xref: /petsc/src/snes/interface/noise/snesdnest.c (revision 047240e14af00aad1ef65e96f6fface8924f7f7e)
1 
2 /* fnoise/snesdnest.F -- translated by f2c (version 20020314).
3 */
4 #include <petscsys.h>
5 #define FALSE_ 0
6 #define TRUE_ 1
7 
8 /*  Noise estimation routine, written by Jorge More'.  Details are below. */
9 
10 /* Subroutine */ PetscErrorCode SNESNoise_dnest_(PetscInt *nf, double *fval,double *h__,double *fnoise, double *fder2, double *hopt, PetscInt *info, double *eps)
11 {
12   /* Initialized data */
13 
14   static double const__[15] = { .71,.41,.23,.12,.063,.033,.018,.0089,
15                                 .0046,.0024,.0012,6.1e-4,3.1e-4,1.6e-4,8e-5 };
16 
17   /* System generated locals */
18   PetscInt i__1;
19   double   d__1, d__2, d__3, d__4;
20 
21 
22   /* Local variables */
23   static double   emin, emax;
24   static PetscInt dsgn[6];
25   static double   f_max, f_min, stdv;
26   static PetscInt i__, j;
27   static double   scale;
28   static PetscInt mh;
29   static PetscInt cancel[6], dnoise;
30   static double   err2, est1, est2, est3, est4;
31 
32 /*     ********** */
33 
34 /*     Subroutine dnest */
35 
36 /*     This subroutine estimates the noise in a function */
37 /*     and provides estimates of the optimal difference parameter */
38 /*     for a forward-difference approximation. */
39 
40 /*     The user must provide a difference parameter h, and the */
41 /*     function value at nf points centered around the current point. */
42 /*     For example, if nf = 7, the user must provide */
43 
44 /*        f(x-2*h), f(x-h), f(x), f(x+h),  f(x+2*h), */
45 
46 /*     in the array fval. The use of nf = 7 function evaluations is */
47 /*     recommended. */
48 
49 /*     The noise in the function is roughly defined as the variance in */
50 /*     the computed value of the function. The noise in the function */
51 /*     provides valuable information. For example, function values */
52 /*     smaller than the noise should be considered to be zero. */
53 
54 /*     This subroutine requires an initial estimate for h. Under estimates */
55 /*     are usually preferred. If noise is not detected, the user should */
56 /*     increase or decrease h according to the ouput value of info. */
57 /*     In most cases, the subroutine detects noise with the initial */
58 /*     value of h. */
59 
60 /*     The subroutine statement is */
61 
62 /*       subroutine dnest(nf,fval,h,hopt,fnoise,info,eps) */
63 
64 /*     where */
65 
66 /*       nf is an int variable. */
67 /*         On entry nf is the number of function values. */
68 /*         On exit nf is unchanged. */
69 
70 /*       f is a double precision array of dimension nf. */
71 /*         On entry f contains the function values. */
72 /*         On exit f is overwritten. */
73 
74 /*       h is a double precision variable. */
75 /*         On entry h is an estimate of the optimal difference parameter. */
76 /*         On exit h is unchanged. */
77 
78 /*       fnoise is a double precision variable. */
79 /*         On entry fnoise need not be specified. */
80 /*         On exit fnoise is set to an estimate of the function noise */
81 /*            if noise is detected; otherwise fnoise is set to zero. */
82 
83 /*       hopt is a double precision variable. */
84 /*         On entry hopt need not be specified. */
85 /*         On exit hopt is set to an estimate of the optimal difference */
86 /*            parameter if noise is detected; otherwise hopt is set to zero. */
87 
88 /*       info is an int variable. */
89 /*         On entry info need not be specified. */
90 /*         On exit info is set as follows: */
91 
92 /*            info = 1  Noise has been detected. */
93 
94 /*            info = 2  Noise has not been detected; h is too small. */
95 /*                      Try 100*h for the next value of h. */
96 
97 /*            info = 3  Noise has not been detected; h is too large. */
98 /*                      Try h/100 for the next value of h. */
99 
100 /*            info = 4  Noise has been detected but the estimate of hopt */
101 /*                      is not reliable; h is too small. */
102 
103 /*       eps is a double precision work array of dimension nf. */
104 
105 /*     MINPACK-2 Project. April 1997. */
106 /*     Argonne National Laboratory. */
107 /*     Jorge J. More'. */
108 
109 /*     ********** */
110   /* Parameter adjustments */
111   --eps;
112   --fval;
113 
114   /* Function Body */
115   *fnoise = 0.;
116   *fder2  = 0.;
117   *hopt   = 0.;
118 /*     Compute an estimate of the second derivative and */
119 /*     determine a bound on the error. */
120   mh   = (*nf + 1) / 2;
121   est1 = (fval[mh + 1] - fval[mh] * 2 + fval[mh - 1]) / *h__ / *h__;
122   est2 = (fval[mh + 2] - fval[mh] * 2 + fval[mh - 2]) / (*h__ * 2) / (*h__ * 2);
123   est3 = (fval[mh + 3] - fval[mh] * 2 + fval[mh - 3]) / (*h__ * 3) / (*h__ * 3);
124   est4 = (est1 + est2 + est3) / 3;
125 /* Computing MAX */
126 /* Computing PETSCMAX */
127   d__3 = PetscMax(est1,est2);
128 /* Computing MIN */
129   d__4 = PetscMin(est1,est2);
130   d__1 = PetscMax(d__3,est3) - est4, d__2 = est4 - PetscMin(d__4,est3);
131   err2 = PetscMax(d__1,d__2);
132 /*      write (2,123) est1, est2, est3 */
133 /* 123  format ('Second derivative estimates', 3d12.2) */
134   if (err2 <= PetscAbsScalar(est4) * .1) *fder2 = est4;
135   else if (err2 < PetscAbsScalar(est4))  *fder2 = est3;
136   else *fder2 = 0.;
137 
138 /*     Compute the range of function values. */
139   f_min = fval[1];
140   f_max = fval[1];
141   i__1  = *nf;
142   for (i__ = 2; i__ <= i__1; ++i__) {
143     /* Computing MIN */
144     d__1 = f_min, d__2 = fval[i__];
145     f_min = PetscMin(d__1,d__2);
146 
147     /* Computing MAX */
148     d__1 = f_max, d__2 = fval[i__];
149     f_max = PetscMax(d__1,d__2);
150   }
151 /*     Construct the difference table. */
152   dnoise = FALSE_;
153   for (j = 1; j <= 6; ++j) {
154     dsgn[j - 1]   = FALSE_;
155     cancel[j - 1] = FALSE_;
156     scale         = 0.;
157     i__1          = *nf - j;
158     for (i__ = 1; i__ <= i__1; ++i__) {
159       fval[i__] = fval[i__ + 1] - fval[i__];
160       if (fval[i__] == 0.) cancel[j - 1] = TRUE_;
161 
162       /* Computing MAX */
163       d__2 = scale, d__3 = (d__1 = fval[i__], PetscAbsScalar(d__1));
164       scale = PetscMax(d__2,d__3);
165     }
166 
167     /*        Compute the estimates for the noise level. */
168     if (scale == 0.) stdv = 0.;
169     else {
170       stdv = 0.;
171       i__1 = *nf - j;
172       for (i__ = 1; i__ <= i__1; ++i__) {
173         /* Computing 2nd power */
174         d__1 = fval[i__] / scale;
175         stdv += d__1 * d__1;
176       }
177       stdv = scale * PetscSqrtScalar(stdv / (*nf - j));
178     }
179     eps[j] = const__[j - 1] * stdv;
180 /*        Determine differences in sign. */
181     i__1 = *nf - j - 1;
182     for (i__ = 1; i__ <= i__1; ++i__) {
183       /* Computing MIN */
184       d__1 = fval[i__], d__2 = fval[i__ + 1];
185       /* Computing MAX */
186       d__3 = fval[i__], d__4 = fval[i__ + 1];
187       if (PetscMin(d__1,d__2) < 0. && PetscMax(d__3,d__4) > 0.) dsgn[j - 1] = TRUE_;
188     }
189   }
190   /*     First requirement for detection of noise. */
191   dnoise = dsgn[3];
192   /*     Check for h too small or too large. */
193   *info = 0;
194   if (f_max == f_min) *info = 2;
195   else /* if (complicated condition) */ {
196     /* Computing MIN */
197     d__1 = PetscAbsScalar(f_max), d__2 = PetscAbsScalar(f_min);
198     if (f_max - f_min > PetscMin(d__1,d__2) * .1) *info = 3;
199   }
200   if (*info != 0) PetscFunctionReturn(0);
201 
202   /*     Determine the noise level. */
203   /* Computing MIN */
204   d__1 = PetscMin(eps[4],eps[5]);
205   emin = PetscMin(d__1,eps[6]);
206 
207   /* Computing MAX */
208   d__1 = PetscMax(eps[4],eps[5]);
209   emax = PetscMax(d__1,eps[6]);
210 
211   if (emax <= emin * 4 && dnoise) {
212     *fnoise = (eps[4] + eps[5] + eps[6]) / 3;
213     if (*fder2 != 0.) {
214       *info = 1;
215       *hopt = PetscSqrtScalar(*fnoise / PetscAbsScalar(*fder2)) * 1.68;
216     } else {
217       *info = 4;
218       *hopt = *h__ * 10;
219     }
220     PetscFunctionReturn(0);
221   }
222 
223   /* Computing MIN */
224   d__1 = PetscMin(eps[3],eps[4]);
225   emin = PetscMin(d__1,eps[5]);
226 
227   /* Computing MAX */
228   d__1 = PetscMax(eps[3],eps[4]);
229   emax = PetscMax(d__1,eps[5]);
230 
231   if (emax <= emin * 4 && dnoise) {
232     *fnoise = (eps[3] + eps[4] + eps[5]) / 3;
233     if (*fder2 != 0.) {
234       *info = 1;
235       *hopt = PetscSqrtScalar(*fnoise / PetscAbsScalar(*fder2)) * 1.68;
236     } else {
237       *info = 4;
238       *hopt = *h__ * 10;
239     }
240     PetscFunctionReturn(0);
241   }
242 /*     Noise not detected; decide if h is too small or too large. */
243   if (!cancel[3]) {
244     if (dsgn[3]) *info = 2;
245     else *info = 3;
246     PetscFunctionReturn(0);
247   }
248   if (!cancel[2]) {
249     if (dsgn[2]) *info = 2;
250     else *info = 3;
251     PetscFunctionReturn(0);
252   }
253 /*     If there is cancelllation on the third and fourth column */
254 /*     then h is too small */
255   *info = 2;
256   PetscFunctionReturn(0);
257 /*      if (cancel .or. dsgn(3)) then */
258 /*         info = 2 */
259 /*      else */
260 /*         info = 3 */
261 /*      end if */
262 } /* dnest_ */
263 
264