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
equality (definitional, propositional, computational, judgemental, extensional, intensional, decidable)
identity type, equivalence of types, definitional isomorphism
isomorphism, weak equivalence, homotopy equivalence, weak homotopy equivalence, equivalence in an (∞,1)-category
Examples.
(in category theory/type theory/computer science)
of all homotopy types
of (-1)-truncated types/h-propositions
In intensional type theory and homotopy type theory, there are typically two ways to construct a “type of small types”. One way is by Russell universes, where an element $A:U$ of a type universe $U$ is literally a type $A \; \mathrm{type}$. The alternative is by Tarski universes, where elements $A:U$ are not literally types, but rather the universe type $U$ comes with the additional structure of a type family $T$, such that the dependent type $T(A)$ represents the actual type. Usually universes are defined in intensional type theory and homotopy type theory using derivations and inference rules, such as the type reflection rule
for Tarski universes. Furthermore, universes are usually defined as being closed under all type formers inside the type theory, regardless of the type theory.
However, this approach of defining Tarski universes means that the notion of “Tarski universe” differs from type theory to type theory. By this approach, a Tarski universe in the bare intensional Martin-Löf type theory, which has identity types, dependent product types, dependent sum types, an empty type, a unit type, a booleans type, and a natural numbers type, would not be considered a Tarski universe in homotopy type theory, since it is missing important type formers present in homotopy type theory, such as internal pushout types, a torus type?, propositional truncations, W-types, and localizations. On the other hand, one could consider an even more spartan type theory which doesn’t even have the unit type, let alone the empty type, the natural numbers type, and the booleans type, whose Tarski universes are thus consisting of only internal identity types, dependent product types, dependent sum types.
All three notions of Tarski universe described above are in fact definable in all three versions of intensional type theory mentioned above, but Tarski universes are usually described as having, apart from addiitonal internal Tarski universes, exactly the internal type formers that the external type theory has. One usually doesn’t consider Tarski universes with an internal James construction type in bare intensional Martin-Löf type theory, even though a Tarski universe closed under James constructions is definable in Martin-Löf type theory. On the other hand, in homotopy type theory the name “Tarski universe” isn’t usually used for a universe only closed under identity types, dependent product types, dependent sum types.
From a global perspective, however, all of these types should be considered Tarski universes, since they model some notion of type of small types, even though the small types in the universe do not behave as the external types in the type theory do. And that is how we approach Tarski universes in this article: Tarski universes are a very general notion encompassing all of the above notions of Tarski universe and more, and one could describe more specific versions of Tarski universes by adding additional structure and axioms to Tarski universes, in the same way one adds additional structure and axioms to an abelian group to get a ring, a commutative ring, a $R$-module, a $R$-algebra, and an $R$-commutative algebra. We could then study in general Tarski universes and their properties as we do for abelian groups and rings. This requires a shift in the point of view of what Tarski universes are: Tarski universes are actual mathematical structure in intensional type theory, rather than being something in the foundations of mathematics to merely address size? issues or do mathematics in.
A Tarski universe or universe à la Tarski is simply a type $U$ with a type family $T$ whose dependent types $T(A)$ are indexed by elements $A:U$. The elements $A:U$ are usually called $U$-small types, or small types for short, and $T(A)$ is the type of elements of $A$.
A Tarski universe or universe à la Tarski is a type $U$ of all $U$-small types with a type $T'$ of all elements, and a function $\mathrm{typeOf}:T' \to U$ which gets the associated type for every element $a:T'$.
From one direction, the individual dependent types $T(A)$ could be defined as the fiber type of $\mathrm{typeOf}$ at $A$
From the other direction, the entire type of elements $T'$ is just the dependent sum type of all the type reflections of small types
We shall be using the first definition throughout this article, but everything could be translated into the second definition of a Tarski universe.
