7.2 Sensitivity and robustness

Sensitivity and Robustness

Every real control system faces uncertainties: plant parameters drift, disturbances hit, sensors add noise, and your mathematical model never perfectly captures the true dynamics. Sensitivity and robustness give you the formal tools to quantify how much these imperfections affect your closed-loop system and to design controllers that still work well despite them.

This topic covers sensitivity functions and what they measure, robustness metrics and analysis techniques, and the main design strategies for building controllers that tolerate real-world uncertainty.

Sensitivity in Control Systems

Sensitivity quantifies how much a system's output or performance changes when something in the system changes. A parameter shifts by 5%, so how much does your step response degrade? A disturbance enters the loop, so how much does the output deviate from the reference? Sensitivity analysis answers these questions.

Sensitivity Functions

Two transfer functions sit at the core of robustness analysis in linear feedback systems:

• Sensitivity function $S(s)$: relates disturbances at the output to the actual output. For a standard unity-feedback loop with open-loop transfer function $L(s) = G(s)C(s)$, the sensitivity function is:

$S(s) = \frac{1}{1 + L(s)}$

A small $|S(j\omega)|$ at a given frequency means the feedback loop effectively rejects disturbances at that frequency. Ideally you want $|S|$ small across your bandwidth.

• Complementary sensitivity function $T(s)$: relates the reference input to the closed-loop output:

$T(s) = \frac{L(s)}{1 + L(s)}$

$T(s)$ describes how well the system tracks the reference, but it also determines how sensor noise propagates to the output.

These two functions are linked by a fundamental constraint:

$S(s) + T(s) = 1$

This means you cannot make both small at the same frequency. Good disturbance rejection (small $S$) at low frequencies forces poor noise rejection (large $T$) at those same frequencies, and vice versa. This tradeoff is one of the most important ideas in feedback design.

Sensitivity to Parameter Variations

Control systems rarely operate with the exact parameter values assumed during design. Gains drift with temperature, time constants change as components age, and physical properties vary between production units.

Classical sensitivity of a closed-loop transfer function $T$ with respect to a parameter $\alpha$ is defined as:

$S_\alpha^T = \frac{\partial T / T}{\partial \alpha / \alpha} = \frac{\alpha}{T} \frac{\partial T}{\partial \alpha}$

This tells you the percentage change in $T$ for a percentage change in $\alpha$. A key advantage of feedback is that it reduces this sensitivity. For a unity-feedback system, the closed-loop sensitivity to plant gain variations is reduced by the factor $\frac{1}{1 + L(s)}$ compared to the open-loop case.

High sensitivity to parameter variations can cause performance degradation or even instability. Robust and adaptive control techniques are specifically designed to keep the system working despite these variations.

Sensitivity to Disturbances

Disturbances are unwanted inputs that push the system away from its desired behavior: external forces on a mechanical system, load changes in a power grid, or flow-rate variations in a chemical process.

The sensitivity function $S(s)$ directly quantifies disturbance rejection. At frequencies where $|S(j\omega)|$ is small (typically low frequencies where loop gain is high), the feedback loop attenuates disturbances effectively. At frequencies where $|S(j\omega)| \approx 1$ or greater, disturbances pass through largely unattenuated.

Designing for low disturbance sensitivity means shaping $S(s)$ to be small over the frequency range where disturbances are expected to occur.

Sensitivity to Noise

Sensor noise is typically high-frequency, which is why the complementary sensitivity function $T(s)$ matters here. Since $T(s)$ describes how sensor signals propagate through the closed loop, high $|T(j\omega)|$ at high frequencies means noise gets amplified into the control signal and the output.

Because of the $S + T = 1$ constraint, you typically design for:

• Small $|S(j\omega)|$ at low frequencies (good disturbance rejection and tracking)
• Small $|T(j\omega)|$ at high frequencies (good noise rejection)

The crossover region where both are moderate corresponds roughly to the system's bandwidth. Filtering and loop-shaping techniques help manage this tradeoff.

Sensitivity to Unmodeled Dynamics

Every model is a simplification. High-frequency resonances, small time delays, parasitic dynamics, and nonlinearities are often left out of the design model. If the controller pushes the bandwidth too high, these unmodeled dynamics can cause unexpected oscillations or instability.

The complementary sensitivity function $T(s)$ plays a role here too. If $|T(j\omega)|$ is large at frequencies where unmodeled dynamics are significant, the closed-loop system is vulnerable. Robust control techniques explicitly account for model uncertainty by bounding the allowable $|T(j\omega)|$ at high frequencies, keeping the controller from being too aggressive where the model is unreliable.

Robustness of Control Systems

While sensitivity tells you how much performance changes, robustness tells you whether the system still works (remains stable, meets specs) across the full range of expected uncertainties. A robust system tolerates parameter changes, disturbances, noise, and modeling errors without failing.

Robustness Definition

Robustness is the ability of a closed-loop system to maintain stability and acceptable performance despite uncertainties. There are two distinct types:

• Robust stability: the system remains stable for all plants within a defined uncertainty set
• Robust performance: the system meets performance specifications for all plants within that uncertainty set

Robust performance is the stronger requirement. A system can be robustly stable but still have poor tracking or sluggish response for some parameter combinations.

Robustness Metrics

Common ways to quantify robustness include:

• Gain margin: how much the loop gain can increase before instability (typically want at least 6 dB)
• Phase margin: how much additional phase lag the system can tolerate before instability (typically want at least 30–45°)
• Peak sensitivity $M_s = \max_\omega |S(j\omega)|$: the inverse of the shortest distance from the Nyquist plot to the critical point $-1$. Values of $M_s$ between 1.2 and 2.0 are common design targets. Lower $M_s$ means more robust.
• Structured singular value $\mu$: a more general metric for systems with multiple structured uncertainties

Gain and phase margins are single-number summaries that can miss certain types of simultaneous gain and phase perturbations. $M_s$ and $\mu$ provide more complete robustness pictures.

Robustness vs. Performance Tradeoff

This is a fundamental tension in control design. Making a system more robust generally means being more conservative, which reduces nominal performance. Specifically:

• Higher bandwidth gives faster tracking and better disturbance rejection, but increases sensitivity to noise and unmodeled dynamics
• Lower bandwidth is more robust to model uncertainty, but responds more slowly and rejects disturbances less effectively
• The $S + T = 1$ constraint means you cannot simultaneously minimize sensitivity to disturbances and sensitivity to noise at the same frequency

Control engineers navigate this tradeoff by clearly defining which uncertainties matter most, what performance is truly required, and where compromises are acceptable.

Robustness to Parameter Uncertainties

When plant parameters are uncertain but bounded (e.g., a gain known to lie between 0.8 and 1.2 times its nominal value), robust control methods design a single fixed controller that stabilizes and performs well across the entire parameter range.

Techniques include:

• Robust control ($H_\infty$, $\mu$-synthesis): explicitly optimizes over the uncertainty set
• Adaptive control: adjusts controller parameters online as the true plant parameters reveal themselves
• Feedback linearization: cancels known nonlinearities, though it requires reasonably accurate knowledge of the nonlinear terms

Robustness to Disturbances

Disturbance robustness means the system can reject or attenuate unwanted inputs and maintain its desired output. Key approaches:

• Feedback control: the fundamental mechanism for disturbance rejection, using error signals to correct deviations
• Feedforward control: if a disturbance is measurable, feedforward can cancel it before it affects the output
• Disturbance observers: estimate unmeasured disturbances and compensate for them in the control law
• Integral action: guarantees zero steady-state error for constant disturbances (this is why PID controllers include an integral term)

Robustness to Noise

Noise robustness means the system doesn't amplify high-frequency sensor noise into large control actions or output variations. Strategies include:

• Rolling off $|T(j\omega)|$ at high frequencies through loop shaping
• Using low-pass filters on sensor signals (with care not to add too much phase lag)
• Stochastic control methods (like LQG) that explicitly model noise statistics and minimize expected performance loss

Robustness to Unmodeled Dynamics

This is often the most dangerous source of instability. A controller designed for a simplified model can destabilize the real plant if it's too aggressive at frequencies where the model is inaccurate.

• $H_\infty$ control bounds $|T(j\omega)|$ at high frequencies, limiting the controller's authority where the model is uncertain
• $\mu$-synthesis handles structured uncertainty, including unmodeled dynamics represented as bounded perturbations
• Model reduction techniques help identify which dynamics are critical and which can be safely neglected, guiding the choice of controller bandwidth

Sensitivity and Robustness Analysis

Before committing to a controller design, you need to analyze how the closed-loop system behaves under uncertainty. Several techniques exist, ranging from simple to computationally intensive.

Sensitivity Analysis Techniques

• Local sensitivity analysis: compute partial derivatives of performance metrics with respect to parameters at the nominal operating point. This is fast but only valid near that point. Finite differences can approximate these derivatives numerically.
• Global sensitivity analysis: evaluate performance across the entire range of parameter uncertainty. Monte Carlo methods and variance-based methods (like Sobol indices) fall into this category. These are more computationally expensive but give a fuller picture.

Robustness Analysis Techniques

• Gain and phase margin analysis: check classical stability margins from Bode or Nyquist plots. Simple and intuitive, but limited to single-loop perturbations.
• Structured singular value ($\mu$) analysis: the most powerful tool for linear systems with structured uncertainties. Determines the smallest structured perturbation that destabilizes the system.
• Lyapunov-based methods: for nonlinear systems, quadratic stability analysis and parameter-dependent Lyapunov functions can prove robustness over a defined uncertainty set.

Worst-Case Analysis

Worst-case analysis finds the specific combination of uncertainties that causes the worst performance degradation or brings the system closest to instability.

Steps in worst-case analysis:

1. Define the uncertainty set (parameter ranges, disturbance bounds, model error bounds)
2. Formulate the performance metric you care about (e.g., peak tracking error, settling time)
3. Search over the uncertainty set for the combination that maximizes degradation of that metric
4. Evaluate whether the worst-case performance is still acceptable

This is often solved as an optimization problem. For linear systems, $\mu$-analysis provides an efficient way to compute worst-case gains.

Monte Carlo Simulations

When the system is too complex for analytical worst-case analysis (nonlinear, high-dimensional, many uncertainty sources), Monte Carlo simulation is a practical alternative.

1. Define probability distributions for each uncertain parameter
2. Randomly sample parameter combinations (hundreds or thousands of trials)
3. Simulate the closed-loop system for each sample
4. Analyze the distribution of performance metrics across all trials

Monte Carlo methods won't guarantee you've found the absolute worst case, but they reveal the statistical distribution of performance and can uncover failure modes that analytical methods might miss.

Structured Singular Value ($\mu$) Analysis

The structured singular value $\mu$ is defined for a matrix $M$ and an uncertainty structure $\Delta$:

$\mu_\Delta(M) = \frac{1}{\min\{||\Delta|| : \det(I - M\Delta) = 0, \Delta \in \mathbf{\Delta}\}}$

In plain terms, $\mu$ tells you the size of the smallest structured perturbation that makes the system singular (unstable). If $\mu < 1$ for a given uncertainty bound, the system is robustly stable. If $\mu < 1$ for the performance problem, the system achieves robust performance.

$\mu$-analysis is used together with $H_\infty$ control in an iterative process called D-K iteration (used in $\mu$-synthesis) to design controllers that are robust to structured uncertainties.

Designing for Sensitivity and Robustness

Design is where analysis meets synthesis. The goal is to choose a controller that achieves acceptable performance while tolerating the uncertainties your system will face.

Robust Control Design Objectives

A robust controller design typically aims to:

1. Guarantee closed-loop stability for all plants in the uncertainty set
2. Keep $|S(j\omega)|$ small at low frequencies (disturbance rejection, tracking)
3. Keep $|T(j\omega)|$ small at high frequencies (noise rejection, robustness to unmodeled dynamics)
4. Satisfy constraints on control effort (actuator limits)
5. Achieve an acceptable tradeoff between nominal performance and robustness

These objectives are often formulated as a weighted optimization problem. For example, in $H_\infty$ design, you choose weighting functions $W_1(s)$ and $W_2(s)$ and minimize:

$\left\| \begin{bmatrix} W_1 S \\ W_2 T \end{bmatrix} \right\|_\infty$

The weights encode your design priorities: where you need tight disturbance rejection versus where you need robustness.

Sensitivity Reduction Techniques

Feedback control is the most fundamental sensitivity reduction tool. Closing the loop divides the open-loop sensitivity by $1 + L(s)$, which is the core reason feedback exists.

Feedforward control complements feedback by acting on measurable disturbances before they affect the output. It doesn't change the feedback loop's sensitivity properties, but it reduces the disturbance's net effect.

Adaptive control adjusts controller parameters in real time to compensate for changing plant dynamics, effectively reducing sensitivity to slow parameter variations.

Feedback Control for Robustness

Feedback is the primary mechanism for robustness. A well-designed feedback controller:

• Reduces the effect of plant gain variations by the factor $1/(1 + L(s))$
• Rejects disturbances entering at the plant output
• Can provide integral action to eliminate steady-state errors

Common feedback design methods for robustness include:

• PID control: widely used, provides proportional, integral, and derivative action. Tuning methods like Ziegler-Nichols give a starting point, but robustness-oriented tuning (e.g., SIMC method) explicitly targets gain and phase margins.
• Lead-lag compensation: shapes the loop transfer function to improve phase margin (lead) or low-frequency gain (lag).
• State feedback with observers: if a state-space model is available, LQR provides optimal state feedback, and a Kalman filter estimates unmeasured states. The combination (LQG) can be augmented with robustness guarantees using loop transfer recovery (LTR).

Feedforward Control for Sensitivity Reduction

Feedforward control works best when:

• The disturbance is measurable (or at least predictable)
• A reasonable model of how the disturbance affects the output is available

The feedforward controller is typically designed as the inverse of the disturbance-to-output transfer function. In practice, perfect inversion isn't possible (especially for non-minimum-phase systems), so approximate inversions are used.

Combining feedforward with feedback gives you the best of both: feedforward handles the predictable part of the disturbance, and feedback cleans up the rest.

Adaptive Control for Robustness

Adaptive control is useful when uncertainties are too large for a single fixed controller to handle, but the parameters change slowly enough to be tracked.

Two main approaches:

• Gain scheduling: design multiple controllers for different operating points, then switch or interpolate between them based on a measured scheduling variable (e.g., airspeed, temperature). This is widely used in aerospace and process control.
• Model reference adaptive control (MRAC): defines a reference model that represents the desired closed-loop behavior, then adjusts controller parameters online to make the actual system match the reference model. MRAC can handle significant parameter uncertainty but requires careful stability analysis.

Robust Optimal Control Design

$H_\infty$ control minimizes the worst-case (peak) gain from disturbances to a weighted combination of outputs. It produces a controller that guarantees a bound on the sensitivity and complementary sensitivity functions. The design requires choosing frequency-dependent weighting functions that reflect your performance and robustness priorities.

$\mu$-synthesis extends $H_\infty$ by explicitly accounting for structured uncertainties. The D-K iteration procedure alternates between:

1. Solving an $H_\infty$ problem for a fixed scaling matrix $D$
2. Optimizing the scaling matrix $D$ to tighten the $\mu$ bound

This iterative process converges to a controller that is robust to the specified structured uncertainty set.

Both methods produce higher-order controllers, which may need to be reduced for practical implementation using model reduction techniques.

Sensitivity and Robustness in Practice

Real-World Applications

Aerospace: Flight control systems must handle changing aerodynamic conditions (altitude, speed, angle of attack), wind gusts, and significant model uncertainty. Gain-scheduled controllers and $H_\infty$ methods are standard. The F-16 VISTA research aircraft, for example, used robust control techniques to maintain handling qualities across its flight envelope.

Process control: Chemical reactors, distillation columns, and other industrial processes face varying feed compositions, catalyst degradation, and thermal disturbances. Robust PID tuning and model predictive control (MPC) with uncertainty handling are common approaches.

Automotive: Electronic stability control (ESC) must work across different road surfaces (dry, wet, icy), tire conditions, and loading. Adaptive cruise control (ACC) must handle varying vehicle masses and aerodynamic drag. These systems use robust control combined with real-time estimation of uncertain parameters.

Challenges in Implementation

Model uncertainty: Getting an accurate-enough model for robust control design is often the hardest part. System identification experiments, first-principles modeling, and uncertainty quantification all play a role.

Computational complexity: $H_\infty$ and $\mu$-synthesis involve solving matrix Riccati equations or linear matrix inequalities (LMIs) that scale with system order. For large-scale systems (hundreds of states), this can be prohibitive without model reduction.

Sensor and actuator limitations: Real sensors have noise floors, biases, and bandwidth limits. Real actuators saturate and have rate limits. These constraints bound what any controller can achieve, no matter how sophisticated the design method.

Balancing Sensitivity and Robustness

In practice, the design process is iterative:

1. Start with a nominal controller that meets performance specs
2. Analyze sensitivity and robustness (margins, $\mu$-analysis, Monte Carlo)
3. If robustness is insufficient, back off on performance (reduce bandwidth, add roll-off)
4. If performance is insufficient, tighten the design (increase bandwidth, add feedforward)
5. Repeat until both performance and robustness are acceptable

The weighting functions in $H_\infty$ design are the primary knobs for this tradeoff. Adjusting them shifts the balance between tracking performance and uncertainty tolerance.

Sensitivity and Robustness Testing

Before deploying a controller, you need to verify its robustness through testing:

• Simulation testing: sweep parameters across their uncertainty ranges, inject worst-case disturbances, and verify performance metrics
• Hardware-in-the-loop (HIL) testing: run the controller on real hardware connected to a simulated plant, allowing you to test edge cases safely
• Field testing: gradually expand the operating envelope, monitoring for unexpected behavior

Testing should cover not just nominal conditions but also the corners of the uncertainty set, where problems are most likely to appear.

Case Studies

• Robust flight control: Modern fly-by-wire aircraft use robust control to handle the wide variation in aerodynamic parameters across the flight envelope. The controller must maintain stability and handling qualities from low-speed takeoff through high-speed cruise, with margins for turbulence and structural flexibility.
• Robust process control: In semiconductor manufacturing, etch rate and deposition thickness must be controlled to nanometer precision despite variations in chamber conditions and wafer properties. Run-to-run control with robust adaptation keeps yields high.
• Robust automotive control: Anti-lock braking systems (ABS) must work on surfaces ranging from dry asphalt (friction coefficient ~0.8) to ice (friction coefficient ~0.1). The controller adapts its braking strategy based on estimated tire-road friction, maintaining robust stopping performance across conditions.