REJOINDER TO "LEAST ANGLE REGRESSION" BY EFRON ET AL.

The original goal of this project was to explain the striking similarities between models produced by the Lasso and Forward Stagewise algorithms, as exemplified by Figure 1. LARS, the Least Angle Regression algorithm, provided the explanation and proved attractive in its own right, its simple structure permitting theoretical insight into all three methods. In what follows “LAR” will refer to the basic, unmodified form of Least Angle Regression developed in Section 2, while “LARS” is the more general version giving LAR, Lasso, Forward Stagewise and other variants as in Section 3.4. Here is a summary of the principal properties developed in the paper: