math-functions-0.2.1.0: Special functions and Chebyshev polynomials

Copyright(c) 2014 Bryan O'Sullivan
LicenseBSD3
Maintainerbos@serpentine.com
Stabilityexperimental
Portabilityportable
Safe HaskellNone
LanguageHaskell2010

Numeric.Sum

Contents

Description

Functions for summing floating point numbers more accurately than the naive sum function and its counterparts in the vector package and elsewhere.

When used with floating point numbers, in the worst case, the sum function accumulates numeric error at a rate proportional to the number of values being summed. The algorithms in this module implement different methods of /compensated summation/, which reduce the accumulation of numeric error so that it either grows much more slowly than the number of inputs (e.g. logarithmically), or remains constant.

Synopsis

Summation type class

class Summation s where #

A class for summation of floating point numbers.

Minimal complete definition

zero, add

Methods

zero :: s #

The identity for summation.

add :: s -> Double -> s #

Add a value to a sum.

sum :: Foldable f => (s -> Double) -> f Double -> Double #

Sum a collection of values.

Example: foo = sum kbn [1,2,3]

Instances

sumVector :: (Vector v Double, Summation s) => (s -> Double) -> v Double -> Double #

O(n) Sum a vector of values.

Usage

Most of these summation algorithms are intended to be used via the Summation typeclass interface. Explicit type annotations should not be necessary, as the use of a function such as kbn or kb2 to extract the final sum out of a Summation instance gives the compiler enough information to determine the precise type of summation algorithm to use.

As an example, here is a (somewhat silly) function that manually computes the sum of elements in a list.

sillySumList :: [Double] -> Double
sillySumList = loop zero
  where loop s []     = kbn s
        loop s (x:xs) = seq s' loop s' xs
          where s'    = add s x

In most instances, you can simply use the much more general sum function instead of writing a summation function by hand.

-- Avoid ambiguity around which sum function we are using.
import Prelude hiding (sum)
--
betterSumList :: [Double] -> Double
betterSumList xs = sum kbn xs

Kahan-Babuška-Neumaier summation

data KBNSum #

Kahan-Babuška-Neumaier summation. This is a little more computationally costly than plain Kahan summation, but is always at least as accurate.

Constructors

KBNSum !Double !Double 

Instances

Eq KBNSum # 

Methods

(==) :: KBNSum -> KBNSum -> Bool #

(/=) :: KBNSum -> KBNSum -> Bool #

Data KBNSum # 

Methods

gfoldl :: (forall d b. Data d => c (d -> b) -> d -> c b) -> (forall g. g -> c g) -> KBNSum -> c KBNSum #

gunfold :: (forall b r. Data b => c (b -> r) -> c r) -> (forall r. r -> c r) -> Constr -> c KBNSum #

toConstr :: KBNSum -> Constr #

dataTypeOf :: KBNSum -> DataType #

dataCast1 :: Typeable (* -> *) t => (forall d. Data d => c (t d)) -> Maybe (c KBNSum) #

dataCast2 :: Typeable (* -> * -> *) t => (forall d e. (Data d, Data e) => c (t d e)) -> Maybe (c KBNSum) #

gmapT :: (forall b. Data b => b -> b) -> KBNSum -> KBNSum #

gmapQl :: (r -> r' -> r) -> r -> (forall d. Data d => d -> r') -> KBNSum -> r #

gmapQr :: (r' -> r -> r) -> r -> (forall d. Data d => d -> r') -> KBNSum -> r #

gmapQ :: (forall d. Data d => d -> u) -> KBNSum -> [u] #

gmapQi :: Int -> (forall d. Data d => d -> u) -> KBNSum -> u #

gmapM :: Monad m => (forall d. Data d => d -> m d) -> KBNSum -> m KBNSum #

gmapMp :: MonadPlus m => (forall d. Data d => d -> m d) -> KBNSum -> m KBNSum #

gmapMo :: MonadPlus m => (forall d. Data d => d -> m d) -> KBNSum -> m KBNSum #

Show KBNSum # 
NFData KBNSum # 

Methods

rnf :: KBNSum -> () #

Unbox KBNSum # 
Summation KBNSum # 

Methods

zero :: KBNSum #

add :: KBNSum -> Double -> KBNSum #

sum :: Foldable f => (KBNSum -> Double) -> f Double -> Double #

Vector Vector KBNSum # 
MVector MVector KBNSum # 
data Vector KBNSum # 
data MVector s KBNSum # 

kbn :: KBNSum -> Double #

Return the result of a Kahan-Babuška-Neumaier sum.

Order-2 Kahan-Babuška summation

data KB2Sum #

Second-order Kahan-Babuška summation. This is more computationally costly than Kahan-Babuška-Neumaier summation, running at about a third the speed. Its advantage is that it can lose less precision (in admittedly obscure cases).

This method compensates for error in both the sum and the first-order compensation term, hence the use of "second order" in the name.

