1MatrixOps(3)          User Contributed Perl Documentation         MatrixOps(3)
2
3
4

NAME

6       PDL::MatrixOps -- Some Useful Matrix Operations
7

SYNOPSIS

9          $inv = $a->inv;
10
11          $det = $a->det;
12
13          ($lu,$perm,$par) = $a->lu_decomp;
14          $x = lu_backsub($lu,$perm,$b); # solve $a x $x = $b
15

DESCRIPTION

17       PDL::MatrixOps is PDL's built-in matrix manipulation code.  It contains
18       utilities for many common matrix operations: inversion, determinant
19       finding, eigenvalue/vector finding, singular value decomposition, etc.
20       PDL::MatrixOps routines are written in a mixture of Perl and C, so that
21       they are reliably present even when there is no FORTRAN compiler or
22       external library available (e.g.  PDL::Slatec or PDL::GSL).
23
24       Matrix manipulation, particularly with large matrices, is a challenging
25       field and no one algorithm is suitable in all cases.  The utilities
26       here use general-purpose algorithms that work acceptably for many cases
27       but might not scale well to very large or pathological (near-singular)
28       matrices.
29
30       Except as noted, the matrices are PDLs whose 0th dimension ranges over
31       column and whose 1st dimension ranges over row.  The matrices appear
32       correctly when printed.
33
34       These routines should work OK with PDL::Matrix objects as well as with
35       normal PDLs.
36

TIPS ON MATRIX OPERATIONS

38       Like most computer languages, PDL addresses matrices in (column,row)
39       order in most cases; this corresponds to (X,Y) coordinates in the
40       matrix itself, counting rightwards and downwards from the upper left
41       corner.  This means that if you print a PDL that contains a matrix, the
42       matrix appears correctly on the screen, but if you index a matrix
43       element, you use the indices in the reverse order that you would in a
44       math textbook.  If you prefer your matrices indexed in (row, column)
45       order, you can try using the PDL::Matrix object, which includes an
46       implicit exchange of the first two dimensions but should be compatible
47       with most of these matrix operations.  TIMTOWDTI.)
48
49       Matrices, row vectors, and column vectors can be multiplied with the
50       'x' operator (which is, of course, threadable):
51
52               $m3 = $m1 x $m2;
53               $col_vec2 = $m1 x $col_vec1;
54               $row_vec2 = $row_vec1 x $m1;
55               $scalar = $row_vec x $col_vec;
56
57       Because of the (column,row) addressing order, 1-D PDLs are treated as
58       _row_ vectors; if you want a _column_ vector you must add a dummy
59       dimension:
60
61             $rowvec  = pdl(1,2);            # row vector
62             $colvec  = $rowvec->(*1);       # 1x2 column vector
63             $matrix  = pdl([[3,4],[6,2]]);  # 2x2 matrix
64             $rowvec2 = $rowvec x $matrix;   # right-multiplication by matrix
65             $colvec  = $matrix x $colvec;   # left-multiplication by matrix
66             $m2      = $matrix x $rowvec;   # Throws an error
67
68       Implicit threading works correctly with most matrix operations, but you
69       must be extra careful that you understand the dimensionality.  In
70       particular, matrix multiplication and other matrix ops need nx1 PDLs as
71       row vectors and 1xn PDLs as column vectors.  In most cases you must
72       explicitly include the trailing 'x1' dimension in order to get the
73       expected results when you thread over multiple row vectors.
74
75       When threading over matrices, it's very easy to get confused about
76       which dimension goes where. It is useful to include comments with every
77       expression, explaining what you think each dimension means:
78
79               $a = xvals(360)*3.14159/180;        # (angle)
80               $rot = cat(cat(cos($a),sin($a)),    # rotmat: (col,row,angle)
81                          cat(-sin($a),cos($a)));
82

ACKNOWLEDGEMENTS

84       MatrixOps includes algorithms and pre-existing code from several
85       origins.  In particular, "eigens_sym" is the work of Stephen Moshier,
86       "svd" uses an SVD subroutine written by Bryant Marks, and "eigens" uses
87       a subset of the Small Scientific Library by Kenneth Geisshirt.  They
88       are free software, distributable under same terms as PDL itself.
89

NOTES

91       This is intended as a general-purpose linear algebra package for small-
92       to-mid sized matrices.  The algorithms may not scale well to large
93       matrices (hundreds by hundreds) or to near singular matrices.
94
95       If there is something you want that is not here, please add and
96       document it!
97

