是否存在Powerset-over-Reader monad?

时间:2017-02-16 00:20:13

标签: haskell monads reader-monad

用于环境共享和非确定性的规范“Monad实例”如下(使用伪Haskell,因为Haskell的Data.Set当然不是monadic):

eta :: a -> r -> {a} -- '{a}' means the type of a set of a's
eta x = \r -> {x}

bind :: (r -> {a}) -> (a -> r -> {b}) -> r -> {b}
m `bind` f = \r -> {v | x ∈ m r, v ∈ f x r}

通常,当尝试将像Powerset(List,Writer等)这样的'容器'monad与第二个monad m(这里,粗略地,Reader)组合时,一个'包裹'm容器monad,如上所述。

我想知道以下潜在的Powerset-over-Reader规范:

eta' :: a -> {r -> a}
eta' x = {\r -> x}

bind' :: {r -> a} -> (a -> {r -> b}) -> {r -> b}
m `bind'` f = {rb | x <- m, ∀r: ∃rb' ∈ f (x r): rb r == rb' r}

这似乎并不那么疯狂(我确实认识到GHCi无法检查rb r == rb' rrb的{​​{1}},但是rb'很复杂到难以(对我来说)检查monad法律是否成立。

那么,我的问题是,bind'eta'是否真的是一元的 - 如果不是,那么哪些法律被违反,以及这可能是什么样的意外行为到?

假设bind'eta'不是monadic,第二个问题是人们如何确定 是否存在这些类型的函数?

2 个答案:

答案 0 :(得分:8)

有趣的问题。这是我的看法 - 让我们看看我是不是在任何地方蠢蠢欲动!

首先,我将在(稍微不那么伪)Haskell中拼写您的签名:

return :: a -> PSet (r -> a)
(>>=) :: PSet (r -> a) -> (a -> PSet (r -> b)) -> PSet (r -> b))

在继续之前,值得一提的是两个实际的并发症。首先,正如您已经观察到的那样,由于Eq和/或Ord约束,给出集FunctorMonad个实例是非常重要的;无论如何,there are ways around it。其次,更令人担忧的是,根据您为(>>=)提议的类型,有必要从a 中提取PSet (r -> a),而不会明显提供r s < / em> - 或者换句话说,您的(>>=)要求遍历函数functor (->) r。当然,这在一般情况下是不可能的,并且即使在可能的情况下也往往是不切实际的 - 至少就Haskell而言。在任何情况下,出于我们的推测目的,假设我们可以通过将函数应用于所有可能的(->) r值来遍历r。我将通过一个手绢universe :: PSet r集来表明这一点,这个集以this package为荣。我还将使用universe :: PSet (r -> b),并假设我们可以判断两个r -> b函数是否同意某个r,即使不需要Eq约束。 (伪Haskell确实变得非常假!)

初步评论,这里是我的方法的伪Haskell版本:

return :: a -> PSet (r -> a)
return x = singleton (const x)

(>>=) :: PSet (r -> a) -> (a -> PSet (r -> b)) -> PSet (r -> b))
m >>= f = unionMap (\x ->
    intersectionMap (\r ->
        filter (\rb -> 
            any (\rb' -> rb' r == rb r) (f (x r)))
            (universe :: PSet (r -> b)))
        (universe :: PSet r)) m
    where
    unionMap f = unions . map f
    intersectionMap f = intersections . map f

接下来,monad法则:

m >>= return = m
return y >>= f = f y
m >>= f >>= g = m >>= \y -> f y >>= g

(顺便说一句,在做这类事情时,最好记住我们正在使用的课程的其他演示文稿 - 在这种情况下,我们有join(>=>)作为(>>=)的替代方案 - 因为切换演示文稿可能会使您选择的实例更加愉快。在这里,我将坚持(>>=) Monad的演示文稿。)

继续第一部法律......

m >>= return = m
m >>= return -- LHS
unionMap (\x ->
    intersectionMap (\r ->
        filter (\rb -> 
            any (\rb' -> rb' r == rb r) (singleton (const (x r))))
            (universe :: PSet (r -> b)))
        (universe :: PSet r)) m
unionMap (\x ->
    intersectionMap (\r ->
        filter (\rb -> 
            const (x r) r == rb r)
            (universe :: PSet (r -> b)))
        (universe :: PSet r)) m
unionMap (\x ->
    intersectionMap (\r ->
        filter (\rb -> 
            x r == rb r)
            (universe :: PSet (r -> b)))
        (universe :: PSet r)) m
-- In other words, rb has to agree with x for all r. 
unionMap (\x -> singleton x) m
m -- RHS

一个下来,两个去。

return y >>= f = f y
return y -- LHS
unionMap (\x ->
    intersectionMap (\r ->
        filter (\rb -> 
            any (\rb' -> rb' r == rb r) (f (x r)))
            (universe :: PSet (r -> b)))
        (universe :: PSet r)) (singleton (const y))
(\x ->
    intersectionMap (\r ->
        filter (\rb -> 
            any (\rb' -> rb' r == rb r) (f (x r)))
            (universe :: PSet (r -> b)))
        (universe :: PSet r)) (const y)
intersectionMap (\r ->
    filter (\rb -> 
        any (\rb' -> rb' r == rb r) (f (const y r)))
        (universe :: PSet (r -> b)))
    (universe :: PSet r)
intersectionMap (\r ->
    filter (\rb -> 
        any (\rb' -> rb' r == rb r) (f y)))
        (universe :: PSet (r -> b)))
    (universe :: PSet r)
-- This set includes all functions that agree with at least one function
-- from (f y) at each r.
因此,

return y >>= f可能比f y更大。我们违反了第二部法律;因此,我们没有monad - 至少没有这里提出的实例。

附录:这是一个实际的,可运行的函数实现,至少可以用于小型类型。它利用了前面提到的universe包。

{-# LANGUAGE GeneralizedNewtypeDeriving #-}
{-# LANGUAGE ScopedTypeVariables #-}
module FunSet where

import Data.Universe
import Data.Map (Map)
import qualified Data.Map as M
import Data.Set (Set)
import qualified Data.Set as S
import Data.Int
import Data.Bool

-- FunSet and its would-be monad instance

newtype FunSet r a = FunSet { runFunSet :: Set (Fun r a) }
    deriving (Eq, Ord, Show)

fsreturn :: (Finite a, Finite r, Ord r) => a -> FunSet r a
fsreturn x = FunSet (S.singleton (toFun (const x)))

-- Perhaps we should think of a better name for this...
fsbind :: forall r a b.
    (Ord r, Finite r, Ord a, Ord b, Finite b, Eq b)
    => FunSet r a -> (a -> FunSet r b) -> FunSet r b
fsbind (FunSet s) f = FunSet $
    unionMap (\x ->
        intersectionMap (\r ->
            S.filter (\rb ->
                any (\rb' -> funApply rb' r == funApply rb r)
                    ((runFunSet . f) (funApply x r)))
                (universeF' :: Set (Fun r b)))
            (universeF' :: Set r)) s

toFunSet :: (Finite r, Finite a, Ord r, Ord a) => [r -> a] -> FunSet r a
toFunSet = FunSet . S.fromList . fmap toFun

-- Materialised functions

newtype Fun r a = Fun { unFun :: Map r a }
    deriving (Eq, Ord, Show, Functor)

instance (Finite r, Ord r, Universe a) => Universe (Fun r a) where
    universe = fmap (Fun . (\f ->
        foldr (\x m ->
            M.insert x (f x) m) M.empty universe))
        universe

instance (Finite r, Ord r, Finite a) => Finite (Fun r a) where
    universeF = universe

funApply :: Ord r => Fun r a -> r -> a
funApply f r = maybe
    (error "funApply: Partial functions are not fun")
    id (M.lookup r (unFun f))

toFun :: (Finite r, Finite a, Ord r) => (r -> a) -> Fun r a
toFun f = Fun (M.fromList (fmap ((,) <$> id <*> f) universeF))

-- Set utilities

unionMap :: (Ord a, Ord b) => (a -> Set b) -> (Set a -> Set b)
unionMap f = S.foldl S.union S.empty . S.map f

-- Note that this is partial. Since for our immediate purposes the only
-- consequence is that r in FunSet r a cannot be Void, I didn't bother
-- with making it cleaner.
intersectionMap :: (Ord a, Ord b) => (a -> Set b) -> (Set a -> Set b)
intersectionMap f s = case ss of
    [] -> error "intersectionMap: Intersection of empty set of sets"
    _ -> foldl1 S.intersection ss
    where
    ss = S.toList (S.map f s)

universeF' :: (Finite a, Ord a) => Set a
universeF' = S.fromList universeF

-- Demo

main :: IO ()
main = do
    let andor = toFunSet [uncurry (&&), uncurry (||)]
    print andor -- Two truth tables
    print $ funApply (toFun (2+)) (3 :: Int8) -- 5
    print $ (S.map (flip funApply (7 :: Int8)) . runFunSet)
        (fsreturn (Just True)) -- fromList [Just True]
    -- First monad law demo
    print $ fsbind andor fsreturn == andor -- True
    -- Second monad law demo
    let twoToFour = [ bool (Left False) (Left True)
                    , bool (Left False) (Right False)]
        decider b = toFunSet
            (fmap (. bool (uncurry (&&)) (uncurry (||)) b) twoToFour)
    print $ fsbind (fsreturn True) decider == decider True -- False (!)

答案 1 :(得分:4)

验证Kleisli符号中的定律更容易一些。

kleisli' :: (a -> {r -> b}) -> (b -> {r -> c}) -> (a -> {r -> c})
g `kleisli'` f = \z -> {rb | x <- g z, ∀r: ∃rb' ∈ f (x r): rb r == rb' r}

让我们尝试验证return `kleisli'` f = f

(\a -> {\r->a}) `kleisli'` f = 
\z -> {rb | x <- {\r->z}, ∀r: ∃rb' ∈ f (x r): rb r == rb' r} = 
\z -> {rb | ∀r: ∃rb' ∈ f z: rb r == rb' r}

我们所有类型abcr都是Integerf x = {const x, const -x}(return `kleisli'` f) 5中有哪些功能?此设置应为f 5,即{const 5, const -5}

是吗?当然const 5const -5都在,但不仅如此。例如,\r->if even r then 5 else -5也在。