Improved rates for identification of partially observed linear dynamical systems

Identification of a linear time-invariant dynamical system from partial observations is a fundamental problem in control theory. A natural question is how to do so with non-asymptotic statistical rates depending on the inherent dimensionality (order) $d$ of the system, rather than on the sufficient rollout length or on $\frac1{1-\rho(A)}$, where $\rho(A)$ is the spectral radius of the dynamics matrix. We develop the first algorithm that given a single trajectory of length $T$ with gaussian observation noise, achieves a near-optimal rate of $\widetilde O\left(\sqrt\frac{d}{T}\right)$ in $\mathcal{H}_2$ error for the learned system. We also give bounds under process noise and improved bounds for learning a realization of the system. Our algorithm is based on low-rank approximation of Hankel matrices of geometrically increasing sizes.