open import Cat.Functor.Adjoint.Reflective
open import Cat.Functor.Properties
open import Cat.Diagram.Terminal
open import Cat.Functor.Adjoint
open import Cat.Prelude

import Cat.Functor.Reasoning as Func
import Cat.Reasoning as Cr

module Cat.Morphism.Orthogonal where

Orthogonal maps🔗

A pair of maps $f:a\to b$ and $g:ctod$ are called orthogonal, written $f\perp g$¹, if for every $u,v$ fitting into a commutative diagram like

the space of arrows $c\to b$ (dashed) which commute with everything is contractible. We refer to the type of these dashed arrows as a Lifting, and this type is parametrised over all maps in the square.

module _ {o ℓ} (C : Precategory o ℓ) where
  private
    module C = Cr C
    variable
      a b c d : ⌞ C ⌟
      f g h u v : C.Hom a b
  open C using (Ob ; Hom ; _∘_ ; _≅_)

  Lifting
    : (f : Hom a b) (g : Hom c d) (u : Hom a c) (v : Hom b d)
    → Type _
  Lifting f g u v = Σ[ w ∈ Hom _ _ ] w ∘ f ≡ u × g ∘ w ≡ v

  m⊥m : Hom a b → Hom c d → Type _
  m⊥m {b = b} {c = c} f g =
    ∀ {u v} → v ∘ f ≡ g ∘ u
    → is-contr (Lifting f g u v)

  lifts-against : (f : Hom a b) (g : Hom c d) → Type _
  lifts-against f g = ∀ {u v} → v ∘ f ≡ g ∘ u → Lifting f g u v

We also outline concepts of a map being orthogonal to an object, which is informally written $f\perp X\text{,}$ and an object being orthogonal to a map $Y\perp f\text{.}$

  m⊥o : Hom a b → ⌞ C ⌟ → Type _
  m⊥o {a} {b} f X = (a : Hom a X) → is-contr (Σ[ b ∈ Hom b X ] (b ∘ f ≡ a))

  o⊥m : ⌞ C ⌟ → Hom a b → Type _
  o⊥m {a} {b} Y f = (c : Hom Y b) → is-contr (Σ[ d ∈ Hom Y a ] (f ∘ d ≡ c))

  open Terminal

The proof is mostly a calculation, so we present it without a lot of comment.

  object-orthogonal-!-orthogonal
    : ∀ {X} (T : Terminal C) (f : Hom a b) → m⊥o f X ≃ m⊥m f (! T)
  object-orthogonal-!-orthogonal {X = X} T f =
    prop-ext (hlevel ) (hlevel ) to from
    where
      to : m⊥o f X → m⊥m f (! T)
      to f⊥X {u} {v} sq =
        contr (f⊥X u .centre .fst , f⊥X u .centre .snd , !-unique₂ T _ _) λ m →
          Σ-prop-path! (ap fst (f⊥X u .paths (m .fst , m .snd .fst)))

      from : m⊥m f (! T) → m⊥o f X
      from f⊥! a = contr
        ( f⊥! {v = ! T} (!-unique₂ T _ _) .centre .fst
        , f⊥! (!-unique₂ T _ _) .centre .snd .fst ) λ x → Σ-prop-path!
          (ap fst (f⊥! _ .paths (x .fst , x .snd , !-unique₂ T _ _)))

As a passing observation we note that if $f\perp X$ and $X\cong Y\text{,}$ then $f\perp Y\text{.}$ Of course, this is immediate in categories, but it holds in the generality of precategories.

  m⊥-iso : ∀ {a b} {X Y} (f : Hom a b) → X ≅ Y → m⊥o f X → m⊥o f Y

  m⊥-iso f x≅y f⊥X a =
    contr
      ( g.to ∘ contr' .centre .fst
      , C.pullr (contr' .centre .snd) ∙ C.cancell g.invl )
      λ x → Σ-prop-path! $
        ap₂ _∘_ refl (ap fst (contr' .paths (g.from ∘ x .fst , C.pullr (x .snd))))
        ∙ C.cancell g.invl
    where
      module g = C._≅_ x≅y
      contr' = f⊥X (g.from ∘ a)

A slightly more interesting lemma is that, if $f$ is orthogonal to itself, then it is an isomorphism:

  self-orthogonal→invertible : ∀ {a b} (f : Hom a b) → m⊥m f f → C.is-invertible f
  self-orthogonal→invertible f f⊥f =
    C.make-invertible (gpq .fst) (gpq .snd .snd) (gpq .snd .fst)
    where
      gpq = f⊥f (C.idl _ ∙ C.intror refl) .centre

For the next few lemmas, consider a square of the following form, where $l$ and $k$ are both lifts of the outer square.

If $f$ is an epimorphism, then $l=k\text{.}$ In more succinct terms, the type of lifts of such a square is a proposition.

  left-epic→lift-is-prop
    : C.is-epic f → v C.∘ f ≡ g C.∘ u → is-prop (Lifting f g u v)
  left-epic→lift-is-prop f-epi vf=gu (l , lf=u , _) (k , kf=u , _) =
    Σ-prop-path! (f-epi l k (lf=u ∙ sym kf=u))

Dually, if $g$ is a monomorphism, then we the type of lifts is also a proposition.

  right-monic→lift-is-prop
    : C.is-monic g → v C.∘ f ≡ g C.∘ u → is-prop (Lifting f g u v)
  right-monic→lift-is-prop g-mono vf=gu (l , _ , gl=v) (k , _ , gk=v) =
    Σ-prop-path! (g-mono l k (gl=v ∙ sym gk=v))

As a corollary, if $f$ is an epi or $g$ is a mono, then it is sufficient to find any lift to establish that $f\perp g\text{.}$

  left-epic-lift→orthogonal
    : (g : C.Hom c d)
    → C.is-epic f → lifts-against f g → m⊥m f g
  left-epic-lift→orthogonal g f-epi lifts vf=gu =
    is-prop∙→is-contr (left-epic→lift-is-prop f-epi vf=gu) (lifts vf=gu)

  right-monic-lift→orthogonal
    : (f : C.Hom a b)
    → C.is-monic g → lifts-against f g → m⊥m f g
  right-monic-lift→orthogonal f g-mono lifts vf=gu =
    is-prop∙→is-contr (right-monic→lift-is-prop g-mono vf=gu) (lifts vf=gu)

Isomorphisms are left and right orthogonal to every other morphism.

  invertible→left-orthogonal  : (g : C.Hom c d) → C.is-invertible f → m⊥m f g
  invertible→right-orthogonal : (f : C.Hom a b) → C.is-invertible g → m⊥m f g

We will focus our attention on the left orthogonal case, as the proof for right orthogonality is completely dual. Suppose that $f$ is invertible, and $g$ is an arbitrary morphism. Invertible morphisms are epis, so it suffices to establish the existence of lifts to prove that $f$ is orthogonal to $g\text{.}$ Luckily, these lifts are easy to find: if we have some square $v\circ f=u\circ g\text{,}$ then $u\circ f^{-1}$ fits perfectly along the diagonal:

A short calculation verifies that $u\circ f^{-1}$ is indeed a lift.

  invertible→left-orthogonal {f = f} g f-inv =
    left-epic-lift→orthogonal g (C.invertible→epic f-inv) λ {u} {v} vf=gu →
      u ∘ f.inv ,
      C.cancelr f.invr ,
      Equiv.from
        (g ∘ u ∘ f.inv ≡ v   ≃⟨ C.reassocl e⁻¹ ∙e C.pre-invr f-inv ⟩≃
         g ∘ u ≡ v ∘ f       ≃⟨ sym-equiv ⟩≃
         v ∘ f ≡ g ∘ u       ≃∎)
        vf=gu
    where module f = C.is-invertible f-inv

  invertible→right-orthogonal {g = g} f g-inv =
    right-monic-lift→orthogonal f (C.invertible→monic g-inv) λ {u} {v} vf=gu →
      g.inv ∘ v ,
      Equiv.from
        ((g.inv ∘ v) ∘ f ≡ u ≃⟨ C.reassocl ∙e C.pre-invl g-inv ⟩≃
          v ∘ f ≡ g ∘ u      ≃∎)
        vf=gu ,
      C.cancell g.invl
    where module g = C.is-invertible g-inv

  orthogonal→lifts-against : m⊥m f g → lifts-against f g
  orthogonal→lifts-against o p = o p .centre

