// Returns an n by 1 column that represents the likelihoods of
// an estimated univariate MS-AR model for each k where 1 \leq k \leq n where
// beta is switching.
// Note that even if beta is non-switching, setting beta as a s by M matrix with
// repeated column of the original beta will give you the likelihood for
// MS-AR model with non-switching beta.
// TODO: Implement the version where z_dependent/z_independent exist(lagged variables should be implemented for MSM models)
// [[Rcpp::export]]
SEXP LikelihoodsMSMAR (Rcpp::NumericVector y_rcpp,
Rcpp::NumericMatrix y_lagged_rcpp,
Rcpp::NumericMatrix z_dependent_rcpp,
Rcpp::NumericMatrix z_independent_rcpp,
Rcpp::NumericMatrix z_dependent_lagged_rcpp,
Rcpp::NumericMatrix z_independent_lagged_rcpp,
Rcpp::NumericMatrix transition_probs_rcpp,
Rcpp::NumericVector initial_dist_extended_rcpp,
Rcpp::NumericMatrix beta_rcpp,
Rcpp::NumericVector mu_rcpp,
Rcpp::NumericVector sigma_rcpp,
Rcpp::NumericMatrix gamma_dependent_rcpp,
Rcpp::NumericVector gamma_independent_rcpp,
Rcpp::IntegerMatrix state_conversion_mat_rcpp
)
{
int n = y_rcpp.size();
arma::colvec y(y_rcpp.begin(), y_rcpp.size(), false);
arma::mat y_lagged(y_lagged_rcpp.begin(),
y_lagged_rcpp.nrow(),
y_lagged_rcpp.ncol(), false);
arma::mat z_dependent(z_dependent_rcpp.begin(),
z_dependent_rcpp.nrow(),
z_dependent_rcpp.ncol(), false);
arma::mat z_dependent_lagged(z_dependent_lagged_rcpp.begin(),
z_dependent_lagged_rcpp.nrow(),
z_dependent_lagged_rcpp.ncol(), false);
arma::mat z_independent(z_independent_rcpp.begin(),
z_independent_rcpp.nrow(),
z_independent_rcpp.ncol(), false);
arma::mat z_independent_lagged(z_independent_lagged_rcpp.begin(),
z_independent_lagged_rcpp.nrow(),
z_independent_lagged_rcpp.ncol(), false);
arma::mat transition_probs(transition_probs_rcpp.begin(),
transition_probs_rcpp.nrow(),
transition_probs_rcpp.ncol(), false);
arma::colvec initial_dist_extended(initial_dist_extended_rcpp.begin(),
initial_dist_extended_rcpp.size(), false);
arma::mat beta(beta_rcpp.begin(),
beta_rcpp.nrow(), beta_rcpp.ncol(), false);
arma::colvec mu(mu_rcpp.begin(), mu_rcpp.size(), false);
arma::colvec sigma(sigma_rcpp.begin(), sigma_rcpp.size(), false);
arma::mat gamma_dependent(gamma_dependent_rcpp.begin(),
gamma_dependent_rcpp.nrow(),
gamma_dependent_rcpp.ncol(), false);
arma::colvec gamma_independent(gamma_independent_rcpp.begin(),
gamma_independent_rcpp.size(), false);
arma::imat state_conversion_mat(state_conversion_mat_rcpp.begin(),
state_conversion_mat_rcpp.nrow(),
state_conversion_mat_rcpp.ncol(), false);
arma::mat transition_probs_extended = GetExtendedTransitionProbs(
transition_probs, state_conversion_mat);
arma::mat transition_probs_extended_t = transition_probs_extended.t();
arma::colvec likelihoods(n, arma::fill::zeros);
int M_extended = transition_probs_extended_t.n_rows;
int M = gamma_dependent.n_cols;
int s = beta.n_rows;
int M_extended_block = IntPower(M, s);
arma::mat* xi_k_t = new arma::mat(M_extended, n, arma::fill::zeros); // make a transpose first for col operations.
// partition blocks
int p = gamma_dependent.n_rows;
int q = gamma_independent.size();
arma::mat* z_dependent_lagged_blocks = new arma::mat[s];
arma::mat* z_independent_lagged_blocks = new arma::mat[s];
for (int i = 0; i < s; i++)
{
int z_dependent_block_first = i * p;
int z_independent_block_first = i * q;
z_dependent_lagged_blocks[i] = z_dependent_lagged.cols(z_dependent_block_first,
z_dependent_block_first + p - 1);
z_independent_lagged_blocks[i] = z_independent_lagged.cols(z_independent_block_first,
z_independent_block_first + q - 1);
}
for (int k = 0; k < n; k++)
{
// initial setting; keep track of minimum value and its index
// to divide everything by the min. value in order to prevent
// possible numerical errors when computing posterior probs.
int min_index = -1;
double min_value = std::numeric_limits<double>::infinity();
double* ratios = new double[M_extended];
double row_sum = 0;
arma::colvec xi_past;
if (k > 0)
xi_past = transition_probs_extended_t * exp(xi_k_t->col(k-1));
else
xi_past = initial_dist_extended;
xi_past /= arma::sum(xi_past);
for (int j_M = 0; j_M < M; j_M++)
{
for (int j_extra = 0; j_extra < M_extended_block; j_extra++)
{
int j = j_M * M_extended_block + j_extra;
arma::colvec xi_k_t_jk = y.row(k);
// arma::colvec xi_k_t_jk = y.row(k) -
// z_dependent.row(k) * gamma_dependent.col(j_M) -
// z_independent.row(k) * gamma_independent - mu(j_M);
for (int lag = 0; lag < s; lag++)
{
int lagged_index = state_conversion_mat.at((lag + 1), j);
xi_k_t_jk -= beta.at(lag, j_M) *
(y_lagged.at(k, lag) -
mu(lagged_index));
xi_k_t_jk -= beta.at(lag, j_M) *
(y_lagged.at(k, lag) -
z_dependent_lagged_blocks[lag].row(k) * gamma_dependent.col(lagged_index) -
z_independent_lagged_blocks[lag].row(k) * gamma_independent -
mu(lagged_index));
}
xi_k_t->at(j,k) = xi_k_t_jk(0); // explicit gluing
xi_k_t->at(j,k) *= xi_k_t->at(j,k);
xi_k_t->at(j,k) = xi_k_t->at(j,k) / (2 * (sigma(j_M) * sigma(j_M)));
if (min_value > xi_k_t->at(j,k))
{
min_value = xi_k_t->at(j,k);
min_index = j;
}
// SQRT2PI only matters in calculation of eta;
// you can remove it in the final log-likelihood.
ratios[j] = xi_past(j) / sigma(j_M);
}
}
for (int j = 0; j < M_extended; j++)
{
if (j == min_index)
row_sum += 1.0;
else
row_sum += (ratios[j] / ratios[min_index]) *
exp(min_value - xi_k_t->at(j,k));
xi_k_t->at(j,k) += log(ratios[j]);
}
likelihoods(k) = log(row_sum) - min_value + log(ratios[min_index]) - LOG2PI_OVERTWO;
delete[] ratios; // clear memory
}
arma::exp(xi_k_t->cols(1,4)).t().print();
// clear memory for blocks
delete[] z_dependent_lagged_blocks;
delete[] z_independent_lagged_blocks;
delete xi_k_t;
return (wrap(likelihoods));
}
// [[Rcpp::export]]
unsigned int GetNumAlive(const Rcpp::IntegerMatrix& board)
{
return std::accumulate(board.begin(), board.end(), 0.0);
}
// [[Rcpp::export]]
XPtrImage magick_image_readbitmap_native(Rcpp::IntegerMatrix x){
Rcpp::IntegerVector dims(x.attr("dim"));
return magick_image_bitmap(x.begin(), Magick::CharPixel, 4, dims[1], dims[0]);
}
