We hope that we could give some evidence of the complexity of merging. We believe that this complexity is typically underestimated, and that designing development processes based on branch patterns that imply a lot of merging is asking for trouble: it spreads this complexity to realms in which it doesn't belong, making the resulting complexity artificial.
We believe that rolling back changes gives a convincing example: it is incomparably simpler and more reliable to move back a label than to perform a subtractive merge. The apparent problem with the former is that the version belonging to the baseline, the one bearing the label, is after the rollback not the latest on the branch. Why should this be a problem? We can only find bad and artificial reasons: "Doctor, when I hit my head on the wall, it hurts".
If one decides to identify the baseline with labels, one doesn't care anymore on what branch the baseline version of any element is sitting: merging "back" to integration...