open import 1Lab.Path.Groupoid open import 1Lab.Type.Sigma open import 1Lab.HLevel open import 1Lab.Equiv open import 1Lab.Path open import 1Lab.Type module 1Lab.Univalence where
Univalenceπ
In Homotopy Type Theory, univalence is the principle stating that equivalent types can be identified. When the book first came out, there was no widely-accepted computational interpretation of this principle, so it was added to the theory as an axiom: the univalence axiom.
Precisely, the axiom as presented in the book consists of the following data (right under remark Β§2.10.4):
A map which turns equivalences into identifications of types. This map is called ua.
A rule for eliminating identifications of types, by turning them into equivalences: pathβequiv
The propositional computation rule, stating that transport along
ua(f)
is identical to applyingf
: uaΞ².
In the book, there is an extra postulated datum asserting that ua is an inverse to pathβequiv. This datum does not have a name in this development, because itβs proved in-line in the construction of the term univalence.
The point of cubical type theory is to give these terms constructive interpretations, i.e., make them definable in the theory, in terms of constructions that have computational behaviour. Letβs see how this is done.
Glueπ
To even state univalence, we first have to make sure that the concept of βpaths between typesβ makes sense in the first place. In βBook HoTTβ, paths between types are a well-formed concept because the path type is uniformly inductively defined for everything β including universes. This is not the case in Cubical type theory, where for paths in to be well-behaved, must be fibrant.
Since thereβs no obvious choice for how to interpret hcomp in Type, a fine
solution is to make hcomp its own type former. This is the
approach taken by some Cubical type theories in the RedPRL school. Univalence in those type
theories is then achieved by adding a type former, called
V
, which turns an equivalence into a path.
In CCHM β and therefore Cubical Agda β a different approach is taken, which combines proving univalence with defining a fibrancy structure for the universe. The core idea is to define a new type former, Glue, which βgluesβ a partial type, along an equivalence, to a total type.
Glue : (A : Type β) β {Ο : I} β (Te : Partial Ο (Ξ£[ T β Type β' ] (T β A))) β Type β'
The public interface of Glue demands a type , called the base type, a formula , and a partial type which is equivalent to . Since the equivalence is defined inside the partial element, it can also (potentially) vary over the interval, so in reality we have a family of partial types and a family of partial equivalences .
In the specific case where we set
,
we can illustrate Glue A (T, f)
as the dashed line in the
square diagram below. The conceptual idea is that by βgluingβ
onto a totally defined type, we get a type which extends
.
For Glue to extend , we add a computation rule which could be called a boundary condition, since it specifies how Glue behaves on the boundaries of cubes. Concisely, when , we have that Glue evaluates to the partial type. This is exactly what it means for Glue to extend !
module _ {A B : Type} {e : A β B} where private Glue-boundary : Glue B {i1} (Ξ» x β A , e) β‘ A Glue-boundary i = A
Furthermore, since we can turn any family of paths into a family of equivalences, we can use the Glue construct to implement something with precisely the same interface as hcomp for Type:
glue-hfill : β {β} Ο (u : β i β Partial (Ο β¨ ~ i) (Type β)) β β i β Type β [ _ β¦ (Ξ» { (i = i0) β u i0 1=1 ; (Ο = i1) β u i 1=1 }) ]
The type of glue-hfill is the same as that of hfill, but the type is stated much more
verbosely β so that we may define it without previous reference to a
hcomp analogue. Like hfill, glue-hfill extends an open box of types to a
totally-defined cube. The type of glue-hfill
expresses this in terms of extensions: We have a path (thatβs the
β i β
binder) of Types which
agrees with outS u0
on the left endpoint, and with
u
everywhere.
glue-hfill Ο u i = inS ( Glue (u i0 1=1) {Ο = Ο β¨ ~ i} Ξ» { (Ο = i1) β u i 1=1 , lineβequiv (Ξ» j β u (i β§ ~ j) 1=1) ; (i = i0) β u i0 1=1 , lineβequiv (Ξ» i β u i0 1=1) })
In the case for
,
we must glue
onto itself using the identity equivalence. This guarantees that the
boundary of the stated type for glue-hfill
is satisfied. However, since different faces of partial elements must
agree where they are defined, we can not use the identity equivalence
directly, since lineβequiv refl
is not definitionally the
identity equivalence.
When , hence where is defined, we glue the endpoint onto using the equivalence generated by the path provided by itself! Itβs a family of partial paths, after all, and that can be turned into a family of partial equivalences.
Using hcomp-unique and glue-hfill together, we get a internal characterisation of the fibrancy structure of the universe. While hcomp-unique may appear surprising, it is essentially a generalisation of the uniqueness of path compositions: Any open box has a contractible space of fillers.
hcompβ‘Glue : β {β} {Ο} (u : β i β Partial (Ο β¨ ~ i) (Type β)) β hcomp Ο u β‘ Glue (u i0 1=1) (Ξ» { (Ο = i1) β u i1 1=1 , lineβequiv (Ξ» j β u (~ j) 1=1) }) hcompβ‘Glue {Ο = Ο} u = hcomp-unique Ο u (glue-hfill Ο u)
Paths from Glueπ
Since Glue generalises hcomp by allowing a partial equivalence as its βtubeβ, rather than a partial path, it allows us to turn any equivalence into a path, using a sort of βtrickβ: We consider the line with endpoints and as an open cube to be filled. A filler for this line is exactly a path . Since Glue fills open boxes of types using equivalences, this path exists!
ua : {A B : Type β} β A β B β A β‘ B ua {A = A} {B} eqv i = Glue B Ξ» { (i = i0) β A , eqv ; (i = i1) β B , _ , id-equiv }
Semantically, the explanation of ua as
completing a partial line is sufficient. But we can also ask ourselves:
Why does this definition go through, syntactically? Because of
the boundary condition for Glue: when i = i0
, the whole
thing evaluates to A
, meaning that the left endpoint of the
path is correct. The same thing happens with the right endpoint.
The action of transporting along
ua(f)
can be described by chasing an element around the
diagram that illustrates Glue in the
case, specialising to ua. Keep in
mind that, since the right face of the diagram βpoints in the wrong
directionβ, it must be inverted. However, the inverse of the identity
equivalence is the identity equivalence, so nothing changes (for this
example).
- The action that corresponds to the left face of the diagram is to
apply the underlying function of
f
. This contributes thef .fst x
part of the uaΞ² term below.
- For the bottom face, we have a path rather than an equivalence, so
we must transport along it. In this case, the path
is the reflexivity on
B
, but in a more general Glue construction, it might be a non-trivial path.
To compensate for this extra transport, we use coe1βi, which
connects f .fst x
and
transport (Ξ» i β B) (f .fst x)
.
- Finally, we apply the inverse of the identity equivalence, corresponding to the right face in the diagram. This immediately computes away, and thus contributes nothing to the uaΞ² path.
uaΞ² : {A B : Type β} (f : A β B) (x : A) β transport (ua f) x β‘ f .fst x uaΞ² {A = A} {B} f x i = coe1βi (Ξ» _ β B) i (f .fst x)
Since ua is a map that turns equivalences into paths, we can compose it with a function that turns isomorphisms into equivalences to get the map IsoβPath.
IsoβPath : {A B : Type β} β Iso A B β A β‘ B IsoβPath (f , iiso) = ua (f , is-isoβis-equiv iiso)
Paths over uaπ
The introduction and elimination forms for Glue can be
specialised to the case of ua, leading
to the definitions of ua-glue
and ua-unglue below. Their types are written in
terms of interval variables and extensions, rather than using
Path
s, because these typings make
the structure of Glue more explicit.
The first, ua-unglue, tells us that if we have some
x : ua e i
(varying over an interval variable
i
), then we have an element of B
which agrees
with e .fst x
on the left and with x
on the
right.
ua-unglue : β {A B : Type β} (e : A β B) (i : I) (x : ua e i) β B ua-unglue e i x = unglue (i β¨ ~ i) x
We can factor the interval variable out, to get a type in terms of
PathP, leading to an explanation of
ua-unglue
without mentioning extensions: A path
x β‘ y
over ua e
induces a path
e .fst x β‘ y
.
ua-pathpβpath : β {A B : Type β} (e : A β B) {x : A} {y : B} β PathP (Ξ» i β ua e i) x y β e .fst x β‘ y ua-pathpβpath e p i = ua-unglue e i (p i)
In the other direction, we have ua-glue,
which expresses that a path e .fst x β‘ y
implies that
x β‘ y
over ua e
. For the type of ua-glue, suppose that we have a partial
element
defined on the left endpoint of the interval, together with an extension
of
where
is defined. What ua-glue expresses is that we can complete
this to a path in
,
which agrees with
and
where these are defined.
ua-glue : β {A B : Type β} (e : A β B) (i : I) (x : Partial (~ i) A) (y : B [ _ β¦ (Ξ» { (i = i0) β e .fst (x 1=1) }) ]) β ua e i [ _ β¦ (Ξ» { (i = i0) β x 1=1 ; (i = i1) β outS y }) ] ua-glue e i x y = inS (prim^glue {Ο = i β¨ ~ i} (Ξ» { (i = i0) β x 1=1 ; (i = i1) β outS y }) (outS y))
Observe that, since is partially in the image of , this essentially constrains to be a βpartial preimageβ of under the equivalence . Factoring in the type of the interval, we get the promised map between dependent paths over ua and paths in B.
pathβua-pathp : β {A B : Type β} (e : A β B) {x : A} {y : B} β e .fst x β‘ y β PathP (Ξ» i β ua e i) x y pathβua-pathp e {x = x} p i = outS (ua-glue e i (Ξ» { (i = i0) β x }) (inS (p i)))
The βpathp to pathβ versions of the above lemmas are definitionally
inverses, so they provide a characterisation of
PathP (ua f)
in terms of non-dependent paths.
ua-pathpβpath : β {A B : Type β} (e : A β B) {x : A} {y : B} β (e .fst x β‘ y) β (PathP (Ξ» i β ua e i) x y) ua-pathpβpath eqv .fst = pathβua-pathp eqv ua-pathpβpath eqv .snd .is-eqv y .centre = strict-fibres (ua-pathpβpath eqv) y .fst ua-pathpβpath eqv .snd .is-eqv y .paths = strict-fibres (ua-pathpβpath eqv) y .snd
The βaxiomβπ
The actual βunivalence axiomβ, as stated in the HoTT book, says that
the canonical map A β‘ B
, defined using J, is an equivalence. This map is idβequiv, defined right above. In more
intuitive terms, itβs βcastingβ the identity equivalence
A β A
along a proof that A β‘ B
to get an
equivalence A β B
.
module _ where private idβequiv : {A B : Type β} β A β‘ B β A β B idβequiv {A = A} {B} = J (Ξ» x _ β A β x) (_ , id-equiv) idβequiv-refl : {A : Type β} β idβequiv (Ξ» i β A) β‘ (_ , id-equiv) idβequiv-refl {A = A} = J-refl (Ξ» x _ β A β x) (_ , id-equiv)
However, because of efficiency concerns (Agda is a programming language, after all), instead of using idβequiv defined using J, we use pathβequiv, which is defined in an auxiliary module.
pathβequiv : {A B : Type β} β A β‘ B β A β B pathβequiv p = lineβequiv (Ξ» i β p i)
Since identity of equivalences is determined by identity of their underlying functions, to show that pathβequiv of refl is the identity equivalence, we use coe1βi to show that transport by refl is the identity.
pathβequiv-refl : {A : Type β} β pathβequiv (refl {x = A}) β‘ (id , id-equiv) pathβequiv-refl {A = A} = Ξ£-path (Ξ» i x β coe1βi (Ξ» i β A) i x) (is-propβpathp (Ξ» i β is-equiv-is-prop _) _ _)
For the other direction, we must show that ua of id-equiv is refl. We can do this quite efficiently using Glue. Since this is a path between paths, we have two interval variables.
ua-id-equiv : {A : Type β} β ua (_ , id-equiv {A = A}) β‘ refl ua-id-equiv {A = A} i j = Glue A {Ο = i β¨ ~ j β¨ j} (Ξ» _ β A , _ , id-equiv)
We can then prove that the map pathβequiv is an isomorphism, hence an equivalence. Itβs very useful to have explicit names for the proofs that pathβequiv and ua are equivalences without referring to components of PathβEquiv, so we introduce names for them as well.
PathβEquiv : {A B : Type β} β Iso (A β‘ B) (A β B) univalence : {A B : Type β} β is-equiv (pathβequiv {A = A} {B}) univalenceβ»ΒΉ : {A B : Type β} β is-equiv (ua {A = A} {B}) PathβEquiv {A = A} {B = B} = pathβequiv , iiso where iiso : is-iso pathβequiv iiso .is-iso.inv = ua
We show that pathβequiv
inverts ua, which
means proving that one can recover the original equivalence from the
generated path. Because of the computational nature of Cubical Agda, all
we have to do is apply uaΞ²:
iiso .is-iso.rinv (f , is-eqv) = Ξ£-path (funext (uaΞ² (f , is-eqv))) (is-equiv-is-prop f _ _)
For the other direction, we use path induction to reduce the problem from showing that ua inverts pathβequiv for an arbitrary path (which is hard) to showing that pathβequiv takes refl to the identity equivalence (pathβequiv-refl), and that ua takes the identity equivalence to refl (ua-id-equiv).
iiso .is-iso.linv = J (Ξ» _ p β ua (pathβequiv p) β‘ p) (ap ua pathβequiv-refl β ua-id-equiv) univalence {A = A} {B} = is-isoβis-equiv (PathβEquiv .snd) univalenceβ»ΒΉ {A = A} {B} = is-isoβis-equiv (is-iso.inverse (PathβEquiv .snd))
In some situations, it is helpful to have a proof that pathβequiv followed by an adjustment of levels is still an equivalence:
univalence-lift : {A B : Type β} β is-equiv (Ξ» e β lift (pathβequiv {A = A} {B} e)) univalence-lift {β = β} = is-isoβis-equiv morp where morp : is-iso (Ξ» e β lift {β = lsuc β} (pathβequiv e)) morp .is-iso.inv x = ua (x .Lift.lower) morp .is-iso.rinv x = lift (pathβequiv (ua (x .Lift.lower))) β‘β¨ ap lift (PathβEquiv .snd .is-iso.rinv _) β©β‘ x β morp .is-iso.linv x = PathβEquiv .snd .is-iso.linv _
Equivalence Inductionπ
One useful consequence of 1 is that the type of equivalences satisfies the same induction principle as the type of identifications. By analogy with how path induction can be characterised as contractibility of singletons and transport, βequivalence inductionβ can be characterised as transport and contractibility of singletons up to equivalence:
Equiv-is-contr : β {β} (A : Type β) β is-contr (Ξ£[ B β Type β ] A β B) is-contr.centre (Equiv-is-contr A) = A , _ , id-equiv is-contr.paths (Equiv-is-contr A) (B , AβB) i = ua AβB i , p i , q i where p : PathP (Ξ» i β A β ua AβB i) id (AβB .fst) p i x = outS (ua-glue AβB i (Ξ» { (i = i0) β x }) (inS (AβB .fst x))) q : PathP (Ξ» i β is-equiv (p i)) id-equiv (AβB .snd) q = is-propβpathp (Ξ» i β is-equiv-is-prop (p i)) _ _
Combining Equiv-is-contr with subst, we get an induction principle for the type of equivalences based at : To prove for any , it suffices to consider the case where is and is the identity equivalence.
EquivJ : β {β β'} {A : Type β} β (P : (B : Type β) β A β B β Type β') β P A (_ , id-equiv) β {B : Type β} (e : A β B) β P B e EquivJ P pid eqv = subst (Ξ» e β P (e .fst) (e .snd)) (Equiv-is-contr _ .is-contr.paths (_ , eqv)) pid
Equivalence induction simplifies the proofs of many properties about equivalences. For example, if is an equivalence, then so is its action on paths .
is-equivβis-embedding : β {β} {A B : Type β} β (f : A β B) β is-equiv f β {x y : A} β is-equiv (ap f {x = x} {y = y}) is-equivβis-embedding f eqv = EquivJ (Ξ» B e β is-equiv (ap (e .fst))) id-equiv (f , eqv)
The proof can be rendered in English roughly as follows:
Suppose is an equivalence. We want to show that, for any choice of , the map is an equivalence.
By induction, it suffices to cover the case where is , and is the identity function.
But then, we have that is definitionally equal to , which is known to be an equivalence.
Object Classifiersπ
In category theory, the idea of classifiers (or classifying objects) often comes up when categories applied to the study of logic. For example, any elementary topos has a subobject classifier: an object such that maps corresponds to maps with propositional fibres (equivalently, inclusions ). In higher categorical analyses of logic, classifying objects exist for more maps: an elementary 2-topos has a discrete object classifier, which classify maps with discrete fibres.
Since a -topos has classifiers for maps with -truncated fibres, and a -topos has classifiers for maps with -truncated fibres, one might expect that an -topos would have classifiers for maps with fibres that are not truncated at all. This is indeed the case! In HoTT, this fact is internalised using the univalent universes, and we can prove that univalent universes are object classifiers.
As an intermediate step, we prove that the value of a type family at a point is equivalent to the fibre of over . The proof follows from the De Morgan structure on the interval, and the βspreadβ operation coe1βi.
-- HoTT book lemma 4.8.1 Fibre-equiv : (B : A β Type β') (a : A) β fibre (fst {B = B}) a β B a Fibre-equiv B a = IsoβEquiv isom where isom : Iso _ _ isom .fst ((x , y) , p) = subst B p y isom .snd .inv x = (a , x) , refl isom .snd .rinv x i = coe1βi (Ξ» _ β B a) i x isom .snd .linv ((x , y) , p) i = (p (~ i) , coe1βi (Ξ» j β B (p (~ i β§ ~ j))) i y) , Ξ» j β p (~ i β¨ j)
Another fact from homotopy theory that we can import into homotopy type theory is that any map is equivalent to a fibration. More specifically, given a map , the total space is equivalent to the dependent sum of the fibres. The theorems Total-equiv and Fibre-equiv are what justify referring to Ξ£ the βtotal spaceβ of a type family.
Total-equiv : (p : E β B) β E β Ξ£ B (fibre p) Total-equiv p = IsoβEquiv isom where isom : Iso _ (Ξ£ _ (fibre p)) isom .fst x = p x , x , refl isom .snd .inv (_ , x , _) = x isom .snd .rinv (b , x , q) i = q i , x , Ξ» j β q (i β§ j) isom .snd .linv x = refl
Putting these together, we get the promised theorem: The space of
maps
is equivalent to the space of fibrations with base space
and variable total space
,
.
If we allow
and
to live in different universes, then the maps are classified by the
biggest universe in which they both fit, namely
Type (β β β')
. Note that the proof of Fibration-equiv makes fundamental use of
ua, to construct the witnesses that taking
fibres and taking total spaces are inverses. Without ua, we could only get an
βisomorphism-up-to-equivalenceβ of types.
Fibration-equiv : β {β β'} {B : Type β} β (Ξ£[ E β Type (β β β') ] (E β B)) β (B β Type (β β β')) Fibration-equiv {B = B} = IsoβEquiv isom where isom : Iso (Ξ£[ E β Type _ ] (E β B)) (B β Type _) isom .fst (E , p) = fibre p isom .snd .inv pβ»ΒΉ = Ξ£ _ pβ»ΒΉ , fst isom .snd .rinv prep i x = ua (Fibre-equiv prep x) i isom .snd .linv (E , p) i = ua e (~ i) , Ξ» x β fst (ua-unglue e (~ i) x) where e = Total-equiv p
To solidify the explanation that object classifiers generalise the -truncated object classifiers you would find in a -topos, we prove that any class of maps described by a property which holds of all of its fibres (or even structure on all of its fibres!) has a classifying object β the total space .
For instance, if we take
to be the property of being a proposition, this theorem tells us
that Ξ£ is-prop
classifies subobjects. With the
slight caveat that Ξ£ is-prop
is not closed under
impredicative quantification, this corresponds exactly to the notion of
subobject classifier in a
-topos,
since the maps with propositional fibres are precisely the injective
maps.
Since the type of βmaps into B with variable domain and P fibresβ has a very unwieldy description β both in words or in Agda syntax β we abbreviate it by . The notation is meant to evoke the idea of a slice category: The objects of are objects of the category equipped with choices of maps into . Similarly, the objects of are objects of the universe , with a choice of map into , such that holds for all the fibres of .
_/[_]_ : β {β' β''} (β : Level) β (Type (β β β') β Type β'') β Type β' β Type _ _/[_]_ {β} β' P B = Ξ£ (Type (β β β')) Ξ» A β Ξ£ (A β B) Ξ» f β (x : B) β P (fibre f x)
The proof that the slice is classified by follows from elementary properties of types (namely: reassociation, distributivity of over ), and the classification theorem Fibration-equiv. Really, the most complicated part of this proof is rearranging the nested sum and product types to a form to which we can apply Fibration-equiv.
Map-classifier : β {β β' β''} {B : Type β'} (P : Type (β β β') β Type β'') β (β /[ P ] B) β (B β Ξ£ _ P) Map-classifier {β = β} {B = B} P = (Ξ£ _ Ξ» A β Ξ£ _ Ξ» f β (x : B) β P (fibre f x)) ββ¨ Ξ£-assoc β©β (Ξ£ _ Ξ» { (x , f) β (x : B) β P (fibre f x) }) ββ¨ Ξ£-ap-fst (Fibration-equiv {β' = β}) β©β (Ξ£ _ Ξ» A β (x : B) β P (A x)) ββ¨ Ξ£-Ξ -distrib eβ»ΒΉ β©β (B β Ξ£ _ P) ββ
Not the fundamental theorem of engineering!β©οΈ