1 /*
  2  Copyright 2008-2024
  3  Matthias Ehmann,
  4  Carsten Miller,
  5  Reinhard Oldenburg,
  6  Andreas Walter,
  7  Alfred Wassermann
  8 
  9  This file is part of JSXGraph.
 10 
 11  JSXGraph is free software dual licensed under the GNU LGPL or MIT License.
 12 
 13  You can redistribute it and/or modify it under the terms of the
 14 
 15  * GNU Lesser General Public License as published by
 16  the Free Software Foundation, either version 3 of the License, or
 17  (at your option) any later version
 18  OR
 19  * MIT License: https://github.com/jsxgraph/jsxgraph/blob/master/LICENSE.MIT
 20 
 21  JSXGraph is distributed in the hope that it will be useful,
 22  but WITHOUT ANY WARRANTY; without even the implied warranty of
 23  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 24  GNU Lesser General Public License for more details.
 25 
 26  You should have received a copy of the GNU Lesser General Public License and
 27  the MIT License along with JSXGraph. If not, see <https://www.gnu.org/licenses/>
 28  and <https://opensource.org/licenses/MIT/>.
 29 
 30  This is a port of jcobyla
 31 
 32  - to JavaScript by Reihard Oldenburg and
 33  - to JSXGraph by Alfred Wassermann
 34  - optimized by Andreas Walter
 35  */
 36 /*
 37  * jcobyla
 38  *
 39  * The MIT License
 40  *
 41  * Copyright (c) 2012 Anders Gustafsson, Cureos AB.
 42  *
 43  * Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files
 44  * (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge,
 45  * publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so,
 46  * subject to the following conditions:
 47  *
 48  * The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
 49  *
 50  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 51  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE
 52  * FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
 53  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 54  *
 55  * Remarks:
 56  *
 57  * The original Fortran 77 version of this code was by Michael Powell (M.J.D.Powell @ damtp.cam.ac.uk)
 58  * The Fortran 90 version was by Alan Miller (Alan.Miller @ vic.cmis.csiro.au). Latest revision - 30 October 1998
 59  */
 60 
 61 /**
 62  * Constrained Optimization BY Linear Approximation in Java.
 63  *
 64  * COBYLA2 is an implementation of Powell's nonlinear derivative free constrained optimization that uses
 65  * a linear approximation approach. The algorithm is a sequential trust region algorithm that employs linear
 66  * approximations to the objective and constraint functions, where the approximations are formed by linear
 67  * interpolation at n + 1 points in the space of the variables and tries to maintain a regular shaped simplex
 68  * over iterations.
 69  *
 70  * It solves nonsmooth NLP with a moderate number of variables (about 100). Inequality constraints only.
 71  *
 72  * The initial point X is taken as one vertex of the initial simplex with zero being another, so, X should
 73  * not be entered as the zero vector.
 74  *
 75  * @author Anders Gustafsson, Cureos AB. Translation to Javascript by Reinhard Oldenburg, Goethe-University
 76  */
 77 
 78 /*global JXG: true, define: true*/
 79 /*jslint nomen: true, plusplus: true, continue: true*/
 80 
 81 import JXG from "../jxg.js";
 82 import Mat from "./math.js";
 83 // import Type from "../utils/type.js";
 84 
 85 /**
 86  * The JXG.Math.Nlp namespace holds numerical algorithms for non-linear optimization.
 87  * @name JXG.Math.Nlp
 88  * @namespace
 89  *
 90  */
 91 JXG.Math.Nlp = {
 92     arr: function (n) {
 93         // Is 0 initialized
 94         return new Float64Array(n);
 95     },
 96 
 97     arr2: function (n, m) {
 98         var i = 0,
 99             a = new Array(n);
100 
101         while (i < n) {
102             a[i] = this.arr(m);
103             i++;
104         }
105         return a;
106     },
107 
108     arraycopy: function (x, a, iox, b, n) {
109         var i = 0;
110         while (i < n) {
111             iox[i + b] = x[i + a];
112             i++;
113         }
114     },
115 
116     lastNumberOfEvaluations: 0,
117     GetLastNumberOfEvaluations: function () {
118         return this.lastNumberOfEvaluations;
119     },
120     // status Variables
121     Normal: 0,
122     MaxIterationsReached: 1,
123     DivergingRoundingErrors: 2,
124 
125     /**
126      * Minimizes the objective function F with respect to a set of inequality constraints CON,
127      * and returns the optimal variable array. F and CON may be non-linear, and should preferably be smooth.
128      * Calls {@link JXG.Math.Nlp#cobylb}.
129      *
130      * @param calcfc Interface implementation for calculating objective function and constraints.
131      * @param n Number of variables.
132      * @param m Number of constraints.
133      * @param x On input initial values of the variables (zero-based array). On output
134      * optimal values of the variables obtained in the COBYLA minimization.
135      * @param rhobeg Initial size of the simplex.
136      * @param rhoend Final value of the simplex.
137      * @param iprint Print level, 0 <= iprint <= 3, where 0 provides no output and
138      * 3 provides full output to the console.
139      * @param maxfun Maximum number of function evaluations before terminating.
140      * @param [testForRoundingErrors=false]
141      * @returns {Number} Exit status of the COBYLA2 optimization.
142      */
143     FindMinimum: function (calcfc, n, m, x, rhobeg, rhoend, iprint, maxfun, testForRoundingErrors) {
144         // CobylaExitStatus FindMinimum(final Calcfc calcfc, int n, int m, double[] x, double rhobeg, double rhoend, int iprint, int maxfun)
145         //     This subroutine minimizes an objective function F(X) subject to M
146         //     inequality constraints on X, where X is a vector of variables that has
147         //     N components.  The algorithm employs linear approximations to the
148         //     objective and constraint functions, the approximations being formed by
149         //     linear interpolation at N+1 points in the space of the variables.
150         //     We regard these interpolation points as vertices of a simplex.  The
151         //     parameter RHO controls the size of the simplex and it is reduced
152         //     automatically from RHOBEG to RHOEND.  For each RHO the subroutine tries
153         //     to achieve a good vector of variables for the current size, and then
154         //     RHO is reduced until the value RHOEND is reached.  Therefore RHOBEG and
155         //     RHOEND should be set to reasonable initial changes to and the required
156         //     accuracy in the variables respectively, but this accuracy should be
157         //     viewed as a subject for experimentation because it is not guaranteed.
158         //     The subroutine has an advantage over many of its competitors, however,
159         //     which is that it treats each constraint individually when calculating
160         //     a change to the variables, instead of lumping the constraints together
161         //     into a single penalty function.  The name of the subroutine is derived
162         //     from the phrase Constrained Optimization BY Linear Approximations.
163 
164         //     The user must set the values of N, M, RHOBEG and RHOEND, and must
165         //     provide an initial vector of variables in X.  Further, the value of
166         //     IPRINT should be set to 0, 1, 2 or 3, which controls the amount of
167         //     printing during the calculation. Specifically, there is no output if
168         //     IPRINT=0 and there is output only at the end of the calculation if
169         //     IPRINT=1.  Otherwise each new value of RHO and SIGMA is printed.
170         //     Further, the vector of variables and some function information are
171         //     given either when RHO is reduced or when each new value of F(X) is
172         //     computed in the cases IPRINT=2 or IPRINT=3 respectively. Here SIGMA
173         //     is a penalty parameter, it being assumed that a change to X is an
174         //     improvement if it reduces the merit function
175         //                F(X)+SIGMA*MAX(0.0, - C1(X), - C2(X),..., - CM(X)),
176         //     where C1,C2,...,CM denote the constraint functions that should become
177         //     nonnegative eventually, at least to the precision of RHOEND. In the
178         //     printed output the displayed term that is multiplied by SIGMA is
179         //     called MAXCV, which stands for 'MAXimum Constraint Violation'.  The
180         //     argument ITERS is an integer variable that must be set by the user to a
181         //     limit on the number of calls of CALCFC, the purpose of this routine being
182         //     given below.  The value of ITERS will be altered to the number of calls
183         //     of CALCFC that are made.
184         //     In order to define the objective and constraint functions, we require
185         //     a subroutine that has the name and arguments
186         //                SUBROUTINE CALCFC (N,M,X,F,CON)
187         //                DIMENSION X(:),CON(:)  .
188         //     The values of N and M are fixed and have been defined already, while
189         //     X is now the current vector of variables. The subroutine should return
190         //     the objective and constraint functions at X in F and CON(1),CON(2),
191         //     ...,CON(M).  Note that we are trying to adjust X so that F(X) is as
192         //     small as possible subject to the constraint functions being nonnegative.
193 
194         // Local variables
195         var mpp = m + 2,
196             status,
197             // Internal base-1 X array
198             iox = this.arr(n + 1),
199             that = this,
200             fcalcfc;
201 
202         this.lastNumberOfEvaluations = 0;
203 
204         if (testForRoundingErrors) {
205             console.log("Experimental feature 'testForRoundingErrors' is activated.");
206         }
207 
208         iox[0] = 0.0;
209         this.arraycopy(x, 0, iox, 1, n);
210 
211         // Internal representation of the objective and constraints calculation method,
212         // accounting for that X and CON arrays in the cobylb method are base-1 arrays.
213         fcalcfc = function (n, m, thisx, con) {
214             // int n, int m, double[] x, double[] con
215             var ix = that.arr(n),
216                 ocon, f;
217 
218             that.lastNumberOfEvaluations = that.lastNumberOfEvaluations + 1;
219             that.arraycopy(thisx, 1, ix, 0, n);
220             ocon = that.arr(m);
221             f = calcfc(n, m, ix, ocon);
222             that.arraycopy(ocon, 0, con, 1, m);
223             return f;
224         };
225 
226         status = this.cobylb(fcalcfc, n, m, mpp, iox, rhobeg, rhoend, iprint, maxfun, testForRoundingErrors);
227         this.arraycopy(iox, 1, x, 0, n);
228 
229         return status;
230     },
231 
232     //    private static CobylaExitStatus cobylb(Calcfc calcfc, int n, int m, int mpp, double[] x,
233     //      double rhobeg, double rhoend, int iprint, int maxfun)
234     /**
235      * JavaScript implementation of the non-linear optimization method COBYLA.
236      * @param {Function} calcfc
237      * @param {Number} n
238      * @param {Number} m
239      * @param {Number} mpp
240      * @param {Number} x
241      * @param {Number} rhobeg
242      * @param {Number} rhoend
243      * @param {Number} iprint
244      * @param {Number} maxfun
245      * @param {Boolean} [testForRoundingErrors=false]
246      * @returns {Number} Exit status of the COBYLA2 optimization
247      */
248     cobylb: function (calcfc, n, m, mpp, x, rhobeg, rhoend, iprint, maxfun, testForRoundingErrors) {
249         // calcf ist funktion die aufgerufen wird wie calcfc(n, m, ix, ocon)
250         // N.B. Arguments CON, SIM, SIMI, DATMAT, A, VSIG, VETA, SIGBAR, DX, W & IACT
251         //      have been removed.
252 
253         //     Set the initial values of some parameters. The last column of SIM holds
254         //     the optimal vertex of the current simplex, and the preceding N columns
255         //     hold the displacements from the optimal vertex to the other vertices.
256         //     Further, SIMI holds the inverse of the matrix that is contained in the
257         //     first N columns of SIM.
258 
259         // Local variables
260         var status = -1,
261             alpha = 0.25,
262             beta = 2.1,
263             gamma = 0.5,
264             delta = 1.1,
265             f = 0.0,
266             resmax = 0.0,
267             total,
268             np = n + 1,
269             mp = m + 1,
270             rho = rhobeg,
271             parmu = 0.0,
272             iflag = false,
273             ifull = false,
274             parsig = 0.0,
275             prerec = 0.0,
276             prerem = 0.0,
277             con = this.arr(1 + mpp),
278             sim = this.arr2(1 + n, 1 + np),
279             simi = this.arr2(1 + n, 1 + n),
280             datmat = this.arr2(1 + mpp, 1 + np),
281             a = this.arr2(1 + n, 1 + mp),
282             vsig = this.arr(1 + n),
283             veta = this.arr(1 + n),
284             sigbar = this.arr(1 + n),
285             dx = this.arr(1 + n),
286             w = this.arr(1 + n),
287             i, j, k, l,
288             temp, tempa,
289             nfvals, jdrop, ibrnch, skipVertexIdent,
290             phimin, nbest, error, pareta, wsig, weta,
291             cvmaxp, cvmaxm, dxsign, resnew, barmu,
292             phi, vmold, vmnew, trured, ratio, edgmax,
293             cmin, cmax, denom,
294             endless = true;
295 
296         if (iprint >= 2) {
297             console.log("The initial value of RHO is " + rho + " and PARMU is set to zero.");
298         }
299 
300         nfvals = 0;
301         temp = 1.0 / rho;
302 
303         for (i = 1; i <= n; ++i) {
304             sim[i][np] = x[i];
305             sim[i][i] = rho;
306             simi[i][i] = temp;
307         }
308 
309         jdrop = np;
310         ibrnch = false;
311 
312         //     Make the next call of the user-supplied subroutine CALCFC. These
313         //     instructions are also used for calling CALCFC during the iterations of
314         //     the algorithm.
315         //alert("Iteration "+nfvals+" x="+x);
316         L_40: do {
317             if (nfvals >= maxfun && nfvals > 0) {
318                 status = this.MaxIterationsReached;
319                 break L_40;
320             }
321 
322             ++nfvals;
323             f = calcfc(n, m, x, con);
324             resmax = 0.0;
325             for (k = 1; k <= m; ++k) {
326                 resmax = Math.max(resmax, -con[k]);
327             }
328             //alert(    "   f="+f+"  resmax="+resmax);
329 
330             if (nfvals === iprint - 1 || iprint === 3) {
331                 this.PrintIterationResult(nfvals, f, resmax, x, n, iprint);
332             }
333 
334             con[mp] = f;
335             con[mpp] = resmax;
336 
337             //     Set the recently calculated function values in a column of DATMAT. This
338             //     array has a column for each vertex of the current simplex, the entries of
339             //     each column being the values of the constraint functions (if any)
340             //     followed by the objective function and the greatest constraint violation
341             //     at the vertex.
342             skipVertexIdent = true;
343             if (!ibrnch) {
344                 skipVertexIdent = false;
345 
346                 for (i = 1; i <= mpp; ++i) {
347                     datmat[i][jdrop] = con[i];
348                 }
349 
350                 if (nfvals <= np) {
351                     //     Exchange the new vertex of the initial simplex with the optimal vertex if
352                     //     necessary. Then, if the initial simplex is not complete, pick its next
353                     //     vertex and calculate the function values there.
354 
355                     if (jdrop <= n) {
356                         if (datmat[mp][np] <= f) {
357                             x[jdrop] = sim[jdrop][np];
358                         } else {
359                             sim[jdrop][np] = x[jdrop];
360                             for (k = 1; k <= mpp; ++k) {
361                                 datmat[k][jdrop] = datmat[k][np];
362                                 datmat[k][np] = con[k];
363                             }
364                             for (k = 1; k <= jdrop; ++k) {
365                                 sim[jdrop][k] = -rho;
366                                 temp = 0.0;
367                                 for (i = k; i <= jdrop; ++i) {
368                                     temp -= simi[i][k];
369                                 }
370                                 simi[jdrop][k] = temp;
371                             }
372                         }
373                     }
374                     if (nfvals <= n) {
375                         jdrop = nfvals;
376                         x[jdrop] += rho;
377                         continue L_40;
378                     }
379                 }
380                 ibrnch = true;
381             }
382 
383             L_140: do {
384                 L_550: do {
385                     if (!skipVertexIdent) {
386                         //     Identify the optimal vertex of the current simplex.
387                         phimin = datmat[mp][np] + parmu * datmat[mpp][np];
388                         nbest = np;
389 
390                         for (j = 1; j <= n; ++j) {
391                             temp = datmat[mp][j] + parmu * datmat[mpp][j];
392                             if (temp < phimin) {
393                                 nbest = j;
394                                 phimin = temp;
395                             } else if (
396                                 temp === phimin &&
397                                 parmu === 0.0 &&
398                                 datmat[mpp][j] < datmat[mpp][nbest]
399                             ) {
400                                 nbest = j;
401                             }
402                         }
403 
404                         //     Switch the best vertex into pole position if it is not there already,
405                         //     and also update SIM, SIMI and DATMAT.
406                         if (nbest <= n) {
407                             for (i = 1; i <= mpp; ++i) {
408                                 temp = datmat[i][np];
409                                 datmat[i][np] = datmat[i][nbest];
410                                 datmat[i][nbest] = temp;
411                             }
412                             for (i = 1; i <= n; ++i) {
413                                 temp = sim[i][nbest];
414                                 sim[i][nbest] = 0.0;
415                                 sim[i][np] += temp;
416 
417                                 tempa = 0.0;
418                                 for (k = 1; k <= n; ++k) {
419                                     sim[i][k] -= temp;
420                                     tempa -= simi[k][i];
421                                 }
422                                 simi[nbest][i] = tempa;
423                             }
424                         }
425 
426                         //     Make an error return if SIGI is a poor approximation to the inverse of
427                         //     the leading N by N submatrix of SIG.
428                         error = 0.0;
429                         if (testForRoundingErrors) {
430                             for (i = 1; i <= n; ++i) {
431                                 for (j = 1; j <= n; ++j) {
432                                     temp =
433                                         this.DOT_PRODUCT_ROW_COL(simi, i, sim, j, 1, n) -
434                                         (i === j ? 1.0 : 0.0);
435                                     // temp = this.DOT_PRODUCT(
436                                     //     this.PART(this.ROW(simi, i), 1, n),
437                                     //     this.PART(this.COL(sim, j), 1, n)
438                                     // ) - (i === j ? 1.0 : 0.0);
439 
440                                     error = Math.max(error, Math.abs(temp));
441                                 }
442                             }
443                         }
444                         if (error > 0.1) {
445                             status = this.DivergingRoundingErrors;
446                             break L_40;
447                         }
448 
449                         //     Calculate the coefficients of the linear approximations to the objective
450                         //     and constraint functions, placing minus the objective function gradient
451                         //     after the constraint gradients in the array A. The vector W is used for
452                         //     working space.
453                         for (k = 1; k <= mp; ++k) {
454                             con[k] = -datmat[k][np];
455                             for (j = 1; j <= n; ++j) {
456                                 w[j] = datmat[k][j] + con[k];
457                             }
458 
459                             for (i = 1; i <= n; ++i) {
460                                 a[i][k] =
461                                     (k === mp ? -1.0 : 1.0) *
462                                     this.DOT_PRODUCT_ROW_COL(w, -1, simi, i, 1, n);
463                                 // this.DOT_PRODUCT(this.PART(w, 1, n), this.PART(this.COL(simi, i), 1, n));
464                             }
465                         }
466 
467                         //     Calculate the values of sigma and eta, and set IFLAG = 0 if the current
468                         //     simplex is not acceptable.
469                         iflag = true;
470                         parsig = alpha * rho;
471                         pareta = beta * rho;
472 
473                         for (j = 1; j <= n; ++j) {
474                             wsig = 0.0;
475                             weta = 0.0;
476                             for (k = 1; k <= n; ++k) {
477                                 wsig += simi[j][k] * simi[j][k];
478                                 weta += sim[k][j] * sim[k][j];
479                             }
480                             vsig[j] = 1.0 / Math.sqrt(wsig);
481                             veta[j] = Math.sqrt(weta);
482                             if (vsig[j] < parsig || veta[j] > pareta) {
483                                 iflag = false;
484                             }
485                         }
486 
487                         //     If a new vertex is needed to improve acceptability, then decide which
488                         //     vertex to drop from the simplex.
489                         if (!ibrnch && !iflag) {
490                             jdrop = 0;
491                             temp = pareta;
492                             for (j = 1; j <= n; ++j) {
493                                 if (veta[j] > temp) {
494                                     jdrop = j;
495                                     temp = veta[j];
496                                 }
497                             }
498                             if (jdrop === 0) {
499                                 for (j = 1; j <= n; ++j) {
500                                     if (vsig[j] < temp) {
501                                         jdrop = j;
502                                         temp = vsig[j];
503                                     }
504                                 }
505                             }
506 
507                             //     Calculate the step to the new vertex and its sign.
508                             temp = gamma * rho * vsig[jdrop];
509                             for (k = 1; k <= n; ++k) {
510                                 dx[k] = temp * simi[jdrop][k];
511                             }
512                             cvmaxp = 0.0;
513                             cvmaxm = 0.0;
514                             total = 0.0;
515                             for (k = 1; k <= mp; ++k) {
516                                 // total = this.DOT_PRODUCT(this.PART(this.COL(a, k), 1, n), this.PART(dx, 1, n));
517                                 total = this.DOT_PRODUCT_ROW_COL(dx, -1, a, k, 1, n);
518                                 if (k < mp) {
519                                     temp = datmat[k][np];
520                                     cvmaxp = Math.max(cvmaxp, -total - temp);
521                                     cvmaxm = Math.max(cvmaxm, total - temp);
522                                 }
523                             }
524                             dxsign = parmu * (cvmaxp - cvmaxm) > 2.0 * total ? -1.0 : 1.0;
525 
526                             //     Update the elements of SIM and SIMI, and set the next X.
527                             temp = 0.0;
528                             for (i = 1; i <= n; ++i) {
529                                 dx[i] = dxsign * dx[i];
530                                 sim[i][jdrop] = dx[i];
531                                 temp += simi[jdrop][i] * dx[i];
532                             }
533                             for (k = 1; k <= n; ++k) {
534                                 simi[jdrop][k] /= temp;
535                             }
536 
537                             for (j = 1; j <= n; ++j) {
538                                 if (j !== jdrop) {
539                                     // temp = this.DOT_PRODUCT(this.PART(this.ROW(simi, j), 1, n), this.PART(dx, 1, n));
540                                     temp = this.DOT_PRODUCT_ROW_COL(simi, j, dx, -1, 1, n);
541                                     for (k = 1; k <= n; ++k) {
542                                         simi[j][k] -= temp * simi[jdrop][k];
543                                     }
544                                 }
545                                 x[j] = sim[j][np] + dx[j];
546                             }
547                             continue L_40;
548                         }
549 
550                         //     Calculate DX = x(*)-x(0).
551                         //     Branch if the length of DX is less than 0.5*RHO.
552                         ifull = this.trstlp(n, m, a, con, rho, dx);
553                         if (!ifull) {
554                             temp = 0.0;
555                             for (k = 1; k <= n; ++k) {
556                                 temp += dx[k] * dx[k];
557                             }
558                             if (temp < 0.25 * rho * rho) {
559                                 ibrnch = true;
560                                 break L_550;
561                             }
562                         }
563 
564                         //     Predict the change to F and the new maximum constraint violation if the
565                         //     variables are altered from x(0) to x(0) + DX.
566                         total = 0.0;
567                         resnew = 0.0;
568                         con[mp] = 0.0;
569                         for (k = 1; k <= mp; ++k) {
570                             //total = con[k] - this.DOT_PRODUCT(this.PART(this.COL(a, k), 1, n), this.PART(dx, 1, n));
571                             total = con[k] - this.DOT_PRODUCT_ROW_COL(dx, -1, a, k, 1, n);
572                             if (k < mp) {
573                                 resnew = Math.max(resnew, total);
574                             }
575                         }
576 
577                         //     Increase PARMU if necessary and branch back if this change alters the
578                         //     optimal vertex. Otherwise PREREM and PREREC will be set to the predicted
579                         //     reductions in the merit function and the maximum constraint violation
580                         //     respectively.
581                         prerec = datmat[mpp][np] - resnew;
582                         barmu = prerec > 0.0 ? total / prerec : 0.0;
583                         if (parmu < 1.5 * barmu) {
584                             parmu = 2.0 * barmu;
585                             if (iprint >= 2) {
586                                 console.log("Increase in PARMU to " + parmu);
587                             }
588                             phi = datmat[mp][np] + parmu * datmat[mpp][np];
589                             for (j = 1; j <= n; ++j) {
590                                 temp = datmat[mp][j] + parmu * datmat[mpp][j];
591                                 if (
592                                     temp < phi ||
593                                     (temp === phi &&
594                                         parmu === 0.0 &&
595                                         datmat[mpp][j] < datmat[mpp][np])
596                                 ) {
597                                     continue L_140;
598                                 }
599                             }
600                         }
601                         prerem = parmu * prerec - total;
602 
603                         //     Calculate the constraint and objective functions at x(*).
604                         //     Then find the actual reduction in the merit function.
605                         for (k = 1; k <= n; ++k) {
606                             x[k] = sim[k][np] + dx[k];
607                         }
608                         ibrnch = true;
609                         continue L_40;
610                     }
611 
612                     skipVertexIdent = false;
613                     vmold = datmat[mp][np] + parmu * datmat[mpp][np];
614                     vmnew = f + parmu * resmax;
615                     trured = vmold - vmnew;
616                     if (parmu === 0.0 && f === datmat[mp][np]) {
617                         prerem = prerec;
618                         trured = datmat[mpp][np] - resmax;
619                     }
620 
621                     //     Begin the operations that decide whether x(*) should replace one of the
622                     //     vertices of the current simplex, the change being mandatory if TRURED is
623                     //     positive. Firstly, JDROP is set to the index of the vertex that is to be
624                     //     replaced.
625                     ratio = trured <= 0.0 ? 1.0 : 0.0;
626                     jdrop = 0;
627                     for (j = 1; j <= n; ++j) {
628                         // temp = Math.abs(this.DOT_PRODUCT(this.PART(this.ROW(simi, j), 1, n), this.PART(dx, 1, n)));
629                         temp = Math.abs(this.DOT_PRODUCT_ROW_COL(simi, j, dx, -1, 1, n));
630                         if (temp > ratio) {
631                             jdrop = j;
632                             ratio = temp;
633                         }
634                         sigbar[j] = temp * vsig[j];
635                     }
636 
637                     //     Calculate the value of ell.
638                     edgmax = delta * rho;
639                     l = 0;
640                     for (j = 1; j <= n; ++j) {
641                         if (sigbar[j] >= parsig || sigbar[j] >= vsig[j]) {
642                             temp = veta[j];
643                             if (trured > 0.0) {
644                                 temp = 0.0;
645                                 for (k = 1; k <= n; ++k) {
646                                     temp += Math.pow(dx[k] - sim[k][j], 2.0);
647                                 }
648                                 temp = Math.sqrt(temp);
649                             }
650                             if (temp > edgmax) {
651                                 l = j;
652                                 edgmax = temp;
653                             }
654                         }
655                     }
656                     if (l > 0) {
657                         jdrop = l;
658                     }
659 
660                     if (jdrop !== 0) {
661                         //     Revise the simplex by updating the elements of SIM, SIMI and DATMAT.
662                         temp = 0.0;
663                         for (i = 1; i <= n; ++i) {
664                             sim[i][jdrop] = dx[i];
665                             temp += simi[jdrop][i] * dx[i];
666                         }
667                         for (k = 1; k <= n; ++k) {
668                             simi[jdrop][k] /= temp;
669                         }
670                         for (j = 1; j <= n; ++j) {
671                             if (j !== jdrop) {
672                                 // temp = this.DOT_PRODUCT(this.PART(this.ROW(simi, j), 1, n), this.PART(dx, 1, n));
673                                 temp = this.DOT_PRODUCT_ROW_COL(simi, j, dx, -1, 1, n);
674                                 for (k = 1; k <= n; ++k) {
675                                     simi[j][k] -= temp * simi[jdrop][k];
676                                 }
677                             }
678                         }
679                         for (k = 1; k <= mpp; ++k) {
680                             datmat[k][jdrop] = con[k];
681                         }
682 
683                         //     Branch back for further iterations with the current RHO.
