On Tue, Apr 19, 2011 at 11:48 PM, Tyson Whitehead <twhitehead@gmail.com> wrote:
On April 19, 2011 23:22:12 Tyson Whitehead wrote:
> ArrowLoop from MonadFix
>
>   loop' f = fst' .' loop'' (f .' arr' (second snd))
>     where loop'' f = mfix (\y -> f .' arr' (,y))

BTW haskellers, I've been wondering if mfix would better be defined as

 mfix' :: (m a -> m a) -> m a
 
where "mfix' f = mfix (f . pure)" for the computational monads.  The advantage
being you can give a useful definition for structural monads as well.

Note: This does not generalize the signature of mfix, it only overlaps slightly, as not every monad m permits the extraction of the value a injected (consider Cont r), so you necessarily change the meaning or obliterate a number of instances.  

Recall the main motivation for mfix was to support Erkoek and Launchbury's recursive do:

http://www.google.com/search?sourceid=chrome&ie=UTF-8&q=mfix+recursive+do
http://www.haskell.org/haskellwiki/MonadFix

This necessitates 4 laws for mfix, which don't translate nicely.
  • mfix (return . h) = return (fix h)
  • mfix (\x -> a >>= \y -> f x y) = a >>= \y -> mfix (\x -> f x y)
  • if 
    h
     is strict, 
    mfix (liftM h . f) = liftM h (mfix (. h))
  • mfix (\x -> mfix (\y -> f x y)) = mfix (\x -> f x x)
The other commonly proposed mfix replacement is to define it once, as guided by the types, but while this works for fix and the the comonadic equivalent, it doesn't generate a useful mfix for recursive do either.

-Edward