Koopman Mode Decomposition Explained

Updated 30 March 2026

Koopman Mode Decomposition is a spectral method that uses linear operator theory to analyze nonlinear systems via eigenfunctions and modes.
It extends classical modal analysis by enabling data-driven reconstruction and forecasting through techniques like DMD, EDMD, and Kernel DMD.
KMD finds practical applications in fields such as climate science, fluid dynamics, and power systems by isolating coherent dynamic structures.

Koopman Mode Decomposition (KMD) provides a spectral framework for analyzing and reconstructing the evolution of observables in potentially nonlinear and high-dimensional dynamical systems. By representing nonlinear dynamics through the linear but infinite-dimensional Koopman operator acting on observables, KMD generalizes classical modal analysis and underpins a range of modern data-driven techniques including Dynamic Mode Decomposition (DMD), Extended DMD (EDMD), Kernel DMD, and several probabilistic or robust extensions.

1. Koopman Operator and Spectral Decomposition

Given a discrete-time dynamical system $x_{t+1}=F(x_t)$ with $x_t\in\mathbb{R}^n$ and an observable $g:\mathbb{R}^n\to\mathbb{C}^p$ , the Koopman operator $\mathcal{K}$ acts linearly on observables via

$(\mathcal{K}g)(x) = g(F(x)).$

Although $F$ is generally nonlinear, $\mathcal{K}$ is linear (but infinite-dimensional). If $\mathcal{K}$ admits eigenfunctions $\varphi_j$ with corresponding eigenvalues $\lambda_j$ , i.e.,

$\mathcal{K}\varphi_j = \lambda_j \varphi_j,$

then any $g$ in the span of $\{\varphi_j\}$ can be expanded as

$g(x) = \sum_j \varphi_j(x)\,\xi_j,$

where $\xi_j\in\mathbb{C}^p$ are Koopman modes. The time evolution becomes

$g(x_k) = \sum_j \lambda_j^k \varphi_j(x_0)\,\xi_j.$

This modal expansion is referred to as the Koopman Mode Decomposition (KMD) (Williams et al., 2014, Hogg et al., 2019, Masuda et al., 2019).

2. Data-Driven Computation: DMD, EDMD, and Kernel Approaches

2.1 Dynamic Mode Decomposition (DMD)

DMD is a data-driven algorithm that approximates KMD by seeking a best-fit finite-dimensional linear operator $A$ such that $X' = AX$ , with $X, X'$ constructed from time-shifted sequence data. Using a truncated Singular Value Decomposition (SVD), a reduced operator $\tilde{A}$ is formed for efficient eigendecomposition. The DMD modes $\Phi$ approximate the Koopman modes, and the eigenvalues $\lambda_j$ approximate Koopman spectral values (Hogg et al., 2019).

2.2 Extended DMD (EDMD)

EDMD generalizes DMD using a user-specified dictionary $\{\psi_i\}_{i=1}^N$ of scalar observables, lifting the data to a higher-dimensional feature space. The finite-dimensional Koopman approximation is constructed by forming Gram-type matrices $G$ and $A$ over the data and solving the least-squares problem $GK = A$ , leading to eigenpairs $(\mu_j, v_j)$ for $K$ . The eigenfunctions and modes are explicitly represented in the dictionary basis, yielding

$\varphi_j(x) = \sum_{i=1}^N v_{j,i} \psi_i(x)$

and associated modes for vector-valued observables (Williams et al., 2014).

2.3 Kernel DMD/EDMD

Kernel methods embed the data in a (potentially infinite-dimensional) Reproducing Kernel Hilbert Space (RKHS) with a positive-definite kernel $k(x,y)$ . The kernel trick enables construction of the Gram matrices $K_{XX}$ and $K_{XY}$ directly from kernel evaluations, and the operator $U = (K_{XX} + \sigma^2 I)^{-1} K_{XY}$ is subjected to eigendecomposition. Koopman eigenfunctions are represented in RKHS via

$\varphi_j(x) = \sum_{i=1}^M w_{j,i} k(x, x_i),$

where $w_j$ is an eigenvector of $U$ (Zanini et al., 2021).

3. Robust and Adaptive KMD Algorithms

3.1 Stochastic and Noisy Systems

In stochastic or noisy dynamical systems, the spectral quantities of the stochastic Koopman operator $\mathcal{K}_\Omega$ must be estimated robustly. Subspace DMD addresses this by projecting future snapshot blocks onto the space of past snapshots, followed by an SVD and a compact eigenproblem. Convergence to the true spectrum of the stochastic Koopman operator is proved under ergodic and whiteness assumptions on the noise (Takeishi et al., 2017).

3.2 Kalman Filter-Based KMD

An ensemble Kalman filter architecture tracks time-varying Koopman modes and eigenvalues from noisy, possibly non-autonomous systems. The state vector includes eigenfunctions, eigenvalues, and expansion coefficients. At each step, the filter updates the hidden state by assimilating new measurements, allowing real-time online updating of the modal decomposition under both autonomous and slowly-varying conditions. Theoretical misfit bounds are established for long-term performance (Liu et al., 2024).

3.3 Sparse KMD

Sparsity-promoting DMD (SP-DMD) leverages $\ell_1$ optimization to extract a minimal set of dynamically significant Koopman modes for interpretable, reduced-order representations, particularly in high-dimensional climate applications. Optimization seeks growth rates, frequencies, and modes subject to minimal reconstruction error and high interpretability (Zhang et al., 8 Jul 2025).

4. Spectral Structure, Interpretation, and Applications

4.1 Spectral Interpretation

Koopman eigenvalues correspond to modal growth/decay and oscillation frequencies. Modes near the unit circle represent persistent oscillations; those with $|\lambda_j|<1$ describe exponential decay. In complex systems (e.g. climate dynamics, post-transient fluid flows), the spectral signature (discrete lattice vs. continuous band) encodes transitions between periodic, quasi-periodic, and chaotic regimes (Hogg et al., 2019, Arbabi et al., 2017, Zhang et al., 8 Jul 2025).

4.2 Real-World Applications

Climate science: KMD isolates spatially- and temporally-coherent modes such as the annual cycle, El Niño–Southern Oscillation (ENSO), and multidecadal variability from massive SST data, supporting long-range forecasting and targeted control (Zhang et al., 8 Jul 2025).
Fluid dynamics: Discrete and continuous spectra, as revealed by KMD, correspond to coherent structures and turbulent fluctuations, respectively. KMD outperforms orthogonal decompositions when strong oscillatory behaviors are present (Arbabi et al., 2017).
Power systems: Nonlinear KMD yields robust estimates of system inertia and modal participation using time-series data, even in the absence of explicit models, by leveraging the inherent modal structure of the swing equations (Susuki et al., 2018).
High-dimensional data and control: Input-Koopman frameworks and deep learning extensions permit computation of Koopman Gramians, enabling partitioning of high-dimensional nonlinear systems by maximizing subsystem observability and disturbance rejection (Liu et al., 2017).

5. Extensions, Theoretical Considerations, and Algorithmic Developments

5.1 Numerical Conditioning and Schur Decomposition

Koopman-Schur frameworks replace potentially ill-conditioned eigenbases with unitary Schur decompositions, yielding orthonormal modal subspaces and well-conditioned expansions—particularly critical for non-normal or nearly defective system matrices. Integration with EDMD and the kernel trick enables Schur-based KMD in general nonlinear and high-dimensional settings (Drmač et al., 2023).

5.2 Analytic and Function-Theoretic KMD

Analytic EDMD operates in an RKHS of analytic functions, exploits block-triangular structure in polynomial bases, and computes lattice-structured Koopman spectra and eigenfunctions with no spectral pollution. This approach achieves arbitrary spectral accuracy with fixed polynomial dimension and sufficient data (Mauroy et al., 2024).

5.3 Probabilistic and Nonparametric KMD

Gaussian process regression and nonlinear probabilistic generative models estimate Koopman eigenvalues, eigenfunctions, and modes, along with latent variables of unknown dynamical systems, directly handling noisy and partially observed data. These methods provide principled uncertainty quantification and robust performance under finite samples (Masuda et al., 2019, Kawashima et al., 2022).

5.4 Feature Learning and Metric Optimization

Featurized KMD (FKMD) coordinates delay embeddings and learns Mahalanobis metrics tailored to the system's infinitesimal dynamics, optimizing feature representations for high-dimensional and partially observed systems. This approach leverages kernel-based and random Fourier features for scalability, with empirical gains in clustering, forecasting, and noise robustness (Aristoff et al., 2023).

5.5 Connection to Tensor Decomposition

A generic correspondence between KMD and canonical polyadic decomposition (CPD) of third-order tensors is established. Under linear-consistency and rank-conditions on the data, the theoretical KMD and CPD decompositions coincide, providing a bridge between dynamical systems and tensor/optimization communities (Redman, 2021).

6. Limitations, Practical Considerations, and Future Directions

Standard DMD and EDMD assume full-state observability and may generate spurious modes if observables/dictionaries are inadequate. In practice, the choice of kernel, dictionary, embedding, regularization, and feature selection are key to extracting physically meaningful modes and avoiding numerical errors. Recent algorithmic advances target robust extraction (e.g., subspace DMD for noise, Kalman/DMD hybrids for adaptivity), improved conditioning (Koopman-Schur), high-dimensional and partially observed settings (probabilistic and feature-learned KMD), and interpretability (sparsity-promoting formulations).

Open problems include scalable computation in very high dimensions, development of systematic feature/dictionary selection strategies, guarantees for metric-learning iterations, and robust Koopman spectral analysis in systems with significant continuous spectrum or non-normal dynamics.

7. Summary Table: KMD Algorithmic Variants

Algorithm	Core Principle	Robustness/Scalability
DMD	Linear map in state observable, SVD+eig.	Sensitive to noise, low-dim
EDMD	Arbitrary dictionary/lifting	High-dim, dictionary choice
Kernel DMD/EDMD	RKHS embedding via kernels	Data-adaptive, scalable
Subspace DMD	Orthogonal projections for noise/uncertainty	Robust to noise, stochastic
SP-DMD	$\ell_1$ penalty for sparse mode selection	Interpretability, low-order
Koopman-Schur	Unitary Schur, modal subspaces	Numerically stable
Analytic EDMD	RKHS analytic functions, block-triangular	Exactness, structured spectra
EnKF/Kalman DMD	Adaptive filtering for time-varying spectra	Online, non-autonomous
GP-KMD/Probabilistic	GP regression/latent-variable models	Uncertainty quantification
FKMD	Delay-embed, learned Mahalanobis metric	High-dim, partial obs.