A mistake was made. The optimal algorithm is only optimal for the case where cardinality constraints are not considered. Once considered, the algorithm omits some possible choices and hence may/will result in a good heuristic, but not an optimal choice. We are currently rerunning all the results with a truly optimal split algorithm. The initial results indicate that our SHIFT results will not be much affected, but until we know for sure we can not conclude anything. A corrected version of this paper will be made available as soon as done.