Christian Soeller >
PDL-2.4.3 >
PDL::MatrixOps

PDL::MatrixOps -- Some Useful Matrix Operations

$inv = $a->inv; $det = $a->det; ($lu,$perm,$par) = $a->lu_decomp; $x = lu_backsub($lu,$perm,$b); # solve $a x $x = $b

PDL::MatrixOps is PDL's built-in matrix manipulation code. It contains utilities for many common matrix operations: inversion, determinant finding, eigenvalue/vector finding, singular value decomposition, etc. PDL::MatrixOps routines are written in a mixture of Perl and C, so that they are reliably present even when there no FORTRAN compiler or external library available (e.g. PDL::Slatec or PDL::GSL).

Matrix manipulation, particularly with large matrices, is a challenging field and no one algorithm is suitable in all cases. The utilities here use general-purpose algorithms that work acceptably for many cases but might not scale well to very large or pathological (near-singular) matrices.

Except as noted, the matrices are PDLs whose 0th dimension ranges over column and whose 1st dimension ranges over row. The matrices appear correctly when printed.

These routines should work OK with PDL::Matrix objects as well as with normal PDLs.

Like most computer languages, PDL addresses matrices in (column,row) order in most cases; this corresponds to (X,Y) coordinates in the matrix itself, counting rightwards and downwards from the upper left corner. This means that if you print a PDL that contains a matrix, the matrix appears correctly on the screen, but if you index a matrix element, you use the indices in the reverse order that you would in a math textbook. If you prefer your matrices indexed in (row, column) order, you can try using the PDL::Matrix object, which includes an implicit exchange of the first two dimensions but should be compatible with most of these matrix operations. TIMTOWDTI.)

Matrices, row vectors, and column vectors can be multiplied with the 'x' operator (which is, of course, threadable):

$m3 = $m1 x $m2; $col_vec2 = $m1 x $col_vec1; $row_vec2 = $row_vec1 x $m1; $scalar = $row_vec x $col_vec;

Because of the (column,row) addressing order, 1-D PDLs are treated as _row_ vectors; if you want a _column_ vector you must add a dummy dimension:

$rowvec = pdl(1,2); # row vector $colvec = $rowvec->(*1); # 1x2 column vector $matrix = pdl([[3,4],[6,2]]); # 2x2 matrix $rowvec2 = $rowvec x $matrix; # right-multiplication by matrix $colvec = $matrix x $colvec; # left-multiplication by matrix $m2 = $matrix x $rowvec; # Throws an error

Implicit threading works correctly with most matrix operations, but you must be extra careful that you understand the dimensionality. In particular, matrix multiplication and other matrix ops need nx1 PDLs as row vectors and 1xn PDLs as column vectors. In most cases you must explicitly include the trailing 'x1' dimension in order to get the expected results when you thread over multiple row vectors.

When threading over matrices, it's very easy to get confused about which dimension goes where. It is useful to include comments with every expression, explaining what you think each dimension means:

$a = xvals(360)*3.14159/180; # (angle) $rot = cat(cat(cos($a),sin($a)), # rotmat: (col,row,angle) cat(-sin($a),cos($a)));

MatrixOps includes algorithms and pre-existing code from several origins. In particular, `eigens_sym`

is the work of Stephen Moshier, `svd`

uses an SVD subroutine written by Bryant Marks, and `eigens`

uses a subset of the Small Scientific Library by Kenneth Geisshirt. All are free software, distributable under same terms as PDL itself.

This is intended as a general-purpose linear algebra package for small-to-mid sized matrices. The algorithms may not scale well to large matrices (hundreds by hundreds) or to near singular matrices.

If there is something you want that is not here, please add and document it!

Signature: (n; [o]a(n,n))

Return an identity matrix of the specified size. If you hand in a scalar, its value is the size of the identity matrix; if you hand in a dimensioned PDL, the 0th dimension is the size of the matrix.

Signature: (a(n); [o]b(n,n))

$mat = stretcher($eigenvalues);

Return a diagonal matrix with the specified diagonal elements

Signature: (a(m,m); sv opt )

$a1 = inv($a, {$opt});

Invert a square matrix.

