{-# LANGUAGE CPP, DeriveFunctor #-} {- | Sparse matrices. We assume the matrices to be very sparse, so we just implement them as sorted association lists. -} module Agda.Termination.SparseMatrix ( -- * Basic data types Matrix , matrixInvariant , Size(..) , sizeInvariant , MIx (..) , mIxInvariant -- * Generating and creating matrices , fromLists , fromIndexList , toLists -- , Agda.Termination.Matrix.zipWith , matrix , matrixUsingRowGen -- * Combining and querying matrices , size , square , isEmpty , isSingleton , add, intersectWith , mul , transpose , diagonal -- * Modifying matrices , addRow , addColumn -- * Tests , Agda.Termination.SparseMatrix.tests ) where import Data.Array import qualified Data.List as List import Data.Maybe import Data.Monoid import Agda.Utils.Pretty hiding (isEmpty) import Agda.Utils.QuickCheck import Agda.Utils.TestHelpers import Agda.Termination.Semiring (HasZero(..), SemiRing, Semiring) import qualified Agda.Termination.Semiring as Semiring #include "../undefined.h" import Agda.Utils.Impossible ------------------------------------------------------------------------ -- Basic data types -- | This matrix type is used for tests. type TM = Matrix Integer Integer -- | Size of a matrix. data Size i = Size { rows :: i, cols :: i } deriving (Eq, Ord, Show) sizeInvariant :: (Ord i, Num i) => Size i -> Bool sizeInvariant sz = rows sz >= 0 && cols sz >= 0 instance (Arbitrary i, Integral i) => Arbitrary (Size i) where arbitrary = do r <- natural c <- natural return $ Size { rows = fromInteger r, cols = fromInteger c } instance CoArbitrary i => CoArbitrary (Size i) where coarbitrary (Size rs cs) = coarbitrary rs . coarbitrary cs prop_Arbitrary_Size :: Size Integer -> Bool prop_Arbitrary_Size = sizeInvariant -- | Converts a size to a set of bounds suitable for use with -- the matrices in this module. toBounds :: Num i => Size i -> (MIx i, MIx i) toBounds sz = (MIx { row = 1, col = 1 }, MIx { row = rows sz, col = cols sz }) -- | Type of matrix indices (row, column). data MIx i = MIx { row, col :: i } deriving (Eq, Show, Ix, Ord) instance (Arbitrary i, Integral i) => Arbitrary (MIx i) where arbitrary = do r <- positive c <- positive return $ MIx { row = r, col = c } instance CoArbitrary i => CoArbitrary (MIx i) where coarbitrary (MIx r c) = coarbitrary r . coarbitrary c -- | No nonpositive indices are allowed. mIxInvariant :: (Ord i, Num i) => MIx i -> Bool mIxInvariant i = row i >= 1 && col i >= 1 prop_Arbitrary_MIx :: MIx Integer -> Bool prop_Arbitrary_MIx = mIxInvariant -- | Type of matrices, parameterised on the type of values. data Matrix i b = M { size :: Size i, unM :: [(MIx i, b)] } deriving (Eq, Ord, Functor) matrixInvariant :: (Num i, Ix i) => Matrix i b -> Bool matrixInvariant m = all (\ (MIx i j, b) -> 1 <= i && i <= rows sz && 1 <= j && j <= cols sz) (unM m) && strictlySorted (MIx 0 0) (unM m) && sizeInvariant sz where sz = size m -- matrix indices are lexicographically sorted with no duplicates -- Ord MIx should be the lexicographic one already (Haskell report) strictlySorted :: (Ord i) => i -> [(i, b)] -> Bool strictlySorted i [] = True strictlySorted i ((i', b) : l) = i < i' && strictlySorted i' l {- strictlySorted (MIx i j) [] = True strictlySorted (MIx i j) ((MIx i' j', b) : l) = (i < i' || i == i' && j < j' ) && strictlySorted (MIx i' j') b -} instance (Ord i, Integral i, Enum i, Ix i, Show i, Show b, HasZero b) => Show (Matrix i b) where showsPrec _ m = showString "Agda.Termination.Matrix.fromLists " . shows (size m) . showString " " . shows (toLists m) instance (Show i, Integral i, Ix i, HasZero b, Pretty b) => Pretty (Matrix i b) where pretty = vcat . map (hsep . map pretty) . toLists instance (Arbitrary i, Num i, Integral i, Arbitrary b, HasZero b) => Arbitrary (Matrix i b) where arbitrary = matrix =<< arbitrary instance (Show i, Ord i, Integral i, Enum i, Ix i, CoArbitrary b, HasZero b) => CoArbitrary (Matrix i b) where coarbitrary m = coarbitrary (toLists m) prop_Arbitrary_Matrix :: TM -> Bool prop_Arbitrary_Matrix = matrixInvariant ------------------------------------------------------------------------ -- Generating and creating matrices -- | Generates a matrix of the given size, using the given generator -- to generate the rows. matrixUsingRowGen :: (Arbitrary i, Integral i, Arbitrary b, HasZero b) => Size i -> (i -> Gen [b]) -- ^ The generator is parameterised on the size of the row. -> Gen (Matrix i b) matrixUsingRowGen sz rowGen = do rows <- vectorOf (fromIntegral $ rows sz) (rowGen $ cols sz) return $ fromLists sz rows -- | Generates a matrix of the given size. matrix :: (Arbitrary i, Integral i, Arbitrary b, HasZero b) => Size i -> Gen (Matrix i b) matrix sz = matrixUsingRowGen sz (\n -> vectorOf (fromIntegral n) arbitrary) prop_matrix sz = forAll (matrix sz :: Gen TM) $ \m -> -- matrixInvariant m && size m == sz -- | Constructs a matrix from a list of (index, value)-pairs. -- compareElt = (\ (i,_) (j,_) -> compare i j) -- normalize = filter (\ (i,b) -> b /= zeroElement) fromIndexList :: (Ord i, HasZero b) => Size i -> [(MIx i, b)] -> Matrix i b fromIndexList sz = M sz . List.sortBy (\ (i,_) (j,_) -> compare i j) . filter (\ (i,b) -> b /= zeroElement) prop_fromIndexList :: TM -> Bool prop_fromIndexList m = matrixInvariant m' && m' == m where vs = unM m m' = fromIndexList (size m) vs -- | @'fromLists' sz rs@ constructs a matrix from a list of lists of -- values (a list of rows). -- -- Precondition: @'length' rs '==' 'rows' sz '&&' 'all' (('==' 'cols' sz) . 'length') rs@. fromLists :: (Ord i, Num i, Enum i, HasZero b) => Size i -> [[b]] -> Matrix i b fromLists sz bs = fromIndexList sz $ zip ([ MIx i j | i <- [1..rows sz] , j <- [1..cols sz]]) (concat bs) -- | Converts a sparse matrix to a sparse list of rows toSparseRows :: (Eq i, Num i, Enum i) => Matrix i b -> [(i,[(i,b)])] toSparseRows m = aux 1 [] (unM m) where aux i' [] [] = [] aux i' row [] = [(i', reverse row)] aux i' row ((MIx i j, b) : m) | i' == i = aux i' ((j,b):row) m | otherwise = (i', reverse row) : aux i [(j,b)] m -- sparse vectors cannot have two entries in one column blowUpSparseVec :: (Show i, Ord i, Num i, Enum i) => b -> i -> [(i,b)] -> [b] blowUpSparseVec zero n l = aux 1 l where aux i [] | i > n = [] | otherwise = zero : aux (i+1) [] aux i ((j,b):l) | i <= n && j == i = b : aux (succ i) l aux i ((j,b):l) | i <= n && j >= i = zero : aux (succ i) ((j,b):l) aux i l = __IMPOSSIBLE__ -- error $ "blowUpSparseVec (n = " ++ show n ++ ") aux i=" ++ show i ++ " j=" ++ show (fst (head l)) ++ " length l = " ++ show (length l) -- | Converts a matrix to a list of row lists. toLists :: (Show i, Ord i, Integral i, Enum i, Ix i, HasZero b) => Matrix i b -> [[b]] toLists m = -- if not $ matrixInvariant m then __IMPOSSIBLE__ else blowUpSparseVec emptyRow nr $ map (\ (i,r) -> (i, blowUpSparseVec zeroElement nc r)) $ toSparseRows m where sz = size m nr = rows sz nc = cols sz emptyRow = take (fromIntegral nc) $ repeat zeroElement prop_fromLists_toLists :: TM -> Bool prop_fromLists_toLists m = fromLists (size m) (toLists m) == m ------------------------------------------------------------------------ -- Combining and querying matrices -- | The size of a matrix. {- size :: Ix i => Matrix i b -> Size i size m = Size { rows = row b, cols = col b } where (_, b) = bounds $ unM m -} prop_size :: TM -> Bool prop_size m = sizeInvariant (size m) prop_size_fromIndexList :: Size Int -> Bool prop_size_fromIndexList sz = size (fromIndexList sz ([] :: [(MIx Int, Integer)])) == sz -- | 'True' iff the matrix is square. square :: Ix i => Matrix i b -> Bool square m = rows (size m) == cols (size m) -- | Returns 'True' iff the matrix is empty. isEmpty :: (Num i, Ix i) => Matrix i b -> Bool isEmpty m = rows sz <= 0 || cols sz <= 0 where sz = size m -- | Returns 'Just b' iff it is a 1x1 matrix with just one entry 'b'. isSingleton :: (Num i, Ix i) => Matrix i b -> Maybe b isSingleton m = if (rows sz == 1 || cols sz == 1) then case unM m of [(_,b)] -> Just b _ -> __IMPOSSIBLE__ else Nothing where sz = size m -- | Transposition transposeSize (Size { rows = n, cols = m }) = Size { rows = m, cols = n } transpose m = M { size = transposeSize (size m) , unM = List.sortBy (\ (i,a) (j,b) -> compare i j) $ map (\(MIx i j, b) -> (MIx j i, b)) $ unM m } prop_transpose :: TM -> Bool prop_transpose m = matrixInvariant m' && m == transpose m' where m' = transpose m -- | @'add' (+) m1 m2@ adds @m1@ and @m2@. Uses @(+)@ to add values. -- -- No longer precondition: @'size' m1 == 'size' m2@. add :: (Ord i) => (a -> a -> a) -> Matrix i a -> Matrix i a -> Matrix i a add plus m1 m2 = M (supSize m1 m2) $ mergeAssocWith plus (unM m1) (unM m2) -- | Compute the matrix size of the union of two matrices. supSize m1 m2 = Size { rows = r, cols = c } where sz1 = size m1 sz2 = size m2 r = max (rows sz1) (rows sz2) c = max (cols sz1) (cols sz2) -- | assoc list union mergeAssocWith :: (Ord i) => (a -> a -> a) -> [(i,a)] -> [(i,a)] -> [(i,a)] mergeAssocWith f [] m = m mergeAssocWith f l [] = l mergeAssocWith f l@((i,a):l') m@((j,b):m') | i < j = (i,a) : mergeAssocWith f l' m | i > j = (j,b) : mergeAssocWith f l m' | otherwise = (i, f a b) : mergeAssocWith f l' m' -- | @'intersectWith' f m1 m2@ build the pointwise conjunction @m1@ and @m2@. -- Uses @f@ to combine non-zero values. -- -- No longer precondition: @'size' m1 == 'size' m2@. intersectWith :: (Ord i) => (a -> a -> a) -> Matrix i a -> Matrix i a -> Matrix i a intersectWith f m1 m2 = M (infSize m1 m2) $ interAssocWith f (unM m1) (unM m2) -- | Compute the matrix size of the intersection of two matrices. infSize m1 m2 = Size { rows = r, cols = c } where sz1 = size m1 sz2 = size m2 r = min (rows sz1) (rows sz2) c = min (cols sz1) (cols sz2) -- | assoc list intersection interAssocWith :: (Ord i) => (a -> a -> a) -> [(i,a)] -> [(i,a)] -> [(i,a)] interAssocWith f [] m = [] interAssocWith f l [] = [] interAssocWith f l@((i,a):l') m@((j,b):m') | i < j = interAssocWith f l' m | i > j = interAssocWith f l m' | otherwise = (i, f a b) : interAssocWith f l' m' prop_add sz = forAll (three (matrix sz :: Gen TM)) $ \(m1, m2, m3) -> let m' = add (+) m1 m2 in associative (add (+)) m1 m2 m3 && commutative (add (+)) m1 m2 && matrixInvariant m' && size m' == size m1 -- | @'mul' semiring m1 m2@ multiplies @m1@ and @m2@. Uses the -- operations of the semiring @semiring@ to perform the -- multiplication. -- -- Precondition: @'cols' ('size' m1) == rows ('size' m2)@. {- mul A B works as follows: * turn A into a list of sparse rows and the transposed B as well * form the crossproduct using the inner vector product to compute els * the inner vector product is summing up after intersecting with the muliplication op of the semiring -} mul :: (Enum i, Num i, Ix i, Eq a) => Semiring a -> Matrix i a -> Matrix i a -> Matrix i a mul semiring m1 m2 = M (Size { rows = rows (size m1), cols = cols (size m2) }) $ filter (\ (i,b) -> b /= Semiring.zero semiring) $ [ (MIx i j, List.foldl' (Semiring.add semiring) (Semiring.zero semiring) $ map snd $ interAssocWith (Semiring.mul semiring) v w) | (i,v) <- toSparseRows m1 , (j,w) <- toSparseRows $ transpose m2 ] prop_mul sz = sized $ \n -> resize (n `div` 2) $ forAll (two natural) $ \(c2, c3) -> forAll (matrix sz :: Gen TM) $ \m1 -> forAll (matrix (Size { rows = cols sz, cols = c2 })) $ \m2 -> forAll (matrix (Size { rows = c2, cols = c3 })) $ \m3 -> let m' = mult m1 m2 in associative mult m1 m2 m3 && matrixInvariant m' && size m' == Size { rows = rows sz, cols = c2 } where mult = mul Semiring.integerSemiring -- | @'diagonal' m@ extracts the diagonal of @m@. -- -- No longer precondition: @'square' m@. diagonal :: (Show i, Enum i, Num i, Ix i, HasZero b) => Matrix i b -> Array i b diagonal m = -- if r /= c then __IMPOSSIBLE__ else -- works also for non-square listArray (1, n) $ blowUpSparseVec zeroElement n $ mapMaybe (\ ((MIx i j),b) -> if i==j then Just (i,b) else Nothing) $ unM m -- map (\ ((MIx i j),b) -> (i,b)) $ filter (\ ((MIx i j),b) -> i==j) (unM m) where sz = size m r = rows sz c = cols sz n = max r c prop_diagonal = forAll natural $ \n -> forAll (matrix (Size n n) :: Gen TM) $ \m -> bounds (diagonal m) == (1, n) ------------------------------------------------------------------------ -- Modifying matrices -- | @'addColumn' x m@ adds a new column to @m@, after the columns -- already existing in the matrix. All elements in the new column get -- set to @x@. addColumn :: (Num i, HasZero b) => b -> Matrix i b -> Matrix i b addColumn x m | x == zeroElement = m { size = (size m) { cols = cols (size m) + 1 }} | otherwise = __IMPOSSIBLE__ prop_addColumn :: TM -> Bool prop_addColumn m = matrixInvariant m' && map init (toLists m') == toLists m where m' = addColumn zeroElement m -- | @'addRow' x m@ adds a new row to @m@, after the rows already -- existing in the matrix. All elements in the new row get set to @x@. addRow :: (Num i, HasZero b) => b -> Matrix i b -> Matrix i b addRow x m | x == zeroElement = m { size = (size m) { rows = rows (size m) + 1 }} | otherwise = __IMPOSSIBLE__ prop_addRow :: TM -> Bool prop_addRow m = matrixInvariant m' && init (toLists m') == toLists m where m' = addRow zeroElement m ------------------------------------------------------------------------ -- Zipping (assumes non-empty matrices) {- use mergeAssocList or interAssocList instead zipWith :: (a -> b -> c) -> Matrix Integer a -> Matrix Integer b -> Matrix Integer c zipWith f m1 m2 = fromLists (Size { rows = toInteger $ length ll, cols = toInteger $ length (head ll) }) ll where ll = List.zipWith (List.zipWith f) (toLists m1) (toLists m2) -} ------------------------------------------------------------------------ -- All tests tests :: IO Bool tests = runTests "Agda.Termination.Matrix" [ quickCheck' prop_transpose , quickCheck' prop_Arbitrary_Size , quickCheck' prop_Arbitrary_Matrix , quickCheck' prop_Arbitrary_MIx , quickCheck' prop_fromIndexList , quickCheck' prop_matrix , quickCheck' prop_size , quickCheck' prop_size_fromIndexList , quickCheck' prop_fromLists_toLists , quickCheck' prop_add , quickCheck' prop_mul , quickCheck' prop_diagonal , quickCheck' prop_addColumn , quickCheck' prop_addRow ]