On Mon, 6 Jun 2011, Olivier Ga wrote:good
Hello everyone,
Sorry for the long delay. As I explained in my first post, I does not have access directly to the code and I need to ask help by email to have a way to reproduce the problem.
After applying your fix, ie, replace "for (k=0; k<r-m; k++) {" by "for (k=0; k<r-m-1; k++) {", purify does not find anymore a memory corruption.
Clearly, in order to reproduce the problem, we need to have a copy of testarma.c and the txt file you're using. However, the message you're seeing does not necessarily imply a bug: the message refers to the fact that the matrix you're trying to invert is not positive definite, as it should be. This may happen for several reasons, one of them being numerical (as opposed to analytical) computation of the Hessian matrix, but there may be other causes: for example, if you specify an overparametrised model, the Hessian is going to be singular at maximum. In ARMA modelling, these occurrences need to be tackled on a case-by-case basis.
Nevertheless, there is a new problem at the line 377 "*err = gretl_invert_symmetric_matrix(H);" in gretl_bfgs.c
gretl_invert_symmetric_matrix:
dpotrf failed with info = 4 (n = 9)
To reproduce the problem, you can compile testarma.c and run with the following parameter
"a.out EntreesArmasJDD58_precision20.txt 21228"
We try this test on an old linux (not reproduce with 1.9.1), sun Solaris ( reproduce with 1.9.1). It is possible there is a bug inside gretl_invert_symmetric_matrix ?
It is possible the first fix is not correct ?
Is there any requirements on the arma input dataset to work correctly ?
That said: it could well be a bug in the code. ;-)
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Universitą Politecnica delle Marche
r.lucchetti@univpm.it
http://www.econ.univpm.it/lucchetti
_______________________________________________
Gretl-devel mailing list
Gretl-devel@lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel