• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1[section:igamma_inv Incomplete Gamma Function Inverses]
2
3[h4 Synopsis]
4
5``
6#include <boost/math/special_functions/gamma.hpp>
7``
8
9   namespace boost{ namespace math{
10
11   template <class T1, class T2>
12   ``__sf_result`` gamma_q_inv(T1 a, T2 q);
13
14   template <class T1, class T2, class ``__Policy``>
15   ``__sf_result`` gamma_q_inv(T1 a, T2 q, const ``__Policy``&);
16
17   template <class T1, class T2>
18   ``__sf_result`` gamma_p_inv(T1 a, T2 p);
19
20   template <class T1, class T2, class ``__Policy``>
21   ``__sf_result`` gamma_p_inv(T1 a, T2 p, const ``__Policy``&);
22
23   template <class T1, class T2>
24   ``__sf_result`` gamma_q_inva(T1 x, T2 q);
25
26   template <class T1, class T2, class ``__Policy``>
27   ``__sf_result`` gamma_q_inva(T1 x, T2 q, const ``__Policy``&);
28
29   template <class T1, class T2>
30   ``__sf_result`` gamma_p_inva(T1 x, T2 p);
31
32   template <class T1, class T2, class ``__Policy``>
33   ``__sf_result`` gamma_p_inva(T1 x, T2 p, const ``__Policy``&);
34
35   }} // namespaces
36
37[h4 Description]
38
39There are four [@http://mathworld.wolfram.com/IncompleteGammaFunction.html incomplete gamma function]
40inverses which either compute
41/x/ given /a/ and /p/ or /q/,
42or else compute /a/ given /x/ and either /p/ or /q/.
43
44The return type of these functions is computed using the __arg_promotion_rules
45when T1 and T2 are different types, otherwise the return type is simply T1.
46
47[optional_policy]
48
49[tip When people normally talk about the inverse of the incomplete
50gamma function, they are talking about inverting on parameter /x/.
51These are implemented here as `gamma_p_inv` and `gamma_q_inv`, and are by
52far the most efficient of the inverses presented here.
53
54The inverse on the /a/ parameter finds use in some statistical
55applications but has to be computed by rather brute force numerical
56techniques and is consequently several times slower.
57These are implemented here as `gamma_p_inva` and `gamma_q_inva`.]
58
59
60   template <class T1, class T2>
61   ``__sf_result`` gamma_q_inv(T1 a, T2 q);
62
63   template <class T1, class T2, class ``__Policy``>
64   ``__sf_result`` gamma_q_inv(T1 a, T2 q, const ``__Policy``&);
65
66Returns a value x such that: `q = gamma_q(a, x);`
67
68Requires: /a > 0/ and /1 >= p,q >= 0/.
69
70   template <class T1, class T2>
71   ``__sf_result`` gamma_p_inv(T1 a, T2 p);
72
73   template <class T1, class T2, class ``__Policy``>
74   ``__sf_result`` gamma_p_inv(T1 a, T2 p, const ``__Policy``&);
75
76Returns a value x such that: `p = gamma_p(a, x);`
77
78Requires: /a > 0/ and /1 >= p,q >= 0/.
79
80   template <class T1, class T2>
81   ``__sf_result`` gamma_q_inva(T1 x, T2 q);
82
83   template <class T1, class T2, class ``__Policy``>
84   ``__sf_result`` gamma_q_inva(T1 x, T2 q, const ``__Policy``&);
85
86Returns a value a such that: `q = gamma_q(a, x);`
87
88Requires: /x > 0/ and /1 >= p,q >= 0/.
89
90   template <class T1, class T2>
91   ``__sf_result`` gamma_p_inva(T1 x, T2 p);
92
93   template <class T1, class T2, class ``__Policy``>
94   ``__sf_result`` gamma_p_inva(T1 x, T2 p, const ``__Policy``&);
95
96Returns a value a such that: `p = gamma_p(a, x);`
97
98Requires: /x > 0/ and /1 >= p,q >= 0/.
99
100[h4 Accuracy]
101
102The accuracy of these functions doesn't vary much by platform or by
103the type T.  Given that these functions are computed by iterative methods,
104they are deliberately "detuned" so as not to be too accurate: it is in
105any case impossible for these function to be more accurate than the
106regular forward incomplete gamma functions.  In practice, the accuracy
107of these functions is very similar to that of __gamma_p and __gamma_q
108functions:
109
110[table_gamma_p_inv]
111
112[table_gamma_q_inv]
113
114[table_gamma_p_inva]
115
116[table_gamma_q_inva]
117
118[h4 Testing]
119
120There are two sets of tests:
121
122* Basic sanity checks attempt to "round-trip" from
123/a/ and /x/ to /p/ or /q/ and back again.  These tests have quite
124generous tolerances: in general both the incomplete gamma, and its
125inverses, change so rapidly that round tripping to more than a couple
126of significant digits isn't possible.  This is especially true when
127/p/ or /q/ is very near one: in this case there isn't enough
128"information content" in the input to the inverse function to get
129back where you started.
130* Accuracy checks using high precision test values.  These measure
131the accuracy of the result, given exact input values.
132
133[h4 Implementation]
134
135The functions `gamma_p_inv` and [@http://functions.wolfram.com/GammaBetaErf/InverseGammaRegularized/ `gamma_q_inv`]
136share a common implementation.
137
138First an initial approximation is computed using the methodology described
139in:
140
141[@http://portal.acm.org/citation.cfm?id=23109&coll=portal&dl=ACM
142A. R. Didonato and A. H. Morris, Computation of the Incomplete Gamma
143Function Ratios and their Inverse, ACM Trans. Math. Software 12 (1986), 377-393.]
144
145Finally, the last few bits are cleaned up using Halley iteration, the iteration
146limit is set to 2/3 of the number of bits in T, which by experiment is
147sufficient to ensure that the inverses are at least as accurate as the normal
148incomplete gamma functions.  In testing, no more than 3 iterations are required
149to produce a result as accurate as the forward incomplete gamma function, and
150in many cases only one iteration is required.
151
152The functions `gamma_p_inva` and `gamma_q_inva` also share a common implementation
153but are handled separately from `gamma_p_inv` and `gamma_q_inv`.
154
155An initial approximation for /a/ is computed very crudely so that
156/gamma_p(a, x) ~ 0.5/, this value is then used as a starting point
157for a generic derivative-free root finding algorithm.  As a consequence,
158these two functions are rather more expensive to compute than the
159`gamma_p_inv` or `gamma_q_inv` functions.  Even so, the root is usually found
160in fewer than 10 iterations.
161
162[endsect] [/section The Incomplete Gamma Function Inverses]
163
164[/
165  Copyright 2006 John Maddock and Paul A. Bristow.
166  Distributed under the Boost Software License, Version 1.0.
167  (See accompanying file LICENSE_1_0.txt or copy at
168  http://www.boost.org/LICENSE_1_0.txt).
169]
170