| 1 |
// Copyright John Maddock 2017. |
| 2 |
// Copyright Paul A. Bristow 2016, 2017, 2018. |
| 3 |
// Copyright Nicholas Thompson 2018 |
| 4 |
|
| 5 |
// Distributed under the Boost Software License, Version 1.0. |
| 6 |
// (See accompanying file LICENSE_1_0.txt or |
| 7 |
// copy at http ://www.boost.org/LICENSE_1_0.txt). |
| 8 |
|
| 9 |
#ifndef BOOST_MATH_SF_LAMBERT_W_HPP |
| 10 |
#define BOOST_MATH_SF_LAMBERT_W_HPP |
| 11 |
|
| 12 |
#ifdef _MSC_VER |
| 13 |
#pragma warning(disable : 4127) |
| 14 |
#endif |
| 15 |
|
| 16 |
/* |
| 17 |
Implementation of an algorithm for the Lambert W0 and W-1 real-only functions. |
| 18 |
|
| 19 |
This code is based in part on the algorithm by |
| 20 |
Toshio Fukushima, |
| 21 |
"Precise and fast computation of Lambert W-functions without transcendental function evaluations", |
| 22 |
J.Comp.Appl.Math. 244 (2013) 77-89, |
| 23 |
and on a C/C++ version by Darko Veberic, [email protected] |
| 24 |
based on the Fukushima algorithm and Toshio Fukushima's FORTRAN version of his algorithm. |
| 25 |
|
| 26 |
First derivative of Lambert_w is derived from |
| 27 |
Princeton Companion to Applied Mathematics, 'The Lambert-W function', Section 1.3: Series and Generating Functions. |
| 28 |
|
| 29 |
*/ |
| 30 |
|
| 31 |
/* |
| 32 |
TODO revise this list of macros. |
| 33 |
Some macros that will show some (or much) diagnostic values if #defined. |
| 34 |
//[boost_math_instrument_lambert_w_macros |
| 35 |
|
| 36 |
// #define-able macros |
| 37 |
BOOST_MATH_INSTRUMENT_LAMBERT_W_HALLEY // Halley refinement diagnostics. |
| 38 |
BOOST_MATH_INSTRUMENT_LAMBERT_W_PRECISION // Precision. |
| 39 |
BOOST_MATH_INSTRUMENT_LAMBERT_WM1 // W1 branch diagnostics. |
| 40 |
BOOST_MATH_INSTRUMENT_LAMBERT_WM1_HALLEY // Halley refinement diagnostics only for W-1 branch. |
| 41 |
BOOST_MATH_INSTRUMENT_LAMBERT_WM1_TINY // K > 64, z > -1.0264389699511303e-26 |
| 42 |
BOOST_MATH_INSTRUMENT_LAMBERT_WM1_LOOKUP // Show results from W-1 lookup table. |
| 43 |
BOOST_MATH_INSTRUMENT_LAMBERT_W_SCHROEDER // Schroeder refinement diagnostics. |
| 44 |
BOOST_MATH_INSTRUMENT_LAMBERT_W_TERMS // Number of terms used for near-singularity series. |
| 45 |
BOOST_MATH_INSTRUMENT_LAMBERT_W_SINGULARITY_SERIES // Show evaluation of series near branch singularity. |
| 46 |
BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 47 |
BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES_ITERATIONS // Show evaluation of series for small z. |
| 48 |
//] [/boost_math_instrument_lambert_w_macros] |
| 49 |
*/ |
| 50 |
|
| 51 |
#include <boost/math/tools/config.hpp> |
| 52 |
#include <boost/math/policies/error_handling.hpp> |
| 53 |
#include <boost/math/policies/policy.hpp> |
| 54 |
#include <boost/math/tools/promotion.hpp> |
| 55 |
#include <boost/math/special_functions/fpclassify.hpp> |
| 56 |
#include <boost/math/special_functions/log1p.hpp> // for log (1 + x) |
| 57 |
#include <boost/math/constants/constants.hpp> // For exp_minus_one == 3.67879441171442321595523770161460867e-01. |
| 58 |
#include <boost/math/special_functions/next.hpp> // for has_denorm_now |
| 59 |
#include <boost/math/special_functions/pow.hpp> // powers with compile time exponent, used in arbitrary precision code. |
| 60 |
#include <boost/math/tools/series.hpp> // series functor. |
| 61 |
//#include <boost/math/tools/polynomial.hpp> // polynomial. |
| 62 |
#include <boost/math/tools/rational.hpp> // evaluate_polynomial. |
| 63 |
#include <boost/math/tools/precision.hpp> // boost::math::tools::max_value(). |
| 64 |
#include <boost/math/tools/big_constant.hpp> |
| 65 |
#include <boost/math/tools/cxx03_warn.hpp> |
| 66 |
|
| 67 |
#ifndef BOOST_MATH_STANDALONE |
| 68 |
#include <boost/lexical_cast.hpp> |
| 69 |
#endif |
| 70 |
|
| 71 |
#include <limits> |
| 72 |
#include <cmath> |
| 73 |
#include <type_traits> |
| 74 |
#include <cstdint> |
| 75 |
|
| 76 |
// Needed for testing and diagnostics only. |
| 77 |
//#include <iostream> |
| 78 |
//#include <typeinfo> |
| 79 |
#include <boost/math/special_functions/next.hpp> // For float_distance. |
| 80 |
|
| 81 |
using lookup_t = double; // Type for lookup table (double or float, or even long double?) |
| 82 |
|
| 83 |
//#include "J:\Cpp\Misc\lambert_w_lookup_table_generator\lambert_w_lookup_table.ipp" |
| 84 |
// #include "lambert_w_lookup_table.ipp" // Boost.Math version. |
| 85 |
#include <boost/math/special_functions/detail/lambert_w_lookup_table.ipp> |
| 86 |
|
| 87 |
#if defined(__GNUC__) && defined(BOOST_MATH_USE_FLOAT128) |
| 88 |
// |
| 89 |
// This is the only way we can avoid |
| 90 |
// warning: non-standard suffix on floating constant [-Wpedantic] |
| 91 |
// when building with -Wall -pedantic. Neither __extension__ |
| 92 |
// nor #pragma diagnostic ignored work :( |
| 93 |
// |
| 94 |
#pragma GCC system_header |
| 95 |
#endif |
| 96 |
|
| 97 |
namespace boost { |
| 98 |
namespace math { |
| 99 |
namespace lambert_w_detail { |
| 100 |
|
| 101 |
//! \brief Applies a single Halley step to make a better estimate of Lambert W. |
| 102 |
//! \details Used the simplified formulae obtained from |
| 103 |
//! http://www.wolframalpha.com/input/?i=%5B2(z+exp(z)-w)+d%2Fdx+(z+exp(z)-w)%5D+%2F+%5B2+(d%2Fdx+(z+exp(z)-w))%5E2+-+(z+exp(z)-w)+d%5E2%2Fdx%5E2+(z+exp(z)-w)%5D |
| 104 |
//! [2(z exp(z)-w) d/dx (z exp(z)-w)] / [2 (d/dx (z exp(z)-w))^2 - (z exp(z)-w) d^2/dx^2 (z exp(z)-w)] |
| 105 |
|
| 106 |
//! \tparam T floating-point (or fixed-point) type. |
| 107 |
//! \param w_est Lambert W estimate. |
| 108 |
//! \param z Argument z for Lambert_w function. |
| 109 |
//! \returns New estimate of Lambert W, hopefully improved. |
| 110 |
//! |
| 111 |
template <typename T> |
| 112 |
inline T lambert_w_halley_step(T w_est, const T z) |
| 113 |
{ |
| 114 |
BOOST_MATH_STD_USING |
| 115 |
T e = exp(w_est); |
| 116 |
w_est -= 2 * (w_est + 1) * (e * w_est - z) / (z * (w_est + 2) + e * (w_est * (w_est + 2) + 2)); |
| 117 |
return w_est; |
| 118 |
} // template <typename T> lambert_w_halley_step(T w_est, T z) |
| 119 |
|
| 120 |
//! \brief Halley iterate to refine Lambert_w estimate, |
| 121 |
//! taking at least one Halley_step. |
| 122 |
//! Repeat Halley steps until the *last step* had fewer than half the digits wrong, |
| 123 |
//! the step we've just taken should have been sufficient to have completed the iteration. |
| 124 |
|
| 125 |
//! \tparam T floating-point (or fixed-point) type. |
| 126 |
//! \param z Argument z for Lambert_w function. |
| 127 |
//! \param w_est Lambert w estimate. |
| 128 |
template <typename T> |
| 129 |
inline T lambert_w_halley_iterate(T w_est, const T z) |
| 130 |
{ |
| 131 |
BOOST_MATH_STD_USING |
| 132 |
static const T max_diff = boost::math::tools::root_epsilon<T>() * fabs(w_est); |
| 133 |
|
| 134 |
T w_new = lambert_w_halley_step(w_est, z); |
| 135 |
T diff = fabs(w_est - w_new); |
| 136 |
while (diff > max_diff) |
| 137 |
{ |
| 138 |
w_est = w_new; |
| 139 |
w_new = lambert_w_halley_step(w_est, z); |
| 140 |
diff = fabs(w_est - w_new); |
| 141 |
} |
| 142 |
return w_new; |
| 143 |
} // template <typename T> lambert_w_halley_iterate(T w_est, T z) |
| 144 |
|
| 145 |
// Two Halley function versions that either |
| 146 |
// single step (if std::false_type) or iterate (if std::true_type). |
| 147 |
// Selected at compile-time using parameter 3. |
| 148 |
template <typename T> |
| 149 |
inline T lambert_w_maybe_halley_iterate(T z, T w, std::false_type const&) |
| 150 |
{ |
| 151 |
return lambert_w_halley_step(z, w); // Single step. |
| 152 |
} |
| 153 |
|
| 154 |
template <typename T> |
| 155 |
inline T lambert_w_maybe_halley_iterate(T z, T w, std::true_type const&) |
| 156 |
{ |
| 157 |
return lambert_w_halley_iterate(z, w); // Iterate steps. |
| 158 |
} |
| 159 |
|
| 160 |
//! maybe_reduce_to_double function, |
| 161 |
//! Two versions that have a compile-time option to |
| 162 |
//! reduce argument z to double precision (if true_type). |
| 163 |
//! Version is selected at compile-time using parameter 2. |
| 164 |
|
| 165 |
template <typename T> |
| 166 |
inline double maybe_reduce_to_double(const T& z, const std::true_type&) |
| 167 |
{ |
| 168 |
return static_cast<double>(z); // Reduce to double precision. |
| 169 |
} |
| 170 |
|
| 171 |
template <typename T> |
| 172 |
inline T maybe_reduce_to_double(const T& z, const std::false_type&) |
| 173 |
{ // Don't reduce to double. |
| 174 |
return z; |
| 175 |
} |
| 176 |
|
| 177 |
template <typename T> |
| 178 |
inline double must_reduce_to_double(const T& z, const std::true_type&) |
| 179 |
{ |
| 180 |
return static_cast<double>(z); // Reduce to double precision. |
| 181 |
} |
| 182 |
|
| 183 |
template <typename T> |
| 184 |
inline double must_reduce_to_double(const T& z, const std::false_type&) |
| 185 |
{ // try a lexical_cast and hope for the best: |
| 186 |
#ifndef BOOST_MATH_STANDALONE |
| 187 |
|
| 188 |
#ifdef BOOST_MATH_USE_CHARCONV_FOR_CONVERSION |
| 189 |
|
| 190 |
// Catches the C++23 floating point types |
| 191 |
if constexpr (std::is_arithmetic_v<T>) |
| 192 |
{ |
| 193 |
return static_cast<double>(z); |
| 194 |
} |
| 195 |
else |
| 196 |
{ |
| 197 |
return boost::lexical_cast<double>(z); |
| 198 |
} |
| 199 |
|
| 200 |
#else |
| 201 |
|
| 202 |
return boost::lexical_cast<double>(z); |
| 203 |
|
| 204 |
#endif |
| 205 |
|
| 206 |
#else |
| 207 |
static_assert(sizeof(T) == 0, "Unsupported in standalone mode: don't know how to cast your number type to a double."); |
| 208 |
return 0.0; |
| 209 |
#endif |
| 210 |
} |
| 211 |
|
| 212 |
//! \brief Schroeder method, fifth-order update formula, |
| 213 |
//! \details See T. Fukushima page 80-81, and |
| 214 |
//! A. Householder, The Numerical Treatment of a Single Nonlinear Equation, |
| 215 |
//! McGraw-Hill, New York, 1970, section 4.4. |
| 216 |
//! Fukushima algorithm switches to @c schroeder_update after pre-computed bisections, |
| 217 |
//! chosen to ensure that the result will be achieve the +/- 10 epsilon target. |
| 218 |
//! \param w Lambert w estimate from bisection or series. |
| 219 |
//! \param y bracketing value from bisection. |
| 220 |
//! \returns Refined estimate of Lambert w. |
| 221 |
|
| 222 |
// Schroeder refinement, called unless NOT required by precision policy. |
| 223 |
template<typename T> |
| 224 |
inline T schroeder_update(const T w, const T y) |
| 225 |
{ |
| 226 |
// Compute derivatives using 5th order Schroeder refinement. |
| 227 |
// Since this is the final step, it will always use the highest precision type T. |
| 228 |
// Example of Call: |
| 229 |
// result = schroeder_update(w, y); |
| 230 |
//where |
| 231 |
// w is estimate of Lambert W (from bisection or series). |
| 232 |
// y is z * e^-w. |
| 233 |
|
| 234 |
BOOST_MATH_STD_USING // Aid argument dependent lookup of abs. |
| 235 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SCHROEDER |
| 236 |
std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 237 |
using boost::math::float_distance; |
| 238 |
T fd = float_distance<T>(w, y); |
| 239 |
std::cout << "Schroder "; |
| 240 |
if (abs(fd) < 214748000.) |
| 241 |
{ |
| 242 |
std::cout << " Distance = "<< static_cast<int>(fd); |
| 243 |
} |
| 244 |
else |
| 245 |
{ |
| 246 |
std::cout << "Difference w - y = " << (w - y) << "."; |
| 247 |
} |
| 248 |
std::cout << std::endl; |
| 249 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SCHROEDER |
| 250 |
// Fukushima equation 18, page 6. |
| 251 |
const T f0 = w - y; // f0 = w - y. |
| 252 |
const T f1 = 1 + y; // f1 = df/dW |
| 253 |
const T f00 = f0 * f0; |
| 254 |
const T f11 = f1 * f1; |
| 255 |
const T f0y = f0 * y; |
| 256 |
const T result = |
| 257 |
w - 4 * f0 * (6 * f1 * (f11 + f0y) + f00 * y) / |
| 258 |
(f11 * (24 * f11 + 36 * f0y) + |
| 259 |
f00 * (6 * y * y + 8 * f1 * y + f0y)); // Fukushima Page 81, equation 21 from equation 20. |
| 260 |
|
| 261 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SCHROEDER |
| 262 |
std::cout << "Schroeder refined " << w << " " << y << ", difference " << w-y << ", change " << w - result << ", to result " << result << std::endl; |
| 263 |
std::cout.precision(saved_precision); // Restore. |
| 264 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SCHROEDER |
| 265 |
|
| 266 |
return result; |
| 267 |
} // template<typename T = double> T schroeder_update(const T w, const T y) |
| 268 |
|
| 269 |
//! \brief Series expansion used near the singularity/branch point z = -exp(-1) = -3.6787944. |
| 270 |
//! Wolfram InverseSeries[Series[sqrt[2(p Exp[1 + p] + 1)], {p,-1, 20}]] |
| 271 |
//! Wolfram command used to obtain 40 series terms at 50 decimal digit precision was |
| 272 |
//! N[InverseSeries[Series[Sqrt[2(p Exp[1 + p] + 1)], { p,-1,40 }]], 50] |
| 273 |
//! -1+p-p^2/3+(11 p^3)/72-(43 p^4)/540+(769 p^5)/17280-(221 p^6)/8505+(680863 p^7)/43545600 ... |
| 274 |
//! Decimal values of specifications for built-in floating-point types below |
| 275 |
//! are at least 21 digits precision == max_digits10 for long double. |
| 276 |
//! Longer decimal digits strings are rationals evaluated using Wolfram. |
| 277 |
|
| 278 |
template<typename T> |
| 279 |
T lambert_w_singularity_series(const T p) |
| 280 |
{ |
| 281 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SINGULARITY_SERIES |
| 282 |
std::size_t saved_precision = std::cout.precision(3); |
| 283 |
std::cout << "Singularity_series Lambert_w p argument = " << p << std::endl; |
| 284 |
std::cout |
| 285 |
//<< "Argument Type = " << typeid(T).name() |
| 286 |
//<< ", max_digits10 = " << std::numeric_limits<T>::max_digits10 |
| 287 |
//<< ", epsilon = " << std::numeric_limits<T>::epsilon() |
| 288 |
<< std::endl; |
| 289 |
std::cout.precision(saved_precision); |
| 290 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SINGULARITY_SERIES |
| 291 |
|
| 292 |
static const T q[] = |
| 293 |
{ |
| 294 |
-static_cast<T>(1), // j0 |
| 295 |
+T(1), // j1 |
| 296 |
-T(1) / 3, // 1/3 j2 |
| 297 |
+T(11) / 72, // 0.152777777777777778, // 11/72 j3 |
| 298 |
-T(43) / 540, // 0.0796296296296296296, // 43/540 j4 |
| 299 |
+T(769) / 17280, // 0.0445023148148148148, j5 |
| 300 |
-T(221) / 8505, // 0.0259847148736037625, j6 |
| 301 |
//+T(0.0156356325323339212L), // j7 |
| 302 |
//+T(0.015635632532333921222810111699000587889476778365667L), // j7 from Wolfram N[680863/43545600, 50] |
| 303 |
+T(680863uLL) / 43545600uLL, // +0.0156356325323339212, j7 |
| 304 |
//-T(0.00961689202429943171L), // j8 |
| 305 |
-T(1963uLL) / 204120uLL, // 0.00961689202429943171, j8 |
| 306 |
//-T(0.0096168920242994317068391142465216539290613364687439L), // j8 from Wolfram N[1963/204120, 50] |
| 307 |
+T(226287557uLL) / 37623398400uLL, // 0.00601454325295611786, j9 |
| 308 |
-T(5776369uLL) / 1515591000uLL, // 0.00381129803489199923, j10 |
| 309 |
//+T(0.00244087799114398267L), j11 0.0024408779911439826658968585286437530215699919795550 |
| 310 |
+T(169709463197uLL) / 69528040243200uLL, // j11 |
| 311 |
// -T(0.00157693034468678425L), // j12 -0.0015769303446867842539234095399314115973161850314723 |
| 312 |
-T(1118511313uLL) / 709296588000uLL, // j12 |
| 313 |
+T(667874164916771uLL) / 650782456676352000uLL, // j13 |
| 314 |
//+T(0.00102626332050760715L), // j13 0.0010262633205076071544375481533906861056468041465973 |
| 315 |
-T(500525573uLL) / 744761417400uLL, // j14 |
| 316 |
// -T(0.000672061631156136204L), j14 |
| 317 |
//+T(1003663334225097487uLL) / 234281684403486720000uLL, // j15 0.00044247306181462090993020760858473726479232802068800 error C2177: constant too big |
| 318 |
//+T(0.000442473061814620910L, // j15 |
| 319 |
BOOST_MATH_BIG_CONSTANT(T, 64, +0.000442473061814620910), // j15 |
| 320 |
// -T(0.000292677224729627445L), // j16 |
| 321 |
BOOST_MATH_BIG_CONSTANT(T, 64, -0.000292677224729627445), // j16 |
| 322 |
//+T(0.000194387276054539318L), // j17 |
| 323 |
BOOST_MATH_BIG_CONSTANT(T, 64, 0.000194387276054539318), // j17 |
| 324 |
//-T(0.000129574266852748819L), // j18 |
| 325 |
BOOST_MATH_BIG_CONSTANT(T, 64, -0.000129574266852748819), // j18 |
| 326 |
//+T(0.0000866503580520812717L), // j19 N[+1150497127780071399782389/13277465363600276402995200000, 50] 0.000086650358052081271660451590462390293190597827783288 |
| 327 |
BOOST_MATH_BIG_CONSTANT(T, 64, +0.0000866503580520812717), // j19 |
| 328 |
//-T(0.0000581136075044138168L) // j20 N[2853534237182741069/49102686267859224000000, 50] 0.000058113607504413816772205464778828177256611844221913 |
| 329 |
// -T(2853534237182741069uLL) / 49102686267859224000000uLL // j20 // error C2177: constant too big, |
| 330 |
// so must use BOOST_MATH_BIG_CONSTANT(T, ) format in hope of using suffix Q for quad or decimal digits string for others. |
| 331 |
//-T(0.000058113607504413816772205464778828177256611844221913L), // j20 N[2853534237182741069/49102686267859224000000, 50] 0.000058113607504413816772205464778828177256611844221913 |
| 332 |
BOOST_MATH_BIG_CONSTANT(T, 113, -0.000058113607504413816772205464778828177256611844221913) // j20 - last used by Fukushima |
| 333 |
// More terms don't seem to give any improvement (worse in fact) and are not use for many z values. |
| 334 |
//BOOST_MATH_BIG_CONSTANT(T, +0.000039076684867439051635395583044527492132109160553593), // j21 |
| 335 |
//BOOST_MATH_BIG_CONSTANT(T, -0.000026338064747231098738584082718649443078703982217219), // j22 |
| 336 |
//BOOST_MATH_BIG_CONSTANT(T, +0.000017790345805079585400736282075184540383274460464169), // j23 |
| 337 |
//BOOST_MATH_BIG_CONSTANT(T, -0.000012040352739559976942274116578992585158113153190354), // j24 |
| 338 |
//BOOST_MATH_BIG_CONSTANT(T, +8.1635319824966121713827512573558687050675701559448E-6), // j25 |
| 339 |
//BOOST_MATH_BIG_CONSTANT(T, -5.5442032085673591366657251660804575198155559225316E-6) // j26 |
| 340 |
// -T(5.5442032085673591366657251660804575198155559225316E-6L) // j26 |
| 341 |
// 21 to 26 Added for long double. |
| 342 |
}; // static const T q[] |
| 343 |
|
| 344 |
/* |
| 345 |
// Temporary copy of original double values for comparison; these are reproduced well. |
| 346 |
static const T q[] = |
| 347 |
{ |
| 348 |
-1L, // j0 |
| 349 |
+1L, // j1 |
| 350 |
-0.333333333333333333L, // 1/3 j2 |
| 351 |
+0.152777777777777778L, // 11/72 j3 |
| 352 |
-0.0796296296296296296L, // 43/540 |
| 353 |
+0.0445023148148148148L, |
| 354 |
-0.0259847148736037625L, |
| 355 |
+0.0156356325323339212L, |
| 356 |
-0.00961689202429943171L, |
| 357 |
+0.00601454325295611786L, |
| 358 |
-0.00381129803489199923L, |
| 359 |
+0.00244087799114398267L, |
| 360 |
-0.00157693034468678425L, |
| 361 |
+0.00102626332050760715L, |
| 362 |
-0.000672061631156136204L, |
| 363 |
+0.000442473061814620910L, |
| 364 |
-0.000292677224729627445L, |
| 365 |
+0.000194387276054539318L, |
| 366 |
-0.000129574266852748819L, |
| 367 |
+0.0000866503580520812717L, |
| 368 |
-0.0000581136075044138168L // j20 |
| 369 |
}; |
| 370 |
*/ |
| 371 |
|
| 372 |
// Decide how many series terms to use, increasing as z approaches the singularity, |
| 373 |
// balancing run-time versus computational noise from round-off. |
| 374 |
// In practice, we truncate the series expansion at a certain order. |
| 375 |
// If the order is too large, not only does the amount of computation increase, |
| 376 |
// but also the round-off errors accumulate. |
| 377 |
// See Fukushima equation 35, page 85 for logic of choice of number of series terms. |
| 378 |
|
| 379 |
BOOST_MATH_STD_USING // Aid argument dependent lookup (ADL) of abs. |
| 380 |
|
| 381 |
const T absp = abs(p); |
| 382 |
|
| 383 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_TERMS |
| 384 |
{ |
| 385 |
int terms = 20; // Default to using all terms. |
| 386 |
if (absp < 0.01159) |
| 387 |
{ // Very near singularity. |
| 388 |
terms = 6; |
| 389 |
} |
| 390 |
else if (absp < 0.0766) |
| 391 |
{ // Near singularity. |
| 392 |
terms = 10; |
| 393 |
} |
| 394 |
std::streamsize saved_precision = std::cout.precision(3); |
| 395 |
std::cout << "abs(p) = " << absp << ", terms = " << terms << std::endl; |
| 396 |
std::cout.precision(saved_precision); |
| 397 |
} |
| 398 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_TERMS |
| 399 |
|
| 400 |
if (absp < T(0.01159)) |
| 401 |
{ // Only 6 near-singularity series terms are useful. |
| 402 |
return -1 + p * (1 + p * (q[2] + p * (q[3] + p * (q[4] + p * (q[5] + p * q[6]))))); |
| 403 |
} |
| 404 |
else if (absp < T(0.0766)) // Use 10 near-singularity series terms. |
| 405 |
{ // Use 10 near-singularity series terms. |
| 406 |
return -1 + p * (1 + p * (q[2] + p * (q[3] + p * (q[4] + p * (q[5] + p * (q[6] + p * (q[7] + p * (q[8] + p * (q[9] + p * q[10]))))))))); |
| 407 |
} |
| 408 |
// Use all 20 near-singularity series terms. |
| 409 |
return -1 + p * (1 + p * (q[2] + p * (q[3] + p * (q[4] + p * (q[5] + p * (q[6] + p * (q[7] + p * (q[8] + p * (q[9] + p * (q[10] + p * (q[11] + p * (q[12] + p * (q[13] + p * (q[14] + p * (q[15] + p * (q[16] + p * (q[17] + p * (q[18] + p * (q[19] + p * q[20] /* Last Fukushima term.*/))))))))))))))))))); |
| 410 |
// + // more terms for more precise T: long double ... |
| 411 |
//// but makes almost no difference, so don't use more terms? |
| 412 |
// p*q[21] + |
| 413 |
// p*q[22] + |
| 414 |
// p*q[23] + |
| 415 |
// p*q[24] + |
| 416 |
// p*q[25] |
| 417 |
// ))))))))))))))))))); |
| 418 |
|
| 419 |
} // template<typename T = double> T lambert_w_singularity_series(const T p) |
| 420 |
|
| 421 |
|
| 422 |
///////////////////////////////////////////////////////////////////////////////////////////// |
| 423 |
|
| 424 |
//! \brief Series expansion used near zero (abs(z) < 0.05). |
| 425 |
//! \details |
| 426 |
//! Coefficients of the inverted series expansion of the Lambert W function around z = 0. |
| 427 |
//! Tosio Fukushima always uses all 17 terms of a Taylor series computed using Wolfram with |
| 428 |
//! InverseSeries[Series[z Exp[z],{z,0,17}]] |
| 429 |
//! Tosio Fukushima / Journal of Computational and Applied Mathematics 244 (2013) page 86. |
| 430 |
|
| 431 |
//! Decimal values of specifications for built-in floating-point types below |
| 432 |
//! are 21 digits precision == max_digits10 for long double. |
| 433 |
//! Care! Some coefficients might overflow some fixed_point types. |
| 434 |
|
| 435 |
//! This version is intended to allow use by user-defined types |
| 436 |
//! like Boost.Multiprecision quad and cpp_dec_float types. |
| 437 |
//! The three specializations below for built-in float, double |
| 438 |
//! (and perhaps long double) will be chosen in preference for these types. |
| 439 |
|
| 440 |
//! This version uses rationals computed by Wolfram as far as possible, |
| 441 |
//! limited by maximum size of uLL integers. |
| 442 |
//! For higher term, uses decimal digit strings computed by Wolfram up to the maximum possible using uLL rationals, |
| 443 |
//! and then higher coefficients are computed as necessary using function lambert_w0_small_z_series_term |
| 444 |
//! until the precision required by the policy is achieved. |
| 445 |
//! InverseSeries[Series[z Exp[z],{z,0,34}]] also computed. |
| 446 |
|
| 447 |
// Series evaluation for LambertW(z) as z -> 0. |
| 448 |
// See http://functions.wolfram.com/ElementaryFunctions/ProductLog/06/01/01/0003/ |
| 449 |
// http://functions.wolfram.com/ElementaryFunctions/ProductLog/06/01/01/0003/MainEq1.L.gif |
| 450 |
|
| 451 |
//! \brief lambert_w0_small_z uses a tag_type to select a variant depending on the size of the type. |
| 452 |
//! The Lambert W is computed by lambert_w0_small_z for small z. |
| 453 |
//! The cutoff for z smallness determined by Tosio Fukushima by trial and error is (abs(z) < 0.05), |
| 454 |
//! but the optimum might be a function of the size of the type of z. |
| 455 |
|
| 456 |
//! \details |
| 457 |
//! The tag_type selection is based on the value @c std::numeric_limits<T>::max_digits10. |
| 458 |
//! This allows distinguishing between long double types that commonly vary between 64 and 80-bits, |
| 459 |
//! and also compilers that have a float type using 64 bits and/or long double using 128-bits. |
| 460 |
//! It assumes that max_digits10 is defined correctly or this might fail to make the correct selection. |
| 461 |
//! causing very small differences in computing lambert_w that would be very difficult to detect and diagnose. |
| 462 |
//! Cannot switch on @c std::numeric_limits<>::max() because comparison values may overflow the compiler limit. |
| 463 |
//! Cannot switch on @c std::numeric_limits<long double>::max_exponent10() |
| 464 |
//! because both 80 and 128 bit floating-point implementations use 11 bits for the exponent. |
| 465 |
//! So must rely on @c std::numeric_limits<long double>::max_digits10. |
| 466 |
|
| 467 |
//! Specialization of float zero series expansion used for small z (abs(z) < 0.05). |
| 468 |
//! Specializations of lambert_w0_small_z for built-in types. |
| 469 |
//! These specializations should be chosen in preference to T version. |
| 470 |
//! For example: lambert_w0_small_z(0.001F) should use the float version. |
| 471 |
//! (Parameter Policy is not used by built-in types when all terms are used during an inline computation, |
| 472 |
//! but for the tag_type selection to work, they all must include Policy in their signature. |
| 473 |
|
| 474 |
// Forward declaration of variants of lambert_w0_small_z. |
| 475 |
template <typename T, typename Policy> |
| 476 |
T lambert_w0_small_z(T x, const Policy&, std::integral_constant<int, 0> const&); // for float (32-bit) type. |
| 477 |
|
| 478 |
template <typename T, typename Policy> |
| 479 |
T lambert_w0_small_z(T x, const Policy&, std::integral_constant<int, 1> const&); // for double (64-bit) type. |
| 480 |
|
| 481 |
template <typename T, typename Policy> |
| 482 |
T lambert_w0_small_z(T x, const Policy&, std::integral_constant<int, 2> const&); // for long double (double extended 80-bit) type. |
| 483 |
|
| 484 |
template <typename T, typename Policy> |
| 485 |
T lambert_w0_small_z(T x, const Policy&, std::integral_constant<int, 3> const&); // for long double (128-bit) type. |
| 486 |
|
| 487 |
template <typename T, typename Policy> |
| 488 |
T lambert_w0_small_z(T x, const Policy&, std::integral_constant<int, 4> const&); // for float128 quadmath Q type. |
| 489 |
|
| 490 |
template <typename T, typename Policy> |
| 491 |
T lambert_w0_small_z(T x, const Policy&, std::integral_constant<int, 5> const&); // Generic multiprecision T. |
| 492 |
// Set tag_type depending on max_digits10. |
| 493 |
template <typename T, typename Policy> |
| 494 |
T lambert_w0_small_z(T x, const Policy& pol) |
| 495 |
{ //std::numeric_limits<T>::max_digits10 == 36 ? 3 : // 128-bit long double. |
| 496 |
using tag_type = std::integral_constant<int, |
| 497 |
std::numeric_limits<T>::is_specialized == 0 ? 5 : |
| 498 |
#ifndef BOOST_NO_CXX11_NUMERIC_LIMITS |
| 499 |
std::numeric_limits<T>::max_digits10 <= 9 ? 0 : // for float 32-bit. |
| 500 |
std::numeric_limits<T>::max_digits10 <= 17 ? 1 : // for double 64-bit. |
| 501 |
std::numeric_limits<T>::max_digits10 <= 22 ? 2 : // for 80-bit double extended. |
| 502 |
std::numeric_limits<T>::max_digits10 < 37 ? 4 // for both 128-bit long double (3) and 128-bit quad suffix Q type (4). |
| 503 |
#else |
| 504 |
std::numeric_limits<T>::radix != 2 ? 5 : |
| 505 |
std::numeric_limits<T>::digits <= 24 ? 0 : // for float 32-bit. |
| 506 |
std::numeric_limits<T>::digits <= 53 ? 1 : // for double 64-bit. |
| 507 |
std::numeric_limits<T>::digits <= 64 ? 2 : // for 80-bit double extended. |
| 508 |
std::numeric_limits<T>::digits <= 113 ? 4 // for both 128-bit long double (3) and 128-bit quad suffix Q type (4). |
| 509 |
#endif |
| 510 |
: 5>; // All Generic multiprecision types. |
| 511 |
// std::cout << "\ntag type = " << tag_type << std::endl; // error C2275: 'tag_type': illegal use of this type as an expression. |
| 512 |
return lambert_w0_small_z(x, pol, tag_type()); |
| 513 |
} // template <typename T> T lambert_w0_small_z(T x) |
| 514 |
|
| 515 |
//! Specialization of float (32-bit) series expansion used for small z (abs(z) < 0.05). |
| 516 |
// Only 9 Coefficients are computed to 21 decimal digits precision, ample for 32-bit float used by most platforms. |
| 517 |
// Taylor series coefficients used are computed by Wolfram to 50 decimal digits using instruction |
| 518 |
// N[InverseSeries[Series[z Exp[z],{z,0,34}]],50], |
| 519 |
// as proposed by Tosio Fukushima and implemented by Darko Veberic. |
| 520 |
|
| 521 |
template <typename T, typename Policy> |
| 522 |
T lambert_w0_small_z(T z, const Policy&, std::integral_constant<int, 0> const&) |
| 523 |
{ |
| 524 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 525 |
std::streamsize prec = std::cout.precision(std::numeric_limits<T>::max_digits10); // Save. |
| 526 |
std::cout << "\ntag_type 0 float lambert_w0_small_z called with z = " << z << " using " << 9 << " terms of precision " |
| 527 |
<< std::numeric_limits<float>::max_digits10 << " decimal digits. " << std::endl; |
| 528 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 529 |
T result = |
| 530 |
z * (1 - // j1 z^1 term = 1 |
| 531 |
z * (1 - // j2 z^2 term = -1 |
| 532 |
z * (static_cast<float>(3uLL) / 2uLL - // 3/2 // j3 z^3 term = 1.5. |
| 533 |
z * (2.6666666666666666667F - // 8/3 // j4 |
| 534 |
z * (5.2083333333333333333F - // -125/24 // j5 |
| 535 |
z * (10.8F - // j6 |
| 536 |
z * (23.343055555555555556F - // j7 |
| 537 |
z * (52.012698412698412698F - // j8 |
| 538 |
z * 118.62522321428571429F)))))))); // j9 |
| 539 |
|
| 540 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 541 |
std::cout << "return w = " << result << std::endl; |
| 542 |
std::cout.precision(prec); // Restore. |
| 543 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 544 |
|
| 545 |
return result; |
| 546 |
} // template <typename T> T lambert_w0_small_z(T x, std::integral_constant<int, 0> const&) |
| 547 |
|
| 548 |
//! Specialization of double (64-bit double) series expansion used for small z (abs(z) < 0.05). |
| 549 |
// 17 Coefficients are computed to 21 decimal digits precision suitable for 64-bit double used by most platforms. |
| 550 |
// Taylor series coefficients used are computed by Wolfram to 50 decimal digits using instruction |
| 551 |
// N[InverseSeries[Series[z Exp[z],{z,0,34}]],50], as proposed by Tosio Fukushima and implemented by Veberic. |
| 552 |
|
| 553 |
template <typename T, typename Policy> |
| 554 |
T lambert_w0_small_z(const T z, const Policy&, std::integral_constant<int, 1> const&) |
| 555 |
{ |
| 556 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 557 |
std::streamsize prec = std::cout.precision(std::numeric_limits<T>::max_digits10); // Save. |
| 558 |
std::cout << "\ntag_type 1 double lambert_w0_small_z called with z = " << z << " using " << 17 << " terms of precision, " |
| 559 |
<< std::numeric_limits<double>::max_digits10 << " decimal digits. " << std::endl; |
| 560 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 561 |
T result = |
| 562 |
z * (1. - // j1 z^1 |
| 563 |
z * (1. - // j2 z^2 |
| 564 |
z * (1.5 - // 3/2 // j3 z^3 |
| 565 |
z * (2.6666666666666666667 - // 8/3 // j4 |
| 566 |
z * (5.2083333333333333333 - // -125/24 // j5 |
| 567 |
z * (10.8 - // j6 |
| 568 |
z * (23.343055555555555556 - // j7 |
| 569 |
z * (52.012698412698412698 - // j8 |
| 570 |
z * (118.62522321428571429 - // j9 |
| 571 |
z * (275.57319223985890653 - // j10 |
| 572 |
z * (649.78717234347442681 - // j11 |
| 573 |
z * (1551.1605194805194805 - // j12 |
| 574 |
z * (3741.4497029592385495 - // j13 |
| 575 |
z * (9104.5002411580189358 - // j14 |
| 576 |
z * (22324.308512706601434 - // j15 |
| 577 |
z * (55103.621972903835338 - // j16 |
| 578 |
z * 136808.86090394293563)))))))))))))))); // j17 z^17 |
| 579 |
|
| 580 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 581 |
std::cout << "return w = " << result << std::endl; |
| 582 |
std::cout.precision(prec); // Restore. |
| 583 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 584 |
|
| 585 |
return result; |
| 586 |
} // T lambert_w0_small_z(const T z, std::integral_constant<int, 1> const&) |
| 587 |
|
| 588 |
//! Specialization of long double (80-bit double extended) series expansion used for small z (abs(z) < 0.05). |
| 589 |
// 21 Coefficients are computed to 21 decimal digits precision suitable for 80-bit long double used by some |
| 590 |
// platforms including GCC and Clang when generating for Intel X86 floating-point processors with 80-bit operations enabled (the default). |
| 591 |
// (This is NOT used by Microsoft Visual Studio where double and long always both use only 64-bit type. |
| 592 |
// Nor used for 128-bit float128.) |
| 593 |
template <typename T, typename Policy> |
| 594 |
T lambert_w0_small_z(const T z, const Policy&, std::integral_constant<int, 2> const&) |
| 595 |
{ |
| 596 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 597 |
std::streamsize precision = std::cout.precision(std::numeric_limits<T>::max_digits10); // Save. |
| 598 |
std::cout << "\ntag_type 2 long double (80-bit double extended) lambert_w0_small_z called with z = " << z << " using " << 21 << " terms of precision, " |
| 599 |
<< std::numeric_limits<long double>::max_digits10 << " decimal digits. " << std::endl; |
| 600 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 601 |
// T result = |
| 602 |
// z * (1.L - // j1 z^1 |
| 603 |
// z * (1.L - // j2 z^2 |
| 604 |
// z * (1.5L - // 3/2 // j3 |
| 605 |
// z * (2.6666666666666666667L - // 8/3 // j4 |
| 606 |
// z * (5.2083333333333333333L - // -125/24 // j5 |
| 607 |
// z * (10.800000000000000000L - // j6 |
| 608 |
// z * (23.343055555555555556L - // j7 |
| 609 |
// z * (52.012698412698412698L - // j8 |
| 610 |
// z * (118.62522321428571429L - // j9 |
| 611 |
// z * (275.57319223985890653L - // j10 |
| 612 |
// z * (649.78717234347442681L - // j11 |
| 613 |
// z * (1551.1605194805194805L - // j12 |
| 614 |
// z * (3741.4497029592385495L - // j13 |
| 615 |
// z * (9104.5002411580189358L - // j14 |
| 616 |
// z * (22324.308512706601434L - // j15 |
| 617 |
// z * (55103.621972903835338L - // j16 |
| 618 |
// z * (136808.86090394293563L - // j17 z^17 last term used by Fukushima double. |
| 619 |
// z * (341422.050665838363317L - // z^18 |
| 620 |
// z * (855992.9659966075514633L - // z^19 |
| 621 |
// z * (2.154990206091088289321e6L - // z^20 |
| 622 |
// z * 5.4455529223144624316423e6L // z^21 |
| 623 |
// )))))))))))))))))))); |
| 624 |
// |
| 625 |
|
| 626 |
T result = |
| 627 |
z * (1.L - // z j1 |
| 628 |
z * (1.L - // z^2 |
| 629 |
z * (1.500000000000000000000000000000000L - // z^3 |
| 630 |
z * (2.666666666666666666666666666666666L - // z ^ 4 |
| 631 |
z * (5.208333333333333333333333333333333L - // z ^ 5 |
| 632 |
z * (10.80000000000000000000000000000000L - // z ^ 6 |
| 633 |
z * (23.34305555555555555555555555555555L - // z ^ 7 |
| 634 |
z * (52.01269841269841269841269841269841L - // z ^ 8 |
| 635 |
z * (118.6252232142857142857142857142857L - // z ^ 9 |
| 636 |
z * (275.5731922398589065255731922398589L - // z ^ 10 |
| 637 |
z * (649.7871723434744268077601410934744L - // z ^ 11 |
| 638 |
z * (1551.160519480519480519480519480519L - // z ^ 12 |
| 639 |
z * (3741.449702959238549516327294105071L - //z ^ 13 |
| 640 |
z * (9104.500241158018935796713574491352L - // z ^ 14 |
| 641 |
z * (22324.308512706601434280005708577137L - // z ^ 15 |
| 642 |
z * (55103.621972903835337697771560205422L - // z ^ 16 |
| 643 |
z * (136808.86090394293563342215789305736L - // z ^ 17 |
| 644 |
z * (341422.05066583836331735491399356945L - // z^18 |
| 645 |
z * (855992.9659966075514633630250633224L - // z^19 |
| 646 |
z * (2.154990206091088289321708745358647e6L // z^20 distance -5 without term 20 |
| 647 |
)))))))))))))))))))); |
| 648 |
|
| 649 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 650 |
std::cout << "return w = " << result << std::endl; |
| 651 |
std::cout.precision(precision); // Restore. |
| 652 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 653 |
return result; |
| 654 |
} // long double lambert_w0_small_z(const T z, std::integral_constant<int, 1> const&) |
| 655 |
|
| 656 |
//! Specialization of 128-bit long double series expansion used for small z (abs(z) < 0.05). |
| 657 |
// 34 Taylor series coefficients used are computed by Wolfram to 50 decimal digits using instruction |
| 658 |
// N[InverseSeries[Series[z Exp[z],{z,0,34}]],50], |
| 659 |
// and are suffixed by L as they are assumed of type long double. |
| 660 |
// (This is NOT used for 128-bit quad boost::multiprecision::float128 type which required a suffix Q |
| 661 |
// nor multiprecision type cpp_bin_float_quad that can only be initialized at full precision of the type |
| 662 |
// constructed with a decimal digit string like "2.6666666666666666666666666666666666666666666666667".) |
| 663 |
|
| 664 |
template <typename T, typename Policy> |
| 665 |
T lambert_w0_small_z(const T z, const Policy&, std::integral_constant<int, 3> const&) |
| 666 |
{ |
| 667 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 668 |
std::streamsize precision = std::cout.precision(std::numeric_limits<T>::max_digits10); // Save. |
| 669 |
std::cout << "\ntag_type 3 long double (128-bit) lambert_w0_small_z called with z = " << z << " using " << 17 << " terms of precision, " |
| 670 |
<< std::numeric_limits<double>::max_digits10 << " decimal digits. " << std::endl; |
| 671 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 672 |
T result = |
| 673 |
z * (1.L - // j1 |
| 674 |
z * (1.L - // j2 |
| 675 |
z * (1.5L - // 3/2 // j3 |
| 676 |
z * (2.6666666666666666666666666666666666L - // 8/3 // j4 |
| 677 |
z * (5.2052083333333333333333333333333333L - // -125/24 // j5 |
| 678 |
z * (10.800000000000000000000000000000000L - // j6 |
| 679 |
z * (23.343055555555555555555555555555555L - // j7 |
| 680 |
z * (52.0126984126984126984126984126984126L - // j8 |
| 681 |
z * (118.625223214285714285714285714285714L - // j9 |
| 682 |
z * (275.57319223985890652557319223985890L - // * z ^ 10 - // j10 |
| 683 |
z * (649.78717234347442680776014109347442680776014109347L - // j11 |
| 684 |
z * (1551.1605194805194805194805194805194805194805194805L - // j12 |
| 685 |
z * (3741.4497029592385495163272941050718828496606274384L - // j13 |
| 686 |
z * (9104.5002411580189357967135744913522691300469078247L - // j14 |
| 687 |
z * (22324.308512706601434280005708577137148565719994291L - // j15 |
| 688 |
z * (55103.621972903835337697771560205422639285073147507L - // j16 |
| 689 |
z * 136808.86090394293563342215789305736395683485630576L // j17 |
| 690 |
)))))))))))))))); |
| 691 |
|
| 692 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 693 |
std::cout << "return w = " << result << std::endl; |
| 694 |
std::cout.precision(precision); // Restore. |
| 695 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 696 |
return result; |
| 697 |
} // T lambert_w0_small_z(const T z, std::integral_constant<int, 3> const&) |
| 698 |
|
| 699 |
//! Specialization of 128-bit quad series expansion used for small z (abs(z) < 0.05). |
| 700 |
// 34 Taylor series coefficients used were computed by Wolfram to 50 decimal digits using instruction |
| 701 |
// N[InverseSeries[Series[z Exp[z],{z,0,34}]],50], |
| 702 |
// and are suffixed by Q as they are assumed of type quad. |
| 703 |
// This could be used for 128-bit quad (which requires a suffix Q for full precision). |
| 704 |
// But experiments with GCC 7.2.0 show that while this gives full 128-bit precision |
| 705 |
// when the -f-ext-numeric-literals option is in force and the libquadmath library available, |
| 706 |
// over the range -0.049 to +0.049, |
| 707 |
// it is slightly slower than getting a double approximation followed by a single Halley step. |
| 708 |
|
| 709 |
#ifdef BOOST_HAS_FLOAT128 |
| 710 |
template <typename T, typename Policy> |
| 711 |
T lambert_w0_small_z(const T z, const Policy&, std::integral_constant<int, 4> const&) |
| 712 |
{ |
| 713 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 714 |
std::streamsize precision = std::cout.precision(std::numeric_limits<T>::max_digits10); // Save. |
| 715 |
std::cout << "\ntag_type 4 128-bit quad float128 lambert_w0_small_z called with z = " << z << " using " << 34 << " terms of precision, " |
| 716 |
<< std::numeric_limits<float128>::max_digits10 << " max decimal digits." << std::endl; |
| 717 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 718 |
T result = |
| 719 |
z * (1.Q - // z j1 |
| 720 |
z * (1.Q - // z^2 |
| 721 |
z * (1.500000000000000000000000000000000Q - // z^3 |
| 722 |
z * (2.666666666666666666666666666666666Q - // z ^ 4 |
| 723 |
z * (5.208333333333333333333333333333333Q - // z ^ 5 |
| 724 |
z * (10.80000000000000000000000000000000Q - // z ^ 6 |
| 725 |
z * (23.34305555555555555555555555555555Q - // z ^ 7 |
| 726 |
z * (52.01269841269841269841269841269841Q - // z ^ 8 |
| 727 |
z * (118.6252232142857142857142857142857Q - // z ^ 9 |
| 728 |
z * (275.5731922398589065255731922398589Q - // z ^ 10 |
| 729 |
z * (649.7871723434744268077601410934744Q - // z ^ 11 |
| 730 |
z * (1551.160519480519480519480519480519Q - // z ^ 12 |
| 731 |
z * (3741.449702959238549516327294105071Q - //z ^ 13 |
| 732 |
z * (9104.500241158018935796713574491352Q - // z ^ 14 |
| 733 |
z * (22324.308512706601434280005708577137Q - // z ^ 15 |
| 734 |
z * (55103.621972903835337697771560205422Q - // z ^ 16 |
| 735 |
z * (136808.86090394293563342215789305736Q - // z ^ 17 |
| 736 |
z * (341422.05066583836331735491399356945Q - // z^18 |
| 737 |
z * (855992.9659966075514633630250633224Q - // z^19 |
| 738 |
z * (2.154990206091088289321708745358647e6Q - // 20 |
| 739 |
z * (5.445552922314462431642316420035073e6Q - // 21 |
| 740 |
z * (1.380733000216662949061923813184508e7Q - // 22 |
| 741 |
z * (3.511704498513923292853869855945334e7Q - // 23 |
| 742 |
z * (8.956800256102797693072819557780090e7Q - // 24 |
| 743 |
z * (2.290416846187949813964782641734774e8Q - // 25 |
| 744 |
z * (5.871035041171798492020292225245235e8Q - // 26 |
| 745 |
z * (1.508256053857792919641317138812957e9Q - // 27 |
| 746 |
z * (3.882630161293188940385873468413841e9Q - // 28 |
| 747 |
z * (1.001394313665482968013913601565723e10Q - // 29 |
| 748 |
z * (2.587356736265760638992878359024929e10Q - // 30 |
| 749 |
z * (6.696209709358073856946120522333454e10Q - // 31 |
| 750 |
z * (1.735711659599198077777078238043644e11Q - // 32 |
| 751 |
z * (4.505680465642353886756098108484670e11Q - // 33 |
| 752 |
z * (1.171223178256487391904047636564823e12Q //z^34 |
| 753 |
)))))))))))))))))))))))))))))))))); |
| 754 |
|
| 755 |
|
| 756 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 757 |
std::cout << "return w = " << result << std::endl; |
| 758 |
std::cout.precision(precision); // Restore. |
| 759 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 760 |
|
| 761 |
return result; |
| 762 |
} // T lambert_w0_small_z(const T z, std::integral_constant<int, 4> const&) float128 |
| 763 |
|
| 764 |
#else |
| 765 |
|
| 766 |
template <typename T, typename Policy> |
| 767 |
inline T lambert_w0_small_z(const T z, const Policy& pol, std::integral_constant<int, 4> const&) // LCOV_EXCL_LINE body is covered, strangley this line is not. |
| 768 |
{ |
| 769 |
return lambert_w0_small_z(z, pol, std::integral_constant<int, 5>()); |
| 770 |
} |
| 771 |
|
| 772 |
#endif // BOOST_HAS_FLOAT128 |
| 773 |
|
| 774 |
//! Series functor to compute series term using pow and factorial. |
| 775 |
//! \details Functor is called after evaluating polynomial with the coefficients as rationals below. |
| 776 |
template <typename T> |
| 777 |
struct lambert_w0_small_z_series_term |
| 778 |
{ |
| 779 |
using result_type = T; |
| 780 |
//! \param _z Lambert W argument z. |
| 781 |
//! \param -term -pow<18>(z) / 6402373705728000uLL |
| 782 |
//! \param _k number of terms == initially 18 |
| 783 |
|
| 784 |
// Note *after* evaluating N terms, its internal state has k = N and term = (-1)^N z^N. |
| 785 |
|
| 786 |
lambert_w0_small_z_series_term(T _z, T _term, int _k) |
| 787 |
: k(_k), z(_z), term(_term) { } |
| 788 |
|
| 789 |
T operator()() |
| 790 |
{ // Called by sum_series until needs precision set by factor (policy::get_epsilon). |
| 791 |
using std::pow; |
| 792 |
++k; |
| 793 |
term *= -z / k; |
| 794 |
//T t = pow(z, k) * pow(T(k), -1 + k) / factorial<T>(k); // (z^k * k(k-1)^k) / k! |
| 795 |
T result = term * pow(T(k), T(-1 + k)); // term * k^(k-1) |
| 796 |
// std::cout << " k = " << k << ", term = " << term << ", result = " << result << std::endl; |
| 797 |
return result; // |
| 798 |
} |
| 799 |
private: |
| 800 |
int k; |
| 801 |
T z; |
| 802 |
T term; |
| 803 |
}; // template <typename T> struct lambert_w0_small_z_series_term |
| 804 |
|
| 805 |
//! Generic variant for T a User-defined types like Boost.Multiprecision. |
| 806 |
template <typename T, typename Policy> |
| 807 |
inline T lambert_w0_small_z(T z, const Policy& pol, std::integral_constant<int, 5> const&) |
| 808 |
{ |
| 809 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 810 |
std::streamsize precision = std::cout.precision(std::numeric_limits<T>::max_digits10); // Save. |
| 811 |
std::cout << "Generic lambert_w0_small_z called with z = " << z << " using as many terms needed for precision." << std::endl; |
| 812 |
std::cout << "Argument z is of type " << typeid(T).name() << std::endl; |
| 813 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 814 |
|
| 815 |
// First several terms of the series are tabulated and evaluated as a polynomial: |
| 816 |
// this will save us a bunch of expensive calls to pow. |
| 817 |
// Then our series functor is initialized "as if" it had already reached term 18, |
| 818 |
// enough evaluation of built-in 64-bit double and float (and 80-bit long double?) types. |
| 819 |
|
| 820 |
// Coefficients should be stored such that the coefficients for the x^i terms are in poly[i]. |
| 821 |
static const T coeff[] = |
| 822 |
{ |
| 823 |
0, // z^0 Care: zeroth term needed by tools::evaluate_polynomial, but not in the Wolfram equation, so indexes are one different! |
| 824 |
1, // z^1 term. |
| 825 |
-1, // z^2 term |
| 826 |
static_cast<T>(3uLL) / 2uLL, // z^3 term. |
| 827 |
-static_cast<T>(8uLL) / 3uLL, // z^4 |
| 828 |
static_cast<T>(125uLL) / 24uLL, // z^5 |
| 829 |
-static_cast<T>(54uLL) / 5uLL, // z^6 |
| 830 |
static_cast<T>(16807uLL) / 720uLL, // z^7 |
| 831 |
-static_cast<T>(16384uLL) / 315uLL, // z^8 |
| 832 |
static_cast<T>(531441uLL) / 4480uLL, // z^9 |
| 833 |
-static_cast<T>(156250uLL) / 567uLL, // z^10 |
| 834 |
static_cast<T>(2357947691uLL) / 3628800uLL, // z^11 |
| 835 |
-static_cast<T>(2985984uLL) / 1925uLL, // z^12 |
| 836 |
static_cast<T>(1792160394037uLL) / 479001600uLL, // z^13 |
| 837 |
-static_cast<T>(7909306972uLL) / 868725uLL, // z^14 |
| 838 |
static_cast<T>(320361328125uLL) / 14350336uLL, // z^15 |
| 839 |
-static_cast<T>(35184372088832uLL) / 638512875uLL, // z^16 |
| 840 |
static_cast<T>(2862423051509815793uLL) / 20922789888000uLL, // z^17 term |
| 841 |
-static_cast<T>(5083731656658uLL) / 14889875uLL, |
| 842 |
// z^18 term. = 136808.86090394293563342215789305735851647769682393 |
| 843 |
|
| 844 |
// z^18 is biggest that can be computed as rational using the largest possible uLL integers, |
| 845 |
// so higher terms cannot be potentially compiler-computed as uLL rationals. |
| 846 |
// Wolfram (5083731656658 z ^ 18) / 14889875 or |
| 847 |
// -341422.05066583836331735491399356945575432970390954 z^18 |
| 848 |
|
| 849 |
// See note below calling the functor to compute another term, |
| 850 |
// sufficient for 80-bit long double precision. |
| 851 |
// Wolfram -341422.05066583836331735491399356945575432970390954 z^19 term. |
| 852 |
// (5480386857784802185939 z^19)/6402373705728000 |
| 853 |
// But now this variant is not used to compute long double |
| 854 |
// as specializations are provided above. |
| 855 |
}; // static const T coeff[] |
| 856 |
|
| 857 |
/* |
| 858 |
Table of 19 computed coefficients: |
| 859 |
|
| 860 |
#0 0 |
| 861 |
#1 1 |
| 862 |
#2 -1 |
| 863 |
#3 1.5 |
| 864 |
#4 -2.6666666666666666666666666666666665382713370408509 |
| 865 |
#5 5.2083333333333333333333333333333330765426740817019 |
| 866 |
#6 -10.800000000000000000000000000000000616297582203915 |
| 867 |
#7 23.343055555555555555555555555555555076212991619177 |
| 868 |
#8 -52.012698412698412698412698412698412659282693193402 |
| 869 |
#9 118.62522321428571428571428571428571146835390992496 |
| 870 |
#10 -275.57319223985890652557319223985891400375196748314 |
| 871 |
#11 649.7871723434744268077601410934743969785223845882 |
| 872 |
#12 -1551.1605194805194805194805194805194947599566007429 |
| 873 |
#13 3741.4497029592385495163272941050719510009019331763 |
| 874 |
#14 -9104.5002411580189357967135744913524243896052869184 |
| 875 |
#15 22324.308512706601434280005708577137322392070452582 |
| 876 |
#16 -55103.621972903835337697771560205423203318720697224 |
| 877 |
#17 136808.86090394293563342215789305735851647769682393 |
| 878 |
136808.86090394293563342215789305735851647769682393 == Exactly same as Wolfram computed value. |
| 879 |
#18 -341422.05066583836331735491399356947486381600607416 |
| 880 |
341422.05066583836331735491399356945575432970390954 z^19 Wolfram value differs at 36 decimal digit, as expected. |
| 881 |
*/ |
| 882 |
|
| 883 |
using boost::math::policies::get_epsilon; // for type T. |
| 884 |
using boost::math::tools::sum_series; |
| 885 |
using boost::math::tools::evaluate_polynomial; |
| 886 |
// http://www.boost.org/doc/libs/release/libs/math/doc/html/math_toolkit/roots/rational.html |
| 887 |
|
| 888 |
// std::streamsize prec = std::cout.precision(std::numeric_limits <T>::max_digits10); |
| 889 |
|
| 890 |
T result = evaluate_polynomial(coeff, z); // LCOV_EXCL_LINE next line covered but not this one strangely - GCOV SNAFU? |
| 891 |
// template <std::size_t N, typename T, typename V> |
| 892 |
// V evaluate_polynomial(const T(&poly)[N], const V& val); |
| 893 |
// Size of coeff found from N |
| 894 |
//std::cout << "evaluate_polynomial(coeff, z); == " << result << std::endl; |
| 895 |
//std::cout << "result = " << result << std::endl; |
| 896 |
// It's an artefact of the way I wrote the functor: *after* evaluating N |
| 897 |
// terms, its internal state has k = N and term = (-1)^N z^N. So after |
| 898 |
// evaluating 18 terms, we initialize the functor to the term we've just |
| 899 |
// evaluated, and then when it's called, it increments itself to the next term. |
| 900 |
// So 18!is 6402373705728000, which is where that comes from. |
| 901 |
|
| 902 |
// The 19th coefficient of the polynomial is actually, 19 ^ 18 / 19!= |
| 903 |
// 104127350297911241532841 / 121645100408832000 which after removing GCDs |
| 904 |
// reduces down to Wolfram rational 5480386857784802185939 / 6402373705728000. |
| 905 |
// Wolfram z^19 term +(5480386857784802185939 z^19) /6402373705728000 |
| 906 |
// +855992.96599660755146336302506332246623424823099755 z^19 |
| 907 |
|
| 908 |
//! Evaluate Functor. |
| 909 |
lambert_w0_small_z_series_term<T> s(z, -pow<18>(z) / 6402373705728000uLL, 18); |
| 910 |
|
| 911 |
// Temporary to list the coefficients. |
| 912 |
//std::cout << " Table of coefficients" << std::endl; |
| 913 |
//std::streamsize saved_precision = std::cout.precision(50); |
| 914 |
//for (size_t i = 0; i != 19; i++) |
| 915 |
//{ |
| 916 |
// std::cout << "#" << i << " " << coeff[i] << std::endl; |
| 917 |
//} |
| 918 |
//std::cout.precision(saved_precision); |
| 919 |
|
| 920 |
std::uintmax_t max_iter = policies::get_max_series_iterations<Policy>(); // Max iterations from policy. |
| 921 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 922 |
std::cout << "max iter from policy = " << max_iter << std::endl; |
| 923 |
// // max iter from policy = 1000000 is default. |
| 924 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES |
| 925 |
|
| 926 |
result = sum_series(s, get_epsilon<T, Policy>(), max_iter, result); |
| 927 |
// result == evaluate_polynomial. |
| 928 |
//sum_series(Functor& func, int bits, std::uintmax_t& max_terms, const U& init_value) |
| 929 |
// std::cout << "sum_series(s, get_epsilon<T, Policy>(), max_iter, result); = " << result << std::endl; |
| 930 |
|
| 931 |
//T epsilon = get_epsilon<T, Policy>(); |
| 932 |
//std::cout << "epsilon from policy = " << epsilon << std::endl; |
| 933 |
// epsilon from policy = 1.93e-34 for T == quad |
| 934 |
// 5.35e-51 for t = cpp_bin_float_50 |
| 935 |
|
| 936 |
// std::cout << " get eps = " << get_epsilon<T, Policy>() << std::endl; // quad eps = 1.93e-34, bin_float_50 eps = 5.35e-51 |
| 937 |
policies::check_series_iterations<T>("boost::math::lambert_w0_small_z<%1%>(%1%)", max_iter, pol); |
| 938 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES_ITERATIONS |
| 939 |
std::cout << "z = " << z << " needed " << max_iter << " iterations." << std::endl; |
| 940 |
std::cout.precision(prec); // Restore. |
| 941 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W_SMALL_Z_SERIES_ITERATIONS |
| 942 |
return result; |
| 943 |
} // template <typename T, typename Policy> inline T lambert_w0_small_z_series(T z, const Policy& pol) |
| 944 |
|
| 945 |
// Approximate lambert_w0 (used for z values that are outside range of lookup table or rational functions) |
| 946 |
// Corless equation 4.19, page 349, and Chapeau-Blondeau equation 20, page 2162. |
| 947 |
template <typename T> |
| 948 |
inline T lambert_w0_approx(T z) |
| 949 |
{ |
| 950 |
BOOST_MATH_STD_USING |
| 951 |
T lz = log(z); |
| 952 |
T llz = log(lz); |
| 953 |
T w = lz - llz + (llz / lz); // Corless equation 4.19, page 349, and Chapeau-Blondeau equation 20, page 2162. |
| 954 |
return w; |
| 955 |
// std::cout << "w max " << max_w << std::endl; // double 703.227 |
| 956 |
} |
| 957 |
|
| 958 |
////////////////////////////////////////////////////////////////////////////////////////// |
| 959 |
|
| 960 |
//! \brief Lambert_w0 implementations for float, double and higher precisions. |
| 961 |
//! 3rd parameter used to select which version is used. |
| 962 |
|
| 963 |
//! /details Rational polynomials are provided for several range of argument z. |
| 964 |
//! For very small values of z, and for z very near the branch singularity at -e^-1 (~= -0.367879), |
| 965 |
//! two other series functions are used. |
| 966 |
|
| 967 |
//! float precision polynomials are used for 32-bit (usually float) precision (for speed) |
| 968 |
//! double precision polynomials are used for 64-bit (usually double) precision. |
| 969 |
//! For higher precisions, a 64-bit double approximation is computed first, |
| 970 |
//! and then refined using Halley iterations. |
| 971 |
|
| 972 |
template <typename T> |
| 973 |
inline T do_get_near_singularity_param(T z) |
| 974 |
{ |
| 975 |
BOOST_MATH_STD_USING |
| 976 |
const T p2 = 2 * (boost::math::constants::e<T>() * z + 1); |
| 977 |
const T p = sqrt(p2); |
| 978 |
return p; |
| 979 |
} |
| 980 |
template <typename T, typename Policy> |
| 981 |
inline T get_near_singularity_param(T z, const Policy) |
| 982 |
{ |
| 983 |
using value_type = typename policies::evaluation<T, Policy>::type; |
| 984 |
return static_cast<T>(do_get_near_singularity_param(static_cast<value_type>(z))); |
| 985 |
} |
| 986 |
|
| 987 |
// Forward declarations: |
| 988 |
|
| 989 |
//template <typename T, typename Policy> T lambert_w0_small_z(T z, const Policy& pol); |
| 990 |
//template <typename T, typename Policy> |
| 991 |
//T lambert_w0_imp(T w, const Policy& pol, const std::integral_constant<int, 0>&); // 32 bit usually float. |
| 992 |
//template <typename T, typename Policy> |
| 993 |
//T lambert_w0_imp(T w, const Policy& pol, const std::integral_constant<int, 1>&); // 64 bit usually double. |
| 994 |
//template <typename T, typename Policy> |
| 995 |
//T lambert_w0_imp(T w, const Policy& pol, const std::integral_constant<int, 2>&); // 80-bit long double. |
| 996 |
|
| 997 |
template <typename T> |
| 998 |
T lambert_w_positive_rational_float(T z) |
| 999 |
{ |
| 1000 |
BOOST_MATH_STD_USING |
| 1001 |
if (z < 2) |
| 1002 |
{ |
| 1003 |
if (z < T(0.5)) |
| 1004 |
{ // 0.05 < z < 0.5 |
| 1005 |
// Maximum Deviation Found: 2.993e-08 |
| 1006 |
// Expected Error Term : 2.993e-08 |
| 1007 |
// Maximum Relative Change in Control Points : 7.555e-04 Y offset : -8.196592331e-01 |
| 1008 |
// LCOV_EXCL_START |
| 1009 |
static const T Y = 8.196592331e-01f; |
| 1010 |
static const T P[] = { |
| 1011 |
1.803388345e-01f, |
| 1012 |
-4.820256838e-01f, |
| 1013 |
-1.068349741e+00f, |
| 1014 |
-3.506624319e-02f, |
| 1015 |
}; |
| 1016 |
static const T Q[] = { |
| 1017 |
1.000000000e+00f, |
| 1018 |
2.871703469e+00f, |
| 1019 |
1.690949264e+00f, |
| 1020 |
}; |
| 1021 |
// LCOV_EXCL_STOP |
| 1022 |
return z * (Y + boost::math::tools::evaluate_polynomial(P, z) / boost::math::tools::evaluate_polynomial(Q, z)); |
| 1023 |
} |
| 1024 |
else |
| 1025 |
{ // 0.5 < z < 2 |
| 1026 |
// Max error in interpolated form: 1.018e-08 |
| 1027 |
// LCOV_EXCL_START |
| 1028 |
static const T Y = 5.503368378e-01f; |
| 1029 |
static const T P[] = { |
| 1030 |
4.493332766e-01f, |
| 1031 |
2.543432707e-01f, |
| 1032 |
-4.808788799e-01f, |
| 1033 |
-1.244425316e-01f, |
| 1034 |
}; |
| 1035 |
static const T Q[] = { |
| 1036 |
1.000000000e+00f, |
| 1037 |
2.780661241e+00f, |
| 1038 |
1.830840318e+00f, |
| 1039 |
2.407221031e-01f, |
| 1040 |
}; |
| 1041 |
// LCOV_EXCL_STOP |
| 1042 |
return z * (Y + boost::math::tools::evaluate_rational(P, Q, z)); |
| 1043 |
} |
| 1044 |
} |
| 1045 |
else if (z < 6) |
| 1046 |
{ |
| 1047 |
// 2 < z < 6 |
| 1048 |
// Max error in interpolated form: 2.944e-08 |
| 1049 |
// LCOV_EXCL_START |
| 1050 |
static const T Y = 1.162393570e+00f; |
| 1051 |
static const T P[] = { |
| 1052 |
-1.144183394e+00f, |
| 1053 |
-4.712732855e-01f, |
| 1054 |
1.563162512e-01f, |
| 1055 |
1.434010911e-02f, |
| 1056 |
}; |
| 1057 |
static const T Q[] = { |
| 1058 |
1.000000000e+00f, |
| 1059 |
1.192626340e+00f, |
| 1060 |
2.295580708e-01f, |
| 1061 |
5.477869455e-03f, |
| 1062 |
}; |
| 1063 |
// LCOV_EXCL_STOP |
| 1064 |
return Y + boost::math::tools::evaluate_rational(P, Q, z); |
| 1065 |
} |
| 1066 |
else if (z < 18) |
| 1067 |
{ |
| 1068 |
// 6 < z < 18 |
| 1069 |
// Max error in interpolated form: 5.893e-08 |
| 1070 |
// LCOV_EXCL_START |
| 1071 |
static const T Y = 1.809371948e+00f; |
| 1072 |
static const T P[] = { |
| 1073 |
-1.689291769e+00f, |
| 1074 |
-3.337812742e-01f, |
| 1075 |
3.151434873e-02f, |
| 1076 |
1.134178734e-03f, |
| 1077 |
}; |
| 1078 |
static const T Q[] = { |
| 1079 |
1.000000000e+00f, |
| 1080 |
5.716915685e-01f, |
| 1081 |
4.489521292e-02f, |
| 1082 |
4.076716763e-04f, |
| 1083 |
}; |
| 1084 |
// LCOV_EXCL_STOP |
| 1085 |
return Y + boost::math::tools::evaluate_rational(P, Q, z); |
| 1086 |
} |
| 1087 |
else if (z < T(9897.12905874)) // 2.8 < log(z) < 9.2 |
| 1088 |
{ |
| 1089 |
// Max error in interpolated form: 1.771e-08 |
| 1090 |
// LCOV_EXCL_START |
| 1091 |
static const T Y = -1.402973175e+00f; |
| 1092 |
static const T P[] = { |
| 1093 |
1.966174312e+00f, |
| 1094 |
2.350864728e-01f, |
| 1095 |
-5.098074353e-02f, |
| 1096 |
-1.054818339e-02f, |
| 1097 |
}; |
| 1098 |
static const T Q[] = { |
| 1099 |
1.000000000e+00f, |
| 1100 |
4.388208264e-01f, |
| 1101 |
8.316639634e-02f, |
| 1102 |
3.397187918e-03f, |
| 1103 |
-1.321489743e-05f, |
| 1104 |
}; |
| 1105 |
// LCOV_EXCL_STOP |
| 1106 |
T log_w = log(z); |
| 1107 |
return log_w + Y + boost::math::tools::evaluate_polynomial(P, log_w) / boost::math::tools::evaluate_polynomial(Q, log_w); |
| 1108 |
} |
| 1109 |
else if (z < T(7.896296e+13)) // 9.2 < log(z) <= 32 |
| 1110 |
{ |
| 1111 |
// Max error in interpolated form: 5.821e-08 |
| 1112 |
// LCOV_EXCL_START |
| 1113 |
static const T Y = -2.735729218e+00f; |
| 1114 |
static const T P[] = { |
| 1115 |
3.424903470e+00f, |
| 1116 |
7.525631787e-02f, |
| 1117 |
-1.427309584e-02f, |
| 1118 |
-1.435974178e-05f, |
| 1119 |
}; |
| 1120 |
static const T Q[] = { |
| 1121 |
1.000000000e+00f, |
| 1122 |
2.514005579e-01f, |
| 1123 |
6.118994652e-03f, |
| 1124 |
-1.357889535e-05f, |
| 1125 |
7.312865624e-08f, |
| 1126 |
}; |
| 1127 |
// LCOV_EXCL_STOP |
| 1128 |
T log_w = log(z); |
| 1129 |
return log_w + Y + boost::math::tools::evaluate_polynomial(P, log_w) / boost::math::tools::evaluate_polynomial(Q, log_w); |
| 1130 |
} |
| 1131 |
|
| 1132 |
// Max error in interpolated form: 1.491e-08 |
| 1133 |
// LCOV_EXCL_START |
| 1134 |
static const T Y = -4.012863159e+00f; |
| 1135 |
static const T P[] = { |
| 1136 |
4.431629226e+00f, |
| 1137 |
2.756690487e-01f, |
| 1138 |
-2.992956930e-03f, |
| 1139 |
-4.912259384e-05f, |
| 1140 |
}; |
| 1141 |
static const T Q[] = { |
| 1142 |
1.000000000e+00f, |
| 1143 |
2.015434591e-01f, |
| 1144 |
4.949426142e-03f, |
| 1145 |
1.609659944e-05f, |
| 1146 |
-5.111523436e-09f, |
| 1147 |
}; |
| 1148 |
// LCOV_EXCL_STOP |
| 1149 |
T log_w = log(z); |
| 1150 |
return log_w + Y + boost::math::tools::evaluate_polynomial(P, log_w) / boost::math::tools::evaluate_polynomial(Q, log_w); |
| 1151 |
|
| 1152 |
} |
| 1153 |
|
| 1154 |
template <typename T, typename Policy> |
| 1155 |
T lambert_w_negative_rational_float(T z, const Policy& pol) |
| 1156 |
{ |
| 1157 |
BOOST_MATH_STD_USING |
| 1158 |
if (z > T(-0.27)) |
| 1159 |
{ |
| 1160 |
if (z < T(-0.051)) |
| 1161 |
{ |
| 1162 |
// -0.27 < z < -0.051 |
| 1163 |
// Max error in interpolated form: 5.080e-08 |
| 1164 |
// LCOV_EXCL_START |
| 1165 |
static const T Y = 1.255809784e+00f; |
| 1166 |
static const T P[] = { |
| 1167 |
-2.558083412e-01f, |
| 1168 |
-2.306524098e+00f, |
| 1169 |
-5.630887033e+00f, |
| 1170 |
-3.803974556e+00f, |
| 1171 |
}; |
| 1172 |
static const T Q[] = { |
| 1173 |
1.000000000e+00f, |
| 1174 |
5.107680783e+00f, |
| 1175 |
7.914062868e+00f, |
| 1176 |
3.501498501e+00f, |
| 1177 |
}; |
| 1178 |
// LCOV_EXCL_STOP |
| 1179 |
return z * (Y + boost::math::tools::evaluate_rational(P, Q, z)); |
| 1180 |
} |
| 1181 |
else |
| 1182 |
{ |
| 1183 |
// Very small z so use a series function. |
| 1184 |
return lambert_w0_small_z(z, pol); |
| 1185 |
} |
| 1186 |
} |
| 1187 |
else if (z > T(-0.3578794411714423215955237701)) |
| 1188 |
{ // Very close to branch singularity. |
| 1189 |
// Max error in interpolated form: 5.269e-08 |
| 1190 |
// LCOV_EXCL_START |
| 1191 |
static const T Y = 1.220928431e-01f; |
| 1192 |
static const T P[] = { |
| 1193 |
-1.221787446e-01f, |
| 1194 |
-6.816155875e+00f, |
| 1195 |
7.144582035e+01f, |
| 1196 |
1.128444390e+03f, |
| 1197 |
}; |
| 1198 |
static const T Q[] = { |
| 1199 |
1.000000000e+00f, |
| 1200 |
6.480326790e+01f, |
| 1201 |
1.869145243e+02f, |
| 1202 |
-1.361804274e+03f, |
| 1203 |
1.117826726e+03f, |
| 1204 |
}; |
| 1205 |
// LCOV_EXCL_STOP |
| 1206 |
T d = z + 0.367879441171442321595523770161460867445811f; |
| 1207 |
return -d / (Y + boost::math::tools::evaluate_polynomial(P, d) / boost::math::tools::evaluate_polynomial(Q, d)); |
| 1208 |
} |
| 1209 |
|
| 1210 |
return lambert_w_singularity_series(get_near_singularity_param(z, pol)); |
| 1211 |
} |
| 1212 |
|
| 1213 |
//! Lambert_w0 @b 'float' implementation, selected when T is 32-bit precision. |
| 1214 |
template <typename T, typename Policy> |
| 1215 |
inline T lambert_w0_imp(T z, const Policy& pol, const std::integral_constant<int, 1>&) |
| 1216 |
{ |
| 1217 |
static const char* function = "boost::math::lambert_w0<%1%>"; // For error messages. |
| 1218 |
BOOST_MATH_STD_USING // Aid ADL of std functions. |
| 1219 |
|
| 1220 |
if ((boost::math::isnan)(z)) |
| 1221 |
{ |
| 1222 |
return boost::math::policies::raise_domain_error<T>(function, "Expected a value > -e^-1 (-0.367879...) but got %1%.", z, pol); |
| 1223 |
} |
| 1224 |
if ((boost::math::isinf)(z)) |
| 1225 |
{ |
| 1226 |
return boost::math::policies::raise_overflow_error<T>(function, "Expected a finite value but got %1%.", z, pol); |
| 1227 |
} |
| 1228 |
|
| 1229 |
if (z >= T(0.05)) // Fukushima switch point. |
| 1230 |
// if (z >= 0.045) // 34 terms makes 128-bit 'exact' below 0.045. |
| 1231 |
{ // Normal ranges using several rational polynomials. |
| 1232 |
return lambert_w_positive_rational_float(z); |
| 1233 |
} |
| 1234 |
else if (z <= -0.3678794411714423215955237701614608674458111310f) |
| 1235 |
{ |
| 1236 |
if (z < -0.3678794411714423215955237701614608674458111310f) |
| 1237 |
return boost::math::policies::raise_domain_error<T>(function, "Expected z >= -e^-1 (-0.367879...) but got %1%.", z, pol); |
| 1238 |
return -1; |
| 1239 |
} |
| 1240 |
|
| 1241 |
return lambert_w_negative_rational_float(z, pol); |
| 1242 |
} // T lambert_w0_imp(T z, const Policy& pol, const std::integral_constant<int, 1>&) for 32-bit usually float. |
| 1243 |
|
| 1244 |
template <typename T> |
| 1245 |
T lambert_w_positive_rational_double(T z) |
| 1246 |
{ |
| 1247 |
BOOST_MATH_STD_USING |
| 1248 |
if (z < 2) |
| 1249 |
{ |
| 1250 |
if (z < 0.5) |
| 1251 |
{ |
| 1252 |
// Max error in interpolated form: 2.255e-17 |
| 1253 |
// LCOV_EXCL_START |
| 1254 |
static const T offset = 8.19659233093261719e-01; |
| 1255 |
static const T P[] = { |
| 1256 |
1.80340766906685177e-01, |
| 1257 |
3.28178241493119307e-01, |
| 1258 |
-2.19153620687139706e+00, |
| 1259 |
-7.24750929074563990e+00, |
| 1260 |
-7.28395876262524204e+00, |
| 1261 |
-2.57417169492512916e+00, |
| 1262 |
-2.31606948888704503e-01 |
| 1263 |
}; |
| 1264 |
static const T Q[] = { |
| 1265 |
1.00000000000000000e+00, |
| 1266 |
7.36482529307436604e+00, |
| 1267 |
2.03686007856430677e+01, |
| 1268 |
2.62864592096657307e+01, |
| 1269 |
1.59742041380858333e+01, |
| 1270 |
4.03760534788374589e+00, |
| 1271 |
2.91327346750475362e-01 |
| 1272 |
}; |
| 1273 |
// LCOV_EXCL_STOP |
| 1274 |
return z * (offset + boost::math::tools::evaluate_polynomial(P, z) / boost::math::tools::evaluate_polynomial(Q, z)); |
| 1275 |
} |
| 1276 |
else |
| 1277 |
{ |
| 1278 |
// Max error in interpolated form: 3.806e-18 |
| 1279 |
// LCOV_EXCL_START |
| 1280 |
static const T offset = 5.50335884094238281e-01; |
| 1281 |
static const T P[] = { |
| 1282 |
4.49664083944098322e-01, |
| 1283 |
1.90417666196776909e+00, |
| 1284 |
1.99951368798255994e+00, |
| 1285 |
-6.91217310299270265e-01, |
| 1286 |
-1.88533935998617058e+00, |
| 1287 |
-7.96743968047750836e-01, |
| 1288 |
-1.02891726031055254e-01, |
| 1289 |
-3.09156013592636568e-03 |
| 1290 |
}; |
| 1291 |
static const T Q[] = { |
| 1292 |
1.00000000000000000e+00, |
| 1293 |
6.45854489419584014e+00, |
| 1294 |
1.54739232422116048e+01, |
| 1295 |
1.72606164253337843e+01, |
| 1296 |
9.29427055609544096e+00, |
| 1297 |
2.29040824649748117e+00, |
| 1298 |
2.21610620995418981e-01, |
| 1299 |
5.70597669908194213e-03 |
| 1300 |
};// LCOV_EXCL_STOP |
| 1301 |
return z * (offset + boost::math::tools::evaluate_rational(P, Q, z)); |
| 1302 |
} |
| 1303 |
} |
| 1304 |
else if (z < 6) |
| 1305 |
{ |
| 1306 |
// 2 < z < 6 |
| 1307 |
// Max error in interpolated form: 1.216e-17 |
| 1308 |
// LCOV_EXCL_START |
| 1309 |
static const T Y = 1.16239356994628906e+00; |
| 1310 |
static const T P[] = { |
| 1311 |
-1.16230494982099475e+00, |
| 1312 |
-3.38528144432561136e+00, |
| 1313 |
-2.55653717293161565e+00, |
| 1314 |
-3.06755172989214189e-01, |
| 1315 |
1.73149743765268289e-01, |
| 1316 |
3.76906042860014206e-02, |
| 1317 |
1.84552217624706666e-03, |
| 1318 |
1.69434126904822116e-05, |
| 1319 |
}; |
| 1320 |
static const T Q[] = { |
| 1321 |
1.00000000000000000e+00, |
| 1322 |
3.77187616711220819e+00, |
| 1323 |
4.58799960260143701e+00, |
| 1324 |
2.24101228462292447e+00, |
| 1325 |
4.54794195426212385e-01, |
| 1326 |
3.60761772095963982e-02, |
| 1327 |
9.25176499518388571e-04, |
| 1328 |
4.43611344705509378e-06, |
| 1329 |
}; |
| 1330 |
// LCOV_EXCL_STOP |
| 1331 |
return Y + boost::math::tools::evaluate_rational(P, Q, z); |
| 1332 |
} |
| 1333 |
else if (z < 18) |
| 1334 |
{ |
| 1335 |
// 6 < z < 18 |
| 1336 |
// Max error in interpolated form: 1.985e-19 |
| 1337 |
// LCOV_EXCL_START |
| 1338 |
static const T offset = 1.80937194824218750e+00; |
| 1339 |
static const T P[] = |
| 1340 |
{ |
| 1341 |
-1.80690935424793635e+00, |
| 1342 |
-3.66995929380314602e+00, |
| 1343 |
-1.93842957940149781e+00, |
| 1344 |
-2.94269984375794040e-01, |
| 1345 |
1.81224710627677778e-03, |
| 1346 |
2.48166798603547447e-03, |
| 1347 |
1.15806592415397245e-04, |
| 1348 |
1.43105573216815533e-06, |
| 1349 |
3.47281483428369604e-09 |
| 1350 |
}; |
| 1351 |
static const T Q[] = { |
| 1352 |
1.00000000000000000e+00, |
| 1353 |
2.57319080723908597e+00, |
| 1354 |
1.96724528442680658e+00, |
| 1355 |
5.84501352882650722e-01, |
| 1356 |
7.37152837939206240e-02, |
| 1357 |
3.97368430940416778e-03, |
| 1358 |
8.54941838187085088e-05, |
| 1359 |
6.05713225608426678e-07, |
| 1360 |
8.17517283816615732e-10 |
| 1361 |
}; |
| 1362 |
// LCOV_EXCL_STOP |
| 1363 |
return offset + boost::math::tools::evaluate_rational(P, Q, z); |
| 1364 |
} |
| 1365 |
else if (z < 9897.12905874) // 2.8 < log(z) < 9.2 |
| 1366 |
{ |
| 1367 |
// Max error in interpolated form: 1.195e-18 |
| 1368 |
// LCOV_EXCL_START |
| 1369 |
static const T Y = -1.40297317504882812e+00; |
| 1370 |
static const T P[] = { |
| 1371 |
1.97011826279311924e+00, |
| 1372 |
1.05639945701546704e+00, |
| 1373 |
3.33434529073196304e-01, |
| 1374 |
3.34619153200386816e-02, |
| 1375 |
-5.36238353781326675e-03, |
| 1376 |
-2.43901294871308604e-03, |
| 1377 |
-2.13762095619085404e-04, |
| 1378 |
-4.85531936495542274e-06, |
| 1379 |
-2.02473518491905386e-08, |
| 1380 |
}; |
| 1381 |
static const T Q[] = { |
| 1382 |
1.00000000000000000e+00, |
| 1383 |
8.60107275833921618e-01, |
| 1384 |
4.10420467985504373e-01, |
| 1385 |
1.18444884081994841e-01, |
| 1386 |
2.16966505556021046e-02, |
| 1387 |
2.24529766630769097e-03, |
| 1388 |
9.82045090226437614e-05, |
| 1389 |
1.36363515125489502e-06, |
| 1390 |
3.44200749053237945e-09, |
| 1391 |
}; |
| 1392 |
// LCOV_EXCL_STOP |
| 1393 |
T log_w = log(z); |
| 1394 |
return log_w + Y + boost::math::tools::evaluate_rational(P, Q, log_w); |
| 1395 |
} |
| 1396 |
else if (z < 7.896296e+13) // 9.2 < log(z) <= 32 |
| 1397 |
{ |
| 1398 |
// Max error in interpolated form: 6.529e-18 |
| 1399 |
// LCOV_EXCL_START |
| 1400 |
static const T Y = -2.73572921752929688e+00; |
| 1401 |
static const T P[] = { |
| 1402 |
3.30547638424076217e+00, |
| 1403 |
1.64050071277550167e+00, |
| 1404 |
4.57149576470736039e-01, |
| 1405 |
4.03821227745424840e-02, |
| 1406 |
-4.99664976882514362e-04, |
| 1407 |
-1.28527893803052956e-04, |
| 1408 |
-2.95470325373338738e-06, |
| 1409 |
-1.76662025550202762e-08, |
| 1410 |
-1.98721972463709290e-11, |
| 1411 |
}; |
| 1412 |
static const T Q[] = { |
| 1413 |
1.00000000000000000e+00, |
| 1414 |
6.91472559412458759e-01, |
| 1415 |
2.48154578891676774e-01, |
| 1416 |
4.60893578284335263e-02, |
| 1417 |
3.60207838982301946e-03, |
| 1418 |
1.13001153242430471e-04, |
| 1419 |
1.33690948263488455e-06, |
| 1420 |
4.97253225968548872e-09, |
| 1421 |
3.39460723731970550e-12, |
| 1422 |
}; |
| 1423 |
// LCOV_EXCL_STOP |
| 1424 |
T log_w = log(z); |
| 1425 |
return log_w + Y + boost::math::tools::evaluate_rational(P, Q, log_w); |
| 1426 |
} |
| 1427 |
else if (z < 2.6881171e+43) // 32 < log(z) < 100 |
| 1428 |
{ |
| 1429 |
// Max error in interpolated form: 2.015e-18 |
| 1430 |
// LCOV_EXCL_START |
| 1431 |
static const T Y = -4.01286315917968750e+00; |
| 1432 |
static const T P[] = { |
| 1433 |
5.07714858354309672e+00, |
| 1434 |
-3.32994414518701458e+00, |
| 1435 |
-8.61170416909864451e-01, |
| 1436 |
-4.01139705309486142e-02, |
| 1437 |
-1.85374201771834585e-04, |
| 1438 |
1.08824145844270666e-05, |
| 1439 |
1.17216905810452396e-07, |
| 1440 |
2.97998248101385990e-10, |
| 1441 |
1.42294856434176682e-13, |
| 1442 |
}; |
| 1443 |
static const T Q[] = { |
| 1444 |
1.00000000000000000e+00, |
| 1445 |
-4.85840770639861485e-01, |
| 1446 |
-3.18714850604827580e-01, |
| 1447 |
-3.20966129264610534e-02, |
| 1448 |
-1.06276178044267895e-03, |
| 1449 |
-1.33597828642644955e-05, |
| 1450 |
-6.27900905346219472e-08, |
| 1451 |
-9.35271498075378319e-11, |
| 1452 |
-2.60648331090076845e-14, |
| 1453 |
}; |
| 1454 |
// LCOV_EXCL_STOP |
| 1455 |
T log_w = log(z); |
| 1456 |
return log_w + Y + boost::math::tools::evaluate_rational(P, Q, log_w); |
| 1457 |
} |
| 1458 |
else // 100 < log(z) < 710 |
| 1459 |
{ |
| 1460 |
// Max error in interpolated form: 5.277e-18 |
| 1461 |
// LCOV_EXCL_START |
| 1462 |
static const T Y = -5.70115661621093750e+00; |
| 1463 |
static const T P[] = { |
| 1464 |
6.42275660145116698e+00, |
| 1465 |
1.33047964073367945e+00, |
| 1466 |
6.72008923401652816e-02, |
| 1467 |
1.16444069958125895e-03, |
| 1468 |
7.06966760237470501e-06, |
| 1469 |
5.48974896149039165e-09, |
| 1470 |
-7.00379652018853621e-11, |
| 1471 |
-1.89247635913659556e-13, |
| 1472 |
-1.55898770790170598e-16, |
| 1473 |
-4.06109208815303157e-20, |
| 1474 |
-2.21552699006496737e-24, |
| 1475 |
}; |
| 1476 |
static const T Q[] = { |
| 1477 |
1.00000000000000000e+00, |
| 1478 |
3.34498588416632854e-01, |
| 1479 |
2.51519862456384983e-02, |
| 1480 |
6.81223810622416254e-04, |
| 1481 |
7.94450897106903537e-06, |
| 1482 |
4.30675039872881342e-08, |
| 1483 |
1.10667669458467617e-10, |
| 1484 |
1.31012240694192289e-13, |
| 1485 |
6.53282047177727125e-17, |
| 1486 |
1.11775518708172009e-20, |
| 1487 |
3.78250395617836059e-25, |
| 1488 |
}; |
| 1489 |
// LCOV_EXCL_STOP |
| 1490 |
T log_w = log(z); |
| 1491 |
return log_w + Y + boost::math::tools::evaluate_rational(P, Q, log_w); |
| 1492 |
} |
| 1493 |
} |
| 1494 |
|
| 1495 |
template <typename T, typename Policy> |
| 1496 |
T lambert_w_negative_rational_double(T z, const Policy& pol) |
| 1497 |
{ |
| 1498 |
BOOST_MATH_STD_USING |
| 1499 |
if (z > -0.1) |
| 1500 |
{ |
| 1501 |
if (z < -0.051) |
| 1502 |
{ |
| 1503 |
// -0.1 < z < -0.051 |
| 1504 |
// Maximum Deviation Found: 4.402e-22 |
| 1505 |
// Expected Error Term : 4.240e-22 |
| 1506 |
// Maximum Relative Change in Control Points : 4.115e-03 |
| 1507 |
// LCOV_EXCL_START |
| 1508 |
static const T Y = 1.08633995056152344e+00; |
| 1509 |
static const T P[] = { |
| 1510 |
-8.63399505615014331e-02, |
| 1511 |
-1.64303871814816464e+00, |
| 1512 |
-7.71247913918273738e+00, |
| 1513 |
-1.41014495545382454e+01, |
| 1514 |
-1.02269079949257616e+01, |
| 1515 |
-2.17236002836306691e+00, |
| 1516 |
}; |
| 1517 |
static const T Q[] = { |
| 1518 |
1.00000000000000000e+00, |
| 1519 |
7.44775406945739243e+00, |
| 1520 |
2.04392643087266541e+01, |
| 1521 |
2.51001961077774193e+01, |
| 1522 |
1.31256080849023319e+01, |
| 1523 |
2.11640324843601588e+00, |
| 1524 |
}; |
| 1525 |
// LCOV_EXCL_STOP |
| 1526 |
return z * (Y + boost::math::tools::evaluate_rational(P, Q, z)); |
| 1527 |
} |
| 1528 |
else |
| 1529 |
{ |
| 1530 |
// Very small z > 0.051: |
| 1531 |
return lambert_w0_small_z(z, pol); |
| 1532 |
} |
| 1533 |
} |
| 1534 |
else if (z > -0.2) |
| 1535 |
{ |
| 1536 |
// -0.2 < z < -0.1 |
| 1537 |
// Maximum Deviation Found: 2.898e-20 |
| 1538 |
// Expected Error Term : 2.873e-20 |
| 1539 |
// Maximum Relative Change in Control Points : 3.779e-04 |
| 1540 |
// LCOV_EXCL_START |
| 1541 |
static const T Y = 1.20359611511230469e+00; |
| 1542 |
static const T P[] = { |
| 1543 |
-2.03596115108465635e-01, |
| 1544 |
-2.95029082937201859e+00, |
| 1545 |
-1.54287922188671648e+01, |
| 1546 |
-3.81185809571116965e+01, |
| 1547 |
-4.66384358235575985e+01, |
| 1548 |
-2.59282069989642468e+01, |
| 1549 |
-4.70140451266553279e+00, |
| 1550 |
}; |
| 1551 |
static const T Q[] = { |
| 1552 |
1.00000000000000000e+00, |
| 1553 |
9.57921436074599929e+00, |
| 1554 |
3.60988119290234377e+01, |
| 1555 |
6.73977699505546007e+01, |
| 1556 |
6.41104992068148823e+01, |
| 1557 |
2.82060127225153607e+01, |
| 1558 |
4.10677610657724330e+00, |
| 1559 |
}; |
| 1560 |
// LCOV_EXCL_STOP |
| 1561 |
return z * (Y + boost::math::tools::evaluate_rational(P, Q, z)); |
| 1562 |
} |
| 1563 |
else if (z > -0.3178794411714423215955237) |
| 1564 |
{ |
| 1565 |
// Max error in interpolated form: 6.996e-18 |
| 1566 |
// LCOV_EXCL_START |
| 1567 |
static const T Y = 3.49680423736572266e-01; |
| 1568 |
static const T P[] = { |
| 1569 |
-3.49729841718749014e-01, |
| 1570 |
-6.28207407760709028e+01, |
| 1571 |
-2.57226178029669171e+03, |
| 1572 |
-2.50271008623093747e+04, |
| 1573 |
1.11949239154711388e+05, |
| 1574 |
1.85684566607844318e+06, |
| 1575 |
4.80802490427638643e+06, |
| 1576 |
2.76624752134636406e+06, |
| 1577 |
}; |
| 1578 |
static const T Q[] = { |
| 1579 |
1.00000000000000000e+00, |
| 1580 |
1.82717661215113000e+02, |
| 1581 |
8.00121119810280100e+03, |
| 1582 |
1.06073266717010129e+05, |
| 1583 |
3.22848993926057721e+05, |
| 1584 |
-8.05684814514171256e+05, |
| 1585 |
-2.59223192927265737e+06, |
| 1586 |
-5.61719645211570871e+05, |
| 1587 |
6.27765369292636844e+04, |
| 1588 |
}; |
| 1589 |
// LCOV_EXCL_STOP |
| 1590 |
T d = z + 0.367879441171442321595523770161460867445811; |
| 1591 |
return -d / (Y + boost::math::tools::evaluate_polynomial(P, d) / boost::math::tools::evaluate_polynomial(Q, d)); |
| 1592 |
} |
| 1593 |
else if (z > -0.3578794411714423215955237701) |
| 1594 |
{ |
| 1595 |
// Max error in interpolated form: 1.404e-17 |
| 1596 |
// LCOV_EXCL_START |
| 1597 |
static const T Y = 5.00126481056213379e-02; |
| 1598 |
static const T P[] = { |
| 1599 |
-5.00173570682372162e-02, |
| 1600 |
-4.44242461870072044e+01, |
| 1601 |
-9.51185533619946042e+03, |
| 1602 |
-5.88605699015429386e+05, |
| 1603 |
-1.90760843597427751e+06, |
| 1604 |
5.79797663818311404e+08, |
| 1605 |
1.11383352508459134e+10, |
| 1606 |
5.67791253678716467e+10, |
| 1607 |
6.32694500716584572e+10, |
| 1608 |
}; |
| 1609 |
static const T Q[] = { |
| 1610 |
1.00000000000000000e+00, |
| 1611 |
9.08910517489981551e+02, |
| 1612 |
2.10170163753340133e+05, |
| 1613 |
1.67858612416470327e+07, |
| 1614 |
4.90435561733227953e+08, |
| 1615 |
4.54978142622939917e+09, |
| 1616 |
2.87716585708739168e+09, |
| 1617 |
-4.59414247951143131e+10, |
| 1618 |
-1.72845216404874299e+10, |
| 1619 |
}; |
| 1620 |
// LCOV_EXCL_STOP |
| 1621 |
T d = z + 0.36787944117144232159552377016146086744581113103176804; |
| 1622 |
return -d / (Y + boost::math::tools::evaluate_polynomial(P, d) / boost::math::tools::evaluate_polynomial(Q, d)); |
| 1623 |
} |
| 1624 |
else |
| 1625 |
{ // z is very close (within 0.01) of the singularity at -e^-1, |
| 1626 |
// so use a series expansion from R. M. Corless et al. |
| 1627 |
const T p2 = 2 * (boost::math::constants::e<T>() * z + 1); |
| 1628 |
const T p = sqrt(p2); |
| 1629 |
return lambert_w_detail::lambert_w_singularity_series(p); |
| 1630 |
} |
| 1631 |
} |
| 1632 |
|
| 1633 |
//! Lambert_w0 @b 'double' implementation, selected when T is 64-bit precision. |
| 1634 |
template <typename T, typename Policy> |
| 1635 |
inline T lambert_w0_imp(T z, const Policy& pol, const std::integral_constant<int, 2>&) |
| 1636 |
{ |
| 1637 |
static const char* function = "boost::math::lambert_w0<%1%>"; |
| 1638 |
BOOST_MATH_STD_USING // Aid ADL of std functions. |
| 1639 |
|
| 1640 |
// Detect unusual case of 32-bit double with a wider/64-bit long double |
| 1641 |
static_assert(std::numeric_limits<double>::digits >= 53, |
| 1642 |
"Our double precision coefficients will be truncated, " |
| 1643 |
"please file a bug report with details of your platform's floating point types " |
| 1644 |
"- or possibly edit the coefficients to have " |
| 1645 |
"an appropriate size-suffix for 64-bit floats on your platform - L?"); |
| 1646 |
|
| 1647 |
if ((boost::math::isnan)(z)) |
| 1648 |
{ |
| 1649 |
return boost::math::policies::raise_domain_error<T>(function, "Expected a value > -e^-1 (-0.367879...) but got %1%.", z, pol); |
| 1650 |
} |
| 1651 |
if ((boost::math::isinf)(z)) |
| 1652 |
{ |
| 1653 |
return boost::math::policies::raise_overflow_error<T>(function, "Expected a finite value but got %1%.", z, pol); |
| 1654 |
} |
| 1655 |
|
| 1656 |
if (z >= 0.05) |
| 1657 |
{ |
| 1658 |
return lambert_w_positive_rational_double(z); |
| 1659 |
} |
| 1660 |
else if (z <= -0.36787944117144232159552377016146086744581113103176804) // Precision is max_digits10(cpp_bin_float_50). |
| 1661 |
{ |
| 1662 |
if (z < -0.36787944117144232159552377016146086744581113103176804) |
| 1663 |
{ |
| 1664 |
return boost::math::policies::raise_domain_error<T>(function, "Expected z >= -e^-1 (-0.367879...) but got %1%.", z, pol); |
| 1665 |
} |
| 1666 |
return -1; |
| 1667 |
} |
| 1668 |
else |
| 1669 |
{ |
| 1670 |
return lambert_w_negative_rational_double(z, pol); |
| 1671 |
} |
| 1672 |
} // T lambert_w0_imp(T z, const Policy& pol, const std::integral_constant<int, 2>&) 64-bit precision, usually double. |
| 1673 |
|
| 1674 |
//! lambert_W0 implementation for extended precision types including |
| 1675 |
//! long double (80-bit and 128-bit), ??? |
| 1676 |
//! quad float128, Boost.Multiprecision types like cpp_bin_float_quad, cpp_bin_float_50... |
| 1677 |
|
| 1678 |
template <typename T, typename Policy> |
| 1679 |
inline T lambert_w0_imp(T z, const Policy& pol, const std::integral_constant<int, 0>&) |
| 1680 |
{ |
| 1681 |
static const char* function = "boost::math::lambert_w0<%1%>"; |
| 1682 |
BOOST_MATH_STD_USING // Aid ADL of std functions. |
| 1683 |
|
| 1684 |
// Filter out special cases first: |
| 1685 |
if ((boost::math::isnan)(z)) |
| 1686 |
{ |
| 1687 |
return boost::math::policies::raise_domain_error<T>(function, "Expected z >= -e^-1 (-0.367879...) but got %1%.", z, pol); |
| 1688 |
} |
| 1689 |
if (fabs(z) <= 0.05f) |
| 1690 |
{ |
| 1691 |
// Very small z: |
| 1692 |
return lambert_w0_small_z(z, pol); |
| 1693 |
} |
| 1694 |
if (z > (std::numeric_limits<double>::max)()) |
| 1695 |
{ |
| 1696 |
if ((boost::math::isinf)(z)) |
| 1697 |
{ |
| 1698 |
return policies::raise_overflow_error<T>(function, nullptr, pol); |
| 1699 |
// Or might return infinity if available else max_value, |
| 1700 |
// but other Boost.Math special functions raise overflow. |
| 1701 |
} |
| 1702 |
// z is larger than the largest double, so cannot use the polynomial to get an approximation, |
| 1703 |
// so use the asymptotic approximation and Halley iterate: |
| 1704 |
|
| 1705 |
T w = lambert_w0_approx(z); // Make an inline function as also used elsewhere. |
| 1706 |
//T lz = log(z); |
| 1707 |
//T llz = log(lz); |
| 1708 |
//T w = lz - llz + (llz / lz); // Corless equation 4.19, page 349, and Chapeau-Blondeau equation 20, page 2162. |
| 1709 |
return lambert_w_halley_iterate(w, z); |
| 1710 |
} |
| 1711 |
if (z < -0.3578794411714423215955237701) |
| 1712 |
{ // Very close to branch point so rational polynomials are not usable. |
| 1713 |
if (z <= -boost::math::constants::exp_minus_one<T>()) |
| 1714 |
{ |
| 1715 |
if (z == -boost::math::constants::exp_minus_one<T>()) |
| 1716 |
{ // Exactly at the branch point singularity. |
| 1717 |
return -1; |
| 1718 |
} |
| 1719 |
return boost::math::policies::raise_domain_error<T>(function, "Expected z >= -e^-1 (-0.367879...) but got %1%.", z, pol); |
| 1720 |
} |
| 1721 |
// z is very close (within 0.01) of the branch singularity at -e^-1 |
| 1722 |
// so use a series approximation proposed by Corless et al. |
| 1723 |
const T p2 = 2 * (boost::math::constants::e<T>() * z + 1); |
| 1724 |
const T p = sqrt(p2); |
| 1725 |
T w = lambert_w_detail::lambert_w_singularity_series(p); |
| 1726 |
return lambert_w_halley_iterate(w, z); |
| 1727 |
} |
| 1728 |
|
| 1729 |
// Phew! If we get here we are in the normal range of the function, |
| 1730 |
// so get a double precision approximation first, then iterate to full precision of T. |
| 1731 |
// We define a tag_type that is: |
| 1732 |
// true_type if there are so many digits precision wanted that iteration is necessary. |
| 1733 |
// false_type if a single Halley step is sufficient. |
| 1734 |
|
| 1735 |
using precision_type = typename policies::precision<T, Policy>::type; |
| 1736 |
using tag_type = std::integral_constant<bool, |
| 1737 |
(precision_type::value == 0) || (precision_type::value > 113) ? |
| 1738 |
true // Unknown at compile-time, variable/arbitrary, or more than float128 or cpp_bin_quad 128-bit precision. |
| 1739 |
: false // float, double, float128, cpp_bin_quad 128-bit, so single Halley step. |
| 1740 |
>; |
| 1741 |
|
| 1742 |
// For speed, we also cast z to type double when that is possible |
| 1743 |
// if (std::is_constructible<double, T>() == true). |
| 1744 |
T w = lambert_w0_imp(maybe_reduce_to_double(z, std::is_constructible<double, T>()), pol, std::integral_constant<int, 2>()); |
| 1745 |
|
| 1746 |
return lambert_w_maybe_halley_iterate(w, z, tag_type()); |
| 1747 |
|
| 1748 |
} // T lambert_w0_imp(T z, const Policy& pol, const std::integral_constant<int, 0>&) all extended precision types. |
| 1749 |
|
| 1750 |
// Lambert w-1 implementation |
| 1751 |
// ============================================================================================== |
| 1752 |
|
| 1753 |
//! Lambert W for W-1 branch, -max(z) < z <= -1/e. |
| 1754 |
// TODO is -max(z) allowed? |
| 1755 |
template<typename T, typename Policy> |
| 1756 |
T lambert_wm1_imp(const T z, const Policy& pol) |
| 1757 |
{ |
| 1758 |
// Catch providing an integer value as parameter x to lambert_w, for example, lambert_w(1). |
| 1759 |
// Need to ensure it is a floating-point type (of the desired type, float 1.F, double 1., or long double 1.L), |
| 1760 |
// or static_casted integer, for example: static_cast<float>(1) or static_cast<cpp_dec_float_50>(1). |
| 1761 |
// Want to allow fixed_point types too, so do not just test for floating-point. |
| 1762 |
// Integral types should be promoted to double by user Lambert w functions. |
| 1763 |
// If integral type provided to user function lambert_w0 or lambert_wm1, |
| 1764 |
// then should already have been promoted to double. |
| 1765 |
static_assert(!std::is_integral<T>::value, |
| 1766 |
"Must be floating-point or fixed type (not integer type), for example: lambert_wm1(1.), not lambert_wm1(1)!"); |
| 1767 |
|
| 1768 |
BOOST_MATH_STD_USING // Aid argument dependent lookup (ADL) of abs. |
| 1769 |
|
| 1770 |
const char* function = "boost::math::lambert_wm1<RealType>(<RealType>)"; // Used for error messages. |
| 1771 |
|
| 1772 |
// Check for edge and corner cases first: |
| 1773 |
if ((boost::math::isnan)(z)) |
| 1774 |
{ |
| 1775 |
return policies::raise_domain_error(function, "Argument z is NaN!", z, pol); |
| 1776 |
} // isnan |
| 1777 |
|
| 1778 |
if ((boost::math::isinf)(z)) |
| 1779 |
{ |
| 1780 |
return policies::raise_domain_error(function, "Argument z is infinite!", z, pol); |
| 1781 |
} // isinf |
| 1782 |
|
| 1783 |
if (z == static_cast<T>(0)) |
| 1784 |
{ // z is exactly zero so return -std::numeric_limits<T>::infinity(); |
| 1785 |
return -policies::raise_overflow_error(function, nullptr, z, pol); |
| 1786 |
} |
| 1787 |
if (boost::math::detail::has_denorm_now<T>()) |
| 1788 |
{ // All real types except arbitrary precision. |
| 1789 |
if (!(boost::math::isnormal)(z)) |
| 1790 |
{ // Almost zero - might also just return infinity like z == 0 or max_value? |
| 1791 |
return -policies::raise_overflow_error(function, "Argument z = %1% is denormalized! (must be z > (std::numeric_limits<RealType>::min)() or z == 0)", z, pol); |
| 1792 |
} |
| 1793 |
} |
| 1794 |
|
| 1795 |
if (z > static_cast<T>(0)) |
| 1796 |
{ // |
| 1797 |
return policies::raise_domain_error(function, "Argument z = %1% is out of range (z <= 0) for Lambert W-1 branch! (Try Lambert W0 branch?)", z, pol); |
| 1798 |
} |
| 1799 |
if (z == -boost::math::constants::exp_minus_one<T>()) // == singularity/branch point z = -exp(-1) = -0.36787944. |
| 1800 |
{ // At singularity, so return exactly -1. |
| 1801 |
return -static_cast<T>(1); |
| 1802 |
} |
| 1803 |
// z is too negative for the W-1 (or W0) branch. |
| 1804 |
if (z < -boost::math::constants::exp_minus_one<T>()) // > singularity/branch point z = -exp(-1) = -0.36787944. |
| 1805 |
{ |
| 1806 |
return policies::raise_domain_error(function, "Argument z = %1% is out of range (require -exp(-1) = -0.36787944... < z <= 0) for Lambert W-1 (or W0) branch!", z, pol); |
| 1807 |
} |
| 1808 |
if (z < static_cast<T>(-0.35)) |
| 1809 |
{ // Close to singularity/branch point z = -0.3678794411714423215955237701614608727 but on W-1 branch. |
| 1810 |
const T p2 = 2 * (boost::math::constants::e<T>() * z + 1); |
| 1811 |
// Commented out, requires z = -1 / 2e which is greater than -0.35 so this whole branch is not taken. |
| 1812 |
//if (p2 == 0) |
| 1813 |
//{ // At the singularity at branch point. |
| 1814 |
// return -1; |
| 1815 |
// } |
| 1816 |
BOOST_MATH_ASSERT(p2 > 0); |
| 1817 |
T w_series = lambert_w_singularity_series(T(-sqrt(p2))); |
| 1818 |
if (boost::math::tools::digits<T>() > 53) |
| 1819 |
{ // Multiprecision, so try a Halley refinement. |
| 1820 |
w_series = lambert_w_detail::lambert_w_halley_iterate(w_series, z); |
| 1821 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_WM1_NOT_BUILTIN |
| 1822 |
std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 1823 |
std::cout << "Lambert W-1 Halley updated to " << w_series << std::endl; |
| 1824 |
std::cout.precision(saved_precision); |
| 1825 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_WM1_NOT_BUILTIN |
| 1826 |
} |
| 1827 |
return w_series; |
| 1828 |
} // if (z < -0.35) |
| 1829 |
|
| 1830 |
using lambert_w_lookup::wm1es; |
| 1831 |
using lambert_w_lookup::wm1zs; |
| 1832 |
using lambert_w_lookup::noof_wm1zs; // size == 64 |
| 1833 |
|
| 1834 |
// std::cout <<" Wm1zs[63] (== G[64]) = " << " " << wm1zs[63] << std::endl; // Wm1zs[63] (== G[64]) = -1.0264389699511283e-26 |
| 1835 |
// Check that z argument value is not smaller than lookup_table G[64] |
| 1836 |
// std::cout << "(z > wm1zs[63]) = " << std::boolalpha << (z > wm1zs[63]) << std::endl; |
| 1837 |
|
| 1838 |
if (z >= T(wm1zs[63])) // wm1zs[63] = -1.0264389699511282259046957018510946438e-26L W = 64.00000000000000000 |
| 1839 |
{ // z >= -1.0264389699511303e-26 (but z != 0 and z >= std::numeric_limits<T>::min() and so NOT denormalized). |
| 1840 |
|
| 1841 |
// Some info on Lambert W-1 values for extreme values of z. |
| 1842 |
// std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 1843 |
// std::cout << "-std::numeric_limits<float>::min() = " << -(std::numeric_limits<float>::min)() << std::endl; |
| 1844 |
// std::cout << "-std::numeric_limits<double>::min() = " << -(std::numeric_limits<double>::min)() << std::endl; |
| 1845 |
// -std::numeric_limits<float>::min() = -1.1754943508222875e-38 |
| 1846 |
// -std::numeric_limits<double>::min() = -2.2250738585072014e-308 |
| 1847 |
// N[productlog(-1, -1.1754943508222875 * 10^-38 ), 50] = -91.856775324595479509567756730093823993834155027858 |
| 1848 |
// N[productlog(-1, -2.2250738585072014e-308 * 10^-308 ), 50] = -1424.8544521230553853558132180518404363617968042942 |
| 1849 |
// N[productlog(-1, -1.4325445274604020119111357113179868158* 10^-27), 37] = -65.99999999999999999999999999999999955 |
| 1850 |
|
| 1851 |
// R.M.Corless, G.H.Gonnet, D.E.G.Hare, D.J.Jeffrey, and D.E.Knuth, |
| 1852 |
// On the Lambert W function, Adv.Comput.Math., vol. 5, pp. 329, 1996. |
| 1853 |
// Francois Chapeau-Blondeau and Abdelilah Monir |
| 1854 |
// Numerical Evaluation of the Lambert W Function |
| 1855 |
// IEEE Transactions On Signal Processing, VOL. 50, NO. 9, Sep 2002 |
| 1856 |
// https://pdfs.semanticscholar.org/7a5a/76a9369586dd0dd34dda156d8f2779d1fd59.pdf |
| 1857 |
// Estimate Lambert W using ln(-z) ... |
| 1858 |
// This is roughly the power of ten * ln(10) ~= 2.3. n ~= 10^n |
| 1859 |
// and improve by adding a second term -ln(ln(-z)) |
| 1860 |
T guess; // bisect lowest possible Gk[=64] (for lookup_t type) |
| 1861 |
T lz = log(-z); |
| 1862 |
T llz = log(-lz); |
| 1863 |
guess = lz - llz + (llz / lz); // Chapeau-Blondeau equation 20, page 2162. |
| 1864 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_WM1_TINY |
| 1865 |
std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 1866 |
std::cout << "z = " << z << ", guess = " << guess << ", ln(-z) = " << lz << ", ln(-ln(-z) = " << llz << ", llz/lz = " << (llz / lz) << std::endl; |
| 1867 |
// z = -1.0000000000000001e-30, guess = -73.312782616731482, ln(-z) = -69.077552789821368, ln(-ln(-z) = 4.2352298269101114, llz/lz = -0.061311231447304194 |
| 1868 |
// z = -9.9999999999999999e-91, guess = -212.56650048504233, ln(-z) = -207.23265836946410, ln(-ln(-z) = 5.3338421155782205, llz/lz = -0.025738424423764311 |
| 1869 |
// >z = -2.2250738585072014e-308, guess = -714.95942238244606, ln(-z) = -708.39641853226408, ln(-ln(-z) = 6.5630038501819854, llz/lz = -0.0092645920821846622 |
| 1870 |
int d10 = policies::digits_base10<T, Policy>(); // policy template parameter digits10 |
| 1871 |
int d2 = policies::digits<T, Policy>(); // digits base 2 from policy. |
| 1872 |
std::cout << "digits10 = " << d10 << ", digits2 = " << d2 // For example: digits10 = 1, digits2 = 5 |
| 1873 |
<< std::endl; |
| 1874 |
std::cout.precision(saved_precision); |
| 1875 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_WM1_TINY |
| 1876 |
if (policies::digits<T, Policy>() < 12) |
| 1877 |
{ // For the worst case near w = 64, the error in the 'guess' is ~0.008, ratio ~ 0.0001 or 1 in 10,000 digits 10 ~= 4, or digits2 ~= 12. |
| 1878 |
return guess; // LCOV_EXCL_LINE We don't have a test type with few enough digits to trigger this. |
| 1879 |
} |
| 1880 |
T result = lambert_w_detail::lambert_w_halley_iterate(guess, z); |
| 1881 |
return result; |
| 1882 |
|
| 1883 |
// Was Fukushima |
| 1884 |
// G[k=64] == g[63] == -1.02643897e-26 |
| 1885 |
//return policies::raise_domain_error(function, |
| 1886 |
// "Argument z = %1% is too small (< -1.02643897e-26) ! (Should not occur, please report.", |
| 1887 |
// z, pol); |
| 1888 |
} // Z too small so use approximation and Halley. |
| 1889 |
// Else Use a lookup table to find the nearest integer part of Lambert W-1 as starting point for Bisection. |
| 1890 |
|
| 1891 |
if (boost::math::tools::digits<T>() > 53) |
| 1892 |
{ // T is more precise than 64-bit double (or long double, or ?), |
| 1893 |
// so compute an approximate value using only one Schroeder refinement, |
| 1894 |
// (avoiding any double-precision Halley refinement from policy double_digits2<50> 53 - 3 = 50 |
| 1895 |
// because are next going to use Halley refinement at full/high precision using this as an approximation). |
| 1896 |
using boost::math::policies::precision; |
| 1897 |
using boost::math::policies::digits10; |
| 1898 |
using boost::math::policies::digits2; |
| 1899 |
using boost::math::policies::policy; |
| 1900 |
// Compute a 50-bit precision approximate W0 in a double (no Halley refinement). |
| 1901 |
T double_approx(static_cast<T>(lambert_wm1_imp(must_reduce_to_double(z, std::is_constructible<double, T>()), policy<digits2<50>>()))); |
| 1902 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_WM1_NOT_BUILTIN |
| 1903 |
std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 1904 |
std::cout << "Lambert_wm1 Argument Type " << typeid(T).name() << " approximation double = " << double_approx << std::endl; |
| 1905 |
std::cout.precision(saved_precision); |
| 1906 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_WM1 |
| 1907 |
// Perform additional Halley refinement(s) to ensure that |
| 1908 |
// get a near as possible to correct result (usually +/- one epsilon). |
| 1909 |
T result = lambert_w_halley_iterate(double_approx, z); |
| 1910 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_WM1 |
| 1911 |
std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 1912 |
std::cout << "Result " << typeid(T).name() << " precision Halley refinement = " << result << std::endl; |
| 1913 |
std::cout.precision(saved_precision); |
| 1914 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_WM1 |
| 1915 |
return result; |
| 1916 |
} // digits > 53 - higher precision than double. |
| 1917 |
else // T is double or less precision. |
| 1918 |
{ // Use a lookup table to find the nearest integer part of Lambert W as starting point for Bisection. |
| 1919 |
using namespace boost::math::lambert_w_detail::lambert_w_lookup; |
| 1920 |
// Bracketing sequence n = (2, 4, 8, 16, 32, 64) for W-1 branch. (0 is -infinity) |
| 1921 |
// Since z is probably quite small, start with lowest n (=2). |
| 1922 |
int n = 2; |
| 1923 |
if (T(wm1zs[n - 1]) > z) |
| 1924 |
{ |
| 1925 |
goto bisect; |
| 1926 |
} |
| 1927 |
for (int j = 1; j <= 5; ++j) |
| 1928 |
{ |
| 1929 |
n *= 2; |
| 1930 |
if (T(wm1zs[n - 1]) > z) |
| 1931 |
{ |
| 1932 |
goto overshot; |
| 1933 |
} |
| 1934 |
} |
| 1935 |
// else z < g[63] == -1.0264389699511303e-26, so Lambert W-1 integer part > 64. |
| 1936 |
// This should not now occur (should be caught by test and code above) so should be a logic_error? |
| 1937 |
return policies::raise_evaluation_error(function, "Argument z = %1% is too small (< -1.026439e-26) (logic error - please report!)", z, pol); // LCOV_EXCL_LINE |
| 1938 |
overshot: |
| 1939 |
{ |
| 1940 |
int nh = n / 2; |
| 1941 |
for (int j = 1; j <= 5; ++j) |
| 1942 |
{ |
| 1943 |
nh /= 2; // halve step size. |
| 1944 |
if (nh <= 0) |
| 1945 |
{ |
| 1946 |
break; // goto bisect; |
| 1947 |
} |
| 1948 |
if (T(wm1zs[n - nh - 1]) > z) |
| 1949 |
{ |
| 1950 |
n -= nh; |
| 1951 |
} |
| 1952 |
} |
| 1953 |
} |
| 1954 |
bisect: |
| 1955 |
--n; |
| 1956 |
// g[n] now holds lambert W of floor integer n and g[n+1] the ceil part; |
| 1957 |
// these are used as initial values for bisection. |
| 1958 |
#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_WM1_LOOKUP |
| 1959 |
std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 1960 |
std::cout << "Result lookup W-1(" << z << ") bisection between wm1zs[" << n - 1 << "] = " << wm1zs[n - 1] << " and wm1zs[" << n << "] = " << wm1zs[n] |
| 1961 |
<< ", bisect mean = " << (wm1zs[n - 1] + wm1zs[n]) / 2 << std::endl; |
| 1962 |
std::cout.precision(saved_precision); |
| 1963 |
#endif // BOOST_MATH_INSTRUMENT_LAMBERT_WM1_LOOKUP |
| 1964 |
|
| 1965 |
// Compute bisections is the number of bisections computed from n, |
| 1966 |
// such that a single application of the fifth-order Schroeder update formula |
| 1967 |
// after the bisections is enough to evaluate Lambert W-1 with (near?) 53-bit accuracy. |
| 1968 |
// Fukushima established these by trial and error? |
| 1969 |
int bisections = 11; // Assume maximum number of bisections will be needed (most common case). |
| 1970 |
if (n >= 8) |
| 1971 |
{ |
| 1972 |
bisections = 8; |
| 1973 |
} |
| 1974 |
else if (n >= 3) |
| 1975 |
{ |
| 1976 |
bisections = 9; |
| 1977 |
} |
| 1978 |
else if (n >= 2) |
| 1979 |
{ |
| 1980 |
bisections = 10; |
| 1981 |
} |
| 1982 |
// Bracketing, Fukushima section 2.3, page 82: |
| 1983 |
// (Avoiding using exponential function for speed). |
| 1984 |
// Only use @c lookup_t precision, default double, for bisection (again for speed), |
| 1985 |
// and use later Halley refinement for higher precisions. |
| 1986 |
using lambert_w_lookup::halves; |
| 1987 |
using lambert_w_lookup::sqrtwm1s; |
| 1988 |
|
| 1989 |
using calc_type = typename std::conditional<std::is_constructible<lookup_t, T>::value, lookup_t, T>::type; |
| 1990 |
|
| 1991 |
calc_type w = -static_cast<calc_type>(n); // Equation 25, |
| 1992 |
calc_type y = static_cast<calc_type>(z * T(wm1es[n - 1])); // Equation 26, |
| 1993 |
// Perform the bisections fractional bisections for necessary precision. |
| 1994 |
for (int j = 0; j < bisections; ++j) |
| 1995 |
{ // Equation 27. |
| 1996 |
calc_type wj = w - halves[j]; // Subtract 1/2, 1/4, 1/8 ... |
| 1997 |
calc_type yj = y * sqrtwm1s[j]; // Multiply by sqrt(1/e), ... |
| 1998 |
if (wj < yj) |
| 1999 |
{ |
| 2000 |
w = wj; |
| 2001 |
y = yj; |
| 2002 |
} |
| 2003 |
} // for j |
| 2004 |
return static_cast<T>(schroeder_update(w, y)); // Schroeder 5th order method refinement. |
| 2005 |
|
| 2006 |
// else // Perform additional Halley refinement(s) to ensure that |
| 2007 |
// // get a near as possible to correct result (usually +/- epsilon). |
| 2008 |
// { |
| 2009 |
// // result = lambert_w_halley_iterate(result, z); |
| 2010 |
// result = lambert_w_halley_step(result, z); // Just one Halley step should be enough. |
| 2011 |
//#ifdef BOOST_MATH_INSTRUMENT_LAMBERT_WM1_HALLEY |
| 2012 |
// std::streamsize saved_precision = std::cout.precision(std::numeric_limits<T>::max_digits10); |
| 2013 |
// std::cout << "Halley refinement estimate = " << result << std::endl; |
| 2014 |
// std::cout.precision(saved_precision); |
| 2015 |
//#endif // BOOST_MATH_INSTRUMENT_LAMBERT_W1_HALLEY |
| 2016 |
// return result; // Halley |
| 2017 |
// } // Schroeder or Schroeder and Halley. |
| 2018 |
} |
| 2019 |
} // template<typename T = double> T lambert_wm1_imp(const T z) |
| 2020 |
} // namespace lambert_w_detail |
| 2021 |
|
| 2022 |
///////////////////////////// User Lambert w functions. ////////////////////////////// |
| 2023 |
|
| 2024 |
//! Lambert W0 using User-defined policy. |
| 2025 |
template <typename T, typename Policy> |
| 2026 |
inline |
| 2027 |
typename boost::math::tools::promote_args<T>::type |
| 2028 |
lambert_w0(T z, const Policy& pol) |
| 2029 |
{ |
| 2030 |
// Promote integer or expression template arguments to double, |
| 2031 |
// without doing any other internal promotion like float to double. |
| 2032 |
using result_type = typename tools::promote_args<T>::type; |
| 2033 |
|
| 2034 |
// Work out what precision has been selected, |
| 2035 |
// based on the Policy and the number type. |
| 2036 |
using precision_type = typename policies::precision<result_type, Policy>::type; |
| 2037 |
// and then select the correct implementation based on that precision (not the type T): |
| 2038 |
using tag_type = std::integral_constant<int, |
| 2039 |
(precision_type::value == 0) || (precision_type::value > 53) ? |
| 2040 |
0 // either variable precision (0), or greater than 64-bit precision. |
| 2041 |
: (precision_type::value <= 24) ? 1 // 32-bit (probably float) precision. |
| 2042 |
: 2 // 64-bit (probably double) precision. |
| 2043 |
>; |
| 2044 |
|
| 2045 |
return lambert_w_detail::lambert_w0_imp(result_type(z), pol, tag_type()); // |
| 2046 |
} // lambert_w0(T z, const Policy& pol) |
| 2047 |
|
| 2048 |
//! Lambert W0 using default policy. |
| 2049 |
template <typename T> |
| 2050 |
inline |
| 2051 |
typename tools::promote_args<T>::type |
| 2052 |
lambert_w0(T z) |
| 2053 |
{ |
| 2054 |
// Promote integer or expression template arguments to double, |
| 2055 |
// without doing any other internal promotion like float to double. |
| 2056 |
using result_type = typename tools::promote_args<T>::type; |
| 2057 |
|
| 2058 |
// Work out what precision has been selected, based on the Policy and the number type. |
| 2059 |
// For the default policy version, we want the *default policy* precision for T. |
| 2060 |
using precision_type = typename policies::precision<result_type, policies::policy<>>::type; |
| 2061 |
// and then select the correct implementation based on that (not the type T): |
| 2062 |
using tag_type = std::integral_constant<int, |
| 2063 |
(precision_type::value == 0) || (precision_type::value > 53) ? |
| 2064 |
0 // either variable precision (0), or greater than 64-bit precision. |
| 2065 |
: (precision_type::value <= 24) ? 1 // 32-bit (probably float) precision. |
| 2066 |
: 2 // 64-bit (probably double) precision. |
| 2067 |
>; |
| 2068 |
return lambert_w_detail::lambert_w0_imp(result_type(z), policies::policy<>(), tag_type()); |
| 2069 |
} // lambert_w0(T z) using default policy. |
| 2070 |
|
| 2071 |
//! W-1 branch (-max(z) < z <= -1/e). |
| 2072 |
|
| 2073 |
//! Lambert W-1 using User-defined policy. |
| 2074 |
template <typename T, typename Policy> |
| 2075 |
inline |
| 2076 |
typename tools::promote_args<T>::type |
| 2077 |
lambert_wm1(T z, const Policy& pol) |
| 2078 |
{ |
| 2079 |
// Promote integer or expression template arguments to double, |
| 2080 |
// without doing any other internal promotion like float to double. |
| 2081 |
using result_type = typename tools::promote_args<T>::type; |
| 2082 |
return lambert_w_detail::lambert_wm1_imp(result_type(z), pol); // |
| 2083 |
} |
| 2084 |
|
| 2085 |
//! Lambert W-1 using default policy. |
| 2086 |
template <typename T> |
| 2087 |
inline |
| 2088 |
typename tools::promote_args<T>::type |
| 2089 |
lambert_wm1(T z) |
| 2090 |
{ |
| 2091 |
using result_type = typename tools::promote_args<T>::type; |
| 2092 |
return lambert_w_detail::lambert_wm1_imp(result_type(z), policies::policy<>()); |
| 2093 |
} // lambert_wm1(T z) |
| 2094 |
|
| 2095 |
// First derivative of Lambert W0 and W-1. |
| 2096 |
namespace lambert_w_detail { |
| 2097 |
template <typename T, typename Policy> |
| 2098 |
inline typename tools::promote_args<T>::type |
| 2099 |
lambert_w0_prime(T z, const Policy& pol) |
| 2100 |
{ |
| 2101 |
using result_type = typename tools::promote_args<T>::type; |
| 2102 |
using std::numeric_limits; |
| 2103 |
if (z == 0) |
| 2104 |
{ |
| 2105 |
return static_cast<result_type>(1); |
| 2106 |
} |
| 2107 |
// This is the sensible choice if we regard the Lambert-W function as complex analytic. |
| 2108 |
// Of course on the real line, it's just undefined. |
| 2109 |
if (z == -boost::math::constants::exp_minus_one<result_type>()) |
| 2110 |
{ |
| 2111 |
return boost::math::policies::raise_overflow_error("lambert_w0_prime", nullptr, z, pol); |
| 2112 |
} |
| 2113 |
// if z < -1/e, we'll let lambert_w0 do the error handling: |
| 2114 |
result_type w = lambert_w0(result_type(z), pol); |
| 2115 |
// If w ~ -1, then presumably this can get inaccurate. |
| 2116 |
// Is there an accurate way to evaluate 1 + W(-1/e + eps)? |
| 2117 |
// Yes: This is discussed in the Princeton Companion to Applied Mathematics, |
| 2118 |
// 'The Lambert-W function', Section 1.3: Series and Generating Functions. |
| 2119 |
// 1 + W(-1/e + x) ~ sqrt(2ex). |
| 2120 |
// Nick is not convinced this formula is more accurate than the naive one. |
| 2121 |
// However, for z != -1/e, we never get rounded to w = -1 in any precision I've tested (up to cpp_bin_float_100). |
| 2122 |
return w / (z * (1 + w)); |
| 2123 |
} // lambert_w0_prime(T z) |
| 2124 |
} |
| 2125 |
// First derivative of Lambert W0 and W-1. |
| 2126 |
template <typename T, typename Policy> |
| 2127 |
inline typename tools::promote_args<T>::type |
| 2128 |
lambert_w0_prime(T z, const Policy& pol) |
| 2129 |
{ |
| 2130 |
using result_type = typename tools::promote_args<T>::type; |
| 2131 |
return lambert_w_detail::lambert_w0_prime(static_cast<result_type>(z), pol); |
| 2132 |
} |
| 2133 |
|
| 2134 |
template <typename T> |
| 2135 |
inline typename tools::promote_args<T>::type |
| 2136 |
lambert_w0_prime(T z) |
| 2137 |
{ |
| 2138 |
return lambert_w0_prime(z, policies::policy<>()); |
| 2139 |
} |
| 2140 |
|
| 2141 |
template <typename T, typename Policy> |
| 2142 |
inline typename tools::promote_args<T>::type |
| 2143 |
lambert_wm1_prime(T z, const Policy& pol) |
| 2144 |
{ |
| 2145 |
using std::numeric_limits; |
| 2146 |
using result_type = typename tools::promote_args<T>::type; |
| 2147 |
//if (z == 0) |
| 2148 |
//{ |
| 2149 |
// return static_cast<result_type>(1); |
| 2150 |
//} |
| 2151 |
//if (z == - boost::math::constants::exp_minus_one<result_type>()) |
| 2152 |
if (z == 0 || z == - boost::math::constants::exp_minus_one<result_type>()) |
| 2153 |
{ |
| 2154 |
return -boost::math::policies::raise_overflow_error("lambert_wm1_prime", nullptr, z, pol); |
| 2155 |
} |
| 2156 |
|
| 2157 |
result_type w = lambert_wm1(z, pol); |
| 2158 |
return w/(z*(1+w)); |
| 2159 |
} // lambert_wm1_prime(T z) |
| 2160 |
|
| 2161 |
template <typename T> |
| 2162 |
inline typename tools::promote_args<T>::type |
| 2163 |
lambert_wm1_prime(T z) |
| 2164 |
{ |
| 2165 |
return lambert_wm1_prime(z, policies::policy<>()); |
| 2166 |
} |
| 2167 |
|
| 2168 |
}} //boost::math namespaces |
| 2169 |
|
| 2170 |
#endif // #ifdef BOOST_MATH_SF_LAMBERT_W_HPP |
| 2171 |
|