[section:f_dist F Distribution]

``#include <boost/math/distributions/fisher_f.hpp>``

   namespace boost{ namespace math{ 
      
   template <class RealType = double, 
             class ``__Policy``   = ``__policy_class`` >
   class fisher_f_distribution;
   
   typedef fisher_f_distribution<> fisher_f;

   template <class RealType, class ``__Policy``>
   class fisher_f_distribution
   {
   public:
      typedef RealType value_type;
      
      // Construct:
      fisher_f_distribution(const RealType& i, const RealType& j);
      
      // Accessors:
      RealType degrees_of_freedom1()const;
      RealType degrees_of_freedom2()const;
   };
   
   }} //namespaces

The F distribution is a continuous distribution that arises when testing
whether two samples have the same variance.  If [chi][super 2][sub m][space] and
[chi][super 2][sub n][space] are independent variates each distributed as 
Chi-Squared with /m/ and /n/ degrees of freedom, then the test statistic:

F[sub n,m][space] = ([chi][super 2][sub n][space] / n) / ([chi][super 2][sub m][space] / m)

Is distributed over the range \[0, [infin]\] with an F distribution, and
has the PDF:

[equation fisher_pdf]

The following graph illustrates how the PDF varies depending on the
two degrees of freedom parameters.

[graph fisher_f_pdf]


[h4 Member Functions]

      fisher_f_distribution(const RealType& df1, const RealType& df2);
      
Constructs an F-distribution with numerator degrees of freedom /df1/
and denominator degrees of freedom /df2/.

Requires that /df1/ and /df2/ are both greater than zero, otherwise __domain_error
is called.
      
      RealType degrees_of_freedom1()const;
      
Returns the numerator degrees of freedom parameter of the distribution.

      RealType degrees_of_freedom2()const;
      
Returns the denominator degrees of freedom parameter of the distribution.

[h4 Non-member Accessors]

All the [link math_toolkit.dist.dist_ref.nmp usual non-member accessor functions]
that are generic to all distributions are supported: __usual_accessors.

The domain of the random variable is \[0, +[infin]\].

[h4 Examples]

Various [link math_toolkit.dist.stat_tut.weg.f_eg worked examples] are 
available illustrating the use of the F Distribution.

[h4 Accuracy]

The normal distribution is implemented in terms of the 
[link math_toolkit.special.sf_beta.ibeta_function incomplete beta function]
and it's [link math_toolkit.special.sf_beta.ibeta_inv_function inverses], 
refer to those functions for accuracy data.

[h4 Implementation]

In the following table /v1/ and /v2/ are the first and second
degrees of freedom parameters of the distribution,
/x/ is the random variate, /p/ is the probability, and /q = 1-p/.

[table
[[Function][Implementation Notes]]
[[pdf][The usual form of the PDF is given by:

[equation fisher_pdf]

However, that form is hard to evaluate directly without incurring problems with
either accuracy or numeric overflow.

Direct differentiation of the CDF expressed in terms of the incomplete beta function

led to the following two formulas:

f[sub v1,v2](x) = y * __ibeta_derivative(v2 \/ 2, v1 \/ 2, v2 \/ (v2 + v1 * x))

with y = (v2 * v1) \/ ((v2 + v1 * x) * (v2 + v1 * x))

and

f[sub v1,v2](x) = y * __ibeta_derivative(v1 \/ 2, v2 \/ 2, v1 * x \/ (v2 + v1 * x))

with y = (z * v1 - x * v1 * v1) \/ z[super 2]

and z = v2 + v1 * x

The first of these is used for v1 * x > v2, otherwise the second is used.

The aim is to keep the /x/ argument to __ibeta_derivative away from 1 to avoid
rounding error. ]]
[[cdf][Using the relations:

p = __ibeta(v1 \/ 2, v2 \/ 2, v1 * x \/ (v2 + v1 * x))

and

p = __ibetac(v2 \/ 2, v1 \/ 2, v2 \/ (v2 + v1 * x))

The first is used for v1 * x > v2, otherwise the second is used.

The aim is to keep the /x/ argument to __ibeta well away from 1 to
avoid rounding error. ]]

[[cdf complement][Using the relations:

p = __ibetac(v1 \/ 2, v2 \/ 2, v1 * x \/ (v2 + v1 * x))

and

p = __ibeta(v2 \/ 2, v1 \/ 2, v2 \/ (v2 + v1 * x))

The first is used for v1 * x < v2, otherwise the second is used.

The aim is to keep the /x/ argument to __ibeta well away from 1 to
avoid rounding error. ]]
[[quantile][Using the relation: 

x = v2 * a \/ (v1 * b)

where:

a = __ibeta_inv(v1 \/ 2, v2 \/ 2, p)

and

b = 1 - a

Quantities /a/ and /b/ are both computed by __ibeta_inv without the
subtraction implied above.]]
[[quantile

from the complement][Using the relation:

x = v2 * a \/ (v1 * b)

where

a = __ibetac_inv(v1 \/ 2, v2 \/ 2, p)

and

b = 1 - a

Quantities /a/ and /b/ are both computed by __ibetac_inv without the
subtraction implied above.]]
[[mean][v2 \/ (v2 - 2)]]
[[variance][2 * v2[super 2 ] * (v1 + v2 - 2) \/ (v1 * (v2 - 2) * (v2 - 2) * (v2 - 4))]]
[[mode][v2 * (v1 - 2) \/ (v1 * (v2 + 2))]]
[[skewness][2 * (v2 + 2 * v1 - 2) * sqrt((2 * v2 - 8) \/ (v1 * (v2 + v1 - 2))) \/ (v2 - 6)]]
[[kurtosis and kurtosis excess]
    [Refer to, [@http://mathworld.wolfram.com/F-Distribution.html
    Weisstein, Eric W. "F-Distribution." From MathWorld--A Wolfram Web Resource.]  ]]
]

[endsect][/section:f_dist F distribution]

[/ fisher.qbk
  Copyright 2006 John Maddock and Paul A. Bristow.
  Distributed under the Boost Software License, Version 1.0.
  (See accompanying file LICENSE_1_0.txt or copy at
  http://www.boost.org/LICENSE_1_0.txt).
]