Model Predictive Control (MPC) is a highly successful paradigm because it transforms a control problem into an optimization problem. At each sampling instance, the controller looks ahead into the future, solves a constrained optimization problem to find the best sequence of control inputs, applies the very first input, and repeats the process at the next time step.
However, this elegant framework faces a harsh reality in real-time applications: optimization takes time. If a physical system operates at a high frequency (such as autonomous racing or medical drug infusion), the optimization solver may not converge to the exact global optimum before the next control action must be taken.
Historically, engineers had to choose between two bad options: let the solver run to convergence and miss the real-time deadline, or terminate the solver early and risk destabilizing the system. As shown by Schildbach et al., utilizing suboptimal approximations or early-exit strategies often invalidates traditional stability and robustness guarantees, leaving systems vulnerable to catastrophic failure (Input-to-state stabilization of feasible model predictive controllers for linear systems).
A rapidly growing body of research addresses this challenge by answering a fundamental question: How can we guarantee closed-loop stability and safety even when the underlying optimization solver is stopped early?
Why Early Termination Breaks Classical Stability
To understand why early termination is dangerous, we must look at how standard MPC proves stability.
In classical MPC, the optimal cost function $V^*(x)$ acts as a Lyapunov function—a mathematical measure of the system's total energy. In an ideal world, if you are at state $x_t$ and move to $x_{t+1}$, the optimal cost is guaranteed to decrease:
$$V^*(x_{t+1}) - V^*(x_t) \le -q(x_t)$$
where $q(x_t)$ is a positive cost associated with the current state. This monotonic decrease proves that the system will safely converge to the desired equilibrium.
When a solver is terminated early, it returns a suboptimal cost $V_{\text{sub}}(x)$ instead of $V^*(x)$. Because the solver hasn't fully converged, $V_{\text{sub}}(x)$ can fluctuate unpredictably. The suboptimality introduces a numerical error that acts as an internal disturbance to the control loop. Without a guaranteed decrease between $V_{\text{sub}}(x_{t+1})$ and $V_{\text{sub}}(x_t)$, the mathematical proof of stability evaporates, and the system can oscillate or diverge (Fedorová et al.; Srikanthan et al.).
Mechanism 1: Stability-Based Stopping Criteria
Traditional optimization solvers use convergence-based stopping criteria, such as checking if the gradient of the cost function is close to zero ($\|\nabla f\| \le \epsilon$). However, in control systems, our primary goal is not finding the absolute mathematical minimum; it is ensuring the physical system does not crash.
This insight led researchers to develop stability-based stopping criteria. Instead of waiting for the gradient to vanish, the solver is stopped the exact moment the current iteration satisfies a Lyapunov descent condition.
Traditional: [Iterate] ---> Is Gradient Zero? ---> [Stop & Apply]
Stability: [Iterate] ---> Is Lyapunov Decreasing? ---> [Stop & Apply] (Much Faster!)
Hall et al. applied this logic to nonlinear MPC for anesthesia control, deriving a theoretical stopping criterion based on Lyapunov theory (Real-time Projected Gradient-based Nonlinear Model Predictive Control with an Application to Anesthesia Control). By stopping the solver as soon as stability was mathematically guaranteed rather than waiting for full convergence, they achieved significant computational savings while maintaining patient safety.
To avoid the real-time overhead of calculating these stability bounds on the fly, Fedorová et al. demonstrated that a minimum iteration threshold can be computed offline (A Generalized Stopping Criterion for Real-Time MPC with Guaranteed Stability). By proving that a fixed, finite number of iterations is sufficient to guarantee asymptotic stability, they reduced real-time computational demand by over 80% with negligible performance loss.
Mechanism 2: Anytime Feasibility and Warm Starts
Stopping a solver early introduces a second, equally dangerous problem: constraint violation. If an optimization algorithm is interrupted mid-calculation, its intermediate solution might command an actuator to exceed its physical limits or plan a path through an obstacle.
To solve this, modern algorithms prioritize anytime feasibility, ensuring that every single iteration of the solver yields a valid, constraint-satisfying trajectory.
Iteration 1: [Feasible Trajectory] <--- Safe to terminate here!
