1: \begin{abstract}%

2: 	\noindent The Koopman operator has become an essential tool for data-driven approximation of dynamical (control) systems, e.g., via extended dynamic mode decomposition. Despite its popularity, convergence results and, in particular, error bounds are still %quite 

3: 	scarce. In this paper, we 

4: 	derive probabilistic bounds for the approximation error and the %(multi-step) 

5: 	prediction error depending on the number of training data points; for both ordinary and stochastic differential equations \MS{while using either ergodic trajectories or i.i.d.\ samples}. %

6: 	\MS{We illustrate these bounds by means of an example with the Ornstein-Uhlenbeck process. 

7: 	Moreover, we extend our analysis to (stochastic) nonlinear control-affine systems.

8: 	We prove error estimates for a previously proposed approach that exploits the linearity of the Koopman generator to obtain a bilinear surrogate control system and, thus, circumvents

9: 	the curse of dimensionality since the system is not autonomized by augmenting the state by the control inputs. 

10: 	To the best of our knowledge, this is the first finite-data error analysis in the stochastic and/or control setting. 

11: 	Finally, we demonstrate the effectiveness of the bilinear approach by comparing it with state-of-the-art techniques showing its superiority whenever state and control are coupled.} %

12: \end{abstract}

13: