⚡ electrical engineering

Optimal $\mathbb{H}_2$ Control with Passivity-Constrained Feedback: Convex Approach

This paper demonstrates that the $\mathcal{H}_2$ -optimal feedback control problem for passive plants with output-strictly passive constraints can be reformulated as a convex, infinite-dimensional optimization over the Youla parameter, which is effectively approximated by converging finite-dimensional truncations to yield sub-optimal controllers and lower bounds.

Original authors: J. T. Scruggs

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: J. T. Scruggs

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 trying to keep a wobbly, dancing robot upright while a storm of wind pushes it around. You want the robot to stay as still as possible, using the least amount of energy. This is the world of control theory: the science of making machines behave exactly how we want them to, even when things get messy. Usually, engineers build "smart" controllers that act like tiny, super-fast computers. They measure the robot's wobble, calculate a perfect counter-move, and send a command to the motors. But these smart systems need batteries, chips, and electricity. If the power goes out, the robot falls.

Now, imagine a different kind of controller: one made entirely of springs, rubber bands, and gears. This is a "passive" controller. It has no brain and no battery. It works purely because of the laws of physics. If you push it, it pushes back. If you pull it, it pulls back. The beauty of passive systems is that they are incredibly reliable; they can't crash because they don't need to be turned on. However, there's a catch. Because they are limited to simple physics, they can't always be as good at stopping the wobble as the fancy computer-controlled ones. The big question for engineers has always been: "How close can a simple, battery-free spring system get to the performance of a super-computer?" This paper dives right into that question, trying to find the absolute best possible performance for a passive controller, without cheating.

The authors, led by J.T. Scruggs, tackle a specific puzzle: finding the "perfect" passive controller for a system that is already somewhat stable. They aren't just guessing; they are trying to solve a massive math problem to find the theoretical limit of what a passive system can do. Think of it like trying to find the fastest possible time a human could run a marathon without using any shoes or technology, just pure biology. You can't just guess; you have to model the muscles, the wind, and the track to find the absolute limit.

In this paper, the team treats the controller as a "black box" that must follow strict rules of passivity (it can't create energy, it can only absorb or store it). They want to minimize the "wobble" (mathematically called the H2 objective) caused by random disturbances. The problem is that finding this perfect controller is like trying to solve a maze that has an infinite number of paths. It's too big for any computer to solve all at once.

Here is where the paper gets clever. The authors realized that instead of trying to design the controller directly, they could design a "shadow" version of it, called the Youla parameter. This is like trying to solve a puzzle by looking at its shadow on the wall instead of the puzzle pieces themselves. By switching to this shadow view, the impossible, infinite maze suddenly turns into a smooth, bowl-shaped valley. In math, this is called a "convex" problem. It means that if you roll a ball down the hill, it will always roll to the very bottom, the perfect solution, without getting stuck in a fake valley.

The paper shows that while we can't solve the infinite problem instantly, we can build a series of smaller, simpler problems that get closer and closer to the perfect answer. Imagine trying to draw a perfect circle. You can't do it in one stroke, but you can draw a square, then an octagon, then a 16-sided shape. Each time you add more sides, the shape looks more like a circle. The authors did exactly this: they created a method to add more "sides" to their controller design. As they added more sides (mathematically, by increasing the number of parameters), their solution got better and better, inching closer to the true, infinite perfect controller.

To make sure they were actually getting close to the truth and not just fooling themselves, they built a second, opposite kind of test. This test provided a "floor" or a lower bound. It was like saying, "No matter what, the best you can do is at least this good." By running both the "getting better" test and the "floor" test, they could watch the two numbers squeeze together. When the top number (the best they found) and the bottom number (the guaranteed minimum) get very close, they know they have found the answer.

The paper demonstrates this with two examples. The first is a simple vibration problem, like a single weight on a spring. The second is a more complex model of a car's suspension system, trying to keep the car cabin smooth while the wheels hit bumps. In both cases, they showed that their method works. They found that with just a moderate amount of complexity (about 30 to 100 "sides" on their shape), they could get almost as good a result as the theoretical perfect controller. They also showed that the math they used is efficient; it doesn't take forever to compute, even for these complex shapes.

The authors are careful to note what they didn't do. They didn't invent a new type of spring or a new battery. They didn't claim that passive controllers are always better than active ones. In fact, they showed that for some very specific, simple cases, there are known shortcuts, but for the general case, the old ways of guessing or using approximations were missing the mark. They proved that by using their new "shadow" method, you can find the true best passive controller, not just a guess.

One of the most exciting parts of their discovery is how they managed to make the math simpler. Usually, as you try to make a solution more accurate, the math gets exponentially harder, like trying to solve a Rubik's cube that keeps getting bigger. The authors found a way to use a mathematical trick called "duality" to keep the difficulty growing only in a straight line. It's like realizing that instead of counting every single grain of sand on a beach, you can just measure the length of the shoreline and multiply by a constant. This makes it possible to solve these problems on a regular laptop, rather than needing a supercomputer.

In the end, this paper gives engineers a powerful new tool. If you are designing a system where reliability is key—like a bridge that needs to withstand earthquakes, or a robot that needs to work in a place where batteries are hard to change—you can now calculate exactly how well a passive system can perform. You won't have to guess if a spring is "good enough." You can calculate the absolute limit, design a controller that hits that limit, and know that it is the best possible passive solution nature allows. The paper doesn't just say "it's possible"; it provides the map to get there, showing that with the right mathematical perspective, even the most complex control problems can be tamed.

Technical Summary: Optimal H2 Control with Passivity-Constrained Feedback

Problem Statement
This paper addresses the $H_2$ -optimal feedback control problem for a linear time-invariant (LTI) plant $P$ where the input-output channel $u \mapsto y$ is passive and the plant has a bounded $L_2$ gain (i.e., $P \in H_\infty$ ). The objective is to minimize the standard $H_2$ norm of the closed-loop mapping from disturbance $w$ to performance output $z$ . Crucially, the feedback controller $K$ is constrained to be output-strictly passive (OSP). This constraint arises from practical engineering requirements in fields such as vibration suppression and robotics, where controllers may need to be implemented using only passive components (requiring no external power) or where stability robustness is guaranteed by the passivity of the loop.

While passive control is robust, it is often less performant than active control. The paper seeks to determine the theoretical performance limit of passive feedback by solving the optimization problem where $K$ is constrained to be OSP, rather than treating passivity as a secondary constraint or using heuristic approximations.

Methodology
The authors propose a convex approach to solve this problem by reformulating it in terms of the Youla parameter. The methodology proceeds through the following steps:

Bilinear Transformation: The continuous-time problem is mapped to a discrete-time domain using a bilinear transformation. This converts the $H_2$ and $H_\infty$ norms and the passivity constraints into their discrete-time equivalents ( $\tilde{H}_2$ , $\tilde{H}_\infty$ , and Positive Real conditions).
Spectral Factorization: The authors utilize spectral factorization results (related to the Kalman-Yakubovich-Popov lemma) to characterize the passivity constraint of the plant and the controller in the discrete domain.
Youla Parametrization: The closed-loop system is expressed using the Youla parameter $\tilde{Q}$ . The authors demonstrate that the $H_2$ objective becomes a convex quadratic function of $\tilde{Q}$ , and, significantly, the OSP constraint on the controller $K$ translates into a convex Linear Matrix Inequality (LMI) constraint on $\tilde{Q}$ .
Truncation and Finite-Dimensionalization: Since the optimal Youla parameter resides in an infinite-dimensional space, the authors propose truncating the Markov parameters of $\tilde{Q}$ to a finite number $n$ . This yields a sequence of finite-dimensional convex optimization problems.
Primal and Dual Formulations:
- Primal (OP4): A direct optimization over the truncated Markov parameters. The authors show that a naive implementation scales quadratically ( $O(n^2)$ ) with the truncation order due to auxiliary Lyapunov variables.
- Dual (OP5): By applying duality techniques, the authors derive a dual formulation where the optimization domain scales linearly ( $O(n)$ ) with $n$ , significantly improving computational efficiency.
Lower Bound Construction (OP6): A second sequence of optimizations is devised to compute a monotonically increasing lower bound on the optimal objective. This allows for the verification of convergence and the determination of when a truncated solution is within a desired tolerance of the global optimum.

Key Contributions
The paper makes the following specific contributions:

Convex Reformulation: It is shown that the $H_2$ optimal control problem with an OSP constraint is equivalent to an infinite-dimensional convex optimization problem over the Youla parameter domain.
Convergence of Truncated Solutions: The optimal controller is proven to be the limit of a sequence of finite-dimensional convex optimizations. As the dimension of the parametric domain increases, the objective value converges monotonically to the global optimum.
Dimensionality Reduction: Using duality, the dimension of the optimization domain for the $n$ -th truncation is reduced from $O(n^2)$ to $O(n)$ .
Lower Bounds: A method is provided to compute a lower bound on the optimal objective that converges monotonically from below, enabling the assessment of solution quality.
Sub-optimal Reference: A computationally simple, certainty-equivalent sub-optimal solution is presented to serve as a reference point, though the paper emphasizes that the proposed convex method yields the true optimum without the conservatism inherent in certainty-equivalence assumptions.

Results
The methodology is demonstrated on two vibration suppression examples:

Example 1: A single-degree-of-freedom oscillator coupled to a piezoelectric transducer. The results show that the optimal passive controller ( $\hat{K}^*_{100}$ ) differs significantly from the sub-optimal certainty-equivalent controller ( $\hat{K}_0$ ), particularly in phase characteristics. The convergence of the upper bound (truncated solution) and lower bound to the true optimum is verified.
Example 2: A two-dimensional automotive suspension model. Similar convergence behavior is observed, with the optimal controller achieving significant performance improvements over the sub-optimal design while maintaining passivity.

In both cases, the numerical experiments confirm that the truncated solutions converge to the infinite-dimensional optimum and that the lower bounds provide a tight estimate of the achievable performance.

Significance
The paper claims significance by resolving the $H_2$ optimal control problem under passivity constraints in a rigorous, convex framework. Unlike previous approaches that relied on Linear Matrix Inequalities (LMIs) with certainty-equivalence assumptions (introducing conservatism) or Bilinear Matrix Inequalities (BMIs) which are non-convex and computationally expensive, this work provides a method that is:

Convex: Ensuring global optimality for any finite truncation.
Asymptotically Exact: Proving convergence to the true infinite-dimensional optimum.
Computationally Efficient: Through the dual formulation, the complexity scales linearly with the approximation order.

The authors conclude that this technique allows engineers to determine the precise performance limits of passive feedback systems, providing a benchmark against which active control strategies can be fairly evaluated. The work also notes that while extensions to generalized $H_2$ and $H_\infty$ objectives are possible, the extension to robust performance bounds (rather than just robust stability) remains an open problem.

More like this