Additive models with trend filtering (Q2284363)

Consider additive regression models \(Y^i = \mu + \sum_{j=1}^d f_{0j}(X_j^i) + \epsilon^i, i=1,\dots,n\), with the input, output variables \(X_j^i,Y^i\), the i.i.d. mean zero error terms \(\epsilon^i\), and the univariate functions \(f_{0j}(\cdot)\). Here, instead of the known linear smoothers, wavelets, etc., additive trend filters are considered by using appropriate estimation/optimization methods. Basic properties of the additive trend filtering model are examined. Moreover, error bounds for additive trend filtering are given. Several empirical experiments and comparisons are presented.