distribution 0.2.0 → 0.3.0

data/History.txt

@@ -1,3 +1,12 @@
+=== 0.3.0 / 2011-01-28
+
+* Included support for binomial distribution. p_value is not accurate
+* Included alias for ruby exact methods on discrete distributions, when they are available
+* Works on Jruby and Ruby 1.8.7 
+* Binomial coefficient optimized. Falling factorial method with Swing Prime on factorial provides a 10x improvement over naive version.
+* Working on binomial coefficient using gamma. The same problem as using Sterling: faster than exact version, but when it requires BigDecimal, is 2-3x slower
+* Deleted Fixnum extension on Hypergeometric
+
 === 0.2.0 / 2011-01-27
 * Shorthands for F and T are 'fdist' and 'tdist' now, to avoid confussions 
 * Added Hypergeometric distribution with pdf, cdf, p_value. The Ruby version uses a very slow factorial process, but also has pdf_with_stirling if you don't care about accuracy at all. This latter method needs to be improved, actually, in terms of which choose function it calls for the numerator and denominator. Once that's done, it can replace the slower pdf function. [John Woods]

data/Manifest.txt

@@ -3,35 +3,47 @@ History.txt
 Manifest.txt
 README.txt
 Rakefile
+benchmark/binomial_coefficient.rb
 benchmark/factorial_method.rb
+benchmark/odd.rb
 bin/distribution
 lib/distribution.rb
 lib/distribution/binomial.rb
+lib/distribution/binomial/gsl.rb
+lib/distribution/binomial/java.rb
+lib/distribution/binomial/ruby.rb
 lib/distribution/bivariatenormal.rb
 lib/distribution/bivariatenormal/gsl.rb
+lib/distribution/bivariatenormal/java.rb
 lib/distribution/bivariatenormal/ruby.rb
 lib/distribution/bivariatenormal/statistics2.rb
 lib/distribution/chisquare.rb
 lib/distribution/chisquare/gsl.rb
+lib/distribution/chisquare/java.rb
 lib/distribution/chisquare/ruby.rb
 lib/distribution/chisquare/statistics2.rb
 lib/distribution/f.rb
 lib/distribution/f/gsl.rb
+lib/distribution/f/java.rb
 lib/distribution/f/ruby.rb
 lib/distribution/f/statistics2.rb
 lib/distribution/hypergeometric.rb
 lib/distribution/hypergeometric/gsl.rb
+lib/distribution/hypergeometric/java.rb
 lib/distribution/hypergeometric/ruby.rb
 lib/distribution/math_extension.rb
 lib/distribution/normal.rb
 lib/distribution/normal/gsl.rb
+lib/distribution/normal/java.rb
 lib/distribution/normal/ruby.rb
 lib/distribution/normal/statistics2.rb
 lib/distribution/normalmultivariate.rb
 lib/distribution/t.rb
 lib/distribution/t/gsl.rb
+lib/distribution/t/java.rb
 lib/distribution/t/ruby.rb
 lib/distribution/t/statistics2.rb
+spec/binomial_spec.rb
 spec/bivariatenormal_spec.rb
 spec/chisquare_spec.rb
 spec/distribution_spec.rb

data/README.txt

@@ -11,17 +11,26 @@ Includes code from statistics2
 
 == FEATURES/PROBLEMS:
 
-* Base API
-* Shorthand for easy use
+* Very fast ruby 1.8.7/1.9.+ implementation, with improved method to calculate factorials and others common functions
+* All methods tested on several ranges. See spec/
+* On Jruby, BivariateNormal returns incorrect pdf
 
 == API structure
 
   Distribution::<name>.(cdf|pdf|p_value|rng)
 
+On discrete distributions, exact Ruby implementations of pdf, cdf and p_value could be provided, using
+
+  Distribution::<name>.exact_(cdf|pdf|p_value)
+
 module Distribution::Shorthand provides (you guess?) shortands method to call all methods
 
   <Distribution shortname>_(cdf|pdf|p|r)
 
+On discrete distributions, exact cdf, pdf and p_value are
+
+  <Distribution shortname>_(ecdf|epdf|ep)
+  
 Shortnames are:
 
 * Normal: norm
@@ -29,14 +38,17 @@ Shortnames are:
 * T: tdist
 * F: fdist
 * Chi Square: chisq
+* Binomial: bino
 * Hypergeometric: hypg
+
 For example
 
-  Distribution::T.rng
+  Distribution::T.cdf
 
 could be called after including Distribution::Shorthand
-
-  t_r
+  
+  tdist_cdf
+  
 
 == SYNOPSIS:
   # Returns Gaussian PDF for x
@@ -48,15 +60,20 @@ could be called after including Distribution::Shorthand
   
 == REQUIREMENTS:
 
-I try to provide a Ruby version for each method. But to increase (notably!) the speed, please install
+I try to provide a Ruby version for each method. To increase (notably!) the speed, please install
 
 * Ruby 1.8-1.9: gsl or statistics2
-* Java: Apache Math
+* Java: Apache Math (not yet implemented)
 
 == INSTALL:
 
   gem install distribution
 
+To speep up
+  
+  gem install gsl
+  gem install statistics
+  
 == DEVELOPERS:
 
 After checking out the source, run:

data/Rakefile

@@ -14,6 +14,8 @@ require 'rubyforge'
 Hoe.spec 'distribution' do
   self.developer('Claudio Bustos', 'clbustos_at_gmail.com')
   self.version=Distribution::VERSION
+  self.extra_dev_deps << ["rspec",">=2.0"] << ["rubyforge",">=0"]
+
 end
 
 # vim: syntax=ruby

data/benchmark/binomial_coefficient.rb

@@ -0,0 +1,55 @@
+$:.unshift(File.dirname(__FILE__)+"/../lib")
+require 'distribution'
+require 'bench_press'
+
+extend BenchPress
+
+name 'binomial coefficient: multiplicative, factorial and optimized factorial methods'
+author 'Claudio Bustos'
+date '2011-01-27'
+summary "Exact calculation of Binomial Coefficient could be obtained using multiplicative, pure factorial or optimized factorial algorithm. 
+Which one is faster?
+
+Lower k is the best for all 
+k=n/2 is the worst case for optimized algorithm
+k near n is the worst for multiplicative
+
+The factorial method uses the fastest Swing Prime Algorithm."
+
+reps 10 #number of repetitions
+
+x=100
+
+n=100
+k=50
+
+samples=10.times.map {|i| 2**(i+1)}
+
+
+
+measure "Multiplicative" do
+  samples.each do |n|
+    [5,n/2,n-1].each do |k|
+      k=[k,n-k].min
+      (1..k).inject(1) {|ac, i| (ac*(n-k+i).quo(i))}
+    end
+  end
+end
+
+measure "Factorial" do
+  samples.each do |n|
+    [5,n/2,n-1].each do |k|
+      k=[k,n-k].min
+      Math.factorial(n).quo(Math.factorial(k) * Math.factorial(n - k))
+    end
+  end
+end
+
+measure "Optimized Factorial" do
+  samples.each do |n|
+    [5,n/2,n-1].each do |k|
+      k=[k,n-k].min
+      (((n-k+1)..n).inject(1) {|ac,v| ac * v}).quo(Math.factorial(k))
+    end
+  end
+end

data/benchmark/factorial_method.rb

@@ -8,7 +8,9 @@ name 'aprox vs exact factorization method'
 author 'Claudio Bustos'
 date '2011-01-27'
 summary "
-Factorization requires a lot of processing, so approximation method are required. But for greats value, bigdecimal are required and our 
+Factorization requires a lot of processing, so approximation method could be  required. But for greats value, bigdecimal are required and things start to get harder.
+* Approximation (fast_factorial): Luschny f.3
+* Exact (factorial): Luschny Swing Prime 
 "
 
 reps 10 #number of repetitions
@@ -20,6 +22,7 @@ measure "Math.factorial(#{x})" do
 end
 
 measure "Math.fast_factorial(#{x})" do
+  
   Math.fast_factorial(x)
 end
 

data/benchmark/odd.rb

@@ -0,0 +1,22 @@
+
+require 'bench_press'
+
+extend BenchPress
+
+name 'n&1==1 vs n%2==1 to detect odd numbers'
+author 'Claudio Bustos'
+date '2011-01-28'
+summary "
+Which is faster, n%1==1 or n%2==1
+"
+
+reps 10_000 #number of repetitions
+n=100000
+measure "Using &" do
+  n%1==1
+end
+
+measure "Using %" do
+  n%2==1
+end
+

data/lib/distribution.rb

@@ -39,6 +39,7 @@ end
 end
 require 'distribution/math_extension'
 
