🔢 mathematics

Certified Model Predictive Control for Switched Evolution Equations using Model Order Reduction

This paper presents a certified Model Predictive Control framework for linear switched parabolic PDEs that utilizes Galerkin reduced-order modeling to create computationally efficient surrogates while providing explicit, recursive a-posteriori error estimates to guarantee the closed-loop trajectory remains within a controllable neighborhood of the true solution.

Original authors: Michael Kartmann, Mattia Manucci, Benjamin Unger, Stefan Volkwein

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Michael Kartmann, Mattia Manucci, Benjamin Unger, Stefan Volkwein

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive, complex ship navigating through a storm. The ship's behavior changes depending on whether a door between two rooms is open or closed (like a switch). Your goal is to steer the ship perfectly to a destination while using as little fuel as possible, all while obeying strict rules about how hard you can turn the wheel.

This is the real-world problem the paper tackles, but in the language of mathematics. Here is the breakdown of their solution using simple analogies.

The Problem: The "Heavy" Ship

The ship represents a Parabolic Partial Differential Equation (PDE). In plain English, this is a mathematical model describing how heat (or fluid, or electricity) spreads over time in a complex shape.

The Switch: The "switched" part means the rules of physics change suddenly. In the paper's example, it's two rooms connected by a door. If the door is closed, heat stays in one room. If it opens, heat rushes into the other. The math has to handle these sudden jumps.
The Challenge: To steer this ship perfectly (Optimal Control), you need to solve a massive, heavy calculation every few seconds. It's like trying to solve a 10,000-piece puzzle in your head every time you turn the wheel. If you try to do this in real-time, your brain (the computer) will crash.

The Solution: The "Lightweight" Map (Model Order Reduction)

To make the ship steerable in real-time, the authors create a Model Order Reduction (ROM).

The Analogy: Instead of carrying the full 10,000-piece puzzle, you create a simplified, low-resolution sketch of the ship. This sketch captures the most important movements (the "energy") but ignores the tiny, unimportant details.
The Benefit: Steering based on the sketch is incredibly fast. You can make decisions instantly.

The Catch: Is the Sketch Accurate Enough?

Here is the paper's biggest contribution. Usually, when you use a sketch instead of the real thing, you don't know how wrong you might be. You might steer the ship into a rock because your sketch missed a hidden reef.

The authors developed a Certified Error Estimator.

The Analogy: Imagine you have a magical "truth detector" attached to your sketch. Every time you make a decision based on the sketch, this detector instantly calculates: "How far off is this sketch from the real ship right now?"
The Safety Net:
- If the detector says, "You are very close to the real ship," you keep using the fast sketch.
- If the detector says, "Whoa, the sketch is getting blurry and dangerous," it triggers an alarm. The system then pauses, runs the heavy, slow, 10,000-piece calculation just once to get the truth, updates the sketch to be more accurate, and then goes back to using the fast sketch.

The Two Driving Modes

The paper tests two ways of using this system:

The "Hybrid" Driver (FOM-ROM-MPC): You make decisions using the fast sketch, but you apply those decisions to the real heavy ship.
- Pros: You always know exactly where the real ship is.
- Cons: You still have to simulate the heavy ship to check your position, which takes some time.
The "Pure Sketch" Driver (ROM-ROM-MPC): You make decisions using the sketch, and you pretend the ship is the sketch.
- Pros: It is the fastest possible method.
- Cons: You need a very high-quality sketch. If the sketch drifts too far from reality, you could get lost. The authors' "truth detector" ensures this doesn't happen by forcing a reality check whenever the sketch gets too sloppy.

The Results: Speed vs. Safety

The authors tested this on a computer simulation of the two-room heat problem.

Speed: Using their "certified sketch" method, they were able to steer the ship 4 to 11 times faster than using the heavy, full calculation every time.
Accuracy: Even though they used a simplified model, the "truth detector" ensured the ship stayed on the correct path. The error was kept within a tiny, predictable neighborhood of the perfect path.
The Trade-off: The "Pure Sketch" driver was the fastest but required the sketch to be updated more often (because the detector was more sensitive). The "Hybrid" driver was slightly slower but needed fewer updates.

Summary

The paper presents a new way to control complex, changing systems (like heat in rooms with moving doors) in real-time. They don't just speed up the math; they build a safety certification system that guarantees the fast, simplified math stays close enough to the slow, perfect math. It's like having a GPS that not only gives you a fast route but also constantly checks a high-definition map to ensure you haven't taken a wrong turn, only stopping to check the high-definition map when absolutely necessary.

Technical Summary: Certified Model Predictive Control for Switched Evolution Equations Using Model Order Reduction

Problem Formulation
The paper addresses the computational challenges associated with Model Predictive Control (MPC) for linear switched evolution equations arising from parabolic partial differential equations (PDEs). Specifically, the authors consider a finite-horizon optimal control problem (OCP) where the system dynamics are governed by a switched bilinear form corresponding to a time derivative. The control problem includes convex, lower-semicontinuous regularization functions, allowing for control box constraints and sparse ( $L_1$ ) regularization. The primary difficulty lies in the high computational cost of solving the full-order model (FOM) optimality conditions repeatedly within the receding horizon framework, which hinders real-time applicability.

Methodology
To mitigate this computational burden, the authors propose a framework integrating Galerkin Model Order Reduction (MOR) with MPC. The methodology proceeds through the following steps:

Optimality Conditions: The authors derive first-order necessary optimality conditions for the finite-horizon OCP. A key analytical finding is that the switching in the bilinear form associated with the time derivative induces a switched adjoint system with specific state transition maps at switching times. This analysis accommodates general convex regularizations, including non-differentiable terms like $L_1$ penalties.
Reduced-Order Modeling (ROM): A Galerkin projection is employed to construct a low-dimensional surrogate for the state and adjoint systems. The reduced system is solved to obtain a ROM-MPC feedback law.
A-Posteriori Error Estimation: The core of the methodology involves deriving recursive a-posteriori error estimates. These estimates bound the error between the ROM-MPC trajectory and the true FOM-MPC trajectory. The analysis accounts for:
- Errors in the initial values (propagated from previous MPC iterations).
- Errors in the optimal control, state, and adjoint state.
- The specific contributions of switching times, characterized by constants derived from the equivalence of energy norms across different switching modes.
Certified Algorithms: Two adaptive algorithms are formulated based on these error bounds:
- FOM-ROM-MPC: The controller is designed using the ROM but applied to the FOM. This approach utilizes expensive-to-evaluate error bounds that require solving FOM equations for intermediate states.
- ROM-ROM-MPC: Both the design and application of the controller occur entirely within the ROM framework. This approach uses cheap-to-evaluate error bounds based on residual norms, making it fully online-efficient.
- Certification Mechanism: Both algorithms include a restart mechanism. If the computed error estimators exceed user-defined thresholds, the reduced control is rejected, a full FOM-MPC step is performed, and the ROM is updated (e.g., via Proper Orthogonal Decomposition) before restarting.

Key Contributions

First-Order Optimality for Switched Systems: The derivation of optimality conditions for linear switched parabolic PDEs with general convex regularization, explicitly characterizing the discontinuous adjoint dynamics caused by switching in the time-derivative operator.
Recursive Error Bounds: The development of a-posteriori error estimates for the optimal control, state, and adjoint that explicitly account for perturbed initial values and error accumulation over multiple MPC iterations. These bounds are shown to control the size of the neighborhood in which the ROM-MPC trajectory evolves relative to the FOM-MPC trajectory.
Certified MPC Framework: The formulation of two distinct certified MPC algorithms (FOM-ROM-MPC and ROM-ROM-MPC) that guarantee the approximation error remains within computable bounds, controlled by the quality of the ROM.
Extension to Non-Smooth Control: The extension of existing error estimation techniques (previously limited to smooth or specific cases) to handle general convex, lower-semicontinuous regularization functions.

Results
Numerical experiments were conducted on a switched heat equation model representing two adjacent rooms with a door that opens and closes (switching the domain connectivity).

Error Estimator Performance: The study compared "expensive-to-evaluate" bounds (requiring FOM solves) and "cheap-to-evaluate" bounds (residual-based). The expensive bounds were found to be sharp (effectivity near 1), while the cheap bounds were conservative (effectivity around $10^{-3}$ ) but significantly faster to compute.
Algorithm Performance:
- The ROM-ROM-MPC algorithm achieved the highest speed-ups (up to 13.7x) but required more frequent ROM updates and larger basis sizes to maintain accuracy due to the conservatism of its cheap error estimators.
- The FOM-ROM-MPC algorithm offered a balanced speed-up (approx. 4x) with fewer updates and smaller basis sizes, demonstrating robustness against error overestimation.
- Both algorithms successfully kept the closed-loop trajectories within the certified error neighborhoods.
Sensitivity: The performance was sensitive to the POD energy parameter; tighter tolerances (retaining more energy) reduced the need for updates and improved speed-ups, while looser tolerances led to excessive updates, negating computational benefits.

Significance and Claims
The authors claim this work is the first contribution to combine optimal control/MPC for switched systems, MPC with MOR, and MOR for switched systems within an infinite-dimensional setting. The significance of the paper lies in providing a certified framework where the approximation error is not merely assumed to be small but is explicitly bounded and controlled. This allows for the safe application of reduced-order models in real-time control scenarios for switched PDEs, ensuring that the reduced controller behaves predictably relative to the full-order system. The framework is designed to be applicable to both stabilizing and economic MPC settings. The authors modestly note that while the current work focuses on linear switched systems, the theoretical framework could potentially be extended to other settings, and future work aims to integrate these estimators with balanced truncation and contour integral methods for time integration.