684                         if (trured > 0.0 && trured >= 0.1 * prerem) {
685                             continue L_140;
686                         }
687                     }
688                     // If we end up here, we drop out.
689                 } while (!endless);
690 
691                 if (!iflag) {
692                     ibrnch = false;
693                     continue L_140;
694                 }
695 
696                 if (rho <= rhoend) {
697                     status = this.Normal;
698                     break L_40;
699                 }
700 
701                 //     Otherwise reduce RHO if it is not at its least value and reset PARMU.
702                 cmin = 0.0;
703                 cmax = 0.0;
704                 rho *= 0.5;
705                 if (rho <= 1.5 * rhoend) {
706                     rho = rhoend;
707                 }
708                 if (parmu > 0.0) {
709                     denom = 0.0;
710                     for (k = 1; k <= mp; ++k) {
711                         cmin = datmat[k][np];
712                         cmax = cmin;
713                         for (i = 1; i <= n; ++i) {
714                             cmin = Math.min(cmin, datmat[k][i]);
715                             cmax = Math.max(cmax, datmat[k][i]);
716                         }
717                         if (k <= m && cmin < 0.5 * cmax) {
718                             temp = Math.max(cmax, 0.0) - cmin;
719                             denom = denom <= 0.0 ? temp : Math.min(denom, temp);
720                         }
721                     }
722                     if (denom === 0.0) {
723                         parmu = 0.0;
724                     } else if (cmax - cmin < parmu * denom) {
725                         parmu = (cmax - cmin) / denom;
726                     }
727                 }
728                 if (iprint >= 2) {
729                     console.log("Reduction in RHO to " + rho + "  and PARMU = " + parmu);
730                 }
731                 if (iprint === 2) {
732                     this.PrintIterationResult(
733                         nfvals,
734                         datmat[mp][np],
735                         datmat[mpp][np],
736                         this.COL(sim, np),
737                         n,
738                         iprint
739                     );
740                 }
741             } while (endless);
742         } while (endless);
743 
744         switch (status) {
745             case this.Normal:
746                 if (iprint >= 1) {
747                     console.log("%nNormal return from subroutine COBYLA%n");
748                 }
749                 if (ifull) {
750                     if (iprint >= 1) {
751                         this.PrintIterationResult(nfvals, f, resmax, x, n, iprint);
752                     }
753                     return status;
754                 }
755                 break;
756             case this.MaxIterationsReached:
757                 if (iprint >= 1) {
758                     console.log(
759                         "%nReturn from subroutine COBYLA because the MAXFUN limit has been reached.%n"
760                     );
761                 }
762                 break;
763             case this.DivergingRoundingErrors:
764                 if (iprint >= 1) {
765                     console.log(
766                         "%nReturn from subroutine COBYLA because rounding errors are becoming damaging.%n"
767                     );
768                 }
769                 break;
770         }
771 
772         for (k = 1; k <= n; ++k) {
773             x[k] = sim[k][np];
774         }
775         f = datmat[mp][np];
776         resmax = datmat[mpp][np];
777         if (iprint >= 1) {
778             this.PrintIterationResult(nfvals, f, resmax, x, n, iprint);
779         }
780 
781         return status;
782     },
783 
784     trstlp: function (n, m, a, b, rho, dx) {
785         //(int n, int m, double[][] a, double[] b, double rho, double[] dx)
786         // N.B. Arguments Z, ZDOTA, VMULTC, SDIRN, DXNEW, VMULTD & IACT have been removed.
787 
788         //     This subroutine calculates an N-component vector DX by applying the
789         //     following two stages. In the first stage, DX is set to the shortest
790         //     vector that minimizes the greatest violation of the constraints
791         //       A(1,K)*DX(1)+A(2,K)*DX(2)+...+A(N,K)*DX(N) .GE. B(K), K = 2,3,...,M,
792         //     subject to the Euclidean length of DX being at most RHO. If its length is
793         //     strictly less than RHO, then we use the resultant freedom in DX to
794         //     minimize the objective function
795         //              -A(1,M+1)*DX(1) - A(2,M+1)*DX(2) - ... - A(N,M+1)*DX(N)
796         //     subject to no increase in any greatest constraint violation. This
797         //     notation allows the gradient of the objective function to be regarded as
798         //     the gradient of a constraint. Therefore the two stages are distinguished
799         //     by MCON .EQ. M and MCON .GT. M respectively. It is possible that a
800         //     degeneracy may prevent DX from attaining the target length RHO. Then the
801         //     value IFULL = 0 would be set, but usually IFULL = 1 on return.
802         //
803         //     In general NACT is the number of constraints in the active set and
804         //     IACT(1),...,IACT(NACT) are their indices, while the remainder of IACT
805         //     contains a permutation of the remaining constraint indices.  Further, Z
806         //     is an orthogonal matrix whose first NACT columns can be regarded as the
807         //     result of Gram-Schmidt applied to the active constraint gradients.  For
808         //     J = 1,2,...,NACT, the number ZDOTA(J) is the scalar product of the J-th
809         //     column of Z with the gradient of the J-th active constraint.  DX is the
810         //     current vector of variables and here the residuals of the active
811         //     constraints should be zero. Further, the active constraints have
812         //     nonnegative Lagrange multipliers that are held at the beginning of
813         //     VMULTC. The remainder of this vector holds the residuals of the inactive
814         //     constraints at DX, the ordering of the components of VMULTC being in
815         //     agreement with the permutation of the indices of the constraints that is
816         //     in IACT. All these residuals are nonnegative, which is achieved by the
817         //     shift RESMAX that makes the least residual zero.
818 
819         //     Initialize Z and some other variables. The value of RESMAX will be
820         //     appropriate to DX = 0, while ICON will be the index of a most violated
821         //     constraint if RESMAX is positive. Usually during the first stage the
822         //     vector SDIRN gives a search direction that reduces all the active
823         //     constraint violations by one simultaneously.
824 
825         // Local variables
826 
827         var temp = 0,
828             nactx = 0,
829             resold = 0.0,
830             z = this.arr2(1 + n, 1 + n),
831             zdota = this.arr(2 + m),
832             vmultc = this.arr(2 + m),
833             sdirn = this.arr(1 + n),
834             dxnew = this.arr(1 + n),
835             vmultd = this.arr(2 + m),
836             iact = this.arr(2 + m),
837             mcon = m,
838             nact = 0,
839             icon, resmax,
840             i, k, first, optold, icount,
841             step, stpful, optnew, ratio,
842             isave, vsave, total,
843             kp, kk, sp, alpha, beta, tot, spabs,
844             acca, accb, zdotv, zdvabs, kw,
845             dd, ss, sd,
846             zdotw, zdwabs, kl, sumabs, tempa,
847             endless = true;
848 
849         for (i = 1; i <= n; ++i) {
850             z[i][i] = 1.0;
851             dx[i] = 0.0;
852         }
853 
854         icon = 0;
855         resmax = 0.0;
856         if (m >= 1) {
857             for (k = 1; k <= m; ++k) {
858                 if (b[k] > resmax) {
859                     resmax = b[k];
860                     icon = k;
861                 }
862             }
863             for (k = 1; k <= m; ++k) {
864                 iact[k] = k;
865                 vmultc[k] = resmax - b[k];
866             }
867         }
868 
869         //     End the current stage of the calculation if 3 consecutive iterations
870         //     have either failed to reduce the best calculated value of the objective
871         //     function or to increase the number of active constraints since the best
872         //     value was calculated. This strategy prevents cycling, but there is a
873         //     remote possibility that it will cause premature termination.
874 
875         first = true;
876         do {
877             L_60: do {
878                 if (!first || (first && resmax === 0.0)) {
879                     mcon = m + 1;
880                     icon = mcon;
881                     iact[mcon] = mcon;
882                     vmultc[mcon] = 0.0;
883                 }
884                 first = false;
885 
886                 optold = 0.0;
887                 icount = 0;
888                 step = 0;
889                 stpful = 0;
890 
891                 L_70: do {
892                     // optnew = (mcon === m) ? resmax : -this.DOT_PRODUCT(this.PART(dx, 1, n), this.PART(this.COL(a, mcon), 1, n));
893                     optnew =
894                         mcon === m ? resmax : -this.DOT_PRODUCT_ROW_COL(dx, -1, a, mcon, 1, n);
895 
896                     if (icount === 0 || optnew < optold) {
897                         optold = optnew;
898                         nactx = nact;
899                         icount = 3;
900                     } else if (nact > nactx) {
901                         nactx = nact;
902                         icount = 3;
903                     } else {
904                         --icount;
905                     }
906                     if (icount === 0) {
907                         break L_60;
908                     }
909 
910                     //     If ICON exceeds NACT, then we add the constraint with index IACT(ICON) to
911                     //     the active set. Apply Givens rotations so that the last N-NACT-1 columns
912                     //     of Z are orthogonal to the gradient of the new constraint, a scalar
913                     //     product being set to zero if its nonzero value could be due to computer
914                     //     rounding errors. The array DXNEW is used for working space.
915                     ratio = 0;
916                     if (icon <= nact) {
917                         if (icon < nact) {
918                             //     Delete the constraint that has the index IACT(ICON) from the active set.
919 
920                             isave = iact[icon];
921                             vsave = vmultc[icon];
922                             k = icon;
923                             do {
924                                 kp = k + 1;
925                                 kk = iact[kp];
926                                 sp = this.DOT_PRODUCT(
927                                     this.PART(this.COL(z, k), 1, n),
928                                     this.PART(this.COL(a, kk), 1, n)
929                                 );
930                                 temp = Mat.hypot(sp, zdota[kp]);
931                                 alpha = zdota[kp] / temp;
932                                 beta = sp / temp;
933                                 zdota[kp] = alpha * zdota[k];
934                                 zdota[k] = temp;
935                                 for (i = 1; i <= n; ++i) {
936                                     temp = alpha * z[i][kp] + beta * z[i][k];
937                                     z[i][kp] = alpha * z[i][k] - beta * z[i][kp];
938                                     z[i][k] = temp;
939                                 }
940                                 iact[k] = kk;
941                                 vmultc[k] = vmultc[kp];
942                                 k = kp;
943                             } while (k < nact);
944 
945                             iact[k] = isave;
946                             vmultc[k] = vsave;
947                         }
948                         --nact;
949 
950                         //     If stage one is in progress, then set SDIRN to the direction of the next
951                         //     change to the current vector of variables.
952                         if (mcon > m) {
953                             //     Pick the next search direction of stage two.
954                             temp = 1.0 / zdota[nact];
955                             for (k = 1; k <= n; ++k) {
956                                 sdirn[k] = temp * z[k][nact];
957                             }
958                         } else {
959                             // temp = this.DOT_PRODUCT(this.PART(sdirn, 1, n), this.PART(this.COL(z, nact + 1), 1, n));
960                             temp = this.DOT_PRODUCT_ROW_COL(sdirn, -1, z, nact + 1, 1, n);
961                             for (k = 1; k <= n; ++k) {
962                                 sdirn[k] -= temp * z[k][nact + 1];
963                             }
964                         }
965                     } else {
966                         kk = iact[icon];
967                         for (k = 1; k <= n; ++k) {
968                             dxnew[k] = a[k][kk];
969                         }
970                         tot = 0.0;
971 
972                         // {
973                         k = n;
974                         while (k > nact) {
975                             sp = 0.0;
976                             spabs = 0.0;
977                             for (i = 1; i <= n; ++i) {
978                                 temp = z[i][k] * dxnew[i];
979                                 sp += temp;
980                                 spabs += Math.abs(temp);
981                             }
982                             acca = spabs + 0.1 * Math.abs(sp);
983                             accb = spabs + 0.2 * Math.abs(sp);
984                             if (spabs >= acca || acca >= accb) {
985                                 sp = 0.0;
986                             }
987                             if (tot === 0.0) {
988                                 tot = sp;
989                             } else {
990                                 kp = k + 1;
991                                 temp = Mat.hypot(sp, tot);
992                                 alpha = sp / temp;
993                                 beta = tot / temp;
994                                 tot = temp;
995                                 for (i = 1; i <= n; ++i) {
996                                     temp = alpha * z[i][k] + beta * z[i][kp];
997                                     z[i][kp] = alpha * z[i][kp] - beta * z[i][k];
998                                     z[i][k] = temp;
999                                 }
1000                             }
1001                             --k;
1002                         }
1003                         // }
1004 
1005                         if (tot === 0.0) {
1006                             //     The next instruction is reached if a deletion has to be made from the
1007                             //     active set in order to make room for the new active constraint, because
1008                             //     the new constraint gradient is a linear combination of the gradients of
1009                             //     the old active constraints.  Set the elements of VMULTD to the multipliers
1010                             //     of the linear combination.  Further, set IOUT to the index of the
1011                             //     constraint to be deleted, but branch if no suitable index can be found.
1012 
1013                             ratio = -1.0;
1014                             //{
1015                             k = nact;
1016                             do {
1017                                 zdotv = 0.0;
1018                                 zdvabs = 0.0;
1019 
1020                                 for (i = 1; i <= n; ++i) {
1021                                     temp = z[i][k] * dxnew[i];
1022                                     zdotv += temp;
1023                                     zdvabs += Math.abs(temp);
1024                                 }
1025                                 acca = zdvabs + 0.1 * Math.abs(zdotv);
1026                                 accb = zdvabs + 0.2 * Math.abs(zdotv);
1027                                 if (zdvabs < acca && acca < accb) {
1028                                     temp = zdotv / zdota[k];
1029                                     if (temp > 0.0 && iact[k] <= m) {
1030                                         tempa = vmultc[k] / temp;
1031                                         if (ratio < 0.0 || tempa < ratio) {
1032                                             ratio = tempa;
1033                                         }
1034                                     }
1035 
1036                                     if (k >= 2) {
1037                                         kw = iact[k];
1038                                         for (i = 1; i <= n; ++i) {
1039                                             dxnew[i] -= temp * a[i][kw];
1040                                         }
1041                                     }
1042                                     vmultd[k] = temp;
1043                                 } else {
1044                                     vmultd[k] = 0.0;
1045                                 }
1046                             } while (--k > 0);
1047                             //}
1048                             if (ratio < 0.0) {
1049                                 break L_60;
1050                             }
1051 
1052                             //     Revise the Lagrange multipliers and reorder the active constraints so
1053                             //     that the one to be replaced is at the end of the list. Also calculate the
1054                             //     new value of ZDOTA(NACT) and branch if it is not acceptable.
1055 
1056                             for (k = 1; k <= nact; ++k) {
1057                                 vmultc[k] = Math.max(0.0, vmultc[k] - ratio * vmultd[k]);
1058                             }
1059                             if (icon < nact) {
1060                                 isave = iact[icon];
1061                                 vsave = vmultc[icon];
1062                                 k = icon;
1063                                 do {
1064                                     kp = k + 1;
1065                                     kw = iact[kp];
1066                                     sp = this.DOT_PRODUCT(
1067                                         this.PART(this.COL(z, k), 1, n),
1068                                         this.PART(this.COL(a, kw), 1, n)
1069                                     );
1070                                     temp = Mat.hypot(sp, zdota[kp]);
1071                                     alpha = zdota[kp] / temp;
1072                                     beta = sp / temp;
1073                                     zdota[kp] = alpha * zdota[k];
1074                                     zdota[k] = temp;
1075                                     for (i = 1; i <= n; ++i) {
1076                                         temp = alpha * z[i][kp] + beta * z[i][k];
1077                                         z[i][kp] = alpha * z[i][k] - beta * z[i][kp];
1078                                         z[i][k] = temp;
1079                                     }
1080                                     iact[k] = kw;
1081                                     vmultc[k] = vmultc[kp];
1082                                     k = kp;
1083                                 } while (k < nact);
1084                                 iact[k] = isave;
1085                                 vmultc[k] = vsave;
1086                             }
1087                             temp = this.DOT_PRODUCT(
1088                                 this.PART(this.COL(z, nact), 1, n),
1089                                 this.PART(this.COL(a, kk), 1, n)
1090                             );
1091                             if (temp === 0.0) {
1092                                 break L_60;
1093                             }
1094                             zdota[nact] = temp;
1095                             vmultc[icon] = 0.0;
1096                             vmultc[nact] = ratio;
1097                         } else {
1098                             //     Add the new constraint if this can be done without a deletion from the
1099                             //     active set.
1100 
1101                             ++nact;
1102                             zdota[nact] = tot;
1103                             vmultc[icon] = vmultc[nact];
1104                             vmultc[nact] = 0.0;
1105                         }
1106 
1107                         //     Update IACT and ensure that the objective function continues to be
1108                         //     treated as the last active constraint when MCON>M.
1109 
1110                         iact[icon] = iact[nact];
1111                         iact[nact] = kk;
1112                         if (mcon > m && kk !== mcon) {
1113                             k = nact - 1;
1114                             sp = this.DOT_PRODUCT(
1115                                 this.PART(this.COL(z, k), 1, n),
1116                                 this.PART(this.COL(a, kk), 1, n)
1117                             );
1118                             temp = Mat.hypot(sp, zdota[nact]);
1119                             alpha = zdota[nact] / temp;
1120                             beta = sp / temp;
1121                             zdota[nact] = alpha * zdota[k];
1122                             zdota[k] = temp;
1123                             for (i = 1; i <= n; ++i) {
1124                                 temp = alpha * z[i][nact] + beta * z[i][k];
1125                                 z[i][nact] = alpha * z[i][k] - beta * z[i][nact];
1126                                 z[i][k] = temp;
1127                             }
1128                             iact[nact] = iact[k];
1129                             iact[k] = kk;
1130                             temp = vmultc[k];
1131                             vmultc[k] = vmultc[nact];
1132                             vmultc[nact] = temp;
1133                         }
1134 
1135                         //     If stage one is in progress, then set SDIRN to the direction of the next
1136                         //     change to the current vector of variables.
1137                         if (mcon > m) {
1138                             //     Pick the next search direction of stage two.
1139                             temp = 1.0 / zdota[nact];
1140                             for (k = 1; k <= n; ++k) {
1141                                 sdirn[k] = temp * z[k][nact];
1142                             }
1143                         } else {
1144                             kk = iact[nact];
1145                             // temp = (this.DOT_PRODUCT(this.PART(sdirn, 1, n),this.PART(this.COL(a, kk), 1, n)) - 1.0) / zdota[nact];
1146                             temp =
1147                                 (this.DOT_PRODUCT_ROW_COL(sdirn, -1, a, kk, 1, n) - 1.0) /
1148                                 zdota[nact];
1149                             for (k = 1; k <= n; ++k) {
1150                                 sdirn[k] -= temp * z[k][nact];
1151                             }
1152                         }
1153                     }
1154 
1155                     //     Calculate the step to the boundary of the trust region or take the step
1156                     //     that reduces RESMAX to zero. The two statements below that include the
1157                     //     factor 1.0E-6 prevent some harmless underflows that occurred in a test
1158                     //     calculation. Further, we skip the step if it could be zero within a
1159                     //     reasonable tolerance for computer rounding errors.
1160                     dd = rho * rho;
1161                     sd = 0.0;
1162                     ss = 0.0;
1163                     for (i = 1; i <= n; ++i) {
1164                         if (Math.abs(dx[i]) >= 1.0e-6 * rho) {
1165                             dd -= dx[i] * dx[i];
1166                         }
1167                         sd += dx[i] * sdirn[i];
1168                         ss += sdirn[i] * sdirn[i];
1169                     }
1170                     if (dd <= 0.0) {
1171                         break L_60;
1172                     }
1173                     temp = Math.sqrt(ss * dd);
1174                     if (Math.abs(sd) >= 1.0e-6 * temp) {
1175                         temp = Math.sqrt(ss * dd + sd * sd);
1176                     }
1177                     stpful = dd / (temp + sd);
1178                     step = stpful;
1179                     if (mcon === m) {
1180                         acca = step + 0.1 * resmax;
1181                         accb = step + 0.2 * resmax;
1182                         if (step >= acca || acca >= accb) {
1183                             break L_70;
1184                         }
1185                         step = Math.min(step, resmax);
1186                     }
1187 
1188                     //     Set DXNEW to the new variables if STEP is the steplength, and reduce
1189                     //     RESMAX to the corresponding maximum residual if stage one is being done.
1190                     //     Because DXNEW will be changed during the calculation of some Lagrange
1191                     //     multipliers, it will be restored to the following value later.
