[Gretl-users] Re: A question on Kalman Filter/MLE
On Mon, 20 Apr 2020, Alecos Papadopoulos wrote:
Because the Kalman Filter is an occasion where large numbers of
matrices get
together to celebrate their existence, can somebody please clarify the
following for me:
Given starting values for the parameters, and all the rest, the Kalman filter
runs, and it gives the first log-likelihood value series. Then the MLE takes
over, but it immediately finds that
/"matrix is not positive definite"./
*The question is: which matrix is the message referring to?*
[...]
Your gradients are all zero. This means that your parameters are not
entering the likelihood at all, for some reason, and therefore the
likelihood function is flat everywhere, so it's probably the curvature
matrix.
That said, we could be a little bit more explicit in the error message.
-------------------------------------------------------
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
-------------------------------------------------------