‘These types of dimensionality techniques [PCA and ICA methods for finding parameter combinations within a dataset], can be useful for identifying classes of objects, for detecting rare or outlying objects, and for constructing compact representations of the distribution of observed objects. One potential weakness of dimensionality reduction algorithms is that the components are defined statistically, and as such have no guarantee of reflecting true physical aspects of the systems being observed. Because of this, one must be careful when making physical inferences from such results.’
—Introduction to astroML: Machine Learning forAstrophysics, Jacob T. VanderPlas, Andrew J. Connolly, Zeljko Ivezic, Alex Gray, 2014
