1: \begin{abstract}

2: Extremal optimization is a new general-purpose method for approximating

3: solutions to hard optimization problems.  We study the method in detail

4: by way of the NP-hard graph partitioning problem.  We discuss the

5: scaling behavior of extremal optimization, focusing on the convergence

6: of the average run as a function of runtime and system size.  The method

7: has a single free parameter, which we determine numerically and justify

8: using a simple argument.  Our numerical results demonstrate that on

9: random graphs, extremal optimization maintains consistent accuracy for

10: increasing system sizes, with an approximation error decreasing over

11: runtime roughly as a power law $t^{-0.4}$.  On geometrically structured

12: graphs, the scaling of results from the {\em average\/} run suggests

13: that these are far from optimal, with large fluctuations between

14: individual trials.  But when only the {\em best\/} runs are considered,

15: results consistent with theoretical arguments are recovered.

16: \hfil\break PACS number(s): 02.60.Pn, 05.65.+b, 75.10.Nr, 64.60.Cn.

17: \end{abstract}

18: