boost/math/special_functions/detail/bessel_jy_zero.hpp
// Copyright (c) 2013 Christopher Kormanyos // Use, modification and distribution are subject to the // Boost Software License, Version 1.0. (See accompanying file // LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) // // This work is based on an earlier work: // "Algorithm 910: A Portable C++ Multiple-Precision System for Special-Function Calculations", // in ACM TOMS, {VOL 37, ISSUE 4, (February 2011)} (C) ACM, 2011. http://doi.acm.org/10.1145/1916461.1916469 // // This header contains implementation details for estimating the zeros // of cylindrical Bessel and Neumann functions on the positive real axis. // Support is included for both positive as well as negative order. // Various methods are used to estimate the roots. These include // empirical curve fitting and McMahon's asymptotic approximation // for small order, uniform asymptotic expansion for large order, // and iteration and root interlacing for negative order. // #ifndef BOOST_MATH_BESSEL_JY_ZERO_2013_01_18_HPP_ #define BOOST_MATH_BESSEL_JY_ZERO_2013_01_18_HPP_ #include <boost/math/tools/config.hpp> #include <boost/math/tools/tuple.hpp> #include <boost/math/tools/precision.hpp> #include <boost/math/tools/cstdint.hpp> #include <boost/math/tools/roots.hpp> #include <boost/math/constants/constants.hpp> #include <boost/math/special_functions/cbrt.hpp> #include <boost/math/special_functions/detail/airy_ai_bi_zero.hpp> #ifndef BOOST_MATH_HAS_NVRTC #include <boost/math/special_functions/math_fwd.hpp> #endif #ifdef BOOST_MATH_ENABLE_CUDA # pragma nv_diag_suppress 20012 #endif namespace boost { namespace math { namespace detail { namespace bessel_zero { template<class T> BOOST_MATH_GPU_ENABLED T equation_nist_10_21_19(const T& v, const T& a) { // Get the initial estimate of the m'th root of Jv or Yv. // This subroutine is used for the order m with m > 1. // The order m has been used to create the input parameter a. // This is Eq. 10.21.19 in the NIST Handbook. const T mu = (v * v) * 4U; const T mu_minus_one = mu - T(1); const T eight_a_inv = T(1) / (a * 8U); const T eight_a_inv_squared = eight_a_inv * eight_a_inv; const T term3 = ((mu_minus_one * 4U) * ((mu * 7U) - T(31U) )) / 3U; const T term5 = ((mu_minus_one * 32U) * ((((mu * 83U) - T(982U) ) * mu) + T(3779U) )) / 15U; const T term7 = ((mu_minus_one * 64U) * ((((((mu * 6949U) - T(153855UL)) * mu) + T(1585743UL)) * mu) - T(6277237UL))) / 105U; return a + (((( - term7 * eight_a_inv_squared - term5) * eight_a_inv_squared - term3) * eight_a_inv_squared - mu_minus_one) * eight_a_inv); } template<typename T> class equation_as_9_3_39_and_its_derivative { public: BOOST_MATH_GPU_ENABLED explicit equation_as_9_3_39_and_its_derivative(const T& zt) : zeta(zt) { } BOOST_MATH_GPU_ENABLED equation_as_9_3_39_and_its_derivative(const equation_as_9_3_39_and_its_derivative&) = default; BOOST_MATH_GPU_ENABLED boost::math::tuple<T, T> operator()(const T& z) const { BOOST_MATH_STD_USING // ADL of std names, needed for acos, sqrt. // Return the function of zeta that is implicitly defined // in A&S Eq. 9.3.39 as a function of z. The function is // returned along with its derivative with respect to z. const T zsq_minus_one_sqrt = sqrt((z * z) - T(1)); const T the_function( zsq_minus_one_sqrt - ( acos(T(1) / z) + ((T(2) / 3U) * (zeta * sqrt(zeta))))); const T its_derivative(zsq_minus_one_sqrt / z); return boost::math::tuple<T, T>(the_function, its_derivative); } private: const equation_as_9_3_39_and_its_derivative& operator=(const equation_as_9_3_39_and_its_derivative&) = delete; const T zeta; }; template<class T, class Policy> BOOST_MATH_GPU_ENABLED T equation_as_9_5_26(const T& v, const T& ai_bi_root, const Policy& pol) { BOOST_MATH_STD_USING // ADL of std names, needed for log, sqrt. // Obtain the estimate of the m'th zero of Jv or Yv. // The order m has been used to create the input parameter ai_bi_root. // Here, v is larger than about 2.2. The estimate is computed // from Abramowitz and Stegun Eqs. 9.5.22 and 9.5.26, page 371. // // The inversion of z as a function of zeta is mentioned in the text // following A&S Eq. 9.5.26. Here, we accomplish the inversion by // performing a Taylor expansion of Eq. 9.3.39 for large z to order 2 // and solving the resulting quadratic equation, thereby taking // the positive root of the quadratic. // In other words: (2/3)(-zeta)^(3/2) approx = z + 1/(2z) - pi/2. // This leads to: z^2 - [(2/3)(-zeta)^(3/2) + pi/2]z + 1/2 = 0. // // With this initial estimate, Newton-Raphson iteration is used // to refine the value of the estimate of the root of z // as a function of zeta. const T v_pow_third(boost::math::cbrt(v, pol)); const T v_pow_minus_two_thirds(T(1) / (v_pow_third * v_pow_third)); // Obtain zeta using the order v combined with the m'th root of // an airy function, as shown in A&S Eq. 9.5.22. const T zeta = v_pow_minus_two_thirds * (-ai_bi_root); const T zeta_sqrt = sqrt(zeta); // Set up a quadratic equation based on the Taylor series // expansion mentioned above. const T b = -((((zeta * zeta_sqrt) * 2U) / 3U) + boost::math::constants::half_pi<T>()); // Solve the quadratic equation, taking the positive root. const T z_estimate = (-b + sqrt((b * b) - T(2))) / 2U; // Establish the range, the digits, and the iteration limit // for the upcoming root-finding. const T range_zmin = (std::max<T>)(z_estimate - T(1), T(1)); const T range_zmax = z_estimate + T(1); const auto my_digits10 = static_cast<int>(static_cast<float>(boost::math::tools::digits<T>() * 0.301F)); // Select the maximum allowed iterations based on the number // of decimal digits in the numeric type T, being at least 12. const auto iterations_allowed = static_cast<boost::math::uintmax_t>(BOOST_MATH_GPU_SAFE_MAX(12, my_digits10 * 2)); boost::math::uintmax_t iterations_used = iterations_allowed; // Calculate the root of z as a function of zeta. const T z = boost::math::tools::newton_raphson_iterate( boost::math::detail::bessel_zero::equation_as_9_3_39_and_its_derivative<T>(zeta), z_estimate, range_zmin, range_zmax, BOOST_MATH_GPU_SAFE_MIN(boost::math::tools::digits<T>(), boost::math::tools::digits<float>()), iterations_used); static_cast<void>(iterations_used); // Continue with the implementation of A&S Eq. 9.3.39. const T zsq_minus_one = (z * z) - T(1); const T zsq_minus_one_sqrt = sqrt(zsq_minus_one); // This is A&S Eq. 9.3.42. const T b0_term_5_24 = T(5) / ((zsq_minus_one * zsq_minus_one_sqrt) * 24U); const T b0_term_1_8 = T(1) / ( zsq_minus_one_sqrt * 8U); const T b0_term_5_48 = T(5) / ((zeta * zeta) * 48U); const T b0 = -b0_term_5_48 + ((b0_term_5_24 + b0_term_1_8) / zeta_sqrt); // This is the second line of A&S Eq. 9.5.26 for f_k with k = 1. const T f1 = ((z * zeta_sqrt) * b0) / zsq_minus_one_sqrt; // This is A&S Eq. 9.5.22 expanded to k = 1 (i.e., one term in the series). return (v * z) + (f1 / v); } namespace cyl_bessel_j_zero_detail { template<class T, class Policy> BOOST_MATH_GPU_ENABLED T equation_nist_10_21_40_a(const T& v, const Policy& pol) { const T v_pow_third(boost::math::cbrt(v, pol)); const T v_pow_minus_two_thirds(T(1) / (v_pow_third * v_pow_third)); return v * ((((( + T(0.043) * v_pow_minus_two_thirds - T(0.0908)) * v_pow_minus_two_thirds - T(0.00397)) * v_pow_minus_two_thirds + T(1.033150)) * v_pow_minus_two_thirds + T(1.8557571)) * v_pow_minus_two_thirds + T(1)); } template<class T, class Policy> class function_object_jv { public: BOOST_MATH_GPU_ENABLED function_object_jv(const T& v, const Policy& pol) : my_v(v), my_pol(pol) { } BOOST_MATH_GPU_ENABLED function_object_jv(const function_object_jv&) = default; BOOST_MATH_GPU_ENABLED T operator()(const T& x) const { return boost::math::cyl_bessel_j(my_v, x, my_pol); } private: const T my_v; const Policy& my_pol; const function_object_jv& operator=(const function_object_jv&) = delete; }; template<class T, class Policy> class function_object_jv_and_jv_prime { public: BOOST_MATH_GPU_ENABLED function_object_jv_and_jv_prime( const T& v, const bool order_is_zero, const Policy& pol) : my_v(v), my_order_is_zero(order_is_zero), my_pol(pol) { } function_object_jv_and_jv_prime(const function_object_jv_and_jv_prime&) = default; BOOST_MATH_GPU_ENABLED boost::math::tuple<T, T> operator()(const T& x) const { // Obtain Jv(x) and Jv'(x). // Chris's original code called the Bessel function implementation layer direct, // but that circumvented optimizations for integer-orders. Call the documented // top level functions instead, and let them sort out which implementation to use. T j_v; T j_v_prime; if(my_order_is_zero) { j_v = boost::math::cyl_bessel_j(0, x, my_pol); j_v_prime = -boost::math::cyl_bessel_j(1, x, my_pol); } else { j_v = boost::math::cyl_bessel_j( my_v, x, my_pol); const T j_v_m1 (boost::math::cyl_bessel_j(T(my_v - 1), x, my_pol)); j_v_prime = j_v_m1 - ((my_v * j_v) / x); } // Return a tuple containing both Jv(x) and Jv'(x). return boost::math::make_tuple(j_v, j_v_prime); } private: const T my_v; const bool my_order_is_zero; const Policy& my_pol; const function_object_jv_and_jv_prime& operator=(const function_object_jv_and_jv_prime&) = delete; }; template<class T> BOOST_MATH_GPU_ENABLED bool my_bisection_unreachable_tolerance(const T&, const T&) { return false; } template<class T, class Policy> BOOST_MATH_GPU_ENABLED T initial_guess(const T& v, const int m, const Policy& pol) { BOOST_MATH_STD_USING // ADL of std names, needed for floor. // Compute an estimate of the m'th root of cyl_bessel_j. T guess; // There is special handling for negative order. if(v < 0) { if((m == 1) && (v > -0.5F)) { // For small, negative v, use the results of empirical curve fitting. // Mathematica(R) session for the coefficients: // Table[{n, BesselJZero[n, 1]}, {n, -(1/2), 0, 1/10}] // N[%, 20] // Fit[%, {n^0, n^1, n^2, n^3, n^4, n^5, n^6}, n] guess = ((((( - T(0.2321156900729) * v - T(0.1493247777488)) * v - T(0.15205419167239)) * v + T(0.07814930561249)) * v - T(0.17757573537688)) * v + T(1.542805677045663)) * v + T(2.40482555769577277); return guess; } // Create the positive order and extract its positive floor integer part. const T vv(-v); const T vv_floor(floor(vv)); // The to-be-found root is bracketed by the roots of the // Bessel function whose reflected, positive integer order // is less than, but nearest to vv. T root_hi = boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::initial_guess(vv_floor, m, pol); T root_lo; if(m == 1) { // The estimate of the first root for negative order is found using // an adaptive range-searching algorithm. root_lo = T(root_hi - 0.1F); const bool hi_end_of_bracket_is_negative = (boost::math::cyl_bessel_j(v, root_hi, pol) < 0); while((root_lo > boost::math::tools::epsilon<T>())) { const bool lo_end_of_bracket_is_negative = (boost::math::cyl_bessel_j(v, root_lo, pol) < 0); if(hi_end_of_bracket_is_negative != lo_end_of_bracket_is_negative) { break; } root_hi = root_lo; // Decrease the lower end of the bracket using an adaptive algorithm. if(root_lo > 0.5F) { root_lo -= 0.5F; } else { root_lo *= 0.75F; // LCOV_EXCL_LINE probably unreachable, but hard to prove? } } } else { root_lo = boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::initial_guess(vv_floor, m - 1, pol); } // Perform several steps of bisection iteration to refine the guess. boost::math::uintmax_t number_of_iterations(12U); // Do the bisection iteration. const boost::math::tuple<T, T> guess_pair = boost::math::tools::bisect( boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::function_object_jv<T, Policy>(v, pol), root_lo, root_hi, boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::my_bisection_unreachable_tolerance<T>, number_of_iterations); return (boost::math::get<0>(guess_pair) + boost::math::get<1>(guess_pair)) / 2U; } if(m == 1U) { // Get the initial estimate of the first root. if(v < 2.2F) { // For small v, use the results of empirical curve fitting. // Mathematica(R) session for the coefficients: // Table[{n, BesselJZero[n, 1]}, {n, 0, 22/10, 1/10}] // N[%, 20] // Fit[%, {n^0, n^1, n^2, n^3, n^4, n^5, n^6}, n] guess = ((((( - T(0.0008342379046010) * v + T(0.007590035637410)) * v - T(0.030640914772013)) * v + T(0.078232088020106)) * v - T(0.169668712590620)) * v + T(1.542187960073750)) * v + T(2.4048359915254634); } else { // For larger v, use the first line of Eqs. 10.21.40 in the NIST Handbook. guess = boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::equation_nist_10_21_40_a(v, pol); } } else { if(v < 2.2F) { // Use Eq. 10.21.19 in the NIST Handbook. const T a(((v + T(m * 2U)) - T(0.5)) * boost::math::constants::half_pi<T>()); guess = boost::math::detail::bessel_zero::equation_nist_10_21_19(v, a); } else { // Get an estimate of the m'th root of airy_ai. const T airy_ai_root(boost::math::detail::airy_zero::airy_ai_zero_detail::initial_guess<T>(m, pol)); // Use Eq. 9.5.26 in the A&S Handbook. guess = boost::math::detail::bessel_zero::equation_as_9_5_26(v, airy_ai_root, pol); } } return guess; } } // namespace cyl_bessel_j_zero_detail namespace cyl_neumann_zero_detail { template<class T, class Policy> BOOST_MATH_GPU_ENABLED T equation_nist_10_21_40_b(const T& v, const Policy& pol) { const T v_pow_third(boost::math::cbrt(v, pol)); const T v_pow_minus_two_thirds(T(1) / (v_pow_third * v_pow_third)); return v * ((((( - T(0.001) * v_pow_minus_two_thirds - T(0.0060)) * v_pow_minus_two_thirds + T(0.01198)) * v_pow_minus_two_thirds + T(0.260351)) * v_pow_minus_two_thirds + T(0.9315768)) * v_pow_minus_two_thirds + T(1)); } template<class T, class Policy> class function_object_yv { public: BOOST_MATH_GPU_ENABLED function_object_yv(const T& v, const Policy& pol) : my_v(v), my_pol(pol) { } BOOST_MATH_GPU_ENABLED function_object_yv(const function_object_yv&) = default; BOOST_MATH_GPU_ENABLED T operator()(const T& x) const { return boost::math::cyl_neumann(my_v, x, my_pol); } private: const T my_v; const Policy& my_pol; const function_object_yv& operator=(const function_object_yv&) = delete; }; template<class T, class Policy> class function_object_yv_and_yv_prime { public: BOOST_MATH_GPU_ENABLED function_object_yv_and_yv_prime(const T& v, const Policy& pol) : my_v(v), my_pol(pol) { } BOOST_MATH_GPU_ENABLED function_object_yv_and_yv_prime(const function_object_yv_and_yv_prime&) = default; BOOST_MATH_GPU_ENABLED boost::math::tuple<T, T> operator()(const T& x) const { const T half_epsilon(boost::math::tools::epsilon<T>() / 2U); const bool order_is_zero = ((my_v > -half_epsilon) && (my_v < +half_epsilon)); // Obtain Yv(x) and Yv'(x). // Chris's original code called the Bessel function implementation layer direct, // but that circumvented optimizations for integer-orders. Call the documented // top level functions instead, and let them sort out which implementation to use. T y_v; T y_v_prime; if(order_is_zero) { y_v = boost::math::cyl_neumann(0, x, my_pol); y_v_prime = -boost::math::cyl_neumann(1, x, my_pol); } else { y_v = boost::math::cyl_neumann( my_v, x, my_pol); const T y_v_m1 (boost::math::cyl_neumann(T(my_v - 1), x, my_pol)); y_v_prime = y_v_m1 - ((my_v * y_v) / x); } // Return a tuple containing both Yv(x) and Yv'(x). return boost::math::make_tuple(y_v, y_v_prime); } private: const T my_v; const Policy& my_pol; const function_object_yv_and_yv_prime& operator=(const function_object_yv_and_yv_prime&) = delete; }; template<class T> BOOST_MATH_GPU_ENABLED bool my_bisection_unreachable_tolerance(const T&, const T&) { return false; } template<class T, class Policy> BOOST_MATH_GPU_ENABLED T initial_guess(const T& v, const int m, const Policy& pol) { BOOST_MATH_STD_USING // ADL of std names, needed for floor. // Compute an estimate of the m'th root of cyl_neumann. T guess; // There is special handling for negative order. if(v < 0) { // Create the positive order and extract its positive floor and ceiling integer parts. const T vv(-v); const T vv_floor(floor(vv)); // The to-be-found root is bracketed by the roots of the // Bessel function whose reflected, positive integer order // is less than, but nearest to vv. // The special case of negative, half-integer order uses // the relation between Yv and spherical Bessel functions // in order to obtain the bracket for the root. // In these special cases, cyl_neumann(-n/2, x) = sph_bessel_j(+n/2, x) // for v = -n/2. T root_hi; T root_lo; if(m == 1) { // The estimate of the first root for negative order is found using // an adaptive range-searching algorithm. // Take special precautions for the discontinuity at negative, // half-integer orders and use different brackets above and below these. if(T(vv - vv_floor) < 0.5F) { root_hi = boost::math::detail::bessel_zero::cyl_neumann_zero_detail::initial_guess(vv_floor, m, pol); } else { root_hi = boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::initial_guess(T(vv_floor + 0.5F), m, pol); } root_lo = T(root_hi - 0.1F); const bool hi_end_of_bracket_is_negative = (boost::math::cyl_neumann(v, root_hi, pol) < 0); while((root_lo > boost::math::tools::epsilon<T>())) { const bool lo_end_of_bracket_is_negative = (boost::math::cyl_neumann(v, root_lo, pol) < 0); if(hi_end_of_bracket_is_negative != lo_end_of_bracket_is_negative) { break; } root_hi = root_lo; // Decrease the lower end of the bracket using an adaptive algorithm. if(root_lo > 0.5F) { root_lo -= 0.5F; } else { root_lo *= 0.75F; // LCOV_EXCL_LINE probably unreachable, but hard to prove? } } } else { if(T(vv - vv_floor) < 0.5F) { root_lo = boost::math::detail::bessel_zero::cyl_neumann_zero_detail::initial_guess(vv_floor, m - 1, pol); root_hi = boost::math::detail::bessel_zero::cyl_neumann_zero_detail::initial_guess(vv_floor, m, pol); root_lo += 0.01F; root_hi += 0.01F; } else { root_lo = boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::initial_guess(T(vv_floor + 0.5F), m - 1, pol); root_hi = boost::math::detail::bessel_zero::cyl_bessel_j_zero_detail::initial_guess(T(vv_floor + 0.5F), m, pol); root_lo += 0.01F; root_hi += 0.01F; } } // Perform several steps of bisection iteration to refine the guess. boost::math::uintmax_t number_of_iterations(12U); // Do the bisection iteration. const boost::math::tuple<T, T> guess_pair = boost::math::tools::bisect( boost::math::detail::bessel_zero::cyl_neumann_zero_detail::function_object_yv<T, Policy>(v, pol), root_lo, root_hi, boost::math::detail::bessel_zero::cyl_neumann_zero_detail::my_bisection_unreachable_tolerance<T>, number_of_iterations); return (boost::math::get<0>(guess_pair) + boost::math::get<1>(guess_pair)) / 2U; } if(m == 1U) { // Get the initial estimate of the first root. if(v < 2.2F) { // For small v, use the results of empirical curve fitting. // Mathematica(R) session for the coefficients: // Table[{n, BesselYZero[n, 1]}, {n, 0, 22/10, 1/10}] // N[%, 20] // Fit[%, {n^0, n^1, n^2, n^3, n^4, n^5, n^6}, n] guess = ((((( - T(0.0025095909235652) * v + T(0.021291887049053)) * v - T(0.076487785486526)) * v + T(0.159110268115362)) * v - T(0.241681668765196)) * v + T(1.4437846310885244)) * v + T(0.89362115190200490); } else { // For larger v, use the second line of Eqs. 10.21.40 in the NIST Handbook. guess = boost::math::detail::bessel_zero::cyl_neumann_zero_detail::equation_nist_10_21_40_b(v, pol); } } else { if(v < 2.2F) { // Use Eq. 10.21.19 in the NIST Handbook. const T a(((v + T(m * 2U)) - T(1.5)) * boost::math::constants::half_pi<T>()); guess = boost::math::detail::bessel_zero::equation_nist_10_21_19(v, a); } else { // Get an estimate of the m'th root of airy_bi. const T airy_bi_root(boost::math::detail::airy_zero::airy_bi_zero_detail::initial_guess<T>(m, pol)); // Use Eq. 9.5.26 in the A&S Handbook. guess = boost::math::detail::bessel_zero::equation_as_9_5_26(v, airy_bi_root, pol); } } return guess; } } // namespace cyl_neumann_zero_detail } // namespace bessel_zero } } } // namespace boost::math::detail #ifdef BOOST_MATH_ENABLE_CUDA # pragma nv_diag_default 20012 #endif #endif // BOOST_MATH_BESSEL_JY_ZERO_2013_01_18_HPP_