+
 # Several distributions modules to calculate pdf, cdf, inverse cdf and generate
 # pseudo-random numbers for several statistical distributions
 # 
@@ -48,15 +49,15 @@ require 'distribution/math_extension'
 #    Distribution::Normal.p_value(0.95)
 #    => 1.64485364660836
 module Distribution
-  VERSION="0.2.0"
+  VERSION="0.3.0"
   
   module Shorthand
-    EQUIVALENCES={:p_value=>:p, :cdf=>:cdf, :pdf=>:pdf, :rng=>:r}
-    def self.add_shortcut(sh,m,&block)
+    EQUIVALENCES={:p_value=>:p, :cdf=>:cdf, :pdf=>:pdf, :rng=>:r, :exact_pdf=>:epdf, :exact_cdf=>:ecdf, :exact_p_value=>:ep}
+    def self.add_shortcut(sh,m, &block)
       if EQUIVALENCES.include? m.to_sym 
         sh_name=sh+"_#{m}"
         define_method(sh_name,&block)
-        sh_name=sh+"_#{EQUIVALENCES[m]}"
+        sh_name=sh+"_#{EQUIVALENCES[m.to_sym]}"
         define_method(sh_name,&block)
         
       end
@@ -107,9 +108,10 @@ module Distribution
     #
     # Kids: Metaprogramming trickery! Don't do at work.
     # This section was created between a very long reunion
-    # and a travel of 456 Km.
-    def create_distribution_methods()  
-    Distribution.libraries_order.each do |l_name|
+    # and a 456 Km. travel
+    def create_distribution_methods()
+      
+      Distribution.libraries_order.each do |l_name|
       if const_defined? l_name
         l =const_get(l_name)
         # Add methods from engine to base base, if not yet included
@@ -141,6 +143,8 @@ module Distribution
   autoload(:T, 'distribution/t')
   autoload(:F, 'distribution/f')
   autoload(:BivariateNormal, 'distribution/bivariatenormal')
+  autoload(:Binomial, 'distribution/binomial')
+
   autoload(:Hypergeometric, 'distribution/hypergeometric')
 end
 

data/lib/distribution/binomial.rb

@@ -1,26 +1,28 @@
 require 'distribution/binomial/ruby'
 require 'distribution/binomial/gsl'
+require 'distribution/binomial/java'
 module Distribution
 
-    # Calculate statisticals for T Distribution.
-    module T
-      SHORTHAND='bino'  
-      extend Distributable
-      create_distribution_methods
-        
-      ##
-      # :singleton-method: pdf(x,k)
-      # Returns the integral of T distribution 
-      # with +k+ degrees of freedom over [0, +x+]    
-  
-      ##
-      # :singleton-method: p_value(qn, k)
-      # Return the P-value of the corresponding integral +qn+ with 
-      # +k+ degrees of freedom    
+  # Calculate statisticals for T Distribution.
+  module Binomial
+    SHORTHAND = 'bino'  
+    
+    extend Distributable
+    create_distribution_methods
       
