• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1[section:error_inv Error Function Inverses]
2
3[h4 Synopsis]
4
5``
6#include <boost/math/special_functions/erf.hpp>
7``
8
9   namespace boost{ namespace math{
10
11   template <class T>
12   ``__sf_result`` erf_inv(T p);
13
14   template <class T, class ``__Policy``>
15   ``__sf_result`` erf_inv(T p, const ``__Policy``&);
16
17   template <class T>
18   ``__sf_result`` erfc_inv(T p);
19
20   template <class T, class ``__Policy``>
21   ``__sf_result`` erfc_inv(T p, const ``__Policy``&);
22
23   }} // namespaces
24
25The return type of these functions is computed using the __arg_promotion_rules:
26the return type is `double` if T is an integer type, and T otherwise.
27
28[optional_policy]
29
30[h4 Description]
31
32   template <class T>
33   ``__sf_result`` erf_inv(T z);
34
35   template <class T, class ``__Policy``>
36   ``__sf_result`` erf_inv(T z, const ``__Policy``&);
37
38Returns the [@http://functions.wolfram.com/GammaBetaErf/InverseErf/ inverse error function]
39of z, that is a value x such that:
40
41[expression ['p = erf(x);]]
42
43[graph erf_inv]
44
45   template <class T>
46   ``__sf_result`` erfc_inv(T z);
47
48   template <class T, class ``__Policy``>
49   ``__sf_result`` erfc_inv(T z, const ``__Policy``&);
50
51Returns the inverse of the complement of the error function of z, that is a
52value x such that:
53
54[expression ['p = erfc(x);]]
55
56[graph erfc_inv]
57
58[h4 Accuracy]
59
60For types up to and including 80-bit long doubles the approximations used
61are accurate to less than ~ 2 epsilon.  For higher precision types these
62functions have the same accuracy as the
63[link math_toolkit.sf_erf.error_function forward error functions].
64
65[table_erf_inv]
66
67[table_erfc_inv]
68
69The following error plot are based on an exhaustive search of the functions domain, MSVC-15.5 at `double` precision,
70and GCC-7.1/Ubuntu for `long double` and `__float128`.
71
72[graph erfc__double]
73
74[graph erfc__80_bit_long_double]
75
76[graph erfc____float128]
77
78[h4 Testing]
79
80There are two sets of tests:
81
82* Basic sanity checks attempt to "round-trip" from
83/x/ to /p/ and back again.  These tests have quite
84generous tolerances: in general both the error functions and their
85inverses change so rapidly in some places that round tripping to more than a couple
86of significant digits isn't possible.  This is especially true when
87/p/ is very near one: in this case there isn't enough
88"information content" in the input to the inverse function to get
89back where you started.
90* Accuracy checks using high-precision test values.  These measure
91the accuracy of the result, given /exact/ input values.
92
93[h4 Implementation]
94
95These functions use a rational approximation [jm_rationals]
96to calculate an initial
97approximation to the result that is accurate to ~10[super -19],
98then only if that has insufficient accuracy compared to the epsilon for T,
99do we clean up the result using
100[@http://en.wikipedia.org/wiki/Simple_rational_approximation Halley iteration].
101
102Constructing rational approximations to the erf/erfc functions is actually
103surprisingly hard, especially at high precision.  For this reason no attempt
104has been made to achieve 10[super -34 ] accuracy suitable for use with 128-bit
105reals.
106
107In the following discussion, /p/ is the value passed to erf_inv, and /q/ is
108the value passed to erfc_inv, so that /p = 1 - q/ and /q = 1 - p/ and in both
109cases we want to solve for the same result /x/.
110
111For /p < 0.5/ the inverse erf function is reasonably smooth and the approximation:
112
113[expression ['x = p(p + 10)(Y + R(p))]]
114
115Gives a good result for a constant Y, and R(p) optimised for low absolute error
116compared to |Y|.
117
118For q < 0.5 things get trickier, over the interval /0.5 > q > 0.25/
119the following approximation works well:
120
121[expression ['x = sqrt(-2log(q)) / (Y + R(q))]]
122
123While for q < 0.25, let
124
125[expression ['z = sqrt(-log(q))]]
126
127Then the result is given by:
128
129[expression ['x = z(Y + R(z - B))]]
130
131As before Y is a constant and the rational function R is optimised for low
132absolute error compared to |Y|.  B is also a constant: it is the smallest value
133of /z/ for which each approximation is valid.  There are several approximations
134of this form each of which reaches a little further into the tail of the erfc
135function (at `long double` precision the extended exponent range compared to
136`double` means that the tail goes on for a very long way indeed).
137
138[endsect] [/ :error_inv The Error Function Inverses]
139
140[/
141  Copyright 2006 John Maddock and Paul A. Bristow.
142  Distributed under the Boost Software License, Version 1.0.
143  (See accompanying file LICENSE_1_0.txt or copy at
144  http://www.boost.org/LICENSE_1_0.txt).
145]
146