Sheaves of Structures, Heyting-Valued Structures, and a Generalization of Ło´s’sTheorem
Hisashi Aratake*
Research Institute for Mathematical Sciences, Kyoto University, Kyoto, Japan
Abstract Sheaves of structures are useful to give constructions in universal al- gebra and model theory. We can describe their logical behavior in terms of Heyting-valued structures. In this paper, we first provide a systematic treatment of sheaves of structures and Heyting-valued structures from the viewpoint of categorical logic. We then prove a form of Ło´s’sthe- orem for Heyting-valued structures. We also give a characterization of Heyting-valued structures for which Ło´s’stheorem holds with respect to any maximal filter.
0 Introduction
Sheaf-theoretic constructions have been used in universal algebra and model theory. In this context, sheaves of abelian groups or rings in geometry are generalized to sheaves of structures. We can obtain, for example, the product (resp. an ultraproduct) of a family of structures from some sheaf by taking the set of global sections (resp. a stalk). This viewpoint originated from the early literature [Com74], [Ell74] and [Maci73]. In combination with the theory of sheaf representations of algebras, Macintyre [Maci73] succeeded in giv- ing model-companions of some theories of commutative rings by transferring model-theoretic properties from stalks to global sections. On the other hand, sheaves have another description as Heyting-valued sets. The notion of Heyting-valued sets originally arises from that of Boolean- valued models of set theory, which was introduced in relation to Cohen’s forcing. The development of topos theory in the early seventies, mainly due arXiv:2012.04317v1 [math.LO] 8 Dec 2020 to Lawvere & Tierney, revealed profound relationships between toposes and models of set theory; objects in a topos can be regarded as “generalized sets” in a universe. Subsequently, Fourman & Scott [FS79] and Higgs 1 indepen- dently established the categorical treatment of Heyting-valued sets (see Re- mark 1.6). The category Sh(X) of sheaves of sets on a space X and the cate- gory Set(O(X)) of O(X)-valued sets turned out to be categorically equivalent.
*E-mail address: [email protected]
1Originally in his preprint written in 1973, part of which was later published as [Hig84].
1 Some model-theorists of that era immediately applied Heyting-valued sets to concrete problems in sheaf-theoretic model theory. However, general meth- ods of Heyting-valued model theory have not been explored enough, though Fourman & Scott mentioned such a direction in the preamble of [FS79]. Even worse, we are not aware of any clear explanation of the relationship between sheaves of structures and Heyting-valued structures. In this paper, employ- ing well-established languages of categorical logic, we will give a gentle and coherent account of Heyting-valued semantics of first-order logic from the categorical point of view, and will apply that framework to obtain a gener- alization of Ło´s’stheorem for Heyting-valued structures. We also provide a characterization of Heyting-valued structures for which Ło´s’stheorem holds w.r.t. any maximal filter. Our theorems improve the works of Caicedo [Cai95] and Pierobon & Viale [PV20]. While our principal examples of Heyting-valued sets are sheaves on topological spaces, other natural examples include sheaves on the complete Boolean algebra of regular open sets and Boolean-valued sets on the measure algebra. Therefore, we will develop our theory based on any complete Heyting algebra (a.k.a. a frame or a locale) rather than a topological space. The intended audience for this paper is anyone who has interests both in model theory and in categorical logic. We assume some familiarity with topos theory and first-order categorical logic. Most categorical prerequisites are covered by [SGL]. In §2.1, we will recall some elements of first-order categorical logic. The areas related to this paper (and its sequels in the future) are diverse, including model theory, universal algebra, set theory, categorical logic, topos theory, and their applications to ordinary mathematics. The author did his best to ensure that the reader can follow the scattered literature (especially in model theory and topos theory) on each occasion during the course.
The structure of this paper: In §1, we will begin with preliminaries on sheaves and Heyting-valued sets. After we see basic properties of Heyting- valued sets and morphisms between them, we will give an outline of the equivalence of sheaves and Heyting-valued sets. We also provide some de- tails on the topos Set(O(X)) of O(X)-valued sets. In §2, we will study struc- tures in the toposes Sh(X) and Set(O(X)) and the relationship between them. We will also introduce forcing values of formulas categorically. In §3, ob- serving that sheaves of structures generalize some model-theoretic construc- tions, we will introduce a further generalization of filter-quotients of sheaves to Heyting-valued structures and prove Ło´s’stheorem and the characteriza- tion theorem. In §4, we will indicate some possible future directions with an expanded list of previous works.
