Search

This article presents the views of author on a paper by Emanuel Parzen which introduced medians and hinges to reproducing kernel Hilbert spaces. Professor Parzen has attempted to provide a scenario for the analysis of a batch of data which does not depend completely on the traditional parametric probability models such as Gaussian, exponential, etc. Unfortunately, Professor Parzen claims that these techniques replace exploratory data analysis and also provide considerable robustness. These two claims deserve a serious challenge. Much is to be gained by looking at quantile (representing) functions and quantile-density (sparsity) functions, since they often put more attention on the data and the structure of estimators, and less on rigid probability models. The robustness of the autoregressive procedure has not been discussed. Even after a model has been built from the data, it is reasonable to ask that it be robust in a neighborhood of that model. Professor Parzen has not put exploratory data analysis on a firm probability base, for that is logically inconsistent. Indeed, he has given even more reason and opportunity to use both exploratory and confirmatory techniques.