Agda type-safe cast / coercion - agda

I found handy a function:
coerce : ∀ {ℓ} {A B : Set ℓ} → A ≡ B → A → B
coerce refl x = x
when defining functions with indexed types. In situations where indexes are not definitionally equal i,e, one have to use lemma, to show the types match.
zipVec : ∀ {a b n m } {A : Set a} {B : Set b} → Vec A n → Vec B m → Vec (A × B) (n ⊓ m)
zipVec [] _ = []
zipVec {n = n} _ [] = coerce (cong (Vec _) (0≡n⊓0 n)) []
zipVec (x ∷ xs) (y ∷ ys) = (x , y) ∷ zipVec xs ys
Note, yet this example is easy to rewrite so one don't need to coerce:
zipVec : ∀ {a b n m } {A : Set a} {B : Set b} → Vec A n → Vec B m → Vec (A × B) (n ⊓ m)
zipVec [] _ = []
zipVec (_ ∷ _) [] = []
zipVec (x ∷ xs) (y ∷ ys) = (x , y) ∷ zipVec xs ys
Sometimes pattern matching doesn't help though.
The question: But I wonder, whether something like that functions is already in agda-stdlib? And is there something like hoogle for Agda, or something like SearchAbout?

I don't think there is exactly your coerce function. However, it's a special case of a more general function - subst (the substitutive property of equality) from Relation.Binary.PropositionalEquality:
subst : ∀ {a p} {A : Set a} (P : A → Set p) {x y : A}
→ x ≡ y → P x → P y
subst P refl p = p
If you choose P = id (from Data.Function, or just write λ x → x), you get:
coerce : ∀ {ℓ} {A B : Set ℓ} → A ≡ B → A → B
coerce = subst id
By the way, the most likely reason you won't find this function predefined, is that Agda deals with coerces like that through rewrite:
postulate
n⊓0≡0 : ∀ n → n ⊓ 0 ≡ 0
zipVec : ∀ {a b n m} {A : Set a} {B : Set b}
→ Vec A n → Vec B m → Vec (A × B) (n ⊓ m)
zipVec [] _ = []
zipVec {n = n} _ [] rewrite n⊓0≡0 n = []
zipVec (x ∷ xs) (y ∷ ys) = (x , y) ∷ zipVec xs ys
This is a syntactic sugar for the more complicated:
zipVec {n = n} _ [] with n ⊓ 0 | n⊓0≡0 n
... | ._ | refl = []
If you are familiar with how with works, try to figure out how rewrite works; it's quite enlightening.

Related

How is Agda inferring the implicit argument to `Vec.foldl`?

foldl : ∀ {a b} {A : Set a} (B : ℕ → Set b) {m} →
(∀ {n} → B n → A → B (suc n)) →
B zero →
Vec A m → B m
foldl b _⊕_ n [] = n
foldl b _⊕_ n (x ∷ xs) = foldl (λ n → b (suc n)) _⊕_ (n ⊕ x) xs
When translating the above function to Lean, I was shocked to find out that its true form is actually like...
def foldl : ∀ (P : ℕ → Type a) {n : nat}
(f : ∀ {n}, P n → α → P (n+1)) (s : P 0)
(l : Vec α n), P n
| P 0 f s (nil _) := s
| P (n+1) f s (cons x xs) := foldl (fun n, P (n+1)) (λ n, #f (n+1)) (#f 0 s x) xs
I find it really impressive that Agda is able to infer the implicit argument to f correctly. How is it doing that?
foldl : ∀ {a b} {A : Set a} (B : ℕ → Set b) {m} →
(∀ {n} → B n → A → B (suc n)) →
B zero →
Vec A m → B m
foldl b _⊕_ n [] = n
foldl b _⊕_ n (x ∷ xs) = foldl (λ n → b (suc n)) _⊕_ (_⊕_ {0} n x) xs
If I pass it 0 explicitly as in the Lean version, I get a hint as to the answer. What is going on is that Agda is doing the same thing as in the Lean version, namely wrapping the implicit arg so it is suc'd.
This is surprising as I thought that implicit arguments just means that Agda should provide them on its own. I did not think it would change the function when it is passed as an argument.

How to convert the J axiom to the fixed-argument form?

I was trying to prove that true ≡ false -> Empty assuming the J axiom. It is defined as:
J : Type
J = forall
{A : Set}
{C : (x y : A) → (x ≡ y) → Set} →
(c : ∀ x → C x x refl) →
(x y : A) →
(p : x ≡ y) →
C x y p
My attempt went like this:
bad : J → true ≡ false -> Empty
bad j e = j Bool (λ { true _ _ => Unit; false _ _ => Empty }) _
Now, to proceed with the proof, I needed a term c : ∀ x -> C x x refl. Since I instantiated C, it becomes c : ∀ x -> (λ { true _ _ => Unit; false _ _ => Empty } x x refl. Then I got stuck. c can't reduce further because we don't know the value of x. I wasn't able to complete this proof. But there is a different version of J:
J' : Type
J' = forall
{A : Set}
{x : A}
{C : (y : A) → (x ≡ y) → Set} →
(c : C x refl) →
(y : A) →
(p : x ≡ y) →
C y p
With this one, this problem is solved, because t can be fixed to be true. This makes the c argument reduce to Unit, which we can provide. My question is: can we convert the former version to the later? That is, can we build a term fix_x : J → J'? Does that hold in general (i.e., can indices be converted to parameters)?
First, regarding true ≡ false -> Empty: this is unprovable if you can only eliminate into Set0 with J, so you need an universe polymorphic or large definition. I write some preliminaries here:
{-# OPTIONS --without-K #-}
open import Relation.Binary.PropositionalEquality
open import Level
data Bool : Set where true false : Bool
data Empty : Set where
record Unit : Set where
constructor tt
JTy : ∀ {i j} → Set _
JTy {i}{j} =
{A : Set i}
(P : (x y : A) → (x ≡ y) → Set j) →
(pr : ∀ x → P x x refl) →
{x y : A} →
(p : x ≡ y) →
P x y p
J : ∀ {i}{j} → JTy {i}{j}
J P pr {x} refl = pr x
J₀ = J {zero}{zero}
Now, transport or subst is the only needed thing for true ≡ false -> Empty:
transp : ∀ {i j}{A : Set i}(P : A → Set j){x y} → x ≡ y → P x → P y
transp P = J (λ x y _ → P x -> P y) (λ _ px → px)
true≢false : true ≡ false → Empty
true≢false e = transp (λ {true → Unit; false → Empty}) e tt
Considering now proving the pointed J' from J, I know about three solutions, and each uses different features from the ambient theory.
The simplest one is to use universes to abstract over the induction motive:
JTy' : ∀ {i j} → Set _
JTy' {i}{j} =
{A : Set i}
{x : A}
(P : ∀ y → x ≡ y → Set j)
(pr : P x refl)
{y : A}
(p : x ≡ y)
→ P y p
JTy→JTy' : (∀ {i j} → JTy {i}{j}) → ∀ {i}{j} → JTy' {i}{j}
JTy→JTy' J {i} {j} {A} {x} P pr {y} e =
J (λ x y e → (P : ∀ y → x ≡ y → Set j) → P x refl → P y e)
(λ x P pr → pr) e P pr
If we only want to use a fixed universe level, then it is a bit more complicated. The following solution, sometimes called "contractible singletons", needs Σ-types, but nothing else:
open import Data.Product
JTy→JTy'withΣ : JTy {zero}{zero} → JTy' {zero}{zero}
JTy→JTy'withΣ J {A} {x} P pr {y} e =
J (λ {(x , r) (y , e) _ → P x r → P y e})
(λ _ px → px)
(J (λ x y e → (x , refl) ≡ (y , e))
(λ _ → refl)
e)
pr
There is a solution which doesn't even need Σ-s, but requires the beta rule for J, which says that J P pr {x} refl = pr x. It doesn't matter whether this rule holds definitionally or just as a propositional equality, but the construction is simpler when it holds definitionally, so let's do that. Note that I don't use any universe other than Set0.
transp₀ = transp {zero}{zero}
transp2 : ∀ {A : Set}{B : A → Set}(C : ∀ a → B a → Set)
{x y : A}(e : x ≡ y){b} → C x b → C y (transp₀ B e b)
transp2 {A}{B} C {x}{y} e {b} cxb =
J₀ (λ x y e → ∀ b → C x b → C y (transp₀ B e b)) (λ _ _ cxb → cxb) e b cxb
JTy→JTy'noΣU : JTy' {zero}{zero}
JTy→JTy'noΣU {A} {x} P pr {y} e =
transp₀ (P y) (J₀ (λ x y e → transp₀ (x ≡_) e refl ≡ e) (λ _ → refl) e)
(transp2 {A} {λ y → x ≡ y} P e pr)
Philosophically, the third version is the most "conservative", since it only assumes J. The addition of the beta rule is not really an extra thing, since it is always assumed to hold (definitionally or propositionally) for _≡_.
can indices be converted to parameters?
If you have propositional equality, then all indices can be converted to parameters, and fixed in constructors using equality proofs.

