Mercurial > repos > vipints > rdiff
comparison rDiff/src/locfit/Source/liblocf.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 /* | |
10 * Integration for hazard rate estimation. The functions in this | |
11 * file are used to evaluate | |
12 * sum int_0^{Ti} W_i(t,x) A()A()' exp( P() ) dt | |
13 * for hazard rate models. | |
14 * | |
15 * These routines assume the weight function is supported on [-1,1]. | |
16 * hasint_sph multiplies by exp(base(lf,i)), which allows estimating | |
17 * the baseline in a proportional hazards model, when the covariate | |
18 * effect base(lf,i) is known. | |
19 * | |
20 * TODO: | |
21 * hazint_sph, should be able to reduce mint in some cases with | |
22 * small integration range. onedint could be used for beta-family | |
23 * (RECT,EPAN,BISQ,TRWT) kernels. | |
24 * hazint_prod, restrict terms from the sum based on x values. | |
25 * I should count obs >= max, and only do that integration once. | |
26 */ | |
27 | |
28 #include "locf.h" | |
29 | |
30 static double ilim[2*MXDIM], *ff, tmax; | |
31 static lfdata *haz_lfd; | |
32 static smpar *haz_sp; | |
33 | |
34 /* | |
35 * hrao returns 0 if integration region is empty. | |
36 * 1 otherwise. | |
37 */ | |
38 int haz_sph_int(dfx,cf,h,r1) | |
39 double *dfx, *cf, h, *r1; | |
40 { double s, t0, t1, wt, th; | |
41 int j, dim, p; | |
42 s = 0; p = npar(haz_sp); | |
43 dim = haz_lfd->d; | |
44 for (j=1; j<dim; j++) s += SQR(dfx[j]/(h*haz_lfd->sca[j])); | |
45 if (s>1) return(0); | |
46 | |
47 setzero(r1,p*p); | |
48 t1 = sqrt(1-s)*h*haz_lfd->sca[0]; | |
49 t0 = -t1; | |
50 if (t0<ilim[0]) t0 = ilim[0]; | |
51 if (t1>ilim[dim]) t1 = ilim[dim]; | |
52 if (t1>dfx[0]) t1 = dfx[0]; | |
53 if (t1<t0) return(0); | |
54 | |
55 /* Numerical integration by Simpson's rule. | |
56 */ | |
57 for (j=0; j<=de_mint; j++) | |
58 { dfx[0] = t0+(t1-t0)*j/de_mint; | |
59 wt = weight(haz_lfd, haz_sp, dfx, NULL, h, 0, 0.0); | |
60 fitfun(haz_lfd, haz_sp, dfx,NULL,ff,NULL); | |
61 th = innerprod(cf,ff,p); | |
62 if (link(haz_sp)==LLOG) th = exp(th); | |
63 wt *= 2+2*(j&1)-(j==0)-(j==de_mint); | |
64 addouter(r1,ff,ff,p,wt*th); | |
65 } | |
66 multmatscal(r1,(t1-t0)/(3*de_mint),p*p); | |
67 | |
68 return(1); | |
69 } | |
70 | |
71 int hazint_sph(t,resp,r1,cf,h) | |
72 double *t, *resp, *r1, *cf, h; | |
73 { int i, j, n, p, st; | |
74 double dfx[MXDIM], eb, sb; | |
75 p = npar(haz_sp); | |
76 setzero(resp,p*p); | |
77 sb = 0.0; | |
78 | |
79 n = haz_lfd->n; | |
80 for (i=0; i<=n; i++) | |
81 { | |
82 if (i==n) | |
83 { dfx[0] = tmax-t[0]; | |
84 for (j=1; j<haz_lfd->d; j++) dfx[j] = 0.0; | |
85 eb = exp(sb/n); | |
86 } | |
87 else | |
88 { eb = exp(base(haz_lfd,i)); sb += base(haz_lfd,i); | |
89 for (j=0; j<haz_lfd->d; j++) dfx[j] = datum(haz_lfd,j,i)-t[j]; | |
90 } | |
91 | |
92 st = haz_sph_int(dfx,cf,h,r1); | |
93 if (st) | |
94 for (j=0; j<p*p; j++) resp[j] += eb*r1[j]; | |
95 } | |
96 return(LF_OK); | |
97 } | |
98 | |
99 int hazint_prod(t,resp,x,cf,h) | |
100 double *t, *resp, *x, *cf, h; | |
101 { int d, p, i, j, k, st; | |
102 double dfx[MXDIM], t_prev, | |
103 hj, hs, ncf[MXDEG], ef, il1; | |
104 double prod_wk[MXDIM][2*MXDEG+1], eb, sb; | |
105 | |
106 p = npar(haz_sp); | |
107 d = haz_lfd->d; | |
108 setzero(resp,p*p); | |
109 hj = hs = h*haz_lfd->sca[0]; | |
110 | |
111 ncf[0] = cf[0]; | |
112 for (i=1; i<=deg(haz_sp); i++) | |
113 { ncf[i] = hj*cf[(i-1)*d+1]; hj *= hs; | |
114 } | |
115 | |
116 /* for i=0..n.... | |
117 * First we compute prod_wk[j], j=0..d. | |
118 * For j=0, this is int_0^T_i (u-t)^k W((u-t)/h) exp(b0*(u-t)) du | |
119 * For remaining j, (x(i,j)-x(j))^k Wj exp(bj*(x..-x.)) | |
120 * | |
121 * Second, we add to the integration (exp(a) incl. in integral) | |
122 * with the right factorial denominators. | |
123 */ | |
124 t_prev = ilim[0]; sb = 0.0; | |
125 for (i=0; i<=haz_lfd->n; i++) | |
126 { if (i==haz_lfd->n) | |
127 { dfx[0] = tmax-t[0]; | |
128 for (j=1; j<d; j++) dfx[j] = 0.0; | |
129 eb = exp(sb/haz_lfd->n); | |
130 } | |
131 else | |
132 { eb = exp(base(haz_lfd,i)); sb += base(haz_lfd,i); | |
133 for (j=0; j<d; j++) dfx[j] = datum(haz_lfd,j,i)-t[j]; | |
134 } | |
135 | |
136 if (dfx[0]>ilim[0]) /* else it doesn't contribute */ | |
137 { | |
138 /* time integral */ | |
139 il1 = (dfx[0]>ilim[d]) ? ilim[d] : dfx[0]; | |
140 if (il1 != t_prev) /* don't repeat! */ | |
141 { st = onedint(haz_sp,ncf,ilim[0]/hs,il1/hs,prod_wk[0]); | |
142 if (st>0) return(st); | |
143 hj = eb; | |
144 for (j=0; j<=2*deg(haz_sp); j++) | |
145 { hj *= hs; | |
146 prod_wk[0][j] *= hj; | |
147 } | |
148 t_prev = il1; | |
149 } | |
150 | |
151 /* covariate terms */ | |
152 for (j=1; j<d; j++) | |
153 { | |
154 ef = 0.0; | |
155 for (k=deg(haz_sp); k>0; k--) ef = (ef+dfx[j])*cf[1+(k-1)*d+j]; | |
156 ef = exp(ef); | |
157 prod_wk[j][0] = ef * W(dfx[j]/(h*haz_lfd->sca[j]),ker(haz_sp)); | |
158 for (k=1; k<=2*deg(haz_sp); k++) | |
159 prod_wk[j][k] = prod_wk[j][k-1] * dfx[j]; | |
160 } | |
161 | |
162 /* add to the integration. */ | |
163 prodintresp(resp,prod_wk,d,deg(haz_sp),p); | |
164 } /* if dfx0 > ilim0 */ | |
165 } /* n loop */ | |
166 | |
167 /* symmetrize */ | |
168 for (k=0; k<p; k++) | |
169 for (j=k; j<p; j++) | |
170 resp[j*p+k] = resp[k*p+j]; | |
171 return(LF_OK); | |
172 } | |
173 | |
174 int hazint(t,resp,resp1,cf,h) | |
175 double *t, *resp, *resp1, *cf, h; | |
176 { if (haz_lfd->d==1) return(hazint_prod(t,resp,resp1,cf,h)); | |
177 if (kt(haz_sp)==KPROD) return(hazint_prod(t,resp,resp1,cf,h)); | |
178 | |
179 return(hazint_sph(t,resp,resp1,cf,h)); | |
180 } | |
181 | |
182 void haz_init(lfd,des,sp,il) | |
183 lfdata *lfd; | |
184 design *des; | |
185 smpar *sp; | |
186 double *il; | |
187 { int i; | |
188 | |
189 haz_lfd = lfd; | |
190 haz_sp = sp; | |
191 | |
192 tmax = datum(lfd,0,0); | |
193 for (i=1; i<lfd->n; i++) tmax = MAX(tmax,datum(lfd,0,i)); | |
194 ff = des->xtwx.wk; | |
195 for (i=0; i<2*lfd->d; i++) ilim[i] = il[i]; | |
196 } | |
197 /* | |
198 * Copyright 1996-2006 Catherine Loader. | |
199 */ | |
200 /* | |
201 * | |
202 * Routines for one-dimensional numerical integration | |
203 * in density estimation. The entry point is | |
204 * | |
205 * onedint(cf,mi,l0,l1,resp) | |
206 * | |
207 * which evaluates int W(u)u^j exp( P(u) ), j=0..2*deg. | |
208 * P(u) = cf[0] + cf[1]u + cf[2]u^2/2 + ... + cf[deg]u^deg/deg! | |
209 * l0 and l1 are the integration limits. | |
210 * The results are returned through the vector resp. | |
211 * | |
212 */ | |
213 | |
214 #include "locf.h" | |
215 | |
216 static int debug; | |
217 | |
218 int exbctay(b,c,n,z) /* n-term taylor series of e^(bx+cx^2) */ | |
219 double b, c, *z; | |
220 int n; | |
221 { double ec[20]; | |
222 int i, j; | |
223 z[0] = 1; | |
224 for (i=1; i<=n; i++) z[i] = z[i-1]*b/i; | |
225 if (c==0.0) return(n); | |
226 if (n>=40) | |
227 { WARN(("exbctay limit to n<40")); | |
228 n = 39; | |
229 } | |
230 ec[0] = 1; | |
231 for (i=1; 2*i<=n; i++) ec[i] = ec[i-1]*c/i; | |
232 for (i=n; i>1; i--) | |
233 for (j=1; 2*j<=i; j++) | |
234 z[i] += ec[j]*z[i-2*j]; | |
235 return(n); | |
236 } | |
237 | |
238 double explinjtay(l0,l1,j,cf) | |
239 /* int_l0^l1 x^j e^(a+bx+cx^2); exbctay aroud l1 */ | |
240 double l0, l1, *cf; | |
241 int j; | |
242 { double tc[40], f, s; | |
243 int k, n; | |
244 if ((l0!=0.0) | (l1!=1.0)) WARN(("explinjtay: invalid l0, l1")); | |
245 n = exbctay(cf[1]+2*cf[2]*l1,cf[2],20,tc); | |
246 s = tc[0]/(j+1); | |
247 f = 1/(j+1); | |
248 for (k=1; k<=n; k++) | |
249 { f *= -k/(j+k+1.0); | |
250 s += tc[k]*f; | |
251 } | |
252 return(f); | |
253 } | |
254 | |
255 void explint1(l0,l1,cf,I,p) /* int x^j exp(a+bx); j=0..p-1 */ | |
256 double l0, l1, *cf, *I; | |
257 int p; | |
258 { double y0, y1, f; | |
259 int j, k, k1; | |
260 y0 = mut_exp(cf[0]+l0*cf[1]); | |
261 y1 = mut_exp(cf[0]+l1*cf[1]); | |
262 if (p<2*fabs(cf[1])) k = p; else k = (int)fabs(cf[1]); | |
263 | |
264 if (k>0) | |
265 { I[0] = (y1-y0)/cf[1]; | |
266 for (j=1; j<k; j++) /* forward steps for small j */ | |
267 { y1 *= l1; y0 *= l0; | |
268 I[j] = (y1-y0-j*I[j-1])/cf[1]; | |
269 } | |
270 if (k==p) return; | |
271 y1 *= l1; y0 *= l0; | |
272 } | |
273 | |
274 f = 1; k1 = k; | |
275 while ((k<50) && (f>1.0e-8)) /* initially Ik = diff(x^{k+1}e^{a+bx}) */ | |
276 { y1 *= l1; y0 *= l0; | |
277 I[k] = y1-y0; | |
278 if (k>=p) f *= fabs(cf[1])/(k+1); | |
279 k++; | |
280 } | |
281 if (k==50) WARN(("explint1: want k>50")); | |
282 I[k] = 0.0; | |
283 for (j=k-1; j>=k1; j--) /* now do back step recursion */ | |
284 I[j] = (I[j]-cf[1]*I[j+1])/(j+1); | |
285 } | |
286 | |
287 void explintyl(l0,l1,cf,I,p) /* small c, use taylor series and explint1 */ | |
288 double l0, l1, *cf, *I; | |
289 int p; | |
290 { int i; | |
291 double c; | |
292 explint1(l0,l1,cf,I,p+8); | |
293 c = cf[2]; | |
294 for (i=0; i<p; i++) | |
295 I[i] = (((I[i+8]*c/4+I[i+6])*c/3+I[i+4])*c/2+I[i+2])*c+I[i]; | |
296 } | |
297 | |
298 void solvetrid(X,y,m) | |
299 double *X, *y; | |
300 int m; | |
301 { int i; | |
302 double s; | |
303 for (i=1; i<m; i++) | |
304 { s = X[3*i]/X[3*i-2]; | |
305 X[3*i] = 0; X[3*i+1] -= s*X[3*i-1]; | |
306 y[i] -= s*y[i-1]; | |
307 } | |
308 for (i=m-2; i>=0; i--) | |
309 { s = X[3*i+2]/X[3*i+4]; | |
310 X[3*i+2] = 0; | |
311 y[i] -= s*y[i+1]; | |
312 } | |
313 for (i=0; i<m; i++) y[i] /= X[3*i+1]; | |
314 } | |
315 | |
316 void initi0i1(I,cf,y0,y1,l0,l1) | |
317 double *I, *cf, y0, y1, l0, l1; | |
318 { double a0, a1, c, d, bi; | |
319 d = -cf[1]/(2*cf[2]); c = sqrt(2*fabs(cf[2])); | |
320 a0 = c*(l0-d); a1 = c*(l1-d); | |
321 if (cf[2]<0) | |
322 { bi = mut_exp(cf[0]+cf[1]*d+cf[2]*d*d)/c; | |
323 if (a0>0) | |
324 { if (a0>6) I[0] = (y0*ptail(-a0)-y1*ptail(-a1))/c; | |
325 else I[0] = S2PI*(mut_pnorm(-a0)-mut_pnorm(-a1))*bi; | |
326 } | |
327 else | |
328 { if (a1< -6) I[0] = (y1*ptail(a1)-y0*ptail(a0))/c; | |
329 else I[0] = S2PI*(mut_pnorm(a1)-mut_pnorm(a0))*bi; | |
330 } | |
331 } | |
332 else | |
333 I[0] = (y1*mut_daws(a1)-y0*mut_daws(a0))/c; | |
334 I[1] = (y1-y0)/(2*cf[2])+d*I[0]; | |
335 } | |
336 | |
337 void explinsid(l0,l1,cf,I,p) /* large b; don't use fwd recursion */ | |
338 double l0, l1, *cf, *I; | |
339 int p; | |
340 { int k, k0, k1, k2; | |
341 double y0, y1, Z[150]; | |
342 if (debug) mut_printf("side: %8.5f %8.5f %8.5f limt %8.5f %8.5f p %2d\n",cf[0],cf[1],cf[2],l0,l1,p); | |
343 | |
344 k0 = 2; | |
345 k1 = (int)(fabs(cf[1])+fabs(2*cf[2])); | |
346 if (k1<2) k1 = 2; | |
347 if (k1>p+20) k1 = p+20; | |
348 k2 = p+20; | |
349 | |
350 if (k2>50) { mut_printf("onedint: k2 warning\n"); k2 = 50; } | |
351 if (debug) mut_printf("k0 %2d k1 %2d k2 %2d p %2d\n",k0,k1,k2,p); | |
352 | |
353 y0 = mut_exp(cf[0]+l0*(cf[1]+l0*cf[2])); | |
354 y1 = mut_exp(cf[0]+l1*(cf[1]+l1*cf[2])); | |
355 initi0i1(I,cf,y0,y1,l0,l1); | |
356 if (debug) mut_printf("i0 %8.5f i1 %8.5f\n",I[0],I[1]); | |
357 | |
358 y1 *= l1; y0 *= l0; /* should be x^(k1)*exp(..) */ | |
359 if (k0<k1) /* center steps; initially x^k*exp(...) */ | |
360 for (k=k0; k<k1; k++) | |
361 { y1 *= l1; y0 *= l0; | |
362 I[k] = y1-y0; | |
363 Z[3*k] = k; Z[3*k+1] = cf[1]; Z[3*k+2] = 2*cf[2]; | |
364 } | |
365 | |
366 y1 *= l1; y0 *= l0; /* should be x^(k1)*exp(..) */ | |
367 if (debug) mut_printf("k1 %2d y0 %8.5f y1 %8.5f\n",k1,y0,y1); | |
368 for (k=k1; k<k2; k++) | |
369 { y1 *= l1; y0 *= l0; | |
370 I[k] = y1-y0; | |
371 } | |
372 I[k2] = I[k2+1] = 0.0; | |
373 for (k=k2-1; k>=k1; k--) | |
374 I[k] = (I[k]-cf[1]*I[k+1]-2*cf[2]*I[k+2])/(k+1); | |
375 | |
376 if (k0<k1) | |
377 { I[k0] -= k0*I[k0-1]; | |
378 I[k1-1] -= 2*cf[2]*I[k1]; | |
379 Z[3*k0] = Z[3*k1-1] = 0; | |
380 solvetrid(&Z[3*k0],&I[k0],k1-k0); | |
381 } | |
382 if (debug) | |
383 { mut_printf("explinsid:\n"); | |
384 for (k=0; k<p; k++) mut_printf(" %8.5f\n",I[k]); | |
385 } | |
386 } | |
387 | |
388 void explinbkr(l0,l1,cf,I,p) /* small b,c; use back recursion */ | |
389 double l0, l1, *cf, *I; | |
390 int p; | |
391 { int k, km; | |
392 double y0, y1; | |
393 y0 = mut_exp(cf[0]+l0*(cf[1]+cf[2]*l0)); | |
394 y1 = mut_exp(cf[0]+l1*(cf[1]+cf[2]*l1)); | |
395 km = p+10; | |
396 for (k=0; k<=km; k++) | |
397 { y1 *= l1; y0 *= l0; | |
398 I[k] = y1-y0; | |
399 } | |
400 I[km+1] = I[km+2] = 0; | |
401 for (k=km; k>=0; k--) | |
402 I[k] = (I[k]-cf[1]*I[k+1]-2*cf[2]*I[k+2])/(k+1); | |
403 } | |
404 | |
405 void explinfbk0(l0,l1,cf,I,p) /* fwd and bac recur; b=0; c<0 */ | |
406 double l0, l1, *cf, *I; | |
407 int p; | |
408 { double y0, y1, f1, f2, f, ml2; | |
409 int k, ks; | |
410 | |
411 y0 = mut_exp(cf[0]+l0*l0*cf[2]); | |
412 y1 = mut_exp(cf[0]+l1*l1*cf[2]); | |
413 initi0i1(I,cf,y0,y1,l0,l1); | |
414 | |
415 ml2 = MAX(l0*l0,l1*l1); | |
416 ks = 1+(int)(2*fabs(cf[2])*ml2); | |
417 if (ks<2) ks = 2; | |
418 if (ks>p-3) ks = p; | |
419 | |
420 /* forward recursion for k < ks */ | |
421 for (k=2; k<ks; k++) | |
422 { y1 *= l1; y0 *= l0; | |
423 I[k] = (y1-y0-(k-1)*I[k-2])/(2*cf[2]); | |
424 } | |
425 if (ks==p) return; | |
426 | |
427 y1 *= l1*l1; y0 *= l0*l0; | |
428 for (k=ks; k<p; k++) /* set I[k] = x^{k+1}e^(a+cx^2) | {l0,l1} */ | |
429 { y1 *= l1; y0 *= l0; | |
430 I[k] = y1-y0; | |
431 } | |
432 | |
433 /* initialize I[p-2] and I[p-1] */ | |
434 f1 = 1.0/p; f2 = 1.0/(p-1); | |
435 I[p-1] *= f1; I[p-2] *= f2; | |
436 k = p; f = 1.0; | |
437 while (f>1.0e-8) | |
438 { y1 *= l1; y0 *= l0; | |
439 if ((k-p)%2==0) /* add to I[p-2] */ | |
440 { f2 *= -2*cf[2]/(k+1); | |
441 I[p-2] += (y1-y0)*f2; | |
442 } | |
443 else /* add to I[p-1] */ | |
444 { f1 *= -2*cf[2]/(k+1); | |
445 I[p-1] += (y1-y0)*f1; | |
446 f *= 2*fabs(cf[2])*ml2/(k+1); | |
447 } | |
448 k++; | |
449 } | |
450 | |
451 /* use back recursion for I[ks..(p-3)] */ | |
452 for (k=p-3; k>=ks; k--) | |
453 I[k] = (I[k]-2*cf[2]*I[k+2])/(k+1); | |
454 } | |
455 | |
456 void explinfbk(l0,l1,cf,I,p) /* fwd and bac recur; b not too large */ | |
457 double l0, l1, *cf, *I; | |
458 int p; | |
459 { double y0, y1; | |
460 int k, ks, km; | |
461 | |
462 y0 = mut_exp(cf[0]+l0*(cf[1]+l0*cf[2])); | |
463 y1 = mut_exp(cf[0]+l1*(cf[1]+l1*cf[2])); | |
464 initi0i1(I,cf,y0,y1,l0,l1); | |
465 | |
466 ks = (int)(3*fabs(cf[2])); | |
467 if (ks<3) ks = 3; | |
468 if (ks>0.75*p) ks = p; /* stretch the forward recurs as far as poss. */ | |
469 /* forward recursion for k < ks */ | |
470 for (k=2; k<ks; k++) | |
471 { y1 *= l1; y0 *= l0; | |
472 I[k] = (y1-y0-cf[1]*I[k-1]-(k-1)*I[k-2])/(2*cf[2]); | |
473 } | |
474 if (ks==p) return; | |
475 | |
476 km = p+15; | |
477 y1 *= l1*l1; y0 *= l0*l0; | |
478 for (k=ks; k<=km; k++) | |
479 { y1 *= l1; y0 *= l0; | |
480 I[k] = y1-y0; | |
481 } | |
482 I[km+1] = I[km+2] = 0.0; | |
483 for (k=km; k>=ks; k--) | |
484 I[k] = (I[k]-cf[1]*I[k+1]-2*cf[2]*I[k+2])/(k+1); | |
485 } | |
486 | |
487 void recent(I,resp,wt,p,s,x) | |
488 double *I, *resp, *wt, x; | |
489 int p, s; | |
490 { int i, j; | |
491 | |
492 /* first, use W taylor series I -> resp */ | |
493 for (i=0; i<=p; i++) | |
494 { resp[i] = 0.0; | |
495 for (j=0; j<s; j++) resp[i] += wt[j]*I[i+j]; | |
496 } | |
497 | |
498 /* now, recenter x -> 0 */ | |
499 if (x==0) return; | |
500 for (j=0; j<=p; j++) for (i=p; i>j; i--) resp[i] += x*resp[i-1]; | |
501 } | |
502 | |
503 void recurint(l0,l2,cf,resp,p,ker) | |
504 double l0, l2, *cf, *resp; | |
505 int p, ker; | |
506 { int i, s; | |
507 double l1, d0, d1, d2, dl, z0, z1, z2, wt[20], ncf[3], I[50], r1[5], r2[5]; | |
508 if (debug) mut_printf("\nrecurint: %8.5f %8.5f %8.5f %8.5f %8.5f\n",cf[0],cf[1],cf[2],l0,l2); | |
509 | |
510 if (cf[2]==0) /* go straight to explint1 */ | |
511 { s = wtaylor(wt,0.0,ker); | |
512 if (debug) mut_printf("case 1\n"); | |
513 explint1(l0,l2,cf,I,p+s); | |
514 recent(I,resp,wt,p,s,0.0); | |
515 return; | |
516 } | |
517 | |
518 dl = l2-l0; | |
519 d0 = cf[1]+2*l0*cf[2]; | |
520 d2 = cf[1]+2*l2*cf[2]; | |
521 z0 = cf[0]+l0*(cf[1]+l0*cf[2]); | |
522 z2 = cf[0]+l2*(cf[1]+l2*cf[2]); | |
523 | |
524 if ((fabs(cf[1]*dl)<1) && (fabs(cf[2]*dl*dl)<1)) | |
525 { ncf[0] = z0; ncf[1] = d0; ncf[2] = cf[2]; | |
526 if (debug) mut_printf("case 2\n"); | |
527 s = wtaylor(wt,l0,ker); | |
528 explinbkr(0.0,dl,ncf,I,p+s); | |
529 recent(I,resp,wt,p,s,l0); | |
530 return; | |
531 } | |
532 | |
533 if (fabs(cf[2]*dl*dl)<0.001) /* small c, use explint1+tay.ser */ | |
534 { ncf[0] = z0; ncf[1] = d0; ncf[2] = cf[2]; | |
535 if (debug) mut_printf("case small c\n"); | |
536 s = wtaylor(wt,l0,ker); | |
537 explintyl(0.0,l2-l0,ncf,I,p+s); | |
538 recent(I,resp,wt,p,s,l0); | |
539 return; | |
540 } | |
541 | |
542 if (d0*d2<=0) /* max/min in [l0,l2] */ | |
543 { l1 = -cf[1]/(2*cf[2]); | |
544 z1 = cf[0]+l1*(cf[1]+l1*cf[2]); | |
545 d1 = 0.0; | |
546 if (cf[2]<0) /* peak, integrate around l1 */ | |
547 { s = wtaylor(wt,l1,ker); | |
548 ncf[0] = z1; ncf[1] = 0.0; ncf[2] = cf[2]; | |
549 if (debug) mut_printf("case peak p %2d s %2d\n",p,s); | |
550 explinfbk0(l0-l1,l2-l1,ncf,I,p+s); | |
551 recent(I,resp,wt,p,s,l1); | |
552 return; | |
553 } | |
554 } | |
555 | |
556 if ((d0-2*cf[2]*dl)*(d2+2*cf[2]*dl)<0) /* max/min is close to [l0,l2] */ | |
557 { l1 = -cf[1]/(2*cf[2]); | |
558 z1 = cf[0]+l1*(cf[1]+l1*cf[2]); | |
559 if (l1<l0) { l1 = l0; z1 = z0; } | |
560 if (l1>l2) { l1 = l2; z1 = z2; } | |
561 | |
562 if ((z1>=z0) & (z1>=z2)) /* peak; integrate around l1 */ | |
563 { s = wtaylor(wt,l1,ker); | |
564 if (debug) mut_printf("case 4\n"); | |
565 d1 = cf[1]+2*l1*cf[2]; | |
566 ncf[0] = z1; ncf[1] = d1; ncf[2] = cf[2]; | |
567 explinfbk(l0-l1,l2-l1,ncf,I,p+s); | |
568 recent(I,resp,wt,p,s,l1); | |
569 return; | |
570 } | |
571 | |
572 /* trough; integrate [l0,l1] and [l1,l2] */ | |
573 for (i=0; i<=p; i++) r1[i] = r2[i] = 0.0; | |
574 if (l0<l1) | |
575 { s = wtaylor(wt,l0,ker); | |
576 if (debug) mut_printf("case 5\n"); | |
577 ncf[0] = z0; ncf[1] = d0; ncf[2] = cf[2]; | |
578 explinfbk(0.0,l1-l0,ncf,I,p+s); | |
579 recent(I,r1,wt,p,s,l0); | |
580 } | |
581 if (l1<l2) | |
582 { s = wtaylor(wt,l2,ker); | |
583 if (debug) mut_printf("case 6\n"); | |
584 ncf[0] = z2; ncf[1] = d2; ncf[2] = cf[2]; | |
585 explinfbk(l1-l2,0.0,ncf,I,p+s); | |
586 recent(I,r2,wt,p,s,l2); | |
587 } | |
588 for (i=0; i<=p; i++) resp[i] = r1[i]+r2[i]; | |
589 return; | |
590 } | |
591 | |
592 /* Now, quadratic is monotone on [l0,l2]; big b; moderate c */ | |
593 if (z2>z0+3) /* steep increase, expand around l2 */ | |
594 { s = wtaylor(wt,l2,ker); | |
595 if (debug) mut_printf("case 7\n"); | |
596 | |
597 | |
598 ncf[0] = z2; ncf[1] = d2; ncf[2] = cf[2]; | |
599 explinsid(l0-l2,0.0,ncf,I,p+s); | |
600 recent(I,resp,wt,p,s,l2); | |
601 if (debug) mut_printf("7 resp: %8.5f %8.5f %8.5f %8.5f\n",resp[0],resp[1],resp[2],resp[3]); | |
602 return; | |
603 } | |
604 | |
605 /* bias towards expansion around l0, because it's often 0 */ | |
606 if (debug) mut_printf("case 8\n"); | |
607 s = wtaylor(wt,l0,ker); | |
608 ncf[0] = z0; ncf[1] = d0; ncf[2] = cf[2]; | |
609 explinsid(0.0,l2-l0,ncf,I,p+s); | |
610 recent(I,resp,wt,p,s,l0); | |
611 return; | |
612 } | |
613 | |
614 int onedexpl(cf,deg,resp) | |
615 double *cf, *resp; | |
616 int deg; | |
617 { int i; | |
618 double f0, fr, fl; | |
619 if (deg>=2) LERR(("onedexpl only valid for deg=0,1")); | |
620 if (fabs(cf[1])>=EFACT) return(LF_BADP); | |
621 | |
622 f0 = exp(cf[0]); fl = fr = 1.0; | |
623 for (i=0; i<=2*deg; i++) | |
624 { f0 *= i+1; | |
625 fl /=-(EFACT+cf[1]); | |
626 fr /= EFACT-cf[1]; | |
627 resp[i] = f0*(fr-fl); | |
628 } | |
629 return(LF_OK); | |
630 } | |
631 | |
632 int onedgaus(cf,deg,resp) | |
633 double *cf, *resp; | |
634 int deg; | |
635 { int i; | |
636 double f0, mu, s2; | |
637 if (deg==3) | |
638 { LERR(("onedgaus only valid for deg=0,1,2")); | |
639 return(LF_ERR); | |
640 } | |
641 if (2*cf[2]>=GFACT*GFACT) return(LF_BADP); | |
642 | |
643 s2 = 1/(GFACT*GFACT-2*cf[2]); | |
644 mu = cf[1]*s2; | |
645 resp[0] = 1.0; | |
646 if (deg>=1) | |
647 { resp[1] = mu; | |
648 resp[2] = s2+mu*mu; | |
649 if (deg==2) | |
650 { resp[3] = mu*(3*s2+mu*mu); | |
651 resp[4] = 3*s2*s2 + mu*mu*(6*s2+mu*mu); | |
652 } | |
653 } | |
654 f0 = S2PI * exp(cf[0]+mu*mu/(2*s2))*sqrt(s2); | |
655 for (i=0; i<=2*deg; i++) resp[i] *= f0; | |
656 return(LF_OK); | |
657 } | |
658 | |
659 int onedint(sp,cf,l0,l1,resp) /* int W(u)u^j exp(..), j=0..2*deg */ | |
660 smpar *sp; | |
661 double *cf, l0, l1, *resp; | |
662 { double u, uj, y, ncf[4], rr[5]; | |
663 int i, j; | |
664 | |
665 if (debug) mut_printf("onedint: %f %f %f %f %f\n",cf[0],cf[1],cf[2],l0,l1); | |
666 | |
667 if (deg(sp)<=2) | |
668 { for (i=0; i<3; i++) ncf[i] = (i>deg(sp)) ? 0.0 : cf[i]; | |
669 ncf[2] /= 2; | |
670 | |
671 if (ker(sp)==WEXPL) return(onedexpl(ncf,deg(sp),resp)); | |
672 if (ker(sp)==WGAUS) return(onedgaus(ncf,deg(sp),resp)); | |
673 | |
674 if (l1>0) | |
675 recurint(MAX(l0,0.0),l1,ncf,resp,2*deg(sp),ker(sp)); | |
676 else for (i=0; i<=2*deg(sp); i++) resp[i] = 0; | |
677 | |
678 if (l0<0) | |
679 { ncf[1] = -ncf[1]; | |
680 l0 = -l0; l1 = -l1; | |
681 recurint(MAX(l1,0.0),l0,ncf,rr,2*deg(sp),ker(sp)); | |
682 } | |
683 else for (i=0; i<=2*deg(sp); i++) rr[i] = 0.0; | |
684 | |
685 for (i=0; i<=2*deg(sp); i++) | |
686 resp[i] += (i%2==0) ? rr[i] : -rr[i]; | |
687 | |
688 return(LF_OK); | |
689 } | |
690 | |
691 /* For degree >= 3, we use Simpson's rule. */ | |
692 for (j=0; j<=2*deg(sp); j++) resp[j] = 0.0; | |
693 for (i=0; i<=de_mint; i++) | |
694 { u = l0+(l1-l0)*i/de_mint; | |
695 y = cf[0]; uj = 1; | |
696 for (j=1; j<=deg(sp); j++) | |
697 { uj *= u; | |
698 y += cf[j]*uj/fact[j]; | |
699 } | |
700 y = (4-2*(i%2==0)-(i==0)-(i==de_mint)) * | |
701 W(fabs(u),ker(sp))*exp(MIN(y,300.0)); | |
702 for (j=0; j<=2*deg(sp); j++) | |
703 { resp[j] += y; | |
704 y *= u; | |
705 } | |
706 } | |
707 for (j=0; j<=2*deg(sp); j++) resp[j] = resp[j]*(l1-l0)/(3*de_mint); | |
708 return(LF_OK); | |
709 } | |
710 /* | |
711 * Copyright 1996-2006 Catherine Loader. | |
712 */ | |
713 #include "locf.h" | |
714 | |
715 extern int lf_status; | |
716 static double u[MXDIM], ilim[2*MXDIM], *ff, hh, *cff; | |
717 static lfdata *den_lfd; | |
718 static design *den_des; | |
719 static smpar *den_sp; | |
720 int fact[] = {1, 1, 2, 6, 24, 120, 720, 5040, 40320, 362880, 3628800}; | |
721 int de_mint = 20; | |
722 int de_itype = IDEFA; | |
723 int de_renorm= 0; | |
724 | |
725 int multint(), prodint(), gausint(), mlinint(); | |
726 | |
727 #define NITYPE 7 | |
728 static char *itype[NITYPE] = { "default", "multi", "product", "mlinear", | |
729 "hazard", "sphere", "monte" }; | |
730 static int ivals[NITYPE] = | |
731 { IDEFA, IMULT, IPROD, IMLIN, IHAZD, ISPHR, IMONT }; | |
732 int deitype(char *z) | |
733 { return(pmatch(z, itype, ivals, NITYPE, IDEFA)); | |
734 } | |
735 | |
736 void prresp(coef,resp,p) | |
737 double *coef, *resp; | |
738 int p; | |
739 { int i, j; | |
740 mut_printf("Coefficients:\n"); | |
741 for (i=0; i<p; i++) mut_printf("%8.5f ",coef[i]); | |
742 mut_printf("\n"); | |
743 mut_printf("Response matrix:\n"); | |
744 for (i=0; i<p; i++) | |
745 { for (j=0; j<p; j++) mut_printf("%9.6f, ",resp[i+j*p]); | |
746 mut_printf("\n"); | |
747 } | |
748 } | |
749 | |
750 int mif(u,d,resp,M) | |
751 double *u, *resp, *M; | |
752 int d; | |
753 { double wt; | |
754 int i, j, p; | |
755 | |
756 p = den_des->p; | |
757 wt = weight(den_lfd, den_sp, u, NULL, hh, 0, 0.0); | |
758 if (wt==0) | |
759 { setzero(resp,p*p); | |
760 return(p*p); | |
761 } | |
762 | |
763 fitfun(den_lfd, den_sp, u,NULL,ff,NULL); | |
764 if (link(den_sp)==LLOG) | |
765 wt *= mut_exp(innerprod(ff,cff,p)); | |
766 for (i=0; i<p; i++) | |
767 for (j=0; j<p; j++) | |
768 resp[i*p+j] = wt*ff[i]*ff[j]; | |
769 return(p*p); | |
770 } | |
771 | |
772 int multint(t,resp1,resp2,cf,h) | |
773 double *t, *resp1, *resp2, *cf, h; | |
774 { int d, i, mg[MXDIM]; | |
775 | |
776 if (ker(den_sp)==WGAUS) return(gausint(t,resp1,resp2,cf,h,den_lfd->sca)); | |
777 | |
778 d = den_lfd->d; | |
779 for (i=0; i<d; i++) mg[i] = de_mint; | |
780 | |
781 hh = h; | |
782 cff= cf; | |
783 simpsonm(mif,ilim,&ilim[d],d,resp1,mg,resp2); | |
784 return(LF_OK); | |
785 } | |
786 | |
787 int mlinint(t,resp1,resp2,cf,h) | |
788 double *t, *resp1, *resp2, *cf, h; | |
789 { | |
790 double hd, nb, wt, wu, g[4], w0, w1, v, *sca; | |
791 int d, p, i, j, jmax, k, l, z, jj[2]; | |
792 | |
793 d = den_lfd->d; p = den_des->p; sca = den_lfd->sca; | |
794 hd = 1; | |
795 for (i=0; i<d; i++) hd *= h*sca[i]; | |
796 | |
797 if (link(den_sp)==LIDENT) | |
798 { setzero(resp1,p*p); | |
799 resp1[0] = wint(d,NULL,0,ker(den_sp))*hd; | |
800 if (deg(den_sp)==0) return(LF_OK); | |
801 jj[0] = 2; w0 = wint(d,jj,1,ker(den_sp))*hd*h*h; | |
802 for (i=0; i<d; i++) resp1[(i+1)*p+i+1] = w0*sca[i]*sca[i]; | |
803 if (deg(den_sp)==1) return(LF_OK); | |
804 for (i=0; i<d; i++) | |
805 { j = p-(d-i)*(d-i+1)/2; | |
806 resp1[j] = resp1[p*j] = w0*sca[i]*sca[i]/2; | |
807 } | |
808 if (d>1) | |
809 { jj[1] = 2; | |
810 w0 = wint(d,jj,2,ker(den_sp)) * hd*h*h*h*h; | |
811 } | |
812 jj[0] = 4; | |
813 w1 = wint(d,jj,1,ker(den_sp)) * hd*h*h*h*h/4; | |
814 z = d+1; | |
815 for (i=0; i<d; i++) | |
816 { k = p-(d-i)*(d-i+1)/2; | |
817 for (j=i; j<d; j++) | |
818 { l = p-(d-j)*(d-j+1)/2; | |
819 if (i==j) resp1[z*p+z] = w1*SQR(sca[i])*SQR(sca[i]); | |
820 else | |
821 { resp1[z*p+z] = w0*SQR(sca[i])*SQR(sca[j]); | |
822 resp1[k*p+l] = resp1[k+p*l] = w0/4*SQR(sca[i])*SQR(sca[j]); | |
823 } | |
824 z++; | |
825 } } | |
826 return(LF_OK); | |
827 } | |
828 switch(deg(den_sp)) | |
829 { case 0: | |
830 resp1[0] = mut_exp(cf[0])*wint(d,NULL,0,ker(den_sp))*hd; | |
831 return(LF_OK); | |
832 case 1: | |
833 nb = 0.0; | |
834 for (i=1; i<=d; i++) | |
835 { v = h*cf[i]*sca[i-1]; | |
836 nb += v*v; | |
837 } | |
838 if (ker(den_sp)==WGAUS) | |
839 { w0 = 1/(GFACT*GFACT); | |
840 g[0] = mut_exp(cf[0]+w0*nb/2+d*log(S2PI/2.5)); | |
841 g[1] = g[3] = g[0]*w0; | |
842 g[2] = g[0]*w0*w0; | |
843 } | |
844 else | |
845 { wt = wu = mut_exp(cf[0]); | |
846 w0 = wint(d,NULL,0,ker(den_sp)); g[0] = wt*w0; | |
847 g[1] = g[2] = g[3] = 0.0; | |
848 j = 0; jmax = (d+2)*de_mint; | |
849 while ((j<jmax) && (wt*w0/g[0]>1.0e-8)) | |
850 { j++; | |
851 jj[0] = 2*j; w0 = wint(d,jj,1,ker(den_sp)); | |
852 if (d==1) g[3] += wt * w0; | |
853 else | |
854 { jj[0] = 2; jj[1] = 2*j-2; w1 = wint(d,jj,2,ker(den_sp)); | |
855 g[3] += wt*w1; | |
856 g[2] += wu*(w0-w1); | |
857 } | |
858 wt /= (2*j-1.0); g[1] += wt*w0; | |
859 wt *= nb/(2*j); g[0] += wt*w0; | |
860 wu /= (2*j-1.0)*(2*j); | |
861 if (j>1) wu *= nb; | |
862 } | |
863 if (j==jmax) WARN(("mlinint: series not converged")); | |
864 } | |
865 g[0] *= hd; g[1] *= hd; | |
866 g[2] *= hd; g[3] *= hd; | |
867 resp1[0] = g[0]; | |
868 for (i=1; i<=d; i++) | |
869 { resp1[i] = resp1[(d+1)*i] = cf[i]*SQR(h*sca[i-1])*g[1]; | |
870 for (j=1; j<=d; j++) | |
871 { resp1[(d+1)*i+j] = (i==j) ? g[3]*SQR(h*sca[i-1]) : 0; | |
872 resp1[(d+1)*i+j] += g[2]*SQR(h*h*sca[i-1]*sca[j-1])*cf[i]*cf[j]; | |
873 } | |
874 } | |
875 return(LF_OK); | |
876 } | |
877 LERR(("mlinint: deg=0,1 only")); | |
878 return(LF_ERR); | |
879 } | |
880 | |
881 void prodintresp(resp,prod_wk,dim,deg,p) | |
882 double *resp, prod_wk[MXDIM][2*MXDEG+1]; | |
883 int dim, deg, p; | |
884 { double prod; | |
885 int i, j, k, j1, k1; | |
886 | |
887 prod = 1.0; | |
888 for (i=0; i<dim; i++) prod *= prod_wk[i][0]; | |
889 resp[0] += prod; | |
890 if (deg==0) return; | |
891 | |
892 for (j1=1; j1<=deg; j1++) | |
893 { for (j=0; j<dim; j++) | |
894 { prod = 1.0; | |
895 for (i=0; i<dim; i++) prod *= prod_wk[i][j1*(j==i)]; | |
896 prod /= fact[j1]; | |
897 resp[1 + (j1-1)*dim +j] += prod; | |
898 } | |
899 } | |
900 | |
901 for (k1=1; k1<=deg; k1++) | |
902 for (j1=k1; j1<=deg; j1++) | |
903 { for (k=0; k<dim; k++) | |
904 for (j=0; j<dim; j++) | |
905 { prod = 1.0; | |
906 for (i=0; i<dim; i++) prod *= prod_wk[i][k1*(k==i) + j1*(j==i)]; | |
907 prod /= fact[k1]*fact[j1]; | |
908 resp[ (1+(k1-1)*dim+k)*p + 1+(j1-1)*dim+j] += prod; | |
909 } | |
910 } | |
911 } | |
912 | |
913 int prodint(t,resp,resp2,coef,h) | |
914 double *t, *resp, *resp2, *coef, h; | |
915 { int dim, p, i, j, k, st; | |
916 double cf[MXDEG+1], hj, hs, prod_wk[MXDIM][2*MXDEG+1]; | |
917 | |
918 dim = den_lfd->d; | |
919 p = den_des->p; | |
920 for (i=0; i<p*p; i++) resp[i] = 0.0; | |
921 cf[0] = coef[0]; | |
922 | |
923 /* compute the one dimensional terms | |
924 */ | |
925 for (i=0; i<dim; i++) | |
926 { hj = 1; hs = h*den_lfd->sca[i]; | |
927 for (j=0; j<deg(den_sp); j++) | |
928 { hj *= hs; | |
929 cf[j+1] = hj*coef[ j*dim+i+1 ]; | |
930 } | |
931 st = onedint(den_sp,cf,ilim[i]/hs,ilim[i+dim]/hs,prod_wk[i]); | |
932 if (st==LF_BADP) return(st); | |
933 hj = 1; | |
934 for (j=0; j<=2*deg(den_sp); j++) | |
935 { hj *= hs; | |
936 prod_wk[i][j] *= hj; | |
937 } | |
938 cf[0] = 0.0; /* so we only include it once, when d>=2 */ | |
939 } | |
940 | |
941 /* transfer to the resp array | |
942 */ | |
943 prodintresp(resp,prod_wk,dim,deg(den_sp),p); | |
944 | |
945 /* Symmetrize. | |
946 */ | |
947 for (k=0; k<p; k++) | |
948 for (j=k; j<p; j++) | |
949 resp[j*p+k] = resp[k*p+j]; | |
950 | |
951 return(st); | |
952 } | |
953 | |
954 int gausint(t,resp,C,cf,h,sca) | |
955 double *t, *resp, *C, *cf, h, *sca; | |
956 { double nb, det, z, *P; | |
957 int d, p, i, j, k, l, m1, m2, f; | |
958 d = den_lfd->d; p = den_des->p; | |
959 m1 = d+1; nb = 0; | |
960 P = &C[d*d]; | |
961 resp[0] = 1; | |
962 for (i=0; i<d; i++) | |
963 { C[i*d+i] = SQR(GFACT/(h*sca[i]))-cf[m1++]; | |
964 for (j=i+1; j<d; j++) C[i*d+j] = C[j*d+i] = -cf[m1++]; | |
965 } | |
966 eig_dec(C,P,d); | |
967 det = 1; | |
968 for (i=1; i<=d; i++) | |
969 { det *= C[(i-1)*(d+1)]; | |
970 if (det <= 0) return(LF_BADP); | |
971 resp[i] = cf[i]; | |
972 for (j=1; j<=d; j++) resp[j+i*p] = 0; | |
973 resp[i+i*p] = 1; | |
974 svdsolve(&resp[i*p+1],u,P,C,P,d,0.0); | |
975 } | |
976 svdsolve(&resp[1],u,P,C,P,d,0.0); | |
977 det = sqrt(det); | |
978 for (i=1; i<=d; i++) | |
979 { nb += cf[i]*resp[i]; | |
980 resp[i*p] = resp[i]; | |
981 for (j=1; j<=d; j++) | |
982 resp[i+p*j] += resp[i]*resp[j]; | |
983 } | |
984 m1 = d; | |
985 for (i=1; i<=d; i++) | |
986 for (j=i; j<=d; j++) | |
987 { m1++; f = 1+(i==j); | |
988 resp[m1] = resp[m1*p] = resp[i*p+j]/f; | |
989 m2 = d; | |
990 for (k=1; k<=d; k++) | |
991 { resp[m1+k*p] = resp[k+m1*p] = | |
992 ( resp[i]*resp[j*p+k] + resp[j]*resp[i*p+k] | |
993 + resp[k]*resp[i*p+j] - 2*resp[i]*resp[j]*resp[k] )/f; | |
994 for (l=k; l<=d; l++) | |
995 { m2++; f = (1+(i==j))*(1+(k==l)); | |
996 resp[m1+m2*p] = resp[m2+m1*p] = ( resp[i+j*p]*resp[k+l*p] | |
997 + resp[i+k*p]*resp[j+l*p] + resp[i+l*p]*resp[j+k*p] | |
998 - 2*resp[i]*resp[j]*resp[k]*resp[l] )/f; | |
999 } } } | |
1000 z = mut_exp(d*0.918938533+cf[0]+nb/2)/det; | |
1001 multmatscal(resp,z,p*p); | |
1002 return(LF_OK); | |
1003 } | |
1004 | |
1005 int likeden(coef, lk0, f1, A) | |
1006 double *coef, *lk0, *f1, *A; | |
1007 { double lk, r; | |
1008 int i, j, p, rstat; | |
1009 | |
1010 lf_status = LF_OK; | |
1011 p = den_des->p; | |
1012 if ((link(den_sp)==LIDENT) && (coef[0] != 0.0)) return(NR_BREAK); | |
1013 lf_status = (den_des->itype)(den_des->xev,A,den_des->xtwx.Q,coef,den_des->h); | |
1014 if (lf_error) lf_status = LF_ERR; | |
1015 if (lf_status==LF_BADP) | |
1016 { *lk0 = -1.0e300; | |
1017 return(NR_REDUCE); | |
1018 } | |
1019 if (lf_status!=LF_OK) return(NR_BREAK); | |
1020 if (lf_debug>2) prresp(coef,A,p); | |
1021 | |
1022 den_des->xtwx.p = p; | |
1023 rstat = NR_OK; | |
1024 switch(link(den_sp)) | |
1025 { case LLOG: | |
1026 r = den_des->ss[0]/A[0]; | |
1027 coef[0] += log(r); | |
1028 multmatscal(A,r,p*p); | |
1029 A[0] = den_des->ss[0]; | |
1030 lk = -A[0]; | |
1031 if (fabs(coef[0]) > 700) | |
1032 { lf_status = LF_OOB; | |
1033 rstat = NR_REDUCE; | |
1034 } | |
1035 for (i=0; i<p; i++) | |
1036 { lk += coef[i]*den_des->ss[i]; | |
1037 f1[i] = den_des->ss[i]-A[i]; | |
1038 } | |
1039 break; | |
1040 case LIDENT: | |
1041 lk = 0.0; | |
1042 for (i=0; i<p; i++) | |
1043 { f1[i] = den_des->ss[i]; | |
1044 for (j=0; j<p; j++) | |
1045 den_des->res[i] -= A[i*p+j]*coef[j]; | |
1046 } | |
1047 break; | |
1048 } | |
1049 *lk0 = den_des->llk = lk; | |
1050 | |
1051 return(rstat); | |
1052 } | |
1053 | |
1054 int inre(x,bound,d) | |
1055 double *x, *bound; | |
1056 int d; | |
1057 { int i, z; | |
1058 z = 1; | |
1059 for (i=0; i<d; i++) | |
1060 if (bound[i]<bound[i+d]) | |
1061 z &= (x[i]>=bound[i]) & (x[i]<=bound[i+d]); | |
1062 return(z); | |
1063 } | |
1064 | |
1065 int setintlimits(lfd, x, h, ang, lset) | |
1066 lfdata *lfd; | |
1067 int *ang, *lset; | |
1068 double *x, h; | |
1069 { int d, i; | |
1070 d = lfd->d; | |
1071 *ang = *lset = 0; | |
1072 for (i=0; i<d; i++) | |
1073 { if (lfd->sty[i]==STANGL) | |
1074 { ilim[i+d] = ((h<2) ? 2*asin(h/2) : PI)*lfd->sca[i]; | |
1075 ilim[i] = -ilim[i+d]; | |
1076 *ang = 1; | |
1077 } | |
1078 else | |
1079 { ilim[i+d] = h*lfd->sca[i]; | |
1080 ilim[i] = -ilim[i+d]; | |
1081 | |
1082 if (lfd->sty[i]==STLEFT) { ilim[i+d] = 0; *lset = 1; } | |
1083 if (lfd->sty[i]==STRIGH) { ilim[i] = 0; *lset = 1; } | |
1084 | |
1085 if (lfd->xl[i]<lfd->xl[i+d]) /* user limits for this variable */ | |
1086 { if (lfd->xl[i]-x[i]> ilim[i]) | |
1087 { ilim[i] = lfd->xl[i]-x[i]; *lset=1; } | |
1088 if (lfd->xl[i+d]-x[i]< ilim[i+d]) | |
1089 { ilim[i+d] = lfd->xl[i+d]-x[i]; *lset=1; } | |
1090 } | |
1091 } | |
1092 if (ilim[i]==ilim[i+d]) return(LF_DEMP); /* empty integration */ | |
1093 } | |
1094 return(LF_OK); | |
1095 } | |
1096 | |
1097 int selectintmeth(itype,lset,ang) | |
1098 int itype, lset, ang; | |
1099 { | |
1100 if (itype==IDEFA) /* select the default method */ | |
1101 { if (fam(den_sp)==THAZ) | |
1102 { if (ang) return(IDEFA); | |
1103 return( IHAZD ); | |
1104 } | |
1105 | |
1106 if (ubas(den_sp)) return(IMULT); | |
1107 | |
1108 if (ang) return(IMULT); | |
1109 | |
1110 if (iscompact(ker(den_sp))) | |
1111 { if (kt(den_sp)==KPROD) return(IPROD); | |
1112 if (lset) | |
1113 return( (den_lfd->d==1) ? IPROD : IMULT ); | |
1114 if (deg(den_sp)<=1) return(IMLIN); | |
1115 if (den_lfd->d==1) return(IPROD); | |
1116 return(IMULT); | |
1117 } | |
1118 | |
1119 if (ker(den_sp)==WGAUS) | |
1120 { if (lset) WARN(("Integration for Gaussian weights ignores limits")); | |
1121 if ((den_lfd->d==1)|(kt(den_sp)==KPROD)) return(IPROD); | |
1122 if (deg(den_sp)<=1) return(IMLIN); | |
1123 if (deg(den_sp)==2) return(IMULT); | |
1124 } | |
1125 | |
1126 return(IDEFA); | |
1127 } | |
1128 | |
1129 /* user provided an integration method, check it is valid */ | |
1130 | |
1131 if (fam(den_sp)==THAZ) | |
1132 { if (ang) return(INVLD); | |
1133 if (!iscompact(ker(den_sp))) return(INVLD); | |
1134 return( ((kt(den_sp)==KPROD) | (kt(den_sp)==KSPH)) ? IHAZD : INVLD ); | |
1135 } | |
1136 | |
1137 if ((ang) && (itype != IMULT)) return(INVLD); | |
1138 | |
1139 switch(itype) | |
1140 { case IMULT: | |
1141 if (ker(den_sp)==WGAUS) return(deg(den_sp)==2); | |
1142 return( iscompact(ker(den_sp)) ? IMULT : INVLD ); | |
1143 case IPROD: return( ((den_lfd->d==1) | (kt(den_sp)==KPROD)) ? IPROD : INVLD ); | |
1144 case IMLIN: return( ((kt(den_sp)==KSPH) && (!lset) && | |
1145 (deg(den_sp)<=1)) ? IMLIN : INVLD ); | |
1146 } | |
1147 | |
1148 return(INVLD); | |
1149 } | |
1150 | |
1151 extern double lf_tol; | |
1152 | |
1153 int densinit(lfd,des,sp) | |
1154 lfdata *lfd; | |
1155 design *des; | |
1156 smpar *sp; | |
1157 { int p, i, ii, j, nnz, rnz, ang, lset, status; | |
1158 double w, *cf; | |
1159 | |
1160 den_lfd = lfd; | |
1161 den_des = des; | |
1162 den_sp = sp; | |
1163 cf = des->cf; | |
1164 | |
1165 lf_tol = (link(sp)==LLOG) ? 1.0e-6 : 0.0; | |
1166 | |
1167 p = des->p; | |
1168 ff = des->xtwx.wk; | |
1169 cf[0] = NOSLN; | |
1170 for (i=1; i<p; i++) cf[i] = 0.0; | |
1171 | |
1172 if (!inre(des->xev,lfd->xl,lfd->d)) return(LF_XOOR); | |
1173 | |
1174 status = setintlimits(lfd,des->xev,des->h,&ang,&lset); | |
1175 if (status != LF_OK) return(status); | |
1176 | |
1177 switch(selectintmeth(de_itype,lset,ang)) | |
1178 { case IMULT: des->itype = multint; break; | |
1179 case IPROD: des->itype = prodint; break; | |
1180 case IMLIN: des->itype = mlinint; break; | |
1181 case IHAZD: des->itype = hazint; break; | |
1182 case INVLD: LERR(("Invalid integration method %d",de_itype)); | |
1183 break; | |
1184 case IDEFA: LERR(("No integration type available for this model")); | |
1185 break; | |
1186 default: LERR(("densinit: unknown integral type")); | |
1187 } | |
1188 | |
1189 switch(deg(den_sp)) | |
1190 { case 0: rnz = 1; break; | |
1191 case 1: rnz = 1; break; | |
1192 case 2: rnz = lfd->d+1; break; | |
1193 case 3: rnz = lfd->d+2; break; | |
1194 default: LERR(("densinit: invalid degree %d",deg(den_sp))); | |
1195 } | |
1196 if (lf_error) return(LF_ERR); | |
1197 | |
1198 setzero(des->ss,p); | |
1199 nnz = 0; | |
1200 for (i=0; i<des->n; i++) | |
1201 { ii = des->ind[i]; | |
1202 if (!cens(lfd,ii)) | |
1203 { w = wght(des,ii)*prwt(lfd,ii); | |
1204 for (j=0; j<p; j++) des->ss[j] += d_xij(des,ii,j)*w; | |
1205 if (wght(des,ii)>0.00001) nnz++; | |
1206 } } | |
1207 | |
1208 if (fam(den_sp)==THAZ) haz_init(lfd,des,sp,ilim); | |
1209 /* this should really only be done once. Not sure how to enforce that, | |
