Re: [Gretl-users] psdroot() questions
On Thu, 27 Jul 2017, Sven Schreiber wrote:
Hi,
I'm using the psdroot() (generalized Choleski) for the first time
"seriously"
and have some questions.
- In one case there is an eigenvalue -1.2906e-006 in the input matrix, but
psdroot still goes to work. So I guess there is no check on whether the
matrix is actually PSD?
- In another case I have an eigenvalue 3.6950e-016 (so zero apart from
machine precision). There psdroot() produces a 'nan' output for one element
and gretl spits out a warning. But since psdroot does exist exactly for the
case of zero eigenvalues (otherwise one could use cholesky), why does that
happen? Any hints/ideas?
Testing for positive semi-definiteness is not trivial. But it sounds
like we should take another look at the psdroot code: it appears to
be over-simple and I'm experimenting with lapack's dpstrf.
Allin