Estimates mixing proportions (theta) using an EM algorithm with a simple acceleration step. The acceleration is attempted based on changes in theta over iterations and log-likelihood improvement. This is a simplified interpretation of HISEA's ACCEL subroutine.
Usage
accel_em_algorithm(
likelihood,
np,
freq = NULL,
max_iter = 100,
tol = 1e-06,
verbose = FALSE,
save_theta_path = NULL
)Arguments
- likelihood
A numeric matrix (N_rows x N_populations) where `likelihood[i, j]` is $P(Data_i | Stock_j)$ or a value proportional to it. N_rows can be N_samples (for Theta5) or N_populations (for Theta4).
- np
Integer, the number of populations (stocks). Should match `ncol(likelihood)`.
- freq
Optional numeric vector of frequencies/weights for each "observation" (row in `likelihood`). Length must match `nrow(likelihood)`. Defaults to `rep(1, nrow(likelihood))`.
- max_iter
Integer, the maximum number of iterations for the EM algorithm.
- tol
Numeric, the convergence tolerance. The algorithm stops if the maximum absolute change in any `theta` component is less than `tol`.
- verbose
Logical, if TRUE, prints iteration number and current theta estimates.
- save_theta_path
Optional character string. If provided, the history of theta estimates at each iteration is saved to a CSV file at this path.