Constructors

KB2Sum !Double !Double !Double 

Instances

Eq KB2Sum # 

Methods

(==) :: KB2Sum -> KB2Sum -> Bool #

(/=) :: KB2Sum -> KB2Sum -> Bool #

Data KB2Sum # 

Methods

gfoldl :: (forall d b. Data d => c (d -> b) -> d -> c b) -> (forall g. g -> c g) -> KB2Sum -> c KB2Sum #

gunfold :: (forall b r. Data b => c (b -> r) -> c r) -> (forall r. r -> c r) -> Constr -> c KB2Sum #

toConstr :: KB2Sum -> Constr #

dataTypeOf :: KB2Sum -> DataType #

dataCast1 :: Typeable (* -> *) t => (forall d. Data d => c (t d)) -> Maybe (c KB2Sum) #

dataCast2 :: Typeable (* -> * -> *) t => (forall d e. (Data d, Data e) => c (t d e)) -> Maybe (c KB2Sum) #

gmapT :: (forall b. Data b => b -> b) -> KB2Sum -> KB2Sum #

gmapQl :: (r -> r' -> r) -> r -> (forall d. Data d => d -> r') -> KB2Sum -> r #

gmapQr :: (r' -> r -> r) -> r -> (forall d. Data d => d -> r') -> KB2Sum -> r #

gmapQ :: (forall d. Data d => d -> u) -> KB2Sum -> [u] #

gmapQi :: Int -> (forall d. Data d => d -> u) -> KB2Sum -> u #

gmapM :: Monad m => (forall d. Data d => d -> m d) -> KB2Sum -> m KB2Sum #

gmapMp :: MonadPlus m => (forall d. Data d => d -> m d) -> KB2Sum -> m KB2Sum #

gmapMo :: MonadPlus m => (forall d. Data d => d -> m d) -> KB2Sum -> m KB2Sum #

Show KB2Sum # 
NFData KB2Sum # 

Methods

rnf :: KB2Sum -> () #

Unbox KB2Sum # 
Summation KB2Sum # 

Methods

zero :: KB2Sum #

add :: KB2Sum -> Double -> KB2Sum #

sum :: Foldable f => (KB2Sum -> Double) -> f Double -> Double #

Vector Vector KB2Sum # 
MVector MVector KB2Sum # 
data Vector KB2Sum # 
data MVector s KB2Sum # 

kb2 :: KB2Sum -> Double #

Return the result of an order-2 Kahan-Babuška sum.

Less desirable approaches

Kahan summation

data KahanSum #

Kahan summation. This is the least accurate of the compensated summation methods. In practice, it only beats naive summation for inputs with large magnitude. Kahan summation can be less accurate than naive summation for small-magnitude inputs.

This summation method is included for completeness. Its use is not recommended. In practice, KBNSum is both 30% faster and more accurate.

Constructors

KahanSum !Double !Double 

Instances

Eq KahanSum # 
Data KahanSum # 

Methods

gfoldl :: (forall d b. Data d => c (d -> b) -> d -> c b) -> (forall g. g -> c g) -> KahanSum -> c KahanSum #

gunfold :: (forall b r. Data b => c (b -> r) -> c r) -> (forall r. r -> c r) -> Constr -> c KahanSum #

toConstr :: KahanSum -> Constr #

dataTypeOf :: KahanSum -> DataType #

dataCast1 :: Typeable (* -> *) t => (forall d. Data d => c (t d)) -> Maybe (c KahanSum) #

dataCast2 :: Typeable (* -> * -> *) t => (forall d e. (Data d, Data e) => c (t d e)) -> Maybe (c KahanSum) #

gmapT :: (forall b. Data b => b -> b) -> KahanSum -> KahanSum #

gmapQl :: (r -> r' -> r) -> r -> (forall d. Data d => d -> r') -> KahanSum -> r #

gmapQr :: (r' -> r -> r) -> r -> (forall d. Data d => d -> r') -> KahanSum -> r #

gmapQ :: (forall d. Data d => d -> u) -> KahanSum -> [u] #

gmapQi :: Int -> (forall d. Data d => d -> u) -> KahanSum -> u #

gmapM :: Monad m => (forall d. Data d => d -> m d) -> KahanSum -> m KahanSum #

gmapMp :: MonadPlus m => (forall d. Data d => d -> m d) -> KahanSum -> m KahanSum #

gmapMo :: MonadPlus m => (forall d. Data d => d -> m d) -> KahanSum -> m KahanSum #

Show KahanSum # 
NFData KahanSum # 

Methods

rnf :: KahanSum -> () #

Unbox KahanSum # 
Summation KahanSum # 

Methods

zero :: KahanSum #

add :: KahanSum -> Double -> KahanSum #

sum :: Foldable f => (KahanSum -> Double) -> f Double -> Double #

Vector Vector KahanSum # 
MVector MVector KahanSum # 
data Vector KahanSum # 
data MVector s KahanSum # 

kahan :: KahanSum -> Double #

Return the result of a Kahan sum.

Pairwise summation

pairwiseSum :: Vector v Double => v Double -> Double #

O(n) Sum a vector of values using pairwise summation.

This approach is perhaps 10% faster than KBNSum, but has poorer bounds on its error growth. Instead of having roughly constant error regardless of the size of the input vector, in the worst case its accumulated error grows with O(log n).

References

  • Kahan, W. (1965), Further remarks on reducing truncation errors. Communications of the ACM 8(1):40.
  • Neumaier, A. (1974), Rundungsfehleranalyse einiger Verfahren zur Summation endlicher Summen. Zeitschrift für Angewandte Mathematik und Mechanik 54:39–51.
  • Klein, A. (2006), A Generalized Kahan-Babuška-Summation-Algorithm. Computing 76(3):279-293.
  • Higham, N.J. (1993), The accuracy of floating point summation. SIAM Journal on Scientific Computing 14(4):783–799.