On the Convergence of Boosting Procedures

A boosting algorithm seeks to minimize empirically a loss function in a greedy fashion. The resulted estimator takes an additive function form and is built iteratively by applying a base estimator (or learner) to updated samples depending on the previous iterations. This paper studies convergence of boosting when it is carried out over the linear span of a family of basis functions. For general loss functions, we prove the convergence of boosting's greedy optimization to the infinimum of the loss function over the linear span. As a side product, these results reveal the importance of restricting the greedy search step sizes, as known in practice through the works of Friedman and others.