Many problems in statistical pattern recognition begin with the preprocessing of multidimensional signals, such as images of faces or spectrograms of speech.