Transfer function matrix

1 Definition and notation

A transfer function matrix is a compact description of how a linear time-invariant multivariable system maps an input signal vector to an output signal vector in the frequency domain. For an \(m\)-input, \(p\)-output system, it is represented as a \(p\times m\) matrix whose entries are scalar transfer functions. The \((i,j)\) element specifies the transfer from the \(j\)-th input to the \(i\)-th output.

1.1 Linear time-invariant multivariable systems

Consider a linear time-invariant (LTI) system with input vector \(u(t)\in\mathbb{C}^{m}\) and output vector \(y(t)\in\mathbb{C}^{p}\). Linearity allows superposition, and time invariance ensures that shifting inputs shifts outputs in the same way. Under these conditions, frequency-domain methods yield an algebraic relation between transforms of \(u(t)\) and \(y(t)\).

1.2 Transfer function matrix as an operator

The transfer function matrix functions as an operator that converts input spectra to output spectra. In the Laplace domain, the relationship is expressed as \[ Y(s)=G(s)\,U(s), \] where \(G(s)\) is the matrix-valued transfer function. This operator view emphasizes that the mapping is defined for each complex frequency \(s\) in a domain where the model is valid.

1.3 Common notation and dimensionality

A standard notation is \[ G(s)\in\mathbb{C}^{p\times m}, \] with elements \(G_{ij}(s)\). When the system is square, \(p=m\), the matrix is often central for invertibility and feedback calculations. In general, rectangular matrices describe over- or under-actuated signal paths, which affects which inversions and channel reconstructions are feasible.

2 Mathematical foundations

The use of transfer function matrices is grounded in the correspondence between linear differential equations (or state-space dynamics) and rational functions in the complex frequency variable. Matrix-valued rational functions inherit many properties from scalar transfer functions, but their interactions are captured through matrix operations.

2.1 Frequency-domain derivation

A transfer function matrix can be obtained from time-domain linear systems by applying transforms and using algebraic elimination of internal signals.

2.1.1 Laplace transform relationship

For LTI systems with zero initial conditions, Laplace transforms convert convolution-type relations into multiplication. If the system behavior can be written as a multiple-input/multiple-output convolution, \[ y(t)=g(t)*u(t), \] then after transforming, \[ Y(s)=G(s)\,U(s), \] where \(G(s)\) is the Laplace transform of the matrix impulse response.

2.1.2 State-space to transfer matrix conversion

In state-space form, \[ \dot{x}(t)=Ax(t)+Bu(t),\qquad y(t)=Cx(t)+Du(t), \] with \(x(t)\in\mathbb{C}^{n}\). Under zero initial conditions, the Laplace transform yields an explicit expression for \(G(s)\) in terms of matrices \(A,B,C,D\).

2.1.3 Convolution interpretation in the time domain

Each scalar element \(G_{ij}(s)\) corresponds to a convolution kernel \(g_{ij}(t)\) such that \[ y_i(t)=\sum_{j=1}^{m} g_{ij}(t) * u_j(t). \] This makes the matrix representation more than a shorthand: it encodes a network of impulse responses that collectively determine the output vector.

2.2 Properties of rational matrix functions

For many physical models, each entry of \(G(s)\) is a ratio of polynomials, yielding a rational matrix function. The matrix structure influences how poles and zeros interact across channels.

2.2.1 Poles, zeros, and pole-zero cancellations

Poles of \(G(s)\) arise from common factors in denominators associated with the system dynamics. Zeros relate to frequencies where the mapping from inputs to outputs loses rank in certain directions. In rational matrix functions, pole-zero cancellations may occur entrywise or through structured cancellations involving multiple elements, sometimes affecting minimality and observability/controllability.

2.2.2 McMillan degree and model order

The McMillan degree generalizes the idea of system order to transfer functions. For a matrix transfer function, it reflects the dimension of a minimal realization consistent with \(G(s)\). Two different state-space descriptions may realize the same input-output behavior, but their minimal degrees coincide.

2.3 System causality and properness

Causality and implementability impose restrictions on how transfer functions behave as \(s\to\infty\).

2.3.1 Strictly proper vs proper matrix elements

A scalar transfer function is strictly proper if its numerator degree is strictly less than its denominator degree; it is proper if degrees are equal or numerator degree is less. For matrices, properness applies to each entry. A strictly proper element implies that its high-frequency gain tends to zero, corresponding to dynamics that cannot respond instantaneously.

2.3.2 Strictly proper condition and realizability constraints

When the direct feedthrough matrix \(D\) is zero, the overall transfer matrix is strictly proper under typical conventions. If nonzero \(D\) exists, the system exhibits a direct, instantaneous mapping from input to output. These distinctions matter when constructing realizations from measured data and when ensuring a model can be physically implemented without algebraic loops in feedback interconnections.

3 Matrix structure and signal coupling

The matrix form reveals how signals interact across channels, distinguishing between pathways that act independently and those that exchange information dynamically.

3.1 Diagonal vs off-diagonal channels

A diagonal transfer function matrix indicates that each output depends only on the corresponding input (under a specific channel ordering). Off-diagonal entries represent coupling.

3.1.1 Decoupled systems

If \(G(s)\) is diagonal (or can be made diagonal by simple relabeling), the system behaves like a set of independent SISO channels. In this case, analysis and controller design can often be carried out channel-by-channel without considering cross-effects.

3.1.2 Coupled multivariable dynamics

When off-diagonal elements are nonzero, input changes in one channel can influence multiple outputs. Coupling can be strong or weak, frequency-dependent, and sometimes localized to certain frequency bands, which complicates controller design because naive independent tuning may fail.

3.2 Relative gain and interaction measures

Coupling is often assessed through measures that reflect how effectively inputs can drive outputs along particular directions.

3.2.1 Input-output pairing interpretation

The \((i,j)\) element \(G_{ij}(s)\) describes how input component \(u_j\) contributes to output component \(y_i\). However, system-level behavior is more accurately described by how the entire matrix maps input vectors to output vectors, not by single entries alone.

3.2.2 Channel significance and dominance

Dominance can be frequency dependent: a channel pair with small \(	G_{ij}(j\omega)	\) at one frequency can become influential at another. For this reason, interaction assessments often rely on norms, singular values, or directional gains rather than a single scalar magnitude.

4 Connections to state-space models

State-space realizations provide a dynamical model whose transfer function matrix is obtained through algebraic operations. Conversely, a transfer matrix can be realized by constructing state-space matrices that reproduce it.

4.1 State-space to transfer function matrix

A standard relationship links the state-space system matrices to the transfer function matrix.

4.1.1 Deriving G(s) = C(sI − A)⁻¹B + D

Using Laplace transforms with zero initial conditions, one obtains \[ G(s)=C(sI-A)^{-1}B + D. \] The term \(C(sI-A)^{-1}B\) represents the dynamic portion through the internal state, while \(D\) captures any direct input-to-output feedthrough.

4.2 Transfer function matrix to state-space realizations

Constructing a state-space model from a transfer matrix is central in system identification and model reduction.

4.2.1 Minimal realizations concept

A realization is minimal if it uses the smallest possible state dimension while reproducing the same transfer matrix. Minimality corresponds to controllability and observability conditions in the state-space representation, which also connect to the McMillan degree.

4.2.2 Realization non-uniqueness

Even when the transfer matrix is fixed, the matrices \(A,B,C,D\) are not unique: different minimal realizations are related by similarity transformations. This non-uniqueness does not change input-output behavior, but it influences numerical aspects and parameter interpretation.

4.3 Feedthrough terms and direct paths

The matrix \(D\) determines whether outputs depend instantaneously on inputs. In frequency response, \(D\) contributes directly and can dominate high-frequency behavior when the dynamic part decays. In feedback interconnections, feedthrough terms can also affect well-posedness by creating algebraic coupling.

5 Frequency response forms

In frequency-domain analysis, the transfer function matrix is evaluated on the imaginary axis to study steady-state sinusoidal response.

5.1 Evaluating at s = jω

For real frequency \(\omega\), the frequency response is typically \[ G(j\omega), \] with evaluation performed wherever the model is defined. This form supports visualization and comparison to experimental frequency measurements.

5.2 Steady-state gain and resonance behavior

Magnitude patterns of \(G(j\omega)\) reveal how strongly different inputs and frequencies contribute to outputs. Near system poles, the response can become large (or indicate sensitivity), while zeros can suppress certain output components for particular input directions.

5.3 Magnitude/phase interpretations for matrices

Matrix-valued frequency responses contain both amplitude and phase relationships across channels, which cannot be fully captured by a single magnitude value.

5.3.1 Singular values vs element-wise views

Element-wise magnitudes \(	G_{ij}(j\omega)	\) show pairwise contributions but ignore cancellations between channels. Singular values summarize the worst-case gain over input directions and identify how the matrix stretches or compresses vectors at each frequency.

5.3.2 Directional gains via input/output vectors

For a given input direction \(v\) and corresponding output direction, one can examine how \(G(j\omega)\) maps \(v\) into output space. This directional perspective aligns with robust control and provides a clearer sense of effective excitation and transmission pathways.

6 Analysis tools using the transfer function matrix

The transfer function matrix enables stability and solvability analysis in multivariable settings, where channel interactions must be treated consistently.

6.1 Stability considerations in multivariable context

Multivariable stability analysis connects pole locations and closed-loop interconnections, but the matrix structure determines how modes are excited and observed.

6.1.1 Open-loop poles and closed-loop implications

Open-loop poles of \(G(s)\) indicate inherent dynamics, while closed-loop poles depend on controller structure and interconnection. Because feedback can couple channels, the closed-loop spectrum is not determined solely by individual scalar channels; it requires the combined system description.

6.2 Well-posedness and regularity

Well-posedness ensures that interconnections produce a consistent input-output mapping without algebraic contradictions.

6.2.1 Invertibility of relevant matrices

Certain calculations involve inverses such as \((I + L(s))^{-1}\), where \(L(s)\) is formed from plant and controller transfer matrices. If these matrices lose invertibility at some frequencies, the closed-loop transfer may become undefined or exhibit singular behavior, indicating problematic interconnection structure.

6.3 Right/left invertibility concepts

Invertibility in the multivariable sense depends on whether the matrix mapping can be undone over frequency.

6.3.1 Input reconstruction and solvability

A transfer matrix is right-invertible if an appropriate input can be found to achieve a desired output (within the range of the mapping). Left-invertibility relates to the ability to infer inputs from outputs. Rectangular matrices may be invertible only in one sense, affecting identifiability and reconstruction tasks.

7 Interconnection and block diagram manipulation

Transfer function matrices support algebraic manipulation mirroring block diagram operations such as series, parallel, and feedback combinations.

7.1 Series, parallel, and feedback combinations

These operations are described through matrix products and sums, often with additional attention to dimensions and causality.

7.1.1 Series connection of transfer matrices

If system \(G_1(s)\) maps \(u\) to an intermediate signal, and system \(G_2(s)\) maps that intermediate to the output, then the overall mapping is \[ G(s)=G_2(s)\,G_1(s), \] subject to compatible dimensions.

7.1.2 Parallel superposition in MIMO form

If two subsystems act on the same input and their outputs are summed, the resulting mapping is \[ G(s)=G_1(s)+G_2(s), \] with the understanding that both terms produce outputs of the same dimension.

7.1.3 Feedback interconnection formulas

In feedback, the closed-loop mapping can be expressed using matrix inverses. The exact formula depends on whether the feedback is negative, how signals are routed, and the assumed direct feedthrough structure. The essential requirement is that the relevant “loop gain” combination remains invertible in the frequency region of interest.

7.2 Transformations and signal routing

Mathematical equivalences correspond to operations performed on block diagrams.

7.2.1 Permutations and scaling

Relabeling inputs or outputs corresponds to multiplying by permutation matrices, while scaling relates to multiplication by diagonal gains. These operations change the representation without altering the underlying physical dynamics, provided corresponding routing changes are applied consistently.

7.2.2 Basis changes and equivalence

Changing the basis of internal signals leads to transformed state-space matrices and, accordingly, transformed transfer matrices that preserve input-output behavior under consistent definitions. In multivariable analysis, basis changes are often used to align the model with measured variables or to simplify numerical conditioning.

8 Control and design applications

Transfer function matrices provide the language for multivariable control, where multiple coupled signals must be regulated simultaneously.

8.1 Multivariable (MIMO) control interpretation

In MIMO control, both the plant and controller are typically represented by transfer matrices, enabling systematic derivations of closed-loop behavior.

8.1.1 Controller and plant transfer matrices

A plant \(G(s)\) and controller \(K(s)\) interact through feedback. With matrix-valued signals, controller design must address cross-coupling, especially when actuators affect multiple outputs.

8.1.2 Closed-loop transfer matrix derivations

Closed-loop maps from reference and disturbance inputs to outputs are obtained by combining plant and controller transfer matrices with feedback algebra. The resulting expressions are matrix rational functions, and they reveal how coupling affects tracking, disturbance rejection, and control effort.

8.2 Loop shaping and frequency-domain design

Frequency-domain design leverages how transfer matrices behave across frequencies, often with metrics derived from singular values.

8.2.1 Using singular value plots

Singular value plots illustrate worst-case gains across input directions at each frequency. Designers use these to ensure that closed-loop sensitivity and complementary sensitivity meet performance goals while respecting constraints implied by stability and robustness.

8.2.2 Robustness-oriented frequency metrics

Robustness metrics often require combining plant dynamics, controller dynamics, and interconnection structure into matrix expressions whose invertibility and norms govern performance under uncertainty. These metrics help identify frequency regions where modeling errors or disturbances could amplify.

8.3 Decoupling and diagonalization approaches

Decoupling aims to reduce interaction so each output can be treated with simpler control laws.

8.3.1 Static decouplers

A static decoupler is a frequency-independent transformation applied to inputs or outputs to reduce coupling. It can improve interaction for certain plants but may fail when coupling is strongly frequency dependent or when constraints prevent perfect diagonalization.

8.3.2 Dynamic decoupling intuition

Dynamic decouplers introduce frequency-dependent compensation to cancel or mitigate coupling over selected bands. While more flexible than static methods, dynamic schemes require careful realization and often involve trade-offs between performance and model complexity.

9 Estimation and identification

In system identification, the transfer function matrix is a target model that can be inferred from experimental data.

9.1 Data-driven model fitting

Model fitting methods estimate the matrix-valued transfer function that best matches observed input-output behavior.

9.1.1 Frequency response fitting

Frequency response fitting uses measured \(u(t)\) and \(y(t)\) under sinusoidal or broadband excitations to estimate \(G(j\omega)\) across frequencies. Because the system is multivariable, the output-to-output and input-to-input correlations guide the estimated matrix structure.

9.1.2 Time-domain estimation approaches

Time-domain approaches estimate a model directly from sampled time series, often using least squares or subspace methods. The resulting model can be converted to a transfer function matrix representation for analysis and controller design.

9.2 Regularization and numerical conditioning

Identification in multivariable settings can suffer from ill-conditioning, particularly when inverting approximate models or resolving near-colliding dynamics.

9.2.1 Noise sensitivity in matrix inversions

Estimating rational matrix functions may involve inversions of data-derived matrices. Measurement noise can cause large errors when those matrices are close to singular, leading to unstable estimates and requiring robust numerical techniques.

9.3 Validation of identified transfer matrices

Validation checks whether the identified model reproduces behaviors not used in fitting.

9.3.1 Model checking with independent data

Independent datasets test predictive accuracy, comparing measured and simulated outputs across operating conditions. In MIMO cases, validation should examine not only overall error norms but also channel-specific and direction-specific mismatches.

10 Computational aspects

Computation of transfer function matrices requires careful handling of numerical issues arising near poles and in large-scale systems.

10.1 Numerical evaluation near poles and singularities

Evaluating \(G(j\omega)\) can be challenging when \(sI-A\) becomes nearly singular.

10.1.1 Conditioning of (sI − A)⁻¹

The conditioning of \((sI-A)^{-1}\) governs numerical accuracy. Large condition numbers can amplify floating-point errors, so algorithms may use stable decompositions, regularization, or cautious step sizes in frequency sweeps.

10.2 Efficient handling of large MIMO systems

Large MIMO models can be computationally heavy in both evaluation and optimization loops.

10.2.1 Order reduction for practical computation

Order reduction approximates a high-dimensional model with a smaller one that preserves key input-output dynamics over a frequency range. This reduces computational load while maintaining accuracy for the intended design or identification tasks.

10.3 Symbolic vs numerical transfer matrix workflows

Symbolic manipulation can simplify expressions but may become unwieldy as model size grows. Numerical workflows, in contrast, compute evaluations and fit parameters directly, often providing better scalability for engineering use.

11 Special cases and examples

Special configurations of transfer function matrices provide intuition and help interpret general results.

11.1 SISO as a degenerate case

When there is one input and one output (\(m=p=1\)), the transfer function matrix reduces to a scalar transfer function. Many matrix notions, such as poles and zeros, collapse to the classical SISO definitions.

11.2 Symmetric and reciprocal systems where applicable

Some systems satisfy symmetry relations that constrain \(G(s)\) (for example, reciprocal behavior that links certain transfer elements). Such properties can simplify analysis and reduce the number of independent parameters to estimate, though exact applicability depends on modeling assumptions.

11.3 Example: 2×2 coupled system walkthrough

Consider a \(2\times 2\) transfer matrix with diagonal terms representing direct channel dynamics and off-diagonal terms representing cross-coupling: \[ G(s)=\begin{bmatrix} G_{11}(s) & G_{12}(s)\\ G_{21}(s) & G_{22}(s) \end{bmatrix}. \] If \(G_{12}(s)\) and \(G_{21}(s)\) share resonance peaks with the diagonal terms, then inputs applied to one actuator can strongly affect both outputs near those frequencies. In a decoupling attempt, a controller tuned using only \(G_{11}\) and \(G_{22}\) may show unexpected interactions because the off-diagonal pathways inject additional dynamics into the controlled outputs.

11.4 Interpreting results from example transfer matrices

From such an example, one can interpret the frequency response in two complementary ways. Element-wise inspection highlights which input-output pairs are large at each frequency, while matrix-level analysis (e.g., singular values or directional gains) indicates whether overall transmission is strong along certain input directions and weak along others. Comparing these viewpoints clarifies whether coupling is merely local to a channel pair or indicative of a global interaction structure.