Wednesday, 29 February 2012

Formation Rules for Strictly Positive Types

For my thesis I've needed to use strictly positive types to avoid the problems which can occur with types of the form: νD. D → D, which were discussed briefly in my post System F is Amazing.

In order to ensure types are positive they should be restricted in some way.  It is therefore convenient to have formation rules for types which ensure that we can't make bad types. My starting point was a wonderful paper by Andreas Abel and Thorsten Altenkirch [1]

Specifically, this paper describes interleaved inductive types, which is close to what I want.   However, I also want to be able to have polymorphism using ∀.  Since I'm interested in becoming more familiar with Agda (a language in which I'm still a novice) I've written the following as a prototype of my type formation rules, as an experiment.

First, we have to set up the appropriate libraries.

module StrictlyPositive where 

open import Data.Nat
open import Data.Bool 
open import Data.List 
open import Data.Unit using ()
open import Relation.Nullary.Core
open import Data.Sum 

The first data type we define is a context of type variables. Since we are going to use de Bruijn indices, we can use a trick and say that the type variable context is really just a bound on the number of free variables. Since our types are not themselves typed (as our theory is impredicative) we are justified in doing this as the only information needed formation is that the variable is in a context.

Ctx : Set
Ctx = 

We then define an ∈ relation which is true whenever a variable is below the bound.

data __ :   Ctx  Set where 
  bounded :  {n Δ}  suc n  Δ  n  Δ

It turns out to be convenient to define lemmas which perform inversions on data-types in Agda. In Coq, I would generally just use a tactic to perform the inversion. However, there are advantages to Agda's way of doing things, one of which is that it obviates a lot of playing with equational reasoning. So far I've yet to be forced to learn how equational reasoning works in Agda since I haven't had to use it. I'm a little frightened that things work so well!

inInversion :  {n Δ}  n  Δ  suc n  Δ
inInversion (bounded p) = p

Next we prove that ∈ is decidable, which is pretty straight forward since it is defined using ≤ on natural numbers.

inctx :  n Δ  Dec (n  Δ)
inctx n Δ with (suc n) ? Δ
inctx n Δ | no p = no (λ x  p (inInversion x))
inctx n Δ | yes q  = yes (bounded q)

Next we define the syntax for types. We have ① for unit, ι which creates a type variable, η which creates a fixed-point type variable, ⇒ which is the function type, × which is the type of a pair, ⊕ which is the type of an injection into the left or right, μ which makes a fixed point, and Π which is the type of a type abstraction.

data Ty : Set where
   : Ty
  ι :   Ty
  η :   Ty
  __ : Ty  Ty  Ty
  _×_ : Ty  Ty  Ty 
  __ : Ty  Ty  Ty
  μ : Ty  Ty
  Π : Ty  Ty

Next we demonstrate the formation rules. Here we use two contexts. One is for ι variables, the other for η. The reason for segregating them is that we can allow ι variables to be placed more freely, while η variables need to be tightly controlled to ensure strictly positive types. The restriction manifests itself in the rule for ⇒, which allows formation with a type to the left of the arrow, only if the context for η vars is empty. The reader will notice that every time we go under a Π binder, the first context is increased in size, and every time we go under a μ, we increase the second. When we check that variables are in context, we refer to these segregated contexts with our two different type variables.

data _·__ : Ctx  Ctx  Ty  Set where 
     UnitValid :  {Δ Δ+}  
       Δ · Δ+  
     VarValid :  {n Δ Δ+}  
       (n  Δ)  
       Δ · Δ+  (ι n) 
     FixValid :  {n Δ Δ+}  
       (n  Δ+)  
       Δ · Δ+  (η n) 
     ImpValid :  {Δ Δ+ α β}  
       Δ · 0  α  
       Δ · Δ+  β  
       Δ · Δ+  (α  β)
     AndValid :  {Δ Δ+ α β}  
       Δ · Δ+  α  
       Δ · Δ+  β  
       Δ · Δ+  (α × β)
     OrValid :  {Δ Δ+ α β}  
       Δ · Δ+  α  
       Δ · Δ+  β  
       Δ · Δ+  (α  β)
     MuValid :  {Δ Δ+ α}  
       Δ · (suc Δ+)  α  
       Δ · Δ+  μ α
     AllValid :  {Δ Δ+ α}  
       (suc Δ) · Δ+  α  
       Δ · Δ+  Π α

Again, we find it convenient to define a number of inversion rules, this type on type formation.

varValidInversion :  {Δ Δ+ n}  Δ · Δ+  ι n  n  Δ
varValidInversion {Δ} {Δ+} {n} (VarValid p) = p

fixValidInversion :  {Δ Δ+ n}  Δ · Δ+  η n  n  Δ+
fixValidInversion {Δ} {Δ+} {n} (FixValid p) = p

orValidInversionL :  {Δ Δ+ α β}  Δ · Δ+  (α  β)  (Δ · Δ+  α)
orValidInversionL {Δ} {Δ+} {α} {β} (OrValid p q) = p

orValidInversionR :  {Δ Δ+ α β}  Δ · Δ+  (α  β)  (Δ · Δ+  β)
orValidInversionR {Δ} {Δ+} {α} {β} (OrValid p q) = q

impValidInversionL :  {Δ Δ+ α β}  Δ · Δ+  (α  β)  (Δ · 0  α)
impValidInversionL {Δ} {Δ+} {α} {β} (ImpValid p q) = p

impValidInversionR :  {Δ Δ+ α β}  Δ · Δ+  (α  β)  (Δ · Δ+  β)
impValidInversionR {Δ} {Δ+} {α} {β} (ImpValid p q) = q

andValidInversionL :  {Δ Δ+ α β}  Δ · Δ+  (α × β)  (Δ · Δ+  α)
andValidInversionL {Δ} {Δ+} {α} {β} (AndValid p q) = p

andValidInversionR :  {Δ Δ+ α β}  Δ · Δ+  (α × β)  (Δ · Δ+  β)
andValidInversionR {Δ} {Δ+} {α} {β} (AndValid p q) = q

allValidInversion :  {Δ Δ+ α}  Δ · Δ+  Π α  ((suc Δ) · Δ+  α)
allValidInversion {Δ} {Δ+} {α} (AllValid p) = p

muValidInversion :  {Δ Δ+ α}  Δ · Δ+  μ α  (Δ · (suc Δ+)  α)
muValidInversion {Δ} {Δ+} {α} (MuValid p) = p

Now we would like to show that we can demonstrate whether or not a type is valid or not. This program will construct a proof of the validity of a given type, or reject it with a proof that one can not be constructed. Intermediate computations which attempt to construct subproofs or proofs that no subproof exists are done using with. This allows us to further pattern much, just as we would do with parameters of functions in Haskell. The notation (λ { (VarValid q) p q }) is especially cool. It is a pattern matching lambda, which allows us to destruct the argument with each possible case. In this case, because of the type, there is exactly one case which allows us to prove that this type can not be valid.

isValid :  Δ Δ+ α  Dec (Δ · Δ+  α)
isValid Δ Δ+        = yes UnitValid
isValid Δ Δ+ (ι n)   with (inctx n Δ)
isValid Δ Δ+ (ι n)   | yes p = yes (VarValid p)
isValid Δ Δ+ (ι n)   | no p  = no (λ { (VarValid q)  p q })
isValid Δ Δ+ (η n)   with (inctx n Δ+)
isValid Δ Δ+ (η n)   | yes p = yes (FixValid p)
isValid Δ Δ+ (η n)   | no p  = no (λ { (FixValid q)  p q })
isValid Δ Δ+ (α  β) with isValid Δ 0 α
isValid Δ Δ+ (α  β) | yes p with isValid Δ Δ+ β
isValid Δ Δ+ (α  β) | yes p | yes q = yes (ImpValid p q) 
isValid Δ Δ+ (α  β) | yes p | no q  = no (λ x  q (impValidInversionR x))
isValid Δ Δ+ (α  β) | no p  = no (λ x  p (impValidInversionL x))
isValid Δ Δ+ (α × β) with isValid Δ Δ+ α
isValid Δ Δ+ (α × β) | yes p with isValid Δ Δ+ β
isValid Δ Δ+ (α × β) | yes p | yes q = yes (AndValid p q) 
isValid Δ Δ+ (α × β) | yes p | no q  = no (λ x  q (andValidInversionR x))
isValid Δ Δ+ (α × β) | no p  = no (λ x  p (andValidInversionL x))
isValid Δ Δ+ (α  β) with isValid Δ Δ+ α
isValid Δ Δ+ (α  β) | yes p with isValid Δ Δ+ β
isValid Δ Δ+ (α  β) | yes p | yes q = yes (OrValid p q) 
isValid Δ Δ+ (α  β) | yes p | no q  = no (λ x  q (orValidInversionR x))
isValid Δ Δ+ (α  β) | no p  = no (λ x  p (orValidInversionL x))
isValid Δ Δ+ (Π α)   with isValid (suc Δ) Δ+ α
isValid Δ Δ+ (Π α)   | yes p = yes (AllValid p) 
isValid Δ Δ+ (Π α)   | no p  = no (λ x  p (allValidInversion x))
isValid Δ Δ+ (μ α)   with isValid Δ (suc Δ+) α
isValid Δ Δ+ (μ α)   | yes p = yes (MuValid p) 
isValid Δ Δ+ (μ α)   | no p  = no (λ x  p (muValidInversion x))

Finally we give an example of some terms which can be typed, and some which can not. When there is no inhabitant of a pattern in Agda, we can use (). Agda is smart enough to know that there is no sense in a right-hand-side of a pattern which has no inhabitants because the type is indexed in such a way that inversion would yield no cases. It's very clever!

idTy : Ty
idTy = Π ((ι zero)  (ι zero))

idValid : 0 · 0  idTy 
idValid = AllValid
            (ImpValid (VarValid (bounded (ss zn)))
             (VarValid (bounded (ss zn))))

listTy : Ty
listTy = Π (μ (  ((ι 0) × (η 0))))

listValid : 0 · 0  listTy
listValid = AllValid
               (OrValid UnitValid
                (AndValid (VarValid (bounded (ss zn)))
                 (FixValid (bounded (ss zn))))))

notPos : Ty 
notPos = μ ((ι 0)  (ι 0))

notPosInvalid : ¬ (0 · 0  notPos)
notPosInvalid x with muValidInversion x
notPosInvalid x | p with impValidInversionL p
notPosInvalid x | p | q with varValidInversion q
notPosInvalid x | p | q | bounded ()

dodgy : Ty
dodgy = μ (μ ((ι 1)  (ι 0)))

dodgyInvalid :  ¬ (0 · 0  dodgy)
dodgyInvalid x with muValidInversion x
dodgyInvalid x | p with muValidInversion p 
dodgyInvalid x | p | q with impValidInversionL q
dodgyInvalid x | p | q | r with varValidInversion r 
dodgyInvalid x | p | q | r | bounded ()

isThisDodgy : Ty 
isThisDodgy = Π (μ (μ ((((ι 0)  (η 1))  (η 0)))))

isThisDodgyValid : ¬ (0 · 0  isThisDodgy)
isThisDodgyValid x with allValidInversion x
isThisDodgyValid x | p with muValidInversion p
isThisDodgyValid x | p | q with muValidInversion q
isThisDodgyValid x | p | q | r with impValidInversionL r
isThisDodgyValid x | p | q | r | s with impValidInversionR s
isThisDodgyValid x | p | q | r | s | t with fixValidInversion t
isThisDodgyValid x | p | q | r | s | t | bounded ()

[1] A Predicative Strong Normalisation Proof for a λ-calculus with Interleaving Inductive Types

Thursday, 23 February 2012

System F is Amazing - Part II

As I mentioned in my last post System-F is amazing. I haven't written in a very long time because I have been busy ensuring that I graduate. I passed my viva finally! So now that I'm back in the real world I thought I'd talk a bit about things that I found while I was in my cave. I'll put my dissertation up with some notes when I've completed the revisions.

Those who have dealt with proof assistants based on type theories (Such as Agda and Coq) might have noticed that we often require types to have a positivity restriction. This restriction essentially states that you can't have non-positive occurrences of the recursive type variable. As I described earlier, System F avoids this whole positivity restriction by forcing the programmer to demonstrate that constructors themselves (as a Church encoding) can be implemented, avoiding the problem of uninhabited types being inadvertently asserted.

So this means that in some sense positivity is probably the same thing as saying that there is an algorithmic method of translation of a data-type into System-F. I suspect (but can not yet prove) it also means that we should be allowed to write down datatypes as long as we can show that the inductive or coinductive types are inhabited by a Church encoding! If we could also do this trick for the Calculus of Constructions it might give a tricky way to increase the number of (co)inductive types we are allowed to write down.

Now, I'll write down an example of what I'm talking about in System-F, embedded in the propositional fragment of Coq (which is suitably impredicative), so that you can see a very straightforwardly non-positive type which has a Church encoding. This example of the LamMu type came from Andreas Abel's paper on sized types (which one I forget!) where he demonstrates that his system allows the definition.

In the very final entry we attempt to enter LamMu into the Coq Inductive type framework and watch it fail (for reasons of positivity violation). The original type definition gives us three constructors, each of which is proved in turn below by a "constructor".

Definition LamMu := forall (X : Prop),
  (Nat -> X) ->
  (Nat -> List X -> X) ->
  ((forall (Y : Prop), ((X -> Y) -> Y)) -> X) -> X.

Definition var : Nat -> LamMu :=
  fun (n : Nat) =>
    fun (X : Prop)
      (v : Nat -> X)
      (f : Nat -> List X -> X)
      (m : (forall (Y : Prop), ((X -> Y) -> Y)) -> X) =>
      v n.

Definition func : Nat -> List LamMu -> LamMu :=
  fun (n : Nat) =>
    fun (t : List LamMu) =>
      fun (X : Prop)
        (v : Nat -> X)
        (f : Nat -> List X -> X)
        (m : (forall (Y : Prop), ((X -> Y) -> Y)) -> X) =>
        f n (t (List X) (nil X) (fun (x : LamMu) (y : List X) => cons (x X v f m) y)).

Definition mu : (forall (Y : Prop), (LamMu -> Y) -> Y) -> LamMu.
  unfold LamMu.
    (fun (zi : (forall (Y : Prop), (LamMu -> Y) -> Y)) =>
      fun (X : Prop)
        (v : Nat -> X)
        (f : Nat -> List X -> X)
        (m : (forall (Y : Prop), ((X -> Y) -> Y)) -> X) =>
        m (fun (Y : Prop) (g : X -> Y) =>
          g (zi X (fun e : LamMu => e X v f m)))).

Inductive LamMu2 : Type :=
| Var : Nat -> LamMu2
| Fun : Nat -> LamMu2 -> LamMu2
| Mu : ((forall Y, ((LamMu2 -> Y) -> Y)) -> LamMu2) -> LamMu2