[Agda] Re: Associativity for free!

[Alan Jeffrey]

I've now finished up the NBE algorithm for the simply-typed 
lambda-calculus. It defines a reduction semantics, and uses NBE to show 
that every term reduces to one in normal form.

At this point, I'm afraid I think of this as an interesting idea, but 
one where some pragmatic nastiness gets in the way of it being useful. 
The main problems are:

- Huge normal forms, which slow Agda down horribly.

- Incomprehensible normal forms, which make debugging remarkably difficult.

- Less unification -- from unifying (a :: as) with (b :: bs) Agda 
doesn't unify a with b or as with bs.

The implementation is up at https://github.com/agda/agda-assoc-free

A.

On 11/02/2011 11:26 AM, Jeffrey, Alan S A (Alan) wrote:
> Hi everyone,
>
> My original motivation for "associativity for free" was properties about
> lambda-calculi. I've now worked through some of the details, and here is
> the syntax of the simply-typed lambda-calculus:
>
>> https://github.com/agda/agda-frp-js/blob/master/src/agda/FRP/JS/Model/STLambdaC/Exp.agda
>
> The interesting thing is weakening, with type:
>
>     weaken+ : ∀ B Γ Δ {T} → Exp (B ++ Δ) T → Exp (B ++ Γ ++ Δ) T
>
> and (news just in!) weakening commutes with itself (up to propositional
> equality):
>
>     weaken+ Φ Ψ (B ++ Γ ++ Δ) (weaken+ (Φ ++ B) Γ Δ M)
>       ≡ weaken+ (Φ ++ Ψ ++ B) Γ Δ (weaken+ Φ Ψ (B ++ Δ) M)
>
> Note that you can't even state this property without associativity on
> the nose, since the lhs has type (Φ ++ Ψ ++ (B ++ Γ ++ Δ)) and the rhs
> has type ((Φ ++ Ψ ++ B) ++ Γ ++ Δ).
>
> The proof makes use of a list-membership type, defined in:
>
>> https://github.com/agda/agda-frp-js/blob/master/src/agda/FRP/JS/DList.agda
>
> It satisfies weakening on the left and right:
>
>     _≪_ : ∀ {A a as} → (a ∈ as) → ∀ bs → (a ∈ (as ++ bs))
>     _≫_ : ∀ {A a bs} → ∀ as → (a ∈ bs) → (a ∈ (as ++ bs))
>
> and moreover has left and right units and associativity properties, all
> up to beta-equivalence (that is, all these properties have proof refl):
>
>     (a∈as ≪ []) ≡ a∈as
>     ([] ≫ a∈as) ≡ a∈as
>     (a∈as ≪ bs ≪ cs) ≡ (a∈as ≪ bs ++ cs)
>     ((as ≫ a∈bs) ≪ cs) ≡ (as ≫ (a∈bs ≪ cs))
>     (as ≫ bs ≫ a∈cs) ≡ (as ≫ bs ≫ a∈cs)
>
> Under the hood, this is implemented in terms of difference naturals
> (that is an implementation of ℕ where + and ♯0 form a monoid up to
> beta-equivalence), and an implementation of difference lists such that
> the length function is a monoid homomorphism up to beta-equivalence:
>
>     (len (as ++ bs) ≡ (len as + len bs))
>     (len [] ≡ ♯0)
>
> With all that in place, the proof of weakening is pretty
> straightforward. The only case which requires any work is variables, and
> a typical case is:
>
>     begin
>       xweaken+ Φ Ψ (B ++ Γ ++ Δ) (case Φ (B ++ Γ ++ Δ) (Φ ≫ x ≪ Γ ≪ Δ))
>     ≡⟨ cong (xweaken+ Φ Ψ (B ++ Γ ++ Δ)) (case-≫ Φ (x ≪ Γ ≪ Δ)) ⟩
>       (Φ ≫ Ψ ≫ x ≪ Γ ≪ Δ)
>     ≡⟨ cong (xweaken+ (Φ ++ Ψ ++ B) Γ Δ) (sym (case-≪ (Φ ≫ Ψ ≫ x) Δ)) ⟩
>       xweaken+ (Φ ++ Ψ ++ B) Γ Δ (case (Φ ++ Ψ ++ B) Δ (Φ ≫ Ψ ≫ x ≪ Δ))
>     ∎
>
> I don't really want to think about what this proof would look like with
> all the associativity wiring made explicit.
>
> There is still a bit of annoying associativity left, because I can't see
> how to get coproducts to associate on the nose, so there's lots of gore
> needed to define a case function:
>
>     data Case {A} a (as bs : List A) : Set where
>       inj₁ : (a∈as : a ∈ as) → Case a as bs
>       inj₂ : (a∈bs : a ∈ bs) → Case a as bs
>
>     case : ∀ {A a} as bs → (a ∈ (as ++ bs)) → Case {A} a as bs
>
> and then prove that it satisfies associativity (up to propositional
> equality, not beta-equivalence).
>
> A.
>
> On 10/26/2011 09:48 AM, Jeffrey, Alan S A (Alan) wrote:
>> Hi everyone,
>>
>> A quick note about a trick for representing lists (or any other monoid)
>> up to associativity. Am I reinventing a wheel here?
>>
>> I've been doing some mucking around recently with proofs about
>> lambda-calculi and was hitting annoying problems with associativity
>> on lists. A prototypical example is:
>>
>>      weakening : ∀ {A a} (as bs cs : List A) →
>>       (a ∈ (as ++ cs)) → (a ∈ (as ++ bs ++ cs))
>>
>> It's easy to prove that ++ is associative, but making use of that fact
>> is rather painful, as there's lots of manual wiring with subst and cong,
>> for example:
>>
>>      weaken2 : ∀ {A a} (as bs cs ds : List A) →
>>        (a ∈ (as ++ bs ++ ds)) → (a ∈ (as ++ bs ++ cs ++ ds))
>>      weaken2 {A} {a} as bs cs ds a∈es =
>>        subst (λ X → (a ∈ X)) (++-assoc as bs (cs ++ ds))
>>          (weakening (as ++ bs) cs ds
>>            (subst (λ X → (a ∈ X)) (sym (++-assoc as bs ds)) a∈es))
>>
>> This gets tiresome pretty rapidly. Oh if only there were a data
>> structure which represented lists, but for which ++ was associative up
>> to beta reduction... Oh look there already is one, it's called
>> difference lists. Quoting the standard library:
>>
>>      DiffList : Set → Set
>>      DiffList a = List a → List a
>>
>>      _++_ : ∀ {a} → DiffList a → DiffList a → DiffList a
>>      xs ++ ys = λ k → xs (ys k)
>>
>> Unfortunately, difference lists are not isomorphic to lists: there's a
>> lot more functions on lists than just concatenation functions. This
>> causes headaches (e.g. writing inductive functions on difference lists).
>>
>> A first shot at fixing this is to write a data structure for those
>> difference lists which are concatenations:
>>
>>      record Seq (A : Set) : Set where
>>        constructor seq
>>        field
>>          list : List A
>>          fun : List A → List A
>>          ok : fun ≡ _++_ list  -- Broken
>>
>> It's pretty straightforward to define the monoidal structure for Seq:
>>
>>      ε : ∀ {A} → Seq A
>>      ε = seq [] id refl
>>
>>       _+_ : ∀ {A} → Seq A → Seq A → Seq A
>>      (seq as f f✓) + (seq bs g g✓) = seq (f bs) (f ∘ g) (...easy lemma...)
>>
>> Unfortunately, we're back to square one again, because + is only
>> associative up to propositional equality, not beta reduction. This time,
>> it's the ok field that gets in the way. So we can fix that, and just
>> make the ok field irrelevant, that is:
>>
>>      record Seq (A : Set) : Set where
>>        constructor seq
>>        field
>>          list : List A
>>          fun : List A → List A
>>          .ok : fun ≡ _++_ list  -- Fixed
>>
>>      open Seq public renaming (list to [_])
>>
>> Hooray, + is now associative up to beta reduction:
>>
>>      +-assoc : ∀ {A} (as bs cs : Seq A) →
>>        ((as + bs) + cs) ≡ (as + (bs + cs))
>>      +-assoc as bs cs = refl
>>
>> Moreover, given (http://thread.gmane.org/gmane.comp.lang.agda/3034):
>>
>>      ≡-relevant : ∀ {A : Set} {a b : A} → .(a ≡ b) → (a ≡ b)
>>
>> we can define:
>>
>>      data Case {A : Set} : Seq A → Set where
>>        [] : Case ε
>>        _∷_ : ∀ a as → Case (a ◁ as)
>>
>>      case : ∀ {A} (as : Seq A) → Case as
>>
>> and this gives us recursive functions over Seq (which somewhat to my
>> surprise actually get past the termination checker), for example:
>>
>>      ids : ∀ {A} → Seq A → Seq A
>>      ids as        with case as
>>      ids .(a ◁ as) | a ∷ as = a ◁ ids as
>>      ids .ε        | []     = ε
>>
>> For example, we can now define weakening:
>>
>>      weakening : ∀ {A a} (as bs cs : Seq A) →
>>        (a ∈ [ as + cs ]) → (a ∈ [ as + bs + cs ])
>>
>> and use it up to associativity without worrying hand-coded wiring:
>>
>>      weaken2 : ∀ {A a} (as bs cs ds : Seq A) →
>>        (a ∈ [ as + bs + ds ]) → (a ∈ [ as + bs + cs + ds ])
>>      weaken2 as bs = weakening (as + bs)
>>
>> Hooray, cake! Eating it!
>>
>> I suspect this trick works for most any monoid, so you can do
>> associativity on naturals, vectors, partial orders with join, etc. etc.
>>
>> The idea of using difference lists to represent lists is standard, but
>> as far as I know, this use of irrelevant equalities to get structural
>> recursion is new. Anyone know of any prior art?
>>
>> A.