[ index.php ] [ docs.php ] [ package.php ] [ test.php ] [ example.php ] [ download.php ]

Source Listing:


Viewing: EigenvalueDecomposition.php
<?php
/**
 *    @package JAMA
 *
 *    Class to obtain eigenvalues and eigenvectors of a real matrix.
 *
 *    If A is symmetric, then A = V*D*V' where the eigenvalue matrix D
 *    is diagonal and the eigenvector matrix V is orthogonal (i.e.
 *    A = V.times(D.times(V.transpose())) and V.times(V.transpose())
 *    equals the identity matrix).
 *
 *    If A is not symmetric, then the eigenvalue matrix D is block diagonal
 *    with the real eigenvalues in 1-by-1 blocks and any complex eigenvalues,
 *    lambda + i*mu, in 2-by-2 blocks, [lambda, mu; -mu, lambda].  The
 *    columns of V represent the eigenvectors in the sense that A*V = V*D,
 *    i.e. A.times(V) equals V.times(D).  The matrix V may be badly
 *    conditioned, or even singular, so the validity of the equation
 *    A = V*D*inverse(V) depends upon V.cond().
 *
 *    @author  Paul Meagher
 *    @license PHP v3.0
 *    @version 1.1
 */
class EigenvalueDecomposition {

    
/**
     *    Row and column dimension (square matrix).
     *    @var int
     */
    
private $n;

    
/**
     *    Internal symmetry flag.
     *    @var int
     */
    
private $issymmetric;

    
/**
     *    Arrays for internal storage of eigenvalues.
     *    @var array
     */
    
private $d = array();
    private 
$e = array();

    
/**
     *    Array for internal storage of eigenvectors.
     *    @var array
     */
    
private $V = array();

    
/**
    *    Array for internal storage of nonsymmetric Hessenberg form.
    *    @var array
    */
    
private $H = array();

    
/**
    *    Working storage for nonsymmetric algorithm.
    *    @var array
    */
    
private $ort;

    
/**
    *    Used for complex scalar division.
    *    @var float
    */
    
private $cdivr;
    private 
$cdivi;


    
/**
     *    Symmetric Householder reduction to tridiagonal form.
     *
     *    @access private
     */
    
private function tred2 () {
        
//  This is derived from the Algol procedures tred2 by
        //  Bowdler, Martin, Reinsch, and Wilkinson, Handbook for
        //  Auto. Comp., Vol.ii-Linear Algebra, and the corresponding
        //  Fortran subroutine in EISPACK.
        
$this->$this->V[$this->n-1];
        
// Householder reduction to tridiagonal form.
        
for ($i $this->n-1$i 0; --$i) {
            
$i_ $i -1;
            
// Scale to avoid under/overflow.
            
$h $scale 0.0;
            
$scale += array_sum(array_map(abs$this->d));
            if (
$scale == 0.0) {
                
$this->e[$i] = $this->d[$i_];
                
$this->array_slice($this->V[$i_], 0$i_);
                for (
$j 0$j $i; ++$j) {
                    
$this->V[$j][$i] = $this->V[$i][$j] = 0.0;
                }
            } else {
                
// Generate Householder vector.
                
for ($k 0$k $i; ++$k) {
                    
$this->d[$k] /= $scale;
                    
$h += pow($this->d[$k], 2);
                }
                
$f $this->d[$i_];
                
$g sqrt($h);
                if (
$f 0) {
                    
$g = -$g;
                }
                
$this->e[$i] = $scale $g;
                
$h $h $f $g;
                
$this->d[$i_] = $f $g;
                for (
$j 0$j $i; ++$j) {
                    
$this->e[$j] = 0.0;
                }
                
// Apply similarity transformation to remaining columns.
                
for ($j 0$j $i; ++$j) {
                    
$f $this->d[$j];
                    
$this->V[$j][$i] = $f;
                    
$g $this->e[$j] + $this->V[$j][$j] * $f;
                    for (
$k $j+1$k <= $i_; ++$k) {
                        
$g += $this->V[$k][$j] * $this->d[$k];
                        
$this->e[$k] += $this->V[$k][$j] * $f;
                    }
                    
$this->e[$j] = $g;
                }
                
$f 0.0;
                for (
$j 0$j $i; ++$j) {
                    
$this->e[$j] /= $h;
                    
$f += $this->e[$j] * $this->d[$j];
                }
                
$hh $f / ($h);
                for (
$j=0$j $i; ++$j) {
                    
$this->e[$j] -= $hh $this->d[$j];
                }
                for (
$j 0$j $i; ++$j) {
                    
$f $this->d[$j];
                    
$g $this->e[$j];
                    for (
$k $j$k <= $i_; ++$k) {
                        
$this->V[$k][$j] -= ($f $this->e[$k] + $g $this->d[$k]);
                    }
                    
$this->d[$j] = $this->V[$i-1][$j];
                    
$this->V[$i][$j] = 0.0;
                }
            }
            
$this->d[$i] = $h;
        }

        
// Accumulate transformations.
        
for ($i 0$i $this->n-1; ++$i) {
            
$this->V[$this->n-1][$i] = $this->V[$i][$i];
            
$this->V[$i][$i] = 1.0;
            
$h $this->d[$i+1];
            if (
$h != 0.0) {
                for (
$k 0$k <= $i; ++$k) {
                    
$this->d[$k] = $this->V[$k][$i+1] / $h;
                }
                for (
$j 0$j <= $i; ++$j) {
                    
$g 0.0;
                    for (
$k 0$k <= $i; ++$k) {
                        
$g += $this->V[$k][$i+1] * $this->V[$k][$j];
                    }
                    for (
$k 0$k <= $i; ++$k) {
                        
$this->V[$k][$j] -= $g $this->d[$k];
                    }
                }
            }
            for (
$k 0$k <= $i; ++$k) {
                
$this->V[$k][$i+1] = 0.0;
            }
        }

        
$this->$this->V[$this->n-1];
        
$this->V[$this->n-1] = array_fill(0$j0.0);
        
$this->V[$this->n-1][$this->n-1] = 1.0;
        
$this->e[0] = 0.0;
    }


    
/**
     *    Symmetric tridiagonal QL algorithm.
     *
     *    This is derived from the Algol procedures tql2, by
     *    Bowdler, Martin, Reinsch, and Wilkinson, Handbook for
     *    Auto. Comp., Vol.ii-Linear Algebra, and the corresponding
     *    Fortran subroutine in EISPACK.
     *
     *    @access private
     */
    
private function tql2() {
        for (
$i 1$i $this->n; ++$i) {
            
$this->e[$i-1] = $this->e[$i];
        }
        
$this->e[$this->n-1] = 0.0;
        
$f 0.0;
        
$tst1 0.0;
        
$eps  pow(2.0,-52.0);

        for (
$l 0$l $this->n; ++$l) {
            
// Find small subdiagonal element
            
$tst1 max($tst1abs($this->d[$l]) + abs($this->e[$l]));
            
$m $l;
            while (
$m $this->n) {
                if (
abs($this->e[$m]) <= $eps $tst1)
                    break;
                ++
$m;
            }
            
// If m == l, $this->d[l] is an eigenvalue,
            // otherwise, iterate.
            
if ($m $l) {
                
$iter 0;
                do {
                    
// Could check iteration count here.
                    
$iter += 1;
                    
// Compute implicit shift
                    
$g $this->d[$l];
                    
$p = ($this->d[$l+1] - $g) / (2.0 $this->e[$l]);
                    
$r hypo($p1.0);
                    if (
$p 0)
                        
$r *= -1;
                    
$this->d[$l] = $this->e[$l] / ($p $r);
                    
$this->d[$l+1] = $this->e[$l] * ($p $r);
                    
$dl1 $this->d[$l+1];
                    
$h $g $this->d[$l];
                    for (
$i $l 2$i $this->n; ++$i)
                        
$this->d[$i] -= $h;
                    
$f += $h;
                    
// Implicit QL transformation.
                    
