| 1 | // (C) Copyright John Maddock 2006. |
| 2 | // Use, modification and distribution are subject to the |
| 3 | // Boost Software License, Version 1.0. (See accompanying file |
| 4 | // LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) |
| 5 | |
| 6 | #ifndef BOOST_MATH_SF_DIGAMMA_HPP |
| 7 | #define BOOST_MATH_SF_DIGAMMA_HPP |
| 8 | |
| 9 | #ifdef _MSC_VER |
| 10 | #pragma once |
| 11 | #pragma warning(push) |
| 12 | #pragma warning(disable:4702) // Unreachable code (release mode only warning) |
| 13 | #endif |
| 14 | |
| 15 | #include <boost/math/special_functions/math_fwd.hpp> |
| 16 | #include <boost/math/tools/rational.hpp> |
| 17 | #include <boost/math/tools/series.hpp> |
| 18 | #include <boost/math/tools/promotion.hpp> |
| 19 | #include <boost/math/policies/error_handling.hpp> |
| 20 | #include <boost/math/constants/constants.hpp> |
| 21 | #include <boost/mpl/comparison.hpp> |
| 22 | #include <boost/math/tools/big_constant.hpp> |
| 23 | |
| 24 | #if defined(__GNUC__) && defined(BOOST_MATH_USE_FLOAT128) |
| 25 | // |
| 26 | // This is the only way we can avoid |
| 27 | // warning: non-standard suffix on floating constant [-Wpedantic] |
| 28 | // when building with -Wall -pedantic. Neither __extension__ |
| 29 | // nor #pragma diagnostic ignored work :( |
| 30 | // |
| 31 | #pragma GCC system_header |
| 32 | #endif |
| 33 | |
| 34 | namespace boost{ |
| 35 | namespace math{ |
| 36 | namespace detail{ |
| 37 | // |
| 38 | // Begin by defining the smallest value for which it is safe to |
| 39 | // use the asymptotic expansion for digamma: |
| 40 | // |
| 41 | inline unsigned digamma_large_lim(const boost::integral_constant<int, 0>*) |
| 42 | { return 20; } |
| 43 | inline unsigned digamma_large_lim(const boost::integral_constant<int, 113>*) |
| 44 | { return 20; } |
| 45 | inline unsigned digamma_large_lim(const void*) |
| 46 | { return 10; } |
| 47 | // |
| 48 | // Implementations of the asymptotic expansion come next, |
| 49 | // the coefficients of the series have been evaluated |
| 50 | // in advance at high precision, and the series truncated |
| 51 | // at the first term that's too small to effect the result. |
| 52 | // Note that the series becomes divergent after a while |
| 53 | // so truncation is very important. |
| 54 | // |
| 55 | // This first one gives 34-digit precision for x >= 20: |
| 56 | // |
| 57 | template <class T> |
| 58 | inline T digamma_imp_large(T x, const boost::integral_constant<int, 113>*) |
| 59 | { |
| 60 | BOOST_MATH_STD_USING // ADL of std functions. |
| 61 | static const T P[] = { |
| 62 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.083333333333333333333333333333333333333333333333333), |
| 63 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.0083333333333333333333333333333333333333333333333333), |
| 64 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.003968253968253968253968253968253968253968253968254), |
| 65 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.0041666666666666666666666666666666666666666666666667), |
| 66 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.0075757575757575757575757575757575757575757575757576), |
| 67 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.021092796092796092796092796092796092796092796092796), |
| 68 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.083333333333333333333333333333333333333333333333333), |
| 69 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.44325980392156862745098039215686274509803921568627), |
| 70 | BOOST_MATH_BIG_CONSTANT(T, 113, 3.0539543302701197438039543302701197438039543302701), |
| 71 | BOOST_MATH_BIG_CONSTANT(T, 113, -26.456212121212121212121212121212121212121212121212), |
| 72 | BOOST_MATH_BIG_CONSTANT(T, 113, 281.4601449275362318840579710144927536231884057971), |
| 73 | BOOST_MATH_BIG_CONSTANT(T, 113, -3607.510546398046398046398046398046398046398046398), |
| 74 | BOOST_MATH_BIG_CONSTANT(T, 113, 54827.583333333333333333333333333333333333333333333), |
| 75 | BOOST_MATH_BIG_CONSTANT(T, 113, -974936.82385057471264367816091954022988505747126437), |
| 76 | BOOST_MATH_BIG_CONSTANT(T, 113, 20052695.796688078946143462272494530559046688078946), |
| 77 | BOOST_MATH_BIG_CONSTANT(T, 113, -472384867.72162990196078431372549019607843137254902), |
| 78 | BOOST_MATH_BIG_CONSTANT(T, 113, 12635724795.916666666666666666666666666666666666667) |
| 79 | }; |
| 80 | x -= 1; |
| 81 | T result = log(x); |
| 82 | result += 1 / (2 * x); |
| 83 | T z = 1 / (x*x); |
| 84 | result -= z * tools::evaluate_polynomial(P, z); |
| 85 | return result; |
| 86 | } |
| 87 | // |
| 88 | // 19-digit precision for x >= 10: |
| 89 | // |
| 90 | template <class T> |
| 91 | inline T digamma_imp_large(T x, const boost::integral_constant<int, 64>*) |
| 92 | { |
| 93 | BOOST_MATH_STD_USING // ADL of std functions. |
| 94 | static const T P[] = { |
| 95 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.083333333333333333333333333333333333333333333333333), |
| 96 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.0083333333333333333333333333333333333333333333333333), |
| 97 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.003968253968253968253968253968253968253968253968254), |
| 98 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.0041666666666666666666666666666666666666666666666667), |
| 99 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.0075757575757575757575757575757575757575757575757576), |
| 100 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.021092796092796092796092796092796092796092796092796), |
| 101 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.083333333333333333333333333333333333333333333333333), |
| 102 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.44325980392156862745098039215686274509803921568627), |
| 103 | BOOST_MATH_BIG_CONSTANT(T, 64, 3.0539543302701197438039543302701197438039543302701), |
| 104 | BOOST_MATH_BIG_CONSTANT(T, 64, -26.456212121212121212121212121212121212121212121212), |
| 105 | BOOST_MATH_BIG_CONSTANT(T, 64, 281.4601449275362318840579710144927536231884057971), |
| 106 | }; |
| 107 | x -= 1; |
| 108 | T result = log(x); |
| 109 | result += 1 / (2 * x); |
| 110 | T z = 1 / (x*x); |
| 111 | result -= z * tools::evaluate_polynomial(P, z); |
| 112 | return result; |
| 113 | } |
| 114 | // |
| 115 | // 17-digit precision for x >= 10: |
| 116 | // |
| 117 | template <class T> |
| 118 | inline T digamma_imp_large(T x, const boost::integral_constant<int, 53>*) |
| 119 | { |
| 120 | BOOST_MATH_STD_USING // ADL of std functions. |
| 121 | static const T P[] = { |
| 122 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.083333333333333333333333333333333333333333333333333), |
| 123 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.0083333333333333333333333333333333333333333333333333), |
| 124 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.003968253968253968253968253968253968253968253968254), |
| 125 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.0041666666666666666666666666666666666666666666666667), |
| 126 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.0075757575757575757575757575757575757575757575757576), |
| 127 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.021092796092796092796092796092796092796092796092796), |
| 128 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.083333333333333333333333333333333333333333333333333), |
| 129 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.44325980392156862745098039215686274509803921568627) |
| 130 | }; |
| 131 | x -= 1; |
| 132 | T result = log(x); |
| 133 | result += 1 / (2 * x); |
| 134 | T z = 1 / (x*x); |
| 135 | result -= z * tools::evaluate_polynomial(P, z); |
| 136 | return result; |
| 137 | } |
| 138 | // |
| 139 | // 9-digit precision for x >= 10: |
| 140 | // |
| 141 | template <class T> |
| 142 | inline T digamma_imp_large(T x, const boost::integral_constant<int, 24>*) |
| 143 | { |
| 144 | BOOST_MATH_STD_USING // ADL of std functions. |
| 145 | static const T P[] = { |
| 146 | BOOST_MATH_BIG_CONSTANT(T, 24, 0.083333333333333333333333333333333333333333333333333), |
| 147 | BOOST_MATH_BIG_CONSTANT(T, 24, -0.0083333333333333333333333333333333333333333333333333), |
| 148 | BOOST_MATH_BIG_CONSTANT(T, 24, 0.003968253968253968253968253968253968253968253968254) |
| 149 | }; |
| 150 | x -= 1; |
| 151 | T result = log(x); |
| 152 | result += 1 / (2 * x); |
| 153 | T z = 1 / (x*x); |
| 154 | result -= z * tools::evaluate_polynomial(P, z); |
| 155 | return result; |
| 156 | } |
| 157 | // |
| 158 | // Fully generic asymptotic expansion in terms of Bernoulli numbers, see: |
| 159 | // http://functions.wolfram.com/06.14.06.0012.01 |
| 160 | // |
| 161 | template <class T> |
| 162 | struct digamma_series_func |
| 163 | { |
| 164 | private: |
| 165 | int k; |
| 166 | T xx; |
| 167 | T term; |
| 168 | public: |
| 169 | digamma_series_func(T x) : k(1), xx(x * x), term(1 / (x * x)) {} |
| 170 | T operator()() |
| 171 | { |
| 172 | T result = term * boost::math::bernoulli_b2n<T>(k) / (2 * k); |
| 173 | term /= xx; |
| 174 | ++k; |
| 175 | return result; |
| 176 | } |
| 177 | typedef T result_type; |
| 178 | }; |
| 179 | |
| 180 | template <class T, class Policy> |
| 181 | inline T digamma_imp_large(T x, const Policy& pol, const boost::integral_constant<int, 0>*) |
| 182 | { |
| 183 | BOOST_MATH_STD_USING |
| 184 | digamma_series_func<T> s(x); |
| 185 | T result = log(x) - 1 / (2 * x); |
| 186 | boost::uintmax_t max_iter = policies::get_max_series_iterations<Policy>(); |
| 187 | result = boost::math::tools::sum_series(s, boost::math::policies::get_epsilon<T, Policy>(), max_iter, -result); |
| 188 | result = -result; |
| 189 | policies::check_series_iterations<T>("boost::math::digamma<%1%>(%1%)" , max_iter, pol); |
| 190 | return result; |
| 191 | } |
| 192 | // |
| 193 | // Now follow rational approximations over the range [1,2]. |
| 194 | // |
| 195 | // 35-digit precision: |
| 196 | // |
| 197 | template <class T> |
| 198 | T digamma_imp_1_2(T x, const boost::integral_constant<int, 113>*) |
| 199 | { |
| 200 | // |
| 201 | // Now the approximation, we use the form: |
| 202 | // |
| 203 | // digamma(x) = (x - root) * (Y + R(x-1)) |
| 204 | // |
| 205 | // Where root is the location of the positive root of digamma, |
| 206 | // Y is a constant, and R is optimised for low absolute error |
| 207 | // compared to Y. |
| 208 | // |
| 209 | // Max error found at 128-bit long double precision: 5.541e-35 |
| 210 | // Maximum Deviation Found (approximation error): 1.965e-35 |
| 211 | // |
| 212 | static const float Y = 0.99558162689208984375F; |
| 213 | |
| 214 | static const T root1 = T(1569415565) / 1073741824uL; |
| 215 | static const T root2 = (T(381566830) / 1073741824uL) / 1073741824uL; |
| 216 | static const T root3 = ((T(111616537) / 1073741824uL) / 1073741824uL) / 1073741824uL; |
| 217 | static const T root4 = (((T(503992070) / 1073741824uL) / 1073741824uL) / 1073741824uL) / 1073741824uL; |
| 218 | static const T root5 = BOOST_MATH_BIG_CONSTANT(T, 113, 0.52112228569249997894452490385577338504019838794544e-36); |
| 219 | |
| 220 | static const T P[] = { |
| 221 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.25479851061131551526977464225335883769), |
| 222 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.18684290534374944114622235683619897417), |
| 223 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.80360876047931768958995775910991929922), |
| 224 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.67227342794829064330498117008564270136), |
| 225 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.26569010991230617151285010695543858005), |
| 226 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.05775672694575986971640757748003553385), |
| 227 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.0071432147823164975485922555833274240665), |
| 228 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.00048740753910766168912364555706064993274), |
| 229 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.16454996865214115723416538844975174761e-4), |
| 230 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.20327832297631728077731148515093164955e-6) |
| 231 | }; |
| 232 | static const T Q[] = { |
| 233 | BOOST_MATH_BIG_CONSTANT(T, 113, 1.0), |
| 234 | BOOST_MATH_BIG_CONSTANT(T, 113, 2.6210924610812025425088411043163287646), |
| 235 | BOOST_MATH_BIG_CONSTANT(T, 113, 2.6850757078559596612621337395886392594), |
| 236 | BOOST_MATH_BIG_CONSTANT(T, 113, 1.4320913706209965531250495490639289418), |
| 237 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.4410872083455009362557012239501953402), |
| 238 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.081385727399251729505165509278152487225), |
| 239 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.0089478633066857163432104815183858149496), |
| 240 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.00055861622855066424871506755481997374154), |
| 241 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.1760168552357342401304462967950178554e-4), |
| 242 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.20585454493572473724556649516040874384e-6), |
| 243 | BOOST_MATH_BIG_CONSTANT(T, 113, -0.90745971844439990284514121823069162795e-11), |
| 244 | BOOST_MATH_BIG_CONSTANT(T, 113, 0.48857673606545846774761343500033283272e-13), |
| 245 | }; |
| 246 | T g = x - root1; |
| 247 | g -= root2; |
| 248 | g -= root3; |
| 249 | g -= root4; |
| 250 | g -= root5; |
| 251 | T r = tools::evaluate_polynomial(P, T(x-1)) / tools::evaluate_polynomial(Q, T(x-1)); |
| 252 | T result = g * Y + g * r; |
| 253 | |
| 254 | return result; |
| 255 | } |
| 256 | // |
| 257 | // 19-digit precision: |
| 258 | // |
| 259 | template <class T> |
