[Octave-bug-tracker] [bug #59804] statistics-1.4.2 - fitgmdist - dimension mismatch check inappropriately fails for non-shared Sigma

[Tasos Papastylianou]

URL:
  <https://savannah.gnu.org/bugs/?59804>

                 Summary: statistics-1.4.2 - fitgmdist - dimension mismatch
check inappropriately fails for non-shared Sigma
                 Project: GNU Octave
            Submitted by: tpapastylianou
            Submitted on: Sat 02 Jan 2021 05:15:07 PM UTC
                Category: Octave Forge Package
                Severity: 3 - Normal
                Priority: 5 - Normal
              Item Group: Unexpected Error or Warning
                  Status: None
             Assigned to: None
         Originator Name: Tasos Papastylianou
        Originator Email: 
             Open/Closed: Open
                 Release: 6.1.0
         Discussion Lock: Any
        Operating System: Any

    _______________________________________________________

Details:

Please see https://stackoverflow.com/a/65530418/4183191 for context and more
details.
Note, my 'fix' in that SO article is only a hacky workaround, not really a
proper mismatch check in that sense; I assume the maintainer will want to
conduct a more comprehensive validation...

In summary, line 197 performs the following check:


if (any (size (data, 2) ~= [size(mu,2), size(Sigma)]) || any (k ~=
[size(mu,1), size(p,2)]))
  error ('fitgmdist: Start parameter has mismatched dimensions');             
                                     
endif


This line is a bit convoluted, but effectively it says the following:
Fail the mismatch check if ANY of the following is true:
1. There is a mismatch between number of features (columns) 'n' in the data,
and features (columns) in the mean.
2. There is a mismatch between number of features (columns) 'n' in the data,
and ANY of the dimensions of Sigma
3. There is a mismatch between number of components 'k', and number of
components (rows) in the mean
4. There is a mismatch between number of components 'k', and number of
components (columns) in the component weights (i.e. ComponentProportion)

The error in logic lies in step 2. This would only pass the check in the case
of a shared covariance, in which case you would expect Sigma to be a 2D array,
of size n×n, and therefore the check any( n ~= [n,n] ) will be false (i.e. no
evidence of mismatch). However, in the typical case where Sigma is expected to
be n×n×k (one 2D covariance matrix per component considered, stacked along
the 3rd dimension), this check will be flagged as true (i.e. evidence of
mismatch, since n ~= k), and therefore the code will exit with a dimension
mismatch error, instead of accepting an otherwise valid n×n×k Sigma.





    _______________________________________________________

Reply to this item at:

  <https://savannah.gnu.org/bugs/?59804>

_______________________________________________
  Message sent via Savannah
  https://savannah.gnu.org/