You feed in an NxN matrix in $a, and get back its inverse (if it exists). The code is inplace-aware, so you can get back the inverse in $a itself if you want -- though temporary storage is used either way. You can cache the LU decomposition in an output option variable.

`inv`

uses lu_decomp by default; that is a numerically stable (pivoting) LU decomposition method. If you ask it to thread then a numerically unstable (non-pivoting) method is used instead, so avoid threading over collections of large (say, more than 4x4) or near-singular matrices unless precision is not important.

OPTIONS:

- s
Boolean value indicating whether to complain if the matrix is singular. If this is false, singular matrices cause inverse to barf. If it is true, then singular matrices cause inverse to return undef. In the threading case, no checking for singularity is performed, if any of the matrices in your threaded collection are singular, they receive NaN entries.

- lu (I/O)
This value contains a list ref with the LU decomposition, permutation, and parity values for $a. If you do not mention the key, or if the value is undef, then inverse calls lu_decomp. If the key exists with an undef value, then the output of lu_decomp is stashed here (unless the matrix is singular). If the value exists, then it is assumed to hold the lu decomposition.

- det (Output)
If this key exists, then the determinant of

`$a`

get stored here, whether or not the matrix is singular.

Signature: (a(m,m); sv opt)

$det = det($a,{opt});

Determinant of a square matrix using LU decomposition (for large matrices)

You feed in a square matrix, you get back the determinant. Some options exist that allow you to cache the LU decomposition of the matrix (note that the LU decomposition is invalid if the determinant is zero!). The LU decomposition is cacheable, in case you want to re-use it. This method of determinant finding is more rapid than recursive-descent on large matrices, and if you reuse the LU decomposition it's essentially free.

If you ask det to thread (by giving it a 3-D or higher dim piddle) then lu_decomp drops you through to lu_decomp2, which is numerically unstable (and hence not useful for very large matrices) but quite fast.

If you want to use threading on a matrix that's less than, say, 10x10, and might be near singular, then you might want to use determinant, which is a more robust (but slower) determinant finder, instead.

OPTIONS:

- lu (I/O)
Provides a cache for the LU decomposition of the matrix. If you provide the key but leave the value undefined, then the LU decomposition goes in here; if you put an LU decomposition here, it will be used and the matrix will not be decomposed again.

Signature: (a(m,m))

$det = determinant($a);

Determinant of a square matrix, using recursive descent (threadable).

This is the traditional, robust recursive determinant method taught in most linear algebra courses. It scales like `O(n!)`

(and hence is pitifully slow for large matrices) but is very robust because no division is involved (hence no division-by-zero errors for singular matrices). It's also threadable, so you can find the determinants of a large collection of matrices all at once if you want.

Matrices up to 3x3 are handled by direct multiplication; larger matrices are handled by recursive descent to the 3x3 case.

The LU-decomposition method det is faster in isolation for single matrices larger than about 4x4, and is much faster if you end up reusing the LU decomposition of $a, but does not thread well.

Eigenvalues and -vectors of a symmetric square matrix. If passed an asymmetric matrix, the routine will warn and symmetrize it, by taking the average value. That is, it will solve for 0.5*($a+$a->mv(0,1)).

It\'s threadable, so if $a is 3x3x100, it\'s treated as 100 separate 3x3 matrices, and both $ev and $e get extra dimensions accordingly.

If called in scalar context it hands back only the eigenvalues. Ultimately, it should switch to a faster algorithm in this case (as discarding the eigenvectors is wasteful).

The algorithm used is due to J. vonNeumann, which was a rediscovery of Jacobi\'s method. http://en.wikipedia.org/wiki/Jacobi_eigenvalue_algorithm

The eigenvectors are returned in COLUMNS of the returned PDL. That makes it slightly easier to access individual eigenvectors, since the 0th dim of the output PDL runs across the eigenvectors and the 1st dim runs across their components.

