Re: [OctDev] regress and rcoplot (partially written, please help)

[William P. Y. H. <wil...@gm...>]

On 1/2/06, Paul Kienzle <pki...@us...> wrote:
>
> On Jan 1, 2006, at 4:20 AM, William Poetra Yoga Hadisoeseno wrote:
>
> > OK, today I finally finished the function (I think). I used Paul's
> > notes at the end of wsolve.m. Take a look at the attachment ;)
>
> You are not using the economy QR decomposition.  This can lead to
> very large matrices for over determined systems.
>
> Also, you should generally avoid computing the full inverse as
> part of solving a system.  It is more expensive and I believe
> less stable than working directly off the QR decomposition that
> you have already computed.
>
> Any reason you did not use my wsolve code as it stands?
>
> Similarly, do you have a reason for not using the code I sent
> for computing confidence intervals, which also avoids direct
> computation of the inverse?
>
> If there is some reason to prefer your code I would like
> to know what it is.
>

No reason, it's just that I didn't study my linear algebra course well
(never went to class, for example :( ), and I haven't retaken it yet,
so that part of Mathematics is my "blurry spot" (not quite a blind
spot...). As such, I don't really understand your code well; for my
dataset, my code works, so for the time being I haven't used your code
yet.

I'm having my exams now, so I wouldn't have too much time for that.
But I'll look at it again (so don't include it first). Alternatively,
if you would like to modify the function, I don't mind :) (although it
wouldn't be very educative for me).

--
William Poetra Yoga Hadisoeseno