[valse.git] / src / sources / EMGLLF.c

#include "utils.h"
#include <stdlib.h>
#include <gsl/gsl_linalg.h>

// TODO: don't recompute indexes every time......
void EMGLLF_core(
	// IN parameters
	const Real* phiInit, // parametre initial de moyenne renormalisé
	const Real* rhoInit, // parametre initial de variance renormalisé
	const Real* piInit,	 // parametre initial des proportions
	const Real* gamInit, // paramètre initial des probabilités a posteriori de chaque échantillon
	int mini, // nombre minimal d'itérations dans l'algorithme EM
	int maxi, // nombre maximal d'itérations dans l'algorithme EM
	Real gamma, // puissance des proportions dans la pénalisation pour un Lasso adaptatif
	Real lambda, // valeur du paramètre de régularisation du Lasso
	const Real* X, // régresseurs
	const Real* Y, // réponse
	Real tau, // seuil pour accepter la convergence
	// OUT parameters (all pointers, to be modified)
	Real* phi, // parametre de moyenne renormalisé, calculé par l'EM
	Real* rho, // parametre de variance renormalisé, calculé par l'EM
	Real* pi, // parametre des proportions renormalisé, calculé par l'EM
	Real* LLF, // log vraisemblance associée à cet échantillon, pour les valeurs estimées des paramètres
	Real* S,
	// additional size parameters
	int n, // nombre d'echantillons
	int p, // nombre de covariables
	int m, // taille de Y (multivarié)
	int k) // nombre de composantes dans le mélange
{
	//Initialize outputs
	copyArray(phiInit, phi, p*m*k);
	copyArray(rhoInit, rho, m*m*k);
	copyArray(piInit, pi, k);
	zeroArray(LLF, maxi);
	//S is already allocated, and doesn't need to be 'zeroed'

	//Other local variables
	Real* gam = (Real*)malloc(n*k*sizeof(Real));
	copyArray(gamInit, gam, n*k);
	Real* b = (Real*)malloc(k*sizeof(Real));
	Real* Phi = (Real*)malloc(p*m*k*sizeof(Real));
	Real* Rho = (Real*)malloc(m*m*k*sizeof(Real));
	Real* Pi = (Real*)malloc(k*sizeof(Real));
	Real* gam2 = (Real*)malloc(k*sizeof(Real));
	Real* pi2 = (Real*)malloc(k*sizeof(Real));
	Real* Gram2 = (Real*)malloc(p*p*k*sizeof(Real));
	Real* ps = (Real*)malloc(m*k*sizeof(Real));
	Real* nY2 = (Real*)malloc(m*k*sizeof(Real));
	Real* ps1 = (Real*)malloc(n*m*k*sizeof(Real));
	Real* ps2 = (Real*)malloc(p*m*k*sizeof(Real));
	Real* nY21 = (Real*)malloc(n*m*k*sizeof(Real));
	Real* Gam = (Real*)malloc(n*k*sizeof(Real));
	Real* X2 = (Real*)malloc(n*p*k*sizeof(Real));
	Real* Y2 = (Real*)malloc(n*m*k*sizeof(Real));
	Real* sqNorm2 = (Real*)malloc(k*sizeof(Real));
	gsl_matrix* matrix = gsl_matrix_alloc(m, m);
	gsl_permutation* permutation = gsl_permutation_alloc(m);
	Real* YiRhoR = (Real*)malloc(m*sizeof(Real));
	Real* XiPhiR = (Real*)malloc(m*sizeof(Real));
	Real dist = 0.;
	Real dist2 = 0.;
	int ite = 0;
	const Real EPS = 1e-15;
	const Real gaussConstM = pow(2.*M_PI,m/2.);

	while (ite < mini || (ite < maxi && (dist >= tau || dist2 >= sqrt(tau))))
	{
		copyArray(phi, Phi, p*m*k);
		copyArray(rho, Rho, m*m*k);
		copyArray(pi, Pi, k);

		// Calculs associés a Y et X
		for (int r=0; r<k; r++)
		{
			for (int mm=0; mm<m; mm++)
			{
				//Y2[,mm,r] = sqrt(gam[,r]) * Y[,mm]
				for (int u=0; u<n; u++)
					Y2[ai(u,mm,r,n,m,k)] = sqrt(gam[mi(u,r,n,k)]) * Y[mi(u,mm,m,n)];
			}
			for (int i=0; i<n; i++)
			{
				//X2[i,,r] = sqrt(gam[i,r]) * X[i,]
				for (int u=0; u<p; u++)
					X2[ai(i,u,r,n,p,k)] = sqrt(gam[mi(i,r,n,k)]) * X[mi(i,u,n,p)];
			}
			for (int mm=0; mm<m; mm++)
			{
				//ps2[,mm,r] = crossprod(X2[,,r],Y2[,mm,r])
				for (int u=0; u<p; u++)
				{
					Real dotProduct = 0.;
					for (int v=0; v<n; v++)
						dotProduct += X2[ai(v,u,r,n,p,k)] * Y2[ai(v,mm,r,n,m,k)];
					ps2[ai(u,mm,r,p,m,k)] = dotProduct;
				}
			}
			for (int j=0; j<p; j++)
			{
				for (int s=0; s<p; s++)
				{
					//Gram2[j,s,r] = crossprod(X2[,j,r], X2[,s,r])
					Real dotProduct = 0.;
					for (int u=0; u<n; u++)
						dotProduct += X2[ai(u,j,r,n,p,k)] * X2[ai(u,s,r,n,p,k)];
					Gram2[ai(j,s,r,p,p,k)] = dotProduct;
				}
			}
		}

