Non-linear Partition of Unity Method (Q6926907)

Radial basis function methods are among the best approximation schemes for multivariate functions that are given at many scattered data. This is possible for any dimension if radial basis functions such as (inverse) multiquadrics or thin-plate splines for example are employed.\N\NThere is a large set of different methods available to approximate those data from spaces spanned by shifts of the radial basis functions convolved with a distance function, usually the Euclidean norm on multivariable real spaces. So the basis functions that span the approximating vector space are formed from a radial basis functions \(\phi\), say, convolved with a distance function, call it \(d(\cdot,\cdot)\), so that we approximate by linear combinations of \(\phi(d(x,y))\). An example is \(d(x,y)=\|x-y\|_2\). These methods include interpolation and the most useful scheme of quasi-interpolation.\N\NIn this article, approximations of radial basis function spaces are combined with a ENO and WENO algorithm that go back in particular to Ami Harten and co-authors. These methods haved been designed for minimising unwelcome oscillations of approximants, especially if the data lack smoothness.\N\NIn this fashion, very useful partition of unity methods are developed in this interesting article that are suitable for general approximation problems. This is because they divide up domains on whose individual parts useful local approximations are applied. Those individually adapted approximations are then summed up via the partition of unity methods. A particular improvement that comes from this paper is that nonlinear partitions of unity are now allowed which seems to be a complete novelty.