ii) .eq. zero) go to 110 ji = j0 + i x(i) = xplus * l(ji) 10 continue c c *** solve (l**t)*x = b, where the components of b have randomly c *** chosen magnitudes in (.5,1) with signs chosen to make x large. c c do j = p-1 to 1 by -1... do 50 jjj = 1, pm1 j = p - jjj c *** determine x(j) in this iteration. note for i = 1,2,...,j c *** that x(i) holds the current partial sum for row i. ix = mod(3432*ix, 9973) b = half*(one + float(ix)/r9973) xplus = (b - x(j)) xminus = (-b - x(j)) splus = abs(xplus) sminus = abs(xminus) jm1 = j - 1 j0 = j*jm1/2 jj = j0 + j xplus = xplus/l(jj) xminus = xminus/l(jj) if (jm1 .eq. 0) go to 30 do 20 i = 1, jm1 ji = j0 + i splus = splus + abs(x(i) + l(ji)*xplus) sminus = sminus + abs(x(i) + l(ji)*xminus) 20 continue 30 if (sminus .gt. splus) xplus = xminus x(j) = xplus c *** update partial sums *** if (jm1 .gt. 0) call vaxpy(jm1, x, xplus, l(j0+1), x) 50 continue c c *** normalize x *** c 60 t = one/v2norm(p, x) do 70 i = 1, p 70 x(i) = t*x(i) c c *** solve l*y = x and return lsvmin = 1/twonorm(y) *** c do 100 j = 1, p jm1 = j - 1 j0 = j*jm1/2 jj = j0 + j t = zero if (jm1 .gt. 0) t = dotprd(jm1, l(j0+1), y) y(j) = (x(j) - t) / l(jj) 100 continue c lsvmin = one/v2norm(p, y) go to 999 c 110 lsvmin = zero 999 return c *** last card of lsvmin follows *** end subroutine slvmul(p, y, s, x) c c *** set y = s * x, s = p x p symmetric matrix. *** c *** lower triangle of s stored rowwise. *** c c *** parameter declarations *** c integer p real s(1), x(p), y(p) c dimension s(p*(p+1)/2) c c *** local variables *** c integer i, im1, j, k real xi c c *** no intrinsic functions *** c c *** external function *** c external dotprd real dotprd c c----------------------------------------------------------------------- c j = 1 do 10 i = 1, p y(i) = dotprd(i, s(j), x) j = j + i 10 continue c if (p .le. 1) go to 999 j = 1 do 40 i = 2, p xi = x(i) im1 = i - 1 j = j + 1 do 30 k = 1, im1 y(k) = y(k) + s(j)*xi j = j + 1 30 continue 40 continue c 999 return c *** last card of slvmul follows *** end c *** simple test program for smsno, sumsl, and humsl *** c c in these examples, n = 4 and f(x) = (1.0 + 0.5*(x1**t)*a*x1)**0.5, c where x1(i) = d1(i)*x(i) - i, **t denotes transpose, and a is a c matrix having fives on the main diagonal and ones everywhere else. c the scale vector d1 is passed to qdrtf, the subroutine that c evaluates f, as part of urparm. specifically, the matrix urp c declared below is passed for ufparm, and d1 is urp(*,1), the first c column of urp. this main program repeatedly minimizes f, starting c from x = 0, by calling smsno, sumsl, and humsl. we actually c use two different objective functions, since we change d1 after c the first call on sumsl. all runs but the last use d = d1. c c f(x) is minimized at x1 = 0 (a vector of zeros), i.e., at c x(i) = i/d1(i). c external deflt, humsl, imdcon, qdrtf, qdrtg, qdrtgh, smsno, 1 sumsl integer imdcon c c imdcon - supplies nout, the output unit number. c qdrtf - passed for calcf to smsno, sumsl, and humsl. c qdrtg - passed for calcg to sumsl. c qdrtgh - passed for calcgh to humsl. c integer i, iv(60), liv, lv, nout, uip(1) real d(4), urp(4,3), v(150), x(4) c c the length of v is dictated by humsl. c data liv/60/, lv/150/ c c initialize nout, d, d1, and x. c nout = imdcon(1) c do 10 i = 1, 4 d(i) = 1.e+0 urp(i,1) = d(i) x(i) = 0.e+0 10 continue write(nout,20) 20 format(/16h smsno on qdrtf ) c c before this first call, we set iv(1) to 0 so that all input c components of iv and v will be given default values. before c subsequent calls, we set iv(1) to 12 so that the old input values c of iv and v are used. c c qdrtf does not make use of ufparm. in calling smsno, we c arbitrarily pass qdrtf for ufparm to satifsy the calling sequence. c iv(1) = 0 call smsno(4, d, x, qdrtf, iv, liv, lv, v, uip, urp, qdrtf) c c we reinitialize x and minimize f again, this time using sumsl. c qdrtg, the subroutine passed for calcg, assumes that ufparm is qdrtf. c do 30 i = 1, 4 30 x(i) = 0.e+0 write(nout,40) 40 format(/16h sumsl on qdrtf ) c iv(1) = 12 call sumsl(4, d, x, qdrtf, qdrtg, iv, liv, lv, v, uip,urp,qdrtf) c c now we modify f by using a different choice of d1. we still use c d = d1, so the performance of sumsl should stay the same -- only d c and the final x and gradient should be affected. c