I was looking at the Eq2 instances for Data.Map and Data.HashMap, and
the Eq1 instances for Set and HashSet, and I realized that they're a bit ... weird. My instinct is to remove these instances immediately, but I figure I should first check with the community (and Oleg, who added them to begin with) to make sure they don't make some sort of sense I don't understand. In particular, these instances compare keys in the order in which they appear in the container. That order may be completely unrelated to the given key comparison function. Data.Map example: suppose I have a Map k () and a Map (Down k) (), where Down is from Data.Ord. If I call liftEq2 (\x (Down y) -> x == y) (==), I will get what looks to me like a totally meaningless result. Data.HashMap example: suppose I have a HashMap Int () and a Hashmap String (). If I call liftEq2 (\x y -> show x == y) (==), that won't return True even if the strings in the second map are actually the results of applying show to the numbers in the first map. Intuitively, I think Eq1 f only makes sense if the shape of f x does not depend on the values of type x, and Eq2 p only makes sense if the shape of p x y does not depend on the values of types x and y. Is there a way to formalize this intuition with class laws? I believe that in the case of a Functor, parametricity will guarantee that liftEq eq (f <$> xs) (g <$> ys) == liftEq (\x y -> eq (f x) (g y)) xs ys Are there *any* legitimate instances of Eq1 or Ord1 that are not Functors? Are there *any* legitimate instances of Eq2 or Ord2 that are not Bifunctors? My intuition says no. We might wish we could write instances for unboxed arrays or vectors, but I believe that is totally impossible anyway. David _______________________________________________ Libraries mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries |
My hand-wavy intuition:
Eq witnesses an equality. a -> a -> Bool type isn't enough to guarantee that, so we have laws. The Eq1 class has member liftEq :: (a -> b -> Bool) -> f a -> f b -> Bool the type `a -> b -> Bool` is not an equality decision procedure, but rather an isomorphism one. It looks like for me, that liftEq gives a natural transformation from `Iso a b` to `Iso (f a) (f b)`. Does this make sense? The liftEq2 is the same. We have to think Map as not of Hask * Hask -> Hask (wrong) Functor but OrdHask * Hask -> Hask (OrdHask: Ord-types and monotone functions) Then, comparing Map k () with Map (Down k) () with \k (Down k') = k == k' is not well defined, as that "isomorphism candidate" doesn't respect ordering. (as an example, for k = Integer, \k (Down k') = k == negate k' would work, or even \n m -> n == succ m, for n m :: Integer) Similar argument can be used on (Eq k, Hashable k) requirement of HashMap k v, with constraints giving an additional Hashable structure. The motivational example is being able to compare HashMap Text Foo and HashMap BusinessId Foo' directly, where newtype BusinessId = BusinessId Text deriving newtype (Eq, Hashable) As Hashable is somewhat arbitrary, I cannot think of other use-cases (preserving Hashable is next to impossible otherwise then via newtype deriving). However, for Map there might be A, B such that `Map A v` and `Map B v` are comparable, i.e. there is monotone f :: A -> B, and compareAB a b = f a `compare` b which we can use as a first argument for liftCompare2 (Ord2). (There is mapKeysMonotonic which can be used to achieve the same effect! unordered-containers could have similar unsafe function too). That said, I didn't have such use case myself (yet!?). When writing the instances, I was thinking what's (a -> b -> Ordering), and at this point I have to wave hands even more. I don't know how to complete the following sentence The ??? to total order is as isomorphism to equality. As a side-note: we have - `~` witnessing nominal equality - `Coercible` witnessing structural equality - It would be very cool to being to say `CoercibleInstance Hashable`, which is satisfied when `Hashable` is `newtype` derived. Is it roles, refining kinds, or something else, I don't know. But I feel it would be useful. It can be used to give `mapKeysMonotonic` a safe type (at least in some cases!) Maybe if `Constraint ~ Type`, we could simply require `Coercible (Hashable k) (Hashable k')`, but how it would fit the roles? So I do agree, that Eq1 f has functorial feel and it's "natural", this is what its type tells. But if you imply that we should add `Functor` or `Bifunctor` super-classes, then I disagree, based on above arguments. Yes, the functions are unsafe to use (to get meaningful results), but I see a value for them. An example of legitimate (in above sence) instance of Eq1 which isn't a functor is Eq1 Set. As long as `eq :: a -> b -> Bool` respects the ordering of `a` and `b`, we can compare `Set a` with `Set b`. <Insert a riddle about comparing sets of apples and oranges, if one can compare apples with oranges, so fruit comparison diagram commutes> Cheers, Oleg > I was looking at the Eq2 instances for Data.Map and Data.HashMap, and > the Eq1 instances for Set and HashSet, and I realized that they're a > bit ... weird. My instinct is to remove these instances immediately, > but I figure I should first check with the community (and Oleg, who > added them to begin with) to make sure they don't make some sort of > sense I don't understand. In particular, these instances compare keys > in the order in which they appear in the container. That order may be > completely unrelated to the given key comparison function. > > Data.Map example: suppose I have a Map k () and a Map (Down k) (), > where Down is from Data.Ord. If I call liftEq2 (\x (Down y) -> x == y) > (==), I will get what looks to me like a totally meaningless result. > > Data.HashMap example: suppose I have a HashMap Int () and a Hashmap > String (). If I call liftEq2 (\x y -> show x == y) (==), that won't > return True even if the strings in the second map are actually the > results of applying show to the numbers in the first map. > > Intuitively, I think Eq1 f only makes sense if the shape of f x does > not depend on the values of type x, and Eq2 p only makes sense if the > shape of p x y does not depend on the values of types x and y. Is > there a way to formalize this intuition with class laws? I believe > that in the case of a Functor, parametricity will guarantee that > > liftEq eq (f <$> xs) (g <$> ys) == liftEq (\x y -> eq (f x) (g y)) xs ys > > Are there *any* legitimate instances of Eq1 or Ord1 that are not > Functors? Are there *any* legitimate instances of Eq2 or Ord2 that are > not Bifunctors? My intuition says no. We might wish we could write > instances for unboxed arrays or vectors, but I believe that is totally > impossible anyway. > > David On 15.03.2018 23:48, David Feuer wrote: > I was looking at the Eq2 instances for Data.Map and Data.HashMap, and > the Eq1 instances for Set and HashSet, and I realized that they're a > bit ... weird. My instinct is to remove these instances immediately, > but I figure I should first check with the community (and Oleg, who > added them to begin with) to make sure they don't make some sort of > sense I don't understand. In particular, these instances compare keys > in the order in which they appear in the container. That order may be > completely unrelated to the given key comparison function. > > Data.Map example: suppose I have a Map k () and a Map (Down k) (), > where Down is from Data.Ord. If I call liftEq2 (\x (Down y) -> x == y) > (==), I will get what looks to me like a totally meaningless result. > > Data.HashMap example: suppose I have a HashMap Int () and a Hashmap > String (). If I call liftEq2 (\x y -> show x == y) (==), that won't > return True even if the strings in the second map are actually the > results of applying show to the numbers in the first map. > > Intuitively, I think Eq1 f only makes sense if the shape of f x does > not depend on the values of type x, and Eq2 p only makes sense if the > shape of p x y does not depend on the values of types x and y. Is > there a way to formalize this intuition with class laws? I believe > that in the case of a Functor, parametricity will guarantee that > > liftEq eq (f <$> xs) (g <$> ys) == liftEq (\x y -> eq (f x) (g y)) xs ys > > Are there *any* legitimate instances of Eq1 or Ord1 that are not > Functors? Are there *any* legitimate instances of Eq2 or Ord2 that are > not Bifunctors? My intuition says no. We might wish we could write > instances for unboxed arrays or vectors, but I believe that is totally > impossible anyway. > > David _______________________________________________ Libraries mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries signature.asc (836 bytes) Download Attachment |
I believe we should not add a Functor constraint, but for a different reason. One could write data Arr a where I :: Unboxed.Vector Int -> Arr Int C :: Unboxed.Vector Char -> Arr Char This has no valid Functor instance, but it has a perfectly sensible Ord1 instance. As for Map and HashMap, I'd be inclined to remove the "dangerous" instances and add functions (names to be determined) people can use when they know what they're doing. The trouble with the instances is that there's no apparent way to express in the *class* laws what makes an argument valid or invalid. It becomes ad hoc overloading, and I don't think most Haskellers are really into that. On Mar 15, 2018 7:55 PM, "Oleg Grenrus" <[hidden email]> wrote: My hand-wavy intuition: _______________________________________________ Libraries mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries |
In reply to this post by Oleg Grenrus
So looking at the “bad” examples they both seem to only be bad for Eq2, and be bad because they “abuse” a design choice in the class. In particular, we have class Eq2 f where liftEq2 :: (a -> b -> Bool) -> (c -> d -> Bool) -> f a c -> f b d -> Bool Source# — Lift equality tests through the type constructor. — The function will usually be applied to equality functions, but the more general type ensures that the implementation uses them to compare elements of the first container with elements of the second. This is to say that the two types, `a` and `b`, are intended to be equal. However, to prevent implementations from making the error of comparing two values from the first container, they are made distinct. So in the intended usage, this is not called directly but rather via eq2. eq2 :: (Eq2 f, Eq a, Eq b) => f a b -> f a b -> Bool What David has pointed out, I think, is that this clever trick, which prevents some bad implementations, is open to abuse in the case of _good_ implementations, since the assumption of the two types being equal is necessary for some structural invariants to hold. (In particular, if the `a` and `b` differ, they will have different ordering properties). So I see two options here, neither of which involves removing any instances. A) change the types in the Eq2 and Ord2 classes to not use this clever trick. I think this should not break anything, since according to the documentation, nothing should be making use of the possibility of these types varying. B) change the documentation of the classes to make explicit that the laws are _only_ expected to apply when the types are equal, and when they are not, the behavior is undefined. Cheers, Gershom On March 15, 2018 at 7:57:34 PM, Oleg Grenrus ([hidden email]) wrote:
_______________________________________________ Libraries mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries |
For Map and HashMap, yes, Eq2 and Ord2 are the problem. For Set and HashSet, Eq1 and Ord1 are trouble. The advantage of the flexibility is that for a Functor or Bifunctor you get an optimization: instead of mapping a function or two over each container to give them the same type and then comparing the results, you can fuse it all into one operation. Of course, the same thing could also be done for each type using rewrite rules, but that's kind of gross. On Mar 15, 2018 11:36 PM, "Gershom B" <[hidden email]> wrote:
_______________________________________________ Libraries mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries |
On March 15, 2018 at 11:50:45 PM, David Feuer ([hidden email]) wrote:
Right. But it is the possibility of this optimization that breaks things. In particular, `Set` is not a functor (and does not permit mapping over it), but it _should_ be able to be an instance of Eq1. -g _______________________________________________ Libraries mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries |
Ah, I think I see what you mean. One option: class EqFoo1 f where liftEqSame :: (a -> a -> Bool) -> f a -> f a -> Bool -- and/or the more principled liftEqOnSame :: Eq b => (a -> b) -> f a -> f a -> Bool class (Functor f, EqFoo1) => EqBar1 where liftEqDifferent :: (a -> b -> Bool) -> f a -> f b -> Bool -- and/or the more principled liftEqOnDifferent :: Eq c => (a -> c) -> (b -> c) -> f a -> f b -> Bool This wouldn't be the best for my hypothetical Arr type, but I don't know that it's an important enough case to worry about. On Mar 15, 2018 11:55 PM, "Gershom B" <[hidden email]> wrote:
_______________________________________________ Libraries mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries |
Free forum by Nabble | Edit this page |