2:55 p.m.
I would like to support Sven's and Jack's views on this. It is really
convenient to work with matrices with zero columns or rows. I agree that
the number of rows should be same same for ~ and columns for |.
I think it would be a good thing if Gretl accepted zero dimension matrices
more generally. For example, letting zeros(5,0)*zeros(0,5) return a 5x5
zero matrix and mols(mnormal(10,2),zeros(10,0),&res) return a zeros(0,2)
matrix a overwrite res with the values from mnormal(10,2). It really works
well in Matlab. It would simplify my code enormously as I could remove many
if/else statements.
Best
Andreas Noack Jensen
10:10 p.m.
On Wed, 19 Dec 2012, Andreas Noack Jensen wrote:
I would like to support Sven's and Jack's views on this.
Their views do differ on some points ;-)
It is really convenient to work with matrices with zero columns or
rows. I agree that the number of rows should be same same for ~
and columns for |.
I think it would be a good thing if Gretl accepted zero dimension matrices
more generally.
Well, I guess Sven argued me out of my opposition to matrices with
one dimension 0 and the other non-zero. And if we're going to
support these, maybe we should support them more fully than at
present.
For example, letting zeros(5,0)*zeros(0,5) return a 5x5
zero matrix
I can see that in principle m1 (5x0) * m2 (0x5) should yield a 5x5
result. I'm not sure that I can see a principled way of determining
what the numerical content of the result should be, since neither of
the operands has any numerical content, but if all-zero is a
convention that Matlabers find useful and expect, we should at least
consider it.
and mols(mnormal(10,2),zeros(10,0),&res) return a zeros(0,2)
matrix a overwrite res with the values from mnormal(10,2). It really works
well in Matlab. It would simplify my code enormously as I could remove many
if/else statements.
Configuring mols() to do this would not be very hard, but of course
the implication is that any arbitrary matrix function should do
something analogous, and that would be a lot of work.
I don't doubt that you have good reason for requesting this, but
could you give a little example of the sort of script that would be
simplified if we went in that direction? Thanks.
Allin Cottrell
2:33 p.m.
When I write code for reduced rank regression in Gretl I need to make
special cases for rank zero and one lag. I think the only possible value
for the entries of matrix(5,0)*matrix(0,5) is zeros and the argument is
reduced rank. The matrix matrix(5,r)*matrix(r,5) generally has rank
min(5,r) and the only zero rank matrix is the zero matrix. For mols, the
acceptance of zero column matrices would ease the concentration step.
Sometimes there is nothing to concentrate out and now I need to check for
this. Of course all this amounts to move checking from my code to Gretl's
code but I really think that zero column/rows matrices make sense in
many occasions. Maybe, just like the zero was a useful addition to the
natural numbers. :-)
--
Andreas Noack Jensen
2:53 p.m.
On Thu, 20 Dec 2012, Andreas Noack Jensen wrote:
When I write code for reduced rank regression in Gretl I need to
make
special cases for rank zero and one lag. I think the only possible value
for the entries of matrix(5,0)*matrix(0,5) is zeros and the argument is
reduced rank. The matrix matrix(5,r)*matrix(r,5) generally has rank
min(5,r) and the only zero rank matrix is the zero matrix.
By my amateurish command of linear algebra, this is a very convincing
argument. If any algebraist is around, contributions are welcome.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
8:41 p.m.
On Thu, 20 Dec 2012, Riccardo (Jack) Lucchetti wrote:
On Thu, 20 Dec 2012, Andreas Noack Jensen wrote:
> When I write code for reduced rank regression in Gretl I need to make
> special cases for rank zero and one lag. I think the only possible value
> for the entries of matrix(5,0)*matrix(0,5) is zeros and the argument is
> reduced rank. The matrix matrix(5,r)*matrix(r,5) generally has rank
> min(5,r) and the only zero rank matrix is the zero matrix.
By my amateurish command of linear algebra, this is a very convincing
argument.
It convinces me too. I've made a small start in CVS:
<hansl>
eval zeros(5,0) * zeros(0,5)
eval zeros(0,3) * zeros(3,0)
eval zeros(0,5)'zeros(0,5)
eval zeros(0,5) * zeros(5,2)
</hansl>
Is the last result OK, do we think?
Allin
9:31 p.m.
On Wed, 19 Dec 2012, Andreas Noack Jensen wrote:
I think it would be a good thing if Gretl accepted zero dimension
matrices
more generally. For example, letting zeros(5,0)*zeros(0,5) return a 5x5
zero matrix and mols(mnormal(10,2),zeros(10,0),&res) return a zeros(0,2)
matrix a overwrite res with the values from mnormal(10,2).
The mols() function now does this in CVS. Example:
<hansl>
Y = mnormal(10,2)
X = zeros(10,0)
U = zeros(10,2)
V = zeros(4,4)
B = mols(Y, X, &U, &V)
print B Y U V
</hansl>
The residuals == Y, and the parameter estimates and their variance
are both null.
Allin Cottrell
4383
days inactive
4385
days old
5 comments
3 participants
participants (3)
-
Allin Cottrell -
Andreas Noack Jensen -
Riccardo (Jack) Lucchetti