Stabilizing Stochastic Systems Under Bounded Control Authority

33 concepts27 papers1 proposal

Every physical control system operates within limits. Actuators can only exert a finite amount of force, valves have maximum flow rates, and motors have voltage caps. In control theory, this is known as bounded control authority or input saturation. At the same time, real-world systems are constantly buffeted by unpredictable, often unbounded environmental disturbances, typically modeled as stochastic noise.

When designing controllers under these dual constraints—limited power and random disturbances—a sharp mathematical boundary emerges. This boundary is dictated entirely by the system's open-loop stability (how the system behaves when no control is applied). Understanding this boundary explains why certain complex systems can be safely managed with minimal power, while others are fundamentally impossible to stabilize.

The Intuition: Flat Tables vs. Curved Domes

To understand why open-loop stability dictates control feasibility, consider two different physical scenarios under the influence of random, unbounded wind gusts (stochastic noise) and a controller with a strict speed limit (bounded control authority).

1. The Marginally Stable Regime (A Flat Table)

Imagine a ball on a perfectly flat table. If the wind blows, the ball rolls. If the wind stops, the ball stays put. Left alone, repeated random gusts will cause the ball to wander aimlessly via a random walk, and its variance (or mean-square distance from the center) will grow linearly over time. This is a marginally stable (or Lyapunov stable) system: it does not naturally return to zero, but it also does not exponentially explode.

If you try to keep this ball near the center using a hand that has a strict maximum speed limit, can you do it?

The answer is yes. Because the ball's natural dynamics do not accelerate it away from the center, even a tiny, capped amount of control effort is enough to counteract the cumulative drift caused by the wind. As shown by Ramponi et al., linear systems that are marginally stable can maintain a bounded mean-square state even when subjected to unbounded noise and restricted by physical actuator limits.

Furthermore, Ganguly et al. demonstrated that as long as the system satisfies basic reachability conditions, simple saturated feedback controllers are sufficient to ensure stability. You do not need an artificially high minimum control authority to prevent the state from drifting to infinity; a small, bounded effort can successfully fight off unbounded noise over time.

2. The Inherently Unstable Regime (A Curved Dome)

Now, imagine the ball balanced on top of a smooth, curved dome. If a gust of wind pushes the ball slightly off the peak, gravity takes over. The further the ball rolls, the faster it accelerates downward. This is an inherently unstable system, characterized by eigenvalues with positive real parts (or a spectral radius greater than one in discrete time).

If the wind noise is unbounded (e.g., Gaussian noise), there is always a non-zero probability of a massive, sudden gust. If such a gust knocks the ball too far down the slope, its downward acceleration will exceed the maximum acceleration your hand can provide. Because your control authority is strictly capped, you can never catch up. The ball flies off to infinity.

This is a fundamental theoretical boundary. Hokayem et al. proved that in stochastic linear systems where noise is unbounded and at least one eigenvalue is strictly unstable, it is mathematically impossible to maintain mean-square stability using only bounded control inputs. No matter how clever the algorithm, Chatterjee et al. confirmed that if a plant's spectral radius exceeds unity, no bounded control policy can prevent the state from diverging under long-tailed stochastic disturbances.

Open-Loop System Type Natural Behavior Feasibility under Bounded Control & Unbounded Noise
Stable / Marginally Stable (Eigenvalues $\le 0$) Linear drift (Random Walk) Possible. Saturated feedback can guarantee mean-square boundedness (Ramponi et al.).
Inherently Unstable (Eigenvalues $> 0$) Exponential growth Impossible. Large noise shocks outpace the control cap, causing inevitable divergence (Hokayem et al.).

The Open-Loop Stability Constraint in Modern Control

Because stabilizing inherently unstable plants under bounded control and unbounded noise is a mathematical dead end, a vast majority of advanced control frameworks explicitly bake in an open-loop stability assumption. When researchers attempt to optimize for performance, communication efficiency, or computational speed, they must restrict their domain to systems that do not exponentially explode by default.

Model Predictive Control (MPC)

Model Predictive Control (MPC) solves an online optimization problem at each time step to find the best control sequence over a future horizon. When dealing with hard input constraints and noise, proving that the controller will remain stable over time is notoriously difficult.

To make these proofs tractable, frameworks often assume the underlying system matrix is already stable. For instance, the stochastic receding horizon control framework proposed by Chatterjee et al. requires the system matrix to be Schur stable, meaning it is primarily effective for regulating systems that are already stable rather than actively stabilizing unstable ones. Similarly, Hokayem et al. rely on the assumption that the system matrix is Lyapunov stable to extend rigorous guarantees to output feedback settings. When Chatterjee et al. showed that mean-square stability can be guaranteed for any positive control bound, removing the requirement for high control authority, they explicitly noted that this holds provided the underlying system is already Lyapunov stable.

Distributed Systems and Consensus

In multi-agent systems (like drone swarms or power grids), individual agents must reach a consensus via a communication network. If the agents themselves are open-loop unstable, delays and saturation become fatal. As noted by Jiang et al., the requirement for agents to possess open-loop stability severely limits consensus algorithms; current methods cannot reliably stabilize networked agents if their individual state matrices contain eigenvalues with positive real parts.

Similarly, when designing anti-windup structures to handle actuator saturation across large-scale systems, Yu et al. point out that formal stability guarantees remain limited to systems that are open-loop stable or exhibit specific natural decay characteristics.

Data-Driven Modeling and Learning

The constraint also rears its head in system identification, online learning, and adaptive control:

  • Targeted Exploration: To learn a system's dynamics safely, you must inject exploration signals. Venkatasubramanian et al. developed an energy-efficient non-stochastic exploration strategy but limited its scope to inherently stable systems because it relies on open-loop inputs. Applying it to an unstable system would cause the system to diverge during the learning phase itself.
  • Adaptive Economic MPC: Frameworks like the one from Degner et al., which learns system parameters on the fly while optimizing for economic efficiency, must assume the underlying system is open-loop stable to guarantee long-term performance without accidental divergence.
  • Sequence Modeling: Modern machine learning architectures used for predictive control, such as state-space models designed for long-range dependencies, often explicitly enforce stability by restricting their internal eigenvalues to the negative domain. As Li et al. observe, this structural restriction prevents these models from accurately capturing or controlling inherently unstable physical systems common in aerospace engineering and robotics.

Bypassing the Constraint

If a system is inherently unstable, how do engineers actually control it in the real world? To bypass the impossibility barrier, you must alter the core assumptions of the problem:

  1. Pre-Stabilization Layers: Before applying advanced optimization or distributed architectures, a fast, unconstrained, or high-authority local feedback loop is applied to "wrap" the unstable plant, making it appear open-loop stable to the outer-loop controllers. For example, Rey et al. note that unstable systems require pre-stabilization techniques before their structural properties can be effectively exploited by distributed optimization algorithms like ADMM. Similarly, the reduced-order models developed by Alora et al. require an initial stabilization layer before they can be safely used in real-time MPC.
  2. Bounded Noise Assumptions: If the environmental noise is strictly bounded (i.e., there is a guaranteed maximum possible gust of wind), a controller can calculate the exact amount of control authority needed to overpower the worst-case disturbance.
  3. Information-Centric Control: In networked control systems, the ability to stabilize an unstable plant is deeply tied to the communication channel's capacity. Khina et al. demonstrated that by using non-linear analog joint source-channel coding, controllers can achieve higher signal-to-distortion ratios than traditional linear methods, allowing them to stabilize plants with significantly larger unstable eigenvalues over noisy communication links. This aligns with fundamental information-theoretic limits, such as those explored by Saberi et al., which dictate exactly how much information capacity a feedback loop must possess to achieve uniformly bounded stabilization.

Key Foundational Work

On stochastic receding horizon control with bounded control inputs (2009) — contributed 2 concept elements to this direction.
Attaining Mean Square Boundedness of a Marginally Stable Stochastic Linear System With a Bounded Control Input (2010) — contributed 1 concept elements to this direction.
Stochastic receding horizon control with output feedback and bounded control inputs (2010) — contributed 1 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_open_loop_stability_constraint_e1889635,
  title = {open-loop stability constraint},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/open-loop-stability-constraint/rdr_e1889635ec9445e59054d39a32b8400f},
  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_open_loop_stability_constraint_e1889635_v1,
  title = {open-loop stability constraint},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/open-loop-stability-constraint/rdr_e1889635ec9445e59054d39a32b8400f/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