Dynamical systems, filtering, smoothing, and identification in Jax
pip install git+https://github.com/PredictiveScienceLab/dax.git
StochasticDifferentialEquation.diffusion returns the vector of diagonal
diffusion amplitudes. For a time step dt, EulerMaruyama therefore uses a
Gaussian transition with mean x + drift * dt and diagonal variance
diffusion**2 * dt. The Gaussian initial-state and observation models include
their normalizing constants, so reported log densities and marginal-likelihood
estimates are on the correct absolute scale.
ExpectationMaximization implements a particle Monte Carlo approximation. It
draws smoothing trajectories once at the beginning of each M-step and holds
them fixed during the inner optimization. Finite particle approximations and
numerical M-steps do not guarantee monotonic ascent of the observed-data
likelihood.
pip install -e ".[test]"
pytest -q