Introduction
Interpretations between first-order theories provide a fundamental tool for understanding relationships between mathematical structures and theories. In the time since Tarski’s introduction of his relative interpretations in Undecidable Theories (1953), various generalizations of the classical notion have arisen [1 ] . These notions, and their philosophical applications, have received significant attention and are relatively well understood.1 The goal of this paper is to systematically study the far less thoroughly investigated piecewise interpretations , introduced by Visser (2004, 2021) in a single-sorted setting. As opposed to ordinary interpretations, these allow one to interpret formulae using disjoint unions of definable sets. The theory developed in the paper is quite general, encompassing ordinary interpretations. It also introduces a natural notion of equivalence of theories, namely piecewise bi-interpretability, and remedies certain defects of the ordinary notion of interpretation, to do with the reduction of many-sorted to single-sorted logic.
The main focus will be the development of the mathematical theory of piecewise interpretations in the many-sorted setting. In Section 2, we define many-sorted piecewise translations and interpretations, discuss the projective completion of an affine plane as an example, develop semantic duals of translations, and introduce morphisms between piecewise translations (i i i -morphisms) and their semantics. In Section 3, we show that theories, piecewise interpretations, and i i i -morphisms form a bicategory T h p w \mathrm{Th}^{\mathrm{pw}} Th pw by semantic means, using the induced semantic duals of interpretations. In particular, we show that the developed notion of composition of piecewise interpretations is associative (up to coherent i i i -isomorphism), thereby answering a small question due to Visser [24 ] . In Section 4, we present a (pseudo)functorial reduction of many-sorted to single-sorted logic, based on the familiar idea that one can replace sorts with sortal predicates. This reduction will take the form of a biequivalence between the developed bicategory of theories (with finitely many sorts) and piecewise interpretations and the single-sorted analogue of that bicategory.
Our decision to place the study of interpretations against the background of category theory follows Visser [24 ] , while our approach to interpretations in many-sorted logic is inspired by Halvorson [8 ] . Meadows [15 ] pursues a structurally similar but more general and semantic project, studying functorial comparisons between theories mediated through their categories of models, organized through 2-category theory. Taking a mostly syntactic route, D’Arienzo et al. [5 ] systematically develop a bicategory of interpretations in coherent logic, establishing a tight relationship to categorical logic. Beyond Visser, piecewise interpretations have been studied independently by Li [13 ] in his master’s thesis: he presents a definition extending Visser’s against a many-sorted background, connecting piecewise interpretations to Meadows’ Morita interpretations and providing a notion of composition. The definition we have chosen differs in presentation from his and Visser’s, taking a cue from categorical logic, and we develop the theory further by placing our notion of composition into a bicategorical framework.
Notation and Preliminaries
Given a set X X X , we denote by X ∗ X^{*} X ∗ the set of finite sequences of elements of that set. Elements of X ∗ X^{*} X ∗ will be denoted by x ˉ = ( x 1 , … , x n ) \bar{x}=(x_{1},\ldots,x_{n}) x ˉ = ( x 1 , … , x n ) . We also write x ˉ = x 1 × x 2 × ⋯ × x n \bar{x}=x_{1}\times x_{2}\times\cdots\times x_{n} x ˉ = x 1 × x 2 × ⋯ × x n , identifying x ∈ X x\in X x ∈ X with the sequence ( x ) (x) ( x ) , where × \times × denotes concatenation of sequences. Concatenation with the empty sequence ε \varepsilon ε is the identity. For a sequence x ˉ \bar{x} x ˉ , we denote by ∣ x ˉ ∣ ∈ N |\bar{x}|\in\mathbb{N} ∣ x ˉ ∣ ∈ N its length and by x i x_{i} x i its i i i -th component. For n ∈ N n\in\mathbb{N} n ∈ N we write [ n ] : = { 1 , … , n } [n]:=\{1,\ldots,n\} [ n ] := { 1 , … , n } , with [ 0 ] : = ∅ [0]:=\varnothing [ 0 ] := ∅ . Identifying a sequence x ˉ = ( x 1 , … , x n ) \bar{x}=(x_{1},\ldots,x_{n}) x ˉ = ( x 1 , … , x n ) over a set X X X with the function [ n ] → X [n]\to X [ n ] → X sending j j j to x j x_{j} x j and given a function s : [ m ] → [ n ] s:[m]\to[n] s : [ m ] → [ n ] , we write x ˉ s \bar{x}^{s} x ˉ s for the composite x ˉ ∘ s \bar{x}\circ s x ˉ ∘ s , componentwise described as x ˉ s : = ( x s ( 1 ) , … , x s ( m ) ) \bar{x}^{s}:=(x_{s(1)},\ldots,x_{s(m)}) x ˉ s := ( x s ( 1 ) , … , x s ( m ) ) . This is functorial: x ˉ i d = x ˉ \bar{x}^{\mathrm{id}}=\bar{x} x ˉ id = x ˉ , and ( x ˉ s ) t = x ˉ s ∘ t (\bar{x}^{s})^{t}=\bar{x}^{s\circ t} ( x ˉ s ) t = x ˉ s ∘ t for t : [ l ] → [ m ] t:[l]\to[m] t : [ l ] → [ m ] , by associativity of composition.
A finite family over a set X X X is a function p : I p → X p:I_{p}\to X p : I p → X where I p I_{p} I p is a finite, non-empty set, called the index set of the family. We denote by F ( X ) : = { p : I p → X ∣ I p is finite and non-empty } \mathcal{F}(X):=\{p:I_{p}\to X\mid I_{p}\text{ is finite and non-empty}\} F ( X ) := { p : I p → X ∣ I p is finite and non-empty } the collection of all finite families over X X X . We will frequently need to consider F ( X ∗ ) \mathcal{F}(X^{*}) F ( X ∗ ) , the collection of finite families of finite sequences over X X X . Given a finite family of sets p : I p → X p:I_{p}\to X p : I p → X , we denote by ⨆ k ∈ I p p ( k ) \bigsqcup_{k\in I_{p}}p(k) ⨆ k ∈ I p p ( k ) its disjoint union, set-theoretically constructed as the set of ordered pairs ( k , a ) (k,a) ( k , a ) such that a ∈ p ( k ) a\in p(k) a ∈ p ( k ) .
Next, we fix our logical vocabulary. A relational many-sorted first-order signature Σ = ( S , R , V ) \Sigma= (S,\mathcal{R},\mathcal{V}) Σ = ( S , R , V ) consists of a non-empty set S S S of sort symbols, a set R \mathcal{R} R of sorted relation symbols, and a family V = ( V σ ) σ ∈ S \mathcal{V} = (\mathcal{V}_{\sigma})_{\sigma\in S} V = ( V σ ) σ ∈ S of sets of sorted variables.3 To ensure we have enough variables available, we fix an infinite cardinal κ \kappa κ and restrict attention to signatures with ∣ R ∪ S ∣ ≤ κ |\mathcal{R} \cup S| \le\kappa ∣ R ∪ S ∣ ≤ κ and ∣ V σ ∣ = κ |\mathcal{V}_{\sigma}| = \kappa ∣ V σ ∣ = κ for all σ ∈ S \sigma\in S σ ∈ S . Each relation symbol R ∈ R R \in\mathcal{R} R ∈ R has an arity σ ˉ ∈ S ∗ \bar{\sigma} \in S^* σ ˉ ∈ S ∗ , written R : σ ˉ R : \bar{\sigma} R : σ ˉ . Among the logical vocabulary, we count verum ⊤ \top ⊤ and falsum ⊥ \bot ⊥ , taking their arity to be the empty sequence ε \varepsilon ε . For sentential operators, we have { ¬ , ∧ , ∨ , → } \{\neg, \wedge, \vee, \to\} { ¬ , ∧ , ∨ , → } . Moreover, each variable x ∈ V x \in\mathcal{V} x ∈ V (resp. sequence x ˉ ∈ V ∗ \bar{x} \in\mathcal{V}^* x ˉ ∈ V ∗ of variables) comes with an arity σ ∈ S \sigma\in S σ ∈ S (resp. sequence of sorts σ ˉ \bar{\sigma} σ ˉ of the same length), expressed by x : σ x : \sigma x : σ (resp. x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ ). Quantifiers are sorted as well, with one quantifier ∀ σ \forall_{\sigma} ∀ σ for each sort σ ∈ S \sigma\in S σ ∈ S , although we will only explicitly represent this sorting with a subscript where helpful. For a sequence of variables y ˉ = ( y 1 , … , y n ) : σ ˉ \bar{y} = (y_1,\ldots,y_n) : \bar{\sigma} y ˉ = ( y 1 , … , y n ) : σ ˉ we define ∀ y ˉ ψ : = ∀ σ 1 y 1 ⋯ ∀ σ n y n ψ \forall\bar{y}\,\psi:= \forall_{\sigma_1}y_1 \cdots\forall_{\sigma_n}y_n\psi ∀ y ˉ ψ := ∀ σ 1 y 1 ⋯ ∀ σ n y n ψ . In particular, ∀ ε ψ = ψ \forall\varepsilon\psi= \psi ∀ ε ψ = ψ . Officially, we will treat the existential quantifier ∃ \exists ∃ as the dual of the universal quantifier. The signature Σ \Sigma Σ generates a sorted first-order language L Σ \mathcal{L}_{\Sigma} L Σ in the standard inductive fashion: given a relation symbol R : σ ˉ R : \bar{\sigma} R : σ ˉ and a sequence of variables x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ , one can form an atomic formula R ( x ˉ ) R(\bar{x}) R ( x ˉ ) , and general formulae are generated from these via the connectives and the quantifiers, taking sorting into account.4
For a syntactic calculus for many-sorted logic, we direct the reader to Halvorson [8 ] ; starting from a syntactic calculus for single-sorted logic, one only needs to modify the rules for the quantifiers so as to take the sorting into account. We will denote the resulting consequence relation by ⊢ \vdash ⊢ . A theory T T T in a signature Σ \Sigma Σ is a deductively closed set of sentences in the language L Σ \mathcal{L}_{\Sigma} L Σ .
We will also work with formulae-in-context, that is, pairs of a formula φ ∈ L Σ \varphi\in\mathcal{L}_{\Sigma} φ ∈ L Σ and a context x ˉ \bar{x} x ˉ for φ \varphi φ . A context for a formula is a duplicate-free list x ˉ \bar{x} x ˉ of variables which contains all the free variables of φ \varphi φ . The notation ( x ˉ ∣ φ ) (\bar{x} \mid\varphi) ( x ˉ ∣ φ ) denotes the formula-in-context consisting of φ \varphi φ and the context x ˉ \bar{x} x ˉ for φ \varphi φ . Of particular importance will be the canonical context5 x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) of a formula φ \varphi φ , which contains exactly the free variables of φ \varphi φ in order of first occurrence. The sort sequence σ ˉ ( φ ) \bar{\sigma}(\varphi) σ ˉ ( φ ) of x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) is called the arity of φ \varphi φ ; we write φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ .
Given a context y ˉ = ( y 1 , … , y m ) : τ ˉ \bar{y} = (y_1,\ldots,y_m) : \bar{\tau} y ˉ = ( y 1 , … , y m ) : τ ˉ for a formula φ \varphi φ with canonical context x ˉ ( φ ) = ( x 1 , … , x n ) : σ ˉ \bar{x}(\varphi) = (x_1,\ldots,x_n) : \bar{\sigma} x ˉ ( φ ) = ( x 1 , … , x n ) : σ ˉ , there is a unique injection c : [ n ] → [ m ] c : [n] \to[m] c : [ n ] → [ m ] such that x j = y c ( j ) x_j = y_{c(j)} x j = y c ( j ) for 1 ≤ j ≤ n 1 \le j \le n 1 ≤ j ≤ n . We call c c c the context injection for y ˉ \bar{y} y ˉ relative to φ \varphi φ . Given a formula φ \varphi φ and a sequence of variables y ˉ : σ ˉ ( φ ) \bar{y} : \bar{\sigma}(\varphi) y ˉ : σ ˉ ( φ ) , we write φ ( y ˉ ) \varphi(\bar{y}) φ ( y ˉ ) for the result of the capture-avoiding simultaneous substitution of y ˉ \bar{y} y ˉ for x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) in φ \varphi φ .6
Lemma 1.1.1 (Canonical Context under Substitution). Let φ \varphi φ be a formula with canonical context x ˉ = ( x 1 , … , x n ) : σ ˉ \bar{x} = (x_1,\ldots,x_n) : \bar{\sigma} x ˉ = ( x 1 , … , x n ) : σ ˉ , and let y ˉ = ( y 1 , … , y n ) : σ ˉ \bar{y} = (y_1,\ldots,y_n) : \bar{\sigma} y ˉ = ( y 1 , … , y n ) : σ ˉ . Then the canonical context of φ ( y ˉ ) \varphi(\bar{y}) φ ( y ˉ ) is y ˉ \bar{y} y ˉ , the sequence resulting from y ˉ \bar{y} y ˉ by deleting repeated members from right to left.
Proof. See Appendix A.
Let us record some other useful abbreviations: given a sequence of variables x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ , we write ⊤ ‾ x ˉ : = ⋀ i = 1 ∣ x ˉ ∣ ( x i = σ i x i ) \overline{\top}_{\bar{x}} := \bigwedge_{i=1}^{|\bar{x}|}(x_i =_{\sigma_i} x_i) ⊤ x ˉ := ⋀ i = 1 ∣ x ˉ ∣ ( x i = σ i x i ) and ⊥ ‾ x ˉ : = ¬ ⊤ ‾ x ˉ \overline{\bot}_{\bar{x}} := \neg\overline{\top}_{\bar{x}} ⊥ x ˉ := ¬ ⊤ x ˉ , where the empty conjunction is the nullary ⊤ \top ⊤ . We write ⊤ σ ˉ \top_{\bar{\sigma}} ⊤ σ ˉ and ⊥ σ ˉ \bot_{\bar{\sigma}} ⊥ σ ˉ if the choice of duplicate-free variable sequence does not matter. Moreover, given another sequence y ˉ : σ ˉ \bar{y}:\bar{\sigma} y ˉ : σ ˉ , we define identity for sequences of the same arity as x ˉ = σ ˉ y ˉ : = ⊤ x ˉ × y ˉ ∧ ⋀ i = 1 ∣ x ˉ ∣ ( x i = σ i y i ) \bar{x} =_{\bar{\sigma}} \bar{y} := \top_{\bar{x}\times\bar{y}} \land\bigwedge_{i=1}^{|\bar{x}|}(x_i =_{\sigma_i} y_i) x ˉ = σ ˉ y ˉ := ⊤ x ˉ × y ˉ ∧ ⋀ i = 1 ∣ x ˉ ∣ ( x i = σ i y i ) , with the first conjunct serving as padding to get the arities right.
Piecewise Interpretations: Basic Definitions and Semantics
In this section, we will define piecewise translations and interpretations between theories, study their semantic duals, present some basic examples, and introduce the notion of an i i i -morphism between piecewise translations.
Piecewise Translations and Interpretations
The following definition is inspired by Visser’s definition of piecewise translation and extends it to the many-sorted setting, using Halvorson’s approach to many-sorted translations as a basis (Visser, 2004, 2021; Halvorson, 2019). The main idea behind the definition is that sorts can be taken to finite families of sequences of sorts, indexed by the set of pieces for the sort, and relation symbols are taken to finite families of accordingly sorted formulae.
Definition 2.1.1 (Piecewise Reconstrual). Let Σ 1 \Sigma_1 Σ 1 and Σ 2 \Sigma_2 Σ 2 be two first-order signatures. A piecewise reconstrual f : Σ 1 → Σ 2 f:\Sigma_1\to\Sigma_2 f : Σ 1 → Σ 2 is a tuple ( f S , f V , f D , f R ) (f_S,f_V,f_D,f_R) ( f S , f V , f D , f R ) , consisting of the following data: The sort function f S f_S f S assigns each sort σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 a finite family of finite sequences of sorts ( σ f k ) k ∈ P σ (\sigma_f^k)_{k\in P_\sigma} ( σ f k ) k ∈ P σ of S 2 S_2 S 2 with hereditarily finite index set P σ P_\sigma P σ , called the set of pieces for σ \sigma σ .7 The variable function f V f_V f V assigns to each variable x : σ x:\sigma x : σ in V 1 V_1 V 1 a finite family of sequences of variables ( x f k ) k ∈ P σ (x_f^k)_{k\in P_\sigma} ( x f k ) k ∈ P σ in V 2 V_2 V 2 such that x f k : σ f k x_f^k:\sigma_f^k x f k : σ f k is duplicate-free and appropriately sorted, and such that x f k x_f^k x f k and y f l y_f^l y f l are pairwise disjoint if x ≠ y x\ne y x = y (no-overlap). The domain function f D f_D f D assigns to each sort σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 a finite family of formulae ( D σ k ) k ∈ P σ (D_\sigma^k)_{k\in P_\sigma} ( D σ k ) k ∈ P σ of L Σ 2 L_{\Sigma_2} L Σ 2 , such that D σ k : σ f k D_\sigma^k:\sigma_f^k D σ k : σ f k . The relation function f R f_R f R assigns to each relation symbol R : σ ˉ = σ 1 × ⋯ × σ n R:\bar{\sigma}=\sigma_1\times\cdots\times\sigma_n R : σ ˉ = σ 1 × ⋯ × σ n of Σ 1 \Sigma_1 Σ 1 a family of L Σ 2 L_{\Sigma_2} L Σ 2 -formulae ( R σ ˉ k ˉ ) k ˉ ∈ P σ ˉ (R_{\bar{\sigma}}^{\bar{k}})_{\bar{k}\in P_{\bar{\sigma}}} ( R σ ˉ k ˉ ) k ˉ ∈ P σ ˉ , where P σ ˉ : = P σ 1 × ⋯ × P σ n P_{\bar{\sigma}}:=P_{\sigma_1}\times\cdots\times P_{\sigma_n} P σ ˉ := P σ 1 × ⋯ × P σ n and R σ ˉ k ˉ : ( σ 1 ) f k 1 × ⋯ × ( σ n ) f k n R_{\bar{\sigma}}^{\bar{k}}:(\sigma_1)_f^{k_1}\times\cdots\times(\sigma_n)_f^{k_n} R σ ˉ k ˉ : ( σ 1 ) f k 1 × ⋯ × ( σ n ) f k n . Note that we take m m m -fold Cartesian products to consist of m m m -tuples.
We extend the data of a reconstrual to sequences as follows: For a sequence σ ˉ = ( σ 1 , … , σ n ) ∈ S 1 ∗ \bar{\sigma}=(\sigma_1,\ldots,\sigma_n)\in S_1^* σ ˉ = ( σ 1 , … , σ n ) ∈ S 1 ∗ , we set σ ˉ f k ˉ : = σ 1 , f k 1 × ⋯ × σ n , f k n \bar{\sigma}_f^{\bar{k}}:=\sigma_{1,f}^{k_1}\times\cdots\times\sigma_{n,f}^{k_n} σ ˉ f k ˉ := σ 1 , f k 1 × ⋯ × σ n , f k n for k ˉ = ( k 1 , … , k n ) ∈ P σ ˉ \bar{k}=(k_1,\ldots,k_n)\in P_{\bar{\sigma}} k ˉ = ( k 1 , … , k n ) ∈ P σ ˉ . For the empty sequence ε \varepsilon ε , we set P ε = { ε } P_\varepsilon=\{\varepsilon\} P ε = { ε } . Given a sequence x ˉ = ( x 1 , … , x n ) \bar{x}=(x_1,\ldots,x_n) x ˉ = ( x 1 , … , x n ) with x ˉ : σ ˉ \bar{x}:\bar{\sigma} x ˉ : σ ˉ and k ˉ ∈ P σ ˉ \bar{k}\in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ , we set ( x ˉ f k ˉ ) : = ( x 1 ) f k 1 × ⋯ × ( x n ) f k n (\bar{x}_f^{\bar{k}}):=(x_1)_f^{k_1}\times\cdots\times(x_n)_f^{k_n} ( x ˉ f k ˉ ) := ( x 1 ) f k 1 × ⋯ × ( x n ) f k n . In particular, if σ f k = ε \sigma_f^k=\varepsilon σ f k = ε , then x f k : ε x_f^k:\varepsilon x f k : ε . To extend the domain function to sequences of sorts, given σ ˉ = ( σ 1 , … , σ n ) \bar{\sigma}=(\sigma_1,\ldots,\sigma_n) σ ˉ = ( σ 1 , … , σ n ) we take some duplicate-free sequence of variables x ˉ : σ ˉ \bar{x}:\bar{\sigma} x ˉ : σ ˉ and set D σ ˉ k ˉ : = ⋀ i = 1 n D σ i k i ( ( x i ) f k i ) D_{\bar{\sigma}}^{\bar{k}}:=\bigwedge_{i=1}^n D_{\sigma_i}^{k_i}((x_i)_f^{k_i}) D σ ˉ k ˉ := ⋀ i = 1 n D σ i k i (( x i ) f k i ) , to ensure that there is no accidental sharing of free variables. For the empty sequence, we set D ε ε : = ⊤ D_\varepsilon^\varepsilon:=\top D ε ε := ⊤ .
Remark 2.1.2 (Convention for Choice of Variables). From now on, whenever a statement about translations or, later, i i i -morphisms involves the choice of some sequences of variables, we tacitly assume that these sequences are chosen to be fresh, pairwise distinct/disjoint, duplicate-free and such that their images under all pertinent translations and pieces are pairwise disjoint.
Where necessary, we will make the dependence on f f f of the data above explicit by using sub-/superscripts (P σ f , D σ f P^{f}_{\sigma},D^{f}_{\sigma} P σ f , D σ f ). Two comments on this definition: Due to the requirement on arities, replacing a translated formula with a logically equivalent one in a reconstrual will not necessarily result in a reconstrual. Moreover, in practice, one does not always work with a globally fixed variable function, but instead just locally chooses fresh variables when translating any given formula. To partially alleviate the strictness of the definition and dependence on the globally chosen variable function, we will make use of the following equivalence relation on reconstruals:
Definition 2.1.3 (Provable Equivalence of Reconstruals). Let f , g : Σ 1 → Σ 2 f,g:\Sigma_1\to\Sigma_2 f , g : Σ 1 → Σ 2 be two piecewise reconstruals and let T T T be a Σ 2 \Sigma_2 Σ 2 -theory. Then f f f and g g g are T T T -provably equivalent, denoted by f ≡ T g f\equiv_Tg f ≡ T g , iff σ f = σ g \sigma_f=\sigma_g σ f = σ g for all σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 , and for all σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 , k ∈ P σ f k\in P^f_\sigma k ∈ P σ f , and u ˉ : σ f k \bar{u}:\sigma^k_f u ˉ : σ f k , T ⊢ D σ f , k ( u ˉ ) ↔ D σ g , k ( u ˉ ) T\vdash D^{f,k}_\sigma(\bar{u})\leftrightarrow D^{g,k}_\sigma(\bar{u}) T ⊢ D σ f , k ( u ˉ ) ↔ D σ g , k ( u ˉ ) , and, for all relation symbols, including identity, R : σ ˉ R:\bar{\sigma} R : σ ˉ , pieces k ˉ ∈ P σ ˉ f \bar{k}\in P^f_{\bar{\sigma}} k ˉ ∈ P σ ˉ f and u ˉ : σ ˉ k ˉ \bar{u}:\bar{\sigma}^{\bar{k}} u ˉ : σ ˉ k ˉ , T ⊢ D σ ˉ f , k ˉ ( u ˉ ) → ( R f k ˉ ( u ˉ ) ↔ R g k ˉ ( u ˉ ) ) T\vdash D^{f,\bar{k}}_{\bar{\sigma}}(\bar{u})\to\left(R^{\bar{k}}_f(\bar{u})\leftrightarrow R^{\bar{k}}_g(\bar{u})\right) T ⊢ D σ ˉ f , k ˉ ( u ˉ ) → ( R f k ˉ ( u ˉ ) ↔ R g k ˉ ( u ˉ ) ) .
To define a proper notion of translation between the full languages L 1 L_1 L 1 and L 2 L_2 L 2 generated by two signatures Σ 1 \Sigma_1 Σ 1 and Σ 2 \Sigma_2 Σ 2 , we recursively extend any given reconstrual. As an intermediate step, this will require a definition of the translation for formulae-in-context, which then restricts to the desired translation via the canonical context. The recursion reduces the translation of a formula in an arbitrary context to its translation in canonical context and the translation of complex formulae in canonical context to the translation of their components in general contexts; the atomic constituents in canonical context are then handled by the underlying reconstrual.
Definition 2.1.5 (Piecewise Translation). Let f : Σ 1 → Σ 2 f:\Sigma_1\to\Sigma_2 f : Σ 1 → Σ 2 be a piecewise reconstrual. It extends to a function f f f from formulae-in-context of L 1 L_1 L 1 to families of formulae of L 2 L_2 L 2 by the following recursive clauses. If the context x ˉ \bar{x} x ˉ is the canonical context x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) , we write φ \varphi φ for ( x ˉ ( φ ) ∣ φ ) (\bar{x}(\varphi)\mid\varphi) ( x ˉ ( φ ) ∣ φ ) .
(i) Truth Values: ⊤ f ε = ⊤ \top^\varepsilon_f=\top ⊤ f ε = ⊤ and ⊥ f ε = ⊥ \bot^\varepsilon_f=\bot ⊥ f ε = ⊥ .
(ii) Contexts: Given a formula φ \varphi φ with canonical context x ˉ ( φ ) = ( x 1 , … , x n ) : σ ˉ \bar{x}(\varphi)=(x_1,\ldots,x_n):\bar{\sigma} x ˉ ( φ ) = ( x 1 , … , x n ) : σ ˉ and a context y ˉ = ( y 1 , … , y m ) : τ ˉ \bar{y}=(y_1,\ldots,y_m):\bar{\tau} y ˉ = ( y 1 , … , y m ) : τ ˉ for φ \varphi φ with x ˉ ≠ y ˉ \bar{x}\ne\bar{y} x ˉ = y ˉ , consider the unique injection c : [ n ] → [ m ] c:[n]\to[m] c : [ n ] → [ m ] such that x i = y c ( i ) x_i=y_{c(i)} x i = y c ( i ) for all 1 ≤ i ≤ n 1\leq i\leq n 1 ≤ i ≤ n . Via c c c , a piece k ˉ ∈ P τ ˉ \bar{k}\in P_{\bar{\tau}} k ˉ ∈ P τ ˉ induces a piece k ˉ c ∈ P σ ˉ \bar{k}^{c}\in P_{\bar{\sigma}} k ˉ c ∈ P σ ˉ by ( k ˉ c ) i : = k c ( i ) (\bar{k}^{c})_i:=k_{c(i)} ( k ˉ c ) i := k c ( i ) for 1 ≤ i ≤ n 1\leq i\leq n 1 ≤ i ≤ n . We can hence set ( y ˉ ∣ φ ) f k : = φ f k ˉ c (\bar{y}\mid\varphi)^k_f:=\varphi^{\bar{k}^{c}}_f ( y ˉ ∣ φ ) f k := φ f k ˉ c .
(iii) Atomic: Given an atomic formula R ( z ˉ 1 , … , z ˉ m ) R(\bar{z}_1,\ldots,\bar{z}_m) R ( z ˉ 1 , … , z ˉ m ) with z ˉ : τ ˉ \bar{z}:\bar{\tau} z ˉ : τ ˉ and canonical context x ˉ = ( x 1 , … , x n ) : σ ˉ \bar{x}=(x_1,\ldots,x_n):\bar{\sigma} x ˉ = ( x 1 , … , x n ) : σ ˉ , note that there is a unique function d : [ m ] → [ n ] d:[m]\to[n] d : [ m ] → [ n ] such that
x d ( i ) = z i x_{d(i)} = z_i x d ( i ) = z i . Via d d d , a piece k ˉ ∈ P σ ˉ \bar{k} \in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ induces a piece k ˉ d ∈ P τ ˉ \bar{k}^{d} \in P_{\bar{\tau}} k ˉ d ∈ P τ ˉ by ( k ˉ d ) i = k d ( i ) (\bar{k}^{d})_i = k_{d(i)} ( k ˉ d ) i = k d ( i ) for 1 ≤ i ≤ m 1 \leq i \leq m 1 ≤ i ≤ m . We can hence set: R ( z ˉ ) f ˉ k ˉ : = R f ˉ k ˉ d ( z ˉ k ˉ d ) R(\bar{z})_{\bar{f}}^{\bar{k}} := R_{\bar{f}}^{\bar{k}^{d}}(\bar{z}\bar{k}^{d}) R ( z ˉ ) f ˉ k ˉ := R f ˉ k ˉ d ( z ˉ k ˉ d ) .
(iv) Negation: Given a negated formula ¬ φ : σ ˉ \neg\varphi: \bar{\sigma} ¬ φ : σ ˉ and a piece k ˉ ∈ P σ ˉ \bar{k} \in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ , we set ( ¬ φ ) f ˉ k ˉ : = ¬ ( φ f ˉ k ˉ ) (\neg\varphi)_{\bar{f}}^{\bar{k}} := \neg(\varphi_{\bar{f}}^{\bar{k}}) ( ¬ φ ) f ˉ k ˉ := ¬ ( φ f ˉ k ˉ )
(v) Binary: Given a complex formula φ = φ 1 ⋆ φ 2 : σ ˉ \varphi= \varphi_1 \star\varphi_2 : \bar{\sigma} φ = φ 1 ⋆ φ 2 : σ ˉ where ⋆ ∈ { ∧ , ∨ , → } \star\in\{\land, \lor, \to\} ⋆ ∈ { ∧ , ∨ , → } with canonical context x ˉ = x ˉ ( φ 1 ⋆ φ 2 ) \bar{x} = \bar{x}(\varphi_1 \star\varphi_2) x ˉ = x ˉ ( φ 1 ⋆ φ 2 ) and a piece k ˉ ∈ P σ ˉ \bar{k} \in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ , we set φ f ˉ k ˉ : = ( x ˉ ∣ φ 1 ) f ˉ k ˉ ⋆ ( x ˉ ∣ φ 2 ) f ˉ k ˉ \varphi_{\bar{f}}^{\bar{k}} := (\bar{x} \mid\varphi_1)_{\bar{f}}^{\bar{k}} \star(\bar{x} \mid\varphi_2)_{\bar{f}}^{\bar{k}} φ f ˉ k ˉ := ( x ˉ ∣ φ 1 ) f ˉ k ˉ ⋆ ( x ˉ ∣ φ 2 ) f ˉ k ˉ .
(vi) Quantifier: Given a quantified formula φ = ∀ τ y ψ \varphi= \forall_{\tau} y \ \psi φ = ∀ τ y ψ with y : τ y : \tau y : τ and canonical context x ˉ : = x ˉ ( φ ) : σ ˉ \bar{x} := \bar{x}(\varphi) : \bar{\sigma} x ˉ := x ˉ ( φ ) : σ ˉ , and a piece k ˉ ∈ P σ ˉ \bar{k} \in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ , note that x ˉ × y : σ ˉ × τ \bar{x} \times y : \bar{\sigma} \times\tau x ˉ × y : σ ˉ × τ is a context for ψ \psi ψ . We can now set:
φ f ˉ k ˉ : = ⋀ h ∈ P τ ˉ ∀ y f ˉ h ( D τ ˉ h ( y f ˉ h ) → ( x ˉ × y ∣ ψ ) f ˉ k ˉ × h ) (1) \varphi_{\bar{f}}^{\bar{k}} := \bigwedge_{h \in P_{\bar{\tau}}} \forall y_{\bar{f}}^{h} \left(D_{\bar{\tau}}^{h}(y_{\bar{f}}^{h}) \to(\bar{x} \times y \mid\psi)_{\bar{f}}^{\bar{k} \times h}\right) \tag*{(1)} φ f ˉ k ˉ := h ∈ P τ ˉ ⋀ ∀ y f ˉ h ( D τ ˉ h ( y f ˉ h ) → ( x ˉ × y ∣ ψ ) f ˉ k ˉ × h ) (1)
When forming a conjunction indexed by a set of pieces, we will generally assume that the conjuncts are ordered in accordance with the well-order on the hereditarily finite sets.
In general, a (piecewise) translation f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 is a function from L 1 \mathcal{L}_1 L 1 to F ( L 2 ) \mathcal{F}(\mathcal{L}_2) F ( L 2 ) induced in this way by a specific reconstrual from Σ 1 \Sigma_1 Σ 1 to Σ 2 \Sigma_2 Σ 2 . Two piecewise translations f , g f, g f , g are called provably equivalent (with respect to an L 2 \mathcal{L}_2 L 2 -theory T T T ) if the underlying reconstruals are provably equivalent (with respect to T T T ).8
Equivalently, a translation can be seen as taking a formula φ : σ ˉ ∈ L 1 \varphi: \bar{\sigma} \in\mathcal{L}_1 φ : σ ˉ ∈ L 1 and a piece k ˉ ∈ P σ ˉ \bar{k} \in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ to a formula φ f ˉ k ˉ ∈ L 2 \varphi_{\bar{f}}^{\bar{k}} \in\mathcal{L}_2 φ f ˉ k ˉ ∈ L 2 . The behaviour of this translation with respect to free variables and arities is as expected:
Lemma 2.1.6 (Canonical Context of Translations). Let f : Σ 1 → Σ 2 f : \Sigma_1 \to\Sigma_2 f : Σ 1 → Σ 2 be a piecewise reconstrual and let φ \varphi φ be an L 1 \mathcal{L}_1 L 1 -formula with canonical context x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ . Then, for any piece k ˉ ∈ P σ ˉ \bar{k} \in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ , the canonical context of φ f ˉ k ˉ \varphi_{\bar{f}}^{\bar{k}} φ f ˉ k ˉ is x ˉ f ˉ k ˉ \bar{x}_{\bar{f}}^{\bar{k}} x ˉ f ˉ k ˉ . In particular, the formula φ f ˉ k ˉ \varphi_{\bar{f}}^{\bar{k}} φ f ˉ k ˉ has exactly the free variables x ˉ f ˉ k ˉ \bar{x}_{\bar{f}}^{\bar{k}} x ˉ f ˉ k ˉ .
Proof. See Appendix A.
Given an L 1 \mathcal{L}_1 L 1 -sentence φ \varphi φ , its canonical context is the empty sequence ε \varepsilon ε with set of pieces P ε = { ε } P_{\varepsilon} = \{\varepsilon\} P ε = { ε } . By the preceding lemma, the translation φ f ˉ ε \varphi_{\bar{f}}^{\varepsilon} φ f ˉ ε of φ \varphi φ at the unique piece ε \varepsilon ε has canonical context ε \varepsilon ε , and is hence a sentence. More generally, for any context y ˉ : σ ˉ \bar{y} : \bar{\sigma} y ˉ : σ ˉ and k ˉ ∈ P σ ˉ \bar{k} \in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ , ( y ˉ ∣ φ ) f ˉ k ˉ = φ f ˉ k ˉ (\bar{y} \mid\varphi)_{\bar{f}}^{\bar{k}} = \varphi_{\bar{f}}^{\bar{k}} ( y ˉ ∣ φ ) f ˉ k ˉ = φ f ˉ k ˉ . Given a translation f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 we hence obtain a well-defined function between sets of sentences:
f : L 1 S E N T ⟶ L 2 S E N T (2) f : \mathcal{L}_1^{\mathrm{SENT}} \longrightarrow\mathcal{L}_2^{\mathrm{SENT}} \tag*{(2)} f : L 1 SENT ⟶ L 2 SENT (2)
φ ↦ φ f : = φ f ˉ ε (3) \varphi\mapsto\varphi_f := \varphi_{\bar{f}}^{\varepsilon} \tag*{(3)} φ ↦ φ f := φ f ˉ ε (3)
This provides a first indication that piecewise translations can serve as the basis of an appropriate notion of interpretation between theories. In analogy to the ordinary theory of interpretations and relying on piecewise translations mapping sentences to sentences, we define:
Definition 2.1.7 (Piecewise Interpretation). Let S S S be a Σ 1 \Sigma_1 Σ 1 -theory and T T T be a Σ 2 \Sigma_2 Σ 2 -theory. Then a piecewise translation f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 is an interpretation of S S S in T T T iff for all sentences φ ∈ L 1 \varphi\in\mathcal{L}_1 φ ∈ L 1 :
S ⊢ φ ⇒ T ⊢ φ f . (4) S \vdash\varphi\Rightarrow T \vdash\varphi_f. \tag*{(4)} S ⊢ φ ⇒ T ⊢ φ f . (4)
In that case, we write f : S → T f : S \to T f : S → T , and say that T T T interprets S S S (piecewise) via f f f .
If one specializes to the case where the set of pieces has exactly one piece for each sort, one recovers the ordinary notion of interpretation between many-sorted theories, as developed in Halvorson [8 ] , in turn generalizing the usual notion between single-sorted theories.
Identity under a Piecewise Interpretation
An interesting feature of the ordinary notion of interpretation is that it allows for the interpretation of identity by more general equivalence relations. To make room for this feature, we follow Halvorson in treating identity as a non-logical symbol, and hence tacitly think of each L \mathcal{L} L -theory S S S as proving the usual axiom schemata for identity:
S ⊢ ∀ σ x ( x = σ x ) S \vdash\forall_{\sigma}x\,(x =_{\sigma} x) S ⊢ ∀ σ x ( x = σ x )
S ⊢ ∀ x ˉ ∀ y ˉ ( φ ( x ˉ ) ∧ ( x ˉ = σ ˉ y ˉ ) → φ ( y ˉ ) ) S \vdash\forall\bar{x}\forall\bar{y}\,(\varphi(\bar{x}) \land(\bar{x} =_{\bar{\sigma}} \bar{y}) \to\varphi(\bar{y})) S ⊢ ∀ x ˉ ∀ y ˉ ( φ ( x ˉ ) ∧ ( x ˉ = σ ˉ y ˉ ) → φ ( y ˉ ))
for all variables x : σ x : \sigma x : σ , x ˉ , y ˉ : σ ˉ \bar{x},\bar{y} : \bar{\sigma} x ˉ , y ˉ : σ ˉ , and L \mathcal{L} L -formulae φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ . Now, a piecewise interpretation f : S → T f : S \to T f : S → T takes the identity = σ =_{\sigma} = σ on σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 to a family ( = σ ) f : P σ × P σ → L 2 (=_{\sigma})_f : P_{\sigma} \times P_{\sigma} \to\mathcal{L}_2 ( = σ ) f : P σ × P σ → L 2 with ( = σ ) f : σ f k × σ f l (=_{\sigma})_f : \sigma_f^k \times\sigma_f^l ( = σ ) f : σ f k × σ f l . Since f f f preserves theoremhood and the identity axioms are theorems of S S S , the translations of identity introduction and identity elimination are theorems of T T T :
T ⊢ ⋀ k ∈ P σ f ∀ x f k ( D x k → ( x = σ x ) f k ) T \vdash\bigwedge_{k \in P_{\sigma}^{f}} \forall x_f^k\left(D_x^k \to(x =_{\sigma} x)_f^k\right) T ⊢ k ∈ P σ f ⋀ ∀ x f k ( D x k → ( x = σ x ) f k )
T ⊢ ⋀ k ∈ P σ ˉ f ⋀ l ∈ P σ ˉ f ( ∀ x ˉ f k ∀ y ˉ f l ( D x ˉ k ˉ ∧ D y ˉ l ˉ ∧ φ ( x ˉ ) f k ˉ ∧ ( x ˉ = σ ˉ y ˉ ) f k ˉ × l ˉ → φ ( y ˉ ) f l ˉ ) ) T \vdash\bigwedge_{k \in P_{\bar{\sigma}}^{f}}\bigwedge_{l \in P_{\bar{\sigma}}^{f}}\left(\forall\bar{x}_f^k\forall\bar{y}_f^l\left(D_{\bar{x}}^{\bar{k}} \land D_{\bar{y}}^{\bar{l}} \land\varphi(\bar{x})_f^{\bar{k}} \land(\bar{x} =_{\bar{\sigma}} \bar{y})_f^{\bar{k}\times\bar{l}} \to\varphi(\bar{y})_f^{\bar{l}}\right)\right) T ⊢ k ∈ P σ ˉ f ⋀ l ∈ P σ ˉ f ⋀ ( ∀ x ˉ f k ∀ y ˉ f l ( D x ˉ k ˉ ∧ D y ˉ l ˉ ∧ φ ( x ˉ ) f k ˉ ∧ ( x ˉ = σ ˉ y ˉ ) f k ˉ × l ˉ → φ ( y ˉ ) f l ˉ ) )
where we set D x ˉ k ˉ : = D σ ˉ k ˉ ( x ˉ f k ˉ ) D_{\bar{x}}^{\bar{k}} := D_{\bar{\sigma}}^{\bar{k}}(\bar{x}_f^{\bar{k}}) D x ˉ k ˉ := D σ ˉ k ˉ ( x ˉ f k ˉ ) to facilitate readability. In particular, this implies, for all k , l , m ∈ P σ f k,l,m \in P_{\sigma}^{f} k , l , m ∈ P σ f :
T ⊢ ∀ x f k ( D x k → ( x = σ x ) f k ) T \vdash\forall x_f^k\left(D_x^k \to(x =_{\sigma} x)_f^k\right) T ⊢ ∀ x f k ( D x k → ( x = σ x ) f k )
T ⊢ ∀ x f k ∀ y f l ( D x , y k , l → ( ( x = σ y ) f k , l → ( y = σ x ) f l , k ) ) T \vdash\forall x_f^k\forall y_f^l\left(D_{x,y}^{k,l} \to\left((x =_{\sigma} y)_f^{k,l} \to(y =_{\sigma} x)_f^{l,k}\right)\right) T ⊢ ∀ x f k ∀ y f l ( D x , y k , l → ( ( x = σ y ) f k , l → ( y = σ x ) f l , k ) )
T ⊢ ∀ x f k ∀ y f l ∀ z f m ( D x , y , z k , l , m → ( ( x = σ y ) f k , l ∧ ( y = σ z ) f l , m → ( x = σ z ) f k , m ) ) T \vdash\forall x_f^k\forall y_f^l\forall z_f^m\left(D_{x,y,z}^{k,l,m} \to\left((x =_{\sigma} y)_f^{k,l} \land(y =_{\sigma} z)_f^{l,m} \to(x =_{\sigma} z)_f^{k,m}\right)\right) T ⊢ ∀ x f k ∀ y f l ∀ z f m ( D x , y , z k , l , m → ( ( x = σ y ) f k , l ∧ ( y = σ z ) f l , m → ( x = σ z ) f k , m ) )
Reflexivity follows from the translation of identity introduction; in its presence, symmetry and transitivity are jointly equivalent to the translation of identity elimination for identity. Semantically, the family ( = σ ) f (=_{\sigma})_f ( = σ ) f defines an equivalence relation on a disjoint union of extensions of sequences of sorts, as we will see. The summands of this disjoint union will be given by the extensions of the domain formulae of the interpretation. For these, identity introduction implies that S ⊢ ∃ x ( x = σ x ) S \vdash\exists x(x =_{\sigma} x) S ⊢ ∃ x ( x = σ x ) for all sorts σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 , and the translation of this sentence follows from T T T . This translation can easily be seen to be equivalent to the following, stating joint non-emptiness of the domain formulae for f f f for each σ \sigma σ :
T ⊢ ⋁ k ∈ Σ f σ ∃ x f k D σ k ( x f k ) T \vdash\bigvee_{k \in\Sigma_f^\sigma} \exists x_f^k D_\sigma^k(x_f^k) T ⊢ k ∈ Σ f σ ⋁ ∃ x f k D σ k ( x f k )
We now know how piecewise interpretations treat identity. It is worth singling out those translations which behave like interpretations in this respect:
Definition 2.2.1 (Piecewise Translation Respecting Identity). Let f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 be a piecewise translation. We say that f f f respects identity with respect to an L 2 \mathcal{L}_2 L 2 -theory T T T iff f f f and T T T satisfy the non-emptiness of domains condition (Domain) and the identity conditions ( = f -Intro ) (=_{f}\text{-Intro}) ( = f -Intro ) and ( = f -Elim ) (=_{f}\text{-Elim}) ( = f -Elim ) .
Note that a translation respects identity with respect to a theory T T T if it satisfies the domain condition (Domain), ( = f -Intro ) (=_{f}\text{-Intro}) ( = f -Intro ) and ( = f -Elim ) (=_{f}\text{-Elim}) ( = f -Elim ) for atomic formulae, by induction. It follows that if f ≡ T g f \equiv_T g f ≡ T g and f f f respects identity with respect to T T T , then so does g g g .
Such translations are the ones for which we will develop a theory of semantic duals, from which the following preservation of provability result will easily follow, generalizing the result for ordinary interpretations, proved syntactically as Theorem 5.4.7 in Halvorson [8 ] :
Theorem 2.2.2 (Preservation of Provability). Let f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 be a piecewise translation which respects identity with respect to an L 2 \mathcal{L}_2 L 2 -theory T T T . For a set of L 1 \mathcal{L}_1 L 1 -sentences Δ \Delta Δ , denote its image under f f f by Δ f : = { δ f ∣ δ ∈ Δ } \Delta_f := \{\delta_f \mid\delta\in\Delta\} Δ f := { δ f ∣ δ ∈ Δ } , and let φ \varphi φ be an L 1 \mathcal{L}_1 L 1 -sentence. Then:
Δ ⊢ φ ⇒ T ∪ Δ f ⊢ φ f (5) \Delta\vdash\varphi\Rightarrow T \cup\Delta_f \vdash\varphi_f \tag*{(5)} Δ ⊢ φ ⇒ T ∪ Δ f ⊢ φ f (5)
In particular, this entails that one can test whether a translation is an interpretation using a set of axioms for the source theory in question:
Corollary 2.2.3 (Syntactic Criterion for Interpretationhood). Let S S S be an L 1 \mathcal{L}_1 L 1 -theory with a set of axioms Δ \Delta Δ and T T T an L 2 \mathcal{L}_2 L 2 -theory. Then a piecewise translation f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 is an interpretation of S S S in T T T iff f f f respects identity with respect to T T T and T ⊢ δ f T \vdash\delta_f T ⊢ δ f for each δ ∈ Δ \delta\in\Delta δ ∈ Δ .
In particular, this tells us that a translation f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 which respects identity with respect to an L 2 \mathcal{L}_2 L 2 -theory T T T can be seen as an interpretation of the theory of identity L ( Σ 1 ) L(\Sigma_1) L ( Σ 1 ) in the signature Σ 1 \Sigma_1 Σ 1 , axiomatized by ( = − Intro ) (= - \text{Intro}) ( = − Intro ) and ( = − Elim ) (= - \text{Elim}) ( = − Elim ) , into T T T .
Two Basic Examples: n n n Elements and Projective Completions
We are now ready for two concrete examples of piecewise interpretations. The first elementary example shows that piecewise interpretations induce a coarser interpretability relation than ordinary interpretations.
Example 2.3.1 (n n n -Element Theories). Let n ≥ 1 n \ge1 n ≥ 1 be a natural number and let E n E_n E n be the theory in a signature with a single sort σ \sigma σ and no relation symbols, except for the identity, saying that there are exactly n n n elements of sort σ \sigma σ . For n ≥ 2 n \ge2 n ≥ 2 , there is no ordinary interpretation E n → E 1 E_n \to E_1 E n → E 1 , as such an interpretation would need to define a quotient structure with n n n elements inside a structure with a single element.
There is, however, a piecewise interpretation f n : E n → E 1 f_n : E_n \to E_1 f n : E n → E 1 . Set P σ : = [ n ] P_\sigma:= [n] P σ := [ n ] and, for k , l ∈ [ n ] k,l \in[n] k , l ∈ [ n ] , σ f n k : = ε \sigma_{f_n}^{k} := \varepsilon σ f n k := ε and D f n σ , k : = ⊤ D_{f_n}^{\sigma,k} := \top D f n σ , k := ⊤ , with identity being translated as
( x = σ y ) f n k , l : = { ⊤ if k = l , ⊥ if k ≠ l . (6) (x =_\sigma y)_{f_n}^{k,l} := \begin{cases} \top& \text{if } k=l, \\ \bot& \text{if } k\ne l. \end{cases} \tag*{(6)} ( x = σ y ) f n k , l := { ⊤ ⊥ if k = l , if k = l . (6) It is easy to check that this is indeed an interpretation by noting that it respects identity and computing the translation of the single axiom of E n E_n E n . Indeed, f n f_n f n can be used to provide an interpretation of E n E_n E n into any theory.
In the following example, we capture the construction of the projective completion of an affine plane as a piecewise interpretation, following Hartshorne’s presentation of the geometric situation [10 ] .9
Example 2.3.2 (Projective Completion of an Affine Plane). Consider the two-sorted signature Σ = ( { p , l } , { I } ) \Sigma= (\{p,l\}, \{I\}) Σ = ({ p , l } , { I }) with a single non-identity relation symbol I : p × l I : p \times l I : p × l for incidence. We abbreviate parallelism by ℓ 1 ∥ ℓ 2 : = ℓ 1 = l ℓ 2 ∨ ¬ ∃ p ( I ( p , ℓ 1 ) ∧ I ( p , ℓ 2 ) ) \ell_1 \mathbin{\|} \ell_2 := \ell_1 =_l \ell_2 \lor\neg\exists p\,(I(p,\ell_1) \land I(p,\ell_2)) ℓ 1 ∥ ℓ 2 := ℓ 1 = l ℓ 2 ∨ ¬∃ p ( I ( p , ℓ 1 ) ∧ I ( p , ℓ 2 )) .
The theory of affine planes T a f f T_{\mathrm{aff}} T aff over Σ \Sigma Σ has axioms stating that any two distinct points lie on a unique line, that for every point p p p and line ℓ \ell ℓ , there is a unique line through p p p parallel to ℓ \ell ℓ and that there exist three non-collinear points. One can show that T a f f T_{\mathrm{aff}} T aff proves parallelism to be an equivalence relation. The theory of projective planes T p r o j T_{\mathrm{proj}} T proj over the same signature exchanges the axiom about unique parallels with an axiom saying that any two lines meet, and adds that every line contains at least three points.
We construct a piecewise interpretation f : T p r o j → T a f f f : T_{\mathrm{proj}} \to T_{\mathrm{aff}} f : T proj → T aff . The sort function assigns two pieces to each sort:
P p = { 1 , 2 } , p f 1 = ( p ) , p f 2 = ( l ) , P_p = \{1,2\}, \qquad p_f^1 = (p), \qquad p_f^2 = (l), P p = { 1 , 2 } , p f 1 = ( p ) , p f 2 = ( l ) , P l = { 1 , 2 } , l f 1 = ( l ) , l f 2 = ε . P_l = \{1,2\}, \qquad l_f^1 = (l), \qquad l_f^2 = \varepsilon. P l = { 1 , 2 } , l f 1 = ( l ) , l f 2 = ε . Piece 1 of p p p represents the affine points, and piece 2 represents the points at infinity, each given by a line up to parallelism. Piece 1 of l l l represents the affine lines, while piece 2 represents the line at infinity.
All domain formulae give the full domain of the respective sorts: D p 1 = ⊤ p D_p^1 = \top_p D p 1 = ⊤ p , D p 2 = ⊤ l D_p^2 = \top_l D p 2 = ⊤ l , D l 1 = ⊤ l D_l^1 = \top_l D l 1 = ⊤ l , D l 2 = ⊤ ε = ⊤ D_l^2 = \top_\varepsilon= \top D l 2 = ⊤ ε = ⊤ .
The translation of identity on p p p determines when two projective points (possibly from different pieces) are identified:
( = p ) f 1 , 1 ( x 1 , x 2 ) = ( x 1 = p x 2 ) , ( = p ) f 1 , 2 = ⊥ p × l , (=_p)_f^{1,1}(x_1,x_2) = (x_1 =_p x_2), \qquad(=_p)_f^{1,2} = \bot_{p\times l}, ( = p ) f 1 , 1 ( x 1 , x 2 ) = ( x 1 = p x 2 ) , ( = p ) f 1 , 2 = ⊥ p × l , ( = p ) f 2 , 1 = ⊥ l × p , ( = p ) f 2 , 2 ( ℓ 1 , ℓ 2 ) = ( ℓ 1 ∥ ℓ 2 ) . (=_p)_f^{2,1} = \bot_{l\times p}, \qquad(=_p)_f^{2,2}(\ell_1,\ell_2) = (\ell_1 \mathbin{\|} \ell_2). ( = p ) f 2 , 1 = ⊥ l × p , ( = p ) f 2 , 2 ( ℓ 1 , ℓ 2 ) = ( ℓ 1 ∥ ℓ 2 ) .
That is, affine points are identified via ordinary equality, points at infinity are identified when their representing lines are parallel, and no affine point is ever identified with a point at infinity. The required reflexivity, symmetry, and transitivity conditions follow from the fact that T a f f T_{\mathrm{aff}} T aff proves parallelism to be an equivalence relation.
The identity on I \mathrm{I} I is translated as follows:
( = f ) f 1 , 1 ( ℓ 1 , ℓ 2 ) = ( ℓ 1 = l ℓ 2 ) , ( = f ) f 1 , 2 = ⊥ l , (=_{\mathrm{f}})_{f}^{1,1}(\ell_1,\ell_2)=(\ell_1 =_{\mathrm{l}} \ell_2), \qquad(=_{\mathrm{f}})_{f}^{1,2}=\bot_{\mathrm{l}}, ( = f ) f 1 , 1 ( ℓ 1 , ℓ 2 ) = ( ℓ 1 = l ℓ 2 ) , ( = f ) f 1 , 2 = ⊥ l ,
( = f ) f 2 , 1 = ⊥ l , ( = f ) f 2 , 2 = ⊤ . (=_{\mathrm{f}})_{f}^{2,1}=\bot_{\mathrm{l}}, \qquad(=_{\mathrm{f}})_{f}^{2,2}=\top. ( = f ) f 2 , 1 = ⊥ l , ( = f ) f 2 , 2 = ⊤.
Note that ( = f ) f 1 , 2 (=_{\mathrm{f}})_{f}^{1,2} ( = f ) f 1 , 2 and ( = f ) f 2 , 1 (=_{\mathrm{f}})_{f}^{2,1} ( = f ) f 2 , 1 have arity ( l ) (\mathrm{l}) ( l ) , while ( = f ) f 2 , 2 (=_{\mathrm{f}})_{f}^{2,2} ( = f ) f 2 , 2 has arity ε \varepsilon ε and is hence a sentence; it expresses that the line at infinity is equal to itself. Finally, the incidence relation I : p × l I:\mathrm{p}\times\mathrm{l} I : p × l is translated on the set of pieces P p × l = { 1 , 2 } × { 1 , 2 } P_{\mathrm{p}\times\mathrm{l}}=\{1,2\}\times\{1,2\} P p × l = { 1 , 2 } × { 1 , 2 } :
I f 1 , 1 ( p , ℓ ) = I ( p , ℓ ) , I f 1 , 2 = ⊥ p , I_{f}^{1,1}(p,\ell)=I(p,\ell), \qquad I_{f}^{1,2}=\bot_{\mathrm{p}}, I f 1 , 1 ( p , ℓ ) = I ( p , ℓ ) , I f 1 , 2 = ⊥ p ,
I f 2 , 1 ( a , ℓ ) = ( a ∥ ℓ ) , I f 2 , 2 = ⊤ l . I_{f}^{2,1}(a,\ell)=(a\parallel\ell), \qquad I_{f}^{2,2}=\top_{\mathrm{l}}. I f 2 , 1 ( a , ℓ ) = ( a ∥ ℓ ) , I f 2 , 2 = ⊤ l .
In prose: an affine point lies on an affine line according to ordinary incidence; no affine point lies on the line at infinity; a point at infinity [ ℓ 1 ] ∥ [\ell_1]_{\parallel} [ ℓ 1 ] ∥ lies on an affine line ℓ 2 \ell_2 ℓ 2 iff ℓ 1 ∥ ℓ 2 \ell_1\parallel\ell_2 ℓ 1 ∥ ℓ 2 ; and every point at infinity lies on the line at infinity.
To see that this translation respects identity, note that it clearly satisfies the domain condition and ( = f (=_{\mathrm{f}} ( = f - Intro) and ( = f (=_{\mathrm{f}} ( = f - Elim) for the identities. For ( = f (=_{\mathrm{f}} ( = f - Elim) for incidence, the only interesting case is for I f 2 , 1 I_{f}^{2,1} I f 2 , 1 , where it follows from symmetry and transitivity of parallelism. To see that it is indeed an interpretation into T a f f T_{\mathrm{aff}} T aff , one could now go through the axioms, or, instead, use the semantic means introduced in the next subsection.
Three notes on this example: Firstly, one could also produce somewhat more involved interpretations T p r o j → T a f f T_{\mathrm{proj}}\to T_{\mathrm{aff}} T proj → T aff without pieces. Secondly, nothing hinges on allowing the empty sequence as the value of the sort function here; without it, the interpretations would just look a bit clunkier. Thirdly, there is also an interpretation g : T a f f → T p r o j g:T_{\mathrm{aff}}\to T_{\mathrm{proj}} g : T aff → T proj , which, however, uses parameters to single out a line.
The Semantic Dual of a Piecewise Interpretation
In this subsection, we will see that each piecewise interpretation has an induced semantic dual: if f : S → T f:S\to T f : S → T is a piecewise interpretation, there is a corresponding functor f ∗ : Mod ( T ) → Mod ( S ) f^{*}:\operatorname{Mod}(T)\to\operatorname{Mod}(S) f ∗ : Mod ( T ) → Mod ( S ) .10 The right morphisms between models of classical theories for our purposes will be the elementary embeddings between them.
Definition 2.4.1 (Category of Models of a Theory). Let T T T be a classical first-order theory. Then Mod ( T ) \operatorname{Mod}(T) Mod ( T ) denotes the category with objects models M M M of T T T and morphisms elementary embeddings between models.11
At this stage, we need to be a bit more precise about our model theory for many-sorted first-order logic. To define the semantic dual, we will introduce a notion of extension not only for atomic, but also for complex formulae. Let us fix some standard definitions:
Definition 2.4.2 (Σ \Sigma Σ -Structures and Assignments). A Σ \Sigma Σ -structure M M M assigns to each sort σ ∈ S \sigma\in\mathcal{S} σ ∈ S a non-empty set M ( σ ) M(\sigma) M ( σ ) and to each relation symbol R : σ ˉ = σ 1 × ⋯ × σ n R : \bar{\sigma} = \sigma_{1} \times\cdots\times\sigma_{n} R : σ ˉ = σ 1 × ⋯ × σ n a subset M ( R ) M(R) M ( R ) of M ( σ ˉ ) : = M ( σ 1 ) × ⋯ × M ( σ n ) M(\bar{\sigma}) := M(\sigma_{1}) \times\cdots\times M(\sigma_{n}) M ( σ ˉ ) := M ( σ 1 ) × ⋯ × M ( σ n ) . Accordingly, the verum ⊤ : ε \top: \varepsilon ⊤ : ε is assigned the subset { ε } ⊆ M ( ε ) = { ε } \{\varepsilon\} \subseteq M(\varepsilon) = \{\varepsilon\} { ε } ⊆ M ( ε ) = { ε } and the falsum ⊥ : ε \bot: \varepsilon ⊥ : ε is assigned the subset ∅ ⊂ M ( ε ) \varnothing\subset M(\varepsilon) ∅ ⊂ M ( ε ) . We will assume that the identity = σ =_{\sigma} = σ is assigned the extension M ( = σ ) = { ( x , y ) ∈ M ( σ ) × M ( σ ) ∣ x = y } M(=_{\sigma}) = \{(x,y) \in M(\sigma) \times M(\sigma) \mid x=y\} M ( = σ ) = {( x , y ) ∈ M ( σ ) × M ( σ ) ∣ x = y } .
Given a Σ \Sigma Σ -structure M M M , a variable assignment α : V Σ → M \alpha: V_{\Sigma} \to M α : V Σ → M for M M M is a function assigning to each variable x : σ x : \sigma x : σ an element α ( x ) ∈ M ( σ ) \alpha(x) \in M(\sigma) α ( x ) ∈ M ( σ ) . The set of all variable assignments for M M M is denoted by A ( M ) A(M) A ( M ) . For a sequence x ˉ : σ 1 × ⋯ × σ n \bar{x} : \sigma_{1} \times\cdots\times\sigma_{n} x ˉ : σ 1 × ⋯ × σ n , we extend any given assignment α \alpha α by α ( x ˉ ) = ( α ( x 1 ) , … , α ( x n ) ) \alpha(\bar{x}) = (\alpha(x_{1}), \ldots, \alpha(x_{n})) α ( x ˉ ) = ( α ( x 1 ) , … , α ( x n )) . For a duplicate-free x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ , a sequence of elements a ˉ ∈ M ( σ ˉ ) \bar{a} \in M(\bar{\sigma}) a ˉ ∈ M ( σ ˉ ) , and an assignment α \alpha α , we define a variant α [ x ˉ ↦ a ˉ ] \alpha[\bar{x} \mapsto\bar{a}] α [ x ˉ ↦ a ˉ ] by α [ x ˉ ↦ a ˉ ] ( x i ) = a i \alpha[\bar{x} \mapsto\bar{a}](x_i) = a_i α [ x ˉ ↦ a ˉ ] ( x i ) = a i and α \alpha α otherwise.
Satisfaction of formulae with respect to a structure and an assignment, denoted by M , α ⊨ φ M,\alpha\models\varphi M , α ⊨ φ , is then defined recursively in the standard Tarskian fashion. We get the following expected lemma on substitution and satisfaction, whose proof is an easy induction on complexity:
Lemma 2.4.3 (Substitution and Satisfaction). Let M M M be a Σ \Sigma Σ -structure, φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ be a Σ \Sigma Σ -formula with canonical context x ˉ \bar{x} x ˉ , and y ˉ : σ ˉ \bar{y} : \bar{\sigma} y ˉ : σ ˉ . Then, for any variable assignment α \alpha α for M M M , the following holds:
M , α ⊨ φ ( y ˉ ) ⟺ M , α [ x ˉ ↦ α ( y ˉ ) ] ⊨ φ (7) M,\alpha\models\varphi(\bar{y}) \Longleftrightarrow M,\alpha[\bar{x} \mapsto\alpha(\bar{y})] \models\varphi \tag*{(7)} M , α ⊨ φ ( y ˉ ) ⟺ M , α [ x ˉ ↦ α ( y ˉ )] ⊨ φ (7)
Satisfaction of sentences does not depend on the variable assignment chosen, and we say that a Σ \Sigma Σ -structure is a model of a theory T T T , denoted by M ⊨ T M \models T M ⊨ T iff M ⊨ φ M \models\varphi M ⊨ φ for all φ ∈ T \varphi\in T φ ∈ T . Using assignments, we can define a notion of extension for general formulae:
Definition 2.4.4 (Extension of a Formula in a Structure). Given a Σ \Sigma Σ -structure M M M and a Σ \Sigma Σ -formula φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ with canonical context x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ , the extension of φ \varphi φ in M M M is denoted by M ( φ ) M(\varphi) M ( φ ) and defined by
M ( φ ) = { α ( x ˉ ) ∣ α ∈ A ( M ) s.t. M , α ⊨ φ } ⊆ M ( σ ˉ ) . (8) M(\varphi) = \{\alpha(\bar{x}) \mid\alpha\in A(M)\ \text{s.t.}\ M,\alpha\models\varphi\} \subseteq M(\bar{\sigma}). \tag*{(8)} M ( φ ) = { α ( x ˉ ) ∣ α ∈ A ( M ) s.t. M , α ⊨ φ } ⊆ M ( σ ˉ ) . (8)
Intuitively, the extension of a formula consists of all those elements of the given structure satisfying the formula, generalizing from the case of relation symbols. Note that the extension of a sentence is M ( ε ) M(\varepsilon) M ( ε ) if it is true in M M M and ∅ \varnothing ∅ otherwise. With the notion of extension at hand, we can define the semantic dual of a piecewise translation:
Definition 2.4.5 (Semantic Dual of a Piecewise Translation). Let f : L 1 → L 2 f : \mathcal{L}_{1} \to\mathcal{L}_{2} f : L 1 → L 2 be a piecewise translation respecting identity with respect to an L 2 \mathcal{L}_{2} L 2 -theory T T T . Given a model M M M of T T T , we define a Σ 1 \Sigma_{1} Σ 1 -structure f ∗ M f^{*}M f ∗ M as follows: For a sort σ \sigma σ of Σ 1 \Sigma_{1} Σ 1 , the extension of σ \sigma σ in f ∗ M f^{*}M f ∗ M is
f ∗ M ( σ ) : = ( ∐ k ∈ P σ f M ( D σ f , k ) ) / ∼ σ f (9) f^{*}M(\sigma) := \left(\coprod_{k \in P_{\sigma}^{f}} M(D_{\sigma}^{f,k})\right) \big/ {\sim_{\sigma}^{f}} \tag*{(9)} f ∗ M ( σ ) := k ∈ P σ f ∐ M ( D σ f , k ) / ∼ σ f (9) where ( k , a ˉ ) ∼ σ f ( l , b ˉ ) (k,\bar{a}) \sim_{\sigma}^{f} (l,\bar{b}) ( k , a ˉ ) ∼ σ f ( l , b ˉ ) iff a ˉ × b ˉ ∈ M ( ( = σ f ) k , l ) \bar{a} \times\bar{b} \in M((=_{\sigma}^{f})^{k,l}) a ˉ × b ˉ ∈ M (( = σ f ) k , l ) . Note that ∼ σ f \sim_{\sigma}^{f} ∼ σ f is an equivalence relation on a non-empty set by the considerations in Section 2.2, and denote by $\pi_{\sigma}^{f} : \bigsqcup_{k \in P_{\sigma}^{f}} M(D_{\sigma}^{f,k}) \to
f ∗ M ( σ ) f^{*}M(\sigma) f ∗ M ( σ ) the associated projection. To extend this to sequences of sorts, note that for σ ˉ = ( σ 1 , … , σ n ) \bar{\sigma} = (\sigma_{1}, \ldots, \sigma_{n}) σ ˉ = ( σ 1 , … , σ n ) ,
M ( D σ ˉ f , k ˉ ) = { a ˉ 1 × ⋯ × a ˉ n ∣ a ˉ i ∈ M ( D σ i f , k i ) for 1 ≤ i ≤ n } , (10) M(D_{\bar{\sigma}}^{f,\bar{k}}) = \{\bar{a}_{1} \times\cdots\times\bar{a}_{n} \mid\bar{a}_{i} \in M(D_{\sigma_{i}}^{f,k_{i}}) \text{ for } 1 \leq i \leq n\}, \tag*{(10)} M ( D σ ˉ f , k ˉ ) = { a ˉ 1 × ⋯ × a ˉ n ∣ a ˉ i ∈ M ( D σ i f , k i ) for 1 ≤ i ≤ n } , (10)
whence
π σ ˉ f : ⨆ k ˉ ∈ P σ ˉ f M ( D σ ˉ f , k ˉ ) ⟶ ∏ i = 1 ∣ σ ˉ ∣ f ∗ M ( σ i ) (11) \pi_{\bar{\sigma}}^{f} : \bigsqcup_{\bar{k} \in P_{\bar{\sigma}}^{f}} M(D_{\bar{\sigma}}^{f,\bar{k}}) \longrightarrow\prod_{i=1}^{|\bar{\sigma}|} f^{*}M(\sigma_{i}) \tag*{(11)} π σ ˉ f : k ˉ ∈ P σ ˉ f ⨆ M ( D σ ˉ f , k ˉ ) ⟶ i = 1 ∏ ∣ σ ˉ ∣ f ∗ M ( σ i ) (11)
( k ˉ , a ˉ ) ↦ ( π σ 1 f ( k 1 , a ˉ 1 ) , … , π σ n f ( k n , a ˉ n ) ) (12) (\bar{k}, \bar{a}) \mapsto(\pi_{\sigma_{1}}^{f}(k_{1},\bar{a}_{1}),\ldots,\pi_{\sigma_{n}}^{f}(k_{n},\bar{a}_{n})) \tag*{(12)} ( k ˉ , a ˉ ) ↦ ( π σ 1 f ( k 1 , a ˉ 1 ) , … , π σ n f ( k n , a ˉ n )) (12)
is well-defined and surjective, where a ˉ = a ˉ 1 × ⋯ × a ˉ n \bar{a} = \bar{a}_{1} \times\cdots\times\bar{a}_{n} a ˉ = a ˉ 1 × ⋯ × a ˉ n is such that ∣ a ˉ i ∣ = ∣ ( σ i ) f k i ∣ |\bar{a}_{i}| = |(\sigma_{i})_{f}^{k_{i}}| ∣ a ˉ i ∣ = ∣ ( σ i ) f k i ∣ for 1 ≤ i ≤ n 1 \leq i \leq n 1 ≤ i ≤ n . Given a relation symbol R : σ ˉ R : \bar{\sigma} R : σ ˉ of Σ 1 \Sigma_{1} Σ 1 , its extension is defined by
f ∗ M ( R ) : = π σ ˉ f ( ⨆ k ˉ ∈ P σ ˉ f ( M ( R σ ˉ k ˉ ) ∩ M ( D σ ˉ f , k ˉ ) ) ) (13) f^{*}M(R) := \pi_{\bar{\sigma}}^{f}\left(\bigsqcup_{\bar{k} \in P_{\bar{\sigma}}^{f}}\left(M(R_{\bar{\sigma}}^{\bar{k}}) \cap M(D_{\bar{\sigma}}^{f,\bar{k}})\right)\right) \tag*{(13)} f ∗ M ( R ) := π σ ˉ f k ˉ ∈ P σ ˉ f ⨆ ( M ( R σ ˉ k ˉ ) ∩ M ( D σ ˉ f , k ˉ ) ) (13)
The structures defined via the duals of piecewise interpretations are not interpretable structures in the usual sense, as these only allow for the taking of products, subsets, and quotients under equivalence relations, but not for disjoint unions. However, the formation of disjoint unions does not seem to be more problematic than these in any obvious way. In particular, it does not rely on problematic infinitary or nonconstructive resources.
Schematically, equation (2.4.7), suitably understood, also applies to general formulae. To make this precise, we need a notion of variable assignment for a translation for a structure:
Definition 2.4.6 (f f f -Assignment for a Structure M M M ). Let f : L 1 → L 2 f : \mathcal{L}_{1} \to\mathcal{L}_{2} f : L 1 → L 2 be a piecewise translation and let M M M be a Σ 2 \Sigma_{2} Σ 2 -structure. An f f f -assignment for M M M is a tuple ⟨ γ , α ⟩ \langle\gamma, \alpha\rangle ⟨ γ , α ⟩ , where γ \gamma γ is a function assigning to each Σ 1 \Sigma_{1} Σ 1 -variable x : σ x : \sigma x : σ a piece γ ( x ) ∈ P σ f \gamma(x) \in P_{\sigma}^{f} γ ( x ) ∈ P σ f , and α \alpha α is an M M M -assignment satisfying α ( x f γ ( x ) ) ∈ M ( D γ ( x ) f , γ ( x ) ) \alpha(x_{f}^{\gamma(x)}) \in M(D_{\gamma(x)}^{f,\gamma(x)}) α ( x f γ ( x ) ) ∈ M ( D γ ( x ) f , γ ( x ) ) for every Σ 1 \Sigma_{1} Σ 1 -variable x : σ x : \sigma x : σ . If f f f respects identity with respect to a Σ 2 \Sigma_{2} Σ 2 -theory T T T and M ⊨ T M \models T M ⊨ T , we define an f ∗ M f^{*}M f ∗ M -assignment π ⟨ γ , α ⟩ \pi\langle\gamma, \alpha\rangle π ⟨ γ , α ⟩ by x ↦ π σ f ( γ ( x ) , α ( x f γ ( x ) ) ) ∈ f ∗ M ( σ ) x \mapsto\pi_{\sigma}^{f}(\gamma(x), \alpha(x_{f}^{\gamma(x)})) \in f^{*}M(\sigma) x ↦ π σ f ( γ ( x ) , α ( x f γ ( x ) )) ∈ f ∗ M ( σ ) .
Three small notes on this definition: Firstly, if a translation f : L 1 → L 2 f : \mathcal{L}_{1} \to\mathcal{L}_{2} f : L 1 → L 2 respects identity with respect to an L 2 \mathcal{L}_{2} L 2 -theory T T T and M ⊨ T M \models T M ⊨ T , then there is an f f f -assignment for M M M . Secondly, surjectivity of π σ f \pi_{\sigma}^{f} π σ f implies that every assignment for f ∗ M f^{*}M f ∗ M is of the form π ⟨ γ , α ⟩ \pi\langle\gamma, \alpha\rangle π ⟨ γ , α ⟩ for an f f f -assignment ⟨ γ , α ⟩ \langle\gamma, \alpha\rangle ⟨ γ , α ⟩ for M M M . Thirdly, given a duplicate-free x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ , a piece k ˉ ∈ P σ ˉ f \bar{k} \in P_{\bar{\sigma}}^{f} k ˉ ∈ P σ ˉ f and a sequence of elements a ˉ ∈ M ( D σ ˉ f , k ˉ ) \bar{a} \in M(D_{\bar{\sigma}}^{f,\bar{k}}) a ˉ ∈ M ( D σ ˉ f , k ˉ ) , there is always an f f f -assignment ⟨ γ , α ⟩ \langle\gamma, \alpha\rangle ⟨ γ , α ⟩ such that γ ( x ˉ ) = k ˉ \gamma(\bar{x}) = \bar{k} γ ( x ˉ ) = k ˉ and α ( x ˉ f k ˉ ) = a ˉ \alpha(\bar{x}_{f}^{\bar{k}}) = \bar{a} α ( x ˉ f k ˉ ) = a ˉ , since we can extend γ \gamma γ and α \alpha α arbitrarily.
More substantially, we can now formulate a proposition on how satisfaction and extension in the semantic dual are related to satisfaction and extension in the original model. From a technical point of view, this is perhaps the most important proposition in this paper. It also justifies the name semantic dual for f ∗ f^{*} f ∗ :
Proposition 2.4.7 (Satisfaction of Translated Formulae). Let f : L 1 → L 2 f : \mathcal{L}_{1} \to\mathcal{L}_{2} f : L 1 → L 2 be a piecewise translation respecting identity with respect to an L 2 \mathcal{L}_{2} L 2 -theory T T T , and let M M M be a model of T T T .
Then, for every L 1 \mathcal{L}_1 L 1 -formula φ \varphi φ with canonical context x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ and every f f f -assignment ⟨ γ , α ⟩ \langle\gamma, \alpha\rangle ⟨ γ , α ⟩ for M M M ,
M , α ⊨ φ f γ ( x ˉ ) ⟺ f ∗ M , π ⟨ γ , α ⟩ ⊨ φ . (14) M,\alpha\models\varphi_f^{\gamma(\bar{x})} \Longleftrightarrow f^*M,\pi\langle\gamma,\alpha\rangle\models\varphi. \tag*{(14)} M , α ⊨ φ f γ ( x ˉ ) ⟺ f ∗ M , π ⟨ γ , α ⟩ ⊨ φ . (14)
In particular, for every L 1 \mathcal{L}_1 L 1 -sentence φ \varphi φ ,
M ⊨ φ f ⟺ f ∗ M ⊨ φ . (15) M \models\varphi_f \Longleftrightarrow f^*M \models\varphi. \tag*{(15)} M ⊨ φ f ⟺ f ∗ M ⊨ φ . (15)
Proof. We proceed by induction on complexity of φ \varphi φ . A preliminary remark regarding translations in context is expedient: given a Σ 2 \Sigma_2 Σ 2 -structure M M M and an f f f -assignment ⟨ γ , α ⟩ \langle\gamma,\alpha\rangle ⟨ γ , α ⟩ for M M M , let y ˉ : τ ˉ \bar{y} : \bar{\tau} y ˉ : τ ˉ be a context for φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ and denote by x ˉ \bar{x} x ˉ the canonical context of φ \varphi φ . Then, by the context clause for translations,
( y ˉ ∣ φ ) f γ ( y ˉ ) = φ f γ ( x ˉ ) . (16) (\bar{y}\mid\varphi)_f^{\gamma(\bar{y})}=\varphi_f^{\gamma(\bar{x})}. \tag*{(16)} ( y ˉ ∣ φ ) f γ ( y ˉ ) = φ f γ ( x ˉ ) . (16)
The cases for the truth values and negation are immediate. The case for the binary connectives follows from the Tarskian clauses, the binary clauses for translations, and the observation (2.4.10).
Turning to atomic formulae, let φ = R ( z ˉ ) \varphi=R(\bar{z}) φ = R ( z ˉ ) with z ˉ = ( z 1 , … , z m ) : τ ˉ \bar{z}=(z_1,\ldots,z_m):\bar{\tau} z ˉ = ( z 1 , … , z m ) : τ ˉ and canonical context x ˉ = ( x 1 , … , x n ) = z ˉ ∘ : σ ˉ \bar{x}=(x_1,\ldots,x_n)=\bar{z}^{\circ}:\bar{\sigma} x ˉ = ( x 1 , … , x n ) = z ˉ ∘ : σ ˉ , and let d : [ m ] → [ n ] d:[m]\to[n] d : [ m ] → [ n ] be such that x d ( i ) = z i x_{d(i)}=z_i x d ( i ) = z i , as in the atomic clause for translations. Then the atomic clause for translation together with ( γ ( x ˉ ) ) d = γ ( x ˉ d ) = γ ( z ˉ ) (\gamma(\bar{x}))^d=\gamma(\bar{x}^d)=\gamma(\bar{z}) ( γ ( x ˉ ) ) d = γ ( x ˉ d ) = γ ( z ˉ ) imply φ f γ ( x ˉ ) = R f γ ( z ˉ ) ( z ˉ f γ ( z ˉ ) ) \varphi_f^{\gamma(\bar{x})}=R_f^{\gamma(\bar{z})}(\bar{z}_f^{\gamma(\bar{z})}) φ f γ ( x ˉ ) = R f γ ( z ˉ ) ( z ˉ f γ ( z ˉ ) ) , whence Lemma 2.4.3 yields
M , α ⊨ φ f γ ( x ˉ ) ⟺ α ( z ˉ f γ ( z ˉ ) ) ∈ M ( R f γ ( z ˉ ) ) . (17) M,\alpha\models\varphi_f^{\gamma(\bar{x})}\Longleftrightarrow\alpha(\bar{z}_f^{\gamma(\bar{z})})\in M(R_f^{\gamma(\bar{z})}). \tag*{(17)} M , α ⊨ φ f γ ( x ˉ ) ⟺ α ( z ˉ f γ ( z ˉ ) ) ∈ M ( R f γ ( z ˉ ) ) . (17)
On the right-hand side, by definition, we get
f ∗ M , π ⟨ γ , α ⟩ ⊨ R ( z ˉ ) ⟺ ( π ⟨ γ , α ⟩ ( z 1 ) , … , π ⟨ γ , α ⟩ ( z m ) ) ∈ f ∗ M ( R ) . (18) f^*M,\pi\langle\gamma,\alpha\rangle\models R(\bar{z})\Longleftrightarrow\bigl(\pi\langle\gamma,\alpha\rangle(z_1),\ldots,\pi\langle\gamma,\alpha\rangle(z_m)\bigr)\in f^*M(R). \tag*{(18)} f ∗ M , π ⟨ γ , α ⟩ ⊨ R ( z ˉ ) ⟺ ( π ⟨ γ , α ⟩ ( z 1 ) , … , π ⟨ γ , α ⟩ ( z m ) ) ∈ f ∗ M ( R ) . (18)
Now, from left to right, α ( z ˉ f γ ( z ˉ ) ) ∈ M ( R f γ ( z ˉ ) ) \alpha(\bar{z}_f^{\gamma(\bar{z})})\in M(R_f^{\gamma(\bar{z})}) α ( z ˉ f γ ( z ˉ ) ) ∈ M ( R f γ ( z ˉ ) ) by assumption and α ( z ˉ f γ ( z ˉ ) ) ∈ M ( D f τ ˉ , z ˉ ) \alpha(\bar{z}_f^{\gamma(\bar{z})})\in M(D_f^{\bar{\tau},\bar{z}}) α ( z ˉ f γ ( z ˉ ) ) ∈ M ( D f τ ˉ , z ˉ ) by the definition of an f f f -assignment, together with (2.4.4). Therefore, π ⟨ γ , α ⟩ ( z ˉ ) = π f τ ˉ ( γ ( z ˉ ) , α ( z ˉ f γ ( z ˉ ) ) ) ∈ f ∗ M ( R ) \pi\langle\gamma,\alpha\rangle(\bar{z})=\pi_f^{\bar{\tau}}(\gamma(\bar{z}),\alpha(\bar{z}_f^{\gamma(\bar{z})}))\in f^*M(R) π ⟨ γ , α ⟩ ( z ˉ ) = π f τ ˉ ( γ ( z ˉ ) , α ( z ˉ f γ ( z ˉ ) )) ∈ f ∗ M ( R ) , by the definition of the semantic dual (2.4.7). From right to left, choose l ˉ ∈ P τ ˉ f \bar{l}\in P_{\bar{\tau}}^f l ˉ ∈ P τ ˉ f and a ˉ = a ˉ 1 × ⋯ × a ˉ m ∈ M ( R l ˉ f ) ∩ M ( D f τ ˉ , l ˉ ) \bar{a}=\bar{a}_1\times\cdots\times\bar{a}_m\in M(R_{\bar{l}}^f)\cap M(D_f^{\bar{\tau},\bar{l}}) a ˉ = a ˉ 1 × ⋯ × a ˉ m ∈ M ( R l ˉ f ) ∩ M ( D f τ ˉ , l ˉ ) such that π f τ ˉ ( l ˉ , a ˉ ) = ( π ⟨ γ , α ⟩ ( z 1 ) , … , π ⟨ γ , α ⟩ ( z m ) ) \pi_f^{\bar{\tau}}(\bar{l},\bar{a})=\bigl(\pi\langle\gamma,\alpha\rangle(z_1),\ldots,\pi\langle\gamma,\alpha\rangle(z_m)\bigr) π f τ ˉ ( l ˉ , a ˉ ) = ( π ⟨ γ , α ⟩ ( z 1 ) , … , π ⟨ γ , α ⟩ ( z m ) ) ; by construction, a ˉ i × α ( z f γ ( z i ) ) ∈ M ( = τ i l i , γ ( z i ) ) \bar{a}_i\times\alpha(z_{f}^{\gamma(z_i)})\in M(=_{\tau_i}^{l_i,\gamma(z_i)}) a ˉ i × α ( z f γ ( z i ) ) ∈ M ( = τ i l i , γ ( z i ) ) , so instantiating ( = f (=_{f} ( = f -Elim) for R R R at a ˉ \bar{a} a ˉ and α ( z ˉ f γ ( z ˉ ) ) \alpha(\bar{z}_f^{\gamma(\bar{z})}) α ( z ˉ f γ ( z ˉ ) ) yields the right-hand side of (2.4.11).
For the universal quantifier, let φ = ∀ τ y ψ \varphi=\forall_{\tau}y\ \psi φ = ∀ τ y ψ with canonical context x ˉ \bar{x} x ˉ . Applying the clause for the universal quantifier and reducing contexts, we get
φ f γ ( x ˉ ) = ⋀ h ∈ P τ f ∀ y f h ( D τ , f h ( y f h ) → ψ f γ [ y ↦ h ] ( x ˉ ( ψ ) ) ) . (19) \varphi_f^{\gamma(\bar{x})}=\bigwedge_{h\in P_{\tau}^f}\forall y_f^h\left(D_{\tau,f}^{h}(y_f^h)\to\psi_f^{\gamma[y\mapsto h](\bar{x}(\psi))}\right). \tag*{(19)} φ f γ ( x ˉ ) = h ∈ P τ f ⋀ ∀ y f h ( D τ , f h ( y f h ) → ψ f γ [ y ↦ h ] ( x ˉ ( ψ )) ) . (19)
Therefore,
M , α ⊨ φ f γ ( x ˉ ) ⟺ ∀ h ∈ P τ f ∀ a ˉ ∈ M ( D τ f , h ) : M , α [ y f h ↦ a ˉ ] ⊨ ψ f γ [ y ↦ h ] ( x ˉ ( ψ ) ) . (20) M,\alpha\models\varphi_f^{\gamma(\bar{x})}\Longleftrightarrow\forall h\in P_{\tau}^f\forall\bar{a}\in M(D_{\tau}^{f,h}):M,\alpha[y_f^h\mapsto\bar{a}]\models\psi_f^{\gamma[y\mapsto h](\bar{x}(\psi))}. \tag*{(20)} M , α ⊨ φ f γ ( x ˉ ) ⟺ ∀ h ∈ P τ f ∀ a ˉ ∈ M ( D τ f , h ) : M , α [ y f h ↦ a ˉ ] ⊨ ψ f γ [ y ↦ h ] ( x ˉ ( ψ )) . (20)
Applying the induction hypothesis with f f f -assignment ⟨ γ ′ , α ′ ⟩ : = ⟨ γ [ y ↦ h ] , α [ y f h ↦ a ˉ ] ⟩ \langle\gamma',\alpha'\rangle:=\langle\gamma[y\mapsto h],\alpha[y_f^h\mapsto\bar{a}]\rangle ⟨ γ ′ , α ′ ⟩ := ⟨ γ [ y ↦ h ] , α [ y f h ↦ a ˉ ]⟩ , together with surjectivity of π τ f \pi_{\tau}^f π τ f , we see that this is equivalent to
∀ b ∈ f ∗ M ( τ ) : f ∗ M , π ⟨ γ , α ⟩ [ y ↦ b ] ⊨ ψ (21) \forall b\in f^*M(\tau):f^*M,\pi\langle\gamma,\alpha\rangle[y\mapsto b]\models\psi \tag*{(21)} ∀ b ∈ f ∗ M ( τ ) : f ∗ M , π ⟨ γ , α ⟩ [ y ↦ b ] ⊨ ψ (21)
which is equivalent to the desired right-hand side by the Tarskian clause for the quantifier.
We can now fulfil the promise to prove Theorem 2.2.2, on preservation of provability, by semantic means:
Proof of Theorem 2.2.2. Let f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 be a piecewise translation respecting identity with respect to an L 2 \mathcal{L}_2 L 2 -theory T T T . Moreover, let Δ \Delta Δ be a set of L 1 \mathcal{L}_1 L 1 -sentences, and let φ \varphi φ be an L 1 \mathcal{L}_1 L 1 -sentence. Assume that Δ ⊢ φ \Delta\vdash\varphi Δ ⊢ φ . Let M ⊨ T M \models T M ⊨ T such that M ⊨ Δ f M \models\Delta_f M ⊨ Δ f . Then, by Proposition 2.4.7, f ∗ M ⊨ Δ f^*M \models\Delta f ∗ M ⊨ Δ , whence, by soundness, f ∗ M ⊨ φ f^*M \models\varphi f ∗ M ⊨ φ . By Proposition 2.4.7 again, M ⊨ φ f M \models\varphi_f M ⊨ φ f , whence, since M ⊨ T M \models T M ⊨ T was arbitrary, completeness allows us to conclude that T ∪ Δ f ⊢ φ f T \cup\Delta_f \vdash\varphi_f T ∪ Δ f ⊢ φ f .
From the observations immediately following Definition 2.4.6, we get an essentially equivalent reformulation of Proposition 2.4.7 in terms of extensions:
Corollary 2.4.8 (Extensions of Translated Formulae). In the setting of Proposition 2.4.7, the following holds: for every L 1 \mathcal{L}_1 L 1 -formula φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ , every piece k ˉ ∈ P f σ ˉ \bar{k} \in P^{\bar{\sigma}}_f k ˉ ∈ P f σ ˉ and every tuple a ˉ ∈ M ( D f σ ˉ , k ˉ ) \bar{a} \in M(D^{\bar{\sigma},\bar{k}}_f) a ˉ ∈ M ( D f σ ˉ , k ˉ )
a ˉ ∈ M ( φ f k ˉ ) ⟺ π f σ ˉ ( k ˉ , a ˉ ) ∈ f ∗ M ( φ ) . (22) \bar{a} \in M(\varphi_f^{\bar{k}}) \Longleftrightarrow\pi^{\bar{\sigma}}_f(\bar{k},\bar{a}) \in f^*M(\varphi). \tag*{(22)} a ˉ ∈ M ( φ f k ˉ ) ⟺ π f σ ˉ ( k ˉ , a ˉ ) ∈ f ∗ M ( φ ) . (22) In particular, taking images under π f σ ˉ \pi_f^{\bar{\sigma}} π f σ ˉ ,
f ∗ M ( φ ) = π f σ ˉ ( ⋃ k ˉ ∈ P f σ ˉ M ( φ f k ˉ ) ∩ M ( D f σ ˉ , k ˉ ) ) . (23) f^*M(\varphi) = \pi_f^{\bar{\sigma}}\left(\bigcup_{\bar{k}\in P_f^{\bar{\sigma}}} M(\varphi_f^{\bar{k}}) \cap M(D_f^{\bar{\sigma},\bar{k}})\right). \tag*{(23)} f ∗ M ( φ ) = π f σ ˉ k ˉ ∈ P f σ ˉ ⋃ M ( φ f k ˉ ) ∩ M ( D f σ ˉ , k ˉ ) . (23)
Putting this together with adequacy of the semantics allows us to state a semantic characterization of interpretationhood, paving the way for us to understand f ∗ f^* f ∗ as a functor:
Corollary 2.4.9 (Characterization of Interpretations via Model Map). Let f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 be a piecewise translation respecting identity with respect to an L 2 \mathcal{L}_2 L 2 -theory T T T . The translation f f f is a piecewise interpretation f : S → T f : S \to T f : S → T iff for any model M M M of T T T , the L 1 \mathcal{L}_1 L 1 -structure f ∗ M f^*M f ∗ M is a model of S S S . In particular, for a piecewise interpretation, f ∗ f^* f ∗ induces a function f ∗ : Mod ( T ) → Mod ( S ) f^* : \operatorname{Mod}(T) \to\operatorname{Mod}(S) f ∗ : Mod ( T ) → Mod ( S ) .
Proof. By adequacy of the semantics, f f f is an interpretation f : S → T f : S \to T f : S → T iff, for every L 1 \mathcal{L}_1 L 1 -sentence φ \varphi φ made true in every model of S S S , φ f \varphi_f φ f is made true in every model of T T T . By Proposition 2.4.7, this is the case iff for every such sentence φ \varphi φ and model M M M of T T T , f ∗ M ⊨ φ f^*M \models\varphi f ∗ M ⊨ φ . But the latter just says that f ∗ M f^*M f ∗ M is a model of S S S .
Moreover, these elementary properties of the semantic dual are enough to also characterize provable equivalence semantically, thereby establishing invariance of interpretationhood under it.
Proposition 2.4.10 (Provable Equivalence and Model Map). Let f , g : L 1 → L 2 f,g : \mathcal{L}_1 \to\mathcal{L}_2 f , g : L 1 → L 2 be piecewise translations respecting identity with respect to an L 2 \mathcal{L}_2 L 2 -theory T T T , and assume that σ f = σ g \sigma_f = \sigma_g σ f = σ g for every σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 . Then the following are equivalent.
(i) f ≡ T g f \equiv_T g f ≡ T g .
(ii) f ∗ M = g ∗ M f^*M = g^*M f ∗ M = g ∗ M for every model M M M of T T T .
(iii) The domain functions of f f f and g g g are provably equivalent, and for every L 1 \mathcal{L}_1 L 1 -formula φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ , every k ˉ ∈ P f σ ˉ \bar{k} \in P_f^{\bar{\sigma}} k ˉ ∈ P f σ ˉ and every u ˉ : σ ˉ f k ˉ \bar{u} : \bar{\sigma}^{\bar{k}}_f u ˉ : σ ˉ f k ˉ ,
T ⊢ D f σ ˉ , k ˉ ( u ˉ ) → ( φ f k ˉ ( u ˉ ) ↔ φ g k ˉ ( u ˉ ) ) . (24) T \vdash D_f^{\bar{\sigma},\bar{k}}(\bar{u}) \to\left(\varphi_f^{\bar{k}}(\bar{u}) \leftrightarrow\varphi_g^{\bar{k}}(\bar{u})\right). \tag*{(24)} T ⊢ D f σ ˉ , k ˉ ( u ˉ ) → ( φ f k ˉ ( u ˉ ) ↔ φ g k ˉ ( u ˉ ) ) . (24)
Therefore, if f : S → T f : S \to T f : S → T and f ≡ T g f \equiv_T g f ≡ T g , then g : S → T g : S \to T g : S → T .
Proof. (i) to (ii) follows by inspection of Definition 2.4.5, soundness and Lemma 2.4.3. For (ii) to (iii), let M ⊨ T M \models T M ⊨ T and σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 . The elements of f ∗ M ( σ ) f^*M(\sigma) f ∗ M ( σ ) are the ∼ σ f \sim_\sigma^f ∼ σ f -classes, which partition ∐ k ∈ P σ f M ( D σ f , k ) \coprod_{k\in P_\sigma^f} M(D_\sigma^{f,k}) ∐ k ∈ P σ f M ( D σ f , k ) ; hence ⋃ f ∗ M ( σ ) = ∐ k ∈ P σ f M ( D σ f , k ) \bigcup f^*M(\sigma)=\coprod_{k\in P_\sigma^f} M(D_\sigma^{f,k}) ⋃ f ∗ M ( σ ) = ∐ k ∈ P σ f M ( D σ f , k ) , and likewise for g g g . Since P σ f = P σ g P_\sigma^f=P_\sigma^g P σ f = P σ g , the equality f ∗ M ( σ ) = g ∗ M ( σ ) f^*M(\sigma)=g^*M(\sigma) f ∗ M ( σ ) = g ∗ M ( σ ) thus yields ∐ k ∈ P σ f M ( D σ f , k ) = ∐ k ∈ P σ g M ( D σ g , k ) \coprod_{k\in P_\sigma^f} M(D_\sigma^{f,k})=\coprod_{k\in P_\sigma^g} M(D_\sigma^{g,k}) ∐ k ∈ P σ f M ( D σ f , k ) = ∐ k ∈ P σ g M ( D σ g , k ) , and by comparison of the fibres over each k ∈ P σ f k\in P_\sigma^f k ∈ P σ f , M ( D σ f , k ) = M ( D σ g , k ) M(D_\sigma^{f,k})=M(D_\sigma^{g,k}) M ( D σ f , k ) = M ( D σ g , k ) . Since M ⊨ T M \models T M ⊨ T was arbitrary, completeness yields the provable equivalence of the domain formulae. Moreover, f ∗ M ( σ ) f^*M(\sigma) f ∗ M ( σ ) and g ∗ M ( σ ) g^*M(\sigma) g ∗ M ( σ ) are the same partition of the same set, so ∼ σ f = ∼ σ g \sim_\sigma^f=\sim_\sigma^g ∼ σ f = ∼ σ g and π σ f ˉ = π σ g ˉ \pi_\sigma^{\bar{f}}=\pi_\sigma^{\bar{g}} π σ f ˉ = π σ g ˉ . Claim (2.4.18) now follows from (2.4.16) for f f f and g g g , together with completeness. (iii) to (i) merely requires specializing to atomic formulae. The last claim then follows from Corollary 2.4.9. □ \square □
Example 2.4.11 (n n n -Element Theories, continued). Consider the piecewise interpretation f n : E n → E 1 f_n : E_n \to E_1 f n : E n → E 1 from Example 2.3.1. In a model M M M of E 1 E_1 E 1 , the model f n ∗ M f_n^*M f n ∗ M is constructed using the set M ( ⊤ ) = { ε } M(\top)=\{\varepsilon\} M ( ⊤ ) = { ε } , since σ f n k = ε \sigma_{f_n}^k=\varepsilon σ f n k = ε and D σ f n , k = ⊤ D_\sigma^{f_n,k}=\top D σ f n , k = ⊤ . Set c k : = π σ f n ( k , ε ) c_k:=\pi_\sigma^{f_n}(k,\varepsilon) c k := π σ f n ( k , ε ) . Then
f n ∗ M ( σ ) = { c k ∣ k ∈ [ n ] } with f n ∗ M ( = σ ) = { ( c k , c k ) ∣ k ∈ [ n ] } . (25) f_n^*M(\sigma)=\{c_k\mid k\in[n]\}\text{ with }f_n^*M(=_{\sigma})=\{(c_k,c_k)\mid k\in[n]\}. \tag*{(25)} f n ∗ M ( σ ) = { c k ∣ k ∈ [ n ]} with f n ∗ M ( = σ ) = {( c k , c k ) ∣ k ∈ [ n ]} . (25)
This is clearly a model of E n E_n E n .
Example 2.4.12 (Projective Completion, Continued). Consider the piecewise translation f f f from Example 2.3.2, which respects identity with respect to T a f f T_{\mathrm{aff}} T aff . Given a model M ⊨ T a f f M\models T_{\mathrm{aff}} M ⊨ T aff and going through the definition of the semantic dual, we see that we can identify the carriers of f ∗ M f^*M f ∗ M with
f ∗ M ( p ) = M ( p ) ⊔ ( M ( l ) / ∥ M ) , f ∗ M ( l ) = M ( l ) ⊔ { ℓ ∞ } . f^*M(p)=M(p)\mathbin{\sqcup}\bigl(M(l)/\mathord{\parallel_M}\bigr),\qquad f^*M(l)=M(l)\mathbin{\sqcup}\{\ell_\infty\}. f ∗ M ( p ) = M ( p ) ⊔ ( M ( l ) / ∥ M ) , f ∗ M ( l ) = M ( l ) ⊔ { ℓ ∞ } .
where ∥ M \parallel_M ∥ M denotes parallelism in M M M . Here an affine point p p p corresponds to π p f ( 1 , p ) \pi_p^f(1,p) π p f ( 1 , p ) , an equivalence class under parallelism [ a ] ∥ M [a]_{\parallel_M} [ a ] ∥ M to π p f ( 2 , a ) \pi_p^f(2,a) π p f ( 2 , a ) , an affine line ℓ \ell ℓ to π l f ( 1 , ℓ ) \pi_l^f(1,\ell) π l f ( 1 , ℓ ) , and ℓ ∞ \ell_\infty ℓ ∞ to π l f ( 2 , ε ) \pi_l^f(2,\varepsilon) π l f ( 2 , ε ) .
Under these identifications, the interpreted identity relations are ordinary equality, and incidence is given by
f ∗ M ( I ) = { ( p , ℓ ) ∣ ( p , ℓ ) ∈ M ( I ) } ∪ { ( [ a ] ∥ M , ℓ ) ∣ a , ℓ ∈ M ( l ) , a ∥ M ℓ } ∪ { ( [ a ] ∥ M , ℓ ∞ ) ∣ a ∈ M ( l ) } . f^*M(I)=\{(p,\ell)\mid(p,\ell)\in M(I)\}
\cup\{([a]_{\parallel_M},\ell)\mid a,\ell\in M(l),\ a\parallel_M\ell\}
\cup\{([a]_{\parallel_M},\ell_\infty)\mid a\in M(l)\}. f ∗ M ( I ) = {( p , ℓ ) ∣ ( p , ℓ ) ∈ M ( I )} ∪ {([ a ] ∥ M , ℓ ) ∣ a , ℓ ∈ M ( l ) , a ∥ M ℓ } ∪ {([ a ] ∥ M , ℓ ∞ ) ∣ a ∈ M ( l )} .
