This is because PCA is a linear algorithm, and so all that the principal components can be is weighted pixel-by-pixel averages of real faces. (Also known as eigenfaces because they’re eigenvectors of the centered covariance matrix of the data—but I digress.)