We also have the following two properties, which state that “lifting against” is, as a property of morphisms, closed under composition on both the left and the right. To understand the proof, it’s helpful to visualise the inputs in a diagram. Suppose we have $f:b\to c\text{,}$ $g:a\to b\text{,}$ $h:x\to y\text{,}$ and assume that both $f$ and $g$ lift against $h\text{.}$ Showing that $fg$ lifts against $h$ amounts to finding a diagonal for the rectangle

under the assumption that $vfg=hu\text{.}$ We’ll populate this diagram a bit by observing that, by composing the $f$ and $v$ edges together, we have a commutative square with faces $g\text{,}$ $vf\text{,}$ $u$ and $h$ — and since $g$ lifts against $h\text{,}$ this implies that we have a diagonal map $w\text{,}$ which appears dashed in

This map satisfies $wg=u\text{,}$ and, importantly, $hw=vf\text{.}$ This latter equation implies that we can now treat the right half of the diagram as another square, with faces $f\text{,}$ $h\text{,}$ $w\text{,}$ and $v\text{.}$ Since $f$ also lifts against $h\text{,}$ this implies that we can find the dotted map $t$ in the diagram

satisfying $tf=w$ and $mt=v\text{.}$ The map $t$ fits the type of fillers for our original rectangle, but we must still show that it makes both triangles commute. But this is easy: we have $tfg=wg=u$ by a short calculation and $ht=w$ immediately.

  ∘l-lifts-against : lifts-against f h → lifts-against g h → lifts-against (f ∘ g) h
  ∘l-lifts-against f-lifts g-lifts vfg=hu =
    let (w , wg=u , hw=vf) = g-lifts (C.reassocl.from vfg=hu)
        (t , tf=w , ht=v)  = f-lifts (sym hw=vf)
    in t , C.pulll tf=w ∙ wg=u , ht=v

This proof dualises almost term-for-term the case where we’re composing on the bottom face, i.e., when we have some $f$ which lifts against both $g$ and $h\text{,}$ and we want to show $f$ lifts against $gh\text{.}$

  ∘r-lifts-against : lifts-against f g → lifts-against f h → lifts-against f (g ∘ h)
  ∘r-lifts-against f-lifts g-lifts ve=fgu =
    let (w , we=gu , fw=v) = f-lifts (C.reassocr.to ve=fgu)
        (t , te=u , gt=w)  = g-lifts we=gu
    in t , te=u , C.pullr gt=w ∙ fw=v

Regarding reflections🔗

module
  _ {o ℓ o' ℓ'} {C : Precategory o ℓ} {D : Precategory o' ℓ'}
    {r : Functor C D} {ι : Functor D C}
    (r⊣ι : r ⊣ ι) (ι-ff : is-fully-faithful ι)
  where
  private
    module C = Cr C
    module D = Cr D
    module ι = Func ι
    module r = Func r
    module rι = Func (r F∘ ι)
    module ιr = Func (ι F∘ r)
  open _⊣_ r⊣ι

Let $r⊣\iota :\mathcal{D}\rightleftarrows \mathcal{C}$ be an arbitrary reflective subcategory. Speaking abstractly, there is a “universal” choice of test for whether an object is “in” the subcategory: Whether the adjunction unit: ${\eta }_x:x\to \iota r(x)$ is an isomorphism. The theory of orthogonality gives a second way to detect this situation. The proof here is from (Borceux 1994, vol. 1, sec. 5.4).

The first thing we observe is that any map $f$ such that $r(f)$ is an isomorphism is orthogonal to every object in the subcategory: $f\perp \iota (X)\text{.}$ Let $f:a\to b$ be inverted by $r\text{,}$ and $X$ be the object. Given a map $a:a\to \iota X\text{,}$

  in-subcategory→orthogonal-to-inverted
    : ∀ {X} {a b} {f : C.Hom a b} → D.is-invertible (r.₁ f) → m⊥o C f (ι.₀ X)
  in-subcategory→orthogonal-to-inverted {X} {A} {B} {f} rf-inv a→x =
    contr (fact , factors) λ { (g , factors') →
      Σ-prop-path! (h≡k factors factors') }
    where
      module rf = D.is-invertible rf-inv
      module η⁻¹ {a} = C.is-invertible (is-reflective→unit-G-is-iso r⊣ι ι-ff {a})

