Predicting and controlling complex, nonlinear systems—such as fluid flows, neural trajectories, or climate patterns—is notoriously difficult. For decades, the gold standard for analyzing dynamical systems has been linear system theory, which offers powerful tools like spectral decomposition and optimal control. To bring these linear tools to nonlinear systems, researchers frequently turn to Koopman operator theory.
The core intuition behind the Koopman approach is a trade-off: you can transform a low-dimensional, highly nonlinear dynamical system into a linear one, provided you are willing to lift the system into an infinite-dimensional space of "observable" functions.
$$\mathbf{x}_{t+1} = f(\mathbf{x}_t) \quad \xrightarrow{\text{Lift via } \psi(\mathbf{x})} \quad \psi(\mathbf{x}_{t+1}) = \mathcal{K} \psi(\mathbf{x}_t)$$
While the Koopman operator $\mathcal{K}$ is linear, computing with an infinite-dimensional operator is practically impossible. To make this framework tractable, data-driven methods like Extended Dynamic Mode Decomposition (EDMD) project the operator onto a finite-dimensional subspace spanned by a set of basis functions, often called a dictionary of observables.
Historically, choosing this dictionary was an art form. Researchers manually selected combinations of polynomials, Fourier bases, or radial basis functions based on trial and error or domain expertise. However, a rapidly growing body of literature demonstrates that this manual selection process is the primary bottleneck of operator-theoretic modeling. If the chosen dictionary fails to align with the system's intrinsic geometry, the resulting linear model fails to generalize, yields slow convergence, or breaks down entirely.
To overcome this limitation, modern research is shifting away from heuristic dictionary selection toward automated, adaptive basis function discovery.
The Vulnerability of Fixed Dictionaries
Why does manual selection fail so easily? The fundamental challenge is that the accuracy of finite-dimensional operator approximations is highly sensitive to whether the chosen basis functions can compactly represent the system's true eigenfunctions.
As Nicolas Boullé demonstrates in a study on On the Convergence of Hermitian Dynamic Mode Decomposition, if a dictionary fails to adequately span the relevant Hilbert space, the resulting spectral measures suffer from slow convergence or incomplete representations of the underlying dynamics. This sensitivity manifests as a severe engineering bottleneck across multiple domains:
- Control Design: In lifting-based control frameworks, Robin Strässer et al. show in Koopman-based feedback design with stability guarantees that controller performance and formal stability guarantees are highly sensitive to the dictionary choice, making manual engineering a major liability.
- System Identification: Samuel E. Otto et al. point out in Learning Bilinear Models of Actuated Koopman Generators from Partially Observed Trajectories that there is no universal or systematic way to pre-define basis functions for complex, non-measure-preserving systems.
- Spectral Accuracy: A comprehensive review by Frank Noe et al. in Data-Driven Model Reduction and Transfer Operator Approximation highlights that while simple bases yield standard linear models, precise approximations of transfer operator eigenfunctions require complex, system-specific basis sets that human intuition rarely uncovers.
When human intuition fails, the model is forced to use an excessively large number of generic basis functions to approximate the dynamics, leading to overfitting, high computational costs, and numerical instability.
Learning the Dictionary via Neural Networks
The most direct solution to the manual bottleneck is to treat the dictionary itself as a learnable parameter. Instead of fixing the basis functions a priori, researchers use deep neural networks to parameterize the lifting functions, optimizing them simultaneously with the linear operator.
Nonlinear State (x) ---> [ Neural Network ] ---> Learned Observables ψ(x) ---> [ Linear Matrix K ] ---> Next Observables ψ(x')
A pioneering framework for this approach is EDMD with Dictionary Learning (EDMD-DL), introduced by Qianxiao Li et al. in Extended dynamic mode decomposition with dictionary learning: A data-driven adaptive spectral decomposition of the Koopman operator. By making the dictionary adaptive rather than fixed, the network automatically discovers a compact, low-dimensional set of observables that captures complex nonlinearities far more accurately than thousands of randomly chosen polynomials.
This adaptive philosophy has been successfully tailored to specific physical domains. For instance, in highly complex multiphase flow problems, Yating Wang et al. developed AMS-Net: Adaptive Multiscale Sparse Neural Network with Interpretable Basis Expansion for Multiphase Flow Problems. AMS-Net adaptively selects and utilizes a sparse set of precomputed basis functions, allowing the model to remain accurate even as underlying physical parameters shift dynamically.
Optimization and Subspace Pruning
While neural networks offer immense flexibility, they can be prone to local minima and lack formal guarantees. A parallel branch of research focuses on finding optimal dictionaries through convex optimization or algebraic pruning of an initial candidate pool.
Instead of relying on deep learning, Mario Sznaier proposed a convex optimization alternative in A Data Driven, Convex Optimization Approach to Learning Koopman Operators. By recasting the simultaneous search for observables and linear dynamics as a rank-constrained semi-definite program (SDP), this approach automatically extracts the simplest possible linear representation directly from data, bypassing manual dictionary selection entirely.
Isolating Invariant Subspaces
Another elegant strategy is to start with a large, over-complete dictionary of candidate functions and systematically prune it down to a Koopman invariant subspace—a space of functions that, when mapped by the Koopman operator, stays within the same finite-dimensional space.
In Learning Koopman Eigenfunctions and Invariant Subspaces From Data: Symmetric Subspace Decomposition, Masih Haseli introduced the Symmetric Subspace Decomposition (SSD) algorithm. SSD leverages a powerful mathematical intuition: true Koopman eigenfunctions must govern the system's evolution consistently whether the data is viewed forward or backward in time. By enforcing this temporal consistency, the algorithm isolates exact eigenfunctions from a noisy dictionary.
However, as Masih Haseli et al. note in Efficient Identification of Linear Evolutions in Nonlinear Vector Fields: Koopman Invariant Subspaces, these pruning methods are fundamentally bounded by the span of the initial candidate pool. If the true underlying eigenfunctions cannot be represented by any combination of the starting functions, the algorithm will return a trivial or incomplete representation.
Why Automated Selection Works
Automated dictionary discovery works because it aligns the mathematical representation with the intrinsic geometry of the data. When we manually choose a basis, we force the data into an arbitrary coordinate system (like polynomials), which often requires an infinite number of terms to resolve sharp gradients or non-local features.
By letting optimization algorithms or neural networks discover the basis functions, the framework identifies the true coordinates along which the system evolves linearly. This drastically reduces the dimensionality of the lifted space, eliminates human trial-and-error, and provides a mathematically rigorous path toward interpreting and controlling complex nonlinear systems.
Key Foundational Work
Dynamic Mode Decomposition with Reproducing Kernels for Koopman Spectral Analysis (2016) — contributed 1 concept elements to this direction.
Interpretable Nonlinear Dynamic Modeling of Neural Trajectories (2016) — contributed 1 concept elements to this direction.
Learning Additive Exponential Family Graphical Models via \ell_{2, 1}-norm Regularized M-Estimation (2016) — contributed 1 concept elements to this direction.