\documentclass{article} %include polycode.fmt \usepackage{fontspec} \usepackage{amsmath} \usepackage{unicode-math} \usepackage{lualatex-math} \setmainfont{latinmodern-math.otf} \setmathfont{latinmodern-math.otf} \usepackage{verbatim} \author{Sophie Taylor} \title{haskell-clifford: A Haskell Clifford algebra dynamics library} \begin{document} So yeah. This is a Clifford number representation. I will fill out the documentation more fully and stuff once the design has stabilised. I am basing the design of this on Issac Trotts' geometric algebra library.\cite{hga} Let us begin. We are going to use the Numeric Prelude because it is (shockingly) nicer for numeric stuff. \begin{code} {-# LANGUAGE NoImplicitPrelude, FlexibleContexts, RankNTypes, ScopedTypeVariables, DeriveDataTypeable #-} {-# LANGUAGE NoMonomorphismRestriction, UnicodeSyntax, GADTs#-} {-# LANGUAGE FlexibleInstances, StandaloneDeriving, KindSignatures, DataKinds #-} {-# LANGUAGE TemplateHaskell #-} {-# LANGUAGE MultiParamTypeClasses #-} \end{code} %if False \begin{code} {-# OPTIONS_GHC -fllvm -fexcess-precision -optlo-O3 -O3 -optlc-O=3 -Wall #-} -- OPTIONS_GHC -Odph -fvectorise -package dph-lifted-vseg -- LANGUAGE ParallelArrays \end{code} %endif Clifford algebras are a module over a ring. They also support all the usual transcendental functions. \begin{code} module Numeric.Clifford.Multivector where import Numeric.Clifford.Blade import NumericPrelude hiding (iterate, head, map, tail, reverse, scanl, zipWith, drop, (++), filter, null, length, foldr, foldl1, zip, foldl, concat, (!!), concatMap,any, repeat, replicate, elem, replicate, all) --import Algebra.Laws import Algebra.Absolute import Algebra.Algebraic import Algebra.Additive import Algebra.Ring import Algebra.OccasionallyScalar import Algebra.ToInteger import Algebra.Transcendental import Algebra.ZeroTestable import Algebra.Module import Algebra.Field import Data.Serialize import MathObj.Polynomial.Core (progression) import System.IO import Data.List.Stream import Data.Permute (sort, isEven) import Data.List.Ordered import Data.Ord import Data.Maybe --import Number.NonNegative import Numeric.Natural import qualified Data.Vector as V import NumericPrelude.Numeric (sum) import qualified NumericPrelude.Numeric as NPN import Test.QuickCheck import Math.Sequence.Converge (convergeBy) import Control.DeepSeq import Number.Ratio hiding (scale) import Algebra.ToRational import qualified GHC.Num as PNum import Control.Lens hiding (indices) import Control.Exception (assert) import Data.Maybe import Data.Monoid import Data.Data import Data.DeriveTH import GHC.TypeLits import Data.Word import Debug.Trace --trace _ a = a \end{code} A multivector is nothing but a linear combination of basis blades. \begin{code} data Multivector (n::Nat) f where BladeSum :: forall n f . (SingI n, Algebra.Field.C f, Ord f) => { _terms :: [Blade n f]} -> Multivector n f deriving instance Eq (Multivector n f) deriving instance Ord (Multivector n f) deriving instance (Show f) => Show (Multivector n f) dimension :: forall (n::Nat) f. SingI n => Multivector n f -> Natural dimension _ = toNatural ((fromIntegral $ fromSing (sing :: Sing n))::Word) terms :: Lens' (Multivector n f) [Blade n f] terms = lens _terms (\bladeSum v -> bladeSum {_terms = v}) mvNormalForm (BladeSum terms) = BladeSum $ if null resultant then [scalarBlade Algebra.Additive.zero] else resultant where resultant = filter bladeNonZero $ addLikeTerms' $ Data.List.Ordered.sortBy compare $ map bladeNormalForm $ terms mvTerms m = m^.terms addLikeTerms' = sumLikeTerms . groupLikeTerms groupLikeTerms ::Eq f => [Blade n f] -> [[Blade n f]] groupLikeTerms = groupBy (\a b -> a^.indices == b^.indices) compensatedSum' :: (Algebra.Additive.C f) => [f] -> f compensatedSum' xs = kahan zero zero xs where kahan s _ [] = s kahan s c (x:xs) = let y = x - c t = s + y in kahan t ((t-s)-y) xs --use this to sum