Re: In opposition of Functor as super-class of Monad
http://haskell.1045720.n5.nabble.com/Interfaces-the-Golden-Path-of-Haskell-tt5732208.html With this Haskell extension it could be possibe to implement both (Monad m) and (Applicative m = Monad m). Not only both, we could define a lot of Monad for many users. -- View this message in context: http://haskell.1045720.n5.nabble.com/Re-In-opposition-of-Functor-as-super-class-of-Monad-tp5719389p5732573.html Sent from the Haskell - Haskell-prime mailing list archive at Nabble.com. ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
The major advantage that I see of making Applicative a superclass of Monad is actually not anywhere to do with the data types or instances, but rather to do with functions that operate on the class. For example, the liftM* family of functions becomes entirely redundant, and can be eliminated in favour of liftA*, which reduces the number of arbitrary choices we need to make and understand. As another example, we sometimes want to use functions (like void or fmap) that work with Functor and functions (like join) that need Monad in the same (polymorphic) code. We then have to state *both* constraints in the context for the code, despite the fact that if one is satisfied, the other must surely be. For example, defining instance Functor (StateT s m) you have to choose whether to use the context (Monad m) or (Functor m): the latter is more general, but requires your programs that use fmap *and* bind to specify (Monad m, Functor m), and the former is almost always what's used in practice anyway. This tension just doesn't exist if one is a superclass of the other. Some code will break, yes, but I think most people would agree that the presence of a Monad instance with no corresponding Applicative instance is a bug anyway, and should be fixed. (Note that if Applicative and Monad are both defined for the same type, it is required by the Applicative docs that return = pure and * = ap, at least extensionally. Of course the implementation can be cleverer, but if you want it to be actually /different/ in behaviour, please use a newtype). ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
I should have been clearer sorry. I should hope not that Functor - Applicative - Monad. Perhaps I do not understand the purpose of this thread, but fixing the hierarchy in this way is a mistake of similar magnitude to the original position -- one that I would cringe at seeing repeated. That is why I thought such a discussion was on-topic. On 25/10/12 10:12, Ben Franksen wrote: Tony Morris wrote: I should hope not. The identity element (return, coreturn, mempty, pure, Category.id) is almost never needed. * http://hackage.haskell.org/package/semigroupoids * https://gist.github.com/3871764 Off-topic. Feel free to start a new thread named The bombastic one-and-true class hierarchy I always wanted to have. These proposals have their merits, and I greatly respect the category theoretic knowledge that went into them -- but this is another discussion. This thread refers to a rather modest correction in the standard libraries, not a complete re-design. The idea is to fix something that is widely accepted as an unfortunate ommision (in fact, Oleg's comment is one of the very few that question the idea of adding super class constraints to Monad in principle). BTW, it is unclear what your I hope not refers to, since in both of the hierarchies you linked to Applicative *is* a super class of Monad. Cheers On 25/10/12 04:49, Ben Franksen wrote: First, let me make it clear that nowadays we are of course (I hope!) talking about making not only Functor, but Applicative a super-class of Monad (so Functor becomes a super-class by transitivity). Petr P wrote: The main objections were that it would break existing code and that it would lead to code duplication. The former is serious, [...] To address the first objection: I don't buy this it breaks lots of code argument. Adding the missing instances is a complete no-brainer; as you wrote: instance Applicative ... where pure = return (*) = ap instance Functor ... where fmap = liftM I do not think it is unreasonable to expect people to add such a simple and practically automatic fix to some old programs in the interest of cleaning up an old wart (and almost everyone agrees that this would be a good thing, in principle). BTW, I guess most programs already contain the Functor instances (but maybe not Applicative, as it is newer). I agree with Petr Pudlak that code duplication is not an issue, see above. And yes, these automatic instances may have stronger super-class constraints than strictly necessary. So what? The program didn't need the Functor (or Applicative) instance anyway (or it already would have defined it, in which case no change would be needed at all). Default superclass instances strike me as a complicated proposal for solving trivial problems. The switch in Control.Exception (from data Exception to class Exception) was much more disrupting, adapting programs meant lots of changes everywhere exceptions are handled, not just adding some trivial instances. Still people managed the transition. Cheers -- Tony Morris http://tmorris.net/ ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
Tony, I think you misparsed the proposal. The ...'s were for specific monads indicating the additional work required for each Monad. I think the only real proposal on the table is the obvious one of adding Applicative as a superclass of monad. From there there are a couple of incremental improvements that could be made like adding the unimplemented superclass defaults or adding the equivalent of DefaultSignatures to the language spec to reduce the burden on Monad implementors. In practice I think either of those extensions would be premature to add to the language specification at this time. I would be 100% behind adding the Applicative constraint as a superclass of Monad and even perhaps of some bikeshedding, like exporting Applicative from the Prelude, because otherwise you can't define a Monad without an import, while you can now. I would be strongly against requiring superclass defaults or DefaultSignatures in the haskell standard, however. The former is a largely untested point in the design space and the latter has issues where it tightly couples classes with their dependencies, leading to unbreakable cycles between classes that all have to be defined together and poor engineering practices. Best, --Edward On Oct 25, 2012, at 5:46 PM, Tony Morris tonymor...@gmail.com wrote: I should have been clearer sorry. I should hope not that Functor - Applicative - Monad. Perhaps I do not understand the purpose of this thread, but fixing the hierarchy in this way is a mistake of similar magnitude to the original position -- one that I would cringe at seeing repeated. That is why I thought such a discussion was on-topic. On 25/10/12 10:12, Ben Franksen wrote: Tony Morris wrote: I should hope not. The identity element (return, coreturn, mempty, pure, Category.id) is almost never needed. * http://hackage.haskell.org/package/semigroupoids * https://gist.github.com/3871764 Off-topic. Feel free to start a new thread named The bombastic one-and-true class hierarchy I always wanted to have. These proposals have their merits, and I greatly respect the category theoretic knowledge that went into them -- but this is another discussion. This thread refers to a rather modest correction in the standard libraries, not a complete re-design. The idea is to fix something that is widely accepted as an unfortunate ommision (in fact, Oleg's comment is one of the very few that question the idea of adding super class constraints to Monad in principle). BTW, it is unclear what your I hope not refers to, since in both of the hierarchies you linked to Applicative *is* a super class of Monad. Cheers On 25/10/12 04:49, Ben Franksen wrote: First, let me make it clear that nowadays we are of course (I hope!) talking about making not only Functor, but Applicative a super-class of Monad (so Functor becomes a super-class by transitivity). Petr P wrote: The main objections were that it would break existing code and that it would lead to code duplication. The former is serious, [...] To address the first objection: I don't buy this it breaks lots of code argument. Adding the missing instances is a complete no-brainer; as you wrote: instance Applicative ... where pure = return (*) = ap instance Functor ... where fmap = liftM I do not think it is unreasonable to expect people to add such a simple and practically automatic fix to some old programs in the interest of cleaning up an old wart (and almost everyone agrees that this would be a good thing, in principle). BTW, I guess most programs already contain the Functor instances (but maybe not Applicative, as it is newer). I agree with Petr Pudlak that code duplication is not an issue, see above. And yes, these automatic instances may have stronger super-class constraints than strictly necessary. So what? The program didn't need the Functor (or Applicative) instance anyway (or it already would have defined it, in which case no change would be needed at all). Default superclass instances strike me as a complicated proposal for solving trivial problems. The switch in Control.Exception (from data Exception to class Exception) was much more disrupting, adapting programs meant lots of changes everywhere exceptions are handled, not just adding some trivial instances. Still people managed the transition. Cheers -- Tony Morris http://tmorris.net/ ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
OK, sorry about the noise! On 26/10/12 09:41, Edward Kmett wrote: Tony, I think you misparsed the proposal. The ...'s were for specific monads indicating the additional work required for each Monad. I think the only real proposal on the table is the obvious one of adding Applicative as a superclass of monad. From there there are a couple of incremental improvements that could be made like adding the unimplemented superclass defaults or adding the equivalent of DefaultSignatures to the language spec to reduce the burden on Monad implementors. In practice I think either of those extensions would be premature to add to the language specification at this time. I would be 100% behind adding the Applicative constraint as a superclass of Monad and even perhaps of some bikeshedding, like exporting Applicative from the Prelude, because otherwise you can't define a Monad without an import, while you can now. I would be strongly against requiring superclass defaults or DefaultSignatures in the haskell standard, however. The former is a largely untested point in the design space and the latter has issues where it tightly couples classes with their dependencies, leading to unbreakable cycles between classes that all have to be defined together and poor engineering practices. Best, --Edward On Oct 25, 2012, at 5:46 PM, Tony Morris tonymor...@gmail.com wrote: I should have been clearer sorry. I should hope not that Functor - Applicative - Monad. Perhaps I do not understand the purpose of this thread, but fixing the hierarchy in this way is a mistake of similar magnitude to the original position -- one that I would cringe at seeing repeated. That is why I thought such a discussion was on-topic. On 25/10/12 10:12, Ben Franksen wrote: Tony Morris wrote: I should hope not. The identity element (return, coreturn, mempty, pure, Category.id) is almost never needed. * http://hackage.haskell.org/package/semigroupoids * https://gist.github.com/3871764 Off-topic. Feel free to start a new thread named The bombastic one-and-true class hierarchy I always wanted to have. These proposals have their merits, and I greatly respect the category theoretic knowledge that went into them -- but this is another discussion. This thread refers to a rather modest correction in the standard libraries, not a complete re-design. The idea is to fix something that is widely accepted as an unfortunate ommision (in fact, Oleg's comment is one of the very few that question the idea of adding super class constraints to Monad in principle). BTW, it is unclear what your I hope not refers to, since in both of the hierarchies you linked to Applicative *is* a super class of Monad. Cheers On 25/10/12 04:49, Ben Franksen wrote: First, let me make it clear that nowadays we are of course (I hope!) talking about making not only Functor, but Applicative a super-class of Monad (so Functor becomes a super-class by transitivity). Petr P wrote: The main objections were that it would break existing code and that it would lead to code duplication. The former is serious, [...] To address the first objection: I don't buy this it breaks lots of code argument. Adding the missing instances is a complete no-brainer; as you wrote: instance Applicative ... where pure = return (*) = ap instance Functor ... where fmap = liftM I do not think it is unreasonable to expect people to add such a simple and practically automatic fix to some old programs in the interest of cleaning up an old wart (and almost everyone agrees that this would be a good thing, in principle). BTW, I guess most programs already contain the Functor instances (but maybe not Applicative, as it is newer). I agree with Petr Pudlak that code duplication is not an issue, see above. And yes, these automatic instances may have stronger super-class constraints than strictly necessary. So what? The program didn't need the Functor (or Applicative) instance anyway (or it already would have defined it, in which case no change would be needed at all). Default superclass instances strike me as a complicated proposal for solving trivial problems. The switch in Control.Exception (from data Exception to class Exception) was much more disrupting, adapting programs meant lots of changes everywhere exceptions are handled, not just adding some trivial instances. Still people managed the transition. Cheers -- Tony Morris http://tmorris.net/ ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime -- Tony Morris http://tmorris.net/ ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
There are very good reasons for not following this road; indeed everything which is a Monad can also be made an instance of Applicative. But more often than not we want to have a more specific implementation. Because Applicative is less general, there is in general more that you can do with it. An analogue is the relation between regular grammars and context-free grammars; indeed, once we have the latter concept we might argue that we do not need the first one any more. But if we know that something is in the first category we can do all kins of nice things which we cannot do with conxet-free grammars, such as constructing a finite state machine for recognising sentences. You proposal would introduce overlapping instances is such cases where we want to give a ``better'' implementation in case we know we are dealing with the more restricted case. I have explained this phenomenon for the first time in: @inproceedings{SwieDupo96, Author = {Swierstra, S. D. and Duponcheel, L.