Capítulo II Marco Teórico
2.5 Los 14 puntos de Deming
Initial Algebras and Final Coalgebras via Factorization Systems
Remarks and notation: We assume the reader have elementary knowledge about ordinal num-bers and transfinite induction. Ordinal numnum-bers will be ranged over by α, β, γ, κ, . . . and the class of ordinal numbers will be denoted by Ord. By a little abuse of notation, Ord will also denote the category of ordinal numbers with arrows α → β iff α ≤ β, and by α we will also denote the full sub-category of Ord of all ordinal numbers less or equal than α. For a functor F : Ord → C, we define the functor F α : α → C as the composite ι ◦ F , where ι : α ,→ Ord is the inclusion functor.
4.1 Initial and Final Sequences
In this section we recall the definition of initial and final sequences for an endofunctor. These structures, which are dual to each other, were first explicitly given by Barr [13] in order to inves-tigate the relationship between the initial and final coalgebra, and in order to provide sufficient conditions for a functor to be algebraically compact (i.e., when the unique arrow from the initial algebra to the final coalgebra is an isomorphism). These sequences have been successfully used in order to infer properties about the initial algebra and final coalgebra and, moreover, to pro-vide sufficient conditions for a functor to admit such initial and terminal objects (see for example, Barr [14], Ad`amek [9, 8, 7, 6], Smyth and Plotkin [77], and Worrell [94, 92]).
4.1.1 Initial Sequences Leads to Initial Algebras
In this section we recall the definition and the main results about initial sequences. The expo-sition is slightly nonstandard and puts the light on some results which were implicit in previous presentations (e.g., in [13, 9]).
Definition 4.1.1 (Initial Sequence) Let C be a category with initial object 0 and colimits of ordinal-indexed diagrams, and assume T : C → C be an endofunctor on C. The initial sequence of T is a limit-preserving functor A : Ord → C such that, for all ordinals γ ≤ β,
i. A(0) = 0;
ii. A(β+1) = T A(β);
iii. A(γ+1 → β+1) = T A(γ → β).
The initial sequence is said to stabilize at some α ∈ Ord, if A(α → α+1) is an isomorphism.
Note that, for all limit ordinals β, (γ → β)γ<β is a colimit in Ord, and since A preserves colimits, (A(γ → β))γ<β is a colimit in C for the diagram Aα.
Historical note. In [13], Barr gave a more direct construction of the initial sequence as an ordinal-indexed sequence of objects (Aβ)β∈Ordwith arrows (fβγ: Aγ → Aβ)γ≤β, uniquely defined by the following conditions, for δ ≤ γ ≤ β:
(IS-1) Aβ+1= T Aβ; (IS-2) fβ+1γ+1= T fβγ; (IS-3) fββ= idAβ: (IS-4) fβγ◦ fγδ= fβδ;
(IS-5) if β is a limit ordinal, the cocone (fβγ: Aγ→ Aβ)γ<β is a colimit.
The sequence is defined by transfinite induction on α ∈ Ord, defining Aα and fαβ: Aβ → Aα, for all β ≤ α, and checking at each stage that conditions (IS-1)–(IS-5) hold for the portion of sequence already defined.
First step: Let α = 0. The sequence begins with A0= 0 and f00= idA0.
Isolated step: Let α = α0+ 1 and assume by inductive hypothesis that Aα0 and the arrows fαγ0
have been given and satisfy (IS-1)–(IS-5), for all γ < α0. We define Aα = T Aα0, and the arrows fαβ are defined by induction on β ≤ α. We distinguish three cases. If β = α, then we define fαβ= idAα. If β is a successor ordinal, say β = β0+ 1, then we define fαβ= T fαβ00. If β is a limit ordinal, (fβγ: Aγ → Aβ)γ<β is a colimit, by (IS-5). By inductive hypothesis on β, we can consider (fαγ: Aγ → Aα)γ<β, which turns out to be a compatible cocone by (IS-2) and (IS-4). Now, we define fαβ as the unique map factorizing the cocone.
Limit step: Let α be a limit ordinal. By inductive hypothesis we are given all arrows fβγ, for γ ≤ β < α, which, indeed, form a chain. We define Aα to be the colimit of this chain and (fαβ: Aβ→ Aα)β<α are its injections.
An initial sequence constitutes an ordinal-index diagram A : Ord → C in C, and it turns out that for any T -algebra (X, h) one can define a cocone over it.
Lemma 4.1.2 ([13]) Assume A : Ord → C be the initial sequence of T , and (X, h) be a T -algebra. Then, there is a cocone (hα: A(α) → X)α∈Ord from A to X, such that, for any ordinal α, the following diagram commutes:
T X X
A(α+1) A(α)
h
T hα hα
Proof. We define hα: A(α) → X by transfinite induction on α ∈ Ord, checking at each step that hα= h ◦ T hα◦ A(α → α+1) and hβ= hα◦ A(β → α), for all β ≤ α.
We define h0: A(0) = 0 → X as the unique arrow from the initial object, therefore, by unique-ness, h0 = h ◦ T h0◦ A(0 → 1) holds. Assume by inductive hypothesis that, for all β ≤ α, the arrows hβ are given, and they are such that hβ= h ◦ T hβ◦ A(β → β+1) and hβ= hα◦ A(β → α).
define hα+1, h ◦ T hα. From this it follows that
hα+1= h ◦ T hα (by def. hα+1)
= h ◦ T (h ◦ T hα◦ A(α → α+1)) (by inductive hp.)
= h ◦ T (h ◦ T hα) ◦ T A(α → α+1) (by funct. T )
= h ◦ T hα+1◦ T A(α → α+1) (by def. hα+1)
= h ◦ T hα+1◦ A(α+1 → α+2) ; (by def. A)
hα+1= hα+1◦ A(α+1 → α+1), since A(α+1 → α+1) = idA(α+1), and, for all β ≤ α,
hβ= hα◦ A(β → α) (by inductive hp.)
= h ◦ T hα◦ A(α → α+1) ◦ A(β → α) (by inductive hp.)
= h ◦ T hα◦ A(β → α+1) (by funct. A)
= hα+1◦ A(β → α+1) . (by def. hα+1)
Let α be a limit ordinal, and assume by inductive hypothesis that, for all β < α, the arrows hβ are given, and they are such that hβ= h ◦ T hβ◦ A(β → β+1) and hβ= hα◦ A(β → α). By definition of initial sequence, (A(β → α))β<αis a colimit over Aα, and by (the second part of the) inductive hypothesis we have that (hβ: A(β) → X)β<α is a cocone over Aα. From this, we define hαas the unique arrow such that hβ = hα◦ A(β → α), for all β < α. Trivially, hα= hα◦ A(α → α), since A(α → α) = idA(α). We have that, for all β < α,
hβ = h ◦ T hβ◦ A(β → β+1) (by inductive hp.)
= h ◦ T (hα◦ A(β → α)) ◦ A(β → β+1) (by inductive hp.)
= h ◦ T hα◦ T A(β → α) ◦ A(β → β+1) (by funct. T )
= h ◦ T hα◦ A(β+1 → α+1) ◦ A(β → β+1) (by def. A)
= h ◦ T hα◦ A(β → α+1) (by funct. A)
= h ◦ T hα◦ A(α → α+1) ◦ A(β → α) . (by funct. A) By uniqueness of the colimiting arrow, the above proves that hα= h ◦ T hα◦ A(α → α+1).
Definition 4.1.3 (Initial co-projection) Let T : C → C be a functor, (X, h) be a T -algebra, and A : Ord → C be the initial sequence of T . An arrow k : A(α) → X is an initial co-projection at α for (X, h), if the following diagram commutes:
T X X
A(α+1) A(α)
h
T k k
Given a T -algebra (X, h), Lemma 4.1.2 states that each arrow in the (canonical) cocone (hα: A(α) → X)α∈Ord for (X, h) over the initial sequence A, is an initial co-projection. It turns out that, for any ordinal α and T -algebra (X, h), initial co-projections at α for (X, h) are unique.
Lemma 4.1.4 (Uniqueness of initial co-projections) Let T : C → C be a functor, (X, h) be a T -algebra, A : Ord → C be the initial sequence of T , and (hα: A(α) → X)α∈Ord be the cocone for (X, h) over A given by Lemma 4.1.9. If k : A(α) → X is an initial co-projection at α for (X, h), then k = hα.
Proof. For β ≤ α, let kβ = k ◦ A(β → α). We will show by transfinite induction on β, that kβ = hβ, for all β ≤ α. Certainly k0 = h0 since their domain is the initial object. Assume by inductive hypothesis that kβ = hβ, then we have
kβ+1= k ◦ A(β+1 → α) (by def. kβ+1)
= h ◦ T k ◦ A(α → α+1) ◦ A(β+1 → α) (by hp.)
= h ◦ T k ◦ A(β+1 → α+1) (by funct. A)
= h ◦ T k ◦ T A(β → α) (by def. A)
= h ◦ T (k ◦ A(β → α)) (by funct. T )
= h ◦ T (kβ) (by def. kβ)
= h ◦ T (hβ) (by inductive hp.)
= hβ+1. (by def. hβ+1)
Assume β is a limit ordinal and, by inductive hypothesis, that kγ = hγ, for every γ < β. By definition of kγ and by the fact that (hγ: A(γ) → X)γ≤α is a cocone over Aα, we have that kγ = k ◦ A(γ → α) and hγ = hα◦ A(γ → α). Therefore, by inductive hypothesis, we obtain k ◦ A(γ → α) = hα◦ A(γ → α). By functoriality of A, this implies that
k ◦ A(β → α) ◦ A(γ → β) = hα◦ A(β → α) ◦ A(γ → β) .
This holds for all γ < β, and since (A(γ → β))γ<β is a colimit over Aβ, by uniqueness of the colimiting arrow, we have k ◦ A(β → α) = hα◦ A(β → α). Thus, by definition of kβ and compatibility of the cocone, we conclude that kβ= hβ.
Remark 4.1.5 Uniqueness of initial co-projections was not explicitly recognized in [13]. Indeed, in [13, Theorem 1.2], uniqueness for the arrows in initial co-projections was only proved assuming their targets are isomorphisms in the initial sequence. Lemma 4.1.4 extends this result and drop the assumption of existence of isomorphisms in the initial sequence. Thank to this lemma, the proof of [13, Theorem 1.2] can be made easier (a restatement of it is given in Theorem 4.1.6).
Initial sequences gives sufficient conditions for the existence of initial algebras. Indeed, if the initial sequence of T stabilizes at some ordinal α, then there exists an initial T -algebra.
Theorem 4.1.6 ([13]) Let α be an ordinal number, and suppose the initial sequence A of T stabilizes at α, then (A(α), A(α → α+1)−1) is an initial T -algebra.
Proof. Let (X, h) be a T -algebra. By Lemma 4.1.2, there exists a cocone (hβ: A(β) → X)β∈Ord
over A. Moreover, hα= h ◦ T hα◦ A(α → α+1). Since, by hypothesis, A(α → α+1) is an isomor-phism, hα is a homomorphism of T -algebras from (A(α), A(α → α+1)−1) to (X, h). Therefore, (A(α), A(α → α+1)−1) is weakly terminal. Uniqueness follows by Lemma 4.1.4.
Remark 4.1.7 One may be tempted to think that if there exists an initial algebra for some endofunctor T , then the initial sequence must lead to it. This is not true in general, even for categories which are complete and cocomplete, as noted by Barr [13].
For example, let C be the category whose objects are all ordinals, ordered by inclusion, plus one more object > greater than all the ordinals. Then C is complete and cocomplete. Let T : C → C be the endofunctor defined by T (α) = α+1, when α is an ordinal, and T (>) = >. Then, the initial sequence of T consists of all the ordinals and never stabilizes. There is only one T -algebra, namely (>, id>), and it is initial.