Internal Stability in Feedback Control Systems

16 concepts2 papers1 proposal

In control theory, a system is often judged by its performance: how well it tracks a reference or rejects a disturbance. However, a system that appears to behave well from the outside—maintaining a bounded output for a bounded input—can still be fundamentally broken on the inside. This discrepancy between input-output stability and internal stability is a central challenge in both modern controller learning and classical robust control.

The Hidden Danger of Pole-Zero Cancellations

The most common way internal stability is lost while maintaining performance is through unstable pole-zero cancellations. If a controller is designed to "cancel out" an unstable pole in the plant by placing a zero at the exact same location, the transfer function from input to output may look perfectly stable.

However, this cancellation is a mathematical mirage. In a real physical system, the unstable dynamics still exist within the feedback loop. As Bozdag et al. demonstrate, model-agnostic learning methods are particularly susceptible to this failure. When an algorithm optimizes solely for a performance metric like tracking error, it may inadvertently discover that "canceling" the plant's instability is the mathematically easiest way to minimize the error.

The result is a closed-loop system that achieves good performance for a specific training input but suffers from unbounded internal states. Even a tiny perturbation or a slight mismatch between the model and the real world will cause these internal signals to explode, rendering the system practically unusable. Bozdag et al. argue that simply optimizing performance is insufficient; internal stability must be explicitly constrained, perhaps through noise injection during training or two-step prestabilization.

Preserving Stability under Transformations

While learning-based methods struggle to avoid instability, classical robust control often relies on loop transformations to simplify analysis. For instance, a complex feedback system might be transformed into a simpler structure to apply the small-gain theorem or analyze uncertainties using Linear Fractional Transformations (LFTs).

The risk here is similar: a transformation might simplify the math but hide an internal instability. To address this, Szabó et al. investigate the conditions under which these transformations are "safe." They focus on Möbius transformations, which are widely used to map between different system representations (such as chain scattering and LFT frameworks).

The authors establish that unimodular Möbius transformations are the key to preserving internal stability. If a transformation is unimodular, the stability of the entire interconnected system—including all internal signals—remains unchanged between the original and transformed versions. By providing explicit formulas to relate these loops, Szabó et al. ensure that engineers can manipulate system representations for easier design without accidentally introducing the kind of hidden instabilities warned about in learning contexts.

Bridging Performance and Stability

The shared insight across these perspectives is that internal stability is a structural property, not just a performance outcome.

  1. In Learning: Performance-centric optimization (e.g., minimizing tracking error) is "blind" to internal states, often leading to fragile pole-zero cancellations Bozdag et al..
  2. In Analysis: Mathematical transformations must be carefully chosen (e.g., unimodular Möbius transforms) to ensure that the internal boundedness of signals is not lost during the mapping Szabó et al..

Ultimately, whether designing a controller via gradient descent or analyzing a system via LFTs, the goal is the same: ensuring that the "black box" of the controller does not just produce the right numbers, but maintains a stable equilibrium for every signal within the loop.

Key Foundational Work

Challenges in Model Agnostic Controller Learning for Unstable Systems (2025) — contributed 8 concept elements to this direction.
Internal stability and loop-transformations: an overview on LFTs, Möbius transforms and chain scattering (2017) — contributed 8 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_performance_stability_trade_off_16476802,
  title = {performance-stability trade-off},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/performance-stability-trade-off/rdr_164768026fe041689f3736470edc6569},
  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_performance_stability_trade_off_16476802_v1,
  title = {performance-stability trade-off},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/performance-stability-trade-off/rdr_164768026fe041689f3736470edc6569/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