}, Booktitle = {Advanced Functional Programming}, Date-Added = {2009-01-04 17:21:54 +0100}, Date-Modified = {2009-01-04 17:21:54 +0100}, Editor = {Launchbury, John and Meijer, Erik and Sheard, Tim}, Pages = {184-207}, Publisher = {Springer-Verlag}, Series = {LNCS-Tutorial}, Title = {Deterministic, Error-Correcting Combinator Parsers}, Urlpdf = {http://www.cs.uu.nl/people/doaitse/Papers/1996/DetErrCorrComPars.pdf}, Volume = {1129}, Year = {1996}} If you look at the uu-parsinglib library you will see that the Applicative instance of the parsers used there is definitely more involved that what you can do with the monadic interface. Your proposal would ruin this library. Unless we have things like e.g. named instances, the possibility to choose between overlapping instances, etc. I think we should leave things the way they are; the only reason I see for having superclasses is to be able to use functions from those classes in the default implementations of functions in the new class, and to group functionality coming from several classes. Doaitse On Oct 24, 2012, at 10:01 , Petr P petr@gmail.com wrote: Hi, I was thinking lately about the well known problem that Monad is neither Functor nor Applicative. As I understand, it is caused by some historical issues. What I like about Haskell is that it allows to describe very nicely what different objects actually are - something that I find very important for programming. And this issue violates this principle. This has been discussed here more than year ago in http://www.haskell.org/pipermail/haskell-prime/2011-January/003312.html : On 1/4/11 11:24, oleg at okmij.org wrote: I'd like to argue in opposition of making Functor a super-class of Monad. I would argue that superclass constraints are not the right tool for expressing mathematical relationship such that all monads are functors and applicatives. Then argument is practical. It seems that making Functor a superclass of Monad makes defining new monad instances more of a chore, leading to code duplication. To me, code duplication is a sign that an abstraction is missing or misused. ... The main objections were that it would break existing code and that it would lead to code duplication. The former is serious, the second can be easily solved by standard Haskell, since one can define instance Applicative ... where pure = return (*) = ap instance Functor ... where fmap = liftM To address the first objection: AFAIK nobody mentioned the Default superclass instances proposal: http://hackage.haskell.org/trac/ghc/wiki/DefaultSuperclassInstances To give an example how it would work: class Applicative f = Monad f where (=) :: f a - (a - f b) - f b ... instance Applicative f where ff * fs = ff = \ f - fs = \ s - return (f s) ... This says that if somebody defines an instance of Monad it automatically becomes an instance of Applicative as defined in the nested instance block. So there is no need to define Applicative/Functor explicitly, making existing code work. Implementing this proposal would allow making Monad to extend Functor and Applicative without breaking existing code. Moreover, this would simplify other things, for example it would be possible to define an instance of Traversable and the instances for Functor and Foldable would be defined implicitly using fmapDefault and foldMapDefault. I'm sure there are many other cases where splitting type classes into a more fine-grained hierarchy would be beneficial, and the main reason against it is simply not to break compatibility with existing code. IMHO this would be worthwhile to consider for some future revision of Haskell. Best regards, Petr Pudlak ___ Haskell-prime mailing list
Re: In opposition of Functor as super-class of Monad
On 24 October 2012 11:16, S. Doaitse Swierstra doai...@swierstra.net wrote: There are very good reasons for not following this road; indeed everything which is a Monad can also be made an instance of Applicative. But more often than not we want to have a more specific implementation. Because Applicative is less general, there is in general more that you can do with it. I don't think anyone is suggesting that we force all type that are both Monad and Applicative to use (*) = ap as the implementation. As you say, that'd be crazy. The details and differences between the various superclass proposals are to do with how you provide the explicit instance vs getting the default. The wiki page explains it and links to the other similar proposals: http://hackage.haskell.org/trac/ghc/wiki/DefaultSuperclassInstances Duncan ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
First, let me make it clear that nowadays we are of course (I hope!) talking about making not only Functor, but Applicative a super-class of Monad (so Functor becomes a super-class by transitivity). Petr P wrote: The main objections were that it would break existing code and that it would lead to code duplication. The former is serious, [...] To address the first objection: I don't buy this it breaks lots of code argument. Adding the missing instances is a complete no-brainer; as you wrote: instance Applicative ... where pure = return (*) = ap instance Functor ... where fmap = liftM I do not think it is unreasonable to expect people to add such a simple and practically automatic fix to some old programs in the interest of cleaning up an old wart (and almost everyone agrees that this would be a good thing, in principle). BTW, I guess most programs already contain the Functor instances (but maybe not Applicative, as it is newer). I agree with Petr Pudlak that code duplication is not an issue, see above. And yes, these automatic instances may have stronger super-class constraints than strictly necessary. So what? The program didn't need the Functor (or Applicative) instance anyway (or it already would have defined it, in which case no change would be needed at all). Default superclass instances strike me as a complicated proposal for solving trivial problems. The switch in Control.Exception (from data Exception to class Exception) was much more disrupting, adapting programs meant lots of changes everywhere exceptions are handled, not just adding some trivial instances. Still people managed the transition. Cheers -- Ben Franksen () ascii ribbon campaign - against html e-mail /\ www.asciiribbon.org - against proprietary attachments ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
I should hope not. The identity element (return, coreturn, mempty, pure, Category.id) is almost never needed. * http://hackage.haskell.org/package/semigroupoids * https://gist.github.com/3871764 On 25/10/12 04:49, Ben Franksen wrote: First, let me make it clear that nowadays we are of course (I hope!) talking about making not only Functor, but Applicative a super-class of Monad (so Functor becomes a super-class by transitivity). Petr P wrote: The main objections were that it would break existing code and that it would lead to code duplication. The former is serious, [...] To address the first objection: I don't buy this it breaks lots of code argument. Adding the missing instances is a complete no-brainer; as you wrote: instance Applicative ... where pure = return (*) = ap instance Functor ... where fmap = liftM I do not think it is unreasonable to expect people to add such a simple and practically automatic fix to some old programs in the interest of cleaning up an old wart (and almost everyone agrees that this would be a good thing, in principle). BTW, I guess most programs already contain the Functor instances (but maybe not Applicative, as it is newer). I agree with Petr Pudlak that code duplication is not an issue, see above. And yes, these automatic instances may have stronger super-class constraints than strictly necessary. So what? The program didn't need the Functor (or Applicative) instance anyway (or it already would have defined it, in which case no change would be needed at all). Default superclass instances strike me as a complicated proposal for solving trivial problems. The switch in Control.Exception (from data Exception to class Exception) was much more disrupting, adapting programs meant lots of changes everywhere exceptions are handled, not just adding some trivial instances. Still people managed the transition. Cheers -- Tony Morris http://tmorris.net/ ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
Tony Morris wrote: I should hope not. The identity element (return, coreturn, mempty, pure, Category.id) is almost never needed. * http://hackage.haskell.org/package/semigroupoids * https://gist.github.com/3871764 Off-topic. Feel free to start a new thread named The bombastic one-and-true class hierarchy I always wanted to have. These proposals have their merits, and I greatly respect the category theoretic knowledge that went into them -- but this is another discussion. This thread refers to a rather modest correction in the standard libraries, not a complete re-design. The idea is to fix something that is widely accepted as an unfortunate ommision (in fact, Oleg's comment is one of the very few that question the idea of adding super class constraints to Monad in principle). BTW, it is unclear what your I hope not refers to, since in both of the hierarchies you linked to Applicative *is* a super class of Monad. Cheers On 25/10/12 04:49, Ben Franksen wrote: First, let me make it clear that nowadays we are of course (I hope!) talking about making not only Functor, but Applicative a super-class of Monad (so Functor becomes a super-class by transitivity). Petr P wrote: The main objections were that it would break existing code and that it would lead to code duplication. The former is serious, [...] To address the first objection: I don't buy this it breaks lots of code argument. Adding the missing instances is a complete no-brainer; as you wrote: instance Applicative ... where pure = return (*) = ap instance Functor ... where fmap = liftM I do not think it is unreasonable to expect people to add such a simple and practically automatic fix to some old programs in the interest of cleaning up an old wart (and almost everyone agrees that this would be a good thing, in principle). BTW, I guess most programs already contain the Functor instances (but maybe not Applicative, as it is newer). I agree with Petr Pudlak that code duplication is not an issue, see above. And yes, these automatic instances may have stronger super-class constraints than strictly necessary. So what? The program didn't need the Functor (or Applicative) instance anyway (or it already would have defined it, in which case no change would be needed at all). Default superclass instances strike me as a complicated proposal for solving trivial problems. The switch in Control.Exception (from data Exception to class Exception) was much more disrupting, adapting programs meant lots of changes everywhere exceptions are handled, not just adding some trivial instances. Still people managed the transition. Cheers -- Ben Franksen () ascii ribbon campaign - against html e-mail /\ www.asciiribbon.org - against proprietary attachments ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
Tony, you're missing the point... Alexey isn't making a complete patch to GHC/base libraries, just a hacky-looking demonstration. Alexey is saying that in a class hierarchy (such as if Functor = Monad were a hierarchy, or for that matter XFunctor=XMonad or Eq = Ord), it is still possible to define the superclass functions (fmap) in terms of the subclass functions (return and =) (such as writing a functor instance in which fmap f m = m = (return . f)). This has always been true in Haskell, it just might not have been obvious. ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
On 06/01/11 04:58, Isaac Dupree wrote: Tony, you're missing the point... Alexey isn't making a complete patch to GHC/base libraries, just a hacky-looking demonstration. Alexey is saying that in a class hierarchy (such as if Functor = Monad were a hierarchy, or for that matter XFunctor=XMonad or Eq = Ord), it is still possible to define the superclass functions (fmap) in terms of the subclass functions (return and =) (such as writing a functor instance in which fmap f m = m = (return . f)). This has always been true in Haskell, it just might not have been obvious. ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime Oh right sorry. I thought a stronger point was being made. Then perhaps it's also worth pointing out that (*) can be written using (=) and return: f * a = f = \ff - a = \aa - return (ff aa) -- Tony Morris http://tmorris.net/ ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
Hi, indeed, this is called ap in Control.Monad. So if we have an instance of Monad, all that needs to be done to support the other instances is: instance (SameContextAsTheMonadInstance) = Functor MyType where fmap = liftM instance (SameContextAsTheMonadInstance) = Applicative MyType where pure = return; (*) = ap Furthermore, this is only in the cases where we are defining the type from scratch, and not using a library like monadLib or MTL, otherwise a simple deriving is sufficient. -Iavor On Wed, Jan 5, 2011 at 12:29 PM, Tony Morris tonymor...@gmail.com wrote: On 06/01/11 04:58, Isaac Dupree wrote: Tony, you're missing the point... Alexey isn't making a complete patch to GHC/base libraries, just a hacky-looking demonstration. Alexey is saying that in a class hierarchy (such as if Functor = Monad were a hierarchy, or for that matter XFunctor=XMonad or Eq = Ord), it is still possible to define the superclass functions (fmap) in terms of the subclass functions (return and =) (such as writing a functor instance in which fmap f m = m = (return . f)). This has always been true in Haskell, it just might not have been obvious. ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime Oh right sorry. I thought a stronger point was being made. Then perhaps it's also worth pointing out that (*) can be written using (=) and return: f * a = f = \ff - a = \aa - return (ff aa) -- Tony Morris http://tmorris.net/ ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
In opposition of Functor as super-class of Monad
I'd like to argue in opposition of making Functor a super-class of Monad. I would argue that superclass constraints are not the right tool for expressing mathematical relationship such that all monads are functors and applicatives. Then argument is practical. It seems that making Functor a superclass of Monad makes defining new monad instances more of a chore, leading to code duplication. To me, code duplication is a sign that an abstraction is missing or misused. For the sake of the argument, let us suppose that Functor is a superclass of Monad. Let us see how to define a new Monad instance. For the sake of a better illustration, I'll use a complex monad. I just happen to have an example of that: Iteratee. The data type Iteratee is defined as follows: type ErrMsg = String-- simplifying data Stream el = EOF (Maybe ErrMsg) | Chunk [el] deriving Show data Iteratee el m a = IE_done a | IE_cont (Maybe ErrMsg) (Stream el - m (Iteratee el m a, Stream el)) We wish to define an instance for Monad (Iteratee el m). Since Functor is a superclass of Monad, we must define a functor instance: instance Functor m = Functor (Iteratee el m) where fmap f (IE_done a) = IE_done (f a) fmap f (IE_cont e k) = IE_cont e (\s - fmap docase (k s)) where docase (IE_done a,s) = (IE_done (f a), s) docase (i, s)= (fmap f i, s) There are two ways to proceed with the Monad instance -- to be precise, there are two ways of defining bind. Method A: just define bind as usual instance (Functor (Iteratee el m),Monad m) = Monad (Iteratee el m) where return = IE_done IE_done a = f = f a IE_cont e k = f = IE_cont e (\s - k s = docase) where docase (IE_done a, stream) = case f a of IE_cont Nothing k - k stream i - return (i,stream) docase (i, s) = return (i = f, s) Although we must state the constraint (Functor (Iteratee el m)) to satisfy the super-class constraint, we have not made any use of the constraint. We defined bind without resorting to fmap. That seems like a waste. What makes it seem more like a waste is that the code for fmap and for bind is almost the same. We had to repeat essentially the same algorithm, analysing Iteratee and the continuation. Method B: define bind in terms of fmap Alas, just fmap is not sufficient to define bind. We need join: joinIter :: Monad m = Iteratee el m (Iteratee el m a) - Iteratee el m a joinIter (IE_done i) = i joinIter (IE_cont e k) = IE_cont e (\s - k s = docase) where docase (IE_done (IE_cont Nothing k), s) = k s docase (IE_done i, s) = return (i, s) docase (i, s) = return (joinIter i, s) Only after defining join we can write bind m f = joinIter $ fmap f m Again we see code duplication: the code for join resembles the code for fmap. The code for join follows the same pattern of analysing Iteratee and the continuation. In either way, Functor as a super-class of Monad leads to code duplication. That gives a bad feeling practically -- and theoretically. The experiment has led me wonder if a superclass constraint is the right way to state the relationship between Monads and Functors. It _almost_ makes me wish the constraint go the other way: instance Monad m = Functor m where fmap f m = m = (return . f) That is, we need an instance rather than a superclass constraint, and in the other direction. The instance constraint says that every monad is a functor. Moreover, \f m = m = (return . f) is a _proof term_ that every monad is a functor. We can state it once and for all, for all present and future monads. Alas, the instance ``instance Monad m = Functor m'' above has several drawbacks (for one, requiring overlapping instances everywhere). This makes me wonder if something is amiss. In the meanwhile, there is a practical work-around. Introduce a TemplateHaskell operation generating an instance such as instance Monad (Iteratee el m) = Functor (Iteratee el m) where fmap f m = m = (return . f) (the code for the method remains the same; only the type in the instance head varies). Alas, that requires undecidable instances. All the code before was Haskell98. ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
Method C: Define fmap in terms of bind instance Monad m = Functor (Iteratee el m) where fmap = liftM Now you need to do the inspection of Iteratee only once: in the definition of the bind. However, to use liftM as implementation of fmap the superclass constraint of the Functor instance has changed from Functor to Monad. Is this a problem? If so, method A seems the way to go: you could argue that `Functor m = fmap :: (a - b) - Iteratee el m a - Iteratee el m b' is more general than the `Monad m =' version (works for more `m's) and therefore deserves to redo the analysis of Iteratee. Martijn. On 1/4/11 11:24, o...@okmij.org wrote: I'd like to argue in opposition of making Functor a super-class of Monad. I would argue that superclass constraints are not the right tool for expressing mathematical relationship such that all monads are functors and applicatives. ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
On Tuesday 04 January 2011 5:24:21 am o...@okmij.org wrote: Method A: just define bind as usual instance (Functor (Iteratee el m),Monad m) = Monad (Iteratee el m) where return = IE_done IE_done a = f = f a IE_cont e k = f = IE_cont e (\s - k s = docase) where docase (IE_done a, stream) = case f a of IE_cont Nothing k - k stream i - return (i,stream) docase (i, s) = return (i = f, s) Although we must state the constraint (Functor (Iteratee el m)) to satisfy the super-class constraint, we have not made any use of the constraint. This, at least, is false. If Functor is a superclass of Monad, then Monad m implies Functor m, which implies Functor (Iteratee el m). So Monad m is a sufficient constraint for the instance. As for the other concerns, I think the closest fix I've seen is to allow subclasses to specify defaults for superclasses, and allow instances for subclasses to include methods for superclasses. So: class Functor m = Monad m where return :: a - m a (=) :: m a - (a - m b) - m b fmap f x = x = return . f This has its own caveats of course. And in this case, it seems to overconstrain the functor instance, since presumably we'd end up with: instance Monad m = Monad (Iteratee el m) where ... == instance Monad m = Functor (Iterate el m) where ... I'm not sure what to do about that. -- Dan ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
On 04.01.2011 13:24, o...@okmij.org wrote: I'd like to argue in opposition of making Functor a super-class of Monad. I would argue that superclass constraints are not the right tool for expressing mathematical relationship such that all monads are functors and applicatives. Then argument is practical. It seems that making Functor a superclass of Monad makes defining new monad instances more of a chore, leading to code duplication. To me, code duplication is a sign that an abstraction is missing or misused. I think I understood your point. But it looks like that it's possible to use subclass's function in superclass instance. At very least GHC is able to do it. Following example works just fine without any language extensions in GHC6.12.3 import Prelude hiding (Monad(..), Functor(..)) class Functor f where fmap :: (a - b) - f a - f b class Functor m = Monad m where return :: a - m a (=) :: m a - (a - m b) - m b instance Functor Maybe where fmap f m = m = (return . f) instance Monad Maybe where return = Just Nothing = _ = Nothing Just x = f = f x ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
I think you'll find a problem using do-notation with your Monad. Tony Morris On 04/01/2011 11:33 PM, Alexey Khudyakov alexey.sklad...@gmail.com wrote: On 04.01.2011 13:24, o...@okmij.org wrote: I'd like to argue in opposition of making Functor a... I think I understood your point. But it looks like that it's possible to use subclass's function in superclass instance. At very least GHC is able to do it. Following example works just fine without any language extensions in GHC6.12.3 import Prelude hiding (Monad(..), Functor(..)) class Functor f where fmap :: (a - b) - f a - f b class Functor m = Monad m where return :: a - m a (=) :: m a - (a - m b) - m b instance Functor Maybe where fmap f m = m = (return . f) instance Monad Maybe where return = Just Nothing = _ = Nothing Just x = f = f x ___ Haskell-prime mailing list haskell-pr...@haskell.o... ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime
Re: In opposition of Functor as super-class of Monad
On Tue, Jan 04, 2011 at 02:24:21AM -0800, o...@okmij.org wrote: I'd like to argue in opposition of making Functor a super-class of Monad. I would argue that superclass constraints are not the right tool for expressing mathematical relationship such that all monads are functors and applicatives. Then argument is practical. It seems that making Functor a superclass of Monad makes defining new monad instances more of a chore, leading to code duplication. To me, code duplication is a sign that an abstraction is missing or misused. The argument about code duplication somehow seems to assume that class member instances need to be defined as part of the instance declaration. This is not the case, and in fact I am arguing in general against putting any interesting code into instance declarations, especially into declarations of instances with constraints (since, in ML terminology, they are functors, and putting their definition inside an instance declaration constrains their applicability). In my opinion, the better approach is to define (generalised versions of) the functions mentioned in the class interface, and then just throw together the instances from those functions. This also makes it easier to adapt to the ``class hierarchy du jour''. The point for the situation here is that although we eventually need definitions of all the functions declared as class members, there is absolutely nothing that constrains the dependency relation between the definitions of these functions to be conforming in any way to the class hierarchy. For a simpler example, assume that I have some arbitrary data type data T a = One a | Two a a and assume that I am interested only in Ord instances, since I want to use T with Data.Set, and I am not really interested in Eq instances. Assume that the order will depend on that for |a|, so I will define a function: compareT :: (a - a - Ordering) - T a - T a - Ordering Then I can thow together the necessary instances from that: instance Ord a = Ord (T a) where compare = compareT compare instance Ord a = Eq (T a) where (==) = eqFromCompare compare assuming I have (preferably from the exporter of Eq and Ord): eqFromCompare :: (a - a - Ordering) - (a - a - Bool) eqFromCompare cmp x y = case cmp x y of EQ - True _ - False The same approach works for Oleg's example: For the sake of the argument, let us suppose that Functor is a superclass of Monad. Let us see how to define a new Monad instance. For the sake of a better illustration, I'll use a complex monad. I just happen to have an example of that: Iteratee. The data type Iteratee is defined as follows: type ErrMsg = String-- simplifying data Stream el = EOF (Maybe ErrMsg) | Chunk [el] deriving Show data Iteratee el m a = IE_done a | IE_cont (Maybe ErrMsg) (Stream el - m (Iteratee el m a, Stream el)) [...] It _almost_ makes me wish the constraint go the other way: instance Monad m = Functor m where fmap f m = m = (return . f) That is, we need an instance rather than a superclass constraint, and in the other direction. The instance constraint says that every monad is a functor. Moreover, \f m = m = (return . f) is a _proof term_ that every monad is a functor. We can state it once and for all, for all present and future monads. I would expect that proof term to exported by the package exporting Functor and Monad; let us define it here: fmapFromBind (=) f m = m = (return . f) Now you can write, no matter which class is a superclass of which: bindIt return (=) (IE_done a) f = f a bindIt return (=) (IE_cont e k) f = IE_cont e (\s - k s = docase) where docase (IE_done a, stream) = case f a of IE_cont Nothing k - k stream i - return (i,stream) docase (i, s) = return (bindIt return (=) i f, s) instance Monad m = Monad (Iteratee el m) where return = IE_done (=) = bindIt return (=) instance Monad m = Functor (Iteratee el m) where fmap = fmapFromBind (=) Of course this assumes that you are not actually interested in an instance of shape: instance (Functor ...) = Functor (Iteratee el m), but this seems to be a plausible assumption. Defining the functionality really has nothing to do with declaring an instance of a type class, and it is normally better to keep the two separated. And that does not lead to any real code duplication, only extremely boring instance declarations. Wolfram ___ Haskell-prime mailing list Haskell-prime@haskell.org http://www.haskell.org/mailman/listinfo/haskell-prime