Automatic Differentiation
 
Loading...
Searching...
No Matches
std_normal_lcdf.hpp
Go to the documentation of this file.
1#ifndef STAN_MATH_PRIM_PROB_STD_NORMAL_LCDF_HPP
2#define STAN_MATH_PRIM_PROB_STD_NORMAL_LCDF_HPP
3
18#include <cmath>
19#include <limits>
20
21namespace stan {
22namespace math {
23
24template <
25 typename T_y,
26 require_all_not_nonscalar_prim_or_rev_kernel_expression_t<T_y>* = nullptr>
27inline return_type_t<T_y> std_normal_lcdf(const T_y& y) {
28 using T_partials_return = partials_return_t<T_y>;
29 using std::exp;
30 using std::fabs;
31 using std::log;
32 using std::pow;
33 using T_y_ref = ref_type_t<T_y>;
34 static constexpr const char* function = "std_normal_lcdf";
35 T_y_ref y_ref = y;
36 check_not_nan(function, "Random variable", y_ref);
37
38 if (size_zero(y)) {
39 return 0;
40 }
41
42 T_partials_return lcdf(0.0);
43 auto ops_partials = make_partials_propagator(y_ref);
44
45 scalar_seq_view<T_y_ref> y_vec(y_ref);
46 size_t N = stan::math::size(y);
47
48 for (size_t n = 0; n < N; n++) {
49 const T_partials_return y_dbl = y_vec.val(n);
50 const T_partials_return scaled_y = y_dbl * INV_SQRT_TWO;
51 const T_partials_return x2 = square(scaled_y);
52
53 // Rigorous numerical approximations are applied here to deal with values
54 // of |scaled_y|>>0. This is needed to deal with rare base-rate
55 // logistic regression problems where it is useful to use an alternative
56 // link function instead.
57 //
58 // use erfc() instead of erf() in order to retain precision
59 // since for x>0 erfc()->0
60 if (scaled_y > 0.0) {
61 // CDF(x) = 1/2 + 1/2erf(x) = 1 - 1/2erfc(x)
62 lcdf += log1p(-0.5 * erfc(scaled_y));
63 if (!is_not_nan(lcdf)) {
64 lcdf = 0;
65 }
66 } else if (scaled_y > -20.0) {
67 // CDF(x) = 1/2 - 1/2erf(-x) = 1/2erfc(-x)
68 lcdf += log(erfc(-scaled_y)) + LOG_HALF;
69 } else if (10.0 * log(fabs(scaled_y))
70 < log(std::numeric_limits<T_partials_return>::max())) {
71 // entering territory where erfc(-x)~0
72 // need to use direct numerical approximation of lcdf instead
73 // the following based on W. J. Cody, Math. Comp. 23(107):631-638 (1969)
74 // CDF(x) = 1/2erfc(-x)
75 const T_partials_return x4 = pow(scaled_y, 4);
76 const T_partials_return x6 = pow(scaled_y, 6);
77 const T_partials_return x8 = pow(scaled_y, 8);
78 const T_partials_return x10 = pow(scaled_y, 10);
79 const T_partials_return temp_p
80 = 0.000658749161529837803157 + 0.0160837851487422766278 / x2
81 + 0.125781726111229246204 / x4 + 0.360344899949804439429 / x6
82 + 0.305326634961232344035 / x8 + 0.0163153871373020978498 / x10;
83 const T_partials_return temp_q
84 = -0.00233520497626869185443 - 0.0605183413124413191178 / x2
85 - 0.527905102951428412248 / x4 - 1.87295284992346047209 / x6
86 - 2.56852019228982242072 / x8 - 1.0 / x10;
87 lcdf += LOG_HALF + log(INV_SQRT_PI + (temp_p / temp_q) / x2)
88 - log(-scaled_y) - x2;
89 } else {
90 // scaled_y^10 term will overflow
92 }
93
95 // compute partial derivatives
96 // based on analytic form given by:
97 // dln(CDF)/dx = exp(-x^2)/(sqrt(pi)*(1/2+erf(x)/2)
98 T_partials_return dnlcdf = 0.0;
99 T_partials_return t = 0.0;
100 T_partials_return t2 = 0.0;
101 T_partials_return t4 = 0.0;
102
103 // calculate using piecewise function
104 // (due to instability / inaccuracy in the various approximations)
105 if (scaled_y > 2.9) {
106 // approximation derived from Abramowitz and Stegun (1964) 7.1.26
107 t = 1.0 / (1.0 + 0.3275911 * scaled_y);
108 t2 = square(t);
109 t4 = pow(t, 4);
110 dnlcdf = INV_SQRT_PI
111 / (exp(x2) - 0.254829592 + 0.284496736 * t - 1.421413741 * t2
112 + 1.453152027 * t2 * t - 1.061405429 * t4);
113 } else if (scaled_y > 2.5) {
114 // in the trouble area where all of the standard numerical
115 // approximations are unstable - bridge the gap using Taylor
116 // expansions of the analytic function
117 // use Taylor expansion centred around x=2.7
118 t = scaled_y - 2.7;
119 t2 = square(t);
120 t4 = pow(t, 4);
121 dnlcdf = 0.0003849882382 - 0.002079084702 * t + 0.005229340880 * t2
122 - 0.008029540137 * t2 * t + 0.008232190507 * t4
123 - 0.005692364250 * t4 * t + 0.002399496363 * pow(t, 6);
124 } else if (scaled_y > 2.1) {
125 // use Taylor expansion centred around x=2.3
126 t = scaled_y - 2.3;
127 t2 = square(t);
128 t4 = pow(t, 4);
129 dnlcdf = 0.002846135439 - 0.01310032351 * t + 0.02732189391 * t2
130 - 0.03326906904 * t2 * t + 0.02482478940 * t4
131 - 0.009883071924 * t4 * t - 0.0002771362254 * pow(t, 6);
132 } else if (scaled_y > 1.5) {
133 // use Taylor expansion centred around x=1.85
134 t = scaled_y - 1.85;
135 t2 = square(t);
136 t4 = pow(t, 4);
137 dnlcdf = 0.01849212058 - 0.06876280470 * t + 0.1099906382 * t2
138 - 0.09274533184 * t2 * t + 0.03543327418 * t4
139 + 0.005644855518 * t4 * t - 0.01111434424 * pow(t, 6);
140 } else if (scaled_y > 0.8) {
141 // use Taylor expansion centred around x=1.15
142 t = scaled_y - 1.15;
143 t2 = square(t);
144 t4 = pow(t, 4);
145 dnlcdf = 0.1585747034 - 0.3898677543 * t + 0.3515963775 * t2
146 - 0.09748053605 * t2 * t - 0.04347986191 * t4
147 + 0.02182506378 * t4 * t + 0.01074751427 * pow(t, 6);
148 } else if (scaled_y > 0.1) {
149 // use Taylor expansion centred around x=0.45
150 t = scaled_y - 0.45;
151 t2 = square(t);
152 t4 = pow(t, 4);
153 dnlcdf = 0.6245634904 - 0.9521866949 * t + 0.3986215682 * t2
154 + 0.04700850676 * t2 * t - 0.03478651979 * t4
155 - 0.01772675404 * t4 * t + 0.0006577254811 * pow(t, 6);
156 } else if (10.0 * log(fabs(scaled_y))
157 < log(std::numeric_limits<T_partials_return>::max())) {
158 // approximation derived from Abramowitz and Stegun (1964) 7.1.26
159 // use fact that erf(x)=-erf(-x)
160 // Abramowitz and Stegun define this for -inf<x<0 but seems to be
161 // accurate for -inf<x<0.1
162 t = 1.0 / (1.0 - 0.3275911 * scaled_y);
163 t2 = square(t);
164 t4 = pow(t, 4);
165 dnlcdf = 2.0 * INV_SQRT_PI
166 / (0.254829592 * t - 0.284496736 * t2 + 1.421413741 * t2 * t
167 - 1.453152027 * t4 + 1.061405429 * t4 * t);
168 // check if we need to add a correction term
169 // (from cubic fit of residuals)
170 if (scaled_y < -29.0) {
171 dnlcdf += 0.0015065154280332 * x2 - 0.3993154819705530 * scaled_y
172 - 4.2919418242931700;
173 } else if (scaled_y < -17.0) {
174 dnlcdf += 0.0001263257217272 * x2 * scaled_y + 0.0123586859488623 * x2
175 - 0.0860505264736028 * scaled_y - 1.252783383752970;
176 } else if (scaled_y < -7.0) {
177 dnlcdf += 0.000471585349920831 * x2 * scaled_y
178 + 0.0296839305424034 * x2 + 0.207402143352332 * scaled_y
179 + 0.425316974683324;
180 } else if (scaled_y < -3.9) {
181 dnlcdf += -0.0006972280656443 * x2 * scaled_y
182 + 0.0068218494628567 * x2 + 0.0585761964460277 * scaled_y
183 + 0.1034397670201370;
184 } else if (scaled_y < -2.1) {
185 dnlcdf += -0.0018742199480885 * x2 * scaled_y
186 - 0.0097119598291202 * x2 - 0.0170137970924080 * scaled_y
187 - 0.0100428567412041;
188 }
189 } else {
191 }
192
194 partials<0>(ops_partials)[n] += dnlcdf * INV_SQRT_TWO;
195 }
196 }
197 }
198
199 return ops_partials.build(lcdf);
200}
201
202} // namespace math
203} // namespace stan
204#endif
scalar_seq_view provides a uniform sequence-like wrapper around either a scalar or a sequence of scal...
return_type_t< T_y_cl > std_normal_lcdf(const T_y_cl &y)
Returns the log standard normal complementary cumulative distribution function.
typename return_type< Ts... >::type return_type_t
Convenience type for the return type of the specified template parameters.
int64_t size(const T &m)
Returns the size (number of the elements) of a matrix_cl or var_value<matrix_cl<T>>.
Definition size.hpp:19
static constexpr double negative_infinity()
Return negative infinity.
static constexpr double LOG_HALF
The natural logarithm of 0.5, .
Definition constants.hpp:92
static constexpr double positive_infinity()
Return positive infinity.
bool size_zero(const T &x)
Returns 1 if input is of length 0, returns 0 otherwise.
Definition size_zero.hpp:19
auto pow(const T1 &x1, const T2 &x2)
Definition pow.hpp:32
fvar< T > log(const fvar< T > &x)
Definition log.hpp:18
bool is_not_nan(const T_y &y)
Return true if y is not NaN.
static constexpr double INV_SQRT_TWO
The value of 1 over the square root of 2, .
fvar< T > erfc(const fvar< T > &x)
Definition erfc.hpp:16
fvar< T > log1p(const fvar< T > &x)
Definition log1p.hpp:12
void check_not_nan(const char *function, const char *name, const T_y &y)
Check if y is not NaN.
static constexpr double INV_SQRT_PI
The value of 1 over the square root of , .
auto make_partials_propagator(Ops &&... ops)
Construct an partials_propagator.
fvar< T > fabs(const fvar< T > &x)
Definition fabs.hpp:16
fvar< T > square(const fvar< T > &x)
Definition square.hpp:12
fvar< T > exp(const fvar< T > &x)
Definition exp.hpp:15
typename ref_type_if< true, T >::type ref_type_t
Definition ref_type.hpp:55
typename partials_return_type< Args... >::type partials_return_t
The lgamma implementation in stan-math is based on either the reentrant safe lgamma_r implementation ...
Extends std::true_type when instantiated with zero or more template parameters, all of which extend t...