Classical MDS turns a table of distances into a map: square the distances, double-centre them, B = −½H(D•D)H, and use the leading eigenvectors of B scaled by the square roots of the eigenvalues as coordinates. Negative eigenvalues mean the distances are not Euclidean. Nonmetric MDS keeps only the order of the dissimilarities: it alternates a monotone regression (the disparities) with a Guttman transform of the map, and Kruskal's stress can only go down.
Ukraine: great-circle distances (km) between 33 places. Ekman (1954): 14 colours, dissimilarity 1 − similarity. Nonmetric MDS starts from the classical solution; ties in the dissimilarities are broken by the current map distances (Kruskal's primary approach).
code/ch09_heterogeneous_data_a.R (Figures 9.1 to 9.12).