示例#1
0
Status EmbeddedExplicitRungeKuttaNyströmIntegrator<Position,
                                                   higher_order,
                                                   lower_order,
                                                   stages,
                                                   first_same_as_last>::
Instance::Solve(Instant const& t_final) {
  using Displacement = typename ODE::Displacement;
  using Velocity = typename ODE::Velocity;
  using Acceleration = typename ODE::Acceleration;

  auto const& a = integrator_.a_;
  auto const& b_hat = integrator_.b_hat_;
  auto const& b_prime_hat = integrator_.b_prime_hat_;
  auto const& b = integrator_.b_;
  auto const& b_prime = integrator_.b_prime_;
  auto const& c = integrator_.c_;

  auto& current_state = this->current_state_;
  auto& append_state = this->append_state_;
  auto& adaptive_step_size = this->adaptive_step_size_;
  auto const& equation = this->equation_;

  // |current_state| gets updated as the integration progresses to allow
  // restartability.

  // State before the last, truncated step.
  std::experimental::optional<typename ODE::SystemState> final_state;

  // Argument checks.
  int const dimension = current_state.positions.size();
  Sign const integration_direction =
      Sign(adaptive_step_size.first_time_step);
  if (integration_direction.Positive()) {
    // Integrating forward.
    CHECK_LT(current_state.time.value, t_final);
  } else {
    // Integrating backward.
    CHECK_GT(current_state.time.value, t_final);
  }

  // Time step.
  Time h = adaptive_step_size.first_time_step;
  // Current time.  This is a non-const reference whose purpose is to make the
  // equations more readable.
  DoublePrecision<Instant>& t = current_state.time;

  // Position increment (high-order).
  std::vector<Displacement> Δq_hat(dimension);
  // Velocity increment (high-order).
  std::vector<Velocity> Δv_hat(dimension);
  // Current position.  This is a non-const reference whose purpose is to make
  // the equations more readable.
  std::vector<DoublePrecision<Position>>& q_hat = current_state.positions;
  // Current velocity.  This is a non-const reference whose purpose is to make
  // the equations more readable.
  std::vector<DoublePrecision<Velocity>>& v_hat = current_state.velocities;

  // Difference between the low- and high-order approximations.
  typename ODE::SystemStateError error_estimate;
  error_estimate.position_error.resize(dimension);
  error_estimate.velocity_error.resize(dimension);

  // Current Runge-Kutta-Nyström stage.
  std::vector<Position> q_stage(dimension);
  // Accelerations at each stage.
  // TODO(egg): this is a rectangular container, use something more appropriate.
  std::vector<std::vector<Acceleration>> g(stages);
  for (auto& g_stage : g) {
    g_stage.resize(dimension);
  }

  bool at_end = false;
  double tolerance_to_error_ratio;

  // The first stage of the Runge-Kutta-Nyström iteration.  In the FSAL case,
  // |first_stage == 1| after the first step, since the first RHS evaluation has
  // already occured in the previous step.  In the non-FSAL case and in the
  // first step of the FSAL case, |first_stage == 0|.
  int first_stage = 0;

  // The number of steps already performed.
  std::int64_t step_count = 0;

  // No step size control on the first step.
  goto runge_kutta_nyström_step;

  while (!at_end) {
    // Compute the next step with decreasing step sizes until the error is
    // tolerable.
    do {
      // Adapt step size.
      // TODO(egg): find out whether there's a smarter way to compute that root,
      // especially since we make the order compile-time.
      h *= adaptive_step_size.safety_factor *
               std::pow(tolerance_to_error_ratio, 1.0 / (lower_order + 1));
      // TODO(egg): should we check whether it vanishes in double precision
      // instead?
      if (t.value + (t.error + h) == t.value) {
        return Status(termination_condition::VanishingStepSize,
                      "At time " + DebugString(t.value) +
                          ", step size is effectively zero.  Singularity or "
                          "stiff system suspected.");
      }

    runge_kutta_nyström_step:
      // Termination condition.
      Time const time_to_end = (t_final - t.value) - t.error;
      at_end = integration_direction * h >= integration_direction * time_to_end;
      if (at_end) {
        // The chosen step size will overshoot.  Clip it to just reach the end,
        // and terminate if the step is accepted.
        h = time_to_end;
        final_state = current_state;
      }

      // Runge-Kutta-Nyström iteration; fills |g|.
      for (int i = first_stage; i < stages; ++i) {
        Instant const t_stage = t.value + c[i] * h;
        for (int k = 0; k < dimension; ++k) {
          Acceleration Σj_a_ij_g_jk{};
          for (int j = 0; j < i; ++j) {
            Σj_a_ij_g_jk += a[i][j] * g[j][k];
          }
          q_stage[k] = q_hat[k].value +
                           h * (c[i] * v_hat[k].value + h * Σj_a_ij_g_jk);
        }
        equation.compute_acceleration(t_stage, q_stage, g[i]);
      }

