Spectral decomposition of \(H^1 (\mu)\) and Poincaré inequality on a compact interval -- application to kernel quadrature (Q6560735)

Let \(\mathcal{B}\) be the set of probability distributions \(\mu\) whose density \(\rho\) is a positive \(C^1\) function on a finite real interval \([a,b]\). These are bounded perturbations of the uniform distribution \(\nu\) on \([a,b]\). The authors consider the integral \(I_{\mu}[f]:=\int_a^b f(x)d\mu(x)\) and the quadrature formulas (QF) \(I_{\mu_n}[f]:=\sum_{i=1}^n w_i f(x_i)\) where \(\mu_n\) is the discrete distribution with nodes \(X=(x_1,\ldots,x_n)\), \(x_i\in[a,b]\) and (positive) weights \(\mathbf{w}=(w_1,\ldots,w_n)\). In particular, they want to find the optimal QF minimising the worst case error \(\mathrm{wce}(\mu,\mu_n,\mathcal{H})= \sup_{f\in\mathcal{H}, \|f\|_{\mathcal{H}}\le1}|(I_\mu-I_{\mu_n})[f]|\). The integrand \(f\) belongs to some space \(\mathcal{H}\) which can be a Sobolev space \(H^1(\mu)=\{f: f,f'\in L^2(\mu)\}\) which is shown to be a reproducing kernel Hilbert space (RKHS), and the QF are exact on finite dimensional subspaces of \(\mathcal{H}\).\N\NTwo approaches to construct the QF are considered and the links between both are proved, implying their equivalence:\N\begin{itemize}\N\item[(1)] A Chebyshev system (or T-system) \(\{u_k\}_{k\in\mathbb{N}}\) of eigenfunctions of the Sturm-Liouville operator \(L(f)=-(f'\rho)'+f\rho\) with boundary conditions \(f(a)=f(b)=0\) has properties similar to monomials \(x^k\) so that there is a unique \(n\)-point Gaussian-type formula with positive weights, exact in \(\mathrm{span}\{u_k\}_{k=0}^{2n-1}\). This is the (optimal) Poincaré QF in this subspace.\N\item[(2)] Given the reproducing kernel \(K\), and the nodes \(X\), a basis \(K(x_i,\cdot)\) can be used and there exists an expression for the wce, so that the corresponding optimal \(\mathbf{w}\) can be obtained. The optimal kernel QF is the one that minimizes over all \(X\).\N\end{itemize}\NThe expression for the error shows that convergence is proportional to \(n^{-1}\) and that the Poincaré QF is asymptotically optimal in \(\mathcal{H}\). More explicit formulas are obtained when \(\mu\) is the uniform distribution \(\nu\).\N\NNumerical experiments suggest that the weighted Poincaré QF asymptotically behaves like for the uniform distribution with uniformly spaced nodes and \(\mathrm{wce}\sim \frac{1}{\sqrt{12}} n^{-1}\).