Tcl Library Source Code

View Ticket
Login
Ticket UUID: 3035986
Title: determineSVD gives wrong dimensions in U for m<n
Type: Bug Version: None
Submitter: anyeli Created on: 2010-07-28 14:14:30
Subsystem: math :: linearalgebra Assigned To: arjenmarkus
Priority: 7 High Severity: Severe
Status: Pending Last Modified: 2015-04-26 12:29:56
Resolution: None Closed By:
    Closed on:
Description:
Given an m-by-n matrix A, lassign "[::math::linearalgebra::determineSVD $A] U S V" will return an m-by-n matrix U.  While this is often what you want for overdetermined systems (m>n), this is not what you want for underdetermined systems (m<n) and is not consistent with any definition of the SVD I know of.  If (m<n), the S and V matrices should be truncated if truncation is desired.

Incidentally, sometimes I don't want the U matrix to be truncated when m>n, for example when I am as interested in the residuals than the actual fit.
User Comments: arjenmarkus added on 2015-04-26 12:29:56:
I have been trying to fully understand the issue, but the description has one flaw at least: the matrix U as returned by the procedure is actually square (mxm). This is in accordance with the definition in http://mathworld.wolfram.com/SingularValueDecomposition.html, but this page also mentions a different convention.

The S matrix is in fact returned as a vector, whose first m values are meaningful, and the remaining (if any) are essentially zero.

The V matrix is square too, nxn.

Here is a small script to verify this behaviour:

set A {{1 1 1 1}
       {1 1 1 2}}
lassign [::math::linearalgebra::determineSVD $A] U S V

puts U:
puts [join $U \n]
puts S:
puts $S
puts V:
puts [join $V]

I have put this ticket on "pending".

aku added on 2014-12-02 20:22:34:
bump

aku added on 2014-07-21 18:38:18:
Any progress here?