($ev,$e) = eigens_sym $a; # Make eigenvector matrix $vector = $ev->($n); # Select nth eigenvector as a column-vector $vector = $ev->(($n)); # Select nth eigenvector as a row-vector

($ev, $e) = eigens_sym($a); # e\'vects & e\'vals $e = eigens_sym($a); # just eigenvalues

',);

###################################################################### ### eigens ### pp_def("eigens", HandleBad => 0, Pars => '[phys]a(m); [o,phys]ev(l,n,n); [o,phys]e(l,n)', GenericTypes => ['D'], Code => ' #include "complex.h" #include "eigen.h"

register int sn = $SIZE(n); int i,j; void **foo; void **bar; New(42, foo, sn, void*); New(111, bar, sn, void*); if($SIZE (l) != 2) { barf("eigens internal error..."); } if($SIZE (m) != (sn * sn )) { fprintf(stderr,"m=%d, sn=%d\n",$SIZE(m),sn); barf("Wrong sized args for eigens"); } for( i=j=0; i<$SIZE(m); i += sn ) { foo[j] = &( ($P(a))[ i ] ); bar[j] = &( ($P(ev))[ i+i ] ); j++; } Eigen( sn, 0, (double **) foo, 20*sn, 1e-13, 0, (SSL_Complex *)( $P(e) ), (SSL_Complex **) bar ); Safefree(foo); Safefree(bar); /* Check for invalid values: convenience block */ { int k; char ok, flag; PDL_Double eps = 0; /* First: find maximum eigenvalue */ for(i=0; i< sn; i++ ) { PDL_Double z = fabs ( ($P(e))[i*2] ); if( z > eps ) eps = z; } eps *= 1e-10; /* Next: scan for non-real terms and parallel vectors */ for( i=0; i < sn; i++ ) { ok = ( fabs( ($P(e))[i*2+1] ) < eps ); for( j=0; ok && j< sn; j++) ok &= fabs( ($P(ev))[ 2* (i*sn + j) + 1] ) < eps; for( k=0; ok && k < i; k++ ) { if( finite( ($P(ev))[ 2 * (k*sn) ] ) ) { for( flag=1, j=0; ok && flag && j< sn; j++) flag &= ( fabs( ($P(ev))[ 2 * (i*sn + j) ] - ($P(ev))[ 2 * (k*sn + j) ] ) < 1e-10 * ( fabs(($P(ev))[ 2 * (k*sn + j) ]) + fabs(($P(ev))[ 2 * (i*sn + j) ]) ) ); ok &= !flag; } } if (ok) { for(j=0; ok && j<sn; j++) { double ex = ($P(e))[i*2] * ($P(ev))[2 * (i*sn + j)]; double ac = 0.0; /*actual matrix product*/ for(k=0; k<sn; k++) ac += ($P(a))[i*sn + k] * ($P(ev))[2 * (i*sn + k)]; ok &= fabs(ac-ex)<eps; } } /* Set column to NaN if necessary */ if( !ok ) { /* TODO: We should set the badval bit if in use. */ for(j=0; j< sn; j++) ($P(ev))[2 * (i*sn+j)] = PDL->NaN_double; ($P(e))[i*2] = PDL->NaN_double; } } } ', PMCode =>' sub PDL::eigens { my ($a) = @_; my (@d) = $a->dims; my $n = $d[0]; barf "Need real square matrix for eigens" if $#d < 1 or $d[0] != $d[1]; my $deviation = PDL::max(abs($a - $a->mv(0,1)))/PDL::max(abs($a)); if ( $deviation <= 1e-5 ) { #taken from eigens_sym code my $lt = PDL::indexND($a, scalar(PDL::whichND(PDL->xvals($n,$n) <= PDL->yvals($n,$n))) )->copy; my $ev = PDL->zeroes($a->dims); my $e = PDL->zeroes($a->index(0)->dims); &PDL::_eigens_sym_int($lt, $ev, $e); return $ev->xchg(0,1), $e if wantarray; return $e; #just eigenvalues } else { barf "eigens currently disabled for asymmatric matrices\n"; my $ev = PDL->zeroes(2, $a->dims); my $e = PDL->zeroes(2, $a->index(0)->dims); &PDL::_eigens_int($a->clump(0,1), $ev, $e); return $ev->index(0)->xchg(0,1)->sever, $e->index(0)->sever if(wantarray); return $e->index(0)->sever; #just eigenvalues } } ' , Doc => ' =for ref

Real eigenvalues and -vectors of a real square matrix.

