5 * For an m-by-n matrix A with m >= n, the singular value decomposition is
6 * an m-by-n orthogonal matrix U, an n-by-n diagonal matrix S, and
7 * an n-by-n orthogonal matrix V so that A = U*S*V'.
9 * The singular values, sigma[$k] = S[$k][$k], are ordered so that
10 * sigma[0] >= sigma[1] >= ... >= sigma[n-1].
12 * The singular value decompostion always exists, so the constructor will
13 * never fail. The matrix condition number and the effective numerical
14 * rank can be computed from this decomposition.
16 * @author Paul Meagher
20 class SingularValueDecomposition
{
23 * Internal storage of U.
29 * Internal storage of V.
35 * Internal storage of singular values.
54 * Construct the singular value decomposition
56 * Derived from LINPACK code.
58 * @param $A Rectangular matrix
59 * @return Structure to access U, S and V.
61 public function __construct($Arg) {
64 $A = $Arg->getArrayCopy();
65 $this->m
= $Arg->getRowDimension();
66 $this->n
= $Arg->getColumnDimension();
67 $nu = min($this->m
, $this->n
);
72 $nct = min($this->m
- 1, $this->n
);
73 $nrt = max(0, min($this->n
- 2, $this->m
));
75 // Reduce A to bidiagonal form, storing the diagonal elements
76 // in s and the super-diagonal elements in e.
77 for ($k = 0; $k < max($nct,$nrt); ++
$k) {
80 // Compute the transformation for the k-th column and
81 // place the k-th diagonal in s[$k].
82 // Compute 2-norm of k-th column without under/overflow.
84 for ($i = $k; $i < $this->m
; ++
$i) {
85 $this->s
[$k] = hypo($this->s
[$k], $A[$i][$k]);
87 if ($this->s
[$k] != 0.0) {
88 if ($A[$k][$k] < 0.0) {
89 $this->s
[$k] = -$this->s
[$k];
91 for ($i = $k; $i < $this->m
; ++
$i) {
92 $A[$i][$k] /= $this->s
[$k];
96 $this->s
[$k] = -$this->s
[$k];
99 for ($j = $k +
1; $j < $this->n
; ++
$j) {
100 if (($k < $nct) & ($this->s
[$k] != 0.0)) {
101 // Apply the transformation.
103 for ($i = $k; $i < $this->m
; ++
$i) {
104 $t +
= $A[$i][$k] * $A[$i][$j];
106 $t = -$t / $A[$k][$k];
107 for ($i = $k; $i < $this->m
; ++
$i) {
108 $A[$i][$j] +
= $t * $A[$i][$k];
110 // Place the k-th row of A into e for the
111 // subsequent calculation of the row transformation.
116 if ($wantu AND ($k < $nct)) {
117 // Place the transformation in U for subsequent back
119 for ($i = $k; $i < $this->m
; ++
$i) {
120 $this->U
[$i][$k] = $A[$i][$k];
125 // Compute the k-th row transformation and place the
126 // k-th super-diagonal in e[$k].
127 // Compute 2-norm without under/overflow.
129 for ($i = $k +
1; $i < $this->n
; ++
$i) {
130 $e[$k] = hypo($e[$k], $e[$i]);
133 if ($e[$k+
1] < 0.0) {
136 for ($i = $k +
1; $i < $this->n
; ++
$i) {
142 if (($k+
1 < $this->m
) AND ($e[$k] != 0.0)) {
143 // Apply the transformation.
144 for ($i = $k+
1; $i < $this->m
; ++
$i) {
147 for ($j = $k+
1; $j < $this->n
; ++
$j) {
148 for ($i = $k+
1; $i < $this->m
; ++
$i) {
149 $work[$i] +
= $e[$j] * $A[$i][$j];
152 for ($j = $k +
1; $j < $this->n
; ++
$j) {
153 $t = -$e[$j] / $e[$k+
1];
154 for ($i = $k +
1; $i < $this->m
; ++
$i) {
155 $A[$i][$j] +
= $t * $work[$i];
160 // Place the transformation in V for subsequent
161 // back multiplication.
162 for ($i = $k +
1; $i < $this->n
; ++
$i) {
163 $this->V
[$i][$k] = $e[$i];
169 // Set up the final bidiagonal matrix or order p.
170 $p = min($this->n
, $this->m +
1);
171 if ($nct < $this->n
) {
172 $this->s
[$nct] = $A[$nct][$nct];
175 $this->s
[$p-1] = 0.0;
178 $e[$nrt] = $A[$nrt][$p-1];
181 // If required, generate U.