The definitions above are already enough to build an internal model of dependent type theory inside the Tarski universe:
There are two notions of equivalence in a Tarski universe $(U, T)$, equality $A =_U B$ and equivalence $T(A) \simeq T(B)$. By the properties of the identity type, there is a canonical transport dependent function
A Tarski universe is a univalent Tarski universe if for all elements $A:U$ and $B:U$ the function $\mathrm{transport}^T(A, B)$ is an equivalence of types
This is the extensionality principle for any Tarski universe $(U, T)$.
Equivalently, by the fundamental theorem of identity types, a Tarski universe is a univalent Tarski universe if it comes with a dependent function
and a dependent function
When the Tarski universe is closed under identity types, dependent sum types, and dependent product types (see below for more on closure of Tarski universes under type formers), one is able to internalize the type of equivalences in the universe as $A \simeq_U B$. It can be proven that $T(A \simeq_U B)$ is equivalent to $T(A) \simeq T(B)$, and that the definition of univalence using transport is equivalent to the usual definition given by the internal type of equivalences and the canonical function $\mathrm{idtoequiv}$. See at Univalence axiom#For Tarski universes for more details on this.
Let $(U, T)$ be a Tarski universe. Then one could construct a univalent Tarski universe as the higher inductive type $(U', T')$ generated by
A univalent Tarski universe $(U, T)$ is regular if it is closed under dependent sum types: namely, for all $U$-small types $A:U$ and $U$-small type families $B:T(A) \to U$, the dependent sum type $\sum_{x:T(A)} T(B(x))$ is essentially $U$-small.
Regular univalent Tarski universes are the type theoretic equivalent of regular cardinals in set theory.
A univalent Tarski universe $(U, T)$ is product-regular if it is closed under dependent product types: namely, for all $U$-small types $A:U$ and $U$-small type families $B:T(A) \to U$, the dependent product type $\sum_{x:T(A)} T(B(x))$ is essentially $U$-small.
Product-regular univalent Tarski universes are the type theoretic equivalent of product-regular cardinals in set theory.
Since product-regularity implies that function types between $U$-small types are essentially $U$-small, and regularity implies that subtypes of essentially $U$-small types are essentially $U$-small, this implies that in a regular and product-regular universe, equivalence types between $U$-small types are essentially $U$-small, because equivalence types are subtypes of function types. In combination with univalence, regularity and product-regularity implies that the identity types between $U$-small types are also essentially $U$-small.
A univalent Tarski universe $(U, T)$ is closed under identity types if for all $U$-small types $A:U$ and elements $x:T(A)$ and $y:T(A)$, the identity type $x =_{T(A)} y$ is essentially $U$-small.
Regularity and closure under identity types implies that the universe is closed under pullbacks.
A univalent Tarski universe satisfies the axiom of singletons if it is closed under the (weak) unit type: if the unit type $\mathbb{1}$ is essentially $U$-small. The axiom of singletons comes from the formation rule, the introduction rules, and the dependent universal property of the natural numbers, which are represented by the following elements:
formation rules: $\mathbb{1}:U$
introduction rules: $0:T(\mathbb{1})$
and either an universal property or a dependent universal property
universal property:
dependent universal property:
where $\exists!x:A.B(x)$ is the uniqueness quantifier for the type family $x:A \vdash B(x)$.
By using dependent sum types, these can be combined into a single element of the following types
Alternatively, since the notion of truth could be defined using the type of $U$-small contractible types $\sum_{A:U} \mathrm{isContr}(T(A))$, the axiom of singletons could be represented as a resizing axiom, similar to propositional resizing.
A univalent Tarski universe $(U, T)$ satisfies the axiom of singletons if it is closed under the type of all $U$-small contractible types: namely, the dependent sum type $\sum_{A:U} \mathrm{isContr}(T(A))$ is essentially $U$-small:
A univalent Tarski universe satisfies the axiom of empty type if it is closed under the (weak) empty type: if the empty type $\emptyset$ is essentially $U$-small. The axiom of empty type comes from the formation rule and the dependent universal property of the empty type, which are represented by the following elements:
and either an universal property or a dependent universal property
By using dependent sum types, these can be combined into a single element of the following types
Alternatively, since falsehood or the empty proposition could be defined using the type of $U$-small propositions $\sum_{A:U} \mathrm{isProp}(T(A))$, the axiom of empty set could be represented as a resizing axiom, similar to propositional resizing.
A univalent Tarski universe $(U, T)$ satisfies the axiom of empty type if it is closed under the empty proposition or falsehood: namely, the dependent product type $\prod_{A:\mathrm{Prop}}_U T(A)$ is essentially $U$-small:
where
Product regularity and the axiom of empty type imply the axiom of singletons, because the dependent universal property of the empty set states that for every type family $C:T(\mathbb{0}) \to U$ the dependent function type $\prod_{x:T(\mathbb{0})} T(C(x))$ is a singleton, and product regularity implies that $\prod_{x:T(\mathbb{0})} T(C(x))$ is essentially $U$-small.
A univalent Tarski universe $(U, T)$ satisfies propositional resizing if it is closed under the type of all $U$-small propositions: namely, the dependent sum type $\sum_{A:U} \mathrm{isProp}(T(A))$ is essentially $U$-small:
This is a version internal to the universe $U$ of having a type of all propositions in the type theory itself. While propositional resizing implies that the universe is impredicative, it does not imply that the type theory as a whole is impredicative; the latter requires an actual type of all propositions in the type theory.
Product regularity and propositional resizing imply the axiom of empty set, because propositional resizing implies that $\mathrm{Prop}_U$ is essentially $U$-small, and then product regularity implies that $\prod_{A:\mathrm{Prop}_U} T(A)$ is essentially $U$-small.
A univalent Tarski universe satisfies the axiom of infinity if it is closed under the (weak) natural numbers type: if the natural numbers type $\mathbb{N}$ is essentially $U$-small. The axiom of infinity comes from the formation rule, the introduction rules, and the dependent universal property of the natural numbers, which are represented by the following elements:
formation rules: $\mathbb{N}:U$
introduction rules: $0:T(\mathbb{N})$ and $s:T(\mathbb{N}) \to T(\mathbb{N})$
and either an universal property or a dependent universal property
universal property:
dependent universal property:
where $\exists!x:A.B(x)$ is the uniqueness quantifier for the type family $x:A \vdash B(x)$.
By using dependent sum types, these can be combined into a single element of the following types:
or
Alternatively, since the type of all $U$-small finite types and set truncations could be defined using the type of $U$-small propositions $\sum_{A:U} \mathrm{isProp}(T(A))$, the axiom of infinity could be represented as a resizing axiom, similar to propositional resizing.
A univalent Tarski universe $(U, T)$ satisfies the axiom of infinity if it is closed under the set truncation of the type of all $U$-small finite types: namely, the type $\left[\sum_{A:U} \mathrm{isFinite}(T(A))\right]_0$ is essentially $U$-small:
where
and
The membership relation and the subtype operations used above are defined in the nLab article on subtypes.
A univalent Tarski universe satisfies the axiom of replacement if for every essentially $U$-small type $A$, every locally $U$-small type $B$, and every function $f:A \to B$, the image of $f$, $\mathrm{im}(f)$, is essentially $U$-small.
Equivalently, given types $A$ and $B$, there is a function
A univalent Tarski universe $(U, T)$ satisfies excluded middle if all $U$-small propositions are decidable propositions:
Suppose that there is a univalent Tarski universe closed under dependent product types, dependent sum types, and identity types, and satisfying the axiom of infinity and excluded middle. Then the limited principle of omniscience for natural numbers $\mathrm{LPO}_\mathbb{N}$ holds in the type theory itself.
The type of all $U$-small propositions $\mathrm{Prop}_U$ is a $\sigma$-frame, and thus the homotopy-initial $\sigma$-frame $\Sigma$ is a sub-$\sigma$-frame of $\mathrm{Prop}_U$, with the following embedding of types
The first embedding is a unique distributive lattice homomorphism, since the boolean domain is the homotopy-initial distributive lattice, and the second embedding is a unique $\sigma$-frame homomorphism, by definition of homotopy-initial $\sigma$-frame. Excluded middle for $U$ implies that $\mathrm{Prop}_U$ is equivalent to the boolean domain and to $\Sigma$, implying that the boolean domain is the initial $\sigma$-frame, which then implies $\mathrm{LPO}_\mathbb{N}$.
A univalent Tarski universe satisfies the axiom of choice if for every $U$-small type $A:U$ and every $U$-small type family $B:T(A) \to U$, and for every every $U$-small type family $C:\prod_{x:T(A)} T(B(x)) \to U$, if $T(A)$ is a set and each $T(B(x))$ is a set for all $x:T(A)$, and each $T(C(x, y))$ is a proposition for all $x:T(A)$ and $y:T(B(x))$, then if for all $x:A$ there merely exists a $y:T(B(x))$ such that $T(C(x, y))$, then there merely exists a dependent function $g:\prod_{x:T(A)} T(B(x))$ such that for all $x:T(A)$, $T(C(x, g(x)))$.
A choice operator on a type is a function from its propositional truncation to the type itself, and represents the concept that if there exists an element of the set (i.e. the propositional truncation has an element), then the set itself has an element chosen by the choice operator. A Tarski universe satisfies the type-theoretic axiom of existence if every $U$-small type in the Tarski universe has a choice operator, represented by the following dependent function
$U$ satisfying the type-theoretic axiom of existence implies that $U$ satisfies axiom K or UIP. If $U$ is also univalent, then it is an h-groupoid.
There are many ways of defining type formers internally in a universe:
by an equivalence or definitional equality with an existing global type former for the entire type theory.
by translating the rules for the type former into internal structure on the universe - the formation rules, the introduction rules, and the dependent universal property of the structure.
by using the universal properties corresponding to the categorical semantics of the types
Using a definitional equality with an existing global type former for each type former results in a strict Tarski universe, while using equivalences for each type former results in a weak Tarski universe. There are also various Tarski universes where some type formers are strictly closed, and some type formers are only weakly closed, resulting in Tarski universes which are intermediate between strict Tarski universes and weak Tarski universes.
A Tarski universe $(U, T)$ has all propositions if given a type $A$ with a family of identities $x:A, y:A \vdash \tau_{-1}(x, y):x =_A y$, the universe $U$ comes with the structure of an element $A_U:U$ and an equivalence $\delta_A:T(A_U) \simeq A$.
The rules for this condition on Tarski universes is as follows:
Given a Tarski universe which has all propositions, the type $\sum_{A:U} \mathrm{isProp}(T(A))$ is the type of all propositions. This is a form of strong impredicativity in the type theory, as given a type $A$, the type $A \to \sum_{A:U} \mathrm{isProp}(T(A))$ is the power set on $A$.
Let $(U, T_U)$ be a Tarski universe. A Tarski universe $(V, T_V)$ is a Tarski subuniverse of $(U, T_U)$ if it comes with an embedding $t:V \hookrightarrow U$ and a dependent function
This is equivalent to a family of Tarski universes $(U(i), T(i))$ indexed by elements $i:\mathbb{2}$ of the interval preorder $(\mathbb{2}, \leq_\mathbb{2})$, which comes with
a dependent function
a dependent function
indicating that lifting each small type results in a type equivalent to the original small type.
A modal operator on a Tarski universe $(U, T)$ is just an endofunction $L:U \to U$. Given a Tarski universe, the modal unit family is a family of functions
and the comodal counit family is a family of functions
Given a modal operator $L:U \to U$, $\eta(L)$ is called the modal unit of $L$ and $\epsilon(L)$ is called the comodal counit of $L$. A small type $A:U$ is $L$-modal if the function
is an equivalence of types, and $L$-comodal if the function
is an equivalence of types.
A modality is…
…and a comodality is…
(Section under construction see reflective subuniverse for the Russell universe variant for the time being)
A Tarski universe $(V, T_V)$ inside of a Tarski universe $(U, T_U)$ consists of an element $V:U$ and a function $T_V:T_U(V) \to U$.
Given a Tarski universe $(U, T_U)$, a Tarski universe $(V, T_V)$ is essentially $U$-small if it comes with
an element $V':U$
a function $T_{V'}:T_U(V') \to U$
an equivalence
a dependent function
Equivalently, the entire structure could be regarded as a family of Tarski universes $(U(a), T(a))$ indexed by elements $a:\mathbb{2}$ of the interval preorder $(\mathbb{2}, \leq_\mathbb{2})$, which comes with
a dependent function
a dependent function
an dependent function
a dependent function
Let $P$ be a preorder. Then, a $P$-indexed hierarchy of Tarski universes is a type family $U$ such that for all indices $a:P$, there is a Tarski universe $(U(a), T(a))$, such that for all indices $a:P$ and $b:P$ and witnesses $p:a \leq_P b$, the Tarski universe $(U(a), T(a))$ is an essentially $U(b)$-small Tarski subuniverse of the Tarski universe $(U(b), T(b))$.
Expanded out, the family of Tarski universes $(U(a), T(a))$ indexed by elements $a:P$ additionally come with
a dependent function
a dependent function
an dependent function
a dependent function
a dependent function
a dependent function
Usually, hierarchies of Tarski universes are indexed by the type of natural numbers. A type theory may have multiple hierarchies of type universes.
Any cumulative hierarchy $V$ with a membership relation $\in$ is a Tarski universe with universal type family
This is not a univalent universe if the von Neumann universe satisfies the set-theoretic axiom of extensionality, which states that
However, the axiom of extensionality implies that $V$ is an h-sets, since the membership relation is an h-proposition, which in turn implies that each $\mathrm{El}(x)$ is an h-set.
Examples include ZFC and its large cardinal, choice-less, constructive, predicative, and non-foundational/anti-foundational variants, such as ZF, Zermelo set theory, CZF, IZF, Mac Lane set theory?, Kripke–Platek set theory?, Tarski-Grothendieck set theory, New Foundations, Mostowski set theory, and so forth.
Any well pointed category $\mathcal{E}$ with hom-set type family $\mathrm{hom}(x, y)$ is a Tarski universe with universal type family
where $1:\mathcal{E}$ is the terminal object and indecomposable projective separator of the category $\mathcal{E}$.
Examples of this include ETCS.
Every concrete category and every concrete dagger category? is a Tarski universe, as given a concrete category or concrete dagger category $C$ by definition for each object $A:Ob(C)$ there is an h-set $El(A)$. This includes models of Mike Shulman‘s SEAR, which is a concrete power allegory, as well as ETCS with elements, which is a concrete elementary topos.
The empty type, unit type, boolean domain, and FinSet are all regular univalent Tarski universes. The types of propositions in a type theory are univalent Tarski universes: they model a dependent type theoretic model of propositional logic with function types, product types, disjunction higher inductive types, empty type, and unit type.
For universes as a replacement for type judgments, see
Regarding the general notion of types of small types, see
Other mathematical structures and their univalent counterparts:
Historical review of Alfred Tarski‘s original notion in set theory:
The terminology “universe à la Tarski” — and now in the context of type universes for Martin-Löf dependent type theory — is due to:
Further discussion:
See also:
Zhaohui Luo, Notes on Universes in Type Theory, 2012 (pdf)
Cesare Gallozzi, Constructive Set Theory from a Weak Tarski Universe, MSc thesis (2014) (arXiv:1411.5591)
Egbert Rijke, Introduction to Homotopy Type Theory, Cambridge Studies in Advanced Mathematics, Cambridge University Press (arXiv:2212.11082)
Homotopy Type Theory: Univalent Foundations of Mathematics, The Univalent Foundations Program, Institute for Advanced Study, 2013. (web, pdf)
Egbert Rijke, Mike Shulman, Bas Spitters, Modalities in homotopy type theory (arXiv:1706.07526)
Last revised on July 26, 2024 at 21:44:01. See the history of this page for a list of all contributions to it.