There is a well-known modification of the initializing step of K-means, named K-means++. It chooses cluster centers with probability proportional to its squared distance from the point's closest existing cluster center.

If the data are coming from a mixture distribution of normals, what is the distribution of the cluster centers after the initializing step?