xref: /petsc/src/tao/leastsquares/impls/pounders/gqt.c (revision 31d78bcd2b98084dc1368b20eb1129c8b9fb39fe)
1 #include <petscsys.h>
2 #include <petscblaslapack.h>
3 
estsv(PetscInt n,PetscReal * r,PetscInt ldr,PetscReal * svmin,PetscReal * z)4 static PetscErrorCode estsv(PetscInt n, PetscReal *r, PetscInt ldr, PetscReal *svmin, PetscReal *z)
5 {
6   PetscBLASInt blas1 = 1, blasn, blasnmi, blasj, blasldr;
7   PetscInt     i, j;
8   PetscReal    e, temp, w, wm, ynorm, znorm, s, sm;
9 
10   PetscFunctionBegin;
11   PetscCall(PetscBLASIntCast(n, &blasn));
12   PetscCall(PetscBLASIntCast(ldr, &blasldr));
13   for (i = 0; i < n; i++) z[i] = 0.0;
14   e = PetscAbs(r[0]);
15   if (e == 0.0) {
16     *svmin = 0.0;
17     z[0]   = 1.0;
18   } else {
19     /* Solve R'*y = e */
20     for (i = 0; i < n; i++) {
21       /* Scale y. The scaling factor (0.01) reduces the number of scalings */
22       if (z[i] >= 0.0) e = -PetscAbs(e);
23       else e = PetscAbs(e);
24 
25       if (PetscAbs(e - z[i]) > PetscAbs(r[i + ldr * i])) {
26         temp = PetscMin(0.01, PetscAbs(r[i + ldr * i])) / PetscAbs(e - z[i]);
27         PetscCallBLAS("BLASscal", BLASscal_(&blasn, &temp, z, &blas1));
28         e = temp * e;
29       }
30 
31       /* Determine the two possible choices of y[i] */
32       if (r[i + ldr * i] == 0.0) {
33         w = wm = 1.0;
34       } else {
35         w  = (e - z[i]) / r[i + ldr * i];
36         wm = -(e + z[i]) / r[i + ldr * i];
37       }
38 
39       /*  Chose y[i] based on the predicted value of y[j] for j>i */
40       s  = PetscAbs(e - z[i]);
41       sm = PetscAbs(e + z[i]);
42       for (j = i + 1; j < n; j++) sm += PetscAbs(z[j] + wm * r[i + ldr * j]);
43       if (i < n - 1) {
44         PetscCall(PetscBLASIntCast(n - i - 1, &blasnmi));
45         PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasnmi, &w, &r[i + ldr * (i + 1)], &blasldr, &z[i + 1], &blas1));
46         PetscCallBLAS("BLASasum", s += BLASasum_(&blasnmi, &z[i + 1], &blas1));
47       }
48       if (s < sm) {
49         temp = wm - w;
50         w    = wm;
51         if (i < n - 1) PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasnmi, &temp, &r[i + ldr * (i + 1)], &blasldr, &z[i + 1], &blas1));
52       }
53       z[i] = w;
54     }
55 
56     PetscCallBLAS("BLASnrm2", ynorm = BLASnrm2_(&blasn, z, &blas1));
57 
58     /* Solve R*z = y */
59     for (j = n - 1; j >= 0; j--) {
60       /* Scale z */
61       if (PetscAbs(z[j]) > PetscAbs(r[j + ldr * j])) {
62         temp = PetscMin(0.01, PetscAbs(r[j + ldr * j] / z[j]));
63         PetscCallBLAS("BLASscal", BLASscal_(&blasn, &temp, z, &blas1));
64         ynorm *= temp;
65       }
66       if (r[j + ldr * j] == 0) {
67         z[j] = 1.0;
68       } else {
69         z[j] = z[j] / r[j + ldr * j];
70       }
71       temp = -z[j];
72       PetscCall(PetscBLASIntCast(j, &blasj));
73       PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasj, &temp, &r[0 + ldr * j], &blas1, z, &blas1));
74     }
75 
76     /* Compute svmin and normalize z */
77     PetscCallBLAS("BLASnrm2", znorm = 1.0 / BLASnrm2_(&blasn, z, &blas1));
78     *svmin = ynorm * znorm;
79     PetscCallBLAS("BLASscal", BLASscal_(&blasn, &znorm, z, &blas1));
80   }
81   PetscFunctionReturn(PETSC_SUCCESS);
82 }
83 
84 /*
85 c     ***********
86 c
87 c     Subroutine gqt
88 c
89 c     Given an n by n symmetric matrix A, an n-vector b, and a
90 c     positive number delta, this subroutine determines a vector
91 c     x which approximately minimizes the quadratic function
92 c
93 c           f(x) = (1/2)*x'*A*x + b'*x
94 c
95 c     subject to the Euclidean norm constraint
96 c
97 c           norm(x) <= delta.
98 c
99 c     This subroutine computes an approximation x and a Lagrange
100 c     multiplier par such that either par is zero and
101 c
102 c            norm(x) <= (1+rtol)*delta,
103 c
104 c     or par is positive and
105 c
106 c            abs(norm(x) - delta) <= rtol*delta.
107 c
108 c     If xsol is the solution to the problem, the approximation x
109 c     satisfies
110 c
111 c            f(x) <= ((1 - rtol)**2)*f(xsol)
112 c
113 c     The subroutine statement is
114 c
115 c       subroutine gqt(n,a,lda,b,delta,rtol,atol,itmax,
116 c                        par,f,x,info,z,wa1,wa2)
117 c
118 c     where
119 c
120 c       n is an integer variable.
121 c         On entry n is the order of A.
122 c         On exit n is unchanged.
123 c
124 c       a is a double precision array of dimension (lda,n).
125 c         On entry the full upper triangle of a must contain the
126 c            full upper triangle of the symmetric matrix A.
127 c         On exit the array contains the matrix A.
128 c
129 c       lda is an integer variable.
130 c         On entry lda is the leading dimension of the array a.
131 c         On exit lda is unchanged.
132 c
133 c       b is an double precision array of dimension n.
134 c         On entry b specifies the linear term in the quadratic.
135 c         On exit b is unchanged.
136 c
137 c       delta is a double precision variable.
138 c         On entry delta is a bound on the Euclidean norm of x.
139 c         On exit delta is unchanged.
140 c
141 c       rtol is a double precision variable.
142 c         On entry rtol is the relative accuracy desired in the
143 c            solution. Convergence occurs if
144 c
145 c              f(x) <= ((1 - rtol)**2)*f(xsol)
146 c
147 c         On exit rtol is unchanged.
148 c
149 c       atol is a double precision variable.
150 c         On entry atol is the absolute accuracy desired in the
151 c            solution. Convergence occurs when
152 c
153 c              norm(x) <= (1 + rtol)*delta
154 c
155 c              max(-f(x),-f(xsol)) <= atol
156 c
157 c         On exit atol is unchanged.
158 c
159 c       itmax is an integer variable.
160 c         On entry itmax specifies the maximum number of iterations.
161 c         On exit itmax is unchanged.
162 c
163 c       par is a double precision variable.
164 c         On entry par is an initial estimate of the Lagrange
165 c            multiplier for the constraint norm(x) <= delta.
166 c         On exit par contains the final estimate of the multiplier.
167 c
168 c       f is a double precision variable.
169 c         On entry f need not be specified.
170 c         On exit f is set to f(x) at the output x.
171 c
172 c       x is a double precision array of dimension n.
173 c         On entry x need not be specified.
174 c         On exit x is set to the final estimate of the solution.
175 c
176 c       info is an integer variable.
177 c         On entry info need not be specified.
178 c         On exit info is set as follows:
179 c
180 c            info = 1  The function value f(x) has the relative
181 c                      accuracy specified by rtol.
182 c
183 c            info = 2  The function value f(x) has the absolute
184 c                      accuracy specified by atol.
185 c
186 c            info = 3  Rounding errors prevent further progress.
187 c                      On exit x is the best available approximation.
188 c
189 c            info = 4  Failure to converge after itmax iterations.
190 c                      On exit x is the best available approximation.
191 c
192 c       z is a double precision work array of dimension n.
193 c
194 c       wa1 is a double precision work array of dimension n.
195 c
196 c       wa2 is a double precision work array of dimension n.
197 c
198 c     Subprograms called
199 c
200 c       MINPACK-2  ......  destsv
201 c
202 c       LAPACK  .........  dpotrf
203 c
204 c       Level 1 BLAS  ...  daxpy, dcopy, ddot, dnrm2, dscal
205 c
206 c       Level 2 BLAS  ...  dtrmv, dtrsv
207 c
208 c     MINPACK-2 Project. October 1993.
