| 1 |
// (C) Copyright John Maddock 2006. |
| 2 |
// (C) Copyright Matt Borland 2024. |
| 3 |
// Use, modification and distribution are subject to the |
| 4 |
// Boost Software License, Version 1.0. (See accompanying file |
| 5 |
// LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) |
| 6 |
|
| 7 |
#ifndef BOOST_MATH_TOOLS_NEWTON_SOLVER_HPP |
| 8 |
#define BOOST_MATH_TOOLS_NEWTON_SOLVER_HPP |
| 9 |
|
| 10 |
#ifdef _MSC_VER |
| 11 |
#pragma once |
| 12 |
#endif |
| 13 |
|
| 14 |
#include <boost/math/tools/config.hpp> |
| 15 |
#include <boost/math/tools/complex.hpp> // test for multiprecision types in complex Newton |
| 16 |
#include <boost/math/tools/type_traits.hpp> |
| 17 |
#include <boost/math/tools/cstdint.hpp> |
| 18 |
#include <boost/math/tools/numeric_limits.hpp> |
| 19 |
#include <boost/math/tools/tuple.hpp> |
| 20 |
#include <boost/math/special_functions/sign.hpp> |
| 21 |
#include <boost/math/policies/policy.hpp> |
| 22 |
#include <boost/math/policies/error_handling.hpp> |
| 23 |
|
| 24 |
#ifndef BOOST_MATH_HAS_GPU_SUPPORT |
| 25 |
#include <boost/math/special_functions/next.hpp> |
| 26 |
#include <boost/math/tools/toms748_solve.hpp> |
| 27 |
#endif |
| 28 |
|
| 29 |
namespace boost { |
| 30 |
namespace math { |
| 31 |
namespace tools { |
| 32 |
|
| 33 |
namespace detail { |
| 34 |
|
| 35 |
namespace dummy { |
| 36 |
|
| 37 |
template<int n, class T> |
| 38 |
BOOST_MATH_GPU_ENABLED typename T::value_type get(const T&) BOOST_MATH_NOEXCEPT(T); |
| 39 |
} |
| 40 |
|
| 41 |
template <class Tuple, class T> |
| 42 |
BOOST_MATH_GPU_ENABLED void unpack_tuple(const Tuple& t, T& a, T& b) BOOST_MATH_NOEXCEPT(T) |
| 43 |
{ |
| 44 |
using dummy::get; |
| 45 |
// Use ADL to find the right overload for get: |
| 46 |
a = get<0>(t); |
| 47 |
b = get<1>(t); |
| 48 |
} |
| 49 |
template <class Tuple, class T> |
| 50 |
BOOST_MATH_GPU_ENABLED void unpack_tuple(const Tuple& t, T& a, T& b, T& c) BOOST_MATH_NOEXCEPT(T) |
| 51 |
{ |
| 52 |
using dummy::get; |
| 53 |
// Use ADL to find the right overload for get: |
| 54 |
a = get<0>(t); |
| 55 |
b = get<1>(t); |
| 56 |
c = get<2>(t); |
| 57 |
} |
| 58 |
|
| 59 |
template <class Tuple, class T> |
| 60 |
BOOST_MATH_GPU_ENABLED inline void unpack_0(const Tuple& t, T& val) BOOST_MATH_NOEXCEPT(T) |
| 61 |
{ |
| 62 |
using dummy::get; |
| 63 |
// Rely on ADL to find the correct overload of get: |
| 64 |
val = get<0>(t); |
| 65 |
} |
| 66 |
|
| 67 |
template <class T, class U, class V> |
| 68 |
BOOST_MATH_GPU_ENABLED inline void unpack_tuple(const boost::math::pair<T, U>& p, V& a, V& b) BOOST_MATH_NOEXCEPT(T) |
| 69 |
{ |
| 70 |
a = p.first; |
| 71 |
b = p.second; |
| 72 |
} |
| 73 |
template <class T, class U, class V> |
| 74 |
BOOST_MATH_GPU_ENABLED inline void unpack_0(const boost::math::pair<T, U>& p, V& a) BOOST_MATH_NOEXCEPT(T) |
| 75 |
{ |
| 76 |
a = p.first; |
| 77 |
} |
| 78 |
|
| 79 |
template <class F, class T> |
| 80 |
BOOST_MATH_GPU_ENABLED void handle_zero_derivative(F f, |
| 81 |
T& last_f0, |
| 82 |
const T& f0, |
| 83 |
T& delta, |
| 84 |
T& result, |
| 85 |
T& guess, |
| 86 |
const T& min, |
| 87 |
const T& max) noexcept(BOOST_MATH_IS_FLOAT(T) |
| 88 |
#ifndef BOOST_MATH_HAS_GPU_SUPPORT |
| 89 |
&& noexcept(std::declval<F>()(std::declval<T>())) |
| 90 |
#endif |
| 91 |
) |
| 92 |
{ |
| 93 |
if (last_f0 == 0) |
| 94 |
{ |
| 95 |
// this must be the first iteration, pretend that we had a |
| 96 |
// previous one at either min or max: |
| 97 |
if (result == min) |
| 98 |
{ |
| 99 |
guess = max; |
| 100 |
} |
| 101 |
else |
| 102 |
{ |
| 103 |
guess = min; |
| 104 |
} |
| 105 |
unpack_0(f(guess), last_f0); |
| 106 |
delta = guess - result; |
| 107 |
} |
| 108 |
if (sign(last_f0) * sign(f0) < 0) |
| 109 |
{ |
| 110 |
// we've crossed over so move in opposite direction to last step: |
| 111 |
if (delta < 0) |
| 112 |
{ |
| 113 |
delta = (result - min) / 2; |
| 114 |
} |
| 115 |
else |
| 116 |
{ |
| 117 |
delta = (result - max) / 2; |
| 118 |
} |
| 119 |
} |
| 120 |
else |
| 121 |
{ |
| 122 |
// move in same direction as last step: |
| 123 |
if (delta < 0) |
| 124 |
{ |
| 125 |
delta = (result - max) / 2; |
| 126 |
} |
| 127 |
else |
| 128 |
{ |
| 129 |
delta = (result - min) / 2; |
| 130 |
} |
| 131 |
} |
| 132 |
} |
| 133 |
|
| 134 |
} // namespace |
| 135 |
|
| 136 |
template <class F, class T, class Tol, class Policy> |
| 137 |
BOOST_MATH_GPU_ENABLED boost::math::pair<T, T> bisect(F f, T min, T max, Tol tol, boost::math::uintmax_t& max_iter, const Policy& pol) noexcept(policies::is_noexcept_error_policy<Policy>::value && BOOST_MATH_IS_FLOAT(T) |
| 138 |
#ifndef BOOST_MATH_HAS_GPU_SUPPORT |
| 139 |
&& noexcept(std::declval<F>()(std::declval<T>())) |
| 140 |
#endif |
| 141 |
) |
| 142 |
{ |
| 143 |
T fmin = f(min); |
| 144 |
T fmax = f(max); |
| 145 |
if (fmin == 0) |
| 146 |
{ |
| 147 |
max_iter = 2; |
| 148 |
return boost::math::make_pair(min, min); |
| 149 |
} |
| 150 |
if (fmax == 0) |
| 151 |
{ |
| 152 |
max_iter = 2; |
| 153 |
return boost::math::make_pair(max, max); |
| 154 |
} |
| 155 |
|
| 156 |
// |
| 157 |
// Error checking: |
| 158 |
// |
| 159 |
constexpr auto function = "boost::math::tools::bisect<%1%>"; |
| 160 |
if (min >= max) |
| 161 |
{ |
| 162 |
return boost::math::detail::pair_from_single(policies::raise_evaluation_error(function, |
| 163 |
"Arguments in wrong order in boost::math::tools::bisect (first arg=%1%)", min, pol)); |
| 164 |
} |
| 165 |
if (fmin * fmax >= 0) |
| 166 |
{ |
| 167 |
return boost::math::detail::pair_from_single(policies::raise_evaluation_error(function, |
| 168 |
"No change of sign in boost::math::tools::bisect, either there is no root to find, or there are multiple roots in the interval (f(min) = %1%).", fmin, pol)); |
| 169 |
} |
| 170 |
|
| 171 |
// |
| 172 |
// Three function invocations so far: |
| 173 |
// |
| 174 |
std::uintmax_t count = max_iter; |
| 175 |
if (count < 3) |
| 176 |
count = 0; |
| 177 |
else |
| 178 |
count -= 3; |
| 179 |
|
| 180 |
while (count && (0 == tol(min, max))) |
| 181 |
{ |
| 182 |
T mid = (min + max) / 2; |
| 183 |
T fmid = f(mid); |
| 184 |
if ((mid == max) || (mid == min)) |
| 185 |
break; |
| 186 |
if (fmid == 0) |
| 187 |
{ |
| 188 |
min = max = mid; |
| 189 |
break; |
| 190 |
} |
| 191 |
else if (sign(fmid) * sign(fmin) < 0) |
| 192 |
{ |
| 193 |
max = mid; |
| 194 |
} |
| 195 |
else |
| 196 |
{ |
| 197 |
min = mid; |
| 198 |
fmin = fmid; |
| 199 |
} |
| 200 |
--count; |
| 201 |
} |
| 202 |
|
| 203 |
max_iter -= count; |
| 204 |
|
| 205 |
#ifdef BOOST_MATH_INSTRUMENT |
| 206 |
std::cout << "Bisection required " << max_iter << " iterations.\n"; |
| 207 |
#endif |
| 208 |
|
| 209 |
return boost::math::make_pair(min, max); |
| 210 |
} |
| 211 |
|
| 212 |
template <class F, class T, class Tol> |
| 213 |
BOOST_MATH_GPU_ENABLED inline boost::math::pair<T, T> bisect(F f, T min, T max, Tol tol, boost::math::uintmax_t& max_iter) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value && BOOST_MATH_IS_FLOAT(T) |
| 214 |
#ifndef BOOST_MATH_HAS_GPU_SUPPORT |
| 215 |
&& noexcept(std::declval<F>()(std::declval<T>())) |
| 216 |
#endif |
| 217 |
) |
| 218 |
{ |
| 219 |
return bisect(f, min, max, tol, max_iter, policies::policy<>()); |
| 220 |
} |
| 221 |
|
| 222 |
template <class F, class T, class Tol> |
| 223 |
BOOST_MATH_GPU_ENABLED inline boost::math::pair<T, T> bisect(F f, T min, T max, Tol tol) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value && BOOST_MATH_IS_FLOAT(T) |
| 224 |
#ifndef BOOST_MATH_HAS_GPU_SUPPORT |
| 225 |
&& noexcept(std::declval<F>()(std::declval<T>())) |
| 226 |
#endif |
| 227 |
) |
| 228 |
{ |
| 229 |
boost::math::uintmax_t m = (boost::math::numeric_limits<boost::math::uintmax_t>::max)(); |
| 230 |
return bisect(f, min, max, tol, m, policies::policy<>()); |
| 231 |
} |
| 232 |
|
| 233 |
|
| 234 |
template <class F, class T> |
| 235 |
BOOST_MATH_GPU_ENABLED T newton_raphson_iterate(F f, T guess, T min, T max, int digits, boost::math::uintmax_t& max_iter) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value && BOOST_MATH_IS_FLOAT(T) |
| 236 |
#ifndef BOOST_MATH_HAS_GPU_SUPPORT |
| 237 |
&& noexcept(std::declval<F>()(std::declval<T>())) |
| 238 |
#endif |
| 239 |
) |
| 240 |
{ |
| 241 |
BOOST_MATH_STD_USING |
| 242 |
|
| 243 |
constexpr auto function = "boost::math::tools::newton_raphson_iterate<%1%>"; |
| 244 |
if (min > max) |
| 245 |
{ |
| 246 |
return policies::raise_evaluation_error(function, "Range arguments in wrong order in boost::math::tools::newton_raphson_iterate(first arg=%1%)", min, boost::math::policies::policy<>()); |
| 247 |
} |
| 248 |
|
| 249 |
T f0(0), f1, last_f0(0); |
| 250 |
T result = guess; |
| 251 |
|
| 252 |
T factor = static_cast<T>(ldexp(1.0, 1 - digits)); |
| 253 |
T delta = tools::max_value<T>(); |
| 254 |
T delta1 = tools::max_value<T>(); |
| 255 |
T delta2 = tools::max_value<T>(); |
| 256 |
|
| 257 |
// |
| 258 |
// We use these to sanity check that we do actually bracket a root, |
| 259 |
// we update these to the function value when we update the endpoints |
| 260 |
// of the range. Then, provided at some point we update both endpoints |
| 261 |
// checking that max_range_f * min_range_f <= 0 verifies there is a root |
| 262 |
// to be found somewhere. Note that if there is no root, and we approach |
| 263 |
// a local minima, then the derivative will go to zero, and hence the next |
| 264 |
// step will jump out of bounds (or at least past the minima), so this |
| 265 |
// check *should* happen in pathological cases. |
| 266 |
// |
| 267 |
T max_range_f = 0; |
| 268 |
T min_range_f = 0; |
| 269 |
|
| 270 |
boost::math::uintmax_t count(max_iter); |
| 271 |
|
| 272 |
#ifdef BOOST_MATH_INSTRUMENT |
| 273 |
std::cout << "Newton_raphson_iterate, guess = " << guess << ", min = " << min << ", max = " << max |
| 274 |
<< ", digits = " << digits << ", max_iter = " << max_iter << "\n"; |
| 275 |
#endif |
| 276 |
|
| 277 |
do { |
| 278 |
last_f0 = f0; |
| 279 |
delta2 = delta1; |
| 280 |
delta1 = delta; |
| 281 |
detail::unpack_tuple(f(result), f0, f1); |
| 282 |
--count; |
| 283 |
if (0 == f0) |
| 284 |
break; |
| 285 |
if (f1 == 0) |
| 286 |
{ |
| 287 |
// Oops zero derivative!!! |
| 288 |
detail::handle_zero_derivative(f, last_f0, f0, delta, result, guess, min, max); |
| 289 |
} |
| 290 |
else |
| 291 |
{ |
| 292 |
delta = f0 / f1; |
| 293 |
} |
| 294 |
#ifdef BOOST_MATH_INSTRUMENT |
| 295 |
std::cout << "Newton iteration " << max_iter - count << ", delta = " << delta << ", residual = " << f0 << "\n"; |
| 296 |
#endif |
| 297 |
if (fabs(delta * 2) > fabs(delta2)) |
| 298 |
{ |
| 299 |
// Last two steps haven't converged. |
| 300 |
T shift = (delta > 0) ? (result - min) / 2 : (result - max) / 2; |
| 301 |
if ((result != 0) && (fabs(shift) > fabs(result))) |
| 302 |
{ |
| 303 |
delta = sign(delta) * fabs(result); // protect against huge jumps! |
| 304 |
} |
| 305 |
else |
| 306 |
delta = shift; |
| 307 |
// reset delta1/2 so we don't take this branch next time round: |
| 308 |
delta1 = 3 * delta; |
| 309 |
delta2 = 3 * delta; |
| 310 |
} |
| 311 |
guess = result; |
| 312 |
result -= delta; |
| 313 |
if (result <= min) |
| 314 |
{ |
| 315 |
delta = 0.5F * (guess - min); |
| 316 |
result = guess - delta; |
| 317 |
if ((result == min) || (result == max)) |
| 318 |
break; |
| 319 |
} |
| 320 |
else if (result >= max) |
| 321 |
{ |
| 322 |
delta = 0.5F * (guess - max); |
| 323 |
result = guess - delta; |
| 324 |
if ((result == min) || (result == max)) |
| 325 |
break; |
| 326 |
} |
| 327 |
// Update brackets: |
| 328 |
if (delta > 0) |
| 329 |
{ |
| 330 |
max = guess; |
| 331 |
max_range_f = f0; |
| 332 |
} |
| 333 |
else |
| 334 |
{ |
| 335 |
min = guess; |
| 336 |
min_range_f = f0; |
| 337 |
} |
| 338 |
// |
| 339 |
// Sanity check that we bracket the root: |
| 340 |
// |
| 341 |
if (max_range_f * min_range_f > 0) |
| 342 |
{ |
| 343 |
return policies::raise_evaluation_error(function, "There appears to be no root to be found in boost::math::tools::newton_raphson_iterate, perhaps we have a local minima near current best guess of %1%", guess, boost::math::policies::policy<>()); |
| 344 |
} |
| 345 |
}while(count && (fabs(result * factor) < fabs(delta))); |
| 346 |
|
| 347 |
max_iter -= count; |
| 348 |
|
| 349 |
#ifdef BOOST_MATH_INSTRUMENT |
| 350 |
std::cout << "Newton Raphson required " << max_iter << " iterations\n"; |
| 351 |
#endif |
| 352 |
|
| 353 |
return result; |
| 354 |
} |
| 355 |
|
| 356 |
template <class F, class T> |
| 357 |
BOOST_MATH_GPU_ENABLED inline T newton_raphson_iterate(F f, T guess, T min, T max, int digits) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value && BOOST_MATH_IS_FLOAT(T) |
| 358 |
#ifndef BOOST_MATH_HAS_GPU_SUPPORT |
| 359 |
&& noexcept(std::declval<F>()(std::declval<T>())) |
| 360 |
#endif |
| 361 |
) |
| 362 |
{ |
| 363 |
boost::math::uintmax_t m = (boost::math::numeric_limits<boost::math::uintmax_t>::max)(); |
| 364 |
return newton_raphson_iterate(f, guess, min, max, digits, m); |
| 365 |
} |
| 366 |
|
| 367 |
// TODO(mborland): Disabled for now |
| 368 |
// Recursion needs to be removed, but there is no demand at this time |
| 369 |
#ifdef BOOST_MATH_HAS_NVRTC |
| 370 |
}}} // Namespaces |
| 371 |
#else |
| 372 |
|
| 373 |
namespace detail { |
| 374 |
|
| 375 |
struct halley_step |
| 376 |
{ |
| 377 |
template <class T> |
| 378 |
static T step(const T& /*x*/, const T& f0, const T& f1, const T& f2) noexcept(BOOST_MATH_IS_FLOAT(T)) |
| 379 |
{ |
| 380 |
using std::fabs; |
| 381 |
T denom = 2 * f0; |
| 382 |
T num = 2 * f1 - f0 * (f2 / f1); |
| 383 |
T delta; |
| 384 |
|
| 385 |
BOOST_MATH_INSTRUMENT_VARIABLE(denom); |
| 386 |
BOOST_MATH_INSTRUMENT_VARIABLE(num); |
| 387 |
|
| 388 |
if ((fabs(num) < 1) && (fabs(denom) >= fabs(num) * tools::max_value<T>())) |
| 389 |
{ |
| 390 |
// possible overflow, use Newton step: |
| 391 |
delta = f0 / f1; |
| 392 |
} |
| 393 |
else |
| 394 |
delta = denom / num; |
| 395 |
return delta; |
| 396 |
} |
| 397 |
}; |
| 398 |
|
| 399 |
template <class F, class T> |
| 400 |
T bracket_root_towards_min(F f, T guess, const T& f0, T& min, T& max, std::uintmax_t& count) noexcept(BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))); |
| 401 |
|
| 402 |
template <class F, class T> |
| 403 |
T bracket_root_towards_max(F f, T guess, const T& f0, T& min, T& max, std::uintmax_t& count) noexcept(BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 404 |
{ |
| 405 |
using std::fabs; |
| 406 |
using std::ldexp; |
| 407 |
using std::abs; |
| 408 |
using std::frexp; |
| 409 |
if(count < 2) |
| 410 |
return guess - (max + min) / 2; // Not enough counts left to do anything!! |
| 411 |
// |
| 412 |
// Move guess towards max until we bracket the root, updating min and max as we go: |
| 413 |
// |
| 414 |
int e; |
| 415 |
frexp(max / guess, &e); |
| 416 |
e = abs(e); |
| 417 |
T guess0 = guess; |
| 418 |
T multiplier = e < 64 ? static_cast<T>(2) : static_cast<T>(ldexp(T(1), e / 32)); |
| 419 |
T f_current = f0; |
| 420 |
if (fabs(min) < fabs(max)) |
| 421 |
{ |
| 422 |
while (--count && ((f_current < 0) == (f0 < 0))) |
| 423 |
{ |
| 424 |
min = guess; |
| 425 |
guess *= multiplier; |
| 426 |
if (guess > max) |
| 427 |
{ |
| 428 |
guess = max; |
| 429 |
f_current = -f_current; // There must be a change of sign! |
| 430 |
break; |
| 431 |
} |
| 432 |
multiplier *= e > 1024 ? 8 : 2; |
| 433 |
unpack_0(f(guess), f_current); |
| 434 |
} |
| 435 |
} |
| 436 |
else |
| 437 |
{ |
| 438 |
// |
| 439 |
// If min and max are negative we have to divide to head towards max: |
| 440 |
// |
| 441 |
while (--count && ((f_current < 0) == (f0 < 0))) |
| 442 |
{ |
| 443 |
min = guess; |
| 444 |
guess /= multiplier; |
| 445 |
if (guess > max) |
| 446 |
{ |
| 447 |
guess = max; |
| 448 |
f_current = -f_current; // There must be a change of sign! |
| 449 |
break; |
| 450 |
} |
| 451 |
multiplier *= e > 1024 ? 8 : 2; |
| 452 |
unpack_0(f(guess), f_current); |
| 453 |
} |
| 454 |
} |
| 455 |
|
| 456 |
if (count) |
| 457 |
{ |
| 458 |
max = guess; |
| 459 |
if (multiplier > 16) |
| 460 |
return (guess0 - guess) + bracket_root_towards_min(f, guess, f_current, min, max, count); |
| 461 |
} |
| 462 |
return guess0 - (max + min) / 2; |
| 463 |
} |
| 464 |
|
| 465 |
template <class F, class T> |
| 466 |
T bracket_root_towards_min(F f, T guess, const T& f0, T& min, T& max, std::uintmax_t& count) noexcept(BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 467 |
{ |
| 468 |
using std::fabs; |
| 469 |
using std::ldexp; |
| 470 |
using std::abs; |
| 471 |
using std::frexp; |
| 472 |
if (count < 2) |
| 473 |
return guess - (max + min) / 2; // Not enough counts left to do anything!! |
| 474 |
// |
| 475 |
// Move guess towards min until we bracket the root, updating min and max as we go: |
| 476 |
// |
| 477 |
int e; |
| 478 |
frexp(guess / min, &e); |
| 479 |
e = abs(e); |
| 480 |
T guess0 = guess; |
| 481 |
T multiplier = e < 64 ? static_cast<T>(2) : static_cast<T>(ldexp(T(1), e / 32)); |
| 482 |
T f_current = f0; |
| 483 |
|
| 484 |
if (fabs(min) < fabs(max)) |
| 485 |
{ |
| 486 |
while (--count && ((f_current < 0) == (f0 < 0))) |
| 487 |
{ |
| 488 |
max = guess; |
| 489 |
guess /= multiplier; |
| 490 |
if (guess < min) |
| 491 |
{ |
| 492 |
guess = min; |
| 493 |
f_current = -f_current; // There must be a change of sign! |
| 494 |
break; |
| 495 |
} |
| 496 |
multiplier *= e > 1024 ? 8 : 2; |
| 497 |
unpack_0(f(guess), f_current); |
| 498 |
} |
| 499 |
} |
| 500 |
else |
| 501 |
{ |
| 502 |
// |
| 503 |
// If min and max are negative we have to multiply to head towards min: |
| 504 |
// |
| 505 |
while (--count && ((f_current < 0) == (f0 < 0))) |
| 506 |
{ |
| 507 |
max = guess; |
| 508 |
guess *= multiplier; |
| 509 |
if (guess < min) |
| 510 |
{ |
| 511 |
guess = min; |
| 512 |
f_current = -f_current; // There must be a change of sign! |
| 513 |
break; |
| 514 |
} |
| 515 |
multiplier *= e > 1024 ? 8 : 2; |
| 516 |
unpack_0(f(guess), f_current); |
| 517 |
} |
| 518 |
} |
| 519 |
|
| 520 |
if (count) |
| 521 |
{ |
| 522 |
min = guess; |
| 523 |
if (multiplier > 16) |
| 524 |
return (guess0 - guess) + bracket_root_towards_max(f, guess, f_current, min, max, count); |
| 525 |
} |
| 526 |
return guess0 - (max + min) / 2; |
| 527 |
} |
| 528 |
|
| 529 |
template <class Stepper, class F, class T> |
| 530 |
T second_order_root_finder(F f, T guess, T min, T max, int digits, std::uintmax_t& max_iter) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value&& BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 531 |
{ |
| 532 |
BOOST_MATH_STD_USING |
| 533 |
|
| 534 |
#ifdef BOOST_MATH_INSTRUMENT |
| 535 |
std::cout << "Second order root iteration, guess = " << guess << ", min = " << min << ", max = " << max |
| 536 |
<< ", digits = " << digits << ", max_iter = " << max_iter << "\n"; |
| 537 |
#endif |
| 538 |
static const char* function = "boost::math::tools::halley_iterate<%1%>"; |
| 539 |
if (min >= max) |
| 540 |
{ |
| 541 |
return policies::raise_evaluation_error(function, "Range arguments in wrong order in boost::math::tools::halley_iterate(first arg=%1%)", min, boost::math::policies::policy<>()); |
| 542 |
} |
| 543 |
|
| 544 |
T f0(0), f1, f2; |
| 545 |
T result = guess; |
| 546 |
|
| 547 |
T factor = ldexp(static_cast<T>(1.0), 1 - digits); |
| 548 |
T delta = (std::max)(T(10000000 * guess), T(10000000)); // arbitrarily large delta |
| 549 |
T last_f0 = 0; |
| 550 |
T delta1 = delta; |
| 551 |
T delta2 = delta; |
| 552 |
bool out_of_bounds_sentry = false; |
| 553 |
|
| 554 |
#ifdef BOOST_MATH_INSTRUMENT |
| 555 |
std::cout << "Second order root iteration, limit = " << factor << "\n"; |
| 556 |
#endif |
| 557 |
|
| 558 |
// |
| 559 |
// We use these to sanity check that we do actually bracket a root, |
| 560 |
// we update these to the function value when we update the endpoints |
| 561 |
// of the range. Then, provided at some point we update both endpoints |
| 562 |
// checking that max_range_f * min_range_f <= 0 verifies there is a root |
| 563 |
// to be found somewhere. Note that if there is no root, and we approach |
| 564 |
// a local minima, then the derivative will go to zero, and hence the next |
| 565 |
// step will jump out of bounds (or at least past the minima), so this |
| 566 |
// check *should* happen in pathological cases. |
| 567 |
// |
| 568 |
T max_range_f = 0; |
| 569 |
T min_range_f = 0; |
| 570 |
|
| 571 |
std::uintmax_t count(max_iter); |
| 572 |
|
| 573 |
do { |
| 574 |
last_f0 = f0; |
| 575 |
delta2 = delta1; |
| 576 |
delta1 = delta; |
| 577 |
#ifndef BOOST_MATH_NO_EXCEPTIONS |
| 578 |
try |
| 579 |
#endif |
| 580 |
{ |
| 581 |
detail::unpack_tuple(f(result), f0, f1, f2); |
| 582 |
} |
| 583 |
#ifndef BOOST_MATH_NO_EXCEPTIONS |
| 584 |
catch (const std::overflow_error&) |
| 585 |
{ |
| 586 |
f0 = max > 0 ? tools::max_value<T>() : -tools::min_value<T>(); |
| 587 |
f1 = f2 = 0; |
| 588 |
} |
| 589 |
#endif |
| 590 |
--count; |
| 591 |
|
| 592 |
BOOST_MATH_INSTRUMENT_VARIABLE(f0); |
| 593 |
BOOST_MATH_INSTRUMENT_VARIABLE(f1); |
| 594 |
BOOST_MATH_INSTRUMENT_VARIABLE(f2); |
| 595 |
|
| 596 |
if (0 == f0) |
| 597 |
break; |
| 598 |
if (f1 == 0) |
| 599 |
{ |
| 600 |
// Oops zero derivative!!! |
| 601 |
detail::handle_zero_derivative(f, last_f0, f0, delta, result, guess, min, max); |
| 602 |
} |
| 603 |
else |
| 604 |
{ |
| 605 |
if (f2 != 0) |
| 606 |
{ |
| 607 |
delta = Stepper::step(result, f0, f1, f2); |
| 608 |
if (delta * f1 / f0 < 0) |
| 609 |
{ |
| 610 |
// Oh dear, we have a problem as Newton and Halley steps |
| 611 |
// disagree about which way we should move. Probably |
| 612 |
// there is cancelation error in the calculation of the |
| 613 |
// Halley step, or else the derivatives are so small |
| 614 |
// that their values are basically trash. We will move |
| 615 |
// in the direction indicated by a Newton step, but |
| 616 |
// by no more than twice the current guess value, otherwise |
| 617 |
// we can jump way out of bounds if we're not careful. |
| 618 |
// See https://svn.boost.org/trac/boost/ticket/8314. |
| 619 |
delta = f0 / f1; |
| 620 |
if (fabs(delta) > 2 * fabs(result)) |
| 621 |
delta = (delta < 0 ? -1 : 1) * 2 * fabs(result); |
| 622 |
} |
| 623 |
} |
| 624 |
else |
| 625 |
delta = f0 / f1; |
| 626 |
} |
| 627 |
#ifdef BOOST_MATH_INSTRUMENT |
| 628 |
std::cout << "Second order root iteration, delta = " << delta << ", residual = " << f0 << "\n"; |
| 629 |
#endif |
| 630 |
// We need to avoid delta/delta2 overflowing here: |
| 631 |
T convergence = (fabs(delta2) > 1) || (fabs(tools::max_value<T>() * delta2) > fabs(delta)) ? fabs(delta / delta2) : tools::max_value<T>(); |
| 632 |
if ((convergence > 0.8) && (convergence < 2)) |
| 633 |
{ |
| 634 |
// last two steps haven't converged. |
| 635 |
if (fabs(min) < 1 ? fabs(1000 * min) < fabs(max) : fabs(max / min) > 1000) |
| 636 |
{ |
| 637 |
if(delta > 0) |
| 638 |
delta = bracket_root_towards_min(f, result, f0, min, max, count); |
| 639 |
else |
| 640 |
delta = bracket_root_towards_max(f, result, f0, min, max, count); |
| 641 |
} |
| 642 |
else |
| 643 |
{ |
| 644 |
delta = (delta > 0) ? (result - min) / 2 : (result - max) / 2; |
| 645 |
if ((result != 0) && (fabs(delta) > result)) |
| 646 |
delta = sign(delta) * fabs(result) * 0.9f; // protect against huge jumps! |
| 647 |
} |
| 648 |
// reset delta2 so that this branch will *not* be taken on the |
| 649 |
// next iteration: |
| 650 |
delta2 = delta * 3; |
| 651 |
delta1 = delta * 3; |
| 652 |
BOOST_MATH_INSTRUMENT_VARIABLE(delta); |
| 653 |
} |
| 654 |
guess = result; |
| 655 |
result -= delta; |
| 656 |
BOOST_MATH_INSTRUMENT_VARIABLE(result); |
| 657 |
|
| 658 |
// check for out of bounds step: |
| 659 |
if (result < min) |
| 660 |
{ |
| 661 |
T diff = ((fabs(min) < 1) && (fabs(result) > 1) && (tools::max_value<T>() / fabs(result) < fabs(min))) |
| 662 |
? T(1000) |
| 663 |
: (fabs(min) < 1) && (fabs(tools::max_value<T>() * min) < fabs(result)) |
| 664 |
? ((min < 0) != (result < 0)) ? -tools::max_value<T>() : tools::max_value<T>() : T(result / min); |
| 665 |
if (fabs(diff) < 1) |
| 666 |
diff = 1 / diff; |
| 667 |
if (!out_of_bounds_sentry && (diff > 0) && (diff < 3)) |
| 668 |
{ |
| 669 |
// Only a small out of bounds step, lets assume that the result |
| 670 |
// is probably approximately at min: |
| 671 |
delta = 0.99f * (guess - min); |
| 672 |
result = guess - delta; |
| 673 |
out_of_bounds_sentry = true; // only take this branch once! |
| 674 |
} |
| 675 |
else |
| 676 |
{ |
| 677 |
if (fabs(float_distance(min, max)) < 2) |
| 678 |
{ |
| 679 |
result = guess = (min + max) / 2; |
| 680 |
break; |
| 681 |
} |
| 682 |
delta = bracket_root_towards_min(f, guess, f0, min, max, count); |
| 683 |
result = guess - delta; |
| 684 |
if (result <= min) |
| 685 |
result = float_next(min); |
| 686 |
if (result >= max) |
| 687 |
result = float_prior(max); |
| 688 |
guess = min; |
| 689 |
continue; |
| 690 |
} |
| 691 |
} |
| 692 |
else if (result > max) |
| 693 |
{ |
| 694 |
T diff = ((fabs(max) < 1) && (fabs(result) > 1) && (tools::max_value<T>() / fabs(result) < fabs(max))) ? T(1000) : T(result / max); |
| 695 |
if (fabs(diff) < 1) |
| 696 |
diff = 1 / diff; |
| 697 |
if (!out_of_bounds_sentry && (diff > 0) && (diff < 3)) |
| 698 |
{ |
| 699 |
// Only a small out of bounds step, lets assume that the result |
| 700 |
// is probably approximately at min: |
| 701 |
delta = 0.99f * (guess - max); |
| 702 |
result = guess - delta; |
| 703 |
out_of_bounds_sentry = true; // only take this branch once! |
| 704 |
} |
| 705 |
else |
| 706 |
{ |
| 707 |
if (fabs(float_distance(min, max)) < 2) |
| 708 |
{ |
| 709 |
result = guess = (min + max) / 2; |
| 710 |
break; |
| 711 |
} |
| 712 |
delta = bracket_root_towards_max(f, guess, f0, min, max, count); |
| 713 |
result = guess - delta; |
| 714 |
if (result >= max) |
| 715 |
result = float_prior(max); |
| 716 |
if (result <= min) |
| 717 |
result = float_next(min); |
| 718 |
guess = min; |
| 719 |
continue; |
| 720 |
} |
| 721 |
} |
| 722 |
// update brackets: |
| 723 |
if (delta > 0) |
| 724 |
{ |
| 725 |
max = guess; |
| 726 |
max_range_f = f0; |
| 727 |
} |
| 728 |
else |
| 729 |
{ |
| 730 |
min = guess; |
| 731 |
min_range_f = f0; |
| 732 |
} |
| 733 |
// |
| 734 |
// Sanity check that we bracket the root: |
| 735 |
// |
| 736 |
if (max_range_f * min_range_f > 0) |
| 737 |
{ |
| 738 |
return policies::raise_evaluation_error(function, "There appears to be no root to be found in boost::math::tools::newton_raphson_iterate, perhaps we have a local minima near current best guess of %1%", guess, boost::math::policies::policy<>()); |
| 739 |
} |
| 740 |
} while(count && (fabs(result * factor) < fabs(delta))); |
| 741 |
|
| 742 |
max_iter -= count; |
| 743 |
|
| 744 |
#ifdef BOOST_MATH_INSTRUMENT |
| 745 |
std::cout << "Second order root finder required " << max_iter << " iterations.\n"; |
| 746 |
#endif |
| 747 |
|
| 748 |
return result; |
| 749 |
} |
| 750 |
} // T second_order_root_finder |
| 751 |
|
| 752 |
template <class F, class T> |
| 753 |
T halley_iterate(F f, T guess, T min, T max, int digits, std::uintmax_t& max_iter) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value&& BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 754 |
{ |
| 755 |
return detail::second_order_root_finder<detail::halley_step>(f, guess, min, max, digits, max_iter); |
| 756 |
} |
| 757 |
|
| 758 |
template <class F, class T> |
| 759 |
inline T halley_iterate(F f, T guess, T min, T max, int digits) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value&& BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 760 |
{ |
| 761 |
std::uintmax_t m = (std::numeric_limits<std::uintmax_t>::max)(); |
| 762 |
return halley_iterate(f, guess, min, max, digits, m); |
| 763 |
} |
| 764 |
|
| 765 |
namespace detail { |
| 766 |
|
| 767 |
struct schroder_stepper |
| 768 |
{ |
| 769 |
template <class T> |
| 770 |
static T step(const T& x, const T& f0, const T& f1, const T& f2) noexcept(BOOST_MATH_IS_FLOAT(T)) |
| 771 |
{ |
| 772 |
using std::fabs; |
| 773 |
T ratio = f0 / f1; |
| 774 |
T delta; |
| 775 |
if ((x != 0) && (fabs(ratio / x) < 0.1)) |
| 776 |
{ |
| 777 |
delta = ratio + (f2 / (2 * f1)) * ratio * ratio; |
| 778 |
// check second derivative doesn't over compensate: |
| 779 |
if (delta * ratio < 0) |
| 780 |
delta = ratio; |
| 781 |
} |
| 782 |
else |
| 783 |
delta = ratio; // fall back to Newton iteration. |
| 784 |
return delta; |
| 785 |
} |
| 786 |
}; |
| 787 |
|
| 788 |
} |
| 789 |
|
| 790 |
template <class F, class T> |
| 791 |
T schroder_iterate(F f, T guess, T min, T max, int digits, std::uintmax_t& max_iter) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value&& BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 792 |
{ |
| 793 |
return detail::second_order_root_finder<detail::schroder_stepper>(f, guess, min, max, digits, max_iter); |
| 794 |
} |
| 795 |
|
| 796 |
template <class F, class T> |
| 797 |
inline T schroder_iterate(F f, T guess, T min, T max, int digits) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value&& BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 798 |
{ |
| 799 |
std::uintmax_t m = (std::numeric_limits<std::uintmax_t>::max)(); |
| 800 |
return schroder_iterate(f, guess, min, max, digits, m); |
| 801 |
} |
| 802 |
// |
| 803 |
// These two are the old spelling of this function, retained for backwards compatibility just in case: |
| 804 |
// |
| 805 |
template <class F, class T> |
| 806 |
T schroeder_iterate(F f, T guess, T min, T max, int digits, std::uintmax_t& max_iter) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value&& BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 807 |
{ |
| 808 |
return detail::second_order_root_finder<detail::schroder_stepper>(f, guess, min, max, digits, max_iter); |
| 809 |
} |
| 810 |
|
| 811 |
template <class F, class T> |
| 812 |
inline T schroeder_iterate(F f, T guess, T min, T max, int digits) noexcept(policies::is_noexcept_error_policy<policies::policy<> >::value&& BOOST_MATH_IS_FLOAT(T) && noexcept(std::declval<F>()(std::declval<T>()))) |
| 813 |
{ |
| 814 |
std::uintmax_t m = (std::numeric_limits<std::uintmax_t>::max)(); |
| 815 |
return schroder_iterate(f, guess, min, max, digits, m); |
| 816 |
} |
| 817 |
|
| 818 |
#ifndef BOOST_NO_CXX11_AUTO_DECLARATIONS |
| 819 |
/* |
| 820 |
* Why do we set the default maximum number of iterations to the number of digits in the type? |
| 821 |
* Because for double roots, the number of digits increases linearly with the number of iterations, |
| 822 |
* so this default should recover full precision even in this somewhat pathological case. |
| 823 |
* For isolated roots, the problem is so rapidly convergent that this doesn't matter at all. |
| 824 |
*/ |
| 825 |
template<class ComplexType, class F> |
| 826 |
ComplexType complex_newton(F g, ComplexType guess, int max_iterations = std::numeric_limits<typename ComplexType::value_type>::digits) |
| 827 |
{ |
| 828 |
typedef typename ComplexType::value_type Real; |
| 829 |
using std::norm; |
| 830 |
using std::abs; |
| 831 |
using std::max; |
| 832 |
// z0, z1, and z2 cannot be the same, in case we immediately need to resort to Muller's Method: |
| 833 |
ComplexType z0 = guess + ComplexType(1, 0); |
| 834 |
ComplexType z1 = guess + ComplexType(0, 1); |
| 835 |
ComplexType z2 = guess; |
| 836 |
|
| 837 |
do { |
| 838 |
auto pair = g(z2); |
| 839 |
if (norm(pair.second) == 0) |
| 840 |
{ |
| 841 |
// Muller's method. Notation follows Numerical Recipes, 9.5.2: |
| 842 |
ComplexType q = (z2 - z1) / (z1 - z0); |
| 843 |
auto P0 = g(z0); |
| 844 |
auto P1 = g(z1); |
| 845 |
ComplexType qp1 = static_cast<ComplexType>(1) + q; |
| 846 |
ComplexType A = q * (pair.first - qp1 * P1.first + q * P0.first); |
| 847 |
|
| 848 |
ComplexType B = (static_cast<ComplexType>(2) * q + static_cast<ComplexType>(1)) * pair.first - qp1 * qp1 * P1.first + q * q * P0.first; |
| 849 |
ComplexType C = qp1 * pair.first; |
| 850 |
ComplexType rad = sqrt(B * B - static_cast<ComplexType>(4) * A * C); |
| 851 |
ComplexType denom1 = B + rad; |
| 852 |
ComplexType denom2 = B - rad; |
| 853 |
ComplexType correction = (z1 - z2) * static_cast<ComplexType>(2) * C; |
| 854 |
if (norm(denom1) > norm(denom2)) |
| 855 |
{ |
| 856 |
correction /= denom1; |
| 857 |
} |
| 858 |
else |
| 859 |
{ |
| 860 |
correction /= denom2; |
| 861 |
} |
| 862 |
|
| 863 |
z0 = z1; |
| 864 |
z1 = z2; |
| 865 |
z2 = z2 + correction; |
| 866 |
} |
| 867 |
else |
| 868 |
{ |
| 869 |
z0 = z1; |
| 870 |
z1 = z2; |
| 871 |
z2 = z2 - (pair.first / pair.second); |
| 872 |
} |
| 873 |
|
| 874 |
// See: https://math.stackexchange.com/questions/3017766/constructing-newton-iteration-converging-to-non-root |
| 875 |
// If f' is continuous, then convergence of x_n -> x* implies f(x*) = 0. |
| 876 |
// This condition approximates this convergence condition by requiring three consecutive iterates to be clustered. |
| 877 |
Real tol = (max)(abs(z2) * std::numeric_limits<Real>::epsilon(), std::numeric_limits<Real>::epsilon()); |
| 878 |
bool real_close = abs(z0.real() - z1.real()) < tol && abs(z0.real() - z2.real()) < tol && abs(z1.real() - z2.real()) < tol; |
| 879 |
bool imag_close = abs(z0.imag() - z1.imag()) < tol && abs(z0.imag() - z2.imag()) < tol && abs(z1.imag() - z2.imag()) < tol; |
| 880 |
if (real_close && imag_close) |
| 881 |
{ |
| 882 |
return z2; |
| 883 |
} |
| 884 |
|
| 885 |
} while (max_iterations--); |
| 886 |
|
| 887 |
// The idea is that if we can get abs(f) < eps, we should, but if we go through all these iterations |
| 888 |
// and abs(f) < sqrt(eps), then roundoff error simply does not allow that we can evaluate f to < eps |
| 889 |
// This is somewhat awkward as it isn't scale invariant, but using the Daubechies coefficient example code, |
| 890 |
// I found this condition generates correct roots, whereas the scale invariant condition discussed here: |
| 891 |
// https://scicomp.stackexchange.com/questions/30597/defining-a-condition-number-and-termination-criteria-for-newtons-method |
| 892 |
// allows nonroots to be passed off as roots. |
| 893 |
auto pair = g(z2); |
| 894 |
if (abs(pair.first) < sqrt(std::numeric_limits<Real>::epsilon())) |
| 895 |
{ |
| 896 |
return z2; |
| 897 |
} |
| 898 |
|
| 899 |
return { std::numeric_limits<Real>::quiet_NaN(), |
| 900 |
std::numeric_limits<Real>::quiet_NaN() }; |
| 901 |
} |
| 902 |
#endif |
| 903 |
|
| 904 |
|
| 905 |
#if !defined(BOOST_MATH_NO_CXX17_IF_CONSTEXPR) |
| 906 |
// https://stackoverflow.com/questions/48979861/numerically-stable-method-for-solving-quadratic-equations/50065711 |
| 907 |
namespace detail |
| 908 |
{ |
| 909 |
#if defined(BOOST_GNU_STDLIB) && !defined(_GLIBCXX_USE_C99_MATH_TR1) |
| 910 |
inline float fma_workaround(float x, float y, float z) { return ::fmaf(x, y, z); } |
| 911 |
inline double fma_workaround(double x, double y, double z) { return ::fma(x, y, z); } |
| 912 |
#ifndef BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS |
| 913 |
inline long double fma_workaround(long double x, long double y, long double z) { return ::fmal(x, y, z); } |
| 914 |
#endif |
| 915 |
