⚡ electrical engineering

An approach to the LQG/LTR design problem with specifications for finite-dimensional SISO control systems

This expository paper presents a weighting augmentation approach to the LQG/LTR design problem for finite-dimensional SISO systems, enabling practitioners to incorporate specific low- and high-frequency performance and robustness requirements, as demonstrated through a geared DC motor torque control example.

Original authors: Mahyar Mahinzaeim, Kamyar Mehran

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahyar Mahinzaeim, Kamyar Mehran

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

The Big Picture: Tuning a Car's Suspension

Imagine you are trying to tune the suspension of a car. You have two main goals that often fight against each other:

Comfort (Low Frequency): When you hit a slow bump or a pothole, the car should absorb it smoothly so you don't feel it.
Stability (High Frequency): When the car hits tiny, rapid vibrations (like gravel or engine noise), the suspension shouldn't go crazy shaking the car apart.

In the world of engineering, this is called control system design. The paper discusses a specific method called LQG/LTR (Linear Quadratic Gaussian / Loop Transfer Recovery). Think of LQG as a "smart autopilot" for machines. It's great at keeping a machine steady, but it has a known flaw: it's like a driver who is so focused on the road ahead that they get jittery when the radio gets too loud (noise) or when the road gets bumpy in unexpected ways.

The Problem: The "Perfect" Driver is Too Sensitive

The authors explain that while this "smart autopilot" (LQG) is mathematically elegant, it often fails in the real world because it doesn't listen to the engineer's specific rules about where it should be sensitive and where it should be tough.

The Old Way: Engineers used to try to "fix" the autopilot by guessing and checking, tweaking knobs until it felt right. This was like trying to tune a radio by turning the dial randomly until the static stopped. It worked sometimes, but it was messy and hard to explain why it worked.
The New Way (This Paper): The authors propose a systematic, algebraic way to "teach" the autopilot exactly what to do. They don't just guess; they build a custom set of rules (called weightings) that act like a filter for the machine's ears.

The Solution: The "Weighted Glasses" Analogy

The core idea of the paper is Weighting Augmentation.

Imagine the machine is wearing a pair of special glasses.

The Low-Frequency Glasses (Weighting $W_2$ ): These glasses make the machine very sensitive to slow movements (like a gentle push). This ensures the machine reacts quickly to disturbances and tracks its target perfectly.
The High-Frequency Glasses (Weighting $W_1$ ): These glasses make the machine "deaf" to fast, tiny noises. This stops the machine from overreacting to static or sensor glitches.

The authors show how to mathematically design these "glasses" (the weightings) so that when you put them on the machine, the resulting "smart autopilot" automatically meets your specific goals.

How They Did It: The Recipe

Instead of using complex, heavy computer simulations that take hours to solve, the authors turned the problem into a math recipe.

Define the Goal: You say, "At frequency X, the error must be this small," and "At frequency Y, the noise must be this quiet."
The Algebraic Trick: They use a set of equations to figure out exactly what the "glasses" (the weightings) need to look like to achieve those goals.
The Result: Once the glasses are designed, you simply plug them into the standard autopilot formula, and poof—you get a controller that works exactly as you specified.

The Real-World Test: The Wobbly Motor

To prove this works, the authors tested it on a geared DC motor (like the kind used in robot arms or power steering).

The Challenge: This motor has a "springy" shaft. If you push it, it wobbles. It's a tricky machine to control because it wants to vibrate.
The Test: They applied their "glasses" method.
- They told the system: "Be very smooth at low speeds (to track the target)."
- They told the system: "Ignore the high-pitched electrical noise."
The Outcome: The system worked perfectly. The motor tracked the target smoothly without shaking.

However, they also found a catch. If you make the system too perfect at tracking, it can become "nervous" (unstable) when things change quickly. So, they had to tweak the "glasses" slightly to find a balance between being smooth and being stable. This is the classic engineering trade-off: you can't have everything, but this method helps you find the best possible balance.

Why This Matters

The authors wrote this paper for practitioners—the engineers who actually build these systems.

They wanted to move away from "black box" methods where you don't know why a controller works.
They wanted to show that you can design these complex systems using clear, step-by-step algebra (math you can do on paper or with standard software) rather than needing super-complex optimization theories.

In short: The paper provides a clear, step-by-step manual for building a "smart autopilot" that knows exactly how to ignore noise while still reacting perfectly to real commands, using a method that is easy to follow and verify.

Technical Summary: An Approach to the LQG/LTR Design Problem with Specifications for Finite-Dimensional SISO Control Systems

Problem Statement
The paper addresses a fundamental limitation in Linear Quadratic Gaussian (LQG) control theory: while the Kalman–Bucy filter (KBF) guarantees minimal sensitivity and the Linear Quadratic Regulator (LQR) guarantees robustness margins, these properties do not automatically extend to the LQG compensator (state-estimate feedback). Although the Loop Transfer Recovery (LTR) technique was developed to "restore" these robustness properties, the conventional approach often fails to simultaneously meet specific low-frequency performance (tracking/disturbance rejection) and high-frequency robustness (noise attenuation) specifications.

Furthermore, the authors note that while $H_\infty$ and other Hardy space norm optimization methods offer a rigorous framework for frequency-domain specifications, they often involve high computational complexity and require posing complex control optimization problems. Conversely, many existing LQG/LTR applications rely on heuristic, unsystematic parameter adjustments, making it difficult for practitioners to formally verify designs or map design parameters to specific loop-shaping characteristics, particularly for finite-dimensional Single-Input/Single-Output (SISO) systems.

Methodology
The proposed approach reformulates the LQG design problem by integrating frequency-dependent design specifications directly into the LTR framework through weighting augmentation. The core methodology proceeds as follows:

Augmented Plant Formulation: The nominal plant $G_0(s)$ is augmented with two dynamic weighting functions, $W_1(s)$ and $W_2(s)$ , placed at the control input and measurement noise input, respectively.
- $W_1(s)$ is designed to shape the loop for the LQR state feedback controller (affecting high-frequency noise attenuation).
- $W_2(s)$ is designed to shape the loop for the KBF (affecting low-frequency sensitivity).
- The authors utilize higher-order lead and lag type transfer functions for these weightings, defined by coefficients $\tau_{j1}$ and $\tau_{j2}$ .
Separation Principle and Asymptotic Recovery: The design relies on the separation principle and the assumption that the plant is minimum phase. Under the condition that the LQR weighting parameter $\rho \to \infty$ , the LQG open-loop transfer function asymptotically recovers the properties of the KBF loop transfer function.
Algebraic Solution for Specifications: Instead of solving a complex optimization problem, the authors translate closed-loop design specifications (bounds on the sensitivity function $S(s)$ and controller noise sensitivity $K(s)S(s)$ ) into open-loop constraints on the weighting gains.
- Low-frequency specifications (tracking) are mapped to constraints on $|W_2(i\omega)|$ .
- High-frequency specifications (noise attenuation) are mapped to constraints on $|W_1(i\omega)|$ .
- This translation results in a system of non-linear algebraic inequalities (or equalities, if design bounds are treated as strict targets) regarding the weighting coefficients $\tau_{j1}, \tau_{j2}$ .
Iterative Design Procedure: The design process becomes an algebraic procedure to find suitable coefficient pairs that satisfy the derived inequalities. The authors emphasize that the existence of solutions depends on the attainability of the specified bounds, highlighting the iterative nature of the design.

Key Contributions

Systematic Algebraic Procedure: The paper presents a completely algebraic method to determine LQG compensator transfer functions for SISO systems. It avoids the need for iterative control optimization solvers, relying instead on solving algebraic equations derived from frequency-domain specifications.
Explicit Frequency Shaping: The approach explicitly links the coefficients of the weighting functions to specific low- and high-frequency design goals, demystifying the "mapping" of design parameters to loop characteristics that is often opaque in heuristic LTR methods.
SISO Focus: The work specifically targets finite-dimensional SISO systems, a domain where the authors argue practical, theoretically grounded LQG/LTR design examples are scarce in the literature compared to complex MIMO cases.
Attainability Analysis: The method naturally incorporates the concept of attainability, demonstrating that not all combinations of performance and robustness specifications are simultaneously achievable, and providing a framework to identify feasible design regions.

Results
The methodology is validated through a detailed design example: the torque control of a geared DC motor with an elastically mounted output shaft.

Plant Model: A 4th-order rational transfer function derived from experimental data, which is stable and minimum phase.
Design Cases: The authors tested three weighting orders ( $m, p$ $m, p$ ).
- Case 1 ( $m=p=1$ ): No suitable solution pairs were found for the specified bounds, illustrating the limits of attainability with lower-order weightings.
- Case 2 ( $m=p=2$ ): Solutions were found, but the high-frequency noise attenuation specification could not be fully met.
- Case 3 ( $m=3, p=2$ ): A suitable solution was found that satisfied all low- and high-frequency specifications simultaneously.
Performance: The resulting compensator achieved the desired sensitivity ( $|S|$ ) and noise sensitivity ($|KS|$) bounds. However, the initial design exhibited a sensitivity peak in the crossover frequency range, leading to a lightly damped step response and reduced robustness margins (gain margin $\approx 1.9$ dB).
Refinement: By relaxing the high-frequency noise sensitivity bounds (increasing the allowable magnitude), the authors successfully redesigned the weighting coefficients. This modification reduced the sensitivity peak, significantly improving the gain margin (to $\approx 4.8$ dB) and phase margin (to $\approx 35.2^\circ$ ) while maintaining acceptable performance.

Significance and Claims
The authors position this work as an expository contribution aimed at non-specialists and practitioners. They claim that their approach:

Provides a "straightforward and systematic" alternative to both heuristic LTR tuning and computationally intensive $H_\infty$ optimization.
Bridges the gap between classical loop shaping intuition and the LQG/LTR framework.
Offers a practical procedure for SISO systems where the "wealth of experience" for applying LTR is reportedly small.

The paper modestly acknowledges that the mathematical results themselves are not new, as they rely on established LQG/LTR theory and separation principles. The novelty lies in the specific application of weighting augmentation to create a solvable algebraic system for SISO design. The authors explicitly state that the approach assumes minimum-phase plants and does not currently handle time delays or non-minimum phase systems, noting these as natural extensions for future work. They also identify the computational complexity of solving the non-linear algebraic systems for higher-order weightings as a challenge for future study. Finally, they note that experimental verification of the compensator is underway but not yet included in the paper.