homotopy-continuation-chann.../cpp/include/hccd/PathTracker.hpp

157 lines
4.3 KiB
C++

#pragma once
// STL includes
#include <expected>
// Library includes
#include <Eigen/Dense>
#include <spdlog/spdlog.h>
// Project includes
#include "util.hpp"
namespace hccd {
namespace detail {
template <typename T>
concept homotopy_c = requires(T) {
{ T::evaluate_H(Eigen::VectorXd()) } -> std::same_as<Eigen::VectorXd>;
{ T::evaluate_DH(Eigen::VectorXd()) } -> std::same_as<Eigen::MatrixXd>;
};
} // namespace detail
///
/// @brief Settings for PathTracker
///
struct Settings {
double euler_step_size = 0.05;
unsigned euler_max_tries = 5;
unsigned newton_max_iter = 5;
double newton_convergence_threshold = 0.01;
int sigma = 1; ///< Direction in which the path is traced
};
///
/// @brief Path tracker for the homotopy continuation method
/// @details Uses a predictor-corrector scheme to trace a path defined by a
/// homotopy.
/// @details References:
/// [1] T. Chen and T.-Y. Li, “Homotopy continuation method for solving
/// systems of nonlinear and polynomial equations,” Communications in
/// Information and Systems, vol. 15, no. 2, pp. 119-307, 2015
///
/// @tparam homotopy_c Homotopy defining the path
///
template <detail::homotopy_c Homotopy>
class PathTracker {
public:
enum Error { NewtonNotConverged };
PathTracker(Settings settings) : m_settings{settings} {
}
///
/// @brief Perform one predictor-corrector step
///
Eigen::VectorXd step(Eigen::VectorXd y) {
auto res = transparent_step(y);
if (res) {
return res.value().first;
} else {
return std::unexpected(res.error());
}
}
///
/// @brief Perform one predictor-corrector step, returning intermediate
/// results
///
std::expected<std::tuple<Eigen::VectorXd, Eigen::VectorXd, Eigen::VectorXd,
Eigen::VectorXd>,
Error>
transparent_step(Eigen::VectorXd y) {
for (int i = 0; i < m_settings.euler_max_tries; ++i) {
double step_size = m_settings.euler_step_size / (1 << i);
const auto [y_hat, y_prime] =
perform_euler_predictor_step(y, step_size);
auto res = perform_newton_corrector_step(y_hat);
if (res) return {{y, y_prime, y_hat, res.value()}};
}
return std::unexpected(Error::NewtonNotConverged);
}
private:
Settings m_settings;
std::pair<Eigen::VectorXd, Eigen::VectorXd>
perform_euler_predictor_step(Eigen::VectorXd y, double step_size) {
/// Obtain y_prime
Eigen::MatrixXd DH = Homotopy::evaluate_DH(y);
auto qr = DH.transpose().colPivHouseholderQr();
Eigen::MatrixXd Q = qr.matrixQ();
Eigen::MatrixXd R = qr.matrixR();
Eigen::VectorXd y_prime = Q.col(2);
spdlog::debug("Q: \t\t{}x{}; det={}", Q.rows(), Q.cols(),
Q.determinant());
spdlog::debug("R.topRows(2): {}x{}; det={}", R.topRows(2).rows(),
R.topRows(2).cols(), R.topRows(2).determinant());
if (sign(Q.determinant() * R.topRows(2).determinant()) !=
sign(m_settings.sigma))
y_prime = -y_prime;
/// Perform prediction
Eigen::VectorXd y_hat = y + step_size * y_prime;
return {y_hat, y_prime};
}
std::expected<Eigen::VectorXd, Error>
perform_newton_corrector_step(Eigen::VectorXd y) {
Eigen::VectorXd prev_y = y;
for (int i = 0; i < m_settings.newton_max_iter; ++i) {
/// Perform correction
Eigen::MatrixXd DH = Homotopy::evaluate_DH(y);
Eigen::MatrixXd DH_pinv =
DH.completeOrthogonalDecomposition().pseudoInverse();
y = y - DH_pinv * Homotopy::evaluate_H(y);
/// Check stopping criterion
spdlog::debug("Newton iteration {}: ||y-prev_y||={}", i,
(y - prev_y).norm());
if ((y - prev_y).norm() < m_settings.newton_convergence_threshold)
return y;
prev_y = y;
}
return std::unexpected(Error::NewtonNotConverged);
}
template <typename T>
static int sign(T val) {
return -1 * (val < T(0)) + 1 * (val >= T(0));
}
};
} // namespace hccd