Ok, I'm not particularly familiar with this material, so if the question sounds funny, that's why :) These PCA methods choose some number of components, n, for use in reconstruction. How is this number generally chosen? In the paper, they present the results for varying n and the choice of n obviously has an impact on the performance. Given the originaly images, it's fairly easy to see which n perform better, but is this a hard problem in practice? Does the performance always increase as n increases?