如何说服ghc类型级别添加是可交换的(实现依赖类型反向)?

时间:2015-01-02 14:09:40

标签: haskell dependent-type

这不能编译,因为ghc告诉我Add不是单射的。如何告诉编译器Add实际上是可交换的(也许告诉它Add是单射的)?从无人教论文中可以看出,必须以某种方式提供代理。

{-# LANGUAGE DataKinds             #-}
{-# LANGUAGE TypeOperators         #-}
{-# LANGUAGE KindSignatures        #-}
{-# LANGUAGE GADTs                 #-}
{-# LANGUAGE RankNTypes            #-}
{-# LANGUAGE ScopedTypeVariables   #-}
{-# LANGUAGE TypeFamilies          #-}
{-# LANGUAGE UndecidableInstances  #-}

data Nat = Z | S Nat

type family Add a b where
  Add  Z    n = n
  Add  n    Z = n
  Add (S n) k = S (Add n k)

data VecList n a where
  Nil  :: VecList Z a
  Cons :: a -> VecList n a -> VecList (S n) a

safeRev :: forall a n . VecList n a -> VecList n a
safeRev xs = safeRevAux Nil xs
  where
    safeRevAux :: VecList p a -> VecList q a -> VecList (Add p q) a
    safeRevAux acc Nil = acc
    safeRevAux acc (Cons y ys) = safeRevAux (Cons y acc) ys

人们可以做到这一点,但感觉就像我的口味一样,太多了。

{-# LANGUAGE TypeOperators         #-}
{-# LANGUAGE GADTs                 #-}
{-# LANGUAGE RankNTypes            #-}
{-# LANGUAGE ScopedTypeVariables   #-}
{-# LANGUAGE TypeFamilies          #-}

import Data.Proxy
import Data.Type.Equality

data Nat = Z | S Nat

type family n1 + n2 where
  Z + n2 = n2
  (S n1) + n2 = S (n1 + n2)

-- singleton for Nat
data SNat :: Nat -> * where
  SZero :: SNat Z
  SSucc :: SNat n -> SNat (S n)

-- inductive proof of right-identity of +
plus_id_r :: SNat n -> ((n + Z) :~: n)
plus_id_r SZero = Refl
plus_id_r (SSucc n) = gcastWith (plus_id_r n) Refl

-- inductive proof of simplification on the rhs of +
plus_succ_r :: SNat n1 -> Proxy n2 -> ((n1 + (S n2)) :~: (S (n1 + n2)))
plus_succ_r SZero _ = Refl
plus_succ_r (SSucc n1) proxy_n2 = gcastWith (plus_succ_r n1 proxy_n2) Refl

data VecList n a where
  V0  :: VecList Z a
  Cons :: a -> VecList n a -> VecList (S n) a

reverseList :: VecList n a -> VecList n a
reverseList V0 = V0
reverseList list = go SZero V0 list
  where
    go :: SNat n1 -> VecList n1  a-> VecList n2 a -> VecList (n1 + n2) a
    go snat acc V0 = gcastWith (plus_id_r snat) acc
    go snat acc (Cons h (t :: VecList n3 a)) =
      gcastWith (plus_succ_r snat (Proxy :: Proxy n3))
              (go (SSucc snat) (Cons h acc) t)

safeHead :: VecList (S n) a -> a
safeHead (Cons x _) = x

test = safeHead $ reverseList (Cons 'a' (Cons 'b' V0))

请参阅https://www.haskell.org/pipermail/haskell-cafe/2014-September/115919.html了解最初的想法。

编辑:

@ user3237465这是非常有趣的,更多的是我的想法 (虽然经过反思,我的问题可能并不是很好 配制)。

似乎我有“公理”

type family n1 :+ n2 where
  Z :+ n2 = n2
  (S n1) :+ n2 = S (n1 + n2)

因此可以产生类似

的证据
plus_id_r :: SNat n -> ((n + Z) :~: n)
plus_id_r SZero = Refl
plus_id_r (SSucc n) = gcastWith (plus_id_r n) Refl

我觉得这很简洁。我通常会推理这样的事情

  • 在上面的最后一句中,我们有SSucc n :: SNat(S k)所以n :: k
  • 因此我们需要证明S k + Z:〜:S k
  • 通过第二个“公理”S k + Z = S(k + Z)
  • 因此我们需要证明S(k + Z):〜:S k
  • plus_id_r n给出“证明”(k + Z):〜:k
  • 和Refl给出“证据”,即m~n => S m:〜:S n
  • 因此,我们可以使用gcastWith统一这些证明来提供所需的证明 结果

对于你的解决方案,你给出了“公理”

type family n :+ m where
    Z   :+ m = m
    S n :+ m = n :+ S m

使用这些,(n + Z):〜:n的证明将不起作用。

  • 在最后一句中,我们再次认为SSucc x的类型为SNat(S k)
  • 因此我们需要证明S k:+ Z:〜:S k
  • 通过第二个新的“公理”,我们得到S k + Z = k + S Z
  • 因此我们需要证明k + S Z:〜:S k
  • 所以我们有更复杂的东西需要证明: - (

我可以从新的第二个“公理”中产生证明 第二个“公理”(所以我的第二个“公理”现在是一个引理?)。

succ_plus_id :: SNat n1 -> SNat n2 -> (((S n1) :+ n2) :~: (S (n1 :+ n2)))
succ_plus_id SZero _ = Refl
succ_plus_id (SSucc n) m = gcastWith (succ_plus_id n (SSucc m)) Refl

所以现在我应该能够获得原始证据但我是 不确定目前的情况。

到目前为止,我的推理是否正确?

PS:ghc同意我为什么证明存在正确身份的证据不起作用

Could not deduce ((n1 :+ 'S 'Z) ~ 'S n1)
...
or from ((n1 :+ 'Z) ~ n1)

2 个答案:

答案 0 :(得分:4)

{-# LANGUAGE GADTs                #-}
{-# LANGUAGE KindSignatures       #-}
{-# LANGUAGE DataKinds            #-}
{-# LANGUAGE TypeFamilies         #-}
{-# LANGUAGE UndecidableInstances #-}
{-# LANGUAGE ExplicitForAll #-}

import Data.Type.Equality

data Nat = Z | S Nat

type family (n :: Nat) :+ (m :: Nat) :: Nat where
    Z   :+ m = m
    S n :+ m = n :+ S m

-- Singleton for Nat
data SNat :: Nat -> * where
  SZero :: SNat Z
  SSucc :: SNat n -> SNat (S n)

succ_plus_id :: SNat n1 -> SNat n2 -> (((S n1) :+ n2) :~: (S (n1 :+ n2)))
succ_plus_id SZero _ = Refl
succ_plus_id (SSucc n) m = gcastWith (succ_plus_id n (SSucc m)) Refl

plus_id_r :: SNat n -> ((n :+ Z) :~: n)
plus_id_r SZero = Refl
plus_id_r (SSucc x) = gcastWith (plus_id_r x) (succ_plus_id x SZero)

data Vec a n where
    Nil   :: Vec a Z
    (:::) :: a -> Vec a n -> Vec a (S n)

size :: Vec a n -> SNat n
size Nil         = SZero
size (_ ::: xs)  = SSucc $ size xs

elim0 :: SNat n -> (Vec a (n :+ Z) -> Vec a n)
elim0 n x = gcastWith (plus_id_r n) x

accrev :: Vec a n -> Vec a n
accrev x = elim0 (size x) $ go Nil x where
    go :: Vec a m -> Vec a n -> Vec a (n :+ m)
    go acc  Nil       = acc
    go acc (x ::: xs) = go (x ::: acc) xs

safeHead :: Vec a (S n) -> a
safeHead (x ::: _) = x

答案 1 :(得分:2)

您可以简化reverse的定义:

{-# LANGUAGE GADTs, KindSignatures, DataKinds    #-}
{-# LANGUAGE TypeFamilies, UndecidableInstances  #-}
{-# LANGUAGE TypeOperators                       #-}

data Nat = Z | S Nat

data Vec a n where
    Nil   :: Vec a Z
    (:::) :: a -> Vec a n -> Vec a (S n)

type family n :+ m where
    Z   :+ m = m
    S n :+ m = n :+ S m

elim0 :: Vec a (n :+ Z) -> Vec a n
elim0 = undefined

accrev :: Vec a n -> Vec a n
accrev = elim0 . go Nil where
    go :: Vec a m -> Vec a n -> Vec a (n :+ m)
    go acc  Nil       = acc
    go acc (x ::: xs) = go (x ::: acc) xs

(:+)运算符是根据(:::)运算符定义的。 (:::)案例中的统一进行如下:

x ::: xs会导致n成为S n。因此,结果的类型变为Vec a (S n :+ m),或者在β减少后Vec a (n :+ S m)。而

x ::: acc         :: Vec a (S m)
xs                :: Vec a  n
go (x ::: acc) xs :: Vec a (n :+ S m)

所以我们有一场比赛。但是现在您需要定义elim0 :: Vec a (n :+ Z) -> Vec a n,这需要您提问的证明。

Agda中的整个代码:http://lpaste.net/117679


顺便说一下,在任何情况下你都需要证据,这不是事实。以下是Agda标准库中reverse定义的方式:

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

reverse : ∀ {a n} {A : Set a} → Vec A n → Vec A n
reverse {A = A} = foldl (Vec A) (λ rev x → x ∷ rev) []

那是因为foldl带有关于_⊕_行为的其他类型信息,因此您在每一步都满足了类型检查器,并且不需要证明。