c@241: /*********************************/
c@241: /* Principal Components Analysis */
c@241: /*********************************/
c@241: 
c@241: /*********************************************************************/
c@241: /* Principal Components Analysis or the Karhunen-Loeve expansion is a
c@241:    classical method for dimensionality reduction or exploratory data
c@241:    analysis.  One reference among many is: F. Murtagh and A. Heck,
c@241:    Multivariate Data Analysis, Kluwer Academic, Dordrecht, 1987.
c@241: 
c@241:    Author:
c@241:    F. Murtagh
c@241:    Phone:        + 49 89 32006298 (work)
c@241:                  + 49 89 965307 (home)
c@241:    Earn/Bitnet:  fionn@dgaeso51,  fim@dgaipp1s,  murtagh@stsci
c@241:    Span:         esomc1::fionn
c@241:    Internet:     murtagh@scivax.stsci.edu
c@241:    
c@241:    F. Murtagh, Munich, 6 June 1989                                   */   
c@241: /*********************************************************************/
c@241: 
c@241: #include <stdio.h>
c@241: #include <stdlib.h>
c@241: #include <math.h>
c@241: 
c@241: #include "pca.h"
c@241: 
c@241: #define SIGN(a, b) ( (b) < 0 ? -fabs(a) : fabs(a) )
c@241: 
c@241: /**  Variance-covariance matrix: creation  *****************************/
c@241: 
c@241: /* Create m * m covariance matrix from given n * m data matrix. */
c@241: void covcol(double** data, int n, int m, double** symmat)
c@241: {
c@241: double *mean;
c@241: int i, j, j1, j2;
c@241: 
c@241: /* Allocate storage for mean vector */
c@241: 
c@241: mean = (double*) malloc(m*sizeof(double));
c@241: 
c@241: /* Determine mean of column vectors of input data matrix */
c@241: 
c@241: for (j = 0; j < m; j++)
c@241:     {
c@241:     mean[j] = 0.0;
c@241:     for (i = 0; i < n; i++)
c@241:         {
c@241:         mean[j] += data[i][j];
c@241:         }
c@241:     mean[j] /= (double)n;
c@241:     }
c@241: 
c@241: /*
c@241: printf("\nMeans of column vectors:\n");
c@241: for (j = 0; j < m; j++)  {
c@241:     printf("%12.1f",mean[j]);  }   printf("\n");
c@241:  */
c@241: 
c@241: /* Center the column vectors. */
c@241: 
c@241: for (i = 0; i < n; i++)
c@241:     {
c@241:     for (j = 0; j < m; j++)
c@241:         {
c@241:         data[i][j] -= mean[j];
c@241:         }
c@241:     }
c@241: 
c@241: /* Calculate the m * m covariance matrix. */
c@241: for (j1 = 0; j1 < m; j1++)
c@241:     {
c@241:     for (j2 = j1; j2 < m; j2++)
c@241:         {
c@241:         symmat[j1][j2] = 0.0;
c@241:         for (i = 0; i < n; i++)
c@241:             {
c@241:             symmat[j1][j2] += data[i][j1] * data[i][j2];
c@241:             }
c@241:         symmat[j2][j1] = symmat[j1][j2];
c@241:         }
c@241:     }
c@241: 
c@241: free(mean);
c@241: 
c@241: return;
c@241: 
c@241: }
c@241: 
c@241: /**  Error handler  **************************************************/
c@241: 
c@241: void erhand(char* err_msg)
c@241: {
c@241:     fprintf(stderr,"Run-time error:\n");
c@241:     fprintf(stderr,"%s\n", err_msg);
c@241:     fprintf(stderr,"Exiting to system.\n");
c@241:     exit(1);
c@241: }
c@241: 
c@241: 
c@241: /**  Reduce a real, symmetric matrix to a symmetric, tridiag. matrix. */
c@241: 
c@241: /* Householder reduction of matrix a to tridiagonal form.
c@241: Algorithm: Martin et al., Num. Math. 11, 181-195, 1968.
c@241: Ref: Smith et al., Matrix Eigensystem Routines -- EISPACK Guide
c@241: Springer-Verlag, 1976, pp. 489-494.
c@241: W H Press et al., Numerical Recipes in C, Cambridge U P,
c@241: 1988, pp. 373-374.  */
c@241: void tred2(double** a, int n, double* d, double* e)
c@241: {
c@241: 	int l, k, j, i;
c@241: 	double scale, hh, h, g, f;
c@241: 	
c@241: 	for (i = n-1; i >= 1; i--)
c@241:     {
c@241: 		l = i - 1;
c@241: 		h = scale = 0.0;
c@241: 		if (l > 0)
c@241: 		{
c@241: 			for (k = 0; k <= l; k++)
c@241: 				scale += fabs(a[i][k]);
c@241: 			if (scale == 0.0)
c@241: 				e[i] = a[i][l];
c@241: 			else
c@241: 			{
c@241: 				for (k = 0; k <= l; k++)
c@241: 				{
c@241: 					a[i][k] /= scale;
c@241: 					h += a[i][k] * a[i][k];
c@241: 				}
c@241: 				f = a[i][l];
c@241: 				g = f>0 ? -sqrt(h) : sqrt(h);
c@241: 				e[i] = scale * g;
c@241: 				h -= f * g;
c@241: 				a[i][l] = f - g;
c@241: 				f = 0.0;
c@241: 				for (j = 0; j <= l; j++)
c@241: 				{
c@241: 					a[j][i] = a[i][j]/h;
c@241: 					g = 0.0;
c@241: 					for (k = 0; k <= j; k++)
c@241: 						g += a[j][k] * a[i][k];
c@241: 					for (k = j+1; k <= l; k++)
c@241: 						g += a[k][j] * a[i][k];
c@241: 					e[j] = g / h;
c@241: 					f += e[j] * a[i][j];
c@241: 				}
c@241: 				hh = f / (h + h);
c@241: 				for (j = 0; j <= l; j++)
c@241: 				{
c@241: 					f = a[i][j];
c@241: 					e[j] = g = e[j] - hh * f;
c@241: 					for (k = 0; k <= j; k++)
c@241: 						a[j][k] -= (f * e[k] + g * a[i][k]);
c@241: 				}
c@241: 			}
c@241: 		}
c@241: 		else
c@241: 			e[i] = a[i][l];
c@241: 		d[i] = h;
c@241:     }
c@241: 	d[0] = 0.0;
c@241: 	e[0] = 0.0;
c@241: 	for (i = 0; i < n; i++)
c@241:     {
c@241: 		l = i - 1;
c@241: 		if (d[i])
c@241: 		{
c@241: 			for (j = 0; j <= l; j++)
c@241: 			{
c@241: 				g = 0.0;
c@241: 				for (k = 0; k <= l; k++)
c@241: 					g += a[i][k] * a[k][j];
c@241: 				for (k = 0; k <= l; k++)
c@241: 					a[k][j] -= g * a[k][i];
c@241: 			}
c@241: 		}
c@241: 		d[i] = a[i][i];
c@241: 		a[i][i] = 1.0;
c@241: 		for (j = 0; j <= l; j++)
c@241: 			a[j][i] = a[i][j] = 0.0;
c@241:     }
c@241: }
c@241: 
c@241: /**  Tridiagonal QL algorithm -- Implicit  **********************/
c@241: 
c@241: void tqli(double* d, double* e, int n, double** z)
c@241: {
c@241: 	int m, l, iter, i, k;
c@241: 	double s, r, p, g, f, dd, c, b;
c@241: 	
c@241: 	for (i = 1; i < n; i++)
c@241: 		e[i-1] = e[i];
c@241: 	e[n-1] = 0.0;
c@241: 	for (l = 0; l < n; l++)
c@241:     {
c@241: 		iter = 0;
c@241: 		do
c@241: 		{
c@241: 			for (m = l; m < n-1; m++)
c@241: 			{
c@241: 				dd = fabs(d[m]) + fabs(d[m+1]);
c@241: 				if (fabs(e[m]) + dd == dd) break;
c@241: 			}
c@241: 			if (m != l)
c@241: 			{
c@241: 				if (iter++ == 30) erhand("No convergence in TLQI.");
c@241: 				g = (d[l+1] - d[l]) / (2.0 * e[l]);
c@241: 				r = sqrt((g * g) + 1.0);
c@241: 				g = d[m] - d[l] + e[l] / (g + SIGN(r, g));
c@241: 				s = c = 1.0;
c@241: 				p = 0.0;
c@241: 				for (i = m-1; i >= l; i--)
c@241: 				{
c@241: 					f = s * e[i];
c@241: 					b = c * e[i];
c@241: 					if (fabs(f) >= fabs(g))
c@241:                     {
c@241: 						c = g / f;
c@241: 						r = sqrt((c * c) + 1.0);
c@241: 						e[i+1] = f * r;
c@241: 						c *= (s = 1.0/r);
c@241:                     }
c@241: 					else
c@241:                     {
c@241: 						s = f / g;
c@241: 						r = sqrt((s * s) + 1.0);
c@241: 						e[i+1] = g * r;
c@241: 						s *= (c = 1.0/r);
c@241:                     }
c@241: 					g = d[i+1] - p;
c@241: 					r = (d[i] - g) * s + 2.0 * c * b;
c@241: 					p = s * r;
c@241: 					d[i+1] = g + p;
c@241: 					g = c * r - b;
c@241: 					for (k = 0; k < n; k++)
c@241: 					{
c@241: 						f = z[k][i+1];
c@241: 						z[k][i+1] = s * z[k][i] + c * f;
c@241: 						z[k][i] = c * z[k][i] - s * f;
c@241: 					}
c@241: 				}
c@241: 				d[l] = d[l] - p;
c@241: 				e[l] = g;
c@241: 				e[m] = 0.0;
c@241: 			}
c@241: 		}  while (m != l);
c@241: 	}
c@241: }
c@241: 
c@241: /* In place projection onto basis vectors */
c@241: void pca_project(double** data, int n, int m, int ncomponents)
c@241: {
c@241: 	int  i, j, k, k2;
