1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
|
[section:inverse_chi_squared_dist Inverse Chi Squared Distribution]
``#include <boost/math/distributions/inverse_chi_squared.hpp>``
namespace boost{ namespace math{
template <class RealType = double,
class ``__Policy`` = ``__policy_class`` >
class inverse_chi_squared_distribution
{
public:
typedef RealType value_type;
typedef Policy policy_type;
BOOST_MATH_GPU_ENABLED inverse_chi_squared_distribution(RealType df = 1); // Not explicitly scaled, default 1/df.
BOOST_MATH_GPU_ENABLED inverse_chi_squared_distribution(RealType df, RealType scale = 1/df); // Scaled.
BOOST_MATH_GPU_ENABLED RealType degrees_of_freedom()const; // Default 1.
BOOST_MATH_GPU_ENABLED RealType scale()const; // Optional scale [xi] (variance), default 1/degrees_of_freedom.
};
}} // namespace boost // namespace math
The inverse chi squared distribution is a continuous probability distribution
of the *reciprocal* of a variable distributed according to the chi squared distribution.
The sources below give confusingly different formulae
using different symbols for the distribution pdf,
but they are all the same, or related by a change of variable, or choice of scale.
Two constructors are available to implement both the scaled and (implicitly) unscaled versions.
The main version has an explicit scale parameter which implements the
[@http://en.wikipedia.org/wiki/Scaled-inverse-chi-square_distribution scaled inverse chi_squared distribution].
A second version has an implicit scale = 1/degrees of freedom and gives the 1st definition in the
[@http://en.wikipedia.org/wiki/Inverse-chi-square_distribution Wikipedia inverse chi_squared distribution].
The 2nd Wikipedia inverse chi_squared distribution definition can be implemented
by explicitly specifying a scale = 1.
Both definitions are also available in __Mathematica and in __R (geoR) with default scale = 1/degrees of freedom.
See
* Inverse chi_squared distribution [@http://en.wikipedia.org/wiki/Inverse-chi-square_distribution]
* Scaled inverse chi_squared distribution[@http://en.wikipedia.org/wiki/Scaled-inverse-chi-square_distribution]
* R inverse chi_squared distribution functions [@http://hosho.ees.hokudai.ac.jp/~kubo/Rdoc/library/geoR/html/InvChisquare.html R ]
* Inverse chi_squared distribution functions [@http://mathworld.wolfram.com/InverseChi-SquaredDistribution.html Weisstein, Eric W. "Inverse Chi-Squared Distribution." From MathWorld--A Wolfram Web Resource.]
* Inverse chi_squared distribution reference [@http://reference.wolfram.com/mathematica/ref/InverseChiSquareDistribution.html Weisstein, Eric W. "Inverse Chi-Squared Distribution reference." From Wolfram Mathematica.]
The inverse_chi_squared distribution is used in
[@http://en.wikipedia.org/wiki/Bayesian_statistics Bayesian statistics]:
the scaled inverse chi-square is conjugate prior for the normal distribution
with known mean, model parameter [sigma][pow2] (variance).
See [@http://en.wikipedia.org/wiki/Conjugate_prior conjugate priors including a table of distributions and their priors.]
See also __inverse_gamma_distrib and __chi_squared_distrib.
The inverse_chi_squared distribution is a special case of a inverse_gamma distribution
with [nu] (degrees_of_freedom) shape ([alpha]) and scale ([beta]) where
[expression [alpha]= [nu] /2 and [beta] = [frac12]]
[note This distribution *does* provide the typedef:
``typedef inverse_chi_squared_distribution<double> inverse_chi_squared;``
If you want a `double` precision inverse_chi_squared distribution you can use
``boost::math::inverse_chi_squared_distribution<>``
or you can write `inverse_chi_squared my_invchisqr(2, 3);`]
For degrees of freedom parameter [nu],
the (*unscaled*) inverse chi_squared distribution is defined by the probability density function (PDF):
[expression f(x;[nu]) = 2[super -[nu]/2] x[super -[nu]/2-1] e[super -1/2x] / [Gamma]([nu]/2)]
and Cumulative Density Function (CDF)
[expression F(x;[nu]) = [Gamma]([nu]/2, 1/2x) / [Gamma]([nu]/2)]
For degrees of freedom parameter [nu] and scale parameter [xi],
the *scaled* inverse chi_squared distribution is defined by the probability density function (PDF):
[expression f(x;[nu], [xi]) = ([xi][nu]/2)[super [nu]/2] e[super -[nu][xi]/2x] x[super -1-[nu]/2] / [Gamma]([nu]/2)]
and Cumulative Density Function (CDF)
[expression F(x;[nu], [xi]) = [Gamma]([nu]/2, [nu][xi]/2x) / [Gamma]([nu]/2)]
The following graphs illustrate how the PDF and CDF of the inverse chi_squared distribution
varies for a few values of parameters [nu] and [xi]:
[graph inverse_chi_squared_pdf] [/.png or .svg]
[graph inverse_chi_squared_cdf]
[h4 Member Functions]
BOOST_MATH_GPU_ENABLED inverse_chi_squared_distribution(RealType df = 1); // Implicitly scaled 1/df.
BOOST_MATH_GPU_ENABLED inverse_chi_squared_distribution(RealType df = 1, RealType scale); // Explicitly scaled.
Constructs an inverse chi_squared distribution with [nu] degrees of freedom ['df],
and scale ['scale] with default value 1\/df.
Requires that the degrees of freedom [nu] parameter is greater than zero, otherwise calls
__domain_error.
BOOST_MATH_GPU_ENABLED RealType degrees_of_freedom()const;
Returns the degrees_of_freedom [nu] parameter of this distribution.
BOOST_MATH_GPU_ENABLED RealType scale()const;
Returns the scale [xi] parameter of this distribution.
[h4 Non-member Accessors]
All the [link math_toolkit.dist_ref.nmp usual non-member accessor functions] that are generic to all
distributions are supported: __usual_accessors.
For this distribution all non-member accessor functions are marked with `BOOST_MATH_GPU_ENABLED` and can
be run on both host and device.
The domain of the random variate is \[0,+[infin]\].
[note Unlike some definitions, this implementation supports a random variate
equal to zero as a special case, returning zero for both pdf and cdf.]
[h4 Accuracy]
The inverse gamma distribution is implemented in terms of the
incomplete gamma functions like the __inverse_gamma_distrib that use
__gamma_p and __gamma_q and their inverses __gamma_p_inv and __gamma_q_inv:
refer to the accuracy data for those functions for more information.
But in general, gamma (and thus inverse gamma) results are often accurate to a few epsilon,
>14 decimal digits accuracy for 64-bit double.
unless iteration is involved, as for the estimation of degrees of freedom.
[h4 Implementation]
In the following table [nu] is the degrees of freedom parameter and
[xi] is the scale parameter of the distribution,
/x/ is the random variate, /p/ is the probability and /q = 1-p/ its complement.
Parameters [alpha] for shape and [beta] for scale
are used for the inverse gamma function: [alpha] = [nu]/2 and [beta] = [nu] * [xi]/2.
[table
[[Function][Implementation Notes]]
[[pdf][Using the relation: pdf = __gamma_p_derivative([alpha], [beta]/ x, [beta]) / x * x ]]
[[cdf][Using the relation: p = __gamma_q([alpha], [beta] / x) ]]
[[cdf complement][Using the relation: q = __gamma_p([alpha], [beta] / x) ]]
[[quantile][Using the relation: x = [beta]/ __gamma_q_inv([alpha], p) ]]
[[quantile from the complement][Using the relation: x = [alpha]/ __gamma_p_inv([alpha], q) ]]
[[mode][[nu] * [xi] / ([nu] + 2) ]]
[[median][no closed form analytic equation is known, but is evaluated as quantile(0.5)]]
[[mean][[nu][xi] / ([nu] - 2) for [nu] > 2, else a __domain_error]]
[[variance][2 [nu][pow2] [xi][pow2] / (([nu] -2)[pow2] ([nu] -4)) for [nu] >4, else a __domain_error]]
[[skewness][4 [sqrt]2 [sqrt]([nu]-4) /([nu]-6) for [nu] >6, else a __domain_error ]]
[[kurtosis_excess][12 * (5[nu] - 22) / (([nu] - 6) * ([nu] - 8)) for [nu] >8, else a __domain_error]]
[[kurtosis][3 + 12 * (5[nu] - 22) / (([nu] - 6) * ([nu]-8)) for [nu] >8, else a __domain_error]]
] [/table]
[h4 References]
# Bayesian Data Analysis, Andrew Gelman, John B. Carlin, Hal S. Stern, Donald B. Rubin,
ISBN-13: 978-1584883883, Chapman & Hall; 2 edition (29 July 2003).
# Bayesian Computation with R, Jim Albert, ISBN-13: 978-0387922973, Springer; 2nd ed. edition (10 Jun 2009)
[endsect] [/section:inverse_chi_squared_dist Inverse chi_squared Distribution]
[/
Copyright 2010 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).
]
|
