« Competitive Programming in Haskell: tree path decomposition, part II » Retiring BlogLiterately

Decidable equality for indexed data types

Posted on September 9, 2024
Tagged , ,

Recently, as part of a larger project, I wanted to define decidable equality for an indexed data type in Agda. I struggled quite a bit to figure out the right way to encode it to make Agda happy, and wasn’t able to find much help online, so I’m recording the results here.

The tl;dr is to use mutual recursion to define the indexed data type along with a sigma type that hides the index, and to use the sigma type in any recursive positions where we don’t care about the index! Read on for more motivation and details (and wrong turns I took along the way).

This post is literate Agda; you can download it here if you want to play along. I tested everything here with Agda version 2.6.4.3 and version 2.0 of the standard library.

Background

First, some imports and a module declaration. Note that the entire development is parameterized by some abstract set B of base types, which must have decidable equality.

open import Data.Product using (Σ ; _×_ ; _,_ ; -,_ ; proj₁ ; proj₂)
open import Data.Product.Properties using (≡-dec)
open import Function using (__)
open import Relation.Binary using (DecidableEquality)
open import Relation.Binary.PropositionalEquality using (__ ; refl)
open import Relation.Nullary.Decidable using (yes; no; Dec)

module OneLevelTypesIndexed (B : Set) (≟B : DecidableEquality B) where

We’ll work with a simple type system containing base types, function types, and some distinguished type constructor □. So far, this is just to give some context; it is not the final version of the code we will end up with, so we stick it in a local module so it won’t end up in the top-level namespace.

module Unindexed where
  data Ty : Set where
    base : B  Ty
    __ : Ty  Ty  Ty
_ : Ty  Ty

For example, if \(X\) and \(Y\) are base types, then we could write down a type like \(\square ((\square \square X \to Y) \to \square Y)\):

  infixr 2 __
  infix 30_

  postulate
    BX BY : B

  X : Ty
  X = base BX
  Y : Ty
  Y = base BY

  example : Ty
  example =((□ □ X ⇒ Y) ⇒ □ Y)

However, for reasons that would take us too far afield in this blog post, I don’t want to allow immediately nested boxes, like \(\square \square X\). We can still have multiple boxes in a type, and even boxes nested inside of other boxes, as long as there is at least one arrow in between. In other words, I only want to rule out boxes immediately applied to another type with an outermost box. So we don’t want to allow the example type given above (since it contains \(\square \square X\)), but, for example, \(\square ((\square X \to Y) \to \square Y)\) would be OK.

Encoding invariants

How can we encode this invariant so it holds by construction? One way would be to have two mutually recursive data types, like so:

module Mutual where
  data Ty : Set
  data UTy : Set

  data Ty where
_ : UTy  Ty
_ : UTy  Ty

  data UTy where
    base : B  UTy
    __ : Ty  Ty  UTy

UTy consists of types which have no top-level box; the constructors of Ty just inject UTy into Ty by adding either one or zero boxes. This works, and defining decidable equality for Ty and UTy is relatively straightforward (again by mutual recursion). However, it seemed to me that having to deal with Ty and UTy everywhere through the rest of the development was probably going to be super annoying.

The other option would be to index Ty by values indicating whether a type has zero or one top-level boxes; then we can use the indices to enforce the appropriate rules. First, we define a data type Boxity to act as the index for Ty, and show that it has decidable equality:

data Boxity : Set where
  [0] : Boxity
  [1] : Boxity

Boxity-≟ : DecidableEquality Boxity
Boxity-≟ [0] [0] = yes refl
Boxity-≟ [0] [1] = no λ ()
Boxity-≟ [1] [0] = no λ ()
Boxity-≟ [1] [1] = yes refl

My first attempt to write down a version of Ty indexed by Boxity looked like this:

module IndexedTry1 where
  data Ty : Boxity  Set where
    base : B  Ty [0]
    __ : {b₁ b₂ : Boxity}  Ty b₁  Ty b₂  Ty [0]
_ : Ty [0]  Ty [1]

base always introduces a type with no top-level box; the constructor requires a type with no top-level box, and produces a type with one (this is what ensures we cannot nest boxes); and the arrow constructor does not care how many boxes its arguments have, but constructs a type with no top-level box.

This is logically correct, but I found it very difficult to work with. The sticking point for me was injectivity of the arrow constructor. When defining decidable equality we need to prove lemmas that each of the constructors are injective, but I was not even able to write down the type of injectivity for _⇒_. We would want something like this:

-inj :
  {bσ₁ bσ₂ bτ₁ bτ₂ : Boxity}
  {σ₁ : Ty bσ₁} {σ₂ : Ty bσ₂} {τ₁ : Ty bτ₁} {τ₂ : Ty bτ₂} →
  (σ₁ ⇒ σ₂) ≡ (τ₁ ⇒ τ₂) →
  (σ₁ ≡ τ₁) × (σ₂ ≡ τ₂)

but this does not even typecheck! The problem is that, for example, σ₁ and τ₁ have different types, so the equality proposition σ₁ ≡ τ₁ is not well-typed.

At this point I tried turning to heterogeneous equality, but it didn’t seem to help. I won’t record here all the things I tried, but the same issues seemed to persist, just pushed around to different places (for example, I was not able to pattern-match on witnesses of heterogeneous equality because of types that didn’t match).

