]> git.donarmstrong.com Git - rsem.git/blobdiff - boost/math/special_functions/detail/bessel_jy.hpp
Updated boost to v1.55.0
[rsem.git] / boost / math / special_functions / detail / bessel_jy.hpp
diff --git a/boost/math/special_functions/detail/bessel_jy.hpp b/boost/math/special_functions/detail/bessel_jy.hpp
new file mode 100644 (file)
index 0000000..d341449
--- /dev/null
@@ -0,0 +1,578 @@
+//  Copyright (c) 2006 Xiaogang Zhang
+//  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)
+
+#ifndef BOOST_MATH_BESSEL_JY_HPP
+#define BOOST_MATH_BESSEL_JY_HPP
+
+#ifdef _MSC_VER
+#pragma once
+#endif
+
+#include <boost/math/tools/config.hpp>
+#include <boost/math/special_functions/gamma.hpp>
+#include <boost/math/special_functions/sign.hpp>
+#include <boost/math/special_functions/hypot.hpp>
+#include <boost/math/special_functions/sin_pi.hpp>
+#include <boost/math/special_functions/cos_pi.hpp>
+#include <boost/math/special_functions/detail/bessel_jy_asym.hpp>
+#include <boost/math/special_functions/detail/bessel_jy_series.hpp>
+#include <boost/math/constants/constants.hpp>
+#include <boost/math/policies/error_handling.hpp>
+#include <boost/mpl/if.hpp>
+#include <boost/type_traits/is_floating_point.hpp>
+#include <complex>
+
+// Bessel functions of the first and second kind of fractional order
+
+namespace boost { namespace math {
+
+   namespace detail {
+
+      //
+      // Simultaneous calculation of A&S 9.2.9 and 9.2.10
+      // for use in A&S 9.2.5 and 9.2.6.
+      // This series is quick to evaluate, but divergent unless
+      // x is very large, in fact it's pretty hard to figure out
+      // with any degree of precision when this series actually 
+      // *will* converge!!  Consequently, we may just have to
+      // try it and see...
+      //
+      template <class T, class Policy>
+      bool hankel_PQ(T v, T x, T* p, T* q, const Policy& )
+      {
+         BOOST_MATH_STD_USING
+            T tolerance = 2 * policies::get_epsilon<T, Policy>();
+         *p = 1;
+         *q = 0;
+         T k = 1;
+         T z8 = 8 * x;
+         T sq = 1;
+         T mu = 4 * v * v;
+         T term = 1;
+         bool ok = true;
+         do
+         {
+            term *= (mu - sq * sq) / (k * z8);
+            *q += term;
+            k += 1;
+            sq += 2;
+            T mult = (sq * sq - mu) / (k * z8);
+            ok = fabs(mult) < 0.5f;
+            term *= mult;
+            *p += term;
+            k += 1;
+            sq += 2;
+         }
+         while((fabs(term) > tolerance * *p) && ok);
+         return ok;
+      }
+
+      // Calculate Y(v, x) and Y(v+1, x) by Temme's method, see
+      // Temme, Journal of Computational Physics, vol 21, 343 (1976)
+      template <typename T, typename Policy>
+      int temme_jy(T v, T x, T* Y, T* Y1, const Policy& pol)
+      {
+         T g, h, p, q, f, coef, sum, sum1, tolerance;
+         T a, d, e, sigma;
+         unsigned long k;
+
+         BOOST_MATH_STD_USING
+            using namespace boost::math::tools;
+         using namespace boost::math::constants;
+
+         BOOST_ASSERT(fabs(v) <= 0.5f);  // precondition for using this routine
+
+         T gp = boost::math::tgamma1pm1(v, pol);
+         T gm = boost::math::tgamma1pm1(-v, pol);
+         T spv = boost::math::sin_pi(v, pol);
+         T spv2 = boost::math::sin_pi(v/2, pol);
+         T xp = pow(x/2, v);
+
+         a = log(x / 2);
+         sigma = -a * v;
+         d = abs(sigma) < tools::epsilon<T>() ?