1210 * esp. when locfit() has been called directly. | |
1211 */ | |
1212 if (fam(den_sp)==TDEN) | |
1213 des->smwt = (lfd->w==NULL) ? lfd->n : vecsum(lfd->w,lfd->n); | |
1214 | |
1215 if (lf_debug>2) | |
1216 { mut_printf(" LHS: "); | |
1217 for (i=0; i<p; i++) mut_printf(" %8.5f",des->ss[i]); | |
1218 mut_printf("\n"); | |
1219 } | |
1220 | |
1221 switch(link(den_sp)) | |
1222 { case LIDENT: | |
1223 cf[0] = 0.0; | |
1224 return(LF_OK); | |
1225 case LLOG: | |
1226 if (nnz<rnz) { cf[0] = -1000; return(LF_DNOP); } | |
1227 cf[0] = 0.0; | |
1228 return(LF_OK); | |
1229 default: | |
1230 LERR(("unknown link in densinit")); | |
1231 return(LF_ERR); | |
1232 } | |
1233 } | |
1234 /* | |
1235 * Copyright 1996-2006 Catherine Loader. | |
1236 */ | |
1237 #include "locf.h" | |
1238 | |
1239 int bino_vallink(link) | |
1240 int link; | |
1241 { return((link==LLOGIT) | (link==LIDENT) | (link==LASIN)); | |
1242 } | |
1243 | |
1244 int bino_fam(y,p,th,link,res,cens,w) | |
1245 double y, p, th, *res, w; | |
1246 int link, cens; | |
1247 { double wp; | |
1248 if (link==LINIT) | |
1249 { if (y<0) y = 0; | |
1250 if (y>w) y = w; | |
1251 res[ZDLL] = y; | |
1252 return(LF_OK); | |
1253 } | |
1254 wp = w*p; | |
1255 if (link==LIDENT) | |
1256 { if ((p<=0) && (y>0)) return(LF_BADP); | |
1257 if ((p>=1) && (y<w)) return(LF_BADP); | |
1258 res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0.0; | |
1259 if (y>0) | |
1260 { res[ZLIK] += y*log(wp/y); | |
1261 res[ZDLL] += y/p; | |
1262 res[ZDDLL]+= y/(p*p); | |
1263 } | |
1264 if (y<w) | |
1265 { res[ZLIK] += (w-y)*log((w-wp)/(w-y)); | |
1266 res[ZDLL] -= (w-y)/(1-p); | |
1267 res[ZDDLL]+= (w-y)/SQR(1-p); | |
1268 } | |
1269 return(LF_OK); | |
1270 } | |
1271 if (link==LLOGIT) | |
1272 { if ((y<0) | (y>w)) /* goon observation; delete it */ | |
1273 { res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0.0; | |
1274 return(LF_OK); | |
1275 } | |
1276 res[ZLIK] = (th<0) ? th*y-w*log(1+exp(th)) : th*(y-w)-w*log(1+exp(-th)); | |
1277 if (y>0) res[ZLIK] -= y*log(y/w); | |
1278 if (y<w) res[ZLIK] -= (w-y)*log(1-y/w); | |
1279 res[ZDLL] = (y-wp); | |
1280 res[ZDDLL]= wp*(1-p); | |
1281 return(LF_OK); | |
1282 } | |
1283 if (link==LASIN) | |
1284 { if ((p<=0) && (y>0)) return(LF_BADP); | |
1285 if ((p>=1) && (y<w)) return(LF_BADP); | |
1286 if ((th<0) | (th>PI/2)) return(LF_BADP); | |
1287 res[ZDLL] = res[ZDDLL] = res[ZLIK] = 0; | |
1288 if (y>0) | |
1289 { res[ZDLL] += 2*y*sqrt((1-p)/p); | |
1290 res[ZLIK] += y*log(wp/y); | |
1291 } | |
1292 if (y<w) | |
1293 { res[ZDLL] -= 2*(w-y)*sqrt(p/(1-p)); | |
1294 res[ZLIK] += (w-y)*log((w-wp)/(w-y)); | |
1295 } | |
1296 res[ZDDLL] = 4*w; | |
1297 return(LF_OK); | |
1298 } | |
1299 LERR(("link %d invalid for binomial family",link)); | |
1300 return(LF_LNK); | |
1301 } | |
1302 | |
1303 int bino_check(sp,des,lfd) | |
1304 smpar *sp; | |
1305 design *des; | |
1306 lfdata *lfd; | |
1307 { int i, ii; | |
1308 double t0, t1; | |
1309 | |
1310 if (fabs(des->cf[0])>700) return(LF_OOB); | |
1311 | |
1312 /* check for separation. | |
1313 * this won't detect separation if there's boundary points with | |
1314 * both 0 and 1 responses. | |
1315 */ | |
1316 t0 = -1e100; t1 = 1e100; | |
1317 for (i=0; i<des->n; i++) | |
1318 { ii = des->ind[i]; | |
1319 if ((resp(lfd,ii)<prwt(lfd,ii)) && (fitv(des,ii) > t0)) t0 = fitv(des,ii); | |
1320 if ((resp(lfd,ii)>0) && (fitv(des,ii) < t1)) t1 = fitv(des,ii); | |
1321 if (t1 <= t0) return(LF_OK); | |
1322 } | |
1323 mut_printf("separated %8.5f %8.5f\n",t0,t1); | |
1324 return(LF_NSLN); | |
1325 } | |
1326 | |
1327 void setfbino(fam) | |
1328 family *fam; | |
1329 { fam->deflink = LLOGIT; | |
1330 fam->canlink = LLOGIT; | |
1331 fam->vallink = bino_vallink; | |
1332 fam->family = bino_fam; | |
1333 fam->pcheck = bino_check; | |
1334 } | |
1335 | |
1336 int rbin_vallink(link) | |
1337 int link; | |
1338 { return(link==LLOGIT); | |
1339 } | |
1340 | |
1341 int rbin_fam(y,p,th,link,res,cens,w) | |
1342 double y, p, th, *res, w; | |
1343 int link, cens; | |
1344 { double s2y; | |
1345 if (link==LINIT) | |
1346 { res[ZDLL] = y; | |
1347 return(LF_OK); | |
1348 } | |
1349 if ((y<0) | (y>w)) /* goon observation; delete it */ | |
1350 { res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0.0; | |
1351 return(LF_OK); | |
1352 } | |
1353 res[ZLIK] = (th<0) ? th*y-w*log(1+exp(th)) : th*(y-w)-w*log(1+exp(-th)); | |
1354 if (y>0) res[ZLIK] -= y*log(y/w); | |
1355 if (y<w) res[ZLIK] -= (w-y)*log(1-y/w); | |
1356 res[ZDLL] = (y-w*p); | |
1357 res[ZDDLL]= w*p*(1-p); | |
1358 if (-res[ZLIK]>HUBERC*HUBERC/2.0) | |
1359 { s2y = sqrt(-2*res[ZLIK]); | |
1360 res[ZLIK] = HUBERC*(HUBERC/2.0-s2y); | |
1361 res[ZDLL] *= HUBERC/s2y; | |
1362 res[ZDDLL] = HUBERC/s2y*(res[ZDDLL]-1/(s2y*s2y)*w*p*(1-p)); | |
1363 } | |
1364 return(LF_OK); | |
1365 } | |
1366 | |
1367 void setfrbino(fam) | |
1368 family *fam; | |
1369 { fam->deflink = LLOGIT; | |
1370 fam->canlink = LLOGIT; | |
1371 fam->vallink = rbin_vallink; | |
1372 fam->family = rbin_fam; | |
1373 fam->pcheck = bino_check; | |
1374 } | |
1375 /* | |
1376 * Copyright 1996-2006 Catherine Loader. | |
1377 */ | |
1378 #include "locf.h" | |
1379 | |
1380 int circ_vallink(link) | |
1381 int link; | |
1382 { return(link==LIDENT); | |
1383 } | |
1384 | |
1385 int circ_fam(y,mean,th,link,res,cens,w) | |
1386 double y, mean, th, *res, w; | |
1387 int link, cens; | |
1388 { if (link==LINIT) | |
1389 { res[ZDLL] = w*sin(y); | |
1390 res[ZLIK] = w*cos(y); | |
1391 return(LF_OK); | |
1392 } | |
1393 res[ZDLL] = w*sin(y-mean); | |
1394 res[ZDDLL]= w*cos(y-mean); | |
1395 res[ZLIK] = res[ZDDLL]-w; | |
1396 return(LF_OK); | |
1397 } | |
1398 | |
1399 extern double lf_tol; | |
1400 int circ_init(lfd,des,sp) | |
1401 lfdata *lfd; | |
1402 design *des; | |
1403 smpar *sp; | |
1404 { int i, ii; | |
1405 double s0, s1; | |
1406 s0 = s1 = 0.0; | |
1407 for (i=0; i<des->n; i++) | |
1408 { ii = des->ind[i]; | |
1409 s0 += wght(des,ii)*prwt(lfd,ii)*sin(resp(lfd,ii)-base(lfd,ii)); | |
1410 s1 += wght(des,ii)*prwt(lfd,ii)*cos(resp(lfd,ii)-base(lfd,ii)); | |
1411 } | |
1412 des->cf[0] = atan2(s0,s1); | |
1413 for (i=1; i<des->p; i++) des->cf[i] = 0.0; | |
1414 lf_tol = 1.0e-6; | |
1415 return(LF_OK); | |
1416 } | |
1417 | |
1418 | |
1419 void setfcirc(fam) | |
1420 family *fam; | |
1421 { fam->deflink = LIDENT; | |
1422 fam->canlink = LIDENT; | |
1423 fam->vallink = circ_vallink; | |
1424 fam->family = circ_fam; | |
1425 fam->initial = circ_init; | |
1426 } | |
1427 /* | |
1428 * Copyright 1996-2006 Catherine Loader. | |
1429 */ | |
1430 #include "locf.h" | |
1431 | |
1432 int dens_vallink(link) | |
1433 int link; | |
1434 { return((link==LIDENT) | (link==LLOG)); | |
1435 } | |
1436 | |
1437 int dens_fam(y,mean,th,link,res,cens,w) | |
1438 double y, mean, th, *res, w; | |
1439 int link, cens; | |
1440 { if (cens) | |
1441 res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0.0; | |
1442 else | |
1443 { res[ZLIK] = w*th; | |
1444 res[ZDLL] = res[ZDDLL] = w; | |
1445 } | |
1446 return(LF_OK); | |
1447 } | |
1448 | |
1449 void setfdensity(fam) | |
1450 family *fam; | |
1451 { fam->deflink = LLOG; | |
1452 fam->canlink = LLOG; | |
1453 fam->vallink = dens_vallink; | |
1454 fam->family = dens_fam; | |
1455 fam->initial = densinit; | |
1456 fam->like = likeden; | |
1457 } | |
1458 /* | |
1459 * Copyright 1996-2006 Catherine Loader. | |
1460 */ | |
1461 #include "locf.h" | |
1462 | |
1463 int gamma_vallink(link) | |
1464 int link; | |
1465 { return((link==LIDENT) | (link==LLOG) | (link==LINVER)); | |
1466 } | |
1467 | |
1468 int gamma_fam(y,mean,th,link,res,cens,w) | |
1469 double y, mean, th, *res, w; | |
1470 int link, cens; | |
1471 { double lb, pt, dg; | |
1472 if (link==LINIT) | |
1473 { res[ZDLL] = MAX(y,0.0); | |
1474 return(LF_OK); | |
1475 } | |
1476 res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0.0; | |
1477 if (w==0.0) return(LF_OK); | |
1478 if ((mean<=0) & (y>0)) return(LF_BADP); | |
1479 if (link==LIDENT) lb = 1/th; | |
1480 if (link==LINVER) lb = th; | |
1481 if (link==LLOG) lb = mut_exp(-th); | |
1482 if (cens) | |
1483 { if (y<=0) return(LF_OK); | |
1484 pt = 1-igamma(lb*y,w); | |
1485 dg = dgamma(lb*y,w,1.0,0); | |
1486 res[ZLIK] = log(pt); | |
1487 res[ZDLL] = -y*dg/pt; | |
1488 /* | |
1489 * res[ZDLL] = -y*dg/pt * dlb/dth. | |
1490 * res[ZDDLL] = y*dg/pt * (d2lb/dth2 + ((w-1)/lb-y)*(dlb/dth)^2) | |
1491 * + res[ZDLL]^2. | |
1492 */ | |
1493 if (link==LLOG) /* lambda = exp(-theta) */ | |
1494 { res[ZDLL] *= -lb; | |
1495 res[ZDDLL] = dg*y*lb*(w-lb*y)/pt + SQR(res[ZDLL]); | |
1496 return(LF_OK); | |
1497 } | |
1498 if (link==LINVER) /* lambda = theta */ | |
1499 { res[ZDLL] *= 1.0; | |
1500 res[ZDDLL] = dg*y*((w-1)*mean-y)/pt + SQR(res[ZDLL]); | |
1501 return(LF_OK); | |
1502 } | |
1503 if (link==LIDENT) /* lambda = 1/theta */ | |
1504 { res[ZDLL] *= -lb*lb; | |
1505 res[ZDDLL] = dg*y*lb*lb*lb*(1+w-lb*y)/pt + SQR(res[ZDLL]); | |
1506 return(LF_OK); | |
1507 } | |
1508 } | |
1509 else | |
1510 { if (y<0) WARN(("Negative Gamma observation")); | |
1511 if (link==LLOG) | |
1512 { res[ZLIK] = -lb*y+w*(1-th); | |
1513 if (y>0) res[ZLIK] += w*log(y/w); | |
1514 res[ZDLL] = lb*y-w; | |
1515 res[ZDDLL]= lb*y; | |
1516 return(LF_OK); | |
1517 } | |
1518 if (link==LINVER) | |
1519 { res[ZLIK] = -lb*y+w-w*log(mean); | |
1520 if (y>0) res[ZLIK] += w*log(y/w); | |
1521 res[ZDLL] = -y+w*mean; | |
1522 res[ZDDLL]= w*mean*mean; | |
1523 return(LF_OK); | |
1524 } | |
1525 if (link==LIDENT) | |
1526 { res[ZLIK] = -lb*y+w-w*log(mean); | |
1527 if (y>0) res[ZLIK] += w*log(y/w); | |
1528 res[ZDLL] = lb*lb*(y-w*mean); | |
1529 res[ZDDLL]= lb*lb*lb*(2*y-w*mean); | |
1530 return(LF_OK); | |
1531 } | |
1532 } | |
1533 LERR(("link %d invalid for Gamma family",link)); | |
1534 return(LF_LNK); | |
1535 } | |
1536 | |
1537 void setfgamma(fam) | |
1538 family *fam; | |
1539 { fam->deflink = LLOG; | |
1540 fam->canlink = LINVER; | |
1541 fam->vallink = gamma_vallink; | |
1542 fam->family = gamma_fam; | |
1543 } | |
1544 /* | |
1545 * Copyright 1996-2006 Catherine Loader. | |
1546 */ | |
1547 #include "locf.h" | |
1548 | |
1549 int gaus_vallink(link) | |
1550 int link; | |
1551 { return((link==LIDENT) | (link==LLOG) | (link==LLOGIT)); | |
1552 } | |
1553 | |
1554 int gaus_fam(y,mean,th,link,res,cens,w) | |
1555 double y, mean, th, *res, w; | |
1556 int link, cens; | |
1557 { double z, pz, dp; | |
1558 if (link==LINIT) | |
1559 { res[ZDLL] = w*y; | |
1560 return(LF_OK); | |
1561 } | |
1562 z = y-mean; | |
1563 if (cens) | |
1564 { if (link!=LIDENT) | |
1565 { LERR(("Link invalid for censored Gaussian family")); | |
1566 return(LF_LNK); | |
1567 } | |
1568 pz = mut_pnorm(-z); | |
1569 dp = ((z>6) ? ptail(-z) : exp(-z*z/2)/pz)/2.5066283; | |
1570 res[ZLIK] = w*log(pz); | |
1571 res[ZDLL] = w*dp; | |
1572 res[ZDDLL]= w*dp*(dp-z); | |
1573 return(LF_OK); | |
1574 } | |
1575 res[ZLIK] = -w*z*z/2; | |
1576 switch(link) | |
1577 { case LIDENT: | |
1578 res[ZDLL] = w*z; | |
1579 res[ZDDLL]= w; | |
1580 break; | |
1581 case LLOG: | |
1582 res[ZDLL] = w*z*mean; | |
1583 res[ZDDLL]= w*mean*mean; | |
1584 break; | |
1585 case LLOGIT: | |
1586 res[ZDLL] = w*z*mean*(1-mean); | |
1587 res[ZDDLL]= w*mean*mean*(1-mean)*(1-mean); | |
1588 break; | |
1589 default: | |
1590 LERR(("Invalid link for Gaussian family")); | |
1591 return(LF_LNK); | |
1592 } | |
1593 return(LF_OK); | |
1594 } | |
1595 | |
1596 int gaus_check(sp,des,lfd) | |
1597 smpar *sp; | |
1598 design *des; | |
1599 lfdata *lfd; | |
1600 { int i, ii; | |
1601 if (fami(sp)->robust) return(LF_OK); | |
1602 if (link(sp)==LIDENT) | |
1603 { for (i=0; i<des->n; i++) | |
1604 { ii = des->ind[i]; | |
1605 if (cens(lfd,ii)) return(LF_OK); | |
1606 } | |
1607 return(LF_DONE); | |
1608 } | |
1609 return(LF_OK); | |
1610 } | |
1611 | |
1612 void setfgauss(fam) | |
1613 family *fam; | |
1614 { fam->deflink = LIDENT; | |
1615 fam->canlink = LIDENT; | |
1616 fam->vallink = gaus_vallink; | |
1617 fam->family = gaus_fam; | |
1618 fam->pcheck = gaus_check; | |
1619 } | |
1620 /* | |
1621 * Copyright 1996-2006 Catherine Loader. | |
1622 */ | |
1623 #include "locf.h" | |
1624 | |
1625 int geom_vallink(link) | |
1626 int link; | |
1627 { return((link==LIDENT) | (link==LLOG)); | |
1628 } | |
1629 | |
1630 int geom_fam(y,mean,th,link,res,cens,w) | |
1631 double y, mean, th, *res, w; | |
1632 int link, cens; | |
1633 { double p, pt, dp, p1; | |
1634 if (link==LINIT) | |
1635 { res[ZDLL] = MAX(y,0.0); | |
1636 return(LF_OK); | |
1637 } | |
1638 p = 1/(1+mean); | |
1639 if (cens) /* censored observation */ | |
1640 { if (y<=0) | |
1641 { res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0; | |
1642 return(LF_OK); | |
1643 } | |
1644 p1 = (link==LIDENT) ? -p*p : -p*(1-p); | |
1645 pt = 1-ibeta(p,w,y); | |
1646 dp = dbeta(p,w,y,0)/pt; | |
1647 res[ZLIK] = log(pt); | |
1648 res[ZDLL] = -dp*p1; | |
1649 res[ZDDLL] = dp*dp*p1*p1; | |
1650 if (link==LIDENT) | |
1651 res[ZDDLL] += dp*p*p*p*(1+w*(1-p)-p*y)/(1-p); | |
1652 else | |
1653 res[ZDDLL] += dp*p*(1-p)*(w*(1-p)-p*y); | |
1654 return(LF_OK); | |
1655 } | |
1656 else | |
1657 { res[ZLIK] = (y+w)*log((y/w+1)/(mean+1)); | |
1658 if (y>0) res[ZLIK] += y*log(w*mean/y); | |
1659 if (link==LLOG) | |
1660 { res[ZDLL] = (y-w*mean)*p; | |
1661 res[ZDDLL]= (y+w)*p*(1-p); | |
1662 return(LF_OK); | |
1663 } | |
1664 if (link==LIDENT) | |
1665 { res[ZDLL] = (y-w*mean)/(mean*(1+mean)); | |
1666 res[ZDDLL]= w/(mean*(1+mean)); | |
1667 return(LF_OK); | |
1668 } | |
1669 } | |
1670 LERR(("link %d invalid for geometric family",link)); | |
1671 return(LF_LNK); | |
1672 } | |
1673 | |
1674 void setfgeom(fam) | |
1675 family *fam; | |
1676 { fam->deflink = LLOG; | |
1677 fam->canlink = LIDENT; /* this isn't correct. I haven't prog. canon */ | |
1678 fam->vallink = geom_vallink; | |
1679 fam->family = geom_fam; | |
1680 } | |
1681 /* | |
1682 * Copyright 1996-2006 Catherine Loader. | |
1683 */ | |
1684 #include "locf.h" | |
1685 | |
1686 #define HUBERC 2.0 | |
1687 | |
1688 double links_rs; | |
1689 int inllmix=0; | |
1690 | |
1691 /* | |
1692 * lffamily("name") converts family names into a numeric value. | |
1693 * typical usage is fam(&lf->sp) = lffamily("gaussian"); | |
1694 * Note that family can be preceded by q and/or r for quasi, robust. | |
1695 * | |
1696 * link(&lf->sp) = lflink("log") does the same for the link function. | |
1697 */ | |
1698 #define NFAMILY 18 | |
1699 static char *famil[NFAMILY] = | |
1700 { "density", "ate", "hazard", "gaussian", "binomial", | |
1701 "poisson", "gamma", "geometric", "circular", "obust", "huber", | |
1702 "weibull", "cauchy","probab", "logistic", "nbinomial", | |
1703 "vonmises", "quant" }; | |
1704 static int fvals[NFAMILY] = | |
1705 { TDEN, TRAT, THAZ, TGAUS, TLOGT, | |
1706 TPOIS, TGAMM, TGEOM, TCIRC, TROBT, TROBT, | |
1707 TWEIB, TCAUC, TPROB, TLOGT, TGEOM, TCIRC, TQUANT }; | |
1708 int lffamily(z) | |
1709 char *z; | |
1710 { int quasi, robu, f; | |
1711 quasi = robu = 0; | |
1712 while ((z[0]=='q') | (z[0]=='r')) | |
1713 { quasi |= (z[0]=='q'); | |
1714 robu |= (z[0]=='r'); | |
1715 z++; | |
1716 } | |
1717 z[0] = tolower(z[0]); | |
1718 f = pmatch(z,famil,fvals,NFAMILY,-1); | |
1719 if ((z[0]=='o') | (z[0]=='a')) robu = 0; | |
1720 if (f==-1) | |
1721 { WARN(("unknown family %s",z)); | |
1722 f = TGAUS; | |
1723 } | |
1724 if (quasi) f += 64; | |
1725 if (robu) f += 128; | |
1726 return(f); | |
1727 } | |
1728 | |
1729 #define NLINKS 8 | |
1730 static char *ltype[NLINKS] = { "default", "canonical", "identity", "log", | |
1731 "logi", "inverse", "sqrt", "arcsin" }; | |
1732 static int lvals[NLINKS] = { LDEFAU, LCANON, LIDENT, LLOG, | |
1733 LLOGIT, LINVER, LSQRT, LASIN }; | |
1734 int lflink(char *z) | |
1735 { int f; | |
1736 if (z==NULL) return(LDEFAU); | |
1737 z[0] = tolower(z[0]); | |
1738 f = pmatch(z, ltype, lvals, NLINKS, -1); | |
1739 if (f==-1) | |
1740 { WARN(("unknown link %s",z)); | |
1741 f = LDEFAU; | |
1742 } | |
1743 return(f); | |
1744 } | |
1745 | |
1746 int defaultlink(link,fam) | |
1747 int link; | |
1748 family *fam; | |
1749 { if (link==LDEFAU) return(fam->deflink); | |
1750 if (link==LCANON) return(fam->canlink); | |
1751 return(link); | |
1752 } | |
1753 | |
1754 /* | |
1755 void robustify(res,rs) | |
1756 double *res, rs; | |
1757 { double sc, z; | |
1758 sc = rs*HUBERC; | |
1759 if (res[ZLIK] > -sc*sc/2) return; | |
1760 z = sqrt(-2*res[ZLIK]); | |
1761 res[ZDDLL]= -sc*res[ZDLL]*res[ZDLL]/(z*z*z)+sc*res[ZDDLL]/z; | |
1762 res[ZDLL]*= sc/z; | |
1763 res[ZLIK] = sc*sc/2-sc*z; | |
1764 } | |
1765 */ | |
1766 void robustify(res,rs) | |
1767 double *res, rs; | |
1768 { double sc, z; | |
1769 sc = rs*HUBERC; | |
1770 if (res[ZLIK] > -sc*sc/2) | |
1771 { res[ZLIK] /= sc*sc; | |
1772 res[ZDLL] /= sc*sc; | |
1773 res[ZDDLL] /= sc*sc; | |
1774 return; | |
1775 } | |
1776 z = sqrt(-2*res[ZLIK]); | |
1777 res[ZDDLL]= (-sc*res[ZDLL]*res[ZDLL]/(z*z*z)+sc*res[ZDDLL]/z)/(sc*sc); | |
1778 res[ZDLL]*= 1.0/(z*sc); | |
1779 res[ZLIK] = 0.5-z/sc; | |
1780 } | |
1781 | |
1782 double lf_link(y,lin) | |
1783 double y; | |
1784 int lin; | |
1785 { switch(lin) | |
1786 { case LIDENT: return(y); | |
1787 case LLOG: return(log(y)); | |
1788 case LLOGIT: return(logit(y)); | |
1789 case LINVER: return(1/y); | |
1790 case LSQRT: return(sqrt(fabs(y))); | |
1791 case LASIN: return(asin(sqrt(y))); | |
1792 } | |
1793 LERR(("link: unknown link %d",lin)); | |
1794 return(0.0); | |
1795 } | |
1796 | |
1797 double invlink(th,lin) | |
1798 double th; | |
1799 int lin; | |
1800 { switch(lin) | |
1801 { case LIDENT: return(th); | |
1802 case LLOG: return(mut_exp(th)); | |
1803 case LLOGIT: return(expit(th)); | |
1804 case LINVER: return(1/th); | |
1805 case LSQRT: return(th*fabs(th)); | |
1806 case LASIN: return(sin(th)*sin(th)); | |
1807 case LINIT: return(0.0); | |
1808 } | |
1809 LERR(("invlink: unknown link %d",lin)); | |
1810 return(0.0); | |
1811 } | |
1812 | |
1813 /* the link and various related functions */ | |
1814 int links(th,y,fam,link,res,c,w,rs) | |
1815 double th, y, *res, w, rs; | |
1816 int link, c; | |
1817 family *fam; | |
1818 { double mean; | |
1819 int st; | |
1820 | |
1821 mean = res[ZMEAN] = invlink(th,link); | |
1822 if (lf_error) return(LF_LNK); | |
1823 links_rs = rs; | |
1824 /* mut_printf("links: rs %8.5f\n",rs); */ | |
1825 | |
1826 st = fam->family(y,mean,th,link,res,c,w); | |
1827 | |
1828 if (st!=LF_OK) return(st); | |
1829 if (link==LINIT) return(st); | |
1830 if (isrobust(fam)) robustify(res,rs); | |
1831 return(st); | |
1832 } | |
1833 | |
