Am 12.12.2025 um 15:33 schrieb Riccardo (Jack) Lucchetti:

On 12/12/2025 15:03, Sven Schreiber wrote:

I guess that this help predates the introduction of the mreverse() function?
Yes
So, OK to change the help example there? Or am I missing something?

Let's change it, yes. Could you please see to it?

OK

I have verified that the following computations coincide: [...]

That's not necessarily true, because the eigenvectors (and hence the principal components) are identified up to a sign switch.

Yes, I know that it depends on the implementation / normalization, but that's exactly the point of checking concrete code snippets: in the one I gave it seems to be identical. Of course, it could be that gretl doesn't guarantee the signs there in the future.

Moreover, there's an even neater way:

<hansl>
clear
A = mnormal(30,10)

# SVD-based
matrix ps = empty 
l = svd(cdemean(A), &ps)
ps = ps .* l
</hansl>

Neater perhaps, but in fact the following variant seems to be faster:

<hansl>
m = cdemean(A)
v = {}
svd(m'm, &v)
ps2 = m * v # up to sign
</hansl>

 
I think it would be helpful to include an example like this one in the princomp help. Or maybe a variant based on the correlation matrix (default in princomp) instead of the covariance like here. OK?

Fine by me.

Alright. Thanks,

sven