Overview
This research addresses a critical gap in control theory: the stability of continuous-time (CT) plants when controlled by discrete-time (DT) iterative optimization algorithms, such as Model Predictive Control (MPC). In real-world applications, these controllers are often suboptimal because they must provide an output within a fixed sampling interval, frequently before the optimization solver has fully converged. This paper provides a rigorous theoretical framework to prove that such systems can remain stable even when the solver performs only a single iteration per time step, provided the sampling rate is sufficiently high.
The Problem: Sampling and Suboptimality
In modern control, strategies like MPC or Control Barrier Functions (CBF) theoretically solve complex optimization problems at every time step to find an "optimal" control input. However, practical implementations face two major hurdles:
- Sampling: The controller only observes the plant and updates commands at discrete intervals ($T$).
- Suboptimality: Due to limited computational resources, the optimization solver may only run for a few iterations ($n$) before it must output a command, potentially never reaching the "perfect" solution.
The core challenge is understanding how the sampling time $T$ and the number of iterations $n$ jointly affect the stability of the interconnected system. Specifically, the authors ask: can we guarantee stability if we only run a single iteration of our solver?
Methodology: Contraction Theory
Rather than relying on traditional Lyapunov stability analysis, the authors employ Contraction Theory. This approach is particularly powerful for tracking problems and interconnected systems because it focuses on how trajectories evolve relative to one another rather than relative to a fixed equilibrium.
Key steps in their analysis include:
- The Reduced Model (RM): Defining an "ideal" limiting system where sampling is instantaneous ($T \to 0$) and the optimizer converges perfectly ($n \to \infty$).
- Lipschitz Analysis: Using one-sided Lipschitz constants (log-norms) to characterize the "contractivity" (stability) of the continuous-time plant and the discrete-time controller.
- Small-Gain Argument: Deriving conditions under which the interconnection of the plant and the iterative controller remains stable, treating $T$ and $n$ as parameters that induce a "time-scale separation."
Key Insights and Results
The $n=1$ Miracle
One of the most significant findings is that for many systems, solving the MPC problem to full convergence is unnecessary. A single iteration of a gradient-descent-like solver is often sufficient to maintain stability, provided the sampling frequency is high enough.
The $T(n)$ Trade-off
The authors derive an explicit threshold for the sampling period, $T(n)$, and show that it is strictly increasing with $n$. This confirms a practical intuition: if your computational budget is low (low $n$), you must sample very fast (low $T$); if you can afford more iterations, you can tolerate slower sampling rates.
Core Theorem (Theorem 3.6)
This central result establishes that if the Reduced Model is contractive, there exists a sampling period $T(n) > 0$ that guarantees Global Exponential Stability (GES) for the interconnected system. For Linear Time-Invariant (LTI) systems, the authors provide even tighter, more practical bounds on stability.
Practical Implications and Future Work
This work provides a unified framework that allows engineers to design controllers in a continuous-time, "perfect" world and then use the paper's formulas to determine the hardware requirements—specifically the sampling rate and CPU cycles—needed for a stable real-world implementation.
However, some limitations remain:
- Conservatism: The stability bounds based on Lipschitz constants can be conservative; systems may remain stable in practice even if they violate these specific theoretical bounds.
- Hard Constraints: The current analysis focuses on unconstrained or "soft-constrained" optimization. Future research is needed to handle hard state and input constraints where a single iteration might result in an infeasible command.
- Output Feedback: The paper assumes full state measurement. Integrating observers or estimators into this framework is a key area for future development.
Summary updated