1: \begin{abstract}

2: MARS is a popular method for nonparametric regression introduced by

3: Friedman in 1991. MARS fits simple nonlinear and non-additive

4: functions to regression data. We propose and study a natural LASSO

5: variant of the MARS method. Our method is based on least squares

6: estimation over a convex class of functions obtained by considering

7: infinite-dimensional linear combinations of functions in the MARS

8: basis and imposing a variation based complexity constraint. We show

9: that our estimator can be computed via finite-dimensional convex

10: optimization and that it is naturally connected to nonparametric

11: function estimation techniques based on smoothness constraints. Under a

12: simple design assumption, we prove that our estimator achieves a rate

13: of convergence that depends only logarithmically on dimension and thus

14: avoids the usual curse of dimensionality to some extent. We implement

15: our method with a cross-validation scheme for the selection of the involved

16: tuning parameter and show that it has favorable performance

17: compared to the usual MARS method in simulation and real data

18: settings.

19: \end{abstract}