Generalizing Provable Control to Nonlinear Systems

3,225 concepts2,781 papers60 proposals

In control theory and reinforcement learning, we often seek a holy grail: controllers that are guaranteed to be safe, stable, and optimal, even when operating in uncertain environments. For decades, researchers have successfully built these guarantees by anchoring their mathematics to a foundational assumption: linearity.

When a system is linear, its behavior scales predictably, and uncertainty can be tightly bounded. However, real-world systems—from quadrotors battling wind gusts to soft robotics and complex chemical plants—are inherently nonlinear. A massive, coordinated effort across control theory and machine learning is currently underway to break free from the linear constraint and extend formal guarantees to nonlinear systems.

The Linearity Assumption as a Safe Harbor

To understand why generalizing to nonlinear systems is difficult, we must first understand why linear systems are so remarkably well-behaved. Consider two core pillars of modern control: the Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC).

In a linear system, the next state $x_{t+1}$ is a predictable linear combination of the current state $x_t$ and the control input $u_t$, typically expressed as $x_{t+1} = Ax_t + Bu_t$. As Benjamin Recht explains in A Tour of Reinforcement Learning: The View from Continuous Control, this linear structure allows us to derive non-asymptotic stability and sample complexity guarantees. If we introduce noise, we can often model it as Gaussian, keeping the system's future state distribution cleanly analytical.

Furthermore, when designing robust controllers that must satisfy strict physical constraints (like avoiding an obstacle or staying under a temperature limit), linearity ensures that our optimization problems remain convex. Convex problems have a single global optimum that can be found rapidly, making them suitable for real-time, safety-critical applications.

Why Data-Driven Control Breaks Without Superposition

A particularly vivid example of the linear constraint appears in the emerging field of purely data-driven control. Frameworks like Data-Enabled Predictive Control (DeePC) attempt to control a system without ever identifying an explicit mathematical model, relying instead on a library of past trajectories.

This approach relies heavily on Willems' Fundamental Lemma, a cornerstone of behavioral systems theory. The lemma states that for a linear time-invariant (LTI) system, any trajectory the system can possibly take can be represented as a linear combination of a single, sufficiently rich past trajectory. This mechanism is central to the work of Julian Berberich et al. in Data-Driven Model Predictive Control with Stability and Robustness Guarantees and Jeremy Coulson et al. in Regularized and Distributionally Robust Data-Enabled Predictive Control.

Linear System (Superposition Holds):
Past Trajectory A:  ───► Output A
Past Trajectory B:  ───► Output B
New Input (A + B):  ───► Output (A + B)  [Predictable]

Nonlinear System (Superposition Fails):
Past Trajectory A:  ───► Output A
Past Trajectory B:  ───► Output B
New Input (A + B):  ───► Chaotic Transition / Discontinuity [Unpredictable]

The moment we move to nonlinear dynamics, the superposition principle vanishes. If you double the control input to a drone, it might not move twice as fast; it might stall or flip. Because past trajectories can no longer be linearly combined to predict the future, the mathematical bridge between collected data and future guarantees collapses. As a result, proving global stability or robust constraint satisfaction becomes an unsolved challenge.

The Geometry of Safety and Uncertainty

In robust MPC, safety is often guaranteed using "tubes." The controller calculates a nominal trajectory and wraps a geometric tube around it. As long as the real system stays inside this tube despite disturbances, it is guaranteed to remain safe.

In a linear system, propagating an uncertainty ellipsoid forward in time is straightforward: the ellipsoid changes size and orientation, but it remains an ellipsoid. This predictability allows single-author work by Lukas Hewing in Recursively Feasible Stochastic Model Predictive Control using Indirect Feedback to guarantee that the control problem remains recursively feasible—meaning if a safe path exists now, a safe path is guaranteed to exist at the next time step.

Linear Propagation:
[Initial Uncertainty] ───► [Predictable Convex Tube] ───► Safe

Nonlinear Propagation:
[Initial Uncertainty] ───► [Warped, Non-Convex Shape] ───► Unsafe / Intractable

In a nonlinear system, the geometry of uncertainty becomes distorted. An initial ball of uncertainty can stretch, warp, or split into non-convex shapes over time. This makes calculating the exact boundaries of the "tube" computationally intractable for real-time systems. As noted by Johannes Köhler et al. in Linear robust adaptive model predictive control: Computational complexity and conservatism, extending these fixed-complexity robust guarantees to nonlinear regimes is exceptionally difficult because modeling how uncertainty transitions through nonlinear functions introduces immense complexity.

Emerging Pathways to Nonlinear Generalization

To bridge this gap, researchers are attacking the problem from several distinct angles, trading off absolute global guarantees for local validity or using learning to handle the non-convexity.

1. Sample-Based Learning and Adjustable Boundaries

Rather than trying to compute exact analytical bounds for nonlinear dynamics, some frameworks use empirical data to safely map out the state space. For instance, Brijen Thananjeyan et al. developed ABC-LMPC: Safe Sample-Based Learning MPC for Stochastic Nonlinear Dynamical Systems with Adjustable Boundary Conditions. By using a database of successful trajectories and dynamically adjusting boundary conditions, they allow a system to safely explore and expand its operating range without requiring a global linear model.

2. Certified Neural Network Controllers

Another approach replaces traditional linear control laws with highly expressive neural networks, but forces the network to comply with physics-based safety certificates during training. He Yin et al. demonstrate this in Imitation Learning with Stability and Safety Guarantees, using Lyapunov stability theory to constrain neural network training. This transforms the certification process into a structured optimization problem, guaranteeing a safe operating region even when the underlying plant dynamics are nonlinear.

3. Nonstochastic and Adversarial Frameworks

When dealing with unknown systems, online learning algorithms attempt to minimize "regret"—the difference between the performance of the online controller and the best ideal controller in hindsight. While Sarah Dean et al. established robust regret bounds for the linear quadratic regulator in Regret Bounds for Robust Adaptive Control of the Linear Quadratic Regulator, extending this to adversarial, non-stochastic settings requires entirely new mechanisms.

As Elad Hazan et al. point out in The Nonstochastic Control Problem, maintaining stability and convexity when state transitions follow complex, non-linear transformations remains a primary theoretical frontier.

Ultimately, the transition from linear to nonlinear systems is not just a matter of adding complexity; it requires rethinking how we represent data, propagate uncertainty, and certify safety when the principle of superposition no longer applies.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_generalization_to_nonlinear_systems_8664bbe3,
  title = {generalization to nonlinear systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generalization-to-nonlinear-systems/rdr_8664bbe3b2a54b2c851cd8c04b467744},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_generalization_to_nonlinear_systems_8664bbe3_v1,
  title = {generalization to nonlinear systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generalization-to-nonlinear-systems/rdr_8664bbe3b2a54b2c851cd8c04b467744/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake