open import 1Lab.Equiv
open import 1Lab.Path hiding (_∙_)
open import 1Lab.Type

module 1Lab.Path.Groupoid where

Types are Groupoids🔗

The Path types equip every Type with the structure of an $\infty$-groupoid. The higher structure of a type begins with its inhabitants (the 0-cells); Then, there are the paths between inhabitants - these are inhabitants of the type Path A x y, which are the 1-cells in A. Then, we can consider the inhabitants of Path (Path A x y) p q, which are homotopies between paths.

This construction iterates forever - and, at every stage, we have that the $n$-cells in A are the 0-cells of some other type (e.g.: The 1-cells of A are the 0-cells of Path A ...). Furthermore, this structure is weak: The laws, e.g. associativity of composition, only hold up to a homotopy. These laws satisfy their own laws — again using associativity as an example, the associator satisfies the pentagon identity.

These laws can be proved in two ways: Using path induction, or directly with a cubical argument. Here, we do both.

Book-style🔗

This is the approach taken in the HoTT book. We fix a type, and some variables of that type, and some paths between variables of that type, so that each definition doesn’t start with 12 parameters.

module WithJ where
  private variable
    ℓ : Level
    A : Type ℓ
    w x y z : A

First, we (re)define the operations using J. These will be closer to the structure given in the book.

  _∙_ : x ≡ y → y ≡ z → x ≡ z
  _∙_ {x = x} {y} {z} = J (λ y _ → y ≡ z → x ≡ z) (λ x → x)

First we define path composition. Then, we can prove that the identity path - refl - acts as an identity for path composition.

  ∙-id-r : (p : x ≡ y) → p ∙ refl ≡ p
  ∙-id-r {x = x} {y = y} p =
    J (λ _ p → p ∙ refl ≡ p)
      (happly (J-refl (λ y _ → y ≡ y → x ≡ y) (λ x → x)) _)
      p

This isn’t as straightforward as it would be in “Book HoTT” because - remember - J doesn’t compute definitionally, only up to the path J-refl. Now the other identity law:

  ∙-id-l : (p : y ≡ z) → refl ∙ p ≡ p
  ∙-id-l {y = y} {z = z} p = happly (J-refl (λ y _ → y ≡ z → y ≡ z) (λ x → x)) p

This case we get for less since it’s essentially the computation rule for J.

  ∙-assoc : (p : w ≡ x) (q : x ≡ y) (r : y ≡ z)
          → p ∙ (q ∙ r) ≡ (p ∙ q) ∙ r
  ∙-assoc {w = w} {x = x} {y = y} {z = z} p q r =
    J (λ x p → (q : x ≡ y) (r : y ≡ z) → p ∙ (q ∙ r) ≡ (p ∙ q) ∙ r) lemma p q r
    where
      lemma : (q : w ≡ y) (r : y ≡ z)
            → (refl ∙ (q ∙ r)) ≡ ((refl ∙ q) ∙ r)
      lemma q r =
        (refl ∙ (q ∙ r)) ≡⟨ ∙-id-l (q ∙ r) ⟩≡
        q ∙ r            ≡⟨ sym (ap (λ e → e ∙ r) (∙-id-l q)) ⟩≡
        (refl ∙ q) ∙ r   ∎

The associativity rule is trickier to prove, since we do inductive where the motive is a dependent product. What we’re doing can be summarised using words: By induction, it suffices to assume p is refl. Then, what we want to show is (refl ∙ (q ∙ r)) ≡ ((refl ∙ q) ∙ r). But both of those compute to q ∙ r, so we are done. This computation isn’t automatic - it’s expressed by the lemma.

This expresses that the paths behave like morphisms in a category. For a groupoid, we also need inverses and cancellation:

  inv : x ≡ y → y ≡ x
  inv {x = x} = J (λ y _ → y ≡ x) refl

The operation which assigns inverses has to be involutive, which follows from two computations.

  inv-inv : (p : x ≡ y) → inv (inv p) ≡ p
  inv-inv {x = x} =
    J (λ y p → inv (inv p) ≡ p)
      (ap inv (J-refl (λ y _ → y ≡ x) refl) ∙ J-refl (λ y _ → y ≡ x) refl)

And we have to prove that composing with an inverse gives the reflexivity path.

  ∙-inv-l : (p : x ≡ y) → p ∙ inv p ≡ refl
  ∙-inv-l {x = x} = J (λ y p → p ∙ inv p ≡ refl)
                      (∙-id-l (inv refl) ∙ J-refl (λ y _ → y ≡ x) refl)

  ∙-inv-r : (p : x ≡ y) → inv p ∙ p ≡ refl
  ∙-inv-r {x = x} = J (λ y p → inv p ∙ p ≡ refl)
                      (∙-id-r (inv refl) ∙ J-refl (λ y _ → y ≡ x) refl)

Cubically🔗

Now we do the same using hfill instead of path induction.

module _ where
  private variable
    ℓ : Level
    A B : Type ℓ
    w x y z : A

  open 1Lab.Path

The left and right identity laws follow directly from the two fillers for the composition operation.

  ∙-id-r : (p : x ≡ y) → p ∙ refl ≡ p
  ∙-id-r p = sym (∙-filler p refl)

  ∙-id-l : (p : x ≡ y) → refl ∙ p ≡ p
  ∙-id-l p = sym (∙-filler' refl p)

For associativity, we use both:

  ∙-assoc : (p : w ≡ x) (q : x ≡ y) (r : y ≡ z)
          → p ∙ (q ∙ r) ≡ (p ∙ q) ∙ r
  ∙-assoc p q r i = ∙-filler p q i ∙ ∙-filler' q r (~ i)

For cancellation, we need to sketch an open cube where the missing square expresses the equation we’re looking for. Thankfully, we only have to do this once!

  ∙-inv-r : ∀ {x y : A} (p : x ≡ y) → p ∙ sym p ≡ refl
  ∙-inv-r {x = x} p i j =
    hcomp (λ l → λ { (j = i0) → x
                   ; (j = i1) → p (~ l ∧ ~ i)
                   ; (i = i1) → x
                   })
          (p (j ∧ ~ i))

For the other direction, we use the fact that p is definitionally equal to sym (sym p). In that case, we show that sym p ∙ sym (sym p) ≡ refl - which computes to the thing we want!

  ∙-inv-l : (p : x ≡ y) → sym p ∙ p ≡ refl
  ∙-inv-l p = ∙-inv-r (sym p)

In addition to the groupoid identities for paths in a type, it has been established that functions behave like functors: These are the lemmas ap-refl and ap-sym in the 1Lab.Path module.

There, a proof that functions preserve path composition wasn’t included, because it needs hcomp to be defined. We fill a cube where the left face is defined to be ap f (p ∙ q) (that’s the (i = i0) face in the hcomp below), and the remaining structure of the proof mimics the definition of _∙_, so that we indeed get ap f p ∙ ap f q as the right face.

  ap-comp-path : (f : A → B) {x y z : A} (p : x ≡ y) (q : y ≡ z)
               → ap f (p ∙ q) ≡ ap f p ∙ ap f q
  ap-comp-path f {x} p q i j =
    hcomp (λ k → λ { (i = i0) → f (∙-filler p q k j)
                   ; (j = i0) → f x
                   ; (j = i1) → f (q k)
                   })
          (f (p j))

Convenient helpers🔗

Since a lot of Homotopy Type Theory is dealing with paths, this section introduces useful helpers for dealing with $n$-ary compositions. For instance, we know that $p^{-1} ∙ p ∙ q$ is $q$, but this involves more than a handful of intermediate steps:

  ∙-cancel-l : {x y z : A} (p : x ≡ y) (q : y ≡ z)
             → (sym p ∙ p ∙ q) ≡ q
  ∙-cancel-l p q =
    sym p ∙ p ∙ q   ≡⟨ ∙-assoc _ _ _ ⟩≡
    (sym p ∙ p) ∙ q ≡⟨ ap₂ _∙_ (∙-inv-l p) refl ⟩≡
    refl ∙ q        ≡⟨ ∙-id-l q ⟩≡
    q               ∎

  ∙-cancel'-l : {x y z : A} (p : x ≡ y) (q r : y ≡ z)
              → p ∙ q ≡ p ∙ r → q ≡ r
  ∙-cancel'-l = J (λ y p → (q r : y ≡ _) → p ∙ q ≡ p ∙ r → q ≡ r)
                  (λ q r proof → sym (∙-id-l q) ·· proof ·· ∙-id-l r)

  ∙-cancel'-r : {x y z : A} (p : y ≡ z) (q r : x ≡ y)
              → q ∙ p ≡ r ∙ p → q ≡ r
  ∙-cancel'-r = J (λ y p → (q r : _ ≡ _) → q ∙ p ≡ r ∙ p → q ≡ r)
                  (λ q r proof → sym (∙-id-r q) ·· proof ·· ∙-id-r r)

Groupoid structure of types (cont.)🔗

A useful fact is that if $H$ is a homotopy f ~ id, then we can “invert” it as such:

open 1Lab.Path

homotopy-invert : ∀ {a} {A : Type a} {f : A → A}
                → (H : (x : A) → f x ≡ x) {x : A}
                → H (f x) ≡ ap f (H x)
homotopy-invert {f = f} H {x = x} =
  sym (
    ap f (H x)                     ≡⟨ sym (∙-id-r _) ⟩≡
    ap f (H x) ∙ refl              ≡⟨ ap₂ _∙_ refl (sym (∙-inv-r _)) ⟩≡
    ap f (H x) ∙ H x ∙ sym (H x)   ≡⟨ ∙-assoc _ _ _ ⟩≡
    (ap f (H x) ∙ H x) ∙ sym (H x) ≡⟨ ap₂ _∙_ (sym (homotopy-natural H _)) refl ⟩≡
    (H (f x) ∙ H x) ∙ sym (H x)    ≡⟨ sym (∙-assoc _ _ _) ⟩≡
    H (f x) ∙ H x ∙ sym (H x)      ≡⟨ ap₂ _∙_ refl (∙-inv-r _) ⟩≡
    H (f x) ∙ refl                 ≡⟨ ∙-id-r _ ⟩≡
    H (f x)                        ∎
  )