$p $this->d[$m];
                    
$c 1.0;
                    
$c2 $c3 $c;
                    
$el1 $this->e[$l 1];
                    
$s $s2 0.0;
                    for (
$i $m-1$i >= $l; --$i) {
                        
$c3 $c2;
                        
$c2 $c;
                        
$s2 $s;
                        
$g  $c $this->e[$i];
                        
$h  $c $p;
                        
$r  hypo($p$this->e[$i]);
                        
$this->e[$i+1] = $s $r;
                        
$s $this->e[$i] / $r;
                        
$c $p $r;
                        
$p $c $this->d[$i] - $s $g;
                        
$this->d[$i+1] = $h $s * ($c $g $s $this->d[$i]);
                        
// Accumulate transformation.
                        
for ($k 0$k $this->n; ++$k) {
                            
$h $this->V[$k][$i+1];
                            
$this->V[$k][$i+1] = $s $this->V[$k][$i] + $c $h;
                            
$this->V[$k][$i] = $c $this->V[$k][$i] - $s $h;
                        }
                    }
                    
$p = -$s $s2 $c3 $el1 $this->e[$l] / $dl1;
                    
$this->e[$l] = $s $p;
                    
$this->d[$l] = $c $p;
                
// Check for convergence.
                
} while (abs($this->e[$l]) > $eps $tst1);
            }
            
$this->d[$l] = $this->d[$l] + $f;
            
$this->e[$l] = 0.0;
        }

        
// Sort eigenvalues and corresponding vectors.
        
for ($i 0$i $this->1; ++$i) {
            
$k $i;
            
$p $this->d[$i];
            for (
$j $i+1$j $this->n; ++$j) {
                if (
$this->d[$j] < $p) {
                    
$k $j;
                    
$p $this->d[$j];
                }
            }
            if (
$k != $i) {
                
$this->d[$k] = $this->d[$i];
                
$this->d[$i] = $p;
                for (
$j 0$j $this->n; ++$j) {
                    
$p $this->V[$j][$i];
                    
$this->V[$j][$i] = $this->V[$j][$k];
                    
$this->V[$j][$k] = $p;
                }
            }
        }
    }


    
/**
     *    Nonsymmetric reduction to Hessenberg form.
     *
     *    This is derived from the Algol procedures orthes and ortran,
     *    by Martin and Wilkinson, Handbook for Auto. Comp.,
     *    Vol.ii-Linear Algebra, and the corresponding
     *    Fortran subroutines in EISPACK.
     *
     *    @access private
     */
    
private function orthes () {
        
$low  0;
        
$high $this->n-1;

        for (
$m $low+1$m <= $high-1; ++$m) {
            
// Scale column.
            
$scale 0.0;
            for (
$i $m$i <= $high; ++$i) {
                
$scale $scale abs($this->H[$i][$m-1]);
            }
            if (
$scale != 0.0) {
                
// Compute Householder transformation.
                
$h 0.0;
                for (
$i $high$i >= $m; --$i) {
                    
$this->ort[$i] = $this->H[$i][$m-1] / $scale;
                    
$h += $this->ort[$i] * $this->ort[$i];
                }
                
$g sqrt($h);
                if (
$this->ort[$m] > 0) {
                    
$g *= -1;
                }
                
$h -= $this->ort[$m] * $g;
                
$this->ort[$m] -= $g;
                
// Apply Householder similarity transformation
                // H = (I -u * u' / h) * H * (I -u * u') / h)
                
for ($j $m$j $this->n; ++$j) {
                    
$f 0.0;
                    for (
$i $high$i >= $m; --$i) {
                        
$f += $this->ort[$i] * $this->H[$i][$j];
                    }
                    
$f /= $h;
                    for (
$i $m$i <= $high; ++$i) {
                        
$this->H[$i][$j] -= $f $this->ort[$i];
                    }
                }
                for (
$i 0$i <= $high; ++$i) {
                    
$f 0.0;
                    for (
$j $high$j >= $m; --$j) {
                        
$f += $this->ort[$j] * $this->H[$i][$j];
                    }
                    
$f $f $h;
                    for (
$j $m$j <= $high; ++$j) {
                        
$this->H[$i][$j] -= $f $this->ort[$j];
                    }
                }
                
$this->ort[$m] = $scale $this->ort[$m];
                
$this->H[$m][$m-1] = $scale $g;
            }
        }

        
// Accumulate transformations (Algol's ortran).
        
for ($i 0$i $this->n; ++$i) {
            for (
$j 0$j $this->n; ++$j) {
                
$this->V[$i][$j] = ($i == $j 1.0 0.0);
            }
        }
        for (
$m $high-1$m >= $low+1; --$m) {
            if (
$this->H[$m][$m-1] != 0.0) {
                for (
$i $m+1$i <= $high; ++$i) {
                    
$this->ort[$i] = $this->H[$i][$m-1];
                }
                for (
$j $m$j <= $high; ++$j) {
                    
$g 0.0;
                    for (
$i $m$i <= $high; ++$i) {
                        
$g += $this->ort[$i] * $this->V[$i][$j];
                    }
                    
// Double division avoids possible underflow
                    
$g = ($g $this->ort[$m]) / $this->H[$m][$m-1];
                    for (
$i $m$i <= $high; ++$i) {
                        
$this->V[$i][$j] += $g $this->ort[$i];
                    }
                }
            }
        }
    }


    
/**
     *    Performs complex division.
     *
     *    @access private
     */
    
private function cdiv($xr$xi$yr$yi) {
        if (
abs($yr) > abs($yi)) {
            
$r $yi $yr;
            
$d $yr $r $yi;
            
$this->cdivr = ($xr $r $xi) / $d;
            
$this->cdivi = ($xi $r $xr) / $d;
        } else {
            
$r $yr $yi;
            
$d $yi $r $yr;
            
$this->cdivr = ($r $xr $xi) / $d;
            
$this->cdivi = ($r $xi $xr) / $d;
        }
    }


    
/**
     *    Nonsymmetric reduction from Hessenberg to real Schur form.
     *
     *    Code is derived from the Algol procedure hqr2,
     *    by Martin and Wilkinson, Handbook for Auto. Comp.,
     *    Vol.ii-Linear Algebra, and the corresponding
     *    Fortran subroutine in EISPACK.
     *
     *    @access private
     */
    
private function hqr2 () {
        
//  Initialize
        
$nn $this->n;
        
$n  $nn 1;
        
$low 0;
        
$high $nn 1;
        
$eps pow(2.0, -52.0);
        
$exshift 0.0;
        
$p $q $r $s $z 0;
        
// Store roots isolated by balanc and compute matrix norm
        
$norm 0.0;

        for (
$i 0$i $nn; ++$i) {
            if ((
$i $low) OR ($i $high)) {
                
$this->d[$i] = $this->H[$i][$i];
                
$this->e[$i] = 0.0;
            }
            for (
$j max($i-10); $j $nn; ++$j) {
                
$norm $norm abs($this->H[$i][$j]);
            }
        }

        
// Outer loop over eigenvalue index
        
$iter 0;
        while (
$n >= $low) {
            
// Look for single small sub-diagonal element
            
$l $n;
            while (
$l $low) {
                
$s abs($this->H[$l-1][$l-1]) + abs($this->H[$l][$l]);
                if (
$s == 0.0) {
                    
$s $norm;
                }
                if (
abs($this->H[$l][$l-1]) < $eps $s) {
                    break;
                }
                --
$l;
            }
            
// Check for convergence
            // One root found
            
if ($l == $n) {
                
$this->H[$n][$n] = $this->H[$n][$n] + $exshift;
                
$this->d[$n] = $this->H[$n][$n];
                
$this->e[$n] = 0.0;
                --
$n;
                
$iter 0;
            
// Two roots found
            
} else if ($l == $n-1) {
                
$w $this->H[$n][$n-1] * $this->H[$n-1][$n];
                
$p = ($this->H[$n-1][$n-1] - $this->H[$n][$n]) / 2.0;
                
$q $p $p $w;
                
$z sqrt(abs($q));
                
$this->H[$n][$n] = $this->H[$n][$n] + $exshift;
                
$this->H[$n-1][$n-1] = $this->H[$n-1][$n-1] + $exshift;
                
$x $this->H[$n][$n];
                
// Real pair
                
if ($q >= 0) {
                    if (
$p >= 0) {
                        
$z $p $z;
                    } else {
                        
$z $p $z;
                    }
                    
$this->d[$n-1] = $x $z;
                    
$this->d[$n] = $this->d[$n-1];
                    if (
$z != 0.0) {
                        
$this->d[$n] = $x $w $z;
                    }
                    
$this->e[$n-1] = 0.0;
                    
$this->e[$n] = 0.0;
                    
$x $this->H[$n][$n-1];
                    
$s abs($x) + abs($z);
                    
$p $x $s;
                    
$q $z $s;
                    
$r sqrt($p $p $q $q);
                    
$p $p $r;
                    
$q $q $r;
                    
// Row modification
                    
for ($j $n-1$j $nn; ++$j) {
                        
$z $this->H[$n-1][$j];
                        
$this->H[$n-1][$j] = $q $z $p $this->H[$n][$j];
                        
$this->H[$n][$j] = $q $this->H[$n][$j] - $p $z;
                    }
                    
// Column modification
                    
for ($i 0$i <= n; ++$i) {
                        
$z $this->H[$i][$n-1];
                        
$this->H[$i][$n-1] = $q $z $p $this->H[$i][$n];
                        
$this->H[$i][$n] = $q $this->H[$i][$n] - $p $z;
                    }
                    
// Accumulate transformations
                    
for ($i $low$i <= $high; ++$i) {
                        
$z $this->V[$i][$n-1];
                        
$this->V[$i][$n-1] = $q $z $p $this->V[$i][$n];
                        
$this->V[$i][$n] = $q $this->V[$i][$n] - $p $z;
                    }
                
// Complex pair
                
} else {
                    
$this->d[$n-1] = $x $p;
                    
$this->d[$n] = $x $p;
                    
$this->e[$n-1] = $z;
                    
$this->e[$n] = -$z;
                }
                
$n $n 2;
                
$iter 0;
            
// No convergence yet
            
} else {
                
// Form shift
                
$x $this->H[$n][$n];
                
$y 0.0;
                
$w 0.0;
                if (
$l $n) {
                    
$y $this->H[$n-1][$n-1];
                    
$w $this->H[$n][$n-1] * $this->H[$n-1][$n];
                }
                
// Wilkinson's original ad hoc shift
                
if ($iter == 10) {
                    
$exshift += $x;
                    for (
$i $low$i <= $n; ++$i) {
                        
$this->H[$i][$i] -= $x;
                    }
                    
$s abs($this->H[$n][$n-1]) + abs($this->H[$n-1][$n-2]);
                    
$x $y 0.75 $s;
                    
$w = -0.4375 $s $s;
                }
                
// MATLAB's new ad hoc shift
                
if ($iter == 30) {
                    
$s = ($y $x) / 2.0;
                    
$s $s $s $w;
                    if (
$s 0) {
                        
$s sqrt($s);
                        if (
$y $x) {
                            
$s = -$s;
                        }
                        
$s $x $w / (($y $x) / 2.0 $s);
                        for (
$i $low$i <= $n; ++$i) {
                            
$this->H[$i][$i] -= $s;
                        }
                        
$exshift += $s;
                        
$x $y $w 0.964;
                    }
                }
                
// Could check iteration count here.
                
$iter $iter 1;
                
// Look for two consecutive small sub-diagonal elements
                
$m $n 2;
                while (
$m >= $l) {
                    
$z $this->H[$m][$m];
                    
$r $x $z;
                    
$s $y $z;
                    
$p = ($r $s $w) / $this->H[$m+1][$m] + $this->H[$m][$m+1];
                    
$q $this->H[$m+1][$m+1] - $z $r $s;
                    
$r $this->H[$m+2][$m+1];
                    
$s abs($p) + abs($q) + abs($r);
                    
$p $p $s;
                    
$q $q $s;
                    
$r $r $s;
                    if (
$m == $l) {
                        break;
                    }
                    if (
abs($this->H[$m][$m-1]) * (abs($q) + abs($r)) <
                        
$eps * (abs($p) * (abs($this->H[$m-1][$m-1]) + abs($z) + abs($this->H[$m+1][$m+1])))) {
                        break;
                    }
                    --
$m;
                }
                for (
$i $m 2$i <= $n; ++$i) {
                    
$this->H[$i][$i-2] = 0.0;
                    if (
$i $m+2) {
                        
$this->H[$i][$i-3] = 0.0;
                    }
                }
                
// Double QR step involving rows l:n and columns m:n
                
for ($k $m$k <= $n-1; ++$k) {
                    
$notlast = ($k != $n-1);
                    if (
$k != $m) {
                        
$p $this->H[$k][$k-1];
                        
$q $this->H[$k+1][$k-1];
                        
$r = ($notlast $this->H[$k+2][$k-1] : 0.0);
                        
$x abs($p) + abs($q) + abs($r);
                        if (
$x != 0.0) {
                            
$p $p $x;
                            
$q $q $x;
                            
$r $r $x;
                        }
                    }
                    if (
$x == 0.0) {
                        break;
                    }
                    
$s sqrt($p $p $q $q $r $r);
                    if (
$p 0) {
                        
$s = -$s;
                    }
                    if (
$s != 0) {
                        if (
$k != $m) {
                            
$this->H[$k][$k-1] = -$s $x;
                        } elseif (
$l != $m) {
                            
$this->H[$k][$k-1] = -$this->H[$k][$k-1];
                        }
                        
$p $p $s;
                        
$x $p $s;
                        
$y $q $s;
                        
$z $r $s;
                        
$q $q $p;
                        
$r $r $p;
                        
// Row modification
                        
for ($j $k$j $nn; ++$j) {
                            
$p $this->H[$k][$j] + $q $this->H[$k+1][$j];
                            if (
$notlast) {
                                
$p $p $r $this->H[$k+2][$j];
                                
$this->H[$k+2][$j] = $this->H[$k+2][$j] - $p $z;
                            }
                            
$this->H[$k][$j] = $this->H[$k][$j] - $p $x;
                            
$this->H[$k+1][$j] = $this->H[$k+1][$j] - $p $y;
                        }
                        
// Column modification
                        
for ($i 0$i <= min($n$k+3); ++$i) {
                            
$p $x $this->H[$i][$k] + $y $this->H[$i][$k+1];
                            if (
$notlast) {
                                
$p $p $z $this->H[$i][$k+2];
                                
$this->H[$i][$k+2] = $this->H[$i][$k+2] - $p $r;
                            }
                            
$this->H[$i][$k] = $this->H[$i][$k] - $p;
                            
$this->H[$i][$k+1] = $this->H[$i][$k+1] - $p $q;
                        }
                        
// Accumulate transformations
                        
for ($i $low$i <= $high; ++$i) {
                            
$p $x $this->V[$i][$k] + $y $this->V[$i][$k+1];
                            if (
$notlast) {
                                
$p $p $z $this->V[$i][$k+2];
                                
$this->V[$i][$k+2] = $this->V[$i][$k+2] - $p $r;
                            }
                            
$this->V[$i][$k] = $this->V[$i][$k] - $p;
                            
$this->V[$i][$k+1] = $this->V[$i][$k+1] - $p $q;
                        }
                    }  
// ($s != 0)
                
}  // k loop
            
}  // check convergence
        
}  // while ($n >= $low)

        // Backsubstitute to find vectors of upper triangular form
        
if ($norm == 0.0) {
            return;
        }

        for (
$n $nn-1$n >= 0; --$n) {
            
$p $this->d[$n];
            
$q $this->e[$n];
            
// Real vector
            
if ($q == 0) {
                
$l $n;
                
