Forgot one other important distinction: if we wanted to add a new expression subtype that can behave like one of the existing types for existing operations, that is not possible with the discriminated union variant. Say we wanted to add an SmallAddition, that stores its two numbers as bytes, but converts them to int when needed. We could add this without modifying any code in the visitors in the C# version (well, technically we may have a problem because we accessed the fields directly, but that's inconsequential). In the pseudo-Haskell version, we would need to modify all 6 evaluation functions just to handle the new Expression subtype.