1: \begin{abstract}

2: An important consideration for variable selection in interaction models

3: is to

4: design an appropriate penalty that respects hierarchy of the importance

5: of the

6: variables. A common theme is to include an interaction term only after the

7: corresponding main effects are present. In this paper, we study several

8: recently proposed approaches and present a unified analysis on the convergence

9: rate for a class of estimators, when the design satisfies the restricted

10: eigenvalue condition. In particular, we show that with probability

11: tending to

12: one, the resulting estimates have a rate of convergence $s\sqrt{\log p_1/n}$

13: in the $\ell_1$ error, where $p_1$ is the ambient dimension, $s$ is

14: the true

15: dimension and $n$ is the sample size. We give a new proof that the restricted

16: eigenvalue condition holds with high probability, when the variables in the

17: main effects and the errors follow sub-Gaussian distributions. Under this

18: setup, the interactions no longer follow Gaussian or sub-Gaussian

19: distributions even if the main effects follow Gaussian, and thus

20: existing works are not applicable. This result is of independent interest.

21: \end{abstract}