# Difference between revisions of "Type Theoretic Interpretation of the Final Chain"

## Introduction

Barr has shown in [1] that final coalgebras can be constructed as the Cauchy completion of initial algebras. We will show in this article how this construction can be carried out in an intuitionistic type theory like Martin-Löf type theory. Of course, we cannot expect to get a final coalgebra, but only a weakly final one. However, as it turns out, we can only prove that the coinductive extension (corecursion) is a homomorphism up to bisimilarity, and not up to computational equality. The latter can be proved if we had function extensionality.

All definitions and proofs in this article will be given in Agda code.

## Polynomials, W-Types and M-types

For simplicity, we restrict ourselves to polynomial functors. These are given as follows. <definition id="polynomial-functor"> A polynomial P is a pair A ▹ B with A : Set and B : A → Set, that is, if P : Poly:

data Poly : Set₁ where
_▹_ : (A : Set) (B : A → Set) → Poly

We define an interpretation of polynomials as functors by and . In Agda, we define two actions: ⟦_⟧ is the action on types and ⟦_⟧₁ is the action on terms.

⟦_⟧ : Poly → Set → Set
⟦ A ▹ B ⟧ X = Σ[ a ∈ A ] (B a → X)

⟦_⟧₁ : (P : Poly) → ∀{X Y} → (X → Y) → (⟦ P ⟧ X → ⟦ P ⟧ Y)
⟦ A ▹ B ⟧₁ f (a , α) = (a , (λ x → f (α x)))

(Weakly) initial algebras for polynomial functors are called W-types, whereas (weakly) final coalgebras for these are called M-types. </definition>

Given a polynomial P, we interpret it as a signature and define to be the type of terms with one variable, which marks the holes of such a term. We can also see as the set of trees over the signature P, where we may mark leaves with a variable for which we can substitute other trees. Thus is given by

data _* (P : Poly) : Set where
sup : ⟦ P ⟧ (P *) ⊎ ⊤ → P *

That is, has one constructor sup that takes either a label and a branching map for the subtrees or a variable, and assembles these into one tree.

## Cauchy sequences of finite trees

The first step to the construction of M-types from W-types is to define what it means for a sequence of finite trees to converge. We define a relation extends on , such that T extends S if we can find trees that we can substitute at the leaves of S to obtain T. This relation is defined inductively as follows.

_extends_ : ∀{P} → P * → P * → Set
_extends_ {A ▹ B} (sup (inj₁ (a' , β))) (sup (inj₁ (a , α)))
= Σ[ e ∈ a' ≡ a ] (∀ (b : B a') → (β b) extends α (subst B e b))
sup (inj₂ _) extends sup (inj₁ _) = ⊥
sup (inj₁ x) extends sup (inj₂ _) = ⊤
sup (inj₂ _) extends sup (inj₂ _) = ⊥

This data type essentially states that T extends S if either T is non-trivial while S is a variable (so we can substitute T for the variable), if both trees are variables, or both are trees with the same root and related subtrees.

Next, we define PreApprox, which are sequences over , and we characterise the converging sequences and non-empty trees.

PreApprox : Poly → Set
PreApprox P = ℕ → P *

converges : ∀{P} → PreApprox P → Set
converges u = ∀ n → u (1 + n) extends u n

non-empty : ∀{P} → P * → Set
non-empty (sup (inj₁ x)) = ⊤
non-empty (sup (inj₂ _)) = ⊥

This allows us now to define the encoding of the (weakly) final coalgebra for a polynomial P by

_∞ : Poly → Set
P ∞ = Σ[ u ∈ PreApprox P ] (non-empty (u 0) × converges u)

One can then show that this is indeed a coalgebra for ⟦ P ⟧, that is, we have a map

out : ∀{P} → P ∞ → ⟦ P ⟧ (P ∞).
out {A ▹ B} (u , ne , conv) = (a , α)
where ...

The idea to define this map is as follows. First, we can use for a the label of the root of the first tree in the sequence because u(0) this tree is non-empty (witnessed by ne). Second, we know from convergence that all trees in the sequence have a as root label. This allows us to now define : Put , where is the subtree at branch b of . All are non-empty and convergence of follows from convergence of u.

