[英]What are the rules to compose f-algebras in a catamorphism
Here are some simple F-algebras for lists. 这里有一些简单的F代数用于列表。 They work with the cata
function from the recursion-schemes library. 它们使用递归方案库中的cata
函数。
import Data.Functor.Foldable
algFilterSmall :: ListF Int [Int] -> [Int]
algFilterSmall Nil = []
algFilterSmall (Cons x xs) = if x >= 10 then (x:xs) else xs
algFilterBig :: ListF Int [Int] -> [Int]
algFilterBig Nil = []
algFilterBig (Cons x xs) = if x < 100 then (x:xs) else xs
algDouble :: ListF Int [Int] -> [Int]
algDouble Nil = []
algDouble (Cons x xs) = 2*x : xs
algTripple :: ListF Int [Int] -> [Int]
algTripple Nil = []
algTripple (Cons x xs) = 3*x : xs
Now I compose them: 现在我把它们组成:
doubleAndTripple :: [Int] -> [Int]
doubleAndTripple = cata $ algTripple . project . algDouble
-- >>> doubleAndTripple [200,300,20,30,2,3]
-- [1200,1800,120,180,12,18]
doubleAndTriple
works as expected. doubleAndTriple
按预期工作。 Both algebras are structure preserving , they don't change the length of the list, so cata can apply both algebras to every item of the list. 两个代数都是结构保留的 ,它们不会改变列表的长度,因此cata可以将两个代数应用于列表的每个项目。
Next one is filter and double: 下一个是过滤和双重:
filterAndDouble :: [Int] -> [Int]
filterAndDouble = cata $ algDouble . project . algFilterBig
-- >>> filterAndDouble [200,300,20,30,2,3]
-- [160,60,4,6]
It doesn't work properly. 它无法正常工作。 I assume it's because algFilterBig
is not structure preserving. 我认为这是因为algFilterBig
不是结构保留。
Now the last example: 现在是最后一个例子:
filterBoth :: [Int] -> [Int]
filterBoth = cata $ algFilterSmall . project . algFilterBig
-- >>> filterBoth [200,300,20,30,2,3]
-- [20,30]
Here both algebras are not structure preserving, but this example is working. 这两个代数都不是结构保留,但这个例子是有效的。
What are the exact rules for composing f-algebras? 组成f-algebras的确切规则是什么?
"Structure preserving algebras" can be formalized as natural transformations (that can be between different functors). “保持代数的结构”可以形式化为自然变换(可以在不同的算子之间)。
inList :: ListF a [a] -> [a]
inList Nil = []
inList (Cons a as) = a : as
ntDouble, ntTriple :: forall a. ListF Int a -> ListF Int a
algDouble = inList . ntDouble
algTriple = inList . ntTriple
Then, for any algebra f
and natural transformation n
, 然后,对于任何代数f
和自然变换n
,
cata (f . inList . n) = cata f . cata n
The doubleAndTriple
example is an instance of that for f = algTriple
and n = ntDouble
. doubleAndTriple
示例是f = algTriple
和n = ntDouble
。
This doesn't easily generalize to larger classes of algebras. 这并不容易推广到更大类的代数。
I think the case of filter is easier to see without categories, as a consequence of the fact that filter
is a semigroup homomorphism: filter p . filter q = filter (liftA2 (&&) pq)
我认为滤波器的情况更容易看到没有类别,因为filter
是半群同态的事实: filter p . filter q = filter (liftA2 (&&) pq)
filter p . filter q = filter (liftA2 (&&) pq)
. filter p . filter q = filter (liftA2 (&&) pq)
。
I searched for a general but sufficient condition for a "distributive law" on filter-like algebras. 我在类似滤波器的代数上寻找一个通用但充分条件的“分配律”。 Abbreviate a bit afs = algFilterSmall
, afb = algFilterBig
. 缩写为afs = algFilterSmall
, afb = algFilterBig
。 We reason backwards, to find a sufficient condition for: 我们向后推理,找到一个充分条件:
cata (afs . project . afb) = cata afs . cata afb -- Equation (0)
By definition of a catamorphism, z = cata (afs . project . afb)
is the unique solution to this equation (a disguised commutative diagram): 根据catamorphism的定义, z = cata (afs . project . afb)
是这个等式的唯一解(一个伪装的交换图):
z . inList = afs . project . afb . fmap z
Substitute z
with the RHS of the previous equation: 用前一个等式的RHS代替z
:
cata afs . cata afb . inList = afs . project . afb . fmap (cata afs . cata afb)
-- (1), equivalent to (0)
On the LHS, we apply the definition of cata
as a Haskell function, cata afb = afb . fmap (cata afb) . project
在LHS上,我们将cata
的定义cata
为Haskell函数, cata afb = afb . fmap (cata afb) . project
cata afb = afb . fmap (cata afb) . project
cata afb = afb . fmap (cata afb) . project
, and simplify with project . inList = id
cata afb = afb . fmap (cata afb) . project
,并简化project . inList = id
project . inList = id
; project . inList = id
;
on the RHS, we apply a functor law fmap (f . g) = fmap f . fmap g
在RHS上,我们应用了一个fmap (f . g) = fmap f . fmap g
子法fmap (f . g) = fmap f . fmap g
fmap (f . g) = fmap f . fmap g
. fmap (f . g) = fmap f . fmap g
。
This yields: 这会产生:
cata afs . afb . fmap (cata afb) = afs . project . afb . fmap (cata afs) . fmap (cata afb)
-- (2), equivalent to (1)
We "cosimplify" away the last factor fmap (cata afb)
(remember that we are reasoning backwards): 我们“缩小”最后一个因素fmap (cata afb)
(记住我们向后推理):
cata afs . afb = afs . project . afb . fmap (cata afs) -- (3), implies (2)
This is the simplest one I could come up with. 这是我能想到的最简单的一个。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.