In modern control engineering, Model Predictive Control (MPC) is the gold standard for managing complex systems with strict constraints. It works by solving an optimization problem at every time step to find the best possible control action. However, there is a fundamental tension: optimization takes time, but real-world hardware—like a drone or a chemical reactor—cannot wait.
In practice, engineers often use "early termination," stopping the optimization solver after a fixed number of iterations to meet real-time deadlines. Traditionally, this was a dangerous gamble. Standard stability proofs for MPC assume the solver reaches the global optimum; if it doesn't, the resulting "suboptimal" control action might cause the system to oscillate, violate safety constraints, or even crash.
A growing body of research is bridging this gap by developing mathematical frameworks that guarantee stability and safety even when the optimization is incomplete.
The Feasibility Implies Stability Paradigm
The core intuition behind this research direction is that a controller does not need to be perfect to be safe. In optimization-based control, the cost function often acts as a Lyapunov function—a mathematical representation of the system's "energy." As long as the controller can find a solution that reduces this energy compared to the previous step, the system will remain stable.
Matthias A. Müller and Julian D. Schiller have formalized this in the context of state estimation. In Suboptimal Nonlinear Moving Horizon Estimation, they demonstrate that even a single iteration of an optimizer can be sufficient for robust stability, provided it is paired with a "stable auxiliary observer" that acts as a safety net. This shift from "global optimality" to "recursive feasibility" allows controllers to operate safely in high-speed environments where full convergence is impossible.
Termination Guidelines and Suboptimality Thresholds
If we are going to stop a solver early, we need to know exactly when it is safe to do so. Researchers have developed "termination guidelines" that provide a rigorous stopping criterion based on the current state of the system.
Farhad Farokhi and Iman Shames established that distributed systems can maintain global stability as long as the intermediate solution meets a specific suboptimality threshold relative to previous performance. Similarly, Kristína Fedorová and Michal Kvasnica introduced a generalized stopping criterion that calculates a minimum iteration count offline. By adhering to these bounds, they achieved an 80% reduction in computational effort with negligible loss in control performance.
Computational Governors and Reference Tracking
Another approach to managing the real-time gap is to modify the problem itself so the solver can handle it quickly. Jordan Leung and Ilya Kolmanovsky proposed a computational governor for MPC. Instead of forcing the solver to find a difficult optimal path to a distant target, the governor adjusts the reference command to keep the optimization problem within the solver's "comfort zone"—specifically, near the optimization's central path. This ensures that even an inexact solution remains mathematically valid and safe.
Verifying Unrolled and Learned Controllers
The rise of machine learning has introduced a new way to solve these optimization problems: "unrolling" the solver into a neural network. Yihan Zhou and Na Li developed Learned Quadratic Programming (LQP) controllers. These are physics-inspired networks that mimic the structure of an optimization solver but are trained via reinforcement learning.
The challenge with these learned controllers is that they typically use a fixed, low number of iterations for speed. Zhou and colleagues address this by structuring the network so that its stability can be mathematically verified, ensuring that the "learned" shortcuts do not violate the underlying physical constraints of the system.
Stability in Distributed and Networked Systems
When control is distributed across multiple agents (like a swarm of drones), the problem is compounded by communication limits. Yujia Yang and Chris Manzie have explored how quantization noise and limited data rates force early termination in Distributed MPC (DMPC).
Research by Giuseppe Belgioioso and Florian Dorfler shows that these networked systems can remain robustly stable by maintaining a "running estimate" of the solution. Instead of requiring full convergence at every single time step, the agents spread the computational and communication load over time, allowing the collective system to converge toward the optimal behavior even if individual steps are suboptimal.