Thus, in particular, each affine line ℓ \ell ℓ has a corresponding point at infinity [ ℓ ] ∥ M [\ell]_{\parallel_M} [ ℓ ] ∥ M , and the additional line ℓ ∞ \ell_\infty ℓ ∞ contains exactly the points at infinity. This is the usual projective completion of M M M , which is a projective plane [10 ] . As a consequence, f ∗ M ⊨ T p r o j f^*M\models T_{\mathrm{proj}} f ∗ M ⊨ T proj for every M ⊨ T a f f M\models T_{\mathrm{aff}} M ⊨ T aff . Corollary 2.4.9 then entails that f : T p r o j → T a f f f:T_{\mathrm{proj}}\to T_{\mathrm{aff}} f : T proj → T aff is indeed a piecewise interpretation.
The Semantic Dual as a Functor between Categories of Models
Given a piecewise interpretation f : S → T f : S \to T f : S → T , the function f ∗ : Mod ( T ) → Mod ( S ) f^* : \operatorname{Mod}(T) \to\operatorname{Mod}(S) f ∗ : Mod ( T ) → Mod ( S ) can be seen as the object-part of a functor. The definition of f ∗ f^* f ∗ on morphisms, that is, on elementary embeddings between models of the target theory, proceeds in the following fashion: Recall that an elementary embedding between Σ \Sigma Σ -structures j : M → N j : M \to N j : M → N is a function from the carrier of M M M to that of N N N , that is, a family of functions ( j σ : M ( σ ) → N ( σ ) ) σ ∈ S (j_\sigma: M(\sigma) \to N(\sigma))_{\sigma\in S} ( j σ : M ( σ ) → N ( σ ) ) σ ∈ S , such that, if φ : σ ˉ = σ 1 × ⋯ × σ n \varphi: \bar{\sigma} = \sigma_1 \times\cdots\times\sigma_n φ : σ ˉ = σ 1 × ⋯ × σ n is an L Σ \mathcal{L}_\Sigma L Σ -formula and a ˉ ∈ M ( σ ˉ ) \bar{a} \in M(\bar{\sigma}) a ˉ ∈ M ( σ ˉ ) ,
( a 1 , … , a n ) ∈ M ( φ ) ⟺ ( j σ 1 ( a 1 ) , … , j σ n ( a n ) ) ∈ N ( φ ) . (26) (a_1,\ldots,a_n) \in M(\varphi) \Longleftrightarrow(j_{\sigma_1}(a_1),\ldots,j_{\sigma_n}(a_n)) \in N(\varphi). \tag*{(26)} ( a 1 , … , a n ) ∈ M ( φ ) ⟺ ( j σ 1 ( a 1 ) , … , j σ n ( a n )) ∈ N ( φ ) . (26)
We will suppress these subscripts from now on, as the respective sort will always be clear from context.
Definition 2.5.1 (Semantic Dual on Elementary Embeddings). Let f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 be a piecewise translation respecting identity with respect to T T T , and let j : M → N j : M \to N j : M → N be an elementary embedding between models of T T T . Define f ∗ j : f ∗ M → f ∗ N f^*j : f^*M \to f^*N f ∗ j : f ∗ M → f ∗ N by
f ∗ j ( π M , σ f ( k , a ˉ ) ) : = π N , σ f ( k , j ( a ˉ ) ) . (27) f^*j\left(\pi^f_{M,\sigma}(k,\bar{a})\right) := \pi^f_{N,\sigma}(k,j(\bar{a})). \tag*{(27)} f ∗ j ( π M , σ f ( k , a ˉ ) ) := π N , σ f ( k , j ( a ˉ )) . (27) Well-definedness follows from the fact that j j j is an elementary embedding.
Proposition 2.5.2 (Semantic Dual is a Functor). Let f : S → T f : S \to T f : S → T be a piecewise interpretation. Then there is a corresponding functor f ∗ : Mod ( T ) → Mod ( S ) f^* : \operatorname{Mod}(T) \to\operatorname{Mod}(S) f ∗ : Mod ( T ) → Mod ( S ) , sending a model M ⊨ T M \models T M ⊨ T to f ∗ M ⊨ S f^*M \models S f ∗ M ⊨ S and an elementary embedding j : M → N j : M \to N j : M → N to the elementary embedding f ∗ j : f ∗ M → f ∗ N f^*j : f^*M \to f^*N f ∗ j : f ∗ M → f ∗ N .
Proof. Given a model M ⊨ T M \models T M ⊨ T , Corollary 2.4.9 tells us that f ∗ M f^*M f ∗ M is a model of S S S . To see that f ∗ j f^*j f ∗ j is an elementary embedding if j : M → N j : M \to N j : M → N is, let a ˉ ∈ ∏ i f ∗ M ( σ i ) \bar{a} \in\prod_i f^*M(\sigma_i) a ˉ ∈ ∏ i f ∗ M ( σ i ) with a i = π σ i f ( k i , b ˉ i ) a_i = \pi^f_{\sigma_i}(k_i,\bar{b}_i) a i = π σ i f ( k i , b ˉ i ) , and set k ˉ = ( k 1 , … , k n ) \bar{k} = (k_1,\ldots,k_n) k ˉ = ( k 1 , … , k n ) . Then for any L S \mathcal{L}_S L S -formula φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ , Corollary 2.4.8 yields
a ˉ ∈ f ∗ M ( φ ) ⟺ b ˉ 1 × ⋯ × b ˉ n ∈ M ( φ f k ˉ ) ∩ M ( D σ f , k ˉ ) (28) \bar{a} \in f^*M(\varphi) \Longleftrightarrow\bar{b}_1 \times\cdots\times\bar{b}_n \in M(\varphi^{\bar{k}}_f) \cap M(D^{f,\bar{k}}_\sigma) \tag*{(28)} a ˉ ∈ f ∗ M ( φ ) ⟺ b ˉ 1 × ⋯ × b ˉ n ∈ M ( φ f k ˉ ) ∩ M ( D σ f , k ˉ ) (28) ⟺ j ( b ˉ 1 ) × ⋯ × j ( b ˉ n ) ∈ N ( φ f k ˉ ) ∩ N ( D σ f , k ˉ ) (29) \Longleftrightarrow j(\bar{b}_1) \times\cdots\times j(\bar{b}_n) \in N(\varphi^{\bar{k}}_f) \cap N(D^{f,\bar{k}}_\sigma) \tag*{(29)} ⟺ j ( b ˉ 1 ) × ⋯ × j ( b ˉ n ) ∈ N ( φ f k ˉ ) ∩ N ( D σ f , k ˉ ) (29) ⟺ f ∗ j ( a ˉ ) ∈ f ∗ N ( φ ) , (30) \Longleftrightarrow f^*j(\bar{a}) \in f^*N(\varphi), \tag*{(30)} ⟺ f ∗ j ( a ˉ ) ∈ f ∗ N ( φ ) , (30) where the middle step applies elementarity of j j j and the outer steps use Corollary 2.4.8 for M M M and N N N respectively.
Functoriality is easy: f ∗ j f^*j f ∗ j acts on representatives by applying j j j componentwise, so f ∗ ( j 2 ∘ j 1 ) = f ∗ j 2 ∘ f ∗ j 1 f^*(j_2 \circ j_1) = f^*j_2 \circ f^*j_1 f ∗ ( j 2 ∘ j 1 ) = f ∗ j 2 ∘ f ∗ j 1 and f ∗ ( id M ) = id f ∗ M f^*(\operatorname{id}_M) = \operatorname{id}_{f^*M} f ∗ ( id M ) = id f ∗ M .
Note that if f , g : S → T f,g : S \to T f , g : S → T satisfy f ≡ T g f \equiv_T g f ≡ T g , then f ∗ = g ∗ f^* = g^* f ∗ = g ∗ , by Proposition 2.4.10 and Definition 2.5.1.
Morphisms between Piecewise Translations
While the assignment f ↦ f ∗ f \mapsto f^{*} f ↦ f ∗ , sending an interpretation to its semantic dual, is clearly not injective, it becomes injective (on interpretations with the same sort functions) after descending to ≡ T \equiv_{T} ≡ T -equivalence classes by Proposition 2.4.10. However, provable equivalence does not capture other natural relations between the semantic duals of translations: for instance, one might be interested in naturally isomorphic duals or duals which are the same up to a reindexing of the sets of pieces, both of which provide other natural candidates for equivalence of translations.
In this subsection, we develop a notion of equivalence for translations which does not unduly depend on the sets of pieces used, on the basis of a syntactically expressible notion of morphism between piecewise translations. Semantically, such a morphism will correspond to a natural transformation between the semantic duals of the interpretations.
Definition 2.6.1 (i i i -Morphism between Piecewise Translations). Let f , g : L 1 → L 2 f,g:\mathcal{L}_{1}\to\mathcal{L}_{2} f , g : L 1 → L 2 be two piecewise translations. A pre-i i i -morphism η \eta η from f f f to g g g is a collection ( η σ : P σ f × P σ g → L 2 ) σ ∈ S 1 (\eta_{\sigma}:P^{f}_{\sigma}\times P^{g}_{\sigma}\to\mathcal{L}_{2})_{\sigma\in S_{1}} ( η σ : P σ f × P σ g → L 2 ) σ ∈ S 1 of finite families of L 2 \mathcal{L}_{2} L 2 -formulae for each sort of L 1 \mathcal{L}_{1} L 1 such that η σ k , l : = η σ k , l ( x f k , y g l ) \eta_{\sigma}^{k,l}:=\eta_{\sigma}^{k,l}(x_{f}^{k},y_{g}^{l}) η σ k , l := η σ k , l ( x f k , y g l ) . We extend to sequences σ ˉ ∈ S 1 ∗ \bar{\sigma}\in S_{1}^{*} σ ˉ ∈ S 1 ∗ and pieces k ˉ ∈ P σ ˉ f \bar{k}\in P^{f}_{\bar{\sigma}} k ˉ ∈ P σ ˉ f and l ˉ ∈ P σ ˉ g \bar{l}\in P^{g}_{\bar{\sigma}} l ˉ ∈ P σ ˉ g by choosing x ˉ = x ˉ 1 × ⋯ × x ˉ n : σ ˉ → σ ˉ f k ˉ \bar{x}=\bar{x}_{1}\times\cdots\times\bar{x}_{n}:\bar{\sigma}\to\bar{\sigma}^{\bar{k}}_{f} x ˉ = x ˉ 1 × ⋯ × x ˉ n : σ ˉ → σ ˉ f k ˉ and y ˉ = y ˉ 1 × ⋯ × y ˉ n : σ ˉ → σ ˉ g l ˉ \bar{y}=\bar{y}_{1}\times\cdots\times\bar{y}_{n}:\bar{\sigma}\to\bar{\sigma}^{\bar{l}}_{g} y ˉ = y ˉ 1 × ⋯ × y ˉ n : σ ˉ → σ ˉ g l ˉ and setting η σ ˉ k ˉ , l ˉ : = ⊤ x ˉ × y ˉ ∧ ⋀ i = 1 ∣ σ ˉ ∣ η σ i k i , l i ( x i , y i ) \eta_{\bar{\sigma}}^{\bar{k},\bar{l}}:=\top_{\bar{x}\times\bar{y}}\land\bigwedge_{i=1}^{|\bar{\sigma}|}\eta_{\sigma_{i}}^{k_{i},l_{i}}(x_{i},y_{i}) η σ ˉ k ˉ , l ˉ := ⊤ x ˉ × y ˉ ∧ ⋀ i = 1 ∣ σ ˉ ∣ η σ i k i , l i ( x i , y i ) . It is an i i i -morphism with respect to an L 2 \mathcal{L}_{2} L 2 -theory T T T , expressed by η : f ⇒ g \eta:f\Rightarrow g η : f ⇒ g , iff the theory T T T proves the universal closures of the following formulae for all appropriate pieces, writing η σ , x , y k , l \eta_{\sigma,x,y}^{k,l} η σ , x , y k , l for η σ k , l ( x f k , y g l ) \eta_{\sigma}^{k,l}(x_{f}^{k},y_{g}^{l}) η σ k , l ( x f k , y g l ) :
η σ , x , y k , l → ( D x f , k ∧ D y g , l ) (Dom/Range) \eta_{\sigma,x,y}^{k,l}\to\left(D_{x}^{f,k}\land D_{y}^{g,l}\right)
\qquad\text{(Dom/Range)} η σ , x , y k , l → ( D x f , k ∧ D y g , l ) (Dom/Range) ( η σ , x , y k , l ∧ D x ′ f , k ′ ∧ D y ′ g , l ′ ∧ ( x = σ x ′ ) f k , k ′ ∧ ( y = σ y ′ ) g l , l ′ ) → η σ , x ′ , y ′ k ′ , l ′ (Well-Def) \left(\eta_{\sigma,x,y}^{k,l}\land D_{x'}^{f,k'}\land D_{y'}^{g,l'}\land(x=_{\sigma}x')_{f}^{k,k'}\land(y=_{\sigma}y')_{g}^{l,l'}\right)\to\eta_{\sigma,x',y'}^{k',l'}
\qquad\text{(Well-Def)} ( η σ , x , y k , l ∧ D x ′ f , k ′ ∧ D y ′ g , l ′ ∧ ( x = σ x ′ ) f k , k ′ ∧ ( y = σ y ′ ) g l , l ′ ) → η σ , x ′ , y ′ k ′ , l ′ (Well-Def) D x f , k → ⋁ l ∈ P σ g ∃ y g l η σ , x , y k , l (Joint-Exist) D_{x}^{f,k}\to\bigvee_{l\in P_{\sigma}^{g}}\exists y_{g}^{l}\eta_{\sigma,x,y}^{k,l}
\qquad\text{(Joint-Exist)} D x f , k → l ∈ P σ g ⋁ ∃ y g l η σ , x , y k , l (Joint-Exist) ( η σ , x , y k , l 1 ∧ η σ , x , z k , l 2 ) → ( y = z ) g l 1 , l 2 (Unique) \left(\eta_{\sigma,x,y}^{k,l_{1}}\land\eta_{\sigma,x,z}^{k,l_{2}}\right)\to(y=z)_{g}^{l_{1},l_{2}}
\qquad\text{(Unique)} ( η σ , x , y k , l 1 ∧ η σ , x , z k , l 2 ) → ( y = z ) g l 1 , l 2 (Unique) Moreover, for each formula φ : σ ˉ ∈ L 1 \varphi:\bar{\sigma}\in\mathcal{L}_{1} φ : σ ˉ ∈ L 1 , pair of pieces k ˉ ∈ P σ ˉ f \bar{k}\in P_{\bar{\sigma}}^{f} k ˉ ∈ P σ ˉ f and l ˉ ∈ P σ ˉ g \bar{l}\in P_{\bar{\sigma}}^{g} l ˉ ∈ P σ ˉ g , the theory T T T must prove
η σ ˉ , x ˉ , y ˉ k ˉ , l ˉ → ( φ ( x ˉ ) k ˉ f ↔ φ ( y ˉ ) l ˉ g ) (Element-Embed) \eta_{\bar{\sigma},\bar{x},\bar{y}}^{\bar{k},\bar{l}}\to\left(\varphi(\bar{x})_{\bar{k}}^{f}\leftrightarrow\varphi(\bar{y})_{\bar{l}}^{g}\right)
\qquad\text{(Element-Embed)} η σ ˉ , x ˉ , y ˉ k ˉ , l ˉ → ( φ ( x ˉ ) k ˉ f ↔ φ ( y ˉ ) l ˉ g ) (Element-Embed) where η σ ˉ , x ˉ , y ˉ k ˉ , l ˉ : = η σ ˉ k ˉ , l ˉ ( x ˉ f k ˉ , y ˉ g l ˉ ) \eta_{\bar{\sigma},\bar{x},\bar{y}}^{\bar{k},\bar{l}}:=\eta_{\bar{\sigma}}^{\bar{k},\bar{l}}(\bar{x}_{f}^{\bar{k}},\bar{y}_{g}^{\bar{l}}) η σ ˉ , x ˉ , y ˉ k ˉ , l ˉ := η σ ˉ k ˉ , l ˉ ( x ˉ f k ˉ , y ˉ g l ˉ ) . Here σ ˉ = ϵ \bar{\sigma}=\epsilon σ ˉ = ϵ is permitted: in that case the antecedent is the empty conjunction ⊤ \top ⊤ , and since P ϵ = { ϵ } P_{\epsilon}=\{\epsilon\} P ϵ = { ϵ } and φ f ϵ = φ f \varphi_{f}^{\epsilon}=\varphi_{f} φ f ϵ = φ f , instantiation yields T ⊢ φ f ↔ φ g T\vdash\varphi_{f}\leftrightarrow\varphi_{g} T ⊢ φ f ↔ φ g .
For i i i -morphisms from a translation respecting identity, (Unique) already follows from (Element-Embed) applied to the identity, together with (Dom/Range) and (Reflexivity) for f f f , so one does not need to check it in these cases. Typically, we will encounter situations in which we have a (pre-)i i i -morphism η : f ⇒ g \eta:f\Rightarrow g η : f ⇒ g , where we antecedently know that f f f is an interpretation S → T ′ S\to T' S → T ′ . In such cases, we will assume T = T ′ T=T' T = T ′ . Note that if we have translations f ≡ T f ′ f\equiv_{T}f' f ≡ T f ′ and g ≡ T g ′ g\equiv_{T}g' g ≡ T g ′ , then a collection of formulae defining a pre-i i i -morphism/an i i i -morphism from f f f to g g g also defines a pre-i i i -morphism/an i i i -morphism from f ′ f' f ′ to g ′ g' g ′ . We will only ever consider (pre-)i i i -morphisms up to provable equivalence in a target theory.
Definition 2.6.2 (Provable Equivalence of i i i -Morphisms). Two parallel pre-i i i -morphisms η \eta η and ζ \zeta ζ between translations f f f and g g g are provably equivalent with respect to a theory T T T , denoted by η ≡ T ζ \eta\equiv_T \zeta η ≡ T ζ , iff, for every pair of pieces ( k , l ) ∈ P σ f × P σ g (k,l) \in P_\sigma^f \times P_\sigma^g ( k , l ) ∈ P σ f × P σ g and all x , y : σ x,y : \sigma x , y : σ , the following holds:
T ⊢ η x , y k , l ↔ ζ x , y k , l (31) T \vdash\eta_{x,y}^{k,l} \leftrightarrow\zeta_{x,y}^{k,l} \tag*{(31)} T ⊢ η x , y k , l ↔ ζ x , y k , l (31)
Taking the sentential case of condition (Element-Embed) as sufficient for equivalence of translations, we obtain a weaker but still instrumentally useful and interesting notion of equivalence for translations:
Definition 2.6.3 (Sentential Equivalence). Two parallel translations f , g : L 1 → L 2 f,g : \mathcal{L}_1 \to\mathcal{L}_2 f , g : L 1 → L 2 are sententially equivalent with respect to an L 2 \mathcal{L}_2 L 2 -theory T 2 T_2 T 2 iff the target theory T 2 T_2 T 2 proves φ f ↔ φ g \varphi_f \leftrightarrow\varphi_g φ f ↔ φ g for any L 1 \mathcal{L}_1 L 1 -sentence φ \varphi φ .
Sentential equivalence is clearly entailed by provable equivalence. The following proposition provides a useful tool for establishing interpretationhood of translations via i i i -morphisms, established by inspecting the sentential instances of (Element-Embed).
Proposition 2.6.4 (Criterion for Interpretationhood). Let f , g : L 1 → L 2 f,g : \mathcal{L}_1 \to\mathcal{L}_2 f , g : L 1 → L 2 be piecewise translations and T 2 T_2 T 2 an L 2 \mathcal{L}_2 L 2 -theory. If there is an i i i -morphism η : f ⇒ g \eta: f \Rightarrow g η : f ⇒ g with respect to T 2 T_2 T 2 , then f f f and g g g are sententially equivalent with respect to T 2 T_2 T 2 . Therefore, if f f f is an interpretation f : T 1 → T 2 f : T_1 \to T_2 f : T 1 → T 2 , then so is g g g .
For i i i -morphisms between interpretations, we have a sensible notion of semantic dual, viz., piecewise definable collections of elementary embeddings between the induced models.
Proposition 2.6.5 (Dual of an i i i -Morphism). Let f , g : S → T f,g : S \to T f , g : S → T be piecewise interpretations and η : f ⇒ g \eta: f \Rightarrow g η : f ⇒ g an i i i -morphism. Then there is a unique natural transformation η ∗ : f ∗ ⇒ g ∗ \eta^* : f^* \Rightarrow g^* η ∗ : f ∗ ⇒ g ∗ , the semantic dual of η \eta η , such that for all M ⊨ T M \models T M ⊨ T , sorts σ \sigma σ of S S S , pieces k ∈ P σ f k \in P_\sigma^f k ∈ P σ f and l ∈ P σ g l \in P_\sigma^g l ∈ P σ g , and tuples a ˉ ∈ M ( D σ f , k ) \bar{a} \in M(D_\sigma^{f,k}) a ˉ ∈ M ( D σ f , k ) and b ˉ ∈ M ( D σ g , l ) \bar{b} \in M(D_\sigma^{g,l}) b ˉ ∈ M ( D σ g , l ) :
a ˉ × b ˉ ∈ M ( η σ k , l ) ⟺ η M , σ ∗ ( π σ f ( k , a ˉ ) ) = π σ g ( l , b ˉ ) \bar{a} \times\bar{b} \in M(\eta_\sigma^{k,l}) \quad\Longleftrightarrow\quad\eta^*_{M,\sigma}\left(\pi_\sigma^f(k,\bar{a})\right)=\pi_\sigma^g(l,\bar{b}) a ˉ × b ˉ ∈ M ( η σ k , l ) ⟺ η M , σ ∗ ( π σ f ( k , a ˉ ) ) = π σ g ( l , b ˉ ) (Dual)
Proof. Existence, (Dual), and Uniqueness: Let M ⊨ T M \models T M ⊨ T . Every element of f ∗ M ( σ ) f^*M(\sigma) f ∗ M ( σ ) is of the form π σ f ( k , a ˉ ) \pi_\sigma^f(k,\bar{a}) π σ f ( k , a ˉ ) with a ˉ ∈ M ( D σ f , k ) \bar{a} \in M(D_\sigma^{f,k}) a ˉ ∈ M ( D σ f , k ) . By (Joint-Exist) and (Dom/Range) there are l ∈ P σ g l \in P_\sigma^g l ∈ P σ g and b ˉ ∈ M ( D σ g , l ) \bar{b} \in M(D_\sigma^{g,l}) b ˉ ∈ M ( D σ g , l ) with a ˉ × b ˉ ∈ M ( η σ k , l ) \bar{a} \times\bar{b} \in M(\eta_\sigma^{k,l}) a ˉ × b ˉ ∈ M ( η σ k , l ) , and we set η M , σ ∗ ( π σ f ( k , a ˉ ) ) : = π σ g ( l , b ˉ ) \eta^*_{M,\sigma}(\pi_\sigma^f(k,\bar{a})) := \pi_\sigma^g(l,\bar{b}) η M , σ ∗ ( π σ f ( k , a ˉ )) := π σ g ( l , b ˉ ) . By condition (Unique) this is independent of ( l , b ˉ ) (l,\bar{b}) ( l , b ˉ ) , and (Well-Def) ensures independence of the representative ( k , a ˉ ) (k,\bar{a}) ( k , a ˉ ) . The left-to-right direction of (Dual) holds by this definition together with (Unique). For the right-to-left direction, suppose η M , σ ∗ ( π σ f ( k , a ˉ ) ) = π σ g ( l , b ˉ ) \eta^*_{M,\sigma}(\pi_\sigma^f(k,\bar{a})) = \pi_\sigma^g(l,\bar{b}) η M , σ ∗ ( π σ f ( k , a ˉ )) = π σ g ( l , b ˉ ) with b ˉ ∈ M ( D σ g , l ) \bar{b} \in M(D_\sigma^{g,l}) b ˉ ∈ M ( D σ g , l ) , and let ( l ′ , b ˉ ′ ) (l',\bar{b}') ( l ′ , b ˉ ′ ) be such that a ˉ × b ˉ ′ ∈ M ( η σ k , l ′ ) \bar{a} \times\bar{b}' \in M(\eta_\sigma^{k,l'}) a ˉ × b ˉ ′ ∈ M ( η σ k , l ′ ) and b ˉ ′ × b ˉ ∈ M ( ( = σ ) l ′ , l ) \bar{b}' \times\bar{b} \in M((=_{\sigma})^{l',l}) b ˉ ′ × b ˉ ∈ M (( = σ ) l ′ , l ) . Since a ˉ × a ˉ ∈ M ( ( = σ f ) k , k ) \bar{a} \times\bar{a} \in M((=_{\sigma}^{f})^{k,k}) a ˉ × a ˉ ∈ M (( = σ f ) k , k ) by ( = f (=_{f} ( = f -Intro) for f f f , condition (Well-Def) implies a ˉ × b ˉ ∈ M ( η σ k , l ) \bar{a} \times\bar{b} \in M(\eta_\sigma^{k,l}) a ˉ × b ˉ ∈ M ( η σ k , l ) . Uniqueness follows immediately from surjectivity of π σ f \pi_\sigma^f π σ f and (Dual).
Elementarity of the components: Fix φ : σ ˉ ∈ L S \varphi: \bar{\sigma} \in\mathcal{L}_S φ : σ ˉ ∈ L S with ∣ σ ˉ ∣ ≥ 1 |\bar{\sigma}| \ge1 ∣ σ ˉ ∣ ≥ 1 . Then:
∀ k ˉ ∈ P σ ˉ f , ∀ l ˉ ∈ P σ ˉ g : T ⊢ η x ˉ , y ˉ k ˉ , l ˉ → ( φ ( x ˉ ) f k ˉ ↔ φ ( y ˉ ) g l ˉ ) (32) \forall\bar{k} \in P_{\bar{\sigma}}^f,\ \forall\bar{l} \in P_{\bar{\sigma}}^g : T \vdash\eta_{\bar{x},\bar{y}}^{\bar{k},\bar{l}} \to\left(\varphi(\bar{x})_f^{\bar{k}} \leftrightarrow\varphi(\bar{y})_g^{\bar{l}}\right) \tag*{(32)} ∀ k ˉ ∈ P σ ˉ f , ∀ l ˉ ∈ P σ ˉ g : T ⊢ η x ˉ , y ˉ k ˉ , l ˉ → ( φ ( x ˉ ) f k ˉ ↔ φ ( y ˉ ) g l ˉ ) (32)
⟺ ∀ M ⊨ T , ∀ k ˉ ∈ P σ ˉ f , ∀ l ˉ ∈ P σ ˉ g ∀ a ˉ ∈ M ( σ ˉ f , k ˉ ) , ∀ b ˉ ∈ M ( σ ˉ g , l ˉ ) : (33) \Longleftrightarrow\ \forall M \models T,\ \forall\bar{k} \in P_{\bar{\sigma}}^f,\ \forall\bar{l} \in P_{\bar{\sigma}}^g\ \forall\bar{a} \in M(\bar{\sigma}^{f,\bar{k}}),\ \forall\bar{b} \in M(\bar{\sigma}^{g,\bar{l}}): \tag*{(33)} ⟺ ∀ M ⊨ T , ∀ k ˉ ∈ P σ ˉ f , ∀ l ˉ ∈ P σ ˉ g ∀ a ˉ ∈ M ( σ ˉ f , k ˉ ) , ∀ b ˉ ∈ M ( σ ˉ g , l ˉ ) : (33)
a ˉ × b ˉ ∈ M ( η k ˉ , l ˉ ) ⇒ ( a ˉ ∈ M ( φ f k ˉ ) ↔ b ˉ ∈ M ( φ g l ˉ ) ) (34) \bar{a} \times\bar{b} \in M(\eta^{\bar{k},\bar{l}}) \Rightarrow\left(\bar{a} \in M(\varphi_f^{\bar{k}}) \leftrightarrow\bar{b} \in M(\varphi_g^{\bar{l}})\right) \tag*{(34)} a ˉ × b ˉ ∈ M ( η k ˉ , l ˉ ) ⇒ ( a ˉ ∈ M ( φ f k ˉ ) ↔ b ˉ ∈ M ( φ g l ˉ ) ) (34)
⟺ ∀ M ⊨ T , ∀ a ˉ ∈ f ∗ M ( σ ˉ ) : a ˉ ∈ f ∗ M ( φ ) ⟺ η M ∗ ( a ˉ ) ∈ g ∗ M ( φ ) . (35) \Longleftrightarrow\ \forall M \models T,\ \forall\bar{a} \in f^*M(\bar{\sigma}):\ \bar{a} \in f^*M(\varphi) \Longleftrightarrow\eta^*_M(\bar{a}) \in g^*M(\varphi). \tag*{(35)} ⟺ ∀ M ⊨ T , ∀ a ˉ ∈ f ∗ M ( σ ˉ ) : a ˉ ∈ f ∗ M ( φ ) ⟺ η M ∗ ( a ˉ ) ∈ g ∗ M ( φ ) . (35)
The first equivalence is an instance of soundness and completeness. The second uses (Dom/Range), (Joint-Exist), (Dual), and Corollary 2.4.8 . Quantifying over all appropriate φ \varphi φ , the top line ranges over the instances of (Element-Embed) at positive arities, and the bottom line states that every η M ∗ \eta^{*}_{M} η M ∗ is an elementary embedding.
Naturality: For naturality, we need to show that if j : M → N j : M \to N j : M → N is an elementary embedding, we have η N ∗ ∘ f ∗ j = g ∗ j ∘ η M ∗ \eta^{*}_{N} \circ f^{*}j = g^{*}j \circ\eta^{*}_{M} η N ∗ ∘ f ∗ j = g ∗ j ∘ η M ∗ . So let j : M → N j : M \to N j : M → N be elementary and π σ f ( k , a ˉ ) ∈ f ∗ M ( σ ) \pi^{f}_{\sigma}(k,\bar{a}) \in f^{*}M(\sigma) π σ f ( k , a ˉ ) ∈ f ∗ M ( σ ) . Pick ( l , b ˉ ) (l,\bar{b}) ( l , b ˉ ) with η M , σ ∗ ( π σ f ( k , a ˉ ) ) = π σ g ( l , b ˉ ) \eta^{*}_{M,\sigma}(\pi^{f}_{\sigma}(k,\bar{a})) = \pi^{g}_{\sigma}(l,\bar{b}) η M , σ ∗ ( π σ f ( k , a ˉ )) = π σ g ( l , b ˉ ) . Then a ˉ × b ˉ ∈ M ( η σ k , l ) \bar{a}\times\bar{b} \in M(\eta^{k,l}_{\sigma}) a ˉ × b ˉ ∈ M ( η σ k , l ) by (Dual), so j ( a ˉ ) × j ( b ˉ ) ∈ N ( η σ k , l ) j(\bar{a})\times j(\bar{b}) \in N(\eta^{k,l}_{\sigma}) j ( a ˉ ) × j ( b ˉ ) ∈ N ( η σ k , l ) with j ( a ˉ ) ∈ N ( D σ f , k ) j(\bar{a}) \in N(D^{f,k}_{\sigma}) j ( a ˉ ) ∈ N ( D σ f , k ) and j ( b ˉ ) ∈ N ( D σ g , l ) j(\bar{b}) \in N(D^{g,l}_{\sigma}) j ( b ˉ ) ∈ N ( D σ g , l ) , by elementarity of j j j . Applying (Dual) in N N N yields η N , σ ∗ ( π σ f ( k , j ( a ˉ ) ) ) = π σ g ( l , j ( b ˉ ) ) \eta^{*}_{N,\sigma}(\pi^{f}_{\sigma}(k,j(\bar{a}))) = \pi^{g}_{\sigma}(l,j(\bar{b})) η N , σ ∗ ( π σ f ( k , j ( a ˉ ))) = π σ g ( l , j ( b ˉ )) , which is the required equation, since f ∗ j f^{*}j f ∗ j and g ∗ j g^{*}j g ∗ j act on representatives by j j j .
As a translation f : L 1 → L 2 f : L_{1} \to L_{2} f : L 1 → L 2 which respects identity with respect to an L 2 L_{2} L 2 -theory T T T can be seen as an interpretation L ( Σ 1 ) → T L(\Sigma_{1}) \to T L ( Σ 1 ) → T , the proposition above also applies to them. We have a partial converse, allowing us to check semantically whether a pre-i i i -morphism is an i i i -morphism.
Proposition 2.6.6 (Semantic Recognition of i i i -Morphisms). Let f , g : S → T f,g : S \to T f , g : S → T be piecewise interpretations, let θ \theta θ be a pre-i i i -morphism from f f f to g g g satisfying (Dom/Range) with respect to T T T , and suppose there is a family of elementary embeddings ( F M : f ∗ M → g ∗ M ) M ⊨ T (F_{M} : f^{*}M \to g^{*}M)_{M\models T} ( F M : f ∗ M → g ∗ M ) M ⊨ T satisfying (Dual) with F F F in place of η ∗ \eta^{*} η ∗ . Then θ : f ⇒ g \theta: f \Rightarrow g θ : f ⇒ g is an i i i -morphism and θ ∗ = F \theta^{*} = F θ ∗ = F .
Proof. We rely on completeness throughout. For (Well-Def), suppose a ˉ × b ˉ ∈ M ( θ σ k , l ) \bar{a}\times\bar{b} \in M(\theta^{k,l}_{\sigma}) a ˉ × b ˉ ∈ M ( θ σ k , l ) , and that ( k ′ , a ˉ ′ ) ∈ P σ f × M ( D σ f , k ′ ) (k',\bar{a}') \in P^{f}_{\sigma} \times M(D^{f,k'}_{\sigma}) ( k ′ , a ˉ ′ ) ∈ P σ f × M ( D σ f , k ′ ) and ( l ′ , b ˉ ′ ) ∈ P σ g × M ( D σ g , l ′ ) (l',\bar{b}') \in P^{g}_{\sigma} \times M(D^{g,l'}_{\sigma}) ( l ′ , b ˉ ′ ) ∈ P σ g × M ( D σ g , l ′ ) represent the same respective classes. Then F M , σ ( π σ f ( k ′ , a ˉ ′ ) ) = F M , σ ( π σ f ( k , a ˉ ) ) = π σ g ( l , b ˉ ) = π σ g ( l ′ , b ˉ ′ ) F_{M,\sigma}(\pi^{f}_{\sigma}(k',\bar{a}')) = F_{M,\sigma}(\pi^{f}_{\sigma}(k,\bar{a})) = \pi^{g}_{\sigma}(l,\bar{b}) = \pi^{g}_{\sigma}(l',\bar{b}') F M , σ ( π σ f ( k ′ , a ˉ ′ )) = F M , σ ( π σ f ( k , a ˉ )) = π σ g ( l , b ˉ ) = π σ g ( l ′ , b ˉ ′ ) by the left-to-right direction of (Dual), so the right-to-left direction yields a ˉ ′ × b ˉ ′ ∈ M ( θ σ k ′ , l ′ ) \bar{a}'\times\bar{b}' \in M(\theta^{k',l'}_{\sigma}) a ˉ ′ × b ˉ ′ ∈ M ( θ σ k ′ , l ′ ) . For (Joint-Exist), given a ˉ ∈ M ( D σ f , k ) \bar{a} \in M(D^{f,k}_{\sigma}) a ˉ ∈ M ( D σ f , k ) , choose a representative π σ g ( l , b ˉ ) \pi^{g}_{\sigma}(l,\bar{b}) π σ g ( l , b ˉ ) of F M , σ ( π σ f ( k , a ˉ ) ) F_{M,\sigma}(\pi^{f}_{\sigma}(k,\bar{a})) F M , σ ( π σ f ( k , a ˉ )) and apply (Dual). For (Element-Embed), elementarity of the components of F F F allows us to run the elementarity part of the proof of Proposition 2.6.5 backwards, with the sentential instances of (Element-Embed) following from the fact that f ∗ M ≡ g ∗ M f^{*}M \equiv g^{*}M f ∗ M ≡ g ∗ M and completeness. Finally, θ ∗ = F \theta^{*} = F θ ∗ = F by the uniqueness part of Proposition 2.6.5 .
Together with the following corollary, this proposition allows us to show that a given pre-i i i -morphism is equivalent to an i i i -morphism and itself an i i i -morphism at once. The corollary itself is an easy consequence of (Dual) and completeness:
Corollary 2.6.7 (Local Faithfulness). Let η , ζ : f ⇒ g \eta,\zeta: f \Rightarrow g η , ζ : f ⇒ g be i i i -morphisms between piecewise interpretations f , g : S → T f,g : S \to T f , g : S → T . Then η ≡ T ζ \eta\equiv_{T} \zeta η ≡ T ζ iff η ∗ = ζ ∗ \eta^{*} = \zeta^{*} η ∗ = ζ ∗ .
The promised notion of equivalence between translations is realized by the notion of an i i i -isomorphism. Recall that an elementary embedding is automatically injective, so an elementary embedding is an isomorphism iff it is surjective, and that every isomorphism is elementary.
Definition 2.6.8 (i i i -Isomorphism). An i i i -morphism η : f ⇒ g \eta: f \Rightarrow g η : f ⇒ g between piecewise translations with respect to T T T is an i i i -isomorphism iff T T T proves D σ g , l → ⋁ k ∈ P σ f ∃ x ˉ η σ k , l ( x ˉ , y ˉ ) D^{g,l}_{\sigma} \to\bigvee_{k\in P^{f}_{\sigma}}\exists\bar{x}\,\eta^{k,l}_{\sigma}(\bar{x},\bar{y}) D σ g , l → ⋁ k ∈ P σ f ∃ x ˉ η σ k , l ( x ˉ , y ˉ ) . If f f f and g g g respect identity with respect to T T T , this is equivalent to η M ∗ \eta^{*}_{M} η M ∗ being an isomorphism at every model M ⊨ T M \models T M ⊨ T . If η : f ⇒ g \eta: f \Rightarrow g η : f ⇒ g is an i i i -isomorphism, we denote by η − 1 : g ⇒ f \eta^{-1} : g \Rightarrow f η − 1 : g ⇒ f the i i i -isomorphism defined by ( η − 1 ) σ l , k ( x ˉ , y ˉ ) = T ‾ x ˉ y ˉ ∧ η σ k , l ( y ˉ , x ˉ ) (\eta^{-1})^{l,k}_{\sigma}(\bar{x},\bar{y}) = \overline{T}_{\bar{x}\bar{y}} \wedge\eta^{k,l}_{\sigma}(\bar{y},\bar{x}) ( η − 1 ) σ l , k ( x ˉ , y ˉ ) = T x ˉ y ˉ ∧ η σ k , l ( y ˉ , x ˉ ) .
Restricting to i i i -morphisms between interpretations, we can apply Proposition 2.6.6 with F : = ( η ∗ ) − 1 F := (\eta^{*})^{-1} F := ( η ∗ ) − 1 to see that the inverse of an i i i -isomorphism defined above is indeed again an i i i -isomorphism. A particularly important class of i i i -isomorphisms in the following will be those between translations agreeing up to a family of bijections of their sets of pieces. Semantically, the interpretations related by such an i i i -isomorphism will merely differ in their choice of index sets for the respective disjoint unions.
Proposition 2.6.9 (i i i -Isomorphism induced by a Bijection of Pieces). Let f , g : T 1 → T 2 f,g:T_1\to T_2 f , g : T 1 → T 2 be piecewise interpretations and let ( b σ : P σ f → P σ g ) σ ∈ S 1 (b_\sigma:P^f_\sigma\to P^g_\sigma)_{\sigma\in S_1} ( b σ : P σ f → P σ g ) σ ∈ S 1 be a family of bijections between the families of sets of pieces for each sort, satisfying the following conditions for all σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 , σ ˉ ∈ S 1 ∗ \bar{\sigma}\in S_1^* σ ˉ ∈ S 1 ∗ , k ∈ P σ ˉ f k\in P^f_{\bar{\sigma}} k ∈ P σ ˉ f , k ˉ ∈ P σ ˉ f \bar{k}\in P^f_{\bar{\sigma}} k ˉ ∈ P σ ˉ f , R : σ ˉ ∈ R 1 R:\bar{\sigma}\in R_1 R : σ ˉ ∈ R 1 , including the identity symbols = σ =_{\sigma} = σ , and z ˉ : σ f k ∈ V 2 ∗ \bar{z}:\sigma^k_f\in V_2^* z ˉ : σ f k ∈ V 2 ∗ (resp. u ˉ : σ ˉ f k ˉ ∈ V 2 ∗ \bar{u}:\bar{\sigma}^{\bar{k}}_f\in V_2^* u ˉ : σ ˉ f k ˉ ∈ V 2 ∗ ):
σ f k = σ g b σ ( k ) \sigma^k_f=\sigma^{b_\sigma(k)}_g σ f k = σ g b σ ( k ) (Sort Condition)
T 2 ⊢ D σ f , k ( z ˉ ) ↔ D σ g , b σ ( k ) ( z ˉ ) T_2\vdash D^{f,k}_\sigma(\bar{z})\leftrightarrow D^{g,b_\sigma(k)}_\sigma(\bar{z}) T 2 ⊢ D σ f , k ( z ˉ ) ↔ D σ g , b σ ( k ) ( z ˉ ) (Domain Condition)
T 2 ⊢ D σ ˉ f , k ˉ ( u ˉ ) → ( R f k ˉ ( u ˉ ) ↔ R g b ( k ˉ ) ( u ˉ ) ) T_2\vdash D^{f,\bar{k}}_{\bar{\sigma}}(\bar{u})\to\left(R^{\bar{k}}_f(\bar{u})\leftrightarrow R^{b(\bar{k})}_g(\bar{u})\right) T 2 ⊢ D σ ˉ f , k ˉ ( u ˉ ) → ( R f k ˉ ( u ˉ ) ↔ R g b ( k ˉ ) ( u ˉ ) ) (Relation Condition)
where b ( k ˉ ) : = ( b σ 1 ( k 1 ) , … , b σ n ( k n ) ) b(\bar{k}):=(b_{\sigma_1}(k_1),\ldots,b_{\sigma_n}(k_n)) b ( k ˉ ) := ( b σ 1 ( k 1 ) , … , b σ n ( k n )) . Then the family of formulae ( η b ) σ : P σ f × P σ g → L 2 (\eta_b)_\sigma:P^f_\sigma\times P^g_\sigma\to L_2 ( η b ) σ : P σ f × P σ g → L 2 defined by
( η b ) σ k , l ( z ˉ , w ˉ ) : = D σ f , k ( z ˉ ) ∧ D σ g , l ( w ˉ ) ∧ ( = σ ) g b σ ( k ) , l ( z ˉ , w ˉ ) , (36) (\eta_b)^{k,l}_\sigma(\bar{z},\bar{w}):=D^{f,k}_\sigma(\bar{z})\land D^{g,l}_\sigma(\bar{w})\land(=_{\sigma})^{b_\sigma(k),l}_g(\bar{z},\bar{w}), \tag*{(36)} ( η b ) σ k , l ( z ˉ , w ˉ ) := D σ f , k ( z ˉ ) ∧ D σ g , l ( w ˉ ) ∧ ( = σ ) g b σ ( k ) , l ( z ˉ , w ˉ ) , (36) where z ˉ : σ f k \bar{z}:\sigma^k_f z ˉ : σ f k and w ˉ : σ g l \bar{w}:\sigma^l_g w ˉ : σ g l , is an i i i -isomorphism η b : f ⇒ g \eta_b:f\Rightarrow g η b : f ⇒ g with respect to T 2 T_2 T 2 . Furthermore, the dual η b ∗ \eta_b^* η b ∗ acts via ( η b ∗ ) M ( π σ f ( k , a ˉ ) ) = π σ g ( b σ ( k ) , a ˉ ) (\eta_b^*)_M(\pi^f_\sigma(k,\bar{a}))=\pi^g_\sigma(b_\sigma(k),\bar{a}) ( η b ∗ ) M ( π σ f ( k , a ˉ )) = π σ g ( b σ ( k ) , a ˉ ) .
Proof. Apply Proposition 2.6.6 with F M , σ ( π σ f ( k , a ˉ ) ) : = π σ g ( b σ ( k ) , a ˉ ) F_{M,\sigma}(\pi^f_\sigma(k,\bar{a})):=\pi^g_\sigma(b_\sigma(k),\bar{a}) F M , σ ( π σ f ( k , a ˉ )) := π σ g ( b σ ( k ) , a ˉ ) ; well-definedness of F F F uses the (Sort Condition), the (Domain Condition), and the (Relation Condition) for = σ =_{\sigma} = σ . From the latter, we also get injectivity, while surjectivity follows from the (Domain Condition) and bijectivity of b σ b_\sigma b σ . Finally, that F M F_M F M is an isomorphism and hence elementary follows from the (Relation Condition) for all R ∈ R 1 R\in R_1 R ∈ R 1 .
This proposition also supplies us with an easy way to see that if f ≡ T g f\equiv_Tg f ≡ T g , then f f f and g g g are i i i -isomorphic, by taking b = id b=\operatorname{id} b = id in the above proposition. Moreover, given two pairs of translations ( f , g ) (f,g) ( f , g ) and ( f ′ , g ′ ) (f',g') ( f ′ , g ′ ) such that f ≡ T f ′ f\equiv_Tf' f ≡ T f ′ and g ≡ T g ′ g\equiv_Tg' g ≡ T g ′ , and a family ( b σ : P σ f → P σ g ) (b_\sigma:P^f_\sigma\to P^g_\sigma) ( b σ : P σ f → P σ g ) satisfying the conditions above for ( f , g ) (f,g) ( f , g ) , it also satisfies them for ( f ′ , g ′ ) (f',g') ( f ′ , g ′ ) , and the resulting i i i -isomorphisms η b \eta_b η b and η b ′ \eta'_{b} η b ′ are provably equivalent.
The Bicategory of Piecewise Interpretations From a category-theoretic point of view, we have yet to establish that piecewise interpretations make for a good notion of morphism between theories, because we have not explained how they compose. In this section, we will show that piecewise interpretations and i i i -morphisms between them form a bicategory by developing appropriate notions of composition for them. That they form a bicategory means, in particular, that composition of piecewise interpretations is associative only up to a coherent i i i -isomorphism.13 The
overall strategy will be to avoid tedious syntactic verifications of the bicategory properties by employing the developed semantic duals. Let us start by recalling the definition of the notion of bicategory we will be using:
Definition 3.0.1 (Bicategory). A bicategory B \mathcal{B} B consists of: a class of objects ob ( B ) \operatorname{ob}(\mathcal{B}) ob ( B ) ; for all X , Y ∈ ob ( B ) X,Y \in\operatorname{ob}(\mathcal{B}) X , Y ∈ ob ( B ) a hom-category ( B ( X , Y ) , ∙ , id ) (\mathcal{B}(X,Y),\mathbin{\bullet},\operatorname{id}) ( B ( X , Y ) , ∙ , id ) , whose objects are called 1 1 1 -morphisms f : X → Y f:X \to Y f : X → Y and whose morphisms are called 2 2 2 -morphisms η : f ⇒ f ′ : X → Y \eta:f \Rightarrow f':X \to Y η : f ⇒ f ′ : X → Y of B \mathcal{B} B ; composition in B ( X , Y ) \mathcal{B}(X,Y) B ( X , Y ) is called vertical composition ; for each object X X X an identity 1 1 1 -morphism id X : X → X \operatorname{id}_{X}:X \to X id X : X → X , and for all X , Y , Z X,Y,Z X , Y , Z a composition functor
c X , Y , Z : B ( Y , Z ) × B ( X , Y ) ⟶ B ( X , Z ) ; (37) c_{X,Y,Z}:\mathcal{B}(Y,Z) \times\mathcal{B}(X,Y) \longrightarrow\mathcal{B}(X,Z); \tag*{(37)} c X , Y , Z : B ( Y , Z ) × B ( X , Y ) ⟶ B ( X , Z ) ; (37) and, finally, invertible 2 2 2 -morphisms, namely the associators and left and right unitors,
a h , g , f : ( h ∘ g ) ∘ f ⇒ h ∘ ( g ∘ f ) : W → Z (38) a_{h,g,f}:(h \circ g) \circ f \Rightarrow h \circ(g \circ f):W \to Z \tag*{(38)} a h , g , f : ( h ∘ g ) ∘ f ⇒ h ∘ ( g ∘ f ) : W → Z (38) l f : id X ∘ f ⇒ f : W → X r f : f ∘ id W ⇒ f : W → X , (39) l^{f}:\operatorname{id}_{X} \circ f \Rightarrow f:W \to X \qquad r^{f}:f \circ\operatorname{id}_{W} \Rightarrow f:W \to X, \tag*{(39)} l f : id X ∘ f ⇒ f : W → X r f : f ∘ id W ⇒ f : W → X , (39) for all f : W → X f:W \to X f : W → X , g : X → Y g:X \to Y g : X → Y , h : Y → Z h:Y \to Z h : Y → Z , natural in each argument and satisfying the pentagon and triangle identities (see Remark B.0.2). A bicategory in which a a a , l l l and r r r are all identities is called a 2 2 2 -category.
On a pair of 1 1 1 -morphisms ( g , f ) (g,f) ( g , f ) , the action of the composition functor is denoted by g ∘ f g \circ f g ∘ f . On a pair of 2 2 2 -morphisms ( η , ζ ) (\eta,\zeta) ( η , ζ ) it is denoted by η ⋆ ζ \eta\star\zeta η ⋆ ζ and called horizontal composition . Two special cases of horizontal composition are worth singling out. For a 1 1 1 -morphism f : X → Y f:X \to Y f : X → Y and a 2 2 2 -morphism ζ : h ⇒ h ′ : W → X \zeta:h \Rightarrow h':W \to X ζ : h ⇒ h ′ : W → X , the horizontal composite id f ⋆ ζ : f ∘ h ⇒ f ∘ h ′ \operatorname{id}_{f} \star\zeta:f \circ h \Rightarrow f \circ h' id f ⋆ ζ : f ∘ h ⇒ f ∘ h ′ is denoted by f ζ f\zeta f ζ and called the *left whiskering of ζ \zeta ζ by f f f ; dually, for η : g ⇒ g ′ : Y → Z \eta:g \Rightarrow g':Y \to Z η : g ⇒ g ′ : Y → Z the composite η ⋆ id f : g ∘ f ⇒ g ′ ∘ f \eta\star\operatorname{id}_{f}:g \circ f \Rightarrow g' \circ f η ⋆ id f : g ∘ f ⇒ g ′ ∘ f is denoted by η f \eta f η f and called the *right whiskering of η \eta η by f f f .
In the case of piecewise interpretations, the objects of our bicategory T h p w \mathrm{Th}^{\mathrm{pw}} Th pw are given by classical first-order theories. Given two theories S S S and T T T , the collection of 1 1 1 -morphisms T h p w ( S , T ) \mathrm{Th}^{\mathrm{pw}}(S,T) Th pw ( S , T ) between them is given by the set of (equivalence classes under provable equivalence of) piecewise interpretations f : S → T f:S \to T f : S → T . For any theory T T T , the identity interpretation id T : T → T \operatorname{id}_{T}:T \to T id T : T → T is defined as follows: For each sort σ \sigma σ of T T T , the set of pieces is P σ = { 0 } P_{\sigma}=\{0\} P σ = { 0 } , with σ id T 0 : = σ \sigma_{\operatorname{id}_{T}}^{0}:=\sigma σ id T 0 := σ , and x id T 0 = x x_{\operatorname{id}_{T}}^{0}=x x id T 0 = x for each variable x : σ x:\sigma x : σ . The domain formula D σ id T , 0 D_{\sigma}^{\operatorname{id}_{T},0} D σ id T , 0 is given by the verum ⊤ σ \top_{\sigma} ⊤ σ . For an atomic formula φ \varphi φ , we have φ id T 0 : = φ \varphi_{\operatorname{id}_{T}}^{0}:=\varphi φ id T 0 := φ . It follows by induction that for any L \mathcal{L} L -formula φ \varphi φ , x ‾ ( φ id T 0 ) = x ‾ ( φ ) \overline{x}(\varphi_{\operatorname{id}_{T}}^{0})=\overline{x}(\varphi) x ( φ id T 0 ) = x ( φ ) and ⊢ φ id T 0 ↔ φ \vdash\varphi_{\operatorname{id}_{T}}^{0} \leftrightarrow\varphi ⊢ φ id T 0 ↔ φ . The semantic dual of id T \operatorname{id}_{T} id T is naturally isomorphic to id Mod ( T ) \operatorname{id}_{\operatorname{Mod}(T)} id Mod ( T ) via the natural isomorphism
α T : id T ∗ ⟹ ∼ id Mod ( T ) π σ id T ( 0 , a ) ↦ a . (40) \alpha_{T}:\operatorname{id}_{T}^{*} \overset{\sim}{\Longrightarrow} \operatorname{id}_{\operatorname{Mod}(T)}
\qquad
\pi_{\sigma}^{\operatorname{id}_{T}}(0,a) \mapsto a. \tag*{(40)} α T : id T ∗ ⟹ ∼ id Mod ( T ) π σ id T ( 0 , a ) ↦ a . (40)
The theory of semantic duals we have developed will allow us to define a locally faithful contravariant pseudofunctor from the bicategory T h p w \mathrm{Th}^{\mathrm{pw}} Th pw into the 2 2 2 -category of categories C A T \mathbf{CAT} CAT , whose objects are categories, 1 1 1 -morphisms are functors, and 2 2 2 -morphisms are natural transformations between functors, with composition the usual compositions of functors and natural transformations. We will develop the bicategory T h p w \mathrm{Th}^{\mathrm{pw}} Th pw and this pseudofunctor concurrently, using it to show that we are indeed defining a bicategory. Here is the general definition of a pseudofunctor:
Definition 3.0.2 (Pseudofunctor). A pseudofunctor F : B → C F : \mathcal{B} \to\mathcal{C} F : B → C between bicategories consists of a function F : ob ( B ) → ob ( C ) F : \operatorname{ob}(\mathcal{B}) \to\operatorname{ob}(\mathcal{C}) F : ob ( B ) → ob ( C ) , a functor F X , Y : B ( X , Y ) → C ( F X , F Y ) F_{X,Y} : \mathcal{B}(X,Y) \to\mathcal{C}(FX,FY) F X , Y : B ( X , Y ) → C ( F X , F Y ) for all X , Y ∈ B X,Y \in\mathcal{B} X , Y ∈ B , an invertible 2-morphism ϕ X : id F X ⇒ F ( id X ) \phi_X : \operatorname{id}_{FX} \Rightarrow F(\operatorname{id}_X) ϕ X : id F X ⇒ F ( id X ) for each X ∈ B X \in\mathcal{B} X ∈ B (the unitor), an invertible 2-morphism ϕ g , f : F g ∘ F f ⇒ F ( g ∘ f ) \phi_{g,f} : Fg \circ Ff \Rightarrow F(g \circ f) ϕ g , f : F g ∘ F f ⇒ F ( g ∘ f ) for all f : X → Y f : X \to Y f : X → Y and g : Y → Z g : Y \to Z g : Y → Z , natural in f f f and g g g (the compositor), subject to two unit identities and an associativity identity:
F ( l f ) ∙ ϕ id X , f ∙ ( ϕ X F f ) = l F f , F(l^f) \mathbin{\bullet} \phi_{\operatorname{id}_X,f} \mathbin{\bullet} (\phi_X \mathbin{Ff}) = l^{Ff}, F ( l f ) ∙ ϕ id X , f ∙ ( ϕ X F f ) = l F f , F ( r f ) ∙ ϕ f , id W ∙ ( F f ϕ W ) = r F f , F(r^f) \mathbin{\bullet} \phi_{f,\operatorname{id}_W} \mathbin{\bullet} (Ff \mathbin{\phi_W}) = r^{Ff}, F ( r f ) ∙ ϕ f , id W ∙ ( F f ϕ W ) = r F f , ϕ h , g ∘ f ∙ ( F h ϕ g , f ) ∙ a F h , F g , F f = F ( a h , g , f ) ∙ ϕ h ∘ g , f ∙ ( ϕ h , g F f ) , \phi_{h,g\circ f} \mathbin{\bullet} (Fh \mathbin{\phi_{g,f}}) \mathbin{\bullet} a_{Fh,Fg,Ff} = F(a_{h,g,f}) \mathbin{\bullet} \phi_{h\circ g,f} \mathbin{\bullet} (\phi_{h,g} \mathbin{Ff}), ϕ h , g ∘ f ∙ ( F h ϕ g , f ) ∙ a F h , F g , F f = F ( a h , g , f ) ∙ ϕ h ∘ g , f ∙ ( ϕ h , g F f ) , for all f : W → X f : W \to X f : W → X , g : X → Y g : X \to Y g : X → Y and h : Y → Z h : Y \to Z h : Y → Z . A contravariant pseudofunctor G : B → C G : \mathcal{B} \to\mathcal{C} G : B → C is a pseudofunctor G : B o p → C G : \mathcal{B}^{\mathrm{op}} \to\mathcal{C} G : B op → C , where B o p \mathcal{B}^{\mathrm{op}} B op is the bicategory defined by setting B o p ( X , Y ) : = B ( Y , X ) \mathcal{B}^{\mathrm{op}}(X,Y) := \mathcal{B}(Y,X) B op ( X , Y ) := B ( Y , X ) , using the inverse associator as the new associator, and exchanging left and right unitor. Its compositor and unitor are invertible 2-morphisms G f ∘ G g ⇒ G ( g ∘ f ) Gf \circ Gg \Rightarrow G(g \circ f) G f ∘ G g ⇒ G ( g ∘ f ) and id G X ⇒ G ( id X ) \operatorname{id}_{GX} \Rightarrow G(\operatorname{id}_X) id GX ⇒ G ( id X ) .
For the semantic duals, we will present these structural morphisms by their inverses. In this presentation, the axioms are spelled out more explicitly in Appendix B, Definition B.0.3.
Definition 3.0.3 (Local Faithfulness). A (contravariant) pseudofunctor F : B → C F : \mathcal{B} \to\mathcal{C} F : B → C is locally faithful if it is injective on 2-morphisms, that is, iff F ( η ) = F ( ζ ) F(\eta) = F(\zeta) F ( η ) = F ( ζ ) implies η = ζ \eta= \zeta η = ζ for all parallel 2-morphisms η , ζ : f ⇒ g \eta,\zeta: f \Rightarrow g η , ζ : f ⇒ g .
Composition of Piecewise Interpretations
As a next step, we will define the composition of piecewise interpretations. That is, we will define the object part ∘ \circ ∘ of the composition functor one needs to obtain a bicategory.
Definition 3.1.1 (Composition of Translations). Let f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 and g : L 2 → L 3 g : \mathcal{L}_2 \to\mathcal{L}_3 g : L 2 → L 3 be piecewise translations. The reconstrual underlying their composition g ∘ f g \circ f g ∘ f is given by the following data: The sort function ( g ∘ f ) S (g \circ f)_{\mathcal{S}} ( g ∘ f ) S assigns to each sort the family σ g ∘ f ( k , l ˉ ) : = ( σ g k ) l ˉ \sigma_{g \circ f}^{(k,\bar l)} := (\sigma_g^k)_{\bar l} σ g ∘ f ( k , l ˉ ) := ( σ g k ) l ˉ with index set P σ g ∘ f : = { ( k , l ˉ ) ∣ k ∈ P σ f , l ˉ ∈ P σ g k g } P_{\sigma}^{g\circ f} := \{(k,\bar l) \mid k \in P_{\sigma}^{f},\ \bar l \in P_{\sigma_g^k}^{g}\} P σ g ∘ f := {( k , l ˉ ) ∣ k ∈ P σ f , l ˉ ∈ P σ g k g } . The variable and domain functions below are indexed by this index set. The variable function ( g ∘ f ) V (g \circ f)_{\mathcal{V}} ( g ∘ f ) V sends a variable x : σ x : \sigma x : σ to the family x g ∘ f ( k , l ˉ ) : = ( x f k ) g l ˉ x_{g\circ f}^{(k,\bar l)} := (x_f^k)_g^{\bar l} x g ∘ f ( k , l ˉ ) := ( x f k ) g l ˉ . The domain function ( g ∘ f ) D (g \circ f)_{\mathcal{D}} ( g ∘ f ) D assigns to each sort σ ∈ S 1 \sigma\in\mathcal{S}_1 σ ∈ S 1 the family D g ∘ f σ , ( k , l ˉ ) : = ( D g f , k ) g l ˉ ∧ D g , l ˉ g , x ˉ ( D g f , k ) D_{g\circ f}^{\sigma,(k,\bar l)} := (D_g^{f,k})_g^{\bar l} \land D_{g,\bar l}^{g,\bar x(D_g^{f,k})} D g ∘ f σ , ( k , l ˉ ) := ( D g f , k ) g l ˉ ∧ D g , l ˉ g , x ˉ ( D g f , k ) . The composite relation function ( g ∘ f ) R (g \circ f)_{\mathcal{R}} ( g ∘ f ) R sends a relation symbol R ∈ R 1 R \in\mathcal{R}_1 R ∈ R 1 of arity σ ˉ = ( σ 1 , … , σ n ) \bar\sigma= (\sigma_1,\ldots,\sigma_n) σ ˉ = ( σ 1 , … , σ n ) to the family of formulae R g ∘ f ( ( k 1 , l ˉ 1 ) , … , ( k n , l ˉ n ) ) : = ( R f k ˉ ) g l ˉ R_{g\circ f}^{((k_1,\bar l_1),\ldots,(k_n,\bar l_n))} := (R_f^{\bar k})_g^{\bar l} R g ∘ f (( k 1 , l ˉ 1 ) , … , ( k n , l ˉ n )) := ( R f k ˉ ) g l ˉ , where k ˉ : = ( k 1 , … , k n ) \bar k := (k_1,\ldots,k_n) k ˉ := ( k 1 , … , k n ) and l ˉ : = l ˉ 1 × ⋯ × l ˉ n \bar l := \bar l_1 \times\cdots\times\bar l_n l ˉ := l ˉ 1 × ⋯ × l ˉ n .
One can easily check that these data satisfy the no-overlap condition, duplicate-freeness and the sortal conditions of Definition 2.1.1. In the following, we will need that the induced composed translation schematically handles general formulae just as it does relation symbols. To that end, one first needs to establish in a technical lemma that substitution and translation commute in an appropriate sense, which is done in Lemma A.0.7 in Appendix § 6 .
Proposition 3.1.2 (Composition for Complex Formulae). Let f : L 1 → L 2 f : \mathcal{L}_1 \to\mathcal{L}_2 f : L 1 → L 2 and g : L 2 → L 3 g : \mathcal{L}_2 \to\mathcal{L}_3 g : L 2 → L 3 be piecewise translations , and let φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ be an L 1 \mathcal{L}_1 L 1 -formula with ∣ σ ˉ ∣ = n |\bar{\sigma}| = n ∣ σ ˉ ∣ = n . For any composite piece m ˉ = ( ( k 1 , l ˉ 1 ) , … , ( k n , l ˉ n ) ) ∈ P σ ˉ g ∘ f \bar{m} = ((k_1,\bar{l}_1),\ldots,(k_n,\bar{l}_n)) \in P_{\bar{\sigma}}^{g\circ f} m ˉ = (( k 1 , l ˉ 1 ) , … , ( k n , l ˉ n )) ∈ P σ ˉ g ∘ f , we set
m ˉ f : = ( k 1 , … , k n ) ∈ P σ ˉ f , m ˉ g : = l ˉ 1 × ⋯ × l ˉ n ∈ P σ ˉ f g . (41) \bar{m}_f := (k_1,\ldots,k_n) \in P_{\bar{\sigma}}^f,\qquad\bar{m}_g := \bar{l}_1 \times\cdots\times\bar{l}_n \in P_{\bar{\sigma}_f}^{g}. \tag*{(41)} m ˉ f := ( k 1 , … , k n ) ∈ P σ ˉ f , m ˉ g := l ˉ 1 × ⋯ × l ˉ n ∈ P σ ˉ f g . (41) Then the formulae φ g ∘ f m ˉ \varphi_{g\circ f}^{\bar{m}} φ g ∘ f m ˉ and ( φ f m ˉ f ) g m ˉ g (\varphi_f^{\bar{m}_f})_g^{\bar{m}_g} ( φ f m ˉ f ) g m ˉ g have the same canonical context, and
⊢ L ( Σ 3 ) φ g ∘ f m ˉ ↔ ( φ f m ˉ f ) g m ˉ g . (42) \vdash_{\mathcal{L}(\Sigma_3)} \varphi_{g\circ f}^{\bar{m}} \leftrightarrow(\varphi_f^{\bar{m}_f})_g^{\bar{m}_g}. \tag*{(42)} ⊢ L ( Σ 3 ) φ g ∘ f m ˉ ↔ ( φ f m ˉ f ) g m ˉ g . (42) We write φ ⊣ ⊢ T ψ \varphi\dashv\vdash_T \psi φ ⊣ ⊢ T ψ to denote that φ \varphi φ and ψ \psi ψ have the same canonical context and are provably equivalent with respecto to the theory T T T .
Proof. See Appendix A.
Remark 3.1.3 (Composite Pieces). In the situation above, since σ ˉ f m ˉ f = σ 1 , f k 1 × ⋯ × σ n , f k n \bar{\sigma}_f^{\bar{m}_f} = \sigma_{1,f}^{k_1} \times\cdots\times\sigma_{n,f}^{k_n} σ ˉ f m ˉ f = σ 1 , f k 1 × ⋯ × σ n , f k n , the assignment m ˉ ↦ ( m ˉ f , m ˉ g ) \bar{m} \mapsto(\bar{m}_f,\bar{m}_g) m ˉ ↦ ( m ˉ f , m ˉ g ) is a bijection from P σ ˉ g ∘ f P_{\bar{\sigma}}^{g\circ f} P σ ˉ g ∘ f to the set of pairs ( k ˉ , l ˉ ) (\bar{k},\bar{l}) ( k ˉ , l ˉ ) with k ˉ ∈ P σ ˉ f \bar{k} \in P_{\bar{\sigma}}^f k ˉ ∈ P σ ˉ f and l ˉ ∈ P σ ˉ k ˉ g \bar{l} \in P_{\bar{\sigma}_{\bar{k}}^g} l ˉ ∈ P σ ˉ k ˉ g .
For this composition operation to fit into the bicategory we are developing, the composition of two piecewise interpretations also needs to be a piecewise interpretation, which follows from the preceding proposition specialized to sentences:
Corollary 3.1.4 (Composition of Interpretations is an Interpretation). Let f : T 1 → T 2 f : T_1 \to T_2 f : T 1 → T 2 and g : T 2 → T 3 g : T_2 \to T_3 g : T 2 → T 3 be piecewise interpretations . Then the composition g ∘ f g\circ f g ∘ f is an interpretation g ∘ f : T 1 → T 3 g \circ f : T_1 \to T_3 g ∘ f : T 1 → T 3 .
Moreover, Proposition 3.1.2 allows us to explicitly describe the semantic dual of the composition of two piecewise interpretations.
Lemma 3.1.5 (Extensions under Composition). Let f : T 1 → T 2 f : T_1 \to T_2 f : T 1 → T 2 and g : T 2 → T 3 g : T_2 \to T_3 g : T 2 → T 3 be piecewise interpretations . Let M ⊨ T 3 M \models T_3 M ⊨ T 3 , let σ ˉ ∈ S 1 ∗ \bar{\sigma} \in S_1^* σ ˉ ∈ S 1 ∗ , and let m ˉ ∈ P σ ˉ g ∘ f \bar{m} \in P_{\bar{\sigma}}^{g\circ f} m ˉ ∈ P σ ˉ g ∘ f with m ˉ f \bar{m}_f m ˉ f and m ˉ g \bar{m}_g m ˉ g as in Proposition 3.1.2. Then
M ( D σ ˉ g ∘ f , m ˉ ) = { a ˉ ∈ M ( D σ ˉ f g , m ˉ g ) ∣ π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ∈ g ∗ M ( D σ ˉ f , m ˉ f ) } . (43) M\left(D_{\bar{\sigma}}^{g\circ f,\bar{m}}\right)=\left\{\bar{a}\in M\left(D_{\bar{\sigma}_f}^{g,\bar{m}_g}\right)\mathrel{\left|\,\pi_{\bar{\sigma}_{\bar{m}_f}}^g(\bar{m}_g,\bar{a})\in g^*M\left(D_{\bar{\sigma}}^{f,\bar{m}_f}\right)\right.}\right\}. \tag*{(43)} M ( D σ ˉ g ∘ f , m ˉ ) = { a ˉ ∈ M ( D σ ˉ f g , m ˉ g ) π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ∈ g ∗ M ( D σ ˉ f , m ˉ f ) } . (43) and for every L 1 \mathcal{L}_1 L 1 -formula φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ and every a ˉ ∈ M ( D σ ˉ g ∘ f , m ˉ ) \bar{a} \in M(D_{\bar{\sigma}}^{g\circ f,\bar{m}}) a ˉ ∈ M ( D σ ˉ g ∘ f , m ˉ ) ,
π σ ˉ g ∘ f ( m ˉ , a ˉ ) ∈ ( g ∘ f ) ∗ M ( φ ) ⟺ π σ ˉ f ( m ˉ f , π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ) ∈ f ∗ ( g ∗ M ) ( φ ) . (44) \pi_{\bar{\sigma}}^{g\circ f}(\bar{m},\bar{a})\in(g\circ f)^*M(\varphi)\Longleftrightarrow\pi_{\bar{\sigma}}^f\left(\bar{m}_f,\pi_{\bar{\sigma}_{\bar{m}_f}}^g(\bar{m}_g,\bar{a})\right)\in f^*(g^*M)(\varphi). \tag*{(44)} π σ ˉ g ∘ f ( m ˉ , a ˉ ) ∈ ( g ∘ f ) ∗ M ( φ ) ⟺ π σ ˉ f ( m ˉ f , π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ) ∈ f ∗ ( g ∗ M ) ( φ ) . (44) Proof. The domain equation (3.1.3) follows from the definition of the composite domain formula and (2.4.16) for g g g . For the equivalence for formulae (3.1.4), let φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ be an L 1 \mathcal{L}_1 L 1 -formula, m ˉ ∈ P σ ˉ g ∘ f \bar{m} \in P_{\bar{\sigma}}^{g\circ f} m ˉ ∈ P σ ˉ g ∘ f , and a ˉ ∈ M ( D σ ˉ g ∘ f , m ˉ ) \bar{a} \in M(D_{\bar{\sigma}}^{g\circ f,\bar{m}}) a ˉ ∈ M ( D σ ˉ g ∘ f , m ˉ ) . We get the following chain of equivalences:
π σ ˉ g ∘ f ( m ˉ , a ˉ ) ∈ ( g ∘ f ) ∗ M ( φ ) ⟺ ( 2.4.16 ) a ˉ ∈ M ( φ g ∘ f m ˉ ) ⟺ Proposition 3.1.2 a ˉ ∈ M ( ( φ f m ˉ f ) g m ˉ g ) (45) \pi_{\bar{\sigma}}^{g\circ f}(\bar{m},\bar{a})\in(g\circ f)^*M(\varphi)
\underset{(2.4.16)}{\Longleftrightarrow}
\bar{a}\in M(\varphi_{g\circ f}^{\bar{m}})
\underset{\text{Proposition 3.1.2}}{\Longleftrightarrow}
\bar{a}\in M\left((\varphi_f^{\bar{m}_f})_g^{\bar{m}_g}\right) \tag*{(45)} π σ ˉ g ∘ f ( m ˉ , a ˉ ) ∈ ( g ∘ f ) ∗ M ( φ ) ( 2.4.16 ) ⟺ a ˉ ∈ M ( φ g ∘ f m ˉ ) Proposition 3.1.2 ⟺ a ˉ ∈ M ( ( φ f m ˉ f ) g m ˉ g ) (45) ⟺ ( 2.4.16 ) π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ∈ g ∗ M ( φ f m ˉ f ) (46) \underset{(2.4.16)}{\Longleftrightarrow}
\pi_{\bar{\sigma}_{\bar{m}_f}}^g(\bar{m}_g,\bar{a})\in g^*M(\varphi_f^{\bar{m}_f}) \tag*{(46)} ( 2.4.16 ) ⟺ π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ∈ g ∗ M ( φ f m ˉ f ) (46) ⟺ ( 2.4.16 ) π σ ˉ f ( m ˉ f , π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ) ∈ f ∗ ( g ∗ M ) ( φ ) . (47) \underset{(2.4.16)}{\Longleftrightarrow}
\pi_{\bar{\sigma}}^f\left(\bar{m}_f,\pi_{\bar{\sigma}_{\bar{m}_f}}^g(\bar{m}_g,\bar{a})\right)\in f^*(g^*M)(\varphi). \tag*{(47)} ( 2.4.16 ) ⟺ π σ ˉ f ( m ˉ f , π σ ˉ m ˉ f g ( m ˉ g , a ˉ ) ) ∈ f ∗ ( g ∗ M ) ( φ ) . (47)
An immediate corollary, using Proposition 2.4.10 , is that composition descends to provable equivalence classes of interpretations. Moreover, this explicit description allows us to construct the compositor we will need to think of the semantic dual as being part of a pseudofunctor:
Proposition 3.1.6 (Compositor for Semantic Duals). Let f : T 1 → T 2 f:T_{1}\to T_{2} f : T 1 → T 2 and g : T 2 → T 3 g:T_{2}\to T_{3} g : T 2 → T 3 be piecewise interpretations. Then there is a natural isomorphism
α g , f : ( g ∘ f ) ∗ ⇒ f ∗ g ∗ : Mod ( T 3 ) → Mod ( T 1 ) . (48) \alpha_{g,f}:(g\circ f)^{*}\Rightarrow f^{*}g^{*}:\operatorname{Mod}(T_{3})\to\operatorname{Mod}(T_{1}). \tag*{(48)} α g , f : ( g ∘ f ) ∗ ⇒ f ∗ g ∗ : Mod ( T 3 ) → Mod ( T 1 ) . (48) For a T 3 T_{3} T 3 -model M M M , the component α g , f , M : ( g ∘ f ) ∗ M → f ∗ ( g ∗ M ) \alpha_{g,f,M}:(g\circ f)^{*}M\to f^{*}(g^{*}M) α g , f , M : ( g ∘ f ) ∗ M → f ∗ ( g ∗ M ) is the isomorphism whose σ \sigma σ -component is defined by
π σ g ∘ f ( ( k , l ˉ ) , a ˉ ) ⟼ π σ f ( k , π σ k g ( l ˉ , a ˉ ) ) . (49) \pi_{\sigma}^{g\circ f}((k,\bar{l}),\bar{a})\longmapsto\pi_{\sigma}^{f}(k,\pi_{\sigma_{k}}^{g}(\bar{l},\bar{a})). \tag*{(49)} π σ g ∘ f (( k , l ˉ ) , a ˉ ) ⟼ π σ f ( k , π σ k g ( l ˉ , a ˉ )) . (49) Here k ∈ P σ f k\in P_{\sigma}^{f} k ∈ P σ f , l ˉ ∈ P σ k g \bar{l}\in P_{\sigma_{k}}^{g} l ˉ ∈ P σ k g and a ˉ ∈ M ( D σ g ∘ f , ( k , l ˉ ) ) \bar{a}\in M(D_{\sigma}^{g\circ f,(k,\bar{l})}) a ˉ ∈ M ( D σ g ∘ f , ( k , l ˉ ) ) .
Proof. Recall that every x ∈ ( g ∘ f ) ∗ M ( σ ) x\in(g\circ f)^{*}M(\sigma) x ∈ ( g ∘ f ) ∗ M ( σ ) is of the form x = π σ g ∘ f ( ( k , l ˉ ) , a ˉ ) x=\pi_{\sigma}^{g\circ f}((k,\bar{l}),\bar{a}) x = π σ g ∘ f (( k , l ˉ ) , a ˉ ) with ( k , l ˉ ) ∈ P σ g ∘ f (k,\bar{l})\in P_{\sigma}^{g\circ f} ( k , l ˉ ) ∈ P σ g ∘ f and a ˉ ∈ M ( D σ g ∘ f , ( k , l ˉ ) ) \bar{a}\in M(D_{\sigma}^{g\circ f,(k,\bar{l})}) a ˉ ∈ M ( D σ g ∘ f , ( k , l ˉ ) ) , whence α g , f \alpha_{g,f} α g , f is total. Well-definedness and injectivity of ( α g , f , M ) σ (\alpha_{g,f,M})_{\sigma} ( α g , f , M ) σ follow from (3.1.4) applied to = σ =_{\sigma} = σ . Surjectivity follows from construction and (3.1.3). That α g , f , M \alpha_{g,f,M} α g , f , M is an elementary embedding follows immediately from the collection of equivalences given by (3.1.4) for all formulae φ ∈ L 1 \varphi\in L_{1} φ ∈ L 1 .
To demonstrate naturality, consider an elementary embedding j : M → N j:M\to N j : M → N and compute:
f ∗ ( g ∗ j ) ( π σ f ( k , π σ k g ( l ˉ , a ˉ ) ) ) = π σ f ( k , π σ k g ( l ˉ , j ( a ˉ ) ) ) (50) f^{*}(g^{*}j)\left(\pi_{\sigma}^{f}\left(k,\pi_{\sigma_{k}}^{g}(\bar{l},\bar{a})\right)\right)=\pi_{\sigma}^{f}\left(k,\pi_{\sigma_{k}}^{g}(\bar{l},j(\bar{a}))\right) \tag*{(50)} f ∗ ( g ∗ j ) ( π σ f ( k , π σ k g ( l ˉ , a ˉ ) ) ) = π σ f ( k , π σ k g ( l ˉ , j ( a ˉ )) ) (50) = α g , f , N , σ ( π σ g ∘ f ( ( k , l ˉ ) , j ( a ˉ ) ) ) = α g , f , N , σ ∘ ( g ∘ f ) ∗ j ( π σ g ∘ f ( ( k , l ˉ ) , a ˉ ) ) . (51) =\alpha_{g,f,N,\sigma}\left(\pi_{\sigma}^{g\circ f}((k,\bar{l}),j(\bar{a}))\right)=\alpha_{g,f,N,\sigma}\circ(g\circ f)^{*}j\left(\pi_{\sigma}^{g\circ f}((k,\bar{l}),\bar{a})\right). \tag*{(51)} = α g , f , N , σ ( π σ g ∘ f (( k , l ˉ ) , j ( a ˉ )) ) = α g , f , N , σ ∘ ( g ∘ f ) ∗ j ( π σ g ∘ f (( k , l ˉ ) , a ˉ ) ) . (51)
Moving on to the identity of our bicategory, the identity interpretation indeed behaves as an identity with respect to composition, up to canonical i i i -isomorphisms, the unitors of our bicategory:
Definition 3.1.7 (Left and Right Unitors). Let f : T → U f:T\to U f : T → U be an interpretation. The left unitor l f : id U ∘ f ⇒ f l^{f}: \operatorname{id}_{U}\circ f\Rightarrow f l f : id U ∘ f ⇒ f and right unitor r f : f ∘ id T ⇒ f r^{f}:f\circ\operatorname{id}_{T}\Rightarrow f r f : f ∘ id T ⇒ f of f f f are i i i -isomorphisms defined via bijections of pieces according to Proposition 2.6.9 :
l σ f : P σ id U ∘ f → P σ f r σ f : P σ f ∘ id T → P σ f (52) l_{\sigma}^{f}:P_{\sigma}^{\operatorname{id}_{U}\circ f}\to P_{\sigma}^{f}
\qquad
r_{\sigma}^{f}:P_{\sigma}^{f\circ\operatorname{id}_{T}}\to P_{\sigma}^{f} \tag*{(52)} l σ f : P σ id U ∘ f → P σ f r σ f : P σ f ∘ id T → P σ f (52) ( k , 0 ∣ σ k f ∣ ) ⟼ k ( 0 , k ) ⟼ k (53) (k,0^{|\sigma_{k}^{f}|})\longmapsto k
\qquad
(0,k)\longmapsto k \tag*{(53)} ( k , 0 ∣ σ k f ∣ ) ⟼ k ( 0 , k ) ⟼ k (53)
Here 0 ∣ σ k f ∣ 0^{|\sigma_{k}^{f}|} 0 ∣ σ k f ∣ denotes the unique element of ∏ i = 1 ∣ σ k f ∣ P ( σ k f ) i id U \prod_{i=1}^{|\sigma_{k}^{f}|}P_{(\sigma_{k}^{f})_{i}}^{\operatorname{id}_{U}} ∏ i = 1 ∣ σ k f ∣ P ( σ k f ) i id U .
That these bijections satisfy the conditions of Proposition 2.6.9 follows easily from the definition and provable triviality of id U \operatorname{id}_{U} id U and id T \operatorname{id}_{T} id T . Under the semantic dual, the unitors are appropriately coherent with the compositors and unitors of the semantic dual:
Lemma 3.1.8 (Coherence of the Unitors). Let f : T → U f:T \to U f : T → U be a piecewise interpretation and recall that for a theory S S S , α S : ( id S ) ∗ ⇒ id Mod ( S ) \alpha_S:(\operatorname{id}_S)^* \Rightarrow\operatorname{id}_{\operatorname{Mod}(S)} α S : ( id S ) ∗ ⇒ id Mod ( S ) is the natural isomorphism defined by π σ id ( 0 , a ) ↦ a \pi_\sigma^{\operatorname{id}}(0,a) \mapsto a π σ id ( 0 , a ) ↦ a . Then
( l f ) ∗ = ( f ∗ α U ) ∙ α id U , f and ( r f ) ∗ = ( α T f ∗ ) ∙ α f , id T . (54) (l^f)_* = (f^* \alpha_U) \mathbin{\bullet} \alpha_{\operatorname{id}_U,f}
\qquad\text{and}\qquad
(r^f)_* = (\alpha_T f^*) \mathbin{\bullet} \alpha_{f,\operatorname{id}_T}. \tag*{(54)} ( l f ) ∗ = ( f ∗ α U ) ∙ α id U , f and ( r f ) ∗ = ( α T f ∗ ) ∙ α f , id T . (54) Proof. Let M ⊨ U M \models U M ⊨ U and σ ∈ S T \sigma\in\mathcal{S}_T σ ∈ S T . We first consider the left unitor: Given an element x x x of ( id U ∘ f ) ∗ M ( σ ) (\operatorname{id}_U \circ f)^*M(\sigma) ( id U ∘ f ) ∗ M ( σ ) , x x x is of the form π σ id U ∘ f ( ( k , 0 ∣ σ f k ∣ ) , a ˉ ) \pi_\sigma^{\operatorname{id}_U\circ f}((k,0^{|\sigma^k_f|}),\bar a) π σ id U ∘ f (( k , 0 ∣ σ f k ∣ ) , a ˉ ) with ( k , 0 ˉ , a ˉ ) ∈ P σ f × { 0 } ∣ σ f k ∣ × M ( D σ f , k ) (k,\bar0,\bar a) \in P^f_\sigma\times\{0\}^{|\sigma^k_f|} \times M(D^{f,k}_\sigma) ( k , 0 ˉ , a ˉ ) ∈ P σ f × { 0 } ∣ σ f k ∣ × M ( D σ f , k ) . By Proposition 2.6.9 , ( l f ) M ( π σ id U ∘ f ( ( k , 0 ∣ σ f k ∣ ) , a ˉ ) ) = π σ f ( k , a ˉ ) (l^f)_M(\pi_\sigma^{\operatorname{id}_U\circ f}((k,0^{|\sigma^k_f|}),\bar a))=\pi^f_\sigma(k,\bar a) ( l f ) M ( π σ id U ∘ f (( k , 0 ∣ σ f k ∣ ) , a ˉ )) = π σ f ( k , a ˉ ) . On the right-hand side of the left equation, by Proposition 3.1.6 and the definitions of whiskering and α U \alpha_U α U ,
( f ∗ α U ) M ∘ α id U , f , M ( π σ id U ∘ f ( ( k , 0 ∣ σ f k ∣ ) , a ˉ ) ) = ( f ∗ α U ) M ( π σ f ( k , π σ f k id U ( 0 ∣ σ f k ∣ , a ˉ ) ) ) = π σ f ( k , a ˉ ) . (55) (f^*\alpha_U)_M \mathbin{\circ} \alpha_{\operatorname{id}_U,f,M}(\pi_\sigma^{\operatorname{id}_U\circ f}((k,0^{|\sigma^k_f|}),\bar a))
=(f^*\alpha_U)_M(\pi^f_\sigma(k,\pi^{\operatorname{id}_U}_{\sigma^k_f}(0^{|\sigma^k_f|},\bar a)))
=\pi^f_\sigma(k,\bar a). \tag*{(55)} ( f ∗ α U ) M ∘ α id U , f , M ( π σ id U ∘ f (( k , 0 ∣ σ f k ∣ ) , a ˉ )) = ( f ∗ α U ) M ( π σ f ( k , π σ f k id U ( 0 ∣ σ f k ∣ , a ˉ ))) = π σ f ( k , a ˉ ) . (55) Moving on to the right unitor, given an element x x x of ( f ∘ id T ) ∗ M ( σ ) (f \circ\operatorname{id}_T)^*M(\sigma) ( f ∘ id T ) ∗ M ( σ ) , x x x is of the form π σ f ∘ id T ( ( 0 , k ) , a ˉ ) \pi^{f\circ\operatorname{id}_T}_\sigma((0,k),\bar a) π σ f ∘ id T (( 0 , k ) , a ˉ ) with ( k , a ˉ ) ∈ P σ f × M ( D σ f , k ) (k,\bar a) \in P^f_\sigma\times M(D^{f,k}_\sigma) ( k , a ˉ ) ∈ P σ f × M ( D σ f , k ) , and ( r f ) M ( x ) = π σ f ( k , a ˉ ) (r^f)_M(x)=\pi^f_\sigma(k,\bar a) ( r f ) M ( x ) = π σ f ( k , a ˉ ) by Proposition 2.6.9 . On the right-hand side of the right equation, we see that
( α T f ∗ ) M ( α f , id T , M ( π σ f ∘ id T ( ( 0 , k ) , a ˉ ) ) ) = ( α T f ∗ ) M ( π σ id T ( 0 , π σ f ( k , a ˉ ) ) ) = π σ f ( k , a ˉ ) . (56) (\alpha_T f^*)_M\left(\alpha_{f,\operatorname{id}_T,M}\left(\pi^{f\circ\operatorname{id}_T}_\sigma((0,k),\bar a)\right)\right)
=(\alpha_T f^*)_M\left(\pi^{\operatorname{id}_T}_\sigma(0,\pi^f_\sigma(k,\bar a))\right)
=\pi^f_\sigma(k,\bar a). \tag*{(56)} ( α T f ∗ ) M ( α f , id T , M ( π σ f ∘ id T (( 0 , k ) , a ˉ ) ) ) = ( α T f ∗ ) M ( π σ id T ( 0 , π σ f ( k , a ˉ )) ) = π σ f ( k , a ˉ ) . (56)
Associator for Piecewise Interpretations
The next ingredient for our bicategory is an associator for each appropriate triple of interpretations, making composition associative up to a coherent i i i -isomorphism. We thereby answer a small question due to Visser, about the sense in which piecewise interpretations compose associatively (personal communication).
Definition 3.2.1 (Associator for Interpretations). Let f : T 1 → T 2 f:T_1 \to T_2 f : T 1 → T 2 , g : T 2 → T 3 g:T_2 \to T_3 g : T 2 → T 3 and h : T 3 → T 4 h:T_3 \to T_4 h : T 3 → T 4 be piecewise interpretations. Given a sort σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 , set τ i = ( σ f k ) i \tau_i=(\sigma^k_f)_i τ i = ( σ f k ) i for each k ∈ P σ f k \in P^f_\sigma k ∈ P σ f and 1 ≤ i ≤ ∣ σ f k ∣ 1 \le i \le|\sigma^k_f| 1 ≤ i ≤ ∣ σ f k ∣ . Then, iterating the definition of composition yields:
P σ ( h ∘ g ) ∘ f = { ( k , ( ( l 1 , m ˉ 1 ) , … , ( l ∣ σ f k ∣ , m ˉ ∣ σ f k ∣ ) ) ) | k ∈ P σ f , l i ∈ P τ i g , m ˉ i ∈ P ( τ i ) g l i h with 1 ≤ i ≤ ∣ σ f k ∣ } (57) P^{(h\circ g)\circ f}_\sigma
=
\left\{
\left(k,\left((l_1,\bar m_1),\ldots,(l_{|\sigma^k_f|},\bar m_{|\sigma^k_f|})\right)\right)
\;\middle|\;
\begin{array}{l}
k\in P^f_\sigma,\ l_i\in P^g_{\tau_i},\ \bar m_i\in P^h_{(\tau_i)^{l_i}_g}\\
\text{with }1\le i\le|\sigma^k_f|
\end{array}
\right\} \tag*{(57)} P σ ( h ∘ g ) ∘ f = { ( k , ( ( l 1 , m ˉ 1 ) , … , ( l ∣ σ f k ∣ , m ˉ ∣ σ f k ∣ ) ) ) k ∈ P σ f , l i ∈ P τ i g , m ˉ i ∈ P ( τ i ) g l i h with 1 ≤ i ≤ ∣ σ f k ∣ } (57)
For the other bracketing, one computes:
P σ h ∘ ( g ∘ f ) = { ( ( k , l ˉ ) , m ˉ 1 × ⋯ × m ˉ ∣ σ f k ∣ ) | k ∈ P σ f , l i ∈ P τ i g , m ˉ i ∈ P ( τ i ) g l i h with 1 ≤ i ≤ ∣ σ f k ∣ } . (58) P^{h\circ(g\circ f)}_\sigma
=
\left\{
\left((k,\bar l),\bar m_1\times\cdots\times\bar m_{|\sigma^k_f|}\right)
\;\middle|\;
\begin{array}{l}
k\in P^f_\sigma,\ l_i\in P^g_{\tau_i},\ \bar m_i\in P^h_{(\tau_i)^{l_i}_g}\\
\text{with }1\le i\le|\sigma^k_f|
\end{array}
\right\}. \tag*{(58)} P σ h ∘ ( g ∘ f ) = { ( ( k , l ˉ ) , m ˉ 1 × ⋯ × m ˉ ∣ σ f k ∣ ) k ∈ P σ f , l i ∈ P τ i g , m ˉ i ∈ P ( τ i ) g l i h with 1 ≤ i ≤ ∣ σ f k ∣ } . (58)
Inspecting these two sets, we see that they can be matched up bijectively by regrouping and concatenating pieces. The associator a h , g , f a_{h,g,f} a h , g , f is the i i i -isomorphism induced by the following family of bijections, as in Proposition 2.6.9 :
( a h , g , f ) σ : P σ ( h ∘ g ) ∘ f ⟶ P σ h ∘ ( g ∘ f ) (59) (a_{h,g,f})_\sigma:P^{(h\circ g)\circ f}_\sigma\longrightarrow P^{h\circ(g\circ f)}_\sigma \tag*{(59)} ( a h , g , f ) σ : P σ ( h ∘ g ) ∘ f ⟶ P σ h ∘ ( g ∘ f ) (59)
( k , ( ( l 1 , m ˉ 1 ) , … , ( l ∣ σ f k ∣ , m ˉ ∣ σ f k ∣ ) ) ) ⟼ ( ( k , l ˉ ) , m ˉ 1 × ⋯ × m ˉ ∣ σ f k ∣ ) , (60) \left(k,\left((l_1,\bar m_1),\ldots,(l_{|\sigma^k_f|},\bar m_{|\sigma^k_f|})\right)\right)
\longmapsto
\left((k,\bar l),\bar m_1\times\cdots\times\bar m_{|\sigma^k_f|}\right), \tag*{(60)} ( k , ( ( l 1 , m ˉ 1 ) , … , ( l ∣ σ f k ∣ , m ˉ ∣ σ f k ∣ ) ) ) ⟼ ( ( k , l ˉ ) , m ˉ 1 × ⋯ × m ˉ ∣ σ f k ∣ ) , (60)
where l ˉ = ( l 1 , … , l ∣ σ f k ∣ ) \bar l=(l_1,\ldots,l_{|\sigma^k_f|}) l ˉ = ( l 1 , … , l ∣ σ f k ∣ ) .
Proposition 3.2.2 (Associativity of Composition). Let f : T 1 → T 2 f:T_1\to T_2 f : T 1 → T 2 , g : T 2 → T 3 g:T_2\to T_3 g : T 2 → T 3 and h : T 3 → T 4 h:T_3\to T_4 h : T 3 → T 4 be piecewise interpretations. Then the associator a h , g , f : ( h ∘ g ) ∘ f ⇒ h ∘ ( g ∘ f ) a_{h,g,f}:(h\circ g)\circ f\Rightarrow h\circ(g\circ f) a h , g , f : ( h ∘ g ) ∘ f ⇒ h ∘ ( g ∘ f ) is an i i i -isomorphism.
Proof. We verify the instances of the conditions of Proposition 2.6.9 for a bijection ( a h , g , f ) σ (a_{h,g,f})_\sigma ( a h , g , f ) σ for an arbitrary sort σ \sigma σ . Write p = ( k , ( ( l 1 , m ˉ 1 ) , … , ( l n , m ˉ n ) ) ) ∈ P σ ( h ∘ g ) ∘ f p=(k,((l_1,\bar{m}_1),\ldots,(l_n,\bar{m}_n)))\in P^{(h\circ g)\circ f}_\sigma p = ( k , (( l 1 , m ˉ 1 ) , … , ( l n , m ˉ n ))) ∈ P σ ( h ∘ g ) ∘ f with n = ∣ σ f k ∣ n=|\sigma^k_f| n = ∣ σ f k ∣ , and abbreviate τ i : = ( σ f k ) l i \tau_i:=(\sigma^k_f)_{l_i} τ i := ( σ f k ) l i , l ˉ = ( l 1 , … , l n ) \bar{l}=(l_1,\ldots,l_n) l ˉ = ( l 1 , … , l n ) , and m ˉ = m ˉ 1 × ⋯ × m ˉ n \bar{m}=\bar{m}_1\times\cdots\times\bar{m}_n m ˉ = m ˉ 1 × ⋯ × m ˉ n .
(Sort Condition): Iterating the definition of the sort function of a composition on each side yields:
σ ( h ∘ g ) ∘ f p = ( τ 1 ) h ∘ g ( l 1 , m ˉ 1 ) × ⋯ × ( τ n ) h ∘ g ( l n , m ˉ n ) = ( ( τ 1 ) g l 1 ) h m ˉ 1 × ⋯ × ( ( τ n ) g l n ) h m ˉ n (61) \sigma^{p}_{(h\circ g)\circ f}=(\tau_1)^{(l_1,\bar{m}_1)}_{h\circ g}\times\cdots\times(\tau_n)^{(l_n,\bar{m}_n)}_{h\circ g}=((\tau_1)^{l_1}_g)^{\bar{m}_1}_h\times\cdots\times((\tau_n)^{l_n}_g)^{\bar{m}_n}_h \tag*{(61)} σ ( h ∘ g ) ∘ f p = ( τ 1 ) h ∘ g ( l 1 , m ˉ 1 ) × ⋯ × ( τ n ) h ∘ g ( l n , m ˉ n ) = (( τ 1 ) g l 1 ) h m ˉ 1 × ⋯ × (( τ n ) g l n ) h m ˉ n (61) = ( ( τ 1 ) g l 1 × ⋯ × ( τ n ) g l n ) h m ˉ = ( σ g ∘ f ( k , l ˉ ) ) h m ˉ (62) =((\tau_1)^{l_1}_g\times\cdots\times(\tau_n)^{l_n}_g)^{\bar{m}}_h=(\sigma^{(k,\bar{l})}_{g\circ f})^{\bar{m}}_h \tag*{(62)} = (( τ 1 ) g l 1 × ⋯ × ( τ n ) g l n ) h m ˉ = ( σ g ∘ f ( k , l ˉ ) ) h m ˉ (62) = σ h ∘ ( g ∘ f ) a ( p ) (63) =\sigma^{a(p)}_{h\circ(g\circ f)} \tag*{(63)} = σ h ∘ ( g ∘ f ) a ( p ) (63) (Domain Condition): For the domain condition, we compute:
D x ( h ∘ g ) ∘ f , p ⊢ L ( Σ 4 ) ( ( D x f , k ) g l ˉ ) h m ˉ ∧ ⋀ i = 1 n [ ( D ( τ f k ) i g , l i ) h m ˉ i ∧ D ( ( τ f k ) i ) g l i h , m ˉ i ] (64) D^{(h\circ g)\circ f,p}_x\mathrel{\vdash_{L(\Sigma_4)}}((D^{f,k}_x)^{\bar{l}}_g)^{\bar{m}}_h\land\bigwedge_{i=1}^{n}\left[(D^{g,l_i}_{(\tau^k_f)_i})^{\bar{m}_i}_h\land D^{h,\bar{m}_i}_{((\tau^k_f)_i)^{l_i}_g}\right] \tag*{(64)} D x ( h ∘ g ) ∘ f , p ⊢ L ( Σ 4 ) (( D x f , k ) g l ˉ ) h m ˉ ∧ i = 1 ⋀ n [ ( D ( τ f k ) i g , l i ) h m ˉ i ∧ D (( τ f k ) i ) g l i h , m ˉ i ] (64) ⊢ L ( Σ 4 ) ( ( D x f , k ) g l ˉ ∧ ⋀ i = 1 n D ( τ f k ) i g , l i ) h m ˉ ∧ ⋀ i = 1 n D ( ( τ f k ) i ) g l i h , m ˉ i (65) \mathrel{\vdash_{L(\Sigma_4)}}((D^{f,k}_x)^{\bar{l}}_g\land\bigwedge_{i=1}^{n}D^{g,l_i}_{(\tau^k_f)_i})^{\bar{m}}_h\land\bigwedge_{i=1}^{n}D^{h,\bar{m}_i}_{((\tau^k_f)_i)^{l_i}_g} \tag*{(65)} ⊢ L ( Σ 4 ) (( D x f , k ) g l ˉ ∧ i = 1 ⋀ n D ( τ f k ) i g , l i ) h m ˉ ∧ i = 1 ⋀ n D (( τ f k ) i ) g l i h , m ˉ i (65) = D x h ∘ ( g ∘ f ) , a ( p ) , (66) =D^{h\circ(g\circ f),a(p)}_x, \tag*{(66)} = D x h ∘ ( g ∘ f ) , a ( p ) , (66) where the first ⊢ \vdash ⊢ writes out the definition of the composite domain formula and applies Proposition 3.1.2 at the intermediate stage for the left-most conjunct, the second distributes and applies the binary and the context clauses.
(Relation Condition): Let R : σ ˉ R:\bar{\sigma} R : σ ˉ be a relation symbol of L 1 \mathcal{L}_1 L 1 and let q ˉ ∈ P σ ˉ ( h ∘ g ) ∘ f \bar{q}\in P^{(h\circ g)\circ f}_{\bar{\sigma}} q ˉ ∈ P σ ˉ ( h ∘ g ) ∘ f . Applying the piece decomposition from Proposition 3.1.2 first to f f f and h ∘ g h\circ g h ∘ g , then to g g g and h h h , set k ˉ : = q ˉ f \bar{k}:=\bar{q}_f k ˉ := q ˉ f , l ˉ : = ( q ˉ h ∘ g ) g \bar{l}:=(\bar{q}_{h\circ g})_g l ˉ := ( q ˉ h ∘ g ) g and m ˉ : = ( q ˉ h ∘ g ) h \bar{m}:=(\bar{q}_{h\circ g})_h m ˉ := ( q ˉ h ∘ g ) h . By the definition of the extension of a a a on sequences, a ( q ˉ ) a(\bar{q}) a ( q ˉ ) decomposes into the same k ˉ , l ˉ , m ˉ \bar{k},\bar{l},\bar{m} k ˉ , l ˉ , m ˉ . Iterative applications of Proposition 3.1.2 yield
R ( h ∘ g ) ∘ f q ˉ ⊢ L ( Σ 4 ) ( ( R f k ˉ ) g l ˉ ) h m ˉ ⊢ L ( Σ 4 ) R h ∘ ( g ∘ f ) a ( q ˉ ) , (67) R^{\bar{q}}_{(h\circ g)\circ f}\mathrel{\vdash_{L(\Sigma_4)}}((R^{\bar{k}}_f)^{\bar{l}}_g)^{\bar{m}}_h\mathrel{\vdash_{L(\Sigma_4)}}R^{a(\bar{q})}_{h\circ(g\circ f)}, \tag*{(67)} R ( h ∘ g ) ∘ f q ˉ ⊢ L ( Σ 4 ) (( R f k ˉ ) g l ˉ ) h m ˉ ⊢ L ( Σ 4 ) R h ∘ ( g ∘ f ) a ( q ˉ ) , (67) which clearly entails the relation condition.
We can describe the semantic dual of the associator using the compositor from Proposition 3.1.6. Recall that if F : C → D F:C\to D F : C → D and G 1 , G 2 : D → E G_1,G_2:D\to E G 1 , G 2 : D → E are functors and η : G 1 ⇒ G 2 \eta:G_1\Rightarrow G_2 η : G 1 ⇒ G 2 is a natural transformation, then η F : G 1 F ⇒ G 2 F \eta F:G_1F\Rightarrow G_2F η F : G 1 F ⇒ G 2 F , the right whiskering of η \eta η by F F F , denotes the natural transformation defined via ( η F ) X : = η F ( X ) (\eta F)_X:=\eta_{F(X)} ( η F ) X := η F ( X ) .
Proposition 3.2.3 (Coherence of the Associator). Let f : T 1 → T 2 f:T_1\to T_2 f : T 1 → T 2 , g : T 2 → T 3 g:T_2\to T_3 g : T 2 → T 3 and h : T 3 → T 4 h:T_3\to T_4 h : T 3 → T 4 be piecewise interpretations. Then the following holds for the associator a h , g , f a_{h,g,f} a h , g , f :
a h , g , f ∗ = α h , g ∘ f − 1 ∘ α g , f − 1 h ∗ ∘ f ∗ α h , g ∘ α h ∘ g , f (68) a^*_{h,g,f}=\alpha^{-1}_{h,g\circ f}\circ\alpha^{-1}_{g,f}h^*\circ f^*\alpha_{h,g}\circ\alpha_{h\circ g,f} \tag*{(68)} a h , g , f ∗ = α h , g ∘ f − 1 ∘ α g , f − 1 h ∗ ∘ f ∗ α h , g ∘ α h ∘ g , f (68)
Proof. Let M ⊨ T 4 M \models T_{4} M ⊨ T 4 , σ \sigma σ be a sort of L 1 \mathcal{L}_{1} L 1 , and x ∈ ( ( h ∘ g ) ∘ f ) M ∗ ( σ ) x \in((h \circ g) \circ f)^{*}_{M}(\sigma) x ∈ (( h ∘ g ) ∘ f ) M ∗ ( σ ) . Then there is a piece p = ( k , ( ( l 1 , m ˉ 1 ) , … , ( l n , m ˉ n ) ) ) ∈ P σ ( h ∘ g ) ∘ f p=(k,((l_{1},\bar{m}_{1}),\ldots,(l_{n},\bar{m}_{n}))) \in P^{(h\circ g)\circ f}_{\sigma} p = ( k , (( l 1 , m ˉ 1 ) , … , ( l n , m ˉ n ))) ∈ P σ ( h ∘ g ) ∘ f and a tuple a ˉ ∈ M ( D p ( h ∘ g ) ∘ f ) \bar{a} \in M(D^{(h\circ g)\circ f}_{p}) a ˉ ∈ M ( D p ( h ∘ g ) ∘ f ) such that x = π σ ( h ∘ g ) ∘ f ( p , a ˉ ) x=\pi^{(h\circ g)\circ f}_{\sigma}(p,\bar{a}) x = π σ ( h ∘ g ) ∘ f ( p , a ˉ ) . Set τ i : = ( σ f ) i \tau_{i}:=(\sigma^{f})_{i} τ i := ( σ f ) i and n : = ∣ σ f ∣ n:=|\sigma^{f}| n := ∣ σ f ∣ . Then a ˉ \bar{a} a ˉ can be written as a ˉ = b ˉ 1 × ⋯ × b ˉ n \bar{a}=\bar{b}_{1}\times\cdots\times\bar{b}_{n} a ˉ = b ˉ 1 × ⋯ × b ˉ n with b ˉ i ∈ M ( D ( l i , m ˉ i ) h ∘ g ) \bar{b}_{i}\in M(D^{h\circ g}_{(l_{i},\bar{m}_{i})}) b ˉ i ∈ M ( D ( l i , m ˉ i ) h ∘ g ) . Set l ˉ = ( l 1 , … , l n ) \bar{l}=(l_{1},\ldots,l_{n}) l ˉ = ( l 1 , … , l n ) , m ˉ = m ˉ 1 × ⋯ × m ˉ n \bar{m}=\bar{m}_{1}\times\cdots\times\bar{m}_{n} m ˉ = m ˉ 1 × ⋯ × m ˉ n , and q ˉ = ( ( l 1 , m ˉ 1 ) , … , ( l n , m ˉ n ) ) \bar{q}=((l_{1},\bar{m}_{1}),\ldots,(l_{n},\bar{m}_{n})) q ˉ = (( l 1 , m ˉ 1 ) , … , ( l n , m ˉ n )) . For each i i i , define
e i : = π τ i h ∘ g ( ( l i , m ˉ i ) , b ˉ i ) , c ˉ i : = π ( τ i ) g h ( m ˉ i , b ˉ i ) . (69) e_{i}:=\pi^{h\circ g}_{\tau_{i}}((l_{i},\bar{m}_{i}),\bar{b}_{i}), \qquad\bar{c}_{i}:=\pi^{h}_{(\tau_{i})_{g}}(\bar{m}_{i},\bar{b}_{i}). \tag*{(69)} e i := π τ i h ∘ g (( l i , m ˉ i ) , b ˉ i ) , c ˉ i := π ( τ i ) g h ( m ˉ i , b ˉ i ) . (69)
Set c ˉ = c ˉ 1 × ⋯ × c ˉ n \bar{c}=\bar{c}_{1}\times\cdots\times\bar{c}_{n} c ˉ = c ˉ 1 × ⋯ × c ˉ n . We compute the right-hand side evaluated at x x x : 14
( α h , g ∘ f − 1 ∙ ( α g , f − 1 h ∗ ) ∙ ( f ∗ α h , g ) ∙ α h ∘ g , f ) M ( π σ ( h ∘ g ) ∘ f ( p , a ˉ ) ) (70) \left(\alpha^{-1}_{h,g\circ f}\mathbin{\bullet}(\alpha^{-1}_{g,f}h^{*})\mathbin{\bullet}(f^{*}\alpha_{h,g})\mathbin{\bullet}\alpha_{h\circ g,f}\right)_{M}\left(\pi^{(h\circ g)\circ f}_{\sigma}(p,\bar{a})\right) \tag*{(70)} ( α h , g ∘ f − 1 ∙ ( α g , f − 1 h ∗ ) ∙ ( f ∗ α h , g ) ∙ α h ∘ g , f ) M ( π σ ( h ∘ g ) ∘ f ( p , a ˉ ) ) (70)
= α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( f ∗ α h , g ) ( π σ f ( k , π q ˉ h ∘ g ( q ˉ , a ˉ ) ) ) (71) =\alpha^{-1}_{h,g\circ f}\left(\alpha^{-1}_{g,f}h^{*}\right)(f^{*}\alpha_{h,g})\left(\pi^{f}_{\sigma}\left(k,\pi^{h\circ g}_{\bar{q}}(\bar{q},\bar{a})\right)\right) \tag*{(71)} = α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( f ∗ α h , g ) ( π σ f ( k , π q ˉ h ∘ g ( q ˉ , a ˉ ) ) ) (71)
= α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( f ∗ α h , g ) ( π σ f ( k , ( e i ) i = 1 n ) ) (72) =\alpha^{-1}_{h,g\circ f}\left(\alpha^{-1}_{g,f}h^{*}\right)(f^{*}\alpha_{h,g})\left(\pi^{f}_{\sigma}\left(k,(e_{i})_{i=1}^{n}\right)\right) \tag*{(72)} = α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( f ∗ α h , g ) ( π σ f ( k , ( e i ) i = 1 n ) ) (72)
= α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( π σ f ( k , ( ( α h , g ) M ( e i ) ) i = 1 n ) ) = α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( π σ f ( k , ( π τ i g ( l i , c ˉ i ) ) i = 1 n ) ) (73) =\alpha^{-1}_{h,g\circ f}\left(\alpha^{-1}_{g,f}h^{*}\right)\left(\pi^{f}_{\sigma}\left(k,\left((\alpha_{h,g})_{M}(e_{i})\right)_{i=1}^{n}\right)\right)
=\alpha^{-1}_{h,g\circ f}\left(\alpha^{-1}_{g,f}h^{*}\right)\left(\pi^{f}_{\sigma}\left(k,\left(\pi^{g}_{\tau_{i}}(l_{i},\bar{c}_{i})\right)_{i=1}^{n}\right)\right) \tag*{(73)} = α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( π σ f ( k , ( ( α h , g ) M ( e i ) ) i = 1 n ) ) = α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( π σ f ( k , ( π τ i g ( l i , c ˉ i ) ) i = 1 n ) ) (73)
= α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( π σ g ∘ f ( ( k , l ˉ ) , c ˉ ) ) = α h , g ∘ f − 1 ( π σ g ∘ f ( ( k , l ˉ ) , c ˉ ) ) (74) =\alpha^{-1}_{h,g\circ f}\left(\alpha^{-1}_{g,f}h^{*}\right)\left(\pi^{g\circ f}_{\sigma}\left((k,\bar{l}),\bar{c}\right)\right)
=\alpha^{-1}_{h,g\circ f}\left(\pi^{g\circ f}_{\sigma}\left((k,\bar{l}),\bar{c}\right)\right) \tag*{(74)} = α h , g ∘ f − 1 ( α g , f − 1 h ∗ ) ( π σ g ∘ f ( ( k , l ˉ ) , c ˉ ) ) = α h , g ∘ f − 1 ( π σ g ∘ f ( ( k , l ˉ ) , c ˉ ) ) (74)
= α h , g ∘ f − 1 ( π σ g ∘ f ( ( k , l ˉ ) , π ( k , l ˉ ) g h ( m ˉ , a ˉ ) ) ) = π σ h ∘ ( g ∘ f ) ( ( ( k , l ˉ ) , m ˉ ) , a ˉ ) (75) =\alpha^{-1}_{h,g\circ f}\left(\pi^{g\circ f}_{\sigma}\left((k,\bar{l}),\pi^{h}_{(k,\bar{l})_{g}}(\bar{m},\bar{a})\right)\right)
=\pi^{h\circ(g\circ f)}_{\sigma}\left(((k,\bar{l}),\bar{m}),\bar{a}\right) \tag*{(75)} = α h , g ∘ f − 1 ( π σ g ∘ f ( ( k , l ˉ ) , π ( k , l ˉ ) g h ( m ˉ , a ˉ ) ) ) = π σ h ∘ ( g ∘ f ) ( (( k , l ˉ ) , m ˉ ) , a ˉ ) (75)
= a h , g , f , M ∗ ( π σ ( h ∘ g ) ∘ f ( p , a ˉ ) ) . (76) =a^{*}_{h,g,f,M}\left(\pi^{(h\circ g)\circ f}_{\sigma}(p,\bar{a})\right). \tag*{(76)} = a h , g , f , M ∗ ( π σ ( h ∘ g ) ∘ f ( p , a ˉ ) ) . (76)
Here, equalities (3.2.15), (3.2.18), and (3.2.20) use the componentwise structure of the projection maps, while the others apply the definitions of the involved maps.
Vertical and Horizontal Composition of i i i -Morphisms
Having defined the operations we need on 1-morphisms, that is, on piecewise interpretations, we can now move on to define the set of 2-morphisms for each pair of parallel interpretations f , g : T 1 → T 2 f,g:T_{1}\to T_{2} f , g : T 1 → T 2 by Th p w ( f , g ) = { [ η : f ⇒ g ] ≡ T 2 } \operatorname{Th}^{\mathrm{pw}}(f,g)=\{[\eta:f\Rightarrow g]_{\equiv_{T_{2}}}\} Th pw ( f , g ) = {[ η : f ⇒ g ] ≡ T 2 } . Such a 2-morphism is just an equivalence class of i i i -morphisms under provable equivalence (Definition 2.6.2 ). With these 2-morphisms providing the required morphisms, the set of all interpretations between two theories can be equipped with the structure of a category: The identity i i i -morphism id f : f ⇒ f \operatorname{id}_{f}:f\Rightarrow f id f : f ⇒ f is the collection ( id f ) σ : P σ f × P σ f → L 2 (\operatorname{id}_{f})_{\sigma}:P^{f}_{\sigma}\times P^{f}_{\sigma}\to\mathcal{L}_{2} ( id f ) σ : P σ f × P σ f → L 2 defined as follows, where u ˉ : σ f k \bar{u}:\sigma^{k}_{f} u ˉ : σ f k and v ˉ : σ f l \bar{v}:\sigma^{l}_{f} v ˉ : σ f l :
( id f ) σ k , l ( u ˉ , v ˉ ) : = D σ f , k ( u ˉ ) ∧ D σ f , l ( v ˉ ) ∧ ( = σ ) f k , l ( u ˉ , v ˉ ) . (77) (\operatorname{id}_{f})^{k,l}_{\sigma}(\bar{u},\bar{v}):=D^{f,k}_{\sigma}(\bar{u})\land D^{f,l}_{\sigma}(\bar{v})\land(=_{\sigma})^{k,l}_{f}(\bar{u},\bar{v}). \tag*{(77)} ( id f ) σ k , l ( u ˉ , v ˉ ) := D σ f , k ( u ˉ ) ∧ D σ f , l ( v ˉ ) ∧ ( = σ ) f k , l ( u ˉ , v ˉ ) . (77)
It follows from Proposition 2.6.9 with b b b the identity that this is indeed an i i i -morphism. Composition for the category Th p w ( T 1 , T 2 ) \operatorname{Th}^{\mathrm{pw}}(T_{1},T_{2}) Th pw ( T 1 , T 2 ) is given by the following definition:
Definition 3.3.1 (Vertical Composition of i i i -Morphisms). Let η : g ⇒ h \eta:g\Rightarrow h η : g ⇒ h and ζ : f ⇒ g \zeta:f\Rightarrow g ζ : f ⇒ g be i i i -morphisms, where f , g , h : T 1 → T 2 f,g,h:T_{1}\to T_{2} f , g , h : T 1 → T 2 are interpretations. Their vertical composition η ∙ ζ : f ⇒ h \eta\bullet\zeta:f\Rightarrow h η ∙ ζ : f ⇒ h is defined, for each sort σ ∈ S 1 \sigma\in S_{1} σ ∈ S 1 and pieces k ∈ P σ f k\in P^{f}_{\sigma} k ∈ P σ f , l ∈ P σ h l\in P^{h}_{\sigma} l ∈ P σ h , by
( η ∙ ζ ) σ k , l ( x ˉ , y ˉ ) : = ⋁ m ∈ P σ g ∃ z ˉ ( ζ σ k , m ( x ˉ , z ˉ ) ∧ η σ m , l ( z ˉ , y ˉ ) ) , (78) (\eta\bullet\zeta)^{k,l}_{\sigma}(\bar{x},\bar{y}):=\bigvee_{m\in P^{g}_{\sigma}}\exists\bar{z}\left(\zeta^{k,m}_{\sigma}(\bar{x},\bar{z})\land\eta^{m,l}_{\sigma}(\bar{z},\bar{y})\right), \tag*{(78)} ( η ∙ ζ ) σ k , l ( x ˉ , y ˉ ) := m ∈ P σ g ⋁ ∃ z ˉ ( ζ σ k , m ( x ˉ , z ˉ ) ∧ η σ m , l ( z ˉ , y ˉ ) ) , (78)
with x ˉ : σ f k \bar{x}:\sigma^{k}_{f} x ˉ : σ f k , y ˉ : σ h l \bar{y}:\sigma^{l}_{h} y ˉ : σ h l , and z ˉ : σ g m \bar{z}:\sigma^{m}_{g} z ˉ : σ g m . Vertical composition descends to equivalence classes of i i i -morphisms, is associative, and has i d f \mathrm{id}_f id f as identity. All this follows from local faithfulness (Corollary 2.6.7) together with the following proposition, which shows that η ∙ ζ \eta\mathbin{\bullet} \zeta η ∙ ζ is an i i i -morphism whose semantic dual is the vertical composite of the duals:
Proposition 3.3.2 (Semantic Dual of Vertical Composition). Let η : g ⇒ h \eta:g\Rightarrow h η : g ⇒ h , ζ : f ⇒ g \zeta:f\Rightarrow g ζ : f ⇒ g be two i i i -morphisms, where f , g , h : T 1 → T 2 f,g,h:T_1\to T_2 f , g , h : T 1 → T 2 are interpretations. Then η ∙ ζ : f ⇒ h \eta\mathbin{\bullet}\zeta:f\Rightarrow h η ∙ ζ : f ⇒ h is an i i i -morphism satisfying ( η ∙ ζ ) ∗ = η ∗ ∙ ζ ∗ (\eta\mathbin{\bullet}\zeta)^*=\eta^*\mathbin{\bullet}\zeta^* ( η ∙ ζ ) ∗ = η ∗ ∙ ζ ∗ , where ∙ \mathbin{\bullet} ∙ denotes vertical composition of natural transformations. In particular, the component at each model is the usual composition of elementary embeddings.
Proof. Condition (Dom/Range) for the pre-i i i -morphism η ∙ ζ \eta\mathbin{\bullet}\zeta η ∙ ζ follows directly from the corresponding (Dom/Range) condition for ζ \zeta ζ and η \eta η . By Proposition 2.6.6, it therefore suffices to verify (Dual) for η ∙ ζ \eta\mathbin{\bullet}\zeta η ∙ ζ and the natural transformation η ∗ ∙ ζ ∗ : f ∗ → h ∗ \eta^*\mathbin{\bullet}\zeta^*:f^*\to h^* η ∗ ∙ ζ ∗ : f ∗ → h ∗ . So let M ⊨ T 2 M\models T_2 M ⊨ T 2 , ( k , a ˉ ) ∈ P σ f × M ( D σ f , k ) (k,\bar{a})\in P_\sigma^f\times M(D_\sigma^{f,k}) ( k , a ˉ ) ∈ P σ f × M ( D σ f , k ) , and ( l , b ˉ ) ∈ P σ h × M ( D σ h , l ) (l,\bar{b})\in P_\sigma^h\times M(D_\sigma^{h,l}) ( l , b ˉ ) ∈ P σ h × M ( D σ h , l ) . By definition, (Dom/Range) for η \eta η (or ζ \zeta ζ ), and (Dual) for η \eta η and ζ \zeta ζ , a ˉ × b ˉ ∈ M ( ( η ∙ ζ ) σ k , l ) \bar{a}\times\bar{b}\in M((\eta\mathbin{\bullet}\zeta)_\sigma^{k,l}) a ˉ × b ˉ ∈ M (( η ∙ ζ ) σ k , l ) iff there are ( m , c ˉ ) ∈ P σ g × M ( D σ g , m ) (m,\bar{c})\in P_\sigma^g\times M(D_\sigma^{g,m}) ( m , c ˉ ) ∈ P σ g × M ( D σ g , m ) such that ζ M ( π σ f ( k , a ˉ ) ) = π σ g ( m , c ˉ ) \zeta_M(\pi_\sigma^f(k,\bar{a}))=\pi_\sigma^g(m,\bar{c}) ζ M ( π σ f ( k , a ˉ )) = π σ g ( m , c ˉ ) and η M ( π σ g ( m , c ˉ ) ) = π σ h ( l , b ˉ ) \eta_M(\pi_\sigma^g(m,\bar{c}))=\pi_\sigma^h(l,\bar{b}) η M ( π σ g ( m , c ˉ )) = π σ h ( l , b ˉ ) . This is the case iff η M ∗ ( ζ M ∗ ( π σ f ( k , a ˉ ) ) ) = π σ h ( l , b ˉ ) \eta_M^*(\zeta_M^*(\pi_\sigma^f(k,\bar{a})))=\pi_\sigma^h(l,\bar{b}) η M ∗ ( ζ M ∗ ( π σ f ( k , a ˉ ))) = π σ h ( l , b ˉ ) , since ζ M ∗ ( π σ f ( k , a ˉ ) ) = π σ g ( m , c ˉ ) \zeta_M^*(\pi_\sigma^f(k,\bar{a}))=\pi_\sigma^g(m,\bar{c}) ζ M ∗ ( π σ f ( k , a ˉ )) = π σ g ( m , c ˉ ) for some such ( m , c ˉ ) (m,\bar{c}) ( m , c ˉ ) by surjectivity of π σ g \pi_\sigma^g π σ g .
A straightforward consequence of this is that what we have called i i i -isomorphisms in Definition 2.6.8 before are indeed exactly the i i i -morphisms invertible with respect to vertical composition. Summing up, for each pair of theories, there is a 1 1 1 -category T h p w ( T 1 , T 2 ) \mathrm{Th}^{\mathrm{pw}}(T_1,T_2) Th pw ( T 1 , T 2 ) whose objects are piecewise interpretations and whose morphisms are i i i -morphisms between them, both quotiented with respect to provable equivalence.
As a final ingredient, necessary to ensure that composition of piecewise interpretations can be seen as a functor T h p w ( T 2 , T 3 ) × T h p w ( T 1 , T 2 ) → T h p w ( T 1 , T 3 ) \mathrm{Th}^{\mathrm{pw}}(T_2,T_3)\times\mathrm{Th}^{\mathrm{pw}}(T_1,T_2)\to\mathrm{Th}^{\mathrm{pw}}(T_1,T_3) Th pw ( T 2 , T 3 ) × Th pw ( T 1 , T 2 ) → Th pw ( T 1 , T 3 ) , we need another notion of composition of i i i -morphisms, namely horizontal composition. Horizontal composition is a function ⋆ : T h p w ( f 2 , g 2 ) × T h p w ( f 1 , g 1 ) → T h p w ( f 2 ∘ f 1 , g 2 ∘ g 1 ) \star:\mathrm{Th}^{\mathrm{pw}}(f_2,g_2)\times\mathrm{Th}^{\mathrm{pw}}(f_1,g_1)\to\mathrm{Th}^{\mathrm{pw}}(f_2\circ f_1,g_2\circ g_1) ⋆ : Th pw ( f 2 , g 2 ) × Th pw ( f 1 , g 1 ) → Th pw ( f 2 ∘ f 1 , g 2 ∘ g 1 ) , where f 1 , g 1 : T 1 → T 2 f_1,g_1:T_1\to T_2 f 1 , g 1 : T 1 → T 2 and f 2 , g 2 : T 2 → T 3 f_2,g_2:T_2\to T_3 f 2 , g 2 : T 2 → T 3 , satisfying certain coherence conditions.
Definition 3.3.3 (Horizontal Composition). Let η : f 2 ⇒ g 2 \eta:f_2\Rightarrow g_2 η : f 2 ⇒ g 2 and ζ : f 1 ⇒ g 1 \zeta:f_1\Rightarrow g_1 ζ : f 1 ⇒ g 1 be i i i -morphisms, where f 1 , g 1 : T 1 → T 2 f_1,g_1:T_1\to T_2 f 1 , g 1 : T 1 → T 2 and f 2 , g 2 : T 2 → T 3 f_2,g_2:T_2\to T_3 f 2 , g 2 : T 2 → T 3 . Their horizontal composition
η ⋆ ζ : f 2 ∘ f 1 ⇒ g 2 ∘ g 1 (79) \eta\star\zeta:f_2\circ f_1\Rightarrow g_2\circ g_1 \tag*{(79)} η ⋆ ζ : f 2 ∘ f 1 ⇒ g 2 ∘ g 1 (79) is defined, for each sort σ ∈ S 1 \sigma\in S_1 σ ∈ S 1 and pieces m 1 = ( k 1 , l ˉ 1 ) ∈ P σ f 2 ∘ f 1 m_1=(k_1,\bar{l}_1)\in P_\sigma^{f_2\circ f_1} m 1 = ( k 1 , l ˉ 1 ) ∈ P σ f 2 ∘ f 1 , m 2 = ( k 2 , l ˉ 2 ) ∈ P σ g 2 ∘ g 1 m_2=(k_2,\bar{l}_2)\in P_\sigma^{g_2\circ g_1} m 2 = ( k 2 , l ˉ 2 ) ∈ P σ g 2 ∘ g 1 , as follows.
( η ⋆ ζ ) σ m 1 , m 2 ( x ˉ , y ˉ ) : = D σ f 2 ∘ f 1 , m 1 ( x ˉ ) ∧ D σ g 2 ∘ g 1 , m 2 ( y ˉ ) ∧ ⋁ m 3 ∈ P σ g 2 ∘ g 1 ( ⋁ k 4 ∈ P σ g 1 k 3 f 2 ∃ u ˉ ( ⋁ v ˉ ( ( ζ σ k 1 , k 3 ) f 2 l ˉ 1 × k ˉ 4 ( x ˉ , v ˉ ) ∧ η σ g 1 k 3 k ˉ 4 , l ˉ 3 ( v ˉ , u ˉ ) ) ∧ ( = σ m 2 , m 3 ( y ˉ , u ˉ ) ) ) . (80) \begin{aligned}
(\eta\star\zeta)_\sigma^{m_1,m_2}(\bar{x},\bar{y}) &:=
D_\sigma^{f_2\circ f_1,m_1}(\bar{x})\land D_\sigma^{g_2\circ g_1,m_2}(\bar{y})\\
&\land\bigvee_{m_3\in P_\sigma^{g_2\circ g_1}}\left(\bigvee_{k_4\in P_{\sigma_{g_1}^{k_3}}^{f_2}}\exists\bar{u}\left(\bigvee_{\bar{v}}\left((\zeta_\sigma^{k_1,k_3})_{\!f_2}^{\bar{l}_1\times\bar{k}_4}(\bar{x},\bar{v})\land\eta_{\sigma_{g_1}^{k_3}}^{\bar{k}_4,\bar{l}_3}(\bar{v},\bar{u})\right)\land(=_{\sigma}^{m_2,m_3}(\bar{y},\bar{u})\right)\right). \tag*{(80)}
\end{aligned} ( η ⋆ ζ ) σ m 1 , m 2 ( x ˉ , y ˉ ) := D σ f 2 ∘ f 1 , m 1 ( x ˉ ) ∧ D σ g 2 ∘ g 1 , m 2 ( y ˉ ) ∧ m 3 ∈ P σ g 2 ∘ g 1 ⋁ k 4 ∈ P σ g 1 k 3 f 2 ⋁ ∃ u ˉ ( v ˉ ⋁ ( ( ζ σ k 1 , k 3 ) f 2 l ˉ 1 × k ˉ 4 ( x ˉ , v ˉ ) ∧ η σ g 1 k 3 k ˉ 4 , l ˉ 3 ( v ˉ , u ˉ ) ) ∧ ( = σ m 2 , m 3 ( y ˉ , u ˉ ) ) . (80) where m 3 = ( k 3 , l ˉ 3 ) m_3=(k_3,\bar{l}_3) m 3 = ( k 3 , l ˉ 3 ) and x ˉ : σ f 2 ∘ f 1 m 1 \bar{x}:\sigma_{f_2\circ f_1}^{m_1} x ˉ : σ f 2 ∘ f 1 m 1 , y ˉ : σ g 2 ∘ g 1 m 2 \bar{y}:\sigma_{g_2\circ g_1}^{m_2} y ˉ : σ g 2 ∘ g 1 m 2 , u ˉ : σ g 2 ∘ g 1 m 3 \bar{u}:\sigma_{g_2\circ g_1}^{m_3} u ˉ : σ g 2 ∘ g 1 m 3 , and v ˉ : σ f 2 ∘ g 1 ( k 3 , k ˉ 4 ) \bar{v}:\sigma_{f_2\circ g_1}^{(k_3,\bar{k}_4)} v ˉ : σ f 2 ∘ g 1 ( k 3 , k ˉ 4 ) . By construction and Lemma 2.1.6, ( η ⋆ ζ ) σ m 1 , m 2 : σ f 2 ∘ f 1 m 1 × σ g 2 ∘ g 1 m 2 (\eta\star\zeta)_\sigma^{m_1,m_2}:\sigma_{f_2\circ f_1}^{m_1}\times\sigma_{g_2\circ g_1}^{m_2} ( η ⋆ ζ ) σ m 1 , m 2 : σ f 2 ∘ f 1 m 1 × σ g 2 ∘ g 1 m 2 , as required of a pre-i i i -morphism.
As with vertical composition, well-definedness is best established through the semantic dual:
Proposition 3.3.4 (Semantic Dual of Horizontal Composition). Let η : f 2 ⇒ g 2 \eta:f_2 \Rightarrow g_2 η : f 2 ⇒ g 2 and ζ : f 1 ⇒ g 1 \zeta:f_1 \Rightarrow g_1 ζ : f 1 ⇒ g 1 be i i i -morphisms, where f 1 , g 1 : T 1 → T 2 f_1,g_1:T_1 \to T_2 f 1 , g 1 : T 1 → T 2 and f 2 , g 2 : T 2 → T 3 f_2,g_2:T_2 \to T_3 f 2 , g 2 : T 2 → T 3 are interpretations. Then η ⋆ ζ : f 2 ∘ f 1 ⇒ g 2 ∘ g 1 \eta\star\zeta:f_2 \circ f_1 \Rightarrow g_2 \circ g_1 η ⋆ ζ : f 2 ∘ f 1 ⇒ g 2 ∘ g 1 is an i i i -morphism satisfying
( η ⋆ ζ ) ∗ = α g 2 , g 1 − 1 ∙ ( ζ ∗ ⋆ η ∗ ) ∙ α f 2 , f 1 , (81) (\eta\star\zeta)^{*}=\alpha^{-1}_{g_2,g_1}\bullet(\zeta^{*}\star\eta^{*})\bullet\alpha_{f_2,f_1}, \tag*{(81)} ( η ⋆ ζ ) ∗ = α g 2 , g 1 − 1 ∙ ( ζ ∗ ⋆ η ∗ ) ∙ α f 2 , f 1 , (81) where ζ ∗ ⋆ η ∗ \zeta^{*}\star\eta^{*} ζ ∗ ⋆ η ∗ denotes the horizontal composition of natural transformations. At a model M ⊨ T 3 M \models T_3 M ⊨ T 3 , the latter can, by general category theory, be described as
( ζ ∗ ⋆ η ∗ ) M = ζ g 2 M ∗ ∘ f 1 ∗ ( η M ∗ ) = g 1 ∗ ( η M ∗ ) ∘ ζ f 2 M ∗ . (82) (\zeta^{*}\star\eta^{*})_M=\zeta^{*}_{g_2M}\circ f_1^{*}(\eta^{*}_M)=g_1^{*}(\eta^{*}_M)\circ\zeta^{*}_{f_2M}. \tag*{(82)} ( ζ ∗ ⋆ η ∗ ) M = ζ g 2 M ∗ ∘ f 1 ∗ ( η M ∗ ) = g 1 ∗ ( η M ∗ ) ∘ ζ f 2 M ∗ . (82) Proof. Condition (Dom/Range) for the pre-i i i -morphism η ⋆ ζ \eta\star\zeta η ⋆ ζ holds by construction, and F : = α g 2 , g 1 − 1 ∙ ( ζ ∗ ⋆ η ∗ ) ∙ α f 2 , f 1 F:=\alpha^{-1}_{g_2,g_1}\bullet(\zeta^{*}\star\eta^{*})\bullet\alpha_{f_2,f_1} F := α g 2 , g 1 − 1 ∙ ( ζ ∗ ⋆ η ∗ ) ∙ α f 2 , f 1 is a composite of natural transformations, so by Proposition 2.6.6 it suffices to verify (Dual) for η ⋆ ζ \eta\star\zeta η ⋆ ζ and F F F . So let M ⊨ T 3 M \models T_3 M ⊨ T 3 , ( m 1 , a ˉ ) ∈ P σ f 2 ∘ f 1 × M ( D σ f 2 ∘ f 1 , m 1 ) (m_1,\bar{a})\in P^{f_2\circ f_1}_{\sigma}\times M(D^{f_2\circ f_1,m_1}_{\sigma}) ( m 1 , a ˉ ) ∈ P σ f 2 ∘ f 1 × M ( D σ f 2 ∘ f 1 , m 1 ) , and ( m 2 , b ˉ ) ∈ P σ g 2 ∘ g 1 × M ( D σ g 2 ∘ g 1 , m 2 ) (m_2,\bar{b})\in P^{g_2\circ g_1}_{\sigma}\times M(D^{g_2\circ g_1,m_2}_{\sigma}) ( m 2 , b ˉ ) ∈ P σ g 2 ∘ g 1 × M ( D σ g 2 ∘ g 1 , m 2 ) , and let m 1 = ( k 1 , l ˉ 1 ) m_1=(k_1,\bar{l}_1) m 1 = ( k 1 , l ˉ 1 ) .
First by definition, and two applications of (Dual), which requires establishing the respective domain conditions via (Dom/Range) for η \eta η at M M M , ζ \zeta ζ at f 2 ∗ M f_2^{*}M f 2 ∗ M via Corollary 2.4.8, and (Element-Embed) for η \eta η for D σ , k 3 g 1 D^{g_1}_{\sigma,k_3} D σ , k 3 g 1 , we observe that a ˉ × b ˉ ∈ M ( ( η ⋆ ζ ) σ m 1 , m 2 ) \bar{a}\times\bar{b}\in M((\eta\star\zeta)^{m_1,m_2}_{\sigma}) a ˉ × b ˉ ∈ M (( η ⋆ ζ ) σ m 1 , m 2 ) iff there are m 3 = ( k 3 , l ˉ 3 ) ∈ P σ g 2 ∘ g 1 m_3=(k_3,\bar{l}_3)\in P^{g_2\circ g_1}_{\sigma} m 3 = ( k 3 , l ˉ 3 ) ∈ P σ g 2 ∘ g 1 , k ˉ 4 ∈ P σ , k 3 f 2 \bar{k}_4\in P^{f_2}_{\sigma,k_3} k ˉ 4 ∈ P σ , k 3 f 2 , u ˉ ∈ M ( D σ g 2 ∘ g 1 , m 3 ) \bar{u}\in M(D^{g_2\circ g_1,m_3}_{\sigma}) u ˉ ∈ M ( D σ g 2 ∘ g 1 , m 3 ) , and v ˉ ∈ M ( D σ f 2 ∘ g 1 , ( k 3 , k ˉ 4 ) ) \bar{v}\in M(D^{f_2\circ g_1,(k_3,\bar{k}_4)}_{\sigma}) v ˉ ∈ M ( D σ f 2 ∘ g 1 , ( k 3 , k ˉ 4 ) ) such that
ζ f 2 ∗ M ∗ ( π σ f 2 ( k 1 , π f 1 , σ f 2 ( l ˉ 1 , a ˉ ) ) ) = π σ g 1 ( k 3 , π g 1 , σ f 2 ( k ˉ 4 , v ˉ ) ) , η M ∗ ( π g 1 , σ f 2 ( k ˉ 4 , v ˉ ) ) = π g 1 , σ g 2 ( l ˉ 3 , u ˉ ) , (83) \zeta^{*}_{f_2^{*}M}\left(\pi^{f_2}_{\sigma}(k_1,\pi^{f_2}_{f_1,\sigma}(\bar{l}_1,\bar{a}))\right)=\pi^{g_1}_{\sigma}(k_3,\pi^{f_2}_{g_1,\sigma}(\bar{k}_4,\bar{v})),\qquad
\eta^{*}_M\left(\pi^{f_2}_{g_1,\sigma}(\bar{k}_4,\bar{v})\right)=\pi^{g_2}_{g_1,\sigma}(\bar{l}_3,\bar{u}), \tag*{(83)} ζ f 2 ∗ M ∗ ( π σ f 2 ( k 1 , π f 1 , σ f 2 ( l ˉ 1 , a ˉ )) ) = π σ g 1 ( k 3 , π g 1 , σ f 2 ( k ˉ 4 , v ˉ )) , η M ∗ ( π g 1 , σ f 2 ( k ˉ 4 , v ˉ ) ) = π g 1 , σ g 2 ( l ˉ 3 , u ˉ ) , (83) and π σ g 2 ∘ g 1 ( m 3 , u ˉ ) = π σ g 2 ∘ g 1 ( m 2 , b ˉ ) \pi^{g_2\circ g_1}_{\sigma}(m_3,\bar{u})=\pi^{g_2\circ g_1}_{\sigma}(m_2,\bar{b}) π σ g 2 ∘ g 1 ( m 3 , u ˉ ) = π σ g 2 ∘ g 1 ( m 2 , b ˉ ) .
To see that this is the case iff F M ( π σ f 2 ∘ f 1 ( m 1 , a ˉ ) ) = π σ g 2 ∘ g 1 ( m 2 , b ˉ ) F_M(\pi^{f_2\circ f_1}_{\sigma}(m_1,\bar{a}))=\pi^{g_2\circ g_1}_{\sigma}(m_2,\bar{b}) F M ( π σ f 2 ∘ f 1 ( m 1 , a ˉ )) = π σ g 2 ∘ g 1 ( m 2 , b ˉ ) , for the forwards direction, one plugs these equalities into the computation of F M F_M F M . For the backwards direction, one chooses representatives of the intermediate steps, again observing that these satisfy the respective domain conditions using elementarity and Corollary 2.4.8. □
Horizontal composition is not associative on the nose, but this is to be expected, since composition of piecewise interpretations is itself only associative up to the associator.
Bicategory and Mod-Pseudofunctor {#sec:3.4} Having defined the basic operations on piecewise interpretations and i i i -morphisms and verified that they are invariant under provable equivalence, we now have all the basic ingredients for the following central definition:
Definition 3.4.1 (Bicategory of Theories and Piecewise Interpretations). Let T h p w \mathrm{Th}^{\mathrm{pw}} Th pw denote the bicategory whose objects are given by classical first-order theories, sets of 1-morphisms T h p w ( T 1 , T 2 ) \mathrm{Th}^{\mathrm{pw}}(T_1,T_2) Th pw ( T 1 , T 2 ) by equivalence classes of piecewise interpretations f : T 1 → T 2 f:T_1\to T_2 f : T 1 → T 2 under provable equivalence, and whose sets of 2-morphisms T h p w ( f , g ) \mathrm{Th}^{\mathrm{pw}}(f,g) Th pw ( f , g ) are given by equivalence classes of i i i -morphisms η : f ⇒ g \eta:f\Rightarrow g η : f ⇒ g under provable equivalence. The identities are the identity interpretations and i i i -morphisms, composition of 1-morphisms is given by composition of
piecewise interpretations, and vertical and horizontal composition of 2-morphisms is given by vertical and horizontal composition of i i i -morphisms. Unitors and associators are given by unitors and associators for piecewise interpretations.
Remark 3.4.2 (Size of T h p w \mathrm{Th}^{\mathrm{pw}} Th pw ). By design, the bicategory T h p w \mathrm{Th}^{\mathrm{pw}} Th pw is locally small, although its objects form a proper class. Allowing theories in languages of larger cardinalities than the infinite cardinal κ \kappa κ fixed in the introduction only adds more objects (provided one adequately enlarges variable supplies), without essentially changing the sets of 1- or 2-morphisms between them.
At this stage, we have not verified the coherence conditions one needs for a bicategory: naturality of the associators and unitors, the interchange law, and so on. However, we do have all the data one needs for a bicategory, and we have semantic duals for theories, namely their categories of models, and for interpretations and i i i -morphisms. Moreover, we have appropriate unitors and compositors for these duals. So we have what one might call a pre-bicategory (T h p w \mathrm{Th}^{\mathrm{pw}} Th pw ) and something which looks like a locally faithful contravariant pseudofunctor (M o d \mathrm{Mod} Mod ) into a strict 2-category (C A T \mathbf{CAT} CAT ). In this situation, we can apply the following general theorem, whose proof makes use of ideas well known from coherence results for bicategories and can be found in Appendix B:
Theorem 3.4.3 (Reflection Theorem). Let B \mathcal{B} B be a pre-bicategory and C \mathcal{C} C be a strict 2-category. Assume that the associators a a a and the unitors l l l and r r r of B \mathcal{B} B are isomorphisms and that there is a locally faithful contravariant pseudofunctor F : B → C F:\mathcal{B}\to\mathcal{C} F : B → C . Then B \mathcal{B} B is a bicategory and F : B → C F:\mathcal{B}\to\mathcal{C} F : B → C is a locally faithful contravariant pseudofunctor between bicategories.
Applying this to our situation yields the following central theorem, justifying the preceding definition and encapsulating the work we have done on semantic duals.
Theorem 3.4.4 (Mod-Functor for Piecewise Interpretations). There is a locally faithful, contravariant pseudofunctor between bicategories M o d : T h p w → C A T \mathrm{Mod}:\mathrm{Th}^{\mathrm{pw}}\to\mathbf{CAT} Mod : Th pw → CAT , sending a theory T T T to its category of models M o d ( T ) \mathrm{Mod}(T) Mod ( T ) , a piecewise interpretation f : T 1 → T 2 f:T_{1}\to T_{2} f : T 1 → T 2 to its semantic dual f ∗ : M o d ( T 2 ) → M o d ( T 1 ) f^{*}:\mathrm{Mod}(T_{2})\to\mathrm{Mod}(T_{1}) f ∗ : Mod ( T 2 ) → Mod ( T 1 ) , and an i i i -morphism η : f ⇒ g \eta:f\Rightarrow g η : f ⇒ g to its semantic dual η ∗ : f ∗ ⇒ g ∗ \eta^{*}:f^{*}\Rightarrow g^{*} η ∗ : f ∗ ⇒ g ∗ . The compositor for f f f and g g g is given by the inverse of α g , f \alpha_{g,f} α g , f , and the unitor for T T T is given by α T \alpha_{T} α T .
Proof. That T h p w \mathrm{Th}^{\mathrm{pw}} Th pw is a pre-bicategory with well-defined operations was established in the remarks above, and that the associators and unitors are isomorphisms is established by Proposition 3.2.2 and the remarks under Definition 3.1.7 , respectively. To see that M o d \mathrm{Mod} Mod indeed defines a locally faithful, contravariant pseudofunctor M o d : T h p w → C A T \mathrm{Mod}:\mathrm{Th}^{\mathrm{pw}}\to\mathbf{CAT} Mod : Th pw → CAT , note first that the required structure, in particular the unitor and compositor of the semantic dual, are well-defined on equivalence classes under provable equivalence, as a consequence of Proposition 2.4.10 . Functoriality on each hom-category (i.e., ( η ∘ ζ ) ∗ = η ∗ ∘ ζ ∗ (\eta\circ\zeta)^{*}=\eta^{*}\circ\zeta^{*} ( η ∘ ζ ) ∗ = η ∗ ∘ ζ ∗ and ( i d f ) ∗ = i d f ∗ (\mathrm{id}_{f})^{*}=\mathrm{id}_{f^{*}} ( id f ) ∗ = id f ∗ ) was established in Proposition 3.3.2 and is immediate from the definition of i d f \mathrm{id}_{f} id f , respectively. Continuing, the compositor and unitor are natural isomorphisms, cohering appropriately with the unitors and associators of T h p w \mathrm{Th}^{\mathrm{pw}} Th pw (Propositions 3.1.6 , 3.2.3 and 3.3.4 and Lemma 3.1.8 ). Finally, local faithfulness is the content of Corollary 2.6.7, allowing us to apply Theorem 3.4.3. ∎
Let us now consider equivalence in this bicategory. Two theories T 1 T_1 T 1 and T 2 T_2 T 2 are equivalent in T h p w \mathrm{Th}^{\mathrm{pw}} Th pw iff there is a pair of piecewise interpretations f : T 1 → T 2 f:T_1\to T_2 f : T 1 → T 2 and g : T 2 → T 1 g:T_2\to T_1 g : T 2 → T 1 whose compositions are isomorphic to the respective identities, that is, g ∘ f ≅ i d T 1 g\circ f\cong\mathrm{id}_{T_1} g ∘ f ≅ id T 1 and f ∘ g ≅ i d T 2 f\circ g\cong\mathrm{id}_{T_2} f ∘ g ≅ id T 2 . Such theories will be called piecewise bi-interpretable .
Of course, if two theories are bi-interpretable in the ordinary sense, they are piecewise bi-interpretable. The converse does not hold in general, as we will see in § 4 . Because pseudofunctors preserve equivalence, we also have the following immediate corollary of Theorem 3.4.4 .
Corollary 3.4.5 (Piecewise Bi-Interpretability Implies Equivalence of Categories of Models). If two theories are piecewise bi-interpretable, they have equivalent categories of models.
The converse, likewise, does not hold: A counterexample from Barrett and Halvorson [2 ] , demonstrating that categorical equivalence does not imply Morita equivalence, can be repurposed to show this: essentially, one notes that piecewise bi-interpretable theories have isomorphic sentential algebras, and that one of the theories in their example has a sentential algebra with atoms, while the other does not.15 Moreover, we see that for n ≠ m n\ne m n = m , the theories E n E_n E n and E m E_m E m from Example 2.3.1 are mutually piecewise interpretable, but not piecewise bi-interpretable, because their respective models have non-isomorphic automorphism groups, whence they cannot have equivalent categories of models.
A Functorial Reduction of (Finitely) Many-Sorted Logic to Single-Sorted Logic
The formalization of many-sorted first-order logic came in tandem with proposed reductions of many-sorted to single-sorted logic (Schmidt [21 , 22 ] ; Wang [26 ] ). Logicians quickly arrived at a procedure, hereafter called unification , which takes a many-sorted theory and returns a corresponding single-sorted theory. The idea behind this procedure is quite simple: use predicate symbols instead of sorts and relativize quantifiers accordingly.16 Quine was a proponent of the idea that this procedure shows that many-sorted logic is ultimately dispensable (see, e.g., Quine [20 ] ). We will follow Barrett and Halvorson [3 ] in their reconstruction of this idea. First, for any many-sorted language L \mathcal{L} L with finitely many sorts σ 1 , … , σ n \sigma_1,\ldots,\sigma_n σ 1 , … , σ n , there is a single-sorted language L u \mathcal{L}^{u} L u with a single sort τ \tau τ , for each sort σ k \sigma_k σ k of L \mathcal{L} L , a unary predicate Q k Q_k Q k , and, for each non-identity relation symbol R : σ ˉ R:\bar{\sigma} R : σ ˉ , a relation symbol R u : τ ∣ σ ˉ ∣ R^{u}:\tau^{|\bar{\sigma}|} R u : τ ∣ σ ˉ ∣ . Moreover, there is a natural translation r L : L → L u r_{\mathcal{L}}:\mathcal{L}\to\mathcal{L}^{u} r L : L → L u with one piece 0 0 0 for each sort, sending every sort σ k \sigma_k σ k to the single sort τ \tau τ with domain formula Q k ( x ) Q_k(x) Q k ( x ) , every non-identity relation symbol R : σ ˉ R:\bar{\sigma} R : σ ˉ with ∣ σ ˉ ∣ = m |\bar{\sigma}|=m ∣ σ ˉ ∣ = m to R u ( y ˉ ) R^{u}(\bar{y}) R u ( y ˉ ) with y ˉ : τ m \bar{y}:\tau^{m} y ˉ : τ m duplicate-free, and identity to identity. Using this, we can define:
Definition 4.0.1 (Unification of a Theory). Let T T T be a many-sorted first-order theory in a language L \mathcal{L} L with finitely many sorts S T = { σ 1 , … , σ n } S_T = \{\sigma_1,\ldots,\sigma_n\} S T = { σ 1 , … , σ n } . The unification T u T^u T u of T T T is the single-sorted first-order theory defined as follows. The axioms of T u T^u T u are17 :
δ r L for each δ ∈ T (Relativization) \delta_{r_{\mathcal{L}}} \qquad\text{for each }\delta\in T \qquad\text{(Relativization)} δ r L for each δ ∈ T (Relativization) ∀ τ x ( ⋁ i = 1 n Q i ( x ) ) and ∃ τ x ( Q i ( x ) ) ( 1 ≤ i ≤ n ) (Cover) \forall_{\tau}x\left(\bigvee_{i=1}^{n}Q_i(x)\right)\quad\text{and}\quad\exists_{\tau}x\left(Q_i(x)\right)\qquad(1\leq i\leq n)\qquad\text{(Cover)} ∀ τ x ( i = 1 ⋁ n Q i ( x ) ) and ∃ τ x ( Q i ( x ) ) ( 1 ≤ i ≤ n ) (Cover) ∀ τ x ¬ ( Q i ( x ) ∧ Q j ( x ) ) ( i ≠ j ) (Disjoint) \forall_{\tau}x\ \neg\left(Q_i(x)\wedge Q_j(x)\right)\qquad(i\ne j)\qquad\text{(Disjoint)} ∀ τ x ¬ ( Q i ( x ) ∧ Q j ( x ) ) ( i = j ) (Disjoint) ∀ τ x 1 … ∀ τ x m ( R u ( x ˉ ) → ⋀ r = 1 m Q j r ( x r ) ) ( R : σ j 1 × ⋯ × σ j m ) (Sorting) \forall_{\tau}x_1\ldots\forall_{\tau}x_m\left(R^u(\bar{x})\to\bigwedge_{r=1}^{m}Q_{j_r}(x_r)\right)\qquad(R:\sigma_{j_1}\times\cdots\times\sigma_{j_m})\qquad\text{(Sorting)} ∀ τ x 1 … ∀ τ x m ( R u ( x ˉ ) → r = 1 ⋀ m Q j r ( x r ) ) ( R : σ j 1 × ⋯ × σ j m ) (Sorting)
Barrett and Halvorson show that T T T and T u T^u T u are Morita equivalent for any theory T T T with finitely many sorts, taking this to be a partial vindication of Quine’s idea about many-sorted logic. We will show that T T T and T u T^u T u are piecewise bi-interpretable, thereby demonstrating that they are equivalent in another sense, and also address a problem for unification raised by Hook [11 ] : the passage from T T T to T u T^u T u need not preserve interpretability. For instance, let T u n o T_{\mathrm{uno}} T uno say that its single sort has exactly one element, and let T d o s T_{\mathrm{dos}} T dos have two sorts, each with exactly one element. Then T d o s T_{\mathrm{dos}} T dos is interpretable in T u n o T_{\mathrm{uno}} T uno , but T d o s u T^u_{\mathrm{dos}} T dos u is not interpretable in T u n o u T^u_{\mathrm{uno}} T uno u by an ordinary interpretation, since every structure ordinarily interpretable in a one-element structure has at most one element.18 So unification of domains does not even preserve the interpretability relation, let alone act functorially on interpretations. With piecewise interpretations, both defects disappear (Theorem 4.0.3 and Corollary 4.0.4 below). In the following, note that one really only needs the piecewise structure and relativization, and not other features of interpretation, such as modification of identity or many-dimensionality, to achieve the desired goal.
