| 1 |
// (C) Copyright Nick Thompson 2018. |
| 2 |
// Use, modification and distribution are subject to the |
| 3 |
// Boost Software License, Version 1.0. (See accompanying file |
| 4 |
// LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) |
| 5 |
|
| 6 |
#ifndef BOOST_MATH_DIFFERENTIATION_FINITE_DIFFERENCE_HPP |
| 7 |
#define BOOST_MATH_DIFFERENTIATION_FINITE_DIFFERENCE_HPP |
| 8 |
|
| 9 |
/* |
| 10 |
* Performs numerical differentiation by finite-differences. |
| 11 |
* |
| 12 |
* All numerical differentiation using finite-differences are ill-conditioned, and these routines are no exception. |
| 13 |
* A simple argument demonstrates that the error is unbounded as h->0. |
| 14 |
* Take the one sides finite difference formula f'(x) = (f(x+h)-f(x))/h. |
| 15 |
* The evaluation of f induces an error as well as the error from the finite-difference approximation, giving |
| 16 |
* |f'(x) - (f(x+h) -f(x))/h| < h|f''(x)|/2 + (|f(x)|+|f(x+h)|)eps/h =: g(h), where eps is the unit roundoff for the type. |
| 17 |
* It is reasonable to choose h in a way that minimizes the maximum error bound g(h). |
| 18 |
* The value of h that minimizes g is h = sqrt(2eps(|f(x)| + |f(x+h)|)/|f''(x)|), and for this value of h the error bound is |
| 19 |
* sqrt(2eps(|f(x+h) +f(x)||f''(x)|)). |
| 20 |
* In fact it is not necessary to compute the ratio (|f(x+h)| + |f(x)|)/|f''(x)|; the error bound of ~\sqrt{\epsilon} still holds if we set it to one. |
| 21 |
* |
| 22 |
* |
| 23 |
* For more details on this method of analysis, see |
| 24 |
* |
| 25 |
* http://www.uio.no/studier/emner/matnat/math/MAT-INF1100/h08/kompendiet/diffint.pdf |
| 26 |
* http://web.archive.org/web/20150420195907/http://www.uio.no/studier/emner/matnat/math/MAT-INF1100/h08/kompendiet/diffint.pdf |
| 27 |
* |
| 28 |
* |
| 29 |
* It can be shown on general grounds that when choosing the optimal h, the maximum error in f'(x) is ~(|f(x)|eps)^k/k+1|f^(k-1)(x)|^1/k+1. |
| 30 |
* From this we can see that full precision can be recovered in the limit k->infinity. |
| 31 |
* |
| 32 |
* References: |
| 33 |
* |
| 34 |
* 1) Fornberg, Bengt. "Generation of finite difference formulas on arbitrarily spaced grids." Mathematics of computation 51.184 (1988): 699-706. |
| 35 |
* |
| 36 |
* |
| 37 |
* The second algorithm, the complex step derivative, is not ill-conditioned. |
| 38 |
* However, it requires that your function can be evaluated at complex arguments. |
| 39 |
* The idea is that f(x+ih) = f(x) +ihf'(x) - h^2f''(x) + ... so f'(x) \approx Im[f(x+ih)]/h. |
| 40 |
* No subtractive cancellation occurs. The error is ~ eps|f'(x)| + eps^2|f'''(x)|/6; hard to beat that. |
| 41 |
* |
| 42 |
* References: |
| 43 |
* |
| 44 |
* 1) Squire, William, and George Trapp. "Using complex variables to estimate derivatives of real functions." Siam Review 40.1 (1998): 110-112. |
| 45 |
*/ |
| 46 |
|
| 47 |
#include <complex> |
| 48 |
#include <boost/math/special_functions/next.hpp> |
| 49 |
|
| 50 |
namespace boost{ namespace math{ namespace differentiation { |
| 51 |
|
| 52 |
namespace detail { |
| 53 |
template<class Real> |
| 54 |
Real make_xph_representable(Real x, Real h) |
| 55 |
{ |
| 56 |
using std::numeric_limits; |
| 57 |
// Redefine h so that x + h is representable. Not using this trick leads to large error. |
| 58 |
// The compiler flag -ffast-math evaporates these operations . . . |
| 59 |
Real temp = x + h; |
| 60 |
h = temp - x; |
| 61 |
// Handle the case x + h == x: |
| 62 |
if (h == 0) |
| 63 |
{ |
| 64 |
h = boost::math::nextafter(x, (numeric_limits<Real>::max)()) - x; |
| 65 |
} |
| 66 |
return h; |
| 67 |
} |
| 68 |
} |
| 69 |
|
| 70 |
template<class F, class Real> |
| 71 |
Real complex_step_derivative(const F f, Real x) |
| 72 |
{ |
| 73 |
// Is it really this easy? Yes. |
| 74 |
// Note that some authors recommend taking the stepsize h to be smaller than epsilon(), some recommending use of the min(). |
| 75 |
// This idea was tested over a few billion test cases and found the make the error *much* worse. |
| 76 |
// Even 2eps and eps/2 made the error worse, which was surprising. |
| 77 |
using std::complex; |
| 78 |
using std::numeric_limits; |
| 79 |
constexpr const Real step = (numeric_limits<Real>::epsilon)(); |
| 80 |
constexpr const Real inv_step = 1/(numeric_limits<Real>::epsilon)(); |
| 81 |
return f(complex<Real>(x, step)).imag()*inv_step; |
| 82 |
} |
| 83 |
|
| 84 |
namespace detail { |
| 85 |
|
| 86 |
template <unsigned> |
| 87 |
struct fd_tag {}; |
| 88 |
|
| 89 |
template<class F, class Real> |
| 90 |
Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<1>&) |
| 91 |
{ |
| 92 |
using std::sqrt; |
| 93 |
using std::pow; |
| 94 |
using std::abs; |
| 95 |
using std::numeric_limits; |
| 96 |
|
| 97 |
const Real eps = (numeric_limits<Real>::epsilon)(); |
| 98 |
// Error bound ~eps^1/2 |
| 99 |
// Note that this estimate of h differs from the best estimate by a factor of sqrt((|f(x)| + |f(x+h)|)/|f''(x)|). |
| 100 |
// Since this factor is invariant under the scaling f -> kf, then we are somewhat justified in approximating it by 1. |
| 101 |
// This approximation will get better as we move to higher orders of accuracy. |
| 102 |
Real h = 2 * sqrt(eps); |
| 103 |
h = detail::make_xph_representable(x, h); |
| 104 |
|
| 105 |
Real yh = f(x + h); |
| 106 |
Real y0 = f(x); |
| 107 |
Real diff = yh - y0; |
| 108 |
if (error) |
| 109 |
{ |
| 110 |
Real ym = f(x - h); |
| 111 |
Real ypph = abs(yh - 2 * y0 + ym) / h; |
| 112 |
// h*|f''(x)|*0.5 + (|f(x+h)+|f(x)|)*eps/h |
| 113 |
*error = ypph / 2 + (abs(yh) + abs(y0))*eps / h; |
| 114 |
} |
| 115 |
return diff / h; |
| 116 |
} |
| 117 |
|
| 118 |
template<class F, class Real> |
| 119 |
Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<2>&) |
| 120 |
{ |
| 121 |
using std::sqrt; |
| 122 |
using std::pow; |
| 123 |
using std::abs; |
| 124 |
using std::numeric_limits; |
| 125 |
|
| 126 |
const Real eps = (numeric_limits<Real>::epsilon)(); |
| 127 |
// Error bound ~eps^2/3 |
| 128 |
// See the previous discussion to understand determination of h and the error bound. |
