-- We begin by adding some primitive bindings using the PRIMITIVE pragma. -- -- It goes like this: PRIMITIVE primName varName. -- -- If the varName is dropped, then it's taken to be the same as primName. -- -- If there is a previous declaration for the varName, then the type -- is checked against the internally-known "proper" type for the primitive. -- Universe of fibrant types {-# PRIMITIVE Type #-} -- Universe of non-fibrant types {-# PRIMITIVE Pretype #-} -- Fibrant is a fancy word for "has a composition structure". Most types -- we inherit from MLTT are fibrant: -- -- Stuff like products Π, sums Σ, naturals, booleans, lists, etc., all -- have composition structures. -- -- The non-fibrant types are part of the structure of cubical -- categories: The interval, partial elements, cubical subtypes, ... -- The interval --------------- -- The interval has two endpoints i0 and i1. -- These form a de Morgan algebra. I : Pretype {-# PRIMITIVE Interval I #-} i0, i1 : I {-# PRIMITIVE i0 #-} {-# PRIMITIVE i1 #-} -- "minimum" on the interval iand : I -> I -> I {-# PRIMITIVE iand #-} -- "maximum" on the interval. ior : I -> I -> I {-# PRIMITIVE ior #-} -- The interpretation of iand as min and ior as max justifies the fact that -- ior i (inot i) != i1, since that equality only holds for the endpoints. -- inot i = 1 - i is a de Morgan involution. inot : I -> I {-# PRIMITIVE inot #-} -- Paths -------- -- Since every function in type theory is internally continuous, -- and the two endpoints i0 and i1 are equal, we can take the type of -- equalities to be continuous functions out of the interval. -- That is, x ≡ y iff. ∃ f : I -> A, f i0 = x, f i1 = y. -- The type PathP generalises this to dependent products (i : I) -> A i. PathP : (A : I -> Type) -> A i0 -> A i1 -> Type {-# PRIMITIVE PathP #-} -- By taking the first argument to be constant we get the equality type -- Path. Path : {A : Type} -> A -> A -> Type Path {A} = PathP (\i -> A) -- reflexivity is given by constant paths refl : {A : Type} {x : A} -> Path x x refl {A} {x} i = x -- Symmetry (for dpeendent paths) is given by inverting the argument to the path, such that -- sym p i0 = p (inot i0) = p i1 -- sym p i1 = p (inot i1) = p i0 -- This has the correct endpoints. sym : {A : I -> Type} {x : A i0} {y : A i1} -> PathP A x y -> PathP (\i -> A (inot i)) y x sym p i = p (inot i) id : {A : Type} -> A -> A id x = x the : (A : Type) -> A -> A the A x = x -- The eliminator for the interval says that if you have x : A i0 and y : A i1, -- and x ≡ y, then you can get a proof A i for every element of the interval. iElim : {A : I -> Type} {x : A i0} {y : A i1} -> PathP A x y -> (i : I) -> A i iElim p i = p i -- This corresponds to the elimination principle for the HIT -- data I : Pretype where -- i0 i1 : I -- seg : i0 ≡ i1 -- The singleton subtype of A at x is the type of elements of y which -- are equal to x. Singl : (A : Type) -> A -> Type Singl A x = (y : A) * Path x y -- Contractible types are those for which there exists an element to which -- all others are equal. isContr : Type -> Type isContr A = (x : A) * ((y : A) -> Path x y) -- Using the connection \i j -> y.2 (iand i j), we can prove that -- singletons are contracible. Together with transport later on, -- we get the J elimination principle of paths. singContr : {A : Type} {a : A} -> isContr (Singl A a) singContr {A} {a} = ((a, \i -> a), \y i -> (y.2 i, \j -> y.2 (iand i j))) -- Some more operations on paths. By rearranging parentheses we get a -- proof that the images of equal elements are themselves equal. cong : {A : Type} {B : A -> Type} (f : (x : A) -> B x) {x : A} {y : A} (p : Path x y) -> PathP (\i -> B (p i)) (f x) (f y) cong f