xref: /petsc/src/tao/leastsquares/impls/pounders/gqt.c (revision b0a7d7e7f246badab30cb8ce3f95dd1540bfb513)
1 #include "petsc.h"
2 #include "petscblaslapack.h"
3 #include "taolapack.h"
4 
5 
6 
7 #undef __FUNCT__
8 #define __FUNCT__ "estsv"
9 static PetscErrorCode estsv(PetscInt n, PetscReal *r, PetscInt ldr, PetscReal *svmin, PetscReal *z) {
10 
11     PetscBLASInt blas1=1, blasn=n, blasnmi, blasj, blasldr = ldr;
12     PetscInt i,j;
13     PetscReal e,temp,w,wm,ynorm,znorm,s,sm;
14     PetscFunctionBegin;
15     for (i=0;i<n;i++) {
16 	z[i]=0.0;
17     }
18 
19     e = PetscAbs(r[0]);
20     if (e == 0.0) {
21 	*svmin = 0.0;
22 	z[0] = 1.0;
23     } else {
24       /* Solve R'*y = e */
25 	for (i=0;i<n;i++) {
26 
27 	  /* Scale y. The scaling factor (0.01) reduces the number of scalings */
28 	    if (z[i] >= 0.0)
29 		e=-PetscAbs(e);
30 	    else
31 		e = PetscAbs(e);
32 
33 	    if (PetscAbs(e - z[i]) > PetscAbs(r[i + ldr*i])) {
34 		temp = PetscMin(0.01,PetscAbs(r[i + ldr*i]))/
35 		    PetscAbs(e-z[i]);
36 		BLASscal_(&blasn, &temp, z, &blas1);
37 		e = temp*e;
38 	    }
39 
40 	    /* Determine the two possible choices of y[i] */
41 	    if (r[i + ldr*i] == 0.0)
42 		w = wm = 1.0;
43 	    else {
44 		w = (e - z[i]) / r[i + ldr*i];
45 		wm = - (e + z[i]) / r[i + ldr*i];
46 	    }
47 
48 	    /*  Chose y[i] based on the predicted value of y[j] for j>i */
49 	    s = PetscAbs(e - z[i]);
50 	    sm = PetscAbs(e + z[i]);
51 	    for (j=i+1;j<n;j++) {
52 		sm += PetscAbs(z[j] + wm * r[i + ldr*j]);
53 	    }
54 	    if (i < n-1) {
55 		blasnmi = n-i-1;
56 		BLASaxpy_(&blasnmi, &w, &r[i + ldr*(i+1)], &blasldr, &z[i+1], &blas1);
57 		s += BLASasum_(&blasnmi, &z[i+1], &blas1);
58 	    }
59 	    if (s < sm) {
60 		temp = wm - w;
61 		w = wm;
62 		if (i < n-1) {
63 		    BLASaxpy_(&blasnmi, &temp, &r[i + ldr*(i+1)], &blasldr, &z[i+1], &blas1);
64 		}
65 	    }
66 	    z[i] = w;
67 	}
68 
69 	ynorm = BLASnrm2_(&blasn, z, &blas1);
70 
71 	/* Solve R*z = y */
72 	for (j=n-1; j>=0; j--) {
73 
74 	  /* Scale z */
75 
76 	    if (PetscAbs(z[j]) > PetscAbs(r[j + ldr*j])) {
77 		temp = PetscMin(0.01, PetscAbs(r[j + ldr*j] / z[j]));
78 		BLASscal_(&blasn, &temp, z, &blas1);
79 		ynorm *=temp;
80 	    }
81 	    if (r[j + ldr*j] == 0) {
82 		z[j] = 1.0;
83 	    } else {
84 		z[j] = z[j] / r[j + ldr*j];
85 	    }
86 	    temp = -z[j];
87 	    blasj=j;
88 	    BLASaxpy_(&blasj,&temp,&r[0+ldr*j],&blas1,z,&blas1);
89 	}
90 
91 	/* Compute svmin and normalize z */
92 	znorm = 1.0 / BLASnrm2_(&blasn, z, &blas1);
93 	*svmin = ynorm*znorm;
94 	BLASscal_(&blasn, &znorm, z, &blas1);
95 
96     }
97 
98     PetscFunctionReturn(0);
99 }
100 
101 
102 
103 /*
104 c     ***********
105 c
106 c     Subroutine dgqt
107 c
108 c     Given an n by n symmetric matrix A, an n-vector b, and a
109 c     positive number delta, this subroutine determines a vector
110 c     x which approximately minimizes the quadratic function
111 c
112 c           f(x) = (1/2)*x'*A*x + b'*x
113 c
114 c     subject to the Euclidean norm constraint
115 c
116 c           norm(x) <= delta.
117 c
118 c     This subroutine computes an approximation x and a Lagrange
119 c     multiplier par such that either par is zero and
120 c
121 c            norm(x) <= (1+rtol)*delta,
122 c
123 c     or par is positive and
124 c
125 c            abs(norm(x) - delta) <= rtol*delta.
126 c
127 c     If xsol is the solution to the problem, the approximation x
128 c     satisfies
129 c
130 c            f(x) <= ((1 - rtol)**2)*f(xsol)
131 c
132 c     The subroutine statement is
133 c
134 c       subroutine dgqt(n,a,lda,b,delta,rtol,atol,itmax,
135 c                        par,f,x,info,z,wa1,wa2)
136 c
137 c     where
138 c
139 c       n is an integer variable.
140 c         On entry n is the order of A.
141 c         On exit n is unchanged.
142 c
143 c       a is a double precision array of dimension (lda,n).
144 c         On entry the full upper triangle of a must contain the
145 c            full upper triangle of the symmetric matrix A.
146 c         On exit the array contains the matrix A.
147 c
148 c       lda is an integer variable.
149 c         On entry lda is the leading dimension of the array a.
150 c         On exit lda is unchanged.
151 c
152 c       b is an double precision array of dimension n.
153 c         On entry b specifies the linear term in the quadratic.
154 c         On exit b is unchanged.
155 c
156 c       delta is a double precision variable.
157 c         On entry delta is a bound on the Euclidean norm of x.
158 c         On exit delta is unchanged.
159 c
160 c       rtol is a double precision variable.
161 c         On entry rtol is the relative accuracy desired in the
162 c            solution. Convergence occurs if
163 c
164 c              f(x) <= ((1 - rtol)**2)*f(xsol)
165 c
166 c         On exit rtol is unchanged.
167 c
168 c       atol is a double precision variable.
169 c         On entry atol is the absolute accuracy desired in the
170 c            solution. Convergence occurs when
171 c
172 c              norm(x) <= (1 + rtol)*delta
173 c
174 c              max(-f(x),-f(xsol)) <= atol
175 c
176 c         On exit atol is unchanged.
177 c
178 c       itmax is an integer variable.
179 c         On entry itmax specifies the maximum number of iterations.
180 c         On exit itmax is unchanged.
181 c
182 c       par is a double precision variable.
183 c         On entry par is an initial estimate of the Lagrange
184 c            multiplier for the constraint norm(x) <= delta.
185 c         On exit par contains the final estimate of the multiplier.
186 c
187 c       f is a double precision variable.
188 c         On entry f need not be specified.
189 c         On exit f is set to f(x) at the output x.
190 c
191 c       x is a double precision array of dimension n.
192 c         On entry x need not be specified.
193 c         On exit x is set to the final estimate of the solution.
194 c
195 c       info is an integer variable.
196 c         On entry info need not be specified.
197 c         On exit info is set as follows:
198 c
199 c            info = 1  The function value f(x) has the relative
200 c                      accuracy specified by rtol.
201 c
202 c            info = 2  The function value f(x) has the absolute
203 c                      accuracy specified by atol.
204 c
205 c            info = 3  Rounding errors prevent further progress.
206 c                      On exit x is the best available approximation.
207 c
208 c            info = 4  Failure to converge after itmax iterations.
209 c                      On exit x is the best available approximation.
210 c
211 c       z is a double precision work array of dimension n.
212 c
213 c       wa1 is a double precision work array of dimension n.
214 c
215 c       wa2 is a double precision work array of dimension n.
216 c
217 c     Subprograms called
218 c
219 c       MINPACK-2  ......  destsv
220 c
221 c       LAPACK  .........  dpotrf
222 c
223 c       Level 1 BLAS  ...  daxpy, dcopy, ddot, dnrm2, dscal
224 c
225 c       Level 2 BLAS  ...  dtrmv, dtrsv
226 c
227 c     MINPACK-2 Project. October 1993.
228 c     Argonne National Laboratory and University of Minnesota.
229 c     Brett M. Averick, Richard Carter, and Jorge J. More'
230 c
231 c     ***********
232 */
233 
234 #undef __FUNCT__
235 #define __FUNCT__ "gqt"
236 PetscErrorCode gqt(PetscInt n, PetscReal *a, PetscInt lda, PetscReal *b,
237 		   PetscReal delta, PetscReal rtol, PetscReal atol,
238 		   PetscInt itmax, PetscReal *retpar, PetscReal *retf,
239 		   PetscReal *x, PetscInt *retinfo, PetscInt *retits,
240 		   PetscReal *z, PetscReal *wa1, PetscReal *wa2)
241 {
242     PetscErrorCode ierr;
243     PetscReal f=0.0,p001=0.001,p5=0.5,minusone=-1,delta2=delta*delta;
244     PetscInt iter, j, rednc,info;
245     PetscBLASInt indef;
246     PetscBLASInt blas1=1, blasn=n, iblas, blaslda = lda,blasldap1=lda+1,blasinfo;
247     PetscReal alpha, anorm, bnorm, parc, parf, parl, pars, par=*retpar,
248 	paru, prod, rxnorm, rznorm=0.0, temp, xnorm;
249 
250     PetscFunctionBegin;
251     iter = 0;
252     parf = 0.0;
253     xnorm = 0.0;
254     rxnorm = 0.0;
255     rednc = 0;
256     for (j=0; j<n; j++) {
257 	x[j] = 0.0;
258 	z[j] = 0.0;
259     }
260 
261     /* Copy the diagonal and save A in its lower triangle */
262     BLAScopy_(&blasn,a,&blasldap1, wa1, &blas1);
263     CHKMEMQ;
264     for (j=0;j<n-1;j++) {
265 	iblas = n - j - 1;
266 	BLAScopy_(&iblas,&a[j + lda*(j+1)], &blaslda, &a[j+1 + lda*j], &blas1);
267 	CHKMEMQ;
268     }
269 
270     /* Calculate the l1-norm of A, the Gershgorin row sums, and the
271        l2-norm of b */
272     anorm = 0.0;
273     for (j=0;j<n;j++) {
274 	wa2[j] = BLASasum_(&blasn, &a[0 + lda*j], &blas1);
275 	CHKMEMQ;
276 	anorm = PetscMax(anorm,wa2[j]);
277     }
278     for (j=0;j<n;j++) {
279 	wa2[j] = wa2[j] - PetscAbs(wa1[j]);
280     }
281     bnorm = BLASnrm2_(&blasn,b,&blas1);
282     CHKMEMQ;
283     /* Calculate a lower bound, pars, for the domain of the problem.
284        Also calculate an upper bound, paru, and a lower bound, parl,
285        for the Lagrange multiplier. */
286     pars = parl = paru = -anorm;
287     for (j=0;j<n;j++) {
288 	pars = PetscMax(pars, -wa1[j]);
289 	parl = PetscMax(parl, wa1[j] + wa2[j]);
290 	paru = PetscMax(paru, -wa1[j] + wa2[j]);
291     }
292     parl = PetscMax(bnorm/delta - parl,pars);
293     parl = PetscMax(0.0,parl);
294     paru = PetscMax(0.0, bnorm/delta + paru);
295 
296     /* If the input par lies outside of the interval (parl, paru),
297        set par to the closer endpoint. */
298 
299     par = PetscMax(par,parl);
300     par = PetscMin(par,paru);
301 
302     /* Special case: parl == paru */
303     paru = PetscMax(paru, (1.0 + rtol)*parl);
304 
305     /* Beginning of an iteration */
306 
307     info = 0;
308     for (iter=1;iter<=itmax;iter++) {
309 
310       /* Safeguard par */
311 	if (par <= pars && paru > 0) {
312 	    par = PetscMax(p001, PetscSqrtScalar(parl/paru)) * paru;
313 	}
314 
315 	/* Copy the lower triangle of A into its upper triangle and
316 	   compute A + par*I */
317 
318 	for (j=0;j<n-1;j++) {
319 	    iblas = n - j - 1;
320 	    BLAScopy_(&iblas,&a[j+1 + j*lda], &blas1,
321 		      &a[j + (j+1)*lda], &blaslda);
322 	    CHKMEMQ;
323 	}
324 	for (j=0;j<n;j++) {
325 	    a[j + j*lda] = wa1[j] + par;
326 	}
327 
328 	/* Attempt the Cholesky factorization of A without referencing
329 	   the lower triangular part. */
330 	LAPACKpotrf_("U",&blasn,a,&blaslda,&indef);
331 	CHKMEMQ;
332 
333 	/* Case 1: A + par*I is pos. def. */
334 	if (indef == 0) {
335 
336   	  /* Compute an approximate solution x and save the
337 	     last value of par with A + par*I pos. def. */
338 
339 	    parf = par;
340 	    BLAScopy_(&blasn, b, &blas1, wa2, &blas1);
341 	    CHKMEMQ;
342 	    LAPACKtrtrs_("U","T","N",&blasn,&blas1,a,&blaslda,wa2,&blasn,&blasinfo);
343 	    CHKMEMQ;
344 	    rxnorm = BLASnrm2_(&blasn, wa2, &blas1);
345 	    LAPACKtrtrs_("U","N","N",&blasn,&blas1,a,&blaslda,wa2,&blasn,&blasinfo);
346 	    CHKMEMQ;
347 	    BLAScopy_(&blasn, wa2, &blas1, x, &blas1);
348 	    CHKMEMQ;
349 	    BLASscal_(&blasn, &minusone, x, &blas1);
350 	    CHKMEMQ;
351 	    xnorm = BLASnrm2_(&blasn, x, &blas1);
352 	    CHKMEMQ;
353 
354 	    /* Test for convergence */
355 	    if (PetscAbs(xnorm - delta) <= rtol*delta ||
356 		(par == 0  && xnorm <= (1.0+rtol)*delta)) {
357 		info = 1;
358 	    }
359 
360 	    /* Compute a direction of negative curvature and use this
361 	       information to improve pars. */
362 
363 	    iblas=blasn*blasn;
364 
365 	    ierr = estsv(n,a,lda,&rznorm,z); CHKERRQ(ierr);
366 	    CHKMEMQ;
367 	    pars = PetscMax(pars, par-rznorm*rznorm);
368 
369 	    /* Compute a negative curvature solution of the form
370 	       x + alpha*z,  where norm(x+alpha*z)==delta */
371 
372 	    rednc = 0;
373 	    if (xnorm < delta) {
374 
375 	        /* Compute alpha */
376 		prod = BLASdot_(&blasn, z, &blas1, x, &blas1) / delta;
377 		temp = (delta - xnorm)*((delta + xnorm)/delta);
378 		alpha = temp/(PetscAbs(prod) + PetscSqrtScalar(prod*prod + temp/delta));
379 		if (prod >= 0)
380 		    alpha = PetscAbs(alpha);
381 		else
382 		    alpha =-PetscAbs(alpha);
383 
384 
385 		/* Test to decide if the negative curvature step
386 		   produces a larger reduction than with z=0 */
387 
388 		rznorm = PetscAbs(alpha) * rznorm;
389 		if ((rznorm*rznorm + par*xnorm*xnorm)/(delta2) <= par) {
390 		    rednc = 1;
391 		}
392 
393 		/* Test for convergence */
394 		if (p5 * rznorm*rznorm / delta2 <=
395 		    rtol*(1.0-p5*rtol)*(par + rxnorm*rxnorm/delta2)) {
396 		    info = 1;
397 		} else if (info == 0 &&
398 			   (p5*(par + rxnorm*rxnorm/delta2) <= atol/delta2)) {
399 		    info = 2;
400 		}
401 	    }
402 
403 	    /* Compute the Newton correction parc to par. */
404 	    if (xnorm == 0) {
405 		parc = -par;
406 	    } else {
407 		BLAScopy_(&blasn, x, &blas1, wa2, &blas1);
408 		CHKMEMQ;
409 		temp = 1.0/xnorm;
410 		BLASscal_(&blasn, &temp, wa2, &blas1);
411 		CHKMEMQ;
412 		LAPACKtrtrs_("U","T","N",&blasn, &blas1, a, &blaslda, wa2, &blasn, &blasinfo);
413 		CHKMEMQ;
414 		temp = BLASnrm2_(&blasn, wa2, &blas1);
415 		parc = (xnorm - delta)/(delta*temp*temp);
416 	    }
417 
418 	    /* update parl or paru */
419 	    if (xnorm > delta) {
420 		parl = PetscMax(parl, par);
421 	    } else if (xnorm < delta) {
422 		paru = PetscMin(paru, par);
423 	    }
424 	} else {
425 	    /* Case 2: A + par*I is not pos. def. */
426 
427 	    /* Use the rank information from the Cholesky
428 	       decomposition to update par. */
429 
430 	    if (indef > 1) {
431 
432 	        /* Restore column indef to A + par*I. */
433 		iblas = indef - 1;
434 		BLAScopy_(&iblas,&a[indef-1 + 0*lda],&blaslda,
435 			  &a[0 + (indef-1)*lda],&blas1);
436 		CHKMEMQ;
437 		a[indef-1 + (indef-1)*lda] = wa1[indef-1] + par;
438 
439 		/* compute parc. */
440 
441 		BLAScopy_(&iblas,&a[0 + (indef-1)*lda], &blas1, wa2, &blas1);
442 		CHKMEMQ;
443 		LAPACKtrtrs_("U","T","N",&iblas,&blas1,a,&blaslda,wa2,&blasn,&blasinfo);
444 		CHKMEMQ;
445 		BLAScopy_(&iblas,wa2,&blas1,&a[0 + (indef-1)*lda],&blas1);
446 		CHKMEMQ;
447 		temp = BLASnrm2_(&iblas,&a[0 + (indef-1)*lda],&blas1);
448 		CHKMEMQ;
449 		a[indef-1 + (indef-1)*lda] -= temp*temp;
450 		LAPACKtrtrs_("U","N","N",&iblas,&blas1,a,&blaslda,wa2,&blasn,&blasinfo);
451 		CHKMEMQ;
452 	    }
453 
454 	    wa2[indef-1] = -1.0;
455 	    iblas = indef;
456 	    temp = BLASnrm2_(&iblas,wa2,&blas1);
457 	    parc = - a[indef-1 + (indef-1)*lda]/(temp*temp);
458 	    pars = PetscMax(pars,par+parc);
459 
460 	    /* If necessary, increase paru slightly.
461 	       This is needed because in some exceptional situations
462 	       paru is the optimal value of par. */
463 
464 	    paru = PetscMax(paru, (1.0+rtol)*pars);
465 	}
466 
467 	/* Use pars to update parl */
468 	parl = PetscMax(parl,pars);
469 
470 	/* Test for termination. */
471 	if (info == 0) {
472 	    if (iter == itmax) info=4;
473 	    if (paru <= (1.0+p5*rtol)*pars) info=3;
474 	    if (paru == 0.0) info = 2;
475 	}
476 
477 	/* If exiting, store the best approximation and restore
478 	   the upper triangle of A. */
479 
480 	if (info != 0) {
481 
482 	  /* Compute the best current estimates for x and f. */
483 
484 	    par = parf;
485 	    f = -p5 * (rxnorm*rxnorm + par*xnorm*xnorm);
486 	    if (rednc) {
487 		f = -p5 * (rxnorm*rxnorm + par*delta*delta - rznorm*rznorm);
488 		BLASaxpy_(&blasn, &alpha, z, &blas1, x, &blas1);
489 		CHKMEMQ;
490 	    }
491 
492 	    /* Restore the upper triangle of A */
493 
494 	    for (j = 0; j<n; j++) {
495 		iblas = n - j - 1;
496 		BLAScopy_(&iblas,&a[j+1 + j*lda],&blas1, &a[j + (j+1)*lda],&blaslda);
497 		CHKMEMQ;
498 	    }
499 	    iblas = lda+1;
500 	    BLAScopy_(&blasn,wa1,&blas1,a,&iblas);
501 	    CHKMEMQ;
502 	    break;
503 	}
504 	par = PetscMax(parl,par+parc);
505     }
506     *retpar = par;
507     *retf = f;
508     *retinfo = info;
509     *retits = iter;
510     CHKMEMQ;
511     PetscFunctionReturn(0);
512 }
513