| 129 |
// Series[(f[x+h] - f[x-h])/(2*h), {h, 0, 4}] |
| 130 |
Real h = pow(3 * eps, static_cast<Real>(1) / static_cast<Real>(3)); |
| 131 |
h = detail::make_xph_representable(x, h); |
| 132 |
|
| 133 |
Real yh = f(x + h); |
| 134 |
Real ymh = f(x - h); |
| 135 |
Real diff = yh - ymh; |
| 136 |
if (error) |
| 137 |
{ |
| 138 |
Real yth = f(x + 2 * h); |
| 139 |
Real ymth = f(x - 2 * h); |
| 140 |
*error = eps * (abs(yh) + abs(ymh)) / (2 * h) + abs((yth - ymth) / 2 - diff) / (6 * h); |
| 141 |
} |
| 142 |
|
| 143 |
return diff / (2 * h); |
| 144 |
} |
| 145 |
|
| 146 |
template<class F, class Real> |
| 147 |
Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<4>&) |
| 148 |
{ |
| 149 |
using std::sqrt; |
| 150 |
using std::pow; |
| 151 |
using std::abs; |
| 152 |
using std::numeric_limits; |
| 153 |
|
| 154 |
const Real eps = (numeric_limits<Real>::epsilon)(); |
| 155 |
// Error bound ~eps^4/5 |
| 156 |
Real h = pow(Real(11.25)*eps, static_cast<Real>(1) / static_cast<Real>(5)); |
| 157 |
h = detail::make_xph_representable(x, h); |
| 158 |
Real ymth = f(x - 2 * h); |
| 159 |
Real yth = f(x + 2 * h); |
| 160 |
Real yh = f(x + h); |
| 161 |
Real ymh = f(x - h); |
| 162 |
Real y2 = ymth - yth; |
| 163 |
Real y1 = yh - ymh; |
| 164 |
if (error) |
| 165 |
{ |
| 166 |
// Mathematica code to extract the remainder: |
| 167 |
// Series[(f[x-2*h]+ 8*f[x+h] - 8*f[x-h] - f[x+2*h])/(12*h), {h, 0, 7}] |
| 168 |
Real y_three_h = f(x + 3 * h); |
| 169 |
Real y_m_three_h = f(x - 3 * h); |
| 170 |
// Error from fifth derivative: |
| 171 |
*error = abs((y_three_h - y_m_three_h) / 2 + 2 * (ymth - yth) + 5 * (yh - ymh) / 2) / (30 * h); |
| 172 |
// Error from function evaluation: |
| 173 |
*error += eps * (abs(yth) + abs(ymth) + 8 * (abs(ymh) + abs(yh))) / (12 * h); |
| 174 |
} |
| 175 |
return (y2 + 8 * y1) / (12 * h); |
| 176 |
} |
| 177 |
|
| 178 |
template<class F, class Real> |
| 179 |
Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<6>&) |
| 180 |
{ |
| 181 |
using std::sqrt; |
| 182 |
using std::pow; |
| 183 |
using std::abs; |
| 184 |
using std::numeric_limits; |
| 185 |
|
| 186 |
const Real eps = (numeric_limits<Real>::epsilon)(); |
| 187 |
// Error bound ~eps^6/7 |
| 188 |
// Error: h^6f^(7)(x)/140 + 5|f(x)|eps/h |
| 189 |
Real h = pow(eps / 168, static_cast<Real>(1) / static_cast<Real>(7)); |
| 190 |
h = detail::make_xph_representable(x, h); |
| 191 |
|
| 192 |
Real yh = f(x + h); |
| 193 |
Real ymh = f(x - h); |
| 194 |
Real y1 = yh - ymh; |
| 195 |
Real y2 = f(x - 2 * h) - f(x + 2 * h); |
| 196 |
Real y3 = f(x + 3 * h) - f(x - 3 * h); |
| 197 |
|
| 198 |
if (error) |
| 199 |
{ |
| 200 |
// Mathematica code to generate fd scheme for 7th derivative: |
| 201 |
// Sum[(-1)^i*Binomial[7, i]*(f[x+(3-i)*h] + f[x+(4-i)*h])/2, {i, 0, 7}] |
| 202 |
// Mathematica to demonstrate that this is a finite difference formula for 7th derivative: |
| 203 |
// Series[(f[x+4*h]-f[x-4*h] + 6*(f[x-3*h] - f[x+3*h]) + 14*(f[x-h] - f[x+h] + f[x+2*h] - f[x-2*h]))/2, {h, 0, 15}] |
| 204 |
