9.2 H-infinity control

Fundamentals of H-infinity control

H-infinity control is a robust control technique that minimizes worst-case system performance in the presence of uncertainties and disturbances. Rather than optimizing for average behavior, it guarantees that the system stays stable and performs acceptably even under the most adverse conditions. This makes it one of the most important tools in modern robust control design.

Norms in control systems

Norms are mathematical tools that quantify the "size" of signals and systems. In control theory, two norms dominate:

• The H2 norm measures the energy of a system's impulse response. It's closely related to the root-mean-square (RMS) value of the output and captures average performance.
• The H-infinity norm measures the maximum gain of a system across all frequencies. It captures worst-case performance, which is why it's central to robust control.

Think of it this way: the H2 norm tells you how the system behaves on average, while the H-infinity norm tells you the single frequency where the system amplifies inputs the most.

H-infinity norm definition

The H-infinity norm of a stable transfer function $G(s)$ is defined as:

$\|G\|_\infty = \sup_{\omega} \sigma_{\max}(G(j\omega))$

where $\sigma_{\max}$ denotes the maximum singular value and $\sup$ is the supremum (least upper bound) over all frequencies $\omega$.

For SISO systems, the maximum singular value reduces to the magnitude of the frequency response, so $\|G\|_\infty = \sup_\omega |G(j\omega)|$. That's just the peak of the Bode magnitude plot.

For MIMO systems, you need singular values because the gain depends on the direction of the input vector. The H-infinity norm gives you the worst-case amplification across all input directions and all frequencies.

H-infinity norm vs H2 norm

H-infinity control problem formulation

The core problem: design a controller $K(s)$ that minimizes the H-infinity norm of the closed-loop transfer function from exogenous inputs (disturbances, noise, references) to regulated outputs (tracking errors, control effort), while maintaining internal stability.

Standard H-infinity control configuration

The standard configuration has two blocks:

• A generalized plant $P(s)$ that bundles together the nominal system model, performance weighting functions, and uncertainty descriptions.
• A controller $K(s)$ that you're designing.

The generalized plant has two sets of inputs and two sets of outputs:

• Exogenous inputs $w$: disturbances, references, noise
• Control inputs $u$: signals from the controller
• Regulated outputs $z$: performance signals you want to keep small
• Measured outputs $y$: signals available to the controller

The goal is to find a stabilizing $K(s)$ that minimizes $\|T_{zw}\|_\infty$, the closed-loop transfer function from $w$ to $z$.

Generalized plant representation

The generalized plant $P(s)$ is represented in state-space form as:

$\begin{bmatrix} \dot{x} \\ z \\ y \end{bmatrix} = \begin{bmatrix} A & B_1 & B_2 \\ C_1 & D_{11} & D_{12} \\ C_2 & D_{21} & D_{22} \end{bmatrix} \begin{bmatrix} x \\ w \\ u \end{bmatrix}$

where:

• $A$ is the state matrix
• $B_1$ maps exogenous inputs to states, $B_2$ maps control inputs to states
• $C_1$ maps states to regulated outputs, $C_2$ maps states to measured outputs
• The $D$ matrices capture direct feedthrough terms

This compact representation encodes the nominal dynamics, performance specs, and uncertainty structure all in one object.

Weighting functions for performance specifications

Weighting functions translate your engineering requirements into the mathematical optimization problem. They are frequency-dependent transfer functions that shape how much you penalize different error signals at different frequencies.

Typical choices:

• A weight $W_1(s)$ on the sensitivity function $S(s)$: made large at low frequencies to enforce good disturbance rejection and reference tracking.
• A weight $W_2(s)$ on the control signal: used to penalize excessive control effort or to reflect actuator bandwidth limits.
• A weight $W_3(s)$ on the complementary sensitivity function $T(s)$: made large at high frequencies to enforce robustness to unmodeled dynamics and sensor noise.

Selecting these weights is the primary design knob in H-infinity control. The weights encode the fundamental trade-off: you can't make both $S$ and $T$ small at the same frequency, since $S(s) + T(s) = I$.

Sensitivity and complementary sensitivity functions

The sensitivity function is:

$S(s) = (I + G(s)K(s))^{-1}$

It describes how disturbances at the plant output propagate to the error signal. Small $S$ at a given frequency means good disturbance rejection there.

The complementary sensitivity function is:

$T(s) = G(s)K(s)(I + G(s)K(s))^{-1}$

It describes the closed-loop transfer function from reference to output (in a unity-feedback setup). It also determines how sensor noise and unmodeled high-frequency dynamics affect the system.

Since $S(s) + T(s) = I$, you face a fundamental trade-off: pushing $S$ down at some frequency forces $T$ up, and vice versa. The H-infinity framework handles this trade-off systematically through the weighting functions.

H-infinity controller synthesis

Once the problem is formulated, you need to actually compute the controller. Several mathematical techniques exist, each with different trade-offs in terms of computational complexity and generality.

Algebraic Riccati equations

The classical approach to H-infinity synthesis reduces the problem to solving two coupled Algebraic Riccati Equations (AREs):

1. The control ARE, which yields the state-feedback gain (analogous to the LQR problem).
2. The filtering ARE, which yields the observer gain (analogous to the Kalman filter problem).

For a solution to exist, both AREs must have stabilizing solutions, and a coupling condition between the two solutions must be satisfied. Specifically, if $X_\infty$ and $Y_\infty$ are the respective solutions, the spectral radius condition $\rho(X_\infty Y_\infty) < \gamma^2$ must hold, where $\gamma$ is the prescribed performance level.

State-space solutions for H-infinity control

The state-space solution procedure follows these steps:

1. Set up the generalized plant $P(s)$ in state-space form with the appropriate weighting functions.
2. Choose a performance level $\gamma > 0$ (the target upper bound on the H-infinity norm).
3. Solve the control ARE for $X_\infty \geq 0$.
4. Solve the filtering ARE for $Y_\infty \geq 0$.
5. Check the coupling condition $\rho(X_\infty Y_\infty) < \gamma^2$.
6. If all conditions are met, construct the controller from $X_\infty$, $Y_\infty$, and the plant matrices.

The resulting controller is a dynamic output-feedback controller whose order equals the order of the generalized plant. This can lead to high-order controllers, which is one practical drawback of the method.

Suboptimal H-infinity controller design

Finding the exact optimal $\gamma^*$ (the smallest achievable H-infinity norm) is rarely done directly. Instead, a gamma iteration (bisection) procedure is used:

1. Pick an initial range $[\gamma_{\text{low}}, \gamma_{\text{high}}]$.
2. Test $\gamma = (\gamma_{\text{low}} + \gamma_{\text{high}})/2$ by attempting to solve the two AREs.
3. If a stabilizing solution exists, set $\gamma_{\text{high}} = \gamma$. If not, set $\gamma_{\text{low}} = \gamma$.
4. Repeat until $\gamma_{\text{high}} - \gamma_{\text{low}}$ is below a desired tolerance.

The controller obtained at the final feasible $\gamma$ is called a suboptimal H-infinity controller. In practice, suboptimal controllers are perfectly adequate and often preferred because the exact optimum can be numerically sensitive.

Youla parameterization for H-infinity control

Youla parameterization (also called Q-parameterization) provides an alternative route to H-infinity synthesis. The key idea:

• Every stabilizing controller for a given plant can be written in terms of a free stable transfer function $Q(s)$, called the Youla parameter.
• The closed-loop transfer function from $w$ to $z$ becomes an affine function of $Q(s)$.
• Minimizing the H-infinity norm over $Q$ is then a convex optimization problem.

This convexity is a major advantage. It means there are no local minima, and efficient numerical solvers can find the global optimum. The practical challenge is that $Q(s)$ is infinite-dimensional, so it must be approximated (e.g., using a finite-dimensional basis of transfer functions).

Robust stability and performance

A controller that works perfectly for the nominal model but fails when parameters shift slightly is useless in practice. Robust stability and performance analysis tells you how much uncertainty the closed-loop system can tolerate.

Structured and unstructured uncertainties

Unstructured uncertainties are modeled as norm-bounded perturbations without specific internal structure. They're typically represented as frequency-dependent transfer functions:

• Multiplicative uncertainty: the true plant is $G_\Delta(s) = (I + W(s)\Delta(s))G(s)$ where $\|\Delta\|_\infty \leq 1$
• Additive uncertainty: $G_\Delta(s) = G(s) + W(s)\Delta(s)$
• These capture unmodeled dynamics, neglected nonlinearities, and high-frequency behavior

Structured uncertainties have known internal structure, typically parametric:

• Variations in physical parameters like mass, damping, or stiffness
• Represented as a block-diagonal perturbation matrix $\Delta = \text{diag}(\delta_1 I, \delta_2 I, \ldots)$
• Each block corresponds to a specific uncertain parameter

The distinction matters because structured uncertainty analysis (using $\mu$) is less conservative than unstructured analysis (using the small gain theorem).

