Koopman Eigenfunctions

The universe, from the orbit of a planet to the firing of a neuron, is governed by dynamics that are fundamentally nonlinear. For scientists and engineers, understanding and predicting these complex systems represents a persistent and profound challenge. Traditional methods often involve tracking the system's "state"—a complete description of its configuration—as it traces a complicated, often chaotic, path through time. This direct approach can be computationally intensive and analytically intractable. This article explores a revolutionary change in perspective that sidesteps this complexity: Koopman operator theory. Instead of asking how the state evolves, we ask how properties of the state evolve, uncovering a hidden world of linear structure beneath the nonlinear surface.

This article will guide you through this powerful framework in two main parts. In the "Principles and Mechanisms" chapter, we will introduce the core concepts of the Koopman operator, its magical eigenfunctions, and its information-rich eigenvalues. You will learn how this approach transforms daunting nonlinear problems into manageable linear ones. Following that, the "Applications and Interdisciplinary Connections" chapter will demonstrate the immense practical value of this theory, showcasing how it provides a crystal ball for prediction, a master key for control, and a microscope for revealing the hidden architecture of complex systems across physics, engineering, and biology.

A great art in science and engineering is often finding the right point of view, a perspective from which a tangled mess of phenomena suddenly snaps into a picture of elegant simplicity. When we study how things change—a planet in its orbit, the weather, the concentration of a chemical—we typically follow the "state" of the system, a point in some abstract space that represents its complete configuration at a given moment. We watch this point dance and weave through its space, tracing a complex trajectory. But what if we changed our perspective? What if, instead of tracking the point itself, we tracked some property of the point? This is the revolutionary shift at the heart of Koopman operator theory.

Imagine you are watching a complicated dance. Instead of trying to track the precise position of every dancer (the state), you decide to just measure the average distance between all dancers. This single number is an "observable"—a function of the state. As the dancers move, this value will fluctuate. The ​​Koopman operator​​, denoted $\mathcal{K}$, is simply the rule that tells you how the value of any observable changes from one moment to the next. It lifts our focus from the nonlinear evolution of states in the original space to the evolution of functions on that space.

And why is this a good idea? Because the Koopman operator is always ​​linear​​! It doesn't matter if the underlying dynamics are terrifyingly nonlinear; the operator that evolves the functions of the state behaves in a perfectly linear way. This is a remarkable fact. We have traded the complexity of the dynamics for the complexity of the space of observables, which is often a very good bargain.

Within this vast space of all possible observables, we hunt for special ones called ​​Koopman eigenfunctions​​. An eigenfunction is an observable that, as the system evolves, doesn't change its essential character but merely gets scaled by a constant factor at each time step. Think of it as a special "measurement" of the system that has a particularly simple, predictable evolution. If $\phi(x)$ is an eigenfunction, then its value at the next state, $x_{k+1} = F(x_k)$, is just a multiple of its value at the current state:

$\phi(x_{k+1}) = \phi(F(x_k)) = \mu \phi(x_k)$

The constant factor $\mu$ is the ​​Koopman eigenvalue​​, a number that holds the secret to the observable's evolution. Finding these special functions and their corresponding values is the key to unlocking a system's secrets.

Let's see this magic at work with a simple example. Consider a population $x$ that grows by a factor of $\lambda$ each year, so $x_{k+1} = \lambda x_k$. This is a simple linear system, but it's a perfect playground. Let's invent an observable, say, the square of the population, $g(x) = x^2$. How does it evolve?

$(\mathcal{K}g)(x) = g(\lambda x) = (\lambda x)^2 = \lambda^2 x^2 = \lambda^2 g(x)$

Look at that! The observable $g(x) = x^2$ is an eigenfunction with eigenvalue $\mu = \lambda^2$. More generally, for the observable $g(x) = x^n$, the same logic shows it's an eigenfunction with eigenvalue $\mu = \lambda^n$.

So what? The power comes from what this allows us to do. If we know the value of an eigenfunction $\phi$ at the start, $\phi(x_0)$, then after one step it's $\mu \phi(x_0)$. After two steps, it's $\mu^2 \phi(x_0)$. After $k$ steps, it is simply:

$\phi(x_k) = \mu^k \phi(x_0)$

This is astounding. We can predict the value of this observable at any time in the future just by knowing its initial value and its eigenvalue. We don't need to compute the state $x_1, x_2, \dots, x_k$ along the way. The eigenfunction's evolution has been perfectly linearized. Even for incredibly complex, nonlinear dynamics, if we can find an eigenfunction, its personal timeline is always this simple exponential march, whether in discrete time ($\mu^k$) or continuous time ($\exp(\omega t)$). The entire messy dance of the system is distilled into a simple scaling law for this special observable.

The true beauty of this framework is that the eigenvalues are not just random numbers; they are a code that describes the qualitative nature of the dynamics. By looking at the value of $\mu$, we can diagnose the system's behavior.

• ​​Conservation and Invariance (Eigenvalue $\mu=1$):​​ What if the eigenvalue is exactly 1? Then $\phi(x_k) = 1^k \phi(x_0) = \phi(x_0)$. The observable's value never changes. It's a ​​conserved quantity​​, or an "integral of motion". For a Hamiltonian system, energy is an obvious candidate. Any quantity that stays constant along a system's trajectory is a Koopman eigenfunction with eigenvalue 1. In fact, the number of independent conserved quantities corresponds to the number of independent eigenfunctions with eigenvalue 1. A system that explores its entire state space over time (an "ergodic" system) has only one conserved quantity: the trivial constant function.

• ​​Stability and Decay ($|\mu| \ne 1$):​​ The magnitude of the eigenvalue tells us about growth and decay. Consider again the simple system $x_{k+1} = \lambda x_k$. The observable $g(x)=x$ is an eigenfunction with eigenvalue $\mu = \lambda$. The system's fixed point at the origin is stable if and only if trajectories that start near it approach it, which means $|x_k| \to 0$. This happens if and only if $|\lambda| 1$. So, the stability of the system is directly encoded in the magnitude of its Koopman eigenvalue: the fixed point is stable if $|\mu| 1$. An eigenvalue with magnitude less than 1 signals an observable that decays to zero; one with magnitude greater than 1 signals one that blows up.

• ​​Oscillation and Periodicity ($|\mu|=1, \mu \ne 1$):​​ What about complex eigenvalues? They encode oscillations. Consider a system with a period-2 orbit, where it flips between state $x_a$ and $x_b$. We can construct an observable that is positive at $x_a$ and negative at $x_b$. As the system evolves, this observable's sign flips back and forth. This is the behavior of an eigenfunction with eigenvalue $\mu = -1$. The sequence of values is $\phi_0, -\phi_0, \phi_0, -\phi_0, \dots$. An eigenvalue of $-1$ is the signature of a period-2 behavior. This generalizes beautifully. An eigenvalue of $\mu = \exp(i 2\pi/N)$ corresponds to a period-$N$ cycle. For continuous systems, an eigenvalue $\lambda = \alpha + i\beta$ of the Koopman generator encodes both growth (or decay) via the real part $\alpha$ and rotation via the imaginary part $\beta$. In fact, for linear systems, the Koopman eigenvalues corresponding to linear observables are precisely the eigenvalues of the matrix governing the system, providing a perfect bridge to standard linear systems analysis.

The eigenvalues tell us how an observable changes, but the eigenfunctions themselves tell us what is changing. An eigenfunction $\phi(x)$ can be imagined as painting a landscape over the state space. Every point $x$ is given a "height" $\phi(x)$. The lines of constant height are the ​​level sets​​ of the eigenfunction.

Now, here is another beautiful piece of the puzzle: the dynamics of the system must respect this landscape. When the system evolves from state $x$ to $F(x)$, the new value of the eigenfunction is just $\mu \phi(x)$. This means the state $F(x)$ must lie on the level set corresponding to the value $\mu \phi(x)$. The system's evolution is constrained to hop from one level set to another in a perfectly orderly progression. For example, if we have an eigenfunction $g(x,y)=x^2y$ with eigenvalue $\mu=2$, a point starting anywhere on the curve $x^2y=10$ will be mapped, in one step, to some point on the curve $x^2y=20$. The complicated, swirling flow of the dynamics becomes a simple, rigid shift across the coordinate system defined by the eigenfunction. The eigenfunctions are, in a deep sense, the "natural coordinates" for the dynamics.

A system rarely has just one eigenfunction. It usually has a whole family of them, and their corresponding eigenvalues form the ​​Koopman spectrum​​. This spectrum acts like a fingerprint, uniquely identifying the system's long-term behavior. The character of this spectrum tells us everything.

Based on the structure of this spectrum, we can make a grand classification of all dynamical systems:

1. ​​Pure Point Spectrum:​​ The spectrum consists of isolated points (discrete eigenvalues). This is the signature of regular, orderly, quasi-periodic motion. Think of the planets in the solar system, each orbiting with its own frequency. The system's state never repeats exactly (unless the frequencies are rationally related), but it never wanders off unpredictably either. An observable's value will oscillate forever, like a chord played by a perfect musical instrument. Its autocorrelation, which measures how a signal is related to its past self, will never decay to zero.

2. ​​Continuous Spectrum:​​ The spectrum contains continuous bands of frequencies. This is the signature of ​​chaos​​. In such systems, nearby trajectories diverge exponentially, and any memory of the initial state is eventually shredded and mixed throughout the state space. Think of milk being stirred into coffee. An observable's value will fluctuate unpredictably, like white noise. Its autocorrelation will decay to zero over time, signifying a complete loss of memory of its initial state.

This connection between the spectral theory of linear operators and the qualitative behavior of dynamical systems—from celestial mechanics to chaotic turbulence—is one of the most profound and beautiful results in modern mathematics and physics.

By changing our perspective, we have found a hidden world of linear structure underlying all dynamics. The daunting task of solving nonlinear equations is transformed into a search for special functions—the eigenfunctions—that turn the complex dance of states into a simple, linear march through time. This is the power and the beauty of the Koopman operator.

In our previous discussion, we uncovered a remarkable piece of mathematical alchemy: the Koopman operator. It gives us a special pair of glasses that can look at a wild, nonlinear world and see it as a calm, orderly, linear one. We found that instead of tracking the chaotic dance of a system's state, we can follow the simple, predictable evolution of special functions—the Koopman eigenfunctions. Each of these eigenfunctions evolves with the elegant simplicity of exponential growth or decay, like a pure musical note holding its pitch.

This might sound like a beautiful but abstract mathematical trick. It is not. This transformation from nonlinear to linear is one of the most powerful ideas in the modern study of complex systems. It is not just a new way of describing things; it is a new way of doing things. It provides a crystal ball for prediction, a master key for control, and a microscope for revealing hidden structures. Let us now embark on a journey across the landscape of science and engineering to witness the profound and practical impact of this idea.

The most immediate gift of the Koopman framework is the power of prediction. Imagine you are observing a system—it could be the fluctuating price of a stock, the weather, or a chemical reaction—and its behavior seems impenetrably complex. If, however, you could identify the system's fundamental Koopman eigenfunctions, the future would unfold with stunning clarity.

Any quantity you wish to measure, an "observable" in our language, can be thought of as a chord played from the system's fundamental notes (the eigenfunctions). To predict the future value of your measurement, you simply let each pure note evolve according to its simple exponential law and then combine them back together. This is the essence of Koopman spectral decomposition. For a seemingly complicated nonlinear process, knowing its eigenfunctions and eigenvalues allows you to write down a simple, linear formula for its future state.

"But," you might protest, "what if I don't know the equations governing the system? What if all I have is data?" This is where the magic truly begins. The Koopman framework is not confined to systems where we have perfect knowledge. In fact, its greatest strength lies in its synergy with data. Suppose an experimentalist is studying a synthetic biological circuit that flashes with a rhythmic, fluctuating light. The detailed biochemical reactions might be unknown, but the light intensity can be recorded as a time series. By analyzing this data, one can work backward to estimate the very Koopman eigenvalues that govern the dynamics. This is the principle behind a powerful set of techniques known as ​​Dynamic Mode Decomposition (DMD)​​. By performing what is essentially a sophisticated form of linear regression on the time-series data, we can extract the dominant frequencies and growth/decay rates—the $\lambda$ values—that define the system's behavior.

This transforms the Koopman operator from a theoretical entity into a practical, data-driven tool. We can take snapshots of a system in time, whether it's a fluid flow or a financial market, and construct a finite-dimensional matrix that approximates the infinite-dimensional Koopman operator. The eigenvalues of this matrix then give us an approximate Koopman spectrum, providing a linear model that can predict the system's evolution.

Prediction is a passive act, but science and engineering are often about active intervention. We want to design systems that are stable, that can be controlled, and that perform robustly. Here, too, the Koopman perspective offers profound insights, forging a deep connection with the theory of stability and control.

A central question in engineering is whether a system, when perturbed, will return to its desired state of equilibrium. The classical approach to this question is Lyapunov's method, which involves finding an "energy-like" function that always decreases as the system returns to rest. Finding such a Lyapunov function can be a black art. The Koopman framework provides a more systematic path. In some wonderfully elegant cases, a Lyapunov function can be constructed directly from a Koopman eigenfunction. When this happens, the corresponding Koopman eigenvalue, $\lambda$, gives us far more than a simple "yes" or "no" answer about stability. Its real part, $\text{Re}(\lambda)$, tells us precisely the exponential rate at which the system settles down. This provides a quantitative measure of stability, a number an engineer can design with.

We can even turn this idea around and use a set of Koopman eigenfunctions as building blocks to systematically construct a Lyapunov function, allowing us to certify the stability of a complex nonlinear system.

The connection to control theory runs even deeper. In standard linear control theory, a major goal is "pole placement"—designing a feedback controller that places the eigenvalues of the system's matrix at desired locations in the complex plane to ensure stability and performance. For a linear system, the Koopman eigenvalues for any linear observable are precisely the eigenvalues of the system matrix. This means that when an engineer performs pole placement on a linear system, they are, in fact, directly shaping the spectrum of the Koopman operator. This realization is more than a curiosity; it provides a bridge. It suggests a grander vision for controlling nonlinear systems: perhaps we can design controllers by directly manipulating the Koopman spectrum, a technique that is now an active and exciting frontier of research.

Perhaps the most breathtaking application of Koopman theory is its ability to act as a microscope, revealing the hidden architecture of complex systems. The operator's spectrum is not just a collection of numbers; it is a fingerprint of the dynamics' underlying geometric and organizational structure.

Consider a system undergoing a ​​bifurcation​​, a dramatic qualitative change in behavior, like a silent, steady flame suddenly starting to flicker. A classic example is the Hopf bifurcation, where a stable equilibrium point becomes unstable and gives birth to a stable, oscillating limit cycle. The Koopman spectrum mirrors this transformation perfectly. Before the bifurcation, the dominant eigenvalues have negative real parts, corresponding to trajectories spiraling into the stable point. At the moment of bifurcation, the real parts become zero, and after the bifurcation, the spectrum is dominated by purely imaginary eigenvalues. These eigenvalues, $\pm i\omega$, correspond to the persistent oscillation on the newly formed limit cycle. The system's "vital signs," encoded in the Koopman spectrum, tell the full story of its change in character.

This idea generalizes to discovering structure in unimaginably complex networks. Imagine a vast chemical reaction network inside a living cell. It looks like a hopeless tangle of interactions. However, biological systems are often modular, with groups of chemicals that interact strongly among themselves but only weakly with other groups. How can we discover this hidden modularity from the dynamics? The Koopman operator provides the answer. In such a system, there will be very slow dynamical processes corresponding to the slow leakage between modules. These slow processes are captured by Koopman eigenfunctions whose eigenvalues are very close to 1. The function corresponding to the eigenvalue $\mu=1$ is a constant, representing the final, mixed equilibrium state. But the eigenfunctions for the next largest real eigenvalues—the "slowest" modes of the system—are special. They are nearly constant within each module but change their value between modules. By finding these leading eigenfunctions, we can effectively "paint" the different modules of the network, revealing its hierarchical organization from dynamic data alone. This is a revolutionary tool for systems biology and the study of all complex networks.

The reach of the Koopman operator extends even to systems that are not smooth, that exhibit jumps and impacts. Consider the seemingly simple problem of a bouncing ball. Its motion is punctuated by discrete impacts with the ground. We can define a Koopman operator not for the continuous flow, but for a discrete map from the apex of one bounce to the next. The eigenfunctions of this operator then describe quantities that evolve predictably from bounce to bounce. The total energy, for example, is an observable whose corresponding eigenvalue is directly related to the coefficient of restitution, telling us exactly how energy is dissipated in the long run. This demonstrates the framework's versatility in handling the hybrid continuous-discrete systems that are ubiquitous in robotics and the physical world.

This line of thinking even extends to the infinite-dimensional world of fluid dynamics. The flow of a fluid, governed by partial differential equations like the Burgers' or Navier-Stokes equations, can appear chaotic and turbulent. Yet, the Koopman operator can be defined for these systems. Its application allows for the decomposition of a complex, swirling flow field into a set of coherent structures, or "Koopman modes," each evolving with a simple frequency and growth/decay rate. This provides a powerful new language for understanding and modeling turbulence and transport in fluids.

From the intricate dance of molecules in a cell to the swirling of a galaxy, the universe is fundamentally nonlinear. For centuries, our best tools have been approximations and linearizations. The Koopman operator offers a different path. It does not approximate the nonlinearity; it reframes it. It reveals a hidden linear skeleton beneath the complex flesh of the world. By learning to see through these new eyes, we can predict, control, and, most importantly, understand the complex systems that surround us with a new and profound clarity.