On a phase transition in general order spline regression

In the Gaussian sequence model in , we study the fundamental limit of approximating the signal by a class of (generalized) splines with free knots. Here is the degree of the spline, is the order of differentiability at each inner knot, and is the maximal number of pieces. We show that, given any integer and , the minimax rate of estimation over exhibits the following phase transition: \begin{equation*} \begin{aligned} \inf_{\widetilde{\theta}}\sup_{\theta\in\Theta(d,d_0, k)}\mathbb{E}_\theta\|\widetilde{\theta} - \theta\|^2 \asymp_d \begin{cases} k\log\log(16n/k), & 2\leq k\leq k_0,\\ k\log(en/k), & k \geq k_0+1. \end{cases} \end{aligned} \end{equation*} The transition boundary , which takes the form , demonstrates the critical role of the regularity parameter in the separation between a faster and a slower rate. We further show that, once encouraging an additional '-monotonicity' shape constraint (including monotonicity for and convexity for ), the above phase transition is eliminated and the faster rate can be achieved for all . These results provide theoretical support for developing -penalized (shape-constrained) spline regression procedures as useful alternatives to - and -penalized ones.
View on arXiv