#endif |
| 916 |
template<class T> |
| 917 |
inline T discriminant(T const& a, T const& b, T const& c) |
| 918 |
{ |
| 919 |
T w = 4 * a * c; |
| 920 |
#if defined(BOOST_GNU_STDLIB) && !defined(_GLIBCXX_USE_C99_MATH_TR1) |
| 921 |
T e = fma_workaround(-c, 4 * a, w); |
| 922 |
T f = fma_workaround(b, b, -w); |
| 923 |
#else |
| 924 |
T e = std::fma(-c, 4 * a, w); |
| 925 |
T f = std::fma(b, b, -w); |
| 926 |
#endif |
| 927 |
return f + e; |
| 928 |
} |
| 929 |
|
| 930 |
template<class T> |
| 931 |
std::pair<T, T> quadratic_roots_imp(T const& a, T const& b, T const& c) |
| 932 |
{ |
| 933 |
#if defined(BOOST_GNU_STDLIB) && !defined(_GLIBCXX_USE_C99_MATH_TR1) |
| 934 |
using boost::math::copysign; |
| 935 |
#else |
| 936 |
using std::copysign; |
| 937 |
#endif |
| 938 |
using std::sqrt; |
| 939 |
if constexpr (std::is_floating_point<T>::value) |
| 940 |
{ |
| 941 |
T nan = std::numeric_limits<T>::quiet_NaN(); |
| 942 |
if (a == 0) |
| 943 |
{ |
| 944 |
if (b == 0 && c != 0) |
| 945 |
{ |
| 946 |
return std::pair<T, T>(nan, nan); |
| 947 |
} |
| 948 |
else if (b == 0 && c == 0) |
| 949 |
{ |
| 950 |
return std::pair<T, T>(0, 0); |
| 951 |
} |
| 952 |
return std::pair<T, T>(-c / b, -c / b); |
| 953 |
} |
| 954 |
if (b == 0) |
| 955 |
{ |
| 956 |
T x0_sq = -c / a; |
| 957 |
if (x0_sq < 0) { |
| 958 |
return std::pair<T, T>(nan, nan); |
| 959 |
} |
| 960 |
T x0 = sqrt(x0_sq); |
| 961 |
return std::pair<T, T>(-x0, x0); |
| 962 |
} |
| 963 |
T discriminant = detail::discriminant(a, b, c); |
| 964 |
// Is there a sane way to flush very small negative values to zero? |
| 965 |
// If there is I don't know of it. |
| 966 |
if (discriminant < 0) |
| 967 |
{ |
| 968 |
return std::pair<T, T>(nan, nan); |
| 969 |
} |
| 970 |
T q = -(b + copysign(sqrt(discriminant), b)) / T(2); |
| 971 |
T x0 = q / a; |
| 972 |
T x1 = c / q; |
| 973 |
if (x0 < x1) |
| 974 |
{ |
| 975 |
return std::pair<T, T>(x0, x1); |
| 976 |
} |
| 977 |
return std::pair<T, T>(x1, x0); |
| 978 |
} |
| 979 |
else if constexpr (boost::math::tools::is_complex_type<T>::value) |
| 980 |
{ |
| 981 |
typename T::value_type nan = std::numeric_limits<typename T::value_type>::quiet_NaN(); |
| 982 |
if (a.real() == 0 && a.imag() == 0) |
| 983 |
{ |
| 984 |
using std::norm; |
| 985 |
if (b.real() == 0 && b.imag() && norm(c) != 0) |
| 986 |
{ |
| 987 |
return std::pair<T, T>({ nan, nan }, { nan, nan }); |
| 988 |
} |
| 989 |
else if (b.real() == 0 && b.imag() && c.real() == 0 && c.imag() == 0) |
| 990 |
{ |
| 991 |
return std::pair<T, T>({ 0,0 }, { 0,0 }); |
| 992 |
} |
| 993 |
return std::pair<T, T>(-c / b, -c / b); |
| 994 |
} |
| 995 |
if (b.real() == 0 && b.imag() == 0) |
| 996 |
{ |
| 997 |
T x0_sq = -c / a; |
| 998 |
T x0 = sqrt(x0_sq); |
| 999 |
return std::pair<T, T>(-x0, x0); |
| 1000 |
} |
| 1001 |
// There's no fma for complex types: |
| 1002 |
T discriminant = b * b - T(4) * a * c; |
| 1003 |
T q = -(b + sqrt(discriminant)) / T(2); |
| 1004 |
return std::pair<T, T>(q / a, c / q); |
| 1005 |
} |
| 1006 |
else // Most likely the type is a boost.multiprecision. |
| 1007 |
{ //There is no fma for multiprecision, and in addition it doesn't seem to be useful, so revert to the naive computation. |
| 1008 |
T nan = std::numeric_limits<T>::quiet_NaN(); |
| 1009 |
if (a == 0) |
| 1010 |
{ |
| 1011 |
if (b == 0 && c != 0) |
| 1012 |
{ |
| 1013 |
return std::pair<T, T>(nan, nan); |
| 1014 |
} |
| 1015 |
else if (b == 0 && c == 0) |
| 1016 |
{ |
| 1017 |
return std::pair<T, T>(0, 0); |
| 1018 |
} |
| 1019 |
return std::pair<T, T>(-c / b, -c / b); |
| 1020 |
} |
| 1021 |
if (b == 0) |
| 1022 |
{ |
| 1023 |
T x0_sq = -c / a; |
| 1024 |
if (x0_sq < 0) { |
| 1025 |
return std::pair<T, T>(nan, nan); |
| 1026 |
} |
| 1027 |
T x0 = sqrt(x0_sq); |
| 1028 |
return std::pair<T, T>(-x0, x0); |
| 1029 |
} |
| 1030 |
T discriminant = b * b - 4 * a * c; |
| 1031 |
if (discriminant < 0) |
| 1032 |
{ |
| 1033 |
return std::pair<T, T>(nan, nan); |
| 1034 |
} |
| 1035 |
T q = -(b + copysign(sqrt(discriminant), b)) / T(2); |
| 1036 |
T x0 = q / a; |
| 1037 |
T x1 = c / q; |
| 1038 |
if (x0 < x1) |
| 1039 |
{ |
| 1040 |
return std::pair<T, T>(x0, x1); |
| 1041 |
} |
| 1042 |
return std::pair<T, T>(x1, x0); |
| 1043 |
} |
| 1044 |
} |
| 1045 |
} // namespace detail |
| 1046 |
|
| 1047 |
template<class T1, class T2 = T1, class T3 = T1> |
| 1048 |
inline std::pair<typename tools::promote_args<T1, T2, T3>::type, typename tools::promote_args<T1, T2, T3>::type> quadratic_roots(T1 const& a, T2 const& b, T3 const& c) |
| 1049 |
{ |
| 1050 |
typedef typename tools::promote_args<T1, T2, T3>::type value_type; |
| 1051 |
return detail::quadratic_roots_imp(static_cast<value_type>(a), static_cast<value_type>(b), static_cast<value_type>(c)); |
| 1052 |
} |
| 1053 |
|
| 1054 |
#endif |
| 1055 |
|
| 1056 |
} // namespace tools |
| 1057 |
} // namespace math |
| 1058 |
} // namespace boost |
| 1059 |
|
| 1060 |
#endif // BOOST_MATH_HAS_NVRTC |
| 1061 |
|
| 1062 |
#endif // BOOST_MATH_TOOLS_NEWTON_SOLVER_HPP |