Learning structured linear dynamical systems from missing observations
The research tackles estimating a linear dynamical system matrix $A^*$ when observations are missing from a trajectory. It proposes an estimator using a bias-corrected objective function and proves that Projected Gradient Descent converges linearly to the true matrix $A^*$. The findings show that accurate recovery of $A^*$ is possible even with limited data, provided the trajectory length is not too long relative to the system size.