[Numpy-discussion] Change of SVD output
Arnar Flatberg
arnar.flatberg@gmail....
Fri Jul 25 17:01:36 CDT 2008
Hi
In a recent thread there was an error in how a matrix is reconstructed from
its SVD decomposition. I apologize if this is just an old and settled issue
and I am just adding noise, but I got bitten by numpy's unfamiliar output
myself a long time ago and I see others get confused as well. So what is the
issue?
Let the svd decomposition of X be USV:
U, S, V = np.linalg.svd(X),
Then U has X's left singular vectors in its columns, S contains the singular
values, and V has X's right singular vectors in its *rows*. The
reconstruction of X is (matrix notation) will be: U*diag(S)*V (not
U*diag(S)*V.T).
All other high-level software (Matlab, Scilab, Mathematica, R, etc), outputs
the right singular vectors columnwise, that is V = [v1, v2, v3, ... vn],
where vn is a column (eigenvector) in V, thus the reconstruction would be
U*diag(S)*V.T. Also, as far as I know, most linear algebra textbooks operate
with eigenvectors consistent as column vectors in explanations of the SVD. I
think numpy's svd should do so too.
I know lapack's dgesdd returns V.T (or conjugate), and specify that in its
documentation so this is a true interface of the library, but I still think
its wrong and its just too confusing for any beginner who usually has
experience in other software, such as matlab, prior to numpy. Also, ,as I am
typing here, I realize that changing the output would break lots of stuff,
and pass silently through many tests as the shape of V is similar (if
full_matrices=0). Oh well, I guess that proposal is off the table? Perhaps
some *stronger* hints in the documentation are needed.
Arnar
PS:
In the docs at http://www.scipy.org/NumPy_for_Matlab_Users , the svd
equivalents have wrong notation, this is not helping :-). I didnt manage to
change it, perhaps some other may be so kind?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.scipy.org/pipermail/numpy-discussion/attachments/20080726/b3b6bed9/attachment.html
More information about the Numpy-discussion
mailing list