NICE_VisLearning/math/algebra/GenericMatrix.cpp

/** 
* @file GenericMatrix.cpp
* @brief matrix interface for indirect matrix multiplications
* @author Erik Rodner
* @date 05/14/2009

*/

#include "GenericMatrix.h"
#include <assert.h>

using namespace OBJREC;
using namespace NICE;
using namespace std;
	
GMSparse::GMSparse ( const NICE::Matrix & A, double epsilon )
{
    m_rows = A.rows();
    m_cols = A.cols();
    for ( u_int i = 0 ; i < m_rows ; i++ )
	for ( u_int j = 0 ; j < m_cols ; j++ )
	    if (fabs(A(i, j)) > epsilon )
			this->A.insert ( this->A.begin(), NICE::triplet<int, int, double> ( i, j,A(i, j) ) );		
}

void GMSparse::multiply ( NICE::Vector & y, const NICE::Vector & x ) const
{
    assert ( x.size() == m_cols );

    y.resize( m_rows );
    y.set(0.0);
    for ( vector< NICE::triplet<int, int, double> >::const_iterator k = this->A.begin();
	    k != this->A.end(); k++ )
    {
	int i = k->first;
	int j = k->second;
	double value = k->third; // = a_ij
	y[i] += value * x[j];
    }
}

GMCovariance::GMCovariance ( const NICE::Matrix *data )
{
    this->data = data;
    m_rows = data->rows();
    m_cols = m_rows;
}

void GMCovariance::multiply ( NICE::Vector & y, const NICE::Vector & x ) const
{
    assert ( x.size() == data->rows() );

	/*******************************************************************
	 * d_k = data(:,k) vector k in the data set
	 * f_k = d_k^T x 
	 * fSum = \sum_k f_k
	 * In the following we use the following relationship
	 * S = \sum_i d_i ( d_i - \mu )^T = \sum_i (d_i - \mu) (d_i - \mu)^T
	 * Sx = \sum_i d_i ( f_i - fSum/N )
	 * which seems to be very efficient
	 * *****************************************************************/

    NICE::Vector f ( data->cols() );
    f.set(0.0);
    double fSum = 0.0;
    for ( uint k = 0 ; k < data->cols() ; k++ )
    {
		for ( uint d = 0 ; d < data->rows() ; d++ )
			f[k] += (*data)(d, k) * x[d];
		fSum += f[k];
    }
    y.resize(x.size());
    y.set(0.0);
    int N = data->cols();
    for ( uint k = 0 ; k < data->cols() ; k++ )
		for ( int l = 0 ; l < (int)y.size(); l++ )
			y[l] += ( f[k] - fSum/N ) * ( (*data)(l, k) );

    for ( int i = 0 ; i < (int)y.size() ; i++ )
		y[i] /= N;
}