On projection pursuit measures of multivariate location and dispersion (Q794103)

Projection pursuit (PP) techniques are directed to the study of high- dimensional data, starting from the idea of selecting ''interesting'' low dimensional projections by iteratively maximizing an appropriate index, usually with the aid of computer. One use of PP is to construct and estimate measures of location and dispersion for multivariate distributions. The following results are proved for locations: (1) A PP measure is equivariant iff the univariate measure from which it is built is linear; (2) Equivariance holds for all distributions only when the univariate measure is expectation; (3) For a fixed multivariate distribution, all PP measures will be equivariant only when the distribution is centrosymmetric: X-\(\mu =^{d}-(X-\mu)\). For dispersion, the corresponding results are phrased with quadratic, standard deviation, and elliptically contoured distributions, respectively. These theorems pertain to functionals on probability distributions, and the corresponding estimators are constructed by evaluating these functionals at the empirical distribution. The paper contains descriptions and discussions of these results, and references to previous findings in the literature (Sections 1 and 2), comments, discussions, and implications for data analysis (Section 3), and proofs (Sections 4 and 5).