WEIGHTED ANALYSIS OF PRINCIPAL COMPONENTS: TWO APPROXIMATIONS TO STATISTICAL WEIGHTS

Two approximate methods for weighted principal components analysis (WPCA) were devised and testedin numerical experiments using either empirical variances (obtained from replicated data) or assumedvariances (derived from unreplicated data). In the first ('spherical') approximation each data vector wasassigned a weight proportional to the geometrical mean of its variances in all dimensions. Thearithmetical mean of variances was used instead in the other approximation. Both the numericalexperiments with artificial data containing random errors of various kinds (constant, proportional,constant plus proportional, Poisson) and the analysis of two sets of Raman spectra clearly indicated thenecessity of introducing statistical weights. The spherical approximation was found to be slightly betterthan the arithmetical one. The application of statistical weighting was found to improve the performanceof PCA in estimation problems.