$this->H[$n][$n] = 1.0;
                for (
$i $n-1$i >= 0; --$i) {
                    
$w $this->H[$i][$i] - $p;
                    
$r 0.0;
                    for (
$j $l$j <= $n; ++$j) {
                        
$r $r $this->H[$i][$j] * $this->H[$j][$n];
                    }
                    if (
$this->e[$i] < 0.0) {
                        
$z $w;
                        
$s $r;
                    } else {
                        
$l $i;
                        if (
$this->e[$i] == 0.0) {
                            if (
$w != 0.0) {
                                
$this->H[$i][$n] = -$r $w;
                            } else {
                                
$this->H[$i][$n] = -$r / ($eps $norm);
                            }
                        
// Solve real equations
                        
} else {
                            
$x $this->H[$i][$i+1];
                            
$y $this->H[$i+1][$i];
                            
$q = ($this->d[$i] - $p) * ($this->d[$i] - $p) + $this->e[$i] * $this->e[$i];
                            
$t = ($x $s $z $r) / $q;
                            
$this->H[$i][$n] = $t;
                            if (
abs($x) > abs($z)) {
                                
$this->H[$i+1][$n] = (-$r $w $t) / $x;
                            } else {
                                
$this->H[$i+1][$n] = (-$s $y $t) / $z;
                            }
                        }
                        
// Overflow control
                        
$t abs($this->H[$i][$n]);
                        if ((
$eps $t) * $t 1) {
                            for (
$j $i$j <= $n; ++$j) {
                                
$this->H[$j][$n] = $this->H[$j][$n] / $t;
                            }
                        }
                    }
                }
            
// Complex vector
            
} else if ($q 0) {
                
$l $n-1;
                
// Last vector component imaginary so matrix is triangular
                
if (abs($this->H[$n][$n-1]) > abs($this->H[$n-1][$n])) {
                    
$this->H[$n-1][$n-1] = $q $this->H[$n][$n-1];
                    
$this->H[$n-1][$n] = -($this->H[$n][$n] - $p) / $this->H[$n][$n-1];
                } else {
                    
$this->cdiv(0.0, -$this->H[$n-1][$n], $this->H[$n-1][$n-1] - $p$q);
                    
$this->H[$n-1][$n-1] = $this->cdivr;
                    
$this->H[$n-1][$n]   = $this->cdivi;
                }
                
$this->H[$n][$n-1] = 0.0;
                
$this->H[$n][$n] = 1.0;
                for (
$i $n-2$i >= 0; --$i) {
                    
// double ra,sa,vr,vi;
                    
$ra 0.0;
                    
$sa 0.0;
                    for (
$j $l$j <= $n; ++$j) {
                        
$ra $ra $this->H[$i][$j] * $this->H[$j][$n-1];
                        
$sa $sa $this->H[$i][$j] * $this->H[$j][$n];
                    }
                    
$w $this->H[$i][$i] - $p;
                    if (
$this->e[$i] < 0.0) {
                        
$z $w;
                        
$r $ra;
                        
$s $sa;
                    } else {
                        
$l $i;
                        if (
$this->e[$i] == 0) {
                            
$this->cdiv(-$ra, -$sa$w$q);
                            
$this->H[$i][$n-1] = $this->cdivr;
                            
$this->H[$i][$n]   = $this->cdivi;
                        } else {
                            
// Solve complex equations
                            
$x $this->H[$i][$i+1];
                            
$y $this->H[$i+1][$i];
                            
$vr = ($this->d[$i] - $p) * ($this->d[$i] - $p) + $this->e[$i] * $this->e[$i] - $q $q;
                            
$vi = ($this->d[$i] - $p) * 2.0 $q;
                            if (
$vr == 0.0 $vi == 0.0) {
                                
$vr $eps $norm * (abs($w) + abs($q) + abs($x) + abs($y) + abs($z));
                            }
                            
$this->cdiv($x $r $z $ra $q $sa$x $s $z $sa $q $ra$vr$vi);
                            
$this->H[$i][$n-1] = $this->cdivr;
                            
