module Data.Image where

Images and replacement🔗

The image of a function $f : A \to B$ is the subtype of $B$ where points are merely equipped with a fibre of $f$. This has the expected universal property: we can factor any $f$ into

$$A \twoheadrightarrow \operatorname*{im}(f) \hookrightarrow B\text{,}$$

and $\operatorname*{im}(f) \hookrightarrow B$ is universal among factorisations of $f$ through a subtype of its codomain.

In general, we can consider images of arbitrary functions $f : A \to B$, where $A$ is a $\ell_1$-small type, and $B$ is a $\ell_2$-small type; This results in a type $\operatorname*{im}f$, which, according to the rules for assigning universe levels, lives in the $(\ell_1 \sqcup \ell_2)$th universe. While the rules are that way for good reasons, in this case, they defy intuition: the type $\operatorname*{im}f$, in a sense, is “at most the size of $A$”.

Can we do better? It turns out that we can! This generally goes by the name of type-theoretic replacement, after Rijke (Rijke 2017, sec. 5); in turn, the name replacement comes from the axiom of material set theory saying the image of a set under a function is again set.¹ Here we implement a slight generalization of Evan Cavallo’s construction in terms of higher induction-recursion.

We understand this construction as generalising the first isomorphism theorem from the setting of algebraic structures² to the setting of general types. Recall that the first isomorphism theorem says you can compute $\operatorname*{im}f$ as the quotient $A/\ker f$.³

If we’re working with sets, that’s the end of it: We can define a relation $x \sim y$ by $f(x) \sim f(y)$, which can be made to live in the zeroth universe, so $A/\ker f$ lives in the same universe as $A$. That’s because, when working with sets, there are no coherence problems to get stuck on: we can first define $A/\ker f$ as a higher inductive type, then, separately, define the embedding $A/\ker f \hookrightarrow B$.

However, in the case of general types, we would like the h-level of $\operatorname*{im}f$ to match that of $B$, so we certainly can’t set-truncate.

Anyway, The next immediate attempt might be a type which looks like

  data Im1 {ℓa ℓb} {A : Type ℓa} {B : Type ℓb} (f : A → B)
        : Type (ℓa ⊔ ℓb) where
    inc  : A → Im1 f
    quot : ∀ x y → f x ≡ f y → inc x ≡ inc y

But this type isn’t particularly well-behaved. For example, the Im1-image of the identity map on the unit type is the circle. This can be established by writing equivalences back and forth, but in the interest of brevity (and our dependency graph), we show it has a non-trivial path.

  Cover : Im1 {A = ⊤} {⊤} (λ { tt → tt }) → Type
  Cover (inc x)          = Bool
  Cover (quot tt tt _ i) = ua (not , not-is-equiv) i

  Im1-nontriv : subst Cover (quot tt tt refl) true ≡ false
  Im1-nontriv = refl

It’s fairly common that defining higher inductive types like this will, sometimes, introduce more paths than you might expect. For an analogy, if we were interested in computing propositional truncations instead, our type Im1 above would correspond to this “one-step truncation”:

  data Tr1 {ℓa} (A : Type ℓa) : Type ℓa where
    inc  : A → Tr1 A
    quot : (x y : A) → inc x ≡ inc y

And, investigating the case of truncations, we might come up with a couple different fixes. One is to define a process that could be iterated: each step adds new paths, and at the limit — the sequential co-limit — we have our original type. This can be made to work for the propositional truncation, and (Rijke 2017)’s construction of the image is similar in spirit.

But if we have recursive higher inductive types, where the type being defined can appear in the arguments to its own path constructors, we can sometimes do these constructions in a single big, recursive step. That’s the usual propositional truncation. A guiding, heuristic principle would be to avoid “green slime”: do not mention the constructors of the type in the endpoints of the path constructor.

Types without green slime are much more merciful than the judges of coherence hell, because you don’t have to match to apply the path constructors, as you would in Tr1: If you have the arguments handy, you can apply them targeting arbitrary points. Lovely! But In the image case, this still doesn’t really work, because we don’t want to identify arbitrary points. Our requirements are something like this:

1. A tractable description: finitely many constructors, and hopefully “no more than a few”. This guarantees that we can tell Agda Agda about the construction and actually work with it, which is also important.

2. Without any green slime, to hopefully avoid the coherent issues.

3. So that points from $a$ are identified in $\operatorname*{im}f$ exactly how they would be identified in $b$ under $f$.

If we had such a type, we could define an embedding $f' : \operatorname*{im}f \hookrightarrow B$, which tells us how to compute the path spaces in $\operatorname*{im}f$: $(x \equiv y) \simeq (f'(x) \equiv f'(y))$. In fact, since images are almost entirely characterised by having such an embedding, we can use our hypothetical $f'$ to rewrite the third bullet point:

3. With a constructor coherently mapping proofs of $f'(x) \equiv f'(y)$ to $x \equiv y$.

Also, note that, since we were trying to get a smaller image, we can’t just use $B$’s path types: those live at the same universe as $B$. What we can do instead is parametrise over an identity system on $B$, a reflexive family of types $x \sim y$ which is equivalent to $x \equiv y$ but which might be smaller. So, our requirement on identifications should be

3. With a constructor coherently mapping proofs of $f'(x) \sim f'(y)$ to $x \equiv y$.

But, identity system or not, this is hopeless, right? We don’t know how to specify $\operatorname*{im}f$ without first defining $f'$; and it makes no sense to define $f'$ before its domain type. The only way forward would be to somehow define them at the same time. That’s impossible, right?