(See also "eigens_sym", for eigenvalues and -vectors of a real, symmetric, square matrix).

PLEASE NOTE: THE EIGENS FUNCTION HAS BEEN DISABLED FOR ASYMMETRIC MATRICES PENDING FIX TO THE CODE.

The eigens function will attempt to compute the eigenvalues and eigenvectors of a square matrix with real components. If the matrix is symmetric, the same underlying code as "eigens_sym" is used. If asymmetric, the eigenvalues and eigenvectors are computed with algorithms from the sslib library (anyone know what algorithm is used?). If any imaginary components exist in the eigenvalues, the results are currently considered to be invalid, and such eigenvalues are returned as "NaN"s. This is true for eigenvectors also. That is if there are imaginary components to any of the values in the eigenvector, the eigenvalue and corresponding eigenvectors are all set to "NaN". Finally, if there are any repeated eigenvectors, they are replaced with all "NaN"s.

Use of the eigens function on asymmetric matrices should be considered experimental! For asymmetric matrices, nearly all observed matrices with real eigenvalues produce incorrect results, due to errors of the sslib algorithm. If your assymmetric matrix returns all NaNs, do not assume that the values are complex. Also, problems with memory access is known in this library.

Not all square matrices are diagonalizable. If you feed in a non-diagonalizable matrix, then one or more of the eigenvectors will be set to NaN, along with the corresponding eigenvalues.

`eigens`

is threadable, so you can solve 100 eigenproblems by feeding in a 3x3x100 array. Both $ev and $e get extra dimensions accordingly.

If called in scalar context `eigens`

hands back only the eigenvalues. This is somewhat wasteful, as it calculates the eigenvectors anyway.

The eigenvectors are returned in COLUMNS of the returned PDL (ie the the 0 dimension). That makes it slightly easier to access individual eigenvectors, since the 0th dim of the output PDL runs across the eigenvectors and the 1st dim runs across their components.

($ev,$e) = eigens $a; # Make eigenvector matrix $vector = $ev->($n); # Select nth eigenvector as a column-vector $vector = $ev->(($n)); # Select nth eigenvector as a row-vector

DEVEL NOTES:

For now, there is no distinction between a complex eigenvalue and an invalid eigenvalue, although the underlying code generates complex numbers. It might be useful to be able to return complex eigenvalues.

($ev, $e) = eigens($a); # e\'vects & e\'vals $e = eigens($a); # just eigenvalues

',);

###################################################################### ### svd pp_def( "svd", HandleBad => 0, Pars => 'a(n,m); [o]u(n,m); [o,phys]z(n); [o]v(n,n);', GenericTypes => ['D'], Code => ' extern void SVD( double *W, double *Z, int nRow, int nCol ); int sm = $SIZE (m), sn = $SIZE (n), i; double *w, *t, zv; t = w = (double *) malloc(sn*(sm+sn)*sizeof(double)); loop (m) %{ loop(n) %{ *t++ = $a (); %} %} SVD(w, $P (z), sm, sn); t = w; loop (n) %{ zv = sqrt($z ()); $z () = zv; %} loop (m) %{ loop (n) %{ $u () = *t++/$z (); %} %} loop (n) %{ for (i=0;i<sn;i++) { $v (n0=>i, n1=>n) = *t++; } %} free(w); ', , Doc => ' =for usage

($r1, $s, $r2) = svd($a);

Singular value decomposition of a matrix.

`svd`

is threadable.

`$r1`

and `$r2`

are rotation matrices that convert from the original matrix\'s singular coordinates to final coordinates, and from original coordinates to singular coordinates, respectively. `$s`

is the diagonal of the singular value matrix, so that, if `$a`

is square, then you can make an expensive copy of `$a`

by saying:

$ess = zeroes($r1); $ess->diagonal(0,1) .= $s; $a_copy .= $r2 x $ess x $r1;

EXAMPLE

The computing literature has loads of examples of how to use SVD. Here\'s a trivial example (used in PDL::Transform::Map) of how to make a matrix less, er, singular, without changing the orientation of the ellipsoid of transformation:

{ my($r1,$s,$r2) = svd $a; $s++; # fatten all singular values $r2 *= $s; # implicit threading for cheap mult. $a .= $r2 x $r1; # a gets r2 x ess x r1 }

',);

###################################################################### pp_add_exported('','lu_decomp'); pp_addpm(<<'EOD');

Signature: (a(m,m); [o]b(n); [o]c; [o]lu)

LU decompose a matrix, with row permutation

($lu, $perm, $parity) = lu_decomp($a); $lu = lu_decomp($a, $perm, $par); # $perm and $par are outputs! lu_decomp($a->inplace,$perm,$par); # Everything in place.

lu_decomp returns an LU decomposition of a square matrix, using Crout's method with partial pivoting. It's ported from *Numerical Recipes*. The partial pivoting keeps it numerically stable but defeats efficient threading, so if you have a few matrices to decompose accurately, you should use lu_decomp, but if you have a million matrices to decompose and don't mind a higher error budget you probably want to use lu_decomp2, which doesn't do the pivoting (and hence gives wrong answers for near-singular or large matrices), but does do threading.

lu_decomp decomposes the input matrix into matrices L and U such that LU = A, L is a subdiagonal matrix, and U is a superdiagonal matrix. By convention, the diagonal of L is all 1's.

The single output matrix contains all the variable elements of both the L and U matrices, stacked together. Because the method uses pivoting (rearranging the lower part of the matrix for better numerical stability), you have to permute input vectors before applying the L and U matrices. The permutation is returned either in the second argument or, in list context, as the second element of the list. You need the permutation for the output to make any sense, so be sure to get it one way or the other.

LU decomposition is the answer to a lot of matrix questions, including inversion and determinant-finding, and lu_decomp is used by inverse.

If you pass in $perm and $parity, they either must be predeclared PDLs of the correct size ($perm is an n-vector, $parity is a scalar) or scalars.

If the matrix is singular, then the LU decomposition might not be defined; in those cases, lu_decomp silently returns undef. Some singular matrices LU-decompose just fine, and those are handled OK but give a zero determinant (and hence can't be inverted).

lu_decomp uses pivoting, which rearranges the values in the matrix for more numerical stability. This makes it really good for large and even near-singular matrices, but makes it unable to properly vectorize threaded operation. If you have a LOT of small matrices to invert (like, say, a 3x3x1000000 PDL) you should use lu_decomp2, which doesn't pivot and is therefore threadable (and, of course, works in-place).

If you ask lu_decomp to thread (by having a nontrivial third dimension in the matrix) then it will call lu_decomp2 instead. That is a numerically unstable (non-pivoting) method that is mainly useful for smallish, not-so-singular matrices but is threadable.

lu_decomp is ported from _Numerical_Recipes to PDL. It should probably be implemented in C.

Signature: (a(m,m); [0]lu(n)

LU decompose a matrix, with no row permutation (threadable!)

($lu, $perm, $parity) = lu_decomp2($a); $lu = lu_decomp2($a,[$perm,$par]); lu_decomp($a->inplace,[$perm,$par]);

`lu_decomp2`

works just like lu_decomp, but it does no pivoting at all and hence can be usefully threaded. For compatibility with lu_decomp, it will give you a permutation list and a parity scalar if you ask for them -- but they are always trivial.

Because `lu_decomp2`

does not pivot, it is numerically unstable -- that means it is less precise than lu_decomp, particularly for large or near-singular matrices. There are also specific types of non-singular matrices that confuse it (e.g. ([0,-1,0],[1,0,0],[0,0,1]), which is a 90 degree rotation matrix but which confuses lu_decomp2). On the other hand, if you want to invert rapidly a few hundred thousand small matrices and don't mind missing one or two, it's just the ticket.

