Boosting-Enabled Robust System Identification of Partially Observed LTI Systems Under Heavy-Tailed Noise

System identification is a fundamental problem that involve s estimating unknown system parameters using noisy data generated from a dynamical process. It is re levant to various disciplines including control theory, economics, time-series forecasting, and m achine learning. System identification also forms a core sub-routine in data-driven control/model -based reinforcement learning where one uses the estimated system model for downstream decision-ma king. To ensure desired performance of such algorithms, it is crucial to quantify the uncertaint y in the data-driven estimates of the model. It stands to reason that the nature of such estimates, and the uncertainty intervals around them, will depend on the statistics of the data used for estim ation. In this regard, despite the wealth of literature on system identification spanning both classical asymptotic results [ 1 ] and more recent finite-time guarantees [ 2 - 4 ], almost all existing works on the topic crucially rely on th e noise processes being either Gaussian or sub-Gaussian, i.e., "li ght-tailed". In practice, however, such an idealistic assumption may not hold. Furthermore, estimato rs that do not account for non-ideal noise processes might lead to poor statistical guarantees that ar e inadequate for safety-critical real-time feedback control loops. With these points in mind, the goal of this work is to initiate a study of system identification under more realistic noise processes that are potentially heavy-tailed and admit no more than the second moment.