Counterexamples of Type Classes

by Phil Freeman on 2015/12/06

In Haskell and PureScript, we are familiar with certain standard hierarchies of type classes. For example, here are some of the type classes provided by the PureScript core libraries:

• Semigroup, Monoid
• Semiring, Ring, CommutativeRing, EuclideanRing, Field
• Functor, Apply, Applicative, Bind, Monad
• Alt, Plus, Alternative
• Semigroupoid, Category
• Profunctor, Strong, Arrow

For each of these hierarchies, we have examples of each class, and in many cases, free constructions for superclass relationships.

However, it is also instructive to look at counterexamples, of types which inhabit a superclass, but not a subclass, to understand why these refinements are useful.

The Monoid Hierarchy

The Monoid type class from Haskell is refined in PureScript, by splitting out the append operation into the Semigroup class:

class Semigroup m where
append :: m -> m -> m

class (Semigroup m) <= Monoid m where
mempty :: m

A Semigroup which is not a Monoid

Non-empty lists form a Semigroup, where append is given by concatenating two lists:

data NonEmpty a = NonEmpty a (List a)

In fact, NonEmpty a is the free semigroup for the type a.

Notes on Semigroup

In general, structures which lack empty elements might have instances of Semigroup but not of Monoid.

The free construction of a Monoid from a Semigroup is to simply add an empty element:

data FreeMonoid s = Empty | NonEmpty s

instance freeMonoidFromSemigroup :: (Semigroup s) => Monoid (FreeMonoid s)

The two free constructions NonEmpty and FreeMonoid compose to give something isomorphic to List a, the free Monoid for the type a.

An example of an operation which typically uses a Monoid instance, but which only requires a Semigroup is Haskell's foldl1 from Data.List. Since we are always folding at least one element, we only need an append operation, not an empty element.

foldl1 :: forall s. (Semigroup s) => NonEmpty s -> s

The applicative validation functor provides an example of a data structure which can be generalized to work with any Semigroup:

data Validation s a = Validation (Either s a)

instance (Semigroup s) => Applicative (Validation s)

Here, we can use the append operation to collect multiple errors, but we do not require a full Monoid. For example, Validation (NonEmpty String) might be used to collect a non-empty list of validation errors represented as strings.

Validation is not a Monad, unlike Either, since there is no implementation of >>= which is compatible with the Applicative implementation.

The Field Hierarchy

The Field type class from Haskell is refined into a number of type classes in PureScript:

class Semiring a where
add  :: a -> a -> a
zero :: a
mul  :: a -> a -> a
one  :: a

class Semiring a <= Ring a where
sub :: a -> a -> a

class Ring a <= CommutativeRing a where

class CommutativeRing a <= EuclideanRing a where
div :: a -> a -> a
mod :: a -> a -> a
degree :: a -> Int

class Ring a <= DivisionRing a where
recip :: a -> a

class EuclideanRing a <= Field a

The Semiring class specifies the addition and multiplication operations, while subtraction is split into the Ring class.

Commutative rings are specified by the CommutativeRing class.

Division and the modulus operator are broken into the EuclideanRing class.

DivisionRing specifies an alternative to division in terms of a reciprocal function, and requires that recip provides multiplicative inverses. A DivisionRing need not be commutative.

Finally, Field requires that multiplication is commutative.

A Semiring which is not a Ring

Natural numbers form a Semiring, but do not form a Ring, since they lack additive inverses:

data Nat = Zero | Succ Nat

A Ring which is not a DivisionRing

The integers provide an example of a Ring which is not a DivisionRing, since non-zero integers do not have multiplicative inverses.

The integers are also an example of a EuclideanRing which is not a Field.

A CommutativeRing which is not a EuclideanRing

Number theory provides examples of commutative rings which fail to be Euclidean. For example, rings which are not principal ideal domains cannot be Euclidean.

Wikipedia provides some additional counterexamples.

A DivisionRing which is not a CommutativeRing

The quaternions are an example of a structure with multiplicative inverses but non-commutative multiplication. They form a DivisionRing, but not a CommutativeRing.

Notes on Semiring

There is a free Semiring which can be constructed from any base type, formed by using distributivity to represent terms in "+-normal form":

data FreeSemiring a = FreeSemiring (List (List a))

instance freeSemiring :: Semiring (FreeSemiring a)

This type is implemented in the purescript-semirings library.

The free semiring may be interpreted in any semiring:

liftFree :: forall s a. (Semiring s) => (a -> s) -> Free a -> s

giving a form of abstract interpretation of the semiring operations.

Here are two examples of data structures which can be usefully generalized to work with an arbitrary Semiring:

• The familiar probability monad newtype Prob a = Prob (List (Tuple Number a)) can be generalized to use a Semiring of probabilities: data GProb s a = GProb (List (Tuple s a)). Choosing different Semirings recovers other less-familiar monads such as a monad for describing quantum wavefunctions (take s to be the Semiring of complex numbers) and a monad for managing priority queues (take s to be the (Max, +) Semiring).
• The applicative validation functor has an Alternative instance for collecting errors in parallel whenever the errors form a Semiring.

Notes on Ring

We can construct a Ring from any Semiring in the same way as we construct the integers from the naturals, by representing differences:

data Diff a = Diff a a

instance Semiring a => Ring (Diff a)

We identify any two values of type Diff a which represent the same difference.

Notes on DivisionRing

We can formally add divisors to any Ring in much the same way:

data Ratio a = Ratio a a

instance Ring a => DivisionRing (Ratio a)

We disallow zero divisors, and identify any two values of type Ratio a which represent the same ratio.

Haskell's Monad hierarchy is further refined in PureScript, by the addition of the Apply class which contains the <*> operator (but not pure) and the Bind class which specifies the >>= operator:

class Functor f where
map :: forall a b. (a -> b) -> f a -> f b

class Functor f <= Apply f where
apply :: forall a b. f (a -> b) -> f a -> f b

class Apply f <= Applicative f where
pure :: forall a. a -> f a

class Apply m <= Bind m where
bind :: forall a b. m a -> (a -> m b) -> m b

class (Applicative m, Bind m) <= Monad m

A type constructor which is not a Functor

Functor forces the type variable of its argument to appear covariantly, so a simple example of a type constructor which is not a Functor is given by:

data Op a b = Op (b -> a)

Op can be made into a contravariant functor, described by the Contravariant type class, which we will not cover here.

A Functor which is not an Apply

There is a functor for any type constructor, given by the Yoneda lemma:

newtype Yoneda f a = Yoneda (forall r. (a -> r) -> f r)

Yoneda f is always a Functor, but cannot be made into an instance of Apply in general, since to implement <*>, we would require a function of type:

forall a b r. (forall r. ((a -> b) -> r) -> f r) ->
(forall r. (a        -> r) -> f r) ->
(b        -> r) -> f r

An Apply which is not an Applicative

The constant functor

data Const k a = Const k

can be made into an Apply (but not an Applicative) whenever k is an instance of Semigroup (but not Monoid).

An Applicative which is not a Monad

We have already seen that the validation functor is an example of an Applicative which is not a Monad, and therefore also an example of an Apply which is not a Bind.

Another example is given by the ZipList functor, where apply is implemented using zipWith:

newtype ZipList a = ZipList (List a)

A Bind which is not a Monad

The Mapk data type has an implementation of Bind, since we can implement the map, apply and bind functions, but fails to be a Monad, since we cannot implement

pure :: forall k a. a -> Map k a

such that the monad laws will hold.

Notes on Apply

Apply is enough to implement a variant of zip:

zipA :: forall f a b. (Apply f) => f a -> f b -> f (Tuple a b)

where we can recover the original zip function by using the ZipList applicative.

Notes on Bind

Bind is enough to implement Kleisli composition:

(>=>) :: forall f a b c. (Bind f) => (a -> f b) -> (b -> f c) -> a -> f c

PureScript refines the Comonad type class to extract the extend function into its own Extend class:

class Functor w <= Extend w where
extend :: forall b a. (w a -> b) -> w a -> w b

class Extend w <= Comonad w where
extract :: forall a. w a -> a

An Extend which is not a Comonad

The function type gives an example of a Comonad whenever the domain is monoidal. If we relax the constraint to Semigroup, however, we only obtain an Extend:

instance extendFn :: Semigroup w => Extend ((->) w)

For example, the NonEmpty Unit semigroup can be thought of as non-zero natural numbers, and gives us an Extend instance which allows us to reannotate a semi-infinite tape by considering values (strictly) to the right of the current position.

Notes on Extend

Extend is enough to implement co-Kleisli composition:

(=>=) :: forall b a w c. (Extend w) => (w a -> b) -> (w b -> c) -> w a -> c

The Alternative Hierarchy

In PureScript, the Alternative class is refined so that the alternation operator <|> is broken out into the Alt class, the empty structure is defined by the Plus class, and Alternative is redefined as a combination of Applicative and Plus:

class Functor f <= Alt f where
alt :: forall a. f a -> f a -> f a

class Alt f <= Plus f where
empty :: forall a. f a

class (Applicative f, Plus f) <= Alternative f

An Alt which is not a Plus

NonEmpty is an example of an Alt in which alt is given by concatenation, but there is no Plus instance due to the lack of an empty element.

A Plus which is not an Alternative

Map k is an example of a Plus which is not Alternative, due to the lack of an Applicative instance

Notes on Alt

Alt is enough to define oneOf1, a variant on oneOf which requires at least one alternative:

oneOf1 :: forall f a. (Alt f) => NonEmpty (f a) -> f a

Any Alt can be used to construct a Semigroup (but not necessarily a Monoid) using <|>.

Notes on Plus

Plus is enough to define oneOf:

oneOf :: forall f a. (Plus f) => List (f a) -> f a

The Category Hierarchy

In PureScript, the Category type class is refined by splitting the compose operation into the Semigroupoid class:

class Semigroupoid a where
compose :: forall b c d. a c d -> a b c -> a b d

class Semigroupoid a <= Category a where
id :: forall t. a t t

A Semigroupoid which is not a Category

Tuple is an example of a Semigroupoid which is not a Category. We can think of Tuple as a strange function space in which we identify only a single point in the domain and a single point in the codomain.

In the same vein, another example is given by

newtype Relation a b = Relation (List (Tuple a b))

in which we choose a selection of pairs of points from the domain and codomain. Again, there is no Category instance, since we cannot construct

id :: forall a. Relation a a

in general, unless both the domain and codomain are enumerable.

Kleisli arrows form a Semigroupoid whenever f is a Bind, but not a Category in general, unless f is also a Monad:

newtype Kleisli f a b = Kleisli (a -> f b)
newtype Cokleisli f a b = Cokleisli (f a -> b)

A similar construction creates a Semigroupoid on Cokleisli f whenever f has Extend (but not necessarily Comonad).

Static arrows f (a -> b) form a Semigroupoid whenever f is an Apply, but not a Category in general, unless f is also Applicative:

newtype Static f a b = Static (f (a -> b))

Notes on Semigroupoid

Any Semigroupoid supports the composition of a NonEmpty list of morphisms:

compose1 :: forall s a. (Semigroupoid s) => NonEmpty (s a a) -> s a a

The Arrow Hierarchy

In PureScript, the Arrow type class is defined in terms of a combination of Category and strong profunctors:

class Profunctor p where
dimap :: forall a b c d. (a -> b) -> (c -> d) -> p b c -> p a d

class Profunctor p <= Strong p where
first :: forall a b c. p a b -> p (Tuple a c) (Tuple b c)

class (Category a, Strong a) <= Arrow a

A Profunctor which is not Strong

Here is a Profunctor which ignores its first type argument, so parametericity makes it impossible to define first:

newtype Ignore a b = Ignore b

A Strong Profunctor which is not an Arrow

Kleisli f is Strong whenever f is a Functor, but cannot be an Arrow unless f is a Monad, due to the lack of a Category instance.

A Category which is not an Arrow

Isomorphisms form a Category:

data Iso a b = Iso { to :: a -> b, from :: b -> a }

But Iso cannot be a Profunctor since a appears both covariantly and contravariantly.

Notes on Profunctor

A Profunctor supports a change of types using any isomorphism:

data Iso s a = Iso { to :: s -> a, from :: a -> s }

liftIso :: forall p a s. (Profunctor p) => Iso s a -> p a a -> p s s

Notes on Strong

A Strong profunctor supports a focussing operation using any lens:

data Lens s a = Lens { get :: s -> a, set :: s -> a -> s }

liftLens :: forall p a s. (Strong p) => Lens s a -> p a a -> p s s