Let us proceed by constructing the relevant interpretations: By construction, r L r_{\mathcal{L}} r L induces an interpretation r T : T → T u r_T:T\to T^u r T : T → T u . We can also furnish a piecewise interpretation d T : T u → T d_T:T^u\to T d T : T u → T , with set of pieces [ n ] [n] [ n ] for τ \tau τ , with τ d T k : = σ k \tau_{d_T}^k:=\sigma_k τ d T k := σ k and D τ d T , k : = ⊤ σ k D_{\tau}^{d_T,k}:=\top_{\sigma_k} D τ d T , k := ⊤ σ k . For the identity and the sortal predicates, we set
( = τ ) d T i , j ( x , y ) : = { x = σ i y if i = j ⊥ x × y if i ≠ j and ( Q k ) d T i ( x ) : = { ⊤ x if i = k ⊥ x if i ≠ k (84) (=_{\tau})_{d_T}^{i,j}(x,y):=
\begin{cases}
x=_{\sigma_i}y & \text{if }i=j\\
\bot_{x\times y} & \text{if }i\ne j
\end{cases}
\qquad\text{and}\qquad
(Q_k)_{d_T}^{i}(x):=
\begin{cases}
\top_x & \text{if }i=k\\
\bot_x & \text{if }i\ne k
\end{cases} \tag*{(84)} ( = τ ) d T i , j ( x , y ) := { x = σ i y ⊥ x × y if i = j if i = j and ( Q k ) d T i ( x ) := { ⊤ x ⊥ x if i = k if i = k (84)
( R u ) d T i 1 , … , i m ( x ˉ ) : = { R ( x ˉ ) if ( i 1 , … , i m ) = ( j 1 , … , j m ) ⊥ x ˉ otherwise ( R : σ j 1 × ⋯ × σ j m ) (85) (R^u)_{d_T}^{i_1,\ldots,i_m}(\bar{x}):=
\begin{cases}
R(\bar{x}) & \text{if }(i_1,\ldots,i_m)=(j_1,\ldots,j_m)\\
\bot_{\bar{x}} & \text{otherwise}
\end{cases}
\qquad(R:\sigma_{j_1}\times\cdots\times\sigma_{j_m}) \tag*{(85)} ( R u ) d T i 1 , … , i m ( x ˉ ) := { R ( x ˉ ) ⊥ x ˉ if ( i 1 , … , i m ) = ( j 1 , … , j m ) otherwise ( R : σ j 1 × ⋯ × σ j m ) (85)
One can easily verify that d T d_T d T respects identity with respect to T T T , noting that ( = f - Elim ) (=_f\text{ - Elim}) ( = f - Elim ) for d T d_T d T reduces to ( = - Elim ) (=\text{ - Elim}) ( = - Elim ) in L \mathcal{L} L . Let us now describe the semantic duals of both translations: Let N ⊨ T u N \models T^{u} N ⊨ T u . Then r T ∗ N ( σ k ) = { ( 0 , a ) ∣ a ∈ N ( Q k ) } / ∼ k r_{T}^{*}N(\sigma_k)=\{(0,a)\mid a\in N(Q_k)\}/\sim_k r T ∗ N ( σ k ) = {( 0 , a ) ∣ a ∈ N ( Q k )} / ∼ k , where ( 0 , a ) ∼ k ( 0 , b ) (0,a)\sim_k(0,b) ( 0 , a ) ∼ k ( 0 , b ) iff a = τ b a=_{\tau}b a = τ b . For a relation symbol R : σ i 1 × ⋯ × σ i m R:\sigma_{i_1}\times\cdots\times\sigma_{i_m} R : σ i 1 × ⋯ × σ i m ,
r T ∗ N ( R ) = { ( [ 0 , a 1 ] ∼ i 1 , … , [ 0 , a m ] ∼ i m ) ∣ a r ∈ N ( Q i r ) ( 1 ≤ r ≤ m ) , ( a 1 , … , a m ) ∈ N ( R u ) } . (86) r_{T}^{*}N(R)=\left\{\left([0,a_1]_{\sim_{i_1}},\ldots,[0,a_m]_{\sim_{i_m}}\right)\mathrel{\left| \begin{aligned} &a_r\in N(Q_{i_r})\quad(1\le r\le m),\\ &(a_1,\ldots,a_m)\in N(R^u)\end{aligned}\right.}\right\}. \tag*{(86)} r T ∗ N ( R ) = { ( [ 0 , a 1 ] ∼ i 1 , … , [ 0 , a m ] ∼ i m ) a r ∈ N ( Q i r ) ( 1 ≤ r ≤ m ) , ( a 1 , … , a m ) ∈ N ( R u ) } . (86)
Turning to d T d_T d T and a model M ⊨ T M\models T M ⊨ T , we get d T ∗ M ( τ ) = { ( k , a ) ∣ k ∈ [ n ] , a ∈ M ( σ k ) } / ∼ d_T^{*}M(\tau)=\{(k,a)\mid k\in[n],\ a\in M(\sigma_k)\}/\sim d T ∗ M ( τ ) = {( k , a ) ∣ k ∈ [ n ] , a ∈ M ( σ k )} / ∼ , where ( i , a ) ∼ ( j , b ) (i,a)\sim(j,b) ( i , a ) ∼ ( j , b ) iff i = j i=j i = j and a = σ i b a=_{\sigma_i}b a = σ i b . The new sortal predicates Q k Q_k Q k receive extensions d T ∗ M ( Q k ) = { [ k , a ] ∼ ∣ a ∈ M ( σ k ) } d_T^{*}M(Q_k)=\{[k,a]_{\sim}\mid a\in M(\sigma_k)\} d T ∗ M ( Q k ) = {[ k , a ] ∼ ∣ a ∈ M ( σ k )} . Finally, a relation symbol R u : σ i 1 × ⋯ × σ i m R^u:\sigma_{i_1}\times\cdots\times\sigma_{i_m} R u : σ i 1 × ⋯ × σ i m receives the extension
d T ∗ M ( R u ) = { ( [ i 1 , a 1 ] ∼ , … , [ i m , a m ] ∼ ) ∣ ( a 1 , … , a m ) ∈ M ( R ) } . (87) d_T^{*}M(R^u)=\left\{\left([i_1,a_1]_{\sim},\ldots,[i_m,a_m]_{\sim}\right)\mid(a_1,\ldots,a_m)\in M(R)\right\}. \tag*{(87)} d T ∗ M ( R u ) = { ( [ i 1 , a 1 ] ∼ , … , [ i m , a m ] ∼ ) ∣ ( a 1 , … , a m ) ∈ M ( R ) } . (87)
Proposition 4.0.2 (Piecewise Bi-interpretability of T T T and T u T^u T u ). Let T T T be a first-order theory in a language L \mathcal{L} L with finitely many sorts. Then T T T and T u T^u T u are piecewise bi-interpretable via the interpretations r T : T → T u r_T:T\to T^u r T : T → T u and d T : T u → T d_T:T^u\to T d T : T u → T .
Proof. We have yet to verify that d T d_T d T really is an interpretation. To that end, note that d T ∘ r T d_T\circ r_T d T ∘ r T respects identity and can hence be seen to be an interpretation L ( Σ T ) → T L(\Sigma_T)\to T L ( Σ T ) → T , as can id T \operatorname{id}_T id T . We now provide an i i i -isomorphism θ : d T ∘ r T ⇒ id T \theta:d_T\circ r_T\Rightarrow\operatorname{id}_T θ : d T ∘ r T ⇒ id T (with respect to T T T ) and its semantic dual, where one shows that it is indeed an i i i -isomorphism with said dual via Proposition 2.6.6 :
θ σ k ( 0 , i ) , 0 ( x , y ) : = { x = σ k y i = k , ⊥ x × y i ≠ k , θ M , σ k ∗ : ( d T ∘ r T ) ∗ M ( σ k ) ⟶ ( id T ∗ M ) ( σ k ) \theta_{\sigma_k}^{(0,i),0}(x,y):=
\begin{cases}
x=_{\sigma_k}y & i=k,\\
\bot_{x\times y} & i\ne k
\end{cases}
,\qquad
\theta^{*}_{M,\sigma_k}:(d_T\circ r_T)^{*}M(\sigma_k)\longrightarrow(\operatorname{id}^{*}_TM)(\sigma_k) θ σ k ( 0 , i ) , 0 ( x , y ) := { x = σ k y ⊥ x × y i = k , i = k , θ M , σ k ∗ : ( d T ∘ r T ) ∗ M ( σ k ) ⟶ ( id T ∗ M ) ( σ k )
[ ( 0 , k ) , a ] ⟼ [ 0 , a ] (88) [(0,k),a]\longmapsto[0,a] \tag*{(88)} [( 0 , k ) , a ] ⟼ [ 0 , a ] (88)
From Propositions 2.6.4 and 3.1.2, we get that T ⊢ ( δ r T ) d T ↔ δ T\vdash(\delta_{r_T})_{d_T}\leftrightarrow\delta T ⊢ ( δ r T ) d T ↔ δ for all L \mathcal{L} L -sentences δ \delta δ , whence T ⊢ ( δ r T ) d T T\vdash(\delta_{r_T})_{d_T} T ⊢ ( δ r T ) d T for all δ \delta δ such that T ⊢ δ T\vdash\delta T ⊢ δ . It remains to check that d T d_T d T preserves the axioms (Cover), (Disjoint), and (Sorting). Up to logical equivalence the translations of these axioms under d T d_T d T are, using that its domain formulae are just the respective vera:
⋀ h ∈ [ n ] ∀ σ h x ⋁ k ∈ [ n ] ( Q k ) d T h ( x ) , ⋁ h ∈ [ n ] ∃ σ h x ( Q k ) d T h ( x ) \bigwedge_{h\in[n]}\forall_{\sigma_h}x\bigvee_{k\in[n]}(Q_k)^h_{d_T}(x),\qquad\bigvee_{h\in[n]}\exists_{\sigma_h}x\,(Q_k)^h_{d_T}(x) h ∈ [ n ] ⋀ ∀ σ h x k ∈ [ n ] ⋁ ( Q k ) d T h ( x ) , h ∈ [ n ] ⋁ ∃ σ h x ( Q k ) d T h ( x )
(Coverd _d d )
⋀ h ∈ [ n ] ∀ σ h x ¬ ( ( Q k ) d T h ( x ) ∧ ( Q l ) d T h ( x ) ) ( k ≠ l ) \bigwedge_{h\in[n]}\forall_{\sigma_h}x\,\neg\left((Q_k)^h_{d_T}(x)\wedge(Q_l)^h_{d_T}(x)\right)\quad(k\ne l) h ∈ [ n ] ⋀ ∀ σ h x ¬ ( ( Q k ) d T h ( x ) ∧ ( Q l ) d T h ( x ) ) ( k = l )
(Disjointd _d d )
⋀ h ˉ ∈ [ n ] m ∀ x ˉ ( ( R u ) d T h ˉ ( x ˉ ) → ⋀ r = 1 m ( Q j r ) d T h r ( x r ) ) \bigwedge_{\bar h\in[n]^m}\forall\bar x\left((R^u)^{\bar h}_{d_T}(\bar x)\to\bigwedge_{r=1}^{m}(Q_{j_r})^{h_r}_{d_T}(x_r)\right) h ˉ ∈ [ n ] m ⋀ ∀ x ˉ ( ( R u ) d T h ˉ ( x ˉ ) → r = 1 ⋀ m ( Q j r ) d T h r ( x r ) )
(Sortingd _d d )
where in (Sortingd _d d ), R : σ j 1 × ⋯ × σ j m R:\sigma_{j_1}\times\cdots\times\sigma_{j_m} R : σ j 1 × ⋯ × σ j m and x ˉ = ( x 1 , … , x m ) : σ h 1 × ⋯ × σ h m \bar x=(x_1,\ldots,x_m):\sigma_{h_1}\times\cdots\times\sigma_{h_m} x ˉ = ( x 1 , … , x m ) : σ h 1 × ⋯ × σ h m is duplicate-free. Since ( Q k ) d T h (Q_k)^h_{d_T} ( Q k ) d T h is ⊤ x \top_x ⊤ x if h = k h=k h = k , the disjunct with k = h k=h k = h makes each conjunct of the first formula in (Coverd _d d ) provable; for the second, the disjunct with h = k h=k h = k is ∃ σ k x ( x = σ k x ) \exists_{\sigma_k}x\,(x=_{\sigma_k}x) ∃ σ k x ( x = σ k x ) , which is provable by assumption on the calculus; and in each conjunct of (Disjointd _d d ) at least one of ( Q k ) d T h (Q_k)^h_{d_T} ( Q k ) d T h , ( Q l ) d T h (Q_l)^h_{d_T} ( Q l ) d T h is ⊥ x \bot_x ⊥ x . Finally, in (Sortingd _d d ), if h ˉ = j ˉ \bar h=\bar j h ˉ = j ˉ , the consequent is a conjunction of instances of ⊤ x r \top_{x_r} ⊤ x r , and otherwise ( R u ) d T h ˉ = ⊥ x ˉ (R^u)^{\bar h}_{d_T}=\bot_{\bar x} ( R u ) d T h ˉ = ⊥ x ˉ . Therefore, d T : T u → T d_T:T^u\to T d T : T u → T is an interpretation.
Now, we produce the required i i i -isomorphism ζ : r T ∘ d T ⇒ id T u \zeta: r_T \circ d_T \Rightarrow\operatorname{id}_{T^u} ζ : r T ∘ d T ⇒ id T u , again using Proposition 2.6.6:
ζ τ ( i , 0 ) , 0 ( x , y ) : = Q i ( x ) ∧ x = τ y , ζ N , τ ∗ : ( r T ∘ d T ) N ∗ ( τ ) ⟶ ( id T u ∗ N ) ( τ ) \zeta^{(i,0),0}_{\tau}(x,y) := Q_i(x) \land x =_{\tau} y,\qquad\zeta^{*}_{N,\tau} : (r_T \circ d_T)^*_N(\tau) \longrightarrow(\operatorname{id}^*_{T^u}N)(\tau) ζ τ ( i , 0 ) , 0 ( x , y ) := Q i ( x ) ∧ x = τ y , ζ N , τ ∗ : ( r T ∘ d T ) N ∗ ( τ ) ⟶ ( id T u ∗ N ) ( τ )
[ ( i , 0 ) , a ] ⟼ [ 0 , a ] . (89) [(i,0),a] \longmapsto[0,a]. \tag*{(89)} [( i , 0 ) , a ] ⟼ [ 0 , a ] . (89)
Note that this, in particular, shows that T d o s T_{\mathrm{dos}} T dos and T d o s u T^u_{\mathrm{dos}} T dos u are piecewise bi-interpretable, but not bi-interpretable without pieces. The assignment T ↦ T u T \mapsto T^u T ↦ T u can be seen as part of a biequivalence of bicategories, which we will show in a slightly indirect fashion, to save us some work. To that end, consider two full sub-bicategories of T h p w \mathrm{Th}^{\mathrm{pw}} Th pw , namely T h f i n p w \mathrm{Th}^{\mathrm{pw}}_{\mathrm{fin}} Th fin pw , the bicategory of theories with finitely many sorts, and T h s s p w \mathrm{Th}^{\mathrm{pw}}_{\mathrm{ss}} Th ss pw , the bicategory of theories with a single sort.
Theorem 4.0.3 (Unification is Part of a Biequivalence). The inclusion pseudofunctor i : T h s s p w → T h f i n p w i : \mathrm{Th}^{\mathrm{pw}}_{\mathrm{ss}} \to\mathrm{Th}^{\mathrm{pw}}_{\mathrm{fin}} i : Th ss pw → Th fin pw is part of a biequivalence, with a quasi-inverse sending a theory T T T to T u T^u T u and an interpretation f : S → T f : S \to T f : S → T to r T ∘ ( f ∘ d S ) : S u → T u r_T \circ(f \circ d_S) : S^u \to T^u r T ∘ ( f ∘ d S ) : S u → T u .
Proof. Using choice, a pseudofunctor is part of a biequivalence iff it is an equivalence locally, that is, on each hom-category, and biessentially surjective, that is, surjective up to equivalence in the target bicategory (see, for instance, Gurski [7 ] , Lemma 3.1).19 It is clear that the inclusion pseudofunctor is a local equivalence, and it is biessentially surjective by Proposition 4.0.2. That we can choose an inverse as above is part of the proof of the characterization of biequivalences.
Corollary 4.0.4 (Preservation of Interpretability under Unification). Let S S S and T T T be first-order theories with finitely many sorts. Then S S S is piecewise interpretable in T T T iff S u S^u S u is piecewise interpretable in T u T^u T u .
Therefore, if one takes piecewise interpretations as one’s notion of morphism between theories, there is a (pseudo)functorial reduction of many-sorted to single-sorted logic. This vindicates the idea that sorts are, at least from the point of view of the theory of interpretations, ultimately dispensable, if one allows for only finitely many sorts.20
Conclusion & Future Work
In this paper, we have shown that piecewise interpretations are a well-behaved notion of interpretation between many-sorted first-order theories and developed their basic theory by organizing them into a bicategory. The developed theory of semantic duals is encapsulated by a locally faithful, contravariant pseudofunctor on that bicategory. As an application, we established a biequivalence between the bicategories of finitely many-sorted and single-sorted theories. In particular, we showed that unification, that is, replacing sorts by sortal predicates on a single sort, preserves piecewise interpretability. In future work, we will first investigate connections to categorical logic, showing that T h p w \mathrm{Th}^{\mathrm{pw}} Th pw is biequivalent to a 2-category of (suitably small) Boolean pretopoi, thereby allowing for the transfer of results from categorical logic. Following that, we will explain the relationship between ordinary and piecewise interpretations with a bicategorical analogue of Visser’s Kleisli construction (2021, Section 3.8.8), and use this construction to show that piecewise bi-interpretability coincides with Morita equivalence, appropriately defined (regarding the relationships between these notions, see [14 , 15 , 13 ] ).
Acknowledgements Funded by the European Union (ERC, FORMALISM, 101044114). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. The author would like to thank Albert Visser and Mirko Engler for helpful feedback on various parts of the project.
AI Statement The author used Claude Opus and ChatGPT Astra to help find relevant literature on the topic and to assist with developing proofs, drafting text, and identifying typos. When parts of an argument or text were suggested by an LLM, great care was taken by the author to check them, render them understandable, and present them in accordance with the overall structure and conventions of the remainder of the document. The author takes full responsibility for the content of the paper.
Canonical Context and Substitution
Definition A.0.1 (Canonical Context and Arity). For a sequence s ˉ \bar{s} s ˉ , write s ˉ ∘ \bar{s}^{\circ} s ˉ ∘ for the subsequence obtained by deleting entries occurring more than once from right to left, retaining only their first occurrences, which we will also call its deduplication. Note that ( a ˉ ∘ × b ˉ ∘ ) ∘ = ( a ˉ × b ˉ ) ∘ (\bar{a}^{\circ} \times\bar{b}^{\circ})^{\circ} = (\bar{a} \times\bar{b})^{\circ} ( a ˉ ∘ × b ˉ ∘ ) ∘ = ( a ˉ × b ˉ ) ∘ for any sequences a ˉ , b ˉ \bar{a},\bar{b} a ˉ , b ˉ . The canonical context x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) of a formula φ \varphi φ is defined inductively:
(i) Truth Values: x ˉ ( ⊤ ) = x ˉ ( ⊥ ) = ε \bar{x}(\top) = \bar{x}(\bot) = \varepsilon x ˉ ( ⊤ ) = x ˉ ( ⊥ ) = ε .
(ii) Atomic: x ˉ ( R ( z 1 , … , z n ) ) = ( z 1 , … , z n ) ∘ \bar{x}(R(z_1,\ldots,z_n)) = (z_1,\ldots,z_n)^{\circ} x ˉ ( R ( z 1 , … , z n )) = ( z 1 , … , z n ) ∘ .
(iii) Binary: x ˉ ( φ ⋆ ψ ) = ( x ˉ ( φ ) × x ˉ ( ψ ) ) ∘ \bar{x}(\varphi\star\psi) = (\bar{x}(\varphi) \times\bar{x}(\psi))^{\circ} x ˉ ( φ ⋆ ψ ) = ( x ˉ ( φ ) × x ˉ ( ψ ) ) ∘ for ⋆ ∈ { ∧ , ∨ , → } \star\in\{\land,\lor,\to\} ⋆ ∈ { ∧ , ∨ , → } .
(iv) Negation: x ˉ ( ¬ φ ) = x ˉ ( φ ) \bar{x}(\neg\varphi) = \bar{x}(\varphi) x ˉ ( ¬ φ ) = x ˉ ( φ ) .
(v) Quantifier: x ˉ ( ∀ z φ ) \bar{x}(\forall z\varphi) x ˉ ( ∀ z φ ) is x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) with z z z deleted.
The arity of φ \varphi φ is the sequence of arities of its canonical context, written σ ˉ ( φ ) \bar{\sigma}(\varphi) σ ˉ ( φ ) . We also write φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ .
We will make frequent use of the following map induced by deduplication: If x ˉ \bar{x} x ˉ is a sequence, denote by d x ˉ : [ ∣ x ˉ ∣ ] → [ ∣ x ˉ ∘ ∣ ] d_{\bar{x}} : [|\bar{x}|] \to[|\bar{x}^{\circ}|] d x ˉ : [ ∣ x ˉ ∣ ] → [ ∣ x ˉ ∘ ∣ ] the unique surjection such that ( x ˉ ∘ ) d x ˉ ( i ) = x i (\bar{x}^{\circ})_{d_{\bar{x}}(i)} = x_i ( x ˉ ∘ ) d x ˉ ( i ) = x i , sending an index i i i to the index of the first occurrence of x i x_i x i in x ˉ ∘ \bar{x}^{\circ} x ˉ ∘ . We will omit the subscript where the context allows it. The deduplication map d x ˉ d_{\bar{x}} d x ˉ induces a function on sequences of the same length as x ˉ ∘ \bar{x}^{\circ} x ˉ ∘ to sequences of the length of x ˉ \bar{x} x ˉ , via ( y ˉ d x ˉ ) i : = y d x ˉ ( i ) (\bar{y}^{d_{\bar{x}}})_i := y_{d_{\bar{x}}(i)} ( y ˉ d x ˉ ) i := y d x ˉ ( i ) . This function satisfies ( x ˉ ∘ ) d x ˉ = x ˉ (\bar{x}^{\circ})^{d_{\bar{x}}} = \bar{x} ( x ˉ ∘ ) d x ˉ = x ˉ .
Recall that, given a context y ˉ = ( y 1 , … , y m ) : τ ˉ \bar{y} = (y_1,\ldots,y_m) : \bar{\tau} y ˉ = ( y 1 , … , y m ) : τ ˉ for a formula φ \varphi φ with canonical context x ˉ ( φ ) = ( x 1 , … , x n ) : σ ˉ \bar{x}(\varphi) = (x_1,\ldots,x_n) : \bar{\sigma} x ˉ ( φ ) = ( x 1 , … , x n ) : σ ˉ , there is a unique injection c : [ n ] → [ m ] c : [n] \to[m] c : [ n ] → [ m ] such that x j = y c ( j ) x_j = y_{c(j)} x j = y c ( j ) for 1 ≤ j ≤ n 1 \leq j \leq n 1 ≤ j ≤ n . We call c c c the context injection of y ˉ \bar{y} y ˉ relative to φ \varphi φ .
Definition A.0.2 (Simultaneous Substitution). For a formula φ \varphi φ with canonical context x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ and a sequence y ˉ : σ ˉ \bar{y} : \bar{\sigma} y ˉ : σ ˉ of the same arity, the simultaneous substitution φ ( y ˉ ) \varphi(\bar{y}) φ ( y ˉ ) is defined inductively:
(i) Truth Values: ⊤ ( y ˉ ) = ⊤ \top(\bar{y}) = \top ⊤ ( y ˉ ) = ⊤ and ⊥ ( y ˉ ) = ⊥ \bot(\bar{y}) = \bot ⊥ ( y ˉ ) = ⊥ .
(ii) Atomic: R ( z ˉ ) ( y ˉ ) : = R ( y ˉ d z ˉ ) R(\bar{z})(\bar{y}) := R(\bar{y}^{d_{\bar{z}}}) R ( z ˉ ) ( y ˉ ) := R ( y ˉ d z ˉ ) .
(iii) Binary: ( φ 1 ⋆ φ 2 ) ( y ˉ ) : = φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 ) (\varphi_1 \star\varphi_2)(\bar{y}) := \varphi_1(\bar{y}^{c_1}) \star\varphi_2(\bar{y}^{c_2}) ( φ 1 ⋆ φ 2 ) ( y ˉ ) := φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 ) , where c i : [ ∣ x ˉ ( φ i ) ∣ ] → [ ∣ x ˉ ( φ 1 ⋆ φ 2 ) ∣ ] c_i : [|\bar{x}(\varphi_i)|] \to[|\bar{x}(\varphi_1 \star\varphi_2)|] c i : [ ∣ x ˉ ( φ i ) ∣ ] → [ ∣ x ˉ ( φ 1 ⋆ φ 2 ) ∣ ] are the context injections for x ˉ ( φ 1 ⋆ φ 2 ) \bar{x}(\varphi_1 \star\varphi_2) x ˉ ( φ 1 ⋆ φ 2 ) relative to φ i \varphi_i φ i , and y ˉ j c i : = y c i ( j ) \bar{y}^{c_i}_j := y_{c_i(j)} y ˉ j c i := y c i ( j ) .
(iv) Negation: ( ¬ φ ) ( y ˉ ) : = ¬ ( φ ( y ˉ ) ) (\neg\varphi)(\bar{y}) := \neg(\varphi(\bar{y})) ( ¬ φ ) ( y ˉ ) := ¬ ( φ ( y ˉ )) .
(v) Quantifier: Let ∀ τ z φ : σ ˉ \forall_{\tau} z\ \varphi: \bar{\sigma} ∀ τ z φ : σ ˉ and y ˉ : σ ˉ \bar{y} : \bar{\sigma} y ˉ : σ ˉ . Since x ˉ ( ∀ τ z φ ) × z \bar{x}(\forall_{\tau} z\ \varphi) \times z x ˉ ( ∀ τ z φ ) × z is a context for φ \varphi φ , let c : [ ∣ x ˉ ( φ ) ∣ ] ↪ [ ∣ σ ˉ ∣ + 1 ] c : [|\bar{x}(\varphi)|] \hookrightarrow[|\bar{\sigma}|+1] c : [ ∣ x ˉ ( φ ) ∣ ] ↪ [ ∣ σ ˉ ∣ + 1 ] be the context injection of x ˉ ( ∀ τ z φ ) × z \bar{x}(\forall_{\tau} z\ \varphi) \times z x ˉ ( ∀ τ z φ ) × z relative to φ \varphi φ . Choose a variable z ′ : τ z' : \tau z ′ : τ , fresh for y ˉ \bar{y} y ˉ and φ \varphi φ , and set
( ∀ z φ ) ( y ˉ ) : = ∀ z ′ φ ( ( y ˉ × z ′ ) c ) . (90) (\forall z\ \varphi)(\bar{y}) := \forall z'\ \varphi\left((\bar{y} \times z')^c\right). \tag*{(90)} ( ∀ z φ ) ( y ˉ ) := ∀ z ′ φ ( ( y ˉ × z ′ ) c ) . (90)
In the quantifier clause of the last definition, different choices of fresh z ′ z' z ′ will produce α \alpha α -variants of the resulting formulae. This has the consequence that translations, as defined in the paper, are, strictly speaking, only functional up to renaming of bound variables, which will be enough for our purposes. While the following lemmata are essentially bookkeeping and require no special insight, we record them in detail, because the details are a bit finicky. We first need some auxiliary combinatorial lemmata.
Lemma A.0.3 (Deduplication under Surjection). Let a ˉ 1 , … , a ˉ n \bar{a}_1,\ldots,\bar{a}_n a ˉ 1 , … , a ˉ n be finite sequences and let s : [ m ] ↠ [ n ] s:[m]\twoheadrightarrow[n] s : [ m ] ↠ [ n ] be a surjection such that min s − 1 ( i ) < min s − 1 ( j ) \min s^{-1}(i)<\min s^{-1}(j) min s − 1 ( i ) < min s − 1 ( j ) whenever i < j i<j i < j . Then
( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m ) ) ∘ = ( a ˉ 1 × ⋯ × a ˉ n ) ∘ . (91) (\bar{a}_{s(1)}\times\cdots\times\bar{a}_{s(m)})^\circ=(\bar{a}_1\times\cdots\times\bar{a}_n)^\circ. \tag*{(91)} ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m ) ) ∘ = ( a ˉ 1 × ⋯ × a ˉ n ) ∘ . (91) In particular, if the sequences a ˉ i \bar{a}_i a ˉ i are pairwise disjoint and individually duplicate-free, then ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m ) ) ∘ = a ˉ 1 × ⋯ × a ˉ n (\bar{a}_{s(1)}\times\cdots\times\bar{a}_{s(m)})^\circ=\bar{a}_1\times\cdots\times\bar{a}_n ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m ) ) ∘ = a ˉ 1 × ⋯ × a ˉ n .
In applications, the surjection s s s will typically be the deduplication map for a sequence other than a ˉ \bar{a} a ˉ . One should also note that this covers the case where the concatenated sequences each have length one.
Proof. Induction on m ≥ 1 m\geq1 m ≥ 1 . For m , n ≤ 1 m,n\leq1 m , n ≤ 1 , the claim is obvious. Now, let m > 1 m>1 m > 1 and assume the claim holds for all 1 ≤ m ′ < m 1\leq m'<m 1 ≤ m ′ < m . If s ( m ) = s ( j ) s(m)=s(j) s ( m ) = s ( j ) for some 1 ≤ j ≤ m − 1 1\leq j\leq m-1 1 ≤ j ≤ m − 1 , then the restriction of s s s to [ m − 1 ] [m-1] [ m − 1 ] satisfies the hypotheses for [ n ] [n] [ n ] and ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m − 1 ) × a ˉ s ( m ) ) ∘ = ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m − 1 ) ) ∘ = ( a ˉ 1 × ⋯ × a ˉ n ) ∘ (\bar{a}_{s(1)}\times\cdots\times\bar{a}_{s(m-1)}\times\bar{a}_{s(m)})^\circ=(\bar{a}_{s(1)}\times\cdots\times\bar{a}_{s(m-1)})^\circ=(\bar{a}_1\times\cdots\times\bar{a}_n)^\circ ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m − 1 ) × a ˉ s ( m ) ) ∘ = ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m − 1 ) ) ∘ = ( a ˉ 1 × ⋯ × a ˉ n ) ∘ by the induction hypothesis. Otherwise, the hypothesis on s s s forces s ( m ) = n s(m)=n s ( m ) = n , whence the restriction of s s s to [ m − 1 ] [m-1] [ m − 1 ] satisfies the hypothesis for [ n − 1 ] [n-1] [ n − 1 ] . Therefore:
( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m ) ) ∘ = ( ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m − 1 ) ) ∘ × a ˉ s ( m ) ) ∘ (92) \begin{aligned}
(\bar{a}_{s(1)}\times\cdots\times\bar{a}_{s(m)})^\circ&=((\bar{a}_{s(1)}\times\cdots\times\bar{a}_{s(m-1)})^\circ\times\bar{a}_{s(m)})^\circ \tag*{(92)}
\end{aligned} ( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m ) ) ∘ = (( a ˉ s ( 1 ) × ⋯ × a ˉ s ( m − 1 ) ) ∘ × a ˉ s ( m ) ) ∘ (92)
= ( ( a ˉ 1 × ⋯ × a ˉ n − 1 ) ∘ × a ˉ n ) ∘ (93) =((\bar{a}_1\times\cdots\times\bar{a}_{n-1})^\circ\times\bar{a}_n)^\circ \tag*{(93)} = (( a ˉ 1 × ⋯ × a ˉ n − 1 ) ∘ × a ˉ n ) ∘ (93)
= ( a ˉ 1 × ⋯ × a ˉ n ) ∘ (94) =(\bar{a}_1\times\cdots\times\bar{a}_n)^\circ \tag*{(94)} = ( a ˉ 1 × ⋯ × a ˉ n ) ∘ (94)
Corollary A.0.4 (Deduplication Map under Surjection). Let a ˉ \bar{a} a ˉ be a sequence of length n n n and let s : [ m ] ↠ [ n ] s:[m]\twoheadrightarrow[n] s : [ m ] ↠ [ n ] be a surjection with min s − 1 ( i ) < min s − 1 ( j ) \min s^{-1}(i)<\min s^{-1}(j) min s − 1 ( i ) < min s − 1 ( j ) whenever i < j i<j i < j . Then
d a ˉ ∘ s = d a ˉ ∘ s (95) d_{\bar{a}\circ s}=d_{\bar{a}}\circ s \tag*{(95)} d a ˉ ∘ s = d a ˉ ∘ s (95)
We now record a lemma about the interaction of context injections and deduplications.
Lemma A.0.5 (Deduplication Map under Context Injection). Let y ˉ \bar{y} y ˉ be a sequence of length n n n and let c : [ m ] ↪ [ n ] c:[m]\hookrightarrow[n] c : [ m ] ↪ [ n ] be an injection, with y ˉ c \bar{y}^{c} y ˉ c the induced sequence ( y ˉ c ) i = y c ( i ) (\bar{y}^{c})_i=y_{c(i)} ( y ˉ c ) i = y c ( i ) . The entries of ( y ˉ c ) ∘ (\bar{y}^{c})^\circ ( y ˉ c ) ∘ occur among those of the duplicate-free sequence y ˉ ∘ \bar{y}^\circ y ˉ ∘ , so there is a unique injection c ^ : [ ∣ ( y ˉ c ) ∘ ∣ ] ↪ [ ∣ y ˉ ∘ ∣ ] \hat{c}:[|(\bar{y}^{c})^\circ|]\hookrightarrow[|\bar{y}^\circ|] c ^ : [ ∣ ( y ˉ c ) ∘ ∣ ] ↪ [ ∣ y ˉ ∘ ∣ ] with ( y ˉ ∘ ) c ^ ( i ) = ( ( y ˉ c ) ∘ ) i (\bar{y}^\circ)_{\hat{c}(i)}=((\bar{y}^{c})^\circ)_i ( y ˉ ∘ ) c ^ ( i ) = (( y ˉ c ) ∘ ) i . Then
d y ˉ ∘ c = c ^ ∘ d y ˉ c . (96) d_{\bar{y}}\circ c=\hat{c}\circ d_{\bar{y}^{c}}. \tag*{(96)} d y ˉ ∘ c = c ^ ∘ d y ˉ c . (96)
Proof. Both sides are functions [ m ] → [ ∣ y ˉ ∘ ∣ ] [m]\to[|\bar{y}^\circ|] [ m ] → [ ∣ y ˉ ∘ ∣ ] . For i ∈ [ m ] i\in[m] i ∈ [ m ] , ( d y ˉ ∘ c ) ( i ) = d y ˉ ( c ( i ) ) (d_{\bar{y}}\circ c)(i)=d_{\bar{y}}(c(i)) ( d y ˉ ∘ c ) ( i ) = d y ˉ ( c ( i )) is the position of y c ( i ) y_{c(i)} y c ( i ) in y ˉ ∘ \bar{y}^\circ y ˉ ∘ . On the other side d y ˉ c ( i ) d_{\bar{y}^{c}}(i) d y ˉ c ( i ) is the position of ( y ˉ c ) i = y c ( i ) (\bar{y}^{c})_i=y_{c(i)} ( y ˉ c ) i = y c ( i ) in ( y ˉ c ) ∘ (\bar{y}^{c})^\circ ( y ˉ c ) ∘ , and c ^ \hat{c} c ^ maps it to the position of that same element y c ( i ) y_{c(i)} y c ( i ) in y ˉ ∘ \bar{y}^\circ y ˉ ∘ . As y ˉ ∘ \bar{y}^\circ y ˉ ∘ is duplicate-free, this position is unique, so the two maps agree.
Lemma 1.1.1 (Canonical Context under Substitution). Let φ \varphi φ be a formula with canonical context x ˉ = ( x 1 , … , x n ) : σ ˉ \bar{x}=(x_1,\ldots,x_n):\bar{\sigma} x ˉ = ( x 1 , … , x n ) : σ ˉ , and let y ˉ = ( y 1 , … , y n ) : σ ˉ \bar{y}=(y_1,\ldots,y_n):\bar{\sigma} y ˉ = ( y 1 , … , y n ) : σ ˉ . Then the canonical context of φ ( y ˉ ) \varphi(\bar{y}) φ ( y ˉ ) is y ˉ ∘ \bar{y}^\circ y ˉ ∘ , the sequence resulting from y ˉ \bar{y} y ˉ by deleting repeated members from right to left.
Proof. By induction on φ \varphi φ . The truth-value and negation cases are immediate, as both substitution and canonical context commute with negation.
Atomic: For any sequence z ˉ \bar{z} z ˉ , the deduplication map d z ˉ d_{\bar{z}} d z ˉ is surjective and monotone on first occurrences, i.e., it satisfies min d z ˉ − 1 ( i ) < min d z ˉ − 1 ( j ) \min d_{\bar{z}}^{-1}(i)<\min d_{\bar{z}}^{-1}(j) min d z ˉ − 1 ( i ) < min d z ˉ − 1 ( j ) if i < j i<j i < j . Using Lemma A.0.3 and the definition of substitution for atomic formulae, we get:
x ˉ ( R ( z ˉ ) ( y ˉ ) ) = x ˉ ( R ( y ˉ − d z ˉ ) ) = ( y ˉ − d z ˉ ) ∘ = y ˉ ∘ (97) \bar{x}(R(\bar{z})(\bar{y}))=\bar{x}(R(\bar{y}^{-d_{\bar{z}}}))=(\bar{y}^{-d_{\bar{z}}})^\circ=\bar{y}^\circ \tag*{(97)} x ˉ ( R ( z ˉ ) ( y ˉ )) = x ˉ ( R ( y ˉ − d z ˉ )) = ( y ˉ − d z ˉ ) ∘ = y ˉ ∘ (97)
Binary: The substitution clause for the binary connectives yields ( φ 1 ⋆ φ 2 ) ( y ˉ ) = φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 ) (\varphi_1\star\varphi_2)(\bar{y})=\varphi_1(\bar{y}^{c_1})\star\varphi_2(\bar{y}^{c_2}) ( φ 1 ⋆ φ 2 ) ( y ˉ ) = φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 ) . By induction x ˉ ( φ i ( y ˉ c i ) ) = ( y ˉ c i ) ∘ \bar{x}(\varphi_i(\bar{y}^{c_i}))=(\bar{y}^{c_i})^\circ x ˉ ( φ i ( y ˉ c i )) = ( y ˉ c i ) ∘ , so
x ˉ ( ( φ 1 ⋆ φ 2 ) ( y ˉ ) ) = ( ( y ˉ c 1 ) ∘ × ( y ˉ c 2 ) ∘ ) ∘ = ( y ˉ c 1 × y ˉ c 2 ) ∘ = ( y s ( 1 ) , … , y s ( n 1 + n 2 ) ) ∘ (98) \bar{x}((\varphi_1\star\varphi_2)(\bar{y}))=((\bar{y}^{c_1})^\circ\times(\bar{y}^{c_2})^\circ)^\circ=(\bar{y}^{c_1}\times\bar{y}^{c_2})^\circ=(y_{s(1)},\ldots,y_{s(n_1+n_2)})^\circ \tag*{(98)} x ˉ (( φ 1 ⋆ φ 2 ) ( y ˉ )) = (( y ˉ c 1 ) ∘ × ( y ˉ c 2 ) ∘ ) ∘ = ( y ˉ c 1 × y ˉ c 2 ) ∘ = ( y s ( 1 ) , … , y s ( n 1 + n 2 ) ) ∘ (98)
where s s s is the concatenation of c 1 c_1 c 1 and c 2 c_2 c 2 and n i = ∣ x ˉ ( φ i ) ∣ n_i=|\bar{x}(\varphi_i)| n i = ∣ x ˉ ( φ i ) ∣ . By definition, x ˉ ( φ 1 ⋆ φ 2 ) = ( x ˉ ( φ 1 ) × x ˉ ( φ 2 ) ) ∘ \bar{x}(\varphi_1\star\varphi_2)=(\bar{x}(\varphi_1)\times\bar{x}(\varphi_2))^\circ x ˉ ( φ 1 ⋆ φ 2 ) = ( x ˉ ( φ 1 ) × x ˉ ( φ 2 ) ) ∘ , the map s s s is surjective with min s − 1 \min s^{-1} min s − 1 strictly increasing, so Lemma A.0.3 yields ( y s ( 1 ) , … , y s ( n 1 + n 2 ) ) ∘ = y ˉ ∘ (y_{s(1)},\ldots,y_{s(n_1+n_2)})^\circ=\bar{y}^\circ ( y s ( 1 ) , … , y s ( n 1 + n 2 ) ) ∘ = y ˉ ∘ .
Quantifier: Let φ = ∀ z ψ \varphi=\forall z\psi φ = ∀ z ψ and let c c c be the context injection of x ˉ ( φ ) × z \bar{x}(\varphi)\times z x ˉ ( φ ) × z relative to ψ \psi ψ . By the quantifier clause for substitution, φ ( y ˉ ) = ∀ z ′ ψ ( ( y ˉ × z ′ ) c ) \varphi(\bar{y})=\forall z'\psi((\bar{y}\times z')^c) φ ( y ˉ ) = ∀ z ′ ψ (( y ˉ × z ′ ) c ) with z ′ z' z ′ fresh, and by the induction hypothesis x ˉ ( ψ ( ( y ˉ × z ′ ) c ) ) = ( ( y ˉ × z ′ ) c ) ∘ \bar{x}(\psi((\bar{y}\times z')^c))=((\bar{y}\times z')^c)^\circ x ˉ ( ψ (( y ˉ × z ′ ) c )) = (( y ˉ × z ′ ) c ) ∘ . Hence x ˉ ( φ ( y ˉ ) ) \bar{x}(\varphi(\bar{y})) x ˉ ( φ ( y ˉ )) is ( ( y ˉ × z ′ ) c ) ∘ ((\bar{y}\times z')^c)^\circ (( y ˉ × z ′ ) c ) ∘ with z ′ z' z ′ deleted. Now, the position of z ′ z' z ′ in ( y ˉ × z ′ ) c (\bar{y}\times z')^c ( y ˉ × z ′ ) c , if it occurs, is that of z z z in x ˉ ( ψ ) \bar{x}(\psi) x ˉ ( ψ ) , whence deleting z ′ z' z ′ in the former and deduplicating yields y ˉ ∘ \bar{y}^\circ y ˉ ∘ . □
Lemma 2.1.6 (Canonical Context of Translations). Let f : Σ 1 → Σ 2 f:\Sigma_1\to\Sigma_2 f : Σ 1 → Σ 2 be a piecewise reconstrual and let φ \varphi φ be an L 1 \mathcal{L}_1 L 1 -formula with canonical context x ˉ : σ ˉ \bar{x}:\bar{\sigma} x ˉ : σ ˉ . Then, for any piece k ˉ ∈ P σ ˉ \bar{k}\in P_{\bar{\sigma}} k ˉ ∈ P σ ˉ , the canonical context of φ f k ˉ \varphi_f^{\bar{k}} φ f k ˉ is x ˉ f k ˉ \bar{x}_f^{\bar{k}} x ˉ f k ˉ . In particular, the formula φ f k ˉ \varphi_f^{\bar{k}} φ f k ˉ has exactly the free variables x ˉ f k ˉ \bar{x}_f^{\bar{k}} x ˉ f k ˉ .
Proof. By induction on φ \varphi φ , strengthening the inductive hypothesis to: for every context y ˉ : τ ˉ \bar{y}:\bar{\tau} y ˉ : τ ˉ for φ \varphi φ with context injection c c c and every k ˉ ∈ P τ ˉ f \bar{k}\in P_{\bar{\tau}}^f k ˉ ∈ P τ ˉ f , the canonical context of ( y ˉ ∣ φ ) f k ˉ (\bar{y}\mid\varphi)_f^{\bar{k}} ( y ˉ ∣ φ ) f k ˉ is x ˉ ( φ ) f k ˉ \bar{x}(\varphi)_f^{\bar{k}} x ˉ ( φ ) f k ˉ . By the context clause ( y ˉ ∣ φ ) f k ˉ = φ f k ˉ (\bar{y}\mid\varphi)_f^{\bar{k}}=\varphi_f^{\bar{k}} ( y ˉ ∣ φ ) f k ˉ = φ f k ˉ , so it suffices to verify, for each φ \varphi φ in canonical context x ˉ : σ ˉ \bar{x}:\bar{\sigma} x ˉ : σ ˉ and each k ˉ ∈ P σ ˉ f \bar{k}\in P_{\bar{\sigma}}^f k ˉ ∈ P σ ˉ f , that x ˉ ( φ f k ˉ ) = x ˉ f k ˉ \bar{x}(\varphi_f^{\bar{k}})=\bar{x}_f^{\bar{k}} x ˉ ( φ f k ˉ ) = x ˉ f k ˉ . By the sortal and no-overlap conditions on the variable function of a reconstrual, the sequences ( ( x ˉ i ) f k ˉ i ) ((\bar{x}_i)_f^{\bar{k}_i}) (( x ˉ i ) f k ˉ i ) are individually duplicate-free and pairwise disjoint. The truth-value and negation cases are immediate.
Atomic: For R ( z ˉ ) : σ ˉ R(\bar{z}):\bar{\sigma} R ( z ˉ ) : σ ˉ with ∣ z ˉ ∣ = n |\bar{z}|=n ∣ z ˉ ∣ = n , the atomic clause yields R ( z ˉ ) f k ˉ = R f k ˉ d ( ( z ˉ f k ˉ d ) ) R(\bar{z})_f^{\bar{k}}=R_f^{\bar{k}d}((\bar{z}_f^{\bar{k}d})) R ( z ˉ ) f k ˉ = R f k ˉ d (( z ˉ f k ˉ d )) . By Lemma 1.1.1 its canonical context is
( z ˉ f k ˉ d ) ∘ = ( ( z 1 ) f k d ( 1 ) × ⋯ × ( z n ) f k d ( n ) ) ∘ = ( ( ( z ˉ ∘ ) d ( 1 ) ) f k d ( 1 ) × ⋯ × ( ( z ˉ ∘ ) d ( n ) ) f k d ( n ) ) ∘ = ( z ˉ ∘ ) f k ˉ (99) (\bar{z}_f^{\bar{k}d})^\circ=((z_1)_f^{k_{d(1)}}\times\cdots\times(z_n)_f^{k_{d(n)}})^\circ=(((\bar{z}^\circ)_{d(1)})_f^{k_{d(1)}}\times\cdots\times((\bar{z}^\circ)_{d(n)})_f^{k_{d(n)}})^\circ=(\bar{z}^\circ)_f^{\bar{k}} \tag*{(99)} ( z ˉ f k ˉ d ) ∘ = (( z 1 ) f k d ( 1 ) × ⋯ × ( z n ) f k d ( n ) ) ∘ = ((( z ˉ ∘ ) d ( 1 ) ) f k d ( 1 ) × ⋯ × (( z ˉ ∘ ) d ( n ) ) f k d ( n ) ) ∘ = ( z ˉ ∘ ) f k ˉ (99)
where d : = d z ˉ d:=d_{\bar{z}} d := d z ˉ and the first equality expands, the second uses ( z ˉ ∘ ) d z ˉ = z ˉ ∘ (\bar{z}^\circ)^{d_{\bar{z}}}=\bar{z}^\circ ( z ˉ ∘ ) d z ˉ = z ˉ ∘ for concatenated sequences, and the third applies Lemma A.0.3.
Binary: Let φ = φ 1 ⋆ φ 2 \varphi=\varphi_1\star\varphi_2 φ = φ 1 ⋆ φ 2 . By the clause for binary connectives φ f k ˉ = ( x ˉ ∣ φ 1 ) f k ˉ ⋆ ( x ˉ ∣ φ 2 ) f k ˉ \varphi_f^{\bar{k}}=(\bar{x}\mid\varphi_1)_f^{\bar{k}}\star(\bar{x}\mid\varphi_2)_f^{\bar{k}} φ f k ˉ = ( x ˉ ∣ φ 1 ) f k ˉ ⋆ ( x ˉ ∣ φ 2 ) f k ˉ . By induction, the canonical context of ( x ˉ ∣ φ i ) f k ˉ (\bar{x}\mid\varphi_i)_f^{\bar{k}} ( x ˉ ∣ φ i ) f k ˉ is x ˉ ( φ i ) f k ˉ c i \bar{x}(\varphi_i)_f^{\bar{k}c_i} x ˉ ( φ i ) f k ˉ c i . Lemma A.0.3 with s s s the concatenation of c 1 c_1 c 1 and c 2 c_2 c 2 implies ( x ˉ ( φ 1 ) f k ˉ c 1 × x ˉ ( φ 2 ) f k ˉ c 2 ) ∘ = x ˉ f k ˉ (\bar{x}(\varphi_1)_f^{\bar{k}c_1}\times\bar{x}(\varphi_2)_f^{\bar{k}c_2})^\circ=\bar{x}_f^{\bar{k}} ( x ˉ ( φ 1 ) f k ˉ c 1 × x ˉ ( φ 2 ) f k ˉ c 2 ) ∘ = x ˉ f k ˉ . Quantifier: Let φ = ∀ z ψ \varphi= \forall z\,\psi φ = ∀ z ψ and let x ˉ \bar{x} x ˉ be the canonical context of φ \varphi φ . By the quantifier clause for translations,
φ ˉ f k ˉ = ⋀ h ∈ P τ k ˉ ∀ z f h ( D τ h ( z f h ) → ( x ˉ × z ∣ ψ ) f ( k ˉ × h ) ) . (100) \bar{\varphi}_{f}^{\bar{k}} = \bigwedge_{h \in P_{\tau}^{\bar{k}}} \forall z_{f}^{h}\left(D_{\tau}^{h}(z_{f}^{h}) \to\left(\bar{x} \times z \mid\psi\right)_{f}^{(\bar{k} \times h)}\right). \tag*{(100)} φ ˉ f k ˉ = h ∈ P τ k ˉ ⋀ ∀ z f h ( D τ h ( z f h ) → ( x ˉ × z ∣ ψ ) f ( k ˉ × h ) ) . (100)
By induction, the canonical context of ( x ˉ × z ∣ ψ ) f ( k ˉ × h ) \left(\bar{x} \times z \mid\psi\right)_{f}^{(\bar{k} \times h)} ( x ˉ × z ∣ ψ ) f ( k ˉ × h ) is x ˉ ( ψ ) f ( k ˉ × h ) c \bar{x}(\psi)_{f}^{(\bar{k} \times h)c} x ˉ ( ψ ) f ( k ˉ × h ) c , and by Lemma 1.1.1, D τ h ( z f h ) D_{\tau}^{h}(z_{f}^{h}) D τ h ( z f h ) has canonical context z f h z_{f}^{h} z f h . Since x ˉ \bar{x} x ˉ results from x ˉ ( ψ ) \bar{x}(\psi) x ˉ ( ψ ) by deleting z z z , the sequence x ˉ ( ψ ) f ( k ˉ × h ) c \bar{x}(\psi)_{f}^{(\bar{k} \times h)c} x ˉ ( ψ ) f ( k ˉ × h ) c is x ˉ f k ˉ \bar{x}_{f}^{\bar{k}} x ˉ f k ˉ with the subsequence z f h z_{f}^{h} z f h possibly inserted at the position of z z z . As these subsequences are duplicate-free and pairwise disjoint, the formula in the scope of the quantifier has canonical context ( z f h × x ˉ ( ψ ) f ( k ˉ × h ) c ) ∘ = z f h × x ˉ f k ˉ \left(z_{f}^{h} \times\bar{x}(\psi)_{f}^{(\bar{k} \times h)c}\right)^{\circ} = z_{f}^{h} \times\bar{x}_{f}^{\bar{k}} ( z f h × x ˉ ( ψ ) f ( k ˉ × h ) c ) ∘ = z f h × x ˉ f k ˉ . Binding z f h z_{f}^{h} z f h deletes this subsequence, whence each conjunct, and therefore the whole conjunction, has canonical context x ˉ f k ˉ \bar{x}_{f}^{\bar{k}} x ˉ f k ˉ .
For the next lemma, we need a preparatory lemma on iterated substitutions.
Lemma A.0.6 (Iterated Substitution). Let φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ be an L \mathcal{L} L -formula with canonical context x ˉ : σ ˉ \bar{x} : \bar{\sigma} x ˉ : σ ˉ , and let y ˉ : σ ˉ \bar{y} : \bar{\sigma} y ˉ : σ ˉ be another variable sequence of the same arity. By Lemma 1.1.1, the canonical context of φ ( y ˉ ) \varphi(\bar{y}) φ ( y ˉ ) is y ˉ ∘ : τ ˉ \bar{y}^{\circ} : \bar{\tau} y ˉ ∘ : τ ˉ . Then, for every variable sequence z ˉ : τ ˉ \bar{z} : \bar{\tau} z ˉ : τ ˉ ,
( φ ( y ˉ ) ) ( z ˉ ) = φ ( z ˉ d y ˉ ) (101) (\varphi(\bar{y}))(\bar{z}) = \varphi(\bar{z}^{d_{\bar{y}}}) \tag*{(101)} ( φ ( y ˉ )) ( z ˉ ) = φ ( z ˉ d y ˉ ) (101) up to renaming of bound variables.
Proof. By induction on φ \varphi φ . The cases for truth values and negation are clear.
Atomic: Let φ = R ( w ˉ ) : σ ˉ \varphi= R(\bar{w}) : \bar{\sigma} φ = R ( w ˉ ) : σ ˉ be an atomic formula. Then ( φ ( y ˉ ) ) ( z ˉ ) = R ( y ˉ d w ˉ ) ( z ˉ ) (\varphi(\bar{y}))(\bar{z}) = R(\bar{y}^{d_{\bar{w}}})(\bar{z}) ( φ ( y ˉ )) ( z ˉ ) = R ( y ˉ d w ˉ ) ( z ˉ ) . Applying Corollary A.0.4 , d y ˉ d w ˉ = d y ˉ ∘ d w ˉ d_{\bar{y}^{d_{\bar{w}}}} = d_{\bar{y}} \circ d_{\bar{w}} d y ˉ d w ˉ = d y ˉ ∘ d w ˉ . Therefore,
R ( y ˉ d w ˉ ) ( z ˉ ) = R ( z ˉ d y ˉ d w ˉ ) = R ( z ˉ d y ˉ ∘ d w ˉ ) = R ( ( z ˉ d y ˉ ) d w ˉ ) = φ ( z ˉ d y ˉ ) . (102) R(\bar{y}^{d_{\bar{w}}})(\bar{z}) = R(\bar{z}^{d_{\bar{y}^{d_{\bar{w}}}}}) = R(\bar{z}^{d_{\bar{y}} \circ d_{\bar{w}}}) = R((\bar{z}^{d_{\bar{y}}})^{d_{\bar{w}}}) = \varphi(\bar{z}^{d_{\bar{y}}}). \tag*{(102)} R ( y ˉ d w ˉ ) ( z ˉ ) = R ( z ˉ d y ˉ d w ˉ ) = R ( z ˉ d y ˉ ∘ d w ˉ ) = R (( z ˉ d y ˉ ) d w ˉ ) = φ ( z ˉ d y ˉ ) . (102)
Binary: Let φ = φ 1 ⋆ φ 2 \varphi= \varphi_{1} \star\varphi_{2} φ = φ 1 ⋆ φ 2 . Let c i c_{i} c i be the context injection of x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) relative to φ i \varphi_{i} φ i . Then
( φ ( y ˉ ) ) ( z ˉ ) = ( φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 ) ) ( z ˉ ) = ( φ 1 ( y ˉ c 1 ) ) ( z ˉ c ^ 1 ) ⋆ ( φ 2 ( y ˉ c 2 ) ) ( z ˉ c ^ 2 ) (103) (\varphi(\bar{y}))(\bar{z}) = (\varphi_{1}(\bar{y}^{c_{1}}) \star\varphi_{2}(\bar{y}^{c_{2}}))(\bar{z}) = (\varphi_{1}(\bar{y}^{c_{1}}))(\bar{z}^{\hat{c}_{1}}) \star(\varphi_{2}(\bar{y}^{c_{2}}))(\bar{z}^{\hat{c}_{2}}) \tag*{(103)} ( φ ( y ˉ )) ( z ˉ ) = ( φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 )) ( z ˉ ) = ( φ 1 ( y ˉ c 1 )) ( z ˉ c ^ 1 ) ⋆ ( φ 2 ( y ˉ c 2 )) ( z ˉ c ^ 2 ) (103)
= φ 1 ( ( z ˉ c ^ 1 ) d y ˉ c 1 ) ⋆ φ 2 ( ( z ˉ c ^ 2 ) d y ˉ c 2 ) = φ 1 ( ( z ˉ d y ˉ ) c 1 ) ⋆ φ 2 ( ( z ˉ d y ˉ ) c 2 ) (104) = \varphi_{1}((\bar{z}^{\hat{c}_{1}})^{d_{\bar{y}^{c_{1}}}}) \star\varphi_{2}((\bar{z}^{\hat{c}_{2}})^{d_{\bar{y}^{c_{2}}}}) = \varphi_{1}((\bar{z}^{d_{\bar{y}}})^{c_{1}}) \star\varphi_{2}((\bar{z}^{d_{\bar{y}}})^{c_{2}}) \tag*{(104)} = φ 1 (( z ˉ c ^ 1 ) d y ˉ c 1 ) ⋆ φ 2 (( z ˉ c ^ 2 ) d y ˉ c 2 ) = φ 1 (( z ˉ d y ˉ ) c 1 ) ⋆ φ 2 (( z ˉ d y ˉ ) c 2 ) (104)
= φ ( z ˉ d y ˉ ) , (105) = \varphi(\bar{z}^{d_{\bar{y}}}), \tag*{(105)} = φ ( z ˉ d y ˉ ) , (105)
where c ^ i \hat{c}_{i} c ^ i is the context injection of y ˉ ∘ \bar{y}^{\circ} y ˉ ∘ relative to φ i ( y ˉ c i ) \varphi_{i}(\bar{y}^{c_{i}}) φ i ( y ˉ c i ) . The first equation uses the binary clause for substitution for y ˉ \bar{y} y ˉ , the second the same clause for z ˉ \bar{z} z ˉ , the third applies the induction hypothesis, the fourth applies Lemma A.0.5 , and the fifth again uses the definition of substitution for z ˉ d y ˉ \bar{z}^{d_{\bar{y}}} z ˉ d y ˉ .
Quantifier: Let φ = ∀ τ u ψ \varphi= \forall_{\tau}u\,\psi φ = ∀ τ u ψ and let c c c be the context injection of x ˉ ( φ ) × u \bar{x}(\varphi) \times u x ˉ ( φ ) × u relative to ψ \psi ψ . Choose v : τ v : \tau v : τ , fresh for y ˉ \bar{y} y ˉ and ψ \psi ψ , so that φ ( y ˉ ) = ∀ v ψ ( ( y ˉ × v ) c ) \varphi(\bar{y}) = \forall_{v}\,\psi((\bar{y} \times v)^{c}) φ ( y ˉ ) = ∀ v ψ (( y ˉ × v ) c ) by the quantifier clause for substitution, and x ˉ ( ψ ( ( y ˉ × v ) c ) ) = ( ( y ˉ × v ) c ) ∘ \bar{x}(\psi((\bar{y} \times v)^{c})) = ((\bar{y} \times v)^{c})^{\circ} x ˉ ( ψ (( y ˉ × v ) c )) = (( y ˉ × v ) c ) ∘ by Lemma 1.1.1. Let c ^ \hat{c} c ^ be the context injection of y ˉ ∘ × v \bar{y}^{\circ} \times v y ˉ ∘ × v relative to ψ ( ( y ˉ × v ) c ) \psi((\bar{y} \times v)^{c}) ψ (( y ˉ × v ) c ) . Now choose v ′ : τ v' : \tau v ′ : τ , fresh for z ˉ \bar{z} z ˉ , ψ \psi ψ and ψ ( ( y ˉ × v ) c ) \psi((\bar{y} \times v)^{c}) ψ (( y ˉ × v ) c ) . Then, using Lemma A.0.5 for the second and the fact that v v v is fresh for y ˉ \bar{y} y ˉ for the third equality, we get
( ( z ˉ × v ′ ) c ^ ) d ( y ˉ × v ) c = ( z ˉ × v ′ ) c ^ ∘ d ( y ˉ × v ) c = ( z ˉ × v ′ ) d y ˉ ∘ × v ∘ c = ( z ˉ d y ˉ × v ′ ) c . (106) ((\bar{z} \times v')^{\hat{c}})^{d_{(\bar{y} \times v)^{c}}} = (\bar{z} \times v')^{\hat{c} \circ d_{(\bar{y} \times v)^{c}}} = (\bar{z} \times v')^{d_{\bar{y}^{\circ} \times v} \circ c} = (\bar{z}^{d_{\bar{y}}} \times v')^{c}. \tag*{(106)} (( z ˉ × v ′ ) c ^ ) d ( y ˉ × v ) c = ( z ˉ × v ′ ) c ^ ∘ d ( y ˉ × v ) c = ( z ˉ × v ′ ) d y ˉ ∘ × v ∘ c = ( z ˉ d y ˉ × v ′ ) c . (106)
Hence, up to renaming of bound variables,
( φ ( y ˉ ) ) ( z ˉ ) = ∀ v ′ ( ψ ( ( y ˉ × v ) c ) ) ( ( z ˉ × v ′ ) c ^ ) = ∀ v ′ ψ ( ( ( z ˉ × v ′ ) c ^ ) d ( y ˉ × v ) c ) (107) \left(\varphi(\bar{y})\right)(\bar{z}) = \forall v' \left(\psi\left((\bar{y} \times v)^{c}\right)\right)\left((\bar{z} \times v')^{\hat{c}}\right) = \forall v' \psi\left(\left((\bar{z} \times v')^{\hat{c}}\right)^{d_{(\bar{y}\times v)^{c}}}\right) \tag*{(107)} ( φ ( y ˉ ) ) ( z ˉ ) = ∀ v ′ ( ψ ( ( y ˉ × v ) c ) ) ( ( z ˉ × v ′ ) c ^ ) = ∀ v ′ ψ ( ( ( z ˉ × v ′ ) c ^ ) d ( y ˉ × v ) c ) (107)
= ∀ v ′ ψ ( ( z ˉ d y ˉ × v ′ ) c ) = φ ( z ˉ d y ˉ ) . (108) = \forall v' \psi\left((\bar{z}^{d_{\bar{y}}} \times v')^{c}\right) = \varphi(\bar{z}^{d_{\bar{y}}}). \tag*{(108)} = ∀ v ′ ψ ( ( z ˉ d y ˉ × v ′ ) c ) = φ ( z ˉ d y ˉ ) . (108)
The first and last equalities apply the quantifier clause of the definition of substitution to ∀ v ψ ( ( y ˉ × v ) c ) \forall v \psi((\bar{y} \times v)^{c}) ∀ v ψ (( y ˉ × v ) c ) with context injection c ^ \hat{c} c ^ and to ∀ v ψ \forall v \psi ∀ v ψ with context injection c c c , respectively; the second uses the induction hypothesis for ψ \psi ψ , and the third is the result of the preceding computation.
Lemma A.0.7 (Substitution and Translation). Let f : Σ 1 → Σ 2 f : \Sigma_{1} \to\Sigma_{2} f : Σ 1 → Σ 2 be a piecewise reconstrual, φ : σ ˉ \varphi: \bar{\sigma} φ : σ ˉ an L 1 \mathcal{L}_{1} L 1 -formula with canonical context x ˉ = x ˉ ( φ ) \bar{x} = \bar{x}(\varphi) x ˉ = x ˉ ( φ ) , and y ˉ : σ ˉ \bar{y} : \bar{\sigma} y ˉ : σ ˉ a sequence of variables. Let τ ˉ : = σ ˉ ( φ ( y ˉ ) ) \bar{\tau} := \bar{\sigma}(\varphi(\bar{y})) τ ˉ := σ ˉ ( φ ( y ˉ )) , so that x ˉ ( φ ( y ˉ ) ) = y ˉ ∘ : τ ˉ \bar{x}(\varphi(\bar{y})) = \bar{y}^{\circ} : \bar{\tau} x ˉ ( φ ( y ˉ )) = y ˉ ∘ : τ ˉ by Lemma 1.1.1. Then, for any piece k ˉ ∈ P τ ˉ f \bar{k} \in P_{\bar{\tau}}^{f} k ˉ ∈ P τ ˉ f ,
φ ( y ˉ ) f k ˉ = φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) , (109) \varphi(\bar{y})_{\!f}^{\bar{k}} = \varphi_{\!f}^{\bar{k}^{d_{\bar{y}}}}\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right), \tag*{(109)} φ ( y ˉ ) f k ˉ = φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) , (109) up to renaming of bound variables, where ( k ˉ d y ˉ ) i = k d y ˉ ( i ) (\bar{k}^{d_{\bar{y}}})_{i} = k_{d_{\bar{y}}(i)} ( k ˉ d y ˉ ) i = k d y ˉ ( i ) .
Proof. By induction on φ \varphi φ ; the truth-value and negation cases are immediate.
Atomic: Let φ = R ( z ˉ ) : σ ˉ \varphi= R(\bar{z}) : \bar{\sigma} φ = R ( z ˉ ) : σ ˉ be atomic and let y ˉ : σ ˉ \bar{y} : \bar{\sigma} y ˉ : σ ˉ . For a piece k ˉ ∈ P σ ˉ ( φ ( y ˉ ) ) \bar{k} \in P_{\bar{\sigma}(\varphi(\bar{y}))} k ˉ ∈ P σ ˉ ( φ ( y ˉ )) , Corollary A.0.4 implies d y ˉ d z ˉ = d y ˉ ∘ d z ˉ d_{\bar{y}d_{\bar{z}}} = d_{\bar{y}} \circ d_{\bar{z}} d y ˉ d z ˉ = d y ˉ ∘ d z ˉ . Hence
φ ( y ˉ ) f k ˉ = R ( y ˉ d z ˉ ) f k ˉ = R f k ˉ d y ˉ d z ˉ ( ( y ˉ d z ˉ ) f k ˉ d y ˉ d z ˉ ) = R f k ˉ d y ˉ d z ˉ ( ( y ˉ d z ˉ ) f k ˉ d y ˉ d z ˉ ) . (110) \varphi(\bar{y})_{\!f}^{\bar{k}} = R(\bar{y}^{d_{\bar{z}}})_{\!f}^{\bar{k}} = R_{\!f}^{\bar{k}^{d_{\bar{y}d_{\bar{z}}}}}\left((\bar{y}^{d_{\bar{z}}})_{\!f}^{\bar{k}^{d_{\bar{y}d_{\bar{z}}}}}\right) = R_{\!f}^{\bar{k}^{d_{\bar{y}}d_{\bar{z}}}}\left((\bar{y}^{d_{\bar{z}}})_{\!f}^{\bar{k}^{d_{\bar{y}d_{\bar{z}}}}}\right). \tag*{(110)} φ ( y ˉ ) f k ˉ = R ( y ˉ d z ˉ ) f k ˉ = R f k ˉ d y ˉ d z ˉ ( ( y ˉ d z ˉ ) f k ˉ d y ˉ d z ˉ ) = R f k ˉ d y ˉ d z ˉ ( ( y ˉ d z ˉ ) f k ˉ d y ˉ d z ˉ ) . (110)
On the other hand,
φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) = R ( z ˉ ) f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) = ( R f k ˉ d y ˉ d z ˉ ( z ˉ f k ˉ d y ˉ d z ˉ ) ) ( y ˉ f k ˉ d y ˉ ) = R f k ˉ d y ˉ d z ˉ ( ( y ˉ d z ˉ ) f k ˉ d y ˉ d z ˉ ) , (111) \varphi_{\!f}^{\bar{k}^{d_{\bar{y}}}}\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right) = R(\bar{z})_{\!f}^{\bar{k}^{d_{\bar{y}}}}\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right) = \left(R_{\!f}^{\bar{k}^{d_{\bar{y}}d_{\bar{z}}}}\left(\bar{z}_{\!f}^{\bar{k}^{d_{\bar{y}}d_{\bar{z}}}}\right)\right)\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right) = R_{\!f}^{\bar{k}^{d_{\bar{y}}d_{\bar{z}}}}\left((\bar{y}^{d_{\bar{z}}})_{\!f}^{\bar{k}^{d_{\bar{y}d_{\bar{z}}}}}\right), \tag*{(111)} φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) = R ( z ˉ ) f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) = ( R f k ˉ d y ˉ d z ˉ ( z ˉ f k ˉ d y ˉ d z ˉ ) ) ( y ˉ f k ˉ d y ˉ ) = R f k ˉ d y ˉ d z ˉ ( ( y ˉ d z ˉ ) f k ˉ d y ˉ d z ˉ ) , (111)
where the last equality uses Lemma A.0.6. Therefore φ ( y ˉ ) f k ˉ = φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) \varphi(\bar{y})_{\!f}^{\bar{k}} = \varphi_{\!f}^{\bar{k}^{d_{\bar{y}}}}\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right) φ ( y ˉ ) f k ˉ = φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) , up to renaming of bound variables.
Binary: Let φ = φ 1 ⋆ φ 2 \varphi= \varphi_{1} \star\varphi_{2} φ = φ 1 ⋆ φ 2 , and let c i c_{i} c i be the context injection of x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) relative to φ i \varphi_{i} φ i . For a piece k ˉ ∈ P σ ˉ ( φ ( y ˉ ) ) \bar{k} \in P_{\bar{\sigma}(\varphi(\bar{y}))} k ˉ ∈ P σ ˉ ( φ ( y ˉ )) , let c ^ i \hat{c}_{i} c ^ i be the context injection of y ˉ ∘ \bar{y}^{\circ} y ˉ ∘ relative to φ i ( y ˉ c i ) \varphi_{i}(\bar{y}^{c_{i}}) φ i ( y ˉ c i ) . By Lemma A.0.5, d y ˉ ∘ c i = c ^ i ∘ d y ˉ c i d_{\bar{y}} \circ c_{i} = \hat{c}_{i} \circ d_{\bar{y}^{c_{i}}} d y ˉ ∘ c i = c ^ i ∘ d y ˉ c i , and hence
( k ˉ c ^ i ) d y ˉ c i = ( k ˉ d y ˉ ) c i . (112) \left(\bar{k}^{\hat{c}_{i}}\right)^{d_{\bar{y}^{c_{i}}}} = \left(\bar{k}^{d_{\bar{y}}}\right)^{c_{i}}. \tag*{(112)} ( k ˉ c ^ i ) d y ˉ c i = ( k ˉ d y ˉ ) c i . (112)
Thus, by the binary clause for translations and the induction hypothesis,
φ ( y ˉ ) f k ˉ = ( φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 ) ) f k ˉ = φ 1 ( y ˉ c 1 ) f k ˉ c ^ 1 ⋆ φ 2 ( y ˉ c 2 ) f k ˉ c ^ 2 (113) \varphi(\bar{y})_{\!f}^{\bar{k}} = \left(\varphi_{1}(\bar{y}^{c_{1}}) \star\varphi_{2}(\bar{y}^{c_{2}})\right)_{\!f}^{\bar{k}} = \varphi_{1}(\bar{y}^{c_{1}})_{\!f}^{\bar{k}^{\hat{c}_{1}}} \star\varphi_{2}(\bar{y}^{c_{2}})_{\!f}^{\bar{k}^{\hat{c}_{2}}} \tag*{(113)} φ ( y ˉ ) f k ˉ = ( φ 1 ( y ˉ c 1 ) ⋆ φ 2 ( y ˉ c 2 ) ) f k ˉ = φ 1 ( y ˉ c 1 ) f k ˉ c ^ 1 ⋆ φ 2 ( y ˉ c 2 ) f k ˉ c ^ 2 (113)
= φ 1 , f ( k ˉ c 1 ) d y ˉ c 1 ( ( y ˉ c 1 ) f ( k ˉ c 1 ) d y ˉ c 1 ) ⋆ φ 2 , f ( k ˉ c 2 ) d y ˉ c 2 ( ( y ˉ c 2 ) f ( k ˉ c 2 ) d y ˉ c 2 ) (114) = \varphi_{1,f}^{(\bar{k}^{c_{1}})^{d_{\bar{y}^{c_{1}}}}}\left((\bar{y}^{c_{1}})_{\!f}^{(\bar{k}^{c_{1}})^{d_{\bar{y}^{c_{1}}}}}\right) \star\varphi_{2,f}^{(\bar{k}^{c_{2}})^{d_{\bar{y}^{c_{2}}}}}\left((\bar{y}^{c_{2}})_{\!f}^{(\bar{k}^{c_{2}})^{d_{\bar{y}^{c_{2}}}}}\right) \tag*{(114)} = φ 1 , f ( k ˉ c 1 ) d y ˉ c 1 ( ( y ˉ c 1 ) f ( k ˉ c 1 ) d y ˉ c 1 ) ⋆ φ 2 , f ( k ˉ c 2 ) d y ˉ c 2 ( ( y ˉ c 2 ) f ( k ˉ c 2 ) d y ˉ c 2 ) (114)
= φ 1 , f ( k ˉ d y ˉ ) c 1 ( ( y ˉ c 1 ) f ( k ˉ d y ˉ ) c 1 ) ⋆ φ 2 , f ( k ˉ d y ˉ ) c 2 ( ( y ˉ c 2 ) f ( k ˉ d y ˉ ) c 2 ) (115) = \varphi_{1,f}^{(\bar{k}^{d_{\bar{y}}})^{c_{1}}}\left((\bar{y}^{c_{1}})_{\!f}^{(\bar{k}^{d_{\bar{y}}})^{c_{1}}}\right) \star\varphi_{2,f}^{(\bar{k}^{d_{\bar{y}}})^{c_{2}}}\left((\bar{y}^{c_{2}})_{\!f}^{(\bar{k}^{d_{\bar{y}}})^{c_{2}}}\right) \tag*{(115)} = φ 1 , f ( k ˉ d y ˉ ) c 1 ( ( y ˉ c 1 ) f ( k ˉ d y ˉ ) c 1 ) ⋆ φ 2 , f ( k ˉ d y ˉ ) c 2 ( ( y ˉ c 2 ) f ( k ˉ d y ˉ ) c 2 ) (115)
= ( φ 1 , f k ˉ d y ˉ c 1 ⋆ φ 2 , f k ˉ d y ˉ c 2 ) ( y ˉ f k ˉ d y ˉ ) = ( ( x ˉ ( φ ) ∣ φ 1 ) f k ˉ d y ˉ ⋆ ( x ˉ ( φ ) ∣ φ 2 ) f k ˉ d y ˉ ) ( y ˉ f k ˉ d y ˉ ) (116) = \left(\varphi_{1,f}^{\bar{k}^{d_{\bar{y}}c_{1}}} \star\varphi_{2,f}^{\bar{k}^{d_{\bar{y}}c_{2}}}\right)\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right) = \left((\bar{x}(\varphi) \mid\varphi_{1})_{\!f}^{\bar{k}^{d_{\bar{y}}}} \star(\bar{x}(\varphi) \mid\varphi_{2})_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right)\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right) \tag*{(116)} = ( φ 1 , f k ˉ d y ˉ c 1 ⋆ φ 2 , f k ˉ d y ˉ c 2 ) ( y ˉ f k ˉ d y ˉ ) = ( ( x ˉ ( φ ) ∣ φ 1 ) f k ˉ d y ˉ ⋆ ( x ˉ ( φ ) ∣ φ 2 ) f k ˉ d y ˉ ) ( y ˉ f k ˉ d y ˉ ) (116)
= φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) , (117) = \varphi_{\!f}^{\bar{k}^{d_{\bar{y}}}}\left(\bar{y}_{\!f}^{\bar{k}^{d_{\bar{y}}}}\right), \tag*{(117)} = φ f k ˉ d y ˉ ( y ˉ f k ˉ d y ˉ ) , (117)
up to renaming of bound variables. Here, the first and second equalities use the binary clauses of substitution and translation, the third the induction hypothesis, and the fourth Lemma A.0.5. The fifth uses the binary substitution clause in reverse, together with Lemma 2.1.6, while the sixth uses the context clause, with the final equality using the binary clause for translations again.
Quantifier: Let φ = ∀ τ z ψ \varphi= \forall_{\tau}z\ \psi φ = ∀ τ z ψ , and let c c c be the context injection of x ˉ ( φ ) × z \bar{x}(\varphi) \times z x ˉ ( φ ) × z relative to ψ \psi ψ . Let y ˉ : σ ˉ \bar{y}:\bar{\sigma} y ˉ : σ ˉ and l ˉ ∈ P σ ˉ ( φ ( y ˉ ) ) \bar{l} \in P_{\bar{\sigma}}(\varphi(\bar{y})) l ˉ ∈ P σ ˉ ( φ ( y ˉ )) . Choose v : τ v:\tau v : τ , fresh for y ˉ \bar{y} y ˉ , ψ \psi ψ , and such that, for every h ∈ P τ f h \in P_{\tau}^{f} h ∈ P τ f , v f h v_{f}^{h} v f h is fresh for φ f l ˉ d y \varphi_{f}^{\bar{l}d_{y}} φ f l ˉ d y . Then φ ( y ˉ ) = ∀ v ψ ( ( y ˉ × v ) c ) \varphi(\bar{y}) = \forall v\ \psi((\bar{y}\times v)^{c}) φ ( y ˉ ) = ∀ v ψ (( y ˉ × v ) c ) , and let c ^ \hat{c} c ^ be the context injection of y ˉ ∘ × v \bar{y}^{\circ}\times v y ˉ ∘ × v relative to ψ ( ( y ˉ × v ) c ) \psi((\bar{y}\times v)^{c}) ψ (( y ˉ × v ) c ) . Lemma A.0.5 then implies d y ˉ × v ∘ c = c ^ ∘ d ( y ˉ × v ) c d_{\bar{y}\times v}\circ c = \hat{c}\circ d_{(\bar{y}\times v)^{c}} d y ˉ × v ∘ c = c ^ ∘ d ( y ˉ × v ) c . Given a piece h ∈ P τ f h \in P_{\tau}^{f} h ∈ P τ f , this implies, together with freshness of v v v for y ˉ \bar{y} y ˉ ,
( ( l ˉ × h ) c ^ ) d ( y ˉ × v ) c = ( l ˉ × h ) c ^ ∘ d ( y ˉ × v ) c = ( l ˉ × h ) d y ˉ × v ∘ c = ( l ˉ d y ˉ × h ) c . (118) ((\bar{l}\times h)^{\hat{c}})^{d_{(\bar{y}\times v)^{c}}}=(\bar{l}\times h)^{\hat{c}\circ d_{(\bar{y}\times v)^{c}}}=(\bar{l}\times h)^{d_{\bar{y}\times v}\circ c}=(\bar{l}^{d_{\bar{y}}\times h})^{c}. \tag*{(118)} (( l ˉ × h ) c ^ ) d ( y ˉ × v ) c = ( l ˉ × h ) c ^ ∘ d ( y ˉ × v ) c = ( l ˉ × h ) d y ˉ × v ∘ c = ( l ˉ d y ˉ × h ) c . (118)
For the left-hand side of the claim: By the quantifier clause for translations, the context clause, and the induction hypothesis for ψ \psi ψ ,
φ ( y ˉ ) f l ˉ = ⋀ h ∈ P τ f ∀ v f h ( D τ h ( v f h ) → ψ f ( ( y ˉ × v ) c ) ( l ˉ × h ) c ^ ) (119) \varphi(\bar{y})_{f}^{\bar{l}}=\bigwedge_{h\in P_{\tau}^{f}}\forall v_{f}^{h}\left(D_{\tau}^{h}(v_{f}^{h})\to\psi_{f}^{((\bar{y}\times v)^{c})^{(\bar{l}\times h)^{\hat{c}}}}\right) \tag*{(119)} φ ( y ˉ ) f l ˉ = h ∈ P τ f ⋀ ∀ v f h ( D τ h ( v f h ) → ψ f (( y ˉ × v ) c ) ( l ˉ × h ) c ^ ) (119)
= ⋀ h ∈ P τ f ∀ v f h ( D τ h ( v f h ) → ψ f l ˉ d y ˉ × h ( ( ( y ˉ × v ) c ) f ( l ˉ d y ˉ × h ) c ) ) , (120) =\bigwedge_{h\in P_{\tau}^{f}}\forall v_{f}^{h}\left(D_{\tau}^{h}(v_{f}^{h})\to\psi_{f}^{\bar{l}^{d_{\bar{y}}}\times h}\left(((\bar{y}\times v)^{c})_{f}^{(\bar{l}^{d_{\bar{y}}\times h)^{c}}}\right)\right), \tag*{(120)} = h ∈ P τ f ⋀ ∀ v f h ( D τ h ( v f h ) → ψ f l ˉ d y ˉ × h ( (( y ˉ × v ) c ) f ( l ˉ d y ˉ × h ) c ) ) , (120)
up to renaming of bound variables.
For the right-hand side: Applying the quantifier clause for translations and the context clause for the first equality,
\varphi_{f}^{\bar{l}^{d_{\bar{y}}}}(\bar{y}_{f}^{\bar{l}^{d_{\bar{y}}}})=\left(\bigwedge_{h\in P_{\tau}^{f}}\forall z_{f}^{h}\left(D_{\tau}^{h}(z_{f}^{h})\to\psi_{f}^{(\bar{l}^{d_{\bar{y}}\times h)^{c}}\left(((\bar{y}\times v)^{c})_{f}^{(\bar{l}^{d_{\bar{y}}\times h)^{c}}}\right)\right)\right)(\bar{y}_{f}^{\bar{l}^{d_{\bar{y}}}})} \tag*{(121)}
=\bigwedge_{h\in P_{\tau}^{f}}\forall v_{f}^{h}\left(D_{\tau}^{h}(v_{f}^{h})\to\psi_{f}^{(\bar{l}^{d_{\bar{y}}\times h)^{c}}\left(((\bar{y}\times v)^{c})_{f}^{(\bar{l}^{d_{\bar{y}}\times h)^{c}}}\right)\right),} \tag*{(122)}
up to renaming of bound variables, where the second equality first distributes the substitution over the conjunction, and then, using the freshness condition on v v v , over the implication, applying Lemma A.0.6. □
Proposition 3.1.2 (Composition for Complex Formulae). Let f : L 1 → L 2 f:\mathcal{L}_{1}\to\mathcal{L}_{2} f : L 1 → L 2 and g : L 2 → L 3 g:\mathcal{L}_{2}\to\mathcal{L}_{3} g : L 2 → L 3 be piecewise translations , and let φ : σ ˉ \varphi:\bar{\sigma} φ : σ ˉ be an L 1 \mathcal{L}_{1} L 1 -formula with ∣ σ ˉ ∣ = n |\bar{\sigma}|=n ∣ σ ˉ ∣ = n . For any composite piece m ˉ = ( ( k 1 , l ˉ 1 ) , … , ( k n , l ˉ n ) ) ∈ P σ ˉ g ∘ f \bar{m}=((k_{1},\bar{l}_{1}),\ldots,(k_{n},\bar{l}_{n}))\in P_{\bar{\sigma}}^{g\circ f} m ˉ = (( k 1 , l ˉ 1 ) , … , ( k n , l ˉ n )) ∈ P σ ˉ g ∘ f , we set
m ˉ f : = ( k 1 , … , k n ) ∈ P σ ˉ f , m ˉ g : = l ˉ 1 × ⋯ × l ˉ n ∈ P σ ˉ f m ˉ f g . (123) \bar{m}_{f}:=(k_{1},\ldots,k_{n})\in P_{\bar{\sigma}}^{f},\qquad\bar{m}_{g}:=\bar{l}_{1}\times\cdots\times\bar{l}_{n}\in P_{\bar{\sigma}_{f}^{\bar{m}_{f}}}^{g}. \tag*{(123)} m ˉ f := ( k 1 , … , k n ) ∈ P σ ˉ f , m ˉ g := l ˉ 1 × ⋯ × l ˉ n ∈ P σ ˉ f m ˉ f g . (123) Then the formulae φ g ∘ f m ˉ \varphi_{g\circ f}^{\bar{m}} φ g ∘ f m ˉ and ( φ f m ˉ f ) g m ˉ g (\varphi_{f}^{\bar{m}_{f}})_{g}^{\bar{m}_{g}} ( φ f m ˉ f ) g m ˉ g have the same canonical context, and
⊢ L ( Σ 3 ) φ g ∘ f m ˉ ↔ ( φ f m ˉ f ) g m ˉ g . (124) \vdash_{\mathcal{L}(\Sigma_{3})}\varphi_{g\circ f}^{\bar{m}}\leftrightarrow(\varphi_{f}^{\bar{m}_{f}})_{g}^{\bar{m}_{g}}. \tag*{(124)} ⊢ L ( Σ 3 ) φ g ∘ f m ˉ ↔ ( φ f m ˉ f ) g m ˉ g . (124) Proof. We work in the theory of identity L ( Σ 3 ) L(\Sigma_{3}) L ( Σ 3 ) in the third signature throughout, abbreviating ⊢ L ( Σ 3 ) \vdash_{L(\Sigma_{3})} ⊢ L ( Σ 3 ) to ⊢ \vdash ⊢ in this proof. The claim about canonical contexts follows from
the definition of the variable function of the composition and Lemma 2.1.6 with the computation
x ˉ ( φ ˉ m ˉ g ∘ f ) = x ˉ ( φ ) m ˉ g ∘ f = ( x ˉ ( φ ) m ˉ f ) m ˉ g = x ˉ ( ( φ m ˉ f ) m ˉ g ) . (125) \bar{x}(\bar{\varphi}^{\bar{m}_{g\circ f}})=\bar{x}(\varphi)^{\bar{m}_{g\circ f}}=\left(\bar{x}(\varphi)^{\bar{m}_f}\right)^{\bar{m}_g}=\bar{x}\left((\varphi^{\bar{m}_f})^{\bar{m}_g}\right). \tag*{(125)} x ˉ ( φ ˉ m ˉ g ∘ f ) = x ˉ ( φ ) m ˉ g ∘ f = ( x ˉ ( φ ) m ˉ f ) m ˉ g = x ˉ ( ( φ m ˉ f ) m ˉ g ) . (125)
The remainder proceeds by induction on φ \varphi φ . Write m ˉ f = ( k 1 , … , k n ) \bar{m}_f=(k_1,\ldots,k_n) m ˉ f = ( k 1 , … , k n ) and m ˉ g = l ˉ 1 × ⋯ × l ˉ n \bar{m}_g=\bar{l}_1\times\cdots\times\bar{l}_n m ˉ g = l ˉ 1 × ⋯ × l ˉ n as in the statement. The truth-value and negation cases are immediate.
Atomic: Let φ = R ( z ˉ ) \varphi=R(\bar{z}) φ = R ( z ˉ ) be atomic, with canonical context x ˉ = z ˉ ∘ \bar{x}=\bar{z}^{\circ} x ˉ = z ˉ ∘ . Set
v ˉ : = z ˉ m ˉ f d z , and q ˉ : = l ˉ d z ( 1 ) × ⋯ × l ˉ d z ( ∣ z ˉ ∣ ) . (126) \bar{v}:=\bar{z}^{\bar{m}^{d_z}_f}, \qquad\text{and} \qquad\bar{q}:=\bar{l}_{d_z(1)}\times\cdots\times\bar{l}_{d_z(|\bar{z}|)}. \tag*{(126)} v ˉ := z ˉ m ˉ f d z , and q ˉ := l ˉ d z ( 1 ) × ⋯ × l ˉ d z ( ∣ z ˉ ∣ ) . (126)
Then ( m ˉ d z ) f = m ˉ f d z (\bar{m}^{d_z})_f=\bar{m}^{d_z}_f ( m ˉ d z ) f = m ˉ f d z and ( m ˉ d z ) g = q ˉ (\bar{m}^{d_z})_g=\bar{q} ( m ˉ d z ) g = q ˉ . Moreover, v ˉ ∘ = x ˉ m ˉ f \bar{v}^{\circ}=\bar{x}^{\bar{m}_f} v ˉ ∘ = x ˉ m ˉ f and m ˉ g v ˉ ∘ = q ˉ \bar{m}^{\bar{v}^{\circ}}_g=\bar{q} m ˉ g v ˉ ∘ = q ˉ , by the no-overlap and duplicate-freeness conditions on the variable function. Since R m ˉ g ∘ f = ( R m ˉ f ) m ˉ g R^{\bar{m}_{g\circ f}}=(R^{\bar{m}_f})^{\bar{m}_g} R m ˉ g ∘ f = ( R m ˉ f ) m ˉ g ,
R ( z ˉ ) m ˉ g ∘ f = R m ˉ g ∘ f d z ( z ˉ m ˉ g ∘ f d z ) = ( R m ˉ f d z ) q ˉ ( z ˉ m ˉ g ∘ f d z ) . (127) R(\bar{z})^{\bar{m}_{g\circ f}}=R^{\bar{m}^{d_z}_{g\circ f}}\left(\bar{z}^{\bar{m}^{d_z}_{g\circ f}}\right)=\left(R^{\bar{m}^{d_z}_f}\right)^{\bar{q}}\left(\bar{z}^{\bar{m}^{d_z}_{g\circ f}}\right). \tag*{(127)} R ( z ˉ ) m ˉ g ∘ f = R m ˉ g ∘ f d z ( z ˉ m ˉ g ∘ f d z ) = ( R m ˉ f d z ) q ˉ ( z ˉ m ˉ g ∘ f d z ) . (127)
On the other hand, by Lemma A.0.7 for g g g and the definition of the composite variable function,
( R ( z ˉ ) m ˉ f ) m ˉ g = ( R m ˉ f d z ( v ˉ ) ) m ˉ g = ( R m ˉ f d z ) q ˉ ( v ˉ q ˉ ) (128) \left(R(\bar{z})^{\bar{m}_f}\right)^{\bar{m}_g}=\left(R^{\bar{m}^{d_z}_f}(\bar{v})\right)^{\bar{m}_g}=\left(R^{\bar{m}^{d_z}_f}\right)^{\bar{q}}(\bar{v}^{\bar{q}}) \tag*{(128)} ( R ( z ˉ ) m ˉ f ) m ˉ g = ( R m ˉ f d z ( v ˉ ) ) m ˉ g = ( R m ˉ f d z ) q ˉ ( v ˉ q ˉ ) (128)
= ( R m ˉ f d z ) q ˉ ( z ˉ m ˉ g ∘ f d z ) . (129) =\left(R^{\bar{m}^{d_z}_f}\right)^{\bar{q}}\left(\bar{z}^{\bar{m}^{d_z}_{g\circ f}}\right). \tag*{(129)} = ( R m ˉ f d z ) q ˉ ( z ˉ m ˉ g ∘ f d z ) . (129)
The two right-hand sides coincide, so R ( z ˉ ) m ˉ g ∘ f = ( R ( z ˉ ) m ˉ f ) m ˉ g R(\bar{z})^{\bar{m}_{g\circ f}}=\left(R(\bar{z})^{\bar{m}_f}\right)^{\bar{m}_g} R ( z ˉ ) m ˉ g ∘ f = ( R ( z ˉ ) m ˉ f ) m ˉ g .
Binary: Let φ = φ 1 ⋆ φ 2 \varphi=\varphi_1\star\varphi_2 φ = φ 1 ⋆ φ 2 , and let c i c_i c i be the context injection of x ˉ ( φ ) \bar{x}(\varphi) x ˉ ( φ ) relative to φ i \varphi_i φ i . Set
q ˉ i : = l ˉ c i ( 1 ) × ⋯ × l ˉ c i ( ∣ x ˉ ( φ i ) ∣ ) ( i = 1 , 2 ) . (130) \bar{q}_i:=\bar{l}_{c_i(1)}\times\cdots\times\bar{l}_{c_i(|\bar{x}(\varphi_i)|)} \qquad(i=1,2). \tag*{(130)} q ˉ i := l ˉ c i ( 1 ) × ⋯ × l ˉ c i ( ∣ x ˉ ( φ i ) ∣ ) ( i = 1 , 2 ) . (130)
Then ( m ˉ c i ) f = m ˉ f c i (\bar{m}^{c_i})_f=\bar{m}^{c_i}_f ( m ˉ c i ) f = m ˉ f c i and ( m ˉ c i ) g = q ˉ i (\bar{m}^{c_i})_g=\bar{q}_i ( m ˉ c i ) g = q ˉ i . Using the induction hypothesis with these pieces,
φ m ˉ g ∘ f = ( x ˉ ( φ ) ∣ φ 1 ) m ˉ g ∘ f ⋆ ( x ˉ ( φ ) ∣ φ 2 ) m ˉ g ∘ f = ( φ 1 ) m ˉ g ∘ f c 1 ⋆ ( φ 2 ) m ˉ g ∘ f c 2 (131) \varphi^{\bar{m}_{g\circ f}}=(\bar{x}(\varphi)\mid\varphi_1)^{\bar{m}_{g\circ f}}\star(\bar{x}(\varphi)\mid\varphi_2)^{\bar{m}_{g\circ f}}=(\varphi_1)^{\bar{m}^{c_1}_{g\circ f}}\star(\varphi_2)^{\bar{m}^{c_2}_{g\circ f}} \tag*{(131)} φ m ˉ g ∘ f = ( x ˉ ( φ ) ∣ φ 1 ) m ˉ g ∘ f ⋆ ( x ˉ ( φ ) ∣ φ 2 ) m ˉ g ∘ f = ( φ 1 ) m ˉ g ∘ f c 1 ⋆ ( φ 2 ) m ˉ g ∘ f c 2 (131)
⊣ ⊢ ( ( φ 1 ) m ˉ f c 1 ) m ˉ g ⋆ ( ( φ 2 ) m ˉ f c 2 ) m ˉ g . (132) \mathrel{\dashv\vdash}\left((\varphi_1)^{\bar{m}^{c_1}_f}\right)^{\bar{m}_g}\star\left((\varphi_2)^{\bar{m}^{c_2}_f}\right)^{\bar{m}_g}. \tag*{(132)} ⊣⊢ ( ( φ 1 ) m ˉ f c 1 ) m ˉ g ⋆ ( ( φ 2 ) m ˉ f c 2 ) m ˉ g . (132)
Now, for the right-hand side, let c ~ i \tilde{c}_i c ~ i be the context injection of x ˉ ( φ ) m ˉ f \bar{x}(\varphi)^{\bar{m}_f} x ˉ ( φ ) m ˉ f relative to ( φ i ) m ˉ f c i (\varphi_i)^{\bar{m}^{c_i}_f} ( φ i ) m ˉ f c i . By Lemma 2.1.6, m ˉ g c ~ i = q ˉ i \bar{m}^{\tilde{c}_i}_g=\bar{q}_i m ˉ g c ~ i = q ˉ i . Then, by the context clause for f f f , the binary and context translation clauses for g g g , and Lemma 2.1.6 to identify the translated contexts,
( φ m ˉ f ) m ˉ g = ( ( x ˉ ( φ ) ∣ φ 1 ) m ˉ f ⋆ ( x ˉ ( φ ) ∣ φ 2 ) m ˉ f ) m ˉ g = ( ( φ 1 ) m ˉ f c 1 ⋆ ( φ 2 ) m ˉ f c 2 ) m ˉ g (133) \left(\varphi^{\bar{m}_f}\right)^{\bar{m}_g}=\left((\bar{x}(\varphi)\mid\varphi_1)^{\bar{m}_f}\star(\bar{x}(\varphi)\mid\varphi_2)^{\bar{m}_f}\right)^{\bar{m}_g}=\left((\varphi_1)^{\bar{m}^{c_1}_f}\star(\varphi_2)^{\bar{m}^{c_2}_f}\right)^{\bar{m}_g} \tag*{(133)} ( φ m ˉ f ) m ˉ g = ( ( x ˉ ( φ ) ∣ φ 1 ) m ˉ f ⋆ ( x ˉ ( φ ) ∣ φ 2 ) m ˉ f ) m ˉ g = ( ( φ 1 ) m ˉ f c 1 ⋆ ( φ 2 ) m ˉ f c 2 ) m ˉ g (133)
= ( ( φ 1 ) m ˉ f c 1 ) m ˉ g c ~ 1 ⋆ ( ( φ 2 ) m ˉ f c 2 ) m ˉ g c ~ 2 = ( ( φ 1 ) m ˉ f c 1 ) q ˉ 1 ⋆ ( ( φ 2 ) m ˉ f c 2 ) q ˉ 2 . (134) =\left((\varphi_1)^{\bar{m}^{c_1}_f}\right)^{\bar{m}^{\tilde{c}_1}_g}\star\left((\varphi_2)^{\bar{m}^{c_2}_f}\right)^{\bar{m}^{\tilde{c}_2}_g}=\left((\varphi_1)^{\bar{m}^{c_1}_f}\right)^{\bar{q}_1}\star\left((\varphi_2)^{\bar{m}^{c_2}_f}\right)^{\bar{q}_2}. \tag*{(134)} = ( ( φ 1 ) m ˉ f c 1 ) m ˉ g c ~ 1 ⋆ ( ( φ 2 ) m ˉ f c 2 ) m ˉ g c ~ 2 = ( ( φ 1 ) m ˉ f c 1 ) q ˉ 1 ⋆ ( ( φ 2 ) m ˉ f c 2 ) q ˉ 2 . (134)
The two right-hand sides coincide, so φ m ˉ g ∘ f ⊣ ⊢ ( φ m ˉ f ) m ˉ g \varphi^{\bar{m}_{g\circ f}}\dashv\vdash\left(\varphi^{\bar{m}_f}\right)^{\bar{m}_g} φ m ˉ g ∘ f ⊣⊢ ( φ m ˉ f ) m ˉ g .
Quantifier: Let φ = ∀ τ u ψ \varphi= \forall_{\tau}u\ \psi φ = ∀ τ u ψ and let x ˉ = x ˉ ( φ ) \bar{x}=\bar{x}(\varphi) x ˉ = x ˉ ( φ ) . By the translation clause for the universal quantifier for f f f ,
φ f m ˉ f = ⋀ h ∈ P τ f ∀ u f h ( D τ f , h ( u f h ) → ( x ˉ × u ∣ ψ ) f m ˉ f × h ) . (135) \varphi_{f}^{\bar{m}_{f}}=\bigwedge_{h\in P_{\tau}^{f}}\forall u_{f}^{h}\left(D_{\tau}^{f,h}(u_{f}^{h})\to\left(\bar{x}\times u\mid\psi\right)_{f}^{\bar{m}_{f}\times h}\right). \tag*{(135)} φ f m ˉ f = h ∈ P τ f ⋀ ∀ u f h ( D τ f , h ( u f h ) → ( x ˉ × u ∣ ψ ) f m ˉ f × h ) . (135)
Applying g g g at piece m ˉ g \bar{m}_{g} m ˉ g and renaming the bound sequence u f h u_{f}^{h} u f h as needed, the conjunct belonging to piece h h h becomes
⋀ h ˉ ′ ∈ P τ g , h ∀ ( u f h ) g h ˉ ′ ( D τ g , h ˉ ′ ( ( u f h ) g h ˉ ′ ) → ( ( x ˉ f m ˉ f × u f h ∣ ( x ˉ × u ∣ ψ ) f m ˉ f × h ) g m ˉ g × h ˉ ′ ) ) . (136) \bigwedge_{\bar{h}'\in P_{\tau}^{g,h}}\forall\left(u_{f}^{h}\right)_{g}^{\bar{h}'}\left(D_{\tau}^{g,\bar{h}'}\left(\left(u_{f}^{h}\right)_{g}^{\bar{h}'}\right)\to\left(\left(\bar{x}_{f}^{\bar{m}_{f}}\times u_{f}^{h}\mid\left(\bar{x}\times u\mid\psi\right)_{f}^{\bar{m}_{f}\times h}\right)_{g}^{\bar{m}_{g}\times\bar{h}'}\right)\right). \tag*{(136)} h ˉ ′ ∈ P τ g , h ⋀ ∀ ( u f h ) g h ˉ ′ ( D τ g , h ˉ ′ ( ( u f h ) g h ˉ ′ ) → ( ( x ˉ f m ˉ f × u f h ∣ ( x ˉ × u ∣ ψ ) f m ˉ f × h ) g m ˉ g × h ˉ ′ ) ) . (136)
By the binary and context clauses for g g g and Lemma A.0.7 applied to D τ f , h D_{\tau}^{f,h} D τ f , h , the formula in the scope of the quantifier is equivalent to
( D τ g , h ˉ ′ ( ( u f h ) g h ˉ ′ ) ∧ ( D τ f , h ) g h ˉ ′ ( ( u f h ) g h ˉ ′ ) ) → ( x ˉ f m ˉ f × u f h ∣ ( x ˉ × u ∣ ψ ) f m ˉ f × h ) g m ˉ g × h ˉ ′ . (137) \left(D_{\tau}^{g,\bar{h}'}\left(\left(u_{f}^{h}\right)_{g}^{\bar{h}'}\right)\wedge\left(D_{\tau}^{f,h}\right)_{g}^{\bar{h}'}\left(\left(u_{f}^{h}\right)_{g}^{\bar{h}'}\right)\right)\to\left(\bar{x}_{f}^{\bar{m}_{f}}\times u_{f}^{h}\mid\left(\bar{x}\times u\mid\psi\right)_{f}^{\bar{m}_{f}\times h}\right)_{g}^{\bar{m}_{g}\times\bar{h}'}. \tag*{(137)} ( D τ g , h ˉ ′ ( ( u f h ) g h ˉ ′ ) ∧ ( D τ f , h ) g h ˉ ′ ( ( u f h ) g h ˉ ′ ) ) → ( x ˉ f m ˉ f × u f h ∣ ( x ˉ × u ∣ ψ ) f m ˉ f × h ) g m ˉ g × h ˉ ′ . (137)
Since its antecedent is equivalent to the composite domain formula D τ g ∘ f , ( h , h ˉ ′ ) ( ( u f h ) g h ˉ ′ ) D_{\tau}^{g\circ f,(h,\bar{h}')}\left(\left(u_{f}^{h}\right)_{g}^{\bar{h}'}\right) D τ g ∘ f , ( h , h ˉ ′ ) ( ( u f h ) g h ˉ ′ ) , it remains to show that its consequent is equivalent to the consequent in the clause for the quantifier. To that end, let c c c be the context injection of x ˉ × u \bar{x}\times u x ˉ × u relative to ψ \psi ψ , and set p ˉ : = m ˉ × ( h , h ˉ ′ ) \bar{p}:=\bar{m}\times(h,\bar{h}') p ˉ := m ˉ × ( h , h ˉ ′ ) . Let c ~ \tilde{c} c ~ be the context injection of x ˉ f m ˉ f × u f h \bar{x}_{f}^{\bar{m}_{f}}\times u_{f}^{h} x ˉ f m ˉ f × u f h relative to ψ f m ˉ f \psi_{f}^{\bar{m}_{f}} ψ f m ˉ f . By Lemma 2.1.6, as in the binary case, p ˉ g c = ( p ˉ c ) g \bar{p}_{g}^{c}=(\bar{p}^{c})_{g} p ˉ g c = ( p ˉ c ) g and p ˉ f c = ( p ˉ c ) f \bar{p}_{f}^{c}=(\bar{p}^{c})_{f} p ˉ f c = ( p ˉ c ) f . Hence, by the context clauses and the induction hypothesis,
( x ˉ f m ˉ f × u f h ∣ ( x ˉ × u ∣ ψ ) f m ˉ f × h ) g p ˉ g c ~ = ( ψ f p ˉ f c ~ ) g p ˉ g = ( ψ p ˉ f c ~ ) g p ˉ g ⊣ ⊢ ψ g ∘ f p ˉ c ~ = ( x ˉ × u ∣ ψ ) g ∘ f p ˉ . (138) \left(\bar{x}_{f}^{\bar{m}_{f}}\times u_{f}^{h}\mid\left(\bar{x}\times u\mid\psi\right)_{f}^{\bar{m}_{f}\times h}\right)_{g}^{\bar{p}_{g}^{\tilde{c}}}
=\left(\psi_{f}^{\bar{p}_{f}^{\tilde{c}}}\right)_{g}^{\bar{p}_{g}}
=\left(\psi^{\bar{p}^{\tilde{c}}_{f}}\right)_{g}^{\bar{p}_{g}}\dashv\vdash\psi_{g\circ f}^{\bar{p}^{\tilde{c}}}
=\left(\bar{x}\times u\mid\psi\right)_{g\circ f}^{\bar{p}}. \tag*{(138)} ( x ˉ f m ˉ f × u f h ∣ ( x ˉ × u ∣ ψ ) f m ˉ f × h ) g p ˉ g c ~ = ( ψ f p ˉ f c ~ ) g p ˉ g = ( ψ p ˉ f c ~ ) g p ˉ g ⊣⊢ ψ g ∘ f p ˉ c ~ = ( x ˉ × u ∣ ψ ) g ∘ f p ˉ . (138)
delivering the desired consequent. Collecting over all ( h , h ˉ ′ ) ∈ P τ g ∘ f (h,\bar{h}')\in P_{\tau}^{g\circ f} ( h , h ˉ ′ ) ∈ P τ g ∘ f and using u g ∘ f ( h , h ˉ ′ ) = ( u f h ) g h ˉ ′ u_{g\circ f}^{(h,\bar{h}')}=(u_{f}^{h})_{g}^{\bar{h}'} u g ∘ f ( h , h ˉ ′ ) = ( u f h ) g h ˉ ′ , we hence obtain a formula appropriately equivalent to the result of the quantifier clause for g ∘ f g\circ f g ∘ f ; that is, ( φ f m ˉ f ) g m ˉ g ⊣ ⊢ φ g ∘ f m ˉ \left(\varphi_{f}^{\bar{m}_{f}}\right)_{g}^{\bar{m}_{g}}\dashv\vdash\varphi_{g\circ f}^{\bar{m}} ( φ f m ˉ f ) g m ˉ g ⊣⊢ φ g ∘ f m ˉ .
Reflection of Bicategorical Structure
To show that T h p w \mathrm{Th}^{\mathrm{pw}} Th pw is a bicategory, we prove and apply a completely general bicategorical theorem, which should be seen as a generalization of the 1-categorical fact that faithful functors reflect commutative diagrams. This general theorem is hinted at in Leinster [12 ] , but we were not able to find a precise statement in the literature, which is why we state and prove it here. Essentially, it says that bicategorical structure is reflected back along locally faithful (contravariant) pseudofunctors. While this result could have been omitted in favour of a more concrete approach, we think there is some independent value in having it recorded for future use in similar situations, which are likely to occur in the study of notions of interpretations.
Definition B.0.1 (Pre-bicategory). A pre-bicategory B \mathcal{B} B consists of a class of objects ob ( B ) \operatorname{ob}(\mathcal{B}) ob ( B ) ; for all X , Y ∈ ob ( B ) X,Y\in\operatorname{ob}(\mathcal{B}) X , Y ∈ ob ( B ) a class B ( X , Y ) \mathcal{B}(X,Y) B ( X , Y ) of 1-morphisms f : X → Y f:X\to Y f : X → Y , and for all f , f ′ ∈ B ( X , Y ) f,f'\in\mathcal{B}(X,Y) f , f ′ ∈ B ( X , Y ) a class B ( f , f ′ ) \mathcal{B}(f,f') B ( f , f ′ ) of 2-morphisms θ : f ⇒ f ′ : X → Y \theta:f\Rightarrow f':X\to Y θ : f ⇒ f ′ : X → Y ; for each 1-morphism f f f an identity 2-morphism id f : f ⇒ f \operatorname{id}_{f}:f\Rightarrow f id f : f ⇒ f , and for 2-morphisms ζ : f ⇒ f ′ \zeta:f\Rightarrow f' ζ : f ⇒ f ′ and θ : f ′ ⇒ f ′ ′ \theta:f'\Rightarrow f'' θ : f ′ ⇒ f ′′ a
vertical composite θ ∙ ζ : f ⇒ f ′ ′ \theta\mathbin{\bullet} \zeta: f \Rightarrow f^{\prime\prime} θ ∙ ζ : f ⇒ f ′′ ; for each X X X an identity 1 1 1 -morphism i d X : X → X \mathrm{id}_{X} : X \to X id X : X → X ; for 1 1 1 -morphisms f : X → Y f : X \to Y f : X → Y and g : Y → Z g : Y \to Z g : Y → Z a composite g ∘ f : X → Z g \circ f : X \to Z g ∘ f : X → Z ; and for 2 2 2 -morphisms ζ : f ⇒ f ′ \zeta: f \Rightarrow f^{\prime} ζ : f ⇒ f ′ between 1 1 1 -morphisms X → Y X \to Y X → Y and θ : g ⇒ g ′ \theta: g \Rightarrow g^{\prime} θ : g ⇒ g ′ between 1 1 1 -morphisms Y → Z Y \to Z Y → Z a horizontal composite θ ⋆ ζ : g ∘ f ⇒ g ′ ∘ f ′ \theta\star\zeta: g \circ f \Rightarrow g^{\prime} \circ f^{\prime} θ ⋆ ζ : g ∘ f ⇒ g ′ ∘ f ′ ; and 2 2 2 -morphisms
a h , g , f : ( h ∘ g ) ∘ f ⇒ h ∘ ( g ∘ f ) : W → Z (139) a_{h,g,f} : (h \circ g) \circ f \Rightarrow h \circ(g \circ f) : W \to Z \tag*{(139)} a h , g , f : ( h ∘ g ) ∘ f ⇒ h ∘ ( g ∘ f ) : W → Z (139)
l f : i d X ∘ f ⇒ f : W → X r f : f ∘ i d W ⇒ f : W → X (140) l^{f} : \mathrm{id}_{X} \circ f \Rightarrow f : W \to X \qquad r^{f} : f \circ\mathrm{id}_{W} \Rightarrow f : W \to X \tag*{(140)} l f : id X ∘ f ⇒ f : W → X r f : f ∘ id W ⇒ f : W → X (140)
for all f : W → X f : W \to X f : W → X , g : X → Y g : X \to Y g : X → Y and h : Y → Z h : Y \to Z h : Y → Z .
Hence a pre-bicategory has the same underlying data as a bicategory, but with no axioms imposed.
Next, we widen the definition of (contravariant) pseudofunctor to allow for pseudofunctors with pre-bicategories as domains and 2 2 2 -categories as targets. This is all the generality we need.
Definition B.0.3 (Contravariant Pseudofunctor from a Pre-Bicategory). Let B \mathcal{B} B be a pre-bicategory and C \mathcal{C} C a strict 2 2 2 -category. A contravariant pseudofunctor F : B → C F : \mathcal{B} \to\mathcal{C} F : B → C comprises a function F : ob ( B ) → ob ( C ) F : \operatorname{ob}(\mathcal{B}) \to\operatorname{ob}(\mathcal{C}) F : ob ( B ) → ob ( C ) , functions F X , Y : B ( X , Y ) → C ( F Y , F X ) F_{X,Y} : \mathcal{B}(X,Y) \to\mathcal{C}(F Y,F X) F X , Y : B ( X , Y ) → C ( F Y , F X ) and F f , g : B ( f , g ) → C ( F f , F g ) F_{f,g} : \mathcal{B}(f,g) \to\mathcal{C}(F f,F g) F f , g : B ( f , g ) → C ( F f , F g ) for all X , Y ∈ ob ( B ) X,Y \in\operatorname{ob}(\mathcal{B}) X , Y ∈ ob ( B ) and f , g ∈ B ( X , Y ) f,g \in\mathcal{B}(X,Y) f , g ∈ B ( X , Y ) , together with invertible 2 2 2 -morphisms of C \mathcal{C} C
α g , f : F ( g ∘ f ) ⇒ F f ∘ F g and α X : F ( i d X ) ⇒ i d F X , (142) \alpha_{g,f} : F(g \circ f) \Rightarrow F f \circ F g \qquad\text{and} \qquad\alpha_{X} : F(\mathrm{id}_{X}) \Rightarrow\mathrm{id}_{F X}, \tag*{(142)} α g , f : F ( g ∘ f ) ⇒ F f ∘ F g and α X : F ( id X ) ⇒ id F X , (142)
such that
F ( θ ∙ ζ ) = F θ ∙ F ζ , F ( id f ) = id F f , (143) F(\theta\mathbin{\bullet} \zeta)=F\theta\mathbin{\bullet} F\zeta,\qquad F(\operatorname{id}_{f})=\operatorname{id}_{Ff}, \tag*{(143)} F ( θ ∙ ζ ) = F θ ∙ F ζ , F ( id f ) = id F f , (143)
α g 2 , g 1 ∙ F ( θ 2 ⋆ θ 1 ) = ( F θ 1 ⋆ F θ 2 ) ∙ α f 2 , f 1 , (144) \alpha_{g_{2},g_{1}}\mathbin{\bullet}F(\theta_{2}\mathbin{\star}\theta_{1})=(F\theta_{1}\mathbin{\star}F\theta_{2})\mathbin{\bullet}\alpha_{f_{2},f_{1}}, \tag*{(144)} α g 2 , g 1 ∙ F ( θ 2 ⋆ θ 1 ) = ( F θ 1 ⋆ F θ 2 ) ∙ α f 2 , f 1 , (144)
( α g , f F h ) ∙ α h , g ∘ f ∙ F ( a h , g , f ) = ( F f ∙ α h , g ) ∙ α h ∘ g , f , (145) (\alpha_{g,f}Fh)\mathbin{\bullet}\alpha_{h,g\circ f}\mathbin{\bullet}F(a_{h,g,f})=(Ff\mathbin{\bullet}\alpha_{h,g})\mathbin{\bullet}\alpha_{h\circ g,f}, \tag*{(145)} ( α g , f F h ) ∙ α h , g ∘ f ∙ F ( a h , g , f ) = ( F f ∙ α h , g ) ∙ α h ∘ g , f , (145)
F ( l f ) = ( F f ∙ α X ) ∙ α id X , f , (146) F(l^{f})=(Ff\mathbin{\bullet}\alpha_{X})\mathbin{\bullet}\alpha_{\operatorname{id}_{X},f}, \tag*{(146)} F ( l f ) = ( F f ∙ α X ) ∙ α id X , f , (146)
F ( r f ) = ( α W ∙ F f ) ∙ α f , id W . (147) F(r^{f})=(\alpha_{W}\mathbin{\bullet}Ff)\mathbin{\bullet}\alpha_{f,\operatorname{id}_{W}}. \tag*{(147)} F ( r f ) = ( α W ∙ F f ) ∙ α f , id W . (147)
hold for all ζ : u ⇒ v \zeta:u\Rightarrow v ζ : u ⇒ v , θ : v ⇒ w \theta:v\Rightarrow w θ : v ⇒ w , θ 1 : f 1 ⇒ g 1 \theta_{1}:f_{1}\Rightarrow g_{1} θ 1 : f 1 ⇒ g 1 , θ 2 : f 2 ⇒ g 2 \theta_{2}:f_{2}\Rightarrow g_{2} θ 2 : f 2 ⇒ g 2 with f 1 , f 2 f_{1},f_{2} f 1 , f 2 and g 2 , g 1 g_{2},g_{1} g 2 , g 1 composable, and f : W → X f:W\to X f : W → X , g : X → Y g:X\to Y g : X → Y and h : Y → Z h:Y\to Z h : Y → Z . It is locally faithful if it is injective on 2-morphisms, that is, iff the following holds for all parallel 2-morphisms η , ζ : f ⇒ g \eta,\zeta:f\Rightarrow g η , ζ : f ⇒ g :
F η = F ζ ⇒ η = ζ (148) F\eta=F\zeta\Rightarrow\eta=\zeta \tag*{(148)} F η = F ζ ⇒ η = ζ (148)
It is clear that the preceding conditions ensure that any such contravariant pseudofunctor is a contravariant pseudofunctor between bicategories if B \mathcal{B} B is a bicategory, as the conditions are exactly the axioms of Definition 3.0.2 for a pseudofunctor B o p → C \mathcal{B}^{\mathrm{op}}\to\mathcal{C} B op → C , with α \alpha α replaced by its inverses, as fixed after Definition 3.0.2 . We can now precisely formulate the following useful theorem:
Theorem 3.4.3 (Reflection Theorem). Let B \mathcal{B} B be a pre-bicategory and C \mathcal{C} C be a strict 2-category. Assume that the associators a a a and the unitors l l l and r r r of B \mathcal{B} B are isomorphisms and that there is a locally faithful contravariant pseudofunctor F : B → C F:\mathcal{B}\to\mathcal{C} F : B → C . Then B \mathcal{B} B is a bicategory and F : B → C F:\mathcal{B}\to\mathcal{C} F : B → C is a locally faithful contravariant pseudofunctor between bicategories.
Before the proof, note that we could just as easily have formulated a covariant version of the theorem, but the contravariant version is the one we need and avoids the need to define opposite pre-bicategories and other dual notions. Moreover, the strictness assumption on C \mathcal{C} C can be relaxed using the coherence theorem for bicategories (see, e.g., Leinster, 1998, Section 2.3).
Proof. We rely on local faithfulness and strictness of C \mathcal{C} C throughout: to show that an equation between 2-morphisms holds in B \mathcal{B} B , it is enough to apply F F F to it and show that it holds in C \mathcal{C} C .
For objects X , Y ∈ ob ( B ) X,Y\in\operatorname{ob}(\mathcal{B}) X , Y ∈ ob ( B ) , the assumption (CF-HomFunct) and associativity of vertical composition in C \mathcal{C} C imply that ( B ( X , Y ) , ∙ , id ) (\mathcal{B}(X,Y),\mathbin{\bullet},\operatorname{id}) ( B ( X , Y ) , ∙ , id ) is a category. We next show that composition is functorial, that is, that horizontal and vertical composition obey the interchange law, and that id g ⋆ id f = id g ∘ f \operatorname{id}_{g}\star\operatorname{id}_{f}=\operatorname{id}_{g\circ f} id g ⋆ id f = id g ∘ f . The latter follows easily from (CF-CompNat) and (CF-HomFunct), since F ( id g ⋆ id f ) = α g , f − 1 ∙ ( id F f ⋆ id F g ) ∙ α g , f = id F ( g ∘ f ) = F ( id g ∘ f ) F(\operatorname{id}_{g}\star\operatorname{id}_{f})=\alpha_{g,f}^{-1}\mathbin{\bullet}(\operatorname{id}_{Ff}\star\operatorname{id}_{Fg})\mathbin{\bullet}\alpha_{g,f}=\operatorname{id}_{F(g\circ f)}=F(\operatorname{id}_{g\circ f}) F ( id g ⋆ id f ) = α g , f − 1 ∙ ( id F f ⋆ id F g ) ∙ α g , f = id F ( g ∘ f ) = F ( id g ∘ f ) . For the former, let ζ i : f i ⇒ g i \zeta_{i}:f_{i}\Rightarrow g_{i} ζ i : f i ⇒ g i and θ i : g i ⇒ h i \theta_{i}:g_{i}\Rightarrow h_{i} θ i : g i ⇒ h i (i = 1 , 2 i=1,2 i = 1 , 2 ) be 2-morphisms in B \mathcal{B} B . We compute:
F ( ( θ 2 ∙ ζ 2 ) ⋆ ( θ 1 ∙ ζ 1 ) ) = α h 2 , h 1 − 1 ∙ ( ( F θ 1 ∙ F ζ 1 ) ⋆ ( F θ 2 ∙ F ζ 2 ) ) ∙ α f 2 , f 1 (149) F\bigl((\theta_{2}\mathbin{\bullet}\zeta_{2})\star(\theta_{1}\mathbin{\bullet}\zeta_{1})\bigr)=\alpha_{h_{2},h_{1}}^{-1}\mathbin{\bullet}\bigl((F\theta_{1}\mathbin{\bullet}F\zeta_{1})\star(F\theta_{2}\mathbin{\bullet}F\zeta_{2})\bigr)\mathbin{\bullet}\alpha_{f_{2},f_{1}} \tag*{(149)} F ( ( θ 2 ∙ ζ 2 ) ⋆ ( θ 1 ∙ ζ 1 ) ) = α h 2 , h 1 − 1 ∙ ( ( F θ 1 ∙ F ζ 1 ) ⋆ ( F θ 2 ∙ F ζ 2 ) ) ∙ α f 2 , f 1 (149)
= α h 2 , h 1 − 1 ∙ ( F θ 1 ⋆ F θ 2 ) ∙ ( F ζ 1 ⋆ F ζ 2 ) ∙ α f 2 , f 1 (150) =\alpha_{h_{2},h_{1}}^{-1}\mathbin{\bullet}(F\theta_{1}\star F\theta_{2})\mathbin{\bullet}(F\zeta_{1}\star F\zeta_{2})\mathbin{\bullet}\alpha_{f_{2},f_{1}} \tag*{(150)} = α h 2 , h 1 − 1 ∙ ( F θ 1 ⋆ F θ 2 ) ∙ ( F ζ 1 ⋆ F ζ 2 ) ∙ α f 2 , f 1 (150)
= F ( θ 2 ⋆ θ 1 ) ∙ F ( ζ 2 ⋆ ζ 1 ) = F ( ( θ 2 ⋆ θ 1 ) ∙ ( ζ 2 ⋆ ζ 1 ) ) (151) =F(\theta_{2}\star\theta_{1})\mathbin{\bullet}F(\zeta_{2}\star\zeta_{1})=F\bigl((\theta_{2}\star\theta_{1})\mathbin{\bullet}(\zeta_{2}\star\zeta_{1})\bigr) \tag*{(151)} = F ( θ 2 ⋆ θ 1 ) ∙ F ( ζ 2 ⋆ ζ 1 ) = F ( ( θ 2 ⋆ θ 1 ) ∙ ( ζ 2 ⋆ ζ 1 ) ) (151)
Here, the first equality applies (CF-CompNat) and (CF-HomFunct). The second uses the interchange law of C \mathcal{C} C , the third applies (CF-CompNat) twice, and the fourth concludes the computation with (CF-HomFunct).
Moving on to naturality of the associator, let f , f ′ ∈ B ( W , X ) , g , g ′ ∈ B ( X , Y ) f,f' \in\mathcal{B}(W,X), g,g' \in\mathcal{B}(X,Y) f , f ′ ∈ B ( W , X ) , g , g ′ ∈ B ( X , Y ) , and h , h ′ ∈ B ( Y , Z ) h,h' \in\mathcal{B}(Y,Z) h , h ′ ∈ B ( Y , Z ) be 1-morphisms, and let θ 1 : f ⇒ f ′ \theta_1:f \Rightarrow f' θ 1 : f ⇒ f ′ , θ 2 : g ⇒ g ′ \theta_2:g \Rightarrow g' θ 2 : g ⇒ g ′ , and θ 3 : h ⇒ h ′ \theta_3:h \Rightarrow h' θ 3 : h ⇒ h ′ be 2-morphisms. We have to show that a h ′ , g ′ , f ′ ∙ ( ( θ 3 ⋆ θ 2 ) ⋆ θ 1 ) ∙ a h , g , f − 1 = ( θ 3 ⋆ ( θ 2 ⋆ θ 1 ) ) a_{h',g',f'} \mathbin{\bullet} ((\theta_3 \star\theta_2) \star\theta_1) \mathbin{\bullet} a^{-1}_{h,g,f} = (\theta_3 \star(\theta_2 \star\theta_1)) a h ′ , g ′ , f ′ ∙ (( θ 3 ⋆ θ 2 ) ⋆ θ 1 ) ∙ a h , g , f − 1 = ( θ 3 ⋆ ( θ 2 ⋆ θ 1 )) .
Together with (CF-HomFunct), two applications of (CF-CompNat) and the interchange law yield that
F ( a h ′ , g ′ , f ′ ∙ ( ( θ 3 ⋆ θ 2 ) ⋆ θ 1 ) ∙ a h , g , f − 1 ) = (152) F\left(a_{h',g',f'} \mathbin{\bullet} ((\theta_3 \star\theta_2) \star\theta_1) \mathbin{\bullet} a^{-1}_{h,g,f}\right) = \tag*{(152)} F ( a h ′ , g ′ , f ′ ∙ (( θ 3 ⋆ θ 2 ) ⋆ θ 1 ) ∙ a h , g , f − 1 ) = (152)
F ( a h ′ , g ′ , f ′ ) ∙ α h ′ , g ′ , f ′ − 1 ∙ ( F f ′ α h ′ , g ′ − 1 ) ∙ Θ ∙ ( F f α h ′ , g ′ ) ∙ α h g , f ∙ F ( a h , g , f ) − 1 , (153) F(a_{h',g',f'}) \mathbin{\bullet} \alpha^{-1}_{h',g',f'} \mathbin{\bullet} (Ff' \alpha^{-1}_{h',g'}) \mathbin{\bullet} \Theta\mathbin{\bullet} (Ff \alpha_{h',g'}) \mathbin{\bullet} \alpha_{hg,f} \mathbin{\bullet} F(a_{h,g,f})^{-1}, \tag*{(153)} F ( a h ′ , g ′ , f ′ ) ∙ α h ′ , g ′ , f ′ − 1 ∙ ( F f ′ α h ′ , g ′ − 1 ) ∙ Θ ∙ ( F f α h ′ , g ′ ) ∙ α h g , f ∙ F ( a h , g , f ) − 1 , (153)
where Θ = F ( θ 1 ) ⋆ F ( θ 2 ) ⋆ F ( θ 3 ) \Theta= F(\theta_1) \star F(\theta_2) \star F(\theta_3) Θ = F ( θ 1 ) ⋆ F ( θ 2 ) ⋆ F ( θ 3 ) . Likewise, one sees that
F ( θ 3 ⋆ ( θ 2 ⋆ θ 1 ) ) = α h ′ , g ′ ∘ f ′ − 1 ∙ ( α g ′ , f ′ − 1 ∙ F h ′ ) ∙ Θ ∙ ( α g , f ∙ F h ) ∙ α h , g ∘ f (154) F(\theta_3 \star(\theta_2 \star\theta_1)) = \alpha^{-1}_{h',g'\circ f'} \mathbin{\bullet} (\alpha^{-1}_{g',f'} \mathbin{\bullet} Fh') \mathbin{\bullet} \Theta\mathbin{\bullet} (\alpha_{g,f} \mathbin{\bullet} Fh) \mathbin{\bullet} \alpha_{h,g\circ f} \tag*{(154)} F ( θ 3 ⋆ ( θ 2 ⋆ θ 1 )) = α h ′ , g ′ ∘ f ′ − 1 ∙ ( α g ′ , f ′ − 1 ∙ F h ′ ) ∙ Θ ∙ ( α g , f ∙ F h ) ∙ α h , g ∘ f (154)
Using that whiskering preserves inverses and vertical composites, and (CF-Assoc), one sees that the respective factors on the left and right agree.
Turning to naturality of the unitors, let θ : f ⇒ f ′ \theta:f \Rightarrow f' θ : f ⇒ f ′ with f , f ′ ∈ B ( X , Y ) f,f' \in\mathcal{B}(X,Y) f , f ′ ∈ B ( X , Y ) . We need to show that l f ′ ∙ ( id id Y ⋆ θ ) = θ ∙ l f l^{f'} \mathbin{\bullet} (\operatorname{id}_{\operatorname{id}_Y} \star\theta) = \theta\mathbin{\bullet} l^f l f ′ ∙ ( id id Y ⋆ θ ) = θ ∙ l f . We compute:
F ( l f ′ ∙ ( id id Y ⋆ θ ) ) = ( F f ′ α Y ) ∙ ( F θ ⋆ id id F Y ) ∙ α id Y , f (155) F\left(l^{f'} \mathbin{\bullet} (\operatorname{id}_{\operatorname{id}_Y} \star\theta)\right) = (Ff' \alpha_Y) \mathbin{\bullet} (F\theta\star\operatorname{id}_{\operatorname{id}_{FY}}) \mathbin{\bullet} \alpha_{\operatorname{id}_Y,f} \tag*{(155)} F ( l f ′ ∙ ( id id Y ⋆ θ ) ) = ( F f ′ α Y ) ∙ ( F θ ⋆ id id F Y ) ∙ α id Y , f (155)
= ( F θ ⋆ α Y ) ∙ α id Y , f = ( F θ ⋆ id id F Y ) ∙ ( F f α Y ) ∙ α id Y , f = F ( θ ∙ l f ) (156) = (F\theta\star\alpha_Y) \mathbin{\bullet} \alpha_{\operatorname{id}_Y,f} = (F\theta\star\operatorname{id}_{\operatorname{id}_{FY}}) \mathbin{\bullet} (Ff \alpha_Y) \mathbin{\bullet} \alpha_{\operatorname{id}_Y,f} = F\left(\theta\mathbin{\bullet} l^f\right) \tag*{(156)} = ( F θ ⋆ α Y ) ∙ α id Y , f = ( F θ ⋆ id id F Y ) ∙ ( F f α Y ) ∙ α id Y , f = F ( θ ∙ l f ) (156)
The first equality uses (CF-HomFunct) and expands F ( l f ′ ) F(l^{f'}) F ( l f ′ ) via (CF-Left) and F ( id id Y ⋆ θ ) F(\operatorname{id}_{\operatorname{id}_Y} \star\theta) F ( id id Y ⋆ θ ) via (CF-CompNat), cancelling resulting inverse compositors. The second and third result from interchange in C \mathcal{C} C . The final equality uses that F θ ⋆ id id F Y = F θ F\theta\star\operatorname{id}_{\operatorname{id}_{FY}} = F\theta F θ ⋆ id id F Y = F θ because of strictness of C \mathcal{C} C and concludes with (CF-Left) and (CF-HomFunct). The computation for the right unitor is analogous.
We now show that the pentagon identities hold in B \mathcal{B} B . Let f : W → X f:W \to X f : W → X , g : X → Y g:X \to Y g : X → Y , h : Y → Z h:Y \to Z h : Y → Z and k : Z → V k:Z \to V k : Z → V be 1-morphisms of B \mathcal{B} B . The pentagon identity (Pentagon) says that the composites along
( ( k h ) g ) f → a k , h , g f ( k ( h g ) ) f → a k , h g , f k ( ( h g ) f ) → k a h , g , f k ( h ( g f ) ) (157) ((kh)g)f \xrightarrow{a_{k,h,gf}} (k(hg))f \xrightarrow{a_{k,hg,f}} k((hg)f) \xrightarrow{k a_{h,g,f}} k(h(gf)) \tag*{(157)} (( k h ) g ) f a k , h , g f ( k ( h g )) f a k , h g , f k (( h g ) f ) k a h , g , f k ( h ( g f )) (157)
and
( ( k h ) g ) f → a k h , g , f ( k h ) ( g f ) → a k , h , g f k ( h ( g f ) ) (158) ((kh)g)f \xrightarrow{a_{kh,g,f}} (kh)(gf) \xrightarrow{a_{k,h,gf}} k(h(gf)) \tag*{(158)} (( k h ) g ) f a k h , g , f ( k h ) ( g f ) a k , h , g f k ( h ( g f )) (158)
agree. By local faithfulness it suffices to show that their images under F F F agree.
Set N : = F f ∘ F g ∘ F h ∘ F k N := Ff \circ Fg \circ Fh \circ Fk N := F f ∘ F g ∘ F h ∘ F k , a 1-morphism F V → F W FV \to FW F V → F W of C \mathcal{C} C , using strictness to omit the brackets. The image of each of the five 1-morphisms occurring in (B.0.13) and (B.0.14) under F F F is isomorphic to N N N via the following 2-morphisms, which one gets by using compositors and their whiskerings to remove the brackets from the outside inwards, and which are hence clearly invertible.
A 1 : = ( F f F g α k , h ) ∙ ( F f α k h , g ) ∙ α ( k h ) g , f : F ( ( ( k h ) g ) f ) ⇒ N (159) A_{1} := (FfFg\alpha_{k,h}) \mathbin{\bullet} (Ff\alpha_{kh,g}) \mathbin{\bullet} \alpha_{(kh)g,f} : F(((kh)g)f) \Rightarrow N \tag*{(159)} A 1 := ( F f F g α k , h ) ∙ ( F f α k h , g ) ∙ α ( k h ) g , f : F ((( k h ) g ) f ) ⇒ N (159)
A 2 : = ( F f α h , g F k ) ∙ ( F f α k h , g ) ∙ α k ( h g ) , f : F ( ( k ( h g ) ) f ) ⇒ N (160) A_{2} := (Ff\alpha_{h,g}Fk) \mathbin{\bullet} (Ff\alpha_{kh,g}) \mathbin{\bullet} \alpha_{k(hg),f} : F((k(hg))f) \Rightarrow N \tag*{(160)} A 2 := ( F f α h , g F k ) ∙ ( F f α k h , g ) ∙ α k ( h g ) , f : F (( k ( h g )) f ) ⇒ N (160)
A 3 : = ( F f α h , g F k ) ∙ ( α h g , f F k ) ∙ α k , ( h g ) f : F ( k ( ( h g ) f ) ) ⇒ N (161) A_{3} := (Ff\alpha_{h,g}Fk) \mathbin{\bullet} (\alpha_{hg,f}Fk) \mathbin{\bullet} \alpha_{k,(hg)f} : F(k((hg)f)) \Rightarrow N \tag*{(161)} A 3 := ( F f α h , g F k ) ∙ ( α h g , f F k ) ∙ α k , ( h g ) f : F ( k (( h g ) f )) ⇒ N (161)
A 4 : = ( α g , f F h F k ) ∙ ( α h , g f F k ) ∙ α k , h ( g f ) : F ( k ( h ( g f ) ) ) ⇒ N (162) A_{4} := (\alpha_{g,f}FhFk) \mathbin{\bullet} (\alpha_{h,gf}Fk) \mathbin{\bullet} \alpha_{k,h(gf)} : F(k(h(gf))) \Rightarrow N \tag*{(162)} A 4 := ( α g , f F h F k ) ∙ ( α h , g f F k ) ∙ α k , h ( g f ) : F ( k ( h ( g f ))) ⇒ N (162)
A 5 : = ( F f F g α k , h ) ∙ ( α g , f F ( k h ) ) ∙ α k h , g f : F ( ( k h ) ( g f ) ) ⇒ N (163) A_{5} := (FfFg\alpha_{k,h}) \mathbin{\bullet} (\alpha_{g,f}F(kh)) \mathbin{\bullet} \alpha_{kh,gf} : F((kh)(gf)) \Rightarrow N \tag*{(163)} A 5 := ( F f F g α k , h ) ∙ ( α g , f F ( k h )) ∙ α k h , g f : F (( k h ) ( g f )) ⇒ N (163)
The following equations, proved below, express that the A i A_i A i together yield a cocone for the image of the pentagon. Equivalently, the first three give a cocone for the long composite and the last two a cocone for the short composite above.
A 2 ∙ F ( α k , h g , f ) = A 1 , A 3 ∙ F ( α k , h g , f ) = A 2 , A 4 ∙ F ( k α h , g , f ) = A 3 (Cocone-L) A_{2} \mathbin{\bullet} F(\alpha_{k,hg,f}) = A_{1}, \qquad A_{3} \mathbin{\bullet} F(\alpha_{k,hg,f}) = A_{2}, \qquad A_{4} \mathbin{\bullet} F(k\alpha_{h,g,f}) = A_{3} \qquad\text{(Cocone-L)} A 2 ∙ F ( α k , h g , f ) = A 1 , A 3 ∙ F ( α k , h g , f ) = A 2 , A 4 ∙ F ( k α h , g , f ) = A 3 (Cocone-L)
A 5 ∙ F ( α k h , g , f ) = A 1 , A 4 ∙ F ( α k , h g , f ) = A 5 . (Cocone-S) A_{5} \mathbin{\bullet} F(\alpha_{kh,g,f}) = A_{1}, \qquad A_{4} \mathbin{\bullet} F(\alpha_{k,hg,f}) = A_{5}. \qquad\text{(Cocone-S)} A 5 ∙ F ( α k h , g , f ) = A 1 , A 4 ∙ F ( α k , h g , f ) = A 5 . (Cocone-S)
Using this and invertibility of the A i A_i A i , (Cocone-L) and (Cocone-S), one sees that the images of the composites along (B.0.13) and (B.0.14) are telescopic, yielding the desired equality:
F ( ( k α h , g , f ) ∙ α k , h g , f ∙ ( α k h , g , f ) ) = F ( k α h , g , f ) ∙ F ( α k , h g , f ) ∙ F ( α k h , g , f ) (164) F\left((k\alpha_{h,g,f}) \mathbin{\bullet} \alpha_{k,hg,f} \mathbin{\bullet} (\alpha_{kh,g,f})\right) = F(k\alpha_{h,g,f}) \mathbin{\bullet} F(\alpha_{k,hg,f}) \mathbin{\bullet} F(\alpha_{kh,g,f}) \tag*{(164)} F ( ( k α h , g , f ) ∙ α k , h g , f ∙ ( α k h , g , f ) ) = F ( k α h , g , f ) ∙ F ( α k , h g , f ) ∙ F ( α k h , g , f ) (164)
= ( A 4 − 1 ∙ A 3 ) ∙ ( A 3 − 1 ∙ A 2 ) ∙ ( A 2 − 1 ∙ A 1 ) (165) = (A_{4}^{-1} \mathbin{\bullet} A_{3}) \mathbin{\bullet} (A_{3}^{-1} \mathbin{\bullet} A_{2}) \mathbin{\bullet} (A_{2}^{-1} \mathbin{\bullet} A_{1}) \tag*{(165)} = ( A 4 − 1 ∙ A 3 ) ∙ ( A 3 − 1 ∙ A 2 ) ∙ ( A 2 − 1 ∙ A 1 ) (165)
= A 4 − 1 ∙ A 1 = ( A 4 − 1 ∙ A 5 ) ∙ ( A 5 − 1 ∙ A 1 ) (166) = A_{4}^{-1} \mathbin{\bullet} A_{1} = (A_{4}^{-1} \mathbin{\bullet} A_{5}) \mathbin{\bullet} (A_{5}^{-1} \mathbin{\bullet} A_{1}) \tag*{(166)} = A 4 − 1 ∙ A 1 = ( A 4 − 1 ∙ A 5 ) ∙ ( A 5 − 1 ∙ A 1 ) (166)
= F ( α k , h g , f ) ∙ F ( α k h , g , f ) = F ( α k , h g , f ∙ α k h , g , f ) , (167) = F(\alpha_{k,hg,f}) \mathbin{\bullet} F(\alpha_{kh,g,f}) = F(\alpha_{k,hg,f} \mathbin{\bullet} \alpha_{kh,g,f}), \tag*{(167)} = F ( α k , h g , f ) ∙ F ( α k h , g , f ) = F ( α k , h g , f ∙ α k h , g , f ) , (167)
where the first and last equalities follow by (CF-HomFunct).
It remains to prove (Cocone-L) and (Cocone-S). For the identity for ( A 2 , A 1 ) (A_{2}, A_{1}) ( A 2 , A 1 ) , we see that
A 2 ∙ F ( α k , h g , f ) = ( F f [ ( α h , g F k ) ∙ α k , h g ∙ F α k h , g ] ) ∙ α ( k h ) g , f (168) A_{2} \mathbin{\bullet} F(\alpha_{k,hg,f}) = \left(Ff\left[(\alpha_{h,g}Fk) \mathbin{\bullet} \alpha_{k,hg} \mathbin{\bullet} F\alpha_{kh,g}\right]\right) \mathbin{\bullet} \alpha_{(kh)g,f} \tag*{(168)} A 2 ∙ F ( α k , h g , f ) = ( F f [ ( α h , g F k ) ∙ α k , h g ∙ F α k h , g ] ) ∙ α ( k h ) g , f (168)
= ( F f [ ( F g α k , h ) ∙ α k h , g ] ) ∙ α ( k h ) g , f = A 1 , (169) = \left(Ff\left[(Fg\alpha_{k,h}) \mathbin{\bullet} \alpha_{kh,g}\right]\right) \mathbin{\bullet} \alpha_{(kh)g,f} = A_{1}, \tag*{(169)} = ( F f [ ( F g α k , h ) ∙ α k h , g ] ) ∙ α ( k h ) g , f = A 1 , (169)
where we use first the instance of (CF-CompNat) for F ( α k , h ⋆ id f ) F(\alpha_{k,h} \mathbin{\star} \operatorname{id}_{f}) F ( α k , h ⋆ id f ) , (CF-HomFunct), and the fact that whiskering commutes with vertical composition, and then (CF-Assoc) for α k , h , g \alpha_{k,h,g} α k , h , g to conclude. An analogous argument for ( A 4 , A 3 ) (A_{4}, A_{3}) ( A 4 , A 3 ) , combined with a right whiskering by F k Fk F k in the second step, yields
A 4 ∙ F ( k α h , g , f ) = ( [ ( α g , f F h ) ∙ α h , g f ∙ F α h , g , f ] F k ) ∙ α k , ( h g ) f (170) A_{4} \mathbin{\bullet} F(k\alpha_{h,g,f}) = \left(\left[(\alpha_{g,f}Fh) \mathbin{\bullet} \alpha_{h,gf} \mathbin{\bullet} F\alpha_{h,g,f}\right]Fk\right) \mathbin{\bullet} \alpha_{k,(hg)f} \tag*{(170)} A 4 ∙ F ( k α h , g , f ) = ( [ ( α g , f F h ) ∙ α h , g f ∙ F α h , g , f ] F k ) ∙ α k , ( h g ) f (170)
= ( [ ( F f α h , g ) ∙ α h , g f ] F k ) ∙ α k , ( h g ) f = A 3 . (171) = \left(\left[(Ff\alpha_{h,g}) \mathbin{\bullet} \alpha_{h,gf}\right]Fk\right) \mathbin{\bullet} \alpha_{k,(hg)f} = A_{3}. \tag*{(171)} = ( [ ( F f α h , g ) ∙ α h , g f ] F k ) ∙ α k , ( h g ) f = A 3 . (171)
The remaining identities, for ( A 3 , A 2 ) (A_{3}, A_{2}) ( A 3 , A 2 ) , ( A 5 , A 1 ) (A_{5}, A_{1}) ( A 5 , A 1 ) , and ( A 4 , A 5 ) (A_{4}, A_{5}) ( A 4 , A 5 ) , follow from (CF-Assoc) for the triples ( k , h g , f ) (k,hg,f) ( k , h g , f ) , ( k h , g , f ) (kh,g,f) ( k h , g , f ) , and ( k , h , g f ) (k,h,gf) ( k , h , g f ) , respectively, by postcomposing with the leftmost factor in the definition of A 3 A_3 A 3 , A 5 A_5 A 5 , and A 4 A_4 A 4 , respectively, using interchange in the last case.
Finally and more directly, we prove the triangle identities: Let f : W → X f : W \to X f : W → X and g : X → Y g : X \to Y g : X → Y be 1-morphisms in B \mathcal{B} B . By (CF-CompNat) for F ( id ⋆ ∘ l f ) F(\operatorname{id}_\star\circ l^f) F ( id ⋆ ∘ l f ) and F ( r g ∘ id ⋆ ) F(r^g \circ\operatorname{id}_\star) F ( r g ∘ id ⋆ ) , then (CF-Left) for f f f and (CF-Right) for g g g , and finally by the fact that whiskering commutes with vertical composition,
α g , f ∙ F ( g l f ) = ( F l f F g ) ∙ α g , id X f = ( F f α X F g ) ∙ ( α id X , f F g ) ∙ α g , id X f = : B 2 (172) \alpha_{g,f} \mathbin{\bullet} F(g l^f) = (F l^f Fg) \mathbin{\bullet} \alpha_{g,\operatorname{id}_X f} = (F f \alpha_X Fg) \mathbin{\bullet} (\alpha_{\operatorname{id}_X,f} Fg) \mathbin{\bullet} \alpha_{g,\operatorname{id}_X f} =: B_2 \tag*{(172)} α g , f ∙ F ( g l f ) = ( F l f F g ) ∙ α g , id X f = ( F f α X F g ) ∙ ( α id X , f F g ) ∙ α g , id X f =: B 2 (172)
α g , f ∙ F ( r g f ) = ( F f F r g ) ∙ α id X g , f = ( F f α X F g ) ∙ ( F f α g , id X ) ∙ α id X g , f = : B 1 (173) \alpha_{g,f} \mathbin{\bullet} F(r^g f) = (F f Fr^g) \mathbin{\bullet} \alpha_{\operatorname{id}_X g,f} = (F f \alpha_X Fg) \mathbin{\bullet} (F f \alpha_{g,\operatorname{id}_X}) \mathbin{\bullet} \alpha_{\operatorname{id}_X g,f} =: B_1 \tag*{(173)} α g , f ∙ F ( r g f ) = ( F f F r g ) ∙ α id X g , f = ( F f α X F g ) ∙ ( F f α g , id X ) ∙ α id X g , f =: B 1 (173)
both 2-morphisms in C \mathcal{C} C to F f ∘ F g Ff \circ Fg F f ∘ F g , noting that the identity in C \mathcal{C} C is strict. Post-composing the instance of (CF-Assoc) for F ( α g , id X , f ) F(\alpha_{g,\operatorname{id}_X,f}) F ( α g , id X , f ) with ( F f α X F g ) (Ff \alpha_X Fg) ( F f α X F g ) yields B 2 ∙ F ( α g , id X , f ) = B 1 B_2 \mathbin{\bullet} F(\alpha_{g,\operatorname{id}_X,f}) = B_1 B 2 ∙ F ( α g , id X , f ) = B 1 , whence
α g , f ∙ F ( ( g l f ) ∙ α g , id X , f ) = B 2 ∙ F ( α g , id X , f ) = B 1 = α g , f ∙ F ( r g f ) , (174) \alpha_{g,f} \mathbin{\bullet} F\left((g l^f) \mathbin{\bullet} \alpha_{g,\operatorname{id}_X,f}\right) = B_2 \mathbin{\bullet} F(\alpha_{g,\operatorname{id}_X,f}) = B_1 = \alpha_{g,f} \mathbin{\bullet} F(r^g f), \tag*{(174)} α g , f ∙ F ( ( g l f ) ∙ α g , id X , f ) = B 2 ∙ F ( α g , id X , f ) = B 1 = α g , f ∙ F ( r g f ) , (174)
using (CF-HomFunct) for the first equality. One concludes with invertibility of α g , f \alpha_{g,f} α g , f .
□ \square □