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
constructive mathematics, realizability, computability
propositions as types, proofs as programs, computational trinitarianism
basic constructions:
strong axioms
further
Proposed by Vladimir Voevodsky, Homotopy Type System (HTS) is a type theory with two equality types, an “exact” or “strict” one which satisfies a reflection rule?, and a “path” or “homotopical” one which does not. It also distinguishes between “fibrant” and “non-fibrant” types: the path type only eliminates into fibrant types.
Vladimir Voevodsky, A type system with two kinds of identity types (Feb. 2013) [pdf]
Vladimir Voevodsky, A simple type system (Jan 2013) [pdf, Implementation]
Let be the type of pointed fibrant types that are merely equal to , i.e.
Define by
Then we have , since by univalence we have a such that .
However, if , then by for strict equality, we have , where is the image of under the map . But all strict equalities are equal to reflexivity, so , and thus , a contradiction.
Therefore, the fibrant type contains two points such that , but (both provable inside the theory).
The universe of non-fibrant types is itself fibrant. Moreover, in standard models (such as the simplicial set model, and other model-category-theoretic models) it is contractible (in the usual sense of homotopy type theory, with respect to the identity type).
The reason that the universe of non-fibrant types (type families) is contractible is that if you have a type with two objects and an object such that for any two types there is a function whose strict fiber over is isomorphic to and strict fiber over is isomorphic to then the family of fibers of this function gives a map from to the universe of all types which, when applied to , gives an isomorphism in this universe from to . This is sufficient to create a path from a type which is isomorphic to empty to a type which is isomorphic to unit and therefore a path from 0 to 1 in nat.
It is not provably contractible in the theory, since we can take its coproduct with pretty much any other type and still have a universe of non-fibrant types. However, it is “almost” contractible by the following argument. Let and be arbitrary types, and let be as defined above. Define by and , and define (the universe of non-fibrant types) by
In other words, is the fiber of over , computed using strict equality. It’s straightforward to show that and . However, since and are both fibrant, we can apply to the path to obtain a . This is not the same as a path , of course.
It is natural to wonder whether we can have a “fibrant replacement” type former which makes non-fibrant types into fibrant ones. However, surprisingly, this is actually inconsistent, essentially because it cannot be made to respect substitution.
Suppose we had a type forming rule
with introduction rule
and elimination rule
Let and , and let be defined as above, so that and . Now consider
Since this is a fibrant type family over a fibrant type, we can transport the element along the known to obtain an element of . However, by , we can define a map by defining a map , which we have. Thus, we obtain an element of .
It does, however, seem to be possible to allow fibrant replacement of types in the empty context. In other words, we could have a rule
and so on. Of course, this does not fit very well in the usual framework of type theory.
Mike Shulman has argued that rather than assuming the natural numbers type to be fibrant and yet able to eliminate into non-fibrant types, there should be two natural numbers types: a fibrant one which can only eliminate into fibrant types, and a non-fibrant one which can eliminate into all types. The same applies to other positive inductive types such as coproducts (even coproducts of fibrant types) and the empty type.
This is because while these types in simplicial sets happens to be fibrant, that is not the case in other categorical models. It may also pose similar problems for constructing new models of HTS out of old ones.
Vladimir Voevodsky claims that HTS allows a definition of semisimplicial types and other infinite objects, in which diagrams of types commute strictly using the exact equality. See the comments in the file:
Voevodsky has conjectured that any fibrant type definable in HTS is equivalent to one definable in MLTT, i.e. without using the strict equality.
The model invariance problem seems more likely to fail for HTS than for theories that do not include strict equality. In other words, distinct model categories presenting the same -topos might have different “internal languages” according to HTS. This can obviously happen if by “internal language” we include the non-fibrant types, but if we restrict to the fibrant types the answer is unclear. It may be related to Voevodsky’s conjecture above.
The following argument, however, shows that model invariance (even for fibrant types) definitely fails if we include fibrant replacement (of types in the empty context), and split inductive types into fibrant and non-fibrant versions (or simply don’t include inductive types in the system).
Let be the poset with four objects such that , , and are the only nonidentity relations; thus an -diagram is four objects , , , with maps , , and . Since is an inverse category, there is a Reedy model structure on in which the cofibrations are levelwise monomorphisms, the weak equivalences are levelwise weak equivalences, and the fibrant objects are diagrams such that , , and are fibrant and the induced map is a fibration. Moreover, this model has univalent universes induced from those of : if is a univalent universe in , then is a univalent universe in , where , and the fiber of over is the type .
Now we can localize this Reedy model structure at the maps , , and , where , , and denote the corresponding representables. A local object is a fibrant object such that , , and are contractible, and the local weak equivalences are just the maps that induce a weak equivalence on -components. In particular, the homotopy category of the localized model structure is equivalent to that of . Moreover, locality and localization (of fibrant objects) can be represented internally as a subobject and an endomorphism, respectively, of the universe . Because localization is left exact, the subuniverse of local objects is itself local, and thus provides a univalent universe for the localized model structure.
In conclusion, we have two model structures, one on and one on , which both model homotopy type theory with univalent universes, and have equivalent homotopy categories. Both model HTS as well, although their inductive types come in fibrant and non-fibrant versions.
Now, inside HTS, define a (non-fibrant) type to be a “strict hprop” if . Write for the subtype of the universe of non-fibrant types determined by the strict hprops. Consider the following type
Call this type KP (for Kreisel-Putnam). Note that “false” and “or” for strict hprops can be defined impredicatively by quantification over .
For semantics of HTS in a model category whose underlying category is a topos, the type will be interpreted by the subobject classifier of that topos (independently of what the model structure might look like). Thus, KP will be interpreted by the assertion, in the internal language of that topos (again independently of the model structure), that the Kreisel-Putnam axiom holds.
Now as pointed out by Francois Dorais, the Kreisel-Putnam axiom holds in the internal logic of (the 1-topos) ; thus the interpretation of KP there is globally inhabited. However, the Kreisel-Putnam axiom does not hold in the internal logic of (the 1-topos) . For this it suffices to exhibit subterminal objects such that is not contained in . Let , , and . Then , so that , but and , so that . Thus, the interpretation of KP in is not globally inhabited.
It follows that is a fibrant type which is inhabited in the model in , but not in the model in .
Last revised on July 13, 2023 at 14:42:35. See the history of this page for a list of all contributions to it.