$this->H[$i][$n]   = $this->cdivi;
                            if (
abs($x) > (abs($z) + abs($q))) {
                                
$this->H[$i+1][$n-1] = (-$ra $w $this->H[$i][$n-1] + $q $this->H[$i][$n]) / $x;
                                
$this->H[$i+1][$n] = (-$sa $w $this->H[$i][$n] - $q $this->H[$i][$n-1]) / $x;
                            } else {
                                
$this->cdiv(-$r $y $this->H[$i][$n-1], -$s $y $this->H[$i][$n], $z$q);
                                
$this->H[$i+1][$n-1] = $this->cdivr;
                                
$this->H[$i+1][$n]   = $this->cdivi;
                            }
                        }
                        
// Overflow control
                        
$t max(abs($this->H[$i][$n-1]),abs($this->H[$i][$n]));
                        if ((
$eps $t) * $t 1) {
                            for (
$j $i$j <= $n; ++$j) {
                                
$this->H[$j][$n-1] = $this->H[$j][$n-1] / $t;
                                
$this->H[$j][$n]   = $this->H[$j][$n] / $t;
                            }
                        }
                    } 
// end else
                
// end for
            
// end else for complex case
        
// end for

        // Vectors of isolated roots
        
for ($i 0$i $nn; ++$i) {
            if (
$i $low $i $high) {
                for (
$j $i$j $nn; ++$j) {
                    
$this->V[$i][$j] = $this->H[$i][$j];
                }
            }
        }

        
// Back transformation to get eigenvectors of original matrix
        
for ($j $nn-1$j >= $low; --$j) {
            for (
$i $low$i <= $high; ++$i) {
                
$z 0.0;
                for (
$k $low$k <= min($j,$high); ++$k) {
                    
$z $z $this->V[$i][$k] * $this->H[$k][$j];
                }
                
$this->V[$i][$j] = $z;
            }
        }
    } 
// end hqr2


    /**
     *    Constructor: Check for symmetry, then construct the eigenvalue decomposition
     *
     *    @access public
     *    @param A  Square matrix
     *    @return Structure to access D and V.
     */
    
public function __construct($Arg) {
        
$this->$Arg->getArray();
        
$this->$Arg->getColumnDimension();

        
$issymmetric true;
        for (
$j 0; ($j $this->n) & $issymmetric; ++$j) {
            for (
$i 0; ($i $this->n) & $issymmetric; ++$i) {
                
$issymmetric = ($this->A[$i][$j] == $this->A[$j][$i]);
            }
        }

        if (
$issymmetric) {
            
$this->$this->A;
            
// Tridiagonalize.
            
$this->tred2();
            
// Diagonalize.
            
$this->tql2();
        } else {
            
$this->$this->A;
            
$this->ort = array();
            
// Reduce to Hessenberg form.
            
$this->orthes();
            
// Reduce Hessenberg to real Schur form.
            
$this->hqr2();
        }
    }


    
/**
     *    Return the eigenvector matrix
     *
     *    @access public
     *    @return V
     */
    
public function getV() {
        return new 
Matrix($this->V$this->n$this->n);
    }


    
/**
     *    Return the real parts of the eigenvalues
     *
     *    @access public
     *    @return real(diag(D))
     */
    
public function getRealEigenvalues() {
        return 
$this->d;
    }


    
/**
     *    Return the imaginary parts of the eigenvalues
     *
     *    @access public
     *    @return imag(diag(D))
     */
    
public function getImagEigenvalues() {
        return 
$this->e;
    }


    
/**
     *    Return the block diagonal eigenvalue matrix
     *
     *    @access public
     *    @return D
     */
    
public function getD() {
        for (
$i 0$i $this->n; ++$i) {
            
$D[$i] = array_fill(0$this->n0.0);
            
$D[$i][$i] = $this->d[$i];
            if (
$this->e[$i] == 0) {
                continue;
            }
            
$o = ($this->e[$i] > 0) ? $i $i 1;
            
$D[$i][$o] = $this->e[$i];
        }
        return new 
Matrix($D);
    }

}    
//    class EigenvalueDecomposition