如何使二叉树拉链成为Comonad的一个实例?

时间:2014-08-27 05:47:38

标签: haskell zipper comonad

我想将二叉树拉链作为comonad的一个实例,但我无法弄清楚如何正确实现duplicate

这是我的尝试:

{-# LANGUAGE DeriveFunctor #-}
import Data.Function
import Control.Arrow
import Control.Comonad

data BinTree a
    = Leaf a
    | Branch a (BinTree a) (BinTree a)
      deriving (Functor, Show, Eq)

data Dir = L | R
    deriving (Show, Eq)

-- an incomplete binary tree, aka data context
data Partial a = Missing Dir (BinTree a) a
    deriving (Show, Eq, Functor)

-- BTZ for BinTree Zipper
newtype BTZ a = BTZ { getBTZ :: ([Partial a], BinTree a) }
    deriving (Show, Eq)

instance Functor BTZ where
    fmap f (BTZ (cs,t)) = BTZ (map (fmap f) cs, fmap f t)

-- | replace every node label with the zipper focusing on that node
dup :: BinTree a -> BinTree (BTZ a)
dup (Leaf v) = Leaf (BTZ ([], Leaf v))
dup t@(Branch v tl tr) = Branch (BTZ ([],t)) tlZ trZ
    where
        tlZ = fmap (BTZ . first (++ [Missing L tr v]) . getBTZ) (dup tl)
        trZ = fmap (BTZ . first (++ [Missing R tl v]) . getBTZ) (dup tr)

-- | extract root label
rootVal :: BinTree a -> a
rootVal (Leaf v) = v
rootVal (Branch v _ _) = v

-- | move zipper focus around
goUp, goLeft, goRight :: BTZ a -> BTZ a

goUp (BTZ ([], _)) = error "already at root"
goUp (BTZ (Missing wt t2 v:xs, t1)) = case wt of
    L -> BTZ (xs, Branch v t1 t2)
    R -> BTZ (xs, Branch v t2 t1)

goLeft z = let (cs,t) = getBTZ z in
    case t of
      Leaf _ -> error "already at leaf"
      Branch v t1 t2 -> BTZ (Missing L t2 v:cs, t1)

goRight z = let (cs,t) = getBTZ z in
    case t of
      Leaf _ -> error "already at leaf"
      Branch v t1 t2 -> BTZ (Missing R t1 v:cs, t2)

instance Comonad BTZ where
    extract (BTZ (_,t)) =
        case t of
          Leaf v -> v
          Branch v _ _ -> v

    duplicate z@(BTZ (cs, bt)) = case bt of
        Leaf _ -> BTZ (csZ, Leaf z) -- extract . duplicate = id
        Branch v tl tr ->
            -- for each subtree, use "dup" to build zippers,
            -- and attach the current focusing root(bt) and rest of the data context to it
            let tlZ = fmap (BTZ . first (++Missing L tr v :cs) . getBTZ) (dup tl)
                trZ = fmap (BTZ . first (++Missing R tl v :cs) . getBTZ) (dup tr)
             in BTZ (csZ, Branch z tlZ trZ)
        where
            -- go up and duplicate, we'll have a "BTZ (BTZ a)"
            -- from which we can grab "[Partial (BTZ a)]" out
            -- TODO: not sure if it works
            upZippers = take (length cs-1) . tail $ iterate goUp z
            csZ = fmap (head . fst . getBTZ . duplicate) upZippers

main :: IO ()
main = do
   let tr :: BTZ Int
       tr = rootVal $ dup (Branch 1 (Leaf 2) (Branch 3 (Leaf 4) (Leaf 5)))
       equalOnTr :: Eq a => (BTZ Int -> a) -> (BTZ Int -> a) -> Bool
       equalOnTr = (==) `on` ($ tr)
   print $ (extract . duplicate)      `equalOnTr` id
   print $ (fmap extract . duplicate) `equalOnTr` id
   print $ (duplicate . duplicate)    `equalOnTr` (fmap duplicate . duplicate)

一些解释:

  • BinTree a是二叉树数据类型,每个树节点都包含一个标签。
  • Partial a是具有左子树或右子树的二叉树。一堆Partial a 在我的代码中扮演数据上下文的角色。
  • BTZ代表BinTree拉链,我想制作Comonad的实例,它由数据上下文和聚焦子树组成。

要使其成为Comonad的实例,我的计划是实施extractduplicate,并通过获取一些随机二叉树来验证comonad属性是否成立。

extract很简单,只需将聚焦子树取出即可。

函数dup用作替换每个节点标签的辅助函数 树拉链专注于该节点。

对于duplicate z,节点标签本身必须为z,以便extract . duplicate == id成立。对于非叶节点,我使用dup来处理它们的子树,就好像它们没有父节点一样,然后z的当前焦点和其余的数据上下文会附加到这些拉链上。< / p>

到目前为止,前两个comonad属性已满足(extract . duplicate = idfmap extract . duplicate),但我不知道如何处理数据上下文。我目前所做的是拉链z并继续上升。在此过程中,我们采用每个数据上下文堆栈的顶部来构建新的数据上下文堆栈,这听起来是正确的,并且也是正确的类型([Partial (BTZ a)]。但我的方法不能满足第三定律。

鉴于上面的二叉树拉链的数据类型定义, 是否有可能使它成为Comonad的一个实例? 如果答案是肯定的,我的方法是否有问题?

1 个答案:

答案 0 :(得分:33)

在微积分中,莱布尼茨的符号比牛顿更少引起混淆,因为它明确了我们所区分的变量。事物的上下文是通过差异化给出的,因此我们必须注意所处的情境。这里有两个概念&#34;子结构&#34;在工作中:子树元素。他们每个人都有不同的(但相关的)概念&#34; context&#34;因此,拉链是一种东西及其背景。

您的BTZ类型显示为子树的拉链概念。但是,拉链comonadic构造适用于元素的拉链extract意味着&#34;在这里给出元素&#34 ;; duplicate表示&#34;用其上下文&#34;装饰每个元素。所以你需要元素上下文。令人困惑的是,对于这些二元树,元素拉链和子树拉链是同构的,但这是一个非常特殊的原因(即它们形成一个共同的comonad)。

通常,元素和子树拉链不同,例如,列表。如果我们从为列表构建element-zipper comonad开始,当我们回到树时,我们不太可能迷路。让我也尝试为其他人以及你自己填写更多的一般情况。

子列表上下文

[a]子列表 -contexts仅由[a]给出,是我们从子列表到整个列表的路径上传递的元素列表。 [3,4][1,2,3,4]的子列表上下文为[2,1]。递归数据的子节点上下文始终是表示您在从节点到根的路径上看到的内容的列表。每个步骤的类型由相对于递归变量的一个数据节点的公式的偏导数给出。所以这里

[a] = t where             -- t is the recursive variable standing for [a]
  t = 1 + a*t             -- lists of a are either [] or an (a : t) pair
∂/∂t (1 + a*t) = a        -- that's one step on a path from node to root
sublist contexts are [a]  -- a list of such steps

所以sublist-zipper是一对

data LinLZ a = LinLZ
  {  subListCtxt  :: [a]
  ,  subList      :: [a]
  }

我们可以编写将子列表插回其上下文的函数,反转备份路径

plugLinLZ :: LinLZ a -> [a]
plugLinLZ (LinLZ { subListCtxt = [],      subList = ys})  = ys
plugLinLZ (LinLZ { subListCtxt = x : xs,  subList = ys})
  = plugLinLZ (LinLZ { subListCtxt = xs,  subList = x : ys})

但我们无法LinLZ成为Comonad,因为(例如)来自

LinLZ { subListCtxt = [], subList = [] }

我们无法extract 元素(来自a的{​​{1}}),只有子列表。

列表元素上下文

列表元素上下文是一对列表:焦点元素之前的元素以及它之后的元素。递归结构中的元素上下文始终是一对:首先为存储元素的子节点提供子节点上下文,然后为其节点中的元素提供上下文。我们通过将节点的公式与代表元素的变量区分开来获得其节点中的元素。

LinLZ a

因此元素上下文由一对

给出
[a] = t where             -- t is the recursive variable standing for [a]
  t = 1 + a*t             -- lists of a are either [] or an (a : t) pair
∂/∂a (1 + a*t) = t = [a]  -- the context for the head element is the tail list

和元素拉链是通过将这样的上下文与元素&#34;在孔中组合而给出的。

type DL a =
  (  [a]     -- the sublist context for the node where the element is
  ,  [a]     -- the tail of the node where the element is
  )

您可以通过首先重新构建元素所在的子列表,给我们一个子列表拉链,然后将子列表插入其子列表,将这样的拉链转换回列表(从元素开始#34; out&#34;) -context。

data ZL a = ZL
  {  this :: a
  ,  between :: DL a
  }  deriving (Show, Eq, Functor)

将每个元素置于上下文

给定一个列表,我们可以将每个元素与其上下文配对。我们列出了我们可以进入的方式&#34;其中一个要素。我们这样开始,

outZL :: ZL a -> [a]
outZL (ZL { this = x, between = (zs, xs) })
  = plugLinLZ (LinLZ { subListCtxt = zs, subList = x : xs })

但真正的工作是由辅助函数完成的,该函数适用于上下文列表。

into :: [a] -> [ZL a]
into xs = moreInto (LinLZ { subListCtxt = [], subList = xs })

请注意,输出会回显当前moreInto :: LinLZ a -> [ZL a] moreInto (LinLZ { subListCtxt = _, subList = [] }) = [] moreInto (LinLZ { subListCtxt = zs, subList = x : xs }) = ZL { this = x, between = (zs, xs) } : moreInto (LinLZ { subListCtxt = x : zs, subList = xs }) 的形状。此外,subList位置的拉链也有 x。此外,用于装饰this = x的生成拉链具有xs和正确的上下文, 记录我们已经过了subList = xs。检测,

x

列表元素拉链的公共结构

我们已经看过如何从一个元素或一个可用元素中走出来。 comonadic结构告诉我们如何在元素之间移动,要么停留在我们所处的位置,要么移动到其他元素之一。

into [1,2,3,4] =
  [  ZL {this = 1, between = ([],[2,3,4])}
  ,  ZL {this = 2, between = ([1],[3,4])}
  ,  ZL {this = 3, between = ([2,1],[4])}
  ,  ZL {this = 4, between = ([3,2,1],[])}
  ]

instance Comonad ZL where 为我们提供了我们正在访问的元素。

extract

extract = this 拉链,我们将当前元素duplicate替换为整个当前拉链x(其zl)...

this = x

...我们在上下文中工作,展示如何重新聚焦每个元素。我们现有的 duplicate zl@(ZL { this = x, between = (zs, ys) }) = ZL { this = zl 让我们向内移动,但我们也必须移动moreInto ...

outward

...涉及沿着上下文返回,将元素移动到子列表中,如下所示

    ,  between =
         (  outward (LinLZ { subListCtxt = zs, subList = x : ys })
         ,  moreInto (LinLZ { subListCtxt = x : zs, subList = ys })
         )
    }

所以我们得到

    where
      outward (LinLZ { subListCtxt = [], subList = _ }) = []
      outward (LinLZ { subListCtxt = z : zs, subList = ys })
        =  ZL { this = z, between = (zs, ys) }
        :  outward (LinLZ { subListCtxt = zs, subList = z : ys })

其中duplicate ZL {this = 2, between = ([1],[3,4])} = ZL { this = ZL {this = 2, between = ([1],[3,4])} , between = ( [ ZL {this = 1, between = ([],[2,3,4])} ] , [ ZL {this = 3, between = ([2,1],[4])} , ZL {this = 4, between = ([3,2,1],[])} ] ) } 是&#34;留在this&#34;我们2&#34;转移到between&#34;并且&#34;转移到1或转移到3&#34;。

因此,comonadic结构向我们展示了如何在列表中的不同元素之间移动。子列表结构在查找元素所在的节点方面起着关键作用,但拉链结构4 d是元素拉链。

那么树木呢?

Digression:标记的树已经是comonads

让我重构你的二叉树类型来展示一些结构。从字面上看,让我们拉出标记叶子或叉子的元素作为一个共同因素。让我们也分离出解释这个叶子或叉子子树结构的仿函数(duplicate)。

TF

也就是说,每个树节点都有一个标签,无论是叶子还是叉子。

无论我们拥有每个节点都有标签和blob子结构的结构,我们都有一个comonad: cofree comonad 。让我再重构一次,抽象出data TF t = Leaf | Fork (t, t) deriving (Show, Eq, Functor) data BT a = a :& TF (BT a) deriving (Show, Eq, Functor) ......

TF

...所以我们之前有data CoFree f a = a :& f (CoFree f a) deriving (Functor) f。我们可以恢复我们的特定树木。

TF

现在,我们可以为所有人提供cofree comonad构建。由于每个子树都有一个根元素,因此每个元素都可以用树根进行修饰。

data TF t = Leaf | Fork (t, t) deriving (Show, Eq, Functor)
type BT = CoFree TF
deriving instance Show a => Show (BT a)
deriving instance Eq a => Eq (BT a)

让我们举个例子

instance Functor f => Comonad (CoFree f) where
  extract   (a :& _)     = a                         -- extract root element
  duplicate t@(a :& ft)  = t :& fmap duplicate ft    -- replace root element by whole tree

请参阅?每个元素都与其子树配对!

列表不会产生cofree comonad,因为并非每个节点都有一个元素:具体而言,aTree = 0 :& Fork ( 1 :& Fork ( 2 :& Leaf , 3 :& Leaf ) , 4 :& Leaf ) duplicate aTree = (0 :& Fork (1 :& Fork (2 :& Leaf,3 :& Leaf),4 :& Leaf)) :& Fork ( (1 :& Fork (2 :& Leaf,3 :& Leaf)) :& Fork ( (2 :& Leaf) :& Leaf , (3 :& Leaf) :& Leaf ) , (4 :& Leaf) :& Leaf ) 没有元素。在cofree comonad中,总有一个元素,你可以进一步了解树结构,但不能进一步

在元素拉链comonad中,总有一个元素,你可以看到它上下。

二叉树中的子树和元素上下文

代数

[]

所以我们可以定义

d/dt (TF t) = d/dt (1 + t*t) = 0 + (1*t + t*1)

在子结构&#34; blob of substructures&#34;里面写一个子树。在左边或右边。我们可以检查一下&#34;插入&#34;的工作原理。

type DTF t = Either ((), t) (t, ())

如果我们实例化plugF :: t -> DTF t -> TF t plugF t (Left ((), r)) = Fork (t, r) plugF t (Right (l, ())) = Fork (l, t) 并与节点标签配对,我们得到子树上下文的一步

t

在问题中与type BTStep a = (a, DTF (BT a)) 同构。

Partial

因此,plugBTinBT :: BT a -> BTStep a -> BT a plugBTinBT t (a, d) = a :& plugF t d 给出了另一个BT a内的子树 - 上下文。

元素上下文呢?好吧,每个元素都标记了一些子树,所以我们应该记录子树的上下文和元素标记的树的其余部分。

[BTStep a]

令人讨厌的是,我必须推出自己的data DBT a = DBT { below :: TF (BT a) -- the rest of the element's node , above :: [BTStep a] -- the subtree context of the element's node } deriving (Show, Eq) 实例。

Functor

现在我可以说一下元素拉链是什么。

instance Functor DBT where
  fmap f (DBT { above = a, below = b }) = DBT
    {  below = fmap (fmap f) b
    ,  above = fmap (f *** (either
         (Left   . (id *** fmap f))
         (Right  . (fmap f *** id)))) a  
    }

如果您正在思考&#34;新的是什么?&#34;,那么你是对的。我们有一个子树上下文data BTZ a = BTZ { here :: a , ctxt :: DBT a } deriving (Show, Eq, Functor) ,以及abovehere给出的子树。这是因为唯一的元素是那些标记节点的元素。 将节点拆分为元素及其上下文与将其拆分为其标签及其子结构blob相同。也就是说,我们得到了cofree comonads的巧合,但不是一般的。

然而,这种巧合只会分散注意力!正如我们在列表中看到的那样,我们不需要将元素拉链与subnode-zippers相同,以使元素拉链成为comonad。

遵循与上面列表相同的模式,我们可以用它的上下文装饰每个元素。这项工作由辅助函数完成,该函数累积了我们当前正在访问的子树上下文。

below

请注意,down :: BT a -> BT (BTZ a) down t = downIn t [] downIn :: BT a -> [BTStep a] -> BT (BTZ a) downIn (a :& ft) ads = BTZ { here = a, ctxt = DBT { below = ft, above = ads } } :& furtherIn a ft ads 会被专注于a的拉链取代。子树由另一个帮助者处理。

a

请参见furtherIn :: a -> TF (BT a) -> [BTStep a] -> TF (BT (BTZ a)) furtherIn a Leaf ads = Leaf furtherIn a (Fork (l, r)) ads = Fork ( downIn l ((a, Left ((), r)) : ads) , downIn r ((a, Right (l, ())) : ads) ) 保留树结构,但在访问子树时适当增长子树上下文。

让我们仔细检查。

furtherIn

请参阅?每个元素都用它的整个上下文来装饰,而不仅仅是它下面的树。

二叉树拉链形成Comonad

现在我们可以用它们的上下文装饰元素,让我们构建down aTree = BTZ { here = 0, ctxt = DBT { below = Fork (1 :& Fork (2 :& Leaf,3 :& Leaf),4 :& Leaf), above = []}} :& Fork ( BTZ { here = 1, ctxt = DBT { below = Fork (2 :& Leaf,3 :& Leaf), above = [(0,Left ((),4 :& Leaf))]}} :& Fork ( BTZ { here = 2, ctxt = DBT { below = Leaf, above = [(1,Left ((),3 :& Leaf)),(0,Left ((),4 :& Leaf))]}} :& Leaf , BTZ { here = 3, ctxt = DBT { below = Leaf, above = [(1,Right (2 :& Leaf,())),(0,Left ((),4 :& Leaf))]}} :& Leaf ) , BTZ { here = 4, ctxt = DBT { below = Leaf, above = [(0,Right (1 :& Fork (2 :& Leaf,3 :& Leaf),()))]}} :& Leaf) 实例。和以前一样......

Comonad

... instance Comonad BTZ where extract = here 告诉我们关注的元素,我们可以利用现有的机制进一步深入树,但我们需要建立新的工具包来探索我们向外移动的方式。

extract

要向外走,与列表一样,我们必须沿着通向根的路径向后移动。与列表一样,路径上的每个步骤都是我们可以访问的地方。

  duplicate z@(BTZ { here = a, ctxt = DBT { below = ft, above = ads }}) = BTZ
    {  here = z
    ,  ctxt = DBT
         {  below = furtherIn a ft ads  -- move somewhere below a
         ,  above = go_a (a :& ft) ads  -- go above a
         }
    } where

与列表不同,沿着该路径有可供探索的替代分支。无论路径存储未访问的子树,我们必须使用 上下文来装饰 元素。

    go_a t []          = []
    go_a t (ad : ads)  = go_ad t ad ads : go_a (plugBTinBT t ad) ads
    go_ad t (a, d) ads =
      (  BTZ { here = a, ctxt = DBT { below = plugF t d, above = ads } }  -- visit here
      ,  go_d t a d ads                                                   -- try other subtree
      )

所以现在我们已经解释了如何从任何元素位置重新聚焦到任何其他位置。

让我们看看。我们在这里访问了 go_d t a (Left ((), r)) ads = Left ((), downIn r ((a, Right (t, ())) : ads)) go_d t a (Right (l, ())) ads = Right (downIn l ((a, Left ((), t)) : ads), ())

1

通过测试一小部分数据的comonad法律,让我们检查一下:

duplicate (BTZ {here = 1, ctxt = DBT {
                below = Fork (2 :& Leaf,3 :& Leaf),
                above = [(0,Left ((),4 :& Leaf))]}}) =
  BTZ {here = BTZ {here = 1, ctxt = DBT {
                   below = Fork (2 :& Leaf,3 :& Leaf),
                   above = [(0,Left ((),4 :& Leaf))]}}, ctxt = DBT {
       below = Fork (BTZ {here = 2, ctxt = DBT {
                          below = Leaf,
                          above = [(1,Left ((),3 :& Leaf)),(0,Left ((),4 :& Leaf))]}} :& Leaf
                    ,BTZ {here = 3, ctxt = DBT {
                          below = Leaf,
                          above = [(1,Right (2 :& Leaf,())),(0,Left ((),4 :& Leaf))]}} :& Leaf
                   ),
       above = [(BTZ {here = 0, ctxt = DBT {
                      below = Fork (1 :& Fork (2 :& Leaf,3 :& Leaf),4 :& Leaf),
                      above = []}}
                ,Left ((),BTZ {here = 4, ctxt = DBT {
                               below = Leaf,
                               above = [(0,Right (1 :& Fork (2 :& Leaf,3 :& Leaf),()))]}} :& Leaf)
                )
               ]}}