Small gain theorem for robust stability

The small gain theorem provides a sufficient condition for robust stability:

For multiplicative uncertainty with weight $W(s)$, this translates to the condition $\|W(s)T(s)\|_\infty < 1$, meaning the complementary sensitivity must be small wherever the uncertainty is large.

The small gain theorem is simple and powerful, but it can be conservative when the uncertainty has known structure, because it treats $\Delta$ as a full-block perturbation.

Robust performance analysis with H-infinity

Robust performance means the system meets its performance specifications for every plant in the uncertainty set, not just the nominal one. This is a stronger requirement than robust stability alone.

In the H-infinity framework, robust performance can be checked by augmenting the uncertainty structure with a fictitious "performance block." The system achieves robust performance if and only if the augmented system achieves robust stability. This elegant equivalence connects performance analysis directly to stability analysis tools.

μ-analysis for robustness assessment

The structured singular value $\mu$ overcomes the conservatism of the small gain theorem by accounting for the structure of $\Delta$:

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

where $\boldsymbol{\Delta}$ is the set of allowable perturbation structures.

Key properties of $\mu$:

• $\mu_\Delta(M) < 1$ at all frequencies guarantees robust stability for the structured uncertainty set
• Computing $\mu$ exactly is NP-hard in general, so upper and lower bounds are used in practice
• The upper bound comes from solving an LMI (D-scaling), the lower bound from a power iteration
• μ-synthesis (DK-iteration) alternates between synthesizing a controller (K-step) and computing the D-scales (D-step) to minimize the peak $\mu$ value

The gap between upper and lower bounds is usually small for practical problems, making $\mu$-analysis a reliable robustness assessment tool.

H-infinity loop shaping

H-infinity loop shaping bridges the gap between classical frequency-domain design intuition and modern H-infinity optimization. You first shape the open-loop response using familiar techniques, then use H-infinity synthesis to "robustify" the design.

Loop shaping design procedure

The procedure has two stages:

1. Shape the open-loop plant using pre-compensator $W_1(s)$ and post-compensator $W_2(s)$ so that the shaped plant $G_s(s) = W_2(s)G(s)W_1(s)$ has the desired open-loop characteristics (crossover frequency, roll-off rate, low-frequency gain).
2. Synthesize a robustifying controller $K_s(s)$ for the shaped plant using H-infinity optimization, specifically targeting the normalized coprime factor uncertainty problem.

The final controller is $K(s) = W_1(s) K_s(s) W_2(s)$. This two-stage approach lets you use your engineering judgment in step 1 while getting formal robustness guarantees from step 2.

McFarlane-Glover H-infinity loop shaping

This is the most widely used H-infinity loop shaping method. Its appeal comes from several properties:

• The H-infinity problem it solves has a closed-form solution (no gamma iteration needed).
• The achievable robustness margin $\epsilon_{\max}$ can be computed directly from the shaped plant, before synthesizing the controller.
• If $\epsilon_{\max}$ is too small (typically below about 0.3), it signals that the loop shape needs to be modified rather than that the H-infinity synthesis failed.

The robustness margin $\epsilon_{\max}$ is given by:

$\epsilon_{\max} = \left(1 - \|[G_s \ \ I]^T (I + G_s)^{-1}\|_H^2\right)^{1/2}$

where $\|\cdot\|_H$ denotes the Hankel norm. A larger $\epsilon_{\max}$ means the shaped plant is easier to robustly stabilize.

Coprime factorization for loop shaping

The shaped plant $G_s(s)$ is expressed as a ratio of stable transfer functions using a normalized coprime factorization:

$G_s(s) = \tilde{M}(s)^{-1}\tilde{N}(s)$

where $\tilde{M}$ and $\tilde{N}$ are stable and satisfy $\tilde{N}\tilde{N}^* + \tilde{M}\tilde{M}^* = I$.

The uncertainty model then perturbs these coprime factors:

$G_\Delta = (\tilde{M} + \Delta_M)^{-1}(\tilde{N} + \Delta_N)$

This is a very general uncertainty description. It captures simultaneous perturbations to the numerator and denominator dynamics and can represent a wide class of model errors, including changes in the number of right-half-plane zeros or unstable poles.

