//' Parallel C++ implementation of multivariate Normal probability density function
//'
//'Based on the implementation from Nino Hardt and Dicko Ahmadou
//'http://gallery.rcpp.org/articles/dmvnorm_arma/
//'(accessed in August 2014)
//'
//'@param x data matrix
//'@param mean mean vector
//'@param varcovM variance covariance matrix
//'@param logical flag for returning the log of the probability density
//'function. Defaults is \code{TRUE}
//'@return vector of densities
//'
//'@export
//'
//'@examples
//'mvnpdf(x=matrix(1.96), mean=0, varcovM=diag(1), Log=FALSE)
//'mvnpdfC(x=matrix(1.96), mean=0, varcovM=diag(1), Log=FALSE)
//'mvnpdf(x=matrix(1.96), mean=0, varcovM=diag(1))
//'mvnpdfC(x=matrix(1.96), mean=0, varcovM=diag(1))
//'
//'library(microbenchmark)
//'microbenchmark(mvnpdf(x=matrix(1.96), mean=0, varcovM=diag(1), Log=FALSE),
//' mvnpdfC(x=matrix(1.96), mean=0, varcovM=diag(1), Log=FALSE),
//' times=10000L)
//'
// [[Rcpp::export]]
NumericVector mvnpdfC_par(NumericMatrix x,
NumericVector mean,
NumericMatrix varcovM,
bool Log=true,
int ncores=1){
const mat xx = as<mat>(x);
const mat S = as<mat>(varcovM);
const colvec m = as<colvec>(mean);
const int p = xx.n_rows;
const int n = xx.n_cols;
NumericVector y = NumericVector(n);
const mat Rinv = inv(trimatu(chol(S)));
const double logSqrtDetvarcovM = sum(log(Rinv.diag()));
const double constant = - p*log2pi2;
omp_set_num_threads(ncores);
#pragma omp parallel for shared(y)
for (int i=0; i < n; i++) {
colvec x_i = xx.col(i) - m;
rowvec xRinv = trans(x_i)*Rinv;
double quadform = sum(xRinv%xRinv);
y(i) = exp(-0.5*quadform + logSqrtDetvarcovM + constant);
if (!Log) {
y(i) = exp(-0.5*quadform + logSqrtDetvarcovM + constant);
} else{
y(i) = -0.5*quadform + logSqrtDetvarcovM + constant;
}
}
return y;
}
void JacobiRotations(int n, double rho_max, mat &A, uvec diagonalA, mat &R){
double epsilon = pow(10,-8);
//cout << epsilon << endl;
// Starting w/A(1,2)
//cout << A << endl;
int k = 0; int l = 0;
double max_A = 0.0;
findMaximumElementOnNonDiagonal(A, k, l, max_A, n);
//cout << max_A << endl;
int numberOfIterations = 0;
int maxIterations = pow(10, 6);
//==================================================================================
//While-loop to change the non-diagonal elements to be approximately zero.
//==================================================================================
while( fabs(max_A) > epsilon && (double) numberOfIterations < maxIterations){
double c = 0.0;
double s = 0.0;
findSinAndCos(A, k, l, c, s);
// Rotate
rotateMatrix(A, k, l, c, s, n, R);
//Find max off-diagonal matrix element
findMaximumElementOnNonDiagonal(A, k, l, max_A, n);
numberOfIterations++;
}
//Preparing results
diagonalA = sort_index(A.diag());
//Printing
//printMatlabMatrix("A", A);
cout << "Number of iterations: " << numberOfIterations << endl;
//cout << rho_max << endl;
cout << "lambda0: " << A.diag()(diagonalA(0)) << " lambda1: " << A.diag()(diagonalA(1)) << " lambda2: " << A.diag()(diagonalA(2)) << endl;
//cout << "hallo" << endl;
}
mat nnls_solver_without_missing(mat & WtW, mat & WtA,
const mat & A, const mat & W, const mat & W1, const mat & H2, const umat & mask,
const double & eta, const double & beta, int max_iter, double rel_tol, int n_threads)
{
// A = [W, W1, W2] [H, H1, H2]^T.
// Where A has not missing
// WtW and WtA are auxiliary matrices, passed by referenced and can be modified
update_WtW(WtW, W, W1, H2);
update_WtA(WtA, W, W1, H2, A);
if (beta > 0) WtW += beta;
if (eta > 0) WtW.diag() += eta;
return nnls_solver(WtW, WtA, mask, max_iter, rel_tol, n_threads);
}
void fill_matrix(mat &A, vec pot, double h)
{
A.diag() = (2/(h*h) + pot);
A.diag(1).fill(-1/(h*h));
A.diag(-1).fill(-1/(h*h));
}
