[Gretl-devel] was: Re: [Gretl-users] Re: Fwd: Norm of gradient in mle command
Am 27.10.2023 um 16:08 schrieb Cottrell, Allin:
On Fri, Oct 27, 2023 at 2:48 AM Riccardo (Jack) Lucchetti
<p002264(a)staff.univpm.it> wrote:
> Call b the parameter vector and g the gradient vector, both with n
> elements. Then the norm is computed as
>
> gradnorm = sqrt(abs(b'g)/n)
>
> I don't remember where this formula came from, Allin and I worked on
> this piece of code long ago.
The gradient norm code was added to gretl_bfgs.c on 2010-01-26. I
presume there must have been some email discussion at the time but I'm
afraid I can't find it.
[moving my reply to the devel list, without implying that other replies
need to do the same]
I'll exploit this trigger to say that this is an example why I've
repeatedly said that the gretl source could always use more comments.
At the same time, I'm happy to admit that I think the situation has
gotten better since 2010.
As for the formula --and without having looked at the code or anything
else-- isn't it a bit surprising that the parameter and its gradient are
multiplied? I mean, wouldn't you expect a kind of normalized or scaled
gradient, perhaps something like abs(g'(1./b)) or so? But again, maybe
I'm totally missing the context and this is a stupid remark.
cheers
sven
Attachments:
- attachment.htm (text/html — 2.2 KB)