Input-to-State Stability for Robust Nonlinear Control

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 280 papers91 conceptsfrozen 2026-10-04

In classical control theory, stability is often a binary state: a system is either stable (it returns to equilibrium) or unstable (it diverges). However, real-world systems—from autonomous underwater helicopters to power grids—never operate in a vacuum. They are constantly buffeted by sensor noise, modeling errors, and environmental disturbances.

Input-to-State Stability (ISS) provides a more nuanced framework. Instead of asking if a system is stable, ISS asks: How much does the state deviate when the input is noisy? It provides a mathematical guarantee that the system's error remains bounded by the magnitude of the disturbance. If the noise is small, the error is small; if the noise vanishes, the system recovers its equilibrium.

The Intuition of Bounded Degradation

The core power of ISS lies in its ability to treat "messiness" as a bounded input. In A New Model of Fringing Capacitance and its Application to the Control of Parallel-Plate Electrostatic Micro Actuators, Guchuan Zhu and Yves-Alain Peter demonstrate this by modeling complex, non-linear edge effects in MEMS actuators as bounded disturbances. By designing a controller that is ISS with respect to these disturbances, they achieve stable control without needing to solve the exact, computationally heavy equations of the physical system.

This "graceful degradation" is what allows engineers to bridge the gap between idealized simulations and noisy reality. As Bo Pang 0002 and Zhong-Ping Jiang show in Robust Reinforcement Learning: A Case Study in Linear Quadratic Regulation, ISS can even be applied to the learning process itself. By treating policy evaluation errors as external disturbances, they prove that reinforcement learning algorithms can converge to a neighborhood of the optimal solution even when the data is noisy.

Bridging Learning and Safety

A recurring challenge in modern robotics is integrating high-performance "black-box" models, like neural networks, into safety-critical control loops. ISS acts as the formal bridge here.

  1. Safety Envelopes: In A Constructive Method for Designing Safe Multirate Controllers for Differentially-Flat Systems, Devansh R. Agrawal and Dimitra Panagou use ISS to define a "safety buffer." By ensuring the tracking error is ISS relative to disturbances, they can mathematically guarantee that a robot will stay within its safe zone even if it encounters unexpected wind or friction.
  2. Adaptive Learning: When a robot learns new dynamics in real-time, there is a risk that imperfect estimates will lead to instability. Siqi Zhou and Angela P. Schoellig address this in Bridging the Model-Reality Gap with Lipschitz Network Adaptation by enforcing Lipschitz constraints on neural networks. This ensures the system's input-to-state gain remains predictable, preventing catastrophic failure during the learning phase.

Contraction and Non-Euclidean Robustness

While traditional ISS often relies on finding a Lyapunov function (a "scalar energy" that decreases over time), recent work has shifted toward Contraction Theory. The intuition here is simpler: if all trajectories of a system naturally "shrink" toward each other, the system is inherently robust.

Alexander Davydov and Francesco Bullo have expanded this in Non-Euclidean Contraction Theory via Semi-Inner Products, showing that contraction in any norm (not just the standard Euclidean distance) implies ISS. This is particularly useful for large-scale networks, like power grids or neural networks, where the "distance" between states might be better measured in $\ell_1$ or $\ell_\infty$ norms.

This framework is now being used to verify the stability of state-of-the-art architectures. For instance, Davide Scaramuzza and Nikola Zubić applied ISS to selective State-Space Models (SSMs) like Mamba in Regularity and Stability Properties of Selective SSMs with Discontinuous Gating, proving that these models remain stable even when their internal parameters change abruptly.

Robust Estimation: Knowing Despite Noise

ISS is not limited to control; it is equally vital for State Estimation. If a robot doesn't know its exact position due to sensor noise, how can it remain stable?

Researchers like Farhad Ghanipoor and Nathan van de Wouw have developed "ultra-local" models that use linear filters to estimate faults in nonlinear systems. By proving these estimators are ISS, as seen in Robust Fault Estimators for Nonlinear Systems: An Ultra-Local Model Design, they ensure that the estimation error stays bounded even when the system is hit by measurement outliers.

Similarly, Wenhan Cao and Wei Pan introduced $\beta$-MHE in Robust Bayesian Inference for Moving Horizon Estimation, using an incremental input/output-to-state stability (i-IOSS) framework. This allows the estimator to "softly" downweight bad data, ensuring the overall system remains stable even when sensors are momentarily compromised.

Summary of the ISS Mechanism

The reason ISS works across these diverse fields—from fluid mechanics to DC-DC converters—is its fundamental framing of the stability problem:

  • Decoupling: It separates the nominal stability of the system from the impact of external noise.
  • Quantification: It provides a specific function (the ISS gain) that tells engineers exactly how much noise the system can tolerate before performance degrades beyond a specific threshold.
  • Convergence: It guarantees that once the disturbance stops, the system will return to its desired state, providing a mathematical "safety net" for complex, nonlinear, and learning-based engineering.

Papers in this version

Showing 60 of 80 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 102 papers
  2. Version 1 · 2026-10-04 · 80 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_input_to_state_stability_f8eeedba,
  title = {input-to-state stability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/input-to-state-stability/rdr_f8eeedbacbb84084820576e4f9c632a3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_input_to_state_stability_for_robust_nonl_f8eeedba_v1,
  title = {Input-to-State Stability for Robust Nonlinear Control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/input-to-state-stability-for-robust-nonlinear-control/rdr_f8eeedbacbb84084820576e4f9c632a3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}