First take on cleaning up EmpiricalUnivariateDistribution

Author: andreasnoack

src/Distributions.jl

@@ -268,7 +268,6 @@ include("genericfit.jl")
 
 # specific samplers and distributions
 include("univariates.jl")
-include("empirical.jl")
 include("edgeworth.jl")
 include("multivariates.jl")
 include("matrixvariates.jl")

src/empirical.jl

@@ -0,0 +1,30 @@
+struct EmpiricalUnivariateDistribution <: DiscreteUnivariateDistribution
+    values::Vector{Float64}
+    cdf::Function
+end
+
+@distr_support EmpiricalUnivariateDistribution d.values[1] d.values[end]
+
+EmpiricalUnivariateDistribution(x::Vector) = EmpiricalUnivariateDistribution(sort(x), ecdf(x))
+
+for f in (:entropy, :mean, :var, :skewness, :kurtosis)
+    @eval ($f)(d::EmpiricalUnivariateDistribution) = ($f)(d.values)
+end
+
+function median(d::DiscreteUnivariateDistribution)
+    v = d.values
+    n = length(v)
+    return (v[(n + 1) >> 1] + v[(n + 2) >> 1]) / 2
+end
+
+### Evaluation
+
+cdf(d::EmpiricalUnivariateDistribution, x::Real) = d.cdf(x)
+
+pdf(d::EmpiricalUnivariateDistribution, x::Real) = mean(t -> t == x, d.values)
+
+quantile(d::EmpiricalUnivariateDistribution, p::Real) = quantile(d.values, p)
+
+function rand(d::EmpiricalUnivariateDistribution)
+    d.values[rand(1:length(d.values))]
+end

src/univariate/discrete/empirical.jl

@@ -610,6 +610,7 @@ const discrete_distributions = [
     "binomial",
     "categorical",
     "discreteuniform",
+    "empirical",
     "geometric",
     "hypergeometric",
     "negativebinomial",

src/univariates.jl

@@ -0,0 +1,28 @@
+using Distributions, Base.Test
+
+@testset "EmpiricalUnivariateDistribution" begin
+    n = 100
+    r = MersenneTwister(123)
+    @testset "$d data" for (d, x) in (("discrete"  , rand( r, 1:10, n)),
+                                 ("continuous", randn(r,       n)))
+        X = EmpiricalUnivariateDistribution(x)
+
+        @testset "test function: $f" for f in (mean, var, std, skewness, kurtosis, median, entropy)
+            @test f(X) ≈ f(x)
+        end
+
+        ecdfx = StatsBase.ecdf(x)
+        @testset "cdf" for t in linspace(-10, 10, 100)
+            @test cdf(X, t) == ecdfx(t)
+            @test cdf(X, t) == mean(x -> x <= t, x)
+        end
+
+        @testset "quantile" for q in linspace(0, 1, 100)
+            @test quantile(X, q) == quantile(x, q)
+        end
+
+        @testset "pdf" begin
+            @test sum(t -> pdf(X, t), unique(x)) ≈ 1
+        end
+    end
+end

test/empirical.jl

@@ -27,7 +27,8 @@ tests = [
     "gradlogpdf",
     "truncate",
     "noncentralt",
-    "locationscale"]
+    "locationscale",
+    "empirical"]
 
 print_with_color(:blue, "Running tests:\n")