Closing the introduction, we have to mention Loullis’ work [Lou79] on Boolean-valued model theory. The starting point of this research was try- ing to digest his work from a modern categorical viewpoint, though our work is still too immature to give the reader a full explanation of his contribution. If the author had not met his work, this paper would not have existed. The
2 author regrets his untimely death, according to [BR81], in 1978.
Acknowledgment: The author is grateful to his supervisor, Kazushige Terui, for careful reading of an earlier draft and many helpful suggestions on the presentation of this paper. He also thanks Soichiro Fujii for useful comments.
1 Sheaves and Heyting-Valued Sets
Heyting-valued sets were introduced independently by Higgs [Hig84] and by Fourman & Scott [FS79]. In this section, we will review the construction of the category Set(O(X)) of O(X)-valued sets for a locale X, its relation to sheaves on X, and its categorical structures as a topos. Most results are covered by [Hig84], [FS79], [Elephant, §C1.3] and [HoCA3, Chapter 2]. For the reader’s convenience, we will occasionally give brief sketches of proofs. For aspects of Heyting-valued sets in intuitionistic logic, see [TvD88, Chap- ters 13–14]. As a category, Set(O(X)) is a prototypical example of the topos obtained from a tripos (see [HJP80] and [vOos08, Chapter 2]). The internal logic of Set(O(X)) is reduced to the logic of tripos. Walters [Wal81], [Wal82] developed another direction of generalization of Heyting-valued sets.
1.1 Heyting-Valued Sets Definition 1.1. A frame is a complete lattice satisfying the infinitary distribu- tive law: _ _ a ∧ bi = a ∧ bi. i i In particular, any frame has 0 and 1. A frame is the same thing as a complete Heyting algebra: the infinitary distributive law for a frame H says that each monotone map a ∧ (−): H → H preserves arbitrary joins. This happens exactly when each map a ∧ (−) has a right adjoint a → (−): H → H, i.e., a monotone map satisfying
∀b, c ∈ H, [a ∧ b ≤ c ⇐⇒ b ≤ a → c].
This fact follows either from category theory (the General Adjoint Functor The- orem), or from a direct construction _ a → c := { b ∈ H ; a ∧ b ≤ c } .
On the other hand, frame homomorphisms differ from those for complete Heyting algebras (and even those for complete lattices):
Definition 1.2. Let H,H0 be frames. A frame homomorphism h: H → H0 is a map from H to H0 preserving finite meets and arbitrary joins. Let Frm denote the category of frames.
3 Similarly to the above, any frame homomorphism h: H → H0 has a right adjoint k : H0 → H given by _ k(b) = { a ∈ H ; h(a) ≤ b } .
Any continuous map f : X → Y of topological spaces gives rise to a frame homomorphism f ∗ : O(Y ) → O(X) given by f ∗(V ) = f −1(V ), where O(X) (resp. O(Y )) is the frame of open sets of X (resp. Y ). The functor O(−): Topop → Frm is full and faithful on sober spaces (i.e., spaces satisfy- ing a suitable axiom between T0 and T2). Therefore, it translates the language of spaces to that of frames, and we may consider frames as “point-free” spaces. This justifies the following definition: Definition 1.3. A frame considered as an object of Frmop is called a locale. We denote Frmop by Loc and the frame corresponding to a locale X ∈ Loc by O(X). We will write U, V, etc. for elements of O(X) and 0X (resp. 1X ) for the smallest (resp. largest) element. For a morphism f : X → Y of locales, the corresponding frame homomor- ∗ ∗ phism is denoted by f : O(Y ) → O(X). f has a right adjoint f∗ : O(X) → O(Y ). Morphisms of locales are also called continuous maps of locales.
By writing X` for the locale given by a topological space X, i.e., O(X`) = O(X), we now have a functor (−)` : Top → Loc. It has a right adjoint pt: Loc → Top sending a locale X to the space pt(X) of “points of X” (see, e.g., [SGL, Chapter IX]). For more on frames and locales in point-free topology, see [Joh82] and [PP12]. We are now ready to define Heyting-valued sets. In the remainder of this section, we fix a locale X. Definition 1.4. An O(X)-valued set (A, α) is a pair of a set A and a map α: A × A → O(X) such that • ∀a, b ∈ A, α(a, b) = α(b, a)
• ∀a, b, c ∈ A, α(a, b) ∧ α(b, c) ≤ α(a, c) In the logic of tripos (associated with X), α is a “partial equivalence rela- tion” on A. Instead of α(a, b), the notation a = b is frequently used in the literature. We introduce a few conventions: J K • α(a) := α(a, a) is called the extent of a. Note that α(a, b) ≤ α(a) ∧ α(b).
• If α(a) = 1X , then a is called a global element of (A, α). Morphisms of Heyting-valued sets should be “functional relations” (again in the logic of tripos). Definition 1.5. Let (A, α), (B, β) be O(X)-valued sets. A morphism ϕ:(A, α) → (B, β) of O(X)-valued sets is a map A × B → O(X) which satisfies ∀a, a0 ∈ A, ∀b, b0 ∈ B, α(a, a0) ∧ ϕ(a, b) ∧ β(b, b0) ≤ ϕ(a0, b0), ∀a ∈ A, ∀b, b0 ∈ B, ϕ(a, b) ∧ ϕ(a, b0) ≤ β(b, b0), _ ∀a ∈ A, α(a) = ϕ(a, b). b∈B
4 In particular, ϕ(a, b) ≤ α(a) ∧ β(b) always holds. If ψ :(B, β) → (C, γ) is another morphism, we can define the composite ψ ◦ ϕ by _ (ψ ◦ ϕ)(a, c) = ϕ(a, b) ∧ ψ(b, c). b∈B
We write Set(O(X)) for the category of O(X)-valued sets and morphisms, where the identity id(A,α) is given by α itself. Remark 1.6. In set theory, for a frame H, we can construct a model V (H) of intuitionistic set theory ( [Bel14, Chapter IV]). V (H) is called the Heyting- valued universe. The category Set(H) is regarded as a categorical counter- part of V (H) ( [Bel05, Appendix], [ACM19]), and we can take arguments and examples from set theory to investigate Heyting-valued sets (cf. [PV20]). We list useful facts on Heyting-valued sets, some of which will not be used in this paper.
Lemma 1.7. Two morphisms ϕ, ψ :(A, α) ⇒ (B, β) are identical if ∀a ∈ A, ∀b ∈ B, ϕ(a, b) ≤ ψ(a, b).
Proof. Suppose ∀a ∈ A, ∀b ∈ B, ϕ(a, b) ≤ ψ(a, b). Then, _ ψ(a, b) = ψ(a, b) ∧ α(a) = ψ(a, b) ∧ ϕ(a, b0) b0 _ _ ≤ [ϕ(a, b0) ∧ ψ(a, b0) ∧ ψ(a, b)] ≤ [ϕ(a, b0) ∧ β(b0, b)] = ϕ(a, b). b0 b0
Proposition 1.8. Let ϕ:(A, α) → (B, β) be a morphism in Set(O(X)).
(1) ϕ is a monomorphism if and only if
∀a, a0 ∈ A, ∀b ∈ B, ϕ(a, b) ∧ ϕ(a0, b) ≤ α(a, a0).
(2) ϕ is an epimorphism if and only if _ ∀b ∈ B, β(b) = ϕ(a, b). a∈A
(3) ϕ is an isomorphism if and only if it is monic and epic. In other words, Set(O(X)) is a balanced category. If ϕ is an isomorphism, ϕ−1 is given −1 by ϕ (b, a) = ϕ(a, b). Definition 1.9. We say that a morphism ϕ:(A, α) → (B, β) is represented by a map h: A → B when
∀a ∈ A, ∀b ∈ B, ϕ(a, b) = α(a) ∧ β(ha, b).
Proposition 1.10.
5 (1) A morphism ϕ:(A, α) → (B, β) is represented by a map h: A → B if and only if ∀a ∈ A, ∀b ∈ B, ϕ(a, b) ≤ β(ha, b).
(2) A map h: A → B represents some morphism from (A, α) to (B, β) if and only if ∀a, a0 ∈ A, α(a, a0) ≤ β(ha, ha0). Moreover, if h further satisfies α(a) = β(ha) for all a ∈ A, then the morphism ϕ represented by h is given simply by ϕ(a, b) = β(ha, b). (3) Suppose two maps h, k : A → B represent some morphisms. They rep- resent the same morphism if and only if ∀a ∈ A, α(a) ≤ β(ha, ka).
(4) Let (C, γ) be another O(X)-valued set and ψ :(B, β) → (C, γ) be another morphism. If ϕ (resp. ψ) is represented by a map h (resp. k), then ψϕ is represented by kh. Proposition 1.11. Let ϕ:(A, α) → (B, β) be a morphism represented by h. • ϕ is monic ⇐⇒ ∀a, a0 ∈ A, α(a, a0) = α(a) ∧ α(a0) ∧ β(ha, ha0). W • ϕ is epic ⇐⇒ ∀b ∈ B, β(b) = a[α(a) ∧ β(ha, b)]. Further, if α(a) = β(ha) for all a ∈ A, these conditions reduce to • ϕ is monic ⇐⇒ ∀a, a0 ∈ A, α(a, a0) = β(ha, ha0). W • ϕ is epic ⇐⇒ ∀b ∈ B, β(b) = a β(ha, b). Combining the above facts, we obtain Corollary 1.12. If h satisfies α(a) = β(ha) for all a ∈ A, then h represents an isomorphism exactly when the following conditions hold _ ∀a, a0 ∈ A, α(a, a0) = β(ha, ha0), and ∀b ∈ B, β(b) = β(ha, b). a If h satisfies these conditions, the induced isomorphism ϕ(a, b) = β(ha, b) has −1 the inverse ϕ (b, a) = β(ha, b).
1.2 Sheaves on Locales and Complete Heyting-Valued Sets Continuing from the previous section, we fix a locale X. Let us discuss the relationship between sheaves on X and O(X)-valued sets. Our presentation style here is largely due to [Elephant, §C1.3]. Definition 1.13.
(1)A presheaf on X is a functor O(X)op → Set. For a presheaf P and U ∈ O(X), elements of PU (resp. of P 1X ) are called sections of P on U (resp. global sections of P ). If a ∈ PU and W ≤ U, we will write a|W for P (W ≤ U)(a).
6 (2) A presheaf P is said to be a sheaf on X when it satisfies the following W condition: For any covering {Ui}i of U ∈ O(X) (i.e., U = i Ui) and any
family {ai}i of sections ai ∈ PUi, if ai|Ui∧Uj = aj|Ui∧Uj for all i, j, then
there exists a unique a ∈ PU such that a|Ui = ai for all i. (3) Morphisms of presheaves are defined to be natural transformations. op The functor category SetO(X) is also called the category of presheaves. Let Sh(X) denote its full subcategory spanned by sheaves. ` We can associate a presheaf P with an O(X)-valued set Θ(P ) := ( U P U, δP ) ` ` as follows: for (a, b) ∈ PU × PV ⊆ U PU × U PU, _ δP (a, b) := { W ≤ U ∧ V ; a|W = b|W } .
Notice that
• a ∈ PU if and only if δP (a) = U. Hence, global elements of Θ(P ) are exactly global sections of P .
• If P is a sheaf, then δP (a, b) is the largest element on which the restric- tions of a and b coincide. Moreover, if X is a topological space,
δP (a, b) = { x ∈ X ; ax = bx } ,
where ax, bx are the germs of a, b over x. ` For a morphism ξ : P → Q of presheaves on X, the induced map h: U PU → ` U QU satisfies
2 a ∀(a, b) ∈ PU , δP (a, b) ≤ δQ(ha, hb) and δP (a) = δQ(ha). U
Therefore, by Proposition 1.10, h represents a morphism Θ(ξ): Θ(P ) → Θ(Q). op This construction gives a functor Θ: SetO(X) → Set(O(X)). Notice that a presheaf P is separated if and only if, for any a, b, δP (a) = δP (b) = δP (a, b) implies a = b. Fourman & Scott [FS79] say Heyting-valued sets are separated if the latter condition holds. To give a similar characteriza- tion of sheaves, we need a more involved definition.
Definition 1.14. For an O(X)-valued set (A, α), define a preorder v on A by
b v a ⇐def==⇒. α(a, b) = α(b).
(A, α) is said to be complete if the following conditions hold:
• v is a partial order. (This is equivalent to separatedness.)
• For any a ∈ A and U ≤ α(a), there exists b ∈ A such that b v a and α(b) = U. (If v is a partial order, such b is uniquely determined and denoted by a|U .)
7 • If a family {ai}i of elements of A is pairwise compatible, i.e., α(ai, aj) = α(ai)∧α(aj) for all i, j, then it has a supremum w.r.t. v. (The supremum is called an amalgamation of {ai}i.)
Let CSet(O(X)) denote the full subcategory of Set(O(X)) spanned by com- plete O(X)-valued sets. Proposition 1.15. A presheaf P on X is a sheaf if and only if Θ(P ) is complete as an O(X)-valued set. Moreover, for any complete O(X)-valued set (A, α), there exists a sheaf P on X such that (A, α) and Θ(P ) are isomorphic.
Proof. The latter part: if (A, α) is complete, then by putting
PU := { a ∈ A ; α(a) = U } , we have a desired sheaf P on X. We can rephrase completeness in terms of singletons.
Definition 1.16. Let (A, α) be an O(X)-valued set. A singleton on (A, α) is a function σ : A → O(X) such that
∀a, a0 ∈ A, σ(a) ∧ α(a, a0) ≤ σ(a0) and σ(a) ∧ σ(a0) ≤ α(a, a0).
In particular, σ(a) ≤ α(a) always holds.
For each a ∈ A, the map σa := α(a, −) is a singleton of (A, α).
Lemma 1.17. For an O(X)-valued set (A, α), TFAE:
(i) (A, α) is complete.
(ii) Any singleton of (A, α) is of the form σa for a uniquely determined a.
Proof. (i)⇒(ii): Suppose (A, α) is complete. Let σ be a singleton on (A, α). Then the family a|σ(a) ; a ∈ A is pairwise compatible, and its supremum s ∈ A satisfies σ = σs. (ii)⇒(i): Suppose the condition (ii) holds. If α(a) = α(a0) = α(a, a0), then 0 σa = σa0 and hence a = a . Thus v is anti-symmetric. If a ∈ A and U ≤ α(a), the map α(a, −) ∧ U is a singleton, and we then have the restriction a|U . If W a family {ai}i is pairwise compatible, the map i α(ai, −) is a singleton, and we then have the amalgamation. Lemma 1.18. Let (A, α), (B, β) be O(X)-valued sets with (B, β) complete. Each morphism ϕ:(A, α) → (B, β) is represented by a unique map h: A → B which satisfies α(a, a0) ≤ β(ha, ha0) and α(a) = β(ha) for all a, a0 ∈ A.
Proof. For any fixed a ∈ A, the map ϕ(a, −) is a singleton of (B, β). By com- pleteness, we can find a unique ha ∈ B such that ϕ(a, b) = β(ha, b) for every b ∈ B. This defines a map h: A → B representing ϕ and having the desired properties. This lemma and Proposition 1.10(4) yield
8 Proposition 1.19. Θ induces a categorical equivalence between Sh(X) and CSet(O(X)). On the other hand, we can also show that Set(O(X)) and CSet(O(X)) are categorically equivalent. Proposition 1.20. Let A˜ be the set of singletons on (A, α). Define a valuation α˜ on A˜ by _ α˜(σ, τ) := σ(a) ∧ τ(a). a∈A Then (A,˜ α˜) is a complete O(X)-valued set, and the map A˜ × A 3 (σ, a) 7→ σ(a) ∈ O(X) represents an isomorphism (A, α) ' (A,˜ α˜). We call (A,˜ α˜) the completion of (A, α). For a morphism ϕ:(A, α) → (B, β), let ϕ˜ be the composite of ϕ (A,˜ α˜) →∼ (A, α) −→ (B, β) →∼ (B,˜ β˜).
Then (−˜ ): Set(O(X)) → CSet(O(X)) becomes a functor and also gives a quasi-inverse of the inclusion functor. Corollary 1.21. The categories Sh(X), Set(O(X)) and CSet(O(X)) are cate- gorically equivalent. In particular, Set(O(X)) is a Grothendieck topos.
1.3 Topos Structure of Set(O(X)) In the previous section, we saw that Set(O(X)) is a Grothendieck topos. Here, we will give a concrete description of the topos structure of Set(O(X)). Most results here (except for some details on the lattice P(A, α)) are borrowed from [Hig84]. The constructions will be exploited later in this paper. Proposition 1.22 (Finite limits in Set(O(X))).
(1) Let ({∗}, >) be the O(X)-valued set with >(∗, ∗) = 1X . This yields a terminal object in Set(O(X)).
(2) Let {(Ai, αi)}i∈I be a finite family of O(X)-valued sets. Define a val- Q 0 V 0 0 uation δ on i Ai by δ(a, a ) = i αi(ai, ai) for a = {ai}i∈I and a = 0 Q Q {ai}i∈I . Then ( i Ai, δ) equipped with the canonical projections ( i Ai, δ) → (Ai, αi) is a product of {(Ai, αi)}i∈I in Set(O(X)).
(3) Let ϕ, ψ :(A, α) ⇒ (B, β) be morphisms of O(X)-valued sets. Define a 0 0 W valuation δ on A by δ(a, a ) = α(a, a )∧ b∈B ϕ(a, b)∧ψ(a, b). Then (A, δ) equipped with the canonical morphism (A, δ) (A, α) is an equalizer of ϕ and ψ in Set(O(X)).
ϕ ψ (4) Let (A, α) −→ (C, γ) ←− (B, β) be morphisms of O(X)-valued sets. Define a valuation δ on A × B by _ δ((a, b), (a0, b0)) = α(a, a0) ∧ β(b, b0) ∧ ϕ(a, c) ∧ ψ(b, c). c∈C Then (A × B, δ) equipped with the canonical projections is a pullback of that diagram in Set(O(X)).
9 The following notion of strict relation is crucial in handling subobjects of an O(X)-valued set. In the next section, it will enable us to define the “forcing values” of formulas. Definition 1.23. Let (A, α) be an O(X)-valued set. A strict relation on (A, α) is a function σ : A → O(X) such that
∀a, a0 ∈ A, σ(a) ∧ α(a, a0) ≤ σ(a0) and σ(a) ≤ α(a).
Note that a singleton is a strict relation on the same O(X)-valued set. Proposition 1.24. Let P(A, α) be the set of strict relations on (A, α) ordered by σ ≤ τ ⇐def==⇒. ∀a ∈ A, σ(a) ≤ τ(a).
Then, as ordered sets, P(A, α) is isomorphic to the poset Sub(A, α) of subob- jects of (A, α).
Proof. For a strict relation σ, we define a map ασ : A × A → O(X) by
0 0 0 0 ασ(a, a ) := σ(a) ∧ α(a, a ) = σ(a ) ∧ α(a, a ).
Then (A, ασ) is an O(X)-valued set and the identity map on A represents a monomorphism ισ :(A, ασ) (A, α) by Proposition 1.11. Conversely, for a monomorphism ϕ:(B, β) (A, α), we define a strict relation ρϕ : A → O(X) by _ ρϕ(a) := ϕ(b, a). b∈B Then we can check
• for any σ, ρ(ισ ) = σ.
• for any ϕ, ι(ρϕ) ' ϕ as subobjects of (A, α).
In fact, for an arbitrary morphism ϕ, a strict relation ρϕ can be defined as above, and the image factorization of ϕ is given by ϕ (B, β) (A, α)
ϕ ι(ρϕ) (A, α ) (ρϕ) .
Set(O(X)) is a topos and, in particular, a Heyting category ( [Elephant, §A1.4]). The associated operations on subobject lattices are as follows: Proposition 1.25. The operations on the frame P(A, α) are given by
1P(A,α)(a) = α(a), 0P(A,α)(a) = 0X , _ _ (σ ∧ τ)(a) = σ(a) ∧ τ(a), σi (a) = σi(a), i i (σ → τ)(a) = α(a) ∧ (σ(a) → τ(a)).
10 Proposition 1.26. Let ϕ:(B, β) → (A, α) be a morphism. Pulling back sub- objects along ϕ defines a frame homomorphism ϕ∗ : P(A, α) → P(B, β) such that, for σ ∈ P(A, α), _ ^ (ϕ∗σ)(b) = ϕ(b, a) ∧ σ(a) = β(b) ∧ [ϕ(b, a) → σ(a)]. a∈A a∈A Proof. For the last identity, note that ^ _ ^ β(b) ∧ [ϕ(b, a) → σ(a)] = ϕ(b, a) ∧ [ϕ(b, a0) → σ(a0)] a∈A a∈A a0∈A _ ≤ ϕ(b, a) ∧ [ϕ(b, a) → σ(a)] a∈A _ ≤ ϕ(b, a) ∧ σ(a). a∈A ∗ Proposition 1.27. In the same notations as above, ϕ has both a left adjoint ∃ϕ and a right adjoint ∀ϕ: for τ ∈ P(B, β), _ (∃ϕτ)(a) = ϕ(b, a) ∧ τ(b), b∈B ^ (∀ϕτ)(a) = α(a) ∧ [ϕ(b, a) → τ(b)]. b∈B Finally, we describe the higher-order structure of Set(O(X)). Proposition 1.28. Put δ(U, V ) = (U → V ) ∧ (V → U) for U, V ∈ O(X). Then (O(X), δ) is an O(X)-valued set. Let t:({∗}, >) → (O(X), δ) be the morphism defined by t(∗,U) = U. This yields a subobject classifier of Set(O(X)). Proof. Let χ:(A, α) → (O(X), δ) be a morphism. Since t corresponds to the strict relation idO(X) on (O(X), δ), the pullback of t along χ is given W by the strict relation σ(a) = U χ(a, U) ∧ U. Conversely, given a strict re- lation σ on (A, α), then σ itself represents a morphism χ(a, U) = α(a) ∧ (U ↔ σ(a)). These correspondences yield a bijection between P(A, α) and Hom((A, α), (O(X), δ)).
! (A, ασ) ({∗}, >) y ισ t χ (A, α) (O(X), δ) Similarly to O(X), P(A, α) is not only a frame but also an O(X)-valued set. Proposition 1.29. The power object of (A, α) is given by P(A, α) equipped with the valuation 2 ^ α¯(σ, τ) := σ(a) ↔ τ(a). a∈A 2Note that α¯ does not necessarily coincide with α˜ on A˜ in Proposition 1.20. Indeed, while α˜(σ, τ) ≤ α¯(σ, τ) holds for any σ, τ ∈ A˜, the converse inequality does not hold if σ = τ = σa for a non-global element a ∈ A.
11 Proof. We would like to establish the following bijection
Hom((B, β), P(A, α)) 'P((B, β) × (A, α)).
For a morphism ϕ:(B, β) → P(A, α), we have a strict relation θ on (B, β) × (A, α) such that _ θ(b, a) = ϕ(b, τ) ∧ τ(a). τ∈P(A,α) On the other hand, for any strict relation θ, we have a morphism ^ ϕ(b, τ) = β(b) ∧ θ(b, a) ↔ τ(a). a∈A
These correspondences are mutual inverses.
2 Sheaves of Structures and Heyting-Valued Struc- tures
2.1 Structures in a Topos We will be concerned with categorical semantics in the toposes Sh(X) and Set(O(X)). In this subsection, we take a glance at first-order categorical logic, which originates from [MR77]. The main reference is [Elephant], in particular, Chapter D1 in volume 2. For an overview, we refer the reader to Caramello’s account [Cara14], which is a preliminary version of the first two chapters of her book [TST]. Although many fragments of (possibly infinitary) first- order logic are considered in the context of categorical logic, we restrict our attention to single-sorted intuitionistic logic. Examples of other fragments include Horn, cartesian, regular, coherent, classical, and geometric logics. Definition 2.1. A (first-order) language L consists of the following data: • A set L-Func of function symbols. Each function symbol f ∈ L-Func is associated with a natural number n (the arity of f). If n = 0, f is called a constant (symbol). • A set L-Rel of relation symbols. Each relation symbol R ∈ L-Rel is associated with a natural number n (the arity of R). If n = 0, R is called an atomic proposition. L-terms and L-formulas are defined as usual. We need some conventions. Definition 2.2.
(1)A context is a finite list u ≡ u1, . . . , un of distinct variables. If n = 0, it is called the empty context and denoted by []. (2) We say that a context u is suitable for an L-formula ϕ when u contains all the free variables of ϕ. A formula ϕ equipped with a suitable context u is called a formula-in-context and indicated by ϕ(u). Similarly, terms- in-context can be defined.
12 For an L-formula-in-context ϕ(u, v), we abbreviate, e.g., ∃v1 · · · ∃vnϕ(u, v) V as ∃vϕ(u, v). We also abbreviate the L-formula i ui = vi as u = v, where u, v are assumed to have the same length. A formula is closed if it contains no free variables. We now give categorical semantics in an arbitrary elementary topos E, though we will only need the case when E is a Grothendieck topos. Definition 2.3. Let L be a language and E a topos. An L-structure M in E is given by specifying the following data: • We have the underlying object |M| ∈ E and denote the n-ary product by n 0 |M| . In particular, |M| is the terminal object 1E . • To an n-ary function symbol f, we assign a morphism f M : |M|n → |M|. • To an n-ary relation symbol R, we assign a subobject RM of |M|n. As usual, we will not distinguish M and its underlying object |M| in notation. Interpretations of L-terms and L-formulas are defined by using internal operations in E. Definition 2.4 (Interpretations of terms). Let M be an L-structure in a topos E. For an L-term-in-context t(u), we define the interpretation tM : Mn → M inductively. M n • If t is a variable ui, then t is the i-th product projection πi : M → M.
• If interpretations of L-terms ti(u) and s(v) are given, the term s(t1(u), . . . , tm(u)) is interpreted as the composite of the following morphisms:
htM,...,tMi M Mn −−−−−−−→M1 m m −−→Ms ,
M M M where ht1 , . . . , tm i is the morphism obtained from the morphisms ti m by using the universal property of the product M . Definition 2.5 (Interpretations of formulas). Let M be an L-structure in a topos E. For an L-formula-in-context ϕ(u), we define the interpretation u. ϕ M as a subobject of Mn inductively. (We drop the subscript M if no confusionJ K arises.) • If ϕ ≡ (s(u) = t(u)) where s, t are terms, then u. ϕ is defined to be the equalizer of J K sM Mn M tM .
• If ϕ ≡ R(t1(u), . . . , tm(u)), then u. ϕ is the pullback J K u.R(t1, . . . , tm) RM J K
htM, . . . , tMi n 1 m m M M .
13 • If ϕ ≡ >, ⊥, ψ ∧ θ, ψ ∨ θ, ψ → θ or ¬ψ, then u. ϕ is defined as expected by using the Heyting operations on Sub(MJn). K
• If ϕ ≡ ∃vψ(u, v), then u. ϕ is the image as in the following diagram: J K u, v. ψ Mn × M J K π
u. ∃vψ Mn J K , where π is the projection onto Mn.
n • If ϕ ≡ ∀vψ(u, v), then u. ϕ := ∀π u, v. ψ , where ∀π : Sub(M ×M) → n J K ∗ J K Sub(M ) is the right adjoint of π . In this paper, we will not consider the notions of models of a theory in a topos nor homomorphisms between structures.
2.2 Sheaves of Structures and Heyting-Valued Structures We now investigate the relationship between structures in Sh(X) and those in Set(O(X)). We first consider the case of sheaves on a topological space X. Let LH be the category of topological spaces and local homeomorphisms between them. Recall that the slice category LH/X is categorically equivalent to Sh(X). Comer [Com74], Ellerman [Ell74], and Macintyre [Maci73] used the following notion to obtain model-theoretic results: Definition 2.6. A sheaf of L-structures (on X) is a tuple