Abstract: Motivated by a real life example from neuroscience, the authors present a theoretical frame for feature selection in discriminant analysis of very high-dimensional data. In light of a theorem, the authors provide a modification to a procedure, which is commonly-employed, of discriminant analysis of very
high-dimensional data. The modified procedure works are better thantwo other popular procedures in this example in that it needsfewer features and the classification error is smaller

Key words: Discrete wavelet transformation, Discriminant analysis, High-dimensional data, Optimal subsets of features, Principal component analysis