+            T(1) : sinh(sigma) / sigma;
+         e = abs(v) < tools::epsilon<T>() ? T(v*pi<T>()*pi<T>() / 2)
+            : T(2 * spv2 * spv2 / v);
+
+         T g1 = (v == 0) ? T(-euler<T>()) : T((gp - gm) / ((1 + gp) * (1 + gm) * 2 * v));
+         T g2 = (2 + gp + gm) / ((1 + gp) * (1 + gm) * 2);
+         T vspv = (fabs(v) < tools::epsilon<T>()) ? T(1/constants::pi<T>()) : T(v / spv);
+         f = (g1 * cosh(sigma) - g2 * a * d) * 2 * vspv;
+
+         p = vspv / (xp * (1 + gm));
+         q = vspv * xp / (1 + gp);
+
+         g = f + e * q;
+         h = p;
+         coef = 1;
+         sum = coef * g;
+         sum1 = coef * h;
+
+         T v2 = v * v;
+         T coef_mult = -x * x / 4;
+
+         // series summation
+         tolerance = policies::get_epsilon<T, Policy>();
+         for (k = 1; k < policies::get_max_series_iterations<Policy>(); k++)
+         {
+            f = (k * f + p + q) / (k*k - v2);
+            p /= k - v;
+            q /= k + v;
+            g = f + e * q;
+            h = p - k * g;
+            coef *= coef_mult / k;
+            sum += coef * g;
+            sum1 += coef * h;
+            if (abs(coef * g) < abs(sum) * tolerance) 
+            { 
+               break; 
+            }
+         }
+         policies::check_series_iterations<T>("boost::math::bessel_jy<%1%>(%1%,%1%) in temme_jy", k, pol);
+         *Y = -sum;
+         *Y1 = -2 * sum1 / x;
+
+         return 0;
+      }
+
+      // Evaluate continued fraction fv = J_(v+1) / J_v, see
+      // Abramowitz and Stegun, Handbook of Mathematical Functions, 1972, 9.1.73
+      template <typename T, typename Policy>
+      int CF1_jy(T v, T x, T* fv, int* sign, const Policy& pol)
+      {
+         T C, D, f, a, b, delta, tiny, tolerance;
+         unsigned long k;
+         int s = 1;
+
+         BOOST_MATH_STD_USING
+
+            // |x| <= |v|, CF1_jy converges rapidly
+            // |x| > |v|, CF1_jy needs O(|x|) iterations to converge
+
+            // modified Lentz's method, see
+            // Lentz, Applied Optics, vol 15, 668 (1976)
+            tolerance = 2 * policies::get_epsilon<T, Policy>();;
+         tiny = sqrt(tools::min_value<T>());
+         C = f = tiny;                           // b0 = 0, replace with tiny
+         D = 0;
+         for (k = 1; k < policies::get_max_series_iterations<Policy>() * 100; k++)
+         {
+            a = -1;
+            b = 2 * (v + k) / x;
+            C = b + a / C;
+            D = b + a * D;
+            if (C == 0) { C = tiny; }
+            if (D == 0) { D = tiny; }
+            D = 1 / D;
+            delta = C * D;
+            f *= delta;
+            if (D < 0) { s = -s; }
+            if (abs(delta - 1) < tolerance) 
+            { break; }
+         }
+         policies::check_series_iterations<T>("boost::math::bessel_jy<%1%>(%1%,%1%) in CF1_jy", k / 100, pol);
+         *fv = -f;
+         *sign = s;                              // sign of denominator
+
+         return 0;
+      }
+      //
+      // This algorithm was originally written by Xiaogang Zhang
+      // using std::complex to perform the complex arithmetic.
+      // However, that turns out to 10x or more slower than using
+      // all real-valued arithmetic, so it's been rewritten using
+      // real values only.
+      //
+      template <typename T, typename Policy>
+      int CF2_jy(T v, T x, T* p, T* q, const Policy& pol)
+      {
+         BOOST_MATH_STD_USING
+
+            T Cr, Ci, Dr, Di, fr, fi, a, br, bi, delta_r, delta_i, temp;
+         T tiny;
+         unsigned long k;
+
+         // |x| >= |v|, CF2_jy converges rapidly
+         // |x| -> 0, CF2_jy fails to converge
+         BOOST_ASSERT(fabs(x) > 1);
+
+         // modified Lentz's method, complex numbers involved, see
+         // Lentz, Applied Optics, vol 15, 668 (1976)
+         T tolerance = 2 * policies::get_epsilon<T, Policy>();
+         tiny = sqrt(tools::min_value<T>());
+         Cr = fr = -0.5f / x;
+         Ci = fi = 1;
+         //Dr = Di = 0;
+         T v2 = v * v;
+         a = (0.25f - v2) / x; // Note complex this one time only!
+         br = 2 * x;
+         bi = 2;
+         temp = Cr * Cr + 1;
+         Ci = bi + a * Cr / temp;
+         Cr = br + a / temp;
+         Dr = br;
+         Di = bi;
+         if (fabs(Cr) + fabs(Ci) < tiny) { Cr = tiny; }
+         if (fabs(Dr) + fabs(Di) < tiny) { Dr = tiny; }
+         temp = Dr * Dr + Di * Di;
+         Dr = Dr / temp;
+         Di = -Di / temp;
+         delta_r = Cr * Dr - Ci * Di;
+         delta_i = Ci * Dr + Cr * Di;
+         temp = fr;
+         fr = temp * delta_r - fi * delta_i;
+         fi = temp * delta_i + fi * delta_r;
+         for (k = 2; k < policies::get_max_series_iterations<Policy>(); k++)
+         {
+            a = k - 0.5f;
+            a *= a;
+            a -= v2;
+            bi += 2;
+            temp = Cr * Cr + Ci * Ci;
+            Cr = br + a * Cr / temp;
+            Ci = bi - a * Ci / temp;
+            Dr = br + a * Dr;
+            Di = bi + a * Di;
+            if (fabs(Cr) + fabs(Ci) < tiny) { Cr = tiny; }
+            if (fabs(Dr) + fabs(Di) < tiny) { Dr = tiny; }
+            temp = Dr * Dr + Di * Di;
+            Dr = Dr / temp;
+            Di = -Di / temp;
+            delta_r = Cr * Dr - Ci * Di;
+            delta_i = Ci * Dr + Cr * Di;
+            temp = fr;
+            fr = temp * delta_r - fi * delta_i;
+            fi = temp * delta_i + fi * delta_r;
+            if (fabs(delta_r - 1) + fabs(delta_i) < tolerance)
+               break;
+         }
+         policies::check_series_iterations<T>("boost::math::bessel_jy<%1%>(%1%,%1%) in CF2_jy", k, pol);
+         *p = fr;
+         *q = fi;
+
+         return 0;
+      }
+
+      static const int need_j = 1;
+      static const int need_y = 2;
+
+      // Compute J(v, x) and Y(v, x) simultaneously by Steed's method, see
+      // Barnett et al, Computer Physics Communications, vol 8, 377 (1974)
+      template <typename T, typename Policy>
+      int bessel_jy(T v, T x, T* J, T* Y, int kind, const Policy& pol)
+      {
+         BOOST_ASSERT(x >= 0);
+
+         T u, Jv, Ju, Yv, Yv1, Yu, Yu1(0), fv, fu;
+         T W, p, q, gamma, current, prev, next;
+         bool reflect = false;
+         unsigned n, k;
+         int s;
+         int org_kind = kind;
+         T cp = 0;
+         T sp = 0;
+
+         static const char* function = "boost::math::bessel_jy<%1%>(%1%,%1%)";
+
+         BOOST_MATH_STD_USING
+            using namespace boost::math::tools;
+         using namespace boost::math::constants;
+
+         if (v < 0)
+         {
+            reflect = true;
+            v = -v;                             // v is non-negative from here
+         }
+         if(v > static_cast<T>((std::numeric_limits<int>::max)()))
+            policies::raise_evaluation_error<T>(function, "Order of Bessel function is too large to evaluate: got %1%", v, pol);
+         n = iround(v, pol);
+         u = v - n;                              // -1/2 <= u < 1/2
+
+         if(reflect)
