-- | This module computes the histogram by the -- specified data and strategy applied for such computing. -- -- The code in this module is essentially based on the -- <http://hackage.haskell.org/package/Histogram> package -- by Mike Izbicki, who kindly agreed to re-license -- his library under BSD3, which allowed me to use his -- code and comments with some modifications. module Simulation.Aivika.Experiment.Histogram (-- * Creating Histograms Histogram(..), histogram, histogramBinSize, histogramNumBins, -- * Binning Strategies BinningStrategy(..), binSturges, binDoane, binSqrt, binScott) where import Data.List import Data.Monoid import qualified Data.Map as Map import Numeric --- import System.IO.Unsafe --- ------------------------------------------------------------------------------- -- | Holds all the information needed to plot the histogram -- for a list of different series. Each series produces its -- own item in the resuling @[Int]@ list that may contain -- zeros. type Histogram = [(Double, [Int])] ------------------------------------------------------------------------------- -- Bin counters; check out http://en.wikipedia.org/wiki/Histogram#Number_of_bins_and_width -- | The strategy applied to calculate the histogram bins. type BinningStrategy = [Double] -> Int -- | This is only a helper function to convert strategies that -- specify bin width into strategies that specify the number of bins. stratFromBinWidth :: [Double] -> Double -> Int stratFromBinWidth xs h = ceiling $ (maximum xs - minimum xs) / h -- | Sturges' binning strategy is the least computational work, -- but recommended for only normal data. binSturges :: BinningStrategy binSturges xs = ceiling $ logBase 2 n + 1 where n = fromIntegral $ length xs -- | Doane's binning strategy extends Sturges' for non-normal data. -- It takes a little more time because it must calculate the kurtosis -- (peakkiness) of the distribution. binDoane :: BinningStrategy binDoane xs = ceiling $ 1 + log n + log (1 + a * ((n/6)**(1/2))) where n = fromIntegral $ length xs a = kurtosis xs -- | Using the sqrt of the number of samples is not supported by any -- theory, but is commonly used by excel and other histogram making software. binSqrt :: BinningStrategy binSqrt xs = round $ sqrt n where n = fromIntegral $ length xs -- | Scott's rule is the optimal solution for normal data, but requires -- more computation than Sturges'. binScott :: BinningStrategy binScott xs = stratFromBinWidth xs $ 3.5 * stddev xs / (n**(1/3)) where n = fromIntegral $ length xs ------------------------------------------------------------------------------- -- create the histogram data -- | Creates a histogram by specifying the list of series. Call it with one of -- the binning strategies that is appropriate to the type of data you have. -- If you don't know, then try using 'binSturges'. histogram :: BinningStrategy -> [[Double]] -> Histogram histogram strat xs = histogramNumBins (strat $ concat xs) xs -- | Create a histogram by specifying the exact bin size. -- You probably don't want to use this function, and should use histogram -- with an appropriate binning strategy. histogramBinSize :: Double -> [[Double]] -> Histogram histogramBinSize size = combineBins . map (histBins size . bin size) -- | Create a histogram by the specified approximated number of bins. -- You probably don't want to use this function, and should use -- histogram with an appropriate binning strategy. histogramNumBins :: Int -> [[Double]] -> Histogram histogramNumBins n xs = histogramBinSize size xs where size = fromIntegral (firstdigit diff) * (10 ** fromIntegral (exponent10 diff)) diff = if diff_test > 0 then diff_test else 1 diff_test = (maximum (map maximum xs) - minimum (map minimum xs)) / fromIntegral (max 1 n) firstdigit dbl = floor $ dbl / (10 ** fromIntegral (exponent10 dbl)) exponent10 dbl = floor $ logBase 10 dbl -- | It does all the binning for the histogram. histBins :: Double -> [Double] -> [(Double, Int)] histBins size xs = [ (head l, length l) | l <- group (sort xs) ] -- | It "rounds" every number into the closest number below it -- that is divisible by size. bin :: Double -> [Double] -> [Double] bin size xs @ [_] = xs bin size xs = map (\x -> size * fromIntegral (floor (x / size)) + size / 2) xs -- | Combine bins from different histograms (optimized version). combineBins :: [[(Double, Int)]] -> [(Double, [Int])] combineBins [xs] = map (\(t, n) -> (t, [n])) xs combineBins xss = combineBins' xss -- | Combine bins from different histograms (generic version). combineBins' :: [[(Double, Int)]] -> [(Double, [Int])] combineBins' xs = map (merging . sorting) $ groupping $ ordering $ concat $ numbering xs where numbering = zipWith (curry indexing) [1..] indexing (i, x) = map (\(t, n) -> (i, t, n)) x ordering = sortBy $ \(_, t1, _) (_, t2, _) -> compare t1 t2 groupping = groupBy $ \(_, t1, _) (_, t2, _) -> t1 == t2 sorting = sortBy $ \(i1, _, _) (i2, _, _) -> compare i1 i2 merging zs @ ((_, t, _) : _) = merging' zs t 1 [] merging' [] t k acc | k <= count = merging' [] t (k + 1) (0 : acc) | k > count = (t, reverse acc) merging' zs @ ((i, _, n) : xs) t k acc | i == k = merging' xs t (k + 1) (n : acc) | i > k = merging' zs t (k + 1) (0 : acc) count = length xs ------------------------------------------------------------------------------ -- simple math functions -- taken from package hstats, which wouldn't fully compile -- | Numerically stable mean. mean :: Floating a => [a] -> a mean x = fst $ foldl' (\(m, n) x -> (m+(x-m)/(n+1),n+1)) (0,0) x -- | Standard deviation of sample. stddev :: (Floating a) => [a] -> a stddev xs = sqrt $ var xs -- | Sample variance. var :: (Floating a) => [a] -> a var xs = var' 0 0 0 xs / fromIntegral (length xs - 1) where var' _ _ s [] = s var' m n s (x:xs) = var' nm (n + 1) (s + delta * (x - nm)) xs where delta = x - m nm = m + delta / fromIntegral (n + 1) -- | kurtosis is taken from wikipedia's definition. kurtosis :: (Floating a) => [a] -> a kurtosis xs = ((1/n) * sum [(x-x_bar)^4 | x <- xs]) / ((1/n) * sum [(x-x_bar)^2 | x <- xs])^2 -3 where n = fromIntegral $ length xs x_bar = mean xs