open import Cat.Functor.Adjoint.Hom
open import Cat.Instances.Functor
open import Cat.Displayed.Fibre
open import Cat.Functor.Adjoint
open import Cat.Displayed.Base
open import Cat.Prelude

import Cat.Displayed.Cocartesian.Indexing
import Cat.Displayed.Cartesian.Indexing
import Cat.Displayed.Cocartesian.Weak
import Cat.Displayed.Fibre.Reasoning
import Cat.Displayed.Cartesian.Weak
import Cat.Displayed.Cocartesian
import Cat.Displayed.Cartesian
import Cat.Displayed.Reasoning
import Cat.Reasoning

module Cat.Displayed.Bifibration
  {o ℓ o′ ℓ′} {ℬ : Precategory o ℓ} (ℰ : Displayed ℬ o′ ℓ′) where

open Cat.Displayed.Cocartesian ℰ
open Cat.Displayed.Cocartesian.Weak ℰ
open Cat.Displayed.Cartesian ℰ
open Cat.Displayed.Cartesian.Weak ℰ
open Cat.Displayed.Reasoning ℰ

open Cat.Reasoning ℬ
open Displayed ℰ
open Functor
open _⊣_
open _=>_
private
  module Fib = Cat.Displayed.Fibre.Reasoning ℰ

Bifibrations🔗

A displayed category $\mathcal{E} \mathrel{\htmlClass{liesover}{\hspace{1.366em}}} \mathcal{B}$ is a bifibration if is it both a fibration and an opfibration. This means that $\mathcal{E}$ is equipped with both reindexing and opreindexing functors, which allows us to both restrict and extend along morphisms $X \to Y$ in the base.

Note that a bifibration is not the same as a “profunctor valued in categories”. Those are a distinct concept, called two-sided fibrations.

record is-bifibration : Type (o ⊔ ℓ ⊔ o′ ⊔ ℓ′) where
  field
    fibration : Cartesian-fibration
    opfibration : Cocartesian-fibration

  module fibration = Cartesian-fibration fibration
  module opfibration = Cocartesian-fibration opfibration

Bifibrations and Adjoints🔗

If $\mathcal{E}$ is a bifibration, then its opreindexing functors are left adjoints to its reindexing functors. To show this, it will suffice to construct natural isomorphism between $\mathcal{E}_{y}(u_{*}(-),-)$ and $\mathcal{E}_{x}(-,u^{*}(-))$. However, we have already shown that $\mathcal{E}_{y}(u_{*}(-),-)$ and $\mathcal{E}_{x}(-,u^{*}(-))$ are both naturally isomorphic to $\mathcal{E}_{u}(-,-)$¹, so all we need to do is compose these natural isomorphisms!

module _ (bifib : is-bifibration) where
  open is-bifibration bifib
  open Cat.Displayed.Cartesian.Indexing ℰ fibration
  open Cat.Displayed.Cocartesian.Indexing ℰ opfibration

  cobase-change⊣base-change
    : ∀ {x y} (f : Hom x y)
    → cobase-change f ⊣ base-change f
  cobase-change⊣base-change {x} {y} f =
    hom-natural-iso→adjoints $
      (opfibration→hom-iso opfibration f ni⁻¹) ni∘ fibration→hom-iso fibration f

In fact, if $\mathcal{E} \mathrel{\htmlClass{liesover}{\hspace{1.366em}}} \mathcal{B}$ is a cartesian fibration where every reindexing functor has a left adjoint, then $\mathcal{E}$ is a bifibration!

Since $\mathcal{E}$ is a fibration, every $u : x \to y$ in $\mathcal{B}$ induces a natural isomorphism $\mathcal{E}_{u}(x',-) \simeq \mathcal{E}_{x}(x', u^{*}(-))$, by the universal property of cartesian lifts. If $u^{*}$ additionally has a left adjoint $L_{u}$, we have natural isomorphisms

$$\mathcal{E}_{u}(x',-) \simeq \mathcal{E}_{x}(x',u^{*}(-)) \simeq \mathcal{E}_{y}(L_{u}(x')-)\text{,}$$

which implies $\mathcal{E}$ is a weak opfibration; and any weak opfibration that’s also a fibration is a proper opfibration.