1834 /* | |
1835 stdlinks is a version of links when family, link, response e.t.c | |
1836 all come from the standard places. | |
1837 */ | |
1838 int stdlinks(res,lfd,sp,i,th,rs) | |
1839 lfdata *lfd; | |
1840 smpar *sp; | |
1841 double th, rs, *res; | |
1842 int i; | |
1843 { | |
1844 return(links(th,resp(lfd,i),fami(sp),link(sp),res,cens(lfd,i),prwt(lfd,i),rs)); | |
1845 } | |
1846 | |
1847 /* | |
1848 * functions used in variance, skewness, kurtosis calculations | |
1849 * in scb corrections. | |
1850 */ | |
1851 | |
1852 double b2(th,tg,w) | |
1853 double th, w; | |
1854 int tg; | |
1855 { double y; | |
1856 switch(tg&63) | |
1857 { case TGAUS: return(w); | |
1858 case TPOIS: return(w*mut_exp(th)); | |
1859 case TLOGT: | |
1860 y = expit(th); | |
1861 return(w*y*(1-y)); | |
1862 } | |
1863 LERR(("b2: invalid family %d",tg)); | |
1864 return(0.0); | |
1865 } | |
1866 | |
1867 double b3(th,tg,w) | |
1868 double th, w; | |
1869 int tg; | |
1870 { double y; | |
1871 switch(tg&63) | |
1872 { case TGAUS: return(0.0); | |
1873 case TPOIS: return(w*mut_exp(th)); | |
1874 case TLOGT: | |
1875 y = expit(th); | |
1876 return(w*y*(1-y)*(1-2*y)); | |
1877 } | |
1878 LERR(("b3: invalid family %d",tg)); | |
1879 return(0.0); | |
1880 } | |
1881 | |
1882 double b4(th,tg,w) | |
1883 double th, w; | |
1884 int tg; | |
1885 { double y; | |
1886 switch(tg&63) | |
1887 { case TGAUS: return(0.0); | |
1888 case TPOIS: return(w*mut_exp(th)); | |
1889 case TLOGT: | |
1890 y = expit(th); y = y*(1-y); | |
1891 return(w*y*(1-6*y)); | |
1892 } | |
1893 LERR(("b4: invalid family %d",tg)); | |
1894 return(0.0); | |
1895 } | |
1896 | |
1897 int def_check(sp,des,lfd) | |
1898 smpar *sp; | |
1899 design *des; | |
1900 lfdata *lfd; | |
1901 { switch(link(sp)) | |
1902 { case LLOG: if (des->cf[0]>700) return(LF_OOB); | |
1903 break; | |
1904 } | |
1905 return(LF_OK); | |
1906 } | |
1907 extern void setfdensity(), setfgauss(), setfbino(), setfpoisson(); | |
1908 extern void setfgamma(), setfgeom(), setfcirc(), setfweibull(); | |
1909 extern void setfrbino(), setfrobust(), setfcauchy(), setfquant(); | |
1910 | |
1911 void setfamily(sp) | |
1912 smpar *sp; | |
1913 { int tg, lnk; | |
1914 family *f; | |
1915 | |
1916 tg = fam(sp); | |
1917 f = fami(sp); | |
1918 f->quasi = tg&64; | |
1919 f->robust = tg&128; | |
1920 f->initial = reginit; | |
1921 f->like = likereg; | |
1922 f->pcheck = def_check; | |
1923 | |
1924 switch(tg&63) | |
1925 { case TDEN: | |
1926 case THAZ: | |
1927 case TRAT: setfdensity(f); break; | |
1928 case TGAUS: setfgauss(f); break; | |
1929 case TLOGT: setfbino(f); break; | |
1930 case TRBIN: setfrbino(f); break; | |
1931 case TPROB: | |
1932 case TPOIS: setfpoisson(f); break; | |
1933 case TGAMM: setfgamma(f); break; | |
1934 case TGEOM: setfgeom(f); break; | |
1935 case TWEIB: setfweibull(f); | |
1936 case TCIRC: setfcirc(f); break; | |
1937 case TROBT: setfrobust(f); break; | |
1938 case TCAUC: setfcauchy(f); break; | |
1939 case TQUANT: setfquant(f); break; | |
1940 default: LERR(("setfamily: unknown family %d",tg&63)); | |
1941 return; | |
1942 } | |
1943 | |
1944 lnk = defaultlink(link(sp),f); | |
1945 if (!f->vallink(lnk)) | |
1946 { WARN(("setfamily: invalid link %d - revert to default",link(sp))); | |
1947 link(sp) = f->deflink; | |
1948 } | |
1949 else | |
1950 link(sp) = lnk; | |
1951 } | |
1952 /* | |
1953 * Copyright 1996-2006 Catherine Loader. | |
1954 */ | |
1955 #include "locf.h" | |
1956 | |
1957 int pois_vallink(link) | |
1958 int link; | |
1959 { return((link==LLOG) | (link==LIDENT) | (link==LSQRT)); | |
1960 } | |
1961 | |
1962 int pois_fam(y,mean,th,link,res,cens,w) | |
1963 double y, mean, th, *res, w; | |
1964 int link, cens; | |
1965 { double wmu, pt, dp; | |
1966 if (link==LINIT) | |
1967 { res[ZDLL] = MAX(y,0.0); | |
1968 return(LF_OK); | |
1969 } | |
1970 wmu = w*mean; | |
1971 if (inllmix) y = w*y; | |
1972 if (cens) | |
1973 { if (y<=0) | |
1974 { res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0.0; | |
1975 return(LF_OK); | |
1976 } | |
1977 pt = igamma(wmu,y); | |
1978 dp = dgamma(wmu,y,1.0,0)/pt; | |
1979 res[ZLIK] = log(pt); | |
1980 /* | |
1981 * res[ZDLL] = dp * w*dmu/dth | |
1982 * res[ZDDLL]= -dp*(w*d2mu/dth2 + (y-1)/mu*(dmu/dth)^2) + res[ZDLL]^2 | |
1983 */ | |
1984 if (link==LLOG) | |
1985 { res[ZDLL] = dp*wmu; | |
1986 res[ZDDLL]= -dp*wmu*(y-wmu) + SQR(res[ZDLL]); | |
1987 return(LF_OK); | |
1988 } | |
1989 if (link==LIDENT) | |
1990 { res[ZDLL] = dp*w; | |
1991 res[ZDDLL]= -dp*(y-1-wmu)*w/mean + SQR(res[ZDLL]); | |
1992 return(LF_OK); | |
1993 } | |
1994 if (link==LSQRT) | |
1995 { res[ZDLL] = dp*2*w*th; | |
1996 res[ZDDLL]= -dp*w*(4*y-2-4*wmu) + SQR(res[ZDLL]); | |
1997 return(LF_OK); | |
1998 } } | |
1999 if (link==LLOG) | |
2000 { if (y<0) /* goon observation - delete it */ | |
2001 { res[ZLIK] = res[ZDLL] = res[ZDDLL] = 0; | |
2002 return(LF_OK); | |
2003 } | |
2004 res[ZLIK] = res[ZDLL] = y-wmu; | |
2005 if (y>0) res[ZLIK] += y*(th-log(y/w)); | |
2006 res[ZDDLL] = wmu; | |
2007 return(LF_OK); | |
2008 } | |
2009 if (link==LIDENT) | |
2010 { if ((mean<=0) && (y>0)) return(LF_BADP); | |
2011 res[ZLIK] = y-wmu; | |
2012 res[ZDLL] = -w; | |
2013 res[ZDDLL] = 0; | |
2014 if (y>0) | |
2015 { res[ZLIK] += y*log(wmu/y); | |
2016 res[ZDLL] += y/mean; | |
2017 res[ZDDLL]= y/(mean*mean); | |
2018 } | |
2019 return(LF_OK); | |
2020 } | |
2021 if (link==LSQRT) | |
2022 { if ((mean<=0) && (y>0)) return(LF_BADP); | |
2023 res[ZLIK] = y-wmu; | |
2024 res[ZDLL] = -2*w*th; | |
2025 res[ZDDLL]= 2*w; | |
2026 if (y>0) | |
2027 { res[ZLIK] += y*log(wmu/y); | |
2028 res[ZDLL] += 2*y/th; | |
2029 res[ZDDLL]+= 2*y/mean; | |
2030 } | |
2031 return(LF_OK); | |
2032 } | |
2033 LERR(("link %d invalid for Poisson family",link)); | |
2034 return(LF_LNK); | |
2035 } | |
2036 | |
2037 void setfpoisson(fam) | |
2038 family *fam; | |
2039 { fam->deflink = LLOG; | |
2040 fam->canlink = LLOG; | |
2041 fam->vallink = pois_vallink; | |
2042 fam->family = pois_fam; | |
2043 } | |
2044 /* | |
2045 * Copyright 1996-2006 Catherine Loader. | |
2046 */ | |
2047 #include "locf.h" | |
2048 | |
2049 #define QTOL 1.0e-10 | |
2050 extern int lf_status; | |
2051 static double q0; | |
2052 | |
2053 int quant_vallink(int link) { return(1); } | |
2054 | |
2055 int quant_fam(y,mean,th,link,res,cens,w) | |
2056 double y, mean, th, *res, w; | |
2057 int link, cens; | |
2058 { double z, p; | |
2059 if (link==LINIT) | |
2060 { res[ZDLL] = w*y; | |
2061 return(LF_OK); | |
2062 } | |
2063 p = 0.5; /* should be pen(sp) */ | |
2064 z = y-mean; | |
2065 res[ZLIK] = (z<0) ? (w*z/p) : (-w*z/(1-p)); | |
2066 res[ZDLL] = (z<0) ? -w/p : w/(1-p); | |
2067 res[ZDDLL]= w/(p*(1-p)); | |
2068 return(LF_OK); | |
2069 } | |
2070 | |
2071 int quant_check(sp,des,lfd) | |
2072 smpar *sp; | |
2073 design *des; | |
2074 lfdata *lfd; | |
2075 { return(LF_DONE); | |
2076 } | |
2077 | |
2078 void setfquant(fam) | |
2079 family *fam; | |
2080 { fam->deflink = LIDENT; | |
2081 fam->canlink = LIDENT; | |
2082 fam->vallink = quant_vallink; | |
2083 fam->family = quant_fam; | |
2084 fam->pcheck = quant_check; | |
2085 } | |
2086 | |
2087 /* | |
2088 * cycling rule for choosing among ties. | |
2089 */ | |
2090 int tiecycle(ind,i0,i1,oi) | |
2091 int *ind, i0, i1, oi; | |
2092 { int i, ii, im; | |
2093 im = ind[i0]; | |
2094 for (i=i0+1; i<=i1; i++) | |
2095 { ii = ind[i]; | |
2096 if (im<=oi) | |
2097 { if ((ii<im) | (ii>oi)) im = ii; | |
2098 } | |
2099 else | |
2100 { if ((ii<im) & (ii>oi)) im = ii; | |
2101 } | |
2102 } | |
2103 return(im); | |
2104 } | |
2105 | |
2106 /* | |
2107 * move coefficient vector cf, as far as possible, in direction dc. | |
2108 */ | |
2109 int movecoef(lfd,des,p,cf,dc,oi) | |
2110 lfdata *lfd; | |
2111 design *des; | |
2112 double p, *cf, *dc; | |
2113 int oi; | |
2114 { int i, ii, im, i0, i1, j; | |
2115 double *lb, *el, e, sp, sn, sw, sum1, sum2, tol1; | |
2116 | |
2117 lb = des->th; | |
2118 el = des->res; | |
2119 sum1 = sum2 = 0.0; | |
2120 | |
2121 sp = sn = sw = 0.0; | |
2122 for (i=0; i<des->n; i++) | |
2123 { ii = des->ind[i]; | |
2124 lb[ii] = innerprod(dc,d_xi(des,ii),des->p); | |
2125 e = resp(lfd,ii) - innerprod(cf,d_xi(des,ii),des->p); | |
2126 el[ii] = (fabs(lb[ii])<QTOL) ? 1e100 : e/lb[ii]; | |
2127 if (lb[ii]>0) | |
2128 sp += prwt(lfd,ii)*wght(des,ii)*lb[ii]; | |
2129 else | |
2130 sn -= prwt(lfd,ii)*wght(des,ii)*lb[ii]; | |
2131 sw += prwt(lfd,ii)*wght(des,ii); | |
2132 } | |
2133 printf("sp %8.5f sn %8.5f\n",sn,sp); | |
2134 /* if sn, sp are both zero, should return an LF_PF. | |
2135 * but within numerical tolerance? what does it mean? | |
2136 */ | |
2137 if (sn+sp <= QTOL*q0) { lf_status = LF_PF; return(0); } | |
2138 | |
2139 sum1 = sp/(1-p) + sn/p; | |
2140 tol1 = QTOL*(sp+sn); | |
2141 mut_order(el,des->ind,0,des->n-1); | |
2142 | |
2143 for (i=0; i<des->n; i++) | |
2144 { ii = des->ind[i]; | |
2145 sum2 += prwt(lfd,ii)*wght(des,ii)*((lb[ii]>0) ? lb[ii]/p : -lb[ii]/(1-p) ); | |
2146 sum1 -= prwt(lfd,ii)*wght(des,ii)*((lb[ii]>0) ? lb[ii]/(1-p) : -lb[ii]/p ); | |
2147 if (sum1<=sum2+tol1) | |
2148 { | |
2149 /* determine the range of ties [i0,i1] | |
2150 * el[ind[i0..i1]] = el[ind[i]]. | |
2151 * if sum1==sum2, el[ind[i+1]]..el[ind[i1]]] = el[ind[i1]], else i1 = i. | |
2152 */ | |
2153 i0 = i1 = i; | |
2154 while ((i0>0) && (el[des->ind[i0-1]]==el[ii])) i0--; | |
2155 while ((i1<des->n-1) && (el[des->ind[i1+1]]==el[ii])) i1++; | |
2156 if (sum1>=sum2-tol1) | |
2157 while ((i1<des->n-1) && (el[des->ind[i1+1]]==el[des->ind[i+1]])) i1++; | |
2158 | |
2159 if (i0<i1) ii = tiecycle(des->ind,i0,i1,oi); | |
2160 for (j=0; j<des->p; j++) cf[j] += el[ii]*dc[j]; | |
2161 return(ii); | |
2162 } | |
2163 } | |
2164 mut_printf("Big finddlt problem.\n"); | |
2165 ii = des->ind[des->n-1]; | |
2166 for (j=0; j<des->p; j++) cf[j] += el[ii]*dc[j]; | |
2167 return(ii); | |
2168 } | |
2169 | |
2170 /* | |
2171 * special version of movecoef for min/max. | |
2172 */ | |
2173 int movemin(lfd,des,f,cf,dc,oi) | |
2174 design *des; | |
2175 lfdata *lfd; | |
2176 double *cf, *dc, f; | |
2177 int oi; | |
2178 { int i, ii, im, p, s, ssum; | |
2179 double *lb, sum, lb0, lb1, z0, z1; | |
2180 | |
2181 lb = des->th; | |
2182 s = (f<=0.0) ? 1 : -1; | |
2183 | |
2184 /* first, determine whether move should be in positive or negative direction */ | |
2185 p = des->p; | |
2186 sum = 0; | |
2187 for (i=0; i<des->n; i++) | |
2188 { ii = des->ind[i]; | |
2189 lb[ii] = innerprod(dc,d_xi(des,ii),des->p); | |
2190 sum += prwt(lfd,ii)*wght(des,ii)*lb[ii]; | |
2191 } | |
2192 if (fabs(sum) <= QTOL*q0) | |
2193 { lf_status = LF_PF; | |
2194 return(0); | |
2195 } | |
2196 ssum = (sum<=0.0) ? -1 : 1; | |
2197 if (ssum != s) | |
2198 for (i=0; i<p; i++) dc[i] = -dc[i]; | |
2199 | |
2200 /* now, move positively. How far can we move? */ | |
2201 lb0 = 1.0e100; im = oi; | |
2202 for (i=0; i<des->n; i++) | |
2203 { ii = des->ind[i]; | |
2204 lb[ii] = innerprod(dc,d_xi(des,ii),des->p); /* must recompute - signs! */ | |
2205 if (s*lb[ii]>QTOL) /* should have scale-free tolerance here */ | |
2206 { z0 = innerprod(cf,d_xi(des,ii),p); | |
2207 lb1 = (resp(lfd,ii) - z0)/lb[ii]; | |
2208 if (lb1<lb0) | |
2209 { if (fabs(lb1-lb0)<QTOL) /* cycle */ | |
2210 { if (im<=oi) | |
2211 { if ((ii>oi) | (ii<im)) im = ii; } | |
2212 else | |
2213 { if ((ii>oi) & (ii<im)) im = ii; } | |
2214 } | |
2215 else | |
2216 { im = ii; lb0 = lb1; } | |
2217 } | |
2218 } | |
2219 } | |
2220 | |
2221 for (i=0; i<p; i++) cf[i] = cf[i]+lb0*dc[i]; | |
2222 if (im==-1) lf_status = LF_PF; | |
2223 return(im); | |
2224 } | |
2225 | |
2226 double qll(lfd,spr,des,cf) | |
2227 lfdata *lfd; | |
2228 smpar *spr; | |
2229 design *des; | |
2230 double *cf; | |
2231 { int i, ii; | |
2232 double th, sp, sn, p, e; | |
2233 | |
2234 p = pen(spr); | |
2235 sp = sn = 0.0; | |
2236 for (i=0; i<des->n; i++) | |
2237 { ii = des->ind[i]; | |
2238 th = innerprod(d_xi(des,ii),cf,des->p); | |
2239 e = resp(lfd,ii)-th; | |
2240 if (e<0) sn -= prwt(lfd,ii)*wght(des,ii)*e; | |
2241 if (e>0) sp += prwt(lfd,ii)*wght(des,ii)*e; | |
2242 } | |
2243 if (p<=0.0) return((sn<QTOL) ? -sp : -1e300); | |
2244 if (p>=1.0) return((sp<QTOL) ? -sn : -1e300); | |
2245 return(-sp/(1-p)-sn/p); | |
2246 } | |
2247 | |
2248 /* | |
2249 * running quantile smoother. | |
2250 */ | |
2251 void lfquantile(lfd,sp,des,maxit) | |
2252 lfdata *lfd; | |
2253 smpar *sp; | |
2254 design *des; | |
2255 int maxit; | |
2256 { int i, ii, im, j, k, p, *ci, (*mover)(); | |
2257 double *cf, *db, *dc, *cm, f, q1, q2, l0; | |
2258 | |
2259 printf("in lfquantile\n"); | |
2260 f = pen(sp); | |
2261 p = des->p; | |
2262 cf = des->cf; | |
2263 dc = des->oc; | |
2264 db = des->ss; | |
2265 setzero(cf,p); | |
2266 setzero(dc,p); | |
2267 cm = des->V; | |
2268 setzero(cm,p*p); | |
2269 ci = (int *)des->fix; | |
2270 | |
2271 q1 = -qll(lfd,sp,des,cf); | |
2272 if (q1==0.0) { lf_status = LF_PF; return; } | |
2273 for (i=0; i<p; i++) cm[i*(p+1)] = 1; | |
2274 mover = movecoef; | |
2275 if ((f<=0.0) | (f>=1.0)) mover = movemin; | |
2276 | |
2277 dc[0] = 1.0; | |
2278 im = mover(lfd,des,f,cf,dc,-1); | |
2279 if (lf_status != LF_OK) return; | |
2280 ci[0] = im; | |
2281 printf("init const %2d\n",ci[0]); | |
2282 q0 = -qll(lfd,sp,des,cf); | |
2283 if (q0<QTOL*q1) { lf_status = LF_PF; return; } | |
2284 | |
2285 printf("loop 0\n"); fflush(stdout); | |
2286 for (i=1; i<p; i++) | |
2287 { | |
2288 printf("i %2d\n",i); | |
2289 memcpy(&cm[(i-1)*p],d_xi(des,im),p*sizeof(double)); | |
2290 setzero(db,p); | |
2291 db[i] = 1.0; | |
2292 resproj(db,cm,dc,p,i); | |
2293 printf("call mover\n"); fflush(stdout); | |
2294 im = mover(lfd,des,f,cf,dc,-1); | |
2295 if (lf_status != LF_OK) return; | |
2296 printf("mover %2d\n",im); fflush(stdout); | |
2297 ci[i] = im; | |
2298 } | |
2299 printf("call qll\n"); fflush(stdout); | |
2300 q1 = qll(lfd,sp,des,cf); | |
2301 | |
2302 printf("loop 1 %d %d %d %d\n",ci[0],ci[1],ci[2],ci[3]); fflush(stdout); | |
2303 for (k=0; k<maxit; k++) | |
2304 { for (i=0; i<p; i++) | |
2305 { for (j=0; j<p; j++) | |
2306 if (j!=i) memcpy(&cm[(j-(j>i))*p],d_xi(des,ci[j]),p*sizeof(double)); | |
2307 memcpy(db,d_xi(des,ci[i]),p*sizeof(double)); | |
2308 resproj(db,cm,dc,p,p-1); | |
2309 printf("call mover\n"); fflush(stdout); | |
2310 im = mover(lfd,des,f,cf,dc,ci[i]); | |
2311 if (lf_status != LF_OK) return; | |
2312 printf("mover %2d\n",im); fflush(stdout); | |
2313 ci[i] = im; | |
2314 } | |
2315 q2 = qll(lfd,sp,des,cf); | |
2316 /* | |
2317 * convergence: require no change -- reasonable, since discrete? | |
2318 * remember we're maximizing, and q's are negative. | |
2319 */ | |
2320 if (q2 <= q1) return; | |
2321 q1 = q2; | |
2322 } | |
2323 printf("loop 2\n"); | |
2324 mut_printf("Warning: lfquantile not converged.\n"); | |
2325 } | |
2326 /* | |
2327 * Copyright 1996-2006 Catherine Loader. | |
2328 */ | |
2329 #include "locf.h" | |
2330 | |
2331 extern double links_rs; | |
2332 | |
2333 int robust_vallink(link) | |
2334 int link; | |
2335 { return(link==LIDENT); | |
2336 } | |
2337 | |
2338 int robust_fam(y,mean,th,link,res,cens,w) | |
2339 double y, mean, th, *res, w; | |
2340 int link, cens; | |
2341 { double z, sw; | |
2342 if (link==LINIT) | |
2343 { res[ZDLL] = w*y; | |
2344 return(LF_OK); | |
2345 } | |
2346 sw = (w==1.0) ? 1.0 : sqrt(w); /* don't want unnecess. sqrt! */ | |
2347 z = sw*(y-mean)/links_rs; | |
2348 res[ZLIK] = (fabs(z)<HUBERC) ? -z*z/2 : HUBERC*(HUBERC/2.0-fabs(z)); | |
2349 if (z< -HUBERC) | |
2350 { res[ZDLL] = -sw*HUBERC/links_rs; | |
2351 res[ZDDLL]= 0.0; | |
2352 return(LF_OK); | |
2353 } | |
2354 if (z> HUBERC) | |
2355 { res[ZDLL] = sw*HUBERC/links_rs; | |
2356 res[ZDDLL]= 0.0; | |
2357 return(LF_OK); | |
2358 } | |
2359 res[ZDLL] = sw*z/links_rs; | |
2360 res[ZDDLL] = w/(links_rs*links_rs); | |
2361 return(LF_OK); | |
2362 } | |
2363 | |
2364 int cauchy_fam(y,p,th,link,res,cens,w) | |
2365 double y, p, th, *res, w; | |
2366 int link, cens; | |
2367 { double z; | |
2368 if (link!=LIDENT) | |
2369 { LERR(("Invalid link in famcauc")); | |
2370 return(LF_LNK); | |
2371 } | |
2372 z = w*(y-th)/links_rs; | |
2373 res[ZLIK] = -log(1+z*z); | |
2374 res[ZDLL] = 2*w*z/(links_rs*(1+z*z)); | |
2375 res[ZDDLL] = 2*w*w*(1-z*z)/(links_rs*links_rs*(1+z*z)*(1+z*z)); | |
2376 return(LF_OK); | |
2377 } | |
2378 | |
2379 extern double lf_tol; | |
2380 int robust_init(lfd,des,sp) | |
2381 lfdata *lfd; | |
2382 design *des; | |
2383 smpar *sp; | |
2384 { int i; | |
2385 for (i=0; i<des->n; i++) | |
2386 des->res[i] = resp(lfd,(int)des->ind[i]) - base(lfd,(int)des->ind[i]); | |
2387 des->cf[0] = median(des->res,des->n); | |
2388 for (i=1; i<des->p; i++) des->cf[i] = 0.0; | |
2389 lf_tol = 1.0e-6; | |
2390 return(LF_OK); | |
2391 } | |
2392 | |
2393 void setfrobust(fam) | |
2394 family *fam; | |
2395 { fam->deflink = LIDENT; | |
2396 fam->canlink = LIDENT; | |
2397 fam->vallink = robust_vallink; | |
2398 fam->family = robust_fam; | |
2399 fam->initial = robust_init; | |
2400 fam->robust = 0; | |
2401 } | |
2402 | |
2403 void setfcauchy(fam) | |
2404 family *fam; | |
2405 { fam->deflink = LIDENT; | |
2406 fam->canlink = LIDENT; | |
2407 fam->vallink = robust_vallink; | |
2408 fam->family = cauchy_fam; | |
2409 fam->initial = robust_init; | |
2410 fam->robust = 0; | |
2411 } | |
2412 /* | |
2413 * Copyright 1996-2006 Catherine Loader. | |
2414 */ | |
2415 #include "locf.h" | |
2416 | |
2417 int weibull_vallink(link) | |
2418 int link; | |
2419 { return((link==LIDENT) | (link==LLOG) | (link==LLOGIT)); | |
2420 } | |
2421 | |
2422 int weibull_fam(y,mean,th,link,res,cens,w) | |
2423 double y, mean, th, *res, w; | |
2424 int link, cens; | |
2425 { double yy; | |
2426 yy = pow(y,w); | |
2427 if (link==LINIT) | |
2428 { res[ZDLL] = MAX(yy,0.0); | |
2429 return(LF_OK); | |
2430 } | |
2431 if (cens) | |
2432 { res[ZLIK] = -yy/mean; | |
2433 res[ZDLL] = res[ZDDLL] = yy/mean; | |
2434 return(LF_OK); | |
2435 } | |
2436 res[ZLIK] = 1-yy/mean-th; | |
2437 if (yy>0) res[ZLIK] += log(w*yy); | |
2438 res[ZDLL] = -1+yy/mean; | |
2439 res[ZDDLL]= yy/mean; | |
2440 return(LF_OK); | |
2441 } | |
2442 | |
2443 void setfweibull(fam) | |
2444 family *fam; | |
2445 { fam->deflink = LLOG; | |
2446 fam->canlink = LLOG; | |
2447 fam->vallink = weibull_vallink; | |
2448 fam->family = weibull_fam; | |
2449 fam->robust = 0; | |
2450 } | |
2451 /* | |
2452 * Copyright 1996-2006 Catherine Loader. | |
2453 */ | |
2454 /* | |
2455 Functions implementing the adaptive bandwidth selection. | |
2456 Will make the final call to nbhd() to set smoothing weights | |
