Can k-means clustering error ever increase?

Question

I have implemented k-means clustering algorithm for 20-dimensional histograms. It seems to be working well on 2-dimensional histograms.

Occassionally, the sum of squared errors will increase slightly after an iteration. Is this normal, or a sign of an underlying bug?

If it is possible to increase the SSE, can anyone provide an example that I can visualise?

Answer 1

The assignment must not increase.

Seems as if you have a bug somewhere. Maybe you are using something other than squared Euclidean distance?

SSE = sum of squared errors, not sum of distances or something else based on distances.