feat():initial version

install/boost_1_75_0/include/boost/math/differentiation/autodiff_cpp11.hpp

@@ -0,0 +1,383 @@
+//           Copyright Matthew Pulver 2018 - 2019.
+// Distributed under the Boost Software License, Version 1.0.
+//      (See accompanying file LICENSE_1_0.txt or copy at
+//           https://www.boost.org/LICENSE_1_0.txt)
+
+// Contributors:
+//  * Kedar R. Bhat - C++11 compatibility.
+
+// Notes:
+//  * Any changes to this file should always be downstream from autodiff.cpp.
+//    C++17 is a higher-level language and is easier to maintain. For example, a number of functions which are
+//    lucidly read in autodiff.cpp are forced to be split into multiple structs/functions in this file for
+//    C++11.
+//  * Use of typename RootType and SizeType is a hack to prevent Visual Studio 2015 from compiling functions
+//    that are never called, that would otherwise produce compiler errors. Also forces functions to be inline.
+
+#ifndef BOOST_MATH_DIFFERENTIATION_AUTODIFF_HPP
+#error \
+    "Do not #include this file directly. This should only be #included by autodiff.hpp for C++11 compatibility."
+#endif
+
+#include <boost/mp11/integer_sequence.hpp>
+
+namespace boost {
+namespace math {
+namespace differentiation {
+inline namespace autodiff_v1 {
+namespace detail {
+
+template <typename RealType, size_t Order>
+fvar<RealType, Order>::fvar(root_type const& ca, bool const is_variable) {
+  fvar_cpp11(is_fvar<RealType>{}, ca, is_variable);
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+void fvar<RealType, Order>::fvar_cpp11(std::true_type, RootType const& ca, bool const is_variable) {
+  v.front() = RealType(ca, is_variable);
+  if (0 < Order)
+    std::fill(v.begin() + 1, v.end(), static_cast<RealType>(0));
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+void fvar<RealType, Order>::fvar_cpp11(std::false_type, RootType const& ca, bool const is_variable) {
+  v.front() = ca;
+  if (0 < Order) {
+    v[1] = static_cast<root_type>(static_cast<int>(is_variable));
+    if (1 < Order)
+      std::fill(v.begin() + 2, v.end(), static_cast<RealType>(0));
+  }
+}
+
+template <typename RealType, size_t Order>
+template <typename... Orders>
+get_type_at<RealType, sizeof...(Orders)> fvar<RealType, Order>::at_cpp11(std::true_type,
+                                                                         size_t order,
+                                                                         Orders...) const {
+  return v.at(order);
+}
+
+template <typename RealType, size_t Order>
+template <typename... Orders>
+get_type_at<RealType, sizeof...(Orders)> fvar<RealType, Order>::at_cpp11(std::false_type,
+                                                                         size_t order,
+                                                                         Orders... orders) const {
+  return v.at(order).at(orders...);
+}
+
+// Can throw "std::out_of_range: array::at: __n (which is 7) >= _Nm (which is 7)"
+template <typename RealType, size_t Order>
+template <typename... Orders>
+get_type_at<RealType, sizeof...(Orders)> fvar<RealType, Order>::at(size_t order, Orders... orders) const {
+  return at_cpp11(std::integral_constant<bool, sizeof...(orders) == 0>{}, order, orders...);
+}
+
+template <typename T, typename... Ts>
+constexpr T product(Ts...) {
+  return static_cast<T>(1);
+}
+
+template <typename T, typename... Ts>
+constexpr T product(T factor, Ts... factors) {
+  return factor * product<T>(factors...);
+}
+
+// Can throw "std::out_of_range: array::at: __n (which is 7) >= _Nm (which is 7)"
+template <typename RealType, size_t Order>
+template <typename... Orders>
+get_type_at<fvar<RealType, Order>, sizeof...(Orders)> fvar<RealType, Order>::derivative(
+    Orders... orders) const {
+  static_assert(sizeof...(Orders) <= depth,
+                "Number of parameters to derivative(...) cannot exceed fvar::depth.");
+  return at(static_cast<size_t>(orders)...) *
+         product(boost::math::factorial<root_type>(static_cast<unsigned>(orders))...);
+}
+
+template <typename RootType, typename Func>
+class Curry {
+  Func const& f_;
+  size_t const i_;
+
+ public:
+  template <typename SizeType>  // typename SizeType to force inline constructor.
+  Curry(Func const& f, SizeType i) : f_(f), i_(static_cast<std::size_t>(i)) {}
+  template <typename... Indices>
+  RootType operator()(Indices... indices) const {
+    using unsigned_t = typename std::make_unsigned<typename std::common_type<Indices>::type...>::type;
+    return f_(i_, static_cast<unsigned_t>(indices)...);
+  }
+};
+
+template <typename RealType, size_t Order>
+template <typename Func, typename Fvar, typename... Fvars>
+promote<fvar<RealType, Order>, Fvar, Fvars...> fvar<RealType, Order>::apply_coefficients(
+    size_t const order,
+    Func const& f,
+    Fvar const& cr,
+    Fvars&&... fvars) const {
+  fvar<RealType, Order> const epsilon = fvar<RealType, Order>(*this).set_root(0);
+  size_t i = order < order_sum ? order : order_sum;
+  using return_type = promote<fvar<RealType, Order>, Fvar, Fvars...>;
+  return_type accumulator = cr.apply_coefficients(
+      order - i, Curry<typename return_type::root_type, Func>(f, i), std::forward<Fvars>(fvars)...);
+  while (i--)
+    (accumulator *= epsilon) += cr.apply_coefficients(
+        order - i, Curry<typename return_type::root_type, Func>(f, i), std::forward<Fvars>(fvars)...);
+  return accumulator;
+}
+
+template <typename RealType, size_t Order>
+template <typename Func, typename Fvar, typename... Fvars>
+promote<fvar<RealType, Order>, Fvar, Fvars...> fvar<RealType, Order>::apply_coefficients_nonhorner(
+    size_t const order,
+    Func const& f,
+    Fvar const& cr,
+    Fvars&&... fvars) const {
+  fvar<RealType, Order> const epsilon = fvar<RealType, Order>(*this).set_root(0);
+  fvar<RealType, Order> epsilon_i = fvar<RealType, Order>(1);  // epsilon to the power of i
+  using return_type = promote<fvar<RealType, Order>, Fvar, Fvars...>;
+  return_type accumulator = cr.apply_coefficients_nonhorner(
+      order, Curry<typename return_type::root_type, Func>(f, 0), std::forward<Fvars>(fvars)...);
+  size_t const i_max = order < order_sum ? order : order_sum;
+  for (size_t i = 1; i <= i_max; ++i) {
+    epsilon_i = epsilon_i.epsilon_multiply(i - 1, 0, epsilon, 1, 0);
+    accumulator += epsilon_i.epsilon_multiply(
+        i,
+        0,
+        cr.apply_coefficients_nonhorner(
+            order - i, Curry<typename return_type::root_type, Func>(f, i), std::forward<Fvars>(fvars)...),
+        0,
+        0);
+  }
+  return accumulator;
+}
+
+template <typename RealType, size_t Order>
+template <typename Func, typename Fvar, typename... Fvars>
+promote<fvar<RealType, Order>, Fvar, Fvars...> fvar<RealType, Order>::apply_derivatives(
+    size_t const order,
+    Func const& f,
+    Fvar const& cr,
+    Fvars&&... fvars) const {
+  fvar<RealType, Order> const epsilon = fvar<RealType, Order>(*this).set_root(0);
+  size_t i = order < order_sum ? order : order_sum;
+  using return_type = promote<fvar<RealType, Order>, Fvar, Fvars...>;
+  return_type accumulator =
+      cr.apply_derivatives(
+          order - i, Curry<typename return_type::root_type, Func>(f, i), std::forward<Fvars>(fvars)...) /
+      factorial<root_type>(static_cast<unsigned>(i));
+  while (i--)
+    (accumulator *= epsilon) +=
+        cr.apply_derivatives(
+            order - i, Curry<typename return_type::root_type, Func>(f, i), std::forward<Fvars>(fvars)...) /
+        factorial<root_type>(static_cast<unsigned>(i));
+  return accumulator;
+}
+
+template <typename RealType, size_t Order>
+template <typename Func, typename Fvar, typename... Fvars>
+promote<fvar<RealType, Order>, Fvar, Fvars...> fvar<RealType, Order>::apply_derivatives_nonhorner(
+    size_t const order,
+    Func const& f,
+    Fvar const& cr,
+    Fvars&&... fvars) const {
+  fvar<RealType, Order> const epsilon = fvar<RealType, Order>(*this).set_root(0);
+  fvar<RealType, Order> epsilon_i = fvar<RealType, Order>(1);  // epsilon to the power of i
+  using return_type = promote<fvar<RealType, Order>, Fvar, Fvars...>;
+  return_type accumulator = cr.apply_derivatives_nonhorner(
+      order, Curry<typename return_type::root_type, Func>(f, 0), std::forward<Fvars>(fvars)...);
+  size_t const i_max = order < order_sum ? order : order_sum;
+  for (size_t i = 1; i <= i_max; ++i) {
+    epsilon_i = epsilon_i.epsilon_multiply(i - 1, 0, epsilon, 1, 0);
+    accumulator += epsilon_i.epsilon_multiply(
+        i,
+        0,
+        cr.apply_derivatives_nonhorner(
+            order - i, Curry<typename return_type::root_type, Func>(f, i), std::forward<Fvars>(fvars)...) /
+            factorial<root_type>(static_cast<unsigned>(i)),
+        0,
+        0);
+  }
+  return accumulator;
+}
+
+template <typename RealType, size_t Order>
+template <typename SizeType>
+fvar<RealType, Order> fvar<RealType, Order>::epsilon_multiply_cpp11(std::true_type,
+                                                                    SizeType z0,
+                                                                    size_t isum0,
+                                                                    fvar<RealType, Order> const& cr,
+                                                                    size_t z1,
+                                                                    size_t isum1) const {
+  size_t const m0 = order_sum + isum0 < Order + z0 ? Order + z0 - (order_sum + isum0) : 0;
+  size_t const m1 = order_sum + isum1 < Order + z1 ? Order + z1 - (order_sum + isum1) : 0;
+  size_t const i_max = m0 + m1 < Order ? Order - (m0 + m1) : 0;
+  fvar<RealType, Order> retval = fvar<RealType, Order>();
+  for (size_t i = 0, j = Order; i <= i_max; ++i, --j)
+    retval.v[j] = epsilon_inner_product(z0, isum0, m0, cr, z1, isum1, m1, j);
+  return retval;
+}
+
+template <typename RealType, size_t Order>
+template <typename SizeType>
+fvar<RealType, Order> fvar<RealType, Order>::epsilon_multiply_cpp11(std::false_type,
+                                                                    SizeType z0,
+                                                                    size_t isum0,
+                                                                    fvar<RealType, Order> const& cr,
+                                                                    size_t z1,
+                                                                    size_t isum1) const {
+  using ssize_t = typename std::make_signed<std::size_t>::type;
+  RealType const zero(0);
+  size_t const m0 = order_sum + isum0 < Order + z0 ? Order + z0 - (order_sum + isum0) : 0;
+  size_t const m1 = order_sum + isum1 < Order + z1 ? Order + z1 - (order_sum + isum1) : 0;
+  size_t const i_max = m0 + m1 < Order ? Order - (m0 + m1) : 0;
+  fvar<RealType, Order> retval = fvar<RealType, Order>();
+  for (size_t i = 0, j = Order; i <= i_max; ++i, --j)
+    retval.v[j] = std::inner_product(
+        v.cbegin() + ssize_t(m0), v.cend() - ssize_t(i + m1), cr.v.crbegin() + ssize_t(i + m0), zero);
+  return retval;
+}
+
+template <typename RealType, size_t Order>
+fvar<RealType, Order> fvar<RealType, Order>::epsilon_multiply(size_t z0,
+                                                              size_t isum0,
+                                                              fvar<RealType, Order> const& cr,
+                                                              size_t z1,
+                                                              size_t isum1) const {
+  return epsilon_multiply_cpp11(is_fvar<RealType>{}, z0, isum0, cr, z1, isum1);
+}
+
+template <typename RealType, size_t Order>
+template <typename SizeType>
+fvar<RealType, Order> fvar<RealType, Order>::epsilon_multiply_cpp11(std::true_type,
+                                                                    SizeType z0,
+                                                                    size_t isum0,
+                                                                    root_type const& ca) const {
+  fvar<RealType, Order> retval(*this);
+  size_t const m0 = order_sum + isum0 < Order + z0 ? Order + z0 - (order_sum + isum0) : 0;
+  for (size_t i = m0; i <= Order; ++i)
+    retval.v[i] = retval.v[i].epsilon_multiply(z0, isum0 + i, ca);
+  return retval;
+}
+
+template <typename RealType, size_t Order>
+template <typename SizeType>
+fvar<RealType, Order> fvar<RealType, Order>::epsilon_multiply_cpp11(std::false_type,
+                                                                    SizeType z0,
+                                                                    size_t isum0,
+                                                                    root_type const& ca) const {
+  fvar<RealType, Order> retval(*this);
+  size_t const m0 = order_sum + isum0 < Order + z0 ? Order + z0 - (order_sum + isum0) : 0;
+  for (size_t i = m0; i <= Order; ++i)
+    if (retval.v[i] != static_cast<RealType>(0))
+      retval.v[i] *= ca;
+  return retval;
+}
+
+template <typename RealType, size_t Order>
+fvar<RealType, Order> fvar<RealType, Order>::epsilon_multiply(size_t z0,
+                                                              size_t isum0,
+                                                              root_type const& ca) const {
+  return epsilon_multiply_cpp11(is_fvar<RealType>{}, z0, isum0, ca);
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+fvar<RealType, Order>& fvar<RealType, Order>::multiply_assign_by_root_type_cpp11(std::true_type,
+                                                                                 bool is_root,
+                                                                                 RootType const& ca) {
+  auto itr = v.begin();
+  itr->multiply_assign_by_root_type(is_root, ca);
+  for (++itr; itr != v.end(); ++itr)
+    itr->multiply_assign_by_root_type(false, ca);
+  return *this;
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+fvar<RealType, Order>& fvar<RealType, Order>::multiply_assign_by_root_type_cpp11(std::false_type,
+                                                                                 bool is_root,
+                                                                                 RootType const& ca) {
+  auto itr = v.begin();
+  if (is_root || *itr != 0)
+    *itr *= ca;  // Skip multiplication of 0 by ca=inf to avoid nan, except when is_root.
+  for (++itr; itr != v.end(); ++itr)
+    if (*itr != 0)
+      *itr *= ca;
+  return *this;
+}
+
+template <typename RealType, size_t Order>
+fvar<RealType, Order>& fvar<RealType, Order>::multiply_assign_by_root_type(bool is_root,
+                                                                           root_type const& ca) {
+  return multiply_assign_by_root_type_cpp11(is_fvar<RealType>{}, is_root, ca);
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+fvar<RealType, Order>& fvar<RealType, Order>::negate_cpp11(std::true_type, RootType const&) {
+  std::for_each(v.begin(), v.end(), [](RealType& r) { r.negate(); });
+  return *this;
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+fvar<RealType, Order>& fvar<RealType, Order>::negate_cpp11(std::false_type, RootType const&) {
+  std::for_each(v.begin(), v.end(), [](RealType& a) { a = -a; });
+  return *this;
+}
+
+template <typename RealType, size_t Order>
+fvar<RealType, Order>& fvar<RealType, Order>::negate() {
+  return negate_cpp11(is_fvar<RealType>{}, static_cast<root_type>(*this));
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+fvar<RealType, Order>& fvar<RealType, Order>::set_root_cpp11(std::true_type, RootType const& root) {
+  v.front().set_root(root);
+  return *this;
+}
+
+template <typename RealType, size_t Order>
+template <typename RootType>
+fvar<RealType, Order>& fvar<RealType, Order>::set_root_cpp11(std::false_type, RootType const& root) {
+  v.front() = root;
+  return *this;
+}
+
+template <typename RealType, size_t Order>
+fvar<RealType, Order>& fvar<RealType, Order>::set_root(root_type const& root) {
+  return set_root_cpp11(is_fvar<RealType>{}, root);
+}
+
+template <typename RealType, size_t Order, size_t... Is>
+auto make_fvar_for_tuple(mp11::index_sequence<Is...>, RealType const& ca)
+    -> decltype(make_fvar<RealType, zero<Is>::value..., Order>(ca)) {
+  return make_fvar<RealType, zero<Is>::value..., Order>(ca);
+}
+
+template <typename RealType, size_t... Orders, size_t... Is, typename... RealTypes>
+auto make_ftuple_impl(mp11::index_sequence<Is...>, RealTypes const&... ca)
+    -> decltype(std::make_tuple(make_fvar_for_tuple<RealType, Orders>(mp11::make_index_sequence<Is>{},
+                                                                      ca)...)) {
+  return std::make_tuple(make_fvar_for_tuple<RealType, Orders>(mp11::make_index_sequence<Is>{}, ca)...);
+}
+
+}  // namespace detail
+
+template <typename RealType, size_t... Orders, typename... RealTypes>
+auto make_ftuple(RealTypes const&... ca)
+    -> decltype(detail::make_ftuple_impl<RealType, Orders...>(mp11::index_sequence_for<RealTypes...>{},
+                                                              ca...)) {
+  static_assert(sizeof...(Orders) == sizeof...(RealTypes),
+                "Number of Orders must match number of function parameters.");
+  return detail::make_ftuple_impl<RealType, Orders...>(mp11::index_sequence_for<RealTypes...>{}, ca...);
+}
+
+}  // namespace autodiff_v1
+}  // namespace differentiation
+}  // namespace math
+}  // namespace boost

install/boost_1_75_0/include/boost/math/differentiation/finite_difference.hpp

@@ -0,0 +1,266 @@
+//  (C) Copyright Nick Thompson 2018.
+//  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_DIFFERENTIATION_FINITE_DIFFERENCE_HPP
+#define BOOST_MATH_DIFFERENTIATION_FINITE_DIFFERENCE_HPP
+
+/*
+ * Performs numerical differentiation by finite-differences.
+ *
+ * All numerical differentiation using finite-differences are ill-conditioned, and these routines are no exception.
+ * A simple argument demonstrates that the error is unbounded as h->0.
+ * Take the one sides finite difference formula f'(x) = (f(x+h)-f(x))/h.
+ * The evaluation of f induces an error as well as the error from the finite-difference approximation, giving
+ * |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.
+ * It is reasonable to choose h in a way that minimizes the maximum error bound g(h).
+ * 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
+ * sqrt(2eps(|f(x+h) +f(x)||f''(x)|)).
+ * 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.
+ *
+ *
+ * For more details on this method of analysis, see
+ *
+ * http://www.uio.no/studier/emner/matnat/math/MAT-INF1100/h08/kompendiet/diffint.pdf
+ * http://web.archive.org/web/20150420195907/http://www.uio.no/studier/emner/matnat/math/MAT-INF1100/h08/kompendiet/diffint.pdf
+ *
+ *
+ * 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.
+ * From this we can see that full precision can be recovered in the limit k->infinity.
+ *
+ * References:
+ *
+ * 1) Fornberg, Bengt. "Generation of finite difference formulas on arbitrarily spaced grids." Mathematics of computation 51.184 (1988): 699-706.
+ *
+ *
+ * The second algorithm, the complex step derivative, is not ill-conditioned.
+ * However, it requires that your function can be evaluated at complex arguments.
+ * The idea is that f(x+ih) = f(x) +ihf'(x) - h^2f''(x) + ... so f'(x) \approx Im[f(x+ih)]/h.
+ * No subtractive cancellation occurs. The error is ~ eps|f'(x)| + eps^2|f'''(x)|/6; hard to beat that.
+ *
+ * References:
+ *
+ * 1) Squire, William, and George Trapp. "Using complex variables to estimate derivatives of real functions." Siam Review 40.1 (1998): 110-112.
+ */
+
+#include <complex>
+#include <boost/math/special_functions/next.hpp>
+
+namespace boost{ namespace math{ namespace differentiation {
+
+namespace detail {
+    template<class Real>
+    Real make_xph_representable(Real x, Real h)
+    {
+        using std::numeric_limits;
+        // Redefine h so that x + h is representable. Not using this trick leads to large error.
+        // The compiler flag -ffast-math evaporates these operations . . .
+        Real temp = x + h;
+        h = temp - x;
+        // Handle the case x + h == x:
+        if (h == 0)
+        {
+            h = boost::math::nextafter(x, (numeric_limits<Real>::max)()) - x;
+        }
+        return h;
+    }
+}
+
+template<class F, class Real>
+Real complex_step_derivative(const F f, Real x)
+{
+    // Is it really this easy? Yes.
+    // Note that some authors recommend taking the stepsize h to be smaller than epsilon(), some recommending use of the min().
+    // This idea was tested over a few billion test cases and found the make the error *much* worse.
+    // Even 2eps and eps/2 made the error worse, which was surprising.
+    using std::complex;
+    using std::numeric_limits;
+    constexpr const Real step = (numeric_limits<Real>::epsilon)();
+    constexpr const Real inv_step = 1/(numeric_limits<Real>::epsilon)();
+    return f(complex<Real>(x, step)).imag()*inv_step;
+}
+
+namespace detail {
+
+   template <unsigned>
+   struct fd_tag {};
+
+   template<class F, class Real>
+   Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<1>&)
+   {
+      using std::sqrt;
+      using std::pow;
+      using std::abs;
+      using std::numeric_limits;
+
+      const Real eps = (numeric_limits<Real>::epsilon)();
+      // Error bound ~eps^1/2
+      // Note that this estimate of h differs from the best estimate by a factor of sqrt((|f(x)| + |f(x+h)|)/|f''(x)|).
+      // Since this factor is invariant under the scaling f -> kf, then we are somewhat justified in approximating it by 1.
+      // This approximation will get better as we move to higher orders of accuracy.
+      Real h = 2 * sqrt(eps);
+      h = detail::make_xph_representable(x, h);
+
+      Real yh = f(x + h);
+      Real y0 = f(x);
+      Real diff = yh - y0;
+      if (error)
+      {
+         Real ym = f(x - h);
+         Real ypph = abs(yh - 2 * y0 + ym) / h;
+         // h*|f''(x)|*0.5 + (|f(x+h)+|f(x)|)*eps/h
+         *error = ypph / 2 + (abs(yh) + abs(y0))*eps / h;
+      }
+      return diff / h;
+   }
+
+   template<class F, class Real>
+   Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<2>&)
+   {
+      using std::sqrt;
+      using std::pow;
+      using std::abs;
+      using std::numeric_limits;
+
+      const Real eps = (numeric_limits<Real>::epsilon)();
+      // Error bound ~eps^2/3
+      // See the previous discussion to understand determination of h and the error bound.
+      // Series[(f[x+h] - f[x-h])/(2*h), {h, 0, 4}]
+      Real h = pow(3 * eps, static_cast<Real>(1) / static_cast<Real>(3));
+      h = detail::make_xph_representable(x, h);
+
+      Real yh = f(x + h);
+      Real ymh = f(x - h);
+      Real diff = yh - ymh;
+      if (error)
+      {
+         Real yth = f(x + 2 * h);
+         Real ymth = f(x - 2 * h);
+         *error = eps * (abs(yh) + abs(ymh)) / (2 * h) + abs((yth - ymth) / 2 - diff) / (6 * h);
+      }
+
+      return diff / (2 * h);
+   }
+
+   template<class F, class Real>
+   Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<4>&)
+   {
+      using std::sqrt;
+      using std::pow;
+      using std::abs;
+      using std::numeric_limits;
+
+      const Real eps = (numeric_limits<Real>::epsilon)();
+      // Error bound ~eps^4/5
+      Real h = pow(11.25*eps, (Real)1 / (Real)5);
+      h = detail::make_xph_representable(x, h);
+      Real ymth = f(x - 2 * h);
+      Real yth = f(x + 2 * h);
+      Real yh = f(x + h);
+      Real ymh = f(x - h);
+      Real y2 = ymth - yth;
+      Real y1 = yh - ymh;
+      if (error)
+      {
+         // Mathematica code to extract the remainder:
+         // Series[(f[x-2*h]+ 8*f[x+h] - 8*f[x-h] - f[x+2*h])/(12*h), {h, 0, 7}]
+         Real y_three_h = f(x + 3 * h);
+         Real y_m_three_h = f(x - 3 * h);
+         // Error from fifth derivative:
+         *error = abs((y_three_h - y_m_three_h) / 2 + 2 * (ymth - yth) + 5 * (yh - ymh) / 2) / (30 * h);
+         // Error from function evaluation:
+         *error += eps * (abs(yth) + abs(ymth) + 8 * (abs(ymh) + abs(yh))) / (12 * h);
+      }
+      return (y2 + 8 * y1) / (12 * h);
+   }
+
+   template<class F, class Real>
+   Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<6>&)
+   {
+      using std::sqrt;
+      using std::pow;
+      using std::abs;
+      using std::numeric_limits;
+
+      const Real eps = (numeric_limits<Real>::epsilon)();
+      // Error bound ~eps^6/7
+      // Error: h^6f^(7)(x)/140 + 5|f(x)|eps/h
+      Real h = pow(eps / 168, (Real)1 / (Real)7);
+      h = detail::make_xph_representable(x, h);
+
+      Real yh = f(x + h);
+      Real ymh = f(x - h);
+      Real y1 = yh - ymh;
+      Real y2 = f(x - 2 * h) - f(x + 2 * h);
+      Real y3 = f(x + 3 * h) - f(x - 3 * h);
+
+      if (error)
+      {
+         // Mathematica code to generate fd scheme for 7th derivative:
+         // Sum[(-1)^i*Binomial[7, i]*(f[x+(3-i)*h] + f[x+(4-i)*h])/2, {i, 0, 7}]
+         // Mathematica to demonstrate that this is a finite difference formula for 7th derivative:
+         // 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}]
+         Real y7 = (f(x + 4 * h) - f(x - 4 * h) - 6 * y3 - 14 * y1 - 14 * y2) / 2;
+         *error = abs(y7) / (140 * h) + 5 * (abs(yh) + abs(ymh))*eps / h;
+      }
+      return (y3 + 9 * y2 + 45 * y1) / (60 * h);
+   }
+
+   template<class F, class Real>
+   Real finite_difference_derivative(const F f, Real x, Real* error, const fd_tag<8>&)
+   {
+      using std::sqrt;
+      using std::pow;
+      using std::abs;
+      using std::numeric_limits;
+
+      const Real eps = (numeric_limits<Real>::epsilon)();
+      // Error bound ~eps^8/9.
+      // In double precision, we only expect to lose two digits of precision while using this formula, at the cost of 8 function evaluations.
+      // Error: h^8|f^(9)(x)|/630 + 7|f(x)|eps/h assuming 7 unstabilized additions.
+      // Mathematica code to get the error:
+      // 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}]
+      // If we used Kahan summation, we could get the max error down to h^8|f^(9)(x)|/630 + |f(x)|eps/h.
+      Real h = pow(551.25*eps, (Real)1 / (Real)9);
+      h = detail::make_xph_representable(x, h);
+
+      Real yh = f(x + h);
+      Real ymh = f(x - h);
+      Real y1 = yh - ymh;
+      Real y2 = f(x - 2 * h) - f(x + 2 * h);
+      Real y3 = f(x + 3 * h) - f(x - 3 * h);
+      Real y4 = f(x - 4 * h) - f(x + 4 * h);
+
+      Real tmp1 = 3 * y4 / 8 + 4 * y3;
+      Real tmp2 = 21 * y2 + 84 * y1;
+
+      if (error)
+      {
+         // Mathematica code to generate fd scheme for 7th derivative:
+         // Sum[(-1)^i*Binomial[9, i]*(f[x+(4-i)*h] + f[x+(5-i)*h])/2, {i, 0, 9}]
+         // Mathematica to demonstrate that this is a finite difference formula for 7th derivative:
+         // 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}]
+         Real f9 = (f(x + 5 * h) - f(x - 5 * h)) / 2 + 4 * y4 + 27 * y3 / 2 + 24 * y2 + 21 * y1;
+         *error = abs(f9) / (630 * h) + 7 * (abs(yh) + abs(ymh))*eps / h;
+      }
+      return (tmp1 + tmp2) / (105 * h);
+   }
+
+   template<class F, class Real, class tag>
+   Real finite_difference_derivative(const F, Real, Real*, const tag&)
+   {
+      // Always fails, but condition is template-arg-dependent so only evaluated if we get instantiated.
+      BOOST_STATIC_ASSERT_MSG(sizeof(Real) == 0, "Finite difference not implemented for this order: try 1, 2, 4, 6 or 8");
+   }
+
+}
+
+template<class F, class Real, size_t order=6>
+inline Real finite_difference_derivative(const F f, Real x, Real* error = nullptr)
+{
+   return detail::finite_difference_derivative(f, x, error, detail::fd_tag<order>());
+}
+
+}}}  // namespaces
+#endif

install/boost_1_75_0/include/boost/math/differentiation/lanczos_smoothing.hpp

@@ -0,0 +1,580 @@
+//  (C) Copyright Nick Thompson 2019.
+//  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_DIFFERENTIATION_LANCZOS_SMOOTHING_HPP
+#define BOOST_MATH_DIFFERENTIATION_LANCZOS_SMOOTHING_HPP
+#include <cmath> // for std::abs
+#include <limits> // to nan initialize
+#include <vector>
+#include <string>
+#include <stdexcept>
+#include <boost/assert.hpp>
+
+namespace boost::math::differentiation {
+
+namespace detail {
+template <typename Real>
+class discrete_legendre {
+  public:
+    explicit discrete_legendre(std::size_t n, Real x) : m_n{n}, m_r{2}, m_x{x},
+                                                        m_qrm2{1}, m_qrm1{x},
+                                                        m_qrm2p{0}, m_qrm1p{1},
+                                                        m_qrm2pp{0}, m_qrm1pp{0}
+    {
+        using std::abs;
+        BOOST_ASSERT_MSG(abs(m_x) <= 1, "Three term recurrence is stable only for |x| <=1.");
+        // The integer n indexes a family of discrete Legendre polynomials indexed by k <= 2*n
+    }
+
+    Real norm_sq(int r) const
+    {
+        Real prod = Real(2) / Real(2 * r + 1);
+        for (int k = -r; k <= r; ++k) {
+            prod *= Real(2 * m_n + 1 + k) / Real(2 * m_n);
+        }
+        return prod;
+    }
+
+    Real next()
+    {
+        Real N = 2 * m_n + 1;
+        Real num = (m_r - 1) * (N * N - (m_r - 1) * (m_r - 1)) * m_qrm2;
+        Real tmp = (2 * m_r - 1) * m_x * m_qrm1 - num / Real(4 * m_n * m_n);
+        m_qrm2 = m_qrm1;
+        m_qrm1 = tmp / m_r;
+        ++m_r;
+        return m_qrm1;
+    }
+
+    Real next_prime()
+    {
+        Real N = 2 * m_n + 1;
+        Real s = (m_r - 1) * (N * N - (m_r - 1) * (m_r - 1)) / Real(4 * m_n * m_n);
+        Real tmp1 = ((2 * m_r - 1) * m_x * m_qrm1 - s * m_qrm2) / m_r;
+        Real tmp2 = ((2 * m_r - 1) * (m_qrm1 + m_x * m_qrm1p) - s * m_qrm2p) / m_r;
+        m_qrm2 = m_qrm1;
+        m_qrm1 = tmp1;
+        m_qrm2p = m_qrm1p;
+        m_qrm1p = tmp2;
+        ++m_r;
+        return m_qrm1p;
+    }
+
+    Real next_dbl_prime()
+    {
+        Real N = 2*m_n + 1;
+        Real trm1 = 2*m_r - 1;
+        Real s = (m_r - 1) * (N * N - (m_r - 1) * (m_r - 1)) / Real(4 * m_n * m_n);
+        Real rqrpp = 2*trm1*m_qrm1p + trm1*m_x*m_qrm1pp - s*m_qrm2pp;
+        Real tmp1 = ((2 * m_r - 1) * m_x * m_qrm1 - s * m_qrm2) / m_r;
+        Real tmp2 = ((2 * m_r - 1) * (m_qrm1 + m_x * m_qrm1p) - s * m_qrm2p) / m_r;
+        m_qrm2 = m_qrm1;
+        m_qrm1 = tmp1;
+        m_qrm2p = m_qrm1p;
+        m_qrm1p = tmp2;
+        m_qrm2pp = m_qrm1pp;
+        m_qrm1pp = rqrpp/m_r;
+        ++m_r;
+        return m_qrm1pp;
+    }
+
+    Real operator()(Real x, std::size_t k)
+    {
+        BOOST_ASSERT_MSG(k <= 2 * m_n, "r <= 2n is required.");
+        if (k == 0)
+        {
+            return 1;
+        }
+        if (k == 1)
+        {
+            return x;
+        }
+        Real qrm2 = 1;
+        Real qrm1 = x;
+        Real N = 2 * m_n + 1;
+        for (std::size_t r = 2; r <= k; ++r) {
+            Real num = (r - 1) * (N * N - (r - 1) * (r - 1)) * qrm2;
+            Real tmp = (2 * r - 1) * x * qrm1 - num / Real(4 * m_n * m_n);
+            qrm2 = qrm1;
+            qrm1 = tmp / r;
+        }
+        return qrm1;
+    }
+
+    Real prime(Real x, std::size_t k) {
+        BOOST_ASSERT_MSG(k <= 2 * m_n, "r <= 2n is required.");
+        if (k == 0) {
+            return 0;
+        }
+        if (k == 1) {
+            return 1;
+        }
+        Real qrm2 = 1;
+        Real qrm1 = x;
+        Real qrm2p = 0;
+        Real qrm1p = 1;
+        Real N = 2 * m_n + 1;
+        for (std::size_t r = 2; r <= k; ++r) {
+            Real s =
+                (r - 1) * (N * N - (r - 1) * (r - 1)) / Real(4 * m_n * m_n);
+            Real tmp1 = ((2 * r - 1) * x * qrm1 - s * qrm2) / r;
+            Real tmp2 = ((2 * r - 1) * (qrm1 + x * qrm1p) - s * qrm2p) / r;
+            qrm2 = qrm1;
+            qrm1 = tmp1;
+            qrm2p = qrm1p;
+            qrm1p = tmp2;
+        }
+        return qrm1p;
+    }
+
+  private:
+    std::size_t m_n;
+    std::size_t m_r;
+    Real m_x;
+    Real m_qrm2;
+    Real m_qrm1;
+    Real m_qrm2p;
+    Real m_qrm1p;
+    Real m_qrm2pp;
+    Real m_qrm1pp;
+};
+
+template <class Real>
+std::vector<Real> interior_velocity_filter(std::size_t n, std::size_t p) {
+    auto dlp = discrete_legendre<Real>(n, 0);
+    std::vector<Real> coeffs(p+1);
+    coeffs[1] = 1/dlp.norm_sq(1);
+    for (std::size_t l = 3; l < p + 1; l += 2)
+    {
+        dlp.next_prime();
+        coeffs[l] = dlp.next_prime()/ dlp.norm_sq(l);
+    }
+
+    // We could make the filter length n, as f[0] = 0,
+    // but that'd make the indexing awkward when applying the filter.
+    std::vector<Real> f(n + 1);
+    // This value should never be read, but this is the correct value *if it is read*.
+    // Hmm, should it be a nan then? I'm not gonna agonize.
+    f[0] = 0;
+    for (std::size_t j = 1; j < f.size(); ++j)
+    {
+        Real arg = Real(j) / Real(n);
+        dlp = discrete_legendre<Real>(n, arg);
+        f[j] = coeffs[1]*arg;
+        for (std::size_t l = 3; l <= p; l += 2)
+        {
+            dlp.next();
+            f[j] += coeffs[l]*dlp.next();
+        }
+        f[j] /= (n * n);
+    }
+    return f;
+}
+
+template <class Real>
+std::vector<Real> boundary_velocity_filter(std::size_t n, std::size_t p, int64_t s)
+{
+    std::vector<Real> coeffs(p+1, std::numeric_limits<Real>::quiet_NaN());
+    Real sn = Real(s) / Real(n);
+    auto dlp = discrete_legendre<Real>(n, sn);
+    coeffs[0] = 0;
+    coeffs[1] = 1/dlp.norm_sq(1);
+    for (std::size_t l = 2; l < p + 1; ++l)
+    {
+        // Calculation of the norms is common to all filters,
+        // so it seems like an obvious optimization target.
+        // I tried this: The spent in computing the norms time is not negligible,
+        // but still a small fraction of the total compute time.
+        // Hence I'm not refactoring out these norm calculations.
+        coeffs[l] = dlp.next_prime()/ dlp.norm_sq(l);
+    }
+
+    std::vector<Real> f(2*n + 1);
+    for (std::size_t k = 0; k < f.size(); ++k)
+    {
+        Real j = Real(k) - Real(n);
+        Real arg = j/Real(n);
+        dlp = discrete_legendre<Real>(n, arg);
+        f[k] = coeffs[1]*arg;
+        for (std::size_t l = 2; l <= p; ++l)
+        {
+            f[k] += coeffs[l]*dlp.next();
+        }
+        f[k] /= (n * n);
+    }
+    return f;
+}
+
+template <class Real>
+std::vector<Real> acceleration_filter(std::size_t n, std::size_t p, int64_t s)
+{
+    BOOST_ASSERT_MSG(p <= 2*n, "Approximation order must be <= 2*n");
+    BOOST_ASSERT_MSG(p > 2, "Approximation order must be > 2");
+
+    std::vector<Real> coeffs(p+1, std::numeric_limits<Real>::quiet_NaN());
+    Real sn = Real(s) / Real(n);
+    auto dlp = discrete_legendre<Real>(n, sn);
+    coeffs[0] = 0;
+    coeffs[1] = 0;
+    for (std::size_t l = 2; l < p + 1; ++l)
+    {
+        coeffs[l] = dlp.next_dbl_prime()/ dlp.norm_sq(l);
+    }
+
+    std::vector<Real> f(2*n + 1, 0);
+    for (std::size_t k = 0; k < f.size(); ++k)
+    {
+        Real j = Real(k) - Real(n);
+        Real arg = j/Real(n);
+        dlp = discrete_legendre<Real>(n, arg);
+        for (std::size_t l = 2; l <= p; ++l)
+        {
+            f[k] += coeffs[l]*dlp.next();
+        }
+        f[k] /= (n * n * n);
+    }
+    return f;
+}
+
+
+} // namespace detail
+
+template <typename Real, std::size_t order = 1>
+class discrete_lanczos_derivative {
+public:
+    discrete_lanczos_derivative(Real const & spacing,
+                                std::size_t n = 18,
+                                std::size_t approximation_order = 3)
+        : m_dt{spacing}
+    {
+        static_assert(!std::is_integral_v<Real>,
+                      "Spacing must be a floating point type.");
+        BOOST_ASSERT_MSG(spacing > 0,
+                         "Spacing between samples must be > 0.");
+
+        if constexpr (order == 1)
+        {
+            BOOST_ASSERT_MSG(approximation_order <= 2 * n,
+                             "The approximation order must be <= 2n");
+            BOOST_ASSERT_MSG(approximation_order >= 2,
+                             "The approximation order must be >= 2");
+
+            if constexpr (std::is_same_v<Real, float> || std::is_same_v<Real, double>)
+            {
+                auto interior = detail::interior_velocity_filter<long double>(n, approximation_order);
+                m_f.resize(interior.size());
+                for (std::size_t j = 0; j < interior.size(); ++j)
+                {
+                    m_f[j] = static_cast<Real>(interior[j])/m_dt;
+                }
+            }
+            else
+            {
+                m_f = detail::interior_velocity_filter<Real>(n, approximation_order);
+                for (auto & x : m_f)
+                {
+                    x /= m_dt;
+                }
+            }
+
+            m_boundary_filters.resize(n);
+            // This for loop is a natural candidate for parallelization.
+            // But does it matter? Probably not.
+            for (std::size_t i = 0; i < n; ++i)
+            {
+                if constexpr (std::is_same_v<Real, float> || std::is_same_v<Real, double>)
+                {
+                    int64_t s = static_cast<int64_t>(i) - static_cast<int64_t>(n);
+                    auto bf = detail::boundary_velocity_filter<long double>(n, approximation_order, s);
+                    m_boundary_filters[i].resize(bf.size());
+                    for (std::size_t j = 0; j < bf.size(); ++j)
+                    {
+                        m_boundary_filters[i][j] = static_cast<Real>(bf[j])/m_dt;
+                    }
+                }
+                else
+                {
+                    int64_t s = static_cast<int64_t>(i) - static_cast<int64_t>(n);
+                    m_boundary_filters[i] = detail::boundary_velocity_filter<Real>(n, approximation_order, s);
+                    for (auto & bf : m_boundary_filters[i])
+                    {
+                        bf /= m_dt;
+                    }
+                }
+            }
+        }
+        else if constexpr (order == 2)
+        {
+            // High precision isn't warranted for small p; only for large p.
+            // (The computation appears stable for large n.)
+            // But given that the filters are reusable for many vectors,
+            // it's better to do a high precision computation and then cast back,
+            // since the resulting cost is a factor of 2, and the cost of the filters not working is hours of debugging.
+            if constexpr (std::is_same_v<Real, double> || std::is_same_v<Real, float>)
+            {
+                auto f = detail::acceleration_filter<long double>(n, approximation_order, 0);
+                m_f.resize(n+1);
+                for (std::size_t i = 0; i < m_f.size(); ++i)
+                {
+                    m_f[i] = static_cast<Real>(f[i+n])/(m_dt*m_dt);
+                }
+                m_boundary_filters.resize(n);
+                for (std::size_t i = 0; i < n; ++i)
+                {
+                    int64_t s = static_cast<int64_t>(i) - static_cast<int64_t>(n);
+                    auto bf = detail::acceleration_filter<long double>(n, approximation_order, s);
+                    m_boundary_filters[i].resize(bf.size());
+                    for (std::size_t j = 0; j < bf.size(); ++j)
+                    {
+                        m_boundary_filters[i][j] = static_cast<Real>(bf[j])/(m_dt*m_dt);
+                    }
+                }
+            }
+            else
+            {
+                // Given that the purpose is denoising, for higher precision calculations,
+                // the default precision should be fine.
+                auto f = detail::acceleration_filter<Real>(n, approximation_order, 0);
+                m_f.resize(n+1);
+                for (std::size_t i = 0; i < m_f.size(); ++i)
+                {
+                    m_f[i] = f[i+n]/(m_dt*m_dt);
+                }
+                m_boundary_filters.resize(n);
+                for (std::size_t i = 0; i < n; ++i)
+                {
+                    int64_t s = static_cast<int64_t>(i) - static_cast<int64_t>(n);
+                    m_boundary_filters[i] = detail::acceleration_filter<Real>(n, approximation_order, s);
+                    for (auto & bf : m_boundary_filters[i])
+                    {
+                        bf /= (m_dt*m_dt);
+                    }
+                }
+            }
+        }
+        else
+        {
+            BOOST_ASSERT_MSG(false, "Derivatives of order 3 and higher are not implemented.");
+        }
+    }
+
+    Real get_spacing() const
+    {
+        return m_dt;
+    }
+
+    template<class RandomAccessContainer>
+    Real operator()(RandomAccessContainer const & v, std::size_t i) const
+    {
+        static_assert(std::is_same_v<typename RandomAccessContainer::value_type, Real>,
+                      "The type of the values in the vector provided does not match the type in the filters.");
+
+        BOOST_ASSERT_MSG(std::size(v) >= m_boundary_filters[0].size(),
+            "Vector must be at least as long as the filter length");
+
+        if constexpr (order==1)
+        {
+            if (i >= m_f.size() - 1 && i <= std::size(v) - m_f.size())
+            {
+                // The filter has length >= 1:
+                Real dvdt = m_f[1] * (v[i + 1] - v[i - 1]);
+                for (std::size_t j = 2; j < m_f.size(); ++j)
+                {
+                    dvdt += m_f[j] * (v[i + j] - v[i - j]);
+                }
+                return dvdt;
+            }
+
+            // m_f.size() = N+1
+            if (i < m_f.size() - 1)
+            {
+                auto &bf = m_boundary_filters[i];
+                Real dvdt = bf[0]*v[0];
+                for (std::size_t j = 1; j < bf.size(); ++j)
+                {
+                    dvdt += bf[j] * v[j];
+                }
+                return dvdt;
+            }
+
+            if (i > std::size(v) - m_f.size() && i < std::size(v))
+            {
+                int k = std::size(v) - 1 - i;
+                auto &bf = m_boundary_filters[k];
+                Real dvdt = bf[0]*v[std::size(v)-1];
+                for (std::size_t j = 1; j < bf.size(); ++j)
+                {
+                    dvdt += bf[j] * v[std::size(v) - 1 - j];
+                }
+                return -dvdt;
+            }
+        }
+        else if constexpr (order==2)
+        {
+            if (i >= m_f.size() - 1 && i <= std::size(v) - m_f.size())
+            {
+                Real d2vdt2 = m_f[0]*v[i];
+                for (std::size_t j = 1; j < m_f.size(); ++j)
+                {
+                    d2vdt2 += m_f[j] * (v[i + j] + v[i - j]);
+                }
+                return d2vdt2;
+            }
+
+            // m_f.size() = N+1
+            if (i < m_f.size() - 1)
+            {
+                auto &bf = m_boundary_filters[i];
+                Real d2vdt2 = bf[0]*v[0];
+                for (std::size_t j = 1; j < bf.size(); ++j)
+                {
+                    d2vdt2 += bf[j] * v[j];
+                }
+                return d2vdt2;
+            }
+
+            if (i > std::size(v) - m_f.size() && i < std::size(v))
+            {
+                int k = std::size(v) - 1 - i;
+                auto &bf = m_boundary_filters[k];
+                Real d2vdt2 = bf[0] * v[std::size(v) - 1];
+                for (std::size_t j = 1; j < bf.size(); ++j)
+                {
+                    d2vdt2 += bf[j] * v[std::size(v) - 1 - j];
+                }
+                return d2vdt2;
+            }
+        }
+
+        // OOB access:
+        std::string msg = "Out of bounds access in Lanczos derivative.";
+        msg += "Input vector has length " + std::to_string(std::size(v)) + ", but user requested access at index " + std::to_string(i) + ".";
+        throw std::out_of_range(msg);
+        return std::numeric_limits<Real>::quiet_NaN();
+    }
+
+    template<class RandomAccessContainer>
+    void operator()(RandomAccessContainer const & v, RandomAccessContainer & w) const
+    {
+        static_assert(std::is_same_v<typename RandomAccessContainer::value_type, Real>,
+                      "The type of the values in the vector provided does not match the type in the filters.");
+        if (&w[0] == &v[0])
+        {
+            throw std::logic_error("This transform cannot be performed in-place.");
+        }
+
+        if (std::size(v) < m_boundary_filters[0].size())
+        {
+            std::string msg = "The input vector must be at least as long as the filter length. ";
+            msg += "The input vector has length = " + std::to_string(std::size(v)) + ", the filter has length " + std::to_string(m_boundary_filters[0].size());
+            throw std::length_error(msg);
+        }
+
+        if (std::size(w) < std::size(v))
+        {
+            std::string msg = "The output vector (containing the derivative) must be at least as long as the input vector.";
+            msg += "The output vector has length = " + std::to_string(std::size(w)) + ", the input vector has length " + std::to_string(std::size(v));
+            throw std::length_error(msg);
+        }
+
+        if constexpr (order==1)
+        {
+            for (std::size_t i = 0; i < m_f.size() - 1; ++i)
+            {
+                auto &bf = m_boundary_filters[i];
+                Real dvdt = bf[0] * v[0];
+                for (std::size_t j = 1; j < bf.size(); ++j)
+                {
+                    dvdt += bf[j] * v[j];
+                }
+                w[i] = dvdt;
+            }
+
+            for(std::size_t i = m_f.size() - 1; i <= std::size(v) - m_f.size(); ++i)
+            {
+                Real dvdt = m_f[1] * (v[i + 1] - v[i - 1]);
+                for (std::size_t j = 2; j < m_f.size(); ++j)
+                {
+                    dvdt += m_f[j] *(v[i + j] - v[i - j]);
+                }
+                w[i] = dvdt;
+            }
+
+
+            for(std::size_t i = std::size(v) - m_f.size() + 1; i < std::size(v); ++i)
+            {
+                int k = std::size(v) - 1 - i;
+                auto &f = m_boundary_filters[k];
+                Real dvdt = f[0] * v[std::size(v) - 1];;
+                for (std::size_t j = 1; j < f.size(); ++j)
+                {
+                    dvdt += f[j] * v[std::size(v) - 1 - j];
+                }
+                w[i] = -dvdt;
+            }
+        }
+        else if constexpr (order==2)
+        {
+            // m_f.size() = N+1
+            for (std::size_t i = 0; i < m_f.size() - 1; ++i)
+            {
+                auto &bf = m_boundary_filters[i];
+                Real d2vdt2 = 0;
+                for (std::size_t j = 0; j < bf.size(); ++j)
+                {
+                    d2vdt2 += bf[j] * v[j];
+                }
+                w[i] = d2vdt2;
+            }
+
+            for (std::size_t i = m_f.size() - 1; i <= std::size(v) - m_f.size(); ++i)
+            {
+                Real d2vdt2 = m_f[0]*v[i];
+                for (std::size_t j = 1; j < m_f.size(); ++j)
+                {
+                    d2vdt2 += m_f[j] * (v[i + j] + v[i - j]);
+                }
+                w[i] = d2vdt2;
+            }
+
+            for (std::size_t i = std::size(v) - m_f.size() + 1; i < std::size(v); ++i)
+            {
+                int k = std::size(v) - 1 - i;
+                auto &bf = m_boundary_filters[k];
+                Real d2vdt2 = bf[0] * v[std::size(v) - 1];
+                for (std::size_t j = 1; j < bf.size(); ++j)
+                {
+                    d2vdt2 += bf[j] * v[std::size(v) - 1 - j];
+                }
+                w[i] = d2vdt2;
+            }
+        }
+    }
+
+    template<class RandomAccessContainer>
+    RandomAccessContainer operator()(RandomAccessContainer const & v) const
+    {
+        RandomAccessContainer w(std::size(v));
+        this->operator()(v, w);
+        return w;
+    }
+
+
+    // Don't copy; too big.
+    discrete_lanczos_derivative( const discrete_lanczos_derivative & ) = delete;
+    discrete_lanczos_derivative& operator=(const discrete_lanczos_derivative&) = delete;
+
+    // Allow moves:
+    discrete_lanczos_derivative(discrete_lanczos_derivative&&) = default;
+    discrete_lanczos_derivative& operator=(discrete_lanczos_derivative&&) = default;
+
+private:
+    std::vector<Real> m_f;
+    std::vector<std::vector<Real>> m_boundary_filters;
+    Real m_dt;
+};
+
+} // namespaces
+#endif