ModelTree.cc

/*
 * Copyright © 2003-2019 Dynare Team
 *
 * This file is part of Dynare.
 *
 * Dynare is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * Dynare is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with Dynare.  If not, see <http://www.gnu.org/licenses/>.
 */

#include <cstdlib>
#include <cassert>
#include <cmath>
#include <iostream>
#include <fstream>

#include "ModelTree.hh"
#include "MinimumFeedbackSet.hh"
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#include <boost/graph/adjacency_list.hpp>
#include <boost/graph/max_cardinality_matching.hpp>
#include <boost/graph/strong_components.hpp>
#include <boost/graph/topological_sort.hpp>
#pragma GCC diagnostic pop

using namespace MFS;

void
ModelTree::copyHelper(const ModelTree &m)
{
  auto f = [this](expr_t e) { return e->clone(*this); };

  // Equations
  for (const auto & it : m.equations)
    equations.push_back(dynamic_cast<BinaryOpNode *>(f(it)));
  for (const auto & it : m.aux_equations)
    aux_equations.push_back(dynamic_cast<BinaryOpNode *>(f(it)));

  auto convert_deriv_map = [f](map<vector<int>, expr_t> dm)
    {
      map<vector<int>, expr_t> dm2;
      for (const auto &it : dm)
        dm2.emplace(it.first, f(it.second));
      return dm2;
    };

  // Derivatives
  for (const auto &it : m.derivatives)
    derivatives.push_back(convert_deriv_map(it));
  for (const auto &it : m.params_derivatives)
    params_derivatives[it.first] = convert_deriv_map(it.second);

  auto convert_temporary_terms_t = [f](temporary_terms_t tt)
    {
      temporary_terms_t tt2;
      for (const auto &it : tt)
        tt2.insert(f(it));
      return tt2;
    };

  // Temporary terms
  for (const auto & it : m.temporary_terms)
    temporary_terms.insert(f(it));
  for (const auto & it : m.temporary_terms_mlv)
    temporary_terms_mlv[f(it.first)] = f(it.second);
  for (const auto &it : m.temporary_terms_derivatives)
    temporary_terms_derivatives.push_back(convert_temporary_terms_t(it));
  for (const auto & it : m.temporary_terms_idxs)
    temporary_terms_idxs[f(it.first)] = it.second;
  for (const auto & it : m.params_derivs_temporary_terms)
    params_derivs_temporary_terms[it.first] = convert_temporary_terms_t(it.second);
  for (const auto & it : m.params_derivs_temporary_terms_idxs)
    params_derivs_temporary_terms_idxs[f(it.first)] = it.second;

  // Other stuff
  for (const auto & it : m.trend_symbols_map)
    trend_symbols_map[it.first] = f(it.second);
  for (const auto & it : m.nonstationary_symbols_map)
    nonstationary_symbols_map[it.first] = {it.second.first, f(it.second.second)};
}

ModelTree::ModelTree(SymbolTable &symbol_table_arg,
                     NumericalConstants &num_constants_arg,
                     ExternalFunctionsTable &external_functions_table_arg,
                     bool is_dynamic_arg) :
  DataTree {symbol_table_arg, num_constants_arg, external_functions_table_arg, is_dynamic_arg},
  derivatives(4),
  NNZDerivatives(4, 0),
  temporary_terms_derivatives(4)
{
}

ModelTree::ModelTree(const ModelTree &m) :
  DataTree {m},
  equations_lineno {m.equations_lineno},
  equation_tags {m.equation_tags},
  NNZDerivatives {m.NNZDerivatives},
  equation_reordered {m.equation_reordered},
  variable_reordered {m.variable_reordered},
  inv_equation_reordered {m.inv_equation_reordered},
  inv_variable_reordered {m.inv_variable_reordered},
  is_equation_linear {m.is_equation_linear},
  endo2eq {m.endo2eq},
  epilogue {m.epilogue},
  prologue {m.prologue},
  block_lag_lead {m.block_lag_lead},
  cutoff {m.cutoff},
  mfs {m.mfs}
{
  copyHelper(m);
}

ModelTree &
ModelTree::operator=(const ModelTree &m)
{
  DataTree::operator=(m);

  equations.clear();
  equations_lineno = m.equations_lineno;
  aux_equations.clear();
  equation_tags = m.equation_tags;
  NNZDerivatives = m.NNZDerivatives;

  derivatives.clear();
  params_derivatives.clear();

  temporary_terms.clear();
  temporary_terms_mlv.clear();
  temporary_terms_derivatives.clear();
  params_derivs_temporary_terms.clear();
  params_derivs_temporary_terms_idxs.clear();

  trend_symbols_map.clear();
  nonstationary_symbols_map.clear();

  equation_reordered = m.equation_reordered;
  variable_reordered = m.variable_reordered;
  inv_equation_reordered = m.inv_equation_reordered;
  inv_variable_reordered = m.inv_variable_reordered;
  is_equation_linear = m.is_equation_linear;
  endo2eq = m.endo2eq;
  epilogue = m.epilogue;
  prologue = m.prologue;
  block_lag_lead = m.block_lag_lead;
  cutoff = m.cutoff;
  mfs = m.mfs;

  copyHelper(m);

  return *this;
}


bool
ModelTree::computeNormalization(const jacob_map_t &contemporaneous_jacobian, bool verbose)
{
  const int n = equations.size();

  assert(n == symbol_table.endo_nbr());

  using BipartiteGraph = boost::adjacency_list<boost::vecS, boost::vecS, boost::undirectedS>;