Sigma types to the rescue

At ICFP last week I asked Jesper Cockx for advice,which felt a bit like asking Rory McIlroy to give some tips on your mini-golf game

and he suggested trying to prove decidable equality for the sigma type pairing an index with a type having that index, like this:

  ΣTy : Set
  ΣTy = Σ Boxity Ty

This turned out to be the key idea, but it still took me a long time to figure out the right way to make it work. Given the above definitions, if we go ahead and try to define decidable equality for ΣTy, injectivity of the arrow constructor is still a problem.

After days of banging my head against this off and on, I finally realized that the way to solve this is to define Ty and ΣTy by mutual recursion: the arrow constructor should just take two ΣTy arguments! This perfectly captures the idea that we don’t care about the indices of the arrow constructor’s argument types, so we hide them by bundling them up in a sigma type.

ΣTy : Set
data Ty : Boxity  Set

ΣTy = Σ Boxity Ty

data Ty where
_ : Ty [0]  Ty [1]
  base : B  Ty [0]
  __ : ΣTy  ΣTy  Ty [0]

infixr 2 __
infix 30_

Now we’re cooking! We now make quick work of the required injectivity lemmas, which all go through trivially by matching on refl:


□-inj : {τ₁ τ₂ : Ty [0]}  (□ τ₁ ≡ □ τ₂)  (τ₁ ≡ τ₂)
□-inj refl = refl

base-inj : {b₁ b₂ : B}  base b₁ ≡ base b₂  b₁ ≡ b₂
base-inj refl = refl

⇒-inj : {σ₁ σ₂ τ₁ τ₂ : ΣTy}  (σ₁ ⇒ σ₂)(τ₁ ⇒ τ₂)  (σ₁ ≡ τ₁) × (σ₂ ≡ τ₂)
⇒-inj refl = refl , refl

Notice how the type of ⇒-inj is now perfectly fine: we just have a bunch of ΣTy values that hide their indices, so we can talk about propositional equality between them with no trouble.

Finally, we can define decidable equality for Ty and ΣTy by mutual recursion.

ΣTy-≟ : DecidableEquality ΣTy

{-# TERMINATING #-}
Ty-≟ :  {b}  DecidableEquality (Ty b)

Sadly, I had to reassure Agda that the definition of Ty-≟ is terminating—more on this later.

To define ΣTy-≟ we can just use a lemma from Data.Product.Properties which derives decidable equality for a sigma type from decidable equality for both components.

ΣTy-≟ = ≡-dec Boxity-≟ Ty-≟

The only thing left is to define decidable equality for any two values of type Ty b (given a specific boxity b), making use of our injectivity lemmas; now that we have the right definitions, this falls out straightforwardly.

Ty-≟ (□ σ) (□ τ) with Ty-≟ σ τ
... | no σ≢τ = no (σ≢τ ∘ □-inj)
... | yes refl = yes refl
Ty-≟ (base x) (base y) with ≟B x y
... | no x≢y = no (x≢y ∘ base-inj)
... | yes refl = yes refl
Ty-≟ (σ₁ ⇒ σ₂) (τ₁ ⇒ τ₂) with ΣTy-≟ σ₁ τ₁ | ΣTy-≟ σ₂ τ₂
... | no σ₁≢τ₁ | _ = no (σ₁≢τ₁ ∘ proj₁ ∘ ⇒-inj)
... | yes _ | no σ₂≢τ₂ = no (σ₂≢τ₂ ∘ proj₂ ∘ ⇒-inj)
... | yes refl | yes refl = yes refl
Ty-≟ (base _) (__) = no λ ()
Ty-≟ (__) (base _) = no λ ()

Final thoughts

First, the one remaining infelicity is that Agda could not tell that Ty-≟ is terminating. I am not entirely sure why, but I think it may be that the way the recursion works is just too convoluted for it to analyze properly: Ty-≟ calls ΣTy-≟ on structural subterms of its inputs, but then ΣTy-≟ works by providing Ty-≟ as a higher-order parameter to ≡-dec. If you look at the definition of ≡-dec, all it does is call its function parameters on structural subterms of its input, so everything should be nicely terminating, but I guess I am not surprised that Agda is not able to figure this out. If anyone has suggestions on how to make this pass the termination checker without using a TERMINATING pragma, I would love to hear it!

As a final aside, I note that converting back and forth between Ty (with ΣTy arguments to the arrow constructor) and IndexedTry1.Ty (with expanded-out Boxity and Ty arguments to arrow) is trivial:

Ty→Ty1 : {b : Boxity}  Ty b  IndexedTry1.Ty b
Ty→Ty1 (□ σ) = IndexedTry1.(Ty→Ty1 σ)
Ty→Ty1 (base x) = IndexedTry1.base x
Ty→Ty1 ((b₁ , σ₁)(b₂ , σ₂)) = (Ty→Ty1 σ₁) IndexedTry1.(Ty→Ty1 σ₂)

Ty1→Ty : {b : Boxity}  IndexedTry1.Ty b  Ty b
Ty1→Ty (IndexedTry1.base x) = base x
Ty1→Ty (σ₁ IndexedTry1.⇒ σ₂) = -, (Ty1→Ty σ₁) ⇒ -, (Ty1→Ty σ₂)
Ty1→Ty (IndexedTry1.□ σ) =(Ty1→Ty σ)

I expect it is also trivial to prove this is an isomorphism, though I’m not particularly motivated to do it. The point is that, as anyone who has spent any time proving things with proof assistants knows, two types can be completely isomorphic, and yet one can be vastly easier to work with than the other in certain contexts. Often when I’m trying to prove something in Agda it feels like at least half the battle is just coming up with the right representation that makes the proofs go through easily.