Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

Can't find a proper signature for a function using STUArray (neither can GHC)

I built a function for finding the determinant of a matrix using the ST-Monad and unboxed STArrays (STUArray). The type for a matrix is the following:

newtype Matrix e = Matrix (Array Int (UArray Int e))

that is, an immutable array containing immutable unboxed arrays containing the elements. This will require me to add the Predicate IArray UArray e to functions dealing with Matrix, which in turn requires FlexibleContexts. Okay, done.

The function used to calculate the determinant has the following signature:

detST :: (IArray UArray e, MArray (STUArray s) e (ST s),
          Num e, Eq e, Division e)
   => Array Int (UArray Int e) -> ST s e

I am required to also add the Predicate MArray (STUArray s) e (ST s) since internally the arrays are converted into mutable arrays (the outer boxed, the inner unboxed).

This function can be used like so:

main = do
    let m@(Matrix x) = matrix [ [1,-2,3,234]
                              , [5,2,3,-3]
                              , [7,18,3,40]
                              , [2,9,71,0] ]
        d = runST (detST x) :: Int -- needed for type check, ambiguous otherwise

print d

Works, fine. But look how ugly it is! Of course I do not want to give away the internals of of Matrix (at least not further than the predicates attached to my functions already make me to). I would like to define the following function:

det :: Matrix e -> e

And I can't.

I tried without a proper signature:

det (Matrix arr) = runST (detST arr)

fails. Fair enough, I'll put my brain to work: detST requires IArray UArray e, MArray (STUArray s) e (ST s), Num e, Eq e, Division e, so does det, doesn't it?

det :: (IArray UArray e, MArray (STUArray s) e (ST s),
          Num e, Eq e, Division e) => Matrix e -> e

fails. But I don't see how. The message that GHC (7.4.2) gives me is:

Could not deduce (MArray (STUArray s) t (ST s))
  arising from a use of `detST'

but that exact term is in the predicates!

GHC suggests:

add (MArray (STUArray s) t (ST s)) to the context of
  a type expected by the context: ST s t
  or the inferred type of
     det :: (Eq t, Num t, IArray UArray t, Division t) => Matrix t -> t
or add an instance declaration for (MArray (STUArray s) t (ST s))

Okay. As to my understanding I have done that first thing. Also there does exist an instance for that (MArray ...) (otherwise, how could I have used it successfully in main?!).

I am not sure about what is wrong here. I believe it has something to do with the "hidden" ST state in s, and that the s of detST is some other s than the s in det would be, but I don't know how to write a type for this.

What is the proper definition of det - and why?!

The program without det compiles fine by using only FlexibleContexts, no warnings with -Wall. The complete source code can be found as a gist here.

like image 273
scravy Avatar asked Oct 05 '22 12:10

scravy


1 Answers

I managed to get this to work using the trick described by Keegan McAllister in this article:

{-# LANGUAGE FlexibleContexts, ScopedTypeVariables, RankNTypes, GADTs #-}

data Evidence s e where
  Evidence :: (MArray (STUArray s) e (ST s)) => Evidence s e

data ElemType e = ElemType (forall s. Evidence s e)

det :: forall e . (IArray UArray e, Num e, Eq e, Division e)
       => ElemType e -> Matrix e -> e
det (ElemType e) mat = runST (f e mat)
  where
    f :: Evidence s e -> Matrix e -> ST s e
    f Evidence (Matrix arr) = detST arr

Usage:

main :: IO ()
main = do
    let m = matrix [ [1,-2,3,234]
                   , [5,2,3,-3]
                   , [7,18,3,40]
                   , [2,9,71,0] ]
    print $ det (ElemType Evidence) (m :: Matrix Int)

The problem stems from the lack of higher-rank constraints - runST has type (forall s. ST s a) -> a, so you need a constraint like forall s . MArray (STUArray s) e (ST s), which are not supported by GHC. This trick allows you to convince the type checker that the constraint actually holds. A more in-depth discussion of this problem is available in the article I linked above.

like image 159
Mikhail Glushenkov Avatar answered Oct 12 '22 19:10

Mikhail Glushenkov