Adaptive risk bounds in univariate total variation denoising and trend filtering (Q2176616)

A nonparametric regression problem is considered \[Y_i=f(i/n)+\xi_i, \quad i=1,\ldots, n.\] Here \(f:[0, 1]\to\mathbb{R}\) is the unknown regression function, and \(\xi_1,\ldots,\xi_n\) are unobserved i.i.d. as \(N(0,\sigma^2).\) The goal is to recover \(f\) from the measurements \(Y_1,\ldots, Y_n.\) For a given integer \(r\ge 1,\) the \(r\)th-order trend filtering estimator (TFE) is defined as the minimizer of the sum of squared errors when we constrain (or penalize) the sum of absolute \(r\)th-order discrete derivatives of the fitted function at the design points. For \(r=1,\) the penalized estimator reduces to total variation regularization which was first proposed in [\textit{L. I. Rudin} et al., Physica D 60, No. 1--4, 259--268 (1992; Zbl 0780.49028)] and, for general \(r\ge 1,\) the penalized estimator was first proposed in [\textit{G. Steidl} et al., ``Splines in higher order TV regularization'', Int. J. Comput. Vis. 70, 241--255 (2006; \url{doi:10.1007/s11263-006-8066-7})]. The paper under review studies the TFE for each \(r\ge 1,\) both in the constrained and penalized form. The main results show that in the strong sparsity setting when the underlying function is a polynomial spline with few knots, the risk (under the global squared error loss) of the TFE achieves the parametric \(n^{-1}\)-rate, up to a logarithmic factor. Thus, the results provide support for the use of TFE in the strong sparsity setting.