Adaptive piecewise polynomial estimation via trend filtering (Q2448732): Difference between revisions

There are many estimation methods for nonparametric regression; some well-known examples include methods based on local polynomials, kernels, splines, sieves and wavelets. This paper focuses on a relative newcomer in nonparametric regression: the trend filtering method proposed by \textit{S.-J. Kim} et al. [SIAM Rev. 51, No. 2, 339--360 (2009; Zbl 1171.37033)]. The trend filtering estimate is defined as the minimizer of a penalized least squares criterion, in which the penalty term sums the absolute \(k\)-th order discrete derivatives over the input points. It is proved that the trend filtering estimate converges to the true underlying function at the minimax rate for functions whose \(k\)-th derivative is of bounded variation. It is shown that the trend filtering method is both fast and locally adaptive. Convergence rates for a true function with growing total variation are studied. Trend filtering estimates are compared to other methods. It is shown in particular that the trend filtering estimates and locally adaptive regression spline estimates converge to each other asymptotically. It is concluded that trend filtering and locally adaptive regression splines are exactly the same for constant or linear orders, but not for quadratic or higher orders. The trend filtering estimates are much easier to compute, due to the bandedness of the discrete derivative operators. The computations for locally adaptive regression splines, meanwhile, cannot exploit such sparsity or structure, and are considerably slower. Empirical examples show that trend filtering estimates achieve a significantly higher degree of local adaptivity than smoothing splines, which are arguably the standard tool for adaptive spline estimation. Other examples show that trend filtering estimates display a comparable level of adaptivity to locally adaptive regression splines, another well-known technique for adaptive spline estimation, proposed by \textit{E. Mammen} and \textit{S. van de Geer} [Ann. Stat. 25, No. 1, 387--413 (1997; Zbl 0871.62040)] on the basis of being more locally adaptive. An astrophysics data example is considered and the performance of several commonly used nonparametric regression tools is compared.