2457 for selected bandwidth, But will **not** make the | |
2458 final call to locfit(). | |
2459 */ | |
2460 | |
2461 #include "locf.h" | |
2462 | |
2463 static double hmin; | |
2464 | |
2465 #define NACRI 5 | |
2466 static char *atype[NACRI] = { "none", "cp", "ici", "mindex", "ok" }; | |
2467 static int avals[NACRI] = { ANONE, ACP, AKAT, AMDI, AOK }; | |
2468 int lfacri(char *z) | |
2469 { return(pmatch(z, atype, avals, NACRI, ANONE)); | |
2470 } | |
2471 | |
2472 double adcri(lk,t0,t2,pen) | |
2473 double lk, t0, t2, pen; | |
2474 { double y; | |
2475 /* return(-2*lk/(t0*exp(pen*log(1-t2/t0)))); */ | |
2476 /* return((-2*lk+pen*t2)/t0); */ | |
2477 y = (MAX(-2*lk,t0-t2)+pen*t2)/t0; | |
2478 return(y); | |
2479 } | |
2480 | |
2481 double mmse(lfd,sp,dv,des) | |
2482 lfdata *lfd; | |
2483 smpar *sp; | |
2484 deriv *dv; | |
2485 design *des; | |
2486 { int i, ii, j, p, p1; | |
2487 double sv, sb, *l, dp; | |
2488 | |
2489 l = des->wd; | |
2490 wdiag(lfd, sp, des,l,dv,0,1,0); | |
2491 sv = sb = 0; | |
2492 p = npar(sp); | |
2493 for (i=0; i<des->n; i++) | |
2494 { sv += l[i]*l[i]; | |
2495 ii = des->ind[i]; | |
2496 dp = dist(des,ii); | |
2497 for (j=0; j<deg(sp); j++) dp *= dist(des,ii); | |
2498 sb += fabs(l[i])*dp; | |
2499 } | |
2500 p1 = factorial(deg(sp)+1); | |
2501 printf("%8.5f sv %8.5f sb %8.5f %8.5f\n",des->h,sv,sb,sv+sb*sb*pen(sp)*pen(sp)/(p1*p1)); | |
2502 return(sv+sb*sb*pen(sp)*pen(sp)/(p1*p1)); | |
2503 } | |
2504 | |
2505 static double mcp, clo, cup; | |
2506 | |
2507 /* | |
2508 Initial bandwidth will be (by default) | |
2509 k-nearest neighbors for k small, just large enough to | |
2510 get defined estimate (unless user provided nonzero nn or fix-h components) | |
2511 */ | |
2512 | |
2513 int ainitband(lfd,sp,dv,des) | |
2514 lfdata *lfd; | |
2515 smpar *sp; | |
2516 deriv *dv; | |
2517 design *des; | |
2518 { int lf_status, p, z, cri, noit, redo; | |
2519 double ho, t[6]; | |
2520 | |
2521 if (lf_debug >= 2) mut_printf("ainitband:\n"); | |
2522 p = des->p; | |
2523 cri = acri(sp); | |
2524 noit = (cri!=AOK); | |
2525 z = (int)(lfd->n*nn(sp)); | |
2526 if ((noit) && (z<p+2)) z = p+2; | |
2527 redo = 0; ho = -1; | |
2528 do | |
2529 { | |
2530 nbhd(lfd,des,z,redo,sp); | |
2531 if (z<des->n) z = des->n; | |
2532 if (des->h>ho) lf_status = locfit(lfd,des,sp,noit,0,0); | |
2533 z++; | |
2534 redo = 1; | |
2535 } while ((z<=lfd->n) && ((des->h==0)||(lf_status!=LF_OK))); | |
2536 hmin = des->h; | |
2537 | |
2538 switch(cri) | |
2539 { case ACP: | |
2540 local_df(lfd,sp,des,t); | |
2541 mcp = adcri(des->llk,t[0],t[2],pen(sp)); | |
2542 return(lf_status); | |
2543 case AKAT: | |
2544 local_df(lfd,sp,des,t); | |
2545 clo = des->cf[0]-pen(sp)*t[5]; | |
2546 cup = des->cf[0]+pen(sp)*t[5]; | |
2547 return(lf_status); | |
2548 case AMDI: | |
2549 mcp = mmse(lfd,sp,dv,des); | |
2550 return(lf_status); | |
2551 case AOK: return(lf_status); | |
2552 } | |
2553 LERR(("aband1: unknown criterion")); | |
2554 return(LF_ERR); | |
2555 } | |
2556 | |
2557 /* | |
2558 aband2 increases the initial bandwidth until lack of fit results, | |
2559 or the fit is close to a global fit. Increase h by 1+0.3/d at | |
2560 each iteration. | |
2561 */ | |
2562 | |
2563 double aband2(lfd,sp,dv,des,h0) | |
2564 lfdata *lfd; | |
2565 smpar *sp; | |
2566 deriv *dv; | |
2567 design *des; | |
2568 double h0; | |
2569 { double t[6], h1, nu1, cp, ncp, tlo, tup; | |
2570 int d, inc, n, p, done; | |
2571 | |
2572 if (lf_debug >= 2) mut_printf("aband2:\n"); | |
2573 d = lfd->d; n = lfd->n; p = npar(sp); | |
2574 h1 = des->h = h0; | |
2575 done = 0; nu1 = 0.0; | |
2576 inc = 0; ncp = 0.0; | |
2577 while ((!done) & (nu1<(n-p)*0.95)) | |
2578 { fixh(sp) = (1+0.3/d)*des->h; | |
2579 nbhd(lfd,des,0,1,sp); | |
2580 if (locfit(lfd,des,sp,1,0,0) > 0) WARN(("aband2: failed fit")); | |
2581 local_df(lfd,sp,des,t); | |
2582 nu1 = t[0]-t[2]; /* tr(A) */ | |
2583 switch(acri(sp)) | |
2584 { case AKAT: | |
2585 tlo = des->cf[0]-pen(sp)*t[5]; | |
2586 tup = des->cf[0]+pen(sp)*t[5]; | |
2587 /* mut_printf("h %8.5f tlo %8.5f tup %8.5f\n",des->h,tlo,tup); */ | |
2588 done = ((tlo>cup) | (tup<clo)); | |
2589 if (!done) | |
2590 { clo = MAX(clo,tlo); | |
2591 cup = MIN(cup,tup); | |
2592 h1 = des->h; | |
2593 } | |
2594 break; | |
2595 case ACP: | |
2596 cp = adcri(des->llk,t[0],t[2],pen(sp)); | |
2597 /* mut_printf("h %8.5f lk %8.5f t0 %8.5f t2 %8.5f cp %8.5f\n",des->h,des->llk,t[0],t[2],cp); */ | |
2598 if (cp<mcp) { mcp = cp; h1 = des->h; } | |
2599 if (cp>=ncp) inc++; else inc = 0; | |
2600 ncp = cp; | |
2601 done = (inc>=10) | ((inc>=3) & ((t[0]-t[2])>=10) & (cp>1.5*mcp)); | |
2602 break; | |
2603 case AMDI: | |
2604 cp = mmse(lfd,sp,dv,des); | |
2605 if (cp<mcp) { mcp = cp; h1 = des->h; } | |
2606 if (cp>ncp) inc++; else inc = 0; | |
2607 ncp = cp; | |
2608 done = (inc>=3); | |
2609 break; | |
2610 } | |
2611 } | |
2612 return(h1); | |
2613 } | |
2614 | |
2615 /* | |
2616 aband3 does a finer search around best h so far. Try | |
2617 h*(1-0.2/d), h/(1-0.1/d), h*(1+0.1/d), h*(1+0.2/d) | |
2618 */ | |
2619 double aband3(lfd,sp,dv,des,h0) | |
2620 lfdata *lfd; | |
2621 smpar *sp; | |
2622 deriv *dv; | |
2623 design *des; | |
2624 double h0; | |
2625 { double t[6], h1, cp, tlo, tup; | |
2626 int i, i0, d, n; | |
2627 | |
2628 if (lf_debug >= 2) mut_printf("aband3:\n"); | |
2629 d = lfd->d; n = lfd->n; | |
2630 h1 = h0; | |
2631 i0 = (acri(sp)==AKAT) ? 1 : -2; | |
2632 if (h0==hmin) i0 = 1; | |
2633 | |
2634 for (i=i0; i<=2; i++) | |
2635 { if (i==0) i++; | |
2636 fixh(sp) = h0*(1+0.1*i/d); | |
2637 nbhd(lfd,des,0,1,sp); | |
2638 if (locfit(lfd,des,sp,1,0,0) > 0) WARN(("aband3: failed fit")); | |
2639 local_df(lfd,sp,des,t); | |
2640 switch (acri(sp)) | |
2641 { case AKAT: | |
2642 tlo = des->cf[0]-pen(sp)*t[5]; | |
2643 tup = des->cf[0]+pen(sp)*t[5]; | |
2644 if ((tlo>cup) | (tup<clo)) /* done */ | |
2645 i = 2; | |
2646 else | |
2647 { h1 = des->h; | |
2648 clo = MAX(clo,tlo); | |
2649 cup = MIN(cup,tup); | |
2650 } | |
2651 break; | |
2652 case ACP: | |
2653 cp = adcri(des->llk,t[0],t[2],pen(sp)); | |
2654 if (cp<mcp) { mcp = cp; h1 = des->h; } | |
2655 else | |
2656 { if (i>0) i = 2; } | |
2657 break; | |
2658 case AMDI: | |
2659 cp = mmse(lfd,sp,dv,des); | |
2660 if (cp<mcp) { mcp = cp; h1 = des->h; } | |
2661 else | |
2662 { if (i>0) i = 2; } | |
2663 } | |
2664 } | |
2665 return(h1); | |
2666 } | |
2667 | |
2668 int alocfit(lfd,sp,dv,des,cv) | |
2669 lfdata *lfd; | |
2670 smpar *sp; | |
2671 deriv *dv; | |
2672 design *des; | |
2673 int cv; | |
2674 { int lf_status; | |
2675 double h0; | |
2676 | |
2677 lf_status = ainitband(lfd,sp,dv,des); | |
2678 if (lf_error) return(lf_status); | |
2679 if (acri(sp) == AOK) return(lf_status); | |
2680 | |
2681 h0 = fixh(sp); | |
2682 fixh(sp) = aband2(lfd,sp,dv,des,des->h); | |
2683 fixh(sp) = aband3(lfd,sp,dv,des,fixh(sp)); | |
2684 nbhd(lfd,des,0,1,sp); | |
2685 lf_status = locfit(lfd,des,sp,0,0,cv); | |
2686 fixh(sp) = h0; | |
2687 | |
2688 return(lf_status); | |
2689 } | |
2690 /* | |
2691 * Copyright 1996-2006 Catherine Loader. | |
2692 */ | |
2693 /* | |
2694 * | |
2695 * Evaluate the locfit fitting functions. | |
2696 * calcp(sp,d) | |
2697 * calculates the number of fitting functions. | |
2698 * makecfn(sp,des,dv,d) | |
2699 * makes the coef.number vector. | |
2700 * fitfun(lfd, sp, x,t,f,dv) | |
2701 * lfd is the local fit structure. | |
2702 * sp smoothing parameter structure. | |
2703 * x is the data point. | |
2704 * t is the fitting point. | |
2705 * f is a vector to return the results. | |
2706 * dv derivative structure. | |
2707 * designmatrix(lfd, sp, des) | |
2708 * is a wrapper for fitfun to build the design matrix. | |
2709 * | |
2710 */ | |
2711 | |
2712 #include "locf.h" | |
2713 | |
2714 int calcp(sp,d) | |
2715 smpar *sp; | |
2716 int d; | |
2717 { int i, k; | |
2718 | |
2719 if (ubas(sp)) return(npar(sp)); | |
2720 | |
2721 switch (kt(sp)) | |
2722 { case KSPH: | |
2723 case KCE: | |
2724 k = 1; | |
2725 for (i=1; i<=deg(sp); i++) k = k*(d+i)/i; | |
2726 return(k); | |
2727 case KPROD: return(d*deg(sp)+1); | |
2728 case KLM: return(d); | |
2729 case KZEON: return(1); | |
2730 } | |
2731 LERR(("calcp: invalid kt %d",kt(sp))); | |
2732 return(0); | |
2733 } | |
2734 | |
2735 int coefnumber(dv,kt,d,deg) | |
2736 int kt, d, deg; | |
2737 deriv *dv; | |
2738 { int d0, d1, t; | |
2739 | |
2740 if (d==1) | |
2741 { if (dv->nd<=deg) return(dv->nd); | |
2742 return(-1); | |
2743 } | |
2744 | |
2745 if (dv->nd==0) return(0); | |
2746 if (deg==0) return(-1); | |
2747 if (dv->nd==1) return(1+dv->deriv[0]); | |
2748 if (deg==1) return(-1); | |
2749 if (kt==KPROD) return(-1); | |
2750 | |
2751 if (dv->nd==2) | |
2752 { d0 = dv->deriv[0]; d1 = dv->deriv[1]; | |
2753 if (d0<d1) { t = d0; d0 = d1; d1 = t; } | |
2754 return((d+1)*(d0+1)-d0*(d0+3)/2+d1); | |
2755 } | |
2756 if (deg==2) return(-1); | |
2757 | |
2758 LERR(("coefnumber not programmed for nd>=3")); | |
2759 return(-1); | |
2760 } | |
2761 | |
2762 void makecfn(sp,des,dv,d) | |
2763 smpar *sp; | |
2764 design *des; | |
2765 deriv *dv; | |
2766 int d; | |
2767 { int i, nd; | |
2768 | |
2769 nd = dv->nd; | |
2770 | |
2771 des->cfn[0] = coefnumber(dv,kt(sp),d,deg(sp)); | |
2772 des->ncoef = 1; | |
2773 if (nd >= deg(sp)) return; | |
2774 if (kt(sp)==KZEON) return; | |
2775 | |
2776 if (d>1) | |
2777 { if (nd>=2) return; | |
2778 if ((nd>=1) && (kt(sp)==KPROD)) return; | |
2779 } | |
2780 | |
2781 dv->nd = nd+1; | |
2782 for (i=0; i<d; i++) | |
2783 { dv->deriv[nd] = i; | |
2784 des->cfn[i+1] = coefnumber(dv,kt(sp),d,deg(sp)); | |
2785 } | |
2786 dv->nd = nd; | |
2787 | |
2788 des->ncoef = 1+d; | |
2789 } | |
2790 | |
2791 void fitfunangl(dx,ff,sca,cd,deg) | |
2792 double dx, *ff, sca; | |
2793 int deg, cd; | |
2794 { | |
2795 if (deg>=3) WARN(("Can't handle angular model with deg>=3")); | |
2796 | |
2797 switch(cd) | |
2798 { case 0: | |
2799 ff[0] = 1; | |
2800 ff[1] = sin(dx/sca)*sca; | |
2801 ff[2] = (1-cos(dx/sca))*sca*sca; | |
2802 return; | |
2803 case 1: | |
2804 ff[0] = 0; | |
2805 ff[1] = cos(dx/sca); | |
2806 ff[2] = sin(dx/sca)*sca; | |
2807 return; | |
2808 case 2: | |
2809 ff[0] = 0; | |
2810 ff[1] = -sin(dx/sca)/sca; | |
2811 ff[2] = cos(dx/sca); | |
2812 return; | |
2813 default: WARN(("Can't handle angular model with >2 derivs")); | |
2814 } | |
2815 } | |
2816 | |
2817 void fitfun(lfd,sp,x,t,f,dv) | |
2818 lfdata *lfd; | |
2819 smpar *sp; | |
2820 double *x, *t, *f; | |
2821 deriv *dv; | |
2822 { int d, deg, nd, m, i, j, k, ct_deriv[MXDIM]; | |
2823 double ff[MXDIM][1+MXDEG], dx[MXDIM], *xx[MXDIM]; | |
2824 | |
2825 if (ubas(sp)) | |
2826 { for (i=0; i<lfd->d; i++) xx[i] = &x[i]; | |
2827 i = 0; | |
2828 sp->vbasis(xx,t,1,lfd->d,1,npar(sp),f); | |
2829 return; | |
2830 } | |
2831 | |
2832 d = lfd->d; | |
2833 deg = deg(sp); | |
2834 m = 0; | |
2835 nd = (dv==NULL) ? 0 : dv->nd; | |
2836 | |
2837 if (kt(sp)==KZEON) | |
2838 { f[0] = 1.0; | |
2839 return; | |
2840 } | |
2841 | |
2842 if (kt(sp)==KLM) | |
2843 { for (i=0; i<d; i++) f[m++] = x[i]; | |
2844 return; | |
2845 } | |
2846 | |
2847 f[m++] = (nd==0); | |
2848 if (deg==0) return; | |
2849 | |
2850 for (i=0; i<d; i++) | |
2851 { ct_deriv[i] = 0; | |
2852 dx[i] = (t==NULL) ? x[i] : x[i]-t[i]; | |
2853 } | |
2854 for (i=0; i<nd; i++) ct_deriv[dv->deriv[i]]++; | |
2855 | |
2856 for (i=0; i<d; i++) | |
2857 { switch(lfd->sty[i]) | |
2858 { | |
2859 case STANGL: | |
2860 fitfunangl(dx[i],ff[i],lfd->sca[i],ct_deriv[i],deg(sp)); | |
2861 break; | |
2862 default: | |
2863 for (j=0; j<ct_deriv[i]; j++) ff[i][j] = 0.0; | |
2864 ff[i][ct_deriv[i]] = 1.0; | |
2865 for (j=ct_deriv[i]+1; j<=deg; j++) | |
2866 ff[i][j] = ff[i][j-1]*dx[i]/(j-ct_deriv[i]); | |
2867 } | |
2868 } | |
2869 | |
2870 /* | |
2871 * Product kernels. Note that if ct_deriv[i] != nd, that implies | |
2872 * there is differentiation wrt another variable, and all components | |
2873 * involving x[i] are 0. | |
2874 */ | |
2875 if ((d==1) || (kt(sp)==KPROD)) | |
2876 { for (j=1; j<=deg; j++) | |
2877 for (i=0; i<d; i++) | |
2878 f[m++] = (ct_deriv[i]==nd) ? ff[i][j] : 0.0; | |
2879 return; | |
2880 } | |
2881 | |
2882 /* | |
2883 * Spherical kernels with the full polynomial basis. | |
2884 * Presently implemented up to deg=3. | |
2885 */ | |
2886 for (i=0; i<d; i++) | |
2887 f[m++] = (ct_deriv[i]==nd) ? ff[i][1] : 0.0; | |
2888 if (deg==1) return; | |
2889 | |
2890 for (i=0; i<d; i++) | |
2891 { | |
2892 /* xi^2/2 terms. */ | |
2893 f[m++] = (ct_deriv[i]==nd) ? ff[i][2] : 0.0; | |
2894 | |
2895 /* xi xj terms */ | |
2896 for (j=i+1; j<d; j++) | |
2897 f[m++] = (ct_deriv[i]+ct_deriv[j]==nd) ? ff[i][1]*ff[j][1] : 0.0; | |
2898 } | |
2899 if (deg==2) return; | |
2900 | |
2901 for (i=0; i<d; i++) | |
2902 { | |
2903 /* xi^3/6 terms */ | |
2904 f[m++] = (ct_deriv[i]==nd) ? ff[i][3] : 0.0; | |
2905 | |
2906 /* xi^2/2 xk terms */ | |
2907 for (k=i+1; k<d; k++) | |
2908 f[m++] = (ct_deriv[i]+ct_deriv[k]==nd) ? ff[i][2]*ff[k][1] : 0.0; | |
2909 | |
2910 /* xi xj xk terms */ | |
2911 for (j=i+1; j<d; j++) | |
2912 { f[m++] = (ct_deriv[i]+ct_deriv[j]==nd) ? ff[i][1]*ff[j][2] : 0.0; | |
2913 for (k=j+1; k<d; k++) | |
2914 f[m++] = (ct_deriv[i]+ct_deriv[j]+ct_deriv[k]==nd) ? | |
2915 ff[i][1]*ff[j][1]*ff[k][1] : 0.0; | |
2916 } | |
2917 } | |
2918 if (deg==3) return; | |
2919 | |
2920 LERR(("fitfun: can't handle deg=%d for spherical kernels",deg)); | |
2921 } | |
2922 | |
2923 /* | |
2924 * Build the design matrix. Assumes des->ind contains the indices of | |
2925 * the required data points; des->n the number of points; des->xev | |
2926 * the fitting point. | |
2927 */ | |
2928 void designmatrix(lfd,sp,des) | |
2929 lfdata *lfd; | |
2930 smpar *sp; | |
2931 design *des; | |
2932 { int i, ii, j, p; | |
2933 double *X, u[MXDIM]; | |
2934 | |
2935 X = d_x(des); | |
2936 p = des->p; | |
2937 | |
2938 if (ubas(sp)) | |
2939 { | |
2940 sp->vbasis(lfd->x,des->xev,lfd->n,lfd->d,des->n,p,X); | |
2941 return; | |
2942 } | |
2943 | |
2944 for (i=0; i<des->n; i++) | |
2945 { ii = des->ind[i]; | |
2946 for (j=0; j<lfd->d; j++) u[j] = datum(lfd,j,ii); | |
2947 fitfun(lfd,sp,u,des->xev,&X[ii*p],NULL); | |
2948 } | |
2949 } | |
2950 /* | |
2951 * Copyright 1996-2006 Catherine Loader. | |
2952 */ | |
2953 /* | |
2954 * | |
2955 * | |
2956 * Functions for determining bandwidth; smoothing neighborhood | |
2957 * and smoothing weights. | |
2958 */ | |
2959 | |
2960 #include "locf.h" | |
2961 | |
2962 double rho(x,sc,d,kt,sty) /* ||x|| for appropriate distance metric */ | |
2963 double *x, *sc; | |
2964 int d, kt, *sty; | |
2965 { double rhoi[MXDIM], s; | |
2966 int i; | |
2967 for (i=0; i<d; i++) | |
2968 { if (sty!=NULL) | |
2969 { switch(sty[i]) | |
2970 { case STANGL: rhoi[i] = 2*sin(x[i]/(2*sc[i])); break; | |
2971 case STCPAR: rhoi[i] = 0; break; | |
2972 default: rhoi[i] = x[i]/sc[i]; | |
2973 } } | |
2974 else rhoi[i] = x[i]/sc[i]; | |
2975 } | |
2976 | |
2977 if (d==1) return(fabs(rhoi[0])); | |
2978 | |
2979 s = 0; | |
2980 if (kt==KPROD) | |
2981 { for (i=0; i<d; i++) | |
2982 { rhoi[i] = fabs(rhoi[i]); | |
2983 if (rhoi[i]>s) s = rhoi[i]; | |
2984 } | |
2985 return(s); | |
2986 } | |
2987 | |
2988 if (kt==KSPH) | |
2989 { for (i=0; i<d; i++) | |
2990 s += rhoi[i]*rhoi[i]; | |
2991 return(sqrt(s)); | |
2992 } | |
2993 | |
2994 LERR(("rho: invalid kt")); | |
2995 return(0.0); | |
2996 } | |
2997 | |
2998 double kordstat(x,k,n,ind) | |
2999 double *x; | |
3000 int k, n, *ind; | |
3001 { int i, i0, i1, l, r; | |
3002 double piv; | |
3003 if (k<1) return(0.0); | |
3004 i0 = 0; i1 = n-1; | |
3005 while (1) | |
3006 { piv = x[ind[(i0+i1)/2]]; | |
3007 l = i0; r = i1; | |
3008 while (l<=r) | |
3009 { while ((l<=i1) && (x[ind[l]]<=piv)) l++; | |
3010 while ((r>=i0) && (x[ind[r]]>piv)) r--; | |
3011 if (l<=r) ISWAP(ind[l],ind[r]); | |
3012 } /* now, x[ind[i0..r]] <= piv < x[ind[l..i1]] */ | |
3013 if (r<k-1) i0 = l; /* go right */ | |
3014 else /* put pivots in middle */ | |
3015 { for (i=i0; i<=r; ) | |
3016 if (x[ind[i]]==piv) { ISWAP(ind[i],ind[r]); r--; } | |
3017 else i++; | |
3018 if (r<k-1) return(piv); | |
3019 i1 = r; | |
3020 } | |
3021 } | |
3022 } | |
3023 | |
3024 /* check if i'th data point is in limits */ | |
3025 int inlim(lfd,i) | |
3026 lfdata *lfd; | |
3027 int i; | |
3028 { int d, j, k; | |
3029 double *xlim; | |
3030 | |
3031 xlim = lfd->xl; | |
3032 d = lfd->d; | |
3033 k = 1; | |
3034 for (j=0; j<d; j++) | |
3035 { if (xlim[j]<xlim[j+d]) | |
3036 k &= ((datum(lfd,j,i)>=xlim[j]) & (datum(lfd,j,i)<=xlim[j+d])); | |
3037 } | |
3038 return(k); | |
3039 } | |
3040 | |
3041 double compbandwid(di,ind,x,n,d,nn,fxh) | |
3042 double *di, *x, fxh; | |
3043 int n, d, nn, *ind; | |
3044 { int i; | |
3045 double nnh; | |
3046 | |
3047 if (nn==0) return(fxh); | |
3048 | |
3049 if (nn<n) | |
3050 nnh = kordstat(di,nn,n,ind); | |
3051 else | |
3052 { nnh = 0; | |
3053 for (i=0; i<n; i++) nnh = MAX(nnh,di[i]); | |
3054 nnh = nnh*exp(log(1.0*nn/n)/d); | |
3055 } | |
3056 return(MAX(fxh,nnh)); | |
3057 } | |
3058 | |
3059 /* | |
3060 fast version of nbhd for ordered 1-d data | |
3061 */ | |
3062 void nbhd1(lfd,sp,des,k) | |
3063 lfdata *lfd; | |
3064 smpar *sp; | |
3065 design *des; | |
3066 int k; | |
3067 { double x, h, *xd, sc; | |
3068 int i, l, r, m, n, z; | |
3069 | |
3070 n = lfd->n; | |
3071 x = des->xev[0]; | |
3072 xd = dvari(lfd,0); | |
3073 sc = lfd->sca[0]; | |
3074 | |
3075 /* find closest data point to x */ | |
3076 if (x<=xd[0]) z = 0; | |
3077 else | |
3078 if (x>=xd[n-1]) z = n-1; | |
3079 else | |
3080 { l = 0; r = n-1; | |
3081 while (r-l>1) | |
3082 { z = (r+l)/2; | |
3083 if (xd[z]>x) r = z; | |
3084 else l = z; | |
3085 } | |
3086 /* now, xd[0..l] <= x < x[r..n-1] */ | |
3087 if ((x-xd[l])>(xd[r]-x)) z = r; else z = l; | |
3088 } | |
3089 /* closest point to x is xd[z] */ | |
3090 | |
3091 if (nn(sp)<0) /* user bandwidth */ | |
3092 h = sp->vb(des->xev); | |
3093 else | |
3094 { if (k>0) /* set h to nearest neighbor bandwidth */ | |
3095 { l = r = z; | |
3096 if (l==0) r = k-1; | |
3097 if (r==n-1) l = n-k; | |
3098 while (r-l<k-1) | |
3099 { if ((x-xd[l-1])<(xd[r+1]-x)) l--; else r++; | |
3100 if (l==0) r = k-1; | |
3101 if (r==n-1) l = n-k; | |
3102 } | |
3103 h = x-xd[l]; | |
3104 if (h<xd[r]-x) h = xd[r]-x; | |
3105 } | |
3106 else h = 0; | |
3107 h /= sc; | |
3108 if (h<fixh(sp)) h = fixh(sp); | |
3109 } | |
3110 | |
3111 m = 0; | |
3112 if (xd[z]>x) z--; /* so xd[z]<=x */ | |
3113 /* look left */ | |
3114 for (i=z; i>=0; i--) if (inlim(lfd,i)) | |
