Sensitivity of the Lanczos recurrence to Gaussian quadrature data: how malignant can small weights be? (Q1044854)

Let \( \mu=\sum_{i=1}^n \omega_i \delta_{\lambda_i} \) be a nonnegative discrete measure on the real line with \(\lambda_1 < \cdots < \lambda_n \) and \( \omega_1,\ldots,\omega_n > 0 \), and in addition let \( Q_k \) be polynomials of degree \(k\) which are orthogonal with respect to the measure \( \mu \) and which have positive leading coefficients for \( 0 \leq k \leq n-1 \). Those polynomials obey the Lanczos recurrence \[ x Q_k(x) = \beta_{k+1} Q_{k+1}(x) + \alpha_{k+1} Q_{k}(x) + \beta_{k} Q_{k-1}(x) \] for \( 0 \leq k \leq n-2\), where \( \beta_k > 0, Q_{-1} = 0, \) and \( Q_0 = 1/\mu(\mathbb{R})^{1/ 2} \). The author consider the impact of small relative perturbations of the numbers \( \lambda_1, \dots, \lambda_n, \omega_1, \dots, \omega_n \) on the coefficients in the Lanczos recurrence. The presented results make use of Hankel determinants and improve some error estimates considered in earlier papers.