H-infinity control

H-infinity control is an optimization-based framework in robust control that designs controllers to minimize the worst-case gain from disturbances to regulated outputs, measured by the H-infinity norm of the closed-loop transfer function. Introduced by George Zames in 1981, it replaces the classical focus on average performance with a guarantee: no disturbance within the assumed class can produce an output exceeding a specified bound.

The H-infinity norm of a stable transfer function \(G(s)\) is its peak gain over all frequencies:

\[\|G\|_\infty = \sup_\omega |G(j\omega)|\]

This is not an average; it is a worst-case. The H-infinity controller is the controller that minimizes this worst-case gain, making it the natural framework for problems where peak performance matters more than average performance — aircraft flight control, process control with safety constraints, and any system where a single bad event is unacceptable.

The Standard Problem

The H-infinity design problem is typically formulated as a standard plant with generalized inputs and outputs. The plant \(P(s)\) has two sets of inputs — the exogenous disturbances \(w\) and the control inputs \(u\) — and two sets of outputs — the regulated outputs \(z\) (to be minimized) and the measured outputs \(y\) (available to the controller). The controller \(K(s)\) maps measured outputs to control inputs. The design objective is to find \(K(s)\) that stabilizes the closed loop and minimizes the H-infinity norm of the transfer function from \(w\) to \(z\).

This formulation is general enough to subsume nearly every linear control design problem: disturbance rejection, reference tracking, robust stability, and mixed-sensitivity problems can all be cast as instances of the standard problem with appropriate choice of \(P(s)\).

The State-Space Solution

The breakthrough that made H-infinity control practical came in 1989, when Doyle, Glover, Khargonekar, and Francis showed that the optimal H-infinity controller can be computed by solving two algebraic Riccati equations. The solution has the elegant structure of a state-feedback gain combined with a state estimator — the same separation principle that underlies LQG control — but with Riccati equations modified to account for the H-infinity criterion.

The state-space solution revealed that H-infinity control is not merely a computational technique but a structural theory. The controller's order equals the plant's order, and the solution can be obtained in polynomial time. This transformed H-infinity design from a theoretical curiosity into a practical engineering tool.

Loop Shaping and Mixed Sensitivity

In practice, H-infinity design is often applied through mixed-sensitivity minimization, where the regulated output \(z\) combines weighted versions of the sensitivity function \(S(s)\) and the complementary sensitivity function \(T(s)\). The design minimizes:

\[\left\| \begin{matrix} W_S S \\ W_T T \\ W_{KS} KS \end{matrix} \right\|_\infty\]

where \(W_S\), \(W_T\), and \(W_{KS}\) are frequency-dependent weights that encode the designer's priorities. \(W_S\) is large at frequencies where disturbance rejection matters; \(W_T\) is large at frequencies where sensor noise must be attenuated; \(W_{KS}\) limits control effort. The H-infinity optimization finds the controller that best satisfies these competing objectives in the worst-case sense.

This is loop shaping with a guarantee. The classical engineer shapes the loop by intuition and iteration; the H-infinity engineer specifies the desired shape as weights and lets the optimization find the controller that achieves it — or proves that no controller can.

The Gamma Iteration

The optimal H-infinity controller minimizes the norm exactly, but in practice designers often settle for suboptimal controllers that achieve a norm below a specified bound \(\gamma\). The design proceeds by gamma iteration: choose \(\gamma\), test whether a controller achieving \(\|\cdot\|_\infty < \gamma\) exists (via the Riccati equations), and bisect until the infimum is found to desired precision.

This iterative structure is not a computational inconvenience. It is a design feature. The engineer can explore the trade-off between performance and robustness by examining how the controller changes as \(\gamma\) varies, and can select a controller that balances nominal performance (small \(\gamma\)) with robustness (controller not too aggressive).

H-infinity control is the mathematical formalization of engineering pessimism. It does not ask what happens on average; it asks what happens in the worst case. This makes it conservative, but the conservatism is principled — it is the price of a guarantee. The systems that matter — the ones that carry people, control critical infrastructure, manage financial risk — do not need optimality. They need guarantees. H-infinity control is the language of guarantees.