Most of our best tools for controlling the world—from PID loops to Kalman filters—rely on a fundamental assumption: that the system we are managing is linear. Unfortunately, the real world is stubbornly nonlinear. Whether it is the turbulent flow of air over a wing, the jerky gait of a bipedal robot, or the complex switching of an electrical drive, nonlinearity usually forces engineers to choose between oversimplified linear approximations or computationally expensive nonlinear simulations.
A growing body of research suggests a third way. By using Koopman operator theory, researchers are finding that we can represent complex nonlinear dynamics as linear systems, provided we look at them through the right "lens."
The Intuition of Lifting
The core insight of the Koopman operator is a shift in perspective. Instead of tracking the state of a system (like the position and velocity of a pendulum), we track the evolution of observables—functions of that state.
In the original state space, a nonlinear system might evolve along a complex, curved manifold. However, J. Nathan Kutz, Joshua L. Proctor, and Steven L. Brunton explain in Koopman Theory for Partial Differential Equations that there exists an infinite-dimensional linear operator (the Koopman operator) that governs how these observables evolve.
Think of it like a shadow. A 3D object rotating in space might create a complex, nonlinear transformation of its shadow on a 2D wall. But if you "lift" your perspective back to the 3D space, the rotation is a simple linear operation. By "lifting" our state into a higher-dimensional space of observables—using polynomials, sinusoids, or thin-plate splines—we can find a coordinate system where the dynamics look linear.
From Data to Matrices
While the Koopman operator is theoretically infinite-dimensional, we can approximate it using finite-dimensional matrices derived directly from data. The primary tool for this is Dynamic Mode Decomposition (DMD) and its variants.
As Shaowu Pan, Eurika Kaiser, and colleagues describe in their work on PyKoopman, the process involves taking snapshots of a system's state over time. By performing a regression—essentially finding the best linear fit that maps the state at time $t$ to the state at time $t+1$ in the "lifted" feature space—we can identify the system's dynamics without ever writing down a physical equation.
This data-driven approach is powerful because it decouples the complexity of the physics from the cost of the computation. For example, Enoch Yeung and Nathan O. Hodas demonstrate in A Koopman Operator Approach for Computing and Balancing Gramians that once this linear surrogate is learned, we can apply classical control tools like balanced truncation to simplify nonlinear systems that were previously intractable.
Real-Time Control and Robotics
The most immediate benefit of this linearization is speed. Nonlinear Model Predictive Control (MPC) is notoriously slow because it requires solving a difficult optimization problem at every time step. If the model is linear, however, the optimization becomes a simple quadratic program that can be solved in milliseconds.
In Koopman Operators in Robot Learning, Lu Shi, Ian Abraham, and their co-authors highlight how this enables agile control for soft robots and quadrotors. Similarly, Sehoon Ha and Harish Ravichandar show in Learning Koopman Dynamics for Safe Legged Locomotion that "lifting" the dynamics of a bipedal robot allows for real-time obstacle avoidance that combines the agility of reinforcement learning with the safety guarantees of linear control.
This speedup isn't limited to robotics. Sebastian Peitz has applied these techniques to electrical drives and fluid dynamics, achieving simulation speeds up to a million times faster than traditional PDE solvers while maintaining high fidelity.
Expanding the Horizon
The utility of the Koopman framework is extending into unexpected domains:
- Quantum Systems: Anna Hunstig and Sebastian Peitz used Koopman surrogates to accelerate optical quantum system simulations by a factor of 1,000.
- Cryptography: Robin Strässer and Frank Allgöwer have begun treating cryptosystems as dynamical systems, using Extended DMD to "decrypt" the underlying linear structure of nonlinear encryption functions.
- Network Science: Elvin Isufi and Geert Leus leverage subspace identification—a close relative of Koopman theory—to recover the hidden topology of networks by observing how signals propagate across them.
By treating the world not as a set of rigid equations, but as a stream of data that can be "lifted" into a linear form, this research direction provides a unified language for understanding and controlling the complexity of the modern world.