module _ (fib : Cartesian-fibration) where
  open Cartesian-fibration fib
  open Cat.Displayed.Cartesian.Indexing ℰ fib

  left-adjoint-base-change→opfibration
    : (L : ∀ {x y} → (f : Hom x y) → Functor (Fibre ℰ x) (Fibre ℰ y))
    → (∀ {x y} → (f : Hom x y) → (L f ⊣ base-change f))
    → Cocartesian-fibration
  left-adjoint-base-change→opfibration L adj =
    cartesian+weak-opfibration→opfibration fib $
    hom-iso→weak-opfibration L λ u →
      fibration→hom-iso-from fib u ni∘ (adjunct-hom-iso-from (adj u) _ ni⁻¹)

  left-adjoint-base-change→bifibration
    : (L : ∀ {x y} → (f : Hom x y) → Functor (Fibre ℰ x) (Fibre ℰ y))
    → (∀ {x y} → (f : Hom x y) → (L f ⊣ base-change f))
    → is-bifibration
  left-adjoint-base-change→bifibration L adj .is-bifibration.fibration =
    fib
  left-adjoint-base-change→bifibration L adj .is-bifibration.opfibration =
    left-adjoint-base-change→opfibration L adj

Adjoints from cocartesian maps🔗

Let $f : X \to Y$ be a morphism in $\mathcal{B}$, and let $L : \mathcal{E}_{X} \to \mathcal{E}_{Y}$ be a functor. If we are given a natural transformation $\eta : \operatorname{Id}_{} \to f^{*} \circ L$ with $\overline{f} \circ \eta$ cocartesian, then $L$ is a left adjoint to $f^{*}$.

  cocartesian→left-adjoint-base-change
    : ∀ {x y} {L : Functor (Fibre ℰ x) (Fibre ℰ y)} {f : Hom x y}
    → (L-unit : Id => base-change f F∘ L)
    → (∀ x → is-cocartesian (f ∘ id) (has-lift.lifting f (L .F₀ x) ∘′ L-unit .η x))
    → L ⊣ base-change f

We will construct the adjunction by constructing a natural equivalence of $\mathbf{Hom}$-sets

$$\mathcal{E}_{X}{L A, B} \simeq \mathcal{E}_{Y}{A, f^{*}B}\text{.}$$

The map $v \mapsto f^{*}(v) \circ \eta$ gives us the forward direction of this equivalence, so all that remains is to find an inverse. Since $\overline{f}\circ\eta$ is cocartesian, it satisfies a mapping-out universal property: if $v : A \to f^{*} B$ is a vertical map in $\mathcal{E}$, we can construct a vertical map $LA \to B$ by factoring $\overline{f} \circ v$ through the cocartesian $\overline{f} \circ \eta$; The actual universal property says that this factorising process is an equivalence.

  cocartesian→left-adjoint-base-change {x = x} {y = y} {L = L} {f = f} L-unit cocart =
    hom-iso→adjoints (λ v → base-change f .F₁ v Fib.∘ L-unit .η _)
      precompose-equiv equiv-natural where
      module cocart x = is-cocartesian (cocart x)
      module f* = Functor (base-change f)

      precompose-equiv
        : ∀ {x′ : Ob[ x ]} {y′ : Ob[ y ]}
        → is-equiv {A = Hom[ id ] (F₀ L x′) y′} (λ v → f*.₁ v Fib.∘ L-unit .η x′)
      precompose-equiv {x′} {y′} = is-iso→is-equiv $ iso
        (λ v → cocart.universalv _ (has-lift.lifting f _ ∘′ v))
        (λ v → has-lift.uniquep₂ _ _ _ _ refl _ _
          (Fib.pulllf (has-lift.commutesp f _ id-comm _)
          ∙[] symP (assoc′ _ _ _)
          ∙[] cocart.commutesv x′ _)
          refl)
        (λ v → symP $ cocart.uniquep x′ _ _ _ _ $
          assoc′ _ _ _
          ∙[] Fib.pushlf (symP $ has-lift.commutesp f _ id-comm _))

      equiv-natural
        : hom-iso-natural {L = L} {R = base-change f} (λ v → f*.₁ v Fib.∘ L-unit .η _)
      equiv-natural g h k =
        has-lift.uniquep₂ _ _ _ _ _ _ _
          (Fib.pulllf (has-lift.commutesp f _ id-comm _)
           ∙[] pushl[] _ (pushl[] _ (to-pathp⁻ (smashr _ _))))
          (Fib.pulllf (has-lift.commutesp f _ id-comm _)
           ∙[] extendr[] _ (Fib.pulllf (Fib.pulllf (has-lift.commutesp f _ id-comm _)))
           ∙[] extendr[] _ (pullr[] _ (to-pathp (L-unit .is-natural _ _ h)))
           ∙[] pullr[] _ (Fib.pulllf (extendr[] _ (has-lift.commutesp f _ id-comm _))))