TL;DR
This paper leverages the multiplicative group property of Koopman operator eigenfunctions to systematically expand a small set of "principal" eigenfunctions into a much larger, richer dictionary using algebraic powers. Crucially, it introduces methods to handle and analytically continue these eigenfunctions across localized or extended singularities (such as steady states and limit cycles), enabling researchers to learn consistent global representations from fragmented, locally sampled data.
What Problem Does This Paper Address?
The Koopman operator is a powerful framework in dynamical systems because it maps nonlinear, finite-dimensional dynamics into an infinite-dimensional linear space, making complex systems easier to analyze and predict. However, data-driven approximation methods like Extended Dynamic Mode Decomposition (EDMD) rely on finite dictionaries of functions. These methods typically only exploit the linearity of the operator, often failing to capture the full eigenspace efficiently.
Furthermore, real-world dynamical systems frequently feature multiple basins of attraction, limit cycles, or separatrices. In these scenarios, Koopman eigenfunctions inherently exhibit singularities—meaning they blow up to infinity at the boundaries or steady states. Standard numerical methods break down near these infinities, making it exceptionally difficult to learn a cohesive global model from data collected in isolated regions of the state space.
Methods
The authors introduce an algebraic framework to accelerate eigenfunction computation and bridge across singularities:
- Algebraic Power Generation: Since the eigenfunctions of reversible continuous-time systems form a multiplicative group, any integer or real power of an eigenfunction is also an eigenfunction. The authors develop algorithms (Algorithms 1 & 2) to systematically generate higher-order "monomial" eigenfunctions from a minimal set of conventionally computed "principal" eigenfunctions.
- Rigorous Error Tracking: Because taking high powers amplifies numerical noise exponentially, they introduce a trajectory metric ($E_{FG}$) to establish strict upper bounds on approximation errors, automatically determining how far an eigenfunction can be safely extended.
- Log-Space Matching: To cross singularities, the authors compute eigenfunctions locally within separate regions where they are well-behaved. By transforming these functions into logarithmic space, the multiplicative relationships become linear, allowing them to use least-squares matching (with Tikhonov regularization) to stitch local eigenfunctions together across boundaries.
Datasets
The approach is validated entirely on synthetic, analytically tractable dynamical systems where numerical results can be compared directly against exact ground truths:
- A 1D polynomial system with three steady states to demonstrate basic singularity crossing.
- A 2D nonlinear system generated via a coordinate transformation (diffeomorphism) of a linear system.
- The Van der Pol oscillator to evaluate limit cycles, isochrons, and isostables.
- A 2D multistable system featuring a saddle point and two stable nodes.
- The unforced Duffing equation to analyze complex separatrices.
Key Insights and Ideas
- Unlocking Missed Eigenspaces: Standard EDMD is bounded by its dictionary. By utilizing the algebraic property of the Koopman operator, researchers can uncover an infinite family of higher-order eigenfunctions that standard linear methods completely miss (see Figure 2).
- Infinities as Geometric Features: Singularities are not just numerical failures; they explicitly define the boundaries of basins of attraction. By taking negative powers of an eigenfunction that blows up to infinity, the function safely approaches zero, turning a numerical hazard into a manageable feature.
- Jumping the Gap: Complex global dynamics can be reconstructed from sparse, local data by aligning the phases and magnitudes of local eigenfunction segments in log-space, effectively "jumping" over hazardous singularities.
Key Results
- Dictionary Enrichment: The proposed algorithms successfully generated accurate higher-order eigenfunctions ($p \ge 4$) without requiring them to be explicitly represented in the initial EDMD dictionary.
- Predictive Error Bounds: The trajectory error framework successfully predicted the exact threshold where numerical integration and eigenvector errors would begin to dominate as powers increased.
- Successful Global Continuation: In multi-basin simulations, the log-space linear mapping accurately matched eigenfunctions across separate steady states, proving that global behavior can be consistently reconstructed from fragmented local measurements.
Limitations and Future Work
While powerful, the method's reliance on taking powers means that initial numerical errors are amplified exponentially. This places a practical ceiling on how high of a power $p$ can be achieved without exceptionally clean data or high-precision eigensolvers. Additionally, the initial step still depends on EDMD, meaning performance remains sensitive to the choice of the base dictionary (e.g., the bandwidth of Radial Basis Functions).
For future work, the authors point toward exploring how these algebraic representations behave near complex bifurcation points, such as "cubic isochron foliation tangencies," where the representation itself becomes singular.
Summary updated