xref: /petsc/src/snes/impls/qn/qn.c (revision 84df9cb40eca90ea9b18a456fab7a4ecc7f6c1a4)
1 
2 #include <private/snesimpl.h>
3 
4 typedef struct {
5   PetscReal lambda;   /* The default step length for the update */
6   Vec * dX;           /* The change in X */
7   Vec * dF;           /* The change in F */
8   PetscInt m;         /* the number of kept previous steps */
9   PetscScalar * alpha;
10   PetscScalar * beta;
11   PetscScalar * rho;
12 } QNContext;
13 
14 #undef __FUNCT__
15 #define __FUNCT__ "LBGFSApplyJinv_Private"
16 PetscErrorCode LBGFSApplyJinv_Private(SNES snes, PetscInt it, Vec g, Vec z) {
17 
18   PetscErrorCode ierr;
19 
20   QNContext * qn = (QNContext *)snes->data;
21 
22   Vec * dX = qn->dX;
23   Vec * dF = qn->dF;
24 
25   PetscScalar * alpha = qn->alpha;
26   PetscScalar * beta = qn->beta;
27   PetscScalar * rho = qn->rho;
28 
29   PetscInt k, i;
30   PetscInt m = qn->m;
31   PetscScalar t;
32   PetscInt l = m;
33 
34   PetscFunctionBegin;
35 
36   if (it < m) l = it;
37 
38   ierr = VecCopy(g, z);CHKERRQ(ierr);
39 
40   /* outward recursion starting at iteration k's update and working back */
41   for (i = 0; i < l; i++) {
42     k = (it - i - 1) % m;
43     /* k = (it + i - l) % m; */
44     ierr = VecDot(dX[k], z, &t);CHKERRQ(ierr);
45     alpha[k] = t*rho[k];
46     ierr = VecAXPY(z, -alpha[k], dF[k]);CHKERRQ(ierr);
47   }
48 
49   /* inner application of the initial inverse jacobian approximation */
50   /* right now it's just the identity. Nothing needs to go here. */
51 
52   /* inward recursion starting at the first update and working forward*/
53   for (i = 0; i < l; i++) {
54     /* k = (it - i - 1) % m; */
55     k = (it + i - l) % m;
56     ierr = VecDot(dF[k], z, &t);CHKERRQ(ierr);
57     beta[k] = rho[k]*t;
58     ierr = VecAXPY(z, (alpha[k] - beta[k]), dX[k]);
59   }
60   ierr = VecScale(z, 1.0);CHKERRQ(ierr);
61 
62   PetscFunctionReturn(0);
63 }
64 
65 
66 #undef __FUNCT__
67 #define __FUNCT__ "QNLineSearchQuadratic"
68 PetscErrorCode QNLineSearchQuadratic(SNES snes,void *lsctx,Vec X,Vec F,Vec Y,PetscReal fnorm,PetscReal dummyXnorm,Vec G,Vec W,PetscReal *dummyYnorm,PetscReal *gnorm,PetscBool *flag)
69 {
70   PetscInt       i;
71   PetscReal      alphas[3] = {0.0, 0.5, 1.0};
72   PetscReal      norms[3];
73   PetscReal      alpha,a,b;
74   PetscErrorCode ierr;
75   PetscFunctionBegin;
76   norms[0]  = fnorm;
77   /* Calculate trial solutions */
78   for(i=1; i < 3; ++i) {
79     /* Calculate X^{n+1} = (1 - \alpha) X^n + \alpha Y */
80     ierr = VecCopy(X, W);CHKERRQ(ierr);
81     ierr = VecAXPBY(W, alphas[i], 1 - alphas[i], Y);CHKERRQ(ierr);
82     ierr = SNESComputeFunction(snes, W, F);CHKERRQ(ierr);
83     ierr = VecNorm(F, NORM_2, &norms[i]);CHKERRQ(ierr);
84   }
85   for(i = 0; i < 3; ++i) {
86     norms[i] = PetscSqr(norms[i]);
87   }
88   /* Fit a quadratic:
89        If we have x_{0,1,2} = 0, x_1, x_2 which generate norms y_{0,1,2}
90        a = (x_1 y_2 - x_2 y_1 + (x_2 - x_1) y_0)/(x^2_2 x_1 - x_2 x^2_1)
91        b = (x^2_1 y_2 - x^2_2 y_1 + (x^2_2 - x^2_1) y_0)/(x_2 x^2_1 - x^2_2 x_1)
92        c = y_0
93        x_min = -b/2a
94 
95        If we let x_{0,1,2} = 0, 0.5, 1.0
96        a = 2 y_2 - 4 y_1 + 2 y_0
97        b =  -y_2 + 4 y_1 - 3 y_0
98        c =   y_0
99   */
100   a = (alphas[1]*norms[2] - alphas[2]*norms[1] + (alphas[2] - alphas[1])*norms[0])/(PetscSqr(alphas[2])*alphas[1] - alphas[2]*PetscSqr(alphas[1]));
101   b = (PetscSqr(alphas[1])*norms[2] - PetscSqr(alphas[2])*norms[1] + (PetscSqr(alphas[2]) - PetscSqr(alphas[1]))*norms[0])/(alphas[2]*PetscSqr(alphas[1]) - PetscSqr(alphas[2])*alphas[1]);
102   /* Check for positive a (concave up) */
103   if (a >= 0.0) {
104     alpha = -b/(2.0*a);
105     alpha = PetscMin(alpha, alphas[2]);
106     alpha = PetscMax(alpha, alphas[0]);
107   } else {
108     alpha = 1.0;
109   }
110   ierr = VecAXPBY(X, alpha, 1 - alpha, Y);CHKERRQ(ierr);
111   ierr = SNESComputeFunction(snes, X, F);CHKERRQ(ierr);
112   if (alpha != 1.0) {
113     ierr = VecNorm(F, NORM_2, gnorm);CHKERRQ(ierr);
114   } else {
115     *gnorm = PetscSqrtReal(norms[2]);
116   }
117   *flag = PETSC_TRUE;
118   PetscFunctionReturn(0);
119 }
120 
121 
122 #undef __FUNCT__
123 #define __FUNCT__ "SNESSolve_QN"
124 static PetscErrorCode SNESSolve_QN(SNES snes)
125 {
126 
127   PetscErrorCode ierr;
128   QNContext * qn = (QNContext*) snes->data;
129 
130   Vec x, xold;
131   Vec f, fold;
132   Vec w, y;
133 
134   PetscInt i, k;
135 
136   PetscReal fnorm, xnorm;
137   PetscInt m = qn->m;
138   PetscBool ls_OK;
139 
140   PetscScalar rhosc;
141 
142   Vec * dX = qn->dX;
143   Vec * dF = qn->dF;
144   PetscScalar * rho = qn->rho;
145 
146   /* basically just a regular newton's method except for the application of the jacobian */
147   PetscFunctionBegin;
148 
149   x = snes->vec_sol;
150   xold = snes->vec_sol_update; /* dX_k */
151   w = snes->work[1];
152   f = snes->vec_func;
153   fold = snes->work[0];
154   y = snes->work[2];
155 
156   snes->reason = SNES_CONVERGED_ITERATING;
157 
158   ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
159   snes->iter = 0;
160   snes->norm = 0.;
161   ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
162   ierr = SNESComputeFunction(snes,x,f);CHKERRQ(ierr);
163   if (snes->domainerror) {
164     snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
165     PetscFunctionReturn(0);
166   }
167   ierr = VecNorm(f, NORM_2, &fnorm);CHKERRQ(ierr); /* fnorm <- ||F||  */
168   if (PetscIsInfOrNanReal(fnorm)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FP,"Infinite or not-a-number generated in norm");
169   ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
170   snes->norm = fnorm;
171   ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
172   SNESLogConvHistory(snes,fnorm,0);
173   ierr = SNESMonitor(snes,0,fnorm);CHKERRQ(ierr);
174 
175   /* set parameter for default relative tolerance convergence test */
176    snes->ttol = fnorm*snes->rtol;
177   /* test convergence */
178   ierr = (*snes->ops->converged)(snes,0,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
179   if (snes->reason) PetscFunctionReturn(0);
180   ierr = VecCopy(f, fold);CHKERRQ(ierr);
181   ierr = VecCopy(x, xold);CHKERRQ(ierr);
182   for(i = 0; i < snes->max_its; i++) {
183     /* general purpose update */
184     if (snes->ops->update) {
185       ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
186     }
187 
188     /* apply the current iteration of the approximate jacobian */
189     ierr = LBGFSApplyJinv_Private(snes, i, f, y);CHKERRQ(ierr);
190 
191     /* line search for lambda */
192     ierr = VecAYPX(y,-1.0,x);CHKERRQ(ierr);
193     ierr = QNLineSearchQuadratic(snes, PETSC_NULL, x, f, y, fnorm, xnorm, 0, w,&xnorm, &fnorm, &ls_OK);CHKERRQ(ierr);
194     ierr = SNESComputeFunction(snes, x, f);CHKERRQ(ierr);
195     ierr = VecNorm(f, NORM_2, &fnorm);CHKERRQ(ierr);
196     if (PetscIsInfOrNanReal(fnorm)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FP,"Infinite or not-a-number generated in norm");
197     ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
198     snes->norm = fnorm;
199     ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
200     SNESLogConvHistory(snes,fnorm,i+1);
201     ierr = SNESMonitor(snes,i+1,fnorm);CHKERRQ(ierr);
202     /* set parameter for default relative tolerance convergence test */
203     ierr = (*snes->ops->converged)(snes,i+1,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
204     if (snes->reason) PetscFunctionReturn(0);
205 
206     /* set the differences */
207     k = i % m;
208     ierr = VecCopy(f, dF[k]);CHKERRQ(ierr);
209     ierr = VecAXPY(dF[k], -1.0, fold);CHKERRQ(ierr);
210     ierr = VecCopy(x, dX[k]);CHKERRQ(ierr);
211     ierr = VecAXPY(dX[k], -1.0, xold);CHKERRQ(ierr);
212     ierr = VecDot(dX[k], dF[k], &rhosc);CHKERRQ(ierr);
213     rho[k] = 1. / rhosc;
214     ierr = VecCopy(f, fold);CHKERRQ(ierr);
215     ierr = VecCopy(x, xold);CHKERRQ(ierr);
216   }
217   if (i == snes->max_its) {
218     ierr = PetscInfo1(snes, "Maximum number of iterations has been reached: %D\n", snes->max_its);CHKERRQ(ierr);
219     if (!snes->reason) snes->reason = SNES_DIVERGED_MAX_IT;
220   }
221   PetscFunctionReturn(0);
222 }
223 
224 
225 #undef __FUNCT__
226 #define __FUNCT__ "SNESSetUp_QN"
227 static PetscErrorCode SNESSetUp_QN(SNES snes)
228 {
229   QNContext * qn = (QNContext *)snes->data;
230   PetscErrorCode ierr;
231   PetscFunctionBegin;
232   ierr = VecDuplicateVecs(snes->vec_sol, qn->m, &qn->dX);CHKERRQ(ierr);
233   ierr = VecDuplicateVecs(snes->vec_sol, qn->m, &qn->dF);CHKERRQ(ierr);
234   ierr = PetscMalloc3(qn->m, PetscScalar, &qn->alpha, qn->m, PetscScalar, &qn->beta, qn->m, PetscScalar, &qn->rho);CHKERRQ(ierr);
235   ierr = SNESDefaultGetWork(snes,3);CHKERRQ(ierr);
236   PetscFunctionReturn(0);
237 }
238 
239 #undef __FUNCT__
240 #define __FUNCT__ "SNESReset_QN"
241 static PetscErrorCode SNESReset_QN(SNES snes)
242 {
243   PetscErrorCode ierr;
244   QNContext * qn;
245   PetscFunctionBegin;
246   if (snes->data) {
247     qn = (QNContext *)snes->data;
248     if (qn->dX) {
249       ierr = VecDestroyVecs(qn->m, &qn->dX);CHKERRQ(ierr);
250     }
251     if (qn->dF) {
252       ierr = VecDestroyVecs(qn->m, &qn->dF);CHKERRQ(ierr);
253     }
254     ierr = PetscFree3(qn->alpha, qn->beta, qn->rho);CHKERRQ(ierr);
255   }
256   if (snes->work) {ierr = VecDestroyVecs(snes->nwork,&snes->work);CHKERRQ(ierr);}
257   PetscFunctionReturn(0);
258 }
259 
260 #undef __FUNCT__
261 #define __FUNCT__ "SNESDestroy_QN"
262 static PetscErrorCode SNESDestroy_QN(SNES snes)
263 {
264   PetscErrorCode ierr;
265   PetscFunctionBegin;
266   ierr = SNESReset_QN(snes);CHKERRQ(ierr);
267   ierr = PetscFree(snes->data);CHKERRQ(ierr);
268   PetscFunctionReturn(0);
269 }
270 
271 #undef __FUNCT__
272 #define __FUNCT__ "SNESSetFromOptions_QN"
273 static PetscErrorCode SNESSetFromOptions_QN(SNES snes)
274 {
275 
276   PetscErrorCode ierr;
277   QNContext * qn;
278 
279   PetscFunctionBegin;
280 
281   qn = (QNContext *)snes->data;
282 
283   ierr = PetscOptionsHead("SNES QN options");CHKERRQ(ierr);
284   ierr = PetscOptionsReal("-snes_ls_damping", "Damping parameter", "SNES", qn->lambda, &qn->lambda, PETSC_NULL);CHKERRQ(ierr);
285   ierr = PetscOptionsInt("-snes_qn_m", "Number of past states saved for L-Broyden methods", "SNES", qn->m, &qn->m, PETSC_NULL);CHKERRQ(ierr);
286   ierr = PetscOptionsTail();CHKERRQ(ierr);
287   PetscFunctionReturn(0);
288 }
289 
290 /* -------------------------------------------------------------------------- */
291 /*MC
292       SNESQN - Limited-Memory Quasi-Newton methods for the solution of nonlinear systems.
293 
294       Options Database:
295 
296 +     -snes_qn_m - Number of past states saved for the L-Broyden methods.
297 +     -snes_ls_damping - The damping parameter on the update to x.
298 
299       Notes: This implements the L-BFGS algorithm for the solution of F(x) = 0 using previous change in F(x) and x to
300       form the approximate inverse Jacobian using a series of multiplicative rank-one updates.  This will eventually be
301       generalized to implement several limited-memory Broyden methods.
302 
303       References:
304 
305       L-Broyden Methods: a generalization of the L-BFGS method to the limited memory Broyden family, M. B. Reed,
306       International Journal of Computer Mathematics, vol. 86, 2009.
307 
308 
309       Level: beginner
310 
311 .seealso:  SNESCreate(), SNES, SNESSetType(), SNESLS, SNESTR
312 
313 M*/
314 EXTERN_C_BEGIN
315 #undef __FUNCT__
316 #define __FUNCT__ "SNESCreate_QN"
317 PetscErrorCode  SNESCreate_QN(SNES snes)
318 {
319 
320   PetscErrorCode ierr;
321   QNContext * qn;
322 
323   PetscFunctionBegin;
324   snes->ops->setup           = SNESSetUp_QN;
325   snes->ops->solve           = SNESSolve_QN;
326   snes->ops->destroy         = SNESDestroy_QN;
327   snes->ops->setfromoptions  = SNESSetFromOptions_QN;
328   snes->ops->view            = 0;
329   snes->ops->reset           = SNESReset_QN;
330 
331   snes->usespc          = PETSC_TRUE;
332   snes->usesksp         = PETSC_FALSE;
333 
334   ierr = PetscNewLog(snes, QNContext, &qn);CHKERRQ(ierr);
335   snes->data = (void *) qn;
336   qn->m = 100;
337   qn->lambda = 1.;
338   qn->dX = PETSC_NULL;
339   qn->dF = PETSC_NULL;
340   PetscFunctionReturn(0);
341 }
342 EXTERN_C_END
343