Eleven vowel sounds, each described by ten acoustic features: 528 training and 462 test utterances. After sphering with the pooled within-class covariance W, LDA classifies to the nearest class centroid, and the eleven centroids span all ten dimensions. Fisher's discriminant (canonical) coordinates order the directions by how far apart the centroids are relative to the within-class spread: the eigenvectors of W−1/2BW−1/2. Keeping only the first L coordinates and classifying to the nearest centroid there is reduced-rank LDA; on these data L = 2 has the smallest test error.
Computed in the browser from the book's data: class means, W with divisor N − K, B* the covariance of the sphered centroids (weights Nk/N, all equal here), coordinates Zl = vlT(x − x̄) with vl = W−1/2v*l, so each coordinate has unit within-class variance. Signs: the first class (hid) has negative coordinates. Decision regions are drawn only when the classifier uses no coordinate outside the plane shown.
code/ch04_linear_classification.py (reduced_rank_lda, rr_transform, rr_lda_predict, Figures 4.4, 4.8, 4.10, 4.11).