  /*
    Vertices 0 to n-1 are for endogenous (using type specific ID)
    Vertices n to 2*n-1 are for equations (using equation no.)
  */
  BipartiteGraph g(2 * n);

  // Fill in the graph
  set<pair<int, int>> endo;

  for (const auto & it : contemporaneous_jacobian)
    add_edge(it.first.first + n, it.first.second, g);

  // Compute maximum cardinality matching
  vector<int> mate_map(2*n);

#if 1
  bool check = checked_edmonds_maximum_cardinality_matching(g, &mate_map[0]);
#else // Alternative way to compute normalization, by giving an initial matching using natural normalizations
  fill(mate_map.begin(), mate_map.end(), boost::graph_traits<BipartiteGraph>::null_vertex());

  multimap<int, int> natural_endo2eqs;
  computeNormalizedEquations(natural_endo2eqs);

  for (int i = 0; i < symbol_table.endo_nbr(); i++)
    {
      if (natural_endo2eqs.count(i) == 0)
        continue;

      int j = natural_endo2eqs.find(i)->second;

      put(&mate_map[0], i, n+j);
      put(&mate_map[0], n+j, i);
    }

  boost::edmonds_augmenting_path_finder<BipartiteGraph, int *, boost::property_map<BipartiteGraph, boost::vertex_index_t>::type> augmentor(g, &mate_map[0], get(boost::vertex_index, g));
  while (augmentor.augment_matching())
    {
    };

  augmentor.get_current_matching(&mate_map[0]);

  bool check = boost::maximum_cardinality_matching_verifier<BipartiteGraph, int *, boost::property_map<BipartiteGraph, boost::vertex_index_t>::type>::verify_matching(g, &mate_map[0], get(boost::vertex_index, g));
#endif

  assert(check);

#ifdef DEBUG
  for (int i = 0; i < n; i++)
    cout << "Endogenous " << symbol_table.getName(symbol_table.getID(eEndogenous, i))
         << " matched with equation " << (mate_map[i]-n+1) << endl;
#endif

  // Create the resulting map, by copying the n first elements of mate_map, and substracting n to them
  endo2eq.resize(equations.size());
  transform(mate_map.begin(), mate_map.begin() + n, endo2eq.begin(), [=](int i) { return i-n; });

#ifdef DEBUG
  multimap<int, int> natural_endo2eqs;
  computeNormalizedEquations(natural_endo2eqs);

  int n1 = 0, n2 = 0;

  for (int i = 0; i < symbol_table.endo_nbr(); i++)
    {
      if (natural_endo2eqs.count(i) == 0)
        continue;

      n1++;

      auto x = natural_endo2eqs.equal_range(i);
      if (find_if(x.first, x.second, [=](auto y) { return y.second == endo2eq[i]; }) == x.second)
        cout << "Natural normalization of variable " << symbol_table.getName(symbol_table.getID(SymbolType::endogenous, i))
             << " not used." << endl;
      else
        n2++;
    }

  cout << "Used " << n2 << " natural normalizations out of " << n1 << ", for a total of " << n << " equations." << endl;
#endif

  // Check if all variables are normalized
  auto it = find(mate_map.begin(), mate_map.begin() + n, boost::graph_traits<BipartiteGraph>::null_vertex());
  if (it != mate_map.begin() + n)
    {
      if (verbose)
        cerr << "ERROR: Could not normalize the model. Variable "
             << symbol_table.getName(symbol_table.getID(SymbolType::endogenous, it - mate_map.begin()))
             << " is not in the maximum cardinality matching." << endl;
      check = false;
    }
  return check;
}

void
ModelTree::computeNonSingularNormalization(jacob_map_t &contemporaneous_jacobian, double cutoff, jacob_map_t &static_jacobian, dynamic_jacob_map_t &dynamic_jacobian)
{
  bool check = false;

  cout << "Normalizing the model..." << endl;

  int n = equations.size();

  // compute the maximum value of each row of the contemporaneous Jacobian matrix
  //jacob_map normalized_contemporaneous_jacobian;
  jacob_map_t normalized_contemporaneous_jacobian(contemporaneous_jacobian);
  vector<double> max_val(n, 0.0);
  for (const auto &it : contemporaneous_jacobian)
    if (fabs(it.second) > max_val[it.first.first])
      max_val[it.first.first] = fabs(it.second);

  for (auto & iter : normalized_contemporaneous_jacobian)
    iter.second /= max_val[iter.first.first];

  //We start with the highest value of the cutoff and try to normalize the model
  double current_cutoff = 0.99999999;

  int suppressed = 0;
  while (!check && current_cutoff > 1e-19)
    {
      jacob_map_t tmp_normalized_contemporaneous_jacobian;
      int suppress = 0;
      for (auto & iter : normalized_contemporaneous_jacobian)
        if (fabs(iter.second) > max(current_cutoff, cutoff))
          tmp_normalized_contemporaneous_jacobian[{ iter.first.first, iter.first.second }] = iter.second;
        else
          suppress++;

      if (suppress != suppressed)
        check = computeNormalization(tmp_normalized_contemporaneous_jacobian, false);
      suppressed = suppress;
      if (!check)
        {
          current_cutoff /= 2;
          // In this last case try to normalize with the complete jacobian
          if (current_cutoff <= 1e-19)
            check = computeNormalization(normalized_contemporaneous_jacobian, false);
        }
    }

  if (!check)
    {
      cout << "Normalization failed with cutoff, trying symbolic normalization..." << endl;
      //if no non-singular normalization can be found, try to find a normalization even with a potential singularity
      jacob_map_t tmp_normalized_contemporaneous_jacobian;
      set<pair<int, int>> endo;
      for (int i = 0; i < n; i++)
        {
          endo.clear();
          equations[i]->collectEndogenous(endo);
          for (const auto & it : endo)
            tmp_normalized_contemporaneous_jacobian[{ i, it.first }] = 1;
        }
      check = computeNormalization(tmp_normalized_contemporaneous_jacobian, true);
      if (check)
        {
          // Update the jacobian matrix
          for (const auto &it : tmp_normalized_contemporaneous_jacobian)
            {
              if (static_jacobian.find({ it.first.first, it.first.second }) == static_jacobian.end())
                static_jacobian[{ it.first.first, it.first.second }] = 0;
              if (dynamic_jacobian.find({ 0, it.first.first, it.first.second }) == dynamic_jacobian.end())
                dynamic_jacobian[{ 0, it.first.first, it.first.second }] = nullptr;
              if (contemporaneous_jacobian.find({ it.first.first, it.first.second }) == contemporaneous_jacobian.end())
                contemporaneous_jacobian[{ it.first.first, it.first.second }] = 0;
              try
                {
                  if (derivatives[1].find({ it.first.first, getDerivID(symbol_table.getID(SymbolType::endogenous, it.first.second), 0) }) == derivatives[1].end())
                    derivatives[1][{ it.first.first, getDerivID(symbol_table.getID(SymbolType::endogenous, it.first.second), 0) }] = Zero;
                }
              catch (DataTree::UnknownDerivIDException &e)
                {
                  cerr << "The variable " << symbol_table.getName(symbol_table.getID(SymbolType::endogenous, it.first.second))
                       << " does not appear at the current period (i.e. with no lead and no lag); this case is not handled by the 'block' option of the 'model' block." << endl;
                  exit(EXIT_FAILURE);
                }
            }
        }
    }

  if (!check)
    {
      cerr << "No normalization could be computed. Aborting." << endl;
      exit(EXIT_FAILURE);
    }
}

void
ModelTree::computeNormalizedEquations(multimap<int, int> &endo2eqs) const
{
  for (size_t i = 0; i < equations.size(); i++)
    {
      auto *lhs = dynamic_cast<VariableNode *>(equations[i]->arg1);
      if (lhs == nullptr)
        continue;

      int symb_id = lhs->symb_id;
      if (symbol_table.getType(symb_id) != SymbolType::endogenous)
        continue;

      set<pair<int, int>> endo;
      equations[i]->arg2->collectEndogenous(endo);
      if (endo.find({ symbol_table.getTypeSpecificID(symb_id), 0 }) != endo.end())
        continue;

      endo2eqs.emplace(symbol_table.getTypeSpecificID(symb_id), static_cast<int>(i));
      cout << "Endogenous " << symbol_table.getName(symb_id) << " normalized in equation " << (i+1) << endl;
    }
}

void
ModelTree::evaluateAndReduceJacobian(const eval_context_t &eval_context, jacob_map_t &contemporaneous_jacobian, jacob_map_t &static_jacobian, dynamic_jacob_map_t &dynamic_jacobian, double cutoff, bool verbose)
{
  int nb_elements_contemparenous_Jacobian = 0;
  set<vector<int>> jacobian_elements_to_delete;
  for (const auto &it : derivatives[1])
    {
      int deriv_id = it.first[1];
      if (getTypeByDerivID(deriv_id) == SymbolType::endogenous)
        {
          expr_t Id = it.second;
          int eq = it.first[0];
          int symb = getSymbIDByDerivID(deriv_id);
          int var = symbol_table.getTypeSpecificID(symb);
          int lag = getLagByDerivID(deriv_id);
          double val = 0;
          try
            {
              val = Id->eval(eval_context);
            }
          catch (ExprNode::EvalExternalFunctionException &e)
            {
              val = 1;
            }
          catch (ExprNode::EvalException &e)
            {
              cerr << "ERROR: evaluation of Jacobian failed for equation " << eq+1 << " (line " << equations_lineno[eq] << ") and variable " << symbol_table.getName(symb) << "(" << lag << ") [" << symb << "] !" << endl;
              Id->writeOutput(cerr, ExprNodeOutputType::matlabDynamicModelSparse, temporary_terms, {});
              cerr << endl;
              exit(EXIT_FAILURE);
            }
          if (fabs(val) < cutoff)
            {
              if (verbose)
                cout << "the coefficient related to variable " << var << " with lag " << lag << " in equation " << eq << " is equal to " << val << " and is set to 0 in the incidence matrix (size=" << symbol_table.endo_nbr() << ")" << endl;
              jacobian_elements_to_delete.insert({ eq, deriv_id });
            }
          else
            {
              if (lag == 0)
                {
                  nb_elements_contemparenous_Jacobian++;
                  contemporaneous_jacobian[{ eq, var }] = val;
                }
              if (static_jacobian.find({ eq, var }) != static_jacobian.end())
                static_jacobian[{ eq, var }] += val;
              else
                static_jacobian[{ eq, var }] = val;
              dynamic_jacobian[{ lag, eq, var }] = Id;
            }
        }
    }

  // Get rid of the elements of the Jacobian matrix below the cutoff
  for (const auto & it : jacobian_elements_to_delete)
    derivatives[1].erase(it);

  if (jacobian_elements_to_delete.size() > 0)
    {
      cout << jacobian_elements_to_delete.size() << " elements among " << derivatives[1].size() << " in the incidence matrices are below the cutoff (" << cutoff << ") and are discarded" << endl
           << "The contemporaneous incidence matrix has " << nb_elements_contemparenous_Jacobian << " elements" << endl;
    }
}

vector<pair<int, int> >
ModelTree::select_non_linear_equations_and_variables(vector<bool> is_equation_linear, const dynamic_jacob_map_t &dynamic_jacobian, vector<int> &equation_reordered, vector<int> &variable_reordered,
                                                     vector<int> &inv_equation_reordered, vector<int> &inv_variable_reordered,
                                                     lag_lead_vector_t &equation_lag_lead, lag_lead_vector_t &variable_lag_lead,
                                                     vector<unsigned int> &n_static, vector<unsigned int> &n_forward, vector<unsigned int> &n_backward, vector<unsigned int> &n_mixed)
{
  vector<int> eq2endo(equations.size(), 0);
  /*equation_reordered.resize(equations.size());
  variable_reordered.resize(equations.size());*/
  unsigned int num = 0;
  for (auto it : endo2eq)
    if (!is_equation_linear[it])
      num++;
  vector<int> endo2block = vector<int>(endo2eq.size(), 1);
  vector<pair<set<int>, pair<set<int>, vector<int> > > > components_set(num);
  int i = 0, j = 0;
  for (auto it : endo2eq)
    {
      if (!is_equation_linear[it])
        {
          equation_reordered[i] = it;
          variable_reordered[i] = j;
          endo2block[j] = 0;
          components_set[endo2block[j]].first.insert(i);
          /*cout << " -----------------------------------------" << endl;
          cout.flush();
          cout << "equation_reordered[" << *it << "] = " << i << endl;
          cout.flush();
          cout << "variable_reordered[" << j << "] = " << i << endl;
          cout.flush();
          cout << "components_set[" << endo2block[j] << "].first[" << i << "] = " << i << endl;
          cout << "endo2block[" << j << "] = " << 0 << endl;
          cout.flush();
          */
          i++;
          j++;
        }
    }
/*  for (unsigned int j = 0; j < is_equation_linear.size() ; j++)
     cout << "endo2block[" << j << "] = " << endo2block[j] << endl;*/
  /*cout << "before getVariableLeadLAgByBlock\n";
  cout.flush();*/
  getVariableLeadLagByBlock(dynamic_jacobian, endo2block, endo2block.size(), equation_lag_lead, variable_lag_lead, equation_reordered, variable_reordered);
  n_static = vector<unsigned int>(endo2eq.size(), 0);
  n_forward = vector<unsigned int>(endo2eq.size(), 0);
  n_backward = vector<unsigned int>(endo2eq.size(), 0);
  n_mixed = vector<unsigned int>(endo2eq.size(), 0);
  for (unsigned int i = 0; i < endo2eq.size(); i++)
    {
      if      (variable_lag_lead[variable_reordered[i]].first != 0 && variable_lag_lead[variable_reordered[i]].second != 0)
        n_mixed[i]++;
      else if (variable_lag_lead[variable_reordered[i]].first == 0 && variable_lag_lead[variable_reordered[i]].second != 0)
        n_forward[i]++;
      else if (variable_lag_lead[variable_reordered[i]].first != 0 && variable_lag_lead[variable_reordered[i]].second == 0)
        n_backward[i]++;
      else if (variable_lag_lead[variable_reordered[i]].first == 0 && variable_lag_lead[variable_reordered[i]].second == 0)
        n_static[i]++;
    }
  cout.flush();
  int nb_endo = is_equation_linear.size();
  vector<pair<int, int>> blocks(1, {i, i});
  inv_equation_reordered.resize(nb_endo);
  inv_variable_reordered.resize(nb_endo);
  for (int i = 0; i < nb_endo; i++)
    {
      inv_variable_reordered[variable_reordered[i]] = i;
      inv_equation_reordered[equation_reordered[i]] = i;
    }
  return blocks;
}

bool
ModelTree::computeNaturalNormalization()
{
  bool bool_result = true;
  set<pair<int, int> > result;
  endo2eq.resize(equations.size());
  for (int eq = 0; eq < static_cast<int>(equations.size()); eq++)
    if (!is_equation_linear[eq])
      {
        BinaryOpNode *eq_node = equations[eq];
        expr_t lhs = eq_node->arg1;
        result.clear();
        lhs->collectDynamicVariables(SymbolType::endogenous, result);
        if (result.size() == 1 && result.begin()->second == 0)
          {
            //check if the endogenous variable has not been already used in an other match !
             auto it = find(endo2eq.begin(), endo2eq.end(), result.begin()->first);
             if (it == endo2eq.end())
               endo2eq[result.begin()->first] = eq;
             else
              {
                bool_result = false;
                break;
              }
          }
      }
  return bool_result;
}

void
ModelTree::computePrologueAndEpilogue(const jacob_map_t &static_jacobian_arg, vector<int> &equation_reordered, vector<int> &variable_reordered)
{
  vector<int> eq2endo(equations.size(), 0);
  equation_reordered.resize(equations.size());
  variable_reordered.resize(equations.size());
  int n = equations.size();
  vector<bool> IM(n*n);
  int i = 0;
  for (auto it : endo2eq)
    {
      eq2endo[it] = i;
      equation_reordered[i] = i;
      variable_reordered[it] = i;
      i++;
    }
  if (cutoff == 0)
    {
      set<pair<int, int>> endo;
      for (int i = 0; i < n; i++)
        {
          endo.clear();
          equations[i]->collectEndogenous(endo);
          for (const auto & it : endo)
            IM[i * n + endo2eq[it.first]] = true;
        }
    }
  else
    for (const auto & it : static_jacobian_arg)
      IM[it.first.first * n + endo2eq[it.first.second]] = true;
  bool something_has_been_done = true;
  prologue = 0;
  int k = 0;
  // Find the prologue equations and place first the AR(1) shock equations first
  while (something_has_been_done)
    {
      int tmp_prologue = prologue;
      something_has_been_done = false;
      for (int i = prologue; i < n; i++)
        {
          int nze = 0;
          for (int j = tmp_prologue; j < n; j++)
            if (IM[i * n + j])
              {
                nze++;
                k = j;
              }
          if (nze == 1)
            {
              for (int j = 0; j < n; j++)
                {
                  bool tmp_bool = IM[tmp_prologue * n + j];
                  IM[tmp_prologue * n + j] = IM[i * n + j];
                  IM[i * n + j] = tmp_bool;
                }
              int tmp = equation_reordered[tmp_prologue];
              equation_reordered[tmp_prologue] = equation_reordered[i];
              equation_reordered[i] = tmp;
              for (int j = 0; j < n; j++)
                {
                  bool tmp_bool = IM[j * n + tmp_prologue];
                  IM[j * n + tmp_prologue] = IM[j * n + k];
                  IM[j * n + k] = tmp_bool;
                }
              tmp = variable_reordered[tmp_prologue];
              variable_reordered[tmp_prologue] = variable_reordered[k];
              variable_reordered[k] = tmp;
              tmp_prologue++;
              something_has_been_done = true;
            }
        }
      prologue = tmp_prologue;
    }

  something_has_been_done = true;
  epilogue = 0;
  // Find the epilogue equations
  while (something_has_been_done)
    {
      int tmp_epilogue = epilogue;
      something_has_been_done = false;
      for (int i = prologue; i < n - static_cast<int>(epilogue); i++)
        {
          int nze = 0;
          for (int j = prologue; j < n - tmp_epilogue; j++)
            if (IM[j * n + i])
              {
                nze++;
                k = j;
              }
          if (nze == 1)
            {
              for (int j = 0; j < n; j++)
                {
                  bool tmp_bool = IM[(n - 1 - tmp_epilogue) * n + j];
                  IM[(n - 1 - tmp_epilogue) * n + j] = IM[k * n + j];
                  IM[k * n + j] = tmp_bool;
                }
              int tmp = equation_reordered[n - 1 - tmp_epilogue];
              equation_reordered[n - 1 - tmp_epilogue] = equation_reordered[k];
              equation_reordered[k] = tmp;
              for (int j = 0; j < n; j++)
                {
                  bool tmp_bool = IM[j * n + n - 1 - tmp_epilogue];
                  IM[j * n + n - 1 - tmp_epilogue] = IM[j * n + i];
                  IM[j * n + i] = tmp_bool;
                }
              tmp = variable_reordered[n - 1 - tmp_epilogue];
              variable_reordered[n - 1 - tmp_epilogue] = variable_reordered[i];
              variable_reordered[i] = tmp;
              tmp_epilogue++;
              something_has_been_done = true;
            }
        }
      epilogue = tmp_epilogue;
    }
}

equation_type_and_normalized_equation_t
ModelTree::equationTypeDetermination(const map<tuple<int, int, int>, expr_t> &first_order_endo_derivatives, const vector<int> &Index_Var_IM, const vector<int> &Index_Equ_IM, int mfs) const
{
  expr_t lhs;
  BinaryOpNode *eq_node;
  EquationType Equation_Simulation_Type;
  equation_type_and_normalized_equation_t V_Equation_Simulation_Type(equations.size());
  for (unsigned int i = 0; i < equations.size(); i++)
    {
      int eq = Index_Equ_IM[i];
      int var = Index_Var_IM[i];
      eq_node = equations[eq];
      lhs = eq_node->arg1;
      Equation_Simulation_Type = E_SOLVE;
      auto derivative = first_order_endo_derivatives.find({ eq, var, 0 });
      pair<bool, expr_t> res;
      if (derivative != first_order_endo_derivatives.end())
        {
          set<pair<int, int>> result;
          derivative->second->collectEndogenous(result);
          auto d_endo_variable = result.find({ var, 0 });
          //Determine whether the equation could be evaluated rather than to be solved
          if (lhs->isVariableNodeEqualTo(SymbolType::endogenous, Index_Var_IM[i], 0) && derivative->second->isNumConstNodeEqualTo(1))
            Equation_Simulation_Type = E_EVALUATE;
          else
            {
              vector<tuple<int, expr_t, expr_t>> List_of_Op_RHS;
              res =  equations[eq]->normalizeEquation(var, List_of_Op_RHS);
              if (mfs == 2)
                {
                  if (d_endo_variable == result.end() && res.second)
                    Equation_Simulation_Type = E_EVALUATE_S;
                }
              else if (mfs == 3)
                {
                  if (res.second) // The equation could be solved analytically
                    Equation_Simulation_Type = E_EVALUATE_S;
                }
            }
        }
      V_Equation_Simulation_Type[eq] = { Equation_Simulation_Type, dynamic_cast<BinaryOpNode *>(res.second) };
    }
  return V_Equation_Simulation_Type;
}

void
ModelTree::getVariableLeadLagByBlock(const dynamic_jacob_map_t &dynamic_jacobian, const vector<int> &components_set, int nb_blck_sim, lag_lead_vector_t &equation_lead_lag, lag_lead_vector_t &variable_lead_lag, const vector<int> &equation_reordered, const vector<int> &variable_reordered) const
{
  int nb_endo = symbol_table.endo_nbr();
  variable_lead_lag = lag_lead_vector_t(nb_endo, { 0, 0 });
  equation_lead_lag = lag_lead_vector_t(nb_endo, { 0, 0 });
  vector<int> variable_blck(nb_endo), equation_blck(nb_endo);
  for (int i = 0; i < nb_endo; i++)
    {
      if (i < static_cast<int>(prologue))
        {
          variable_blck[variable_reordered[i]] = i;
          equation_blck[equation_reordered[i]] = i;
        }
      else if (i < static_cast<int>(components_set.size() + prologue))
        {
          variable_blck[variable_reordered[i]] = components_set[i-prologue] + prologue;
          equation_blck[equation_reordered[i]] = components_set[i-prologue] + prologue;
        }
      else
        {
          variable_blck[variable_reordered[i]] = i- (nb_endo - nb_blck_sim - prologue - epilogue);
          equation_blck[equation_reordered[i]] = i- (nb_endo - nb_blck_sim - prologue - epilogue);
        }
    }
  for (const auto & it : dynamic_jacobian)
    {
      int lag, j_1, i_1;
      tie(lag, j_1, i_1) = it.first;
      if (variable_blck[i_1] == equation_blck[j_1])
        {
          if (lag > variable_lead_lag[i_1].second)
            variable_lead_lag[i_1] = { variable_lead_lag[i_1].first, lag };
          if (lag < -variable_lead_lag[i_1].first)
            variable_lead_lag[i_1] = { -lag, variable_lead_lag[i_1].second };
          if (lag > equation_lead_lag[j_1].second)
            equation_lead_lag[j_1] = { equation_lead_lag[j_1].first, lag };
          if (lag < -equation_lead_lag[j_1].first)
            equation_lead_lag[j_1] = { -lag, equation_lead_lag[j_1].second };
        }
    }
}

void
ModelTree::computeBlockDecompositionAndFeedbackVariablesForEachBlock(const jacob_map_t &static_jacobian, const dynamic_jacob_map_t &dynamic_jacobian, vector<int> &equation_reordered, vector<int> &variable_reordered, vector<pair<int, int>> &blocks, const equation_type_and_normalized_equation_t &Equation_Type, bool verbose_, bool select_feedback_variable, int mfs, vector<int> &inv_equation_reordered, vector<int> &inv_variable_reordered, lag_lead_vector_t &equation_lag_lead, lag_lead_vector_t &variable_lag_lead, vector<unsigned int> &n_static, vector<unsigned int> &n_forward, vector<unsigned int> &n_backward, vector<unsigned int> &n_mixed) const
{
  int nb_var = variable_reordered.size();
  int n = nb_var - prologue - epilogue;

  AdjacencyList_t G2(n);

  // It is necessary to manually initialize vertex_index property since this graph uses listS and not vecS as underlying vertex container
  auto v_index = get(boost::vertex_index, G2);
  for (int i = 0; i < n; i++)
    put(v_index, vertex(i, G2), i);

  vector<int> reverse_equation_reordered(nb_var), reverse_variable_reordered(nb_var);

  for (int i = 0; i < nb_var; i++)
    {
      reverse_equation_reordered[equation_reordered[i]] = i;
      reverse_variable_reordered[variable_reordered[i]] = i;
    }
  jacob_map_t tmp_normalized_contemporaneous_jacobian;
  if (cutoff == 0)
    {
      set<pair<int, int>> endo;
      for (int i = 0; i < nb_var; i++)
        {
          endo.clear();
          equations[i]->collectEndogenous(endo);
          for (const auto & it : endo)
            tmp_normalized_contemporaneous_jacobian[{ i, it.first }] = 1;

        }
    }
  else
    tmp_normalized_contemporaneous_jacobian = static_jacobian;
  for (const auto &it : tmp_normalized_contemporaneous_jacobian)
    if (reverse_equation_reordered[it.first.first] >= static_cast<int>(prologue) && reverse_equation_reordered[it.first.first] < static_cast<int>(nb_var - epilogue)
        && reverse_variable_reordered[it.first.second] >= static_cast<int>(prologue) && reverse_variable_reordered[it.first.second] < static_cast<int>(nb_var - epilogue)
        && it.first.first != endo2eq[it.first.second])
      add_edge(vertex(reverse_equation_reordered[endo2eq[it.first.second]]-prologue, G2),
               vertex(reverse_equation_reordered[it.first.first]-prologue, G2),
               G2);

  vector<int> endo2block(num_vertices(G2)), discover_time(num_vertices(G2));
  boost::iterator_property_map<int *, boost::property_map<AdjacencyList_t, boost::vertex_index_t>::type, int, int &> endo2block_map(&endo2block[0], get(boost::vertex_index, G2));

  // Compute strongly connected components
  int num = strong_components(G2, endo2block_map);

  blocks = vector<pair<int, int>>(num, { 0, 0 });

  // Create directed acyclic graph associated to the strongly connected components
  using DirectedGraph = boost::adjacency_list<boost::vecS, boost::vecS, boost::directedS>;
  DirectedGraph dag(num);

  for (unsigned int i = 0; i < num_vertices(G2); i++)
    {
      AdjacencyList_t::out_edge_iterator it_out, out_end;
      AdjacencyList_t::vertex_descriptor vi = vertex(i, G2);
      for (tie(it_out, out_end) = out_edges(vi, G2); it_out != out_end; ++it_out)
        {
          int t_b = endo2block_map[target(*it_out, G2)];
          int s_b = endo2block_map[source(*it_out, G2)];
          if (s_b != t_b)
            add_edge(s_b, t_b, dag);
        }
    }

  // Compute topological sort of DAG (ordered list of unordered SCC)
  deque<int> ordered2unordered;
  topological_sort(dag, front_inserter(ordered2unordered)); // We use a front inserter because topological_sort returns the inverse order

  // Construct mapping from unordered SCC to ordered SCC
  vector<int> unordered2ordered(num);
  for (int i = 0; i < num; i++)
    unordered2ordered[ordered2unordered[i]] = i;

  //This vector contains for each block:
  //   - first set = equations belonging to the block,
  //   - second set = the feeback variables,
  //   - third vector = the reordered non-feedback variables.
  vector<tuple<set<int>, set<int>, vector<int>>> components_set(num);
  for (unsigned int i = 0; i < endo2block.size(); i++)
    {
      endo2block[i] = unordered2ordered[endo2block[i]];
      blocks[endo2block[i]].first++;
      get<0>(components_set[endo2block[i]]).insert(i);
    }

  getVariableLeadLagByBlock(dynamic_jacobian, endo2block, num, equation_lag_lead, variable_lag_lead, equation_reordered, variable_reordered);

  vector<int> tmp_equation_reordered(equation_reordered), tmp_variable_reordered(variable_reordered);
  int order = prologue;
  //Add a loop on vertices which could not be normalized or vertices related to lead variables => force those vertices to belong to the feedback set
  if (select_feedback_variable)
    {
      for (int i = 0; i < n; i++)
        if (Equation_Type[equation_reordered[i+prologue]].first == E_SOLVE
            || variable_lag_lead[variable_reordered[i+prologue]].second > 0
            || variable_lag_lead[variable_reordered[i+prologue]].first > 0
            || equation_lag_lead[equation_reordered[i+prologue]].second > 0
            || equation_lag_lead[equation_reordered[i+prologue]].first > 0
            || mfs == 0)
          add_edge(vertex(i, G2), vertex(i, G2), G2);
    }
  else
    {
      for (int i = 0; i < n; i++)
        if (Equation_Type[equation_reordered[i+prologue]].first == E_SOLVE || mfs == 0)
          add_edge(vertex(i, G2), vertex(i, G2), G2);
    }
  //Determines the dynamic structure of each equation
  n_static = vector<unsigned int>(prologue+num+epilogue, 0);
  n_forward = vector<unsigned int>(prologue+num+epilogue, 0);
  n_backward = vector<unsigned int>(prologue+num+epilogue, 0);
  n_mixed = vector<unsigned int>(prologue+num+epilogue, 0);

  for (int i = 0; i < static_cast<int>(prologue); i++)
    {
      if      (variable_lag_lead[tmp_variable_reordered[i]].first != 0 && variable_lag_lead[tmp_variable_reordered[i]].second != 0)
        n_mixed[i]++;
      else if (variable_lag_lead[tmp_variable_reordered[i]].first == 0 && variable_lag_lead[tmp_variable_reordered[i]].second != 0)
        n_forward[i]++;
      else if (variable_lag_lead[tmp_variable_reordered[i]].first != 0 && variable_lag_lead[tmp_variable_reordered[i]].second == 0)
        n_backward[i]++;
      else if (variable_lag_lead[tmp_variable_reordered[i]].first == 0 && variable_lag_lead[tmp_variable_reordered[i]].second == 0)
        n_static[i]++;
    }
  //For each block, the minimum set of feedback variable is computed
  // and the non-feedback variables are reordered to get
  // a sub-recursive block without feedback variables

  for (int i = 0; i < num; i++)
    {
      AdjacencyList_t G = extract_subgraph(G2, get<0>(components_set[i]));
      set<int> feed_back_vertices;
      //Print(G);
      AdjacencyList_t G1 = Minimal_set_of_feedback_vertex(feed_back_vertices, G);
      auto v_index = get(boost::vertex_index, G);
      get<1>(components_set[i]) = feed_back_vertices;
      blocks[i].second = feed_back_vertices.size();
      vector<int> Reordered_Vertice;
      Reorder_the_recursive_variables(G, feed_back_vertices, Reordered_Vertice);

      //First we have the recursive equations conditional on feedback variables
      for (int j = 0; j < 4; j++)
        {
          for (int its : Reordered_Vertice)
            {
              bool something_done = false;
              if      (j == 2 && variable_lag_lead[tmp_variable_reordered[its +prologue]].first != 0 && variable_lag_lead[tmp_variable_reordered[its +prologue]].second != 0)
                {
                  n_mixed[prologue+i]++;
                  something_done = true;
                }
              else if (j == 3 && variable_lag_lead[tmp_variable_reordered[its +prologue]].first == 0 && variable_lag_lead[tmp_variable_reordered[its +prologue]].second != 0)
                {
                  n_forward[prologue+i]++;
                  something_done = true;
                }
              else if (j == 1 && variable_lag_lead[tmp_variable_reordered[its +prologue]].first != 0 && variable_lag_lead[tmp_variable_reordered[its +prologue]].second == 0)
                {
                  n_backward[prologue+i]++;
                  something_done = true;
                }
              else if (j == 0 && variable_lag_lead[tmp_variable_reordered[its +prologue]].first == 0 && variable_lag_lead[tmp_variable_reordered[its +prologue]].second == 0)
                {
                  n_static[prologue+i]++;
                  something_done = true;
                }
              if (something_done)
                {
                  equation_reordered[order] = tmp_equation_reordered[its+prologue];
                  variable_reordered[order] = tmp_variable_reordered[its+prologue];
                  order++;
                }
            }
        }
      get<2>(components_set[i]) = Reordered_Vertice;
      //Second we have the equations related to the feedback variables
      for (int j = 0; j < 4; j++)
        {
          for (int feed_back_vertice : feed_back_vertices)
            {
              bool something_done = false;
              if      (j == 2 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].first != 0 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].second != 0)
                {
                  n_mixed[prologue+i]++;
                  something_done = true;
                }
              else if (j == 3 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].first == 0 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].second != 0)
                {
                  n_forward[prologue+i]++;
                  something_done = true;
                }
              else if (j == 1 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].first != 0 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].second == 0)
                {
                  n_backward[prologue+i]++;
                  something_done = true;
                }
              else if (j == 0 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].first == 0 && variable_lag_lead[tmp_variable_reordered[v_index[vertex(feed_back_vertice, G)]+prologue]].second == 0)
                {
                  n_static[prologue+i]++;
                  something_done = true;
                }
              if (something_done)
                {
                  equation_reordered[order] = tmp_equation_reordered[v_index[vertex(feed_back_vertice, G)]+