--- /dev/null
+// Copyright Nick Thompson, 2020
+// 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)
+
+#include <iostream>
+#include <unordered_map>
+#include <string>
+#include <future>
+#include <thread>
+#include <fstream>
+#include <boost/hana/for_each.hpp>
+#include <boost/hana/ext/std/integer_sequence.hpp>
+#include <boost/math/special_functions/daubechies_scaling.hpp>
+#include <boost/math/special_functions/detail/daubechies_scaling_integer_grid.hpp>
+#include <boost/math/interpolators/cubic_hermite.hpp>
+#include <boost/math/interpolators/quintic_hermite.hpp>
+#include <boost/math/interpolators/quintic_hermite.hpp>
+#include <boost/math/interpolators/septic_hermite.hpp>
+#include <boost/math/interpolators/cardinal_quadratic_b_spline.hpp>
+#include <boost/math/interpolators/cardinal_cubic_b_spline.hpp>
+#include <boost/math/interpolators/cardinal_quintic_b_spline.hpp>
+#include <boost/math/interpolators/whittaker_shannon.hpp>
+#include <boost/math/interpolators/cardinal_trigonometric.hpp>
+#include <boost/math/special_functions/next.hpp>
+#include <boost/math/interpolators/makima.hpp>
+#include <boost/math/interpolators/pchip.hpp>
+#include <boost/multiprecision/float128.hpp>
+#include <boost/core/demangle.hpp>
+
+using boost::multiprecision::float128;
+
+
+template<typename Real, typename PreciseReal, int p>
+void choose_refinement()
+{
+ std::cout << "Choosing refinement for " << boost::core::demangle(typeid(Real).name()) << " precision Daubechies scaling function with " << p << " vanishing moments.\n";
+ using std::abs;
+ int rmax = 22;
+ auto phi_dense = boost::math::daubechies_scaling_dyadic_grid<PreciseReal, p, 0>(rmax);
+ Real dx_dense = (2*p-1)/static_cast<Real>(phi_dense.size()-1);
+
+ for (int r = 2; r <= 18; ++r)
+ {
+ Real dx = Real(1)/ (1 << r);
+ std::cout << "\tdx = 1/" << (1/dx) << " = 1/2^" << r << " = " << dx << "\n";
+ auto phi = boost::math::daubechies_scaling<Real, p>(r);
+ Real max_flt_distance = 0;
+ Real sup = 0;
+ Real rel_sup = 0;
+ Real worst_flt_abscissa = 0;
+ Real worst_flt_value = 0;
+ Real worst_flt_computed = 0;
+
+ Real worst_rel_abscissa = 0;
+ Real worst_rel_value = 0;
+ Real worst_rel_computed = 0;
+
+ Real worst_abs_abscissa = 0;
+ Real worst_abs_computed = 0;
+ Real worst_abs_expected = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real t = i*dx_dense;
+ Real computed = phi(t);
+ Real expected = Real(phi_dense[i]);
+ Real abs_diff = abs(computed - expected);
+ Real rel_diff = abs_diff/abs(expected);
+ Real flt_distance = abs(boost::math::float_distance(computed, expected));
+ if (flt_distance > max_flt_distance)
+ {
+ max_flt_distance = flt_distance;
+ worst_flt_abscissa = t;
+ worst_flt_value = expected;
+ worst_flt_computed = computed;
+ }
+ if (expected != 0 && rel_diff > rel_sup)
+ {
+ rel_sup = rel_diff;
+ worst_rel_abscissa = t;
+ worst_rel_value = expected;
+ worst_rel_computed = computed;
+
+ }
+ if (abs_diff > sup)
+ {
+ sup = abs_diff;
+ worst_abs_abscissa = t;
+ worst_abs_computed = computed;
+ worst_abs_expected = expected;
+ }
+ }
+ std::cout << "\t\tFloat distance at r = " << r << " is " << max_flt_distance << ", sup distance = " << sup << ", max relative error = " << rel_sup << "\n";
+ std::cout << "\t\tWorst flt abscissa = " << worst_flt_abscissa << ", worst expected value = " << worst_flt_value << ", computed = " << worst_flt_computed << "\n";
