1: \begin{abstract}

2: We focus on the dominant poles of the transfer function of a descriptor system. The transfer function

3: typically exhibits large norm at and near the imaginary parts of the dominant poles. Consequently,

4: the dominant poles provide information about the points on the imaginary axis where the ${\mathcal L}_\infty$ 

5: norm of the system is attained, and they are also sometimes useful to obtain crude reduced-order models.

6: For a large-scale descriptor system, we introduce a subspace framework to estimate a prescribed 

7: number of dominant poles. At every iteration, the large-scale system is projected into a small system,

8: whose dominant poles can be computed at ease. Then the projection spaces are expanded so that the 

9: projected system after subspace expansion interpolates the large-scale system at the computed dominant

10: poles. We prove an at-least-quadratic-convergence result for the framework, and provide numerical

11: results confirming this. On real benchmark examples, the proposed framework appears to be more

12: reliable than SAMDP [IEEE Trans. Power Syst. 21, 1471-1483, 2006], one of the widely used algorithms 

13: due to Rommes and Martins for the estimation of the dominant poles.

14: \end{abstract}

15: