18 August, 2010

fractions and proof-irrelevance IV - forgetting proofs

Continuing from previous posts on implementing fractions in Agda.

We've tried to force our representations of fractions to have a canonical form, but requiring the inclusion of a proof of that canonical form breaks canonicity.

So we need a different approach.

Data.Rational does it something like this, which I found hard to understand at first. Trying to understand isCoprime here was the motivation for this sequence of posts:

record ℚ : Set where
  field
    numerator     : ℤ
    denominator-1 : ℕ
    isCoprime     : True (coprime? ( ∣ numerator ∣ ) (suc denominator-1))

isComprime is different. Rather than directly expressing a proposition (Coprime numerator (suc denominator-1)) as the type of isCoprime, we instead use a slightly different proposition (coprime?) and wrap it in True.

coprime? has type Decidable Coprime. It is a decidable version of the coprime proposition. A normal proposition type has values that are proofs of that proposition. If I give you a proof, then you know the proposition is true. However, if I don't give you a proof, you don't know that the proposition is false. Perhaps its true but I haven't given you a proof. A decidable proposition type has values that are either proofs that the proposition is true, or proofs that the proposition is false, wrapped in yes or no respectively. coprime? is a function that decide whether two numbers are coprime or not, and return a corresponding proof.

True is a function that maps a decidable proposition type into another type: either ⊤ (when the proposition is true) or ⊥ (when the proposition is false).

⊤ (top) is the unit type, that has only one member value. ⊥ (bottom) is the empty type with no member values at all.

Both of those seem a bit useless to a traditional programmer: if I have a variable of the unit type ⊤, it can only have one value. So then I'll always know what that value is - its pretty much a constant. So why have a variable? And if I have a variable of the empty type ⊥ then I can't ever invent a value. If that variable is an input parameter, I can never even call my program, because I can't give input parameters. But it turns out this weird behaviour is exactly right for Data.Rational.

Say I'm going to construct a half. I'm going to pass numerator = 1 and denominator = 2. We also need to pass an isCoprime parameter.

To start with, what type of parameter is that?

That depends on the value of numerator and denominator (that's what dependent types are). Once we know the numerator and denominator, we can evaluate the expression True (coprime? ( ∣ numerator ∣ ) (suc denominator-1)) to find out the type of isCoprime. 1 and 2 are coprime, so coprime? evaluates to yes _, and so True (yes _) evaluates to ⊤.

So we need to give a value for coprime that is of type ⊤. ⊤ only has one value, so thats the value that we'll put in isCoprime. The resulting structure has no proof of coprimality in it. If two fractions have the same numerator and denominator, then they'll always have the same value of isCoprime. So we've got equality of equivalent fractions!

But now what happens if we try to construct a non-canonical fraction, for example 3/6 ?
Once we have specified a numerator of 3 and a denominator of 6, we can figure out the type of isCoprime. 3 and 6 are not coprime, and True (no _) will return ⊥.
So we need to provide a value of type ⊥. There are no values in ⊥. So there is no way we can call the constructor in this case, because we can't come up with a value for each parameter.

So now the goal is achieved - a data structure for representing rationals, where equal rationals are definitionally equal.

My goal in picking apart Data.Rational like this was two-fold: i) I wanted to use rationals, and wanted to understand what was going on; ii) I want to implement my own canonical forms (for finitely-generated abelian groups) and understanding how its done for rationals seems like a step in the right direction.

No comments:

Post a Comment