FUNCTIONS

99   identity
100        Signature: (n; [o]a(n,n))
101
102       Return an identity matrix of the specified size.  If you hand in a
103       scalar, its value is the size of the identity matrix; if you hand in a
104       dimensioned PDL, the 0th dimension is the size of the matrix.
105
106   stretcher
107         Signature: (a(n); [o]b(n,n))
108
109         $mat = stretcher($eigenvalues);
110
111       Return a diagonal matrix with the specified diagonal elements
112
113   inv
114        Signature: (a(m,m); sv opt )
115
116         $a1 = inv($a, {$opt});
117
118       Invert a square matrix.
119
120       You feed in an NxN matrix in $a, and get back its inverse (if it
121       exists).  The code is inplace-aware, so you can get back the inverse in
122       $a itself if you want -- though temporary storage is used either way.
123       You can cache the LU decomposition in an output option variable.
124
125       "inv" uses lu_decomp by default; that is a numerically stable
126       (pivoting) LU decomposition method.  If you ask it to thread then a
127       numerically unstable (non-pivoting) method is used instead, so avoid
128       threading over collections of large (say, more than 4x4) or near-
129       singular matrices unless precision is not important.
130
131       OPTIONS:
132
133       ·  s
134
135          Boolean value indicating whether to complain if the matrix is
136          singular.  If this is false, singular matrices cause inverse to
137          barf.  If it is true, then singular matrices cause inverse to return
138          undef.  In the threading case, no checking for singularity is
139          performed, if any of the matrices in your threaded collection are
140          singular, they receive NaN entries.
141
142       ·  lu (I/O)
143
144          This value contains a list ref with the LU decomposition,
145          permutation, and parity values for $a.  If you do not mention the
146          key, or if the value is undef, then inverse calls lu_decomp.  If the
147          key exists with an undef value, then the output of lu_decomp is
148          stashed here (unless the matrix is singular).  If the value exists,
149          then it is assumed to hold the lu decomposition.
150
151       ·  det (Output)
152
153          If this key exists, then the determinant of $a get stored here,
154          whether or not the matrix is singular.
155
156   det
157        Signature: (a(m,m); sv opt)
158
159         $det = det($a,{opt});
160
161       Determinant of a square matrix using LU decomposition (for large
162       matrices)
163
164       You feed in a square matrix, you get back the determinant.  Some
165       options exist that allow you to cache the LU decomposition of the
166       matrix (note that the LU decomposition is invalid if the determinant is
167       zero!).  The LU decomposition is cacheable, in case you want to re-use
168       it.  This method of determinant finding is more rapid than recursive-
169       descent on large matrices, and if you reuse the LU decomposition it's
170       essentially free.
171
172       If you ask det to thread (by giving it a 3-D or higher dim piddle) then
173       lu_decomp drops you through to lu_decomp2, which is numerically
174       unstable (and hence not useful for very large matrices) but quite fast.
175
176       If you want to use threading on a matrix that's less than, say, 10x10,
177       and might be near singular, then you might want to use determinant,
178       which is a more robust (but slower) determinant finder, instead.
179
180       OPTIONS:
181
182       ·  lu (I/O)
183
184          Provides a cache for the LU decomposition of the matrix.  If you
185          provide the key but leave the value undefined, then the LU
186          decomposition goes in here; if you put an LU decomposition here, it
187          will be used and the matrix will not be decomposed again.
188
189   determinant
190        Signature: (a(m,m))
191
192         $det = determinant($a);
193
194       Determinant of a square matrix, using recursive descent (threadable).
195
196       This is the traditional, robust recursive determinant method taught in
197       most linear algebra courses.  It scales like "O(n!)" (and hence is
198       pitifully slow for large matrices) but is very robust because no
199       division is involved (hence no division-by-zero errors for singular
200       matrices).  It's also threadable, so you can find the determinants of a
201       large collection of matrices all at once if you want.
202
203       Matrices up to 3x3 are handled by direct multiplication; larger
204       matrices are handled by recursive descent to the 3x3 case.
205
206       The LU-decomposition method det is faster in isolation for single
207       matrices larger than about 4x4, and is much faster if you end up
208       reusing the LU decomposition of $a, but does not thread well.
209
210   eigens_sym