209 c     Argonne National Laboratory and University of Minnesota.
210 c     Brett M. Averick, Richard Carter, and Jorge J. More'
211 c
212 c     ***********
213 */
gqt(PetscInt n,PetscReal * a,PetscInt lda,PetscReal * b,PetscReal delta,PetscReal rtol,PetscReal atol,PetscInt itmax,PetscReal * retpar,PetscReal * retf,PetscReal * x,PetscInt * retinfo,PetscInt * retits,PetscReal * z,PetscReal * wa1,PetscReal * wa2)214 PetscErrorCode gqt(PetscInt n, PetscReal *a, PetscInt lda, PetscReal *b, PetscReal delta, PetscReal rtol, PetscReal atol, PetscInt itmax, PetscReal *retpar, PetscReal *retf, PetscReal *x, PetscInt *retinfo, PetscInt *retits, PetscReal *z, PetscReal *wa1, PetscReal *wa2)
215 {
216   PetscReal    f = 0.0, p001 = 0.001, p5 = 0.5, minusone = -1, delta2 = delta * delta;
217   PetscInt     iter, j, rednc, info;
218   PetscBLASInt indef;
219   PetscBLASInt blas1 = 1, blasn, iblas, blaslda, blasldap1, blasinfo;
220   PetscReal    alpha, anorm, bnorm, parc, parf, parl, pars, par = *retpar, paru, prod, rxnorm, rznorm = 0.0, temp, xnorm;
221 
222   PetscFunctionBegin;
223   PetscCall(PetscBLASIntCast(n, &blasn));
224   PetscCall(PetscBLASIntCast(lda, &blaslda));
225   PetscCall(PetscBLASIntCast(lda + 1, &blasldap1));
226   parf   = 0.0;
227   xnorm  = 0.0;
228   rxnorm = 0.0;
229   rednc  = 0;
230   for (j = 0; j < n; j++) {
231     x[j] = 0.0;
232     z[j] = 0.0;
233   }
234 
235   /* Copy the diagonal and save A in its lower triangle */
236   PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, a, &blasldap1, wa1, &blas1));
237   for (j = 0; j < n - 1; j++) {
238     PetscCall(PetscBLASIntCast(n - j - 1, &iblas));
239     PetscCallBLAS("BLAScopy", BLAScopy_(&iblas, &a[j + lda * (j + 1)], &blaslda, &a[j + 1 + lda * j], &blas1));
240   }
241 
242   /* Calculate the l1-norm of A, the Gershgorin row sums, and the
243    l2-norm of b */
244   anorm = 0.0;
245   for (j = 0; j < n; j++) {
246     PetscCallBLAS("BLASasum", wa2[j] = BLASasum_(&blasn, &a[0 + lda * j], &blas1));
247     CHKMEMQ;
248     anorm = PetscMax(anorm, wa2[j]);
249   }
250   for (j = 0; j < n; j++) wa2[j] = wa2[j] - PetscAbs(wa1[j]);
251   PetscCallBLAS("BLASnrm2", bnorm = BLASnrm2_(&blasn, b, &blas1));
252   CHKMEMQ;
253   /* Calculate a lower bound, pars, for the domain of the problem.
254    Also calculate an upper bound, paru, and a lower bound, parl,
255    for the Lagrange multiplier. */
256   pars = parl = paru = -anorm;
257   for (j = 0; j < n; j++) {
258     pars = PetscMax(pars, -wa1[j]);
259     parl = PetscMax(parl, wa1[j] + wa2[j]);
260     paru = PetscMax(paru, -wa1[j] + wa2[j]);
261   }
262   parl = PetscMax(bnorm / delta - parl, pars);
263   parl = PetscMax(0.0, parl);
264   paru = PetscMax(0.0, bnorm / delta + paru);
265 
266   /* If the input par lies outside of the interval (parl, paru),
267    set par to the closer endpoint. */
268 
269   par = PetscMax(par, parl);
270   par = PetscMin(par, paru);
271 
272   /* Special case: parl == paru */
273   paru = PetscMax(paru, (1.0 + rtol) * parl);
274 
275   /* Beginning of an iteration */
276 
277   info = 0;
278   for (iter = 1; iter <= itmax; iter++) {
279     /* Safeguard par */
280     if (par <= pars && paru > 0) par = PetscMax(p001, PetscSqrtScalar(parl / paru)) * paru;
281 
282     /* Copy the lower triangle of A into its upper triangle and  compute A + par*I */
283 
284     for (j = 0; j < n - 1; j++) {
285       PetscCall(PetscBLASIntCast(n - j - 1, &iblas));
286       PetscCallBLAS("BLAScopy", BLAScopy_(&iblas, &a[j + 1 + j * lda], &blas1, &a[j + (j + 1) * lda], &blaslda));
287     }
288     for (j = 0; j < n; j++) a[j + j * lda] = wa1[j] + par;
289 
290     /* Attempt the Cholesky factorization of A without referencing the lower triangular part. */
291     PetscCallBLAS("LAPACKpotrf", LAPACKpotrf_("U", &blasn, a, &blaslda, &indef));
292 
293     /* Case 1: A + par*I is pos. def. */
294     if (indef == 0) {
295       /* Compute an approximate solution x and save the last value of par with A + par*I pos. def. */
296 
297       parf = par;
298       PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, b, &blas1, wa2, &blas1));
299       PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "T", "N", &blasn, &blas1, a, &blaslda, wa2, &blasn, &blasinfo));
300       PetscCheck(!blasinfo, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACKtrtrs() returned info %" PetscBLASInt_FMT, blasinfo);
301       PetscCallBLAS("BLASnrm2", rxnorm = BLASnrm2_(&blasn, wa2, &blas1));
302       PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "N", "N", &blasn, &blas1, a, &blaslda, wa2, &blasn, &blasinfo));
303       PetscCheck(!blasinfo, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACKtrtrs() returned info %" PetscBLASInt_FMT, blasinfo);
304 
305       PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, wa2, &blas1, x, &blas1));
306       PetscCallBLAS("BLASscal", BLASscal_(&blasn, &minusone, x, &blas1));
307       PetscCallBLAS("BLASnrm2", xnorm = BLASnrm2_(&blasn, x, &blas1));
308       CHKMEMQ;
309 
310       /* Test for convergence */
311       if (PetscAbs(xnorm - delta) <= rtol * delta || (par == 0 && xnorm <= (1.0 + rtol) * delta)) info = 1;
312 
313       /* Compute a direction of negative curvature and use this information to improve pars. */
314       PetscCall(estsv(n, a, lda, &rznorm, z));
315       CHKMEMQ;
316       pars = PetscMax(pars, par - rznorm * rznorm);
317 
318       /* Compute a negative curvature solution of the form x + alpha*z,  where norm(x+alpha*z)==delta */
319 
320       rednc = 0;
321       if (xnorm < delta) {
322         /* Compute alpha */
323         PetscCallBLAS("BLASdot", prod = BLASdot_(&blasn, z, &blas1, x, &blas1) / delta);
324         temp  = (delta - xnorm) * ((delta + xnorm) / delta);
325         alpha = temp / (PetscAbs(prod) + PetscSqrtScalar(prod * prod + temp / delta));
326         if (prod >= 0) alpha = PetscAbs(alpha);
327         else alpha = -PetscAbs(alpha);
328 
329         /* Test to decide if the negative curvature step produces a larger reduction than with z=0 */
330         rznorm = PetscAbs(alpha) * rznorm;
331         if ((rznorm * rznorm + par * xnorm * xnorm) / (delta2) <= par) rednc = 1;
332         /* Test for convergence */
333         if (p5 * rznorm * rznorm / delta2 <= rtol * (1.0 - p5 * rtol) * (par + rxnorm * rxnorm / delta2)) {
334           info = 1;
335         } else if (info == 0 && (p5 * (par + rxnorm * rxnorm / delta2) <= atol / delta2)) {
336           info = 2;
337         }
338       }
339 
340       /* Compute the Newton correction parc to par. */
341       if (xnorm == 0) {
342         parc = -par;
343       } else {
344         PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, x, &blas1, wa2, &blas1));
345         temp = 1.0 / xnorm;
346         PetscCallBLAS("BLASscal", BLASscal_(&blasn, &temp, wa2, &blas1));
347         PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "T", "N", &blasn, &blas1, a, &blaslda, wa2, &blasn, &blasinfo));
348         PetscCheck(!blasinfo, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACKtrtrs() returned info %" PetscBLASInt_FMT, blasinfo);
349         PetscCallBLAS("BLASnrm2", temp = BLASnrm2_(&blasn, wa2, &blas1));
350         parc = (xnorm - delta) / (delta * temp * temp);
351       }
352 
353       /* update parl or paru */
354       if (xnorm > delta) {
355         parl = PetscMax(parl, par);
356       } else if (xnorm < delta) {
357         paru = PetscMin(paru, par);
358       }
359     } else {
360       /* Case 2: A + par*I is not pos. def. */
361 
362       /* Use the rank information from the Cholesky decomposition to update par. */
363 
364       if (indef > 1) {
365         /* Restore column indef to A + par*I. */
366         iblas = indef - 1;
367         PetscCallBLAS("BLAScopy", BLAScopy_(&iblas, &a[indef - 1 + 0 * lda], &blaslda, &a[0 + (indef - 1) * lda], &blas1));
368         a[indef - 1 + (indef - 1) * lda] = wa1[indef - 1] + par;
369 
370         /* compute parc. */
371         PetscCallBLAS("BLAScopy", BLAScopy_(&iblas, &a[0 + (indef - 1) * lda], &blas1, wa2, &blas1));
372         PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "T", "N", &iblas, &blas1, a, &blaslda, wa2, &blasn, &blasinfo));
373         PetscCheck(!blasinfo, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACKtrtrs() returned info %" PetscBLASInt_FMT, blasinfo);
374         PetscCallBLAS("BLAScopy", BLAScopy_(&iblas, wa2, &blas1, &a[0 + (indef - 1) * lda], &blas1));
375         PetscCallBLAS("BLASnrm2", temp = BLASnrm2_(&iblas, &a[0 + (indef - 1) * lda], &blas1));
376         CHKMEMQ;
377         a[indef - 1 + (indef - 1) * lda] -= temp * temp;
378         PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "N", "N", &iblas, &blas1, a, &blaslda, wa2, &blasn, &blasinfo));
379         PetscCheck(!blasinfo, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACKtrtrs() returned info %" PetscBLASInt_FMT, blasinfo);
380       }
381 
382       wa2[indef - 1] = -1.0;
383       iblas          = indef;
384       PetscCallBLAS("BLASnrm2", temp = BLASnrm2_(&iblas, wa2, &blas1));
385       parc = -a[indef - 1 + (indef - 1) * lda] / (temp * temp);
386       pars = PetscMax(pars, par + parc);
387 
388       /* If necessary, increase paru slightly.
389        This is needed because in some exceptional situations
390        paru is the optimal value of par. */
391 
392       paru = PetscMax(paru, (1.0 + rtol) * pars);
393     }
394 
395     /* Use pars to update parl */
396     parl = PetscMax(parl, pars);
397 
398     /* Test for converged. */
399     if (info == 0) {
400       if (iter == itmax) info = 4;
401       if (paru <= (1.0 + p5 * rtol) * pars) info = 3;
402       if (paru == 0.0) info = 2;
403     }
404 
405     /* If exiting, store the best approximation and restore
406      the upper triangle of A. */
407 
408     if (info != 0) {
409       /* Compute the best current estimates for x and f. */
410       par = parf;
411       f   = -p5 * (rxnorm * rxnorm + par * xnorm * xnorm);
412       if (rednc) {
413         f = -p5 * (rxnorm * rxnorm + par * delta * delta - rznorm * rznorm);
414         PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &alpha, z, &blas1, x, &blas1));
415       }
416       /* Restore the upper triangle of A */
417       for (j = 0; j < n; j++) {
418         PetscCall(PetscBLASIntCast(n - j - 1, &iblas));
419         PetscCallBLAS("BLAScopy", BLAScopy_(&iblas, &a[j + 1 + j * lda], &blas1, &a[j + (j + 1) * lda], &blaslda));
420       }
421       PetscCall(PetscBLASIntCast(lda + 1, &iblas));
422       PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, wa1, &blas1, a, &iblas));
423       break;
424     }
425     par = PetscMax(parl, par + parc);
426   }
427   *retpar  = par;
428   *retf    = f;
429   *retinfo = info;
430   *retits  = iter;
431   CHKMEMQ;
432   PetscFunctionReturn(PETSC_SUCCESS);
433 }
434