+ std::cout << "\t\tWorst rel abscissa = " << worst_rel_abscissa << ", worst expected value = " << worst_rel_value << ", computed = " << worst_rel_computed << "\n";
+ std::cout << "\t\tWorst abs abscissa = " << worst_abs_abscissa << ", worst expected value = " << worst_abs_computed << ", worst abs value (expected) = " << worst_abs_expected << "\n";
+ }
+ std::cout << "\n\n\n";
+}
+
+template<typename Real, typename PreciseReal, int p>
+void find_best_interpolator()
+{
+ std::string filename = "daubechies_" + std::to_string(p) + "_scaling_convergence.csv";
+ std::ofstream fs{filename};
+ static_assert(sizeof(PreciseReal) >= sizeof(Real), "sizeof(PreciseReal) >= sizeof(Real) is required.");
+ using std::abs;
+ int rmax = 18;
+ std::cout << "Computing phi_dense_precise\n";
+ auto phi_dense_precise = boost::math::daubechies_scaling_dyadic_grid<PreciseReal, p, 0>(rmax);
+ std::vector<Real> phi_dense(phi_dense_precise.size());
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ phi_dense[i] = static_cast<Real>(phi_dense_precise[i]);
+ }
+ phi_dense_precise.resize(0);
+ std::cout << "Done\n";
+
+ Real dx_dense = (2*p-1)/static_cast<Real>(phi_dense.size()-1);
+ fs << std::setprecision(std::numeric_limits<Real>::digits10 + 3);
+ fs << std::fixed;
+ fs << "r, matched_holder, linear, quadratic_b_spline, cubic_b_spline, quintic_b_spline, cubic_hermite, pchip, makima, fo_taylor";
+ if (p==2)
+ {
+ fs << "\n";
+ }
+ else
+ {
+ fs << ", quintic_hermite, second_order_taylor";
+ if (p > 3)
+ {
+ fs << ", third_order_taylor, septic_hermite\n";
+ }
+ else
+ {
+ fs << "\n";
+ }
+ }
+ for (int r = 2; r < 13; ++r)
+ {
+ fs << r << ", ";
+ std::map<Real, std::string> m;
+ auto phi = boost::math::daubechies_scaling_dyadic_grid<Real, p, 0>(r);
+ auto phi_prime = boost::math::daubechies_scaling_dyadic_grid<Real, p, 1>(r);
+
+ std::vector<Real> x(phi.size());
+ Real dx = (2*p-1)/static_cast<Real>(x.size()-1);
+ std::cout << "dx = 1/" << (1 << r) << " = " << dx << "\n";
+ for (size_t i = 0; i < x.size(); ++i)
+ {
+ x[i] = i*dx;
+ }
+
+ {
+ auto phi_copy = phi;
+ auto phi_prime_copy = phi_prime;
+ auto mh = boost::math::detail::matched_holder(std::move(phi_copy), std::move(phi_prime_copy), r, Real(0));
+ Real sup = 0;
+ // call to matched_holder is unchecked, so only go to phi_dense.size() -1.
+ for (size_t i = 0; i < phi_dense.size() - 1; ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - mh(x));
+ if (diff > sup)
+ {
+ sup = diff;
+ }
+ }
+ m.insert({sup, "matched_holder"});
+ fs << sup << ", ";
+ }
+
+
+ {
+ auto linear = [&phi, &dx, &r](Real x)->Real {
+ if (x <= 0 || x >= 2*p-1)
+ {
+ return Real(0);
+ }
+ using std::floor;
+
+ Real y = (1<<r)*x;
+ Real k = floor(y);
+
+ size_t kk = static_cast<size_t>(k);
+
+ Real t = y - k;
+ return (1-t)*phi[kk] + t*phi[kk+1];
+ };
+
+ Real linear_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - linear(x));
+ if (diff > linear_sup)
+ {
+ linear_sup = diff;
+ }
+ }
+ m.insert({linear_sup, "linear interpolation"});
+ fs << linear_sup << ", ";
+ }
+
+
+ {
+ auto qbs = boost::math::interpolators::cardinal_quadratic_b_spline(phi.data(), phi.size(), Real(0), dx, phi_prime.front(), phi_prime.back());
+ Real qbs_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - qbs(x));
+ if (diff > qbs_sup) {
+ qbs_sup = diff;
+ }
+ }
+ m.insert({qbs_sup, "quadratic_b_spline"});
+ fs << qbs_sup << ", ";
+ }
+
+ {
+ auto cbs = boost::math::interpolators::cardinal_cubic_b_spline(phi.data(), phi.size(), Real(0), dx, phi_prime.front(), phi_prime.back());
+ Real cbs_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - cbs(x));
+ if (diff > cbs_sup)
+ {
+ cbs_sup = diff;
+ }
+ }
+ m.insert({cbs_sup, "cubic_b_spline"});
+ fs << cbs_sup << ", ";
+ }
+
+ {
+ auto qbs = boost::math::interpolators::cardinal_quintic_b_spline(phi.data(), phi.size(), Real(0), dx, {0,0}, {0,0});
+ Real qbs_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - qbs(x));
+ if (diff > qbs_sup)
+ {
+ qbs_sup = diff;
+ }
+ }
+ m.insert({qbs_sup, "quintic_b_spline"});
+ fs << qbs_sup << ", ";
+ }
+
+ {
+ auto phi_copy = phi;
+ auto phi_prime_copy = phi_prime;
+ auto ch = boost::math::interpolators::cardinal_cubic_hermite(std::move(phi_copy), std::move(phi_prime_copy), Real(0), dx);
+ Real chs_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - ch(x));
+ if (diff > chs_sup)
+ {
+ chs_sup = diff;
+ }
+ }
+ m.insert({chs_sup, "cubic_hermite_spline"});
+ fs << chs_sup << ", ";
+ }
+
+ {
+ auto phi_copy = phi;
+ auto x_copy = x;
+ auto phi_prime_copy = phi_prime;
+ auto pc = boost::math::interpolators::pchip(std::move(x_copy), std::move(phi_copy));
+ Real pchip_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - pc(x));
+ if (diff > pchip_sup)
+ {
+ pchip_sup = diff;
+ }
+ }
+ m.insert({pchip_sup, "pchip"});
+ fs << pchip_sup << ", ";
+ }
+
+ {
+ auto phi_copy = phi;
+ auto x_copy = x;
+ auto pc = boost::math::interpolators::makima(std::move(x_copy), std::move(phi_copy));
+ Real makima_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i) {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - pc(x));
+ if (diff > makima_sup)
+ {
+ makima_sup = diff;
+ }
+ }
+ m.insert({makima_sup, "makima"});
+ fs << makima_sup << ", ";
+ }
+
+ // Whittaker-Shannon interpolation has linear complexity; test over all points and it's quadratic.
+ // I ran this a couple times and found it's not competitive; so comment out for now.
+ /*{
+ auto phi_copy = phi;
+ auto ws = boost::math::interpolators::whittaker_shannon(std::move(phi_copy), Real(0), dx);
+ Real sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i) {
+ Real x = i*dx_dense;
+ using std::abs;
+ Real diff = abs(phi_dense[i] - ws(x));
+ if (diff > sup) {
+ sup = diff;
+ }
+ }
+
+ m.insert({sup, "whittaker_shannon"});
+ }
+
+ // Again, linear complexity of evaluation => quadratic complexity of exhaustive checking.
+ {
+ auto trig = boost::math::interpolators::cardinal_trigonometric(phi, Real(0), dx);
+ Real sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i) {
+ Real x = i*dx_dense;
+ using std::abs;
+ Real diff = abs(phi_dense[i] - trig(x));
+ if (diff > sup) {
+ sup = diff;
+ }
+ }
+ m.insert({sup, "trig"});
+ }*/
+
+ {
+ auto fotaylor = [&phi, &phi_prime, &r](Real x)->Real
+ {
+ if (x <= 0 || x >= 2*p-1)
+ {
+ return 0;
+ }
+ using std::floor;
+
+ Real y = (1<<r)*x;
+ Real k = floor(y);
+
+ size_t kk = static_cast<size_t>(k);
+ if (y - k < k + 1 - y)
+ {
+ Real eps = (y-k)/(1<<r);
+ return phi[kk] + eps*phi_prime[kk];
+ }
+ else {
+ Real eps = (y-k-1)/(1<<r);
+ return phi[kk+1] + eps*phi_prime[kk+1];
+ }
+ };
+ Real fo_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - fotaylor(x));
+ if (diff > fo_sup)
+ {
+ fo_sup = diff;
+ }
+ }
+ m.insert({fo_sup, "First-order Taylor"});
+ if (p==2)
+ {
+ fs << fo_sup << "\n";
+ }
+ else
+ {
+ fs << fo_sup << ", ";
+ }
+ }
+
+ if constexpr (p > 2) {
+ auto phi_dbl_prime = boost::math::daubechies_scaling_dyadic_grid<Real, p, 2>(r);
+
+ {
+ auto phi_copy = phi;
+ auto phi_prime_copy = phi_prime;
+ auto phi_dbl_prime_copy = phi_dbl_prime;
+ auto qh = boost::math::interpolators::cardinal_quintic_hermite(std::move(phi_copy), std::move(phi_prime_copy), std::move(phi_dbl_prime_copy), Real(0), dx);
+ Real qh_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - qh(x));
+ if (diff > qh_sup)
+ {
+ qh_sup = diff;
+ }
+ }
+ m.insert({qh_sup, "quintic_hermite_spline"});
+ fs << qh_sup << ", ";
+ }
+
+ {
+ auto sotaylor = [&phi, &phi_prime, &phi_dbl_prime, &r](Real x)->Real {
+ if (x <= 0 || x >= 2*p-1)
+ {
+ return 0;
+ }
+ using std::floor;
+
+ Real y = (1<<r)*x;
+ Real k = floor(y);
+
+ size_t kk = static_cast<size_t>(k);
+ if (y - k < k + 1 - y)
+ {
+ Real eps = (y-k)/(1<<r);
+ return phi[kk] + eps*phi_prime[kk] + eps*eps*phi_dbl_prime[kk]/2;
+ }
+ else {
+ Real eps = (y-k-1)/(1<<r);
+ return phi[kk+1] + eps*phi_prime[kk+1] + eps*eps*phi_dbl_prime[kk+1]/2;
+ }
+ };
+ Real so_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - sotaylor(x));
+ if (diff > so_sup)
+ {
+ so_sup = diff;
+ }
+ }
+ m.insert({so_sup, "Second-order Taylor"});
+ if (p > 3)
+ {
+ fs << so_sup << ", ";
+ }
+ else
+ {
+ fs << so_sup << "\n";
+ }
+
+ }
+ }
+
+ if constexpr (p > 3)
+ {
+ auto phi_dbl_prime = boost::math::daubechies_scaling_dyadic_grid<Real, p, 2>(r);
+ auto phi_triple_prime = boost::math::daubechies_scaling_dyadic_grid<Real, p, 3>(r);
+
+ {
+ auto totaylor = [&phi, &phi_prime, &phi_dbl_prime, &phi_triple_prime, &r](Real x)->Real {
+ if (x <= 0 || x >= 2*p-1) {
+ return 0;
+ }
+ using std::floor;
+
+ Real y = (1<<r)*x;
+ Real k = floor(y);
+
+ size_t kk = static_cast<size_t>(k);
+ if (y - k < k + 1 - y)
+ {
+ Real eps = (y-k)/(1<<r);
+ return phi[kk] + eps*phi_prime[kk] + eps*eps*phi_dbl_prime[kk]/2 + eps*eps*eps*phi_triple_prime[kk]/6;
+ }
+ else {
+ Real eps = (y-k-1)/(1<<r);
+ return phi[kk+1] + eps*phi_prime[kk+1] + eps*eps*phi_dbl_prime[kk+1]/2 + eps*eps*eps*phi_triple_prime[kk]/6;
+ }
+ };
+ Real to_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - totaylor(x));
+ if (diff > to_sup)
+ {
+ to_sup = diff;
+ }
+ }
+
+ m.insert({to_sup, "Third-order Taylor"});
+ fs << to_sup << ", ";
+ }
+
+ {
+ auto phi_copy = phi;
+ auto phi_prime_copy = phi_prime;
+ auto phi_dbl_prime_copy = phi_dbl_prime;
+ auto phi_triple_prime_copy = phi_triple_prime;
+ auto sh = boost::math::interpolators::cardinal_septic_hermite(std::move(phi_copy), std::move(phi_prime_copy), std::move(phi_dbl_prime_copy), std::move(phi_triple_prime_copy), Real(0), dx);
+ Real septic_sup = 0;
+ for (size_t i = 0; i < phi_dense.size(); ++i)
+ {
+ Real x = i*dx_dense;
+ Real diff = abs(phi_dense[i] - sh(x));
+ if (diff > septic_sup)
+ {
+ septic_sup = diff;
+ }
+ }
+ m.insert({septic_sup, "septic_hermite_spline"});
+ fs << septic_sup << "\n";
+ }
+
+
+ }
+ std::string best = "none";
+ Real best_sup = 1000000000;
+ std::cout << std::setprecision(std::numeric_limits<Real>::digits10 + 3) << std::fixed;
+ for (auto & e : m)
+ {
+ std::cout << "\t" << e.first << " is error of " << e.second << "\n";
+ if (e.first < best_sup)
+ {
+ best = e.second;
+ best_sup = e.first;
+ }
+ }
+ std::cout << "\tThe best method for p = " << p << " is the " << best << "\n";
+ }
+}
+
+int main()
+{
+ //boost::hana::for_each(std::make_index_sequence<4>(), [&](auto i){ choose_refinement<double, float128, i+16>(); });
+ boost::hana::for_each(std::make_index_sequence<12>(), [&](auto i){ find_best_interpolator<double, float128, i+2>(); });
+}