Real y7 = (f(x + 4 * h) - f(x - 4 * h) - 6 * y3 - 14 * y1 - 14 * y2) / 2; |
| 205 |
*error = abs(y7) / (140 * h) + 5 * (abs(yh) + abs(ymh))*eps / h; |
| 206 |
} |
| 207 |
return (y3 + 9 * y2 + 45 * y1) / (60 * h); |
| 208 |
} |
| 209 |
|
| 210 |
template<class F, class Real> |
| 211 |
Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<8>&) |
| 212 |
{ |
| 213 |
using std::sqrt; |
| 214 |
using std::pow; |
| 215 |
using std::abs; |
| 216 |
using std::numeric_limits; |
| 217 |
|
| 218 |
const Real eps = (numeric_limits<Real>::epsilon)(); |
| 219 |
// Error bound ~eps^8/9. |
| 220 |
// In double precision, we only expect to lose two digits of precision while using this formula, at the cost of 8 function evaluations. |
| 221 |
// Error: h^8|f^(9)(x)|/630 + 7|f(x)|eps/h assuming 7 unstabilized additions. |
| 222 |
// Mathematica code to get the error: |
| 223 |
// Series[(f[x+h]-f[x-h])*(4/5) + (1/5)*(f[x-2*h] - f[x+2*h]) + (4/105)*(f[x+3*h] - f[x-3*h]) + (1/280)*(f[x-4*h] - f[x+4*h]), {h, 0, 9}] |
| 224 |
// If we used Kahan summation, we could get the max error down to h^8|f^(9)(x)|/630 + |f(x)|eps/h. |
| 225 |
Real h = pow(Real(551.25)*eps, static_cast<Real>(1) / static_cast<Real>(9)); |
| 226 |
h = detail::make_xph_representable(x, h); |
| 227 |
|
| 228 |
Real yh = f(x + h); |
| 229 |
Real ymh = f(x - h); |
| 230 |
Real y1 = yh - ymh; |
| 231 |
Real y2 = f(x - 2 * h) - f(x + 2 * h); |
| 232 |
Real y3 = f(x + 3 * h) - f(x - 3 * h); |
| 233 |
Real y4 = f(x - 4 * h) - f(x + 4 * h); |
| 234 |
|
| 235 |
Real tmp1 = 3 * y4 / 8 + 4 * y3; |
| 236 |
Real tmp2 = 21 * y2 + 84 * y1; |
| 237 |
|
| 238 |
if (error) |
| 239 |
{ |
| 240 |
// Mathematica code to generate fd scheme for 7th derivative: |
| 241 |
// Sum[(-1)^i*Binomial[9, i]*(f[x+(4-i)*h] + f[x+(5-i)*h])/2, {i, 0, 9}] |
| 242 |
// Mathematica to demonstrate that this is a finite difference formula for 7th derivative: |
| 243 |
// Series[(f[x+5*h]-f[x- 5*h])/2 + 4*(f[x-4*h] - f[x+4*h]) + 27*(f[x+3*h] - f[x-3*h])/2 + 24*(f[x-2*h] - f[x+2*h]) + 21*(f[x+h] - f[x-h]), {h, 0, 15}] |
| 244 |
Real f9 = (f(x + 5 * h) - f(x - 5 * h)) / 2 + 4 * y4 + 27 * y3 / 2 + 24 * y2 + 21 * y1; |
| 245 |
*error = abs(f9) / (630 * h) + 7 * (abs(yh) + abs(ymh))*eps / h; |
| 246 |
} |
| 247 |
return (tmp1 + tmp2) / (105 * h); |
| 248 |
} |
| 249 |
|
| 250 |
template<class F, class Real, class tag> |
| 251 |
Real finite_difference_derivative(const F, Real, Real*, const tag&) |
| 252 |
{ |
| 253 |
// Always fails, but condition is template-arg-dependent so only evaluated if we get instantiated. |
| 254 |
static_assert(sizeof(Real) == 0, "Finite difference not implemented for this order: try 1, 2, 4, 6 or 8"); |
| 255 |
} |
| 256 |
|
| 257 |
} |
| 258 |
|
| 259 |
template<class F, class Real, size_t order=6> |
| 260 |
inline Real finite_difference_derivative(const F f, Real x, Real* error = nullptr) |
| 261 |
{ |
| 262 |
return detail::finite_difference_derivative(f, x, error, detail::fd_tag<order>()); |
| 263 |
} |
| 264 |
|
| 265 |
}}} // namespaces |
| 266 |
#endif |