i.pca

[Agustin Lobo]

M. Shapiro wrote:

> Agustin, I got the three 5x5 test set and ran the 4.1 version of i.pca on 
> it (both with and without rescaling). I get the following from r.covar -r
> 
> with rescaling 0-255
>     1.000000 -0.005759 0.001019 
>     -0.005759 1.000000 0.000495 
>     0.001019 0.000495 1.000000 
> 
> without rescaling
>     1.000000 -0.002383 0.001550 
>     -0.002383 1.000000 0.002323 
>     0.001550 0.002323 1.000000 
> 
> It seems that i.pca has been modified for 4.1 - we (apparently)
> replaced the eigenvector routines with better ones. Would you say that
> these results (considering integer truncation in the pca output maps)
> are acceptable?


Michael:

Your result without rescaling is the same than using m.eigenvectors. The result
using rescaling shows a higher correlation than if you just use r.rescale (4.0)
I would not say this difference is significant, but we also have integer truncation when using r.rescale. Is the rescaling different in i.pca than in r.rescale?
Note that the critic point is that the rescaling has to be linear to conserve the uncorrelations.
Agus.