211         Signature: ([phys]a(m); [o,phys]ev(n,n); [o,phys]e(n))
212
213       Eigenvalues and -vectors of a symmetric square matrix.  If passed an
214       asymmetric matrix, the routine will warn and symmetrize it, by taking
215       the average value.  That is, it will solve for 0.5*($a+$a->mv(0,1)).
216
217       It's threadable, so if $a is 3x3x100, it's treated as 100 separate 3x3
218       matrices, and both $ev and $e get extra dimensions accordingly.
219
220       If called in scalar context it hands back only the eigenvalues.
221       Ultimately, it should switch to a faster algorithm in this case (as
222       discarding the eigenvectors is wasteful).
223
224       The algorithm used is due to J. vonNeumann, which was a rediscovery of
225       Jacobi's method.
226       http://en.wikipedia.org/wiki/Jacobi_eigenvalue_algorithm
227
228       The eigenvectors are returned in COLUMNS of the returned PDL.  That
229       makes it slightly easier to access individual eigenvectors, since the
230       0th dim of the output PDL runs across the eigenvectors and the 1st dim
231       runs across their components.
232
233               ($ev,$e) = eigens_sym $a;  # Make eigenvector matrix
234               $vector = $ev->($n);       # Select nth eigenvector as a column-vector
235               $vector = $ev->(($n));     # Select nth eigenvector as a row-vector
236
237        ($ev, $e) = eigens_sym($a); # e'vects & e'vals
238        $e = eigens_sym($a);        # just eigenvalues
239
240       eigens_sym ignores the bad-value flag of the input piddles.  It will
241       set the bad-value flag of all output piddles if the flag is set for any
242       of the input piddles.
243
244   eigens
245         Signature: ([phys]a(m); [o,phys]ev(l,n,n); [o,phys]e(l,n))
246
247       Real eigenvalues and -vectors of a real square matrix.
248
249       (See also "eigens_sym", for eigenvalues and -vectors of a real,
250       symmetric, square matrix).
251
252       The eigens function will attempt to compute the eigenvalues and
253       eigenvectors of a square matrix with real components.  If the matrix is
254       symmetric, the same underlying code as "eigens_sym" is used.  If
255       asymmetric, the eigenvalues and eigenvectors are computed with
256       algorithms from the sslib library.  If any imaginary components exist
257       in the eigenvalues, the results are currently considered to be invalid,
258       and such eigenvalues are returned as "NaN"s.  This is true for
259       eigenvectors also.  That is if there are imaginary components to any of
260       the values in the eigenvector, the eigenvalue and corresponding
261       eigenvectors are all set to "NaN".  Finally, if there are any repeated
262       eigenvectors, they are replaced with all "NaN"s.
263
264       Use of the eigens function on asymmetric matrices should be considered
265       experimental!  For asymmetric matrices, nearly all observed matrices
266       with real eigenvalues produce incorrect results, due to errors of the
267       sslib algorithm.  If your assymmetric matrix returns all NaNs, do not
268       assume that the values are complex.  Also, problems with memory access
269       is known in this library.
270
271       Not all square matrices are diagonalizable.  If you feed in a non-
272       diagonalizable matrix, then one or more of the eigenvectors will be set
273       to NaN, along with the corresponding eigenvalues.
274
275       "eigens" is threadable, so you can solve 100 eigenproblems by feeding
276       in a 3x3x100 array. Both $ev and $e get extra dimensions accordingly.
277
278       If called in scalar context "eigens" hands back only the eigenvalues.
279       This is somewhat wasteful, as it calculates the eigenvectors anyway.
280
281       The eigenvectors are returned in COLUMNS of the returned PDL (ie the
282       the 0 dimension).  That makes it slightly easier to access individual
283       eigenvectors, since the 0th dim of the output PDL runs across the
284       eigenvectors and the 1st dim runs across their components.
285
286               ($ev,$e) = eigens $a;  # Make eigenvector matrix
287               $vector = $ev->($n);   # Select nth eigenvector as a column-vector
288               $vector = $ev->(($n)); # Select nth eigenvector as a row-vector
289
290       DEVEL NOTES:
291
292       For now, there is no distinction between a complex eigenvalue and an
293       invalid eigenvalue, although the underlying code generates complex
294       numbers.  It might be useful to be able to return complex eigenvalues.
295
296        ($ev, $e) = eigens($a); # e'vects & e'vals
297        $e = eigens($a);        # just eigenvalues
298
299       eigens ignores the bad-value flag of the input piddles.  It will set
300       the bad-value flag of all output piddles if the flag is set for any of
301       the input piddles.
302
303   svd
304         Signature: (a(n,m); [o]u(n,m); [o,phys]z(n); [o]v(n,n))
305
306        ($r1, $s, $r2) = svd($a);
307
308       Singular value decomposition of a matrix.
309
310       "svd" is threadable.
311
312       $r1 and $r2 are rotation matrices that convert from the original
313       matrix's singular coordinates to final coordinates, and from original
314       coordinates to singular coordinates, respectively.  $s is the diagonal
315       of the singular value matrix, so that, if $a is square, then you can
316       make an expensive copy of $a by saying:
317
318        $ess = zeroes($r1); $ess->diagonal(0,1) .= $s;
319        $a_copy .= $r2 x $ess x $r1;
320
321       EXAMPLE
322
323       The computing literature has loads of examples of how to use SVD.
324       Here's a trivial example (used in PDL::Transform::Map) of how to make a
325       matrix less, er, singular, without changing the orientation of the
326       ellipsoid of transformation:
327
328        { my($r1,$s,$r2) = svd $a;
329          $s++;             # fatten all singular values
330          $r2 *= $s;        # implicit threading for cheap mult.
331          $a .= $r2 x $r1;  # a gets r2 x ess x r1
332        }
333
334       svd ignores the bad-value flag of the input piddles.  It will set the
335       bad-value flag of all output piddles if the flag is set for any of the
336       input piddles.
337
338   lu_decomp
339         Signature: (a(m,m); [o]b(n); [o]c; [o]lu)
340
341       LU decompose a matrix, with row permutation
342
343         ($lu, $perm, $parity) = lu_decomp($a);
344
345         $lu = lu_decomp($a, $perm, $par);  # $perm and $par are outputs!
346
347         lu_decomp($a->inplace,$perm,$par); # Everything in place.
348
349       lu_decomp returns an LU decomposition of a square matrix, using Crout's
350       method with partial pivoting.  It's ported from Numerical Recipes.  The
351       partial pivoting keeps it numerically stable but defeats efficient
352       threading, so if you have a few matrices to decompose accurately, you
353       should use lu_decomp, but if you have a million matrices to decompose
354       and don't mind a higher error budget you probably want to use
355       lu_decomp2, which doesn't do the pivoting (and hence gives wrong
356       answers for near-singular or large matrices), but does do threading.
357
358       lu_decomp decomposes the input matrix into matrices L and U such that
359       LU = A, L is a subdiagonal matrix, and U is a superdiagonal matrix.  By
360       convention, the diagonal of L is all 1's.
361
362       The single output matrix contains all the variable elements of both the
363       L and U matrices, stacked together.  Because the method uses pivoting
364       (rearranging the lower part of the matrix for better numerical
365       stability), you have to permute input vectors before applying the L and
366       U matrices.  The permutation is returned either in the second argument
367       or, in list context, as the second element of the list.  You need the
368       permutation for the output to make any sense, so be sure to get it one
369       way or the other.
370
371       LU decomposition is the answer to a lot of matrix questions, including
372       inversion and determinant-finding, and lu_decomp is used by inverse.
373
374       If you pass in $perm and $parity, they either must be predeclared PDLs
375       of the correct size ($perm is an n-vector, $parity is a scalar) or
376       scalars.
377
378       If the matrix is singular, then the LU decomposition might not be
379       defined; in those cases, lu_decomp silently returns undef.  Some
380       singular matrices LU-decompose just fine, and those are handled OK but
381       give a zero determinant (and hence can't be inverted).
382
383       lu_decomp uses pivoting, which rearranges the values in the matrix for
384       more numerical stability.  This makes it really good for large and even
385       near-singular matrices, but makes it unable to properly vectorize
386       threaded operation.  If you have a LOT of small matrices to invert
387       (like, say, a 3x3x1000000 PDL) you should use lu_decomp2, which doesn't
388       pivot and is therefore threadable (and, of course, works in-place).
389
390       If you ask lu_decomp to thread (by having a nontrivial third dimension
391       in the matrix) then it will call lu_decomp2 instead.  That is a
392       numerically unstable (non-pivoting) method that is mainly useful for
393       smallish, not-so-singular matrices but is threadable.
394
395       lu_decomp is ported from _Numerical_Recipes to PDL.  It should probably
396       be implemented in C.
397
398   lu_decomp2
399        Signature: (a(m,m); [0]lu(n)
400
401       LU decompose a matrix, with no row permutation (threadable!)
402
403         ($lu, $perm, $parity) = lu_decomp2($a);
404
405         $lu = lu_decomp2($a,[$perm,$par]);
406
407         lu_decomp($a->inplace,[$perm,$par]);
