In classical control theory, we face a frustrating trade-off. We have incredibly powerful, mature tools for managing linear systems—like Model Predictive Control (MPC) and Kalman filters—but the real world is stubbornly nonlinear. Traditionally, we handle this by "local linearization," approximating a system as linear only in the immediate neighborhood of a specific point. But as soon as a drone tilts too far or a chemical reactor shifts temperatures, these local approximations crumble.
A growing body of research, centered on Koopman Operator Theory, offers a different path: instead of approximating the dynamics, we transform the coordinate system. By "lifting" nonlinear states into a higher-dimensional space of observables, we can represent complex, chaotic motion as a simple linear evolution.
The Intuition of Lifting
The core idea, as explored by J. Nathan Kutz, Joshua L. Proctor, and Steven L. Brunton in Koopman Theory for Partial Differential Equations, is a shift in perspective. Rather than following the state $x$ itself, we follow "observables"—functions of the state, $g(x)$.
Imagine a particle moving along a circle. In $(x, y)$ coordinates, its velocity is a nonlinear function of its position. However, if we "lift" the problem into a space that includes functions like $\sin(\theta)$ and $\cos(\theta)$, the rotation becomes a simple linear shift. The Koopman operator is the infinite-dimensional linear operator that describes how these functions evolve. While the operator itself is infinite, researchers like Yongqian Xiao and Xinglong Zhang use deep neural networks to find a finite-dimensional "lifting function" that captures the most important dynamics, a framework they call Deep EDMD.
Why It Works: Neural Networks as Coordinate Finders
The difficulty of Koopman theory has always been choosing the right observables. If you pick the wrong functions, the system won't look linear. Modern research solves this by using deep learning to learn the mapping.
In Deep Koopman learning of nonlinear time-varying systems, Wenjian Hao and Shaoshuai Mou demonstrate that a neural network can act as an encoder, mapping raw sensor data into a "latent" linear space. Once in this space, the system follows $z_{t+1} = Az_t$. This allows for real-time optimal control of systems with fluctuating properties, such as a cartpole with changing friction, because the "heavy lifting" of understanding the nonlinearity is handled by the encoder, leaving the controller to solve a simple linear problem.
This approach is particularly potent for high-dimensional problems. Chenye Wang and George Em Karniadakis showed in A time-dependent symplectic network for non-convex path planning that by mapping non-convex dynamics into a linear latent space, they could coordinate 100 drones simultaneously—a feat that would normally be impossible due to the "curse of dimensionality."
Applications in Complex Environments
Because Koopman models are data-driven, they excel in areas where the underlying physics are too messy to model by hand:
- Autonomous Vehicles: Xiao et al. used Koopman operators to model vehicle dynamics, enabling real-time MPC that is as accurate as nonlinear models but significantly faster.
- Power Grids: Zihan Wang and colleagues applied this to suppress oscillations in renewable energy grids, transforming chaotic frequency wobbles into predictable linear modes.
- Human Motion: The DenseKoopman framework by Han Jiang treats pedestrian trajectories as a dynamical system, lifting social interactions into a Hilbert space to predict movement in crowded environments with up to 85% higher accuracy.
Handling Uncertainty and Stability
A common critique of data-driven models is their "black box" nature. However, the Koopman framework allows for rigorous mathematical guarantees. Jeremy Morton and Mykel J. Kochenderfer introduced Deep Variational Koopman Models, which infer a distribution over linear models. This allows a controller to know when it is "uncertain" about the dynamics, enabling robust, safety-critical control.
Similarly, Tomoharu Iwata and Yoshinobu Kawahara have integrated pole placement control directly into the learning process. By constraining the eigenvalues of the learned linear operator, they can guarantee that the resulting nonlinear system will be stable and converge at a desired rate.
By shifting the complexity from nonlinear dynamics in low dimensions to linear dynamics in high dimensions, Koopman operator learning provides a bridge between the flexibility of deep learning and the reliability of classical control theory.