Abstract
The multivariate statistical analysis of chemical data in petrology and geochemistry entails principles that are little known among geological practitioners. Such data are compositional, and hence, lie in a sub-space of full space, the simplex. The concept of product-moment correlation of full space cannot be validly applied to them, because of the subcompositional incoherency of that statistic in the simplex. A second type of problem concerns the stability of multivariate estimates. Appropriate techniques, such as principal component cross-validation, are discussed exemplified and contrasted with inappropriate procedures using geochemical data on Swedish Proterozoic dyke swarms. Essential features of compositional covariance structure are exemplified in an Appendix.