where E[•] is the expectation operation with respect to x
and k•k2 is the norm squared value. We know that PCA
transformation u is of size d · h and it is used to do dimensionality
reduction from d-dimensional space to h-dimensional
feature space, i.e. u:x ! y or y = uTx. It should be
noted that in this transformation, x is assumed to be zero
mean, if mean is not zero then y can be represented as