1192                     for (k = 1; k <= n; ++k) {
1193                         dxnew[k] = dx[k] + step * sdirn[k];
1194                     }
1195                     if (mcon === m) {
1196                         resold = resmax;
1197                         resmax = 0.0;
1198                         for (k = 1; k <= nact; ++k) {
1199                             kk = iact[k];
1200                             // temp = b[kk] - this.DOT_PRODUCT(this.PART(this.COL(a, kk), 1, n), this.PART(dxnew, 1, n));
1201                             temp = b[kk] - this.DOT_PRODUCT_ROW_COL(dxnew, -1, a, kk, 1, n);
1202                             resmax = Math.max(resmax, temp);
1203                         }
1204                     }
1205 
1206                     //     Set VMULTD to the VMULTC vector that would occur if DX became DXNEW. A
1207                     //     device is included to force VMULTD(K) = 0.0 if deviations from this value
1208                     //     can be attributed to computer rounding errors. First calculate the new
1209                     //     Lagrange multipliers.
1210                     //{
1211                     k = nact;
1212                     do {
1213                         zdotw = 0.0;
1214                         zdwabs = 0.0;
1215                         for (i = 1; i <= n; ++i) {
1216                             temp = z[i][k] * dxnew[i];
1217                             zdotw += temp;
1218                             zdwabs += Math.abs(temp);
1219                         }
1220                         acca = zdwabs + 0.1 * Math.abs(zdotw);
1221                         accb = zdwabs + 0.2 * Math.abs(zdotw);
1222                         if (zdwabs >= acca || acca >= accb) {
1223                             zdotw = 0.0;
1224                         }
1225                         vmultd[k] = zdotw / zdota[k];
1226                         if (k >= 2) {
1227                             kk = iact[k];
1228                             for (i = 1; i <= n; ++i) {
1229                                 dxnew[i] -= vmultd[k] * a[i][kk];
1230                             }
1231                         }
1232                     } while (k-- >= 2);
1233                     if (mcon > m) {
1234                         vmultd[nact] = Math.max(0.0, vmultd[nact]);
1235                     }
1236                     //}
1237 
1238                     //     Complete VMULTC by finding the new constraint residuals.
1239 
1240                     for (k = 1; k <= n; ++k) {
1241                         dxnew[k] = dx[k] + step * sdirn[k];
1242                     }
1243                     if (mcon > nact) {
1244                         kl = nact + 1;
1245                         for (k = kl; k <= mcon; ++k) {
1246                             kk = iact[k];
1247                             total = resmax - b[kk];
1248                             sumabs = resmax + Math.abs(b[kk]);
1249                             for (i = 1; i <= n; ++i) {
1250                                 temp = a[i][kk] * dxnew[i];
1251                                 total += temp;
1252                                 sumabs += Math.abs(temp);
1253                             }
1254                             acca = sumabs + 0.1 * Math.abs(total);
1255                             accb = sumabs + 0.2 * Math.abs(total);
1256                             if (sumabs >= acca || acca >= accb) {
1257                                 total = 0.0;
1258                             }
1259                             vmultd[k] = total;
1260                         }
1261                     }
1262 
1263                     //     Calculate the fraction of the step from DX to DXNEW that will be taken.
1264 
1265                     ratio = 1.0;
1266                     icon = 0;
1267                     for (k = 1; k <= mcon; ++k) {
1268                         if (vmultd[k] < 0.0) {
1269                             temp = vmultc[k] / (vmultc[k] - vmultd[k]);
1270                             if (temp < ratio) {
1271                                 ratio = temp;
1272                                 icon = k;
1273                             }
1274                         }
1275                     }
1276 
1277                     //     Update DX, VMULTC and RESMAX.
1278 
1279                     temp = 1.0 - ratio;
1280                     for (k = 1; k <= n; ++k) {
1281                         dx[k] = temp * dx[k] + ratio * dxnew[k];
1282                     }
1283                     for (k = 1; k <= mcon; ++k) {
1284                         vmultc[k] = Math.max(0.0, temp * vmultc[k] + ratio * vmultd[k]);
1285                     }
1286                     if (mcon === m) {
1287                         resmax = resold + ratio * (resmax - resold);
1288                     }
1289 
1290                     //     If the full step is not acceptable then begin another iteration.
1291                     //     Otherwise switch to stage two or end the calculation.
1292                 } while (icon > 0);
1293 
1294                 if (step === stpful) {
1295                     return true;
1296                 }
1297             } while (endless);
1298 
1299             //     We employ any freedom that may be available to reduce the objective
1300             //     function before returning a DX whose length is less than RHO.
1301         } while (mcon === m);
1302 
1303         return false;
1304     },
1305 
1306     PrintIterationResult: function (nfvals, f, resmax, x, n, iprint) {
1307         if (iprint > 1) {
1308             console.log("NFVALS = " + nfvals + "  F = " + f + "  MAXCV = " + resmax);
1309         }
1310         if (iprint > 1) {
1311             console.log("X = " + this.PART(x, 1, n));
1312         }
1313     },
1314 
1315     ROW: function (src, rowidx) {
1316         return src[rowidx].slice();
1317         // var col,
1318         //     cols = src[0].length,
1319         //     dest = this.arr(cols);
1320 
1321         // for (col = 0; col < cols; ++col) {
1322         //     dest[col] = src[rowidx][col];
1323         // }
1324         // return dest;
1325     },
1326 
1327     COL: function (src, colidx) {
1328         var row,
1329             rows = src.length,
1330             dest = []; // this.arr(rows);
1331 
1332         for (row = 0; row < rows; ++row) {
1333             dest[row] = src[row][colidx];
1334         }
1335         return dest;
1336     },
1337 
1338     PART: function (src, from, to) {
1339         return src.slice(from, to + 1);
1340         // var srcidx,
1341         //     dest = this.arr(to - from + 1),
1342         //     destidx = 0;
1343         // for (srcidx = from; srcidx <= to; ++srcidx, ++destidx) {
1344         //     dest[destidx] = src[srcidx];
1345         // }
1346         // return dest;
1347     },
1348 
1349     FORMAT: function (x) {
1350         return x.join(",");
1351         // var i, fmt = "";
1352         // for (i = 0; i < x.length; ++i) {
1353         //     fmt += ", " + x[i];
1354         // }
1355         // return fmt;
1356     },
1357 
1358     DOT_PRODUCT: function (lhs, rhs) {
1359         var i,
1360             sum = 0.0,
1361             len = lhs.length;
1362         for (i = 0; i < len; ++i) {
1363             sum += lhs[i] * rhs[i];
1364         }
1365         return sum;
1366     },
1367 
1368     DOT_PRODUCT_ROW_COL: function (lhs, row, rhs, col, start, end) {
1369         var i,
1370             sum = 0.0;
1371 
1372         if (row === -1) {
1373             // lhs is vector
1374             for (i = start; i <= end; ++i) {
1375                 sum += lhs[i] * rhs[i][col];
1376             }
1377         } else {
1378             // lhs is row of matrix
1379             if (col === -1) {
1380                 // rhs is vector
1381                 for (i = start; i <= end; ++i) {
1382                     sum += lhs[row][i] * rhs[i];
1383                 }
1384             } else {
1385                 // rhs is column of matrix
1386                 for (i = start; i <= end; ++i) {
1387                     sum += lhs[row][i] * rhs[i][col];
1388                 }
1389             }
1390         }
1391 
1392         return sum;
1393     }
1394 };
1395 
1396 export default JXG.Math.Nlp;
1397