Re: [Gretl-devel] Re. gretl too liberal with zero-length matrices
On 12/08/2012 05:26 PM, Allin Cottrell wrote:
Sorry: Error, Fail! This "fix" broke things that are stated
to
work in the manual (e.g. the ability to use a null matrix as
the basis for either row or column concatenation). I think
that's now put to rights in CVS.
Ah, so that's the reason why my script suddenly failed in a new and
unexpected place yesterday! (At least I hope that's the reason.)
Maybe not for right now, but... I tend to think that we should
not allow the construction of matrices like the one in Sven's
example above -- that is, with zero rows but a positive number
of columns or vice versa. You can have a proper null matrix,
or you can have a matrix with rows > 0 and cols > 0, but not a
chimera.
Well, this came up for me when I had the following matrix in a loop with
i the loop index:
zeros((i-1)*n,n)
This produced quite elegant code (if I may say so) without any special
casing. The number of rows could be zero (for i=1) in which case the
matrix would vanish in that loop iteration, but the number of columns
was constant (n>0).
So on the principles that Jack argued with, I would say this should be
possible in gretl if it's also possible in other languages. I'm pretty
certain this is possible in Numerical Python, less certain about other
languages.
As I mentioned, doing "zeros(0,9)" would (silently) produce a
proper null matrix (0x0) until recently. That was not very
satisfactory and for the present I've left in place the "fix"
that produces a 0x9 matrix, but I think that shouldn't stand
for long. Thoughts?
See above. I really think it's only necessary to require one dimension
to match.
thanks,
sven