first-order hyperdoctrine

**natural deduction** metalanguage, practical foundations

**type theory** (dependent, intensional, observational type theory, homotopy type theory)

**computational trinitarianism** = **propositions as types** +**programs as proofs** +**relation type theory/category theory**

A first-order hyperdoctrine is a hyperdoctrine with respect to lattices that are Heyting algebras.

A **first-order hyperdoctrine** consists of a category with finite products $C_T$, along with a functor

$P \colon C_T^{op} \to HeytAlg_{AdjCyl}$

(where $HeytAlg_{AdjCyl}$ is the subcategory of the category of Heyting algebras containing only those morphisms with left and right adjoints) such that the following Beck-Chevalley condition is satisfied: for every object $A$, the left adjoints to $P(\pi)$ for the projections $\pi : A \times - \to -$ comprise a natural transformation from $P(A \times -)$ to $P(-)$, and so do the right adjoints. This expresses the Beck-Chevalley condition for pullback squares of the form

$\array{
A \times B & \stackrel{\pi_B}{\to} & B \\
^\mathllap{1_A \times f} \downarrow & & \downarrow^\mathrlap{f} \\
A \times C & \underset{\pi_C}{\to} & C
}$

An element of $P(A)$ is often called a *predicate over $A$* (with respect to the hyperdoctrine).

For any first-order hyperdoctrine, an *equality predicate* can be defined for each type $A \in Ob(C_T)$ as

$\exists(\delta_A)(\top_{P(A)}) \in P(A \times A)$

where for any morphism $f$ in $C_T$, we use $\exists (f)$ to denote the left adjoint of $P(f)$, and $\top_H$ denotes the top element in a Heyting algebra $H$. However, to get good properties for equality, we need to assume a little more. A **first-order hyperdoctrine with equality** is a first-order hyperdoctrine such that the Beck-Chevalley condition is also satisfied for pullback diagrams of the form

$\array{
A & \stackrel{\delta_A}{\to} & A \times A \\
^\mathllap{\delta_A} \downarrow & & \downarrow^\mathrlap{1_A \times \delta_A} \\
A \times A & \underset{\delta_A \times 1_A}{\to} & A \times A \times A
}$

This says $P(1_A \times \delta_A) \circ \exists (\delta_A \times 1_A) = \exists (\delta_A) \circ P(\delta_A)$; this equation is also known as a Frobenius law. By taking right adjoints, a similar equation holds for $\forall (f)$ in place of $\exists (f)$, where $P(f) \dashv \forall (f)$.

With this, we have enough structure to interpret multi-sorted first-order intuitionistic logic with equality, taking the objects of $C_T$ to be sorts and its morphisms to be terms, $P$ to assign to each sort the Lindenbaum algebra? of predicates upon that sort and to each term the operation of substitution of that term into predicates, and the left and right adjoints to $P$ upon projections to provide existential and universal quantification, respectively, with the existence of the further adjoints providing the ability to interpret equality, and the Beck-Chevalley condition ensuring that quantification commutes appropriately with substitution (just as the propositional connectives do).

There are, of course, many variants on this, corresponding straightforwardly to modifications of the kind of logic one wishes to represent. For instance, to represent specifically classical logic, one should use Boolean algebras instead of Heyting algebras. To represent first-order logic without equality, one should no longer require left and right adjoints to every morphism in the range of $P$, but rather only those given by the natural transformations yielding the quantifiers (i.e., only requiring adjoints to substitution along projections). Various higher-order constructs can be added by adding new ways of forming objects in $C_T$ (e.g., adding cartesian closedness). Etc.

One can even extend this into the realm not just of provability, but furthermore of proof theory, by taking the objects in the codomain of $P$ to be categories rather than mere preorders (e.g., by using bicartesian closed categories rather than Heyting algebras); in this case, the objects in a category $P(\sigma)$ would still represent predicates on the sort $\sigma$, but the morphisms in $P(\sigma)$ would represent proofs (rather than mere provability) of entailments between these predicates, with the possibility that not all such proofs would be equal.

Revised on October 25, 2012 11:18:26
by Urs Schreiber
(131.174.189.236)