| 260 | T digamma_imp_1_2(T x, const boost::integral_constant<int, 64>*) |
| 261 | { |
| 262 | // |
| 263 | // Now the approximation, we use the form: |
| 264 | // |
| 265 | // digamma(x) = (x - root) * (Y + R(x-1)) |
| 266 | // |
| 267 | // Where root is the location of the positive root of digamma, |
| 268 | // Y is a constant, and R is optimised for low absolute error |
| 269 | // compared to Y. |
| 270 | // |
| 271 | // Max error found at 80-bit long double precision: 5.016e-20 |
| 272 | // Maximum Deviation Found (approximation error): 3.575e-20 |
| 273 | // |
| 274 | static const float Y = 0.99558162689208984375F; |
| 275 | |
| 276 | static const T root1 = T(1569415565) / 1073741824uL; |
| 277 | static const T root2 = (T(381566830) / 1073741824uL) / 1073741824uL; |
| 278 | static const T root3 = BOOST_MATH_BIG_CONSTANT(T, 64, 0.9016312093258695918615325266959189453125e-19); |
| 279 | |
| 280 | static const T P[] = { |
| 281 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.254798510611315515235), |
| 282 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.314628554532916496608), |
| 283 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.665836341559876230295), |
| 284 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.314767657147375752913), |
| 285 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.0541156266153505273939), |
| 286 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.00289268368333918761452) |
| 287 | }; |
| 288 | static const T Q[] = { |
| 289 | BOOST_MATH_BIG_CONSTANT(T, 64, 1.0), |
| 290 | BOOST_MATH_BIG_CONSTANT(T, 64, 2.1195759927055347547), |
| 291 | BOOST_MATH_BIG_CONSTANT(T, 64, 1.54350554664961128724), |
| 292 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.486986018231042975162), |
| 293 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.0660481487173569812846), |
| 294 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.00298999662592323990972), |
| 295 | BOOST_MATH_BIG_CONSTANT(T, 64, -0.165079794012604905639e-5), |
| 296 | BOOST_MATH_BIG_CONSTANT(T, 64, 0.317940243105952177571e-7) |
| 297 | }; |
| 298 | T g = x - root1; |
| 299 | g -= root2; |
| 300 | g -= root3; |
| 301 | T r = tools::evaluate_polynomial(P, T(x-1)) / tools::evaluate_polynomial(Q, T(x-1)); |
| 302 | T result = g * Y + g * r; |
| 303 | |
| 304 | return result; |
| 305 | } |
| 306 | // |
| 307 | // 18-digit precision: |
| 308 | // |
| 309 | template <class T> |
| 310 | T digamma_imp_1_2(T x, const boost::integral_constant<int, 53>*) |
| 311 | { |
| 312 | // |
| 313 | // Now the approximation, we use the form: |
| 314 | // |
| 315 | // digamma(x) = (x - root) * (Y + R(x-1)) |
| 316 | // |
| 317 | // Where root is the location of the positive root of digamma, |
| 318 | // Y is a constant, and R is optimised for low absolute error |
| 319 | // compared to Y. |
| 320 | // |
| 321 | // Maximum Deviation Found: 1.466e-18 |
| 322 | // At double precision, max error found: 2.452e-17 |
| 323 | // |
| 324 | static const float Y = 0.99558162689208984F; |
| 325 | |
| 326 | static const T root1 = T(1569415565) / 1073741824uL; |
| 327 | static const T root2 = (T(381566830) / 1073741824uL) / 1073741824uL; |
| 328 | static const T root3 = BOOST_MATH_BIG_CONSTANT(T, 53, 0.9016312093258695918615325266959189453125e-19); |
| 329 | |
| 330 | static const T P[] = { |
| 331 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.25479851061131551), |
| 332 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.32555031186804491), |
| 333 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.65031853770896507), |
| 334 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.28919126444774784), |
| 335 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.045251321448739056), |
| 336 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.0020713321167745952) |
| 337 | }; |
| 338 | static const T Q[] = { |
| 339 | BOOST_MATH_BIG_CONSTANT(T, 53, 1.0), |
| 340 | BOOST_MATH_BIG_CONSTANT(T, 53, 2.0767117023730469), |
| 341 | BOOST_MATH_BIG_CONSTANT(T, 53, 1.4606242909763515), |
| 342 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.43593529692665969), |
| 343 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.054151797245674225), |
| 344 | BOOST_MATH_BIG_CONSTANT(T, 53, 0.0021284987017821144), |
| 345 | BOOST_MATH_BIG_CONSTANT(T, 53, -0.55789841321675513e-6) |
| 346 | }; |
| 347 | T g = x - root1; |
| 348 | g -= root2; |
| 349 | g -= root3; |
| 350 | T r = tools::evaluate_polynomial(P, T(x-1)) / tools::evaluate_polynomial(Q, T(x-1)); |
| 351 | T result = g * Y + g * r; |
| 352 | |
| 353 | return result; |
| 354 | } |
| 355 | // |
| 356 | // 9-digit precision: |
| 357 | // |
| 358 | template <class T> |
| 359 | inline T digamma_imp_1_2(T x, const boost::integral_constant<int, 24>*) |
| 360 | { |
| 361 | // |
| 362 | // Now the approximation, we use the form: |
| 363 | // |
| 364 | // digamma(x) = (x - root) * (Y + R(x-1)) |
| 365 | // |
| 366 | // Where root is the location of the positive root of digamma, |
| 367 | // Y is a constant, and R is optimised for low absolute error |
| 368 | // compared to Y. |
| 369 | // |
| 370 | // Maximum Deviation Found: 3.388e-010 |
| 371 | // At float precision, max error found: 2.008725e-008 |
| 372 | // |
| 373 | static const float Y = 0.99558162689208984f; |
| 374 | static const T root = 1532632.0f / 1048576; |
| 375 | static const T root_minor = static_cast<T>(0.3700660185912626595423257213284682051735604e-6L); |
| 376 | static const T P[] = { |
| 377 | 0.25479851023250261e0f, |
| 378 | -0.44981331915268368e0f, |
| 379 | -0.43916936919946835e0f, |
| 380 | -0.61041765350579073e-1f |
| 381 | }; |
| 382 | static const T Q[] = { |
| 383 | 0.1e1, |
| 384 | 0.15890202430554952e1f, |
| 385 | 0.65341249856146947e0f, |
| 386 | 0.63851690523355715e-1f |
| 387 | }; |
| 388 | T g = x - root; |
| 389 | g -= root_minor; |
| 390 | T r = tools::evaluate_polynomial(P, T(x-1)) / tools::evaluate_polynomial(Q, T(x-1)); |
| 391 | T result = g * Y + g * r; |
| 392 | |
| 393 | return result; |
| 394 | } |
| 395 | |
| 396 | template <class T, class Tag, class Policy> |
| 397 | T digamma_imp(T x, const Tag* t, const Policy& pol) |
| 398 | { |
| 399 | // |
| 400 | // This handles reflection of negative arguments, and all our |
| 401 | // error handling, then forwards to the T-specific approximation. |
| 402 | // |
| 403 | BOOST_MATH_STD_USING // ADL of std functions. |
| 404 | |
| 405 | T result = 0; |
| 406 | // |
| 407 | // Check for negative arguments and use reflection: |
| 408 | // |
| 409 | if(x <= -1) |
| 410 | { |
| 411 | // Reflect: |
| 412 | x = 1 - x; |
| 413 | // Argument reduction for tan: |
| 414 | T remainder = x - floor(x); |
| 415 | // Shift to negative if > 0.5: |
| 416 | if(remainder > 0.5) |
| 417 | { |
| 418 | remainder -= 1; |
| 419 | } |
| 420 | // |
| 421 | // check for evaluation at a negative pole: |
| 422 | // |
| 423 | if(remainder == 0) |
| 424 | { |
| 425 | return policies::raise_pole_error<T>("boost::math::digamma<%1%>(%1%)" , 0, (1-x), pol); |
| 426 | } |
| 427 | result = constants::pi<T>() / tan(constants::pi<T>() * remainder); |
| 428 | } |
| 429 | if(x == 0) |
| 430 | return policies::raise_pole_error<T>("boost::math::digamma<%1%>(%1%)" , 0, x, pol); |
| 431 | // |
| 432 | // If we're above the lower-limit for the |
| 433 | // asymptotic expansion then use it: |
| 434 | // |
| 435 | if(x >= digamma_large_lim(t)) |
| 436 | { |
| 437 | result += digamma_imp_large(x, t); |
| 438 | } |
| 439 | else |
| 440 | { |
| 441 | // |
| 442 | // If x > 2 reduce to the interval [1,2]: |
| 443 | // |
| 444 | while(x > 2) |
| 445 | { |
| 446 | x -= 1; |
| 447 | result += 1/x; |
| 448 | } |
| 449 | // |
| 450 | // If x < 1 use recurrence to shift to > 1: |
| 451 | // |
| 452 | while(x < 1) |
| 453 | { |
| 454 | result -= 1/x; |
| 455 | x += 1; |
| 456 | } |
| 457 | result += digamma_imp_1_2(x, t); |
| 458 | } |
| 459 | return result; |
| 460 | } |
| 461 | |
| 462 | template <class T, class Policy> |
| 463 | T digamma_imp(T x, const boost::integral_constant<int, 0>* t, const Policy& pol) |
| 464 | { |
| 465 | // |
| 466 | // This handles reflection of negative arguments, and all our |
| 467 | // error handling, then forwards to the T-specific approximation. |
| 468 | // |
| 469 | BOOST_MATH_STD_USING // ADL of std functions. |
| 470 | |
| 471 | T result = 0; |
| 472 | // |
| 473 | // Check for negative arguments and use reflection: |
| 474 | // |
| 475 | if(x <= -1) |
| 476 | { |
| 477 | // Reflect: |
| 478 | x = 1 - x; |
| 479 | // Argument reduction for tan: |
| 480 | T remainder = x - floor(x); |
| 481 | // Shift to negative if > 0.5: |
| 482 | if(remainder > 0.5) |
| 483 | { |
| 484 | remainder -= 1; |
| 485 | } |
| 486 | // |
| 487 | // check for evaluation at a negative pole: |
| 488 | // |
| 489 | if(remainder == 0) |
| 490 | { |
| 491 | return policies::raise_pole_error<T>("boost::math::digamma<%1%>(%1%)" , 0, (1 - x), pol); |
| 492 | } |
| 493 | result = constants::pi<T>() / tan(constants::pi<T>() * remainder); |
| 494 | } |
| 495 | if(x == 0) |
| 496 | return policies::raise_pole_error<T>("boost::math::digamma<%1%>(%1%)" , 0, x, pol); |
| 497 | // |
| 498 | // If we're above the lower-limit for the |
| 499 | // asymptotic expansion then use it, the |
| 500 | // limit is a linear interpolation with |
| 501 | // limit = 10 at 50 bit precision and |
| 502 | // limit = 250 at 1000 bit precision. |
| 503 | // |
| 504 | int lim = 10 + ((tools::digits<T>() - 50) * 240L) / 950; |
| 505 | T two_x = ldexp(x, 1); |
| 506 | if(x >= lim) |
| 507 | { |
| 508 | result += digamma_imp_large(x, pol, t); |
| 509 | } |
| 510 | else if(floor(x) == x) |
| 511 | { |
| 512 | // |
| 513 | // Special case for integer arguments, see |
| 514 | // http://functions.wolfram.com/06.14.03.0001.01 |
| 515 | // |
| 516 | result = -constants::euler<T, Policy>(); |
| 517 | T val = 1; |
| 518 | while(val < x) |
| 519 | { |
| 520 | result += 1 / val; |
| 521 | val += 1; |
| 522 | } |
| 523 | } |
| 524 | else if(floor(two_x) == two_x) |
| 525 | { |
| 526 | // |
| 527 | // Special case for half integer arguments, see: |
| 528 | // http://functions.wolfram.com/06.14.03.0007.01 |
| 529 | // |
| 530 | result = -2 * constants::ln_two<T, Policy>() - constants::euler<T, Policy>(); |
| 531 | int n = itrunc(x); |
| 532 | if(n) |
| 533 | { |
| 534 | for(int k = 1; k < n; ++k) |
| 535 | result += 1 / T(k); |
| 536 | for(int k = n; k <= 2 * n - 1; ++k) |
| 537 | result += 2 / T(k); |
| 538 | } |
| 539 | } |
| 540 | else |
| 541 | { |
| 542 | // |
| 543 | // Rescale so we can use the asymptotic expansion: |
| 544 | // |
| 545 | while(x < lim) |
| 546 | { |
| 547 | result -= 1 / x; |
| 548 | x += 1; |
| 549 | } |
| 550 | result += digamma_imp_large(x, pol, t); |
| 551 | } |
| 552 | return result; |
| 553 | } |
| 554 | // |
| 555 | // Initializer: ensure all our constants are initialized prior to the first call of main: |
| 556 | // |
| 557 | template <class T, class Policy> |
| 558 | struct digamma_initializer |
| 559 | { |
| 560 | struct init |
| 561 | { |
| 562 | init() |
| 563 | { |
| 564 | typedef typename policies::precision<T, Policy>::type precision_type; |
| 565 | do_init(boost::integral_constant<bool, precision_type::value && (precision_type::value <= 113)>()); |
| 566 | } |
| 567 | void do_init(const boost::true_type&) |
| 568 | { |
| 569 | boost::math::digamma(T(1.5), Policy()); |
| 570 | boost::math::digamma(T(500), Policy()); |
| 571 | } |
| 572 | void do_init(const false_type&){} |
| 573 | void force_instantiate()const{} |
| 574 | }; |
| 575 | static const init initializer; |
| 576 | static void force_instantiate() |
| 577 | { |
| 578 | initializer.force_instantiate(); |
| 579 | } |
| 580 | }; |
| 581 | |
| 582 | template <class T, class Policy> |
| 583 | const typename digamma_initializer<T, Policy>::init digamma_initializer<T, Policy>::initializer; |
| 584 | |
| 585 | } // namespace detail |
| 586 | |
| 587 | template <class T, class Policy> |
| 588 | inline typename tools::promote_args<T>::type |
| 589 | digamma(T x, const Policy&) |
| 590 | { |
| 591 | typedef typename tools::promote_args<T>::type result_type; |
| 592 | typedef typename policies::evaluation<result_type, Policy>::type value_type; |
| 593 | typedef typename policies::precision<T, Policy>::type precision_type; |
| 594 | typedef boost::integral_constant<int, |
| 595 | (precision_type::value <= 0) || (precision_type::value > 113) ? 0 : |
| 596 | precision_type::value <= 24 ? 24 : |
| 597 | precision_type::value <= 53 ? 53 : |
| 598 | precision_type::value <= 64 ? 64 : |
| 599 | precision_type::value <= 113 ? 113 : 0 > tag_type; |
| 600 | typedef typename policies::normalise< |
| 601 | Policy, |
| 602 | policies::promote_float<false>, |
| 603 | policies::promote_double<false>, |
| 604 | policies::discrete_quantile<>, |
| 605 | policies::assert_undefined<> >::type forwarding_policy; |
| 606 | |
| 607 | // Force initialization of constants: |
| 608 | detail::digamma_initializer<value_type, forwarding_policy>::force_instantiate(); |
| 609 | |
| 610 | return policies::checked_narrowing_cast<result_type, Policy>(detail::digamma_imp( |
| 611 | static_cast<value_type>(x), |
| 612 | static_cast<const tag_type*>(0), forwarding_policy()), "boost::math::digamma<%1%>(%1%)" ); |
| 613 | } |
| 614 | |
| 615 | template <class T> |
| 616 | inline typename tools::promote_args<T>::type |
| 617 | digamma(T x) |
| 618 | { |
| 619 | return digamma(x, policies::policy<>()); |
| 620 | } |
| 621 | |
| 622 | } // namespace math |
| 623 | } // namespace boost |
| 624 | |
| 625 | #ifdef _MSC_VER |
| 626 | #pragma warning(pop) |
| 627 | #endif |
| 628 | |
| 629 | #endif |
| 630 | |
| 631 | |