Mean-Square Boundedness Under Unbounded Stochastic Disturbances

24 concepts27 papers1 proposal

In classical robust control and Model Predictive Control (MPC), safety guarantees rely on a comforting assumption: environmental disturbances are strictly bounded. If we know a disturbance $w_t$ satisfies a hard geometric bound (such as $\|w_t\| \le W$), we can construct "invariant sets"—regions of the state space where, if the system starts inside, a controller can guarantee it stays inside forever.

But real-world noise rarely respects hard boundaries. Sensor errors, wind gusts, and thermal noise are routinely modeled as Gaussian distributions. Because Gaussian noise has infinite support, there is always a non-zero probability of an arbitrarily large disturbance. A single extreme outlier can push a system completely out of its predefined safe zone. When this happens, traditional controllers experience a catastrophic failure mode: the underlying optimization problem becomes mathematically infeasible, leaving the system with no valid control action.

To deploy control systems safely in the real world, we must move away from deterministic bounds and design frameworks capable of handling unbounded stochastic disturbances.

Redefining Stability: Mean-Square Boundedness

When disturbances are unbounded, demanding that a system never leaves a safe region is a mathematical impossibility. We must change what it means for a system to be "stable."

Instead of requiring the state to remain within a hard geometric boundary, we can require its average energy to remain finite. This is the core intuition behind mean-square boundedness. As explored by Peter Hokayem et al. in Stochastic receding horizon control with output feedback and bounded control inputs, mean-square boundedness ensures that even though continuous random perturbations constantly buffet the system, the state's variance remains trapped within a finite envelope. This approach allows controllers to maintain system stability without requiring the state to reach a perfect zero equilibrium, which is physically impossible under continuous noise.

Hard Invariance (Bounded Noise)          Mean-Square Boundedness (Unbounded Noise)
         +---------------+                                 .

- - - - - - .
         |  Safe Region  |                                (   State Space   )
  o---o--+--> o          |                         o--o--/--------o-------\---> o
         |               |                              (    Variance      )
         +---------------+                               ' - - Envelope - '
   (State never escapes)                            (State escapes, but energy is bounded)

To achieve this stability without violating physical actuator limits, the authors design nonlinear feedback policies that prevent the state from drifting to infinity. Similarly, for reference tracking over unreliable networks, Prabhat K. Mishra et al. demonstrate in Reference tracking stochastic model predictive control over unreliable channels and bounded control actions that mean-square boundedness ensures the average tracking error stays bounded over long operational horizons, even when control signals are dropped. By replacing terminal invariant sets with "constant negative drift" conditions, Debasish Chatterjee et al. prove in Stabilizing Stochastic Predictive Control under Bernoulli Dropouts that stochastic systems can remain stable under unbounded noise even with a non-zero probability of communication failure.

Loosening the Grip: Chance Constraints and Constraint Tightening

If hard constraints are impossible to satisfy under Gaussian noise, how do we express safety? The standard approach is to use chance constraints, which require that constraints are satisfied with a high probability:

$$\mathbb{P}(x_t \in \mathcal{X}) \ge 1 - \epsilon$$

where $\epsilon \in (0, 1)$ is a user-defined risk tolerance.

To solve an MPC problem with chance constraints, we typically use constraint tightening. Instead of planning a trajectory right up to the edge of the physical wall, we artificially shrink the allowable state space based on the noise covariance.

Physical Boundary: |---------------------------------------------------------|
Chance Constraint: |--------------------------------------------------|      |
Planned Trajectory: o --------> o --------> o                                |
                                              \___(Extreme Noise Spike)___/--> [Safe Escape]

Kai Wang et al. show in Robustifying Model Predictive Control of Uncertain Linear Systems with Chance Constraints that by applying the multivariate Chebyshev inequality, complex probabilistic chance constraints can be converted into a standard, easily solvable Quadratic Program (QP) using only the mean and covariance of the noise.

Another powerful tool is the Probabilistic Reachable Set (PRS). As detailed in Stochastic Model Predictive Control for Linear Systems using Probabilistic Reachable Sets by Lukas Hewing et al., a PRS bounds the propagation of unbounded disturbances over a prediction horizon. This allows the controller to optimize performance while guaranteeing closed-loop chance constraint satisfaction, avoiding the extreme conservativeness of worst-case robust control frameworks.

Ensuring Recursive Feasibility via Algorithmic Flexibility

Even with chance constraints, a massive noise spike can still land the system in a state where the next MPC optimization problem has no solution. To guarantee recursive feasibility—meaning that if the optimization is solvable now, it is guaranteed to be solvable at the next time step—researchers have introduced clever architectural modifications to MPC.

Indirect Feedback

Proposed by Lukas Hewing et al. in Recursively Feasible Stochastic Model Predictive Control using Indirect Feedback, this mechanism decouples performance from feasibility. Instead of forcing the predicted trajectory to start exactly at the corrupted, measured state, the constraints are evaluated relative to a purely nominal trajectory. The actual measured state enters only through the cost function, pulling the system back on track without ever causing constraint infeasibility.

Initial State Optimization

Alternatively, Henning Schlüter et al. introduce initial state optimization in Stochastic Model Predictive Control using Initial State Optimization. If a large disturbance pushes the system into an unsafe zone, the controller is allowed to treat the starting point of its optimization trajectory as a decision variable rather than a fixed measurement. By "nudging" the starting point along a mathematically sound line, the optimization absorbs the shock of the unbounded noise and maintains feasibility.

To reconcile these different philosophies, Johannes Köhler et al. developed a unifying framework called Multi-Step Stochastic MPC (MS-SMPC) in On stochastic MPC formulations with closed-loop guarantees: Analysis and a unifying framework, allowing engineers to systematically navigate the trade-offs between robust constraint tightening and indirect feedback performance.

Beyond MPC: Verification and Safe Learning

The challenge of unbounded noise extends far beyond MPC into system identification, formal verification, and machine learning. For instance, Charis J. Stamouli et al. highlight in Rate-Optimal Non-Asymptotics for the Quadratic Prediction Error Method how classical nonlinear system identification methods require sophisticated truncation arguments to handle unbounded input distributions.

In formal verification, Leonard Berrada et al. point out in Make Sure You're Unsure: A Framework for Verifying Probabilistic Specifications that verifying neural networks (like Bayesian Neural Networks) under pure Gaussian noise breaks standard deterministic verification tools, requiring new probabilistic verification frameworks.

Finally, in safe reinforcement learning and robotics, methods like Stochastic Control Barrier Functions (CBFs) are being adapted. As Ryan Kazuo Cosner et al. argue in Robust Safety under Stochastic Uncertainty with Discrete-Time Control Barrier Functions and Ashish Kapoor et al. demonstrate in Sample-efficient Safe Learning for Online Nonlinear Control with Control Barrier Functions, because absolute safety is impossible under untruncated noise, modern robotic frameworks must pivot toward quantifying risk and ensuring high-probability safety boundaries that remain practically achievable in messy, stochastic environments.

Key Foundational Work

Incorporating state estimation into model predictive control and its application to network traffic control (2005) — contributed 1 concept elements to this direction.
On stochastic control up to a hitting time (2009) — contributed 1 concept elements to this direction.
On stochastic receding horizon control with bounded control inputs (2009) — 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_mean_square_boundedness_9abe3f0b,
  title = {mean-square boundedness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mean-square-boundedness/rdr_9abe3f0b797e4d5398be35f6e708e42b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_mean_square_boundedness_9abe3f0b_v1,
  title = {mean-square boundedness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mean-square-boundedness/rdr_9abe3f0b797e4d5398be35f6e708e42b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake