1 Overview of Sensitivity and Complementary Sensitivity

1.1 Sensitivity matrix (S) and its role

In robust control, the sensitivity matrix \(S\) quantifies how errors in the loop—such as plant/model mismatch or internal dynamics—affect the controlled output when tracking a reference command. In multivariable (MIMO) settings, \(S\) is a matrix-valued transfer function: each entry describes how a particular perturbation channel propagates to each output channel, capturing interactions between multiple inputs and outputs.

1.2 Complementary sensitivity matrix (T) intuition

The complementary sensitivity matrix \(T\) describes how the closed loop routes disturbances and uncertainty-related effects to the output through the “complementary” path of the feedback interconnection. A key intuition is that \(S\) and \(T\) partition the influence of different signals: one path emphasizes the response to loop-breaking errors, while the other emphasizes how signals that enter through other locations are seen at the output.

1.3 Relationship between S and T (loop partitioning)

In standard feedback interconnections, \(S\) and \(T\) are complementary in the sense that their sum relates to an identity operator (often expressed as \(S + T = I\) under common conventions). This identity formalizes that, for a given internal signal location, the closed loop splits the system’s effect into components associated with sensitivity and complementary sensitivity. As a result, reducing sensitivity over a frequency band tends to increase complementary sensitivity there, and vice versa.

1.4 Frequency-domain viewpoint (Bode/transfer behavior)

Frequency-domain analysis treats \(S(j\omega)\) and \(T(j\omega)\) as complex matrix functions. Their magnitudes and singular values characterize how the loop behaves across low, mid, and high frequencies—commonly interpreted through Bode-style plots or matrix singular value plots. Typically, good tracking corresponds to small \(S\) at low frequencies, while good disturbance rejection and noise management depend on shaping \(T\) appropriately over the frequencies where disturbances or measurement noise dominate.

2 Mathematical Definitions in Control Frameworks

2.1 Standard closed-loop interconnection form

Consider a standard negative-feedback MIMO interconnection where a controller \(K\) drives the plant \(P\). The loop gain (often denoted \(L\)) combines the controller and plant dynamics in the forward path. With appropriate signal definitions, the closed-loop system can be written so that transfer matrices from reference, disturbances, or measurement noise to the outputs are expressed directly using \(S\) and \(T\).

2.2 Expression of T in terms of loop transfer L

Under common conventions for negative feedback, the sensitivity and complementary sensitivity matrices satisfy \[ S = (I + L)^{-1}, \quad T = (I + L)^{-1}L. \] Equivalent forms are frequently used, such as \(T = LS\) when dimensions and signal ordering align. These expressions make explicit that \(T\) depends on the same loop gain as \(S\), thereby tying output propagation directly to the closed-loop loop gain across frequencies.

In Youla (doubly coprime) parameterization, many closed-loop transfer functions—including sensitivity-type quantities—can be expressed using a free parameter that stabilizes a chosen nominal plant structure. In that framework, \(S\) and \(T\) become structured functions of the Youla parameter, clarifying which degrees of freedom are available to shape robustness and performance. While explicit formulas depend on the chosen coprime factorization and interconnection, the conceptual point is that \(T\) is not an independent design object; it is constrained by the closed-loop parameterization.

2.4 MIMO formulation and matrix dimensions

For a system with \(m\) control inputs and \(p\) measured outputs (or \(p\) controlled outputs), the complementary sensitivity matrix \(T\) is typically \(p \times p\) when it maps output-related internal signals to outputs. The loop gain \(L\) must have compatible dimensions (often \(p \times p\) in square loop formulations). In non-square cases, additional care is required: complementary sensitivity may be defined relative to particular input-output partitions and generalized inverse structures, depending on how the feedback interconnection is formulated.

3 State-Space Formulation

3.1 Realization of the closed-loop system

State-space models represent the closed-loop dynamics as a combined system containing plant and controller states. A common approach constructs an overall realization whose state vector stacks plant and controller states, and whose output equation yields the plant outputs of interest. Once a consistent closed-loop realization is available, complementary sensitivity can be computed as a transfer matrix from a chosen disturbance/measurement channel to outputs.

3.2 Computing complementary sensitivity from state-space

3.2.1 Transfer from disturbance inputs to outputs

To obtain \(T(s)\) directly, one selects signal locations consistent with the definition of complementary sensitivity (for example, a disturbance injected at a point whose effect at the output corresponds to the “complementary” channel). In state-space form, if the overall closed-loop realization is \[ \dot x = A x + B_w w + B_u u,\quad y = C x + D_w w + D_u u, \] then the transfer from \(w\) to \(y\) is \[ T(s) = C(sI-A)^{-1}B_w + D_w, \] with \(w\) chosen to represent the internal signal corresponding to the complementary path.

3.2.2 Direct-feedthrough and properness considerations

In many control architectures, the complementary sensitivity transfer matrix may include a direct term \(D_w\). Whether the resulting \(T(s)\) is proper (or strictly proper) depends on relative degrees in the chosen state-space realization. Properness matters for interpretation in frequency-weighted norms and for numerical evaluation at high frequencies. If the controller or sensor dynamics introduce algebraic loops, the realization may require algebraic loop resolution to produce a well-defined transfer matrix.

3.3 Numerical aspects (conditioning and realization choices)

Computing \(T\) from state-space data can be numerically sensitive, especially when \(A\) has eigenvalues near the imaginary axis or when the system order is high. Realization choices—minimality, balancing, and elimination of uncontrollable/unobservable modes—affect conditioning. Additionally, direct computation of \((sI-A)^{-1}\) at many frequencies can be costly; state-space-to-transfer conversion should be done with care, and frequency responses are often computed using numerically stable linear solvers rather than explicit inversion.

4 Disturbance Rejection and Robustness Interpretation

4.1 Disturbances at plant input vs measurement paths

Complementary sensitivity is closely tied to where disturbances enter the interconnection. Disturbances injected in different locations can map to different closed-loop transfer paths: one may be governed primarily by \(S\), while another is governed by \(T\). As a result, disturbance rejection design often requires identifying the dominant disturbance channels (e.g., actuator-side vs sensor-side) and shaping the corresponding sensitivity quantity.

4.2 High-frequency behavior and noise propagation

At high frequencies, many controllers are designed to avoid excessive gain that would amplify measurement noise. Because complementary sensitivity describes how certain internal signals reach the output, the high-frequency magnitude of \(T(j\omega)\) often reflects noise amplification risk in those channels. In practice, \(T\) is used to assess whether the closed loop attenuates or transmits high-frequency disturbances to the measured outputs.

4.3 Robustness metrics using T

Robustness analysis frequently uses norms of sensitivity-related transfer matrices. Since complementary sensitivity captures how uncertainty or disturbances affecting the complementary path reach the outputs, its frequency-weighted norms can act as robustness/performance indicators. Typical metrics relate to worst-case amplification measured via induced norms (such as \(H_\infty\)) or structured singular value tools when uncertainty is present.

4.4 Trade-offs with bandwidth and performance

Shaping \(T\) and \(S\) simultaneously reveals fundamental performance trade-offs. Increasing loop gain in a bandwidth can improve tracking (associated with smaller \(S\) at low frequencies) but may increase complementary sensitivity in corresponding ranges, potentially worsening noise transmission or reducing robustness margins at other frequencies. Effective design balances bandwidth, actuator limits, noise characteristics, and uncertainty models by targeting appropriate complementary sensitivity behavior.

5 Loop-Shaping and Design Use

5.1 How T is influenced by controller tuning

Controller tuning changes the loop gain \(L\), which in turn determines \(T = (I+L)^{-1}L\). Adjusting gain, crossover frequency, or compensator pole/zero placement modifies how \(T\) behaves across frequencies. In general, aggressive tuning can force complementary sensitivity to approach the identity over certain bands (indicating strong transmission from the complementary channel to the outputs), while conservative tuning keeps \(T\) smaller but may reduce disturbance rejection or tracking performance.

5.2 Constraints on complementary sensitivity (norm bounds)

5.2.1 H∞ performance interpretation

A common objective is to limit the \(H_\infty\) norm of \(T\) or of a weighted version of \(T\). Such bounds can be interpreted as limiting the worst-case gain from the relevant complementary disturbance channels to the outputs over all frequencies. When paired with \(S\)-based constraints, these objectives help define a feasible performance envelope for the closed loop.

5.2.2 Interpreting singular value plots for MIMO

For MIMO systems, complementary sensitivity is a matrix and its largest singular value \(\bar\sigma(T(j\omega))\) indicates the worst-case amplification across all input directions at a given frequency. Singular value plots can reveal whether certain coupling directions dominate and whether peaks occur at frequencies where the controller is near resonance, saturating, or switching regimes. The distribution among singular values also indicates whether performance degradation is concentrated in a few modes or broadly spread.