Iteration 2: [Better Feasible Trajectory]
Iteration 3: [Optimal Trajectory]
Numerow et al. leveraged this concept for autonomous racing using a specialized Sequential Quadratic Programming (SQP) algorithm (Inherently robust suboptimal MPC for autonomous racing with anytime feasible SQP). Because the trajectory is feasible at every iteration, the solver can be abruptly interrupted during unexpected computational spikes without risking a crash.
A key enabler for anytime feasibility is the warm-start strategy. Instead of starting the optimization from scratch at each time step, the solver is initialized with a shifted version of the solution from the previous time step.
- Srikanthan et al. proved that combining a warm start with the Alternating Direction Method of Multipliers (ADMM) guarantees both feasibility and asymptotic stability within a fixed, finite number of iterations, even when state and input constraints are active (Closed-loop Analysis of ADMM-based Suboptimal Linear Model Predictive Control).
- In the context of state estimation, Schiller et al. utilized an auxiliary observer to generate a warm-start candidate solution for Moving Horizon Estimation (MHE), proving that the estimator remains robustly stable regardless of how many iterations the solver completes (Robust Stability of Suboptimal Moving Horizon Estimation using an Observer-Based Candidate Solution).
Without these careful initializations and recursive structures, approximating the control problem—whether through value function approximations (Kibalama et al.) or infinite-horizon approximations in System Level Synthesis (Alonso et al.)—can compromise recursive feasibility and leave the system without a valid control action at future steps.
Mechanism 3: Bounded Errors and Geometric Compatibility
What happens if the optimization landscape is non-convex, and the solver gets stuck in a local minimum or a stationary point instead of finding the global optimum?
Westenbroek et al. provided a geometric perspective on this issue, proving that Nonlinear Receding Horizon Control can still achieve exponential stability even when optimization methods only find first-order stationary points (On the Stability of Nonlinear Receding Horizon Control: A Geometric Perspective). The key is ensuring that the chosen cost functions are "compatible" with the system's underlying dynamics, meaning that even local minima point the system in a stabilizing direction.
This principle extends naturally to distributed and networked systems, where multiple independent agents must cooperatively optimize a global objective (such as a power grid or a robot swarm). In these setups, agents communicate iteratively to reach a consensus. Forcing them to communicate until full convergence is impractical due to network latency (Lygeros et al.).
Agent 1 [Local Opt] <-\
---> Limited Iterations ---> Bounded Consensus Error ---> Stable Global System
Agent 2 [Local Opt] <-/
As long as the optimization error can be strictly bounded, the global system can be shown to remain stable.
- Belgioioso et al. proved that distributed MPC can maintain robust stability by spreading the computational and communication load over time, treating the non-converged solution as a "running estimate" (Stability and Robustness of Distributed Suboptimal Model Predictive Control).
- Allibhoy et al. demonstrated a similar principle for data-driven predictive control, showing that network systems successfully converge to a desired equilibrium as long as the approximate solution remains within a specific bound of the true optimum (Data-Driven Distributed Predictive Control via Network Optimization; Data-based Receding Horizon Control of Linear Network Systems).
- Even when dealing with complex multi-agent routing via entropy-regularized optimal transport, Ito et al. showed that truncating the algorithm to a finite number of iterations per time step merely results in bounded transient oscillations, while the underlying physical system remains locally asymptotically stable (Entropic Model Predictive Optimal Transport for Underactuated Linear Systems).
Similarly, when replacing exact online optimization entirely with pre-computed explicit maps—such as polynomial chaos expansions—the resulting numerical approximation errors can be bounded mathematically to ensure safe operation across the entire state space (Lefebvre et al.).
Summary
The paradigm shift in real-time MPC can be summarized as follows: we do not need optimal optimization to achieve optimal control.
By shifting the solver's objective from finding the global minimum to satisfying a Lyapunov descent condition, and by enforcing anytime feasibility through structured warm starts, modern predictive control decouples system safety from solver convergence. This theoretical bridge allows high-speed, safety-critical engineering systems to operate with minimal computational overhead without ever sacrificing mathematical guarantees of stability.
Key Foundational Work
APPLICATION OF OPTIMIZATION ALGORITHMS TO NONLINEAR MPC (2002) — contributed 2 concept elements to this direction.
Input-to-state stabilization of feasible model predictive controllers for linear systems (2011) — contributed 1 concept elements to this direction.