      // Increment computation and step size control.
      for (int k = 0; k < dimension; ++k) {
        Acceleration Σi_b_hat_i_g_ik{};
        Acceleration Σi_b_i_g_ik{};
        Acceleration Σi_b_prime_hat_i_g_ik{};
        Acceleration Σi_b_prime_i_g_ik{};
        // Please keep the eight assigments below aligned, they become illegible
        // otherwise.
        for (int i = 0; i < stages; ++i) {
          Σi_b_hat_i_g_ik       += b_hat[i] * g[i][k];
          Σi_b_i_g_ik           += b[i] * g[i][k];
          Σi_b_prime_hat_i_g_ik += b_prime_hat[i] * g[i][k];
          Σi_b_prime_i_g_ik     += b_prime[i] * g[i][k];
        }
        // The hat-less Δq and Δv are the low-order increments.
        Δq_hat[k]               = h * (h * (Σi_b_hat_i_g_ik) + v_hat[k].value);
        Displacement const Δq_k = h * (h * (Σi_b_i_g_ik) + v_hat[k].value);
        Δv_hat[k]               = h * Σi_b_prime_hat_i_g_ik;
        Velocity const Δv_k     = h * Σi_b_prime_i_g_ik;

        error_estimate.position_error[k] = Δq_k - Δq_hat[k];
        error_estimate.velocity_error[k] = Δv_k - Δv_hat[k];
      }
      tolerance_to_error_ratio =
          adaptive_step_size.tolerance_to_error_ratio(h, error_estimate);
    } while (tolerance_to_error_ratio < 1.0);

    if (first_same_as_last) {
      using std::swap;
      swap(g.front(), g.back());
      first_stage = 1;
    }

    // Increment the solution with the high-order approximation.
    t.Increment(h);
    for (int k = 0; k < dimension; ++k) {
      q_hat[k].Increment(Δq_hat[k]);
      v_hat[k].Increment(Δv_hat[k]);
    }
    append_state(current_state);
    ++step_count;
    if (step_count == adaptive_step_size.max_steps && !at_end) {
      return Status(termination_condition::ReachedMaximalStepCount,
                    "Reached maximum step count " +
                        std::to_string(adaptive_step_size.max_steps) +
                        " at time " + DebugString(t.value) +
                        "; requested t_final is " + DebugString(t_final) +
                        ".");
    }
  }
  // The resolution is restartable from the last non-truncated state.
  CHECK(final_state);
  current_state = *final_state;
  return Status(termination_condition::Done, "");
}
void SymplecticRungeKuttaNyströmIntegrator<Position, order, time_reversible,
                                           evaluations, composition>::Solve(
    IntegrationProblem<ODE> const& problem,
    Time const& step) const {
  using Displacement = typename ODE::Displacement;
  using Velocity = typename ODE::Velocity;
  using Acceleration = typename ODE::Acceleration;

  // Argument checks.
  CHECK_NOTNULL(problem.initial_state);
  int const dimension = problem.initial_state->positions.size();
  CHECK_EQ(dimension, problem.initial_state->velocities.size());
  CHECK_NE(Time(), step);
  Sign const integration_direction = Sign(step);
  if (integration_direction.Positive()) {
    // Integrating forward.
    CHECK_LT(problem.initial_state->time.value, problem.t_final);
  } else {
    // Integrating backward.
    CHECK_GT(problem.initial_state->time.value, problem.t_final);
  }

  typename ODE::SystemState current_state = *problem.initial_state;

  // Time step.
  Time const& h = step;
  Time const abs_h = integration_direction * h;
  // Current time.  This is a non-const reference whose purpose is to make the
  // equations more readable.
  DoublePrecision<Instant>& t = current_state.time;

  // Position increment.
  std::vector<Displacement> Δq(dimension);
  // Velocity increment.
  std::vector<Velocity> Δv(dimension);
  // Current position.  This is a non-const reference whose purpose is to make
  // the equations more readable.
  std::vector<DoublePrecision<Position>>& q = current_state.positions;
  // Current velocity.  This is a non-const reference whose purpose is to make
  // the equations more readable.
  std::vector<DoublePrecision<Velocity>>& v = current_state.velocities;

  // Current Runge-Kutta-Nyström stage.
  std::vector<Position> q_stage(dimension);
  // Accelerations at the current stage.
  std::vector<Acceleration> g(dimension);

  // The first full stage of the step, i.e. the first stage where
  // exp(bᵢ h B) exp(aᵢ h A) must be entirely computed.
  // Always 0 in the non-FSAL kBA case, always 1 in the kABA case since b₀ = 0,
  // means the first stage is only exp(a₀ h A), and 1 after the first step
  // in the kBAB case, since the last right-hand-side evaluation can be used for
  // exp(bᵢ h B).
  int first_stage = composition == kABA ? 1 : 0;

  while (abs_h <= Abs((problem.t_final - t.value) - t.error)) {
    std::fill(Δq.begin(), Δq.end(), Displacement{});
    std::fill(Δv.begin(), Δv.end(), Velocity{});

    if (first_stage == 1) {
      for (int k = 0; k < dimension; ++k) {
        if (composition == kBAB) {
          // exp(b₀ h B)
          Δv[k] += h * b_[0] * g[k];
        }
        // exp(a₀ h A)
        Δq[k] += h * a_[0] * (v[k].value + Δv[k]);
      }
    }

    for (int i = first_stage; i < stages_; ++i) {
      for (int k = 0; k < dimension; ++k) {
        q_stage[k] = q[k].value + Δq[k];
      }
      problem.equation.compute_acceleration(t.value + c_[i] * h, q_stage, &g);
      for (int k = 0; k < dimension; ++k) {
        // exp(bᵢ h B)
        Δv[k] += h * b_[i] * g[k];
        // exp(aᵢ h A)
        Δq[k] += h * a_[i] * (v[k].value + Δv[k]);
      }
    }

    if (composition == kBAB) {
      first_stage = 1;
    }

    // Increment the solution.
    t.Increment(h);
    for (int k = 0; k < dimension; ++k) {
      q[k].Increment(Δq[k]);
      v[k].Increment(Δv[k]);
    }
    problem.append_state(current_state);
  }
}