Robust stabilization with loop shaping

The H-infinity controller synthesized in the McFarlane-Glover framework guarantees robust stability against all coprime factor perturbations satisfying:

$\left\|\begin{bmatrix} \Delta_N \\ \Delta_M \end{bmatrix}\right\|_\infty < \epsilon_{\max}$

This provides a quantitative robustness guarantee tied directly to the loop shape you chose. If the robustness margin is insufficient, you go back and reshape the open-loop response. This iterative but intuitive workflow is why loop shaping remains popular in industrial applications where engineers need to understand and trust the design.

Applications of H-infinity control

H-infinity control has been successfully deployed across many engineering domains. Its strength lies in providing formal robustness guarantees for systems where model uncertainty is significant.

H-infinity control in aerospace systems

Aerospace systems are a natural fit for H-infinity control because aerodynamic models are inherently uncertain (they change with flight condition, altitude, and speed). Specific applications include:

• Flight control systems for aircraft operating across wide flight envelopes
• Attitude control of spacecraft with flexible appendages and sloshing propellant
• Flutter suppression in aeroelastic structures where unmodeled dynamics are critical

H-infinity techniques for process control

Chemical plants and refineries deal with significant model uncertainty from time-varying operating conditions, imprecise reaction kinetics, and transport delays. H-infinity controllers provide consistent performance across these variations, which is important for maintaining product quality and safety.

Robust control of power systems using H-infinity

Power grids face uncertainties from fluctuating renewable generation, variable loads, and changing network topology. H-infinity control has been applied to frequency regulation, voltage control, and damping of inter-area oscillations. The worst-case guarantees are particularly valuable here because grid instability can have cascading consequences.

H-infinity methods in automotive control

Active suspension, electronic stability control, and traction control all operate under uncertain conditions (road surface, vehicle loading, tire wear). H-infinity controllers handle these variations while maintaining ride comfort and safety across diverse driving scenarios.

Advanced topics in H-infinity control

Several extensions of the basic H-infinity framework address limitations of the standard approach or expand its applicability.

Mixed H2/H-infinity control

Mixed H2/H-infinity control optimizes both norms simultaneously:

• The H2 component minimizes average performance (e.g., RMS tracking error under stochastic disturbances).
• The H-infinity component bounds worst-case performance (e.g., peak gain under deterministic disturbances).

This is useful when you care about both typical operating conditions and worst-case scenarios. The problem is typically formulated as minimizing the H2 norm subject to an H-infinity constraint (or vice versa), and solved using LMI techniques.

LMI-based H-infinity controller synthesis

Linear Matrix Inequalities (LMIs) reformulate the H-infinity problem as a convex optimization:

• The Riccati equation conditions are replaced by equivalent LMI conditions (via the Bounded Real Lemma).
• Additional constraints like pole placement regions, H2 bounds, or actuator limits can be added as extra LMIs.
• Efficient interior-point solvers (e.g., SeDuMi, MOSEK) handle the resulting semidefinite programs.

The LMI approach is more flexible than the Riccati approach because you can stack multiple objectives and constraints into a single optimization. The trade-off is that computational cost grows with problem size.

Nonlinear H-infinity control

For nonlinear systems, the H-infinity problem is formulated using differential game theory. The controller and the disturbance are treated as opposing players in a zero-sum game:

• The controller tries to minimize a cost functional.
• The disturbance tries to maximize it.
• The solution satisfies the Hamilton-Jacobi-Isaacs (HJI) equation, which is the nonlinear analog of the Riccati equation.

Solving the HJI equation analytically is generally intractable, so approximation methods (Taylor series expansion, neural networks, sum-of-squares programming) are used. This remains an active area of research.

Adaptive H-infinity control strategies

Adaptive H-infinity control addresses systems where the uncertainty is not just bounded but actively changing over time. The controller adjusts its parameters online based on estimated plant behavior:

• Parameter estimation runs concurrently with the H-infinity controller.
• Controller gains are updated to maintain the H-infinity performance bound as the plant changes.
• This combines the formal robustness guarantees of H-infinity control with the flexibility of adaptive control.

The main challenge is ensuring stability during the adaptation transient, which requires careful design of the adaptation law and persistence of excitation conditions.