Convergence rates of a family of barycentric rational Hermite interpolants and their derivatives (Q2029640)

The authors study a family of the so-called \textit{barycentric rational Hermite interpolants} given by \[ r_m(x)=\frac{\sum_{i=9}^{n-d}\,\lambda_{m,i}(x)p_{m,i}(x)}{\sum_{i=0}^{n-d}\,\lambda_{m,i}(x)}, \] with \[ \lambda_{m,i}(x)=\frac{(-1)^{(m+1)i}}{(x-x_i)^{m+1}\cdots (x-x_{i+d})^{m+1}}, \] where the \(p_{m,i}(x)\) denote the unique Hermite polynomials of degree at most \((m+1)(d+1)-1\) that interpolate \(f_i^{(k)}\) (the functions \(f_i\) and their derivatives: \(k=0,1,\dots,m\)) at nodes \(\{x_i,\dots,x_{i+d}\}\). These nodes constitute a subset of a given set of \((n+1)\) interpolation points \[ a=x_0<x_1<\cdots<x_n=b \tag{1} \] and \(f\in C^m[a,b]\) is a real-valued function. In the paper the following notation is used: \begin{itemize} \item[--] \(h=\max_{0\leq i\leq n-1}(x_{i+1}-x_i\)), \item[--] \(C\) is a constant depending only on \(d\), the derivative of \(f\) and the interval \([a,b]\), and only in equidistant points \(h/\delta\) with \(\delta=\min_{0\leq i\leq n-1}(x_{i+1}-x_i)\). \end{itemize} The layout of the paper is as follows: {\S1. Introduction:} Gives a historical overview, indicating the search for a favorable performance of the interpolants (the Lebesgue constants \(\Lambda_n\) increasing only exponential in \(d\)). {\S2. Absence of poles.} Here it is proved: {Theorem 1.} For all integers \(d\), \(0\leq d\leq n\), the functions \(r_m\) have no poles in \(\mathbb{R}\). They interpolate \(f^{(k)}_i\), \(k=0,1,\dots, m\) at the points (1). {\S3. Approximation error.} Let \[ f(x)-r_m(x)=\frac{A_m(x)}{B_m(x)}\hbox{ with }B_m(x)=\sum_{i=0}^{n-d}\lambda_{m,i}(x). \] This section gives the main results in several propositions and theorems. Three of these are given below. {Proposition 1.} If \(f\in C^{(m+1)(d+1)+k}[a,b]\), then \(|A_m^{(k)}(x)|\leq C/h\) for \(x\in [a,b]\). {Theorem 2.} Suppose \(d\), \(0\leq d\leq n\) and \(f\in C^{(m+1)(d+1)}[a,b]\). If \(m\) is odd and the points (1) are quasi-equidistant, then \[ \| f-r_m\| = \|e_m\|\leq Ch^{(m+1)(d+1)-1}. \] {Theorem 3.} Suppose \(d\), \(1\leq d\leq n\) and \(f\in C^{(m+1)(d+1)+1}[a,b]\). If \(m\) is even, then \[ \| f-r_m\| = \|e_m\|\leq Ch^{(m+1)(d+1)}, \] and if \(f\in C^{(m+1)(d+1)}[a,b]\) and the points (1) are quasi-equidistant, then \[ \| f-r_m\| = \|e_m\|\leq Ch^{(m+1)(d+1)-1}. \] {\S4. The barycentric form.} This section discusses more efficient formulae for the evaluation of the rational Hermite interpolants. {\S5. Numerical examples.} Several numerical examples are given with tables for the values and graphs of the functions and their interpolants. The functions considered are: \begin{itemize} \item[1.] \((1+\tanh{(-9x+1)})/2\):\quad for \(m=2\), \(d=1, 2\), \(x_i=i/n\), \([a,b]=[0,1]\). \item[2.] \((3x-1)+\frac{1}{2}(3x-1)-(3x-1)^2\):\quad for \(m=2\), \(d=4\), \(x_i=i/n\), \([a,b]=[0,1]\). \item[3.] \(1/(x+1)^2\):\quad for \(m=1\), \(d=2\), \(x_i=-5+10i/n\), \([a,b]=[-5,5]\). \end{itemize} {\S6. Conclusions.} The authors point out that ony the derivatives of orders \(k=0,1,2\) were studied. They intend to investigate the approximation of \(f^{(k)}\) by \(r_m^{(k)}\) for \(k\geq 3\).