F ⊣ G ⊣ H
F ⊣ G, G ⊣ H

Perhaps part of the reason they are so common is that (co)limits form one:

colim ⊣ Δ ⊣ lim

where Δ : C -> C^J is the diagonal functor, which takes objects in C to the constant functor returning that object. A version of this shows up in Haskell (with some extensions) and dependent type theories, as:

∃ ⊣ Const ⊣ ∀
Σ ⊣ Const ⊣ Π

where, if we only care about quantifying over a single variable, existential and sigma types can be seen as a left adjoint to a diagonal functor that maps types into constant type families (either over * for the first triple in Haskell, or some other type for the second in a dependently typed language), while universal and pi types can be seen as a right adjoint to the same.

It's not uncommon to see the above information in type theory discussion forums. But, there are a few cute properties and examples of adjoint triples that I haven't really seen come up in such contexts.

To begin, we can compose the two adjunctions involved, since the common functor ensures things match up. By calculating on the hom definition, we can see:

Hom(FGA, B)     Hom(GFA, B)
    ~=              ~=
Hom(GA, GB)     Hom(FA, HB)
    ~=              ~=
Hom(A, HGB)     Hom(A, GHB)

So there are two ways to compose the adjunctions, giving two induced adjunctions:

FG ⊣ HG,  GF ⊣ GH

And there is something special about these adjunctions. Note that FG is the comonad for the F ⊣ G adjunction, while HG is the monad for the G ⊣ H adjunction. Similarly, GF is the F ⊣ G monad, and GH is the G ⊣ H comonad. So each adjoint triple gives rise to two adjunctions between monads and comonads.

The second of these has another interesting property. We often want to consider the algebras of a monad, and coalgebras of a comonad. The (co)algebra operations with carrier A have type:

alg   : GFA -> A
coalg : A -> GHA

but these types are isomorphic according to the GF ⊣ GH adjunction. Thus, one might guess that GF monad algebras are also GH comonad coalgebras, and that in such a situation, we actually have some structure that can be characterized both ways. In fact this is true for any monad left adjoint to a comonad; [0] but all adjoint triples give rise to these.

The first adjunction actually turns out to be more familiar for the triple examples above, though. (Edit: [2]) If we consider the Σ ⊣ Const ⊣ Π adjunction, where:

Σ Π : (A -> Type) -> Type
Const : Type -> (A -> Type)

we get:

ΣConst : Type -> Type
ΣConst B = A × B
ΠConst : Type -> Type
ΠConst B = A -> B

So this is the familiar adjunction:

A × - ⊣ A -> -

But, there happens to be a triple that is a bit more interesting for both cases. It refers back to categories of functors vs. bare type constructors mentioned in previous posts. So, suppose we have a category called Con whose objects are (partially applied) type constructors (f, g) with kind * -> *, and arrows are polymorphic functions with types like:

 
forall x. f x -> g x
 

And let us further imagine that there is a similar category, called Func, except its objects are the things with Functor instances. Now, there is a functor:

U : Func -> Con

that 'forgets' the functor instance requirement. This functor is in the middle of an adjoint triple:

F ⊣ U ⊣ C
F, C : Con -> Func

where F creates the free functor over a type constructor, and C creates the cofree functor over a type constructor. These can be written using the types:

 
data F f a = forall e. F (e -> a) (f e)
newtype C f a = C (forall r. (a -> r) -> f r)
 

and these types will also serve as the types involved in the composite adjunctions:

FU ⊣ CU : Func -> Func
UF ⊣ UC : Con -> Con

Now, CU is a monad on functors, and the Yoneda lemma tells us that it is actually the identity monad. Similarly, FU is a comonad, and the co-Yoneda lemma tells us that it is the identity comonad (which makes sense, because identity is self-adjoint; and the above is why F and C are often named (Co)Yoneda in Haskell examples).

On the other hand, UF is a monad on type constructors (note, U isn't represented in the Haskell types; F and C just play triple duty, and the constraints on f control what's going on):

 
eta :: f a -> F f a
eta = F id
 
transform :: (forall x. f x -> g x) -> F f a -> F g a
transform tr (F g x) = F g (tr x)
 
mu :: F (F f) a -> F f a
mu (F g (F h x)) = F (g . h) x
 

and UC is a comonad:

 
epsilon :: C f a -> f a
epsilon (C e) = e id
 
transform' :: (forall x. f x -> g x) -> C f a -> C g a
transform' tr (C e) = C (tr . e)
 
delta :: C f a -> C (C f) a
delta (C e) = C $ \h -> C $ \g -> e (g . h)
 

These are not the identity (co)monad, but this is the case where we have algebras and coalgebras that are equivalent. So, what are the (co)algebras? If we consider UF (and unpack the definitions somewhat):

 
alg :: forall e. (e -> a, f e) -> f a
alg (id, x) = x
alg (g . h, x) = alg (g, alg (h, x))
 

and for UC:

 
coalg :: f a -> forall r. (a -> r) -> f r
coalg x id = x
coalg x (g . h) = coalg (coalg x h) g
 

in other words, (co)algebra actions of these (co)monads are (mangled) fmap implementations, and the commutativity requirements are exactly what is required to be a law abiding instance. So the (co)algebras are exactly the Functors. [1]

There are, of course, many other examples of adjoint triples. And further, there are even adjoint quadruples, which in turn give rise to adjoint triples of (co)monads. Hopefully this has sparked some folks' interest in finding and studying more interesting examples.

[0]: Another exmaple is A × - ⊣ A -> - where the A in question is a monoid. (Co)monad (co)algebras of these correspond to actions of the monoid on the carrier set.

[1]: This shouldn't be too surprising, because having a category of (co)algebraic structures that is equivalent to the category of (co)algebras of the (co)monad that comes from the (co)free-forgetful adjunction is the basis for doing algebra in category theory (with (co)monads, at least). However, it is somewhat unusual for a forgetful functor to have both a left and right adjoint. In many cases, something is either algebraic or coalgebraic, and not both.

[2]: Urs Schreiber informed me of an interesting interpretation of the ConstΣ ⊣ ConstΠ adjunction. If you are familiar with modal logic and the possible worlds semantics thereof, you can probably imagine that we could model it using something like P : W -> Type, where W is the type of possible worlds, and propositions are types. Then values of type Σ P demonstrate that P holds in particular worlds, while values of type Π P demonstrate that it holds in all worlds. Const turns these types back into world-indexed 'propositions,' so ConstΣ is the possibility modality and ConstΠ is the necessity modality.

(Advance note: for some continuous code to look at see this file.)

First, it'll help to talk about how some categories can work in Haskell. For any kind k made of * and (->), [0] we can define a category of type constructors. Objects of the category will be first-class [1] types of that kind, and arrows will be defined by the following type family:

 
newtype Transformer f g = Transform { ($$) :: forall i. f i ~> g i }
 
type family (~>) :: k -> k -> * where
  (~>) = (->)
  (~>) = Transformer
 
type a < -> b = (a -> b, b -> a)
type a < ~> b = (a ~> b, b ~> a)
 

So, for a base case, * has monomorphic functions as arrows, and categories for higher kinds have polymorphic functions that saturate the constructor:

 
  Int ~> Char = Int -> Char
  Maybe ~> [] = forall a. Maybe a -> [a]
  Either ~> (,) = forall a b. Either a b -> (a, b)
  StateT ~> ReaderT = forall s m a. StateT s m a -> ReaderT s m a
 

We can of course define identity and composition for these, and it will be handy to do so:

 
class Morph (p :: k -> k -> *) where
  id :: p a a
  (.) :: p b c -> p a b -> p a c
 
instance Morph (->) where
  id x = x
  (g . f) x = g (f x)
 
instance Morph ((~>) :: k -> k -> *)
      => Morph (Transformer :: (i -> k) -> (i -> k) -> *) where
  id = Transform id
  Transform f . Transform g = Transform $ f . g
 

These categories can be looked upon as the most basic substrates in Haskell. For instance, every type of kind * -> * is an object of the relevant category, even if it's a GADT or has other structure that prevents it from being nicely functorial.

The category for * is of course just the normal category of types and functions we usually call Hask, and it is fairly analogous to the category of sets. One common activity in category theory is to study categories of sets equipped with extra structure, and it turns out we can do this in Haskell, as well. And it even makes some sense to study categories of structures over any of these type categories.

When we equip our types with structure, we often use type classes, so that's how I'll do things here. Classes have a special status socially in that we expect people to only define instances that adhere to certain equational rules. This will take the place of equations that we are not able to state in the Haskell type system, because it doesn't have dependent types. So using classes will allow us to define more structures that we normally would, if only by convention.

So, if we have a kind k, then a corresponding structure will be σ :: k -> Constraint. We can then define the category (k,σ) as having objects t :: k such that there is an instance σ t. Arrows are then taken to be f :: t ~> u such that f "respects" the operations of σ.

As a simple example, we have:

 
  k = *
  σ = Monoid :: * -> Constraint
 
  Sum Integer, Product Integer, [Integer] :: (*, Monoid)
 
  f :: (Monoid m, Monoid n) => m -> n
    if f mempty = mempty
       f (m <> n) = f m <> f n
 

This is just the category of monoids in Haskell.

As a side note, we will sometimes be wanting to quantify over these "categories of structures". There isn't really a good way to package together a kind and a structure such that they work as a unit, but we can just add a constraint to the quantification. So, to quantify over all Monoids, we'll use 'forall m. Monoid m => ...'.

Now, once we have these categories of structures, there is an obvious forgetful functor back into the unadorned category. We can then look for free and cofree functors as adjoints to this. More symbolically:

 
  Forget σ :: (k,σ) -> k
  Free   σ :: k -> (k,σ)
  Cofree σ :: k -> (k,σ)
 
  Free σ ⊣ Forget σ ⊣ Cofree σ
 

However, what would be nicer (for some purposes) than having to look for these is being able to construct them all systematically, without having to think much about the structure σ.

Category theory gives a hint at this, too, in the form of Kan extensions. In category terms they look like:

  p : C -> C'
  f : C -> D
  Ran p f : C' -> D
  Lan p f : C' -> D

  Ran p f c' = end (c : C). Hom_C'(c', p c) ⇒ f c
  Lan p f c' = coend (c : c). Hom_C'(p c, c') ⊗ f c

where ⇒ is a "power" and ⊗ is a copower, which are like being able to take exponentials and products by sets (or whatever the objects of the hom category are), instead of other objects within the category. Ends and coends are like universal and existential quantifiers (as are limits and colimits, but ends and coends involve mixed-variance).

Some handy theorems relate Kan extensions and adjoint functors:

  if L ⊣ R
  then L = Ran R Id and R = Lan L Id

  if Ran R Id exists and is absolute
  then Ran R Id ⊣ R

  if Lan L Id exists and is absolute
  then L ⊣ Lan L Id

  Kan P F is absolute iff forall G. (G . Kan P F) ~= Kan P (G . F)

It turns out we can write down Kan extensions fairly generally in Haskell. Our restricted case is:

 
  p = Forget σ :: (k,σ) -> k
  f = Id :: (k,σ) -> (k,σ)
 
  Free   σ = Ran (Forget σ) Id :: k -> (k,σ)
  Cofree σ = Lan (Forget σ) Id :: k -> (k,σ)
 
  g :: (k,σ) -> j
  g . Free   σ = Ran (Forget σ) g
  g . Cofree σ = Lan (Forget σ) g
 

As long as the final category is like one of our type constructor categories, ends are universal quantifiers, powers are function types, coends are existential quantifiers and copowers are product spaces. This only breaks down for our purposes when g is contravariant, in which case they are flipped. For higher kinds, these constructions occur point-wise. So, we can break things down into four general cases, each with cases for each arity:

 
newtype Ran0 σ p (f :: k -> *) a =
  Ran0 { ran0 :: forall r. σ r => (a ~> p r) -> f r }
 
newtype Ran1 σ p (f :: k -> j -> *) a b =
  Ran1 { ran1 :: forall r. σ r => (a ~> p r) -> f r b }
 
-- ...
 
data RanOp0 σ p (f :: k -> *) a =
  forall e. σ e => RanOp0 (a ~> p e) (f e)
 
-- ...
 
data Lan0 σ p (f :: k -> *) a =
  forall e. σ e => Lan0 (p e ~> a) (f e)
 
data Lan1 σ p (f :: k -> j -> *) a b =
  forall e. σ e => Lan1 (p e ~> a) (f e b)
 
-- ...
 
data LanOp0 σ p (f :: k -> *) a =
  LanOp0 { lan0 :: forall r. σ r => (p r -> a) -> f r }
 
-- ...
 

The more specific proposed (co)free definitions are:

 
type family Free   :: (k -> Constraint) -> k -> k
type family Cofree :: (k -> Constraint) -> k -> k
 
newtype Free0 σ a = Free0 { gratis0 :: forall r. σ r => (a ~> r) -> r }
type instance Free = Free0
 
newtype Free1 σ f a = Free1 { gratis1 :: forall g. σ g => (f ~> g) -> g a }
type instance Free = Free1
 
-- ...
 
data Cofree0 σ a = forall e. σ e => Cofree0 (e ~> a) e
type instance Cofree = Cofree0
 
data Cofree1 σ f a = forall g. σ g => Cofree1 (g ~> f) (g a)
type instance Cofree = Cofree1
 
-- ...
 

We can define some handly classes and instances for working with these types, several of which generalize existing Haskell concepts:

 
class Covariant (f :: i -> j) where
  comap :: (a ~> b) -> (f a ~> f b)
 
class Contravariant f where
  contramap :: (b ~> a) -> (f a ~> f b)
 
class Covariant m => Monad (m :: i -> i) where
  pure :: a ~> m a
  join :: m (m a) ~> m a
 
class Covariant w => Comonad (w :: i -> i) where
  extract :: w a ~> a
  split :: w a ~> w (w a)
 
class Couniversal σ f | f -> σ where
  couniversal :: σ r => (a ~> r) -> (f a ~> r)
 
class Universal σ f | f -> σ where
  universal :: σ e => (e ~> a) -> (e ~> f a)
 
instance Covariant (Free0 σ) where
  comap f (Free0 e) = Free0 (e . (.f))
 
instance Monad (Free0 σ) where
  pure x = Free0 $ \k -> k x
  join (Free0 e) = Free0 $ \k -> e $ \(Free0 e) -> e k
 
instance Couniversal σ (Free0 σ) where
  couniversal h (Free0 e) = e h
 
-- ...
 

The only unfamiliar classes here should be (Co)Universal. They are for witnessing the adjunctions that make Free σ the initial σ and Cofree σ the final σ in the relevant way. Only one direction is given, since the opposite is very easy to construct with the (co)monad structure.

Free σ is a monad and couniversal, Cofree σ is a comonad and universal.

We can now try to convince ourselves that Free σ and Cofree σ are absolute Here are some examples:

 
free0Absolute0 :: forall g σ a. (Covariant g, σ (Free σ a))
               => g (Free0 σ a) < -> Ran σ Forget g a
free0Absolute0 = (l, r)
 where
 l :: g (Free σ a) -> Ran σ Forget g a
 l g = Ran0 $ \k -> comap (couniversal $ remember0 . k) g
 
 r :: Ran σ Forget g a -> g (Free σ a)
 r (Ran0 e) = e $ Forget0 . pure
 
free0Absolute1 :: forall (g :: * -> * -> *) σ a x. (Covariant g, σ (Free σ a))
               => g (Free0 σ a) x < -> Ran σ Forget g a x
free0Absolute1 = (l, r)
 where
 l :: g (Free σ a) x -> Ran σ Forget g a x
 l g = Ran1 $ \k -> comap (couniversal $ remember0 . k) $$ g
 
 r :: Ran σ Forget g a x -> g (Free σ a) x
 r (Ran1 e) = e $ Forget0 . pure
 
free0Absolute0Op :: forall g σ a. (Contravariant g, σ (Free σ a))
                 => g (Free0 σ a) < -> RanOp σ Forget g a
free0Absolute0Op = (l, r)
 where
 l :: g (Free σ a) -> RanOp σ Forget g a
 l = RanOp0 $ Forget0 . pure
 
 r :: RanOp σ Forget g a -> g (Free σ a)
 r (RanOp0 h g) = contramap (couniversal $ remember0 . h) g
 
-- ...
 

As can be seen, the definitions share a lot of structure. I'm quite confident that with the right building blocks these could be defined once for each of the four types of Kan extensions, with types like:

 
freeAbsolute
  :: forall g σ a. (Covariant g, σ (Free σ a))
  => g (Free σ a) < ~> Ran σ Forget g a
 
cofreeAbsolute
  :: forall g σ a. (Covariant g, σ (Cofree σ a))
  => g (Cofree σ a) < ~> Lan σ Forget g a
 
freeAbsoluteOp
  :: forall g σ a. (Contravariant g, σ (Free σ a))
  => g (Free σ a) < ~> RanOp σ Forget g a
 
cofreeAbsoluteOp
  :: forall g σ a. (Contravariant g, σ (Cofree σ a))
  => g (Cofree σ a) < ~> LanOp σ Forget g a
 

However, it seems quite difficult to structure things in a way such that GHC will accept the definitions. I've successfully written freeAbsolute using some axioms, but turning those axioms into class definitions and the like seems impossible.

Anyhow, the punchline is that we can prove absoluteness using only the premise that there is a valid σ instance for Free σ and Cofree σ. This tends to be quite easy; we just borrow the structure of the type we are quantifying over. This means that in all these cases, we are justified in saying that Free σ ⊣ Forget σ ⊣ Cofree σ, and we have a very generic presentations of (co)free structures in Haskell. So let's look at some.

We've already seen Free Monoid, and last time we talked about Free Applicative, and its relation to traversals. But, Applicative is to traversal as Functor is to lens, so it may be interesting to consider constructions on that. Both Free Functor and Cofree Functor make Functors:

 
instance Functor (Free1 Functor f) where
  fmap f (Free1 e) = Free1 $ fmap f . e
 
instance Functor (Cofree1 Functor f) where
  fmap f (Cofree1 h e) = Cofree1 h (fmap f e)
 

And of course, they are (co)monads, covariant functors and (co)universal among Functors. But, it happens that I know some other types with these properties:

 
data CoYo f a = forall e. CoYo (e -> a) (f e)
 
instance Covariant CoYo where
  comap f = Transform $ \(CoYo h e) -> CoYo h (f $$ e)
 
instance Monad CoYo where
  pure = Transform $ CoYo id
  join = Transform $ \(CoYo h (CoYo h' e)) -> CoYo (h . h') e
 
instance Functor (CoYo f) where
  fmap f (CoYo h e) = CoYo (f . h) e
 
instance Couniversal Functor CoYo where
  couniversal tr = Transform $ \(CoYo h e) -> fmap h (tr $$ e)
 
newtype Yo f a = Yo { oy :: forall r. (a -> r) -> f r }
 
instance Covariant Yo where
  comap f = Transform $ \(Yo e) -> Yo $ (f $$) . e
 
instance Comonad Yo where
  extract = Transform $ \(Yo e) -> e id
  split = Transform $ \(Yo e) -> Yo $ \k -> Yo $ \k' -> e $ k' . k
 
instance Functor (Yo f) where
  fmap f (Yo e) = Yo $ \k -> e (k . f)
 
instance Universal Functor Yo where
  universal tr = Transform $ \e -> Yo $ \k -> tr $$ fmap k e
 

These are the types involved in the (co-)Yoneda lemma. CoYo is a monad, couniversal among functors, and CoYo f is a Functor. Yo is a comonad, universal among functors, and is always a Functor. So, are these equivalent types?

 
coyoIso :: CoYo < ~> Free Functor
coyoIso = (Transform $ couniversal pure, Transform $ couniversal pure)
 
yoIso :: Yo < ~> Cofree Functor
yoIso = (Transform $ universal extract, Transform $ universal extract)
 

Indeed they are. And similar identities hold for the contravariant versions of these constructions.

I don't have much of a use for this last example. I suppose to be perfectly precise, I should point out that these uses of (Co)Yo are not actually part of the (co-)Yoneda lemma. They are two different constructions. The (co-)Yoneda lemma can be given in terms of Kan extensions as:

 
yoneda :: Ran Id f < ~> f
 
coyoneda :: Lan Id f < ~> f
 

But, the use of (Co)Yo to make Functors out of things that aren't necessarily is properly thought of in other terms. In short, we have some kind of category of Haskell types with only identity arrows---it is discrete. Then any type constructor, even non-functorial ones, is certainly a functor from said category (call it Haskrete) into the normal one (Hask). And there is an inclusion functor from Haskrete into Hask:

             F
 Haskrete -----> Hask
      |        /|
      |       /
      |      /
Incl  |     /
      |    /  Ran/Lan Incl F
      |   /
      |  /
      v /
    Hask

So, (Co)Free Functor can also be thought of in terms of these Kan extensions involving the discrete category.

To see more fleshed out, loadable versions of the code in this post, see this file. I may also try a similar Agda development at a later date, as it may admit the more general absoluteness constructions easier.

[0]: The reason for restricting ourselves to kinds involving only * and (->) is that they work much more simply than data kinds. Haskell values can't depend on type-level entities without using type classes. For *, this is natural, but for something like Bool -> *, it is more natural for transformations to be able to inspect the booleans, and so should be something more like forall b. InspectBool b => f b -> g b.

[1]: First-class types are what you get by removing type families and synonyms from consideration. The reason for doing so is that these can't be used properly as parameters and the like, except in cases where they reduce to some other type that is first-class. For example, if we define:

 
type I a = a
 

even though GHC will report I :: * -> *, it is not legal to write Transform I I.

This time I'd like to talk about some other examples of this, and point out how doing so can (perhaps) resolve some disagreements that people have about the specific cases.

The first example is not one that I came up with: induction. It's sometimes said that Haskell does not have inductive types at all, or that we cannot reason about functions on its data types by induction. However, I think this is (techincally) inaccurate. What's true is that we cannot simply pretend that that our types are sets and use the induction principles for sets to reason about Haskell programs. Instead, one has to figure out what inductive domains would be, and what their proof principles are.

Fortunately, there are some papers about doing this. The most recent (that I'm aware of) is Generic Fibrational Induction. I won't get too into the details, but it shows how one can talk about induction in a general setting, where one has a category that roughly corresponds to the type theory/programming language, and a second category of proofs that is 'indexed' by the first category's objects. Importantly, it is not required that the second category is somehow 'part of' the type theory being reasoned about, as is often the case with dependent types, although that is also a special case of their construction.

One of the results of the paper is that this framework can be used to talk about induction principles for types that don't make sense as sets. Specifically:

 
newtype Hyp = Hyp ((Hyp -> Int) -> Int)
 

the type of "hyperfunctions". Instead of interpreting this type as a set, where it would effectively require a set that is isomorphic to the power set of its power set, they interpret it in the category of domains and strict functions mentioned earlier. They then construct the proof category in a similar way as one would for sets, except instead of talking about predicates as subsets, we talk about sub-domains instead. Once this is done, their framework gives a notion of induction for this type.

This example is suitable for ML (and suchlike), due to the strict functions, and sort of breaks the idea that we can really get away with only thinking about sets, even there. Sets are good enough for some simple examples (like flat domains where we don't care about ⊥), but in general we have to generalize induction itself to apply to all types in the 'good' language.

While I haven't worked out how the generic induction would work out for Haskell, I have little doubt that it would, because ML actually contains all of Haskell's data types (and vice versa). So the fact that the framework gives meaning to induction for ML implies that it does so for Haskell. If one wants to know what induction for Haskell's 'lazy naturals' looks like, they can study the ML analogue of:

 
data LNat = Zero | Succ (() -> LNat)
 

because function spaces lift their codomain, and make things 'lazy'.

----

The other example I'd like to talk about hearkens back to the previous article. I explained how foldMap is the proper fundamental method of the Foldable class, because it can be massaged to look like:

 
foldMap :: Foldable f => f a -> FreeMonoid a
 

and lists are not the free monoid, because they do not work properly for various infinite cases.

I also mentioned that foldMap looks a lot like traverse:

 
foldMap  :: (Foldable t   , Monoid m)      => (a -> m)   -> t a -> m
traverse :: (Traversable t, Applicative f) => (a -> f b) -> t a -> f (t b)
 

And of course, we have Monoid m => Applicative (Const m), and the functions are expected to agree in this way when applicable.

Now, people like to get in arguments about whether traversals are allowed to be infinite. I know Ed Kmett likes to argue that they can be, because he has lots of examples. But, not everyone agrees, and especially people who have papers proving things about traversals tend to side with the finite-only side. I've heard this includes one of the inventors of Traversable, Conor McBride.

In my opinion, the above disagreement is just another example of a situation where we have a generic notion instantiated in two different ways, and intuition about one does not quite transfer to the other. If you are working in a language like Agda or Coq (for proving), you will be thinking about traversals in the context of sets and total functions. And there, traversals are finite. But in Haskell, there are infinitary cases to consider, and they should work out all right when thinking about domains instead of sets. But I should probably put forward some argument for this position (and even if I don't need to, it leads somewhere else interesting).

One example that people like to give about finitary traversals is that they can be done via lists. Given a finite traversal, we can traverse to get the elements (using Const [a]), traverse the list, then put them back where we got them by traversing again (using State [a]). Usually when you see this, though, there's some subtle cheating in relying on the list to be exactly the right length for the second traversal. It will be, because we got it from a traversal of the same structure, but I would expect that proving the function is actually total to be a lot of work. Thus, I'll use this as an excuse to do my own cheating later.

Now, the above uses lists, but why are we using lists when we're in Haskell? We know they're deficient in certain ways. It turns out that we can give a lot of the same relevant structure to the better free monoid type:

 
newtype FM a = FM (forall m. Monoid m => (a -> m) -> m) deriving (Functor)
 
instance Applicative FM where
  pure x = FM ($ x)
  FM ef < *> FM ex = FM $ \k -> ef $ \f -> ex $ \x -> k (f x)
 
instance Monoid (FM a) where
  mempty = FM $ \_ -> mempty
  mappend (FM l) (FM r) = FM $ \k -> l k <> r k
 
instance Foldable FM where
  foldMap f (FM e) = e f
 
newtype Ap f b = Ap { unAp :: f b }
 
instance (Applicative f, Monoid b) => Monoid (Ap f b) where
  mempty = Ap $ pure mempty
  mappend (Ap l) (Ap r) = Ap $ (<>) < $> l < *> r
 
instance Traversable FM where
  traverse f (FM e) = unAp . e $ Ap . fmap pure . f
 

So, free monoids are Monoids (of course), Foldable, and even Traversable. At least, we can define something with the right type that wouldn't bother anyone if it were written in a total language with the right features, but in Haskell it happens to allow various infinite things that people don't like.

Now it's time to cheat. First, let's define a function that can take any Traversable to our free monoid:

 
toFreeMonoid :: Traversable t => t a -> FM a
toFreeMonoid f = FM $ \k -> getConst $ traverse (Const . k) f
 

Now let's define a Monoid that's not a monoid:

 
data Cheat a = Empty | Single a | Append (Cheat a) (Cheat a)
 
instance Monoid (Cheat a) where
  mempty = Empty
  mappend = Append
 

You may recognize this as the data version of the free monoid from the previous article, where we get the real free monoid by taking a quotient. using this, we can define an Applicative that's not valid:

 
newtype Cheating b a =
  Cheating { prosper :: Cheat b -> a } deriving (Functor)
 
instance Applicative (Cheating b) where
  pure x = Cheating $ \_ -> x
 
  Cheating f < *> Cheating x = Cheating $ \c -> case c of
    Append l r -> f l (x r)
 

Given these building blocks, we can define a function to relabel a traversable using a free monoid:

 
relabel :: Traversable t => t a -> FM b -> t b
relabel t (FM m) = propser (traverse (const hope) t) (m Single)
 where
 hope = Cheating $ \c -> case c of
   Single x -> x
 

And we can implement any traversal by taking a trip through the free monoid:

 
slowTraverse
  :: (Applicative f, Traversable t) => (a -> f b) -> t a -> f (t b)
slowTraverse f t = fmap (relabel t) . traverse f . toFreeMonoid $ t
 

And since we got our free monoid via traversing, all the partiality I hid in the above won't blow up in practice, rather like the case with lists and finite traversals.

Arguably, this is worse cheating. It relies on the exact association structure to work out, rather than just number of elements. The reason is that for infinitary cases, you cannot flatten things out, and there's really no way to detect when you have something infinitary. The finitary traversals have the luxury of being able to reassociate everything to a canonical form, while the infinite cases force us to not do any reassociating at all. So this might be somewhat unsatisfying.

But, what if we didn't have to cheat at all? We can get the free monoid by tweaking foldMap, and it looks like traverse, so what happens if we do the same manipulation to the latter?

It turns out that lens has a type for this purpose, a slight specialization of which is:

 
newtype Bazaar a b t =
  Bazaar { runBazaar :: forall f. Applicative f => (a -> f b) -> f t }
 

Using this type, we can reorder traverse to get:

 
howBizarre :: Traversable t => t a -> Bazaar a b (t b)
howBizarre t = Bazaar $ \k -> traverse k t
 

But now, what do we do with this? And what even is it? [1]

If we continue drawing on intuition from Foldable, we know that foldMap is related to the free monoid. Traversable has more indexing, and instead of Monoid uses Applicative. But the latter are actually related to the former; Applicatives are monoidal (closed) functors. And it turns out, Bazaar has to do with free Applicatives.

If we want to construct free Applicatives, we can use our universal property encoding trick:

 
newtype Free p f a =
  Free { gratis :: forall g. p g => (forall x. f x -> g x) -> g a }
 

This is a higher-order version of the free p, where we parameterize over the constraint we want to use to represent structures. So Free Applicative f is the free Applicative over a type constructor f. I'll leave the instances as an exercise.

Since free monoid is a monad, we'd expect Free p to be a monad, too. In this case, it is a McBride style indexed monad, as seen in The Kleisli Arrows of Outrageous Fortune.

 
type f ~> g = forall x. f x -> g x
 
embed :: f ~> Free p f
embed fx = Free $ \k -> k fx
 
translate :: (f ~> g) -> Free p f ~> Free p g
translate tr (Free e) = Free $ \k -> e (k . tr)
 
collapse :: Free p (Free p f) ~> Free p f
collapse (Free e) = Free $ \k -> e $ \(Free e') -> e' k
 

That paper explains how these are related to Atkey style indexed monads:

 
data At key i j where
  At :: key -> At key i i
 
type Atkey m i j a = m (At a j) i
 
ireturn :: IMonad m => a -> Atkey m i i a
ireturn = ...
 
ibind :: IMonad m => Atkey m i j a -> (a -> Atkey m j k b) -> Atkey m i k b
ibind = ...
 

It turns out, Bazaar is exactly the Atkey indexed monad derived from the Free Applicative indexed monad (with some arguments shuffled) [2]:

 
hence :: Bazaar a b t -> Atkey (Free Applicative) t b a
hence bz = Free $ \tr -> runBazaar bz $ tr . At
 
forth :: Atkey (Free Applicative) t b a -> Bazaar a b t
forth fa = Bazaar $ \g -> gratis fa $ \(At a) -> g a
 
imap :: (a -> b) -> Bazaar a i j -> Bazaar b i j
imap f (Bazaar e) = Bazaar $ \k -> e (k . f)
 
ipure :: a -> Bazaar a i i
ipure x = Bazaar ($ x)
 
(>>>=) :: Bazaar a j i -> (a -> Bazaar b k j) -> Bazaar b k i
Bazaar e >>>= f = Bazaar $ \k -> e $ \x -> runBazaar (f x) k
 
(>==>) :: (s -> Bazaar i o t) -> (i -> Bazaar a b o) -> s -> Bazaar a b t
(f >==> g) x = f x >>>= g
 

As an aside, Bazaar is also an (Atkey) indexed comonad, and the one that characterizes traversals, similar to how indexed store characterizes lenses. A Lens s t a b is equivalent to a coalgebra s -> Store a b t. A traversal is a similar Bazaar coalgebra:

 
  s -> Bazaar a b t
    ~
  s -> forall f. Applicative f => (a -> f b) -> f t
    ~
  forall f. Applicative f => (a -> f b) -> s -> f t
 

It so happens that Kleisli composition of the Atkey indexed monad above (>==>) is traversal composition.

Anyhow, Bazaar also inherits Applicative structure from Free Applicative:

 
instance Functor (Bazaar a b) where
  fmap f (Bazaar e) = Bazaar $ \k -> fmap f (e k)
 
instance Applicative (Bazaar a b) where
  pure x = Bazaar $ \_ -> pure x
  Bazaar ef < *> Bazaar ex = Bazaar $ \k -> ef k < *> ex k
 

This is actually analogous to the Monoid instance for the free monoid; we just delegate to the underlying structure.

The more exciting thing is that we can fold and traverse over the first argument of Bazaar, just like we can with the free monoid:

 
bfoldMap :: Monoid m => (a -> m) -> Bazaar a b t -> m
bfoldMap f (Bazaar e) = getConst $ e (Const . f)
 
newtype Comp g f a = Comp { getComp :: g (f a) } deriving (Functor)
 
instance (Applicative f, Applicative g) => Applicative (Comp g f) where
  pure = Comp . pure . pure
  Comp f < *> Comp x = Comp $ liftA2 (< *>) f x
 
btraverse
  :: (Applicative f) => (a -> f a') -> Bazaar a b t -> Bazaar a' b t
btraverse f (Bazaar e) = getComp $ e (c . fmap ipure . f)
 

This is again analogous to the free monoid code. Comp is the analogue of Ap, and we use ipure in traverse. I mentioned that Bazaar is a comonad:

 
extract :: Bazaar b b t -> t
extract (Bazaar e) = runIdentity $ e Identity
 

And now we are finally prepared to not cheat:

 
honestTraverse
  :: (Applicative f, Traversable t) => (a -> f b) -> t a -> f (t b)
honestTraverse f = fmap extract . btraverse f . howBizarre
 

So, we can traverse by first turning out Traversable into some structure that's kind of like the free monoid, except having to do with Applicative, traverse that, and then pull a result back out. Bazaar retains the information that we're eventually building back the same type of structure, so we don't need any cheating.

To pull this back around to domains, there's nothing about this code to object to if done in a total language. But, if we think about our free Applicative-ish structure, in Haskell, it will naturally allow infinitary expressions composed of the Applicative operations, just like the free monoid will allow infinitary monoid expressions. And this is okay, because some Applicatives can make sense of those, so throwing them away would make the type not free, in the same way that even finite lists are not the free monoid in Haskell. And this, I think, is compelling enough to say that infinite traversals are right for Haskell, just as they are wrong for Agda.

For those who wish to see executable code for all this, I've put a files here and here. The latter also contains some extra goodies at the end that I may talk about in further installments.

[1] Truth be told, I'm not exactly sure.

[2] It turns out, you can generalize Bazaar to have a correspondence for every choice of p

 
newtype Bizarre p a b t =
  Bizarre { bizarre :: forall f. p f => (a -> f b) -> f t }
 

hence and forth above go through with the more general types. This can be seen here.

This latter fact can be seen relatively easily by considering another list-like type:

 
data SL a = Empty | SL a :> a
 
instance Monoid (SL a) where
  mempty = Empty
  mappend ys Empty = ys
  mappend ys (xs :> x) = (mappend ys xs) :> x
 
single :: a -> SL a
single x = Empty :> x
 

So, we have a type SL a of snoc lists, which are a monoid, and a function that embeds a into SL a. If (ordinary) lists were the free monoid, there would be a unique monoid homomorphism from lists to snoc lists. Such a homomorphism (call it h) would have the following properties:

 
h [] = Empty
h (xs <> ys) = h xs <> h ys
h [x] = single x
 

And in fact, this (together with some general facts about Haskell functions) should be enough to define h for our purposes (or any purposes, really). So, let's consider its behavior on two values:

 
h [1] = single 1
 
h [1,1..] = h ([1] <> [1,1..]) -- [1,1..] is an infinite list of 1s
          = h [1] <> h [1,1..]
 

This second equation can tell us what the value of h is at this infinite value, since we can consider it the definition of a possibly infinite value:

 
x = h [1] <> x = fix (single 1 <>)
h [1,1..] = x
 

(single 1 <>) is a strict function, so the fixed point theorem tells us that x = ⊥.

This is a problem, though. Considering some additional equations:

 
[1,1..] <> [n] = [1,1..] -- true for all n
h [1,1..] = ⊥
h ([1,1..] <> [1]) = h [1,1..] <> h [1]
                   = ⊥ <> single 1
                   = ⊥ :> 1
                   ≠ ⊥
 

So, our requirements for h are contradictory, and no such homomorphism can exist.

The issue is that Haskell types are domains. They contain these extra partially defined values and infinite values. The monoid structure on (cons) lists has infinite lists absorbing all right-hand sides, while the snoc lists are just the opposite.

This also means that finite lists (or any method of implementing finite sequences) are not free monoids in Haskell. They, as domains, still contain the additional bottom element, and it absorbs all other elements, which is incorrect behavior for the free monoid:

 
pure x <> ⊥ = ⊥
h ⊥ = ⊥
h (pure x <> ⊥) = [x] <> h ⊥
                = [x] ++ ⊥
                = x:⊥
                ≠ ⊥
 

So, what is the free monoid? In a sense, it can't be written down at all in Haskell, because we cannot enforce value-level equations, and because we don't have quotients. But, if conventions are good enough, there is a way. First, suppose we have a free monoid type FM a. Then for any other monoid m and embedding a -> m, there must be a monoid homomorphism from FM a to m. We can model this as a Haskell type:

 
forall a m. Monoid m => (a -> m) -> FM a -> m
 

Where we consider the Monoid m constraint to be enforcing that m actually has valid monoid structure. Now, a trick is to recognize that this sort of universal property can be used to define types in Haskell (or, GHC at least), due to polymorphic types being first class; we just rearrange the arguments and quantifiers, and take FM a to be the polymorphic type:

 
newtype FM a = FM { unFM :: forall m. Monoid m => (a -> m) -> m }
 

Types defined like this are automatically universal in the right sense. [1] The only thing we have to check is that FM a is actually a monoid over a. But that turns out to be easily witnessed:

 
embed :: a -> FM a
embed x = FM $ \k -> k x
 
instance Monoid (FM a) where
  mempty = FM $ \_ -> mempty
  mappend (FM e1) (FM e2) = FM $ \k -> e1 k <> e2 k
 

Demonstrating that the above is a proper monoid delegates to instances of Monoid being proper monoids. So as long as we trust that convention, we have a free monoid.

However, one might wonder what a free monoid would look like as something closer to a traditional data type. To construct that, first ignore the required equations, and consider only the generators; we get:

 
data FMG a = None | Single a | FMG a :<> FMG a
 

Now, the proper FM a is the quotient of this by the equations:

 
None :<> x = x = x :<> None
x :<> (y :<> z) = (x :<> y) :<> z
 

One way of mimicking this in Haskell is to hide the implementation in a module, and only allow elimination into Monoids (again, using the convention that Monoid ensures actual monoid structure) using the function:

 
unFMG :: forall a m. Monoid m => FMG a -> (a -> m) -> m
unFMG None _ = mempty
unFMG (Single x) k = k x
unFMG (x :<> y) k = unFMG x k <> unFMG y k
 

This is actually how quotients can be thought of in richer languages; the quotient does not eliminate any of the generated structure internally, it just restricts the way in which the values can be consumed. Those richer languages just allow us to prove equations, and enforce properties by proof obligations, rather than conventions and structure hiding. Also, one should note that the above should look pretty similar to our encoding of FM a using universal quantification earlier.