p i = f (p i) -- These satisfy definitional equalities, like congComp and congId, which are -- propositional in vanilla MLTT. congComp : {A : Type} {B : Type} {C : Type} {f : A -> B} {g : B -> C} {x : A} {y : A} (p : Path x y) -> Path (cong g (cong f p)) (cong (\x -> g (f x)) p) congComp p = refl congId : {A : Type} {x : A} {y : A} (p : Path x y) -> Path (cong (id {A}) p) p congId p = refl -- Just like rearranging parentheses gives us cong, swapping the value -- and interval binders gives us function extensionality. funext : {A : Type} {B : A -> Type} {f : (x : A) -> B x} {g : (x : A) -> B x} (h : (x : A) -> Path (f x) (g x)) -> Path f g funext h i x = h x i happly : {A : Type} {B : A -> Type} {f : (x : A) -> B x} {g : (x : A) -> B x} -> (p : Path f g) -> (x : A) -> Path (f x) (g x) happly p x i = p i x -- The proposition associated with an element of the interval ------------------------------------------------------------- -- Associated with every element i : I of the interval, we have the type -- IsOne i which is inhabited only when i = i1. In the model, this -- corresponds to the map [φ] from the interval cubical set to the -- subobject classifier. IsOne : I -> Pretype {-# PRIMITIVE IsOne #-} -- The value itIs1 witnesses the fact that i1 = i1. itIs1 : IsOne i1 -- Furthermore, if either of i or j are one, then so is (i or j). isOneL : {i : I} {j : I} -> IsOne i -> IsOne (ior i j) isOneR : {i : I} {j : I} -> IsOne j -> IsOne (ior i j) {-# PRIMITIVE itIs1 #-} {-# PRIMITIVE isOneL #-} {-# PRIMITIVE isOneR #-} -- Partial elements ------------------- -- -- Since a function I -> A has two endpoints, and a function I -> I -> A -- has four endpoints + four functions I -> A as "sides" (obtained by -- varying argument while holding the other as a bound variable), we -- refer to elements of I^n -> A as "cubes". -- This justifies the existence of partial elements, which are, as the -- name implies, partial cubes. Namely, a Partial φ A is an element of A -- which depends on a proof that IsOne φ. Partial : I -> Type -> Pretype {-# PRIMITIVE Partial #-} -- There is also a dependent version where the type A is itself a -- partial element. PartialP : (phi : I) -> Partial phi Type -> Pretype {-# PRIMITIVE PartialP #-} -- Why is Partial φ A not just defined as φ -> A? The difference is that -- Partial φ A has an internal representation which definitionally relates -- any two partial elements which "agree everywhere", that is, have -- equivalent values for every possible assignment of variables which -- makes IsOne φ hold. -- Cubical Subtypes -------------------- -- Given A : Type, phi : I, and a partial element u : A defined on φ, -- we have the type Sub A phi u, notated A[phi -> u] in the output of -- the type checker, whose elements are "extensions" of u. -- That is, element of A[phi -> u] is an element of A defined everywhere -- (a total element), which, when IsOne φ, agrees with u. Sub : (A : Type) (phi : I) -> Partial phi A -> Pretype {-# PRIMITIVE Sub #-} -- Every total element u : A can be made partial on φ by ignoring the -- constraint. Furthermore, this "totally partial" element agrees with -- the original total element on φ. inS : {A : Type} {phi : I} (u : A) -> Sub A phi (\x -> u) {-# PRIMITIVE inS #-} -- When IsOne φ, outS {A} {φ} {u} x reduces to u itIs1. -- This implements the fact that x agrees with u on φ. outS : {A : Type} {phi : I} {u : Partial phi A} -> Sub A phi u -> A {-# PRIMITIVE outS #-} -- The composition operation ---------------------------- -- Now that we have syntax for specifying partial cubes, -- and specifying that an element agrees with a partial cube, -- we can describe the composition operation. comp : (A : I -> Type) {phi : I} (u : (i : I) -> Partial phi (A i)) -> Sub (A i0) phi (u i0) -> A i1 {-# PRIMITIVE comp #-} -- In particular, when φ is a disjunction of the form -- (j = 0) || (j = 1), we can draw u as being a pair of lines forming a -- "tube", an open square with no floor or roof: -- -- Given u = \j [ (i = i0) -> x, (i = i1) -> q j] on the extent i || ~i, -- we draw: -- -- x q i1 -- | | -- \j -> x | | \j -> q j -- | | -- x q i0 -- -- The composition operation says that, as long as we can provide a -- "floor" connecting x -- q i0, as a total element of A which, on -- phi, extends u i0, then we get the "roof" connecting x and q i1 -- for free. -- -- If we have a path p : x ≡ y, and q : y ≡ z, then we do get the -- "floor", and composition gets us the dotted line: -- -- x..........z -- | | -- x | | q j -- | | -- x----------y -- p i trans : {A : Type} {x : A} {y : A} {z : A} -> PathP (\i -> A) x y -> PathP (\i -> A) y z -> PathP (\i -> A) x z trans {A} {x} p q i = comp (\i -> A) {ior i (inot i)} (\j [ (i = i0) -> x, (i = i1) -> q j ]) (inS (p i)) -- In particular when the formula φ = i0 we get the "opposite face" to a -- single point, which corresponds to transport. transp : (A : I -> Type) (x : A i0) -> A i1 transp A x = comp A (\i [ ]) (inS x) -- Since we have the iand operator, we can also derive the *filler* of a cube, -- which connects the given face and the output of composition. fill : (A : I -> Type) {phi : I} (u : (i : I) -> Partial phi (A i)) -> Sub (A i0) phi (u i0) -> (i : I) -> A i fill A {phi} u a0 i = comp (\j -> A (iand i j)) {ior phi (inot i)} (\j [ (phi = i1) as p -> u (iand i j) p, (i = i0) -> outS a0 ]) (inS (outS a0)) hfill : {A : Type} {phi : I} (u : (i : I) -> Partial phi A) -> Sub A phi (u i0) -> I -> A hfill {A} {phi} u a0 i = fill (\i -> A) {phi} u a0 i hcomp : {A : Type} {phi : I} (u : (i : I) -> Partial phi A) -> Sub A phi (u i0) -> A hcomp {A} {phi} u a0 = comp (\i -> A) {phi} u a0 -- For instance, the filler of the previous composition square -- tells us that trans p refl = p: transRefl : {A : Type} {x : A} {y : A} (p : Path x y) -> Path (trans p refl) p transRefl p j i = fill (\i -> A) {ior i (inot i)} (\k [ (i = i0) -> x, (i = i1) -> y ]) (inS (p i)) (inot j) -- Reduction of composition --------------------------- -- -- Composition reduces on the structure of the family A : I -> Type to create -- the element a1 : (A i1)[phi -> u i1]. -- -- For instance, when filling a cube of functions, the behaviour is to -- first transport backwards along the domain, apply the function, then -- forwards along the codomain. transpFun : {A : Type} {B : Type} {C : Type} {D : Type} (p : Path A B) (q : Path C D) -> (f : A -> C) -> Path (transp (\i -> p i -> q i) f) (\x -> transp (\i -> q i) (f (transp (\i -> p (inot i)) x))) transpFun p q f = refl -- When considering the more general case of a composition respecing sides, -- the outer transport becomes a composition. -- Glueing and Univalence ------------------------- -- First, let's get some definitions out of the way. -- -- The *fiber* of a function f : A -> B at a point y : B is the type of -- inputs x : A which f takes to y, that is, for which there exists a -- path f(x) = y. fiber : {A : Type} {B : Type} -> (A -> B) -> B -> Type fiber f y = (x : A) * Path (f x) y -- An *equivalence* is a function where every fiber is contractible. -- That is, for every point in the codomain y : B, there is exactly one -- point in the input which f maps to y. isEquiv : {A : Type} {B : Type} -> (A -> B) -> Type isEquiv {A} {B} f = (y : B) -> isContr (fiber f y) -- By extracting this point, which must exist because the fiber is contractible, -- we can get an inverse of f: inverse : {A : Type} {B : Type} {f : A -> B} -> isEquiv f -> B -> A inverse eqv y = (eqv y) .1 .1 -- We can prove that «inverse eqv» is a section of f: section : {A : Type} {B : Type} (f : A -> B) (eqv : isEquiv f) -> Path (\x -> f (inverse eqv x)) id section f eqv i y = (eqv y) .1 .2 i -- Proving that it's also a retraction is left as an exercise to the -- reader. We can package together a function and a proof that it's an -- equivalence to get a capital-E Equivalence. Equiv : (A : Type) (B : Type) -> Type Equiv A B = (f : A -> B) * isEquiv f -- The identity function is an equivalence between any type A and -- itself. idEquiv : {A : Type} -> isEquiv (id {A}) idEquiv y = ((y, \i -> y), \u i -> (u.2 (inot i), \j -> u.2 (ior (inot i) j))) -- The glue operation expresses that "extensibility is invariant under -- equivalence". Less concisely, the Glue type and its constructor, -- glue, let us extend a partial element of a partial type to a total -- element of a total type, by "gluing" the partial type T using a -- partial equivalence e onto a total type A. -- In particular, we have that when φ = i1, Glue A [i1 -> (T, f)] = T. primGlue : (A : Type) {phi : I} (T : Partial phi Type) (e : PartialP phi (\o -> Equiv (T o) A)) -> Type {-# PRIMITIVE Glue primGlue #-} -- The glue constructor extends the partial element t : T to a total -- element of Glue A [φ -> (T, e)] as long as we have a total im : A -- which is the image of f(t). -- -- Agreeing with the condition that Glue A [i1 -> (T, e)] = T, -- we have that glue {A} {i1} t im => t. prim'glue : {A : Type} {phi : I} {T : Partial phi Type} {e : PartialP phi (\o -> Equiv (T o) A)} -> (t : PartialP phi T) -> (im : Sub A phi (\o -> (e o).1 (t o))) -> primGlue A T e {-# PRIMITIVE glue prim'glue #-} -- The unglue operation undoes a glueing. Since when φ = i1, -- Glue A [φ -> (T, f)] = T, the argument to primUnglue {A} {i1} ... -- will have type T, and so to get back an A we need to apply the -- partial equivalence f (defined everywhere). primUnglue : {A : Type} {phi : I} {T : Partial phi Type} {e : PartialP phi (\o -> Equiv (T o) A)} -> primGlue A {phi} T e -> A {-# PRIMITIVE unglue primUnglue #-} -- Diagramatically, i : I |- Glue A [(i \/ ~i) -> (T, e)] can be drawn -- as giving us the dotted line in: -- -- T i0 ......... T i1 -- | | -- | | -- e i0 |~ ~| e i1 -- | | -- | | -- A i0 --------- A i1 -- A -- -- Where the the two "e" sides are equivalences, and the bottom side is -- the line i : I |- A. -- -- Thus, by choosing a base type, a set of partial types and partial -- equivalences, we can make a line between two types (T i0) and (T i1). Glue : (A : Type) {phi : I} -> Partial phi ((X : Type) * Equiv X A) -> Type Glue A {phi} u = primGlue A {phi} (\o -> (u o).1) (\o -> (u o).2) -- For example, we can glue together the type A and the type B as long -- as there exists an Equiv A B. -- -- A ............ B -- | | -- | | -- equiv |~ ua equiv ~| idEquiv {B} -- | | -- | | -- B ------------ B -- \i → B -- univalence : {A : Type} {B : Type} -> Equiv A B -> Path A B univalence {A} {B} equiv i = Glue B (\[ (i = i0) -> (A, equiv), (i = i1) -> (B, the B, idEquiv {B}) ]) -- The fact that this diagram has 2 filled-in B sides explains the -- complication in the proof below. -- -- In particular, the actual behaviour of transp (\i -> univalence f i) -- (x : A) is not just to apply f x to get a B (the left side), it also -- needs to: -- -- * For the bottom side, compose along (\i -> B) (the bottom side) -- * For the right side, apply the inverse of the identity, which -- is just identity, to get *some* b : B -- -- But that b : B might not agree with the sides of the composition -- operation in a more general