Perhaps I need to expand on the implications of what I said in my
message.

Because the Burnaby method removes dimensions from a matrix, the
resulting matrix is supposed to be singular. That means that software
that does not use some form of generalized inverse will complain about
the covariance matrix being singular. In NTSYSpc I allow the user to
specify a cutoff criterion. It should be small enough so that real
dimensions are not discarded. Usually something like 1.0e-12 works well
but that will depend on the variances of the variables you are using.
Because of rounding error the last eigenvalue of the within covariance
matrix will cannot be expected to be exactly zero. 

An alternative strategy is to project the data onto its principal axes
and discard axes corresponding to those with eigenvalues that are
essentially zero. There should be as many of them as dimensions that
were removed using the Burnaby method.

Note that the Burnaby approach does not involve dividing by the
geometric mean or adding a constant to the data. 

--------------------
F. James Rohlf - Dept. Ecology & Evolution
SUNY, Stony Brook, NY 11794-5245
==
Replies will be sent to list.
For more information see http://life.bio.sunysb.edu/morph/morphmet.html.

Reply via email to