1: \begin{abstract}

2: We propose a new approach to computing global minimizers

3: of singular value functions in two real variables.

4: Specifically, we present new algorithms to compute the Kreiss constant of a matrix

5: and the distance to uncontrollability of a linear control system, both to arbitrary accuracy.

6: Previous state-of-the-art methods for these two quantities rely on 2D level-set tests 

7: that are based on solving large eigenvalue problems.

8: Consequently, these methods are costly, i.e., $\bigO(n^6)$ work using dense eigensolvers,

9: and often multiple tests are needed before convergence.

10: Divide-and-conquer techniques have been proposed that reduce the work complexity to

11: $\bigO(n^4)$ on average and $\bigO(n^5)$ in the worst case, but

12: these variants are nevertheless still very expensive and can be numerically unreliable.

13: In contrast, our new interpolation-based globality certificates 

14: perform level-set tests by building interpolant approximations to certain one-variable continuous functions

15: that are both relatively cheap and numerically robust to evaluate.

16: Our new approach has a $\bigO(kn^3)$ work complexity and uses $\bigO(n^2)$ memory, 

17: where $k$ is the number of function evaluations necessary to build the interpolants.

18: Not only is this interpolation process mostly ``embarrassingly parallel," 

19: but also low-fidelity approximations typically suffice for all but the final interpolant, which must be built to high accuracy.  

20: Even without taking advantage of the aforementioned parallelism, $k$ is sufficiently small

21: that our new approach is generally orders of magnitude faster than the previous state-of-the-art.

22: \end{abstract}

23: