The Koopman Operator: A Linear Perspective on Nonlinear Dynamics

For centuries, the study of change—from the orbit of planets to the flow of air over a wing—has been dominated by the direct analysis of equations of motion. This state-space approach, while foundational, often leads us into the tangled complexities of nonlinear dynamics, where behavior can be chaotic, unpredictable, and difficult to analyze. The inherent difficulty of solving these nonlinear problems creates a significant knowledge gap, limiting our ability to predict, control, and understand many natural and engineered systems. What if there was a way to sidestep this complexity by changing our perspective entirely?

This article explores the Koopman operator, a revolutionary mathematical framework that achieves just that. Instead of tracking how the state of a system evolves, the Koopman operator focuses on how measurements, or "observables," of the system evolve. This shift in viewpoint transforms the difficult nonlinear problem in state space into an elegant and solvable linear problem in an infinite-dimensional function space. By leveraging the powerful tools of linear spectral theory, we gain unprecedented insight into the heart of nonlinear behavior.

The following chapters will guide you through this powerful paradigm. In ​​Principles and Mechanisms​​, we will delve into the core theory, exploring how the operator is defined, what its spectrum reveals about system properties like ergodicity and mixing, and how it can be approximated from data. Then, in ​​Applications and Interdisciplinary Connections​​, we will witness the theory in action, seeing how it revolutionizes fields from fluid dynamics and stability analysis to machine learning and even quantum mechanics, providing a unified language to decode complexity across science and engineering.

For centuries, we have understood dynamics by watching things move. A planet orbiting the sun, a pendulum swinging, a fluid swirling—we write down equations of motion, like Newton's famous $F=ma$, and we track the position and velocity of objects through time. This is the world of state space, a world that can be maddeningly complex. The rules governing the motion might be simple, but the trajectories they produce can be tangled, chaotic, and seemingly unpredictable.

What if there were another way? A way to trade this thorny, nonlinear world for a simpler, more elegant one? This is the revolutionary idea behind the Koopman operator. It’s a paradigm shift. Instead of watching the ball, we decide to watch a measurement on the ball.

Imagine a complex system—say, the weather. The "state" of the system is the temperature, pressure, and wind velocity at every single point in the atmosphere. Tracking every point is an impossible task. But we can ask a simpler question: what is the temperature at Chicago's O'Hare airport? This is a function, an ​​observable​​, that takes the enormously complex state of the atmosphere and returns a single number.

The Koopman operator describes how the values of these observables evolve in time. Let's say our system's state at time $t$, which started at an initial state $x_0$, is given by a (typically nonlinear) flow map $\Phi^t(x_0)$. And let's say we are interested in some observable, a function $g(x)$. The Koopman operator, which we'll call $\mathcal{K}^t$, is defined by a beautifully simple act of composition:

$(\mathcal{K}^t g)(x) = g(\Phi^t(x))$

In plain English: to find the value of the "evolved observable" $\mathcal{K}^t g$ at point $x$, you first let the system run forward for time $t$ to find the future state $\Phi^t(x)$, and then you evaluate your original observable function $g$ at that new state. You are pulling the value of the observable from the future back to the present.

Here's the magic. Even if the underlying dynamics $\Phi^t$ are horribly nonlinear—think of the turbulent flow of a river—the Koopman operator $\mathcal{K}^t$ itself is perfectly ​​linear​​. This means that the Koopman evolution of a sum of two observables is just the sum of their individual evolutions: $\mathcal{K}^t(g_1 + g_2) = \mathcal{K}^t g_1 + \mathcal{K}^t g_2$. We have traded a finite-dimensional, nonlinear problem for an infinite-dimensional, linear one. This might sound like a bad deal—infinite is bigger than finite, after all!—but it is a spectacular one, because humanity has an immense and powerful toolbox for understanding linear systems: the theory of spectra.

Because the Koopman operator is linear, we can ask about its ​​eigenfunctions​​ and ​​eigenvalues​​. An eigenfunction is a special observable, let's call it $\phi(x)$, that doesn't really change its "shape" as the system evolves; it just gets multiplied by a number, its eigenvalue $\lambda$, at each step in time. For a discrete-time map $T$, this relationship is:

$(U_T \phi)(x) = \phi(T(x)) = \lambda \phi(x)$

These eigenfunctions are, in a profound sense, the "natural coordinates" of the dynamics. If we can decompose any arbitrary observable into a sum of these eigenfunctions, we can predict its evolution with remarkable ease. To evolve the entire function, we just have to multiply each eigenfunction component by its corresponding eigenvalue. The chaotic dance is revealed to be a symphony of simple, oscillating modes.

Consider a simple toy system: a point moving around a circle. The transformation is just a rotation by an angle $a$: $T_a(x) = (x+a) \bmod 2\pi$. What are the natural coordinates for rotation? The answer has been known for centuries: the complex exponentials, $\psi_k(x) = \exp(ikx)$. Let's see if they are Koopman eigenfunctions:

$(U_{T_a} \psi_k)(x) = \psi_k(x+a) = \exp(ik(x+a)) = \exp(ika) \exp(ikx) = \exp(ika) \psi_k(x)$

They are indeed! The eigenfunction $\psi_k(x)$ evolves by being multiplied by the eigenvalue $\lambda_k = \exp(ika)$. The complicated motion of rotation becomes simple multiplication in the "frequency domain" of these eigenfunctions.

This even works for nonlinear systems. Consider a particle whose position obeys the rule $\frac{dx}{dt} = x^2$. This is a nonlinear equation, and its solution, the flow map, is $\Phi^t(x) = \frac{x}{1-tx}$. While finding the eigenfunctions analytically can be hard, we can still see the Koopman operator in action. Its effect on an observable like $f(x) = \exp(-\alpha x)$ is to transform it into an entirely new function, $\exp\left(-\frac{\alpha x}{1-tx}\right)$. The nonlinearity of the state's evolution is encoded in the complex functional form of the evolved observable.

The true power of the Koopman framework is that the collection of all its eigenvalues—its ​​spectrum​​—acts as a fingerprint, a Rosetta Stone that allows us to translate spectral properties into the language of dynamical behavior.

The Anchor Point: Eigenvalue $\lambda=1$

What does an eigenvalue of 1 mean? It means $\phi(T(x)) = \phi(x)$. The value of the observable $\phi$ is unchanged by the dynamics. It is a ​​conserved quantity​​, or an ​​invariant​​. For any system, there is one obvious invariant: a constant function. If $f(x) = c$, then of course $f(T(x)) = c$. So, constant functions are always an eigenfunction with eigenvalue 1. This is the "trivial" invariant.

The crucial question is: are there any other, non-constant eigenfunctions with eigenvalue 1? The answer to this question tells us if the system is ​​ergodic​​.

Ergodicity: The "Well-Stirred" System

An ergodic system is one that, given enough time, will explore every nook and cranny of its available state space. A single drop of cream in a cup of coffee will, with sufficient stirring, eventually visit the neighborhood of every water molecule. The system is "well-stirred." In such a system, the only conserved quantities are the trivial ones—constants. Therefore, ​​a system is ergodic if and only if the only eigenfunction of its Koopman operator for eigenvalue $\lambda=1$ is a constant function.​​

Let's see this in action. Imagine a game of musical chairs with 12 seats, where the rule for moving is a fixed permutation $T = (1\ 5\ 9)(2\ 11)(3\ 4\ 7\ 12\ 8)(6\ 10)$. A person starting in seat 1 is confined to the cycle of seats $\{1, 5, 9\}$. They will never reach seat 2. The system is not well-stirred; it is broken into four disjoint "universes". We can define a non-constant observable that is 1 for seats in the first cycle and 0 elsewhere. This function is an invariant, an eigenfunction with eigenvalue 1. Since there are four such cycles, we can construct four independent, non-constant eigenfunctions for $\lambda=1$. The system is not ergodic. To be ergodic, the permutation would need to be a single, grand cycle involving all 12 seats.

This same principle distinguishes rational and irrational rotations on a circle. For a rational rotation, $T(x) = (x + \frac{p}{q}) \bmod 1$, a point returns to its starting position after $q$ steps. The function $\phi(x) = \exp(2\pi i q x)$ is non-constant, yet it is an eigenfunction with eigenvalue $\lambda=1$. Therefore, the rational rotation is not ergodic. For an irrational rotation, however, a point never exactly returns. The only way $\exp(2\pi i k \alpha)$ can be 1 is if $k=0$, which corresponds to the constant function. The system is ergodic.

Mixing: The Fading of Memory

Ergodicity is just the first level of complexity. Some systems are more "chaotic" than others. A stronger property is ​​mixing​​. In a mixing system, any initial blob of states not only visits all regions, but it gets stretched and thinned out so completely that it eventually becomes uniformly distributed throughout the entire space. It represents a a true "decay of memory."

This, too, has a spectral fingerprint.

• An ergodic but not mixing system (like the irrational circle rotation) still has a sense of memory. Its Koopman spectrum has many eigenvalues other than 1, but they all have magnitude 1 (e.g., $\exp(2\pi i k \alpha)$). The system doesn't "forget" its phase. This type of spectrum, made up of isolated points, is called a ​​pure point spectrum​​.
• A truly mixing system is one that forgets. Its "memory" of an initial state fades over time. This corresponds to the Koopman spectrum being ​​continuous​​ (on the space of zero-mean observables). A continuous spectrum means there are no true eigenfunctions, only "almost eigenfunctions" that dissipate over time. This decay is observable in the system's statistics. For any physical measurement with its mean subtracted, its correlation with its own past will decay to zero on average.

Chaotic systems like the famous "Arnold's Cat Map" are prime examples of mixing systems. Their Koopman operator doesn't have a neat picket fence of eigenvalues; instead, its spectrum is a continuous smear, often covering the entire unit circle in the complex plane. This spectral difference is fundamental. A system with a pure point spectrum, like a simple rotation, has a "clockwork" predictability to it and has zero metric entropy (a measure of chaos). A system with a continuous spectrum, like the Cat Map or a Bernoulli shift, is genuinely chaotic and has positive entropy. The Koopman spectrum provides the ultimate classification tool.

This is all wonderfully elegant, but what if we don't know the exact equations of motion for our system? What if we just have data—snapshots of the system taken at discrete intervals? Can we still find this ghostly Koopman operator and its spectrum?

The astonishing answer is yes, approximately. This is the goal of a powerful modern technique called ​​Dynamic Mode Decomposition (DMD)​​. Given a sequence of data snapshots $x_0, x_1, x_2, \dots$, DMD seeks to find a single linear matrix $A$ that best approximates the evolution, such that $x_{k+1} \approx A x_k$.

Here is the profound connection: this data-driven matrix $A$ is a finite-dimensional projection of the true, infinite-dimensional Koopman operator. Its eigenvalues approximate the dominant Koopman eigenvalues, and its eigenvectors (the "DMD modes") approximate the dominant Koopman eigenfunctions.

DMD bridges the gap between abstract theory and messy reality. It allows us, armed only with data, to extract the fundamental frequencies, growth rates, and dominant spatial patterns of hugely complex nonlinear systems—from fluid dynamics and climate science to neuroscience and financial markets. It is the practical realization of the Koopman dream: to find the hidden linear skeleton that underlies the nonlinear flesh of the world, and in doing so, to make the complex a little more simple.

So, we have this marvelous mathematical machine—the Koopman operator. By shifting our viewpoint from the chaotic dance of system states to the calm, linear evolution of observable functions, we've promised to tame nonlinearity. It's a beautiful idea, elegant and clean. But a physicist, an engineer, or just a curious person should rightly ask: What is it good for? Does this abstract art hang in a gallery of pure mathematics, or is it a practical tool we can use to understand and shape the world around us?

The wonderful answer is that it is profoundly useful. The Koopman perspective is not just a theoretical curiosity; it's a powerful lens that is revolutionizing how we analyze complex systems across a staggering range of disciplines. Let's take a tour and see this operator in action, moving from painting portraits of fluid flows to designing stable systems and even peering into a connection between classical chaos and quantum mechanics.

Perhaps the most immediate and impactful application of the Koopman operator is in making sense of complex data. The universe is constantly bombarding us with information—the shimmering wake behind a boat, the fluctuating price of a stock, the intricate concert of proteins in a living cell. These are all nonlinear dynamical systems, and we often only have snapshots of their behavior. How can we find the underlying rules?

​​From Data to Dynamics​​

The Koopman operator itself is an infinite-dimensional beast, which sounds rather intimidating to work with. But the key insight is that for many systems, the most important dynamics—the dominant patterns, the main rhythms—can be captured by a finite set of observables. This is the foundation of a powerful data-driven technique called ​​Dynamic Mode Decomposition (DMD)​​.

Imagine you're a fluid dynamicist studying the flow of air over a wing. You can't track every single molecule, but you can take high-speed video, capturing snapshots of the pressure or velocity field. DMD takes this sequence of snapshots and seeks to find a single linear operator, a matrix, that best advances the system from one snapshot to the next. The magic is that this data-derived matrix is nothing less than a finite-dimensional approximation of the true Koopman operator!

The eigenvalues of this matrix then reveal the fundamental "notes" of the flow: their imaginary parts give you the frequencies of oscillation (the vortices shedding at a certain rhythm), and their real parts tell you if these patterns are growing (instability), decaying (damping), or stable. Incredibly, under the right conditions where the dynamics are confined to a finite-dimensional subspace of observables, the eigenvalues found by DMD are precisely the eigenvalues of the Koopman operator. We have, in essence, used raw data to construct a linear "portrait" of the underlying nonlinear dynamics. This technique is now a workhorse in fields from fluid mechanics and plasma physics to epidemiology and video processing.

​​Revealing the Hidden Architecture​​

The Koopman framework does more than just extract frequencies; the eigenfunctions themselves tell a deep story about the system's structure. Think of a large chemical factory, or better yet, the metabolic network inside a living cell. These are vast networks of reactions. Some processes happen very quickly within a specific functional unit—a "module"—while the exchange of materials between modules is much slower. The system has a hidden modular architecture.

How can we find these modules from observing the overall system? We look at the Koopman eigenfunctions. An eigenfunction associated with an eigenvalue $\lambda=0$ represents a conserved quantity—something that never changes. For a system with disconnected parts, a function that is constant on each part is a conserved quantity. Now, what happens if we weakly connect these parts? Perturbation theory tells us a beautiful story: the degenerate $\lambda=0$ eigenvalue splits. One eigenvalue remains at zero, corresponding to the global constant observable. But new eigenvalues appear with very small real parts.

The eigenfunctions corresponding to these near-zero eigenvalues are the key. They are no longer perfectly piecewise constant, but they are almost so. Each of these "slow" eigenfunctions is nearly constant inside one of the modules and rapidly changes its value only in the narrow regions between them. By finding these slow eigenfunctions, we can literally draw the boundaries of the system's functional modules. We are using dynamics to reveal geography.

This idea goes even deeper. In a chemical reaction, molecules navigate a high-dimensional energy landscape to get from reactants to products. The central question is: what is the "reaction coordinate" that truly measures progress along this path? Transition Path Theory gives an answer with a special function called the committor, which gives the probability that a molecule at a given state will reach the product state before returning to the reactant state. Finding this committor is typically very hard. Yet, for many systems, this crucial function turns out to be almost perfectly approximated by the "slowest" nontrivial Koopman eigenfunction—the one with the largest eigenvalue less than one. The Koopman operator, once again, provides a direct path to the heart of the system's dynamics.

For over a century, the gold standard for proving that a system is stable has been Lyapunov's second method. The idea is to find a function, a "Lyapunov function" $V(\mathbf{x})$, that acts like an energy bowl. If the system's state is always rolling downhill into the bottom of the bowl (i.e., its time derivative $\dot{V}$ is always negative), then the bottom of the bowl must be a stable equilibrium. The genius of this method is that you don't need to solve the equations of motion. The crippling difficulty is that there's no systematic way to find the function $V(\mathbf{x})$; it's often a matter of artistry and inspired guesswork.

Here, the Koopman operator provides a stunningly elegant new perspective. What if we guess that our Lyapunov function $V(\mathbf{x})$ is also a Koopman eigenfunction $\phi(\mathbf{x})$? The definition of an eigenfunction is that its time derivative is just $\dot{\phi} = \lambda \phi$. So, if we can find a Koopman eigenfunction that is also a valid Lyapunov function (positive definite), then its time derivative is trivially calculated! The stability of the system is then directly determined by the sign of the real part of the Koopman eigenvalue $\lambda$. A negative real part means exponential decay to the equilibrium, and the value of $\text{Re}(\lambda)$ gives the exact exponential decay rate.

This profound connection means we can shift the search for a mysterious Lyapunov function to a more structured search for the eigenfunctions of a linear operator. We can even construct explicit Lyapunov functions by taking quadratic combinations of Koopman eigenfunctions, turning stability analysis from a black art into a systematic procedure.

This applies not just to static fixed points but to dynamic attractors as well. Consider a system with a stable limit cycle, like a self-sustaining chemical oscillator or the steady orbit of a satellite. Trajectories that start near this cycle spiral towards it. How fast do they approach? Again, the Koopman operator has the answer. There will be a set of Koopman eigenvalues whose real parts are negative, and the "leading" eigenvalue—the one with the real part closest to zero—governs the long-term rate of convergence to the cycle. The entire geometry of attraction is encoded in the operator's spectrum.

The unifying power of the Koopman framework becomes most apparent when we see it building bridges to the most modern and exciting areas of science and technology.

​​Koopman, Meet Machine Learning​​

At its heart, much of modern machine learning is about representation learning—finding a transformation of data into a new space where problems become easier to solve. For example, the famous "kernel trick" lifts data into a higher-dimensional space to make it linearly separable. This sounds suspiciously familiar, doesn't it?

Extended Dynamic Mode Decomposition (EDMD) does exactly this: it lifts the state into a higher-dimensional space of observables (for example, polynomials) in the hope that the dynamics become linear in that space. This idea is now being supercharged with the power of deep learning. Advanced architectures like neural state-space models can be interpreted as simultaneously learning an encoder (a map to a latent space) and a linear dynamic model within that space. In essence, these networks are being trained to find a finite-dimensional Koopman-invariant subspace! The theory of the Koopman operator provides a rigorous foundation for these methods, explaining why they work and guiding the development of more powerful and reliable AI for modeling and controlling the physical world.

​​A Quantum Window into Classical Chaos​​

Perhaps the most mind-bending connection is one that links classical chaos and quantum computation. For a system that preserves energy or volume (like Hamiltonian mechanics), the Koopman operator is a unitary operator. In quantum mechanics, the evolution of a quantum state is also described by a unitary operator. This parallelism is not a mere coincidence; it's a deep structural link we can exploit.

It means we can simulate the Koopman operator of a classical system on a quantum computer. Imagine we take a famous chaotic system, like the logistic map, which is a simple model for population dynamics. We can encode an observable of this classical system as a quantum state and apply a quantum algorithm like Phase Estimation to the corresponding Koopman operator. The algorithm's output is a measurement of the operator's spectrum.

What does this quantum measurement tell us about the classical world? It tells us about the system's rate of "memory loss," or the decay of its time-autocorrelation function. For many chaotic systems, this decay follows a power law. The exponent of this power law is directly related to the spectral properties of the Koopman operator near zero frequency. A quantum computer, by measuring the Koopman spectrum, can therefore determine this classical exponent. We are using a quantum device to probe the statistical mechanics of a classical chaotic system. It's a beautiful, surprising bridge between two of the great pillars of 20th-century physics.

From practical engineering to fundamental physics, the Koopman operator offers a unified language. By daring to look at a system not through the evolution of its state, but through the evolution of all the things we could possibly measure about it, we find an astonishing reward: the tangled, nonlinear mess untangles into the clean, predictable world of linear algebra. The lesson, as so often in science, is that finding the right point of view can make all the difference, revealing a hidden simplicity and unity in a world that at first seems overwhelmingly complex.