Now, one might look at the above and have some objections. For one, we'd normally think that the quotient of the above type is just [a]. Second, it seems like the type is revealing something about the associativity of the operations, because defining recursive values via left nesting is different from right nesting, and this difference is observable by extracting into different monoids. But aren't monoids supposed to remove associativity as a concern? For instance:

 
ones1 = embed 1 <> ones1
ones2 = ones2 <> embed 1
 

Shouldn't we be able to prove these are the same, becuase of an argument like:

 
ones1 = embed 1 <> (embed 1 <> ...)
      ... reassociate ...
      = (... <> embed 1) <> embed 1
      = ones2
 

The answer is that the equation we have only specifies the behavior of associating three values:

 
x <> (y <> z) = (x <> y) <> z
 

And while this is sufficient to nail down the behavior of finite values, and finitary reassociating, it does not tell us that infinitary reassociating yields the same value back. And the "... reassociate ..." step in the argument above was decidedly infinitary. And while the rules tell us that we can peel any finite number of copies of embed 1 to the front of ones1 or the end of ones2, it does not tell us that ones1 = ones2. And in fact it is vital for FM a to have distinct values for these two things; it is what makes it the free monoid when we're dealing with domains of lazy values.

Finally, we can come back to Foldable. If we look at foldMap:

 
foldMap :: (Foldable f, Monoid m) => (a -> m) -> f a -> m
 

we can rearrange things a bit, and get the type:

 
Foldable f => f a -> (forall m. Monoid m => (a -> m) -> m)
 

And thus, the most fundamental operation of Foldable is not toList, but toFreeMonoid, and lists are not free monoids in Haskell.

[1]: What we are doing here is noting that (co)limits are objects that internalize natural transformations, but the natural transformations expressible by quantification in GHC are already automatically internalized using quantifiers. However, one has to be careful that the quantifiers are actually enforcing the relevant naturality conditions. In many simple cases they are.

Given all this, It seems post on thinking with applicatives is in order, showing how to build them up and reason about them. One nice thing about the approach we'll be taking is that it uses a "final" encoding of applicatives, rather than building up and then later interpreting a structure. This is in fact how we typically write monads (pace operational, free, etc.), but since we more often only determine our data structures are applicative after the fact, we often get some extra junk lying around (OptParse-Applicative, for example, has a GADT that I think is entirely extraneous).

So the usual throat clearing:

{-# LANGUAGE TypeOperators, MultiParamTypeClasses, FlexibleInstances,
StandaloneDeriving, FlexibleContexts, UndecidableInstances,
GADTs, KindSignatures, RankNTypes #-}
 
module Main where
import Control.Applicative hiding (Const)
import Data.Monoid hiding (Sum, Product)
import Control.Monad.Identity
instance Show a => Show (Identity a) where
    show (Identity x) = "(Identity " ++ show x ++ ")"

And now, let's start with a classic applicative, going back to the Applicative Programming With Effects paper:

data Const mo a = Const mo deriving Show
 
instance Functor (Const mo) where
    fmap _ (Const mo) = Const mo
 
instance Monoid mo => Applicative (Const mo) where
    pure _ = Const mempty
    (Const f) < *> (Const x) = Const (f <> x)

(Const lives in transformers as the Constant functor, or in base as Const)

Note that Const is not a monad. We've defined it so that its structure is independent of the `a` type. Hence if we try to write (>>=) of type Const mo a -> (a -> Const mo b) -> Const mo b, we'll have no way to "get out" the first `a` and feed it to our second argument.

One great thing about Applicatives is that there is no distinction between applicative transformers and applicatives themselves. This is to say that the composition of two applicatives is cleanly and naturally always also an applicative. We can capture this like so:

 
newtype Compose f g a = Compose (f (g a)) deriving Show
 
instance (Functor f, Functor g) => Functor (Compose f g) where
    fmap f (Compose x) = Compose $ (fmap . fmap) f x
 
instance (Applicative f, Applicative g) => Applicative (Compose f g) where
    pure = Compose . pure . pure
    (Compose f) < *> (Compose x) = Compose $ (< *>) < $> f < *> x

(Compose also lives in transformers)

Note that Applicatives compose two ways. We can also write:

data Product f g a = Product (f a) (g a) deriving Show
 
instance (Functor f, Functor g) => Functor (Product f g) where
    fmap f (Product  x y) = Product (fmap f x) (fmap f y)
 
instance (Applicative f, Applicative g) => Applicative (Product f g) where
    pure x = Product (pure x) (pure x)
    (Product f g) < *> (Product  x y) = Product (f < *> x) (g < *> y)

(Product lives in transformers as well)

This lets us now construct an extremely rich set of applicative structures from humble beginnings. For example, we can reconstruct the Writer Applicative.

type Writer mo = Product (Const mo) Identity
 
tell :: mo -> Writer mo ()
tell x = Product (Const x) (pure ())

-- tell [1] *> tell [2]
-- > Product (Const [1,2]) (Identity ())

Note that if we strip away the newtype noise, Writer turns into (mo,a) which is isomorphic to the Writer monad. However, we've learned something along the way, which is that the monoidal component of Writer (as long as we stay within the rules of applicative) is entirely independent from the "identity" component. However, if we went on to write the Monad instance for our writer (by defining >>=), we'd have to "reach in" to the identity component to grab a value to hand back to the function yielding our monoidal component. Which is to say we would destroy this nice seperation of "trace" and "computational content" afforded by simply taking the product of two Applicatives.

Now let's make things more interesting. It turns out that just as the composition of two applicatives may be a monad, so too the composition of two monads may be no stronger than an applicative!

We'll see this by introducing Maybe into the picture, for possibly failing computations.

type FailingWriter mo = Compose (Writer mo) Maybe
 
tellFW :: Monoid mo => mo -> FailingWriter mo ()
tellFW x = Compose (tell x *> pure (Just ()))
 
failFW :: Monoid mo => FailingWriter mo a
failFW = Compose (pure Nothing)

-- tellFW [1] *> tellFW [2]
-- > Compose (Product (Const [1,2]) (Identity Just ()))

-- tellFW [1] *> failFW *> tellFW [2]
-- > Compose (Product (Const [1,2]) (Identity Nothing))

Maybe over Writer gives us the same effects we'd get in a Monad — either the entire computation fails, or we get the result and the trace. But Writer over Maybe gives us new behavior. We get the entire trace, even if some computations have failed! This structure, just like Const, cannot be given a proper Monad instance. (In fact if we take Writer over Maybe as a Monad, we get only the trace until the first point of failure).

This seperation of a monoidal trace from computational effects (either entirely independent of a computation [via a product] or independent between parts of a computation [via Compose]) is the key to lots of neat tricks with applicative functors.

Next, let's look at Gergo Erdi's "Static Analysis with Applicatives" that is built using free applicatives. We can get essentially the same behavior directly from the product of a constant monad with an arbitrary effectful monad representing our ambient environment of information. As long as we constrain ourselves to only querying it with the takeEnv function, then we can either read the left side of our product to statically read dependencies, or the right side to actually utilize them.

type HasEnv k m = Product (Const [k]) m
takeEnv :: (k -> m a) -> k -> HasEnv k m a
takeEnv f x = Product (Const [x]) (f x)

If we prefer, we can capture queries of a static environment directly with the standard Reader applicative, which is just a newtype over the function arrow. There are other varients of this that perhaps come closer to exactly how Erdi's post does things, but I think this is enough to demonstrate the general idea.

data Reader r a = Reader (r -> a)
instance Functor (Reader r) where
    fmap f (Reader x) = Reader (f . x)
instance Applicative (Reader r) where
    pure x = Reader $ pure x
    (Reader f) < *> (Reader x) = Reader (f < *> x)
 
runReader :: (Reader r a) -> r -> a
runReader (Reader f) = f
 
takeEnvNew :: (env -> k -> a) -> k -> HasEnv k (Reader env) a
takeEnvNew f x = Product (Const [x]) (Reader $ flip f x)

So, what then is a full formlet? It's something that can be executed in one context as a monoid that builds a form, and in another as a parser. so the top level must be a product.

type FormletOne mo a = Product (Const mo) Identity a

Below the product, we read from an environment and perhaps get an answer. So that's reader with a maybe.

type FormletTwo mo env a =
    Product (Const mo) (Compose (Reader env) Maybe) a

Now if we fail, we want to have a trace of errors. So we expand out the Maybe into a product as well to get the following, which adds monoidal errors:

type FormletThree mo err env a =
    Product (Const mo)
            (Compose (Reader env) (Product (Const err) Maybe)) a

But now we get errors whether or not the parse succeeds. We want to say either the parse succeeds or we get errors. For this, we can turn to the typical Sum functor, which currently lives as Coproduct in comonad-transformers, but will hopefully be moving as Sum to the transformers library in short order.

data Sum f g a = InL (f a) | InR (g a) deriving Show
 
instance (Functor f, Functor g) => Functor (Sum f g) where
    fmap f (InL x) = InL (fmap f x)
    fmap f (InR y) = InR (fmap f y)

The Functor instance is straightforward for Sum, but the applicative instance is puzzling. What should "pure" do? It needs to inject into either the left or the right, so clearly we need some form of "bias" in the instance. What we really need is the capacity to "work in" one side of the sum until compelled to switch over to the other, at which point we're stuck there. If two functors, F and G are in a relationship such that we can always send f x -> g x in a way that "respects" fmap (that is to say, such that (fmap f . fToG == ftoG . fmap f) then we call this a natural transformation. The action that sends f to g is typically called "eta". (We actually want something slightly stronger called a "monoidal natural transformation" that respects not only the functorial action fmap but the applicative action <*>, but we can ignore that for now).

Now we can assert that as long as there is a natural transformation between g and f, then Sum f g can be made an Applicative, like so:

class Natural f g where
    eta :: f a -> g a
 
instance (Applicative f, Applicative g, Natural g f) =>
  Applicative (Sum f g) where
    pure x = InR $ pure x
    (InL f) < *> (InL x) = InL (f < *> x)
    (InR g) < *> (InR y) = InR (g < *> y)
    (InL f) < *> (InR x) = InL (f < *> eta x)
    (InR g) < *> (InL x) = InL (eta g < *> x)

The natural transformation we'll tend to use simply sends any functor to Const.

instance Monoid mo => Natural f (Const mo) where
    eta = const (Const mempty)

However, there are plenty of other natural transformations that we could potentially make use of, like so:

instance Applicative f =>
  Natural g (Compose f g) where
     eta = Compose . pure
 
instance (Applicative g, Functor f) => Natural f (Compose f g) where
     eta = Compose . fmap pure
 
instance (Natural f g) => Natural f (Product f g) where
    eta x = Product x (eta x)
 
instance (Natural g f) => Natural g (Product f g) where
    eta x = Product (eta x) x
 
instance Natural (Product f g) f where
    eta (Product x _ ) = x
 
instance Natural (Product f g) g where
    eta (Product _ x) = x
 
instance Natural g f => Natural (Sum f g) f where
    eta (InL x) = x
    eta (InR y) = eta y
 
instance Natural Identity (Reader r) where
    eta (Identity x) = pure x

In theory, there should also be a natural transformation that can be built magically from the product of any other two natural transformations, but that will just confuse the Haskell typechecker hopelessly. This is because we know that often different "paths" of typeclass choices will often be isomorphic, but the compiler has to actually pick one "canonical" composition of natural transformations to compute with, although multiple paths will typically be possible.

For similar reasons of avoiding overlap, we can't both have the terminal homomorphism that sends everything to "Const" and the initial homomorphism that sends "Identity" to anything like so:

-- instance Applicative g => Natural Identity g where
--     eta (Identity x) = pure x
 

We choose to keep the terminal transformation around because it is more generally useful for our purposes. As the comments below point out, it turns out that a version of "Sum" with the initial transformation baked in now lives in transformers as Lift.

In any case we can now write a proper Validation applicative:

type Validation mo = Sum (Const mo) Identity
 
validationError :: Monoid mo => mo -> Validation mo a
validationError x = InL (Const x)

This applicative will yield either a single result, or an accumulation of monoidal errors. It exists on hackage in the Validation package.

Now, based on the same principles, we can produce a full Formlet.

type Formlet mo err env a =
    Product (Const mo)
            (Compose (Reader env)
                     (Sum (Const err) Identity))
    a

All the type and newtype noise looks a bit ugly, I'll grant. But the idea is to think with structures built with applicatives, which gives guarantees that we're building applicative structures, and furthermore, structures with certain guarantees in terms of which components can be interpreted independently of which others. So, for example, we can strip away the newtype noise and find the following:

type FormletClean mo err env a = (mo, env -> Either err a)

Because we built this up from our basic library of applicatives, we also know how to write its applicative instance directly.

Now that we've gotten a basic algebraic vocabulary of applicatives, and especially now that we've produced this nifty Sum applicative (which I haven't seen presented before), we've gotten to where I intended to stop.

But lots of other questions arise, on two axes. First, what other typeclasses beyond applicative do our constructions satisfy? Second, what basic pieces of vocabulary are missing from our constructions — what do we need to add to flesh out our universe of discourse? (Fixpoints come to mind).

Also, what statements can we make about "completeness" — what portion of the space of all applicatives can we enumerate and construct in this way? Finally, why is it that monoids seem to crop up so much in the course of working with Applicatives? I plan to tackle at least some of these questions in future blog posts.

In category theory, the codensity monad is given by the rather frightening expression:

Where the integral notation denotes an end, and the square brackets denote a power, which allows us to take what is essentially an exponential of the objects of by objects of , where is enriched in . Provided the above end exists, is a monad regardless of whether has an adjoint, which is the usual way one thinks of functors (in general) giving rise to monads.

It also turns out that this construction is a sort of generalization of the adjunction case. If we do have , this gives rise to a monad . But, in such a case, , so the codensity monad is the same as the monad given by the adjunction when it exists, but codensity may exist when there is no adjunction.

In Haskell, this all becomes a bit simpler (to my eyes, at least). Our category is always , which is enriched in itself, so powers are just function spaces. And all the functors we write will be rather like (objects will come from kinds we can quantify over), so ends of functors will look like forall r. F r r where . Then:

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


As mentioned, we've known for a while that we can write a Monad instance for Codensity f without caring at all about f.

As for the adjunction correspondence, consider the adjunction between products and exponentials:

This gives rise to the monad , the state monad. According to the facts above, we should have that Codensity (s ->) (excuse the sectioning) is the same as state, and if we look, we see:

forall r. (a -> s -> r) -> s -> r


which is the continuation passing, or Church (or Boehm-Berarducci) encoding of the monad.

Now, it's also well known that for any monad, we can construct an adjunction that gives rise to it. There are multiple ways to do this, but the most accessible in Haskell is probably via the Kleisli category. So, given a monad on , there is a category with the same objects, but where . The identity for each object is return and composition of arrows is:

(f >=> g) x = f x >>= g


Our two functors are:

F a = a
F f = return . f

U a = M a
U f = (>>= f)


Verifying that requires only that , but this is just , which is a triviality. Now we should have that .

So, one of the simplest monads is reader, . Now, just takes objects in the Kleisli category (which are objects in ) and applies to them, so we should have Codensity (e ->) is reader. But earlier we had Codensity (e ->) was state. So reader is state, right?

We can actually arrive at this result another way. One of the most famous pieces of category theory is the Yoneda lemma, which states that the following correspondence holds for any functor :

This also works for any functor into and looks like:

F a ~= forall r. (a -> r) -> F r


for . But we also have our functor , which should look more like:

U a ~= forall r. (a -> M r) -> U r
M a ~= forall r. (a -> M r) -> M r

So, we fill in M = (e ->) and get that reader is isomorphic to state, right? What's going on?

To see, we have to take a closer look at natural transformations. Given two categories and , and functors , a natural transformation is a family of maps such that for every the following diagram commutes:

The key piece is what the morphisms look like. It's well known that parametricity ensures the naturality of t :: forall a. F a -> G a for , and it also works when the source is . It should also work for a category, call it , which has Haskell types as objects, but where , which is the sort of category that newtype Endo a = Endo (a -> a) is a functor from. So we should be at liberty to say:

Codensity Endo a = forall r. (a -> r -> r) -> r -> r ~= [a]


However, hom types for are not merely made up of hom types on the same arguments, so naturality turns out not to be guaranteed. A functor must take a Kleisli arrow to an arrow , and transformations must commute with that mapping. So, if we look at our use of Yoneda, we are considering transformations :

Now, and . So

t :: forall r. (a -> M r) -> M r

will get us the right type of maps. But, the above commutative square corresponds to the condition that for all f :: b -> M c:

t . (>=> f) = (>>= f) . t


So, if we have h :: a -> M b, Kleisli composing it with f and then feeding to t is the same as feeding h to t and then binding the result with f.

Now, if we go back to reader, we can consider the reader morphism:

f = const id :: a -> e -> e


For all relevant m and g, m >>= f = id and g >=> f = f. So the
naturality condition here states that t f = id.

Now, t :: forall r. (a -> e -> r) -> e -> r. The general form of these is state actions (I've split e -> (a, e) into two pieces):

t f e = f (v e) (st e)
  where
  rd :: e -> a
  st :: e -> e

If f = const id, then:

t (const id) e = st e
 where
 st :: e -> e

But our naturality condition states that this must be the identity, so we must have st = id. That is, the naturality condition selects t for which the corresponding state action does not change the state, meaning it is equivalent to a reader action! Presumably the definition of an end (which involves dinaturality) enforces a similar condition, although I won't work through it, as it'd be rather more complicated.

However, we have learned a lesson. Quantifiers do not necessarily enforce (di)naturality for every category with objects of the relevant kind. It is important to look at the hom types, not just the objects .In this case, the point of failure seems to be the common, extra s. Even though the type contains nautral transformations for the similar functors over , they can (in general) still manipulate the shared parameter in ways that are not natural for the domain in question.

I am unsure of how exactly one could enforce the above condition in (Haskell's) types. For instance, if we consider:

forall r m. Monad m => (a -> m r) -> m r

This still contains transformations of the form:

t k = k a >> k a

And for this to be natural would require:

(k >=> f) a >> (k >=> f) a = (k a >> k a) >>= f

Which is not true for all possible instantiations of f. It seems as though leaving m unconstrained would be sufficient, as all that could happen is t feeding a value to k and yielding the result, but it seems likely to be over-restrictive.

Fortunately, while it was rather inconvenient that they shut down Google Knol completely, and I'll forever remember a knol as a "unit of abandonment", Google did provide a nice way to download at least your own user content and for that I am grateful.

I have fixed up the internal linkage as much as possible and have placed a copy of the original article below.

Catamorphisms: A Knol

Sadly, as I am not "Dark Magus", I am unable to download the Russian translation. If anyone knows how to contact him, I would love to obtain and preserve a copy of the translation as well.

Constraint Kinds adds a new kind Constraint, such that Eq :: * -> Constraint, Monad :: (* -> *) -> Constraint, but since it is a kind, we can make type families for constraints, and even parameterize constraints on constraints.

So, let's play with them and see what we can come up with!

A Few Extensions

First, we'll need a few language features:

 
{-# LANGUAGE
  CPP,
  ScopedTypeVariables,
  FlexibleInstances,
  FlexibleContexts,
  ConstraintKinds,
  KindSignatures,
  TypeOperators,
  FunctionalDependencies,
  Rank2Types,
  StandaloneDeriving,
  GADTs
  #-}
 

Because of the particular version of GHC I'm using I'll also need

 
{-# LANGUAGE UndecidableInstances #-}
#define UNDECIDABLE
 

but this bug has been fixed in the current version of GHC Head. I'll be explicit about any instances that need UndecidableInstances by surrounding them in an #ifdef UNDECIDABLE block.

Explicit Dictionaries

So with that out of the way, let's import some definitions

 
import Control.Monad
import Control.Monad.Instances
import Control.Applicative
import Data.Monoid
import Data.Complex
import Data.Ratio
import Unsafe.Coerce
 

and make one of our own that shows what we get out of making Constraints into a kind we can manipulate like any other.

 
data Dict a where
  Dict :: a => Dict a
 

Previously, we coud make a Dict like data type for any one particular class constraint that we wanted to capture, but now we can write this type once and for all. The act of pattern matching on the Dict constructor will bring the constraint 'a' into scope.

Of course, in the absence of incoherent and overlapping instances there is at most one dictionary of a given type, so we could make instances, like we can for any other data type, but standalone deriving is smart enough to figure these out for me. (Thanks copumpkin!)

 
deriving instance Eq (Dict a)
deriving instance Ord (Dict a)
deriving instance Show (Dict a)
 

If we're willing to turn on UndecidableInstances to enable the polymorphic constraint we can even add:

 
#ifdef UNDECIDABLE
deriving instance a => Read (Dict a)
instance a => Monoid (Dict a) where
  mappend Dict Dict = Dict
  mempty = Dict
#endif
 

and similar polymorphically constrained instances for Enum, Bounded, etc.

Entailment

For that we'll need a notion of entailment.

 
infixr 9 :-
newtype a :- b = Sub (a => Dict b)
 
instance Eq (a :- b) where
  _ == _ = True
 
instance Ord (a :- b) where
  compare _ _ = EQ
 
instance Show (a :- b) where
  showsPrec d _ = showParen (d > 10) $
    showString "Sub Dict"
 

Here we're saying that Sub takes one argument, which is a computation that when implicitly given a constraint of type a, can give me back a dictionary for the type b. Moreover, as a newtype it adds no overhead that isn't aleady present in manipulating terms of type (a => Dict b) directly.

The simplest thing we can define with this is that entailment is reflexive.

 
refl :: a :- a
refl = Sub Dict
 

Max has already written up a nice restricted monad example using these, but what I want to play with today is the category of substitutability of constraints, but there are a few observations I need to make, first.

ConstraintKinds overloads () and (a,b) to represent the trivial constraint and the product of two constraints respectively.

The latter is done with a bit of a hack, which we'll talk about in a minute, but we can use the former as a terminal object for our category of entailments.

 
weaken1 :: (a, b) :- a
weaken1 = Sub Dict
 
weaken2 :: (a, b) :- b
weaken2 = Sub Dict
 

Constraints are idempotent, so we can duplicate one, perhaps as a prelude to transforming one of them into something else.

 
contract :: a :- (a, a)
contract = Sub Dict
 

But to do much more complicated, we're going to need a notion of substitution, letting us use our entailment relation to satisfy obligations.

 
infixl 1 \\ -- required comment
(\\) :: a => (b => r) -> (a :- b) -> r
r \\ Sub Dict = r
 

The type says that given that a constraint a can be satisfied, a computation that needs a constraint of type b to be satisfied in order to obtain a result, and the fact that a entails b, we can compute the result.

The constraint a is satisfied by the type signature, and the fact that we get quietly passed whatever dictionary is needed. Pattern matching on Sub brings into scope a computation of type (a => Dict b), and we are able to discharge the a obligation, using the dictionary we were passed, Pattern matching on Dict forces that computation to happen and brings b into scope, allowing us to meet the obligation of the computation of r. All of this happens for us behind the scenes just by pattern matching.

So what can we do with this?

We can use \\ to compose constraints.

 
trans :: (b :- c) -> (a :- b) -> a :- c
trans f g = Sub $ Dict \\ f \\ g
 

In fact, the way the dictionaries get plumbed around inside the argument to Sub is rather nice, because we can give that same definition different type signatures, letting us make (,) more product-like, giving us the canonical product morphism to go with the weakenings/projections we defined above.

 
(&&&) :: (a :- b) -> (a :- c) -> a :- (b, c)
f &&& g = Sub $ Dict \\ f \\ g
 

And since we're using it as a product, we can make it act like a bifunctor also using the same definition.

 
(***) :: (a :- b) -> (c :- d) -> (a, c) :- (b, d)
f *** g = Sub $ Dict \\ f \\ g
 

Limited Sub-Superkinding?

Ideally we'd be able to capture something like that bifunctoriality using a type like

 
#if 0
class BifunctorS (p :: Constraint -> Constraint -> Constraint) where
  bimapS :: (a :- b) -> (c :- d) -> p a c :- p b d
#endif
 

In an even more ideal world, it would be enriched using something like

 
#ifdef POLYMORPHIC_KINDS
class Category (k :: x -> x -> *) where
  id :: k a a
  (.) :: k b c -> k a b -> k a c
instance Category (:-) where
  id = refl
  (.) = trans
#endif
 

where x is a kind variable, then we could obtain a more baroque and admittedly far less thought-out bifunctor class like:

 
#if 0
class Bifunctor (p :: x -> y -> z) where
  type Left p :: x -> x -> *
  type Left p = (->)
  type Right p :: y -> y -> *
  type Right p = (->)
  type Cod p :: z -> z -> *
  type Cod p = (->)
  bimap :: Left p a b -> Right p c d -> Cod p (p a c) (p b d)
#endif
 

Or even more more ideally, you could use the fact that we can directly define product categories!

Since they are talking about kind-indexing for classes and type families, we could have separate bifunctors for (,) for both kinds * and Constraint.

The current constraint kind code uses a hack to let (a,b) be used as a type inhabiting * and as the syntax for constraints. This hack is limited however. It only works when the type (,) is fully applied to its arguments. Otherwise you'd wind up with the fact that the type (,) needs to have both of these kinds:

 
-- (,) :: Constraint -> Constraint -> Constraint and
-- (,) :: * -> * -> *
 

What is currently done is that the kind magically switches for () and (,) in certain circumstances. GHC already had some support for this because it parses (Foo a, Bar b) as a type in (Foo a, Bar b) => Baz a b before transforming it into a bunch of constraints.

Since we already have a notion of sub-kinding at the kind level, we could solve this for () by making up a new kind, say, ??? which is the subkind of both * and Constraint, but this would break the nice join lattice properties of the current system.

[Edit: in the initial draft, I had said superkind]

 
--    ?
--   / \
-- (#)  ??
--     /  \
--    #    *  Constraint
--          \ /
--          ???
 

But this doesn't address the kind of (,) above. With the new polymorphic kinds that Brent Yorgey and company have been working on and a limited notion of sub-superkinding, this could be resolved by making a new super-kind @ that is the super-kind of both * and Constraint, and which is a sub-superkind of the usual unnamed Box superkind.

 
-- Box
--  |
--  @
 

Then we can have:

 
-- (,) :: forall (k :: @). k -> k -> k
-- () :: forall (k :: @). k
 

and kind checking/inference will do the right thing about keeping the kind ambiguous for types like (,) () :: forall (k :: @). k

This would get rid of the hack and let me make a proper bifunctor for (,) in the category of entailments.

The version of GHC head I'm working with right now doesn't support polymorphic kinds, so I've only been playing with these in a toy type checker, but I'm really looking forward to being able to have product categories!

Stay Tuned

Next, we'll go over how to reflect the class and instance declarations so we can derive entailment of a superclass for a class, and the entailment of instances.

[Source]

In particular, I want to talk about a novel encoding of linear functionals, polynomials and linear maps in Haskell, but first we're going to have to build up some common terminology.

Having obtained the blessing of Wolfgang Jeltsch, I replaced the algebra package on hackage with something... bigger, although still very much a work in progress.

(Infinite) Modules over Semirings

Recall that a vector space V over a field F is given by an additive Abelian group on V, and a scalar multiplication operator
(.*) :: F -> V -> V subject to distributivity laws

 
s .* (u + v) = s .* u + s .* v
(s + t) .* v = s .* v + t .* v
 

and associativity laws

 
   (s * t) .* v = s .* (t .* v)
 

and respect of the unit of the field.

 
   1 .* v = v
 

Since multiplication on a field is commutative, we can also add

 
  (*.) :: V -> F -> V
  v *. f = f .* v
 

with analogous rules.

But when F is only a Ring, we call the analogous structure a module, and in a ring, we can't rely on the commutativity of multiplication, so we may have to deal left-modules and right-modules, where only one of those products is available.

We can weaken the structure still further. If we lose the negation in our Ring we and go to a Rig (often called a Semiring), now our module is an additive moniod.

If we get rid of the additive and multiplicative unit on our Rig we get down to what some authors call a Ringoid, but which we'll call a Semiring here, because it makes the connection between semiring and semigroup clearer, and the -oid suffix is dangerously overloaded due to category theory.

First we'll define additive semigroups, because I'm going to need both additive and multiplicative monoids over the same types, and Data.Monoid has simultaneously too much and too little structure.

 
-- (a + b) + c = a + (b + c)
class Additive m where
  (+) :: m -> m -> m
  replicate1p :: Whole n => n -> m -> m -- (ignore this for now)
  -- ...
 

their Abelian cousins

 
-- a + b = b + a
class Additive m => Abelian m
 

and Multiplicative semigroups

 
-- (a * b) * c = a * (b * c)
class Multiplicative m where
  (*) :: m -> m -> m
  pow1p :: Whole n => m -> n -> m
  -- ...
 

Then we can define a semirings

 
-- a*(b + c) = a*b + a*c
-- (a + b)*c = a*c + b*c
class (Additive m, Abelian m, Multiplicative m) => Semiring
 

With that we can define modules over a semiring:

 
-- r .* (x + y) = r .* x + r .* y
-- (r + s) .* x = r .* x + s .* x
-- (r * s) .* x = r .* (s .* x)
class (Semiring r, Additive m) => LeftModule r m
   (.*) :: r -> m -> m
 

and analogously:

 
class (Semiring r, Additive m) => RightModule r m
   (*.) :: m -> r -> m
 

For instance every additive semigroup forms a semiring module over the positive natural numbers (1,2..) using replicate1p.

If we know that our addition forms a monoid, then we can form a module over the naturals as well

 
-- | zero + a = a = a + zero
class
    (LeftModule Natural m,
    RightModule Natural m
    ) => AdditiveMonoid m where
   zero :: m
   replicate :: Whole n => n -> m -> m
 

and if our addition forms a group, then we can form a module over the integers

 
-- | a + negate a = zero = negate a + a
class
    (LeftModule Integer m
    , RightModule Integer m
    ) => AdditiveGroup m where
  negate :: m -> m
  times :: Integral n => n -> m -> m
  -- ...
 

Free Modules over Semirings

A free module on a set E, is a module where the basis vectors are elements of E. Basically it is |E| copies of some (semi)ring.

In Haskell we can represent the free module of a ring directly by defining the action of the (semi)group pointwise.

 
instance Additive m => Additive (e -> m) where
   f + g = \x -> f x + g x
 
instance Abelian m => Abelian (e -> m)
 
instance AdditiveMonoid m => AdditiveMonoid (e -> m) where
   zero = const zero
 
instance AdditiveGroup m => AdditveGroup (e -> m) where
   f - g = \x -> f x - g x
 

We could define the following

 
instance Semiring r => LeftModule r (e -> m) where
   r .* f = \x -> r * f x
 

but then we'd have trouble dealing with the Natural and Integer constraints above, so instead we lift modules

 
instance LeftModule r m => LeftModule r (e -> m) where
   (.*) m f e = m .* f e
 
instance RightModule r m => RightModule r (e -> m) where
   (*.) f m e = f e *. m
 

We could go one step further and define multiplication pointwise, but while the direct product of |e| copies of a ring _does_ define a ring, and this ring is the one provided by the Conal Elliot's vector-space package, it isn't the most general ring we could construct. But we'll need to take a detour first.

Linear Functionals

A Linear functional f on a module M is a linear function from a M to its scalars R.

That is to say that, f : M -> R such that

 
f (a .* x + y) = a * f x + f y
 

Consequently linear functionals also form a module over R. We call this module the dual module M*.

Dan Piponi has blogged about these dual vectors (or covectors) in the context of trace diagrams.

If we limit our discussion to free modules, then M = E -> R, so a linear functional on M looks like (E -> R) -> R
subject to additional linearity constraints on the result arrow.

The main thing we're not allowed to do in our function is apply our function from E -> R to two different E's and then multiply the results together. Our pointwise definitions above satisfy those linearity constraints, but for example:

 
bad f = f 0 * f 0
 

does not.

We could capture this invariant in the type by saying that instead we want

 
newtype LinearM r e =
  LinearM {
    runLinearM :: forall r. LeftModule r m => (e -> m) -> m
  }
 

we'd have to make a new such type every time we subclassed Semiring. I'll leave further exploration of this more exotic type to another time. (Using some technically illegal module instances we can recover more structure that you'd expect.)

Now we can package up the type of covectors/linear functionals:

 
infixr 0 $*
newtype Linear r a = Linear { ($*) :: (a -> r) -> r }
 

The sufficiently observant may have already noticed that this type is the same as the Cont monad (subject to the linearity restriction on the result arrow).

In fact the Functor, Monad, Applicative instances for Cont all carry over, and preserve linearity.

(We lose callCC, but that is at least partially due to the fact that callCC has a less than ideal type signature.)

In addition we get a number of additional instances for Alternative, MonadPlus, by exploiting the knowledge that r is ring-like:

 
instance AdditiveMonoid r => Alternative (Linear r a) where
  Linear f < |> Linear g = Linear (f + g)
  empty = Linear zero
 

Note that the (+) and zero there are the ones defined on functions from our earlier free module construction!

Linear Maps

Since Linear r is a monad, Kleisli (Linear r) forms an Arrow:

 
b -> ((a -> r) ~> r)
 

where the ~> denotes the arrow that is constrained to be linear.

If we swap the order of the arguments so that

 
(a -> r) ~> (b -> r)
 

this arrow has a very nice meaning! (See Numeric.Map.Linear)

 
infixr 0 $#
newtype Map r b a = Map { ($#) :: (a -> r) -> (b -> r) }
 

Map r b a represents the type of linear maps from a -> b. Unfortunately due to contravariance the arguments wind up in the "wrong" order.

 
instance Category (Map r) where
  Map f . Map g = Map (g . f)
  id = Map id
 

So we can see that a linear map from a module A with basis a to a vector space with basis b effectively consists of |b| linear functionals on A.

Map r b a provides a lot of structure. It is a valid instance of an insanely large number of classes.

Vectors and Covectors

In physics, we sometimes call linear functionals covectors or covariant vectors, and if we're feeling particularly loquacious, we'll refer to vectors as contravariant vectors.

This has to do with the fact that when you change basis, you change map the change over covariant vectors covariantly, and map the change over vectors contravariantly. (This distinction is beautifully captured by Einstein's summation notation.)

We also have a notion of covariance and contravariance in computer science!

Functions vary covariantly in their result, and contravariant in their argument. E -> R is contravariant in E. But we chose this representation for our free modules, so the vectors in our free vector space (or module) are contravariant in E.

 
class Contravariant f where
  contramap :: (a -> b) -> f a -> f b
 
-- | Dual function arrows.
newtype Op a b = Op { getOp :: b -> a } 
 
instance Contravariant (Op a) where
  contramap f g = Op (getOp g . f)
 

On the other hand (E -> R) ~> R varies covariantly with the change of E.

as witnessed by the fact that it is a Functor.

 
instance Functor (Linear r) where
  fmap f m = Linear $ \k -> m $* k . f
 

We have lots of classes for manipulating covariant structures, and most of them apply to both (Linear r) and (Map r b).

Other Representations and Design Trade-offs

One common representation of vectors in a free vector space is as some kind of normalized list of scalars and basis vectors. In particular, David Amos's wonderful HaskellForMaths uses

 
newtype Vect r a = Vect { runVect :: [(r,a)] }
 

for free vector spaces, only considering them up to linearity, paying for normalization as it goes.

Given the insight above we can see that Vect isn't a representation of vectors in the free vector space, but instead represents the covectors of that space, quite simply because Vect r a varies covariantly with change of basis!

Now the price of using the Monad on Vect r is that the monad denormalizes the representation. In particular, you can have multiple copies of the same basis vector., so any function that uses Vect r a has to merge them together.

On the other hand with the directly encoded linear functionals we've described here, we've placed no obligations on the consumer of a linear functional. They can feed the directly encoded linear functional any vector they want!

In fact, it'll even be quite a bit more efficient to compute,

To see this, just consider:

 
instance MultiplicativeMonoid r => Monad (Vect r) where
   return a = Vect [(1,a)]
   Vect as >>= f = Vect
       [ (p*q, b) | (p,a) < - as, (q,b) <- runVect (f b) ]
 

Every >>= must pay for multiplication. Every return will multiply the element by one. On the other hand, the price of return and bind in Linear r is function application.

 
instance Monad (Linear r) where
  return a = Linear $ \k -> k a
  m >>= f = Linear $ \k -> m $* \a -> f a $* k
 

A Digression on Free Linear Functionals

To wax categorical for a moment, we can construct a forgetful functor U : Vect_F -> Set that takes a vector space over F to just its set of covectors.

 
F E = (E -> F, F,\f g x -> f x + g x ,\r f x -> r * f x)
 

using the pointwise constructions we built earlier.

Then in a classical setting, you can show that F is left adjoint to U.

In particular the witnesses of this adjunction provide the linear map from (E -> F) to V and the function E -> (V ~> F) giving a linear functional on V for each element of E.

In a classical setting you can go a lot farther, and show that all vector spaces (but not all modules) are free.

But in a constructive setting, such as Haskell, we need a fair bit to go back and forth, in particular we wind up need E to be finitely enumerable to go one way, and for it to have decidable equality to go in the other. The latter is fairly easy to see, because even going from E -> (E -> F) requires that we can define and partially apply something like Kronecker's delta:

 
delta :: (Rig r, Eq a) => e -> e -> r
delta i j | i == j = one
             | otherwise = zero
 

The Price of Power

The price we pay is that, given a Rig, we can go from Vect r a to Linear r a but going back requires a to be be finitely enumerable (or for our functional to satisfy other exotic side-conditions).

 
vectMap :: Rig r => Vect r a -> Linear r a
vectMap (Vect as) = Map $ \k -> sum [ r * k a | (r, a) < - as ]
 

You can still probe Linear r a for individual coefficients, or pass it a vector for polynomial evaluation very easily, but for instance determining a degree of a polynomial efficiently requires attaching more structure to your semiring, because the only value you can get out of Linear r a is an r.

Optimizing Linear Functionals

In both the Vect r and Linear r cases, excessive use of (>>=) without somehow normalizing or tabulating your data will cause a lot of repeated work.

This is perhaps easiest to see from the fact that Vect r never used the addition of r, so it distributed everything into a kind of disjunctive normal form. Linear r does the same thing.

If you look at the Kleisli arrows of Vect r or Linear r as linear mappings, then you can see that Kleisli composition is going to explode the number of terms.

So how can we collapse back down?

In the Kleisli (Vect r) case we usually build up a map as we walk through the list then spit the list back out in order having added up like terms.

In the Map r case, we can do better. My representable-tries package provides a readily instantiable HasTrie class, and the method:

 
memo :: HasTrie a => (a -> r) -> a -> r
 

which is responsible for providing a memoized version of the function from a -> r in a purely functional way. This is obviously a linear map!

 
memoMap :: HasTrie a => Map r a a
memoMap = Map memo
 

We can also flip memo around and memoize linear functionals.

 
memoLinear :: HasTrie a => a -> Linear r a
memoLinear = Linear . flip memo
 

Next time, (co)associative (co)algebras and the myriad means of multiplying (co)vectors!