Observe that, since $r⊣\iota$ is a reflective subcategory, every unit morphism ${\eta }_{\iota X}$ is an isomorphism; We define a morphism $b:\iota r(A)\to \iota X$ as the composite

      b : C.Hom (ι.₀ (r.₀ A)) (ι.₀ X)
      b = η⁻¹.inv C.∘ ι.₁ (r.₁ a→x)

satisfying (by naturality of the unit map) the property that $b\eta =a\text{.}$ This is an intermediate step in what we have to do: construct a map $B\to \iota (X)\text{.}$

      p : a→x ≡ b C.∘ unit.η _
      p = sym (C.pullr (sym (unit.is-natural _ _ _)) ∙ C.cancell zag)

We define that using the map $b$ we just constructed. It’s the composite

and a calculation shows us that this map is indeed a factorisation of $a$ through $f\text{.}$

      fact : C.Hom B (ι.₀ X)
      fact = b C.∘ ι.₁ rf.inv C.∘ unit.η _

      factors =
        (b C.∘ ι.₁ rf.inv C.∘ unit.η B) C.∘ f      ≡⟨ C.pullr (C.pullr (unit.is-natural _ _ _)) ⟩≡
        b C.∘ ι.₁ rf.inv C.∘ (ιr.₁ f) C.∘ unit.η A ≡⟨ C.refl⟩∘⟨ C.cancell (ι.annihilate rf.invr) ⟩≡
        b C.∘ unit.η A                             ≡˘⟨ p ⟩≡˘
        a→x                                        ∎

The proof that this factorisation is unique is surprisingly totally independent of the actual map we just constructed: If $hf=a=kf\text{,}$ note that we must have $r(h)=r(k)$ (since $r(f)$ is invertible, it is epic); But then we have

and since ${\eta }_{\iota X}$ is an isomorphism, thus monic, we have $h=k\text{.}$

      module _ {h k : C.Hom B (ι.₀ X)} (p : h C.∘ f ≡ a→x) (q : k C.∘ f ≡ a→x) where
        rh≡rk : r.₁ h ≡ r.₁ k
        rh≡rk = D.invertible→epic rf-inv (r.₁ h) (r.₁ k) (r.weave (p ∙ sym q))

        h≡k = C.invertible→monic (is-reflective→unit-G-is-iso r⊣ι ι-ff) _ _ $
          unit.η (ι.₀ X) C.∘ h ≡⟨ unit.is-natural _ _ _ ⟩≡
          ιr.₁ h C.∘ unit.η B  ≡⟨ ap ι.₁ rh≡rk C.⟩∘⟨refl ⟩≡
          ιr.₁ k C.∘ unit.η B  ≡˘⟨ unit.is-natural _ _ _ ⟩≡˘
          unit.η (ι.₀ X) C.∘ k ∎

As a partial converse, if an object $X$ is orthogonal to every unit map (it suffices to be orthogonal to its own unit map), then it lies in the subcategory:

  orthogonal-to-ηs→in-subcategory
    : ∀ {X} → (∀ B → m⊥o C (unit.η B) X) → C.is-invertible (unit.η X)
  orthogonal-to-ηs→in-subcategory {X} ortho =
    C.make-invertible x lemma (ortho X C.id .centre .snd)
    where
      x = ortho X C.id .centre .fst
      lemma = unit.η _ C.∘ x             ≡⟨ unit.is-natural _ _ _ ⟩≡
              ιr.₁ x C.∘ unit.η (ιr.₀ X) ≡⟨ C.refl⟩∘⟨ η-comonad-commute r⊣ι ι-ff ⟩≡
              ιr.₁ x C.∘ ιr.₁ (unit.η X) ≡⟨ ιr.annihilate (ortho X C.id .centre .snd) ⟩≡
              C.id                       ∎

And the converse to that is a specialisation of the first thing we proved: We established that $f\perp \iota (X)$ if $f$ is invertible by the reflection functor, and we know that $\eta$ is invertible by the reflection functor; It remains to replace $\iota (X)$ with any object for which $\eta$ is an isomorphism.

  in-subcategory→orthogonal-to-ηs
    : ∀ {X B} → C.is-invertible (unit.η X) → m⊥o C (unit.η B) X
  in-subcategory→orthogonal-to-ηs inv =
    m⊥-iso C (unit.η _) (C.invertible→iso _ (C.is-invertible-inverse inv))
      (in-subcategory→orthogonal-to-inverted
        (is-reflective→F-unit-is-iso r⊣ι ι-ff))