1: \begin{abstract}

2:   Randomized methods are becoming increasingly popular in numerical linear algebra.

3:   However, few attempts have been made to use them in developing preconditioners.

4:   Our interest lies in solving large-scale sparse symmetric positive definite linear systems of equations

5:   where the system matrix is preordered to doubly bordered block diagonal form (for example, using a

6:   nested dissection ordering).

7:   We investigate the use of randomized methods to construct high quality preconditioners.

8:   In particular, we propose a new and efficient approach that employs Nystr\"om's method for computing low rank approximations

9:   to develop robust algebraic two-level preconditioners. Construction of the new 

10:   preconditioners involves iteratively solving a smaller but denser symmetric positive definite

11:   Schur complement system with multiple right-hand sides. Numerical experiments 

12:   on problems coming from a range of application areas demonstrate that this inner system

13:   can be solved cheaply using block conjugate gradients and that using a large convergence tolerance 

14:   to limit the cost does

15:   not adversely affect the quality of the resulting Nystr\"om--Schur two-level preconditioner.

16: \end{abstract}

17: