On 7/25/12, Christian Sternagel wrote:

Dear Alexander,

On 07/26/2012 01:09 PM, Alexander Solla wrote:

...

On 7/25/12, Christian Sternagel wrote:

...

On 07/26/2012 11:53 AM, Alexander Solla wrote:

...

The classically valid inference:

(x == y) = _|_ => (y == x) = _|_ Btw: whether this inference is valid or not depends on the semantics of (==) and that's exactly what I was asking about. In Haskell we just have type classes, thus (==) is more or less arbitrary (apart from its type).

Indeed. This is true for the interpretation of any function. But you apparently want to treat (==) as equality. This may or may not be possible, depending on the interpretation you choose. Does _|_ == _|_, or _|_ =/= _|_, or do these questions not even make semantic sense in the object language? That's what you need to answer, and the solution to your problem will become clear. Note that picking any of these commits you to a "philosophical" position, insofar as the commitment will induce a metalanguage which excludes expressions from other metalanguages.

for my specific case (HOLCF) there is already a fixed metalanguage which has logical equality (=) and differentiates between type 'bool' (True, False) for logical truth and type 'tr' (TT, FF, _|_) for truth-valued computable functions (including nontermination/error). Logical equality satisfies '_|_ = _|_'. Now in principle (==) is just an arbitrary function (for each instance) but I gather that there is some intended use for the type class Eq (and I strongly suspect that it is to model equality ;), I merely want to find out to what extend it does so).

Yes, in my opinion, its purpose is to model mathematical/logical equality. That said, you brought up a very important point which you need to address. I'll discuss it more below. I will be using scare-quotes quite a bit. I'm not being sarcastic -- please ask if you don't understand what I mean.

Currently the axioms of the formal eq class include

(_|_ == y) = _|_ (x == _|_) = _|_

(and this decision was just based on how ghci evaluates corresponding expressions).

The equations you use above would be (roughly) written '(_|_ == _|_) = TT' and '(_|_ /= _|_) = TT' in HOLCF and neither of them is satisfied, since both expressions are logically equivalent to _|_ (in HOLCF). (But still both expressions make sense.)

Okay, you have a metalanguage. And it places limits on what you can express. But you can also put limits on what you can express, to induce a sublanguage of the metalanguage. I don't know how "natural" this idea is given your circumstances and goals, but you may want to consider it. I promise I'll get to the meat of my email soon.

During the discussion I revisited the other two axioms I was using for eq... and now I am wondering whether those make sense. The other two axioms where

(x == y) = TT ==> x = y (x == y) = FF ==> not (x = y)

The second one should be okay, but the first one is not true in general, since it assumes that we would only ever use Eq instances implementing syntactic equality (which should be true for deriving?). E.g., for the data type

data Frac = Frac Int Int

we would have 'not (Frac 1 2 = Frac 2 4)' in HOLCF (since we have injectivity of all constructors). But nobody prevents a programmer from writing an Eq instance of Frac that compares 'normal forms' of fractions.

Now, /this/ is an example of what I've been trying to get at. (To be fair, it's a generalization of what I've been trying to get at) You put it in "behavioral" terms -- "nothing stops a programmer", whereas I've been trying to put it in "interpretive" terms (... is ..., but ... means many things -- you have to pick an interpretation) A hypothetical programmer can write:

data Fraction = Fraction Int Int data Rational = Rational Int Int

and have different semantics for each. In particular, we should treat Fraction 1 2 to be distinct from Fraction 2 4, /as fractions/. And Rational 1 2 should be the same as Rational 2 4, /as rational numbers/. These "implied" semantics are just as important as the minimum Haskell demands (in my opinion). The minimum Haskell demands is not (and cannot) be rich enough to distinguish between intentions/intensions. "Obviously", Fraction and Rational are intimately related (the rationals "should" be a ring of fractions in "normal form"). And functions between Fraction and Rational "should" preserve the "correct" relationships. But, what are these? We can come up with theorems, but some things are nothing but the programmer/mathematician's choice. In other words, by choosing a semantic where (Rational 1 2) == (Rational 2 4), a programmer/mathematician has *generated* a sublanguage of Haskell by imposing his *intended* semantics. This is an example of "intensions" versus "extensions" in logic. So, your task is... difficult. From my point of view, you will have to axiomatize at least one set of "best practices" for Haskell programming, and target /that/ Haskell sublanguage in your HOLCF analyzer. You can even do more than one Haskell sublanguage, but as I said before, each set of "best practices" will commit you to a different "philosophical" position.