When analyzing dynamical systems, we often seek to linearize their behavior to make them predictable. The Koopman operator framework achieves this by shifting our focus from the nonlinear evolution of state vectors to the linear evolution of observable functions. If a system is simple or quasiperiodic, its behavior can be cleanly decomposed into a discrete set of frequencies—isolated spikes in a power spectrum, corresponding to classical eigenvalues and eigenvectors.
But what happens when a system is chaotic, mixing, or turbulent?
In these complex systems, states spread out and mix over time. This chaotic spreading manifests as a continuous spectrum—a smooth, unbroken distribution of frequencies rather than isolated spikes. As Zhou notes in Spectra of subdivision operators, distinguishing between discrete eigenvalues and a continuous residual spectrum in infinite-dimensional spaces is a profound challenge that can leave the mapping properties of an operator theoretically ambiguous.
Traditional eigenvalue-eigenvector analysis completely breaks down in this regime because there are no true isolated eigenvectors to find. While we can successfully use particle approximations to compute the principal eigenfunctions of simpler non-negative integral kernels (Calculating Principal Eigen-Functions of Non-Negative Integral Kernels: Particle Approximations and Applications), chaotic systems present a fundamental tension. As Peng explores in Continuous Spectrum for a Class of Smooth Mixing CMV Matrices, the temporal mixing of a system complicates its spectral behavior, making it notoriously difficult to determine how continuous components emerge from underlying dynamics.
The Breakdown of Classical Data-Driven Methods
To extract Koopman representations from data, researchers widely rely on Dynamic Mode Decomposition (DMD). Standard DMD forces an infinite-dimensional operator into a finite-dimensional matrix, implicitly assuming the spectrum consists entirely of discrete points. When applied to systems with continuous spectra, this truncation introduces severe numerical and theoretical failures.
1. Numerical Ill-Conditioning
Direct implementations of DMD often rely on companion matrices, which require inverting notoriously ill-conditioned Vandermonde structures. As Mohr et al. demonstrate in Data Driven Koopman Spectral Analysis in Vandermonde-Cauchy Form via the DFT: Numerical Method and Theoretical Insights, this ill-conditioning severely limits the reliability of spectral analysis under standard numerical precision. Minor data inaccuracies explode into massive errors, making the identification of coherent structures highly unstable when solving the underlying least-squares problems (On Least Squares Problems with Certain Vandermonde-Khatri-Rao Structure with Applications to DMD).
2. Spectral Pollution
More fundamentally, forcing a continuous spectrum into a finite matrix causes spectral pollution—the appearance of spurious discrete eigenvalues that are artifacts of the discretization rather than true physical features.
Discrete Spectrum (Quasiperiodic) Continuous Spectrum (Chaotic/Turbulent)
▲ ▲
│ █ │ ▄▆███▆▄
│ █ │ ▄███████▄
│ █ █ │███████████
└──┴───┴────────► └──┴┴┴┴┴┴┴┴┴►
Frequencies Frequencies
(Captured well by DMD) (Polluted/Shattered by DMD)
As Klus et al. point out in Data-driven approximation of the Koopman generator: Model reduction, system identification, and control, while existing methods effectively approximate systems characterized by discrete point spectra, extending them to handle continuous spectra—such as those found in turbulent fluid flows—remains an unresolved challenge. Takeishi echoes this in Kernel Learning for Data-Driven Spectral Analysis of Koopman Operators, noting that standard kernel techniques struggle to represent non-discrete spectral components because they lack a mechanism to map the continuous spread of chaotic attractors.
Consequently, direct numerical approximations of measure-preserving transformations often become highly problematic (On the Approximation of Koopman Spectra for Measure Preserving Transformations), and standard spectral matching techniques fail because they rely too heavily on real, non-zero discrete eigenvalues (On Matching, and Even Rectifying, Dynamical Systems through Koopman Operator Eigenfunctions).
From Eigenvalues to Spectral Measures and Residuals
To accurately resolve a continuous spectrum, we must abandon the hunt for individual eigenvalues and instead compute spectral measures. Think of a spectral measure as a continuous probability density function over the frequency domain, describing how the system's energy is distributed across a continuum.
Computing Measures via Resolvents
In Computing Spectral Measures of Self-Adjoint Operators, Colbrook et al. developed a high-order algorithm to compute these measures directly. Instead of diagonalizing a truncated matrix, the authors leverage the resolvent operator—the infinite-dimensional analogue of a matrix inverse, expressed as:
$$R(z) = (zI - A)^{-1}$$
By evaluating the resolvent near the continuous spectrum and smoothing the result with high-order rational kernels, they avoid spectral pollution entirely, achieving near machine precision for infinite-dimensional operators.
Residual Dynamic Mode Decomposition (ResDMD)
Building on this foundation, Colbrook et al. introduced Residual Dynamic Mode Decomposition (ResDMD) in Rigorous data‐driven computation of spectral properties of Koopman operators for dynamical systems.
The core insight of ResDMD is that while we cannot compute exact eigenvectors for a continuous spectrum, we can compute a spectral residual—a rigorous mathematical bound on how far a candidate function is from the true spectrum. Instead of guessing eigenvalues blindly, ResDMD solves an optimization problem that minimizes this residual. This provides guaranteed convergence and allows for a posteriori verification, ensuring that the continuous spectrum of continuous-time flows is captured accurately without introducing artificial damping or numerical instabilities (On the Approximation of Koopman Spectra of Measure-Preserving Flows).
Deep Learning with Spectral Residuals
This residual-driven philosophy has recently been translated into deep learning architectures to solve complex physics equations (KoopmanLab: Machine learning for solving complex physics equations).
Specifically, Xu et al. developed ResKoopNet: Learning Koopman Representations for Complex Dynamics with Spectral Residuals. Traditional neural networks often act as low-pass filters, smoothing out high-frequency variations as noise and losing global structural physics. ResKoopNet overcomes this by incorporating the spectral residual directly into its loss function:
$$\mathcal{L}_{\text{residual}} = \| \mathcal{K}\phi(x) - \frac{d}{dt}\phi(x) \|$$
By actively minimizing this residual during training, the network is forced to search for and represent the complete continuous spectrum, making it uniquely capable of capturing the intricate dynamics of turbulence and neural signals.
Alternative Framings: Rigged Spaces and Intermittent Forcing
Rather than computing the continuous spectrum directly, some researchers circumvent the problem by altering the mathematical or physical framing of the system.
Rigged Hilbert Spaces
Traditional Hilbert spaces are often too restrictive to contain the "wavefunctions" or states associated with a continuous spectrum. To resolve this, Chiba et al. utilize a Rigged Hilbert Space (or Gelfand triplet) framework in Generalized Eigenvalues of the Perron-Frobenius Operators of Symbolic Dynamical Systems.
By embedding the standard Hilbert space within a larger space of generalized distributions, the authors extend the very definition of eigenvalues to generalized eigenvalues. This mathematical scaffolding allows them to extract explicit information about convergence to equilibrium and mixing rates—properties that are completely obscured when viewed through the lens of standard spectral theory.
┌─────────────────────────────────────────────────────────┐
│ Generalized Vectors / Distributions (Continuous Spectrum)│
│ ┌───────────────────────────────────────────────────┐ │
│ │ Standard Hilbert Space (Discrete Eigenvalues) │ │
│ └───────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────┘
Rigged Hilbert Space
Intermittently Forced Linear Systems (HAVOK)
From a physical perspective, Brunton et al. proposed the HAVOK (Hankel Alternative View of Koopman) framework in Chaos as an intermittently forced linear system.
Instead of trying to represent the continuous spectrum as an uncountably infinite set of frequencies, HAVOK reframes chaos entirely. It decomposes chaotic dynamics into a closed-form, stable linear system driven by intermittent, rare external "kicks."
While HAVOK is highly effective at predicting sudden behavioral shifts (such as heart arrhythmias or magnetic field reversals) from data, the authors note that a formal mathematical bridge between these empirical data-driven forcing events and the theoretical continuous spectra of Koopman operators remains an open and vital area of research.
Key Foundational Work
Ergodic Theory, Dynamic Mode Decomposition, and Computation of Spectral Properties of the Koopman Operator (2017) — contributed 2 concept elements to this direction.
Chaos as an intermittently forced linear system (2017) — contributed 2 concept elements to this direction.
Spectra of subdivision operators (2000) — contributed 1 concept elements to this direction.