/*******M step, maximazize log-likelihood*/
void Plsa::M_step(MatrixXd &data)
{
MatrixXd X;
for (int i=0;i<K;i++)
{
Pw_z.col(i)=(Pz_wd[i].cwiseProduct(data)).rowwise().sum();//suma de filas[vector(1XN)]
Pd_z.col(i)=(Pz_wd[i].cwiseProduct(data)).colwise().sum();//suma de columnas[vector(1XN)]
}
//normalize
RowVectorXd Temp;
RowVectorXd C;
VectorXd E;
VectorXd T;
Temp=RowVectorXd::Ones(K);
T=VectorXd::Ones(K); // vector of K with ones
P_z=Pd_z.colwise().sum();
cout<<P_z;
C=Pd_z.colwise().sum(); //suma de columnas[vector(1XN)]
Temp=Temp.cwiseQuotient(C);
Pd_z=Pd_z*(Temp.asDiagonal());
C=Pw_z.colwise().sum(); //
Temp=Temp.cwiseQuotient(C);
Pw_z=Pw_z*(Temp.asDiagonal());
E=P_z.rowwise().sum();
P_z=P_z.cwiseProduct(T.cwiseQuotient(E));
}
MatrixXd NumInt::GuassLobattoQaudrature(const int &N) {
int N0=N;
double a=1.0; double b=1.0;
double a1=2.0; double b1 = 2.0;
if (N0>=1)
{
N0=N0-2;
}
// Build the pointers
int* _N = &N0;
double* _a=&a; double* _b=&b;
double* _a1=&a1; double* _b1=&b1;
// Initial Guess - Chebyshev-Gauss-Lobatto points
ArrayXd xu;
xu.setLinSpaced(N0+2,0.0,N0+1);
ArrayXd x=-cos(xu/(N0+1)*M_PI);
// Allocate space for points and weights
VectorXd z = VectorXd::Zero(x.size());
VectorXd w = VectorXd::Zero(x.size());
double x0,x1,del; double* _x0 = &x0; double* _x1 = &x1;
double deps = std::numeric_limits<double>::epsilon();
for (int k=0; k<=x.size()-1; k++)
{
x0=x(k);
del=2.0;
while (fabs(del) > deps)
{
// Polynomial Deflation: Exclude the already determined roots
VectorXd s1 = x.head(k);
VectorXd ones = VectorXd::Constant(s1.size(),1);
double s = (ones.cwiseQuotient((x0-s1.array()).matrix())).sum();
// Compute Jacobi polynomial p(a,b)
JacobiPolynomials J(_N,_a,_b,_x0,false);
VectorXd p = J.getJacobiPolynomials();
// Compute Jacobi polynomial p(a+1,b+1) for derivative dp(a,b)
JacobiPolynomials J1(_N,_a1,_b1,_x0,false);
VectorXd p1 = J1.getJacobiPolynomials();
VectorXd dp=VectorXd::Zero(p1.size()); dp(0)=0;
// Compute derivative of Jacobi polynomial p(a,b)
for (int j=0; j<=*_N-1; j++)
{
dp(j+1) = 0.5*(*_a+*_b+j+2)*p1(j);
}
//Gauss-Lobatto points are roots of (1-x^2)*dp, hence
double nom = (1-x0*x0)*p(N0); double dnom = -2*x0*p(N0)+(1-x0*x0)*dp(N0);
del = - nom/(dnom-nom*s);
x1 = x0+del;
x0=x1;
}
z(k)=x1;
double a2=0; double b2=0; int N1=N0+1;
double* _a2=&a2; double* _b2=&b2; int* _N1 = &N1;
JacobiPolynomials J(_N1,_a2,_b2,_x1,false);
VectorXd p = J.getJacobiPolynomials();
w(k) = 2.0/((N1)*(N1+1)*p(N1)*p(N1));
}
// Store
Matrix<double,Dynamic,Dynamic> zw(N0+2,2);
zw.col(0)=z;
zw.col(1)=w;
return zw;
}
MNEInverseOperator MNEInverseOperator::prepare_inverse_operator(qint32 nave ,float lambda2, bool dSPM, bool sLORETA) const
{
if(nave <= 0)
{
printf("The number of averages should be positive\n");
return MNEInverseOperator();
}
printf("Preparing the inverse operator for use...\n");
MNEInverseOperator inv(*this);
//
// Scale some of the stuff
//
float scale = ((float)inv.nave)/((float)nave);
inv.noise_cov->data *= scale;
inv.noise_cov->eig *= scale;
inv.source_cov->data *= scale;
//
if (inv.eigen_leads_weighted)
inv.eigen_leads->data *= sqrt(scale);
//
printf("\tScaled noise and source covariance from nave = %d to nave = %d\n",inv.nave,nave);
inv.nave = nave;
//
// Create the diagonal matrix for computing the regularized inverse
//
VectorXd tmp = inv.sing.cwiseProduct(inv.sing) + VectorXd::Constant(inv.sing.size(), lambda2);
// if(inv.reginv)
// delete inv.reginv;
inv.reginv = VectorXd(inv.sing.cwiseQuotient(tmp));
printf("\tCreated the regularized inverter\n");
//
// Create the projection operator
//
qint32 ncomp = FiffProj::make_projector(inv.projs, inv.noise_cov->names, inv.proj);
if (ncomp > 0)
printf("\tCreated an SSP operator (subspace dimension = %d)\n",ncomp);
//
// Create the whitener
//
// if(inv.whitener)
// delete inv.whitener;
inv.whitener = MatrixXd::Zero(inv.noise_cov->dim, inv.noise_cov->dim);
qint32 nnzero, k;
if (inv.noise_cov->diag == 0)
{
//
// Omit the zeroes due to projection
//
nnzero = 0;
for (k = ncomp; k < inv.noise_cov->dim; ++k)
{
if (inv.noise_cov->eig[k] > 0)
{
inv.whitener(k,k) = 1.0/sqrt(inv.noise_cov->eig[k]);
++nnzero;
}
}
//
// Rows of eigvec are the eigenvectors
//
inv.whitener *= inv.noise_cov->eigvec;
printf("\tCreated the whitener using a full noise covariance matrix (%d small eigenvalues omitted)\n", inv.noise_cov->dim - nnzero);
}
else
{
//
// No need to omit the zeroes due to projection
//
for (k = 0; k < inv.noise_cov->dim; ++k)
inv.whitener(k,k) = 1.0/sqrt(inv.noise_cov->data(k,0));
printf("\tCreated the whitener using a diagonal noise covariance matrix (%d small eigenvalues discarded)\n",ncomp);
}
//
// Finally, compute the noise-normalization factors
//
if (dSPM || sLORETA)
{
VectorXd noise_norm = VectorXd::Zero(inv.eigen_leads->nrow);
VectorXd noise_weight;
if (dSPM)
{
printf("\tComputing noise-normalization factors (dSPM)...");
noise_weight = VectorXd(inv.reginv);
}
else
{
printf("\tComputing noise-normalization factors (sLORETA)...");
VectorXd tmp = (VectorXd::Constant(inv.sing.size(), 1) + inv.sing.cwiseProduct(inv.sing)/lambda2);
noise_weight = inv.reginv.cwiseProduct(tmp.cwiseSqrt());
}
VectorXd one;
if (inv.eigen_leads_weighted)
{
for (k = 0; k < inv.eigen_leads->nrow; ++k)
{
one = inv.eigen_leads->data.block(k,0,1,inv.eigen_leads->data.cols()).cwiseProduct(noise_weight);
noise_norm[k] = sqrt(one.dot(one));
}
}
else
{
// qDebug() << 32;
double c;
for (k = 0; k < inv.eigen_leads->nrow; ++k)
{
// qDebug() << 321;
c = sqrt(inv.source_cov->data(k,0));
// qDebug() << 322;
// qDebug() << "inv.eigen_leads->data" << inv.eigen_leads->data.rows() << "x" << inv.eigen_leads->data.cols();
// qDebug() << "noise_weight" << noise_weight.rows() << "x" << noise_weight.cols();
one = c*(inv.eigen_leads->data.row(k).transpose()).cwiseProduct(noise_weight);//ToDo eigenleads data -> pointer
noise_norm[k] = sqrt(one.dot(one));
// qDebug() << 324;
}
}
// qDebug() << 4;
//
// Compute the final result
//
VectorXd noise_norm_new;
if (inv.source_ori == FIFFV_MNE_FREE_ORI)
{
//
// The three-component case is a little bit more involved
// The variances at three consequtive entries must be squeared and
// added together
//
// Even in this case return only one noise-normalization factor
// per source location
//
VectorXd* t = MNEMath::combine_xyz(noise_norm.transpose());
noise_norm_new = t->cwiseSqrt();//double otherwise values are getting too small
delete t;
//
// This would replicate the same value on three consequtive
// entries
//
// noise_norm = kron(sqrt(mne_combine_xyz(noise_norm)),ones(3,1));
}
VectorXd vOnes = VectorXd::Ones(noise_norm_new.size());
VectorXd tmp = vOnes.cwiseQuotient(noise_norm_new.cwiseAbs());
// if(inv.noisenorm)
// delete inv.noisenorm;
typedef Eigen::Triplet<double> T;
std::vector<T> tripletList;
tripletList.reserve(noise_norm_new.size());
for(qint32 i = 0; i < noise_norm_new.size(); ++i)
tripletList.push_back(T(i, i, tmp[i]));
inv.noisenorm = SparseMatrix<double>(noise_norm_new.size(),noise_norm_new.size());
inv.noisenorm.setFromTriplets(tripletList.begin(), tripletList.end());
printf("[done]\n");
}
else
{
// if(inv.noisenorm)
// delete inv.noisenorm;
inv.noisenorm = SparseMatrix<double>();
}
return inv;
}