case, so it composes along (\i -> B) -- *again*! -- -- Thus the proof: a simple cubical argument suffices, since -- for any composition, its filler connects either endpoints. So -- we need to come up with a filler for the bottom and right faces. univalenceBeta : {A : Type} {B : Type} (f : Equiv A B) -> Path (transp (\i -> univalence f i)) f.1 univalenceBeta {A} {B} f i a = let -- The bottom left corner botLeft : B botLeft = transp (\i -> B) (f.1 a) -- The "bottom face" filler connects the bottom-left corner, f.1 a, -- and the bottom-right corner, which is the transport of f.1 a -- along \i -> B. botFace : Path (f.1 a) botLeft botFace i = fill (\i -> B) (\j []) (inS (f.1 a)) i -- The "right face" filler connects the bottom-right corner and the -- upper-right corner, which is again a simple transport along -- \i -> B. rightFace : Path (transp (\i -> B) botLeft) botLeft rightFace i = fill (\i -> B) (\j []) (inS botLeft) (inot i) -- The goal is a path between the bottom-left and top-right corners, -- which we can get by composing (in the path sense) the bottom and -- right faces. goal : Path (transp (\i -> B) botLeft) (f.1 a) goal = trans rightFace (\i -> botFace (inot i)) in goal i -- The terms univalence + univalenceBeta suffice to prove the "full" -- univalence axiom of Voevodsky, as can be seen in the paper -- -- Ian Orton, & Andrew M. Pitts. (2017). Decomposing the Univalence Axiom. -- -- Available freely here: https://arxiv.org/abs/1712.04890v3 J : {A : Type} {x : A} (P : (y : A) -> Path x y -> Type) (d : P x (\i -> x)) {y : A} (p : Path x y) -> P y p J P d p = transp (\i -> P (p i) (\j -> p (iand i j))) d -- Isomorphisms --------------- -- -- Since isomorphisms are a much more convenient notion of equivalence -- than contractible fibers, it's natural to ask why the CCHM paper, and -- this implementation following that, decided on the latter for our -- definition of equivalence. isIso : {A : Type} -> {B : Type} -> (A -> B) -> Type isIso {A} {B} f = (g : B -> A) * ((y : B) -> Path (f (g y)) y) * ((x : A) -> Path (g (f x)) x) -- The reason is that the family of types IsIso is not a proposition! -- This means that there can be more than one way for a function to be -- an equivalence. This is Lemma 4.1.1 of the HoTT book. Iso : Type -> Type -> Type Iso A B = (f : A -> B) * isIso f -- Nevertheless, we can prove that any function with an isomorphism -- structure has contractible fibers, using a cubical argument adapted -- from CCHM's implementation of cubical type theory: -- -- https://github.com/mortberg/cubicaltt/blob/master/experiments/isoToEquiv.ctt#L7-L55 IsoToEquiv : {A : Type} {B : Type} -> Iso A B -> Equiv A B IsoToEquiv {A} {B} iso = let f = iso.1 g = iso.2.1 s = iso.2.2.1 t = iso.2.2.2 lemIso : (y : B) (x0 : A) (x1 : A) (p0 : Path (f x0) y) (p1 : Path (f x1) y) -> PathP (\i -> fiber f y) (x0, p0) (x1, p1) lemIso y x0 x1 p0 p1 = let rem0 : Path x0 (g y) rem0 i = hcomp {A} (\k [ (i = i0) -> t x0 k, (i = i1) -> g y ]) (inS (g (p0 i))) rem1 : Path x1 (g y) rem1 i = hcomp {A} (\k [ (i = i0) -> t x1 k, (i = i1) -> g y ]) (inS (g (p1 i))) p : Path x0 x1 p i = hcomp {A} (\k [ (i = i0) -> rem0 (inot k), (i = i1) -> rem1 (inot k) ]) (inS (g y)) fill0 : I -> I -> A fill0 i j = hcomp {A} (\k [ (i = i0) -> t x0 (iand j k) , (i = i1) -> g y , (j = i0) -> g (p0 i) ]) (inS (g (p0 i))) fill1 : I -> I -> A fill1 i j = hcomp {A} (\k [ (i = i0) -> t x1 (iand j k) , (i = i1) -> g y , (j = i0) -> g (p1 i) ]) (inS (g (p1 i))) fill2 : I -> I -> A fill2 i j = hcomp {A} (\k [ (i = i0) -> rem0 (ior j (inot k)) , (i = i1) -> rem1 (ior j (inot k)) , (j = i1) -> g y ]) (inS (g y)) sq : I -> I -> A sq i j = hcomp {A} (\k [ (i = i0) -> fill0 j (inot k) , (i = i1) -> fill1 j (inot k) , (j = i1) -> g y , (j = i0) -> t (p i) (inot k) ]) (inS (fill2 i j)) sq1 : I -> I -> B sq1 i j = hcomp {B} (\k [ (i = i0) -> s (p0 j) k , (i = i1) -> s (p1 j) k , (j = i0) -> s (f (p i)) k , (j = i1) -> s y k ]) (inS (f (sq i j))) in \i -> (p i, \j -> sq1 i j) fCenter : (y : B) -> fiber f y fCenter y = (g y, s y) fIsCenter : (y : B) (w : fiber f y) -> Path (fCenter y) w fIsCenter y w = lemIso y (fCenter y).1 w.1 (fCenter y).2 w.2 in (f, \y -> (fCenter y, fIsCenter y)) -- We can prove that any involutive function is an isomorphism, since -- such a function is its own inverse. involToIso : {A : Type} (f : A -> A) -> ((x : A) -> Path (f (f x)) x) -> isIso f involToIso {A} f inv = (f, inv, inv) -- An example of univalence --------------------------- -- -- The classic example of univalence is the equivalence -- not : Bool \simeq Bool. -- -- We define it here. Bool : Type {-# PRIMITIVE Bool #-} true, false : Bool {-# PRIMITIVE true #-} {-# PRIMITIVE false #-} -- Pattern matching for booleans: If a proposition holds for true and -- for false, then it holds for every bool. elimBool : (P : Bool -> Type) -> P true -> P false -> (b : Bool) -> P b {-# PRIMITIVE if elimBool #-} -- Non-dependent elimination of booleans if : {A : Type} -> A -> A -> Bool -> A if {A} = elimBool (\b -> A) not : Bool -> Bool not = if false true -- By pattern matching it suffices to prove (not (not true)) ≡ true and -- not (not false) ≡ false. Since not (not true) computes to true (resp. -- false), both proofs go through by refl. notInvol : (x : Bool) -> Path (not (not x)) x notInvol = elimBool (\b -> Path (not (not b)) b) refl refl notp : Path Bool Bool notp = univalence (IsoToEquiv (not, involToIso not notInvol)) -- This path actually serves to prove a simple lemma about the universes -- of HoTT, namely, that any univalent universe is not a 0-type. If we -- had HITs, we could prove that this fact holds for any n, but for now, -- proving it's not an h-set is the furthest we can go. -- First we define what it means for something to be false. In type theory, -- we take ¬P = P → ⊥, where the bottom type is the only type satisfying -- the elimination principle -- -- elimBottom : (P : bottom -> Type) -> (b : bottom) -> P b -- -- This follows from setting bottom := ∀ A, A. bottom : Type bottom = {A : Type} -> A elimBottom : (P : bottom -> Type) -> (b : bottom) -> P b elimBottom P x = x -- We prove that true != false by transporting along the path -- -- \i -> if (Bool -> Bool) A (p i) -- (Bool -> Bool) ------------------------------------ A -- -- To verify that this has the correct endpoints, check out the endpoints -- for p: -- -- true ------------------------------------ false -- -- and evaluate the if at either end. trueNotFalse : Path true false -> bottom trueNotFalse p {A} = transp (\i -> if (Bool -> Bool) A (p i)) id -- To be an h-Set is to have no "higher path information". Alternatively, -- -- isHSet A = (x : A) (y : A) -> isHProp (Path x y) -- isHSet : Type -> Type isHSet A = {x : A} {y : A} (p : Path x y) (q : Path x y) -> Path p q -- We can prove *a* contradiction (note: this is a direct proof!) by adversarially -- choosing two paths p, q that we know are not equal. Since "equal" paths have -- equal behaviour when transporting, we can choose two paths p, q and a point x -- such that transporting x along p gives a different result from x along q. -- -- Since transp notp = not but transp refl = id, that's what we go with. The choice -- of false as the point x is just from the endpoints of trueNotFalse. universeNotSet : isHSet Type -> bottom universeNotSet itIs = trueNotFalse (\i -> transp (\j -> itIs notp refl i j) false)