Gap statistic with cosine similarity

[ckolluru]

Hi,

Is it possible to use the gap statistic as is when considering distances based on cosine similarity instead of euclidean L2?

Since cosine similarity is high for low euclidean distances, I wonder if the formulae need to be modified in such a case.

When I try to do so, I get a negative gap statistic value that increases with increasing number of clusters. Is that expected?

Thanks.