Moreover, we also get a corecursion principle for :

corec : ∀{P} {C : Set} → (C → ⟦ P ⟧ C) → (C → P ∞)

The idea here is that corec f c is a sequence such that is the tree obtained by the n-fold iteration of f starting at c. This sequence is clearly converging and starts with a non-empty tree.

## The Problem

Having defined all of this, we would now expect that corec f is actually a coalgebra homomorphism, i.e., that out (corec f c) = ⟦ P ⟧₁ (corec f) (f c). However, it is fairly clear that this equality (up to reduction) cannot hold in an intensional type theory, as both sides are functions. Thus, the best we can hope for is to define equality (say ) on PreApprox to be extensional equality of functions but then we lose replacement . This will force us to restrict to predicates that satisfy replacement for . Hence, we need to reason using Setoids, which becomes very cumbersome.

## Conclusion

The conclusion of this is, so far, that from a practical perspective native coinductive types are better as we obtain from corecursion a coalgebra homomorphism up to syntactic reduction, and not just up to some external equality. This makes them more usable in practice.

## References

1. M. Barr, “Terminal coalgebras in well-founded set theory,” Theoretical Computer Science, vol. 114, no. 2, pp. 299–315, 1993.

## Full Code

{- # --without-K # -}

open import Data.Unit
open import Data.Empty
open import Data.Nat
open import Data.Sum as Sum
open import Data.Product
open import Function
open import Relation.Binary.PropositionalEquality
open import Relation.Unary
open import Relation.Binary
open import Relation.Binary using (module IsEquivalence; Setoid; module Setoid)
open ≡-Reasoning
import Relation.Binary.HeterogeneousEquality as HE
open import Relation.Binary.HeterogeneousEquality using (_≅_; refl)

-- | Polynomials
data Poly : Set₁ where
_▹_ : (A : Set) (B : A → Set) → Poly

-- | Interpretation of polynomials as functor, object part.
⟦_⟧ : Poly → Set → Set
⟦ A ▹ B ⟧ X = Σ[ a ∈ A ] (B a → X)

-- | Interpretation of polynomials as functor, morphism part.
⟦_⟧₁ : (P : Poly) → ∀{X Y} → (X → Y) → (⟦ P ⟧ X → ⟦ P ⟧ Y)
⟦ A ▹ B ⟧₁ f (a , α) = (a , (λ x → f (α x)))

-- | Finite trees over the signature determined by P with one free variable.
data _* (P : Poly) : Set where
sup : ⟦ P ⟧ (P *) ⊎ ⊤ → P *

-- | Inverse of sup.
sup⁻¹ : ∀{P} → P * → ⟦ P ⟧ (P *) ⊎ ⊤
sup⁻¹ (sup x) = x

_extends_ : ∀{P} → P * → P * → Set
_extends_ {A ▹ B} (sup (inj₁ (a' , β))) (sup (inj₁ (a , α)))
= Σ[ e ∈ a' ≡ a ] (∀ (b : B a') → (β b) extends α (subst B e b))
sup (inj₂ _) extends sup (inj₁ _) = ⊥
sup (inj₁ x) extends sup (inj₂ _) = ⊤
sup (inj₂ _) extends sup (inj₂ _) = ⊥

PreApprox : Poly → Set
PreApprox P = ℕ → P *

converges : ∀{P} → PreApprox P → Set
converges u = ∀ n → u (1 + n) extends u n

non-empty : ∀{P} → P * → Set
non-empty (sup (inj₁ x)) = ⊤
non-empty (sup (inj₂ _)) = ⊥

root : ∀{A B} → (t : (A ▹ B) *) → non-empty t → A
root (sup (inj₁ x)) p = proj₁ x
root (sup (inj₂ y)) ()

branch : ∀{A B} → (t : (A ▹ B) *) → (p : non-empty t) → B (root t p) → (A ▹ B) *
branch (sup (inj₁ x)) p  b = proj₂ x b
branch (sup (inj₂ y)) () b

_∞ : Poly → Set
P ∞ = Σ[ u ∈ PreApprox P ] (non-empty (u 0) × converges u)

-- | If u₀ is non-empty and u converges, then uₙ is non-empty for every n.
ne₀→all-ne : ∀{P} →
(u : PreApprox P) → non-empty (u 0) → converges u →
∀ n → non-empty (u n)
ne₀→all-ne u ne₀ conv zero = ne₀
ne₀→all-ne {A ▹ B} u ne₀ conv (suc n)
with u (suc n)   | u n          | conv n
... | sup (inj₁ x) | y            | z       = tt
... | sup (inj₂ x) | sup (inj₁ y) | ()
... | sup (inj₂ x) | sup (inj₂ y) | ()

-- | If t₂ extends t₁ and they are both non-empty, then they have the same
-- root labels.
ext→root≡ : ∀{A B} →
(t₁ t₂ : (A ▹ B) *)
(ne₁ : non-empty t₁)
(ne₂ : non-empty t₂)
(e : t₂ extends t₁) →
root t₁ ne₁ ≡ root t₂ ne₂
ext→root≡ (sup (inj₁ (a , α))) (sup (inj₁ (b , β))) ne₁ ne₂ e = sym (proj₁ e)
ext→root≡ (sup (inj₁ x))       (sup (inj₂ tt))      ne₁ ()  e
ext→root≡ (sup (inj₂ tt))      (sup (inj₁ y))       ()  ne₂ e
ext→root≡ (sup (inj₂ tt))      (sup (inj₂ tt))      ne₁ ne₂ ()

-- | The roots of the first two trees in a converging sequence have the same
-- label. This is used as base case in conv→suc-roots≡.
conv→roots-0-1≡ : ∀{A B} →
(u : PreApprox (A ▹ B))
(ne : non-empty (u 0))
(c : converges u) →
root (u 0) ne ≡ root (u 1) (ne₀→all-ne u ne c 1)
conv→roots-0-1≡ {A} {B} u ne c
= ext→root≡ (u 0) (u 1) ne (ne₀→all-ne u ne c 1) (c 0)

-- | The roots of two successive trees in a converging sequence have the same
-- label. We use this in the induction step in conv→roots≡.
conv→suc-roots≡ : ∀{A B} →
(u : PreApprox (A ▹ B))
(ne : non-empty (u 0))
(c : converges u) →
∀ n → root (u n) (ne₀→all-ne u ne c n)
≡ root (u (1 + n)) (ne₀→all-ne u ne c (1 + n))
conv→suc-roots≡ u ne c zero = conv→roots-0-1≡ u ne c
conv→suc-roots≡ u ne c (suc n)
= ext→root≡ (u (1 + n))
(u (2 + n))
(ne₀→all-ne u ne c (1 + n))
(ne₀→all-ne u ne c (2 + n))
(c (suc n))

-- | The root of any tree in a converging sequence has the same label as
-- the root of the first tree.
conv→roots≡ : ∀{A B} →
(u : PreApprox (A ▹ B))
(ne : non-empty (u 0))
(c : converges u) →
∀ n → root (u 0) ne ≡ root (u n) (ne₀→all-ne u ne c n)
conv→roots≡ u ne c zero = refl
conv→roots≡ u ne c (suc n)
= trans (conv→roots≡ u ne c n) (conv→suc-roots≡ u ne c n)

-- | All trees in a converging sequence have the same root label.
conv→all-roots≡ : ∀{A B} →
(u : PreApprox (A ▹ B))
(ne : non-empty (u 0))
(c : converges u) →
∀ n m → root (u n) (ne₀→all-ne u ne c n) ≡ root (u m) (ne₀→all-ne u ne c m)
conv→all-roots≡ u ne c n m =
begin
root (u n) (ne₀→all-ne u ne c n)
≡⟨ sym (conv→roots≡ u ne c n) ⟩
root (u 0) ne
≡⟨ conv→roots≡ u ne c m ⟩
root (u m) (ne₀→all-ne u ne c m)
∎

-- | If t₁ is non-empty and t₂ extends t₁, then all branches of t₂ are
-- non-empty. This shows that _extends_ defines a strict order.
conv+ne₀→ne-branch : ∀{A B} →
(t₁ t₂ : (A ▹ B)*)
(ne₁ : non-empty t₁)
(ne₂ : non-empty t₂)
(e : t₂ extends t₁) →
∀ b → non-empty (branch t₂ ne₂ b)
conv+ne₀→ne-branch (sup (inj₁ (a , α))) (sup (inj₁ (.a , β))) ne₁ ne₂ (refl , q) b
with α b          | β b           | q b
... | sup (inj₁ x)  | sup (inj₁ y)  | z  = tt
... | sup (inj₁ x)  | sup (inj₂ tt) | ()
... | sup (inj₂ tt) | sup (inj₁ y)  | z  = tt
... | sup (inj₂ tt) | sup (inj₂ tt) | ()
conv+ne₀→ne-branch (sup (inj₁ x)) (sup (inj₂ tt)) ne₁ () e b
conv+ne₀→ne-branch (sup (inj₂ tt)) t₂             () ne₂ e b

-- | If t₂ extends t₂, then all branches of t₂ extend the corresponding one
-- of t₁.
conv→branch-conv : ∀{A B} →
(t₁ t₂ : (A ▹ B)*)
(ne₁ : non-empty t₁)
(ne₂ : non-empty t₂)
(e : t₂ extends t₁) →
∀ b b' → b ≅ b' →
(branch t₂ ne₂ b) extends
(branch t₁ ne₁ b')
conv→branch-conv (sup (inj₁ (a'' , α)))
(sup (inj₁ (.a'' , β))) ne₁ ne₂ (refl , t) b .b refl = t b
conv→branch-conv (sup (inj₁ x)) (sup (inj₂ tt)) ne₁ () e b b' p
conv→branch-conv (sup (inj₂ tt)) (sup y) () ne₂ e b b' p

-- | Technical lemma:
-- The equality proofs for substitution by root equality are the same.
root-conv≡ : ∀ {A B} →
(u : PreApprox (A ▹ B))
(ne : non-empty (u 0))
(c : converges u) →
∀ b n → (subst B (conv→roots≡ u ne c (1 + n)) b)
≅ (subst B (conv→roots≡ u ne c n) b)
root-conv≡ {A} {B} u ne c b n =
HE.trans (HE.≡-subst-removable B (conv→roots≡ u ne c (1 + n)) b)
(HE.sym (HE.≡-subst-removable B (conv→roots≡ u ne c n) b))

-- | Coalgebra structure on P∞.
out : ∀{P} → P ∞ → ⟦ P ⟧ (P ∞)
out {A ▹ B} (u , ne , conv) = (a , α)
where
a : A
a = root (u 0) ne

u' : B a → PreApprox (A ▹ B)
u' b n = branch (u (1 + n))
(ne₀→all-ne u ne conv (1 + n))
(subst B (conv→roots≡ u ne conv (1 + n)) b)

ne' : (b : B a) → non-empty (u' b 0)
ne' b =
conv+ne₀→ne-branch (u 0) (u 1) ne
(ne₀→all-ne u ne conv 1)
(conv 0)
(subst B (conv→roots≡ u ne conv 1) b)

conv' : (b : B a) → converges (u' b)
conv' b n = conv→branch-conv (u (1 + n)) (u (2 + n))
(ne₀→all-ne u ne conv (1 + n))
(ne₀→all-ne u ne conv (2 + n))
(conv (1 + n))
(subst B (conv→roots≡ u ne conv (2 + n)) b)
(subst B (conv→roots≡ u ne conv (1 + n)) b)
(root-conv≡ u ne conv b (suc n))

α : B a → (A ▹ B) ∞
α b = (u' b , ne' b , conv' b)

-- | Get root label.
root∞ : ∀{A B} → (A ▹ B) ∞ → A
root∞ t = proj₁ (out t)

-- | Get branches below root.
branch∞ : ∀{A B} → (t : (A ▹ B) ∞) → B (root∞ t) → (A ▹ B) ∞
branch∞ t = proj₂ (out t)

-- | Corecursion principle for P∞.
corec : ∀{P} {C : Set} → (C → ⟦ P ⟧ C) → (C → P ∞)
corec {A ▹ B} {C} f c = (u , ne , conv)
where
u-aux : C → PreApprox (A ▹ B)
u-aux c zero =
let (a , α) = f c
in sup (inj₁ (a , (λ x → sup (inj₂ tt))))
u-aux c (suc n) =
let (a , α) = f c
in sup (inj₁ (a , (λ b → u-aux (α b) n)))

u : PreApprox (A ▹ B)
u = u-aux c

ne : non-empty (u 0)
ne = tt

conv-aux : (c : C) → converges (u-aux c)
conv-aux c zero    = (refl , (λ b → tt))
conv-aux c (suc n) = (refl , (λ b → conv-aux (proj₂ (f c) b) n))

conv : converges u
conv = conv-aux c