We hypothesize that it is easier to optimize the residual mapping than to optimize the original, unreferenced mapping.