Distribution of alternation points in best rational approximations (Q2517996): Difference between revisions

Let \(\mathcal P_n\) denote the class of real polynomials in one variable of degree at most \(n\) and \(\mathcal R_{n,m}\) be the class of all rational functions of the form \(r=p/q\) with \(p\in\mathcal P_n\) and \(q\in\mathcal P_m\), \(q\not\equiv0\). If \(E=[\alpha,\beta]\) and \(f\in C(E)\), then for each pair of nonnegative integers \((n,m)\) there exists a unique rational function \(r^*_{n,m}\in\mathcal R_{n,m}\) which is the best rational approximation of \(f\) on \(E\). If \(r^*_{n,m}=p^*_n/q^*_m,\) where \(p^*\) and \(q^*_m\) have no common multipliers, set \(d_{n,m}:=\min\{n-\deg p_n^*,m-\deg q_{m}^*\}\). By a classical Chebyshev theorem there exists a (not necessarily unique) collection of \(m+n+2-d_{n,m}\) strictly increasing points \(x_k^{(n,m)}\) in \([\alpha,\beta]\), called alternation points, such that \[ \lambda_{n,m}(-1)^k(f-r^*_{n,m})(x_k^{(n,m)})=\|f-r^*_{n.m}\|,\quad k=1,\dots,m+n+2-d_{n,m}, \] where \(\lambda^2_{n,m}=1\). Denote by \(\mathcal A_{n,m}=\mathcal A_{n,m}(f)= \{x_k^{(n,m)}\}_{k=1}^{m+n+2-d_{n,m}}\) an arbirarily fixed alternation point set in the approximation \(r^*_{n,m}\) of \(f\). Let \[ \nu_{n,m}([a,b]):=\frac{\#\{x_k^{(n,m)}: a\leq x_k^{(n,m)}\leq b\}}{m+n+2-d_{n.m}} \] be the normalized counting measure of the set \(\mathcal A_{n,m}\). By \textit{P. B. Borwein, A. Kroó, R. Grothmann} and \textit{E. B. Saff} [Proc. Am. Math. Soc. 105, No. 4, 881--888 (1989; Zbl 0688.41018)], if \(m_n\leq m_{n+1}\leq m_n+1\) and \(\lim_{n\to\infty}\frac{m_n}{n}=0\), then there exists a subsequence \(\Lambda=\Lambda(f)\subset\mathbb N\) such that \(\nu_{n,m_n}\to\mu\) (weakly), as \(n\in\Lambda, n\to\infty\), where \(\mu\) is the equilibrium measure of the interval \([\alpha,\beta]\). In the case of best polynomial approximation (\(m_k=0\)), \textit{G. G. Lorentz} [Proc. Am. Math. Soc. 92, 401--403 (1984; Zbl 0566.41013)] gave an example of a function for which the distribution of points from \(\mathcal A_{n,0}\) ``pathologically'' differs from a \(\mu\)-uniform one for an infinite sequence of indices \(n\in\Lambda\subset\mathbb N\). In this paper, developing Lorentz's idea the author proves the following. For any \(\varepsilon>0\) and for any nondecreasing sequence \(\{m_k\}_{k=0}^\infty\) of denominator degrees there exists a function \(f\) which is analytic on the interval \([0,1]\) and a sequence \(\{n_k\}_{k=0}^\infty\) of numerator degrees such that \(\mathcal A_{n_k,m_k}(f)\subset[0,\varepsilon]\) if \(k=2l\) and \(\mathcal A_{n_k,m_k}(f)\subset[1-\varepsilon,1]\) if \(k=2l+1\).