do 50 i = 1, 4 x(i) = 0.e+0 d(i) = 1.e2 ** i urp(i,1) = d(i) 50 continue write(nout,40) c iv(1) = 12 call sumsl(4, d, x, qdrtf, qdrtg, iv, liv, lv, v, uip,urp,qdrtf) c c using the last choice of d and d1, we now use humsl to minimize f. c like qdrtg, qdrtgh assumes that ufparm is qdrtf. c do 60 i = 1, 4 60 x(i) = 0.e+0 write(nout,70) 70 format(/16h humsl on qdrtf ) c iv(1) = 12 call humsl(4, d, x, qdrtf, qdrtgh, iv, liv, lv, v, uip,urp,qdrtf) c c we repeat the last run with iv(dtype) = 1 and v(dinit) = 0.0, so c that humsl will determine d from the diagonal of the hessian. c this run also demonstrates the use of subroutine deflt and the c passing of nondefault parameters. (since the iv and v input c components still have their default values at his point, it is not c really necessary to call deflt. it is necessary to reset iv(1) to c 12, however, and deflt does this for us.) c write(nout,80) 80 format(/18h humsl updating d ) c do 90 i = 1, 4 90 x(i) = 0.e+0 c call deflt(2, iv, liv, lv, v) iv(16) = 1 v(38) = 0.e+0 call humsl(4, d, x, qdrtf, qdrtgh, iv, liv, lv, v, uip,urp,qdrtf) c 999 stop end c c*********************************************************************** c c q d r t f c c*********************************************************************** c subroutine qdrtf(n, x, nf, f, uip, urp, ufp) c c this routine evaluates the objective function f(x) described in the c main program above. it stores in urp(*,2) and urp(*,3) some c information useful in evaluating the gradient and hessian of f, and c it stores nf in uip(1) to identify the x corresponding to this c information. f(x) has the form f(x) = phi(q(x)), where q is a c quadratic form and phi(y) = y**0.5. the gradient of f is c g(x) = phiprm(q(x))*gq(x), where phiprm is the derivative of phi c and gq is the gradient of q. this routine stores phiprm(q(x)) in c urp(1,3) and gq(x) in urp(*,2). the hessian of f is c h(x) = phi2prm(q(x))*gq(x)*gq(x)**t + phiprm(q(x))*hq(x), where c phi2prm is the second derivative of phi, **t denotes transpose, c and hq is the hessian of q. this routine stores phi2prm(q(x)) in c urp(2,3). the subroutines qdrtg and qdrtgh given below would work c without change on any other choice of phi. qdrtg would also work c with any other differentiable function q. qdrtgh, on the other c hand, assumes that hq(x) is the matrix a described in the main c program above. c integer n, nf, uip(1) real x(n), f, urp(n,3) external ufp c/+ real float real sqrt c/ integer i real dn, f2, t, t1 c c uip(1) = nf dn = n t = 0.e+0 do 10 i = 1, n urp(i,2) = urp(i,1)*x(i) - float(i) t = t + urp(i,2) 10 continue f2 = 0.e+0 do 20 i = 1, n t1 = dn*urp(i,2) + t f2 = f2 + t1*urp(i,2) urp(i,2) = urp(i,1) * t1 20 continue f2 = 1.e+0 + 0.5e+0 * f2 f = sqrt(f2) urp(1,3) = 0.5e+0 / f urp(2,3) = -0.5e+0 / (f * f2) 999 return end c c*********************************************************************** c c q d r t g c c*********************************************************************** c subroutine qdrtg(n, x, nf, g, uip, urp, qdrtf) c c this routine evaluates the gradient of the objective function f(x) c described in the main program above. see the comments there and in c subroutine qdrtf above. c integer n, nf, uip(1) real x(n), g(n), urp(n,3) external qdrtf c integer i real f c if (nf .ne. uip(1)) call qdrtf(n, x, nf, f, uip, urp, qdrtf) do 10 i = 1, n 10 g(i) = urp(1,3) * urp(i,2) 999 return end c c*********************************************************************** c c q d r t g h c c*********************************************************************** c subroutine qdrtgh(n, x, nf, g, h, uip, urp, qdrtf) c c this routine evaluates the gradient and hessian of the objective c function f(x) described in the main program above. see the comments c there and in subroutine qdrtf above. note that the h returned is c the lower triangle of the hessian, stored row-wise. c integer n, nf, uip(1) real x(n), g(n), h(1), urp(n,3) c dimension h(n*(n+1)/2) external qdrtf c integer i, j, k real dn, f, t1, t2 c if (nf .ne. uip(1)) call qdrtf(n, x, nf, f, uip, urp, qdrtf) k = 0 dn = n do 20 i = 1, n g(i) = urp(1,3) * urp(i,2) t1 = urp(1,3) * urp(i,1) t2 = urp(i,2) * urp(2,3) do 10 j = 1, i k = k + 1 h(k) = t2*urp(j,2) + t1*urp(j,1) 10 continue h(k) = h(k) + dn*urp(i,1)*t1 20 continue 999 return end