According to this question and answer, the sum of variances of all partial least squares (PLS) components is normally less than 100%: Why do all the PLS components together explain only a part of the variance of the original data?
Can somebody provide (further) evidence for this? When would this occur?
When running PLS in SAS using the SIMPLS method, I find that 100% of the variance is explained when I use all components.
I have confirmed that the weight vectors produced by PLS in my case are not orthogonal. However, according to The Elements of Statistical Learning the outputed components themselves are orthogonal:
partial least squares produces a sequence of derived, orthogonal inputs or directions $z_1, z_2, \ldots, z_M$.