align C version of EMGLLF on R version
[valse.git] / pkg / src / sources / EMGLLF.c
CommitLineData
8e92c49c
BA
1#include "utils.h"
2#include <stdlib.h>
1d3c1faa
BA
3#include <gsl/gsl_linalg.h>
4
b42f0f40 5// TODO: don't recompute indexes ai(...) and mi(...) when possible
09ab3c16 6void EMGLLF_core(
1d3c1faa 7 // IN parameters
9ff729fb
BA
8 const Real* phiInit, // parametre initial de moyenne renormalisé
9 const Real* rhoInit, // parametre initial de variance renormalisé
10 const Real* piInit, // parametre initial des proportions
11 const Real* gamInit, // paramètre initial des probabilités a posteriori de chaque échantillon
8e92c49c
BA
12 int mini, // nombre minimal d'itérations dans l'algorithme EM
13 int maxi, // nombre maximal d'itérations dans l'algorithme EM
9ff729fb
BA
14 Real gamma, // puissance des proportions dans la pénalisation pour un Lasso adaptatif
15 Real lambda, // valeur du paramètre de régularisation du Lasso
16 const Real* X, // régresseurs
17 const Real* Y, // réponse
18 Real tau, // seuil pour accepter la convergence
1d3c1faa 19 // OUT parameters (all pointers, to be modified)
9ff729fb
BA
20 Real* phi, // parametre de moyenne renormalisé, calculé par l'EM
21 Real* rho, // parametre de variance renormalisé, calculé par l'EM
22 Real* pi, // parametre des proportions renormalisé, calculé par l'EM
23 Real* LLF, // log vraisemblance associée à cet échantillon, pour les valeurs estimées des paramètres
24 Real* S,
1d3c1faa 25 // additional size parameters
8e92c49c
BA
26 int n, // nombre d'echantillons
27 int p, // nombre de covariables
28 int m, // taille de Y (multivarié)
29 int k) // nombre de composantes dans le mélange
1d3c1faa
BA
30{
31 //Initialize outputs
32 copyArray(phiInit, phi, p*m*k);
33 copyArray(rhoInit, rho, m*m*k);
34 copyArray(piInit, pi, k);
35 zeroArray(LLF, maxi);
36 //S is already allocated, and doesn't need to be 'zeroed'
4cab944a 37
b42f0f40 38 //Other local variables: same as in R
9ff729fb 39 Real* gam = (Real*)malloc(n*k*sizeof(Real));
1d3c1faa 40 copyArray(gamInit, gam, n*k);
b42f0f40
BA
41 Real* Gram2 = (Real*)malloc(p*p*k*sizeof(Real));
42 Real* ps2 = (Real*)malloc(p*m*k*sizeof(Real));
9ff729fb 43 Real* b = (Real*)malloc(k*sizeof(Real));
b42f0f40
BA
44 Real* X2 = (Real*)malloc(n*p*k*sizeof(Real));
45 Real* Y2 = (Real*)malloc(n*m*k*sizeof(Real));
46 Real dist = 0.;
47 Real dist2 = 0.;
48 int ite = 0;
9ff729fb 49 Real* pi2 = (Real*)malloc(k*sizeof(Real));
9ff729fb
BA
50 Real* ps = (Real*)malloc(m*k*sizeof(Real));
51 Real* nY2 = (Real*)malloc(m*k*sizeof(Real));
52 Real* ps1 = (Real*)malloc(n*m*k*sizeof(Real));
9ff729fb 53 Real* Gam = (Real*)malloc(n*k*sizeof(Real));
b42f0f40
BA
54 const Real EPS = 1e-15;
55 // Additional (not at this place, in R file)
56 Real* gam2 = (Real*)malloc(k*sizeof(Real));
57 Real* nY21 = (Real*)malloc(n*m*k*sizeof(Real));
ef67d338 58 Real* sqNorm2 = (Real*)malloc(k*sizeof(Real));
1d3c1faa
BA
59 gsl_matrix* matrix = gsl_matrix_alloc(m, m);
60 gsl_permutation* permutation = gsl_permutation_alloc(m);
9ff729fb
BA
61 Real* YiRhoR = (Real*)malloc(m*sizeof(Real));
62 Real* XiPhiR = (Real*)malloc(m*sizeof(Real));
ef67d338 63 const Real gaussConstM = pow(2.*M_PI,m/2.);
b42f0f40
BA
64 Real* Phi = (Real*)malloc(p*m*k*sizeof(Real));
65 Real* Rho = (Real*)malloc(m*m*k*sizeof(Real));
66 Real* Pi = (Real*)malloc(k*sizeof(Real));
4cab944a 67
1d3c1faa
BA
68 while (ite < mini || (ite < maxi && (dist >= tau || dist2 >= sqrt(tau))))
69 {
70 copyArray(phi, Phi, p*m*k);
71 copyArray(rho, Rho, m*m*k);
72 copyArray(pi, Pi, k);
4cab944a
BA
73
74 // Calculs associés a Y et X
75 for (int r=0; r<k; r++)
1d3c1faa 76 {
4cab944a 77 for (int mm=0; mm<m; mm++)
1d3c1faa 78 {
ef67d338 79 //Y2[,mm,r] = sqrt(gam[,r]) * Y[,mm]
4cab944a
BA
80 for (int u=0; u<n; u++)
81 Y2[ai(u,mm,r,n,m,k)] = sqrt(gam[mi(u,r,n,k)]) * Y[mi(u,mm,m,n)];
1d3c1faa 82 }
4cab944a 83 for (int i=0; i<n; i++)
1d3c1faa 84 {
ef67d338 85 //X2[i,,r] = sqrt(gam[i,r]) * X[i,]
4cab944a 86 for (int u=0; u<p; u++)
e39bc178 87 X2[ai(i,u,r,n,p,k)] = sqrt(gam[mi(i,r,n,k)]) * X[mi(i,u,n,p)];
1d3c1faa 88 }
4cab944a 89 for (int mm=0; mm<m; mm++)
1d3c1faa 90 {
ef67d338 91 //ps2[,mm,r] = crossprod(X2[,,r],Y2[,mm,r])
4cab944a 92 for (int u=0; u<p; u++)
1d3c1faa 93 {
9ff729fb 94 Real dotProduct = 0.;
4cab944a 95 for (int v=0; v<n; v++)
46a2e676 96 dotProduct += X2[ai(v,u,r,n,p,k)] * Y2[ai(v,mm,r,n,m,k)];
e39bc178 97 ps2[ai(u,mm,r,p,m,k)] = dotProduct;
1d3c1faa
BA
98 }
99 }
4cab944a 100 for (int j=0; j<p; j++)
1d3c1faa 101 {
4cab944a 102 for (int s=0; s<p; s++)
1d3c1faa 103 {
ef67d338 104 //Gram2[j,s,r] = crossprod(X2[,j,r], X2[,s,r])
9ff729fb 105 Real dotProduct = 0.;
4cab944a
BA
106 for (int u=0; u<n; u++)
107 dotProduct += X2[ai(u,j,r,n,p,k)] * X2[ai(u,s,r,n,p,k)];
108 Gram2[ai(j,s,r,p,p,k)] = dotProduct;
1d3c1faa
BA
109 }
110 }
111 }
112
113 /////////////
114 // Etape M //
115 /////////////
4cab944a 116
1d3c1faa 117 // Pour pi
4cab944a 118 for (int r=0; r<k; r++)
1d3c1faa 119 {
ef67d338 120 //b[r] = sum(abs(phi[,,r]))
9ff729fb 121 Real sumAbsPhi = 0.;
4cab944a
BA
122 for (int u=0; u<p; u++)
123 for (int v=0; v<m; v++)
124 sumAbsPhi += fabs(phi[ai(u,v,r,p,m,k)]);
1d3c1faa
BA
125 b[r] = sumAbsPhi;
126 }
ef67d338 127 //gam2 = colSums(gam)
4cab944a 128 for (int u=0; u<k; u++)
1d3c1faa 129 {
9ff729fb 130 Real sumOnColumn = 0.;
4cab944a
BA
131 for (int v=0; v<n; v++)
132 sumOnColumn += gam[mi(v,u,n,k)];
1d3c1faa
BA
133 gam2[u] = sumOnColumn;
134 }
ef67d338 135 //a = sum(gam %*% log(pi))
9ff729fb 136 Real a = 0.;
4cab944a 137 for (int u=0; u<n; u++)
1d3c1faa 138 {
9ff729fb 139 Real dotProduct = 0.;
4cab944a
BA
140 for (int v=0; v<k; v++)
141 dotProduct += gam[mi(u,v,n,k)] * log(pi[v]);
1d3c1faa
BA
142 a += dotProduct;
143 }
4cab944a 144
1d3c1faa 145 //tant que les proportions sont negatives
4cab944a 146 int kk = 0;
1d3c1faa 147 int pi2AllPositive = 0;
9ff729fb 148 Real invN = 1./n;
1d3c1faa
BA
149 while (!pi2AllPositive)
150 {
ef67d338
BA
151 //pi2 = pi + 0.1^kk * ((1/n)*gam2 - pi)
152 Real pow_01_kk = pow(0.1,kk);
4cab944a 153 for (int r=0; r<k; r++)
ef67d338
BA
154 pi2[r] = pi[r] + pow_01_kk * (invN*gam2[r] - pi[r]);
155 //pi2AllPositive = all(pi2 >= 0)
1d3c1faa 156 pi2AllPositive = 1;
4cab944a 157 for (int r=0; r<k; r++)
1d3c1faa
BA
158 {
159 if (pi2[r] < 0)
160 {
161 pi2AllPositive = 0;
162 break;
163 }
164 }
165 kk++;
166 }
4cab944a 167
1d3c1faa 168 //(pi.^gamma)*b
9ff729fb 169 Real piPowGammaDotB = 0.;
4cab944a 170 for (int v=0; v<k; v++)
1d3c1faa
BA
171 piPowGammaDotB += pow(pi[v],gamma) * b[v];
172 //(pi2.^gamma)*b
9ff729fb 173 Real pi2PowGammaDotB = 0.;
4cab944a 174 for (int v=0; v<k; v++)
1d3c1faa
BA
175 pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
176 //transpose(gam2)*log(pi2)
9ff729fb 177 Real prodGam2logPi2 = 0.;
4cab944a 178 for (int v=0; v<k; v++)
1d3c1faa 179 prodGam2logPi2 += gam2[v] * log(pi2[v]);
ef67d338 180 //t(m) la plus grande valeur dans la grille O.1^k tel que ce soit décroissante ou constante
8e92c49c
BA
181 while (-invN*a + lambda*piPowGammaDotB < -invN*prodGam2logPi2 + lambda*pi2PowGammaDotB
182 && kk<1000)
1d3c1faa 183 {
ef67d338
BA
184 Real pow_01_kk = pow(0.1,kk);
185 //pi2 = pi + 0.1^kk * (1/n*gam2 - pi)
4cab944a 186 for (int v=0; v<k; v++)
ef67d338 187 pi2[v] = pi[v] + pow_01_kk * (invN*gam2[v] - pi[v]);
1d3c1faa 188 //pi2 was updated, so we recompute pi2PowGammaDotB and prodGam2logPi2
4cab944a
BA
189 pi2PowGammaDotB = 0.;
190 for (int v=0; v<k; v++)
1d3c1faa 191 pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
4cab944a
BA
192 prodGam2logPi2 = 0.;
193 for (int v=0; v<k; v++)
1d3c1faa
BA
194 prodGam2logPi2 += gam2[v] * log(pi2[v]);
195 kk++;
196 }
9ff729fb 197 Real t = pow(0.1,kk);
ef67d338 198 //sum(pi + t*(pi2-pi))
9ff729fb 199 Real sumPiPlusTbyDiff = 0.;
4cab944a 200 for (int v=0; v<k; v++)
1d3c1faa 201 sumPiPlusTbyDiff += (pi[v] + t*(pi2[v] - pi[v]));
ef67d338 202 //pi = (pi + t*(pi2-pi)) / sum(pi + t*(pi2-pi))
4cab944a 203 for (int v=0; v<k; v++)
1d3c1faa 204 pi[v] = (pi[v] + t*(pi2[v] - pi[v])) / sumPiPlusTbyDiff;
4cab944a 205
1d3c1faa 206 //Pour phi et rho
4cab944a 207 for (int r=0; r<k; r++)
1d3c1faa 208 {
4cab944a 209 for (int mm=0; mm<m; mm++)
1d3c1faa 210 {
4cab944a 211 for (int i=0; i<n; i++)
1d3c1faa
BA
212 {
213 //< X2(i,:,r) , phi(:,mm,r) >
ef67d338 214 Real dotProduct = 0.;
4cab944a 215 for (int u=0; u<p; u++)
a2d68d1d 216 dotProduct += X2[ai(i,u,r,n,p,k)] * phi[ai(u,mm,r,p,m,k)];
ef67d338 217 //ps1[i,mm,r] = Y2[i,mm,r] * sum(X2[i,,r] * phi[,mm,r])
4cab944a
BA
218 ps1[ai(i,mm,r,n,m,k)] = Y2[ai(i,mm,r,n,m,k)] * dotProduct;
219 nY21[ai(i,mm,r,n,m,k)] = Y2[ai(i,mm,r,n,m,k)] * Y2[ai(i,mm,r,n,m,k)];
1d3c1faa 220 }
ef67d338
BA
221 //ps[mm,r] = sum(ps1[,mm,r])
222 Real sumPs1 = 0.;
4cab944a
BA
223 for (int u=0; u<n; u++)
224 sumPs1 += ps1[ai(u,mm,r,n,m,k)];
225 ps[mi(mm,r,m,k)] = sumPs1;
ef67d338
BA
226 //nY2[mm,r] = sum(nY21[,mm,r])
227 Real sumNy21 = 0.;
4cab944a
BA
228 for (int u=0; u<n; u++)
229 sumNy21 += nY21[ai(u,mm,r,n,m,k)];
230 nY2[mi(mm,r,m,k)] = sumNy21;
ef67d338 231 //rho[mm,mm,r] = (ps[mm,r]+sqrt(ps[mm,r]^2+4*nY2[mm,r]*(gam2[r]))) / (2*nY2[mm,r])
a2d68d1d 232 rho[ai(mm,mm,r,m,m,k)] = ( ps[mi(mm,r,m,k)] + sqrt( ps[mi(mm,r,m,k)]*ps[mi(mm,r,m,k)]
ef67d338 233 + 4*nY2[mi(mm,r,m,k)] * gam2[r] ) ) / (2*nY2[mi(mm,r,m,k)]);
1d3c1faa
BA
234 }
235 }
4cab944a 236 for (int r=0; r<k; r++)
1d3c1faa 237 {
4cab944a 238 for (int j=0; j<p; j++)
1d3c1faa 239 {
4cab944a 240 for (int mm=0; mm<m; mm++)
1d3c1faa 241 {
b42f0f40 242 //sum(phi[-j,mm,r] * Gram2[j, setdiff(1:p,j),r])
9ff729fb 243 Real dotPhiGram2 = 0.0;
b42f0f40
BA
244 for (int u=0; u<p; u++)
245 {
246 if (u != j)
247 dotPhiGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)];
248 }
249 //S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] + sum(phi[-j,mm,r] * Gram2[j, setdiff(1:p,j),r])
4cab944a 250 S[ai(j,mm,r,p,m,k)] = -rho[ai(mm,mm,r,m,m,k)] * ps2[ai(j,mm,r,p,m,k)] + dotPhiGram2;
ef67d338
BA
251 Real pow_pir_gamma = pow(pi[r],gamma);
252 if (fabs(S[ai(j,mm,r,p,m,k)]) <= n*lambda*pow_pir_gamma)
4cab944a 253 phi[ai(j,mm,r,p,m,k)] = 0;
ef67d338
BA
254 else if (S[ai(j,mm,r,p,m,k)] > n*lambda*pow_pir_gamma)
255 {
256 phi[ai(j,mm,r,p,m,k)] = (n*lambda*pow_pir_gamma - S[ai(j,mm,r,p,m,k)])
4cab944a 257 / Gram2[ai(j,j,r,p,p,k)];
ef67d338 258 }
1d3c1faa 259 else
ef67d338
BA
260 {
261 phi[ai(j,mm,r,p,m,k)] = -(n*lambda*pow_pir_gamma + S[ai(j,mm,r,p,m,k)])
4cab944a 262 / Gram2[ai(j,j,r,p,p,k)];
ef67d338 263 }
1d3c1faa
BA
264 }
265 }
266 }
4cab944a 267
1d3c1faa
BA
268 /////////////
269 // Etape E //
270 /////////////
4cab944a 271
1d3c1faa 272 int signum;
b42f0f40 273 Real sumLogLLF2 = 0.;
4cab944a 274 for (int i=0; i<n; i++)
1d3c1faa 275 {
4cab944a 276 for (int r=0; r<k; r++)
1d3c1faa 277 {
ef67d338 278 //compute Y[i,]%*%rho[,,r]
4cab944a 279 for (int u=0; u<m; u++)
1d3c1faa 280 {
b42f0f40 281 YiRhoR[u] = 0.;
4cab944a 282 for (int v=0; v<m; v++)
aa8df014 283 YiRhoR[u] += Y[mi(i,v,n,m)] * rho[ai(v,u,r,m,m,k)];
1d3c1faa 284 }
4cab944a 285
1d3c1faa 286 //compute X(i,:)*phi(:,:,r)
4cab944a 287 for (int u=0; u<m; u++)
1d3c1faa 288 {
b42f0f40 289 XiPhiR[u] = 0.;
4cab944a
BA
290 for (int v=0; v<p; v++)
291 XiPhiR[u] += X[mi(i,v,n,p)] * phi[ai(v,u,r,p,m,k)];
1d3c1faa 292 }
4cab944a 293
ef67d338 294 //compute sq norm || Y(:,i)*rho(:,:,r)-X(i,:)*phi(:,:,r) ||_2^2
b42f0f40 295 sqNorm2[r] = 0.;
4cab944a 296 for (int u=0; u<m; u++)
ef67d338 297 sqNorm2[r] += (YiRhoR[u]-XiPhiR[u]) * (YiRhoR[u]-XiPhiR[u]);
1d3c1faa 298 }
ef67d338 299
b42f0f40
BA
300 Real sumLLF1 = 0.;
301 Real sumGamI = 0.;
4cab944a 302 for (int r=0; r<k; r++)
1d3c1faa 303 {
ef67d338 304 //compute det(rho[,,r]) [TODO: avoid re-computations]
4cab944a 305 for (int u=0; u<m; u++)
1d3c1faa 306 {
4cab944a
BA
307 for (int v=0; v<m; v++)
308 matrix->data[u*m+v] = rho[ai(u,v,r,m,m,k)];
1d3c1faa
BA
309 }
310 gsl_linalg_LU_decomp(matrix, permutation, &signum);
9ff729fb 311 Real detRhoR = gsl_linalg_LU_det(matrix, signum);
b42f0f40 312 Gam[mi(i,r,n,k)] = pi[r] * exp(-.5*sqNorm2[r]) * detRhoR;
ef67d338 313 sumLLF1 += Gam[mi(i,r,n,k)] / gaussConstM;
4cab944a 314 sumGamI += Gam[mi(i,r,n,k)];
1d3c1faa
BA
315 }
316 sumLogLLF2 += log(sumLLF1);
4cab944a 317 for (int r=0; r<k; r++)
1d3c1faa 318 {
ef67d338
BA
319 //gam[i,] = Gam[i,] / sumGamI
320 gam[mi(i,r,n,k)] = sumGamI > EPS ? Gam[mi(i,r,n,k)] / sumGamI : 0.;
1d3c1faa
BA
321 }
322 }
ef67d338
BA
323
324 //sumPen = sum(pi^gamma * b)
b42f0f40 325 Real sumPen = 0.;
4cab944a 326 for (int r=0; r<k; r++)
1d3c1faa 327 sumPen += pow(pi[r],gamma) * b[r];
ef67d338 328 //LLF[ite] = -sumLogLLF2/n + lambda*sumPen
1d3c1faa 329 LLF[ite] = -invN * sumLogLLF2 + lambda * sumPen;
b42f0f40 330 dist = ite==0 ? LLF[ite] : (LLF[ite] - LLF[ite-1]) / (1. + fabs(LLF[ite]));
ef67d338
BA
331
332 //Dist1 = max( abs(phi-Phi) / (1+abs(phi)) )
b42f0f40 333 Real Dist1 = 0.;
4cab944a 334 for (int u=0; u<p; u++)
1d3c1faa 335 {
4cab944a 336 for (int v=0; v<m; v++)
1d3c1faa 337 {
4cab944a 338 for (int w=0; w<k; w++)
1d3c1faa 339 {
b42f0f40
BA
340 Real tmpDist = fabs(phi[ai(u,v,w,p,m,k)]-Phi[ai(u,v,w,p,m,k)])
341 / (1.+fabs(phi[ai(u,v,w,p,m,k)]));
1d3c1faa
BA
342 if (tmpDist > Dist1)
343 Dist1 = tmpDist;
344 }
345 }
346 }
ef67d338 347 //Dist2 = max( (abs(rho-Rho)) / (1+abs(rho)) )
b42f0f40 348 Real Dist2 = 0.;
4cab944a 349 for (int u=0; u<m; u++)
1d3c1faa 350 {
4cab944a 351 for (int v=0; v<m; v++)
1d3c1faa 352 {
4cab944a 353 for (int w=0; w<k; w++)
1d3c1faa 354 {
b42f0f40
BA
355 Real tmpDist = fabs(rho[ai(u,v,w,m,m,k)]-Rho[ai(u,v,w,m,m,k)])
356 / (1.+fabs(rho[ai(u,v,w,m,m,k)]));
1d3c1faa
BA
357 if (tmpDist > Dist2)
358 Dist2 = tmpDist;
359 }
360 }
361 }
ef67d338 362 //Dist3 = max( (abs(pi-Pi)) / (1+abs(Pi)))
b42f0f40 363 Real Dist3 = 0.;
4cab944a 364 for (int u=0; u<n; u++)
1d3c1faa 365 {
4cab944a 366 for (int v=0; v<k; v++)
1d3c1faa 367 {
b42f0f40 368 Real tmpDist = fabs(pi[v]-Pi[v]) / (1.+fabs(pi[v]));
1d3c1faa
BA
369 if (tmpDist > Dist3)
370 Dist3 = tmpDist;
371 }
372 }
373 //dist2=max([max(Dist1),max(Dist2),max(Dist3)]);
374 dist2 = Dist1;
375 if (Dist2 > dist2)
376 dist2 = Dist2;
377 if (Dist3 > dist2)
378 dist2 = Dist3;
ef67d338 379
1d3c1faa
BA
380 ite++;
381 }
ef67d338 382
1d3c1faa
BA
383 //free memory
384 free(b);
385 free(gam);
386 free(Gam);
387 free(Phi);
388 free(Rho);
389 free(Pi);
390 free(ps);
391 free(nY2);
392 free(ps1);
393 free(nY21);
394 free(Gram2);
395 free(ps2);
396 gsl_matrix_free(matrix);
397 gsl_permutation_free(permutation);
398 free(XiPhiR);
399 free(YiRhoR);
400 free(gam2);
401 free(pi2);
402 free(X2);
403 free(Y2);
ef67d338 404 free(sqNorm2);
1d3c1faa 405}