How to prove unfold-reverse for Vec?

The Agda standard library has a few properties on how reverse and _++_ work on List. Trying to transfer these proofs to Vec appears to be non-trivial (disregarding universes):
open import Data.Nat
open import Data.Vec
open import Relation.Binary.HeterogeneousEquality
unfold-reverse : {A : Set} → (x : A) → {n : ℕ} → (xs : Vec A n) →
reverse (x ∷ xs) ≅ reverse xs ++ [ x ]
TL;DR: How to prove unfold-reverse?
The rest of this question outlines approaches to doing so and explains what problems surface.
The type of this property is very similar to the List counter part in Data.List.Properties. The proof involves a helper which roughly translates to:
open import Function
helper : ∀ {n m} → (xs : Vec A n) → (ys : Vec A m) →
foldl (Vec A ∘ (flip _+_ n)) (flip _∷_) xs ys ≅ reverse ys ++ xs
Trying to insert this helper in unfold-reverse fails, because the left hand reverse is a foldl application with Vec A ∘ suc as first argument whereas the left hand side of helper has a foldl application with Vec A ∘ (flip _+_ 1) as first argument. Even though suc ≗ flip _+_ 1 is readily available from Data.Nat.Properties.Simple, it cannot be used here as cong would need a non-pointwise equality here and we don't have extensionality without further assumptions.
Removing the flip from flip _+_ n in helper yields a type error, so that is no option either.
Any other ideas?
The Data.Vec.Properties module contains this function:
foldl-cong : ∀ {a b} {A : Set a}
{B₁ : ℕ → Set b}
{f₁ : ∀ {n} → B₁ n → A → B₁ (suc n)} {e₁}
{B₂ : ℕ → Set b}
{f₂ : ∀ {n} → B₂ n → A → B₂ (suc n)} {e₂} →
(∀ {n x} {y₁ : B₁ n} {y₂ : B₂ n} →
y₁ ≅ y₂ → f₁ y₁ x ≅ f₂ y₂ x) →
e₁ ≅ e₂ →
∀ {n} (xs : Vec A n) →
foldl B₁ f₁ e₁ xs ≅ foldl B₂ f₂ e₂ xs
foldl-cong _ e₁=e₂ [] = e₁=e₂
foldl-cong {B₁ = B₁} f₁=f₂ e₁=e₂ (x ∷ xs) =
foldl-cong {B₁ = B₁ ∘ suc} f₁=f₂ (f₁=f₂ e₁=e₂) xs
Here is more or less elaborated solution:
unfold-reverse : {A : Set} → (x : A) → {n : ℕ} → (xs : Vec A n) →
reverse (x ∷ xs) ≅ reverse xs ++ (x ∷ [])
unfold-reverse x xs = begin
foldl (Vec _ ∘ _+_ 1) (flip _∷_) (x ∷ []) xs
≅⟨ (foldl-cong
{B₁ = Vec _ ∘ _+_ 1}
{f₁ = flip _∷_}
{e₁ = x ∷ []}
{B₂ = Vec _ ∘ flip _+_ 1}
{f₂ = flip _∷_}
{e₂ = x ∷ []}
(λ {n} {a} {as₁} {as₂} as₁≅as₂ -> {!!})
refl
xs) ⟩
foldl (Vec _ ∘ flip _+_ 1) (flip _∷_) (x ∷ []) xs
≅⟨ helper (x ∷ []) xs ⟩
reverse xs ++ x ∷ []
∎
Note, that only B₁ and B₂ are distinct in the arguments of the foldl-cong function. After simplifying context in the hole we have
Goal: a ∷ as₁ ≅ a ∷ as₂
————————————————————————————————————————————————————————————
as₁≅as₂ : as₁ ≅ as₂
as₂ : Vec A (n + 1)
as₁ : Vec A (1 + n)
a : A
n : ℕ
A : Set
So we need to prove, that at each recursive call adding an element to an accumulator of type Vec A (n + 1) is equal to adding an element to an accumulator of type Vec A (1 + n), and then results of two foldls are equal. The proof itself is simple. Here is the whole code:
open import Function
open import Relation.Binary.HeterogeneousEquality
open import Data.Nat
open import Data.Vec
open import Data.Nat.Properties.Simple
open import Data.Vec.Properties
open ≅-Reasoning
postulate
helper : ∀ {n m} {A : Set} (xs : Vec A n) (ys : Vec A m)
-> foldl (Vec A ∘ flip _+_ n) (flip _∷_) xs ys ≅ reverse ys ++ xs
cong' : ∀ {α β γ} {I : Set α} {i j : I}
-> (A : I -> Set β) {B : {k : I} -> A k -> Set γ} {x : A i} {y : A j}
-> i ≅ j
-> (f : {k : I} -> (x : A k) -> B x)
-> x ≅ y
-> f x ≅ f y
cong' _ refl _ refl = refl
unfold-reverse : {A : Set} → (x : A) → {n : ℕ} → (xs : Vec A n) →
reverse (x ∷ xs) ≅ reverse xs ++ (x ∷ [])
unfold-reverse x xs = begin
foldl (Vec _ ∘ _+_ 1) (flip _∷_) (x ∷ []) xs
≅⟨ (foldl-cong
{B₁ = Vec _ ∘ _+_ 1}
{f₁ = flip _∷_}
{e₁ = x ∷ []}
{B₂ = Vec _ ∘ flip _+_ 1}
{f₂ = flip _∷_}
{e₂ = x ∷ []}
(λ {n} {a} {as₁} {as₂} as₁≅as₂ -> begin
a ∷ as₁
≅⟨ cong' (Vec _) (sym (≡-to-≅ (+-comm n 1))) (_∷_ a) as₁≅as₂ ⟩
a ∷ as₂
∎)
refl
xs) ⟩
foldl (Vec _ ∘ flip _+_ 1) (flip _∷_) (x ∷ []) xs
≅⟨ helper (x ∷ []) xs ⟩
reverse xs ++ x ∷ []
∎