Higher induction-recursion🔗

It’s actually totally possible to, at least in Agda, define higher inductive types at the same time as we define pattern-matching functions on them. And that’s exactly what we’ll do: specify $\operatorname*{im}f$ higher-inductively at the same time as we define $f' : \operatorname*{im}f \hookrightarrow b$ by recursion. Our three bullet points from before become the specification of the inductive type!

module Replacement
    {ℓₐ ℓₜ ℓᵢ} {A : Type ℓₐ} {T : Type ℓₜ}
    {_∼_ : T → T → Type ℓᵢ} {rr : ∀ x → x ∼ x}
    (locally-small : is-identity-system _∼_ rr)
    (f : A → T)
  where

  private module ls = Ids locally-small

  data Image : Type (ℓₐ ⊔ ℓᵢ)
  embed : Image → T

  data Image where
    inc  : A → Image
    quot : ∀ {r r'} → embed r ∼ embed r' → r ≡ r'

  embed (inc a)     = f a
  embed (quot p i)  = locally-small .to-path p i

And, having used inductive-recursion to tie the dependency knot, we’re actually done: the construction above is coherent enough, even if it looks like the quot constructor only says that embed is an injection. We can use the algebraic properties of identity systems to show that it’s actually a proper embedding:

  embed-is-embedding' : ∀ x → is-contr (fibre embed (embed x))
  embed-is-embedding' x .centre = x , refl
  embed-is-embedding' x .paths (y , q) =
    Σ-pathp (quot (ls.from (sym q))) (commutes→square coh)
    where abstract
      coh : ls.to (ls.from (sym q)) ∙ q ≡ refl ∙ refl
      coh = ap (_∙ q) (ls.ε (sym q)) ·· ∙-invl q ·· sym (∙-idl refl)

  embed-is-embedding : is-embedding embed
  embed-is-embedding = embedding-lemma embed-is-embedding'

And it’s possible to pull back the identity system on $b$ to one on $\operatorname*{im}f$, to really drive home the point that points in the image are identified precisely through their identification, under $f$, in the codomain.

  Image-identity-system : is-identity-system (λ x y → embed x ∼ embed y) (λ _ → rr _)
  Image-identity-system = pullback-identity-system locally-small
    (embed , embed-is-embedding)

As usual with these things, we can establish properties of Image without caring about the higher constructors.

  Image-elim-prop
    : ∀ {ℓ'} {P : Image → Type ℓ'}
    → (∀ x → is-prop (P x))
    → (∀ x → P (inc x))
    → ∀ x → P x
  Image-elim-prop pprop pinc (inc x) = pinc x
  Image-elim-prop pprop pinc (quot {x} {y} p i) =
    is-prop→pathp (λ i → pprop (quot p i))
      (Image-elim-prop pprop pinc x)
      (Image-elim-prop pprop pinc y) i

From which surjectivity follows immediately:

  inc-is-surjective : is-surjective inc
  inc-is-surjective = Image-elim-prop (λ _ → squash) (λ x → inc (x , refl))

Comparison🔗

We define a “comparison” map from the resized image to the “ordinary” image. This map is a bit annoying to define, and it relies on being able to easily map from the Image into propositions. It’s definitional that projecting from Image→image gives our embedding, since that’s how we define the map.

  Image→image : Image → image f
  Image→image x .fst = embed x
  Image→image x .snd =
    Image-elim-prop {P = λ x → ∥ fibre f (embed x) ∥} (λ _ → squash)
      (λ x → inc (x , refl))
      x

Establishing that Image→image is an equivalence is one of the rare cases where it’s easier to show that it has contractible fibres. That’s because we can eliminate the propositional truncation as the first step, and deal only with untruncated data from then on.

  Image→image-is-equiv : is-equiv Image→image
  Image→image-is-equiv .is-eqv (x , p) =
    ∥-∥-elim {P = λ p → is-contr (fibre Image→image (x , p))}
      (λ _ → is-contr-is-prop)
      (λ { (i , p) → J
        (λ x p → is-contr (fibre Image→image (x , inc (i , p))))
        (work' i) p })
      p
    where
      work' : (f⁻¹x : A) → is-contr (fibre Image→image (f f⁻¹x , inc (_ , refl)))
      work' f⁻¹x .centre = inc f⁻¹x , refl -- inc f⁻¹x , refl

Contracting the fibres is where we get some mileage out of having gotten the green slime out of quot. We have to show $f^{-1}(x) = i$, as elements of the image, but we have an assumption that $\mathrm{embed}(i) = ff^{-1}(x)$, which, under quot, is exactly what we need.

      work' f⁻¹x .paths (i , α) = Σ-pathp (quot (ls.from (sym (ap fst α)))) $
        Σ-prop-square (λ _ → squash) $ commutes→square $
          (ap₂ _∙_ (ls.ε (sym (ap fst α))) refl ∙ ∙-invl _ ∙ sym (∙-idl _))