Practical improvement of the divide-and-conquer eigenvalue algorithms (Q1192014)

A practical modification of the well known divide-and-conquer algorithms for approximating the eigenvalues of a real symmetric tridiagonal matrix is presented. In this modification version the authors avoid the numerical stability problems of the algorithm but preserve the insensivity to clustering the eigenvalue and possibility to give upper bounds on the computational cost for any input matrix. Theoretical effectiveness of the algorithms is confirmed by numerical experiments.