183 for ($j = $nct; $j < $nu; ++
$j) {
184 for ($i = 0; $i < $this->m
; ++
$i) {
185 $this->U
[$i][$j] = 0.0;
187 $this->U
[$j][$j] = 1.0;
189 for ($k = $nct - 1; $k >= 0; --$k) {
190 if ($this->s
[$k] != 0.0) {
191 for ($j = $k +
1; $j < $nu; ++
$j) {
193 for ($i = $k; $i < $this->m
; ++
$i) {
194 $t +
= $this->U
[$i][$k] * $this->U
[$i][$j];
196 $t = -$t / $this->U
[$k][$k];
197 for ($i = $k; $i < $this->m
; ++
$i) {
198 $this->U
[$i][$j] +
= $t * $this->U
[$i][$k];
201 for ($i = $k; $i < $this->m
; ++
$i ) {
202 $this->U
[$i][$k] = -$this->U
[$i][$k];
204 $this->U
[$k][$k] = 1.0 +
$this->U
[$k][$k];
205 for ($i = 0; $i < $k - 1; ++
$i) {
206 $this->U
[$i][$k] = 0.0;
209 for ($i = 0; $i < $this->m
; ++
$i) {
210 $this->U
[$i][$k] = 0.0;
212 $this->U
[$k][$k] = 1.0;
217 // If required, generate V.
219 for ($k = $this->n
- 1; $k >= 0; --$k) {
220 if (($k < $nrt) AND ($e[$k] != 0.0)) {
221 for ($j = $k +
1; $j < $nu; ++
$j) {
223 for ($i = $k +
1; $i < $this->n
; ++
$i) {
224 $t +
= $this->V
[$i][$k]* $this->V
[$i][$j];
226 $t = -$t / $this->V
[$k+
1][$k];
227 for ($i = $k +
1; $i < $this->n
; ++
$i) {
228 $this->V
[$i][$j] +
= $t * $this->V
[$i][$k];
232 for ($i = 0; $i < $this->n
; ++
$i) {
233 $this->V
[$i][$k] = 0.0;
235 $this->V
[$k][$k] = 1.0;
239 // Main iteration loop for the singular values.
242 $eps = pow(2.0, -52.0);
245 // Here is where a test for too many iterations would go.
246 // This section of the program inspects for negligible
247 // elements in the s and e arrays. On completion the
248 // variables kase and k are set as follows:
249 // kase = 1 if s(p) and e[k-1] are negligible and k<p
250 // kase = 2 if s(k) is negligible and k<p
251 // kase = 3 if e[k-1] is negligible, k<p, and
252 // s(k), ..., s(p) are not negligible (qr step).
253 // kase = 4 if e(p-1) is negligible (convergence).
254 for ($k = $p - 2; $k >= -1; --$k) {
258 if (abs($e[$k]) <= $eps * (abs($this->s
[$k]) +
abs($this->s
[$k+
1]))) {
266 for ($ks = $p - 1; $ks >= $k; --$ks) {
270 $t = ($ks != $p ?
abs($e[$ks]) : 0.) +
($ks != $k +
1 ?
abs($e[$ks-1]) : 0.);
271 if (abs($this->s
[$ks]) <= $eps * $t) {
278 } else if ($ks == $p-1) {
287 // Perform the task indicated by kase.
289 // Deflate negligible s(p).
293 for ($j = $p - 2; $j >= $k; --$j) {
294 $t = hypo($this->s
[$j],$f);
295 $cs = $this->s
[$j] / $t;
299 $f = -$sn * $e[$j-1];
300 $e[$j-1] = $cs * $e[$j-1];
303 for ($i = 0; $i < $this->n
; ++
$i) {
304 $t = $cs * $this->V
[$i][$j] +
$sn * $this->V
[$i][$p-1];
305 $this->V
[$i][$p-1] = -$sn * $this->V
[$i][$j] +
$cs * $this->V
[$i][$p-1];
306 $this->V
[$i][$j] = $t;
311 // Split at negligible s(k).
315 for ($j = $k; $j < $p; ++
$j) {
316 $t = hypo($this->s
[$j], $f);
317 $cs = $this->s
[$j] / $t;
321 $e[$j] = $cs * $e[$j];
323 for ($i = 0; $i < $this->m
; ++
$i) {
324 $t = $cs * $this->U
[$i][$j] +
$sn * $this->U
[$i][$k-1];
325 $this->U
[$i][$k-1] = -$sn * $this->U
[$i][$j] +
$cs * $this->U
[$i][$k-1];
326 $this->U
[$i][$j] = $t;
331 // Perform one qr step.
333 // Calculate the shift.