The output is a single matrix that contains the LU decomposition of $a; you can even do it in-place, thereby destroying $a, if you want. See lu_decomp for more information about LU decomposition.

lu_decomp2 is ported from _Numerical_Recipes_ into PDL. If lu_decomp were implemented in C, then lu_decomp2 might become unnecessary.

Signature: (lu(m,m); perm(m); b(m))

Solve A X = B for matrix A, by back substitution into A's LU decomposition.

($lu,$perm) = lu_decomp($a); $x = lu_backsub($lu,$perm,$par,$b); lu_backsub($lu,$perm,$b->inplace); # modify $b in-place $x = lu_backsub(lu_decomp($a),$b); # (ignores parity value from lu_decomp)

Given the LU decomposition of a square matrix (from lu_decomp), lu_backsub does back substitution into the matrix to solve `A X = B`

for given vector `B`

. It is separated from the lu_decomp method so that you can call the cheap lu_backsub multiple times and not have to do the expensive LU decomposition more than once.

lu_backsub acts on single vectors and threads in the usual way, which means that it treats `$b`

as the *transpose* of the input. If you want to process a matrix, you must hand in the *transpose* of the matrix, and then transpose the output when you get it back. That is because PDLs are indexed by (col,row), and matrices are (row,column) by convention, so a 1-D PDL corresponds to a row vector, not a column vector.

If `$lu`

is dense and you have more than a few points to solve for, it is probably cheaper to find `A^-1`

with inverse, and just multiply `X = A^-1 B`

.) In fact, inverse works by calling lu_backsub with the identity matrix.

lu_backsub is ported from Section 2.3 of *Numerical Recipes*. It is written in PDL but should probably be implemented in C.

Solution of simultaneous linear equations, `a x = b`

.

`$a`

is an `n x n`

matrix (i.e., a vector of length `n*n`

), stored row-wise: that is, `a(i,j) = a[ij]`

, where `ij = i*n + j`

.

While this is the transpose of the normal column-wise storage, this corresponds to normal PDL usage. The contents of matrix a may be altered (but may be required for subsequent calls with flag = -1).

`$b`

, `$x`

, `$ips`

are vectors of length `n`

.

Set `flag=0`

to solve. Set `flag=-1`

to do a new back substitution for different `$b`

vector using the same a matrix previously reduced when `flag=0`

(the `$ips`

vector generated in the previous solution is also required).

See also lu_backsub, which does the same thing with a slightly less opaque interface.

');

###################################################################### ### squaretotri ### # this doesn't need to be changed to support bad values # I could put 'HandleBad => 1', but it would just cause an # unnecessary increase (admittedly small) in the amount of # code # pp_def("squaretotri", Pars => 'a(n,n); b(m)', Code => ' register int mna=0, nb=0, ns = $SIZE (n); #if (PERL_VERSION >= 5) && (PERL_SUBVERSION >= 57) dXSARGS; #endif

if($SIZE (m) != (ns * (ns+1))/2) { barf("Wrong sized args for squaretotri"); } threadloop %{ loop(m) %{ $b () = $a (n0 => mna, n1 => nb); mna++; if(mna > nb) {mna = 0; nb ++;} %} %} ', Doc => ' =for ref

Convert a symmetric square matrix to triangular vector storage.

', );

pp_addpm({At=>'Bot'},<<'EOD');

sub eigen_c { print STDERR "eigen_c is no longer part of PDL::MatrixOps or PDL::Math; use eigens instead.\n";

## my($mat) = @_; ## my $s = $mat->getdim(0); ## my $z = zeroes($s * ($s+1) / 2); ## my $ev = zeroes($s); ## squaretotri($mat,$z); ## my $k = 0 * $mat; ## PDL::eigens($z, $k, $ev); ## return ($ev, $k); }

Copyright (C) 2002 Craig DeForest (deforest@boulder.swri.edu), R.J.R. Williams (rjrw@ast.leeds.ac.uk), Karl Glazebrook (kgb@aaoepp.aao.gov.au). There is no warranty. You are allowed to redistribute and/or modify this work under the same conditions as PDL itself. If this file is separated from the PDL distribution, then the PDL copyright notice should be included in this file.

syntax highlighting: