Mercurial > repos > vipints > rdiff
comparison rDiff/src/locfit/Source/libmut.c @ 0:0f80a5141704
version 0.3 uploaded
author | vipints |
---|---|
date | Thu, 14 Feb 2013 23:38:36 -0500 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:0f80a5141704 |
---|---|
1 /* | |
2 * Copyright 1996-2006 Catherine Loader. | |
3 */ | |
4 | |
5 #include "mex.h" | |
6 /* | |
7 * Copyright 1996-2006 Catherine Loader. | |
8 */ | |
9 #include <math.h> | |
10 #include "mut.h" | |
11 | |
12 /* stirlerr(n) = log(n!) - log( sqrt(2*pi*n)*(n/e)^n ) */ | |
13 | |
14 #define S0 0.083333333333333333333 /* 1/12 */ | |
15 #define S1 0.00277777777777777777778 /* 1/360 */ | |
16 #define S2 0.00079365079365079365079365 /* 1/1260 */ | |
17 #define S3 0.000595238095238095238095238 /* 1/1680 */ | |
18 #define S4 0.0008417508417508417508417508 /* 1/1188 */ | |
19 | |
20 /* | |
21 error for 0, 0.5, 1.0, 1.5, ..., 14.5, 15.0. | |
22 */ | |
23 static double sferr_halves[31] = { | |
24 0.0, /* n=0 - wrong, place holder only */ | |
25 0.1534264097200273452913848, /* 0.5 */ | |
26 0.0810614667953272582196702, /* 1.0 */ | |
27 0.0548141210519176538961390, /* 1.5 */ | |
28 0.0413406959554092940938221, /* 2.0 */ | |
29 0.03316287351993628748511048, /* 2.5 */ | |
30 0.02767792568499833914878929, /* 3.0 */ | |
31 0.02374616365629749597132920, /* 3.5 */ | |
32 0.02079067210376509311152277, /* 4.0 */ | |
33 0.01848845053267318523077934, /* 4.5 */ | |
34 0.01664469118982119216319487, /* 5.0 */ | |
35 0.01513497322191737887351255, /* 5.5 */ | |
36 0.01387612882307074799874573, /* 6.0 */ | |
37 0.01281046524292022692424986, /* 6.5 */ | |
38 0.01189670994589177009505572, /* 7.0 */ | |
39 0.01110455975820691732662991, /* 7.5 */ | |
40 0.010411265261972096497478567, /* 8.0 */ | |
41 0.009799416126158803298389475, /* 8.5 */ | |
42 0.009255462182712732917728637, /* 9.0 */ | |
43 0.008768700134139385462952823, /* 9.5 */ | |
44 0.008330563433362871256469318, /* 10.0 */ | |
45 0.007934114564314020547248100, /* 10.5 */ | |
46 0.007573675487951840794972024, /* 11.0 */ | |
47 0.007244554301320383179543912, /* 11.5 */ | |
48 0.006942840107209529865664152, /* 12.0 */ | |
49 0.006665247032707682442354394, /* 12.5 */ | |
50 0.006408994188004207068439631, /* 13.0 */ | |
51 0.006171712263039457647532867, /* 13.5 */ | |
52 0.005951370112758847735624416, /* 14.0 */ | |
53 0.005746216513010115682023589, /* 14.5 */ | |
54 0.005554733551962801371038690 /* 15.0 */ | |
55 }; | |
56 | |
57 double stirlerr(n) | |
58 double n; | |
59 { double nn; | |
60 | |
61 if (n<15.0) | |
62 { nn = 2.0*n; | |
63 if (nn==(int)nn) return(sferr_halves[(int)nn]); | |
64 return(mut_lgamma(n+1.0) - (n+0.5)*log((double)n)+n - HF_LG_PIx2); | |
65 } | |
66 | |
67 nn = (double)n; | |
68 nn = nn*nn; | |
69 if (n>500) return((S0-S1/nn)/n); | |
70 if (n>80) return((S0-(S1-S2/nn)/nn)/n); | |
71 if (n>35) return((S0-(S1-(S2-S3/nn)/nn)/nn)/n); | |
72 return((S0-(S1-(S2-(S3-S4/nn)/nn)/nn)/nn)/n); | |
73 } | |
74 | |
75 double bd0(x,np) | |
76 double x, np; | |
77 { double ej, s, s1, v; | |
78 int j; | |
79 if (fabs(x-np)<0.1*(x+np)) | |
80 { | |
81 s = (x-np)*(x-np)/(x+np); | |
82 v = (x-np)/(x+np); | |
83 ej = 2*x*v; v = v*v; | |
84 for (j=1; ;++j) | |
85 { ej *= v; | |
86 s1 = s+ej/((j<<1)+1); | |
87 if (s1==s) return(s1); | |
88 s = s1; | |
89 } | |
90 } | |
91 return(x*log(x/np)+np-x); | |
92 } | |
93 | |
94 /* | |
95 Raw binomial probability calculation. | |
96 (1) This has both p and q arguments, when one may be represented | |
97 more accurately than the other (in particular, in df()). | |
98 (2) This should NOT check that inputs x and n are integers. This | |
99 should be done in the calling function, where necessary. | |
100 (3) Does not check for 0<=p<=1 and 0<=q<=1 or NaN's. Do this in | |
101 the calling function. | |
102 */ | |
103 double dbinom_raw(x,n,p,q,give_log) | |
104 double x, n, p, q; | |
105 int give_log; | |
106 { double f, lc; | |
107 | |
108 if (p==0.0) return((x==0) ? D_1 : D_0); | |
109 if (q==0.0) return((x==n) ? D_1 : D_0); | |
110 | |
111 if (x==0) | |
112 { lc = (p<0.1) ? -bd0(n,n*q) - n*p : n*log(q); | |
113 return( DEXP(lc) ); | |
114 } | |
115 | |
116 if (x==n) | |
117 { lc = (q<0.1) ? -bd0(n,n*p) - n*q : n*log(p); | |
118 return( DEXP(lc) ); | |
119 } | |
120 | |
121 if ((x<0) | (x>n)) return( D_0 ); | |
122 | |
123 lc = stirlerr(n) - stirlerr(x) - stirlerr(n-x) | |
124 - bd0(x,n*p) - bd0(n-x,n*q); | |
125 f = (PIx2*x*(n-x))/n; | |
126 | |
127 return( FEXP(f,lc) ); | |
128 } | |
129 | |
130 double dbinom(x,n,p,give_log) | |
131 int x, n; | |
132 double p; | |
133 int give_log; | |
134 { | |
135 if ((p<0) | (p>1) | (n<0)) return(INVALID_PARAMS); | |
136 if (x<0) return( D_0 ); | |
137 | |
138 return( dbinom_raw((double)x,(double)n,p,1-p,give_log) ); | |
139 } | |
140 | |
141 /* | |
142 Poisson probability lb^x exp(-lb) / x!. | |
143 I don't check that x is an integer, since other functions | |
144 that call dpois_raw() (i.e. dgamma) may use a fractional | |
145 x argument. | |
146 */ | |
147 double dpois_raw(x,lambda,give_log) | |
148 int give_log; | |
149 double x, lambda; | |
150 { | |
151 if (lambda==0) return( (x==0) ? D_1 : D_0 ); | |
152 if (x==0) return( DEXP(-lambda) ); | |
153 if (x<0) return( D_0 ); | |
154 | |
155 return(FEXP( PIx2*x, -stirlerr(x)-bd0(x,lambda) )); | |
156 } | |
157 | |
158 double dpois(x,lambda,give_log) | |
159 int x, give_log; | |
160 double lambda; | |
161 { | |
162 if (lambda<0) return(INVALID_PARAMS); | |
163 if (x<0) return( D_0 ); | |
164 | |
165 return( dpois_raw((double)x,lambda,give_log) ); | |
166 } | |
167 | |
168 double dbeta(x,a,b,give_log) | |
169 double x, a, b; | |
170 int give_log; | |
171 { double f, p; | |
172 | |
173 if ((a<=0) | (b<=0)) return(INVALID_PARAMS); | |
174 if ((x<=0) | (x>=1)) return(D_0); | |
175 | |
176 if (a<1) | |
177 { if (b<1) /* a<1, b<1 */ | |
178 { f = a*b/((a+b)*x*(1-x)); | |
179 p = dbinom_raw(a,a+b,x,1-x,give_log); | |
180 } | |
181 else /* a<1, b>=1 */ | |
182 { f = a/x; | |
183 p = dbinom_raw(a,a+b-1,x,1-x,give_log); | |
184 } | |
185 } | |
186 else | |
187 { if (b<1) /* a>=1, b<1 */ | |
188 { f = b/(1-x); | |
189 p = dbinom_raw(a-1,a+b-1,x,1-x,give_log); | |
190 } | |
191 else /* a>=1, b>=1 */ | |
192 { f = a+b-1; | |
193 p = dbinom_raw(a-1,(a-1)+(b-1),x,1-x,give_log); | |
194 } | |
195 } | |
196 | |
197 return( (give_log) ? p + log(f) : p*f ); | |
198 } | |
199 | |
200 /* | |
201 * To evaluate the F density, write it as a Binomial probability | |
202 * with p = x*m/(n+x*m). For m>=2, use the simplest conversion. | |
203 * For m<2, (m-2)/2<0 so the conversion will not work, and we must use | |
204 * a second conversion. Note the division by p; this seems unavoidable | |
205 * for m < 2, since the F density has a singularity as x (or p) -> 0. | |
206 */ | |
207 double df(x,m,n,give_log) | |
208 double x, m, n; | |
209 int give_log; | |
210 { double p, q, f, dens; | |
211 | |
212 if ((m<=0) | (n<=0)) return(INVALID_PARAMS); | |
213 if (x <= 0.0) return(D_0); | |
214 | |
215 f = 1.0/(n+x*m); | |
216 q = n*f; | |
217 p = x*m*f; | |
218 | |
219 if (m>=2) | |
220 { f = m*q/2; | |
221 dens = dbinom_raw((m-2)/2.0, (m+n-2)/2.0, p, q, give_log); | |
222 } | |
223 else | |
224 { f = m*m*q / (2*p*(m+n)); | |
225 dens = dbinom_raw(m/2.0, (m+n)/2.0, p, q, give_log); | |
226 } | |
227 | |
228 return((give_log) ? log(f)+dens : f*dens); | |
229 } | |
230 | |
231 /* | |
232 * Gamma density, | |
233 * lb^r x^{r-1} exp(-lb*x) | |
234 * p(x;r,lb) = ----------------------- | |
235 * (r-1)! | |
236 * | |
237 * If USE_SCALE is defined below, the lb argument will be interpreted | |
238 * as a scale parameter (i.e. replace lb by 1/lb above). Otherwise, | |
239 * it is interpreted as a rate parameter, as above. | |
240 */ | |
241 | |
242 /* #define USE_SCALE */ | |
243 | |
244 double dgamma(x,r,lambda,give_log) | |
245 int give_log; | |
246 double x, r, lambda; | |
247 { double pr; | |
248 | |
249 if ((r<=0) | (lambda<0)) return(INVALID_PARAMS); | |
250 if (x<=0.0) return( D_0 ); | |
251 | |
252 #ifdef USE_SCALE | |
253 lambda = 1.0/lambda; | |
254 #endif | |
255 | |
256 if (r<1) | |
257 { pr = dpois_raw(r,lambda*x,give_log); | |
258 return( (give_log) ? pr + log(r/x) : pr*r/x ); | |
259 } | |
260 | |
261 pr = dpois_raw(r-1.0,lambda*x,give_log); | |
262 return( (give_log) ? pr + log(lambda) : lambda*pr); | |
263 } | |
264 | |
265 double dchisq(x, df, give_log) | |
266 double x, df; | |
267 int give_log; | |
268 { | |
269 return(dgamma(x, df/2.0, | |
270 0.5 | |
271 ,give_log)); | |
272 /* | |
273 #ifdef USE_SCALE | |
274 2.0 | |
275 #else | |
276 0.5 | |
277 #endif | |
278 ,give_log)); | |
279 */ | |
280 } | |
281 | |
282 /* | |
283 * Given a sequence of r successes and b failures, we sample n (\le b+r) | |
284 * items without replacement. The hypergeometric probability is the | |
285 * probability of x successes: | |
286 * | |
287 * dbinom(x,r,p) * dbinom(n-x,b,p) | |
288 * p(x;r,b,n) = --------------------------------- | |
289 * dbinom(n,r+b,p) | |
290 * | |
291 * for any p. For numerical stability, we take p=n/(r+b); with this choice, | |
292 * the denominator is not exponentially small. | |
293 */ | |
294 double dhyper(x,r,b,n,give_log) | |
295 int x, r, b, n, give_log; | |
296 { double p, q, p1, p2, p3; | |
297 | |
298 if ((r<0) | (b<0) | (n<0) | (n>r+b)) | |
299 return( INVALID_PARAMS ); | |
300 | |
301 if (x<0) return(D_0); | |
302 | |
303 if (n==0) return((x==0) ? D_1 : D_0); | |
304 | |
305 p = ((double)n)/((double)(r+b)); | |
306 q = ((double)(r+b-n))/((double)(r+b)); | |
307 | |
308 p1 = dbinom_raw((double)x,(double)r,p,q,give_log); | |
309 p2 = dbinom_raw((double)(n-x),(double)b,p,q,give_log); | |
310 p3 = dbinom_raw((double)n,(double)(r+b),p,q,give_log); | |
311 | |
312 return( (give_log) ? p1 + p2 - p3 : p1*p2/p3 ); | |
313 } | |
314 | |
315 /* | |
316 probability of x failures before the nth success. | |
317 */ | |
318 double dnbinom(x,n,p,give_log) | |
319 double n, p; | |
320 int x, give_log; | |
321 { double prob, f; | |
322 | |
323 if ((p<0) | (p>1) | (n<=0)) return(INVALID_PARAMS); | |
324 | |
325 if (x<0) return( D_0 ); | |
326 | |
327 prob = dbinom_raw(n,x+n,p,1-p,give_log); | |
328 f = n/(n+x); | |
329 | |
330 return((give_log) ? log(f) + prob : f*prob); | |
331 } | |
332 | |
333 double dt(x, df, give_log) | |
334 double x, df; | |
335 int give_log; | |
336 { double t, u, f; | |
337 | |
338 if (df<=0.0) return(INVALID_PARAMS); | |
339 | |
340 /* | |
341 exp(t) = Gamma((df+1)/2) /{ sqrt(df/2) * Gamma(df/2) } | |
342 = sqrt(df/2) / ((df+1)/2) * Gamma((df+3)/2) / Gamma((df+2)/2). | |
343 This form leads to a computation that should be stable for all | |
344 values of df, including df -> 0 and df -> infinity. | |
345 */ | |
346 t = -bd0(df/2.0,(df+1)/2.0) + stirlerr((df+1)/2.0) - stirlerr(df/2.0); | |
347 | |
348 if (x*x>df) | |
349 u = log( 1+ x*x/df ) * df/2; | |
350 else | |
351 u = -bd0(df/2.0,(df+x*x)/2.0) + x*x/2.0; | |
352 | |
353 f = PIx2*(1+x*x/df); | |
354 | |
355 return( FEXP(f,t-u) ); | |
356 } | |
357 /* | |
358 * Copyright 1996-2006 Catherine Loader. | |
359 */ | |
360 /* | |
361 * Provides mut_erf() and mut_erfc() functions. Also mut_pnorm(). | |
362 * Had too many problems with erf()'s built into math libraries | |
363 * (when they existed). Hence the need to write my own... | |
364 * | |
365 * Algorithm from Walter Kr\"{a}mer, Frithjof Blomquist. | |
366 * "Algorithms with Guaranteed Error Bounds for the Error Function | |
367 * and Complementary Error Function" | |
368 * Preprint 2000/2, Bergische Universt\"{a}t GH Wuppertal | |
369 * http://www.math.uni-wuppertal.de/wrswt/preprints/prep_00_2.pdf | |
370 * | |
371 * Coded by Catherine Loader, September 2006. | |
372 */ | |
373 | |
374 #include "mut.h" | |
375 | |
376 double erf1(double x) /* erf; 0 < x < 0.65) */ | |
377 { double p[5] = {1.12837916709551256e0, /* 2/sqrt(pi) */ | |
378 1.35894887627277916e-1, | |
379 4.03259488531795274e-2, | |
380 1.20339380863079457e-3, | |
381 6.49254556481904354e-5}; | |
382 double q[5] = {1.00000000000000000e0, | |
383 4.53767041780002545e-1, | |
384 8.69936222615385890e-2, | |
385 8.49717371168693357e-3, | |
386 3.64915280629351082e-4}; | |
387 double x2, p4, q4; | |
388 x2 = x*x; | |
389 p4 = p[0] + p[1]*x2 + p[2]*x2*x2 + p[3]*x2*x2*x2 + p[4]*x2*x2*x2*x2; | |
390 q4 = q[0] + q[1]*x2 + q[2]*x2*x2 + q[3]*x2*x2*x2 + q[4]*x2*x2*x2*x2; | |
391 return(x*p4/q4); | |
392 } | |
393 | |
394 double erf2(double x) /* erfc; 0.65 <= x < 2.2 */ | |
395 { double p[6] = {9.99999992049799098e-1, | |
396 1.33154163936765307e0, | |
397 8.78115804155881782e-1, | |
398 3.31899559578213215e-1, | |
399 7.14193832506776067e-2, | |
400 7.06940843763253131e-3}; | |
401 double q[7] = {1.00000000000000000e0, | |
402 2.45992070144245533e0, | |
403 2.65383972869775752e0, | |
404 1.61876655543871376e0, | |
405 5.94651311286481502e-1, | |
406 1.26579413030177940e-1, | |
407 1.25304936549413393e-2}; | |
408 double x2, p5, q6; | |
409 x2 = x*x; | |
410 p5 = p[0] + p[1]*x + p[2]*x2 + p[3]*x2*x + p[4]*x2*x2 + p[5]*x2*x2*x; | |
411 q6 = q[0] + q[1]*x + q[2]*x2 + q[3]*x2*x + q[4]*x2*x2 + q[5]*x2*x2*x + q[6]*x2*x2*x2; | |
412 return( exp(-x2)*p5/q6 ); | |
413 } | |
414 | |
415 double erf3(double x) /* erfc; 2.2 < x <= 6 */ | |
416 { double p[6] = {9.99921140009714409e-1, | |
417 1.62356584489366647e0, | |
418 1.26739901455873222e0, | |
419 5.81528574177741135e-1, | |
420 1.57289620742838702e-1, | |
421 2.25716982919217555e-2}; | |
422 double q[7] = {1.00000000000000000e0, | |
423 2.75143870676376208e0, | |
424 3.37367334657284535e0, | |
425 2.38574194785344389e0, | |
426 1.05074004614827206e0, | |
427 2.78788439273628983e-1, | |
428 4.00072964526861362e-2}; | |
429 double x2, p5, q6; | |
430 x2 = x*x; | |
431 p5 = p[0] + p[1]*x + p[2]*x2 + p[3]*x2*x + p[4]*x2*x2 + p[5]*x2*x2*x; | |
432 q6 = q[0] + q[1]*x + q[2]*x2 + q[3]*x2*x + q[4]*x2*x2 + q[5]*x2*x2*x + q[6]*x2*x2*x2; | |
433 return( exp(-x2)*p5/q6 ); | |
434 } | |
435 | |
436 double erf4(double x) /* erfc; x > 6.0 */ | |
437 { double p[5] = {5.64189583547756078e-1, | |
438 8.80253746105525775e0, | |
439 3.84683103716117320e1, | |
440 4.77209965874436377e1, | |
441 8.08040729052301677e0}; | |
442 double q[5] = {1.00000000000000000e0, | |
443 1.61020914205869003e1, | |
444 7.54843505665954743e1, | |
445 1.12123870801026015e2, | |
446 3.73997570145040850e1}; | |
447 double x2, p4, q4; | |
448 if (x>26.5432) return(0.0); | |
449 x2 = 1.0/(x*x); | |
450 p4 = p[0] + p[1]*x2 + p[2]*x2*x2 + p[3]*x2*x2*x2 + p[4]*x2*x2*x2*x2; | |
451 q4 = q[0] + q[1]*x2 + q[2]*x2*x2 + q[3]*x2*x2*x2 + q[4]*x2*x2*x2*x2; | |
452 return(exp(-x*x)*p4/(x*q4)); | |
453 } | |
454 | |
455 double mut_erfc(double x) | |
456 { if (x<0.0) return(2.0-mut_erfc(-x)); | |
457 if (x==0.0) return(1.0); | |
458 if (x<0.65) return(1.0-erf1(x)); | |
459 if (x<2.2) return(erf2(x)); | |
460 if (x<6.0) return(erf3(x)); | |
461 return(erf4(x)); | |
462 } | |
463 | |
464 double mut_erf(double x) | |
465 { | |
466 if (x<0.0) return(-mut_erf(-x)); | |
467 if (x==0.0) return(0.0); | |
468 if (x<0.65) return(erf1(x)); | |
469 if (x<2.2) return(1.0-erf2(x)); | |
470 if (x<6.0) return(1.0-erf3(x)); | |
471 return(1.0-erf4(x)); | |
472 } | |
473 | |
474 double mut_pnorm(double x) | |
475 { if (x<0.0) return(mut_erfc(-x/SQRT2)/2); | |
476 return((1.0+mut_erf(x/SQRT2))/2); | |
477 } | |
478 /* | |
479 * Copyright 1996-2006 Catherine Loader. | |
480 */ | |
481 #include "mut.h" | |
482 | |
483 static double lookup_gamma[21] = { | |
484 0.0, /* place filler */ | |
485 0.572364942924699971, /* log(G(0.5)) = log(sqrt(pi)) */ | |
486 0.000000000000000000, /* log(G(1)) = log(0!) */ | |
487 -0.120782237635245301, /* log(G(3/2)) = log(sqrt(pi)/2)) */ | |
488 0.000000000000000000, /* log(G(2)) = log(1!) */ | |
489 0.284682870472919181, /* log(G(5/2)) = log(3sqrt(pi)/4) */ | |
490 0.693147180559945286, /* log(G(3)) = log(2!) */ | |
491 1.200973602347074287, /* etc */ | |
492 1.791759469228054957, | |
493 2.453736570842442344, | |
494 3.178053830347945752, | |
495 3.957813967618716511, | |
496 4.787491742782045812, | |
497 5.662562059857141783, | |
498 6.579251212010101213, | |
499 7.534364236758732680, | |
500 8.525161361065414667, | |
501 9.549267257300996903, | |
502 10.604602902745250859, | |
503 11.689333420797268559, | |
504 12.801827480081469091 }; | |
505 | |
506 /* | |
507 * coefs are B(2n)/(2n(2n-1)) 2n(2n-1) = | |
508 * 2n B(2n) 2n(2n-1) coef | |
509 * 2 1/6 2 1/12 | |
510 * 4 -1/30 12 -1/360 | |
511 * 6 1/42 30 1/1260 | |
512 * 8 -1/30 56 -1/1680 | |
513 * 10 5/66 90 1/1188 | |
514 * 12 -691/2730 132 691/360360 | |
515 */ | |
516 | |
517 double mut_lgamma(double x) | |
518 { double f, z, x2, se; | |
519 int ix; | |
520 | |
521 /* lookup table for common values. | |
522 */ | |
523 ix = (int)(2*x); | |
524 if (((ix>=1) & (ix<=20)) && (ix==2*x)) return(lookup_gamma[ix]); | |
525 | |
526 f = 1.0; | |
527 while (x <= 15) | |
528 { f *= x; | |
529 x += 1.0; | |
530 } | |
531 | |
532 x2 = 1.0/(x*x); | |
533 z = (x-0.5)*log(x) - x + HF_LG_PIx2; | |
534 se = (13860 - x2*(462 - x2*(132 - x2*(99 - 140*x2))))/(166320*x); | |
535 | |
536 return(z + se - log(f)); | |
537 } | |
538 | |
539 double mut_lgammai(int i) /* log(Gamma(i/2)) for integer i */ | |
540 { if (i>20) return(mut_lgamma(i/2.0)); | |
541 return(lookup_gamma[i]); | |
542 } | |
543 /* | |
544 * Copyright 1996-2006 Catherine Loader. | |
545 */ | |
546 /* | |
547 * A is a n*p matrix, find the cholesky decomposition | |
548 * of the first p rows. In most applications, will want n=p. | |
549 * | |
550 * chol_dec(A,n,p) computes the decomoposition R'R=A. | |
551 * (note that R is stored in the input A). | |
552 * chol_solve(A,v,n,p) computes (R'R)^{-1}v | |
553 * chol_hsolve(A,v,n,p) computes (R')^{-1}v | |
554 * chol_isolve(A,v,n,p) computes (R)^{-1}v | |
555 * chol_qf(A,v,n,p) computes ||(R')^{-1}v||^2. | |
556 * chol_mult(A,v,n,p) computes (R'R)v | |
557 * | |
558 * The solve functions assume that A is already decomposed. | |
559 * chol_solve(A,v,n,p) is equivalent to applying chol_hsolve() | |
560 * and chol_isolve() in sequence. | |
561 */ | |
562 | |
563 #include <math.h> | |
564 #include "mut.h" | |
565 | |
566 void chol_dec(A,n,p) | |
567 double *A; | |
568 int n, p; | |
569 { int i, j, k; | |
570 | |
571 for (j=0; j<p; j++) | |
572 { k = n*j+j; | |
573 for (i=0; i<j; i++) A[k] -= A[n*j+i]*A[n*j+i]; | |
574 if (A[k]<=0) | |
575 { for (i=j; i<p; i++) A[n*i+j] = 0.0; } | |
576 else | |
577 { A[k] = sqrt(A[k]); | |
578 for (i=j+1; i<p; i++) | |
579 { for (k=0; k<j; k++) | |
580 A[n*i+j] -= A[n*i+k]*A[n*j+k]; | |
581 A[n*i+j] /= A[n*j+j]; | |
582 } | |
583 } | |
584 } | |
585 for (j=0; j<p; j++) | |
586 for (i=j+1; i<p; i++) A[n*j+i] = 0.0; | |
587 } | |
588 | |
589 int chol_solve(A,v,n,p) | |
590 double *A, *v; | |
591 int n, p; | |
592 { int i, j; | |
593 | |
594 for (i=0; i<p; i++) | |
595 { for (j=0; j<i; j++) v[i] -= A[i*n+j]*v[j]; | |
596 v[i] /= A[i*n+i]; | |
597 } | |
598 for (i=p-1; i>=0; i--) | |
599 { for (j=i+1; j<p; j++) v[i] -= A[j*n+i]*v[j]; | |
600 v[i] /= A[i*n+i]; | |
601 } | |
602 return(p); | |
603 } | |
604 | |
605 int chol_hsolve(A,v,n,p) | |
606 double *A, *v; | |
607 int n, p; | |
608 { int i, j; | |
609 | |
610 for (i=0; i<p; i++) | |
611 { for (j=0; j<i; j++) v[i] -= A[i*n+j]*v[j]; | |
612 v[i] /= A[i*n+i]; | |
613 } | |
614 return(p); | |
615 } | |
616 | |
617 int chol_isolve(A,v,n,p) | |
618 double *A, *v; | |
619 int n, p; | |
620 { int i, j; | |
621 | |
622 for (i=p-1; i>=0; i--) | |
623 { for (j=i+1; j<p; j++) v[i] -= A[j*n+i]*v[j]; | |
624 v[i] /= A[i*n+i]; | |
625 } | |
626 return(p); | |
627 } | |
628 | |
629 double chol_qf(A,v,n,p) | |
630 double *A, *v; | |
631 int n, p; | |
632 { int i, j; | |
633 double sum; | |
634 | |
635 sum = 0.0; | |
636 for (i=0; i<p; i++) | |
637 { for (j=0; j<i; j++) v[i] -= A[i*n+j]*v[j]; | |
638 v[i] /= A[i*n+i]; | |
639 sum += v[i]*v[i]; | |
640 } | |
641 return(sum); | |
642 } | |
643 | |
644 int chol_mult(A,v,n,p) | |
645 double *A, *v; | |
646 int n, p; | |
647 { int i, j; | |
648 double sum; | |
649 for (i=0; i<p; i++) | |
650 { sum = 0; | |
651 for (j=i; j<p; j++) sum += A[j*n+i]*v[j]; | |
652 v[i] = sum; | |
653 } | |
654 for (i=p-1; i>=0; i--) | |
655 { sum = 0; | |
656 for (j=0; j<=i; j++) sum += A[i*n+j]*v[j]; | |
657 v[i] = sum; | |
658 } | |
659 return(1); | |
660 } | |
661 /* | |
662 * Copyright 1996-2006 Catherine Loader. | |
663 */ | |
664 #include <stdio.h> | |
665 #include <math.h> | |
666 #include "mut.h" | |
667 #define E_MAXIT 20 | |
668 #define E_TOL 1.0e-10 | |
669 #define SQR(x) ((x)*(x)) | |
670 | |
671 double e_tol(D,p) | |
672 double *D; | |
673 int p; | |
674 { double mx; | |
675 int i; | |
676 if (E_TOL <= 0.0) return(0.0); | |
677 mx = D[0]; | |
678 for (i=1; i<p; i++) if (D[i*(p+1)]>mx) mx = D[i*(p+1)]; | |
679 return(E_TOL*mx); | |
680 } | |
681 | |
682 void eig_dec(X,P,d) | |
683 double *X, *P; | |
684 int d; | |
685 { int i, j, k, iter, ms; | |
686 double c, s, r, u, v; | |
687 | |
688 for (i=0; i<d; i++) | |
689 for (j=0; j<d; j++) P[i*d+j] = (i==j); | |
690 | |
691 for (iter=0; iter<E_MAXIT; iter++) | |
692 { ms = 0; | |
693 for (i=0; i<d; i++) | |
694 for (j=i+1; j<d; j++) | |
695 if (SQR(X[i*d+j]) > 1.0e-15*fabs(X[i*d+i]*X[j*d+j])) | |
696 { c = (X[j*d+j]-X[i*d+i])/2; | |
697 s = -X[i*d+j]; | |
698 r = sqrt(c*c+s*s); | |
699 c /= r; | |
700 s = sqrt((1-c)/2)*(2*(s>0)-1); | |
701 c = sqrt((1+c)/2); | |
702 for (k=0; k<d; k++) | |
703 { u = X[i*d+k]; v = X[j*d+k]; | |
704 X[i*d+k] = u*c+v*s; | |
705 X[j*d+k] = v*c-u*s; | |
706 } | |
707 for (k=0; k<d; k++) | |
708 { u = X[k*d+i]; v = X[k*d+j]; | |
709 X[k*d+i] = u*c+v*s; | |
710 X[k*d+j] = v*c-u*s; | |
711 } | |
712 X[i*d+j] = X[j*d+i] = 0.0; | |
713 for (k=0; k<d; k++) | |
714 { u = P[k*d+i]; v = P[k*d+j]; | |
715 P[k*d+i] = u*c+v*s; | |
716 P[k*d+j] = v*c-u*s; | |
717 } | |
718 ms = 1; | |
719 } | |
720 if (ms==0) return; | |
721 } | |
722 mut_printf("eig_dec not converged\n"); | |
723 } | |
724 | |
725 int eig_solve(J,x) | |
726 jacobian *J; | |
727 double *x; | |
728 { int d, i, j, rank; | |
729 double *D, *P, *Q, *w; | |
730 double tol; | |
731 | |
732 D = J->Z; | |
733 P = Q = J->Q; | |
734 d = J->p; | |
735 w = J->wk; | |
736 | |
737 tol = e_tol(D,d); | |
738 | |
739 rank = 0; | |
740 for (i=0; i<d; i++) | |
741 { w[i] = 0.0; | |
742 for (j=0; j<d; j++) w[i] += P[j*d+i]*x[j]; | |
743 } | |
744 for (i=0; i<d; i++) | |
745 if (D[i*d+i]>tol) | |
746 { w[i] /= D[i*(d+1)]; | |
747 rank++; | |
748 } | |
749 for (i=0; i<d; i++) | |
750 { x[i] = 0.0; | |
751 for (j=0; j<d; j++) x[i] += Q[i*d+j]*w[j]; | |
752 } | |
753 return(rank); | |
754 } | |
755 | |
756 int eig_hsolve(J,v) | |
757 jacobian *J; | |
758 double *v; | |
759 { int i, j, p, rank; | |
760 double *D, *Q, *w; | |
761 double tol; | |
762 | |
763 D = J->Z; | |
764 Q = J->Q; | |
765 p = J->p; | |
766 w = J->wk; | |
767 | |
768 tol = e_tol(D,p); | |
769 rank = 0; | |
770 | |
771 for (i=0; i<p; i++) | |
772 { w[i] = 0.0; | |
773 for (j=0; j<p; j++) w[i] += Q[j*p+i]*v[j]; | |
774 } | |
775 for (i=0; i<p; i++) | |
776 { if (D[i*p+i]>tol) | |
777 { v[i] = w[i]/sqrt(D[i*(p+1)]); | |
778 rank++; | |
779 } | |
780 else v[i] = 0.0; | |
781 } | |
782 return(rank); | |
783 } | |
784 | |
785 int eig_isolve(J,v) | |
786 jacobian *J; | |
787 double *v; | |
788 { int i, j, p, rank; | |
789 double *D, *Q, *w; | |
790 double tol; | |
791 | |
792 D = J->Z; | |
793 Q = J->Q; | |
794 p = J->p; | |
795 w = J->wk; | |
796 | |
797 tol = e_tol(D,p); | |
798 rank = 0; | |
799 | |
800 for (i=0; i<p; i++) | |
801 { if (D[i*p+i]>tol) | |
802 { v[i] = w[i]/sqrt(D[i*(p+1)]); | |
803 rank++; | |
804 } | |
805 else v[i] = 0.0; | |
806 } | |
807 | |
808 for (i=0; i<p; i++) | |
809 { w[i] = 0.0; | |
810 for (j=0; j<p; j++) w[i] += Q[i*p+j]*v[j]; | |
811 } | |
812 | |
813 return(rank); | |
814 } | |
815 | |
816 double eig_qf(J,v) | |
817 jacobian *J; | |
818 double *v; | |
819 { int i, j, p; | |
820 double sum, tol; | |
821 | |
822 p = J->p; | |
823 sum = 0.0; | |
824 tol = e_tol(J->Z,p); | |
825 | |
826 for (i=0; i<p; i++) | |
827 if (J->Z[i*p+i]>tol) | |
828 { J->wk[i] = 0.0; | |
829 for (j=0; j<p; j++) J->wk[i] += J->Q[j*p+i]*v[j]; | |
830 sum += J->wk[i]*J->wk[i]/J->Z[i*p+i]; | |
831 } | |
832 return(sum); | |
833 } | |
834 /* | |
835 * Copyright 1996-2006 Catherine Loader. | |
836 */ | |
837 /* | |
838 * Integrate a function f over a circle or disc. | |
839 */ | |
840 | |
841 #include "mut.h" | |
842 | |
843 void setM(M,r,s,c,b) | |
844 double *M, r, s, c; | |
845 int b; | |
846 { M[0] =-r*s; M[1] = r*c; | |
847 M[2] = b*c; M[3] = b*s; | |
848 M[4] =-r*c; M[5] = -s; | |
849 M[6] = -s; M[7] = 0.0; | |
850 M[8] =-r*s; M[9] = c; | |
851 M[10]= c; M[11]= 0.0; | |
852 } | |
853 | |
854 void integ_circ(f,r,orig,res,mint,b) | |
855 int (*f)(), mint, b; | |
856 double r, *orig, *res; | |
857 { double y, x[2], theta, tres[MXRESULT], M[12], c, s; | |
858 int i, j, nr; | |
859 | |
860 y = 0; | |
861 for (i=0; i<mint; i++) | |
862 { theta = 2*PI*(double)i/(double)mint; | |
863 c = cos(theta); s = sin(theta); | |
864 x[0] = orig[0]+r*c; | |
865 x[1] = orig[1]+r*s; | |
866 | |
867 if (b!=0) | |
868 { M[0] =-r*s; M[1] = r*c; | |
869 M[2] = b*c; M[3] = b*s; | |
870 M[4] =-r*c; M[5] = -s; | |
871 M[6] = -s; M[7] = 0.0; | |
872 M[8] =-r*s; M[9] = c; | |
873 M[10]= c; M[11]= 0.0; | |
874 } | |
875 | |
876 nr = f(x,2,tres,M); | |
877 if (i==0) setzero(res,nr); | |
878 for (j=0; j<nr; j++) res[j] += tres[j]; | |
879 } | |
880 y = 2 * PI * ((b==0)?r:1.0) / mint; | |
881 for (j=0; j<nr; j++) res[j] *= y; | |
882 } | |
883 | |
884 void integ_disc(f,fb,fl,res,resb,mg) | |
885 int (*f)(), (*fb)(), *mg; | |
886 double *fl, *res, *resb; | |
887 { double x[2], y, r, tres[MXRESULT], *orig, rmin, rmax, theta, c, s, M[12]; | |
888 int ct, ctb, i, j, k, nr, nrb, w; | |
889 | |
890 orig = &fl[2]; | |
891 rmax = fl[1]; | |
892 rmin = fl[0]; | |
893 y = 0.0; | |
894 ct = ctb = 0; | |
895 | |
896 for (j=0; j<mg[1]; j++) | |
897 { theta = 2*PI*(double)j/(double)mg[1]; | |
898 c = cos(theta); s = sin(theta); | |
899 for (i= (rmin>0) ? 0 : 1; i<=mg[0]; i++) | |
900 { r = rmin + (rmax-rmin)*i/mg[0]; | |
901 w = (2+2*(i&1)-(i==0)-(i==mg[0])); | |
902 x[0] = orig[0] + r*c; | |
903 x[1] = orig[1] + r*s; | |
904 nr = f(x,2,tres,NULL); | |
905 if (ct==0) setzero(res,nr); | |
906 for (k=0; k<nr; k++) res[k] += w*r*tres[k]; | |
907 ct++; | |
908 if (((i==0) | (i==mg[0])) && (fb!=NULL)) | |
909 { setM(M,r,s,c,1-2*(i==0)); | |
910 nrb = fb(x,2,tres,M); | |
911 if (ctb==0) setzero(resb,nrb); | |
912 ctb++; | |
913 for (k=0; k<nrb; k++) resb[k] += tres[k]; | |
914 } | |
915 } | |
916 } | |
917 | |
918 | |
919 /* for (i= (rmin>0) ? 0 : 1; i<=mg[0]; i++) | |
920 { | |
921 r = rmin + (rmax-rmin)*i/mg[0]; | |
922 w = (2+2*(i&1)-(i==0)-(i==mg[0])); | |
923 | |
924 for (j=0; j<mg[1]; j++) | |
925 { theta = 2*PI*(double)j/(double)mg[1]; | |
926 c = cos(theta); s = sin(theta); | |
927 x[0] = orig[0] + r*c; | |
928 x[1] = orig[1] + r*s; | |
929 nr = f(x,2,tres,NULL); | |
930 if (ct==0) setzero(res,nr); | |
931 ct++; | |
932 for (k=0; k<nr; k++) res[k] += w*r*tres[k]; | |
933 | |
934 if (((i==0) | (i==mg[0])) && (fb!=NULL)) | |
935 { setM(M,r,s,c,1-2*(i==0)); | |
936 nrb = fb(x,2,tres,M); | |
937 if (ctb==0) setzero(resb,nrb); | |
938 ctb++; | |
939 for (k=0; k<nrb; k++) resb[k] += tres[k]; | |
940 } | |
941 } | |
942 } */ | |
943 for (j=0; j<nr; j++) res[j] *= 2*PI*(rmax-rmin)/(3*mg[0]*mg[1]); | |
944 for (j=0; j<nrb; j++) resb[j] *= 2*PI/mg[1]; | |
945 } | |
946 /* | |
947 * Copyright 1996-2006 Catherine Loader. | |
948 */ | |
949 /* | |
950 * Multivariate integration of a vector-valued function | |
951 * using Monte-Carlo method. | |
952 * | |
953 * uses drand48() random number generator. Does not seed. | |
954 */ | |
955 | |
956 #include <stdlib.h> | |
957 #include "mut.h" | |
958 extern void setzero(); | |
959 | |
960 static double M[(1+MXIDIM)*MXIDIM*MXIDIM]; | |
961 | |
962 void monte(f,ll,ur,d,res,n) | |
963 int (*f)(), d, n; | |
964 double *ll, *ur, *res; | |
965 { | |
966 int i, j, nr; | |
967 #ifdef WINDOWS | |
968 mut_printf("Sorry, Monte-Carlo Integration not enabled.\n"); | |
969 for (i=0; i<n; i++) res[i] = 0.0; | |
970 #else | |
971 double z, x[MXIDIM], tres[MXRESULT]; | |
972 | |
973 srand48(234L); | |
974 | |
975 for (i=0; i<n; i++) | |
976 { for (j=0; j<d; j++) x[j] = ll[j] + (ur[j]-ll[j])*drand48(); | |
977 nr = f(x,d,tres,NULL); | |
978 if (i==0) setzero(res,nr); | |
979 for (j=0; j<nr; j++) res[j] += tres[j]; | |
980 } | |
981 | |
982 z = 1; | |
983 for (i=0; i<d; i++) z *= (ur[i]-ll[i]); | |
984 for (i=0; i<nr; i++) res[i] *= z/n; | |
985 #endif | |
986 } | |
987 /* | |
988 * Copyright 1996-2006 Catherine Loader. | |
989 */ | |
990 /* | |
991 * Multivariate integration of a vector-valued function | |
992 * using Simpson's rule. | |
993 */ | |
994 | |
995 #include <math.h> | |
996 #include "mut.h" | |
997 extern void setzero(); | |
998 | |
999 static double M[(1+MXIDIM)*MXIDIM*MXIDIM]; | |
1000 | |
1001 /* third order corners */ | |
1002 void simp3(fd,x,d,resd,delta,wt,i0,i1,mg,ct,res2,index) | |
1003 int (*fd)(), d, wt, i0, i1, *mg, ct, *index; | |
1004 double *x, *resd, *delta, *res2; | |
1005 { int k, l, m, nrd; | |
1006 double zb; | |
1007 | |
1008 for (k=i1+1; k<d; k++) if ((index[k]==0) | (index[k]==mg[k])) | |
1009 { | |
1010 setzero(M,d*d); | |
1011 m = 0; zb = 1.0; | |
1012 for (l=0; l<d; l++) | |
1013 if ((l!=i0) & (l!=i1) & (l!=k)) | |
1014 { M[m*d+l] = 1.0; | |
1015 m++; | |
1016 zb *= delta[l]; | |
1017 } | |
1018 M[(d-3)*d+i0] = (index[i0]==0) ? -1 : 1; | |
1019 M[(d-2)*d+i1] = (index[i1]==0) ? -1 : 1; | |
1020 M[(d-1)*d+k] = (index[k]==0) ? -1 : 1; | |
1021 nrd = fd(x,d,res2,M); | |
1022 if ((ct==0) & (i0==0) & (i1==1) & (k==2)) setzero(resd,nrd); | |
1023 for (l=0; l<nrd; l++) | |
1024 resd[l] += wt*zb*res2[l]; | |
1025 } | |
1026 } | |
1027 | |
1028 /* second order corners */ | |
1029 void simp2(fc,fd,x,d,resc,resd,delta,wt,i0,mg,ct,res2,index) | |
1030 int (*fc)(), (*fd)(), d, wt, i0, *mg, ct, *index; | |
1031 double *x, *resc, *resd, *delta, *res2; | |
1032 { int j, k, l, nrc; | |
1033 double zb; | |
1034 for (j=i0+1; j<d; j++) if ((index[j]==0) | (index[j]==mg[j])) | |
1035 { setzero(M,d*d); | |
1036 l = 0; zb = 1; | |
1037 for (k=0; k<d; k++) if ((k!=i0) & (k!=j)) | |
1038 { M[l*d+k] = 1.0; | |
1039 l++; | |
1040 zb *= delta[k]; | |
1041 } | |
1042 M[(d-2)*d+i0] = (index[i0]==0) ? -1 : 1; | |
1043 M[(d-1)*d+j] = (index[j]==0) ? -1 : 1; | |
1044 nrc = fc(x,d,res2,M); | |
1045 if ((ct==0) & (i0==0) & (j==1)) setzero(resc,nrc); | |
1046 for (k=0; k<nrc; k++) resc[k] += wt*zb*res2[k]; | |
1047 | |
1048 if (fd!=NULL) | |
1049 simp3(fd,x,d,resd,delta,wt,i0,j,mg,ct,res2,index); | |
1050 } | |
1051 } | |
1052 | |
1053 /* first order boundary */ | |
1054 void simp1(fb,fc,fd,x,d,resb,resc,resd,delta,wt,mg,ct,res2,index) | |
1055 int (*fb)(), (*fc)(), (*fd)(), d, wt, *mg, ct, *index; | |
1056 double *x, *resb, *resc, *resd, *delta, *res2; | |
1057 { int i, j, k, nrb; | |
1058 double zb; | |
1059 for (i=0; i<d; i++) if ((index[i]==0) | (index[i]==mg[i])) | |
1060 { setzero(M,(1+d)*d*d); | |
1061 k = 0; | |
1062 for (j=0; j<d; j++) if (j!=i) | |
1063 { M[k*d+j] = 1; | |
1064 k++; | |
1065 } | |
1066 M[(d-1)*d+i] = (index[i]==0) ? -1 : 1; | |
1067 nrb = fb(x,d,res2,M); | |
1068 zb = 1; | |
1069 for (j=0; j<d; j++) if (i!=j) zb *= delta[j]; | |
1070 if ((ct==0) && (i==0)) | |
1071 for (j=0; j<nrb; j++) resb[j] = 0.0; | |
1072 for (j=0; j<nrb; j++) resb[j] += wt*zb*res2[j]; | |
1073 | |
1074 if (fc!=NULL) | |
1075 simp2(fc,fd,x,d,resc,resd,delta,wt,i,mg,ct,res2,index); | |
1076 } | |
1077 } | |
1078 | |
1079 void simpson4(f,fb,fc,fd,ll,ur,d,res,resb,resc,resd,mg,res2) | |
1080 int (*f)(), (*fb)(), (*fc)(), (*fd)(), d, *mg; | |
1081 double *ll, *ur, *res, *resb, *resc, *resd, *res2; | |
1082 { int ct, i, j, nr, wt, index[MXIDIM]; | |
1083 double x[MXIDIM], delta[MXIDIM], z; | |
1084 | |
1085 for (i=0; i<d; i++) | |
1086 { index[i] = 0; | |
1087 x[i] = ll[i]; | |
1088 if (mg[i]&1) mg[i]++; | |
1089 delta[i] = (ur[i]-ll[i])/(3*mg[i]); | |
1090 } | |
1091 ct = 0; | |
1092 | |
1093 while(1) | |
1094 { wt = 1; | |
1095 for (i=0; i<d; i++) | |
1096 wt *= (4-2*(index[i]%2==0)-(index[i]==0)-(index[i]==mg[i])); | |
1097 nr = f(x,d,res2,NULL); | |
1098 if (ct==0) setzero(res,nr); | |
1099 for (i=0; i<nr; i++) res[i] += wt*res2[i]; | |
1100 | |
1101 if (fb!=NULL) | |
1102 simp1(fb,fc,fd,x,d,resb,resc,resd,delta,wt,mg,ct,res2,index); | |
1103 | |
1104 /* compute next grid point */ | |
1105 for (i=0; i<d; i++) | |
1106 { index[i]++; | |
1107 if (index[i]>mg[i]) | |
1108 { index[i] = 0; | |
1109 x[i] = ll[i]; | |
1110 if (i==d-1) /* done */ | |
1111 { z = 1.0; | |
1112 for (j=0; j<d; j++) z *= delta[j]; | |
1113 for (j=0; j<nr; j++) res[j] *= z; | |
1114 return; | |
1115 } | |
1116 } | |
1117 else | |
1118 { x[i] = ll[i] + 3*delta[i]*index[i]; | |
1119 i = d; | |
1120 } | |
1121 } | |
1122 ct++; | |
1123 } | |
1124 } | |
1125 | |
1126 void simpsonm(f,ll,ur,d,res,mg,res2) | |
1127 int (*f)(), d, *mg; | |
1128 double *ll, *ur, *res, *res2; | |
1129 { simpson4(f,NULL,NULL,NULL,ll,ur,d,res,NULL,NULL,NULL,mg,res2); | |
1130 } | |
1131 | |
1132 double simpson(f,l0,l1,m) | |
1133 double (*f)(), l0, l1; | |
1134 int m; | |
1135 { double x, sum; | |
1136 int i; | |
1137 sum = 0; | |
1138 for (i=0; i<=m; i++) | |
1139 { x = ((m-i)*l0 + i*l1)/m; | |
1140 sum += (2+2*(i&1)-(i==0)-(i==m)) * f(x); | |
1141 } | |
1142 return( (l1-l0) * sum / (3*m) ); | |
1143 } | |
1144 /* | |
1145 * Copyright 1996-2006 Catherine Loader. | |
1146 */ | |
1147 #include "mut.h" | |
1148 | |
1149 static double *res, *resb, *orig, rmin, rmax; | |
1150 static int ct0; | |
1151 | |
1152 void sphM(M,r,u) | |
1153 double *M, r, *u; | |
1154 { double h, u1[3], u2[3]; | |
1155 | |
1156 /* set the orthogonal unit vectors. */ | |
1157 h = sqrt(u[0]*u[0]+u[1]*u[1]); | |
1158 if (h<=0) | |
1159 { u1[0] = u2[1] = 1.0; | |
1160 u1[1] = u1[2] = u2[0] = u2[2] = 0.0; | |
1161 } | |
1162 else | |
1163 { u1[0] = u[1]/h; u1[1] = -u[0]/h; u1[2] = 0.0; | |
1164 u2[0] = u[2]*u[0]/h; u2[1] = u[2]*u[1]/h; u2[2] = -h; | |
1165 } | |
1166 | |
1167 /* parameterize the sphere as r(cos(t)cos(v)u + sin(t)u1 + cos(t)sin(v)u2). | |
1168 * first layer of M is (dx/dt, dx/dv, dx/dr) at t=v=0. | |
1169 */ | |
1170 M[0] = r*u1[0]; M[1] = r*u1[1]; M[2] = r*u1[2]; | |
1171 M[3] = r*u2[0]; M[4] = r*u2[1]; M[5] = r*u2[2]; | |
1172 M[6] = u[0]; M[7] = u[1]; M[8] = u[2]; | |
1173 | |
1174 /* next layers are second derivative matrix of components of x(r,t,v). | |
1175 * d^2x/dt^2 = d^2x/dv^2 = -ru; d^2x/dtdv = 0; | |
1176 * d^2x/drdt = u1; d^2x/drdv = u2; d^2x/dr^2 = 0. | |
1177 */ | |
1178 | |
1179 M[9] = M[13] = -r*u[0]; | |
1180 M[11]= M[15] = u1[0]; | |
1181 M[14]= M[16] = u2[0]; | |
1182 M[10]= M[12] = M[17] = 0.0; | |
1183 | |
1184 M[18]= M[22] = -r*u[1]; | |
1185 M[20]= M[24] = u1[1]; | |
1186 M[23]= M[25] = u2[1]; | |
1187 M[19]= M[21] = M[26] = 0.0; | |
1188 | |
1189 M[27]= M[31] = -r*u[1]; | |
1190 M[29]= M[33] = u1[1]; | |
1191 M[32]= M[34] = u2[1]; | |
1192 M[28]= M[30] = M[35] = 0.0; | |
1193 | |
1194 } | |
1195 | |
1196 double ip3(a,b) | |
1197 double *a, *b; | |
1198 { return(a[0]*b[0] + a[1]*b[1] + a[2]*b[2]); | |
1199 } | |
1200 | |
1201 void rn3(a) | |
1202 double *a; | |
1203 { double s; | |
1204 s = sqrt(ip3(a,a)); | |
1205 a[0] /= s; a[1] /= s; a[2] /= s; | |
1206 } | |
1207 | |
1208 double sptarea(a,b,c) | |
1209 double *a, *b, *c; | |
1210 { double ea, eb, ec, yab, yac, ybc, sab, sac, sbc; | |
1211 double ab[3], ac[3], bc[3], x1[3], x2[3]; | |
1212 | |
1213 ab[0] = a[0]-b[0]; ab[1] = a[1]-b[1]; ab[2] = a[2]-b[2]; | |
1214 ac[0] = a[0]-c[0]; ac[1] = a[1]-c[1]; ac[2] = a[2]-c[2]; | |
1215 bc[0] = b[0]-c[0]; bc[1] = b[1]-c[1]; bc[2] = b[2]-c[2]; | |
1216 | |
1217 yab = ip3(ab,a); yac = ip3(ac,a); ybc = ip3(bc,b); | |
1218 | |
1219 x1[0] = ab[0] - yab*a[0]; x2[0] = ac[0] - yac*a[0]; | |
1220 x1[1] = ab[1] - yab*a[1]; x2[1] = ac[1] - yac*a[1]; | |
1221 x1[2] = ab[2] - yab*a[2]; x2[2] = ac[2] - yac*a[2]; | |
1222 sab = ip3(x1,x1); sac = ip3(x2,x2); | |
1223 ea = acos(ip3(x1,x2)/sqrt(sab*sac)); | |
1224 | |
1225 x1[0] = ab[0] + yab*b[0]; x2[0] = bc[0] - ybc*b[0]; | |
1226 x1[1] = ab[1] + yab*b[1]; x2[1] = bc[1] - ybc*b[1]; | |
1227 x1[2] = ab[2] + yab*b[2]; x2[2] = bc[2] - ybc*b[2]; | |
1228 sbc = ip3(x2,x2); | |
1229 eb = acos(ip3(x1,x2)/sqrt(sab*sbc)); | |
1230 | |
1231 x1[0] = ac[0] + yac*c[0]; x2[0] = bc[0] + ybc*c[0]; | |
1232 x1[1] = ac[1] + yac*c[1]; x2[1] = bc[1] + ybc*c[1]; | |
1233 x1[2] = ac[2] + yac*c[2]; x2[2] = bc[2] + ybc*c[2]; | |
1234 ec = acos(ip3(x1,x2)/sqrt(sac*sbc)); | |
1235 | |
1236 /* | |
1237 * Euler's formula is a+b+c-PI, except I've cheated... | |
1238 * a=ea, c=ec, b=PI-eb, which is more stable. | |
1239 */ | |
1240 return(ea+ec-eb); | |
1241 } | |
1242 | |
1243 void li(x,f,fb,mint,ar) | |
1244 double *x, ar; | |
1245 int (*f)(), (*fb)(), mint; | |
1246 { int i, j, nr, nrb, ct1, w; | |
1247 double u[3], r, M[36]; | |
1248 double sres[MXRESULT], tres[MXRESULT]; | |
1249 | |
1250 /* divide mint by 2, and force to even (Simpson's rule...) | |
1251 * to make comparable with rectangular interpretation of mint | |
1252 */ | |
1253 mint <<= 1; | |
1254 if (mint&1) mint++; | |
1255 | |
1256 ct1 = 0; | |
1257 for (i= (rmin==0) ? 1 : 0; i<=mint; i++) | |
1258 { | |
1259 r = rmin + (rmax-rmin)*i/mint; | |
1260 w = 2+2*(i&1)-(i==0)-(i==mint); | |
1261 u[0] = orig[0]+x[0]*r; | |
1262 u[1] = orig[1]+x[1]*r; | |
1263 u[2] = orig[2]+x[2]*r; | |
1264 nr = f(u,3,tres,NULL); | |
1265 if (ct1==0) setzero(sres,nr); | |
1266 for (j=0; j<nr; j++) | |
1267 sres[j] += w*r*r*tres[j]; | |
1268 ct1++; | |
1269 | |
1270 if ((fb!=NULL) && (i==mint)) /* boundary */ | |
1271 { sphM(M,rmax,x); | |
1272 nrb = fb(u,3,tres,M); | |
1273 if (ct0==0) for (j=0; j<nrb; j++) resb[j] = 0.0; | |
1274 for (j=0; j<nrb; j++) | |
1275 resb[j] += tres[j]*ar; | |
1276 } | |
1277 } | |
1278 | |
1279 if (ct0==0) for (j=0; j<nr; j++) res[j] = 0.0; | |
1280 ct0++; | |
1281 | |
1282 for (j=0; j<nr; j++) | |
1283 res[j] += sres[j] * ar * (rmax-rmin)/(3*mint); | |
1284 } | |
1285 | |
1286 void sphint(f,fb,a,b,c,lev,mint,cent) | |
1287 double *a, *b, *c; | |
1288 int (*f)(), (*fb)(), lev, mint, cent; | |
1289 { double x[3], ab[3], ac[3], bc[3], ar; | |
1290 int i; | |
1291 | |
1292 if (lev>1) | |
1293 { ab[0] = a[0]+b[0]; ab[1] = a[1]+b[1]; ab[2] = a[2]+b[2]; rn3(ab); | |
1294 ac[0] = a[0]+c[0]; ac[1] = a[1]+c[1]; ac[2] = a[2]+c[2]; rn3(ac); | |
1295 bc[0] = b[0]+c[0]; bc[1] = b[1]+c[1]; bc[2] = b[2]+c[2]; rn3(bc); | |
1296 lev >>= 1; | |
1297 if (cent==0) | |
1298 { sphint(f,fb,a,ab,ac,lev,mint,1); | |
1299 sphint(f,fb,ab,bc,ac,lev,mint,0); | |
1300 } | |
1301 else | |
1302 { sphint(f,fb,a,ab,ac,lev,mint,1); | |
1303 sphint(f,fb,b,ab,bc,lev,mint,1); | |
1304 sphint(f,fb,c,ac,bc,lev,mint,1); | |
1305 sphint(f,fb,ab,bc,ac,lev,mint,1); | |
1306 } | |
1307 return; | |
1308 } | |
1309 | |
1310 x[0] = a[0]+b[0]+c[0]; | |
1311 x[1] = a[1]+b[1]+c[1]; | |
1312 x[2] = a[2]+b[2]+c[2]; | |
1313 rn3(x); | |
1314 ar = sptarea(a,b,c); | |
1315 | |
1316 for (i=0; i<8; i++) | |
1317 { if (i>0) | |
1318 { x[0] = -x[0]; | |
1319 if (i%2 == 0) x[1] = -x[1]; | |
1320 if (i==4) x[2] = -x[2]; | |
1321 } | |
1322 switch(cent) | |
1323 { case 2: /* the reflection and its 120', 240' rotations */ | |
1324 ab[0] = x[0]; ab[1] = x[2]; ab[2] = x[1]; li(ab,f,fb,mint,ar); | |
1325 ab[0] = x[2]; ab[1] = x[1]; ab[2] = x[0]; li(ab,f,fb,mint,ar); | |
1326 ab[0] = x[1]; ab[1] = x[0]; ab[2] = x[2]; li(ab,f,fb,mint,ar); | |
1327 case 1: /* and the 120' and 240' rotations */ | |
1328 ab[0] = x[1]; ab[1] = x[2]; ab[2] = x[0]; li(ab,f,fb,mint,ar); | |
1329 ac[0] = x[2]; ac[1] = x[0]; ac[2] = x[1]; li(ac,f,fb,mint,ar); | |
1330 case 0: /* and the triangle itself. */ | |
1331 li( x,f,fb,mint,ar); | |
1332 } | |
1333 } | |
1334 } | |
1335 | |
1336 void integ_sphere(f,fb,fl,Res,Resb,mg) | |
1337 double *fl, *Res, *Resb; | |
1338 int (*f)(), (*fb)(), *mg; | |
1339 { double a[3], b[3], c[3]; | |
1340 | |
1341 a[0] = 1; a[1] = a[2] = 0; | |
1342 b[1] = 1; b[0] = b[2] = 0; | |
1343 c[2] = 1; c[0] = c[1] = 0; | |
1344 | |
1345 res = Res; | |
1346 resb=Resb; | |
1347 orig = &fl[2]; | |
1348 rmin = fl[0]; | |
1349 rmax = fl[1]; | |
1350 | |
1351 ct0 = 0; | |
1352 sphint(f,fb,a,b,c,mg[1],mg[0],0); | |
1353 } | |
1354 /* | |
1355 * Copyright 1996-2006 Catherine Loader. | |
1356 */ | |
1357 /* | |
1358 * solving symmetric equations using the jacobian structure. Currently, three | |
1359 * methods can be used: cholesky decomposition, eigenvalues, eigenvalues on | |
1360 * the correlation matrix. | |
1361 * | |
1362 * jacob_dec(J,meth) decompose the matrix, meth=JAC_CHOL, JAC_EIG, JAC_EIGD | |
1363 * jacob_solve(J,v) J^{-1}v | |
1364 * jacob_hsolve(J,v) (R')^{-1/2}v | |
1365 * jacob_isolve(J,v) (R)^{-1/2}v | |
1366 * jacob_qf(J,v) v' J^{-1} v | |
1367 * jacob_mult(J,v) (R'R) v (pres. CHOL only) | |
1368 * where for each decomposition, R'R=J, although the different decomp's will | |
1369 * produce different R's. | |
1370 * | |
1371 * To set up the J matrix: | |
1372 * first, allocate storage: jac_alloc(J,p,wk) | |
1373 * where p=dimension of matrix, wk is a numeric vector of length | |
1374 * jac_reqd(p) (or NULL, to allocate automatically). | |
1375 * now, copy the numeric values to J->Z (numeric vector with length p*p). | |
1376 * (or, just set J->Z to point to the data vector. But remember this | |
1377 * will be overwritten by the decomposition). | |
1378 * finally, set: | |
1379 * J->st=JAC_RAW; | |
1380 * J->p = p; | |
1381 * | |
1382 * now, call jac_dec(J,meth) (optional) and the solve functions as required. | |
1383 * | |
1384 */ | |
1385 | |
1386 #include "math.h" | |
1387 #include "mut.h" | |
1388 | |
1389 #define DEF_METH JAC_EIGD | |
1390 | |
1391 int jac_reqd(int p) { return(2*p*(p+1)); } | |
1392 | |
1393 double *jac_alloc(J,p,wk) | |
1394 jacobian *J; | |
1395 int p; | |
1396 double *wk; | |
1397 { if (wk==NULL) | |
1398 wk = (double *)calloc(2*p*(p+1),sizeof(double)); | |
1399 if ( wk == NULL ) { | |
1400 printf("Problem allocating memory for wk\n");fflush(stdout); | |
1401 } | |
1402 J->Z = wk; wk += p*p; | |
1403 J->Q = wk; wk += p*p; | |
1404 J->wk= wk; wk += p; | |
1405 J->dg= wk; wk += p; | |
1406 return(wk); | |
1407 } | |
1408 | |
1409 void jacob_dec(J, meth) | |
1410 jacobian *J; | |
1411 int meth; | |
1412 { int i, j, p; | |
1413 | |
1414 if (J->st != JAC_RAW) return; | |
1415 | |
1416 J->sm = J->st = meth; | |
1417 switch(meth) | |
1418 { case JAC_EIG: | |
1419 eig_dec(J->Z,J->Q,J->p); | |
1420 return; | |
1421 case JAC_EIGD: | |
1422 p = J->p; | |
1423 for (i=0; i<p; i++) | |
1424 J->dg[i] = (J->Z[i*(p+1)]<=0) ? 0.0 : 1/sqrt(J->Z[i*(p+1)]); | |
1425 for (i=0; i<p; i++) | |
1426 for (j=0; j<p; j++) | |
1427 J->Z[i*p+j] *= J->dg[i]*J->dg[j]; | |
1428 eig_dec(J->Z,J->Q,J->p); | |
1429 J->st = JAC_EIGD; | |
1430 return; | |
1431 case JAC_CHOL: | |
1432 chol_dec(J->Z,J->p,J->p); | |
1433 return; | |
1434 default: mut_printf("jacob_dec: unknown method %d",meth); | |
1435 } | |
1436 } | |
1437 | |
1438 int jacob_solve(J,v) /* (X^T W X)^{-1} v */ | |
1439 jacobian *J; | |
1440 double *v; | |
1441 { int i, rank; | |
1442 | |
1443 if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); | |
1444 | |
1445 switch(J->st) | |
1446 { case JAC_EIG: | |
1447 return(eig_solve(J,v)); | |
1448 case JAC_EIGD: | |
1449 for (i=0; i<J->p; i++) v[i] *= J->dg[i]; | |
1450 rank = eig_solve(J,v); | |
1451 for (i=0; i<J->p; i++) v[i] *= J->dg[i]; | |
1452 return(rank); | |
1453 case JAC_CHOL: | |
1454 return(chol_solve(J->Z,v,J->p,J->p)); | |
1455 } | |
1456 mut_printf("jacob_solve: unknown method %d",J->st); | |
1457 return(0); | |
1458 } | |
1459 | |
1460 int jacob_hsolve(J,v) /* J^{-1/2} v */ | |
1461 jacobian *J; | |
1462 double *v; | |
1463 { int i; | |
1464 | |
1465 if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); | |
1466 | |
1467 switch(J->st) | |
1468 { case JAC_EIG: | |
1469 return(eig_hsolve(J,v)); | |
1470 case JAC_EIGD: /* eigenvalues on corr matrix */ | |
1471 for (i=0; i<J->p; i++) v[i] *= J->dg[i]; | |
1472 return(eig_hsolve(J,v)); | |
1473 case JAC_CHOL: | |
1474 return(chol_hsolve(J->Z,v,J->p,J->p)); | |
1475 } | |
1476 mut_printf("jacob_hsolve: unknown method %d\n",J->st); | |
1477 return(0); | |
1478 } | |
1479 | |
1480 int jacob_isolve(J,v) /* J^{-1/2} v */ | |
1481 jacobian *J; | |
1482 double *v; | |
1483 { int i, r; | |
1484 | |
1485 if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); | |
1486 | |
1487 switch(J->st) | |
1488 { case JAC_EIG: | |
1489 return(eig_isolve(J,v)); | |
1490 case JAC_EIGD: /* eigenvalues on corr matrix */ | |
1491 r = eig_isolve(J,v); | |
1492 for (i=0; i<J->p; i++) v[i] *= J->dg[i]; | |
1493 return(r); | |
1494 case JAC_CHOL: | |
1495 return(chol_isolve(J->Z,v,J->p,J->p)); | |
1496 } | |
1497 mut_printf("jacob_hsolve: unknown method %d\n",J->st); | |
1498 return(0); | |
1499 } | |
1500 | |
1501 double jacob_qf(J,v) /* vT J^{-1} v */ | |
1502 jacobian *J; | |
1503 double *v; | |
1504 { int i; | |
1505 | |
1506 if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); | |
1507 | |
1508 switch (J->st) | |
1509 { case JAC_EIG: | |
1510 return(eig_qf(J,v)); | |
1511 case JAC_EIGD: | |
1512 for (i=0; i<J->p; i++) v[i] *= J->dg[i]; | |
1513 return(eig_qf(J,v)); | |
1514 case JAC_CHOL: | |
1515 return(chol_qf(J->Z,v,J->p,J->p)); | |
1516 default: | |
1517 mut_printf("jacob_qf: invalid method\n"); | |
1518 return(0.0); | |
1519 } | |
1520 } | |
1521 | |
1522 int jacob_mult(J,v) /* J v */ | |
1523 jacobian *J; | |
1524 double *v; | |
1525 { | |
1526 if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); | |
1527 switch (J->st) | |
1528 { case JAC_CHOL: | |
1529 return(chol_mult(J->Z,v,J->p,J->p)); | |
1530 default: | |
1531 mut_printf("jacob_mult: invalid method\n"); | |
1532 return(0); | |
1533 } | |
1534 } | |
1535 /* | |
1536 * Copyright 1996-2006 Catherine Loader. | |
1537 */ | |
1538 /* | |
1539 * Routines for maximization of a one dimensional function f() | |
1540 * over an interval [xlo,xhi]. In all cases. the flag argument | |
1541 * controls the return: | |
1542 * flag='x', the maximizer xmax is returned. | |
1543 * otherwise, maximum f(xmax) is returned. | |
1544 * | |
1545 * max_grid(f,xlo,xhi,n,flag) | |
1546 * grid maximization of f() over [xlo,xhi] with n intervals. | |
1547 * | |
1548 * max_golden(f,xlo,xhi,n,tol,err,flag) | |
1549 * golden section maximization. | |
1550 * If n>2, an initial grid search is performed with n intervals | |
1551 * (this helps deal with local maxima). | |
1552 * convergence criterion is |x-xmax| < tol. | |
1553 * err is an error flag. | |
1554 * if flag='x', return value is xmax. | |
1555 * otherwise, return value is f(xmax). | |
1556 * | |
1557 * max_quad(f,xlo,xhi,n,tol,err,flag) | |
1558 * quadratic maximization. | |
1559 * | |
1560 * max_nr() | |
1561 * newton-raphson, handles multivariate case. | |
1562 * | |
1563 * TODO: additional error checking, non-convergence stop. | |
1564 */ | |
1565 | |
1566 #include <math.h> | |
1567 #include "mut.h" | |
1568 | |
1569 #define max_val(a,b) ((flag=='x') ? a : b) | |
1570 | |
1571 double max_grid(f,xlo,xhi,n,flag) | |
1572 double (*f)(), xlo, xhi; | |
1573 int n; | |
1574 char flag; | |
1575 { int i, mi; | |
1576 double x, y, mx, my; | |
1577 for (i=0; i<=n; i++) | |
1578 { x = xlo + (xhi-xlo)*i/n; | |
1579 y = f(x); | |
1580 if ((i==0) || (y>my)) | |
1581 { mx = x; | |
1582 my = y; | |
1583 mi = i; | |
1584 } | |
1585 } | |
1586 if (mi==0) return(max_val(xlo,my)); | |
1587 if (mi==n) return(max_val(xhi,my)); | |
1588 return(max_val(mx,my)); | |
1589 } | |
1590 | |
1591 double max_golden(f,xlo,xhi,n,tol,err,flag) | |
1592 double (*f)(), xhi, xlo, tol; | |
1593 int n, *err; | |
1594 char flag; | |
1595 { double dlt, x0, x1, x2, x3, y0, y1, y2, y3; | |
1596 *err = 0; | |
1597 | |
1598 if (n>2) | |
1599 { dlt = (xhi-xlo)/n; | |
1600 x0 = max_grid(f,xlo,xhi,n,'x'); | |
1601 if (xlo<x0) xlo = x0-dlt; | |
1602 if (xhi>x0) xhi = x0+dlt; | |
1603 } | |
1604 | |
1605 x0 = xlo; y0 = f(xlo); | |
1606 x3 = xhi; y3 = f(xhi); | |
1607 x1 = gold_rat*x0 + (1-gold_rat)*x3; y1 = f(x1); | |
1608 x2 = gold_rat*x3 + (1-gold_rat)*x0; y2 = f(x2); | |
1609 | |
1610 while (fabs(x3-x0)>tol) | |
1611 { if ((y1>=y0) && (y1>=y2)) | |
1612 { x3 = x2; y3 = y2; | |
1613 x2 = x1; y2 = y1; | |
1614 x1 = gold_rat*x0 + (1-gold_rat)*x3; y1 = f(x1); | |
1615 } | |
1616 else if ((y2>=y3) && (y2>=y1)) | |
1617 { x0 = x1; y0 = y1; | |
1618 x1 = x2; y1 = y2; | |
1619 x2 = gold_rat*x3 + (1-gold_rat)*x0; y2 = f(x2); | |
1620 } | |
1621 else | |
1622 { if (y3>y0) { x0 = x2; y0 = y2; } | |
1623 else { x3 = x1; y3 = y1; } | |
1624 x1 = gold_rat*x0 + (1-gold_rat)*x3; y1 = f(x1); | |
1625 x2 = gold_rat*x3 + (1-gold_rat)*x0; y2 = f(x2); | |
1626 } | |
1627 } | |
1628 if (y0>=y1) return(max_val(x0,y0)); | |
1629 if (y3>=y2) return(max_val(x3,y3)); | |
1630 return((y1>y2) ? max_val(x1,y1) : max_val(x2,y2)); | |
1631 } | |
1632 | |
1633 double max_quad(f,xlo,xhi,n,tol,err,flag) | |
1634 double (*f)(), xhi, xlo, tol; | |
1635 int n, *err; | |
1636 char flag; | |
1637 { double x0, x1, x2, xnew, y0, y1, y2, ynew, a, b; | |
1638 *err = 0; | |
1639 | |
1640 if (n>2) | |
1641 { x0 = max_grid(f,xlo,xhi,n,'x'); | |
1642 if (xlo<x0) xlo = x0-1.0/n; | |
1643 if (xhi>x0) xhi = x0+1.0/n; | |
1644 } | |
1645 | |
1646 x0 = xlo; y0 = f(x0); | |
1647 x2 = xhi; y2 = f(x2); | |
1648 x1 = (x0+x2)/2; y1 = f(x1); | |
1649 | |
1650 while (x2-x0>tol) | |
1651 { | |
1652 /* first, check (y0,y1,y2) is a peak. If not, | |
1653 * next interval is the halve with larger of (y0,y2). | |
1654 */ | |
1655 if ((y0>y1) | (y2>y1)) | |
1656 { | |
1657 if (y0>y2) { x2 = x1; y2 = y1; } | |
1658 else { x0 = x1; y0 = y1; } | |
1659 x1 = (x0+x2)/2; | |
1660 y1 = f(x1); | |
1661 } | |
1662 else /* peak */ | |
1663 { a = (y1-y0)*(x2-x1) + (y1-y2)*(x1-x0); | |
1664 b = ((y1-y0)*(x2-x1)*(x2+x1) + (y1-y2)*(x1-x0)*(x1+x0))/2; | |
1665 /* quadratic maximizer is b/a. But first check if a's too | |
1666 * small, since we may be close to constant. | |
1667 */ | |
1668 if ((a<=0) | (b<x0*a) | (b>x2*a)) | |
1669 { /* split the larger halve */ | |
1670 xnew = ((x2-x1) > (x1-x0)) ? (x1+x2)/2 : (x0+x1)/2; | |
1671 } | |
1672 else | |
1673 { xnew = b/a; | |
1674 if (10*xnew < (9*x0+x1)) xnew = (9*x0+x1)/10; | |
1675 if (10*xnew > (9*x2+x1)) xnew = (9*x2+x1)/10; | |
1676 if (fabs(xnew-x1) < 0.001*(x2-x0)) | |
1677 { | |
1678 if ((x2-x1) > (x1-x0)) | |
1679 xnew = (99*x1+x2)/100; | |
1680 else | |
1681 xnew = (99*x1+x0)/100; | |
1682 } | |
1683 } | |
1684 ynew = f(xnew); | |
1685 if (xnew>x1) | |
1686 { if (ynew >= y1) { x0 = x1; y0 = y1; x1 = xnew; y1 = ynew; } | |
1687 else { x2 = xnew; y2 = ynew; } | |
1688 } | |
1689 else | |
1690 { if (ynew >= y1) { x2 = x1; y2 = y1; x1 = xnew; y1 = ynew; } | |
1691 else { x0 = xnew; y0 = ynew; } | |
1692 } | |
1693 } | |
1694 } | |
1695 return(max_val(x1,y1)); | |
1696 } | |
1697 | |
1698 double max_nr(F, coef, old_coef, f1, delta, J, p, maxit, tol, err) | |
1699 double *coef, *old_coef, *f1, *delta, tol; | |
1700 int (*F)(), p, maxit, *err; | |
1701 jacobian *J; | |
1702 { double old_f, f, lambda; | |
1703 int i, j, fr; | |
1704 double nc, nd, cut; | |
1705 int rank; | |
1706 | |
1707 *err = NR_OK; | |
1708 J->p = p; | |
1709 fr = F(coef, &f, f1, J->Z); J->st = JAC_RAW; | |
1710 | |
1711 for (i=0; i<maxit; i++) | |
1712 { memcpy(old_coef,coef,p*sizeof(double)); | |
1713 old_f = f; | |
1714 rank = jacob_solve(J,f1); | |
1715 memcpy(delta,f1,p*sizeof(double)); | |
1716 | |
1717 if (rank==0) /* NR won't move! */ | |
1718 delta[0] = -f/f1[0]; | |
1719 | |
1720 lambda = 1.0; | |
1721 | |
1722 nc = innerprod(old_coef,old_coef,p); | |
1723 nd = innerprod(delta, delta, p); | |
1724 cut = sqrt(nc/nd); | |
1725 if (cut>1.0) cut = 1.0; | |
1726 cut *= 0.0001; | |
1727 do | |
1728 { for (j=0; j<p; j++) coef[j] = old_coef[j] + lambda*delta[j]; | |
1729 f = old_f - 1.0; | |
1730 fr = F(coef, &f, f1, J->Z); J->st = JAC_RAW; | |
1731 if (fr==NR_BREAK) return(old_f); | |
1732 | |
1733 lambda = (fr==NR_REDUCE) ? lambda/2 : lambda/10.0; | |
1734 } while ((lambda>cut) & (f <= old_f - 1.0e-3)); | |
1735 | |
1736 if (f < old_f - 1.0e-3) | |
1737 { *err = NR_NDIV; | |
1738 return(f); | |
1739 } | |
1740 if (fr==NR_REDUCE) return(f); | |
1741 | |
1742 if (fabs(f-old_f) < tol) return(f); | |
1743 | |
1744 } | |
1745 *err = NR_NCON; | |
1746 return(f); | |
1747 } | |
1748 /* | |
1749 * Copyright 1996-2006 Catherine Loader. | |
1750 */ | |
1751 #include <math.h> | |
1752 #include "mut.h" | |
1753 | |
1754 /* qr decomposition of X (n*p organized by column). | |
1755 * Take w for the ride, if not NULL. | |
1756 */ | |
1757 void qr(X,n,p,w) | |
1758 double *X, *w; | |
1759 int n, p; | |
1760 { int i, j, k, mi; | |
1761 double c, s, mx, nx, t; | |
1762 | |
1763 for (j=0; j<p; j++) | |
1764 { mi = j; | |
1765 mx = fabs(X[(n+1)*j]); | |
1766 nx = mx*mx; | |
1767 | |
1768 /* find the largest remaining element in j'th column, row mi. | |
1769 * flip that row with row j. | |
1770 */ | |
1771 for (i=j+1; i<n; i++) | |
1772 { nx += X[j*n+i]*X[j*n+i]; | |
1773 if (fabs(X[j*n+i])>mx) | |
1774 { mi = i; | |
1775 mx = fabs(X[j*n+i]); | |
1776 } | |
1777 } | |
1778 for (i=j; i<p; i++) | |
1779 { t = X[i*n+j]; | |
1780 X[i*n+j] = X[i*n+mi]; | |
1781 X[i*n+mi] = t; | |
1782 } | |
1783 if (w!=NULL) { t = w[j]; w[j] = w[mi]; w[mi] = t; } | |
1784 | |
1785 /* want the diag. element -ve, so we do the `good' Householder reflect. | |
1786 */ | |
1787 if (X[(n+1)*j]>0) | |
1788 { for (i=j; i<p; i++) X[i*n+j] = -X[i*n+j]; | |
1789 if (w!=NULL) w[j] = -w[j]; | |
1790 } | |
1791 | |
1792 nx = sqrt(nx); | |
1793 c = nx*(nx-X[(n+1)*j]); | |
1794 if (c!=0) | |
1795 { for (i=j+1; i<p; i++) | |
1796 { s = 0; | |
1797 for (k=j; k<n; k++) | |
1798 s += X[i*n+k]*X[j*n+k]; | |
1799 s = (s-nx*X[i*n+j])/c; | |
1800 for (k=j; k<n; k++) | |
1801 X[i*n+k] -= s*X[j*n+k]; | |
1802 X[i*n+j] += s*nx; | |
1803 } | |
1804 if (w != NULL) | |
1805 { s = 0; | |
1806 for (k=j; k<n; k++) | |
1807 s += w[k]*X[n*j+k]; | |
1808 s = (s-nx*w[j])/c; | |
1809 for (k=j; k<n; k++) | |
1810 w[k] -= s*X[n*j+k]; | |
1811 w[j] += s*nx; | |
1812 } | |
1813 X[j*n+j] = nx; | |
1814 } | |
1815 } | |
1816 } | |
1817 | |
1818 void qrinvx(R,x,n,p) | |
1819 double *R, *x; | |
1820 int n, p; | |
1821 { int i, j; | |
1822 for (i=p-1; i>=0; i--) | |
1823 { for (j=i+1; j<p; j++) x[i] -= R[j*n+i]*x[j]; | |
1824 x[i] /= R[i*n+i]; | |
1825 } | |
1826 } | |
1827 | |
1828 void qrtinvx(R,x,n,p) | |
1829 double *R, *x; | |
1830 int n, p; | |
1831 { int i, j; | |
1832 for (i=0; i<p; i++) | |
1833 { for (j=0; j<i; j++) x[i] -= R[i*n+j]*x[j]; | |
1834 x[i] /= R[i*n+i]; | |
1835 } | |
1836 } | |
1837 | |
1838 void qrsolv(R,x,n,p) | |
1839 double *R, *x; | |
1840 int n, p; | |
1841 { qrtinvx(R,x,n,p); | |
1842 qrinvx(R,x,n,p); | |
1843 } | |
1844 /* | |
1845 * Copyright 1996-2006 Catherine Loader. | |
1846 */ | |
1847 /* | |
1848 * solve f(x)=c by various methods, with varying stability etc... | |
1849 * xlo and xhi should be initial bounds for the solution. | |
1850 * convergence criterion is |f(x)-c| < tol. | |
1851 * | |
1852 * double solve_bisect(f,c,xmin,xmax,tol,bd_flag,err) | |
1853 * double solve_secant(f,c,xmin,xmax,tol,bd_flag,err) | |
1854 * Bisection and secant methods for solving of f(x)=c. | |
1855 * xmin and xmax are starting values and bound for solution. | |
1856 * tol = convergence criterion, |f(x)-c| < tol. | |
1857 * bd_flag = if (xmin,xmax) doesn't bound a solution, what action to take? | |
1858 * BDF_NONE returns error. | |
1859 * BDF_EXPRIGHT increases xmax. | |
1860 * BDF_EXPLEFT decreases xmin. | |
1861 * err = error flag. | |
1862 * The (xmin,xmax) bound is not formally necessary for the secant method. | |
1863 * But having such a bound vastly improves stability; the code performs | |
1864 * a bisection step whenever the iterations run outside the bounds. | |
1865 * | |
1866 * double solve_nr(f,f1,c,x0,tol,err) | |
1867 * Newton-Raphson solution of f(x)=c. | |
1868 * f1 = f'(x). | |
1869 * x0 = starting value. | |
1870 * tol = convergence criteria, |f(x)-c| < tol. | |
1871 * err = error flag. | |
1872 * No stability checks at present. | |
1873 * | |
1874 * double solve_fp(f,x0,tol) | |
1875 * fixed-point iteration to solve f(x)=x. | |
1876 * x0 = starting value. | |
1877 * tol = convergence criteria, stops when |f(x)-x| < tol. | |
1878 * Convergence requires |f'(x)|<1 in neighborhood of true solution; | |
1879 * f'(x) \approx 0 gives the fastest convergence. | |
1880 * No stability checks at present. | |
1881 * | |
1882 * TODO: additional error checking, non-convergence stop. | |
1883 */ | |
1884 | |
1885 #include <math.h> | |
1886 #include "mut.h" | |
1887 | |
1888 typedef struct { | |
1889 double xmin, xmax, x0, x1; | |
1890 double ymin, ymax, y0, y1; | |
1891 } solvest; | |
1892 | |
1893 int step_expand(f,c,sv,bd_flag) | |
1894 double (*f)(), c; | |
1895 solvest *sv; | |
1896 int bd_flag; | |
1897 { double x, y; | |
1898 if (sv->ymin*sv->ymax <= 0.0) return(0); | |
1899 if (bd_flag == BDF_NONE) | |
1900 { mut_printf("invalid bracket\n"); | |
1901 return(1); /* error */ | |
1902 } | |
1903 if (bd_flag == BDF_EXPRIGHT) | |
1904 { while (sv->ymin*sv->ymax > 0) | |
1905 { x = sv->xmax + 2*(sv->xmax-sv->xmin); | |
1906 y = f(x) - c; | |
1907 sv->xmin = sv->xmax; sv->xmax = x; | |
1908 sv->ymin = sv->ymax; sv->ymax = y; | |
1909 } | |
1910 return(0); | |
1911 } | |
1912 if (bd_flag == BDF_EXPLEFT) | |
1913 { while (sv->ymin*sv->ymax > 0) | |
1914 { x = sv->xmin - 2*(sv->xmax-sv->xmin); | |
1915 y = f(x) - c; | |
1916 sv->xmax = sv->xmin; sv->xmin = x; | |
1917 sv->ymax = sv->ymin; sv->ymin = y; | |
1918 } | |
1919 return(0); | |
1920 } | |
1921 mut_printf("step_expand: unknown bd_flag %d.\n",bd_flag); | |
1922 return(1); | |
1923 } | |
1924 | |
1925 int step_addin(sv,x,y) | |
1926 solvest *sv; | |
1927 double x, y; | |
1928 { sv->x1 = sv->x0; sv->x0 = x; | |
1929 sv->y1 = sv->y0; sv->y0 = y; | |
1930 if (y*sv->ymin > 0) | |
1931 { sv->xmin = x; | |
1932 sv->ymin = y; | |
1933 return(0); | |
1934 } | |
1935 if (y*sv->ymax > 0) | |
1936 { sv->xmax = x; | |
1937 sv->ymax = y; | |
1938 return(0); | |
1939 } | |
1940 if (y==0) | |
1941 { sv->xmin = sv->xmax = x; | |
1942 sv->ymin = sv->ymax = 0; | |
1943 return(0); | |
1944 } | |
1945 return(1); | |
1946 } | |
1947 | |
1948 int step_bisect(f,c,sv) | |
1949 double (*f)(), c; | |
1950 solvest *sv; | |
1951 { double x, y; | |
1952 x = sv->x0 = (sv->xmin + sv->xmax)/2; | |
1953 y = sv->y0 = f(x)-c; | |
1954 return(step_addin(sv,x,y)); | |
1955 } | |
1956 | |
1957 double solve_bisect(f,c,xmin,xmax,tol,bd_flag,err) | |
1958 double (*f)(), c, xmin, xmax, tol; | |
1959 int bd_flag, *err; | |
1960 { solvest sv; | |
1961 int z; | |
1962 *err = 0; | |
1963 sv.xmin = xmin; sv.ymin = f(xmin)-c; | |
1964 sv.xmax = xmax; sv.ymax = f(xmax)-c; | |
1965 *err = step_expand(f,c,&sv,bd_flag); | |
1966 if (*err>0) return(sv.xmin); | |
1967 while(1) /* infinite loop if f is discontinuous */ | |
1968 { z = step_bisect(f,c,&sv); | |
1969 if (z) | |
1970 { *err = 1; | |
1971 return(sv.x0); | |
1972 } | |
1973 if (fabs(sv.y0)<tol) return(sv.x0); | |
1974 } | |
1975 } | |
1976 | |
1977 int step_secant(f,c,sv) | |
1978 double (*f)(), c; | |
1979 solvest *sv; | |
1980 { double x, y; | |
1981 if (sv->y0==sv->y1) return(step_bisect(f,c,sv)); | |
1982 x = sv->x0 + (sv->x1-sv->x0)*sv->y0/(sv->y0-sv->y1); | |
1983 if ((x<=sv->xmin) | (x>=sv->xmax)) return(step_bisect(f,c,sv)); | |
1984 y = f(x)-c; | |
1985 return(step_addin(sv,x,y)); | |
1986 } | |
1987 | |
1988 double solve_secant(f,c,xmin,xmax,tol,bd_flag,err) | |
1989 double (*f)(), c, xmin, xmax, tol; | |
1990 int bd_flag, *err; | |
1991 { solvest sv; | |
1992 int z; | |
1993 *err = 0; | |
1994 sv.xmin = xmin; sv.ymin = f(xmin)-c; | |
1995 sv.xmax = xmax; sv.ymax = f(xmax)-c; | |
1996 *err = step_expand(f,c,&sv,bd_flag); | |
1997 if (*err>0) return(sv.xmin); | |
1998 sv.x0 = sv.xmin; sv.y0 = sv.ymin; | |
1999 sv.x1 = sv.xmax; sv.y1 = sv.ymax; | |
2000 while(1) /* infinite loop if f is discontinuous */ | |
2001 { z = step_secant(f,c,&sv); | |
2002 if (z) | |
2003 { *err = 1; | |
2004 return(sv.x0); | |
2005 } | |
2006 if (fabs(sv.y0)<tol) return(sv.x0); | |
2007 } | |
2008 } | |
2009 | |
2010 double solve_nr(f,f1,c,x0,tol,err) | |
2011 double (*f)(), (*f1)(), c, x0, tol; | |
2012 int *err; | |
2013 { double y; | |
2014 do | |
2015 { y = f(x0)-c; | |
2016 x0 -= y/f1(x0); | |
2017 } while (fabs(y)>tol); | |
2018 return(x0); | |
2019 } | |
2020 | |
2021 double solve_fp(f,x0,tol,maxit) | |
2022 double (*f)(), x0, tol; | |
2023 int maxit; | |
2024 { double x1; | |
2025 int i; | |
2026 for (i=0; i<maxit; i++) | |
2027 { x1 = f(x0); | |
2028 if (fabs(x1-x0)<tol) return(x1); | |
2029 x0 = x1; | |
2030 } | |
2031 return(x1); /* although it hasn't converged */ | |
2032 } | |
2033 /* | |
2034 * Copyright 1996-2006 Catherine Loader. | |
2035 */ | |
2036 #include "mut.h" | |
2037 | |
2038 void svd(x,p,q,d,mxit) /* svd of square matrix */ | |
2039 double *x, *p, *q; | |
2040 int d, mxit; | |
2041 { int i, j, k, iter, ms, zer; | |
2042 double r, u, v, cp, cm, sp, sm, c1, c2, s1, s2, mx; | |
2043 for (i=0; i<d; i++) | |
2044 for (j=0; j<d; j++) p[i*d+j] = q[i*d+j] = (i==j); | |
2045 for (iter=0; iter<mxit; iter++) | |
2046 { ms = 0; | |
2047 for (i=0; i<d; i++) | |
2048 for (j=i+1; j<d; j++) | |
2049 { s1 = fabs(x[i*d+j]); | |
2050 s2 = fabs(x[j*d+i]); | |
2051 mx = (s1>s2) ? s1 : s2; | |
2052 zer = 1; | |
2053 if (mx*mx>1.0e-15*fabs(x[i*d+i]*x[j*d+j])) | |
2054 { if (fabs(x[i*(d+1)])<fabs(x[j*(d+1)])) | |
2055 { for (k=0; k<d; k++) | |
2056 { u = x[i*d+k]; x[i*d+k] = x[j*d+k]; x[j*d+k] = u; | |
2057 u = p[k*d+i]; p[k*d+i] = p[k*d+j]; p[k*d+j] = u; | |
2058 } | |
2059 for (k=0; k<d; k++) | |
2060 { u = x[k*d+i]; x[k*d+i] = x[k*d+j]; x[k*d+j] = u; | |
2061 u = q[k*d+i]; q[k*d+i] = q[k*d+j]; q[k*d+j] = u; | |
2062 } | |
2063 } | |
2064 cp = x[i*(d+1)]+x[j*(d+1)]; | |
2065 sp = x[j*d+i]-x[i*d+j]; | |
2066 r = sqrt(cp*cp+sp*sp); | |
2067 if (r>0) { cp /= r; sp /= r; } | |
2068 else { cp = 1.0; zer = 0;} | |
2069 cm = x[i*(d+1)]-x[j*(d+1)]; | |
2070 sm = x[i*d+j]+x[j*d+i]; | |
2071 r = sqrt(cm*cm+sm*sm); | |
2072 if (r>0) { cm /= r; sm /= r; } | |
2073 else { cm = 1.0; zer = 0;} | |
2074 c1 = cm+cp; | |
2075 s1 = sm+sp; | |
2076 r = sqrt(c1*c1+s1*s1); | |
2077 if (r>0) { c1 /= r; s1 /= r; } | |
2078 else { c1 = 1.0; zer = 0;} | |
2079 if (fabs(s1)>ms) ms = fabs(s1); | |
2080 c2 = cm+cp; | |
2081 s2 = sp-sm; | |
2082 r = sqrt(c2*c2+s2*s2); | |
2083 if (r>0) { c2 /= r; s2 /= r; } | |
2084 else { c2 = 1.0; zer = 0;} | |
2085 for (k=0; k<d; k++) | |
2086 { u = x[i*d+k]; v = x[j*d+k]; | |
2087 x[i*d+k] = c1*u+s1*v; | |
2088 x[j*d+k] = c1*v-s1*u; | |
2089 u = p[k*d+i]; v = p[k*d+j]; | |
2090 p[k*d+i] = c1*u+s1*v; | |
2091 p[k*d+j] = c1*v-s1*u; | |
2092 } | |
2093 for (k=0; k<d; k++) | |
2094 { u = x[k*d+i]; v = x[k*d+j]; | |
2095 x[k*d+i] = c2*u-s2*v; | |
2096 x[k*d+j] = s2*u+c2*v; | |
2097 u = q[k*d+i]; v = q[k*d+j]; | |
2098 q[k*d+i] = c2*u-s2*v; | |
2099 q[k*d+j] = s2*u+c2*v; | |
2100 } | |
2101 if (zer) x[i*d+j] = x[j*d+i] = 0.0; | |
2102 ms = 1; | |
2103 } | |
2104 } | |
2105 if (ms==0) iter=mxit+10; | |
2106 } | |
2107 if (iter==mxit) mut_printf("Warning: svd not converged.\n"); | |
2108 for (i=0; i<d; i++) | |
2109 if (x[i*d+i]<0) | |
2110 { x[i*d+i] = -x[i*d+i]; | |
2111 for (j=0; j<d; j++) p[j*d+i] = -p[j*d+i]; | |
2112 } | |
2113 } | |
2114 | |
2115 int svdsolve(x,w,P,D,Q,d,tol) /* original X = PDQ^T; comp. QD^{-1}P^T x */ | |
2116 double *x, *w, *P, *D, *Q, tol; | |
2117 int d; | |
2118 { int i, j, rank; | |
2119 double mx; | |
2120 if (tol>0) | |
2121 { mx = D[0]; | |
2122 for (i=1; i<d; i++) if (D[i*(d+1)]>mx) mx = D[i*(d+1)]; | |
2123 tol *= mx; | |
2124 } | |
2125 rank = 0; | |
2126 for (i=0; i<d; i++) | |
2127 { w[i] = 0.0; | |
2128 for (j=0; j<d; j++) w[i] += P[j*d+i]*x[j]; | |
2129 } | |
2130 for (i=0; i<d; i++) | |
2131 if (D[i*d+i]>tol) | |
2132 { w[i] /= D[i*(d+1)]; | |
2133 rank++; | |
2134 } | |
2135 for (i=0; i<d; i++) | |
2136 { x[i] = 0.0; | |
2137 for (j=0; j<d; j++) x[i] += Q[i*d+j]*w[j]; | |
2138 } | |
2139 return(rank); | |
2140 } | |
2141 | |
2142 void hsvdsolve(x,w,P,D,Q,d,tol) /* original X = PDQ^T; comp. D^{-1/2}P^T x */ | |
2143 double *x, *w, *P, *D, *Q, tol; | |
2144 int d; | |
2145 { int i, j; | |
2146 double mx; | |
2147 if (tol>0) | |
2148 { mx = D[0]; | |
2149 for (i=1; i<d; i++) if (D[i*(d+1)]>mx) mx = D[i*(d+1)]; | |
2150 tol *= mx; | |
2151 } | |
2152 for (i=0; i<d; i++) | |
2153 { w[i] = 0.0; | |
2154 for (j=0; j<d; j++) w[i] += P[j*d+i]*x[j]; | |
2155 } | |
2156 for (i=0; i<d; i++) if (D[i*d+i]>tol) w[i] /= sqrt(D[i*(d+1)]); | |
2157 for (i=0; i<d; i++) x[i] = w[i]; | |
2158 } | |
2159 /* | |
2160 * Copyright 1996-2006 Catherine Loader. | |
2161 */ | |
2162 /* | |
2163 * Includes some miscellaneous vector functions: | |
2164 * setzero(v,p) sets all elements of v to 0. | |
2165 * unitvec(x,k,p) sets x to k'th unit vector e_k. | |
2166 * innerprod(v1,v2,p) inner product. | |
2167 * addouter(A,v1,v2,p,c) A <- A + c * v_1 v2^T | |
2168 * multmatscal(A,z,n) A <- A*z | |
2169 * matrixmultiply(A,B,C,m,n,p) C(m*p) <- A(m*n) * B(n*p) | |
2170 * transpose(x,m,n) inline transpose | |
2171 * m_trace(x,n) trace | |
2172 * vecsum(x,n) sum elements. | |
2173 */ | |
2174 | |
2175 #include "mut.h" | |
2176 | |
2177 void setzero(v,p) | |
2178 double *v; | |
2179 int p; | |
2180 { int i; | |
2181 for (i=0; i<p; i++) v[i] = 0.0; | |
2182 } | |
2183 | |
2184 void unitvec(x,k,p) | |
2185 double *x; | |
2186 int k, p; | |
2187 { setzero(x,p); | |
2188 x[k] = 1.0; | |
2189 } | |
2190 | |
2191 double innerprod(v1,v2,p) | |
2192 double *v1, *v2; | |
2193 int p; | |
2194 { int i; | |
2195 double s; | |
2196 s = 0; | |
2197 for (i=0; i<p; i++) s += v1[i]*v2[i]; | |
2198 return(s); | |
2199 } | |
2200 | |
2201 void addouter(A,v1,v2,p,c) | |
2202 double *A, *v1, *v2, c; | |
2203 int p; | |
2204 { int i, j; | |
2205 for (i=0; i<p; i++) | |
2206 for (j=0; j<p; j++) | |
2207 A[i*p+j] += c*v1[i]*v2[j]; | |
2208 } | |
2209 | |
2210 void multmatscal(A,z,n) | |
2211 double *A, z; | |
2212 int n; | |
2213 { int i; | |
2214 for (i=0; i<n; i++) A[i] *= z; | |
2215 } | |
2216 | |
2217 /* matrix multiply A (m*n) times B (n*p). | |
2218 * store in C (m*p). | |
2219 * all matrices stored by column. | |
2220 */ | |
2221 void matrixmultiply(A,B,C,m,n,p) | |
2222 double *A, *B, *C; | |
2223 int m, n, p; | |
2224 { int i, j, k, ij; | |
2225 for (i=0; i<m; i++) | |
2226 for (j=0; j<p; j++) | |
2227 { ij = j*m+i; | |
2228 C[ij] = 0.0; | |
2229 for (k=0; k<n; k++) | |
2230 C[ij] += A[k*m+i] * B[j*n+k]; | |
2231 } | |
2232 } | |
2233 | |
2234 /* | |
2235 * transpose() transposes an m*n matrix in place. | |
2236 * At input, the matrix has n rows, m columns and | |
2237 * x[0..n-1] is the is the first column. | |
2238 * At output, the matrix has m rows, n columns and | |
2239 * x[0..m-1] is the first column. | |
2240 */ | |
2241 void transpose(x,m,n) | |
2242 double *x; | |
2243 int m, n; | |
2244 { int t0, t, ti, tj; | |
2245 double z; | |
2246 for (t0=1; t0<m*n-2; t0++) | |
2247 { ti = t0%m; tj = t0/m; | |
2248 do | |
2249 { t = ti*n+tj; | |
2250 ti= t%m; | |
2251 tj= t/m; | |
2252 } while (t<t0); | |
2253 z = x[t]; | |
2254 x[t] = x[t0]; | |
2255 x[t0] = z; | |
2256 } | |
2257 } | |
2258 | |
2259 /* trace of an n*n square matrix. */ | |
2260 double m_trace(x,n) | |
2261 double *x; | |
2262 int n; | |
2263 { int i; | |
2264 double sum; | |
2265 sum = 0; | |
2266 for (i=0; i<n; i++) | |
2267 sum += x[i*(n+1)]; | |
2268 return(sum); | |
2269 } | |
2270 | |
2271 double vecsum(v,n) | |
2272 double *v; | |
2273 int n; | |
2274 { int i; | |
2275 double sum; | |
2276 sum = 0.0; | |
2277 for (i=0; i<n; i++) sum += v[i]; | |
2278 return(sum); | |
2279 } | |
2280 /* | |
2281 * Copyright 1996-2006 Catherine Loader. | |
2282 */ | |
2283 /* | |
2284 miscellaneous functions that may not be defined in the math | |
2285 libraries. The implementations are crude. | |
2286 mut_daws(x) -- dawson's function | |
2287 mut_exp(x) -- exp(x), but it won't overflow. | |
2288 | |
2289 where required, these must be #define'd in header files. | |
2290 | |
2291 also includes | |
2292 ptail(x) -- exp(x*x/2)*int_{-\infty}^x exp(-u^2/2)du for x < -6. | |
2293 logit(x) -- logistic function. | |
2294 expit(x) -- inverse of logit. | |
2295 factorial(n)-- factorial | |
2296 */ | |
2297 | |
2298 #include "mut.h" | |
2299 | |
2300 double mut_exp(x) | |
2301 double x; | |
2302 { if (x>700.0) return(1.014232054735004e+304); | |
2303 return(exp(x)); | |
2304 } | |
2305 | |
2306 double mut_daws(x) | |
2307 double x; | |
2308 { static double val[] = { | |
2309 0, 0.24485619356002, 0.46034428261948, 0.62399959848185, 0.72477845900708, | |
2310 0.76388186132749, 0.75213621001998, 0.70541701910853, 0.63998807456541, | |
2311 0.56917098836654, 0.50187821196415, 0.44274283060424, 0.39316687916687, | |
2312 0.35260646480842, 0.31964847250685, 0.29271122077502, 0.27039629581340, | |
2313 0.25160207761769, 0.23551176224443, 0.22153505358518, 0.20924575719548, | |
2314 0.19833146819662, 0.18855782729305, 0.17974461154688, 0.17175005072385 }; | |
2315 double h, f0, f1, f2, y, z, xx; | |
2316 int j, m; | |
2317 if (x<0) return(-mut_daws(-x)); | |
2318 if (x>6) | |
2319 { /* Tail series: 1/x + 1/x^3 + 1.3/x^5 + 1.3.5/x^7 + ... */ | |
2320 y = z = 1/x; | |
2321 j = 0; | |
2322 while (((f0=(2*j+1)/(x*x))<1) && (y>1.0e-10*z)) | |
2323 { y *= f0; | |
2324 z += y; | |
2325 j++; | |
2326 } | |
2327 return(z); | |
2328 } | |
2329 m = (int) (4*x); | |
2330 h = x-0.25*m; | |
2331 if (h>0.125) | |
2332 { m++; | |
2333 h = h-0.25; | |
2334 } | |
2335 xx = 0.25*m; | |
2336 f0 = val[m]; | |
2337 f1 = 1-xx*f0; | |
2338 z = f0+h*f1; | |
2339 y = h; | |
2340 j = 2; | |
2341 while (fabs(y)>z*1.0e-10) | |
2342 { f2 = -(j-1)*f0-xx*f1; | |
2343 y *= h/j; | |
2344 z += y*f2; | |
2345 f0 = f1; f1 = f2; | |
2346 j++; | |
2347 } | |
2348 return(z); | |
2349 } | |
2350 | |
2351 double ptail(x) /* exp(x*x/2)*int_{-\infty}^x exp(-u^2/2)du for x < -6 */ | |
2352 double x; | |
2353 { double y, z, f0; | |
2354 int j; | |
2355 y = z = -1.0/x; | |
2356 j = 0; | |
2357 while ((fabs(f0= -(2*j+1)/(x*x))<1) && (fabs(y)>1.0e-10*z)) | |
2358 { y *= f0; | |
2359 z += y; | |
2360 j++; | |
2361 } | |
2362 return(z); | |
2363 } | |
2364 | |
2365 double logit(x) | |
2366 double x; | |
2367 { return(log(x/(1-x))); | |
2368 } | |
2369 | |
2370 double expit(x) | |
2371 double x; | |
2372 { double u; | |
2373 if (x<0) | |
2374 { u = exp(x); | |
2375 return(u/(1+u)); | |
2376 } | |
2377 return(1/(1+exp(-x))); | |
2378 } | |
2379 | |
2380 int factorial(n) | |
2381 int n; | |
2382 { if (n<=1) return(1.0); | |
2383 return(n*factorial(n-1)); | |
2384 } | |
2385 /* | |
2386 * Copyright 1996-2006 Catherine Loader. | |
2387 */ | |
2388 /* | |
2389 * Constrained maximization of a bivariate function. | |
2390 * maxbvgrid(f,x,ll,ur,m0,m1) | |
2391 * maximizes over a grid of m0*m1 points. Returns the maximum, | |
2392 * and the maximizer through the array x. Usually this is a starter, | |
2393 * to choose between local maxima, followed by other routines to refine. | |
2394 * | |
2395 * maxbvstep(f,x,ymax,h,ll,ur,err) | |
2396 * essentially multivariate bisection. A 3x3 grid of points is | |
2397 * built around the starting value (x,ymax). This grid is moved | |
2398 * around (step size h[0] and h[1] in the two dimensions) until | |
2399 * the maximum is in the middle. Then, the step size is halved. | |
2400 * Usually, this will be called in a loop. | |
2401 * The error flag is set if the maximum can't be centered in a | |
2402 * reasonable number of steps. | |
2403 * | |
2404 * maxbv(f,x,h,ll,ur,m0,m1,tol) | |
2405 * combines the two previous functions. It begins with a grid search | |
2406 * (if m0>0 and m1>0), followed by refinement. Refines until both h | |
2407 * components are < tol. | |
2408 */ | |
2409 #include "mut.h" | |
2410 | |
2411 #define max(a,b) ((a)>(b) ? (a) : (b)) | |
2412 #define min(a,b) ((a)<(b) ? (a) : (b)) | |
2413 | |
2414 double maxbvgrid(f,x,ll,ur,m0,m1,con) | |
2415 double (*f)(), *x, *ll, *ur; | |
2416 int m0, m1, *con; | |
2417 { int i, j, im, jm; | |
2418 double y, ymax; | |
2419 | |
2420 im = -1; | |
2421 for (i=0; i<=m0; i++) | |
2422 { x[0] = ((m0-i)*ll[0] + i*ur[0])/m0; | |
2423 for (j=0; j<=m1; j++) | |
2424 { x[1] = ((m1-j)*ll[1] + j*ur[1])/m1; | |
2425 y = f(x); | |
2426 if ((im==-1) || (y>ymax)) | |
2427 { im = i; jm = j; | |
2428 ymax = y; | |
2429 } | |
2430 } | |
2431 } | |
2432 | |
2433 x[0] = ((m0-im)*ll[0] + im*ur[0])/m0; | |
2434 x[1] = ((m1-jm)*ll[1] + jm*ur[1])/m1; | |
2435 con[0] = (im==m0)-(im==0); | |
2436 con[1] = (jm==m1)-(jm==0); | |
2437 return(ymax); | |
2438 } | |
2439 | |
2440 double maxbvstep(f,x,ymax,h,ll,ur,err,con) | |
2441 double (*f)(), *x, ymax, *h, *ll, *ur; | |
2442 int *err, *con; | |
2443 { int i, j, ij, imax, steps, cts[2]; | |
2444 double newx, X[9][2], y[9]; | |
2445 | |
2446 imax =4; y[4] = ymax; | |
2447 | |
2448 for (i=(con[0]==-1)-1; i<2-(con[0]==1); i++) | |
2449 for (j=(con[1]==-1)-1; j<2-(con[1]==1); j++) | |
2450 { ij = 3*i+j+4; | |
2451 X[ij][0] = x[0]+i*h[0]; | |
2452 if (X[ij][0] < ll[0]+0.001*h[0]) X[ij][0] = ll[0]; | |
2453 if (X[ij][0] > ur[0]-0.001*h[0]) X[ij][0] = ur[0]; | |
2454 X[ij][1] = x[1]+j*h[1]; | |
2455 if (X[ij][1] < ll[1]+0.001*h[1]) X[ij][1] = ll[1]; | |
2456 if (X[ij][1] > ur[1]-0.001*h[1]) X[ij][1] = ur[1]; | |
2457 if (ij != 4) | |
2458 { y[ij] = f(X[ij]); | |
2459 if (y[ij]>ymax) { imax = ij; ymax = y[ij]; } | |
2460 } | |
2461 } | |
2462 | |
2463 steps = 0; | |
2464 cts[0] = cts[1] = 0; | |
2465 while ((steps<20) && (imax != 4)) | |
2466 { steps++; | |
2467 if ((con[0]>-1) && ((imax/3)==0)) /* shift right */ | |
2468 { | |
2469 cts[0]--; | |
2470 for (i=8; i>2; i--) | |
2471 { X[i][0] = X[i-3][0]; y[i] = y[i-3]; | |
2472 } | |
2473 imax = imax+3; | |
2474 if (X[imax][0]==ll[0]) | |
2475 con[0] = -1; | |
2476 else | |
2477 { newx = X[imax][0]-h[0]; | |
2478 if (newx < ll[0]+0.001*h[0]) newx = ll[0]; | |
2479 for (i=(con[1]==-1); i<3-(con[1]==1); i++) | |
2480 { X[i][0] = newx; | |
2481 y[i] = f(X[i]); | |
2482 if (y[i]>ymax) { ymax = y[i]; imax = i; } | |
2483 } | |
2484 con[0] = 0; | |
2485 } | |
2486 } | |
2487 | |
2488 if ((con[0]<1) && ((imax/3)==2)) /* shift left */ | |
2489 { | |
2490 cts[0]++; | |
2491 for (i=0; i<6; i++) | |
2492 { X[i][0] = X[i+3][0]; y[i] = y[i+3]; | |
2493 } | |
2494 imax = imax-3; | |
2495 if (X[imax][0]==ur[0]) | |
2496 con[0] = 1; | |
2497 else | |
2498 { newx = X[imax][0]+h[0]; | |
2499 if (newx > ur[0]-0.001*h[0]) newx = ur[0]; | |
2500 for (i=6+(con[1]==-1); i<9-(con[1]==1); i++) | |
2501 { X[i][0] = newx; | |
2502 y[i] = f(X[i]); | |
2503 if (y[i]>ymax) { ymax = y[i]; imax = i; } | |
2504 } | |
2505 con[0] = 0; | |
2506 } | |
2507 } | |
2508 | |
2509 if ((con[1]>-1) && ((imax%3)==0)) /* shift up */ | |
2510 { | |
2511 cts[1]--; | |
2512 for (i=9; i>0; i--) if (i%3 > 0) | |
2513 { X[i][1] = X[i-1][1]; y[i] = y[i-1]; | |
2514 } | |
2515 imax = imax+1; | |
2516 if (X[imax][1]==ll[1]) | |
2517 con[1] = -1; | |
2518 else | |
2519 { newx = X[imax][1]-h[1]; | |
2520 if (newx < ll[1]+0.001*h[1]) newx = ll[1]; | |
2521 for (i=3*(con[0]==-1); i<7-(con[0]==1); i=i+3) | |
2522 { X[i][1] = newx; | |
2523 y[i] = f(X[i]); | |
2524 if (y[i]>ymax) { ymax = y[i]; imax = i; } | |
2525 } | |
2526 con[1] = 0; | |
2527 } | |
2528 } | |
2529 | |
2530 if ((con[1]<1) && ((imax%3)==2)) /* shift down */ | |
2531 { | |
2532 cts[1]++; | |
2533 for (i=0; i<9; i++) if (i%3 < 2) | |
2534 { X[i][1] = X[i+1][1]; y[i] = y[i+1]; | |
2535 } | |
2536 imax = imax-1; | |
2537 if (X[imax][1]==ur[1]) | |
2538 con[1] = 1; | |
2539 else | |
2540 { newx = X[imax][1]+h[1]; | |
2541 if (newx > ur[1]-0.001*h[1]) newx = ur[1]; | |
2542 for (i=2+3*(con[0]==-1); i<9-(con[0]==1); i=i+3) | |
2543 { X[i][1] = newx; | |
2544 y[i] = f(X[i]); | |
2545 if (y[i]>ymax) { ymax = y[i]; imax = i; } | |
2546 } | |
2547 con[1] = 0; | |
2548 } | |
2549 } | |
2550 /* if we've taken 3 steps in one direction, try increasing the | |
2551 * corresponding h. | |
2552 */ | |
2553 if ((cts[0]==-2) | (cts[0]==2)) | |
2554 { h[0] = 2*h[0]; cts[0] = 0; } | |
2555 if ((cts[1]==-2) | (cts[1]==2)) | |
2556 { h[1] = 2*h[1]; cts[1] = 0; } | |
2557 } | |
2558 | |
2559 if (steps==40) | |
2560 *err = 1; | |
2561 else | |
2562 { | |
2563 h[0] /= 2.0; h[1] /= 2.0; | |
2564 *err = 0; | |
2565 } | |
2566 | |
2567 x[0] = X[imax][0]; | |
2568 x[1] = X[imax][1]; | |
2569 return(y[imax]); | |
2570 } | |
2571 | |
2572 #define BQMmaxp 5 | |
2573 | |
2574 int boxquadmin(J,b0,p,x0,ll,ur) | |
2575 jacobian *J; | |
2576 double *b0, *x0, *ll, *ur; | |
2577 int p; | |
2578 { double b[BQMmaxp], x[BQMmaxp], L[BQMmaxp*BQMmaxp], C[BQMmaxp*BQMmaxp], d[BQMmaxp]; | |
2579 double f, fmin; | |
2580 int i, imin, m, con[BQMmaxp], rlx; | |
2581 | |
2582 if (p>BQMmaxp) mut_printf("boxquadmin: maxp is 5.\n"); | |
2583 if (J->st != JAC_RAW) mut_printf("boxquadmin: must start with JAC_RAW.\n"); | |
2584 | |
2585 m = 0; | |
2586 setzero(L,p*p); | |
2587 setzero(x,p); | |
2588 memcpy(C,J->Z,p*p*sizeof(double)); | |
2589 for (i=0; i<p; i++) con[i] = 0; | |
2590 | |
2591 do | |
2592 { | |
2593 /* first, keep minimizing and add constraints, one at a time. | |
2594 */ | |
2595 do | |
2596 { | |
2597 matrixmultiply(C,x,b,p,p,1); | |
2598 for (i=0; i<p; i++) b[i] += b0[i]; | |
2599 conquadmin(J,b,p,L,d,m); | |
2600 /* if C matrix is not pd, don't even bother. | |
2601 * this relies on having used cholesky dec. | |
2602 */ | |
2603 if ((J->Z[0]==0.0) | (J->Z[3]==0.0)) return(1); | |
2604 fmin = 1.0; | |
2605 for (i=0; i<p; i++) if (con[i]==0) | |
2606 { f = 1.0; | |
2607 if (x0[i]+x[i]+b[i] < ll[i]) f = (ll[i]-x[i]-x0[i])/b[i]; | |
2608 if (x0[i]+x[i]+b[i] > ur[i]) f = (ur[i]-x[i]-x0[i])/b[i]; | |
2609 if (f<fmin) fmin = f; | |
2610 imin = i; | |
2611 } | |
2612 for (i=0; i<p; i++) x[i] += fmin*b[i]; | |
2613 if (fmin<1.0) | |
2614 { L[m*p+imin] = 1; | |
2615 m++; | |
2616 con[imin] = (b[imin]<0) ? -1 : 1; | |
2617 } | |
2618 } while ((fmin < 1.0) & (m<p)); | |
2619 | |
2620 /* now, can I relax any constraints? | |
2621 * compute slopes at current point. Can relax if: | |
2622 * slope is -ve on a lower boundary. | |
2623 * slope is +ve on an upper boundary. | |
2624 */ | |
2625 rlx = 0; | |
2626 if (m>0) | |
2627 { matrixmultiply(C,x,b,p,p,1); | |
2628 for (i=0; i<p; i++) b[i] += b0[i]; | |
2629 for (i=0; i<p; i++) | |
2630 { if ((con[i]==-1)&& (b[i]<0)) { con[i] = 0; rlx = 1; } | |
2631 if ((con[i]==1) && (b[i]>0)) { con[i] = 0; rlx = 1; } | |
2632 } | |
2633 | |
2634 if (rlx) /* reconstruct the constraint matrix */ | |
2635 { setzero(L,p*p); m = 0; | |
2636 for (i=0; i<p; i++) if (con[i] != 0) | |
2637 { L[m*p+i] = 1; | |
2638 m++; | |
2639 } | |
2640 } | |
2641 } | |
2642 } while (rlx); | |
2643 | |
2644 memcpy(b0,x,p*sizeof(double)); /* this is how far we should move from x0 */ | |
2645 return(0); | |
2646 } | |
2647 | |
2648 double maxquadstep(f,x,ymax,h,ll,ur,err,con) | |
2649 double (*f)(), *x, ymax, *h, *ll, *ur; | |
2650 int *err, *con; | |
2651 { jacobian J; | |
2652 double b[2], c[2], d, jwork[12]; | |
2653 double x0, x1, y0, y1, ym, h0, xl[2], xu[2], xi[2]; | |
2654 int i, m; | |
2655 | |
2656 *err = 0; | |
2657 | |
2658 /* first, can we relax any of the initial constraints? | |
2659 * if so, just do one step away from the boundary, and | |
2660 * return for restart. | |
2661 */ | |
2662 for (i=0; i<2; i++) | |
2663 if (con[i] != 0) | |
2664 { xi[0] = x[0]; xi[1] = x[1]; | |
2665 xi[i] = x[i]-con[i]*h[i]; | |
2666 y0 = f(xi); | |
2667 if (y0>ymax) | |
2668 { memcpy(x,xi,2*sizeof(double)); | |
2669 con[i] = 0; | |
2670 return(y0); | |
2671 } | |
2672 } | |
2673 | |
2674 /* now, all initial constraints remain active. | |
2675 */ | |
2676 | |
2677 m = 9; | |
2678 for (i=0; i<2; i++) if (con[i]==0) | |
2679 { m /= 3; | |
2680 xl[0] = x[0]; xl[1] = x[1]; | |
2681 xl[i] = max(x[i]-h[i],ll[i]); y0 = f(xl); | |
2682 x0 = xl[i]-x[i]; y0 -= ymax; | |
2683 xu[0] = x[0]; xu[1] = x[1]; | |
2684 xu[i] = min(x[i]+h[i],ur[i]); y1 = f(xu); | |
2685 x1 = xu[i]-x[i]; y1 -= ymax; | |
2686 if (x0*x1*(x1-x0)==0) { *err = 1; return(0.0); } | |
2687 b[i] = (x0*x0*y1-x1*x1*y0)/(x0*x1*(x0-x1)); | |
2688 c[i] = 2*(x0*y1-x1*y0)/(x0*x1*(x1-x0)); | |
2689 if (c[i] >= 0.0) { *err = 1; return(0.0); } | |
2690 xi[i] = (b[i]<0) ? xl[i] : xu[i]; | |
2691 } | |
2692 else { c[i] = -1.0; b[i] = 0.0; } /* enforce initial constraints */ | |
2693 | |
2694 if ((con[0]==0) && (con[1]==0)) | |
2695 { x0 = xi[0]-x[0]; | |
2696 x1 = xi[1]-x[1]; | |
2697 ym = f(xi) - ymax - b[0]*x0 - c[0]*x0*x0/2 - b[1]*x1 - c[1]*x1*x1/2; | |
2698 d = ym/(x0*x1); | |
2699 } | |
2700 else d = 0.0; | |
2701 | |
2702 /* now, maximize the quadratic. | |
2703 * y[4] + b0*x0 + b1*x1 + 0.5(c0*x0*x0 + c1*x1*x1 + 2*d*x0*x1) | |
2704 * -ve everything, to call quadmin. | |
2705 */ | |
2706 jac_alloc(&J,2,jwork); | |
2707 J.Z[0] = -c[0]; | |
2708 J.Z[1] = J.Z[2] = -d; | |
2709 J.Z[3] = -c[1]; | |
2710 J.st = JAC_RAW; | |
2711 J.p = 2; | |
2712 b[0] = -b[0]; b[1] = -b[1]; | |
2713 *err = boxquadmin(&J,b,2,x,ll,ur); | |
2714 if (*err) return(ymax); | |
2715 | |
2716 /* test to see if this step successfully increases... | |
2717 */ | |
2718 for (i=0; i<2; i++) | |
2719 { xi[i] = x[i]+b[i]; | |
2720 if (xi[i]<ll[i]+1e-8*h[i]) xi[i] = ll[i]; | |
2721 if (xi[i]>ur[i]-1e-8*h[i]) xi[i] = ur[i]; | |
2722 } | |
2723 y1 = f(xi); | |
2724 if (y1 < ymax) /* no increase */ | |
2725 { *err = 1; | |
2726 return(ymax); | |
2727 } | |
2728 | |
2729 /* wonderful. update x, h, with the restriction that h can't decrease | |
2730 * by a factor over 10, or increase by over 2. | |
2731 */ | |
2732 for (i=0; i<2; i++) | |
2733 { x[i] = xi[i]; | |
2734 if (x[i]==ll[i]) con[i] = -1; | |
2735 if (x[i]==ur[i]) con[i] = 1; | |
2736 h0 = fabs(b[i]); | |
2737 h0 = min(h0,2*h[i]); | |
2738 h0 = max(h0,h[i]/10); | |
2739 h[i] = min(h0,(ur[i]-ll[i])/2.0); | |
2740 } | |
2741 return(y1); | |
2742 } | |
2743 | |
2744 double maxbv(f,x,h,ll,ur,m0,m1,tol) | |
2745 double (*f)(), *x, *h, *ll, *ur, tol; | |
2746 int m0, m1; | |
2747 { double ymax; | |
2748 int err, con[2]; | |
2749 | |
2750 con[0] = con[1] = 0; | |
2751 if ((m0>0) & (m1>0)) | |
2752 { | |
2753 ymax = maxbvgrid(f,x,ll,ur,m0,m1,con); | |
2754 h[0] = (ur[0]-ll[0])/(2*m0); | |
2755 h[1] = (ur[1]-ll[1])/(2*m1); | |
2756 } | |
2757 else | |
2758 { x[0] = (ll[0]+ur[0])/2; | |
2759 x[1] = (ll[1]+ur[1])/2; | |
2760 h[0] = (ur[0]-ll[0])/2; | |
2761 h[1] = (ur[1]-ll[1])/2; | |
2762 ymax = f(x); | |
2763 } | |
2764 | |
2765 while ((h[0]>tol) | (h[1]>tol)) | |
2766 { ymax = maxbvstep(f,x,ymax,h,ll,ur,&err,con); | |
2767 if (err) mut_printf("maxbvstep failure\n"); | |
2768 } | |
2769 | |
2770 return(ymax); | |
2771 } | |
2772 | |
2773 double maxbvq(f,x,h,ll,ur,m0,m1,tol) | |
2774 double (*f)(), *x, *h, *ll, *ur, tol; | |
2775 int m0, m1; | |
2776 { double ymax; | |
2777 int err, con[2]; | |
2778 | |
2779 con[0] = con[1] = 0; | |
2780 if ((m0>0) & (m1>0)) | |
2781 { | |
2782 ymax = maxbvgrid(f,x,ll,ur,m0,m1,con); | |
2783 h[0] = (ur[0]-ll[0])/(2*m0); | |
2784 h[1] = (ur[1]-ll[1])/(2*m1); | |
2785 } | |
2786 else | |
2787 { x[0] = (ll[0]+ur[0])/2; | |
2788 x[1] = (ll[1]+ur[1])/2; | |
2789 h[0] = (ur[0]-ll[0])/2; | |
2790 h[1] = (ur[1]-ll[1])/2; | |
2791 ymax = f(x); | |
2792 } | |
2793 | |
2794 while ((h[0]>tol) | (h[1]>tol)) | |
2795 { /* first, try a quadratric step */ | |
2796 ymax = maxquadstep(f,x,ymax,h,ll,ur,&err,con); | |
2797 /* if the quadratic step fails, move the grid around */ | |
2798 if (err) | |
2799 { | |
2800 ymax = maxbvstep(f,x,ymax,h,ll,ur,&err,con); | |
2801 if (err) | |
2802 { mut_printf("maxbvstep failure\n"); | |
2803 return(ymax); | |
2804 } | |
2805 } | |
2806 } | |
2807 | |
2808 return(ymax); | |
2809 } | |
2810 /* | |
2811 * Copyright 1996-2006 Catherine Loader. | |
2812 */ | |
2813 #include "mut.h" | |
2814 | |
2815 prf mut_printf = (prf)printf; | |
2816 | |
2817 void mut_redirect(newprf) | |
2818 prf newprf; | |
2819 { mut_printf = newprf; | |
2820 } | |
2821 /* | |
2822 * Copyright 1996-2006 Catherine Loader. | |
2823 */ | |
2824 /* | |
2825 * function to find order of observations in an array. | |
2826 * | |
2827 * mut_order(x,ind,i0,i1) | |
2828 * x array to find order of. | |
2829 * ind integer array of indexes. | |
2830 * i0,i1 (integers) range to order. | |
2831 * | |
2832 * at output, ind[i0...i1] are permuted so that | |
2833 * x[ind[i0]] <= x[ind[i0+1]] <= ... <= x[ind[i1]]. | |
2834 * (with ties, output order of corresponding indices is arbitrary). | |
2835 * The array x is unchanged. | |
2836 * | |
2837 * Typically, if x has length n, then i0=0, i1=n-1 and | |
2838 * ind is (any permutation of) 0...n-1. | |
2839 */ | |
2840 | |
2841 #include "mut.h" | |
2842 | |
2843 double med3(x0,x1,x2) | |
2844 double x0, x1, x2; | |
2845 { if (x0<x1) | |
2846 { if (x2<x0) return(x0); | |
2847 if (x1<x2) return(x1); | |
2848 return(x2); | |
2849 } | |
2850 /* x1 < x0 */ | |
2851 if (x2<x1) return(x1); | |
2852 if (x0<x2) return(x0); | |
2853 return(x2); | |
2854 } | |
2855 | |
2856 void mut_order(x,ind,i0,i1) | |
2857 double *x; | |
2858 int *ind, i0, i1; | |
2859 { double piv; | |
2860 int i, l, r, z; | |
2861 | |
2862 if (i1<=i0) return; | |
2863 piv = med3(x[ind[i0]],x[ind[i1]],x[ind[(i0+i1)/2]]); | |
2864 l = i0; r = i0-1; | |
2865 | |
2866 /* at each stage, | |
2867 * x[i0..l-1] < piv | |
2868 * x[l..r] = piv | |
2869 * x[r+1..i-1] > piv | |
2870 * then, decide where to put x[i]. | |
2871 */ | |
2872 for (i=i0; i<=i1; i++) | |
2873 { if (x[ind[i]]==piv) | |
2874 { r++; | |
2875 z = ind[i]; ind[i] = ind[r]; ind[r] = z; | |
2876 } | |
2877 else if (x[ind[i]]<piv) | |
2878 { r++; | |
2879 z = ind[i]; ind[i] = ind[r]; ind[r] = ind[l]; ind[l] = z; | |
2880 l++; | |
2881 } | |
2882 } | |
2883 | |
2884 if (l>i0) mut_order(x,ind,i0,l-1); | |
2885 if (r<i1) mut_order(x,ind,r+1,i1); | |
2886 } | |
2887 /* | |
2888 * Copyright 1996-2006 Catherine Loader. | |
2889 */ | |
2890 #include "mut.h" | |
2891 | |
2892 #define LOG_2 0.6931471805599453094172321214581765680755 | |
2893 #define IBETA_LARGE 1.0e30 | |
2894 #define IBETA_SMALL 1.0e-30 | |
2895 #define IGAMMA_LARGE 1.0e30 | |
2896 #define DOUBLE_EP 2.2204460492503131E-16 | |
2897 | |
2898 double ibeta(x, a, b) | |
2899 double x, a, b; | |
2900 { int flipped = 0, i, k, count; | |
2901 double I = 0, temp, pn[6], ak, bk, next, prev, factor, val; | |
2902 if (x <= 0) return(0); | |
2903 if (x >= 1) return(1); | |
2904 /* use ibeta(x,a,b) = 1-ibeta(1-x,b,z) */ | |
2905 if ((a+b+1)*x > (a+1)) | |
2906 { flipped = 1; | |
2907 temp = a; | |
2908 a = b; | |
2909 b = temp; | |
2910 x = 1 - x; | |
2911 } | |
2912 pn[0] = 0.0; | |
2913 pn[2] = pn[3] = pn[1] = 1.0; | |
2914 count = 1; | |
2915 val = x/(1.0-x); | |
2916 bk = 1.0; | |
2917 next = 1.0; | |
2918 do | |
2919 { count++; | |
2920 k = count/2; | |
2921 prev = next; | |
2922 if (count%2 == 0) | |
2923 ak = -((a+k-1.0)*(b-k)*val)/((a+2.0*k-2.0)*(a+2.0*k-1.0)); | |
2924 else | |
2925 ak = ((a+b+k-1.0)*k*val)/((a+2.0*k)*(a+2.0*k-1.0)); | |
2926 pn[4] = bk*pn[2] + ak*pn[0]; | |
2927 pn[5] = bk*pn[3] + ak*pn[1]; | |
2928 next = pn[4] / pn[5]; | |
2929 for (i=0; i<=3; i++) | |
2930 pn[i] = pn[i+2]; | |
2931 if (fabs(pn[4]) >= IBETA_LARGE) | |
2932 for (i=0; i<=3; i++) | |
2933 pn[i] /= IBETA_LARGE; | |
2934 if (fabs(pn[4]) <= IBETA_SMALL) | |
2935 for (i=0; i<=3; i++) | |
2936 pn[i] /= IBETA_SMALL; | |
2937 } while (fabs(next-prev) > DOUBLE_EP*prev); | |
2938 /* factor = a*log(x) + (b-1)*log(1-x); | |
2939 factor -= mut_lgamma(a+1) + mut_lgamma(b) - mut_lgamma(a+b); */ | |
2940 factor = dbeta(x,a,b,1) + log(x/a); | |
2941 I = exp(factor) * next; | |
2942 return(flipped ? 1-I : I); | |
2943 } | |
2944 | |
2945 /* | |
2946 * Incomplete gamma function. | |
2947 * int_0^x u^{df-1} e^{-u} du / Gamma(df). | |
2948 */ | |
2949 double igamma(x, df) | |
2950 double x, df; | |
2951 { double factor, term, gintegral, pn[6], rn, ak, bk; | |
2952 int i, count, k; | |
2953 if (x <= 0.0) return(0.0); | |
2954 | |
2955 if (df < 1.0) | |
2956 return( dgamma(x,df+1.0,1.0,0) + igamma(x,df+1.0) ); | |
2957 | |
2958 factor = x * dgamma(x,df,1.0,0); | |
2959 /* factor = exp(df*log(x) - x - lgamma(df)); */ | |
2960 | |
2961 if (x > 1.0 && x >= df) | |
2962 { | |
2963 pn[0] = 0.0; | |
2964 pn[2] = pn[1] = 1.0; | |
2965 pn[3] = x; | |
2966 count = 1; | |
2967 rn = 1.0 / x; | |
2968 do | |
2969 { count++; | |
2970 k = count / 2; | |
2971 gintegral = rn; | |
2972 if (count%2 == 0) | |
2973 { bk = 1.0; | |
2974 ak = (double)k - df; | |
2975 } else | |
2976 { bk = x; | |
2977 ak = (double)k; | |
2978 } | |
2979 pn[4] = bk*pn[2] + ak*pn[0]; | |
2980 pn[5] = bk*pn[3] + ak*pn[1]; | |
2981 rn = pn[4] / pn[5]; | |
2982 for (i=0; i<4; i++) | |
2983 pn[i] = pn[i+2]; | |
2984 if (pn[4] > IGAMMA_LARGE) | |
2985 for (i=0; i<4; i++) | |
2986 pn[i] /= IGAMMA_LARGE; | |
2987 } while (fabs(gintegral-rn) > DOUBLE_EP*rn); | |
2988 gintegral = 1.0 - factor*rn; | |
2989 } | |
2990 else | |
2991 { /* For x<df, use the series | |
2992 * dpois(df,x)*( 1 + x/(df+1) + x^2/((df+1)(df+2)) + ... ) | |
2993 * This could be slow if df large and x/df is close to 1. | |
2994 */ | |
2995 gintegral = term = 1.0; | |
2996 rn = df; | |
2997 do | |
2998 { rn += 1.0; | |
2999 term *= x/rn; | |
3000 gintegral += term; | |
3001 } while (term > DOUBLE_EP*gintegral); | |
3002 gintegral *= factor/df; | |
3003 } | |
3004 return(gintegral); | |
3005 } | |
3006 | |
3007 double pf(q, df1, df2) | |
3008 double q, df1, df2; | |
3009 { return(ibeta(q*df1/(df2+q*df1), df1/2, df2/2)); | |
3010 } | |
3011 /* | |
3012 * Copyright 1996-2006 Catherine Loader. | |
3013 */ | |
3014 #include "mut.h" | |
3015 #include <string.h> | |
3016 | |
3017 /* quadmin: minimize the quadratic, | |
3018 * 2<x,b> + x^T A x. | |
3019 * x = -A^{-1} b. | |
3020 * | |
3021 * conquadmin: min. subject to L'x = d (m constraints) | |
3022 * x = -A^{-1}(b+Ly) (y = Lagrange multiplier) | |
3023 * y = -(L'A^{-1}L)^{-1} (L'A^{-1}b) | |
3024 * x = -A^{-1}b + A^{-1}L (L'A^{-1}L)^{-1} [(L'A^{-1})b + d] | |
3025 * (non-zero d to be added!!) | |
3026 * | |
3027 * qprogmin: min. subject to L'x >= 0. | |
3028 */ | |
3029 | |
3030 void quadmin(J,b,p) | |
3031 jacobian *J; | |
3032 double *b; | |
3033 int p; | |
3034 { int i; | |
3035 jacob_dec(J,JAC_CHOL); | |
3036 i = jacob_solve(J,b); | |
3037 if (i<p) mut_printf("quadmin singular %2d %2d\n",i,p); | |
3038 for (i=0; i<p; i++) b[i] = -b[i]; | |
3039 } | |
3040 | |
3041 /* project vector a (length n) onto | |
3042 * columns of X (n rows, m columns, organized by column). | |
3043 * store result in y. | |
3044 */ | |
3045 #define pmaxm 10 | |
3046 #define pmaxn 100 | |
3047 void project(a,X,y,n,m) | |
3048 double *a, *X, *y; | |
3049 int n, m; | |
3050 { double xta[pmaxm], R[pmaxn*pmaxm]; | |
3051 int i; | |
3052 | |
3053 if (n>pmaxn) mut_printf("project: n too large\n"); | |
3054 if (m>pmaxm) mut_printf("project: m too large\n"); | |
3055 | |
3056 for (i=0; i<m; i++) xta[i] = innerprod(a,&X[i*n],n); | |
3057 memcpy(R,X,m*n*sizeof(double)); | |
3058 qr(R,n,m,NULL); | |
3059 qrsolv(R,xta,n,m); | |
3060 | |
3061 matrixmultiply(X,xta,y,n,m,1); | |
3062 } | |
3063 | |
3064 void resproj(a,X,y,n,m) | |
3065 double *a, *X, *y; | |
3066 int n, m; | |
3067 { int i; | |
3068 project(a,X,y,n,m); | |
3069 for (i=0; i<n; i++) y[i] = a[i]-y[i]; | |
3070 } | |
3071 | |
3072 /* x = -A^{-1}b + A^{-1}L (L'A^{-1}L)^{-1} [(L'A^{-1})b + d] */ | |
3073 void conquadmin(J,b,n,L,d,m) | |
3074 jacobian *J; | |
3075 double *b, *L, *d; | |
3076 int m, n; | |
3077 { double bp[10], L0[100]; | |
3078 int i, j; | |
3079 | |
3080 if (n>10) mut_printf("conquadmin: max. n is 10.\n"); | |
3081 memcpy(L0,L,n*m*sizeof(double)); | |
3082 jacob_dec(J,JAC_CHOL); | |
3083 for (i=0; i<m; i++) jacob_hsolve(J,&L[i*n]); | |
3084 jacob_hsolve(J,b); | |
3085 | |
3086 resproj(b,L,bp,n,m); | |
3087 | |
3088 jacob_isolve(J,bp); | |
3089 for (i=0; i<n; i++) b[i] = -bp[i]; | |
3090 | |
3091 qr(L,n,m,NULL); | |
3092 qrsolv(L,d,n,m); | |
3093 for (i=0; i<n; i++) | |
3094 { bp[i] = 0; | |
3095 for (j=0; j<m; j++) bp[i] += L0[j*n+i]*d[j]; | |
3096 } | |
3097 jacob_solve(J,bp); | |
3098 for (i=0; i<n; i++) b[i] += bp[i]; | |
3099 } | |
3100 | |
3101 void qactivemin(J,b,n,L,d,m,ac) | |
3102 jacobian *J; | |
3103 double *b, *L, *d; | |
3104 int m, n, *ac; | |
3105 { int i, nac; | |
3106 double M[100], dd[10]; | |
3107 nac = 0; | |
3108 for (i=0; i<m; i++) if (ac[i]>0) | |
3109 { memcpy(&M[nac*n],&L[i*n],n*sizeof(double)); | |
3110 dd[nac] = d[i]; | |
3111 nac++; | |
3112 } | |
3113 conquadmin(J,b,n,M,dd,nac); | |
3114 } | |
3115 | |
3116 /* return 1 for full step; 0 if new constraint imposed. */ | |
3117 int movefrom(x0,x,n,L,d,m,ac) | |
3118 double *x0, *x, *L, *d; | |
3119 int n, m, *ac; | |
3120 { int i, imin; | |
3121 double c0, c1, lb, lmin; | |
3122 lmin = 1.0; | |
3123 for (i=0; i<m; i++) if (ac[i]==0) | |
3124 { c1 = innerprod(&L[i*n],x,n)-d[i]; | |
3125 if (c1<0.0) | |
3126 { c0 = innerprod(&L[i*n],x0,n)-d[i]; | |
3127 if (c0<0.0) | |
3128 { if (c1<c0) { lmin = 0.0; imin = 1; } | |
3129 } | |
3130 else | |
3131 { lb = c0/(c0-c1); | |
3132 if (lb<lmin) { lmin = lb; imin = i; } | |
3133 } | |
3134 } | |
3135 } | |
3136 for (i=0; i<n; i++) | |
3137 x0[i] = lmin*x[i]+(1-lmin)*x0[i]; | |
3138 if (lmin==1.0) return(1); | |
3139 ac[imin] = 1; | |
3140 return(0); | |
3141 } | |
3142 | |
3143 int qstep(J,b,x0,n,L,d,m,ac,deac) | |
3144 jacobian *J; | |
3145 double *b, *x0, *L, *d; | |
3146 int m, n, *ac, deac; | |
3147 { double x[10]; | |
3148 int i; | |
3149 | |
3150 if (m>10) mut_printf("qstep: too many constraints.\n"); | |
3151 if (deac) | |
3152 { for (i=0; i<m; i++) if (ac[i]==1) | |
3153 { ac[i] = 0; | |
3154 memcpy(x,b,n*sizeof(double)); | |
3155 qactivemin(J,x,n,L,d,m,ac); | |
3156 if (innerprod(&L[i*n],x,n)>d[i]) /* deactivate this constraint; should rem. */ | |
3157 i = m+10; | |
3158 else | |
3159 ac[i] = 1; | |
3160 } | |
3161 if (i==m) return(0); /* no deactivation possible */ | |
3162 } | |
3163 | |
3164 do | |
3165 { if (!deac) | |
3166 { memcpy(x,b,n*sizeof(double)); | |
3167 qactivemin(J,x,n,L,d,m,ac); | |
3168 } | |
3169 i = movefrom(x0,x,n,L,d,m,ac); | |
3170 | |
3171 deac = 0; | |
3172 } while (i==0); | |
3173 return(1); | |
3174 } | |
3175 | |
3176 /* | |
3177 * x0 is starting value; should satisfy constraints. | |
3178 * L is n*m constraint matrix. | |
3179 * ac is active constraint vector: | |
3180 * ac[i]=0, inactive. | |
3181 * ac[i]=1, active, but can be deactivated. | |
3182 * ac[i]=2, active, cannot be deactivated. | |
3183 */ | |
3184 | |
3185 void qprogmin(J,b,x0,n,L,d,m,ac) | |
3186 jacobian *J; | |
3187 double *b, *x0, *L, *d; | |
3188 int m, n, *ac; | |
3189 { int i; | |
3190 for (i=0; i<m; i++) if (ac[i]==0) | |
3191 { if (innerprod(&L[i*n],x0,n) < d[i]) ac[i] = 1; } | |
3192 jacob_dec(J,JAC_CHOL); | |
3193 qstep(J,b,x0,n,L,d,m,ac,0); | |
3194 while (qstep(J,b,x0,n,L,d,m,ac,1)); | |
3195 } | |
3196 | |
3197 void qpm(A,b,x0,n,L,d,m,ac) | |
3198 double *A, *b, *x0, *L, *d; | |
3199 int *n, *m, *ac; | |
3200 { jacobian J; | |
3201 double wk[1000]; | |
3202 jac_alloc(&J,*n,wk); | |
3203 memcpy(J.Z,A,(*n)*(*n)*sizeof(double)); | |
3204 J.p = *n; | |
3205 J.st = JAC_RAW; | |
3206 qprogmin(&J,b,x0,*n,L,d,*m,ac); | |
3207 } |