Abstract
We study the problem of learning mixtures of linear dynamical systems (MLDS)from input-output data. The mixture setting allows us to leverage observationsfrom related dynamical systems to improve the estimation of individual models.Building on spectral methods for mixtures of linear regressions, we propose amoment-based estimator that uses tensor decomposition to estimate the impulseresponse parameters of the mixture models. The estimator improves upon existingtensor decomposition approaches for MLDS by utilizing the entire length of theobserved trajectories. We provide sample complexity bounds for estimating MLDSin the presence of noise, in terms of both the number of trajectories $N$ andthe trajectory length $T$, and demonstrate the performance of the estimatorthrough simulations.