Error bounds for kernel-based numerical differentiation (Q2634608)

From MaRDI portal
scientific article
Language Label Description Also known as
English
Error bounds for kernel-based numerical differentiation
scientific article

    Statements

    Error bounds for kernel-based numerical differentiation (English)
    0 references
    0 references
    0 references
    17 February 2016
    0 references
    Given a set \(X\) of \(N\) interpolation points in \(\mathbb{R}^d\), and a symmetric kernel \(K(x,y)\), \(x,y\in\mathbb{R}^d\), kernel-based interpolation for a function \(f\) constructs an interpolant in the points \(X\) that is of the form \(r=\sum_{j=1}^N a_jK(\cdot,x_j)+\sum_{j=1}^M b_j p_j\), with \(x_j\in X\) and \(\{p_j\}_{j=1}^M\) is a basis for all \(d\)-variate polynomials of degree at most \(s\) and such that \(\sum_{j=1}^N a_jp_i(x_j)=0\), \(1\leq i\leq M\). The whole idea is to approximate \(Df\) by \(Dr\) where \(D\) is some differential operator. Preferably it has polynomial consistency, meaning that the derivative should be exact for all polynomials up to a certain degree. This method gives surprisingly good results when applied in meshless methods and this paper investigates why. First the definition, properties and computation of such interpolants and their derivatives are explained. However, the main objective is to get error bounds for the differentiation and to compare them to other polynomially consistent techniques. This analysis is based on a new definition of a growth function that depends on \(X\) and the differential operator \(D\) and gives an upper bound for \(Dp(z)\) subject to \(|p(x_i)|\leq \|z-x_i\|^q_2\), \(x_i\in X\), and \(p\) a polynomial at most of degree \(q\). This growth function appears in the error bound for the kernel-based method and for other polynomially consistent formulas. Some numerical examples illustrate for example the role of the geometry of the set \(X\). The conclusion is that in general kernel based methods are comparable to the best polynomially consistent methods.
    0 references
    multivariate numerical differentiation
    0 references
    differential operator
    0 references
    kernel-based interpolation
    0 references
    polynomial interpolation
    0 references
    growth function
    0 references
    radial basis functions
    0 references
    meshless methods
    0 references
    error bound
    0 references
    numerical examples
    0 references

    Identifiers

    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references