		/////////////
		// Etape M //
		/////////////

		// Pour pi
		for (int r=0; r<k; r++)
		{
			//b[r] = sum(abs(phi[,,r]))
			Real sumAbsPhi = 0.;
			for (int u=0; u<p; u++)
				for (int v=0; v<m; v++)
					sumAbsPhi += fabs(phi[ai(u,v,r,p,m,k)]);
			b[r] = sumAbsPhi;
		}
		//gam2 = colSums(gam)
		for (int u=0; u<k; u++)
		{
			Real sumOnColumn = 0.;
			for (int v=0; v<n; v++)
				sumOnColumn += gam[mi(v,u,n,k)];
			gam2[u] = sumOnColumn;
		}
		//a = sum(gam %*% log(pi))
		Real a = 0.;
		for (int u=0; u<n; u++)
		{
			Real dotProduct = 0.;
			for (int v=0; v<k; v++)
				dotProduct += gam[mi(u,v,n,k)] * log(pi[v]);
			a += dotProduct;
		}

		//tant que les proportions sont negatives
		int kk = 0;
		int pi2AllPositive = 0;
		Real invN = 1./n;
		while (!pi2AllPositive)
		{
			//pi2 = pi + 0.1^kk * ((1/n)*gam2 - pi)
			Real pow_01_kk = pow(0.1,kk);
			for (int r=0; r<k; r++)
				pi2[r] = pi[r] + pow_01_kk * (invN*gam2[r] - pi[r]);
			//pi2AllPositive = all(pi2 >= 0)
			pi2AllPositive = 1;
			for (int r=0; r<k; r++)
			{
				if (pi2[r] < 0)
				{
					pi2AllPositive = 0;
					break;
				}
			}
			kk++;
		}

		//(pi.^gamma)*b
		Real piPowGammaDotB = 0.;
		for (int v=0; v<k; v++)
			piPowGammaDotB += pow(pi[v],gamma) * b[v];
		//(pi2.^gamma)*b
		Real pi2PowGammaDotB = 0.;
		for (int v=0; v<k; v++)
			pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
		//transpose(gam2)*log(pi2)
		Real prodGam2logPi2 = 0.;
		for (int v=0; v<k; v++)
			prodGam2logPi2 += gam2[v] * log(pi2[v]);
		//t(m) la plus grande valeur dans la grille O.1^k tel que ce soit décroissante ou constante
		while (-invN*a + lambda*piPowGammaDotB < -invN*prodGam2logPi2 + lambda*pi2PowGammaDotB
			&& kk<1000)
		{
			Real pow_01_kk = pow(0.1,kk);
			//pi2 = pi + 0.1^kk * (1/n*gam2 - pi)
			for (int v=0; v<k; v++)
				pi2[v] = pi[v] + pow_01_kk * (invN*gam2[v] - pi[v]);
			//pi2 was updated, so we recompute pi2PowGammaDotB and prodGam2logPi2
			pi2PowGammaDotB = 0.;
			for (int v=0; v<k; v++)
				pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
			prodGam2logPi2 = 0.;
			for (int v=0; v<k; v++)
				prodGam2logPi2 += gam2[v] * log(pi2[v]);
			kk++;
		}
		Real t = pow(0.1,kk);
		//sum(pi + t*(pi2-pi))
		Real sumPiPlusTbyDiff = 0.;
		for (int v=0; v<k; v++)
			sumPiPlusTbyDiff += (pi[v] + t*(pi2[v] - pi[v]));
		//pi = (pi + t*(pi2-pi)) / sum(pi + t*(pi2-pi))
		for (int v=0; v<k; v++)
			pi[v] = (pi[v] + t*(pi2[v] - pi[v])) / sumPiPlusTbyDiff;

		//Pour phi et rho
		for (int r=0; r<k; r++)
		{
			for (int mm=0; mm<m; mm++)
			{
				for (int i=0; i<n; i++)
				{
					//< X2(i,:,r) , phi(:,mm,r) >
					Real dotProduct = 0.;
					for (int u=0; u<p; u++)
						dotProduct += X2[ai(i,u,r,n,p,k)] * phi[ai(u,mm,r,p,m,k)];
					//ps1[i,mm,r] = Y2[i,mm,r] * sum(X2[i,,r] * phi[,mm,r])
					ps1[ai(i,mm,r,n,m,k)] = Y2[ai(i,mm,r,n,m,k)] * dotProduct;
					nY21[ai(i,mm,r,n,m,k)] = Y2[ai(i,mm,r,n,m,k)] * Y2[ai(i,mm,r,n,m,k)];
				}
				//ps[mm,r] = sum(ps1[,mm,r])
				Real sumPs1 = 0.;
				for (int u=0; u<n; u++)
					sumPs1 += ps1[ai(u,mm,r,n,m,k)];
				ps[mi(mm,r,m,k)] = sumPs1;
				//nY2[mm,r] = sum(nY21[,mm,r])
				Real sumNy21 = 0.;
				for (int u=0; u<n; u++)
					sumNy21 += nY21[ai(u,mm,r,n,m,k)];
				nY2[mi(mm,r,m,k)] = sumNy21;
				//rho[mm,mm,r] = (ps[mm,r]+sqrt(ps[mm,r]^2+4*nY2[mm,r]*(gam2[r]))) / (2*nY2[mm,r])
				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)]
					+ 4*nY2[mi(mm,r,m,k)] * gam2[r] ) ) / (2*nY2[mi(mm,r,m,k)]);
			}
		}
		for (int r=0; r<k; r++)
		{
			for (int j=0; j<p; j++)
			{
				for (int mm=0; mm<m; mm++)
				{
					//sum(phi[1:(j-1),mm,r] * Gram2[j,1:(j-1),r])
					Real dotPhiGram2 = 0.0;
					for (int u=0; u<j; u++)
						dotPhiGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)];
					//sum(phi[(j+1):p,mm,r] * Gram2[j,(j+1):p,r])
					for (int u=j+1; u<p; u++)
						dotPhiGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)];
					//S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] +
					//  (if(j>1) sum(phi[1:(j-1),mm,r] * Gram2[j,1:(j-1),r]) else 0) +
					//  (if(j<p) sum(phi[(j+1):p,mm,r] * Gram2[j,(j+1):p,r]) else 0)
					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;
					Real pow_pir_gamma = pow(pi[r],gamma);
					if (fabs(S[ai(j,mm,r,p,m,k)]) <= n*lambda*pow_pir_gamma)
						phi[ai(j,mm,r,p,m,k)] = 0;
					else if (S[ai(j,mm,r,p,m,k)] > n*lambda*pow_pir_gamma)
					{
						phi[ai(j,mm,r,p,m,k)] = (n*lambda*pow_pir_gamma - S[ai(j,mm,r,p,m,k)])
							/ Gram2[ai(j,j,r,p,p,k)];
					}
					else
					{
						phi[ai(j,mm,r,p,m,k)] = -(n*lambda*pow_pir_gamma + S[ai(j,mm,r,p,m,k)])
							/ Gram2[ai(j,j,r,p,p,k)];
					}
				}
			}
		}

		/////////////
		// Etape E //
		/////////////

		int signum;
		Real sumLogLLF2 = 0.0;
		for (int i=0; i<n; i++)
		{
			Real minSqNorm2 = INFINITY;

			for (int r=0; r<k; r++)
			{
				//compute Y[i,]%*%rho[,,r]
				for (int u=0; u<m; u++)
				{
					YiRhoR[u] = 0.0;
					for (int v=0; v<m; v++)
						YiRhoR[u] += Y[mi(i,v,n,m)] * rho[ai(v,u,r,m,m,k)];
				}

				//compute X(i,:)*phi(:,:,r)
				for (int u=0; u<m; u++)
				{
					XiPhiR[u] = 0.0;
					for (int v=0; v<p; v++)
						XiPhiR[u] += X[mi(i,v,n,p)] * phi[ai(v,u,r,p,m,k)];
				}

				//compute sq norm || Y(:,i)*rho(:,:,r)-X(i,:)*phi(:,:,r) ||_2^2
				sqNorm2[r] = 0.0;
				for (int u=0; u<m; u++)
					sqNorm2[r] += (YiRhoR[u]-XiPhiR[u]) * (YiRhoR[u]-XiPhiR[u]);
				if (sqNorm2[r] < minSqNorm2)
					minSqNorm2 = sqNorm2[r];
			}
			Real shift = 0.5*minSqNorm2;

			Real sumLLF1 = 0.0;
			Real sumGamI = 0.0;
			for (int r=0; r<k; r++)
			{
				//compute det(rho[,,r]) [TODO: avoid re-computations]
				for (int u=0; u<m; u++)
				{
					for (int v=0; v<m; v++)
						matrix->data[u*m+v] = rho[ai(u,v,r,m,m,k)];
				}
				gsl_linalg_LU_decomp(matrix, permutation, &signum);
				Real detRhoR = gsl_linalg_LU_det(matrix, signum);

				//FIXME: det(rho[,,r]) too small(?!). See EMGLLF.R
				Gam[mi(i,r,n,k)] = pi[r] * exp(-0.5*sqNorm2[r] + shift) ; //* detRhoR;
				sumLLF1 += Gam[mi(i,r,n,k)] / gaussConstM;
				sumGamI += Gam[mi(i,r,n,k)];
			}
			sumLogLLF2 += log(sumLLF1);
			for (int r=0; r<k; r++)
			{
				//gam[i,] = Gam[i,] / sumGamI
				gam[mi(i,r,n,k)] = sumGamI > EPS ? Gam[mi(i,r,n,k)] / sumGamI : 0.;
			}
		}

		//sumPen = sum(pi^gamma * b)
		Real sumPen = 0.0;
		for (int r=0; r<k; r++)
			sumPen += pow(pi[r],gamma) * b[r];
		//LLF[ite] = -sumLogLLF2/n + lambda*sumPen
		LLF[ite] = -invN * sumLogLLF2 + lambda * sumPen;
		dist = ite==0 ? LLF[ite] : (LLF[ite] - LLF[ite-1]) / (1.0 + fabs(LLF[ite]));

		//Dist1 = max( abs(phi-Phi) / (1+abs(phi)) )
		Real Dist1 = 0.0;
		for (int u=0; u<p; u++)
		{
			for (int v=0; v<m; v++)
			{
				for (int w=0; w<k; w++)
				{
					Real tmpDist = fabs(phi[ai(u,v,w,p,m,k)]-Phi[ai(u,v,w,p,m,k)]) 
						/ (1.0+fabs(phi[ai(u,v,w,p,m,k)]));
					if (tmpDist > Dist1)
						Dist1 = tmpDist;
				}
			}
		}
		//Dist2 = max( (abs(rho-Rho)) / (1+abs(rho)) )
		Real Dist2 = 0.0;
		for (int u=0; u<m; u++)
		{
			for (int v=0; v<m; v++)
			{
				for (int w=0; w<k; w++)
				{
					Real tmpDist = fabs(rho[ai(u,v,w,m,m,k)]-Rho[ai(u,v,w,m,m,k)]) 
						/ (1.0+fabs(rho[ai(u,v,w,m,m,k)]));
					if (tmpDist > Dist2)
						Dist2 = tmpDist;
				}
			}
		}
		//Dist3 = max( (abs(pi-Pi)) / (1+abs(Pi)))
		Real Dist3 = 0.0;
		for (int u=0; u<n; u++)
		{
			for (int v=0; v<k; v++)
			{
				Real tmpDist = fabs(pi[v]-Pi[v]) / (1.0+fabs(pi[v]));
				if (tmpDist > Dist3)
					Dist3 = tmpDist;
			}
		}
		//dist2=max([max(Dist1),max(Dist2),max(Dist3)]);
		dist2 = Dist1;
		if (Dist2 > dist2)
			dist2 = Dist2;
		if (Dist3 > dist2)
			dist2 = Dist3;

		ite++;
	}

	//free memory
	free(b);
	free(gam);
	free(Gam);
	free(Phi);
	free(Rho);
	free(Pi);
	free(ps);
	free(nY2);
	free(ps1);
	free(nY21);
	free(Gram2);
	free(ps2);
	gsl_matrix_free(matrix);
	gsl_permutation_free(permutation);
	free(XiPhiR);
	free(YiRhoR);
	free(gam2);
	free(pi2);
	free(X2);
	free(Y2);
	free(sqNorm2);
}
Commit	Line	Data
8e92c49c BA	1	#include "utils.h"
8e92c49c BA	2	#include <stdlib.h>
1d3c1faa BA	3	#include <gsl/gsl_linalg.h>
1d3c1faa BA	4
4cab944a	5	// TODO: don't recompute indexes every time......
09ab3c16	6	void 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
8e92c49c BA	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, pmk);
	33	copyArray(rhoInit, rho, mmk);
	34	copyArray(piInit, pi, k);
	35	zeroArray(LLF, maxi);
	36	//S is already allocated, and doesn't need to be 'zeroed'
4cab944a	37
1d3c1faa	38	//Other local variables
9ff729fb	39	Real* gam = (Real)malloc(nk*sizeof(Real));
1d3c1faa	40	copyArray(gamInit, gam, n*k);
9ff729fb BA	41	Real* b = (Real)malloc(ksizeof(Real));
	42	Real* Phi = (Real)malloc(pmksizeof(Real));
	43	Real* Rho = (Real)malloc(mmksizeof(Real));
	44	Real* Pi = (Real)malloc(ksizeof(Real));
	45	Real* gam2 = (Real)malloc(ksizeof(Real));
	46	Real* pi2 = (Real)malloc(ksizeof(Real));
	47	Real* Gram2 = (Real)malloc(ppksizeof(Real));
	48	Real* ps = (Real)malloc(mk*sizeof(Real));
	49	Real* nY2 = (Real)malloc(mk*sizeof(Real));
	50	Real* ps1 = (Real)malloc(nmksizeof(Real));
	51	Real* ps2 = (Real)malloc(pmksizeof(Real));
	52	Real* nY21 = (Real)malloc(nmksizeof(Real));
	53	Real* Gam = (Real)malloc(nk*sizeof(Real));
	54	Real* X2 = (Real)malloc(npksizeof(Real));
	55	Real* Y2 = (Real)malloc(nmksizeof(Real));
ef67d338	56	Real* sqNorm2 = (Real)malloc(ksizeof(Real));
1d3c1faa BA	57	gsl_matrix* matrix = gsl_matrix_alloc(m, m);
1d3c1faa BA	58	gsl_permutation* permutation = gsl_permutation_alloc(m);
9ff729fb BA	59	Real* YiRhoR = (Real)malloc(msizeof(Real));
	60	Real* XiPhiR = (Real)malloc(msizeof(Real));
	61	Real dist = 0.;
	62	Real dist2 = 0.;
4cab944a	63	int ite = 0;
ef67d338 BA	64	const Real EPS = 1e-15;
ef67d338 BA	65	const Real gaussConstM = pow(2.*M_PI,m/2.);
4cab944a	66
1d3c1faa BA	67	while (ite < mini \|\| (ite < maxi && (dist >= tau \|\| dist2 >= sqrt(tau))))
	68	{
	69	copyArray(phi, Phi, pmk);
	70	copyArray(rho, Rho, mmk);
	71	copyArray(pi, Pi, k);
4cab944a BA	72
	73	// Calculs associés a Y et X
	74	for (int r=0; r<k; r++)
1d3c1faa	75	{
4cab944a	76	for (int mm=0; mm<m; mm++)
1d3c1faa	77	{
ef67d338	78	//Y2[,mm,r] = sqrt(gam[,r]) * Y[,mm]
4cab944a BA	79	for (int u=0; u<n; u++)
4cab944a BA	80	Y2[ai(u,mm,r,n,m,k)] = sqrt(gam[mi(u,r,n,k)]) * Y[mi(u,mm,m,n)];
1d3c1faa	81	}
4cab944a	82	for (int i=0; i<n; i++)
1d3c1faa	83	{
ef67d338	84	//X2[i,,r] = sqrt(gam[i,r]) * X[i,]
4cab944a	85	for (int u=0; u<p; u++)
e39bc178	86	X2[ai(i,u,r,n,p,k)] = sqrt(gam[mi(i,r,n,k)]) * X[mi(i,u,n,p)];
1d3c1faa	87	}
4cab944a	88	for (int mm=0; mm<m; mm++)
1d3c1faa	89	{
ef67d338	90	//ps2[,mm,r] = crossprod(X2[,,r],Y2[,mm,r])
4cab944a	91	for (int u=0; u<p; u++)
1d3c1faa	92	{
9ff729fb	93	Real dotProduct = 0.;
4cab944a	94	for (int v=0; v<n; v++)
46a2e676	95	dotProduct += X2[ai(v,u,r,n,p,k)] * Y2[ai(v,mm,r,n,m,k)];
e39bc178	96	ps2[ai(u,mm,r,p,m,k)] = dotProduct;
1d3c1faa BA	97	}
1d3c1faa BA	98	}
4cab944a	99	for (int j=0; j<p; j++)
1d3c1faa	100	{
4cab944a	101	for (int s=0; s<p; s++)
1d3c1faa	102	{
ef67d338	103	//Gram2[j,s,r] = crossprod(X2[,j,r], X2[,s,r])
9ff729fb	104	Real dotProduct = 0.;
4cab944a BA	105	for (int u=0; u<n; u++)
	106	dotProduct += X2[ai(u,j,r,n,p,k)] * X2[ai(u,s,r,n,p,k)];
	107	Gram2[ai(j,s,r,p,p,k)] = dotProduct;
1d3c1faa BA	108	}
	109	}
	110	}
	111
	112	/////////////
	113	// Etape M //
	114	/////////////
4cab944a	115
1d3c1faa	116	// Pour pi
4cab944a	117	for (int r=0; r<k; r++)
1d3c1faa	118	{
ef67d338	119	//b[r] = sum(abs(phi[,,r]))
9ff729fb	120	Real sumAbsPhi = 0.;
4cab944a BA	121	for (int u=0; u<p; u++)
	122	for (int v=0; v<m; v++)
	123	sumAbsPhi += fabs(phi[ai(u,v,r,p,m,k)]);
1d3c1faa BA	124	b[r] = sumAbsPhi;
1d3c1faa BA	125	}
ef67d338	126	//gam2 = colSums(gam)
4cab944a	127	for (int u=0; u<k; u++)
1d3c1faa	128	{
9ff729fb	129	Real sumOnColumn = 0.;
4cab944a BA	130	for (int v=0; v<n; v++)
4cab944a BA	131	sumOnColumn += gam[mi(v,u,n,k)];
1d3c1faa BA	132	gam2[u] = sumOnColumn;
1d3c1faa BA	133	}
ef67d338	134	//a = sum(gam %*% log(pi))
9ff729fb	135	Real a = 0.;
4cab944a	136	for (int u=0; u<n; u++)
1d3c1faa	137	{
9ff729fb	138	Real dotProduct = 0.;
4cab944a BA	139	for (int v=0; v<k; v++)
4cab944a BA	140	dotProduct += gam[mi(u,v,n,k)] * log(pi[v]);
1d3c1faa BA	141	a += dotProduct;
1d3c1faa BA	142	}
4cab944a	143
1d3c1faa	144	//tant que les proportions sont negatives
4cab944a	145	int kk = 0;
1d3c1faa	146	int pi2AllPositive = 0;
9ff729fb	147	Real invN = 1./n;
1d3c1faa BA	148	while (!pi2AllPositive)
1d3c1faa BA	149	{
ef67d338 BA	150	//pi2 = pi + 0.1^kk * ((1/n)*gam2 - pi)
ef67d338 BA	151	Real pow_01_kk = pow(0.1,kk);
4cab944a	152	for (int r=0; r<k; r++)
ef67d338 BA	153	pi2[r] = pi[r] + pow_01_kk * (invN*gam2[r] - pi[r]);
ef67d338 BA	154	//pi2AllPositive = all(pi2 >= 0)
1d3c1faa	155	pi2AllPositive = 1;
4cab944a	156	for (int r=0; r<k; r++)
1d3c1faa BA	157	{
	158	if (pi2[r] < 0)
	159	{
	160	pi2AllPositive = 0;
	161	break;
	162	}
	163	}
	164	kk++;
	165	}
4cab944a	166
1d3c1faa	167	//(pi.^gamma)*b
9ff729fb	168	Real piPowGammaDotB = 0.;
4cab944a	169	for (int v=0; v<k; v++)
1d3c1faa BA	170	piPowGammaDotB += pow(pi[v],gamma) * b[v];
1d3c1faa BA	171	//(pi2.^gamma)*b
9ff729fb	172	Real pi2PowGammaDotB = 0.;
4cab944a	173	for (int v=0; v<k; v++)
1d3c1faa BA	174	pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
1d3c1faa BA	175	//transpose(gam2)*log(pi2)
9ff729fb	176	Real prodGam2logPi2 = 0.;
4cab944a	177	for (int v=0; v<k; v++)
1d3c1faa	178	prodGam2logPi2 += gam2[v] * log(pi2[v]);
ef67d338	179	//t(m) la plus grande valeur dans la grille O.1^k tel que ce soit décroissante ou constante
8e92c49c BA	180	while (-invNa + lambdapiPowGammaDotB < -invNprodGam2logPi2 + lambdapi2PowGammaDotB
8e92c49c BA	181	&& kk<1000)
1d3c1faa	182	{
ef67d338 BA	183	Real pow_01_kk = pow(0.1,kk);
ef67d338 BA	184	//pi2 = pi + 0.1^kk * (1/n*gam2 - pi)
4cab944a	185	for (int v=0; v<k; v++)
ef67d338	186	pi2[v] = pi[v] + pow_01_kk * (invN*gam2[v] - pi[v]);
1d3c1faa	187	//pi2 was updated, so we recompute pi2PowGammaDotB and prodGam2logPi2
4cab944a BA	188	pi2PowGammaDotB = 0.;
4cab944a BA	189	for (int v=0; v<k; v++)
1d3c1faa	190	pi2PowGammaDotB += pow(pi2[v],gamma) * b[v];
4cab944a BA	191	prodGam2logPi2 = 0.;
4cab944a BA	192	for (int v=0; v<k; v++)
1d3c1faa BA	193	prodGam2logPi2 += gam2[v] * log(pi2[v]);
	194	kk++;
	195	}
9ff729fb	196	Real t = pow(0.1,kk);
ef67d338	197	//sum(pi + t*(pi2-pi))
9ff729fb	198	Real sumPiPlusTbyDiff = 0.;
4cab944a	199	for (int v=0; v<k; v++)
1d3c1faa	200	sumPiPlusTbyDiff += (pi[v] + t*(pi2[v] - pi[v]));
ef67d338	201	//pi = (pi + t(pi2-pi)) / sum(pi + t(pi2-pi))
4cab944a	202	for (int v=0; v<k; v++)
1d3c1faa	203	pi[v] = (pi[v] + t*(pi2[v] - pi[v])) / sumPiPlusTbyDiff;
4cab944a	204
1d3c1faa	205	//Pour phi et rho
4cab944a	206	for (int r=0; r<k; r++)
1d3c1faa	207	{
4cab944a	208	for (int mm=0; mm<m; mm++)
1d3c1faa	209	{
4cab944a	210	for (int i=0; i<n; i++)
1d3c1faa BA	211	{
1d3c1faa BA	212	//< X2(i,:,r) , phi(:,mm,r) >
ef67d338	213	Real dotProduct = 0.;
4cab944a	214	for (int u=0; u<p; u++)
a2d68d1d	215	dotProduct += X2[ai(i,u,r,n,p,k)] * phi[ai(u,mm,r,p,m,k)];
ef67d338	216	//ps1[i,mm,r] = Y2[i,mm,r] * sum(X2[i,,r] * phi[,mm,r])
4cab944a BA	217	ps1[ai(i,mm,r,n,m,k)] = Y2[ai(i,mm,r,n,m,k)] * dotProduct;
4cab944a BA	218	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	219	}
ef67d338 BA	220	//ps[mm,r] = sum(ps1[,mm,r])
ef67d338 BA	221	Real sumPs1 = 0.;
4cab944a BA	222	for (int u=0; u<n; u++)
	223	sumPs1 += ps1[ai(u,mm,r,n,m,k)];
	224	ps[mi(mm,r,m,k)] = sumPs1;
ef67d338 BA	225	//nY2[mm,r] = sum(nY21[,mm,r])
ef67d338 BA	226	Real sumNy21 = 0.;
4cab944a BA	227	for (int u=0; u<n; u++)
	228	sumNy21 += nY21[ai(u,mm,r,n,m,k)];
	229	nY2[mi(mm,r,m,k)] = sumNy21;
ef67d338	230	//rho[mm,mm,r] = (ps[mm,r]+sqrt(ps[mm,r]^2+4nY2[mm,r](gam2[r]))) / (2*nY2[mm,r])
a2d68d1d	231	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	232	+ 4nY2[mi(mm,r,m,k)] gam2[r] ) ) / (2*nY2[mi(mm,r,m,k)]);
1d3c1faa BA	233	}
1d3c1faa BA	234	}
4cab944a	235	for (int r=0; r<k; r++)
1d3c1faa	236	{
4cab944a	237	for (int j=0; j<p; j++)
1d3c1faa	238	{
4cab944a	239	for (int mm=0; mm<m; mm++)
1d3c1faa	240	{
ef67d338	241	//sum(phi[1:(j-1),mm,r] * Gram2[j,1:(j-1),r])
9ff729fb	242	Real dotPhiGram2 = 0.0;
4cab944a BA	243	for (int u=0; u<j; u++)
4cab944a BA	244	dotPhiGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)];
ef67d338	245	//sum(phi[(j+1):p,mm,r] * Gram2[j,(j+1):p,r])
4cab944a BA	246	for (int u=j+1; u<p; u++)
4cab944a BA	247	dotPhiGram2 += phi[ai(u,mm,r,p,m,k)] * Gram2[ai(j,u,r,p,p,k)];
ef67d338 BA	248	//S[j,mm,r] = -rho[mm,mm,r]*ps2[j,mm,r] +
	249	// (if(j>1) sum(phi[1:(j-1),mm,r] * Gram2[j,1:(j-1),r]) else 0) +
	250	// (if(j<p) sum(phi[(j+1):p,mm,r] * Gram2[j,(j+1):p,r]) else 0)
4cab944a	251	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	252	Real pow_pir_gamma = pow(pi[r],gamma);
ef67d338 BA	253	if (fabs(S[ai(j,mm,r,p,m,k)]) <= nlambdapow_pir_gamma)
4cab944a	254	phi[ai(j,mm,r,p,m,k)] = 0;
ef67d338 BA	255	else if (S[ai(j,mm,r,p,m,k)] > nlambdapow_pir_gamma)
	256	{
	257	phi[ai(j,mm,r,p,m,k)] = (nlambdapow_pir_gamma - S[ai(j,mm,r,p,m,k)])
4cab944a	258	/ Gram2[ai(j,j,r,p,p,k)];
ef67d338	259	}
1d3c1faa	260	else
ef67d338 BA	261	{
ef67d338 BA	262	phi[ai(j,mm,r,p,m,k)] = -(nlambdapow_pir_gamma + S[ai(j,mm,r,p,m,k)])
4cab944a	263	/ Gram2[ai(j,j,r,p,p,k)];
ef67d338	264	}
1d3c1faa BA	265	}
	266	}
	267	}
4cab944a	268
1d3c1faa BA	269	/////////////
	270	// Etape E //
	271	/////////////
4cab944a	272
1d3c1faa	273	int signum;
9ff729fb	274	Real sumLogLLF2 = 0.0;
4cab944a	275	for (int i=0; i<n; i++)
1d3c1faa	276	{
ef67d338	277	Real minSqNorm2 = INFINITY;
4cab944a BA	278
4cab944a BA	279	for (int r=0; r<k; r++)
1d3c1faa	280	{
ef67d338	281	//compute Y[i,]%*%rho[,,r]
4cab944a	282	for (int u=0; u<m; u++)
1d3c1faa BA	283	{
1d3c1faa BA	284	YiRhoR[u] = 0.0;
4cab944a	285	for (int v=0; v<m; v++)
aa8df014	286	YiRhoR[u] += Y[mi(i,v,n,m)] * rho[ai(v,u,r,m,m,k)];
1d3c1faa	287	}
4cab944a	288
1d3c1faa	289	//compute X(i,:)*phi(:,:,r)
4cab944a	290	for (int u=0; u<m; u++)
1d3c1faa BA	291	{
1d3c1faa BA	292	XiPhiR[u] = 0.0;
4cab944a BA	293	for (int v=0; v<p; v++)
4cab944a BA	294	XiPhiR[u] += X[mi(i,v,n,p)] * phi[ai(v,u,r,p,m,k)];
1d3c1faa	295	}
4cab944a	296
ef67d338 BA	297	//compute sq norm \|\| Y(:,i)rho(:,:,r)-X(i,:)phi(:,:,r) \|\|_2^2
ef67d338 BA	298	sqNorm2[r] = 0.0;
4cab944a	299	for (int u=0; u<m; u++)
ef67d338 BA	300	sqNorm2[r] += (YiRhoR[u]-XiPhiR[u]) * (YiRhoR[u]-XiPhiR[u]);
	301	if (sqNorm2[r] < minSqNorm2)
	302	minSqNorm2 = sqNorm2[r];
1d3c1faa	303	}
ef67d338 BA	304	Real shift = 0.5*minSqNorm2;
	305
	306	Real sumLLF1 = 0.0;
	307	Real sumGamI = 0.0;
4cab944a	308	for (int r=0; r<k; r++)
1d3c1faa	309	{
ef67d338	310	//compute det(rho[,,r]) [TODO: avoid re-computations]
4cab944a	311	for (int u=0; u<m; u++)
1d3c1faa	312	{
4cab944a BA	313	for (int v=0; v<m; v++)
4cab944a BA	314	matrix->data[u*m+v] = rho[ai(u,v,r,m,m,k)];
1d3c1faa BA	315	}
1d3c1faa BA	316	gsl_linalg_LU_decomp(matrix, permutation, &signum);
9ff729fb	317	Real detRhoR = gsl_linalg_LU_det(matrix, signum);
4cab944a	318
ef67d338	319	//FIXME: det(rho[,,r]) too small(?!). See EMGLLF.R
46a2e676	320	Gam[mi(i,r,n,k)] = pi[r] * exp(-0.5sqNorm2[r] + shift) ; // detRhoR;
ef67d338	321	sumLLF1 += Gam[mi(i,r,n,k)] / gaussConstM;
4cab944a	322	sumGamI += Gam[mi(i,r,n,k)];
1d3c1faa BA	323	}
1d3c1faa BA	324	sumLogLLF2 += log(sumLLF1);
4cab944a	325	for (int r=0; r<k; r++)
1d3c1faa	326	{
ef67d338 BA	327	//gam[i,] = Gam[i,] / sumGamI
ef67d338 BA	328	gam[mi(i,r,n,k)] = sumGamI > EPS ? Gam[mi(i,r,n,k)] / sumGamI : 0.;
1d3c1faa BA	329	}
1d3c1faa BA	330	}
ef67d338 BA	331
ef67d338 BA	332	//sumPen = sum(pi^gamma * b)
9ff729fb	333	Real sumPen = 0.0;
4cab944a	334	for (int r=0; r<k; r++)
1d3c1faa	335	sumPen += pow(pi[r],gamma) * b[r];
ef67d338	336	//LLF[ite] = -sumLogLLF2/n + lambda*sumPen
1d3c1faa	337	LLF[ite] = -invN * sumLogLLF2 + lambda * sumPen;
ef67d338 BA	338	dist = ite==0 ? LLF[ite] : (LLF[ite] - LLF[ite-1]) / (1.0 + fabs(LLF[ite]));
	339
	340	//Dist1 = max( abs(phi-Phi) / (1+abs(phi)) )
9ff729fb	341	Real Dist1 = 0.0;
4cab944a	342	for (int u=0; u<p; u++)
1d3c1faa	343	{
4cab944a	344	for (int v=0; v<m; v++)
1d3c1faa	345	{
4cab944a	346	for (int w=0; w<k; w++)
1d3c1faa	347	{
9ff729fb	348	Real tmpDist = fabs(phi[ai(u,v,w,p,m,k)]-Phi[ai(u,v,w,p,m,k)])
4cab944a	349	/ (1.0+fabs(phi[ai(u,v,w,p,m,k)]));
1d3c1faa BA	350	if (tmpDist > Dist1)
	351	Dist1 = tmpDist;
	352	}
	353	}
	354	}
ef67d338	355	//Dist2 = max( (abs(rho-Rho)) / (1+abs(rho)) )
9ff729fb	356	Real Dist2 = 0.0;
4cab944a	357	for (int u=0; u<m; u++)
1d3c1faa	358	{
4cab944a	359	for (int v=0; v<m; v++)
1d3c1faa	360	{
4cab944a	361	for (int w=0; w<k; w++)
1d3c1faa	362	{
9ff729fb	363	Real tmpDist = fabs(rho[ai(u,v,w,m,m,k)]-Rho[ai(u,v,w,m,m,k)])
4cab944a	364	/ (1.0+fabs(rho[ai(u,v,w,m,m,k)]));
1d3c1faa BA	365	if (tmpDist > Dist2)
	366	Dist2 = tmpDist;
	367	}
	368	}
	369	}
ef67d338	370	//Dist3 = max( (abs(pi-Pi)) / (1+abs(Pi)))
9ff729fb	371	Real Dist3 = 0.0;
4cab944a	372	for (int u=0; u<n; u++)
1d3c1faa	373	{
4cab944a	374	for (int v=0; v<k; v++)
1d3c1faa	375	{
9ff729fb	376	Real tmpDist = fabs(pi[v]-Pi[v]) / (1.0+fabs(pi[v]));
1d3c1faa BA	377	if (tmpDist > Dist3)
	378	Dist3 = tmpDist;
	379	}
	380	}
	381	//dist2=max([max(Dist1),max(Dist2),max(Dist3)]);
	382	dist2 = Dist1;
	383	if (Dist2 > dist2)
	384	dist2 = Dist2;
	385	if (Dist3 > dist2)
	386	dist2 = Dist3;
ef67d338	387
1d3c1faa BA	388	ite++;
1d3c1faa BA	389	}
ef67d338	390
1d3c1faa BA	391	//free memory
	392	free(b);
	393	free(gam);
	394	free(Gam);
	395	free(Phi);
	396	free(Rho);
	397	free(Pi);
	398	free(ps);
	399	free(nY2);
	400	free(ps1);
	401	free(nY21);
	402	free(Gram2);
	403	free(ps2);
	404	gsl_matrix_free(matrix);
	405	gsl_permutation_free(permutation);
	406	free(XiPhiR);
	407	free(YiRhoR);
	408	free(gam2);
	409	free(pi2);
	410	free(X2);
	411	free(Y2);
ef67d338	412	free(sqNorm2);
1d3c1faa	413	}