taylor series et al with converge --compensatedRunningSum :: (Algebra.Additive.C f) => [f] -> [f] compensatedRunningSum xs=shanksTransformation . map fst $ scanl kahanSum (zero,zero) xs where kahanSum (s,c) b = (t,newc) where y = b - c t = s + y newc = (t - s) - y --multiplyAdd a b c = a*b + c --twoProduct a b = (x,y) where -- x = a*b --z y = multiplyAdd a b (negate x) --multiplyList [] = [] --multiplyList a@(x:[])=a --multiplyList (a:b:xs) = loop a (b:xs) zero where -- loop pm [] ei = pm+ei -- loop pm1 (ai:remaining) eim1= loop pi remaining ei where -- (pi, pii) = twoProduct pm1 ai -- ei = multiplyAdd eim1 ai pii multiplyOutBlades :: (SingI n, Algebra.Ring.C a) => [Blade n a] -> [Blade n a] -> [Blade n a] multiplyOutBlades x y = [bladeMul l r | l <-x, r <- y] --multiplyList :: Algebra.Ring.C t => [Multivector t] -> Multivector t multiplyList [] = error "Empty list" --multiplyList a@(x:[]) = x multiplyList l = mvNormalForm $ BladeSum listOfBlades where expandedBlades a = foldl1 multiplyOutBlades a listOfBlades = expandedBlades $ map mvTerms l --things to test: is 1. adding blades into a map based on indices 2. adding errything together 3. sort results quicker than -- 1. sorting by indices 2. groupBy-ing on indices 3. adding the lists of identical indices sumList xs = mvNormalForm $ BladeSum $ concat $ map mvTerms xs sumLikeTerms :: (Algebra.Field.C f, SingI n) => [[Blade n f]] -> [Blade n f] sumLikeTerms blades = map (\sameIxs -> map bScale sameIxs & compensatedSum' & (\result -> Blade result ((head sameIxs) & bIndices))) blades instance (Algebra.Field.C f, SingI n, Ord f) => Data.Monoid.Monoid (Sum (Multivector n f)) where mempty = Data.Monoid.Sum Algebra.Additive.zero mappend (Data.Monoid.Sum x) (Data.Monoid.Sum y)= Data.Monoid.Sum (x + y) mconcat = Data.Monoid.Sum . sumList . map getSum instance (Algebra.Field.C f, SingI n, Ord f) => Data.Monoid.Monoid (Product (Multivector n f)) where mempty = Product one mappend (Product x) (Product y) = Product (x * y) mconcat = Product . foldl (*) one . map getProduct --Constructs a multivector from a scaled blade. e :: (Algebra.Field.C f, Ord f, SingI n) => f -> [Natural] -> Multivector n f s `e` indices = mvNormalForm $ BladeSum [Blade s indices] scalar s = s `e` [] instance (Control.DeepSeq.NFData f) => Control.DeepSeq.NFData (Multivector n f) instance (Control.DeepSeq.NFData f) => Control.DeepSeq.NFData (Blade n f) instance (Algebra.Field.C f, Ord f, SingI n) => Algebra.Additive.C (Multivector n f) where a + b = mvNormalForm $ BladeSum (mvTerms a ++ mvTerms b) a - b = mvNormalForm $ BladeSum (mvTerms a ++ map bladeNegate (mvTerms b)) zero = BladeSum [scalarBlade Algebra.Additive.zero] \end{code} Now it is time for the Clifford product. :3 \begin{code} instance (Algebra.Field.C f, Ord f, SingI n) => Algebra.Ring.C (Multivector n f) where BladeSum [Blade s []] * b = BladeSum $ map (bladeScaleLeft s) $ mvTerms b a * BladeSum [Blade s []] = BladeSum $ map (bladeScaleRight s) $ mvTerms a a * b = mvNormalForm $ BladeSum [bladeMul x y | x <- mvTerms a, y <- mvTerms b] one = scalar Algebra.Ring.one fromInteger i = scalar $ Algebra.Ring.fromInteger i a ^ 2 = a * a a ^ 0 = one a ^ 1 = a --a ^ n --n < 0 = Clifford.recip $ a ^ (negate n) a ^ n = multiplyList (replicate (NPN.fromInteger n) a) two = fromInteger 2 mul = (Algebra.Ring.