In this recent JMLR paper, David Mease and Abraham Wyner experimentally show that the statistical viewpoint to boosting, which is defined as a stagewise optimization problem, does not give a sufficient interpretation for understanding the behavior of boosting algorithms (I guess they just focus on AdaBoost).
Fortunately, the paper is followed by a series of discussions by people like Andreas Buja, Yoav Freund, Robert Schapire, Jerome Friedman, Trevor Hastie, Robert Tibshirani, Peter Bickel, and other great researchers.
I have to find time and read all these papers closely. Boosting is an interesting subject for me (and probably many others). Its “automatic” and performance-dependent feature selection property deserves better understanding.