+         {
+            T z = (u + n % 2);
+            cp = boost::math::cos_pi(z, pol);
+            sp = boost::math::sin_pi(z, pol);
+            if(u != 0)
+               kind = need_j|need_y;               // need both for reflection formula
+         }
+
+         if(x == 0)
+         {
+            if(v == 0)
+               *J = 1;
+            else if((u == 0) || !reflect)
+               *J = 0;
+            else if(kind & need_j)
+               *J = policies::raise_domain_error<T>(function, "Value of Bessel J_v(x) is complex-infinity at %1%", x, pol); // complex infinity
+            else
+               *J = std::numeric_limits<T>::quiet_NaN();  // any value will do, not using J.
+
+            if((kind & need_y) == 0)
+               *Y = std::numeric_limits<T>::quiet_NaN();  // any value will do, not using Y.
+            else if(v == 0)
+               *Y = -policies::raise_overflow_error<T>(function, 0, pol);
+            else
+               *Y = policies::raise_domain_error<T>(function, "Value of Bessel Y_v(x) is complex-infinity at %1%", x, pol); // complex infinity
+            return 1;
+         }
+
+         // x is positive until reflection
+         W = T(2) / (x * pi<T>());               // Wronskian
+         T Yv_scale = 1;
+         if(((kind & need_y) == 0) && ((x < 1) || (v > x * x / 4) || (x < 5)))
+         {
+            //
+            // This series will actually converge rapidly for all small
+            // x - say up to x < 20 - but the first few terms are large
+            // and divergent which leads to large errors :-(
+            //
+            Jv = bessel_j_small_z_series(v, x, pol);
+            Yv = std::numeric_limits<T>::quiet_NaN();
+         }
+         else if((x < 1) && (u != 0) && (log(policies::get_epsilon<T, Policy>() / 2) > v * log((x/2) * (x/2) / v)))
+         {
+            // Evaluate using series representations.
+            // This is particularly important for x << v as in this
+            // area temme_jy may be slow to converge, if it converges at all.
+            // Requires x is not an integer.
+            if(kind&need_j)
+               Jv = bessel_j_small_z_series(v, x, pol);
+            else
+               Jv = std::numeric_limits<T>::quiet_NaN();
+            if((org_kind&need_y && (!reflect || (cp != 0))) 
+               || (org_kind & need_j && (reflect && (sp != 0))))
+            {
+               // Only calculate if we need it, and if the reflection formula will actually use it:
+               Yv = bessel_y_small_z_series(v, x, &Yv_scale, pol);
+            }
+            else
+               Yv = std::numeric_limits<T>::quiet_NaN();
+         }
+         else if((u == 0) && (x < policies::get_epsilon<T, Policy>()))
+         {
+            // Truncated series evaluation for small x and v an integer,
+            // much quicker in this area than temme_jy below.
+            if(kind&need_j)
+               Jv = bessel_j_small_z_series(v, x, pol);
+            else
+               Jv = std::numeric_limits<T>::quiet_NaN();
+            if((org_kind&need_y && (!reflect || (cp != 0))) 
+               || (org_kind & need_j && (reflect && (sp != 0))))
+            {
+               // Only calculate if we need it, and if the reflection formula will actually use it:
+               Yv = bessel_yn_small_z(n, x, &Yv_scale, pol);
+            }
+            else
+               Yv = std::numeric_limits<T>::quiet_NaN();
+         }
+         else if(asymptotic_bessel_large_x_limit(v, x))
+         {
+            if(kind&need_y)
+            {
+               Yv = asymptotic_bessel_y_large_x_2(v, x);
+            }
+            else
+               Yv = std::numeric_limits<T>::quiet_NaN(); // any value will do, we're not using it.
+            if(kind&need_j)
+            {
+               Jv = asymptotic_bessel_j_large_x_2(v, x);
+            }
+            else
+               Jv = std::numeric_limits<T>::quiet_NaN(); // any value will do, we're not using it.
+         }
+         else if((x > 8) && hankel_PQ(v, x, &p, &q, pol))
+         {
+            //
+            // Hankel approximation: note that this method works best when x 
+            // is large, but in that case we end up calculating sines and cosines
+            // of large values, with horrendous resulting accuracy.  It is fast though
+            // when it works....
+            //
+            // Normally we calculate sin/cos(chi) where:
+            //
+            // chi = x - fmod(T(v / 2 + 0.25f), T(2)) * boost::math::constants::pi<T>();
+            //
+            // But this introduces large errors, so use sin/cos addition formulae to
+            // improve accuracy:
+            //
+            T mod_v = fmod(T(v / 2 + 0.25f), T(2));
+            T sx = sin(x);
+            T cx = cos(x);
+            T sv = sin_pi(mod_v);
+            T cv = cos_pi(mod_v);
+
+            T sc = sx * cv - sv * cx; // == sin(chi);
+            T cc = cx * cv + sx * sv; // == cos(chi);
+            T chi = boost::math::constants::root_two<T>() / (boost::math::constants::root_pi<T>() * sqrt(x)); //sqrt(2 / (boost::math::constants::pi<T>() * x));
+            Yv = chi * (p * sc + q * cc);
+            Jv = chi * (p * cc - q * sc);
+         }
+         else if (x <= 2)                           // x in (0, 2]
+         {
+            if(temme_jy(u, x, &Yu, &Yu1, pol))             // Temme series
+            {
+               // domain error:
+               *J = *Y = Yu;
+               return 1;
+            }
+            prev = Yu;
+            current = Yu1;
+            T scale = 1;
+            policies::check_series_iterations<T>(function, n, pol);
+            for (k = 1; k <= n; k++)            // forward recurrence for Y
+            {
+               T fact = 2 * (u + k) / x;
+               if((tools::max_value<T>() - fabs(prev)) / fact < fabs(current))
+               {
+                  scale /= current;
+                  prev /= current;
+                  current = 1;
+               }
+               next = fact * current - prev;
+               prev = current;
+               current = next;
+            }
+            Yv = prev;
+            Yv1 = current;
+            if(kind&need_j)
+            {
+               CF1_jy(v, x, &fv, &s, pol);                 // continued fraction CF1_jy
+               Jv = scale * W / (Yv * fv - Yv1);           // Wronskian relation
+            }
+            else
+               Jv = std::numeric_limits<T>::quiet_NaN(); // any value will do, we're not using it.
+            Yv_scale = scale;
+         }
+         else                                    // x in (2, \infty)
+         {
+            // Get Y(u, x):
+
+            T ratio;
+            CF1_jy(v, x, &fv, &s, pol);
+            // tiny initial value to prevent overflow
+            T init = sqrt(tools::min_value<T>());
+            prev = fv * s * init;
+            current = s * init;
+            if(v < max_factorial<T>::value)
+            {
+               policies::check_series_iterations<T>(function, n, pol);
+               for (k = n; k > 0; k--)             // backward recurrence for J
+               {
+                  next = 2 * (u + k) * current / x - prev;
+                  prev = current;
+                  current = next;
+               }
+               ratio = (s * init) / current;     // scaling ratio
+               // can also call CF1_jy() to get fu, not much difference in precision
+               fu = prev / current;
+            }
+            else
+            {
+               //
+               // When v is large we may get overflow in this calculation
+               // leading to NaN's and other nasty surprises:
+               //
+               policies::check_series_iterations<T>(function, n, pol);
+               bool over = false;
+               for (k = n; k > 0; k--)             // backward recurrence for J
+               {
+                  T t = 2 * (u + k) / x;
+                  if((t > 1) && (tools::max_value<T>() / t < current))
+                  {
+                     over = true;
+                     break;
+                  }
+                  next = t * current - prev;
+                  prev = current;
+                  current = next;
+               }
+               if(!over)
+               {
+                  ratio = (s * init) / current;     // scaling ratio
+                  // can also call CF1_jy() to get fu, not much difference in precision
+                  fu = prev / current;
+               }
+               else
+               {
+                  ratio = 0;
+                  fu = 1;
+               }
+            }
+            CF2_jy(u, x, &p, &q, pol);                  // continued fraction CF2_jy
+            T t = u / x - fu;                   // t = J'/J
+            gamma = (p - t) / q;
+            //
+            // We can't allow gamma to cancel out to zero competely as it messes up
+            // the subsequent logic.  So pretend that one bit didn't cancel out
+            // and set to a suitably small value.  The only test case we've been able to
+            // find for this, is when v = 8.5 and x = 4*PI.
+            //
+            if(gamma == 0)
+            {
+               gamma = u * tools::epsilon<T>() / x;
+            }
+            Ju = sign(current) * sqrt(W / (q + gamma * (p - t)));
+
+            Jv = Ju * ratio;                    // normalization
+
+            Yu = gamma * Ju;
+            Yu1 = Yu * (u/x - p - q/gamma);
+
+            if(kind&need_y)
+            {
+               // compute Y:
+               prev = Yu;
+               current = Yu1;
+               policies::check_series_iterations<T>(function, n, pol);
+               for (k = 1; k <= n; k++)            // forward recurrence for Y
+               {
+                  T fact = 2 * (u + k) / x;
+                  if((tools::max_value<T>() - fabs(prev)) / fact < fabs(current))
+                  {
+                     prev /= current;
+                     Yv_scale /= current;
+                     current = 1;
+                  }
+                  next = fact * current - prev;
+                  prev = current;
+                  current = next;
+               }
+               Yv = prev;
+            }
+            else
+               Yv = std::numeric_limits<T>::quiet_NaN(); // any value will do, we're not using it.
+         }
+
+         if (reflect)
+         {
+            if((sp != 0) && (tools::max_value<T>() * fabs(Yv_scale) < fabs(sp * Yv)))
+               *J = org_kind & need_j ? T(-sign(sp) * sign(Yv) * sign(Yv_scale) * policies::raise_overflow_error<T>(function, 0, pol)) : T(0);
+            else
+               *J = cp * Jv - (sp == 0 ? T(0) : T((sp * Yv) / Yv_scale));     // reflection formula
+            if((cp != 0) && (tools::max_value<T>() * fabs(Yv_scale) < fabs(cp * Yv)))
+               *Y = org_kind & need_y ? T(-sign(cp) * sign(Yv) * sign(Yv_scale) * policies::raise_overflow_error<T>(function, 0, pol)) : T(0);
+            else
+               *Y = (sp != 0 ? sp * Jv : T(0)) + (cp == 0 ? T(0) : T((cp * Yv) / Yv_scale));
+         }
+         else
+         {
+            *J = Jv;
+            if(tools::max_value<T>() * fabs(Yv_scale) < fabs(Yv))
+               *Y = org_kind & need_y ? T(sign(Yv) * sign(Yv_scale) * policies::raise_overflow_error<T>(function, 0, pol)) : T(0);
+            else
+               *Y = Yv / Yv_scale;
+         }
+
+         return 0;
+      }
+
+   } // namespace detail
+
+}} // namespaces
+
+#endif // BOOST_MATH_BESSEL_JY_HPP
+