Linearly constrained reconstruction of functions by kernels with applications to machine learning (Q2498389)

A particular interesting example of reconstruction by kernels is interpolation with positive definite radial basis functions. The radial basis functions interpolant can be interpreted as the function minimizing a Hilbert space norm of all functions which interpolate the given data. The norm is the native space norm corresponding to the basis function (compare with spline interpolation where the norm is the \(L^2\)-norm of the second derivative). In this paper the interpolation condition is replaced by a uniform error bound \(| s(x_j)-f_j| \leq \eta\) at the data sites \(\{x_j\}_{j=1}^N\subset\mathbb{R}^d\) (the corresponding inequality for functions is also considered). The authors prove that there exists a unique minimal norm solution which satisfies this condition. The function has a particular form: Define the contact set \(Y=\{y\in X: | s(x)-f(x)| =\eta\}\), then \(s\in \text{span}\{K(\cdot,y):y\in Y\}\), where \(K(\cdot,y)\) is the reproducing kernel for the Hilbert space. In addition, some conditions for optimality are proved. There is a trade-off between accurate solutions for small \(\eta\) and more simple solutions for higher values of \(\eta\). This motivates the study of a case where \(\eta\) is not constant but included in the optimization problem. The authors also present an iterative greedy algorithm to find the optimal approximating function. It is based on some monotonicity results and quadratic programming techniques. Connections to machine learning and some results from numerical experiments are also presented.