c@241: 	double  **symmat, **symmat2, *evals, *interm;
c@241: 	
c@241: 	//TODO: assert ncomponents < m
c@241: 	
c@241: 	symmat = (double**) malloc(m*sizeof(double*));
c@241: 	for (i = 0; i < m; i++)
c@241: 		symmat[i] = (double*) malloc(m*sizeof(double));
c@241: 		
c@241: 	covcol(data, n, m, symmat);
c@241: 	
c@241: 	/*********************************************************************
c@241: 		Eigen-reduction
c@241: 		**********************************************************************/
c@241: 	
c@241:     /* Allocate storage for dummy and new vectors. */
c@241:     evals = (double*) malloc(m*sizeof(double));     /* Storage alloc. for vector of eigenvalues */
c@241:     interm = (double*) malloc(m*sizeof(double));    /* Storage alloc. for 'intermediate' vector */
c@241:     //MALLOC_ARRAY(symmat2,m,m,double);    
c@241: 	//for (i = 0; i < m; i++) {
c@241: 	//	for (j = 0; j < m; j++) {
c@241: 	//		symmat2[i][j] = symmat[i][j]; /* Needed below for col. projections */
c@241: 	//	}
c@241: 	//}
c@241:     tred2(symmat, m, evals, interm);  /* Triangular decomposition */
c@241: tqli(evals, interm, m, symmat);   /* Reduction of sym. trid. matrix */
c@241: /* evals now contains the eigenvalues,
c@241: columns of symmat now contain the associated eigenvectors. */	
c@241: 
c@241: /*
c@241: 	printf("\nEigenvalues:\n");
c@241: 	for (j = m-1; j >= 0; j--) {
c@241: 		printf("%18.5f\n", evals[j]); }
c@241: 	printf("\n(Eigenvalues should be strictly positive; limited\n");
c@241: 	printf("precision machine arithmetic may affect this.\n");
c@241: 	printf("Eigenvalues are often expressed as cumulative\n");
c@241: 	printf("percentages, representing the 'percentage variance\n");
c@241: 	printf("explained' by the associated axis or principal component.)\n");
c@241: 	
c@241: 	printf("\nEigenvectors:\n");
c@241: 	printf("(First three; their definition in terms of original vbes.)\n");
c@241: 	for (j = 0; j < m; j++) {
c@241: 		for (i = 1; i <= 3; i++)  {
c@241: 			printf("%12.4f", symmat[j][m-i]);  }
c@241: 		printf("\n");  }
c@241:  */
c@241: 
c@241: /* Form projections of row-points on prin. components. */
c@241: /* Store in 'data', overwriting original data. */
c@241: for (i = 0; i < n; i++) {
c@241: 	for (j = 0; j < m; j++) {
c@241: 		interm[j] = data[i][j]; }   /* data[i][j] will be overwritten */
c@241:         for (k = 0; k < ncomponents; k++) {
c@241: 			data[i][k] = 0.0;
c@241: 			for (k2 = 0; k2 < m; k2++) {
c@241: 				data[i][k] += interm[k2] * symmat[k2][m-k-1]; }
c@241:         }
c@241: }
c@241: 
c@241: /*	
c@241: printf("\nProjections of row-points on first 3 prin. comps.:\n");
c@241:  for (i = 0; i < n; i++) {
c@241: 	 for (j = 0; j < 3; j++)  {
c@241: 		 printf("%12.4f", data[i][j]);  }
c@241: 	 printf("\n");  }
c@241:  */
c@241: 
c@241: /* Form projections of col.-points on first three prin. components. */
c@241: /* Store in 'symmat2', overwriting what was stored in this. */
c@241: //for (j = 0; j < m; j++) {
c@241: //	 for (k = 0; k < m; k++) {
c@241: //		 interm[k] = symmat2[j][k]; }  /*symmat2[j][k] will be overwritten*/
c@241: //  for (i = 0; i < 3; i++) {
c@241: //	symmat2[j][i] = 0.0;
c@241: //		for (k2 = 0; k2 < m; k2++) {
c@241: //			symmat2[j][i] += interm[k2] * symmat[k2][m-i-1]; }
c@241: //		if (evals[m-i-1] > 0.0005)   /* Guard against zero eigenvalue */
c@241: //			symmat2[j][i] /= sqrt(evals[m-i-1]);   /* Rescale */
c@241: //		else
c@241: //			symmat2[j][i] = 0.0;    /* Standard kludge */
c@241: //    }
c@241: // }
c@241: 
c@241: /*
c@241:  printf("\nProjections of column-points on first 3 prin. comps.:\n");
c@241:  for (j = 0; j < m; j++) {
c@241: 	 for (k = 0; k < 3; k++)  {
c@241: 		 printf("%12.4f", symmat2[j][k]);  }
c@241: 	 printf("\n");  }
c@241: 	*/
c@241: 
c@241: 
c@241: for (i = 0; i < m; i++)
c@241: 	free(symmat[i]);
c@241: free(symmat);
c@241: //FREE_ARRAY(symmat2,m);
c@241: free(evals);
c@241: free(interm);
c@241: 
c@241: }
c@241: 
c@241: 
c@241: