Canonical categorical conversions (injections and projections)
Canonical injections and projections. See README.md for more details.
injections
Canonical categorical conversions (injections and projections) between Haskell types.
Problem: We want predictable conversions between types that have equivalent semantics in their domains but not equivalent representations (so we can't use coerce
).
Alternatively: Rust's From
trait with Haskell flavor.
Injection
The class Injection
describes a lossless conversion from one type to another; that is, the sole method of the class,
inject :: from -> into
takes a value input :: from
and returns a value output :: into
which preserves all the information contained in the input. Specifically, each input
is mapped to a uniqueoutput
. In mathematical terminology, inject
is injective:
inject a ≡ inject b → a ≡ b
the outputs of two calls to inject
is the same only if the inputs are the same. The name of the class is derived from the mathematical term.
Injection
models the "is-a" relationship used in languages with subtypes (such as in object-oriented programming), but an explicit cast with inject
is required in Haskell.
Examples
There are many examples of injections scattered throughout Haskell, but the Injection
class collects them in one place. Here we present some examples. Some of these instances have been generalized in the library, but we present the simplified version here for explanatory purposes.
Dynamic
is a dynamically-typed wrapper for values of all Typeable
types. Wrapping a value with toDyn
preserves the value exactly, so we know that it is injective:
instance Typeable a => Injection a Dynamic where
inject = toDyn
The constructor Just :: a -> Maybe a
is an injective function:
instance Injection a (Maybe a) where
inject = Just
Constructors with a single argument are always injective functions. We must use Just
here; if we had written inject _ = Nothing
, that would violate the injective law. Usually, any putative definition of inject
with a wildcard match on the left-hand side will fail to be injective. There are exceptions to this guideline; for example, this instance is injective:
-- BAD!
instance Injection () (Maybe ()) where
-- This is injective because the type () has only one value.
inject _ = Nothing
However, we also require instances to be canonical. This instance isn't canonical because it arbitrarily restricts @from ~ ()@. Actually, there is already the definition of a canonical injection into Maybe
; we may as well write
instance Injection a (Maybe a) where
inject = pure
One consequence of the injectivity law is that the output type must be at least as large as the input type. We can inject a Maybe a
into a [a]
because the latter type is strictly larger:
instance Injection (Maybe a) [a] where
inject = maybeToList
Where Maybe a
contains either zero or one values of a
, [a]
contains zero or more values of a
.
Some common conversions are notably not injective. For example, Data.Map.fromList
returns the same Map
for different lists:
Data.Map.fromList [('a', 'A'), ('b', 'B')] == Data.Map.fromList [('b', 'B'), ('a', 'A')]
Therefore, we cannot define an instance [(k, v)] (Map k v)
.
When there is an equivalence between two types, that equivalence is usually an injection. For example, the class Integral
defines
toInteger :: Integra a => a -> Integer
Where this conversion is a total equivalence, it forms a canonical injection into Integer
:
instance Injection Natural Integer where
inject = toInteger
Likewise, the class Num
defines an equivalence
fromInteger :: Num a => Integer -> a
For types that have total implementations of fromInteger
, this is usually an injection:
instance HasResolution a => Injection Integer (Fixed a) where
inject = fromInteger
-- BAD: No reasonable person would accept this!
instance Injection String Text where
inject = Text.pack . reverse
Retraction
Because Injection
is a lossless conversion, we can define a Retraction
which undoes it. The method
retract :: into -> Maybe from
is the (left) inverse of inject
:
retract (inject x) = Just x
retract
is partial (returns Maybe
) because the type into
may be larger than the type from
; that is, there may be values in into
which are not inject
-ed from from
, and in that case retract
may return Nothing
.
Examples
instance Typeable a => Retraction a Dynamic where
retract = fromDyn
instance Retraction a (Maybe a) where
retract = id
instance Retraction (Maybe a) [a] where
retract [] = Just Nothing
retract [x] = Just (Just x)
retract _ = Nothing
instance Retraction Natural Integer where
retract x
| x < 0 = Nothing
| otherwise = Just (fromInteger x)