Large data sets are often modeled as being noisy samples from probability distributions in R^D, with D large. It has been noticed that oftentimes the support M of these probability distributions seems to be well-approximated ...