{-# LANGUAGE CPP #-} {-# LANGUAGE KindSignatures #-} module BoundLaws where import Bound.Class #if !(MIN_VERSION_base(4,8,0)) import Control.Applicative hiding (Const(..)) #endif import Control.Monad {- What laws should Bound have? We need at least enough to make sure the typical Monad Exp instances are valid. Let's start by writing some generic Bound instances. -} newtype Const x (m :: * -> *) a = Const x instance Bound (Const x) where Const x >>>= _ = Const x newtype Identity (m :: * -> *) a = Id (m a) instance Bound Identity where Id ma >>>= f = Id (ma >>= f) data Product f g (m :: * -> *) a = f m a :*: g m a instance (Bound f, Bound g) => Bound (Product f g) where (fma :*: gma) >>>= f = (fma >>>= f) :*: (gma >>>= f) data Sum f g (m :: * -> *) a = Inl (f m a) | Inr (g m a) instance (Bound f, Bound g) => Bound (Sum f g) where Inl fma >>>= f = Inl (fma >>>= f) Inr gma >>>= f = Inr (gma >>>= f) {- Now we can actually write the typical Monad Exp instance generically (for theory, not practice), since sums and products and all of the above is plenty enough to specify an AST. -} data Exp (f :: (* -> *) -> * -> *) a = Var a | Branch (f (Exp f) a) instance Bound f => Functor (Exp f) where fmap = liftM instance Bound f => Applicative (Exp f) where pure = Var (<*>) = ap instance Bound f => Monad (Exp f) where return = Var Var a >>= f = f a Branch fE >>= f = Branch (fE >>>= f) {- Is this valid? Let's go to Agda and try to prove the Monad laws. left-return : ∀ {A B} (x : A)(f : A -> Exp F B) -> (return x >>= f) ≡ f x left-return x f = refl right-return : ∀ {A}(m : Exp F A) -> (m >>= return) ≡ m right-return (Var x) = refl right-return (Branch m) = cong Branch {!!}0 assoc : ∀ {A B C} (m : Exp F A) (k : A -> Exp F B) (h : B -> Exp F C) -> (m >>= (\ x -> k x >>= h)) ≡ ((m >>= k) >>= h) assoc (Var x) k h = refl assoc (Branch m) k h = cong Branch {!!}1 So the first one is fine, but we have two holes: ?0 : m >>>= return ≡ m ?1 : m >>>= (λ x → k x >>= h) ≡ (m >>>= k) >>>= h But all of the instances above respect these laws, and they are implied by the current law for monad transformers, we could just make them the Bound class laws. Btw these laws correspond to requiring (f m) to be an m-left module for every m [1], so we'd also get a law-abiding fmap for (f m). Bonus: composing pointwise (\m a -> f m (g m a)) would also create a valid Bound [1] Modules over Monads and Initial Semantics - http://web.math.unifi.it/users/maggesi/syn.pdf -}