Hi Mingbo,
The problem is caused by a poor estimate of what the single iteration smoothing kernel ought to be. The program tries to be conservative and pick a small one but sometimes it is too small and convergence cannot be reached. The decrease in FWHM is most likely due to precision errors in the smoothing or the estimation of the smoothness.
If you were using SurfSmooth directly, you can just increase the -sigma value (see -help output for this option). When running SurfSmooth through slow_surf_clustsim.py, you will need to manually edit the resulting script and specify sigma. I don't think sigma can be controlled from slow_surf_clustsim.py directly at the moment.
cheers,
Ziad