334 $scale = max(max(max(max(
335 abs($this->s
[$p-1]),abs($this->s
[$p-2])),abs($e[$p-2])),
336 abs($this->s
[$k])), abs($e[$k]));
337 $sp = $this->s
[$p-1] / $scale;
338 $spm1 = $this->s
[$p-2] / $scale;
339 $epm1 = $e[$p-2] / $scale;
340 $sk = $this->s
[$k] / $scale;
341 $ek = $e[$k] / $scale;
342 $b = (($spm1 +
$sp) * ($spm1 - $sp) +
$epm1 * $epm1) / 2.0;
343 $c = ($sp * $epm1) * ($sp * $epm1);
345 if (($b != 0.0) ||
($c != 0.0)) {
346 $shift = sqrt($b * $b +
$c);
350 $shift = $c / ($b +
$shift);
352 $f = ($sk +
$sp) * ($sk - $sp) +
$shift;
355 for ($j = $k; $j < $p-1; ++
$j) {
362 $f = $cs * $this->s
[$j] +
$sn * $e[$j];
363 $e[$j] = $cs * $e[$j] - $sn * $this->s
[$j];
364 $g = $sn * $this->s
[$j+
1];
365 $this->s
[$j+
1] = $cs * $this->s
[$j+
1];
367 for ($i = 0; $i < $this->n
; ++
$i) {
368 $t = $cs * $this->V
[$i][$j] +
$sn * $this->V
[$i][$j+
1];
369 $this->V
[$i][$j+
1] = -$sn * $this->V
[$i][$j] +
$cs * $this->V
[$i][$j+
1];
370 $this->V
[$i][$j] = $t;
377 $f = $cs * $e[$j] +
$sn * $this->s
[$j+
1];
378 $this->s
[$j+
1] = -$sn * $e[$j] +
$cs * $this->s
[$j+
1];
380 $e[$j+
1] = $cs * $e[$j+
1];
381 if ($wantu && ($j < $this->m
- 1)) {
382 for ($i = 0; $i < $this->m
; ++
$i) {
383 $t = $cs * $this->U
[$i][$j] +
$sn * $this->U
[$i][$j+
1];
384 $this->U
[$i][$j+
1] = -$sn * $this->U
[$i][$j] +
$cs * $this->U
[$i][$j+
1];
385 $this->U
[$i][$j] = $t;
394 // Make the singular values positive.
395 if ($this->s
[$k] <= 0.0) {
396 $this->s
[$k] = ($this->s
[$k] < 0.0 ?
-$this->s
[$k] : 0.0);
398 for ($i = 0; $i <= $pp; ++
$i) {
399 $this->V
[$i][$k] = -$this->V
[$i][$k];
403 // Order the singular values.
405 if ($this->s
[$k] >= $this->s
[$k+
1]) {
409 $this->s
[$k] = $this->s
[$k+
1];
411 if ($wantv AND ($k < $this->n
- 1)) {
412 for ($i = 0; $i < $this->n
; ++
$i) {
413 $t = $this->V
[$i][$k+
1];
414 $this->V
[$i][$k+
1] = $this->V
[$i][$k];
415 $this->V
[$i][$k] = $t;
418 if ($wantu AND ($k < $this->m
-1)) {
419 for ($i = 0; $i < $this->m
; ++
$i) {
420 $t = $this->U
[$i][$k+
1];
421 $this->U
[$i][$k+
1] = $this->U
[$i][$k];
422 $this->U
[$i][$k] = $t;
437 * Return the left singular vectors
442 public function getU() {
443 return new Matrix($this->U
, $this->m
, min($this->m +
1, $this->n
));
448 * Return the right singular vectors
453 public function getV() {
454 return new Matrix($this->V
);
459 * Return the one-dimensional array of singular values
462 * @return diagonal of S.
464 public function getSingularValues() {
470 * Return the diagonal matrix of singular values
475 public function getS() {
476 for ($i = 0; $i < $this->n
; ++
$i) {
477 for ($j = 0; $j < $this->n
; ++
$j) {
480 $S[$i][$i] = $this->s
[$i];
482 return new Matrix($S);
492 public function norm2() {
498 * Two norm condition number
501 * @return max(S)/min(S)
503 public function cond() {
504 return $this->s
[0] / $this->s
[min($this->m
, $this->n
) - 1];
509 * Effective numerical matrix rank
512 * @return Number of nonnegligible singular values.
514 public function rank() {
515 $eps = pow(2.0, -52.0);
516 $tol = max($this->m
, $this->n
) * $this->s
[0] * $eps;
518 for ($i = 0; $i < count($this->s
); ++
$i) {
519 if ($this->s
[$i] > $tol) {
526 } // class SingularValueDecomposition