In classical dynamical systems theory, we face a fundamental trade-off: we can model systems using simple linear mathematics, or we can model the real world, which is overwhelmingly nonlinear. Linear systems are elegant and easy to control, but nonlinear systems quickly become intractable.
Koopman operator theory offers a radical escape hatch from this dilemma by trading nonlinearity for dimensionality. Instead of tracking the state of a system directly, we track functions of that state, known as observables. As reviewed by Steven L. Brunton et al. in Modern Koopman Theory for Dynamical Systems, if you choose the right set of observables, the evolution of those functions becomes perfectly linear.
The catch is that this linear operator is typically infinite-dimensional. To make this theory practically useful for computation, prediction, and control, we must find a way to compress this infinity into a finite-dimensional matrix without losing the underlying physics. This requires identifying a finite-dimensional invariant subspace.
The Intuition of Invariance and the Closure Problem
To understand why an invariant subspace is necessary, consider a simple nonlinear system governed by the differential equation:
$$\dot{x} = x^2$$
If we want to represent this system linearly, we can define our first observable function as the state itself: $g_1(x) = x$. Taking its time derivative yields:
$$\dot{g}_1 = \dot{x} = x^2$$
To write this as a linear equation, we are forced to introduce a second observable, $g_2(x) = x^2$, so that $\dot{g}_1 = g_2$. But now we need to know how $g_2$ evolves:
$$\dot{g}_2 = \frac{d}{dt}(x^2) = 2x\dot{x} = 2x^3$$
This requires a third observable, $g_3(x) = x^3$. If we keep going, we will generate an infinite chain of observables. This infinite cascade is known as the closure problem (Joshua L. Proctor et al., Including inputs and control within equation-free architectures for complex systems). Because the derivatives keep producing new nonlinear terms, the system never "closes."
An invariant subspace is a specific, finite pool of observable functions where the time derivative of every function in the pool can be expressed strictly as a linear combination of functions already in the pool. If we can find such a subspace, the infinite chain stops, the dynamics close perfectly, and the infinite-dimensional Koopman operator collapses into a clean, finite-dimensional matrix.
Discovering the Right Observables
Historically, finding these magical coordinate systems required deep physical intuition or manual trial-and-error. Modern data-driven methods automate this search by combining linear algebra with machine learning.
Sparse Regression
If we suspect that the invariant subspace can be spanned by combinations of standard functions (like polynomials or sinusoids), we can use sparse regression to discover them. In Koopman Invariant Subspaces and Finite Linear Representations of Nonlinear Dynamical Systems for Control, Steven L. Brunton et al. utilize the SINDy (Sparse Identification of Nonlinear Dynamics) algorithm to evaluate a large dictionary of candidate functions. By penalizing complexity, the algorithm strips away the functions that cause the infinite cascade, isolating the precise, sparse subset of observables that form a closed, invariant linear system.
Adaptive Dictionary Learning
When the underlying physics are too complex or completely unknown, fixed dictionaries of functions fail. To bypass this, Qianxiao Li et al. introduced EDMD-DL (Extended dynamic mode decomposition with dictionary learning: A data-driven adaptive spectral decomposition of the Koopman operator). Instead of picking functions beforehand, this approach uses a neural network to continuously warp and optimize the dictionary of observables during training. The network learns to reshape the coordinate system so that the data fits a linear model with minimal variables, making the Koopman framework scalable to high-dimensional data.
Pruning and Regularizing the Subspace
Even with powerful neural networks, blindly searching for an invariant subspace can lead to numerical instability or "mode collapse," where the model learns trivial, uninformative representations.
To enforce strict mathematical invariance, Masih Haseli et al. developed the Symmetric Subspace Decomposition (SSD) algorithm in Efficient Identification of Linear Evolutions in Nonlinear Vector Fields: Koopman Invariant Subspaces. Rather than just minimizing prediction error, SSD takes an arbitrary dictionary of functions and systematically prunes it down to its largest possible, strictly linear invariant subspace. The authors later scaled this concept to high-dimensional systems by introducing a parallelized framework (Parallel Learning of Koopman Eigenfunctions and Invariant Subspaces For Accurate Long-Term Prediction), ensuring that the discovered subspaces remain stable over long prediction horizons.
Another perspective leverages information theory to guide the search. In Information Shapes Koopman Representation, Xiaoyuan Cheng et al. show that the search for a compact representation can be framed as a balance between latent mutual information (which enforces simplicity) and von Neumann entropy (which preserves expressivity). This prevents the learned subspace from collapsing into redundant dimensions, ensuring the model remains predictive over long time horizons.
Handling Inputs and Continuous Spectra
While finding an invariant subspace for an isolated, ideal system is challenging enough, real-world applications introduce control inputs, noise, and chaotic behaviors.
Actuated Systems
When a system has external control inputs, the standard Koopman operator breaks because the inputs continuously push the system out of its invariant subspace. The authors of Koopman form of nonlinear systems with inputs (Lucian Cristian Iacob et al.) address this by embedding the dynamics into a Linear Parameter-Varying (LPV) framework, providing an exact, finite-dimensional representation that accommodates control. For systems where trajectories are only partially observed and corrupted by noise, Samuel E. Otto et al. frame the problem as a hidden Markov model (Learning Bilinear Models of Actuated Koopman Generators from Partially Observed Trajectories), using an Expectation-Maximization algorithm to robustly isolate the underlying linear dynamics.
The Challenge of Chaos
For highly chaotic or turbulent systems, the Koopman operator possesses a continuous spectrum—meaning the system's frequencies are blended together rather than separated into distinct, discrete eigenvalues. In these scenarios, traditional finite approximations fail because they try to fit a continuous phenomenon with a discrete set of functions.
To capture these complex spectral footprints, Yuanchao Xu et al. developed ResKoopNet (ResKoopNet: Learning Koopman Representations for Complex Dynamics with Spectral Residuals). Instead of merely filtering data, this architecture directly minimizes spectral residuals, allowing neural networks to approximate the continuous operator spectrum accurately.
By shifting the focus from approximating the infinite operator to isolating its finite, invariant components, this research direction bridges the gap between abstract operator theory and practical engineering, turning the dream of universal linearization into a computable reality.
Key Foundational Work
Koopman Invariant Subspaces and Finite Linear Representations of Nonlinear Dynamical Systems for Control (2016) — contributed 2 concept elements to this direction.
Extended dynamic mode decomposition with dictionary learning: A data-driven adaptive spectral decomposition of the Koopman operator (2017) — contributed 1 concept elements to this direction.
Including inputs and control within equation-free architectures for complex systems (2016) — contributed 1 concept elements to this direction.