In which case does FCM membership converge to 1/K?

I have tested the fuzzy C-means (FCM) algorithm using the R function fanny from the cluster package and I have wrote my own FCM algorithm to have more control on the distance function. The problem is that my final membership function converges many times to the 1/K, I mean almost all the values in the membership matrix are near to 1/K. I am wondering what are the conditions that influence the final result. I am using fuzzy parameter m=2.

• Thanks for clarifying that mbq! Now I can be sure I'm not qualified to answer the question! – onestop Jan 23 '11 at 14:57