ordered_constrain.hpp

Go to the documentation of this file.

#ifndef STAN_MATH_REV_CONSTRAINT_ORDERED_CONSTRAIN_HPP
#define STAN_MATH_REV_CONSTRAINT_ORDERED_CONSTRAIN_HPP
 
#include <stan/math/rev/meta.hpp>
#include <stan/math/rev/core/reverse_pass_callback.hpp>
#include <stan/math/rev/core/arena_matrix.hpp>
#include <stan/math/rev/fun/value_of.hpp>
#include <stan/math/prim/fun/Eigen.hpp>
#include <cmath>
#include <tuple>
#include <vector>
 
namespace stan {
namespace math {
 
template <typename T, require_rev_col_vector_t<T>* = nullptr>
inline auto ordered_constrain(const T& x) {
  using ret_type = plain_type_t<T>;
 
  using std::exp;
 
  size_t N = x.size();
  if (unlikely(N == 0)) {
    return ret_type(x);
  }
 
  Eigen::VectorXd y_val(N);
  arena_t<T> arena_x = x;
  arena_t<Eigen::VectorXd> exp_x(N - 1);
 
  y_val.coeffRef(0) = arena_x.val().coeff(0);
  for (Eigen::Index n = 1; n < N; ++n) {
    exp_x.coeffRef(n - 1) = exp(arena_x.val().coeff(n));
    y_val.coeffRef(n) = y_val.coeff(n - 1) + exp_x.coeff(n - 1);
  }
 
  arena_t<ret_type> y = y_val;
 
  reverse_pass_callback([arena_x, y, exp_x]() mutable {
    double rolling_adjoint_sum = 0.0;
 
    for (int n = arena_x.size() - 1; n > 0; --n) {
      rolling_adjoint_sum += y.adj().coeff(n);
      arena_x.adj().coeffRef(n) += exp_x.coeff(n - 1) * rolling_adjoint_sum;
    }
    arena_x.adj().coeffRef(0) += rolling_adjoint_sum + y.adj().coeff(0);
  });
 
  return ret_type(y);
}
 
template <typename VarVec, require_var_col_vector_t<VarVec>* = nullptr>
auto ordered_constrain(const VarVec& x, scalar_type_t<VarVec>& lp) {
  if (x.size() > 1) {
    lp += sum(x.tail(x.size() - 1));
  }
  return ordered_constrain(x);
}
 
}  // namespace math
}  // namespace stan
#endif