Abstract
This paper extends Willems' Fundamental Lemma (WFL) to nonlinear systems that admit a Koopman linear embedding. By characterizing the relationship between the trajectory space of a nonlinear system and its high-dimensional linear counterpart, the authors demonstrate that nonlinear dynamics can be represented using only raw data trajectories. This approach, termed Data-Driven Koopman (DD-K), eliminates the need for manually designing lifting functions, thereby bypassing associated bias errors. Experimental results show that DD-K achieves superior prediction accuracy and control performance compared to standard Koopman and deep learning baselines, while maintaining high data efficiency.
Bridging Linear and Nonlinear Control
In control theory, linear systems benefit from robust tools like Willems' Fundamental Lemma, which allows for model-free representation using past trajectories. Nonlinear systems, however, typically require complex "lifting" into higher-dimensional spaces via Koopman operator theory. Choosing these lifting functions is often a heuristic process prone to error. This research bridges the gap by proving that if a Koopman embedding exists, the raw trajectories of the nonlinear system already contain the necessary information to span the lifted space, making explicit lifting redundant.
The DD-K Framework
The proposed DD-K framework relies on "lifted excitation," a condition ensuring that data trajectories are sufficiently diverse to capture the hidden dimensions of the Koopman space. The authors prove that a linear combination of these rich trajectories can perfectly represent the system's evolution. This allows for the construction of predictive models, such as Data-enabled Predictive Control (DeePC), using only raw input-output data matrices.
The Critical Role of Trajectory Depth
A key theoretical insight is the requirement for "depth" in the data. Unlike many data-driven methods that prioritize the number of trajectories, DD-K requires a sufficiently long initial trajectory ($T_{ini}$) to uniquely identify the system's state within the high-dimensional Koopman space. As shown in the simulation results, the prediction only aligns perfectly with the true system once the depth threshold is met.
Performance Benchmarks
The DD-K method was validated against several baselines: data-driven affine representation (DD-A), Extended Dynamic Mode Decomposition (EDMD-K), and Deep Neural Network-based Koopman (DNN-K). In tracking tasks involving sinusoidal and step signals, DD-K demonstrated significantly lower realized control costs and superior tracking performance.
Notably, DD-K achieved these results with far less data than deep learning approaches. While DNN-K required approximately $10^5$ trajectories and exhibited high variance, DD-K succeeded with a small set of well-chosen trajectories, highlighting its efficiency and theoretical exactness for systems with linear embeddings.
Conclusion
By extending WFL to the Koopman framework, this work provides a rigorous foundation for model-free nonlinear control. While the method assumes the existence of a finite-dimensional embedding and requires knowledge of the Koopman space dimension, it offers a powerful alternative to traditional approximation-based methods. Future work will address noise sensitivity and automated dimension estimation.
Summary updated