Koopman Spectrum

The intricate and often unpredictable behavior of nonlinear dynamical systems, from the weather to financial markets, has long posed a significant challenge to scientists and engineers. Traditional methods that track the evolution of a system's state can quickly become intractable due to their inherent complexity. This article addresses a fundamental question: Is there a way to look at these chaotic systems through a simpler, more structured lens? The answer lies in a revolutionary shift in perspective offered by Koopman operator theory. Instead of wrestling with nonlinear equations, this approach lifts the problem into a new space where the dynamics become linear and manageable.

This article will guide you through this powerful framework. First, under "Principles and Mechanisms," we will explore the core concept of the Koopman operator, demystifying how it transforms nonlinear problems into linear ones and what its spectrum—the collection of its eigenvalues—reveals about a system's fundamental nature, from stable orbits to chaotic mixing. Following that, in "Applications and Interdisciplinary Connections," we will bridge theory and practice, discovering how computational methods like Dynamic Mode Decomposition (DMD) have unlocked the Koopman operator's potential, leading to breakthroughs in control theory, data-driven discovery, and our understanding of chaos itself.

Imagine you're watching a leaf caught in a swirling gust of wind. Its path is a frantic, unpredictable dance—a textbook example of nonlinear dynamics. Trying to predict its exact position from one moment to the next is a notoriously difficult problem. The equations governing its motion are complex, and the slightest change in the initial puff of wind could send it on a completely different journey. This is the traditional viewpoint of dynamics, focused on the trajectory of the state itself.

But what if we asked a different kind of question? Instead of "Where is the leaf?", what if we asked, "How does the temperature at the leaf's location change over time?" or "How does its height above the ground evolve?" We've shifted our focus from the object's state (its position and velocity) to an ​​observable​​—a measurement or property of that state. This subtle shift in perspective, from the state space to a space of functions, is the revolutionary idea at the heart of Koopman operator theory.

The genius of this approach lies in a remarkable mathematical fact. Let's say our system evolves from a state $\mathbf{x}$ at time $t$ to a new state $\mathbf{f}(\mathbf{x})$ at time $t+1$. We can define a "time-evolution machine" for our observables, called the ​​Koopman operator​​, $K$. This operator takes any observable function, $g(\mathbf{x})$, and tells us what that function will be at the next time step. Its definition is astonishingly simple:

$(Kg)(\mathbf{x}) = g(\mathbf{f}(\mathbf{x}))$

In plain English, the new value of our observable is just the old observable function evaluated at the new state of the system. Here's the magic trick: while the evolution function $\mathbf{f}(\mathbf{x})$ can be terribly nonlinear and complicated, the Koopman operator $K$ that acts on the space of all possible observables is always, beautifully, ​​linear​​.

This is a game-changer. By lifting our problem from the messy, nonlinear world of states into the pristine, well-ordered world of linear operators, we gain access to a powerful arsenal of tools from linear algebra and spectral theory, tools that were previously off-limits.

The most powerful of these tools is the concept of ​​eigenfunctions​​ and ​​eigenvalues​​. Think of an eigenfunction as a special, "magic" observable. While most observables change their form in complex ways as the system evolves, an eigenfunction maintains its shape perfectly. It is only stretched or shrunk by a constant factor at each time step. This scaling factor is its eigenvalue, $\lambda$. Mathematically, this relationship is:

$K\phi = \lambda \phi$

where $\phi$ is the eigenfunction and $\lambda$ is the eigenvalue.

Let's make this concrete. Consider a toy system where a value $z$ simply doubles at each step: $z_{k+1} = 2z_k$. What if we choose our observable to be $g(z) = z^3$? Applying the Koopman operator gives:

$(Kg)(z) = g(2z) = (2z)^3 = 8z^3 = 8g(z)$

You see? The function $z^3$ is an eigenfunction of this system's Koopman operator, with an eigenvalue of $\lambda=8$. More generally, for a system $z_{k+1} = \alpha z_k$, the observable $g(z) = z^n$ is an eigenfunction with eigenvalue $\lambda = \alpha^n$. The eigenfunction $z^n$ captures a fundamental "mode" of the system, and its eigenvalue $\alpha^n$ tells us exactly how that mode behaves over time—whether it grows, decays, or oscillates.

This set of all eigenvalues, known as the ​​Koopman spectrum​​, acts like a Rosetta Stone. By reading the spectrum, we can decipher the long-term behavior of the dynamical system without ever having to solve the nonlinear equations of motion directly.

The nature of the eigenvalues tells a rich story about the dynamics.

If an eigenvalue $\lambda$ has a magnitude of one, $|\lambda|=1$, the corresponding eigenfunction's magnitude neither grows nor decays; it persists forever. This is the hallmark of ​​periodic or quasi-periodic motion​​. These eigenvalues typically lie on the unit circle in the complex plane. A spectrum composed entirely of such isolated points is called a ​​pure point spectrum​​.

Conversely, if $|\lambda| < 1$, the eigenfunction decays exponentially to zero. This corresponds to ​​transient behavior​​ or convergence to a stable fixed point. If $|\lambda| > 1$, it signals an unstable system where the observable grows without bound.

For continuous-time systems evolving as $\dot{\mathbf{x}} = \mathbf{F}(\mathbf{x})$, it's often easier to work with the ​​infinitesimal generator​​ of the Koopman operator, denoted $\mathcal{L}$. This operator is like the time derivative of the Koopman operator at $t=0$. The two are beautifully related: if a function $\phi$ is an eigenfunction of the generator with eigenvalue $\nu$ (i.e., $\mathcal{L}\phi = \nu\phi$), then it is also an eigenfunction of the Koopman operator $K_t$ with eigenvalue $\lambda = \exp(\nu t)$. The real part of $\nu$ governs the decay or growth rate, while the imaginary part governs the frequency of oscillation. For instance, in a system spiraling into a fixed point, the eigenvalues of the generator might be $\nu = a \pm ib$, where $a < 0$ represents the rate of decay into the center and $b$ represents the frequency of spiraling.

Imagine a planetary system where planets orbit in perfect, stable cycles. The motion is regular and predictable. This is the world of quasi-periodicity. Such systems have a pure point spectrum.

Consider a simple model of this: a point moving on a 2D torus with constant, incommensurate angular velocities $\omega_1$ and $\omega_2$. The trajectory never repeats but densely fills the surface of the torus over time. The natural observables for this system are the Fourier modes, functions like $\phi_{k_1,k_2}(\theta_1, \theta_2) = \exp(i(k_1\theta_1 + k_2\theta_2))$. Each of these is a Koopman eigenfunction, and their corresponding generator eigenvalues are found to be $\nu = i(k_1\omega_1 + k_2\omega_2)$ for all integers $k_1$ and $k_2$. The spectrum is a discrete set of points on the imaginary axis, a lattice formed by the fundamental frequencies $\omega_1$ and $\omega_2$. The dynamics can be decomposed into a sum of these pure, undying oscillations—the "music of the spheres."

This spectral structure has a profound consequence. For observables in such a system, the correlation with their own past values never truly dies out. The autocorrelation function, which measures this "memory," will oscillate indefinitely without decaying to zero. The system is ​​ergodic​​, meaning it explores its entire space over time, but it is not ​​mixing​​. It remembers its origins through these persistent correlations. A special case of this is the irrational rotation on a circle, whose Koopman operator has eigenvalues $\exp(2\pi i k \alpha)$ for an irrational $\alpha$, forming a dense set on the unit circle but remaining discrete. In stark contrast, a rational rotation is not ergodic, which is revealed in its spectrum by the existence of non-constant functions with an eigenvalue of 1, corresponding to periodic behavior that prevents the system from exploring the whole circle uniformly.

What happens when the system is chaotic, like our swirling leaf? In this case, the dynamics are ​​mixing​​. If you take a drop of ink and put it in a glass of water, a mixing system is one where, after stirring, the ink becomes uniformly distributed throughout the water. Any initial patch of ink loses its identity completely.

In spectral terms, this behavior corresponds to a ​​continuous spectrum​​. Instead of discrete points, the eigenvalues form a continuous smear or band. Why? Think about the autocorrelation function again. For a mixing system, any memory of the initial state is eventually lost. The correlation between an observable now and its value in the distant future must decay to zero.

A signal that fades away to nothing, like an echo, cannot be constructed from a finite sum of pure, non-decaying sine waves. Just as a musical chord is made of discrete frequencies, a decaying sound like a cymbal crash is made of a continuous spread of frequencies. Therefore, the decay of correlations implies that the spectral content of the observable cannot be a set of discrete points; it must be a continuous distribution. This is the spectral signature of chaos. For such systems, the time-averaged memory of an observable, as captured by its autocorrelation, vanishes completely.

Perhaps most surprisingly, a continuous spectrum does not always mean chaos. It can also arise in very simple, non-chaotic systems that are merely converging to a stable state.

Imagine a particle spiraling into the origin. Its distance from the origin, $r$, shrinks exponentially, and its angular velocity, $\dot{\theta}$, depends on this distance (e.g., $\dot{\theta} \propto r^n$). As the particle gets closer to the center, it spirals more and more slowly. If we observe its x-coordinate, $x(t) = r(t)\cos(\theta(t))$, we see a wave whose amplitude is decaying and whose frequency is continuously "chirping" down towards zero.

A signal with a time-varying frequency does not have a single, well-defined frequency. Its energy is spread across a continuous band of frequencies. This "chirp" is another source of a continuous Koopman spectrum. It tells us that the system's "rhythm" is changing over time.

Thus, the Koopman spectrum provides a unified framework. Whether a system is perfectly periodic, chaotically mixing, or simply spiraling to its end, the structure of its spectrum—discrete points, a continuous smear, or a mix of both—lays bare the fundamental nature of its dynamics in the universal language of frequency and decay. It transforms the daunting task of predicting nonlinear trajectories into the elegant problem of understanding a linear operator's spectrum.

After our journey through the principles and mechanisms of the Koopman operator, a natural question arises: Is this elegant mathematical framework just a beautiful theoretical construct, or does it have a tangible impact on the world of science and engineering? You might be tempted to think of it as a clever but esoteric piece of mathematics. Nothing could be further from the truth. The Koopman operator is not a museum piece; it is a master key, a versatile tool that unlocks profound insights across an astonishing range of disciplines. It provides a unified language to describe phenomena as diverse as the stability of a planetary orbit, the design of a flight controller, the rhythm of a beating heart, and the very nature of chaos and time itself.

Let's embark on a tour of these applications. We will see how this single idea—of stepping back from the chaotic dance of state variables to watch the orderly, linear evolution of observables—has revolutionized how we analyze, predict, and control the complex systems that surround us.

For many years, the Koopman operator was a concept of great theoretical beauty but limited practical use. The operator itself is infinite-dimensional, a rather intimidating object to work with directly. How could we possibly get our hands on it for a real-world system, like the turbulent wake behind an airplane or the fluctuating price of a stock? The breakthrough came with the development of data-driven methods, most prominently ​​Dynamic Mode Decomposition (DMD)​​.

Imagine you have a movie of a complex process—a series of snapshots taken at regular time intervals. DMD is an algorithm that looks at pairs of consecutive frames and tries to find the single best linear "movie-stepper" that advances the system from one moment to the next. In essence, it assumes that there is a matrix $A$ such that if $x_k$ is the state of the system at time step $k$, then the state at the next step is approximately $x_{k+1} \approx A x_k$. DMD uses the collected data to compute the best possible version of this matrix $A$.

The magic is that this data-derived matrix $A$ is a finite-dimensional, tangible approximation of the elusive, infinite-dimensional Koopman operator $\mathcal{K}_{\Delta t}$ sampled at the time interval $\Delta t$. The eigenvalues of $A$ (the DMD eigenvalues) approximate the eigenvalues of the Koopman operator, and its eigenvectors (the DMD modes) approximate the Koopman eigenfunctions. This connection is not just an analogy; under ideal conditions, one can prove that the eigenvalues $\mu$ of the DMD matrix are directly related to the eigenvalues $\omega$ of the continuous-time Koopman generator by the simple and beautiful formula $\omega = \frac{1}{\Delta t}\ln(\mu)$.

The primary power of Koopman spectral analysis is decomposition. It allows us to take a bewilderingly complex behavior and break it down into a sum of simple, dynamically pure components. Each component, a Koopman mode, has a simple evolution in time: it just oscillates and/or grows/decays at a constant rate. The spectrum of Koopman eigenvalues is the menu of all these possible behaviors.

A beautiful example is the study of ​​limit cycles​​, which are stable, periodic orbits that appear everywhere in nature. Think of the regular beating of a heart, the chirping of a cricket, the predator-prey cycles in an ecosystem, or the oscillating chemical reactions that produce mesmerizing patterns. A system in a limit cycle is dynamic, yet stable. If you push it slightly off its cycle, it will spiral back. How fast does it spiral back? The Koopman spectrum holds the answer. The spectrum will contain purely imaginary eigenvalues that correspond to the frequency of the oscillation itself, but it will also contain eigenvalues with negative real parts. The largest of these real parts (the one closest to zero) dictates the slowest, and therefore dominant, rate of convergence back to the stable cycle.

Another powerful application is in systems with ​​multiple timescales​​. Most real-world systems are a mix of fast and slow processes. In climate science, you have fast-moving weather fronts and slow-moving ocean currents. In molecular biology, you have the rapid vibrations of atomic bonds and the slow folding of a protein. The Koopman spectrum neatly disentangles these. An eigenvalue $\mu$ of the discrete-time Koopman operator with a magnitude $|\mu|$ close to 1 corresponds to a very slow process, something that persists for a long time. An eigenvalue with a small magnitude corresponds to a process that dies out very quickly. By simply looking at the magnitudes of the Koopman eigenvalues, we can identify and isolate the slow, essential dynamics of a system from the fast, transient details, a crucial step in building simplified and understandable models.

The Koopman perspective is not just for passive observation; it is a powerful tool for active engineering. The field of ​​control theory​​ is all about designing inputs to a system to make it behave in a desired way—stabilizing a rocket, guiding a robot, or regulating the temperature in a chemical reactor.

Traditionally, control theory for linear systems is based on shaping the eigenvalues of the system's state matrix. For a linear system $\dot{x} = Ax$, the eigenvalues of $A$ tell you everything about its stability. If we can add a control input $u$, say $\dot{x} = Ax + Bu$, we can design a feedback law $u = -Kx$ to create a new closed-loop system $\dot{x} = (A - BK)x$. The goal is to choose the gain matrix $K$ to place the eigenvalues of the new matrix $A_{cl} = A-BK$ in stable locations.

From the Koopman viewpoint, this process is beautifully re-contextualized. The eigenvalues of the closed-loop matrix $A_{cl}$ are precisely the point spectrum of the Koopman generator for the controlled system, at least when we consider simple linear observables. Thus, the classic engineering task of "pole placement" is, in this language, the task of sculpting the Koopman spectrum. This reframing opens the door to extending these powerful control ideas to nonlinear systems, where the goal becomes shaping the Koopman spectrum to control the system's long-term behavior.

Armed with DMD, the Koopman operator has become a new kind of scientific instrument, allowing us to peer into the dynamics of systems in ways previously unimaginable. It's a "spectral microscope" for time-series data.

One of its most important roles is in helping us distinguish between different kinds of patterns in data. For decades, a dominant method for extracting patterns from complex data, especially in fields like fluid mechanics, was ​​Proper Orthogonal Decomposition (POD)​​. POD is excellent at finding the shapes or modes that contain the most energy or variance in the data. DMD, on the other hand, finds modes that are dynamically coherent—modes that evolve with a single frequency and growth rate. These are not always the same thing. A system might have a very energetic mode that is just a short-lived, transient puff, while a less energetic mode might represent the persistent, underlying oscillation. POD identifies structures based on statistical dominance, whereas DMD identifies structures based on dynamical significance. DMD gives us the actors, while POD just tells us who is standing in the spotlight.

This ability to find hidden dynamic patterns has some surprising applications. Consider a pseudo-random number generator, the kind used in everything from video games to scientific simulations and cryptography. We expect its output to be unpredictable. But what if the algorithm has a subtle flaw, a hidden, long-period cycle? Such a flaw could be disastrous. Using DMD on the time series produced by the generator, we can search for Koopman eigenvalues very close to the unit circle. Such an eigenvalue is the smoking gun for a hidden periodicity. By converting the eigenvalue's angle to a period, we can unmask the generator's secret rhythm, turning DMD into a tool for a kind of digital forensics.

Perhaps the most profound connections revealed by the Koopman operator are those with the fundamental concepts of chaos theory and statistical mechanics.

A hallmark of a ​​chaotic system​​ is its sensitive dependence on initial conditions: infinitesimally close starting points diverge exponentially fast. The rate of this divergence is quantified by the system's Lyapunov exponents. In some archetypal chaotic systems, such as the famous Arnold's Cat Map, the positive Lyapunov exponent—the very number that defines "how chaotic" the system is—is given directly by the logarithm of a Koopman eigenvalue. More generally, the spectrum of a Koopman operator for a chaotic system is not just a set of discrete points (as for periodic motion) but contains continuous bands. The Koopman spectrum, therefore, carries the very signature of chaos.

This leads us to one of the deepest questions in physics: Why do systems approach thermal equilibrium? Why does cream mix into coffee but never spontaneously unmix? This is the question of the ​​arrow of time​​. The microscopic laws of physics are time-reversible, yet the macroscopic world is not. The bridge between these two worlds is built by concepts from ergodic theory, such as "mixing". A system is mixing if, over time, any initial distribution of states gets stirred so thoroughly that it becomes indistinguishable from a uniform equilibrium state.

This abstract property of mixing is encoded directly in the Koopman spectrum. Systems with discrete Koopman spectra (like a simple pendulum) are not mixing; they remember their initial state in the form of conserved quantities. Systems with a continuous Koopman spectrum, on the other hand, are characteristic of mixing and chaotic dynamics. This continuous spectrum ensures that correlations between the past and present decay over time, allowing the system to "forget" its initial state and evolve toward the most probable macroscopic state—equilibrium. The Koopman operator thus provides a direct link between the spectral properties of the underlying dynamics and the justification for the foundational assumptions of statistical mechanics, such as the principle of equal a priori probabilities for isolated systems in equilibrium.

From a shift in perspective, a world of connections has unfolded. By choosing to watch the functions on a space rather than the points themselves, we have found a framework of stunning power and breadth. It shows us that a hidden linear simplicity underlies the world's nonlinear complexity, providing a unified theme that resonates through engineering, data science, and the deepest foundations of physics.