3115 { dist(des,i) = (x-xd[i])/sc; | |
3116 wght(des,i) = weight(lfd, sp, &xd[i], &x, h, 1, dist(des,i)); | |
3117 if (wght(des,i)>0) | |
3118 { des->ind[m] = i; | |
3119 m++; | |
3120 } else i = 0; | |
3121 } | |
3122 /* look right */ | |
3123 for (i=z+1; i<n; i++) if (inlim(lfd,i)) | |
3124 { dist(des,i) = (xd[i]-x)/sc; | |
3125 wght(des,i) = weight(lfd, sp, &xd[i], &x, h, 1, dist(des,i)); | |
3126 if (wght(des,i)>0) | |
3127 { des->ind[m] = i; | |
3128 m++; | |
3129 } else i = n; | |
3130 } | |
3131 | |
3132 des->n = m; | |
3133 des->h = h; | |
3134 } | |
3135 | |
3136 void nbhd_zeon(lfd,des) | |
3137 lfdata *lfd; | |
3138 design *des; | |
3139 { int i, j, m, eq; | |
3140 | |
3141 m = 0; | |
3142 for (i=0; i<lfd->n; i++) | |
3143 { eq = 1; | |
3144 for (j=0; j<lfd->d; j++) eq = eq && (des->xev[j] == datum(lfd,j,i)); | |
3145 if (eq) | |
3146 { wght(des,i) = 1; | |
3147 des->ind[m] = i; | |
3148 m++; | |
3149 } | |
3150 } | |
3151 des->n = m; | |
3152 des->h = 1.0; | |
3153 } | |
3154 | |
3155 void nbhd(lfd,des,nn,redo,sp) | |
3156 lfdata *lfd; | |
3157 design *des; | |
3158 int redo, nn; | |
3159 smpar *sp; | |
3160 { int d, i, j, m, n; | |
3161 double h, u[MXDIM]; | |
3162 | |
3163 if (lf_debug>1) mut_printf("nbhd: nn %d fixh %8.5f\n",nn,fixh(sp)); | |
3164 | |
3165 d = lfd->d; n = lfd->n; | |
3166 | |
3167 if (ker(sp)==WPARM) | |
3168 { for (i=0; i<n; i++) | |
3169 { wght(des,i) = 1.0; | |
3170 des->ind[i] = i; | |
3171 } | |
3172 des->n = n; | |
3173 return; | |
3174 } | |
3175 | |
3176 if (kt(sp)==KZEON) | |
3177 { nbhd_zeon(lfd,des); | |
3178 return; | |
3179 } | |
3180 | |
3181 if (kt(sp)==KCE) | |
3182 { des->h = 0.0; | |
3183 return; | |
3184 } | |
3185 | |
3186 /* ordered 1-dim; use fast searches */ | |
3187 if ((nn<=n) & (lfd->ord) & (ker(sp)!=WMINM) & (lfd->sty[0]!=STANGL)) | |
3188 { nbhd1(lfd,sp,des,nn); | |
3189 return; | |
3190 } | |
3191 | |
3192 if (!redo) | |
3193 { for (i=0; i<n; i++) | |
3194 { for (j=0; j<d; j++) u[j] = datum(lfd,j,i)-des->xev[j]; | |
3195 dist(des,i) = rho(u,lfd->sca,d,kt(sp),lfd->sty); | |
3196 des->ind[i] = i; | |
3197 } | |
3198 } | |
3199 else | |
3200 for (i=0; i<n; i++) des->ind[i] = i; | |
3201 | |
3202 if (ker(sp)==WMINM) | |
3203 { des->h = minmax(lfd,des,sp); | |
3204 return; | |
3205 } | |
3206 | |
3207 if (nn<0) | |
3208 h = sp->vb(des->xev); | |
3209 else | |
3210 h = compbandwid(des->di,des->ind,des->xev,n,lfd->d,nn,fixh(sp)); | |
3211 m = 0; | |
3212 for (i=0; i<n; i++) if (inlim(lfd,i)) | |
3213 { for (j=0; j<d; j++) u[j] = datum(lfd,j,i); | |
3214 wght(des,i) = weight(lfd, sp, u, des->xev, h, 1, dist(des,i)); | |
3215 if (wght(des,i)>0) | |
3216 { des->ind[m] = i; | |
3217 m++; | |
3218 } | |
3219 } | |
3220 des->n = m; | |
3221 des->h = h; | |
3222 } | |
3223 /* | |
3224 * Copyright 1996-2006 Catherine Loader. | |
3225 */ | |
3226 /* | |
3227 * | |
3228 * This file includes functions to solve for the scale estimate in | |
3229 * local robust regression and likelihood. The main entry point is | |
3230 * lf_robust(lfd,sp,des,mxit), | |
3231 * called from the locfit() function. | |
3232 * | |
3233 * The update_rs(x) accepts a residual scale x as the argument (actually, | |
3234 * it works on the log-scale). The function computes the local fit | |
3235 * assuming this residual scale, and re-estimates the scale from this | |
3236 * new fit. The final solution satisfies the fixed point equation | |
3237 * update_rs(x)=x. The function lf_robust() automatically calls | |
3238 * update_rs() through the fixed point iterations. | |
3239 * | |
3240 * The estimation of the scale from the fit is based on the sqrt of | |
3241 * the median deviance of observations with non-zero weights (in the | |
3242 * gaussian case, this is the median absolute residual). | |
3243 * | |
3244 * TODO: | |
3245 * Should use smoothing weights in the median. | |
3246 */ | |
3247 | |
3248 #include "locf.h" | |
3249 | |
3250 extern int lf_status; | |
3251 double robscale; | |
3252 | |
3253 static lfdata *rob_lfd; | |
3254 static smpar *rob_sp; | |
3255 static design *rob_des; | |
3256 static int rob_mxit; | |
3257 | |
3258 double median(x,n) | |
3259 double *x; | |
3260 int n; | |
3261 { int i, j, lt, eq, gt; | |
3262 double lo, hi, s; | |
3263 lo = hi = x[0]; | |
3264 for (i=0; i<n; i++) | |
3265 { lo = MIN(lo,x[i]); | |
3266 hi = MAX(hi,x[i]); | |
3267 } | |
3268 if (lo==hi) return(lo); | |
3269 lo -= (hi-lo); | |
3270 hi += (hi-lo); | |
3271 for (i=0; i<n; i++) | |
3272 { if ((x[i]>lo) & (x[i]<hi)) | |
3273 { s = x[i]; lt = eq = gt = 0; | |
3274 for (j=0; j<n; j++) | |
3275 { lt += (x[j]<s); | |
3276 eq += (x[j]==s); | |
3277 gt += (x[j]>s); | |
3278 } | |
3279 if ((2*(lt+eq)>n) && (2*(gt+eq)>n)) return(s); | |
3280 if (2*(lt+eq)<=n) lo = s; | |
3281 if (2*(gt+eq)<=n) hi = s; | |
3282 } | |
3283 } | |
3284 return((hi+lo)/2); | |
3285 } | |
3286 | |
3287 double nrobustscale(lfd,sp,des,rs) | |
3288 lfdata *lfd; | |
3289 smpar *sp; | |
3290 design *des; | |
3291 double rs; | |
3292 { int i, ii, p; | |
3293 double link[LLEN], sc, sd, sw, e; | |
3294 p = des->p; sc = sd = sw = 0.0; | |
3295 for (i=0; i<des->n; i++) | |
3296 { ii = des->ind[i]; | |
3297 fitv(des,ii) = base(lfd,ii)+innerprod(des->cf,d_xi(des,ii),p); | |
3298 e = resp(lfd,ii)-fitv(des,ii); | |
3299 stdlinks(link,lfd,sp,ii,fitv(des,ii),rs); | |
3300 sc += wght(des,ii)*e*link[ZDLL]; | |
3301 sd += wght(des,ii)*e*e*link[ZDDLL]; | |
3302 sw += wght(des,ii); | |
3303 } | |
3304 | |
3305 /* newton-raphson iteration for log(s) | |
3306 -psi(ei/s) - log(s); s = e^{-th} | |
3307 */ | |
3308 rs *= exp((sc-sw)/(sd+sc)); | |
3309 return(rs); | |
3310 } | |
3311 | |
3312 double robustscale(lfd,sp,des) | |
3313 lfdata *lfd; | |
3314 smpar *sp; | |
3315 design *des; | |
3316 { int i, ii, p, fam, lin, or; | |
3317 double rs, link[LLEN]; | |
3318 p = des->p; | |
3319 fam = fam(sp); | |
3320 lin = link(sp); | |
3321 or = fami(sp)->robust; | |
3322 fami(sp)->robust = 0; | |
3323 | |
3324 for (i=0; i<des->n; i++) | |
3325 { ii = des->ind[i]; | |
3326 fitv(des,ii) = base(lfd,ii) + innerprod(des->cf,d_xi(des,ii),p); | |
3327 links(fitv(des,ii),resp(lfd,ii),fami(sp),lin,link,cens(lfd,ii),prwt(lfd,ii),1.0); | |
3328 des->res[i] = -2*link[ZLIK]; | |
3329 } | |
3330 fami(sp)->robust = or; | |
3331 rs = sqrt(median(des->res,des->n)); | |
3332 | |
3333 if (rs==0.0) rs = 1.0; | |
3334 return(rs); | |
3335 } | |
3336 | |
3337 double update_rs(x) | |
3338 double x; | |
3339 { double nx; | |
3340 if (lf_status != LF_OK) return(x); | |
3341 robscale = exp(x); | |
3342 lfiter(rob_lfd,rob_sp,rob_des,rob_mxit); | |
3343 if (lf_status != LF_OK) return(x); | |
3344 | |
3345 nx = log(robustscale(rob_lfd,rob_sp,rob_des)); | |
3346 if (nx<x-0.2) nx = x-0.2; | |
3347 return(nx); | |
3348 } | |
3349 | |
3350 void lf_robust(lfd,sp,des,mxit) | |
3351 lfdata *lfd; | |
3352 design *des; | |
3353 smpar *sp; | |
3354 int mxit; | |
3355 { double x; | |
3356 rob_lfd = lfd; | |
3357 rob_des = des; | |
3358 rob_sp = sp; | |
3359 rob_mxit = mxit; | |
3360 lf_status = LF_OK; | |
3361 | |
3362 x = log(robustscale(lfd,sp,des)); | |
3363 | |
3364 solve_fp(update_rs, x, 1.0e-6, mxit); | |
3365 } | |
3366 /* | |
3367 * Copyright 1996-2006 Catherine Loader. | |
3368 */ | |
3369 /* | |
3370 * Post-fitting functions to compute the local variance and | |
3371 * influence functions. Also the local degrees of freedom | |
3372 * calculations for adaptive smoothing. | |
3373 */ | |
3374 | |
3375 #include "locf.h" | |
3376 | |
3377 extern double robscale; | |
3378 | |
3379 /* | |
3380 vmat() computes (after the local fit..) the matrix | |
3381 M2 = X^T W^2 V X. | |
3382 M12 = (X^T W V X)^{-1} M2 | |
3383 Also, for convenience, tr[0] = sum(wi) tr[1] = sum(wi^2). | |
3384 */ | |
3385 void vmat(lfd, sp, des, M12, M2) | |
3386 lfdata *lfd; | |
3387 smpar *sp; | |
3388 design *des; | |
3389 double *M12, *M2; | |
3390 { int i, ii, p, nk, ok; | |
3391 double link[LLEN], h, ww, tr0, tr1; | |
3392 p = des->p; | |
3393 setzero(M2,p*p); | |
3394 | |
3395 nk = -1; | |
3396 | |
3397 /* for density estimation, use integral rather than | |
3398 sum form, if W^2 is programmed... | |
3399 */ | |
3400 if ((fam(sp)<=THAZ) && (link(sp)==LLOG)) | |
3401 { switch(ker(sp)) | |
3402 { case WGAUS: nk = WGAUS; h = des->h/SQRT2; break; | |
3403 case WRECT: nk = WRECT; h = des->h; break; | |
3404 case WEPAN: nk = WBISQ; h = des->h; break; | |
3405 case WBISQ: nk = WQUQU; h = des->h; break; | |
3406 case WTCUB: nk = W6CUB; h = des->h; break; | |
3407 case WEXPL: nk = WEXPL; h = des->h/2; break; | |
3408 } | |
3409 } | |
3410 | |
3411 tr0 = tr1 = 0.0; | |
3412 if (nk != -1) | |
3413 { ok = ker(sp); ker(sp) = nk; | |
3414 /* compute M2 using integration. Use M12 as work matrix. */ | |
3415 (des->itype)(des->xev, M2, M12, des->cf, h); | |
3416 ker(sp) = ok; | |
3417 if (fam(sp)==TDEN) multmatscal(M2,des->smwt,p*p); | |
3418 tr0 = des->ss[0]; | |
3419 tr1 = M2[0]; /* n int W e^<a,A> */ | |
3420 } | |
3421 else | |
3422 { for (i=0; i<des->n; i++) | |
3423 { ii = des->ind[i]; | |
3424 stdlinks(link,lfd,sp,ii,fitv(des,ii),robscale); | |
3425 ww = SQR(wght(des,ii))*link[ZDDLL]; | |
3426 tr0 += wght(des,ii); | |
3427 tr1 += SQR(wght(des,ii)); | |
3428 addouter(M2,d_xi(des,ii),d_xi(des,ii),p,ww); | |
3429 } | |
3430 } | |
3431 des->tr0 = tr0; | |
3432 des->tr1 = tr1; | |
3433 | |
3434 memcpy(M12,M2,p*p*sizeof(double)); | |
3435 for (i=0; i<p; i++) | |
3436 jacob_solve(&des->xtwx,&M12[i*p]); | |
3437 } | |
3438 | |
3439 void lf_vcov(lfd,sp,des) | |
3440 lfdata *lfd; | |
3441 smpar *sp; | |
3442 design *des; | |
3443 { int i, j, k, p; | |
3444 double *M12, *M2; | |
3445 M12 = des->V; M2 = des->P; p = des->p; | |
3446 vmat(lfd,sp,des,M12,M2); /* M2 = X^T W^2 V X tr0=sum(W) tr1=sum(W*W) */ | |
3447 des->tr2 = m_trace(M12,p); /* tr (XTWVX)^{-1}(XTW^2VX) */ | |
3448 | |
3449 /* | |
3450 * Covariance matrix is M1^{-1} * M2 * M1^{-1} | |
3451 * We compute this using the cholesky decomposition of | |
3452 * M2; premultiplying by M1^{-1} and squaring. This | |
3453 * is more stable than direct computation in near-singular cases. | |
3454 */ | |
3455 chol_dec(M2,p,p); | |
3456 for (i=0; i<p; i++) | |
3457 for (j=0; j<i; j++) | |
3458 { M2[j*p+i] = M2[i*p+j]; | |
3459 M2[i*p+j] = 0.0; | |
3460 } | |
3461 for (i=0; i<p; i++) jacob_solve(&des->xtwx,&M2[i*p]); | |
3462 for (i=0; i<p; i++) | |
3463 { for (j=0; j<p; j++) | |
3464 { M12[i*p+j] = 0; | |
3465 for (k=0; k<p; k++) | |
3466 M12[i*p+j] += M2[k*p+i]*M2[k*p+j]; /* ith column of covariance */ | |
3467 } | |
3468 } | |
3469 if ((fam(sp)==TDEN) && (link(sp)==LIDENT)) | |
3470 multmatscal(M12,1/SQR(des->smwt),p*p); | |
3471 | |
3472 /* this computes the influence function as des->f1[0]. */ | |
3473 unitvec(des->f1,0,des->p); | |
3474 jacob_solve(&des->xtwx,des->f1); | |
3475 } | |
3476 | |
3477 /* local_df computes: | |
3478 * tr[0] = trace(W) | |
3479 * tr[1] = trace(W*W) | |
3480 * tr[2] = trace( M1^{-1} M2 ) | |
3481 * tr[3] = trace( M1^{-1} M3 ) | |
3482 * tr[4] = trace( (M1^{-1} M2)^2 ) | |
3483 * tr[5] = var(theta-hat). | |
3484 */ | |
3485 void local_df(lfd,sp,des,tr) | |
3486 lfdata *lfd; | |
3487 smpar *sp; | |
3488 design *des; | |
3489 double *tr; | |
3490 { int i, ii, j, p; | |
3491 double *m2, *V, ww, link[LLEN]; | |
3492 | |
3493 tr[0] = tr[1] = tr[2] = tr[3] = tr[4] = tr[5] = 0.0; | |
3494 m2 = des->V; V = des->P; p = des->p; | |
3495 | |
3496 vmat(lfd,sp,des,m2,V); /* M = X^T W^2 V X tr0=sum(W) tr1=sum(W*W) */ | |
3497 tr[0] = des->tr0; | |
3498 tr[1] = des->tr1; | |
3499 tr[2] = m_trace(m2,p); /* tr (XTWVX)^{-1}(XTW^2VX) */ | |
3500 | |
3501 unitvec(des->f1,0,p); | |
3502 jacob_solve(&des->xtwx,des->f1); | |
3503 for (i=0; i<p; i++) | |
3504 for (j=0; j<p; j++) | |
3505 { tr[4] += m2[i*p+j]*m2[j*p+i]; /* tr(M^2) */ | |
3506 tr[5] += des->f1[i]*V[i*p+j]*des->f1[j]; /* var(thetahat) */ | |
3507 } | |
3508 tr[5] = sqrt(tr[5]); | |
3509 | |
3510 setzero(m2,p*p); | |
3511 for (i=0; i<des->n; i++) | |
3512 { ii = des->ind[i]; | |
3513 stdlinks(link,lfd,sp,ii,fitv(des,ii),robscale); | |
3514 ww = wght(des,ii)*wght(des,ii)*wght(des,ii)*link[ZDDLL]; | |
3515 addouter(m2,d_xi(des,ii),d_xi(des,ii),p,ww); | |
3516 } | |
3517 for (i=0; i<p; i++) | |
3518 { jacob_solve(&des->xtwx,&m2[i*p]); | |
3519 tr[3] += m2[i*(p+1)]; | |
3520 } | |
3521 | |
3522 return; | |
3523 } | |
3524 /* | |
3525 * Copyright 1996-2006 Catherine Loader. | |
3526 */ | |
3527 /* | |
3528 * Routines for computing weight diagrams. | |
3529 * wdiag(lf,des,lx,deg,ty,exp) | |
3530 * Must locfit() first, unless ker==WPARM and has par. comp. | |
3531 * | |
3532 */ | |
3533 | |
3534 #include "locf.h" | |
3535 | |
3536 static double *wd; | |
3537 extern double robscale; | |
3538 void nnresproj(lfd,sp,des,u,m,p) | |
3539 lfdata *lfd; | |
3540 smpar *sp; | |
3541 design *des; | |
3542 double *u; | |
3543 int m, p; | |
3544 { int i, ii, j; | |
3545 double link[LLEN]; | |
3546 setzero(des->f1,p); | |
3547 for (j=0; j<m; j++) | |
3548 { ii = des->ind[j]; | |
3549 stdlinks(link,lfd,sp,ii,fitv(des,ii),robscale); | |
3550 for (i=0; i<p; i++) des->f1[i] += link[ZDDLL]*d_xij(des,j,ii)*u[j]; | |
3551 } | |
3552 jacob_solve(&des->xtwx,des->f1); | |
3553 for (i=0; i<m; i++) | |
3554 { ii = des->ind[i]; | |
3555 u[i] -= innerprod(des->f1,d_xi(des,ii),p)*wght(des,ii); | |
3556 } | |
3557 } | |
3558 | |
3559 void wdexpand(l,n,ind,m) | |
3560 double *l; | |
3561 int *ind, n, m; | |
3562 { int i, j, t; | |
3563 double z; | |
3564 for (j=m; j<n; j++) { l[j] = 0.0; ind[j] = -1; } | |
3565 j = m-1; | |
3566 while (j>=0) | |
3567 { if (ind[j]==j) j--; | |
3568 else | |
3569 { i = ind[j]; | |
3570 z = l[j]; l[j] = l[i]; l[i] = z; | |
3571 t = ind[j]; ind[j] = ind[i]; ind[i] = t; | |
3572 if (ind[j]==-1) j--; | |
3573 } | |
3574 } | |
3575 | |
3576 /* for (i=n-1; i>=0; i--) | |
3577 { l[i] = ((j>=0) && (ind[j]==i)) ? l[j--] : 0.0; } */ | |
3578 } | |
3579 | |
3580 int wdiagp(lfd,sp,des,lx,pc,dv,deg,ty,exp) | |
3581 lfdata *lfd; | |
3582 smpar *sp; | |
3583 design *des; | |
3584 paramcomp *pc; | |
3585 deriv *dv; | |
3586 double *lx; | |
3587 int deg, ty, exp; | |
3588 { int i, j, p, nd; | |
3589 double *l1; | |
3590 | |
3591 p = des->p; | |
3592 | |
3593 fitfun(lfd,sp,des->xev,pc->xbar,des->f1,dv); | |
3594 if (exp) | |
3595 { jacob_solve(&pc->xtwx,des->f1); | |
3596 for (i=0; i<lfd->n; i++) | |
3597 lx[i] = innerprod(des->f1,d_xi(des,des->ind[i]),p); | |
3598 return(lfd->n); | |
3599 } | |
3600 jacob_hsolve(&pc->xtwx,des->f1); | |
3601 for (i=0; i<p; i++) lx[i] = des->f1[i]; | |
3602 | |
3603 nd = dv->nd; | |
3604 dv->nd = nd+1; | |
3605 if (deg>=1) | |
3606 for (i=0; i<lfd->d; i++) | |
3607 { dv->deriv[nd] = i; | |
3608 l1 = &lx[(i+1)*p]; | |
3609 fitfun(lfd,sp,des->xev,pc->xbar,l1,dv); | |
3610 jacob_hsolve(&pc->xtwx,l1); | |
3611 } | |
3612 | |
3613 dv->nd = nd+2; | |
3614 if (deg>=2) | |
3615 for (i=0; i<lfd->d; i++) | |
3616 { dv->deriv[nd] = i; | |
3617 for (j=0; j<lfd->d; j++) | |
3618 { dv->deriv[nd+1] = j; | |
3619 l1 = &lx[(i*lfd->d+j+lfd->d+1)*p]; | |
3620 fitfun(lfd,sp,des->xev,pc->xbar,l1,dv); | |
3621 jacob_hsolve(&pc->xtwx,l1); | |
3622 } } | |
3623 dv->nd = nd; | |
3624 return(p); | |
3625 } | |
3626 | |
3627 int wdiag(lfd,sp,des,lx,dv,deg,ty,exp) | |
3628 lfdata *lfd; | |
3629 smpar *sp; | |
3630 design *des; | |
3631 deriv *dv; | |
3632 double *lx; | |
3633 int deg, ty, exp; | |
3634 /* deg=0: l(x) only. | |
3635 deg=1: l(x), l'(x) | |
3636 deg=2: l(x), l'(x), l''(x) | |
3637 ty = 1: e1 (X^T WVX)^{-1} X^T W -- hat matrix | |
3638 ty = 2: e1 (X^T WVX)^{-1} X^T WV^{1/2} -- scb's | |
3639 */ | |
3640 { double w, *X, *lxd, *lxdd, wdd, wdw, *ulx, link[LLEN], h; | |
3641 double dfx[MXDIM], hs[MXDIM]; | |
3642 int i, ii, j, k, l, m, d, p, nd; | |
3643 | |
3644 h = des->h; | |
3645 nd = dv->nd; | |
3646 wd = des->wd; | |
3647 d = lfd->d; p = des->p; X = d_x(des); | |
3648 ulx = des->res; | |
3649 m = des->n; | |
3650 for (i=0; i<d; i++) hs[i] = h*lfd->sca[i]; | |
3651 if (deg>0) | |
3652 { lxd = &lx[m]; | |
3653 setzero(lxd,m*d); | |
3654 if (deg>1) | |
3655 { lxdd = &lxd[d*m]; | |
3656 setzero(lxdd,m*d*d); | |
3657 } } | |
3658 | |
3659 if (nd>0) fitfun(lfd,sp,des->xev,des->xev,des->f1,dv); /* c(0) */ | |
3660 else unitvec(des->f1,0,p); | |
3661 jacob_solve(&des->xtwx,des->f1); /* c(0) (X^TWX)^{-1} */ | |
3662 for (i=0; i<m; i++) | |
3663 { ii = des->ind[i]; | |
3664 lx[i] = innerprod(des->f1,&X[ii*p],p); /* c(0)(XTWX)^{-1}X^T */ | |
3665 if (deg>0) | |
3666 { wd[i] = Wd(dist(des,ii)/h,ker(sp)); | |
3667 for (j=0; j<d; j++) | |
3668 { dfx[j] = datum(lfd,j,ii)-des->xev[j]; | |
3669 lxd[j*m+i] = lx[i]*wght(des,ii)*weightd(dfx[j],lfd->sca[j], | |
3670 d,ker(sp),kt(sp),h,lfd->sty[j],dist(des,ii)); | |
3671 /* c(0) (XTWX)^{-1}XTW' */ | |
3672 } | |
3673 if (deg>1) | |
3674 { wdd = Wdd(dist(des,ii)/h,ker(sp)); | |
3675 for (j=0; j<d; j++) | |
3676 for (k=0; k<d; k++) | |
3677 { w = (dist(des,ii)==0) ? 0 : h/dist(des,ii); | |
3678 w = wdd * (des->xev[k]-datum(lfd,k,ii)) * (des->xev[j]-datum(lfd,j,ii)) | |
3679 * w*w / (hs[k]*hs[k]*hs[j]*hs[j]); | |
3680 if (j==k) w += wd[i]/(hs[j]*hs[j]); | |
3681 lxdd[(j*d+k)*m+i] = lx[i]*w; | |
3682 /* c(0)(XTWX)^{-1}XTW'' */ | |
3683 } | |
3684 } | |
3685 } | |
3686 lx[i] *= wght(des,ii); | |
3687 } | |
3688 | |
3689 dv->nd = nd+1; | |
3690 if (deg==2) | |
3691 { for (i=0; i<d; i++) | |
3692 { dv->deriv[nd] = i; | |
3693 fitfun(lfd,sp,des->xev,des->xev,des->f1,dv); | |
3694 for (k=0; k<m; k++) | |
3695 { ii = des->ind[i]; | |
3696 stdlinks(link,lfd,sp,ii,fitv(des,ii),robscale); | |
3697 for (j=0; j<p; j++) | |
3698 des->f1[j] -= link[ZDDLL]*lxd[i*m+k]*X[ii*p+j]; | |
3699 /* c'(x)-c(x)(XTWX)^{-1}XTW'X */ | |
3700 } | |
3701 jacob_solve(&des->xtwx,des->f1); /* (...)(XTWX)^{-1} */ | |
3702 for (j=0; j<m; j++) | |
3703 { ii = des->ind[j]; | |
3704 ulx[j] = innerprod(des->f1,&X[ii*p],p); /* (...)XT */ | |
3705 } | |
3706 for (j=0; j<d; j++) | |
3707 for (k=0; k<m; k++) | |
3708 { ii = des->ind[k]; | |
3709 dfx[j] = datum(lfd,j,ii)-des->xev[j]; | |
3710 wdw = wght(des,ii)*weightd(dfx[j],lfd->sca[j],d,ker(sp), | |
3711 kt(sp),h,lfd->sty[j],dist(des,ii)); | |
3712 lxdd[(i*d+j)*m+k] += ulx[k]*wdw; | |
3713 lxdd[(j*d+i)*m+k] += ulx[k]*wdw; | |
3714 } /* + 2(c'-c(XTWX)^{-1}XTW'X)(XTWX)^{-1}XTW' */ | |
3715 } | |
3716 for (j=0; j<d*d; j++) nnresproj(lfd,sp,des,&lxdd[j*m],m,p); | |
3717 /* * (I-X(XTWX)^{-1} XTW */ | |
3718 } | |
3719 if (deg>0) | |
3720 { for (j=0; j<d; j++) nnresproj(lfd,sp,des,&lxd[j*m],m,p); | |
3721 /* c(0)(XTWX)^{-1}XTW'(I-X(XTWX)^{-1}XTW) */ | |
3722 for (i=0; i<d; i++) | |
3723 { dv->deriv[nd]=i; | |
3724 fitfun(lfd,sp,des->xev,des->xev,des->f1,dv); | |
3725 jacob_solve(&des->xtwx,des->f1); | |
3726 for (k=0; k<m; k++) | |
3727 { ii = des->ind[k]; | |
3728 for (l=0; l<p; l++) | |
3729 lxd[i*m+k] += des->f1[l]*X[ii*p+l]*wght(des,ii); | |
3730 } /* add c'(0)(XTWX)^{-1}XTW */ | |
3731 } | |
3732 } | |
3733 | |
3734 dv->nd = nd+2; | |
3735 if (deg==2) | |
3736 { for (i=0; i<d; i++) | |
3737 { dv->deriv[nd]=i; | |
3738 for (j=0; j<d; j++) | |
3739 { dv->deriv[nd+1]=j; | |
3740 fitfun(lfd,sp,des->xev,des->xev,des->f1,dv); | |
3741 jacob_solve(&des->xtwx,des->f1); | |
3742 for (k=0; k<m; k++) | |
3743 { ii = des->ind[k]; | |
3744 for (l=0; l<p; l++) | |
3745 lxdd[(i*d+j)*m+k] += des->f1[l]*X[ii*p+l]*wght(des,ii); | |
3746 } /* + c''(x)(XTWX)^{-1}XTW */ | |
3747 } | |
3748 } | |
3749 } | |
3750 dv->nd = nd; | |
3751 | |
3752 k = 1+d*(deg>0)+d*d*(deg==2); | |
3753 | |
3754 if (exp) wdexpand(lx,lfd->n,des->ind,m); | |
3755 | |
3756 if (ty==1) return(m); | |
3757 for (i=0; i<m; i++) | |
3758 { ii = des->ind[i]; | |
3759 stdlinks(link,lfd,sp,ii,fitv(des,ii),robscale); | |
3760 link[ZDDLL] = sqrt(fabs(link[ZDDLL])); | |
3761 for (j=0; j<k; j++) lx[j*m+i] *= link[ZDDLL]; | |
3762 } | |
3763 return(m); | |
3764 } | |
3765 /* | |
3766 * Copyright 1996-2006 Catherine Loader. | |
3767 */ | |
3768 /* | |
3769 * String matching functions. For a given argument string, find | |
3770 * the best match from an array of possibilities. Is there a library | |
3771 * function somewhere to do something like this? | |
3772 * | |
3773 * return values of -1 indicate failure/unknown string. | |
3774 */ | |
3775 | |
3776 #include "locf.h" | |
3777 | |
3778 int ct_match(z1, z2) | |
3779 char *z1, *z2; | |
3780 { int ct = 0; | |
3781 while (z1[ct]==z2[ct]) | |
3782 { if (z1[ct]=='\0') return(ct+1); | |
3783 ct++; | |
3784 } | |
3785 return(ct); | |
3786 } | |
3787 | |
3788 int pmatch(z, strings, vals, n, def) | |
3789 char *z, **strings; | |
3790 int *vals, n, def; | |
3791 { int i, ct, best, best_ct; | |
3792 best = -1; | |
3793 best_ct = 0; | |
3794 | |
3795 for (i=0; i<n; i++) | |
3796 { ct = ct_match(z,strings[i]); | |
3797 if (ct==strlen(z)+1) return(vals[i]); | |
3798 if (ct>best_ct) { best = i; best_ct = ct; } | |
3799 } | |
3800 if (best==-1) return(def); | |
3801 return(vals[best]); | |
3802 } | |
3803 /* | |
3804 * Copyright 1996-2006 Catherine Loader. | |
3805 */ | |
3806 #include "locf.h" | |
3807 | |
3808 int lf_maxit = 20; | |
3809 int lf_debug = 0; | |
3810 int lf_error = 0; | |
3811 | |
3812 double s0, s1; | |
3813 static lfdata *lf_lfd; | |
3814 static design *lf_des; | |
3815 static smpar *lf_sp; | |
3816 int lf_status; | |
3817 int ident=0; | |
3818 double lf_tol; | |
3819 extern double robscale; | |
3820 | |
3821 void lfdata_init(lfd) | |
3822 lfdata *lfd; | |
3823 { int i; | |
3824 for (i=0; i<MXDIM; i++) | |
3825 { lfd->sty[i] = 0; | |
3826 lfd->sca[i] = 1.0; | |
3827 lfd->xl[i] = lfd->xl[i+MXDIM] = 0.0; | |
3828 } | |
3829 lfd->y = lfd->w = lfd->c = lfd->b = NULL; | |
3830 lfd->d = lfd->n = 0; | |
3831 } | |
3832 | |
3833 void smpar_init(sp,lfd) | |
3834 smpar *sp; | |
3835 lfdata *lfd; | |
3836 { nn(sp) = 0.7; | |
3837 fixh(sp)= 0.0; | |
3838 pen(sp) = 0.0; | |
3839 acri(sp)= ANONE; | |
3840 deg(sp) = deg0(sp) = 2; | |
3841 ubas(sp) = 0; | |
3842 kt(sp) = KSPH; | |
3843 ker(sp) = WTCUB; | |
3844 fam(sp) = 64+TGAUS; | |
3845 link(sp)= LDEFAU; | |
3846 npar(sp) = calcp(sp,lfd->d); | |
3847 } | |
3848 | |
3849 void deriv_init(dv) | |
3850 deriv *dv; | |
3851 { dv->nd = 0; | |
3852 } | |
3853 | |
3854 int des_reqd(n,p) | |
3855 int n, p; | |
3856 { | |
3857 return(n*(p+5)+2*p*p+4*p + jac_reqd(p)); | |
3858 } | |
3859 int des_reqi(n,p) | |
3860 int n, p; | |
3861 { return(n+p); | |
3862 } | |
3863 | |
3864 void des_init(des,n,p) | |
3865 design *des; | |
3866 int n, p; | |
3867 { double *z; | |
3868 int k; | |
3869 | |
3870 if (n<=0) WARN(("des_init: n <= 0")); | |
3871 if (p<=0) WARN(("des_init: p <= 0")); | |
3872 | |
3873 if (des->des_init_id != DES_INIT_ID) | |
3874 { des->lwk = des->lind = 0; | |
3875 des->des_init_id = DES_INIT_ID; | |
3876 } | |
3877 | |
3878 k = des_reqd(n,p); | |
3879 if (k>des->lwk) | |
3880 { des->wk = (double *)calloc(k,sizeof(double)); | |
3881 if ( des->wk == NULL ) { | |
3882 printf("Problem allocating memory for des->wk\n");fflush(stdout); | |
3883 } | |
3884 des->lwk = k; | |
3885 } | |
3886 z = des->wk; | |
3887 | |
3888 des->X = z; z += n*p; | |
3889 des->w = z; z += n; | |
3890 des->res=z; z += n; | |
3891 des->di =z; z += n; | |
3892 des->th =z; z += n; | |
3893 des->wd =z; z += n; | |
3894 des->V =z; z += p*p; | |
3895 des->P =z; z += p*p; | |
3896 des->f1 =z; z += p; | |
3897 des->ss =z; z += p; | |
3898 des->oc =z; z += p; | |
3899 des->cf =z; z += p; | |
3900 | |
3901 z = jac_alloc(&des->xtwx,p,z); | |
3902 | |
3903 k = des_reqi(n,p); | |
3904 if (k>des->lind) | |
3905 { | |
3906 des->ind = (int *)calloc(k,sizeof(int)); | |
3907 if ( des->ind == NULL ) { | |
3908 printf("Problem allocating memory for des->ind\n");fflush(stdout); | |
3909 } | |
3910 des->lind = k; | |
3911 } | |
3912 des->fix = &des->ind[n]; | |
3913 for (k=0; k<p; k++) des->fix[k] = 0; | |
3914 | |
3915 des->n = n; des->p = p; | |
3916 des->smwt = n; | |
3917 des->xtwx.p = p; | |
3918 } | |
3919 | |
3920 void deschk(des,n,p) | |
3921 design *des; | |
3922 int n, p; | |
3923 { WARN(("deschk deprecated - use des_init()")); | |
3924 des_init(des,n,p); | |
3925 } | |
3926 | |
3927 int likereg(coef, lk0, f1, Z) | |
3928 double *coef, *lk0, *f1, *Z; | |
3929 { int i, ii, j, p; | |
3930 double lk, ww, link[LLEN], *X; | |
3931 | |
3932 if (lf_debug>2) mut_printf(" likereg: %8.5f\n",coef[0]); | |
3933 lf_status = LF_OK; | |
3934 lk = 0.0; p = lf_des->p; | |
3935 setzero(Z,p*p); | |
3936 setzero(f1,p); | |
3937 for (i=0; i<lf_des->n; i++) | |
3938 { | |
3939 ii = lf_des->ind[i]; | |
3940 X = d_xi(lf_des,ii); | |
3941 fitv(lf_des,ii) = base(lf_lfd,ii)+innerprod(coef,X,p); | |
3942 lf_status = stdlinks(link,lf_lfd,lf_sp,ii,fitv(lf_des,ii),robscale); | |
3943 if (lf_status == LF_BADP) | |
3944 { *lk0 = -1.0e300; | |
3945 return(NR_REDUCE); | |
3946 } | |
3947 if (lf_error) lf_status = LF_ERR; | |
3948 if (lf_status != LF_OK) return(NR_BREAK); | |
3949 | |
3950 ww = wght(lf_des,ii); | |
3951 lk += ww*link[ZLIK]; | |
3952 for (j=0; j<p; j++) | |
3953 f1[j] += X[j]*ww*link[ZDLL]; | |
3954 addouter(Z, X, X, p, ww*link[ZDDLL]); | |
3955 } | |
3956 for (i=0; i<p; i++) if (lf_des->fix[i]) | |
3957 { for (j=0; j<p; j++) Z[i*p+j] = Z[j*p+i] = 0.0; | |
3958 Z[i*p+i] = 1.0; | |
3959 f1[i] = 0.0; | |
3960 } | |
3961 | |
3962 if (lf_debug>4) prresp(coef,Z,p); | |
3963 if (lf_debug>3) mut_printf(" likelihood: %8.5f\n",lk); | |
3964 *lk0 = lf_des->llk = lk; | |
3965 | |
3966 lf_status = fami(lf_sp)->pcheck(lf_sp,lf_des,lf_lfd); | |
3967 switch(lf_status) | |
3968 { case LF_DONE: return(NR_BREAK); | |
3969 case LF_OOB: return(NR_REDUCE); | |
3970 case LF_PF: return(NR_REDUCE); | |
3971 case LF_NSLN: return(NR_BREAK); | |
3972 } | |
3973 | |
3974 return(NR_OK); | |
3975 } | |
3976 | |
3977 int reginit(lfd,des,sp) | |
3978 lfdata *lfd; | |
3979 design *des; | |
3980 smpar *sp; | |
3981 { int i, ii; | |
3982 double sb, link[LLEN]; | |
3983 s0 = s1 = sb = 0; | |
3984 for (i=0; i<des->n; i++) | |
3985 { ii = des->ind[i]; | |
3986 links(base(lfd,ii),resp(lfd,ii),fami(sp),LINIT,link,cens(lfd,ii),prwt(lfd,ii),1.0); | |
3987 s1 += wght(des,ii)*link[ZDLL]; | |
3988 s0 += wght(des,ii)*prwt(lfd,ii); | |
3989 sb += wght(des,ii)*prwt(lfd,ii)*base(lfd,ii); | |
3990 } | |
3991 if (s0==0) return(LF_NOPT); /* no observations with W>0 */ | |
3992 setzero(des->cf,des->p); | |
3993 lf_tol = 1.0e-6*s0; | |
3994 switch(link(sp)) | |
3995 { case LIDENT: | |
3996 des->cf[0] = (s1-sb)/s0; | |
3997 return(LF_OK); | |
3998 case LLOG: | |
3999 if (s1<=0.0) | |
4000 { des->cf[0] = -1000; | |
4001 return(LF_INFA); | |
4002 } | |
4003 des->cf[0] = log(s1/s0) - sb/s0; | |
4004 return(LF_OK); | |
4005 case LLOGIT: | |
4006 if (s1<=0.0) | |
4007 { des->cf[0] = -1000; | |
4008 return(LF_INFA); | |
4009 } | |
4010 if (s1>=s0) | |
4011 { des->cf[0] = 1000; | |
4012 return(LF_INFA); | |
4013 } | |
4014 des->cf[0] = logit(s1/s0)-sb/s0; | |
4015 return(LF_OK); | |
4016 case LINVER: | |
4017 if (s1<=0.0) | |
4018 { des->cf[0] = 1e100; | |
4019 return(LF_INFA); | |
4020 } | |
4021 des->cf[0] = s0/s1-sb/s0; | |
4022 return(LF_OK); | |
4023 case LSQRT: | |
4024 des->cf[0] = sqrt(s1/s0)-sb/s0; | |
4025 return(LF_OK); | |
4026 case LASIN: | |
4027 des->cf[0] = asin(sqrt(s1/s0))-sb/s0; | |
4028 return(LF_OK); | |
4029 default: | |
4030 LERR(("reginit: invalid link %d",link(sp))); | |
4031 return(LF_ERR); | |
4032 } | |
4033 } | |
4034 | |
4035 int lfinit(lfd,sp,des) | |
4036 lfdata *lfd; | |
4037 smpar *sp; | |
4038 design *des; | |
4039 { int initstat; | |
4040 des->xtwx.sm = (deg0(sp)<deg(sp)) ? JAC_CHOL : JAC_EIGD; | |
4041 | |
4042 designmatrix(lfd,sp,des); | |
4043 setfamily(sp); | |
4044 initstat = fami(sp)->initial(lfd,des,sp); | |
4045 | |
4046 return(initstat); | |
4047 } | |
4048 | |
4049 void lfiter(lfd,sp,des,maxit) | |
4050 lfdata *lfd; | |
4051 smpar *sp; | |
4052 design *des; | |
4053 int maxit; | |
4054 { int err; | |
4055 if (lf_debug>1) mut_printf(" lfiter: %8.5f\n",des->cf[0]); | |
4056 | |
4057 lf_des = des; | |
4058 lf_lfd = lfd; | |
4059 lf_sp = sp; | |
4060 | |
4061 max_nr(fami(sp)->like, des->cf, des->oc, des->res, des->f1, | |
4062 &des->xtwx, des->p, maxit, lf_tol, &err); | |
4063 switch(err) | |
4064 { case NR_OK: return; | |
4065 case NR_NCON: | |
4066 WARN(("max_nr not converged")); | |
4067 return; | |
4068 case NR_NDIV: | |
4069 WARN(("max_nr reduction problem")); | |
4070 return; | |
4071 } | |
4072 WARN(("max_nr return status %d",err)); | |
4073 } | |
4074 | |
4075 int use_robust_scale(int tg) | |
4076 { if ((tg&64)==0) return(0); /* not quasi - no scale */ | |
4077 if (((tg&128)==0) & (((tg&63)!=TROBT) & ((tg&63)!=TCAUC))) return(0); | |
4078 return(1); | |
4079 } | |
4080 | |
4081 /* | |
4082 * noit not really needed any more, since | |
4083 * gauss->pcheck returns LF_DONE, and likereg NR_BREAK | |
4084 * in gaussian case. | |
4085 * nb: 0/1: does local neighborhood and weights need computing? | |
4086 * cv: 0/1: is variance/covariance matrix needed? | |
4087 */ | |
4088 int locfit(lfd,des,sp,noit,nb,cv) | |
4089 lfdata *lfd; | |
4090 design *des; | |
4091 smpar *sp; | |
4092 int noit, nb, cv; | |
4093 { int i; | |
4094 | |
4095 if (des->xev==NULL) | |
4096 { LERR(("locfit: NULL evaluation point?")); | |
4097 return(246); | |
4098 } | |
4099 | |
4100 if (lf_debug>0) | |
4101 { mut_printf("locfit: "); | |
4102 for (i=0; i<lfd->d; i++) mut_printf(" %10.6f",des->xev[i]); | |
4103 mut_printf("\n"); | |
4104 } | |
4105 | |
4106 /* the 1e-12 avoids problems that can occur with roundoff */ | |
4107 if (nb) nbhd(lfd,des,(int)(lfd->n*nn(sp)+1e-12),0,sp); | |
4108 | |
4109 lf_status = lfinit(lfd,sp,des); | |
4110 | |
4111 if (lf_status == LF_OK) | |
4112 { if (use_robust_scale(fam(sp))) | |
4113 lf_robust(lfd,sp,des,lf_maxit); | |
4114 else | |
4115 { if ((fam(sp)&63)==TQUANT) | |
4116 lfquantile(lfd,sp,des,lf_maxit); | |
4117 else | |
4118 { robscale = 1.0; | |
4119 lfiter(lfd,sp,des,lf_maxit); | |
4120 } | |
4121 } | |
4122 } | |
4123 | |
4124 if (lf_status == LF_DONE) lf_status = LF_OK; | |
4125 if (lf_status == LF_OOB) lf_status = LF_OK; | |
4126 | |
4127 if ((fam(sp)&63)==TDEN) /* convert from rate to density */ | |
4128 { switch(link(sp)) | |
4129 { case LLOG: | |
4130 des->cf[0] -= log(des->smwt); | |
4131 break; | |
4132 case LIDENT: | |
4133 multmatscal(des->cf,1.0/des->smwt,des->p); | |
4134 break; | |
4135 default: LERR(("Density adjustment; invalid link")); | |
4136 } | |
4137 } | |
4138 | |
4139 /* variance calculations, if requested */ | |
4140 if (cv) | |
4141 { switch(lf_status) | |
4142 { case LF_PF: /* for these cases, variance calc. would likely fail. */ | |
4143 case LF_NOPT: | |
4144 case LF_NSLN: | |
4145 case LF_INFA: | |
4146 case LF_DEMP: | |
4147 case LF_XOOR: | |
4148 case LF_DNOP: | |
4149 case LF_BADP: | |
4150 des->llk = des->tr0 = des->tr1 = des->tr2 = 0.0; | |
4151 setzero(des->V,des->p*des->p); | |
4152 setzero(des->f1,des->p); | |
4153 break; | |
4154 default: lf_vcov(lfd,sp,des); | |
4155 } | |
4156 } | |
4157 | |
4158 return(lf_status); | |
4159 } | |
4160 | |
4161 void lf_status_msg(status) | |
4162 int status; | |
4163 { switch(status) | |
4164 { case LF_OK: return; | |
4165 case LF_NCON: WARN(("locfit did not converge")); return; | |
4166 case LF_OOB: WARN(("parameters out of bounds")); return; | |
4167 case LF_PF: WARN(("perfect fit")); return; | |
4168 case LF_NOPT: WARN(("no points with non-zero weight")); return; | |
4169 case LF_NSLN: WARN(("no solution")); return; | |
4170 case LF_INFA: WARN(("initial value problem")); return; | |
4171 case LF_DEMP: WARN(("density estimate, empty integration region")); return; | |
4172 case LF_XOOR: WARN(("procv: fit point outside xlim region")); return; | |
4173 case LF_DNOP: WARN(("density estimation -- insufficient points in smoothing window")); return; | |
4174 case LF_BADP: WARN(("bad parameters")); return; | |
4175 default: WARN(("procv: unknown return code %d",status)); return; | |
4176 } } | |
4177 /* | |
4178 * Copyright 1996-2006 Catherine Loader. | |
4179 */ | |
4180 /* | |
4181 * Compute minimax weights for local regression. | |
4182 */ | |
4183 | |
4184 #include "locf.h" | |
4185 #define NR_EMPTY 834 | |
4186 | |
4187 int mmsm_ct; | |
4188 | |
4189 static int debug=0; | |
4190 #define CONVTOL 1.0e-8 | |
4191 #define SINGTOL 1.0e-10 | |
4192 #define NR_SINGULAR 100 | |
4193 | |
4194 static lfdata *mm_lfd; | |
4195 static design *mm_des; | |
4196 static double mm_gam, mmf, lb; | |
4197 static int st; | |
4198 | |
4199 double ipower(x,n) /* use for n not too large!! */ | |
4200 double x; | |
4201 int n; | |
4202 { if (n==0) return(1.0); | |
4203 if (n<0) return(1/ipower(x,-n)); | |
4204 return(x*ipower(x,n-1)); | |
4205 } | |
4206 | |
4207 double setmmwt(des,a,gam) | |
4208 design *des; | |
4209 double *a, gam; | |
4210 { double ip, w0, w1, sw, wt; | |
4211 int i; | |
4212 sw = 0.0; | |
4213 for (i=0; i<mm_lfd->n; i++) | |
4214 { ip = innerprod(a,d_xi(des,i),des->p); | |
4215 wt = prwt(mm_lfd,i); | |
4216 w0 = ip - gam*des->wd[i]; | |
4217 w1 = ip + gam*des->wd[i]; | |
4218 wght(des,i) = 0.0; | |
4219 if (w0>0) { wght(des,i) = w0; sw += wt*w0*w0; } | |
4220 if (w1<0) { wght(des,i) = w1; sw += wt*w1*w1; } | |
4221 } | |
4222 return(sw/2-a[0]); | |
4223 } | |
4224 | |
4225 /* compute sum_{w!=0} AA^T; e1-sum wA */ | |
4226 int mmsums(des,coef,f,z,J) | |
4227 design *des; | |
4228 double *coef, *f, *z; | |
4229 jacobian *J; | |
4230 { int ct, i, j, p, sing; | |
4231 double *A; | |
4232 | |
4233 mmsm_ct++; | |
4234 A = J->Z; | |
4235 *f = setmmwt(des,coef,mm_gam); | |
4236 | |
4237 p = des->p; | |
4238 setzero(A,p*p); | |
4239 setzero(z,p); | |
4240 z[0] = 1.0; | |
4241 ct = 0; | |
4242 | |
4243 for (i=0; i<mm_lfd->n; i++) | |
4244 if (wght(des,i)!=0.0) | |
4245 { addouter(A,d_xi(des,i),d_xi(des,i),p,prwt(mm_lfd,i)); | |
4246 for (j=0; j<p; j++) z[j] -= prwt(mm_lfd,i)*wght(des,i)*d_xij(des,i,j); | |
4247 ct++; | |
4248 } | |
4249 if (ct==0) return(NR_EMPTY); | |
4250 | |
4251 J->st = JAC_RAW; | |
4252 J->p = p; | |
4253 jacob_dec(J,JAC_EIGD); | |
4254 | |
4255 sing = 0; | |
4256 for (i=0; i<p; i++) sing |= (J->Z[i*p+i]<SINGTOL); | |
4257 if ((debug) & (sing)) mut_printf("SINGULAR!!!!\n"); | |
4258 | |
4259 return((sing) ? NR_SINGULAR : NR_OK); | |
4260 } | |
4261 | |
4262 int descenddir(des,coef,dlt,f,af) | |
4263 design *des; | |
4264 double *coef, *dlt, *f; | |
4265 int af; | |
4266 { int i, p; | |
4267 double f0, *oc; | |
4268 | |
4269 if (debug) mut_printf("descenddir: %8.5f %8.5f\n",dlt[0],dlt[1]); | |
4270 | |
4271 f0 = *f; | |
4272 oc = des->oc; | |
4273 p = des->p; | |
4274 memcpy(oc,coef,p*sizeof(double)); | |
4275 | |
4276 for (i=0; i<p; i++) coef[i] = oc[i]+lb*dlt[i]; | |
4277 st = mmsums(des,coef,f,des->f1,&des->xtwx); | |
4278 | |
4279 if (*f>f0) /* halve till we drop */ | |
4280 { while (*f>f0) | |
4281 { lb = lb/2.0; | |
4282 for (i=0; i<p; i++) coef[i] = oc[i]+lb*dlt[i]; | |
4283 st = mmsums(des,coef,f,des->f1,&des->xtwx); | |
4284 } | |
4285 return(st); | |
4286 } | |
4287 | |
4288 if (!af) return(st); | |
4289 | |
4290 /* double */ | |
4291 while (*f<f0) | |
4292 { f0 = *f; | |
4293 lb *= 2.0; | |
4294 for (i=0; i<p; i++) coef[i] = oc[i]+lb*dlt[i]; | |
4295 st = mmsums(des,coef,f,des->f1,&des->xtwx); | |
4296 } | |
4297 | |
4298 lb /= 2.0; | |
4299 for (i=0; i<p; i++) coef[i] = oc[i]+lb*dlt[i]; | |
4300 st = mmsums(des,coef,f,des->f1,&des->xtwx); | |
4301 | |
4302 return(st); | |
4303 } | |
4304 | |
4305 int mm_initial(des) | |
4306 design *des; | |
4307 { double *dlt; | |
4308 | |
4309 dlt = des->ss; | |
4310 | |
4311 setzero(des->cf,des->p); | |
4312 st = mmsums(des,des->cf,&mmf,des->f1,&des->xtwx); | |
4313 | |
4314 setzero(dlt,des->p); | |
4315 dlt[0] = 1; | |
4316 lb = 1.0; | |
4317 st = descenddir(des,des->cf,dlt,&mmf,1); | |
4318 return(st); | |
4319 } | |
4320 | |
4321 void getsingdir(des,dlt) | |
4322 design *des; | |
4323 double *dlt; | |
4324 { double f, sw, c0; | |
4325 int i, j, p, sd; | |
4326 | |
4327 sd = -1; p = des->p; | |
4328 setzero(dlt,p); | |
4329 for (i=0; i<p; i++) if (des->xtwx.Z[i*p+i]<SINGTOL) sd = i; | |
4330 if (sd==-1) | |
4331 { mut_printf("getsingdir: nonsing?\n"); | |
4332 return; | |
4333 } | |
4334 if (des->xtwx.dg[sd]>0) | |
4335 for (i=0; i<p; i++) dlt[i] = des->xtwx.Q[p*i+sd]*des->xtwx.dg[i]; | |
4336 else | |
4337 { dlt[sd] = 1.0; | |
4338 } | |
4339 | |
4340 c0 = innerprod(dlt,des->f1,p); | |
4341 if (c0<0) for (i=0; i<p; i++) dlt[i] = -dlt[i]; | |
4342 } | |
4343 | |
4344 void mmax(coef, old_coef, delta, J, p, maxit, tol, err) | |
4345 double *coef, *old_coef, *delta, tol; | |
4346 int p, maxit, *err; | |
4347 jacobian *J; | |
4348 { double old_f, lambda; | |
4349 int i, j; | |
4350 | |
4351 *err = NR_OK; | |
4352 | |
4353 for (j=0; j<maxit; j++) | |
4354 { memcpy(old_coef,coef,p*sizeof(double)); | |
4355 old_f = mmf; | |
4356 | |
4357 if (st == NR_SINGULAR) | |
4358 { | |
4359 getsingdir(mm_des,delta); | |
4360 st = descenddir(mm_des,coef,delta,&mmf,1); | |
4361 } | |
4362 if (st == NR_EMPTY) | |
4363 { | |
4364 setzero(delta,p); | |
4365 delta[0] = 1.0; | |
4366 st = descenddir(mm_des,coef,delta,&mmf,1); | |
4367 } | |
4368 if (st == NR_OK) | |
4369 { | |
4370 lb = 1.0; | |
4371 jacob_solve(J,mm_des->f1); | |
4372 memcpy(delta,mm_des->f1,p*sizeof(double)); | |
4373 st = descenddir(mm_des,coef,delta,&mmf,0); | |
4374 } | |
4375 | |
4376 if ((j>0) & (fabs(mmf-old_f)<tol)) return; | |
4377 } | |
4378 WARN(("findab not converged")); | |
4379 *err = NR_NCON; | |
4380 return; | |
4381 } | |
4382 | |
4383 double findab(gam) | |
4384 double gam; | |
4385 { double sl; | |
4386 int i, p, nr_stat; | |
4387 | |
4388 if (debug) mut_printf(" findab: gam %8.5f\n",gam); | |
4389 mm_gam = gam; | |
4390 p = mm_des->p; | |
4391 lb = 1.0; | |
4392 st = mm_initial(mm_des); | |
4393 | |
4394 mmax(mm_des->cf, mm_des->oc, mm_des->ss, | |
4395 &mm_des->xtwx, p, lf_maxit, CONVTOL, &nr_stat); | |
4396 | |
4397 sl = 0.0; | |
4398 for (i=0; i<mm_lfd->n; i++) sl += fabs(wght(mm_des,i))*mm_des->wd[i]; | |
4399 | |
4400 if (debug) mut_printf(" sl %8.5f gam %8.5f %8.5f %d\n", sl,gam,sl-gam,nr_stat); | |
4401 return(sl-gam); | |
4402 } | |
4403 | |
4404 double weightmm(coef,di,ff,gam) | |
4405 double *coef, di, *ff, gam; | |
4406 { double y1, y2, ip; | |
4407 ip = innerprod(ff,coef,mm_des->p); | |
4408 y1 = ip-gam*di; if (y1>0) return(y1/ip); | |
4409 y2 = ip+gam*di; if (y2<0) return(y2/ip); | |
4410 return(0.0); | |
4411 } | |
4412 | |
4413 double minmax(lfd,des,sp) | |
4414 lfdata *lfd; | |
4415 design *des; | |
4416 smpar *sp; | |
4417 { double h, u[MXDIM], gam; | |
4418 int i, j, m, d1, p1, err_flag; | |
4419 | |
4420 if (debug) mut_printf("minmax: x %8.5f\n",des->xev[0]); | |
4421 mm_lfd = lfd; | |
4422 mm_des = des; | |
4423 | |
4424 mmsm_ct = 0; | |
4425 d1 = deg(sp)+1; | |
4426 p1 = factorial(d1); | |
4427 for (i=0; i<lfd->n; i++) | |
4428 { for (j=0; j<lfd->d; j++) u[j] = datum(lfd,j,i); | |
4429 des->wd[i] = sp->nn/p1*ipower(dist(des,i),d1); | |
4430 des->ind[i] = i; | |
4431 fitfun(lfd, sp, u,des->xev,d_xi(des,i),NULL); | |
4432 } | |
4433 | |
4434 /* find gamma (i.e. solve eqn 13.17 from book), using the secant method. | |
4435 * As a side effect, this finds the other minimax coefficients. | |
4436 * Note that 13.17 is rewritten as | |
4437 * g2 = sum |l_i(x)| (||xi-x||^(p+1) M/(s*(p+1)!)) | |
4438 * where g2 = gamma * s * (p+1)! / M. The gam variable below is g2. | |
4439 * The smoothing parameter is sp->nn == M/s. | |
4440 */ | |
4441 gam = solve_secant(findab, 0.0, 0.0,1.0, 0.0000001, BDF_EXPRIGHT, &err_flag); | |
4442 | |
4443 /* | |
4444 * Set the smoothing weights, in preparation for the actual fit. | |
4445 */ | |
4446 h = 0.0; m = 0; | |
4447 for (i=0; i<lfd->n; i++) | |
4448 { wght(des,i) = weightmm(des->cf, des->wd[i],d_xi(des,i),gam); | |
4449 if (wght(des,i)>0) | |
4450 { if (dist(des,i)>h) h = dist(des,i); | |
4451 des->ind[m] = i; | |
4452 m++; | |
4453 } | |
4454 } | |
4455 des->n = m; | |
4456 return(h); | |
4457 } | |
4458 /* | |
4459 * Copyright 1996-2006 Catherine Loader. | |
4460 */ | |
4461 /* | |
4462 * | |
4463 * Defines the weight functions and related quantities used | |
4464 * in LOCFIT. | |
4465 */ | |
4466 | |
4467 #include "locf.h" | |
4468 | |
4469 /* | |
4470 * convert kernel and kernel type strings to numeric codes. | |
4471 */ | |
4472 #define NWFUNS 13 | |
4473 static char *wfuns[NWFUNS] = { | |
4474 "rectangular", "epanechnikov", "bisquare", "tricube", | |
4475 "triweight", "gaussian", "triangular", "ququ", | |
4476 "6cub", "minimax", "exponential", "maclean", "parametric" }; | |
4477 static int wvals[NWFUNS] = { WRECT, WEPAN, WBISQ, WTCUB, | |
4478 WTRWT, WGAUS, WTRIA, WQUQU, W6CUB, WMINM, WEXPL, WMACL, WPARM }; | |
4479 int lfkernel(char *z) | |
4480 { return(pmatch(z, wfuns, wvals, NWFUNS, WTCUB)); | |
4481 } | |
4482 | |
4483 #define NKTYPE 5 | |
4484 static char *ktype[NKTYPE] = { "spherical", "product", "center", "lm", "zeon" }; | |
4485 static int kvals[NKTYPE] = { KSPH, KPROD, KCE, KLM, KZEON }; | |
4486 int lfketype(char *z) | |
4487 { return(pmatch(z, ktype, kvals, NKTYPE, KSPH)); | |
4488 } | |
4489 | |
4490 /* The weight functions themselves. Used everywhere. */ | |
4491 double W(u,ker) | |
4492 double u; | |
4493 int ker; | |
4494 { u = fabs(u); | |
4495 switch(ker) | |
4496 { case WRECT: return((u>1) ? 0.0 : 1.0); | |
4497 case WEPAN: return((u>1) ? 0.0 : 1-u*u); | |
4498 case WBISQ: if (u>1) return(0.0); | |
4499 u = 1-u*u; return(u*u); | |
4500 case WTCUB: if (u>1) return(0.0); | |
4501 u = 1-u*u*u; return(u*u*u); | |
4502 case WTRWT: if (u>1) return(0.0); | |
4503 u = 1-u*u; return(u*u*u); | |
4504 case WQUQU: if (u>1) return(0.0); | |
4505 u = 1-u*u; return(u*u*u*u); | |
4506 case WTRIA: if (u>1) return(0.0); | |
4507 return(1-u); | |
4508 case W6CUB: if (u>1) return(0.0); | |
4509 u = 1-u*u*u; u = u*u*u; return(u*u); | |
4510 case WGAUS: return(exp(-SQR(GFACT*u)/2.0)); | |
4511 case WEXPL: return(exp(-EFACT*u)); | |
4512 case WMACL: return(1/((u+1.0e-100)*(u+1.0e-100))); | |
4513 case WMINM: LERR(("WMINM in W")); | |
4514 return(0.0); | |
4515 case WPARM: return(1.0); | |
4516 } | |
4517 LERR(("W(): Unknown kernel %d\n",ker)); | |
4518 return(1.0); | |
4519 } | |
4520 | |
4521 int iscompact(ker) | |
4522 int ker; | |
4523 { if ((ker==WEXPL) | (ker==WGAUS) | (ker==WMACL) | (ker==WPARM)) return(0); | |
4524 return(1); | |
4525 } | |
4526 | |
4527 double weightprod(lfd,u,h,ker) | |
4528 lfdata *lfd; | |
4529 double *u, h; | |
4530 int ker; | |
4531 { int i; | |
4532 double sc, w; | |
4533 w = 1.0; | |
4534 for (i=0; i<lfd->d; i++) | |
4535 { sc = lfd->sca[i]; | |
4536 switch(lfd->sty[i]) | |
4537 { case STLEFT: | |
4538 if (u[i]>0) return(0.0); | |
4539 w *= W(-u[i]/(h*sc),ker); | |
4540 break; | |
4541 case STRIGH: | |
4542 if (u[i]<0) return(0.0); | |
4543 w *= W(u[i]/(h*sc),ker); | |
4544 break; | |
4545 case STANGL: | |
4546 w *= W(2*fabs(sin(u[i]/(2*sc)))/h,ker); | |
4547 break; | |
4548 case STCPAR: | |
4549 break; | |
4550 default: | |
4551 w *= W(fabs(u[i])/(h*sc),ker); | |
4552 } | |
4553 if (w==0.0) return(w); | |
4554 } | |
4555 return(w); | |
4556 } | |
4557 | |
4558 double weightsph(lfd,u,h,ker, hasdi,di) | |
4559 lfdata *lfd; | |
4560 double *u, h, di; | |
4561 int ker, hasdi; | |
4562 { int i; | |
4563 | |
4564 if (!hasdi) di = rho(u,lfd->sca,lfd->d,KSPH,lfd->sty); | |
4565 | |
4566 for (i=0; i<lfd->d; i++) | |
4567 { if ((lfd->sty[i]==STLEFT) && (u[i]>0.0)) return(0.0); | |
4568 if ((lfd->sty[i]==STRIGH) && (u[i]<0.0)) return(0.0); | |
4569 } | |
4570 if (h==0) return((di==0.0) ? 1.0 : 0.0); | |
4571 | |
4572 return(W(di/h,ker)); | |
4573 } | |
4574 | |
4575 double weight(lfd,sp,x,t,h, hasdi,di) | |
4576 lfdata *lfd; | |
4577 smpar *sp; | |
4578 double *x, *t, h, di; | |
4579 int hasdi; | |
4580 { double u[MXDIM]; | |
4581 int i; | |
4582 for (i=0; i<lfd->d; i++) u[i] = (t==NULL) ? x[i] : x[i]-t[i]; | |
4583 switch(kt(sp)) | |
4584 { case KPROD: return(weightprod(lfd,u,h,ker(sp))); | |
4585 case KSPH: return(weightsph(lfd,u,h,ker(sp), hasdi,di)); | |
4586 } | |
4587 LERR(("weight: unknown kernel type %d",kt(sp))); | |
4588 return(1.0); | |
4589 } | |
4590 | |
4591 double sgn(x) | |
4592 double x; | |
4593 { if (x>0) return(1.0); | |
4594 if (x<0) return(-1.0); | |
4595 return(0.0); | |
4596 } | |
4597 | |
4598 double WdW(u,ker) /* W'(u)/W(u) */ | |
4599 double u; | |
4600 int ker; | |
4601 { double eps=1.0e-10; | |
4602 if (ker==WGAUS) return(-GFACT*GFACT*u); | |
4603 if (ker==WPARM) return(0.0); | |
4604 if (fabs(u)>=1) return(0.0); | |
4605 switch(ker) | |
4606 { case WRECT: return(0.0); | |
4607 case WTRIA: return(-sgn(u)/(1-fabs(u)+eps)); | |
4608 case WEPAN: return(-2*u/(1-u*u+eps)); | |
4609 case WBISQ: return(-4*u/(1-u*u+eps)); | |
4610 case WTRWT: return(-6*u/(1-u*u+eps)); | |
4611 case WTCUB: return(-9*sgn(u)*u*u/(1-u*u*fabs(u)+eps)); | |
4612 case WEXPL: return((u>0) ? -EFACT : EFACT); | |
4613 } | |
4614 LERR(("WdW: invalid kernel")); | |
4615 return(0.0); | |
4616 } | |
4617 | |
4618 /* deriv. weights .. spherical, product etc | |
4619 u, sc, sty needed only in relevant direction | |
4620 Acutally, returns (d/dx W(||x||/h) ) / W(.) | |
4621 */ | |
4622 double weightd(u,sc,d,ker,kt,h,sty,di) | |
4623 double u, sc, h, di; | |
4624 int d, ker, kt, sty; | |
4625 { if (sty==STANGL) | |
4626 { if (kt==KPROD) | |
4627 return(-WdW(2*sin(u/(2*sc)),ker)*cos(u/(2*sc))/(h*sc)); | |
4628 if (di==0.0) return(0.0); | |
4629 return(-WdW(di/h,ker)*sin(u/sc)/(h*sc*di)); | |
4630 } | |
4631 if (sty==STCPAR) return(0.0); | |
4632 if (kt==KPROD) | |
4633 return(-WdW(u/(h*sc),ker)/(h*sc)); | |
4634 if (di==0.0) return(0.0); | |
4635 return(-WdW(di/h,ker)*u/(h*di*sc*sc)); | |
4636 } | |
4637 | |
4638 double weightdd(u,sc,d,ker,kt,h,sty,di,i0,i1) | |
4639 double *u, *sc, h, di; | |
4640 int d, ker, kt, i0, i1, *sty; | |
4641 { double w; | |
4642 w = 1; | |
4643 if (kt==KPROD) | |
4644 { | |
4645 w = WdW(u[i0]/(h*sc[i0]),ker)*WdW(u[i1]/(h*sc[i1]),ker)/(h*h*sc[i0]*sc[i1]); | |
4646 } | |
4647 return(0.0); | |
4648 } | |
4649 | |
4650 /* Derivatives W'(u)/u. | |
4651 Used in simult. conf. band computations, | |
4652 and kernel density bandwidth selectors. */ | |
4653 double Wd(u,ker) | |
4654 double u; | |
4655 int ker; | |
4656 { double v; | |
4657 if (ker==WGAUS) return(-SQR(GFACT)*exp(-SQR(GFACT*u)/2)); | |
4658 if (ker==WPARM) return(0.0); | |
4659 if (fabs(u)>1) return(0.0); | |
4660 switch(ker) | |
4661 { case WEPAN: return(-2.0); | |
4662 case WBISQ: return(-4*(1-u*u)); | |
4663 case WTCUB: v = 1-u*u*u; | |
4664 return(-9*v*v*u); | |
4665 case WTRWT: v = 1-u*u; | |
4666 return(-6*v*v); | |
4667 default: LERR(("Invalid kernel %d in Wd",ker)); | |
4668 } | |
4669 return(0.0); | |
4670 } | |
4671 | |
4672 /* Second derivatives W''(u)-W'(u)/u. | |
4673 used in simult. conf. band computations in >1 dimension. */ | |
4674 double Wdd(u,ker) | |
4675 double u; | |
4676 int ker; | |
4677 { double v; | |
4678 if (ker==WGAUS) return(SQR(u*GFACT*GFACT)*exp(-SQR(u*GFACT)/2)); | |
4679 if (ker==WPARM) return(0.0); | |
4680 if (u>1) return(0.0); | |
4681 switch(ker) | |
4682 { case WBISQ: return(12*u*u); | |
4683 case WTCUB: v = 1-u*u*u; | |
4684 return(-9*u*v*v+54*u*u*u*u*v); | |
4685 case WTRWT: return(24*u*u*(1-u*u)); | |
4686 default: LERR(("Invalid kernel %d in Wdd",ker)); | |
4687 } | |
4688 return(0.0); | |
4689 } | |
4690 | |
4691 /* int u1^j1..ud^jd W(u) du. | |
4692 Used for local log-linear density estimation. | |
4693 Assume all j_i are even. | |
4694 Also in some bandwidth selection. | |
4695 */ | |
4696 double wint(d,j,nj,ker) | |
4697 int d, *j, nj, ker; | |
4698 { double I, z; | |
4699 int k, dj; | |
4700 dj = d; | |
4701 for (k=0; k<nj; k++) dj += j[k]; | |
4702 switch(ker) /* int_0^1 u^(dj-1) W(u)du */ | |
4703 { case WRECT: I = 1.0/dj; break; | |
4704 case WEPAN: I = 2.0/(dj*(dj+2)); break; | |
4705 case WBISQ: I = 8.0/(dj*(dj+2)*(dj+4)); break; | |
4706 case WTCUB: I = 162.0/(dj*(dj+3)*(dj+6)*(dj+9)); break; | |
4707 case WTRWT: I = 48.0/(dj*(dj+2)*(dj+4)*(dj+6)); break; | |
4708 case WTRIA: I = 1.0/(dj*(dj+1)); break; | |
4709 case WQUQU: I = 384.0/(dj*(dj+2)*(dj+4)*(dj+6)*(dj+8)); break; | |
4710 case W6CUB: I = 524880.0/(dj*(dj+3)*(dj+6)*(dj+9)*(dj+12)*(dj+15)*(dj+18)); break; | |
4711 case WGAUS: switch(d) | |
4712 { case 1: I = S2PI/GFACT; break; | |
4713 case 2: I = 2*PI/(GFACT*GFACT); break; | |
4714 default: I = exp(d*log(S2PI/GFACT)); /* for nj=0 */ | |
4715 } | |
4716 for (k=0; k<nj; k++) /* deliberate drop */ | |
4717 switch(j[k]) | |
4718 { case 4: I *= 3.0/(GFACT*GFACT); | |
4719 case 2: I /= GFACT*GFACT; | |
4720 } | |
4721 return(I); | |
4722 case WEXPL: I = factorial(dj-1)/ipower(EFACT,dj); break; | |
4723 default: LERR(("Unknown kernel %d in exacint",ker)); | |
4724 } | |
4725 if ((d==1) && (nj==0)) return(2*I); /* common case quick */ | |
4726 z = (d-nj)*LOGPI/2-mut_lgammai(dj); | |
4727 for (k=0; k<nj; k++) z += mut_lgammai(j[k]+1); | |
4728 return(2*I*exp(z)); | |
4729 } | |
4730 | |
4731 /* taylor series expansion of weight function around x. | |
4732 0 and 1 are common arguments, so are worth programming | |
4733 as special cases. | |
4734 Used in density estimation. | |
4735 */ | |
4736 int wtaylor(f,x,ker) | |
4737 double *f, x; | |
4738 int ker; | |
4739 { double v; | |
4740 switch(ker) | |
4741 { case WRECT: | |
4742 f[0] = 1.0; | |
4743 return(1); | |
4744 case WEPAN: | |
4745 f[0] = 1-x*x; f[1] = -2*x; f[2] = -1; | |
4746 return(3); | |
4747 case WBISQ: | |
4748 v = 1-x*x; | |
4749 f[0] = v*v; f[1] = -4*x*v; f[2] = 4-6*v; | |
4750 f[3] = 4*x; f[4] = 1; | |
4751 return(5); | |
4752 case WTCUB: | |
4753 if (x==1.0) | |
4754 { f[0] = f[1] = f[2] = 0; f[3] = -27; f[4] = -81; f[5] = -108; | |
4755 f[6] = -81; f[7] = -36; f[8] = -9; f[9] = -1; return(10); } | |
4756 if (x==0.0) | |
4757 { f[1] = f[2] = f[4] = f[5] = f[7] = f[8] = 0; | |
4758 f[0] = 1; f[3] = -3; f[6] = 3; f[9] = -1; return(10); } | |
4759 v = 1-x*x*x; | |
4760 f[0] = v*v*v; f[1] = -9*v*v*x*x; f[2] = x*v*(27-36*v); | |
4761 f[3] = -27+v*(108-84*v); f[4] = -3*x*x*(27-42*v); | |
4762 f[5] = x*(-108+126*v); f[6] = -81+84*v; | |
4763 f[7] = -36*x*x; f[8] = -9*x; f[9] = -1; | |
4764 return(10); | |
4765 case WTRWT: | |
4766 v = 1-x*x; | |
4767 f[0] = v*v*v; f[1] = -6*x*v*v; f[2] = v*(12-15*v); | |
4768 f[3] = x*(20*v-8); f[4] = 15*v-12; f[5] = -6; f[6] = -1; | |
4769 return(7); | |
4770 case WTRIA: | |
4771 f[0] = 1-x; f[1] = -1; | |
4772 return(2); | |
4773 case WQUQU: | |
4774 v = 1-x*x; | |
4775 f[0] = v*v*v*v; f[1] = -8*x*v*v*v; f[2] = v*v*(24-28*v); | |
4776 f[3] = v*x*(56*v-32); f[4] = (70*v-80)*v+16; f[5] = x*(32-56*v); | |
4777 f[6] = 24-28*v; f[7] = 8*x; f[8] = 1; | |
4778 return(9); | |
4779 case W6CUB: | |
4780 v = 1-x*x*x; | |
4781 f[0] = v*v*v*v*v*v; | |
4782 f[1] = -18*x*x*v*v*v*v*v; | |
4783 f[2] = x*v*v*v*v*(135-153*v); | |
4784 f[3] = v*v*v*(-540+v*(1350-816*v)); | |
4785 f[4] = x*x*v*v*(1215-v*(4050-v*3060)); | |
4786 f[5] = x*v*(-1458+v*(9234+v*(-16254+v*8568))); | |
4787 f[6] = 729-v*(10206-v*(35154-v*(44226-v*18564))); | |
4788 f[7] = x*x*(4374-v*(30132-v*(56862-v*31824))); | |
4789 f[8] = x*(12393-v*(61479-v*(92664-v*43758))); | |
4790 f[9] = 21870-v*(89100-v*(115830-v*48620)); | |
4791 f[10]= x*x*(26730-v*(69498-v*43758)); | |
4792 f[11]= x*(23814-v*(55458-v*31824)); | |
4793 f[12]= 15849-v*(34398-v*18564); | |
4794 f[13]= x*x*(7938-8568*v); | |
4795 f[14]= x*(2970-3060*v); | |
4796 f[15]= 810-816*v; | |
4797 f[16]= 153*x*x; | |
4798 f[17]= 18*x; | |
4799 f[18]= 1; | |
4800 return(19); | |
4801 } | |
4802 LERR(("Invalid kernel %d in wtaylor",ker)); | |
4803 return(0); | |
4804 } | |
4805 | |
4806 /* convolution int W(x)W(x+v)dx. | |
4807 used in kde bandwidth selection. | |
4808 */ | |
4809 double Wconv(v,ker) | |
4810 double v; | |
4811 int ker; | |
4812 { double v2; | |
4813 switch(ker) | |
4814 { case WGAUS: return(SQRPI/GFACT*exp(-SQR(GFACT*v)/4)); | |
4815 case WRECT: | |
4816 v = fabs(v); | |
4817 if (v>2) return(0.0); | |
4818 return(2-v); | |
4819 case WEPAN: | |
4820 v = fabs(v); | |
4821 if (v>2) return(0.0); | |
4822 return((2-v)*(16+v*(8-v*(16-v*(2+v))))/30); | |
4823 case WBISQ: | |
4824 v = fabs(v); | |
4825 if (v>2) return(0.0); | |
4826 v2 = 2-v; | |
4827 return(v2*v2*v2*v2*v2*(16+v*(40+v*(36+v*(10+v))))/630); | |
4828 } | |
4829 LERR(("Wconv not implemented for kernel %d",ker)); | |
4830 return(0.0); | |
4831 } | |
4832 | |
4833 /* derivative of Wconv. | |
4834 1/v d/dv int W(x)W(x+v)dx | |
4835 used in kde bandwidth selection. | |
4836 */ | |
4837 double Wconv1(v,ker) | |
4838 double v; | |
4839 int ker; | |
4840 { double v2; | |
4841 v = fabs(v); | |
4842 switch(ker) | |
4843 { case WGAUS: return(-0.5*SQRPI*GFACT*exp(-SQR(GFACT*v)/4)); | |
4844 case WRECT: | |
4845 if (v>2) return(0.0); | |
4846 return(1.0); | |
4847 case WEPAN: | |
4848 if (v>2) return(0.0); | |
4849 return((-16+v*(12-v*v))/6); | |
4850 case WBISQ: | |
4851 if (v>2) return(0.0); | |
4852 v2 = 2-v; | |
4853 return(-v2*v2*v2*v2*(32+v*(64+v*(24+v*3)))/210); | |
4854 } | |
4855 LERR(("Wconv1 not implemented for kernel %d",ker)); | |
4856 return(0.0); | |
4857 } | |
4858 | |
4859 /* 4th derivative of Wconv. | |
4860 used in kde bandwidth selection (BCV, SJPI, GKK) | |
4861 */ | |
4862 double Wconv4(v,ker) | |
4863 double v; | |
4864 int ker; | |
4865 { double gv; | |
4866 switch(ker) | |
4867 { case WGAUS: | |
4868 gv = GFACT*v; | |
4869 return(exp(-SQR(gv)/4)*GFACT*GFACT*GFACT*(12-gv*gv*(12-gv*gv))*SQRPI/16); | |
4870 } | |
4871 LERR(("Wconv4 not implemented for kernel %d",ker)); | |
4872 return(0.0); | |
4873 } | |
4874 | |
4875 /* 5th derivative of Wconv. | |
4876 used in kde bandwidth selection (BCV method only) | |
4877 */ | |
4878 double Wconv5(v,ker) /* (d/dv)^5 int W(x)W(x+v)dx */ | |
4879 double v; | |
4880 int ker; | |
4881 { double gv; | |
4882 switch(ker) | |
4883 { case WGAUS: | |
4884 gv = GFACT*v; | |
4885 return(-exp(-SQR(gv)/4)*GFACT*GFACT*GFACT*GFACT*gv*(60-gv*gv*(20-gv*gv))*SQRPI/32); | |
4886 } | |
4887 LERR(("Wconv5 not implemented for kernel %d",ker)); | |
4888 return(0.0); | |
4889 } | |
4890 | |
4891 /* 6th derivative of Wconv. | |
4892 used in kde bandwidth selection (SJPI) | |
4893 */ | |
4894 double Wconv6(v,ker) | |
4895 double v; | |
4896 int ker; | |
4897 { double gv, z; | |
4898 switch(ker) | |
4899 { case WGAUS: | |
4900 gv = GFACT*v; | |
4901 gv = gv*gv; | |
4902 z = exp(-gv/4)*(-120+gv*(180-gv*(30-gv)))*0.02769459142; | |
4903 gv = GFACT*GFACT; | |
4904 return(z*gv*gv*GFACT); | |
4905 } | |
4906 LERR(("Wconv6 not implemented for kernel %d",ker)); | |
4907 return(0.0); | |
4908 } | |
4909 | |
4910 /* int W(v)^2 dv / (int v^2 W(v) dv)^2 | |
4911 used in some bandwidth selectors | |
4912 */ | |
4913 double Wikk(ker,deg) | |
4914 int ker, deg; | |
4915 { switch(deg) | |
4916 { case 0: | |
4917 case 1: /* int W(v)^2 dv / (int v^2 W(v) dv)^2 */ | |
4918 switch(ker) | |
4919 { case WRECT: return(4.5); | |
4920 case WEPAN: return(15.0); | |
4921 case WBISQ: return(35.0); | |
4922 case WGAUS: return(0.2820947918*GFACT*GFACT*GFACT*GFACT*GFACT); | |
4923 case WTCUB: return(34.152111046847892); /* 59049 / 1729 */ | |
4924 case WTRWT: return(66.083916083916080); /* 9450/143 */ | |
4925 } | |
4926 case 2: | |
4927 case 3: /* 4!^2/8*int(W1^2)/int(v^4W1)^2 | |
4928 W1=W*(n4-v^2n2)/(n0n4-n2n2) */ | |
4929 switch(ker) | |
4930 { case WRECT: return(11025.0); | |
4931 case WEPAN: return(39690.0); | |
4932 case WBISQ: return(110346.9231); | |
4933 case WGAUS: return(14527.43412); | |
4934 case WTCUB: return(126500.5904); | |
4935 case WTRWT: return(254371.7647); | |
4936 } | |
4937 } | |
4938 LERR(("Wikk not implemented for kernel %d, deg %d",ker,deg)); | |
4939 return(0.0); | |
4940 } |