-      ##
-      # :singleton-method: cdf(x,k)
-      # Returns the integral of T distribution 
-      # with +k+ degrees of freedom over [0, +x+]    
-    end
+    ##
+    # :singleton-method: pdf(x,k)
+    # Returns the integral of T distribution 
+    # with +k+ degrees of freedom over [0, +x+]    
+
+    ##
+    # :singleton-method: p_value(qn, k)
+    # Return the P-value of the corresponding integral +qn+ with 
+    # +k+ degrees of freedom    
+    
+    ##
+    # :singleton-method: cdf(x,k)
+    # Returns the integral of T distribution 
+    # with +k+ degrees of freedom over [0, +x+]    
+  end
 end

data/lib/distribution/binomial/gsl.rb

@@ -0,0 +1,14 @@
+module Distribution
+  module Binomial
+    module GSL_
+      class << self
+        def pdf(k,n,prob)
+          GSL::Ran.binomial_pdf(k,prob,n)
+        end
+        def cdf(k,n,prob)
+          GSL::Cdf.binomial_P(k,prob,n)
+        end
+      end
+    end
+  end
+end

data/lib/distribution/binomial/java.rb

@@ -0,0 +1,9 @@
+module Distribution
+  module Binomial
+    # TODO
+    module Java_
+      class << self
+      end
+    end
+  end
+end

data/lib/distribution/binomial/ruby.rb

@@ -0,0 +1,26 @@
+module Distribution
+  module Binomial
+    module Ruby_
+      class << self
+        def pdf(k,n,pr)
+          Math.binomial_coefficient(n,k)*(pr**k)*(1-pr)**(n-k)          
+        end
+        def cdf(k,n,pr)
+          #(0..x.floor).inject(0) {|ac,i| ac+pdf(i,n,pr)}
+          Math.regularized_beta_function(1-pr,n - k,k+1)
+        end
+        def p_value(prob,n,pr)
+          ac=0
+          (0..n).each do |i|
+            ac+=pdf(i,n,pr)
+            return i if ac>=prob
+          end
+        end
+        
+        alias :exact_pdf :pdf 
+
+        
+      end
+    end
+  end
+end

data/lib/distribution/bivariatenormal.rb

@@ -1,5 +1,6 @@
 require 'distribution/bivariatenormal/ruby'
 require 'distribution/bivariatenormal/gsl'
+require 'distribution/bivariatenormal/java'
 module Distribution
   # Calculate pdf and cdf for bivariate normal distribution.
   #
@@ -12,14 +13,15 @@ module Distribution
     create_distribution_methods
     
     ##
-    # :singleton-method: pdf(x,y, rho, s1=1.0, s2=1.0)
-    # Probability density function for a given x, y and rho value.
+    # :singleton-method: pdf(k,n,prob)
+    # Probability density function for exactly +k+ successes in +n+ trials
+    # with success probability +prob+
     # 
     
     ##
-    # :singleton-method: cdf(x,y,rho)
-    # CDF for a given x, y and rho value.
-    #
+    # :singleton-method: cdf(k,n,prob)
+    # Cumulative density function for +k+ or less successes in +n+ trials
+    # with success probability +prob+
     
   end
 end

data/lib/distribution/bivariatenormal/java.rb

@@ -0,0 +1,9 @@
+module Distribution
+  module BivariateNormal
+    # TODO
+    module Java_
+      class << self
+      end
+    end
+  end
+end

data/lib/distribution/chisquare.rb

@@ -1,6 +1,7 @@
 require 'distribution/chisquare/ruby'
 require 'distribution/chisquare/gsl'
 require 'distribution/chisquare/statistics2'
+require 'distribution/chisquare/java'
 module Distribution
   # Calculate cdf and inverse cdf for Chi Square Distribution.
   # 

data/lib/distribution/chisquare/gsl.rb

@@ -6,12 +6,12 @@ module Distribution
         
       end
       def pdf(x,k)
-        GSL::Ran::chisq_pdf(x,k)
+        GSL::Ran::chisq_pdf(x.to_f,k.to_i)
       end
       # Return the P-value of the corresponding integral with 
       # k degrees of freedom
       def p_value(pr,k)
-        GSL::Cdf::chisq_Pinv(pr,k)
+        GSL::Cdf::chisq_Pinv(pr.to_f,k.to_i)
       end
       # Chi-square cumulative distribution function (cdf).
       # 
@@ -19,7 +19,7 @@ module Distribution
       # with k degrees of freedom over [0, x]
       # 
       def cdf(x, k)
-        GSL::Cdf::chisq_P(x,k)
+        GSL::Cdf::chisq_P(x.to_f,k.to_i)
       end
       end
     end