Is there any generic way to preserve loss of numerical accuracy for any given function?