We did the same for vectors. We don't have enough guarantee that Python will return Indet when evaluating float(u"-nan").