Robust control and the small-gain theorem

Use causal operators G and Δ mapping compatible vector-valued signals with zero initial conditions. Assume their truncated L² gains are bounded uniformly on every finite horizon: the output norm through time T is at most a fixed γ times the input norm through T. This extended-signal formulation permits a finite-horizon argument before knowing the full closed-loop output has finite energy. Assume the feedback relation y=G(w+Δy) is well posed, meaning it has a unique causal solution for each admissible external input. A gain inequality alone should not be used to ignore an undefined algebraic loop.

For a stable finite-dimensional proper LTI transfer matrix, its induced L² gain is the supremum over frequency of the largest singular value of G(iω). This is the H-infinity norm in this stable setting. The singular value, rather than a maximum of individual matrix entries, accounts for coordinated multi-input directions. A pure delay has unit L² gain despite changing phase. A static scalar multiplier k has gain |k|. These facts allow certain uncertain dynamics to be bounded without pretending their detailed frequency response is known.

Apply the G bound, then the triangle inequality, then the Δ bound. This yields ||y||≤γ_G||w||+γ_Gγ_Δ||y|| on each finite horizon. If the product is strictly below one, move the last term to the left. The resulting bound is uniform in T, so finite-energy w produces finite-energy y with gain at most γ_G/(1−γ_Gγ_Δ). Other internal connection signals are bounded through the same equations. This proves a finite-gain input–output statement under the chosen zero-state operator assumptions.

The strict inequality is essential to this proof; equality leaves a zero denominator and no bound. It does not follow that every individual system at or above the threshold is unstable. Phase and uncertainty structure may prevent reinforcing feedback, so the norm test can be conservative for a specified loop. For robust claims over a whole uncertainty class, state exactly which causal operators are allowed and how their gains are bounded. A point estimate from a few measured frequencies is not a certified global frequency supremum.

Input–output stability of a transfer map does not reveal an unstable hidden state canceled from that map. If the requirement is internal stability of a finite-dimensional realization, the nominal subsystems must have internally stable realizations and the interconnection must be well posed, with all relevant ports covered by the analysis. This reconnects the first lesson’s hidden-mode warning to robust design. Merely simplifying a rational expression and observing stable remaining poles can erase the very unstable dynamics one intended to control.

Additive plant uncertainty, multiplicative uncertainty and sensor uncertainty enter different block diagrams and generally produce different nominal operators G in the small-gain test. Weighting functions can encode frequency-dependent uncertainty, but their units and stable realizability matter. The certificate concerns the modeled class; unmodeled saturation or a nonlinear bias is covered only if it actually satisfies the assumed gain description. Robust stability also differs from good performance: a finite bound can be very large near product one, so a loop may be certified stable yet amplify disturbances beyond acceptable limits.