The number of components in the feature vector should only be equal to the number of measured values should be chosen as large as the number of training samples could be selected using cross validation should be increased as large as possible until computation time becomes problematic