*) \end{code} Clifford numbers have a magnitude and absolute value: \begin{code} --magnitude :: (Algebra.Algebraic.C f) => Multivector f -> f magnitude = sqrt . compensatedSum' . map (\b -> (bScale b)^ 2) . mvTerms instance (Algebra.Absolute.C f, Algebra.Algebraic.C f, Ord f, SingI n) => Algebra.Absolute.C (Multivector n f) where abs v = magnitude v `e` [] signum (BladeSum [Blade scale []]) = scalar $ signum scale signum (BladeSum []) = scalar Algebra.Additive.zero instance (Algebra.Field.C f, Ord f, SingI n) => Algebra.Module.C f (Multivector n f) where -- (*>) zero v = Algebra.Additive.zero (*>) s v = v & mvTerms & map (bladeScaleLeft s) & BladeSum (/) :: (Algebra.Field.C f, Ord f, SingI n) => Multivector n f -> f -> Multivector n f (/) v d = BladeSum $ map (bladeScaleLeft (NPN.recip d)) $ mvTerms v --Algebra.Field.recip d *> v () n d = n * Numeric.Clifford.Multivector.inverse d () n d = n /> d integratePoly c x = c : zipWith (Numeric.Clifford.Multivector./) x progression --converge :: (Eq f, Show f) => [f] -> f converge [] = error "converge: empty list" converge xs = fromMaybe empty (convergeBy checkPeriodic Just xs) where empty = error "converge: error in implmentation" checkPeriodic (a:b:c:_) | (trace ("Converging at " ++ show a) a) == b = Just a | a == c = Just a checkPeriodic _ = Nothing aitkensAcceleration [] = [] aitkensAcceleration a@(xn:[]) = a aitkensAcceleration a@(xn:xnp1:[]) = a aitkensAcceleration a@(xn:xnp1:xnp2:[]) = a aitkensAcceleration (xn:xnp1:xnp2:xs) | xn == xnp1 = [xnp1] | xn == xnp2 = [xnp2] | otherwise = xn - ((dxn ^ 2) /> ddxn) : aitkensAcceleration (xnp1:xnp2:xs) where dxn = sumList [xnp1,negate xn] ddxn = sumList [xn, (-2) * xnp1, xnp2] shanksTransformation [] = [] shanksTransformation a@(xnm1:[]) = a shanksTransformation a@(xnm1:xn:[]) = a shanksTransformation (xnm1:xn:xnp1:xs) | xnm1 == xn = [xn] | xnm1 == xnp1 = [xnm1] | denominator == zero = [xnp1] | otherwise = trace ("Shanks transformation input = " ++ show xn ++ "\nShanks transformation output = " ++ show out) out:shanksTransformation (xn:xnp1:xs) where out = numerator /> denominator numerator = sumList [xnp1*xnm1, negate (xn^2)] denominator = sumList [xnp1, (-2)*xn, xnm1] --exp ::(Ord f, Show f, Algebra.Transcendental.C f)=> Multivector f -> Multivector f exp (BladeSum [ Blade s []]) = trace ("scalar exponential of " ++ show s) scalar $ Algebra.Transcendental.exp s exp x = trace ("Computing exponential of " ++ show x) convergeTerms x where --(expMag ^ expScaled) where --todo: compute a ^ p via a^n where n = floor p then multiply remaining power expMag = Algebra.Transcendental.exp mag expScaled = converge $ shanksTransformation.shanksTransformation . compensatedRunningSum $ expTerms scaled convergeTerms terms = converge $ shanksTransformation.shanksTransformation.compensatedRunningSum $ expTerms terms mag = trace ("In exponential, magnitude is " ++ show ( magnitude x)) magnitude x scaled = let val = (Numeric.Clifford.Multivector./) x mag in trace ("In exponential, scaled is" ++ show val) val takeEvery nth xs = case drop (nth-1) xs of (y:ys) -> y : takeEvery nth ys [] -> [] cosh x = converge $ shanksTransformation . compensatedRunningSum $ takeEvery 2 $ expTerms x sinh x = converge $ shanksTransformation . compensatedRunningSum $ takeEvery 2 $ tail $ expTerms x seriesPlusMinus (x:y:rest) = x:Algebra.Additive.negate y: seriesPlusMinus rest seriesMinusPlus (x:y:rest) = Algebra.Additive.negate x : y : seriesMinusPlus rest sin x = converge $ shanksTransformation $ compensatedRunningSum $ sinTerms x sinTerms x = seriesPlusMinus $ takeEvery 2 $ expTerms x cos x = converge $ shanksTransformation $ compensatedRunningSum (Algebra.Ring.one : cosTerms x) cosTerms x = seriesMinusPlus $ takeEvery 2 $ tail $ expTerms x expTerms x = map snd $ iterate (\(n,b) -> (n + 1, (x*b) Numeric.Clifford.Multivector./ fromInteger n )) (1::NPN.Integer,one) dot a b = mvNormalForm $ BladeSum [x `bDot` y | x <- mvTerms a, y <- mvTerms b] wedge a b = mvNormalForm $ BladeSum [x `bWedge` y | x <- mvTerms a, y <- mvTerms b] (∧) = wedge (⋅) = dot reverseBlade b = bladeNormalForm $ b & indices %~ reverse reverseMultivector v = mvNormalForm $ v & terms.traverse%~ reverseBlade inverse a = assert (a /= zero) $ reverseMultivector a Numeric.Clifford.Multivector./ bScale (head $ mvTerms (a * reverseMultivector a)) recip=Numeric.Clifford.Multivector.inverse instance (Algebra.Field.C f, Ord f, SingI n) => Algebra.OccasionallyScalar.C f (Multivector n f) where toScalar = bScale . bladeGetGrade 0 . head . mvTerms toMaybeScalar (BladeSum [Blade s []]) = Just s toMaybeScalar (BladeSum []) = Just Algebra.Additive.zero toMaybeScalar _ = Nothing fromScalar = scalar \end{code} Also, we may as well implement the standard prelude Num interface. \begin{code} instance (Algebra.Algebraic.C f, SingI n, Ord f) => PNum.Num (Multivector n f) where (+) = (Algebra.Additive.+) (-) = (Algebra.Additive.-) (*) = (Algebra.Ring.*) negate = NPN.negate abs = scalar . magnitude fromInteger = Algebra.Ring.fromInteger signum m = Numeric.Clifford.Multivector.inverse (scalar $ magnitude m) * m \end{code} Let's use Newton or Halley iteration to find the principal n-th root :3 \begin{code} root :: (Show f, Ord f, Algebra.Algebraic.C f, SingI d) => NPN.Integer -> Multivector d f -> Multivector d f root n (BladeSum [Blade s []]) = scalar $ Algebra.Algebraic.root n s root n a@(BladeSum _) = converge $ rootIterationsStart n a one rootIterationsStart ::(Ord f, Show f, Algebra.Algebraic.C f)=> NPN.Integer -> Multivector d f -> Multivector d f -> [Multivector d f] rootIterationsStart n a@(BladeSum (Blade s [] :xs)) one = rootHalleysIterations n a g where g = if s >= NPN.zero then one else Algebra.Ring.one `e` [1,2] --BladeSum[Blade Algebra.Ring.one [1,2]] rootIterationsStart n a@(BladeSum _) g = rootHalleysIterations n a g rootNewtonIterations :: (Algebra.Field.C f, Ord f, SingI d) => NPN.Integer -> Multivector d f -> Multivector d f -> [Multivector d f] rootNewtonIterations n a = iterate xkplus1 where xkplus1 xk = xk + deltaxk xk deltaxk xk = oneOverN * ((Numeric.Clifford.Multivector.inverse (xk ^ (n - one))* a) - xk) oneOverN = scalar $ NPN.recip $ fromInteger n rootHalleysIterations :: (Show a, Ord a, Algebra.Algebraic.C a, SingI d) => NPN.Integer -> Multivector d a -> Multivector d a -> [Multivector d a] rootHalleysIterations n a = halleysMethod f f' f'' where f x = a - (x^n) f' x = fromInteger (-n) * (x^(n-1)) f'' x = fromInteger (-(n*(n-1))) * (x^(n-2)) pow a p = (a ^ up) Numeric.Clifford.Multivector./> Numeric.Clifford.Multivector.root down a where ratio = toRational p up = numerator ratio down = denominator ratio halleysMethod :: (Show a, Ord a, Algebra.Algebraic.C a, SingI d) => (Multivector d a -> Multivector d a) -> (Multivector d a -> Multivector d a) -> (Multivector d a -> Multivector d a) -> Multivector d a -> [Multivector d a] halleysMethod f f' f'' = iterate update where update x = x - (numerator x * Numeric.Clifford.Multivector.inverse (denominator x)) where numerator x = multiplyList [2, fx, dfx] denominator x = multiplyList [2, dfx, dfx] - (fx * ddfx) fx = f x dfx = f' x ddfx = f'' x secantMethod f x0 x1 = update x1 x0 where update xm1 xm2 | xm1 == xm2 = [xm1] | otherwise = if x == xm1 then [x] else x : update x xm1 where x = xm1 - f xm1 * (xm1-xm2) * Numeric.Clifford.Multivector.inverse (f xm1 - f xm2) \end{code} Now let's try logarithms by fixed point iteration. It's gonna be slow, but whatever! \begin{code} normalised a = a * (scalar $ NPN.recip $ magnitude a) log (BladeSum [Blade s []]) = scalar $ NPN.log s log a = scalar (NPN.log mag) + log' scaled where scaled = normalised a mag = magnitude a log' a = converge $ halleysMethod f f' f'' (one `e` [1,2]) where f x = a - Numeric.Clifford.Multivector.exp x f' x = NPN.negate $ Numeric.Clifford.Multivector.exp x f'' = f' \end{code} Now let's do (slow as fuck probably) numerical integration! :D~! Since this is gonna be used for physical applications, it's we're gonna start off with a Hamiltonian structure and then a symplectic integrator. \begin{code} {- $(derive makeSerialize ''Blade) $(derive makeSerialize ''Multivector) $(derive makeData ''Blade) $(derive makeTypeable ''Blade) $(derive makeData ''Multivector) $(derive makeTypeable ''Multivector)-} -- $(derive makeArbitrary ''Multivector) \end{code} \bibliographystyle{IEEEtran} \bibliography{biblio.bib} \end{document}