5.3 Coupling considerations in MIMO controller design

In MIMO systems, cross-coupling means that complementary sensitivity is not simply a diagonal set of SISO behaviors. The controller may route disturbances in ways that reduce amplification in one output while increasing it in another, depending on the interaction structure. This coupling makes it necessary to interpret \(T\) using matrix norms and singular values rather than relying on individual channel gains.

5.4 Practical loop-shaping workflows

A typical workflow is to (i) identify reference, disturbance, and noise locations; (ii) define shaping goals via acceptable levels of \(S\) and \(T\) (often expressed through weighting functions); (iii) tune or synthesize a controller to meet these targets while ensuring internal stability; and (iv) validate using time/frequency-domain simulations and robustness checks. Complementary sensitivity is then used as a diagnostic tool: if noise is seen at the output, the frequency range where \(\bar\sigma(T)\) is elevated helps pinpoint the cause.

6 Norms, Singular Values, and Frequency Weighting

6.1 Operator norms for complementary sensitivity

Complementary sensitivity is analyzed using operator norms induced by energy measures. In the \(H_\infty\) setting, the relevant quantity is \[

\|T\|_\infty = \sup_{\omega} \bar\sigma(T(j\omega)),

\] which captures the maximum gain from the complementary channel to outputs over all frequencies. For other design contexts, different norms may be used to emphasize different signal classes or constraints.

6.2 Singular values and worst-case directions

At each frequency, singular vectors identify worst-case directions in the complementary input space that produce maximum output amplification. This is particularly important for MIMO plants, where disturbances or uncertainty may have arbitrary correlation structure. Singular value analysis thus provides both magnitude information (peaks and roll-off) and directional insight (which coupled modes are most sensitive).

6.3 Frequency weighting functions and shaping objectives

Weighting functions \(W(s)\) are commonly introduced so that a design objective becomes \(\|W(s)T(s)\|_\infty\). Weights emphasize frequency regions where disturbances are likely (or where performance matters most) and de-emphasize regions where limited improvement is acceptable. This approach aligns mathematical optimization with practical requirements such as bandwidth limits, sensor noise spectra, or actuator dynamics.

6.4 Interpretation of peaks and roll-off regions

Peaks in \(\bar\sigma(T(j\omega))\) indicate frequencies where the complementary channel is strongly transmitted to outputs. In robust designs, such peaks can correlate with reduced margins or with increased sensitivity to unmodeled dynamics near resonant frequencies. Roll-off behavior at high frequencies is often interpreted as attenuation of noise propagation; insufficient roll-off can manifest as high-frequency output jitter or amplified measurement errors.

7 Robustness Analysis Tools Involving Complementary Sensitivity

7.1 Small-gain type reasoning (conceptual)

Small-gain arguments provide a conceptual link between sensitivity-type transfer functions and robustness. The idea is that if the product of a transfer function and an uncertainty bound remains below a critical threshold across frequencies, closed-loop behavior remains stable. Complementary sensitivity enters such reasoning when the uncertainty affects complementary paths, translating bounded uncertainty magnitude into an achievable performance limitation.

Because complementary sensitivity quantifies transmission in the complementary path, reducing \(T\) (in an appropriate norm or weighted norm) directly improves attenuation of those channels. In many designs, one cannot arbitrarily reduce \(T\) everywhere due to the complementary relationship with \(S\) and the structure of the loop gain. Consequently, robustness-oriented design often seeks a balanced reduction: suppress transmission where it matters most while preserving achievable tracking and stability margins.

7.3 Uncertainty handling through robust stability/performance (overview)

Robust stability and robust performance frameworks use sensitivity-related transfers to relate model uncertainty to closed-loop behavior. While the specific tools vary (e.g., norm-based conditions or structured uncertainty analysis), complementary sensitivity typically appears in the expressions determining how output-relevant uncertainty impacts the closed loop. In practice, engineers combine these theoretical conditions with uncertainty models validated by identification or modeling error bounds.

7.4 Controller synthesis objectives expressed via T

In synthesis problems—such as \(H_\infty\) control or mixed-sensitivity optimization—objectives may explicitly involve \(T\), often through weighted norms. The resulting controller is designed so that the complementary sensitivity meets specified gain bounds, shaping disturbance/noise transmission and, indirectly, robustness margins. For MIMO systems, the synthesis typically handles matrix-valued constraints using singular value-based interpretations.