Termination check on list merge

Agda 2.3.2.1 can't see that the following function terminates:
open import Data.Nat
open import Data.List
open import Relation.Nullary
merge : List ℕ → List ℕ → List ℕ
merge (x ∷ xs) (y ∷ ys) with x ≤? y
... | yes p = x ∷ merge xs (y ∷ ys)
... | _ = y ∷ merge (x ∷ xs) ys
merge xs ys = xs ++ ys
Agda wiki says that it's OK for the termination checker if the arguments on recursive calls decrease lexicographically. Based on that it seems that this function should also pass. So what am I missing here? Also, is it maybe OK in previous versions of Agda? I've seen similar code on the Internet and no one mentioned termination issues there.
I cannot give you the reason why exactly this happens, but I can show you how to cure the symptoms. Before I start: This is a known problem with the termination checker. If you are well-versed in Haskell, you could take a look at the source.
One possible solution is to split the function into two: first one for the case where the first argument gets smaller and second for the second one:
mutual
merge : List ℕ → List ℕ → List ℕ
merge (x ∷ xs) (y ∷ ys) with x ≤? y
... | yes _ = x ∷ merge xs (y ∷ ys)
... | no _ = y ∷ merge′ x xs ys
merge xs ys = xs ++ ys
merge′ : ℕ → List ℕ → List ℕ → List ℕ
merge′ x xs (y ∷ ys) with x ≤? y
... | yes _ = x ∷ merge xs (y ∷ ys)
... | no _ = y ∷ merge′ x xs ys
merge′ x xs [] = x ∷ xs
So, the first function chops down xs and once we have to chop down ys, we switch to the second function and vice versa.
Another (perhaps surprising) option, which is also mentioned in the issue report, is to introduce the result of recursion via with:
merge : List ℕ → List ℕ → List ℕ
merge (x ∷ xs) (y ∷ ys) with x ≤? y | merge xs (y ∷ ys) | merge (x ∷ xs) ys
... | yes _ | r | _ = x ∷ r
... | no _ | _ | r = y ∷ r
merge xs ys = xs ++ ys
And lastly, we can perform the recursion on Vectors and then convert back to List:
open import Data.Vec as V
using (Vec; []; _∷_)
merge : List ℕ → List ℕ → List ℕ
merge xs ys = V.toList (go (V.fromList xs) (V.fromList ys))
where
go : ∀ {n m} → Vec ℕ n → Vec ℕ m → Vec ℕ (n + m)
go {suc n} {suc m} (x ∷ xs) (y ∷ ys) with x ≤? y
... | yes _ = x ∷ go xs (y ∷ ys)
... | no _ rewrite lem n m = y ∷ go (x ∷ xs) ys
go xs ys = xs V.++ ys
However, here we need a simple lemma:
open import Relation.Binary.PropositionalEquality
lem : ∀ n m → n + suc m ≡ suc (n + m)
lem zero m = refl
lem (suc n) m rewrite lem n m = refl
We could also have go return List directly and avoid the lemma altogether:
merge : List ℕ → List ℕ → List ℕ
merge xs ys = go (V.fromList xs) (V.fromList ys)
where
go : ∀ {n m} → Vec ℕ n → Vec ℕ m → List ℕ
go (x ∷ xs) (y ∷ ys) with x ≤? y
... | yes _ = x ∷ go xs (y ∷ ys)
... | no _ = y ∷ go (x ∷ xs) ys
go xs ys = V.toList xs ++ V.toList ys
The first trick (i.e. split the function into few mutually recursive ones) is actually quite good to remember. Since the termination checker doesn't look inside the definitions of other functions you use, it rejects a great deal of perfectly fine programs, consider:
data Rose {a} (A : Set a) : Set a where
[] : Rose A
node : A → List (Rose A) → Rose A
And now, we'd like to implement mapRose:
mapRose : ∀ {a b} {A : Set a} {B : Set b} →
(A → B) → Rose A → Rose B
mapRose f [] = []
mapRose f (node t ts) = node (f t) (map (mapRose f) ts)
The termination checker, however, doesn't look inside the map to see if it doesn't do anything funky with the elements and just rejects this definition. We must inline the definition of map and write a pair of mutually recursive functions:
mutual
mapRose : ∀ {a b} {A : Set a} {B : Set b} →
(A → B) → Rose A → Rose B
mapRose f [] = []
mapRose f (node t ts) = node (f t) (mapRose′ f ts)
mapRose′ : ∀ {a b} {A : Set a} {B : Set b} →
(A → B) → List (Rose A) → List (Rose B)
mapRose′ f [] = []
mapRose′ f (t ∷ ts) = mapRose f t ∷ mapRose′ f ts
Usually, you can hide most of the mess in a where declaration:
mapRose : ∀ {a b} {A : Set a} {B : Set b} →
(A → B) → Rose A → Rose B
mapRose {A = A} {B = B} f = go
where
go : Rose A → Rose B
go-list : List (Rose A) → List (Rose B)
go [] = []
go (node t ts) = node (f t) (go-list ts)
go-list [] = []
go-list (t ∷ ts) = go t ∷ go-list ts
Note: Declaring signatures of both functions before they are defined can be used instead of mutual in newer versions of Agda.
Update: The development version of Agda got an update to the termination checker, I'll let the commit message and release notes speak for themselves:
A revision of call graph completion that can deal with arbitrary termination depth.
This algorithm has been sitting around in MiniAgda for some time,
waiting for its great day. It is now here!
Option --termination-depth can now be retired.
And from the release notes:
Termination checking of functions defined by 'with' has been improved.
Cases which previously required --termination-depth (now obsolete!)
to pass the termination checker (due to use of 'with') no longer
need the flag. For example
merge : List A → List A → List A
merge [] ys = ys
merge xs [] = xs
merge (x ∷ xs) (y ∷ ys) with x ≤ y
merge (x ∷ xs) (y ∷ ys) | false = y ∷ merge (x ∷ xs) ys
merge (x ∷ xs) (y ∷ ys) | true = x ∷ merge xs (y ∷ ys)
This failed to termination check previously, since the 'with'
expands to an auxiliary function merge-aux:
merge-aux x y xs ys false = y ∷ merge (x ∷ xs) ys
merge-aux x y xs ys true = x ∷ merge xs (y ∷ ys)
This function makes a call to merge in which the size of one of the
arguments is increasing. To make this pass the termination checker
now inlines the definition of merge-aux before checking, thus
effectively termination checking the original source program.
As a result of this transformation doing 'with' on a variable no
longer preserves termination. For instance, this does not
termination check:
bad : Nat → Nat
bad n with n
... | zero = zero
... | suc m = bad m
And indeed, your original function now passes the termination check!

Implicit arguments and applying a function to the tail-part of fixed-size-vectors

I wrote an Agda-function applyPrefix to apply a fixed-size-vector-function to the initial part of a longer vector where the vector-sizes m, n and k may stay implicit. Here's the definition together with a helper-function split:
split : ∀ {A m n} → Vec A (n + m) → (Vec A n) × (Vec A m)
split {_} {_} {zero} xs = ( [] , xs )
split {_} {_} {suc _} (x ∷ xs) with split xs
... | ( ys , zs ) = ( (x ∷ ys) , zs )
applyPrefix : ∀ {A n m k} → (Vec A n → Vec A m) → Vec A (n + k) → Vec A (m + k)
applyPrefix f xs with split xs
... | ( ys , zs ) = f ys ++ zs
I need a symmetric function applyPostfix which applies a fixed-size-vector-function to the tail-part of a longer vector.
applyPostfix ∀ {A n m k} → (Vec A n → Vec A m) → Vec A (k + n) → Vec A (k + m)
applyPostfix {k = k} f xs with split {_} {_} {k} xs
... | ( ys , zs ) = ys ++ (f zs)
As the definition of applyPrefix already shows, the k-Argument cannot stay implicit when applyPostfix is used. For example:
change2 : {A : Set} → Vec A 2 → Vec A 2
change2 ( x ∷ y ∷ [] ) = (y ∷ x ∷ [] )
changeNpre : {A : Set}{n : ℕ} → Vec A (2 + n) → Vec A (2 + n)
changeNpre = applyPrefix change2
changeNpost : {A : Set}{n : ℕ} → Vec A (n + 2) → Vec A (n + 2)
changeNpost = applyPost change2 -- does not work; n has to be provided
Does anyone know a technique, how to implement applyPostfix so that the k-argument may stay implicit when using applyPostfix?
What I did is proofing / programming:
lem-plus-comm : (n m : ℕ) → (n + m) ≡ (m + n)
and use that lemma when defining applyPostfix:
postfixApp2 : ∀ {A}{n m k : ℕ} → (Vec A n → Vec A m) → Vec A (k + n) → Vec A (k + m)
postfixApp2 {A} {n} {m} {k} f xs rewrite lem-plus-comm n k | lem-plus-comm k n | lem-plus-comm k m | lem-plus-comm m k = reverse (drop {n = n} (reverse xs)) ++ f (reverse (take {n = n} (reverse xs)))
Unfortunately, this didnt help, since I use the k-Parameter for calling the lemma :-(
Any better ideas how to avoid k to be explicit? Maybe I should use a snoc-View on Vectors?
What you can do is to give postfixApp2 the same type as applyPrefix.
The source of the problem is that a natural number n can be unified with p + q only if p is known. This is because + is defined via induction on the first argument.
So this one works (I'm using the standard-library version of commutativity on +):
+-comm = comm
where
open IsCommutativeSemiring isCommutativeSemiring
open IsCommutativeMonoid +-isCommutativeMonoid
postfixApp2 : {A : Set} {n m k : ℕ}
→ (Vec A n → Vec A m)
→ Vec A (n + k) → Vec A (m + k)
postfixApp2 {A} {n} {m} {k} f xs rewrite +-comm n k | +-comm m k =
applyPostfix {k = k} f xs
Yes, I'm reusing the original applyPostfix here and just give it a different type by rewriting twice.
And testing:
changeNpost : {A : Set} {n : ℕ} → Vec A (2 + n) → Vec A (2 + n)
changeNpost = postfixApp2 change2
test : changeNpost (1 ∷ 2 ∷ 3 ∷ 4 ∷ []) ≡ 1 ∷ 2 ∷ 4 ∷ 3 ∷ []
test = refl

Resources