408
409       "lu_decomp2" works just like lu_decomp, but it does no pivoting at all
410       and hence can be usefully threaded.  For compatibility with lu_decomp,
411       it will give you a permutation list and a parity scalar if you ask for
412       them -- but they are always trivial.
413
414       Because "lu_decomp2" does not pivot, it is numerically unstable -- that
415       means it is less precise than lu_decomp, particularly for large or
416       near-singular matrices.  There are also specific types of non-singular
417       matrices that confuse it (e.g. ([0,-1,0],[1,0,0],[0,0,1]), which is a
418       90 degree rotation matrix but which confuses lu_decomp2).  On the other
419       hand, if you want to invert rapidly a few hundred thousand small
420       matrices and don't mind missing one or two, it's just the ticket.
421
422       The output is a single matrix that contains the LU decomposition of $a;
423       you can even do it in-place, thereby destroying $a, if you want.  See
424       lu_decomp for more information about LU decomposition.
425
426       lu_decomp2 is ported from _Numerical_Recipes_ into PDL.  If lu_decomp
427       were implemented in C, then lu_decomp2 might become unnecessary.
428
429   lu_backsub
430        Signature: (lu(m,m); perm(m); b(m))
431
432       Solve A X = B for matrix A, by back substitution into A's LU
433       decomposition.
434
435         ($lu,$perm) = lu_decomp($a);
436         $x = lu_backsub($lu,$perm,$par,$b);
437
438         lu_backsub($lu,$perm,$b->inplace); # modify $b in-place
439
440         $x = lu_backsub(lu_decomp($a),$b); # (ignores parity value from lu_decomp)
441
442       Given the LU decomposition of a square matrix (from lu_decomp),
443       lu_backsub does back substitution into the matrix to solve "A X = B"
444       for given vector "B".  It is separated from the lu_decomp method so
445       that you can call the cheap lu_backsub multiple times and not have to
446       do the expensive LU decomposition more than once.
447
448       lu_backsub acts on single vectors and threads in the usual way, which
449       means that it treats $b as the transpose of the input.  If you want to
450       process a matrix, you must hand in the transpose of the matrix, and
451       then transpose the output when you get it back.  That is because PDLs
452       are indexed by (col,row), and matrices are (row,column) by convention,
453       so a 1-D PDL corresponds to a row vector, not a column vector.
454
455       If $lu is dense and you have more than a few points to solve for, it is
456       probably cheaper to find "A^-1" with inverse, and just multiply "X =
457       A^-1 B".)  In fact, inverse works by calling lu_backsub with the
458       identity matrix.
459
460       lu_backsub is ported from Section 2.3 of Numerical Recipes.  It is
461       written in PDL but should probably be implemented in C.
462
463   simq
464         Signature: ([phys]a(n,n); [phys]b(n); [o,phys]x(n); int [o,phys]ips(n); int flag)
465
466       Solution of simultaneous linear equations, "a x = b".
467
468       $a is an "n x n" matrix (i.e., a vector of length "n*n"), stored row-
469       wise: that is, "a(i,j) = a[ij]", where "ij = i*n + j".
470
471       While this is the transpose of the normal column-wise storage, this
472       corresponds to normal PDL usage.  The contents of matrix a may be
473       altered (but may be required for subsequent calls with flag = -1).
474
475       $b, $x, $ips are vectors of length "n".
476
477       Set "flag=0" to solve.  Set "flag=-1" to do a new back substitution for
478       different $b vector using the same a matrix previously reduced when
479       "flag=0" (the $ips vector generated in the previous solution is also
480       required).
481
482       See also lu_backsub, which does the same thing with a slightly less
483       opaque interface.
484
485       simq ignores the bad-value flag of the input piddles.  It will set the
486       bad-value flag of all output piddles if the flag is set for any of the
487       input piddles.
488
489   squaretotri
490         Signature: (a(n,n); b(m))
491
492       Convert a symmetric square matrix to triangular vector storage.
493
494       squaretotri does not process bad values.  It will set the bad-value
495       flag of all output piddles if the flag is set for any of the input
496       piddles.
497

AUTHOR

499       Copyright (C) 2002 Craig DeForest (deforest@boulder.swri.edu), R.J.R.
500       Williams (rjrw@ast.leeds.ac.uk), Karl Glazebrook
501       (kgb@aaoepp.aao.gov.au).  There is no warranty.  You are allowed to
502       redistribute and/or modify this work under the same conditions as PDL
503       itself.  If this file is separated from the PDL distribution, then the
504       PDL copyright notice should be included in this file.
505
506
507
508perl v5.12.3                      2011-03-31                      MatrixOps(3)
Impressum