8 Special Cases and Example Scenarios

8.1 SISO complementary sensitivity example

In a single-input single-output loop, \(T(s)\) reduces to a scalar transfer function. The complementary relationship can often be visualized with Bode plots: low-frequency behavior reflects how the loop transmits complementary disturbances to the output, while high-frequency roll-off reflects noise attenuation capability. The scalar nature makes trade-offs transparent: improving one frequency region commonly worsens another due to the structural coupling between \(S\) and \(T\).

8.2 Simple MIMO coupling example

For a basic 2×2 MIMO plant with cross-coupling, complementary sensitivity becomes a matrix whose off-diagonal entries capture how disturbances in one channel affect outputs in another. Even if diagonal elements seem acceptable, large off-diagonal interactions can yield a high peak in \(\bar\sigma(T)\), indicating that worst-case disturbances aligned with coupled directions will still propagate strongly. This scenario motivates using matrix norms rather than channel-by-channel inspection.

8.3 Non-minimum phase considerations (conceptual)

When the plant has non-minimum phase behavior, limitations arise in achievable loop shaping. Complementary sensitivity may exhibit constrained peak/roll-off profiles: aggressive disturbance/noise attenuation goals can conflict with the fundamental zeros and phase characteristics of the plant. In such cases, designers rely on robust synthesis methods that incorporate constraints on complementary sensitivity while respecting achievable dynamics.

8.4 Time-delay and actuator dynamics effects (overview)

Actuator dynamics and time delays affect the effective loop gain \(L\), which then reshapes \(T\). Time delay often limits achievable bandwidth and can produce phase margins reduction, which may cause peaks in complementary sensitivity near crossover frequencies. Actuator bandwidth constraints similarly restrict high-frequency control action, influencing the attenuation of complementary disturbances and noise propagation reflected in \(T(j\omega)\).

9 Common Pitfalls and Implementation Notes

9.1 Confusing T with S (sign conventions and definitions)

Different texts and software tools may adopt varying sign conventions, signal definitions, and interconnection structures. A frequent error is to compute a quantity believed to be \(T\) while actually obtaining \(S\), or vice versa, due to swapped signal labels or sign errors in the block diagram. Correct identification requires verifying the definitions from the specific interconnection being used.

9.2 Reference-to-output vs disturbance-to-output interpretation

Complementary sensitivity is sometimes conflated with transfer functions from reference to output. However, \(T\) is typically tied to complementary disturbance/measurement channels rather than directly to reference tracking. Misinterpreting what excitation \(T\) corresponds to can lead to incorrect conclusions about tracking quality or disturbance rejection effectiveness.

9.3 Numerically computing T from interconnections

In computational workflows, \(T\) is often obtained by reducing a closed-loop transfer matrix from an interconnection model. Numerical pitfalls include inconsistent interconnection assumptions, algebraic loop resolution issues, and frequency grid selection that misses resonance peaks. Ensuring that the computed closed-loop model is internally stable and that the transfer extraction matches the intended signal ports helps avoid misleading results.

9.4 Ensuring consistent signal naming and causality assumptions

Transfer matrix computation depends on consistent signal naming (which port is the “complementary” input and which is the output) and on assumptions about causality and properness. If feedthrough terms are present, the model must reflect them correctly; otherwise, the resulting \(T\) may appear to violate expected high-frequency behavior. Careful auditing of the state-space realization and interconnection diagram is essential.

10.1 Sensitivity complement pairs in robust control

The sensitivity/complementary sensitivity pair is a central tool for understanding how feedback distributes the closed-loop effect between uncertainty-related error paths and complementary disturbance/noise transmission paths. Further study often compares how these matrices appear in standard robust control conditions and how their norms relate to performance guarantees.

10.2 Robust performance vs robust stability context

Robust stability focuses on keeping the closed loop stable under uncertainty, while robust performance additionally addresses how well outputs behave. Complementary sensitivity commonly appears more directly in robust performance formulations because it quantifies output propagation from disturbance/noise-like channels.

10.3 MIMO control fundamentals (cross-coupling)

Understanding MIMO complementary sensitivity benefits from familiarity with MIMO controller structure, cross-coupling effects, and how singular values represent worst-case gain over all coupled directions. References on MIMO frequency response and multivariable norm interpretation complement the study of \(T\).

Useful keywords include: “sensitivity function,” “complementary sensitivity,” “loop gain \(L\),” “MIMO singular value plots,” “robust performance,” “\(H_\infty\) control,” “Youla parameterization,” and “doubly coprime factorization.” Reading across these areas helps connect definitions, computations, and design practices involving the complementary sensitivity matrix.