Nonlinear dynamical systems are notoriously difficult to predict and control. While we have a massive toolkit for linear systems—think of the Kalman filter or Linear Quadratic Regulators (LQR)—these tools often fail when applied to the "curvy" reality of fluid dynamics, robotics, or biological systems.
Koopman operator theory offers a radical workaround: instead of trying to solve the nonlinear equations directly, we "lift" the system into a higher-dimensional space where the dynamics appear linear. The catch is that this space is usually infinite-dimensional. Recent research focuses on the closure problem: finding a small, finite set of "golden" coordinates that capture the full complexity of the system without losing information.
The Lifting Trick
Imagine a particle moving along a complex, nonlinear path on a 2D plane. In its original state space $(x, y)$, the equations governing its motion are a mess of sines, cosines, and polynomials.
Koopman theory suggests that if we look at functions of the state (called observables), like $x^2$, $xy$, or $\sin(y)$, we might find that these functions evolve linearly. As Joshua L. Proctor, J. Nathan Kutz, and Steven L. Brunton explain in Including inputs and control within equation-free architectures for complex systems, this transformation allows us to treat a complex system as a simple linear one, provided we pick the right functions.
The Search for Invariant Subspaces
The fundamental challenge is that most nonlinear systems require an infinite number of these observable functions to be perfectly linear. In practice, we need a finite-dimensional invariant subspace—a small collection of functions that, when evolved forward in time, stay within the same collection.
Steven L. Brunton and colleagues demonstrated in Koopman Invariant Subspaces and Finite Linear Representations of Nonlinear Dynamical Systems for Control that we can use sparse regression (SINDy) to discover these subspaces from data. By identifying a sparse set of nonlinear observables that remain closed under the dynamics, they enabled the application of linear control techniques to fully nonlinear problems.
Learning the Dictionary
If we don't know which functions to pick, why not let a machine learn them? This is the core of Extended Dynamic Mode Decomposition with Dictionary Learning (EDMD-DL).
In Extended dynamic mode decomposition with dictionary learning, Qianxiao Li, Felix Dietrich, Erik M. Bollt, and Yannis G. Kevrekidis used neural networks to "learn" the best dictionary of functions. By making the dictionary adaptive, they achieved high accuracy with far fewer variables than traditional fixed-basis methods.
However, learning these representations can be unstable. Deep learning models often suffer from "mode collapse," where the model ignores certain parts of the system's behavior. To combat this, Xiaoliang Cheng, Sibo Cheng, and their team introduced an information-theoretic framework in Information Shapes Koopman Representation. By balancing mutual information (simplicity) with entropy (expressivity), they ensure the learned linear space is both compact and complete.
Is There a Minimum?
A natural question arises: how many functions do we actually need? While the theory often points to infinity, Ido Cohen and Haibin Li provided a theoretical breakthrough in A Minimal Set of Koopman Eigenfunctions -- Analysis and Numerics. They proved that for an $N$-dimensional system, you only need a minimal set of $N$ eigenfunctions to fully reconstruct the dynamics. This "golden set" of coordinates effectively bridges the gap between abstract operator theory and practical engineering.
Beyond Fixed Dictionaries: Kernels and Residuals
For systems with continuous spectra—like turbulence—finding a finite set of functions is even harder. Researchers are now moving toward nonparametric approaches. Alessandro Chiuso in Estimating Koopman operators for nonlinear dynamical systems: a nonparametric approach and Jong‐Hyeon Lee in Kernel Methods for the Approximation of the Eigenfunctions of the Koopman Operator leverage Reproducing Kernel Hilbert Spaces (RKHS). These methods allow the model to operate in an implicit infinite-dimensional space without ever having to define the functions manually.
Finally, to capture the most elusive parts of a system's behavior, Yuanchao Xu developed ResKoopNet in ResKoopNet: Learning Koopman Representations for Complex Dynamics with Spectral Residuals. Instead of just filtering noise, this architecture actively searches for "spectral residuals"—the missing pieces of the system's frequency spectrum—to ensure the linear model is truly representative of the underlying physics.