Rational matrix function

1 Basic definitions and representations

1.1 Polynomial matrix functions

1.1.1 Matrix polynomials and evaluation

A matrix polynomial in one indeterminate \(z\) over a field (often \(\mathbb{C}\)) is an expression of the form \[ P(z)=\sum_{k=0}^m P_k z^k, \] where each coefficient \(P_k\) is a matrix of fixed size (for example, \(n\times n\)). Evaluation is defined entrywise by substituting a scalar value of \(z\) into the polynomial, producing a matrix \(P(\zeta)\) whenever \(\zeta\) is taken from the underlying field.

1.1.2 Degree, leading coefficient, and polynomial matrix properties

The degree of \(P(z)\) is the largest \(k\) for which \(P_k\neq 0\). The leading coefficient is \(P_m\). Basic properties follow from how coefficients multiply under addition and multiplication: the degree of a product is at most the sum of degrees, and the leading term is determined by the product of leading coefficients when no cancellation occurs. For polynomial matrices, rank can vary with \(z\), and this variation becomes important when relating polynomial matrices to zeros and factor structure.

1.2 Rational matrix functions

1.2.1 Quotient form \(F(z)=P(z)Q(z)^{-1}\)

A rational matrix function is a matrix-valued function whose entries are rational functions of \(z\). A standard representation is the quotient form \[ F(z)=P(z)Q(z)^{-1}, \] where \(P(z)\) and \(Q(z)\) are polynomial matrices of compatible sizes, and \(Q(z)\) is square so that an inverse makes sense where it exists. This viewpoint connects rationality to algebra: singularities of \(F\) can arise precisely where \(Q(z)\) fails to be invertible.

1.2.2 Invertibility domain and regular points

The expression \(P(z)Q(z)^{-1}\) is defined on the set of \(z\) for which \(Q(z)\) is invertible. Points where \(Q(z)\) is singular are typically excluded from the literal quotient representation. However, a singularity may be removable if the singular behavior cancels with factors in \(P(z)\). Such issues motivate distinguishing between *a representation’s domain* and the *function’s intrinsic analytic behavior*.

1.3 Equivalent representations

1.3.1 Left/right factorizations

Two quotient representations may represent the same rational matrix function even when their individual factors differ. For example, multiplying a quotient representation by invertible factors on the left or right (with those factors unimodular in the polynomial sense) can leave the resulting function unchanged. Factorization language is therefore central: one can study rational matrix functions through alternative polynomial factorizations that reveal their pole/zero structure more transparently.

1.3.2 Clearing denominators and minimal common denominators

Given a representation with denominators, one can clear denominators to convert expressions into equivalent quotient forms with shared polynomial denominators. In matrix settings, this can be done by finding polynomial matrices that absorb common factors. A minimal common denominator concept aims to avoid introducing unnecessary factors that complicate invariant quantities. Canonical forms (such as Smith–McMillan) formalize the minimality idea by extracting the smallest invariant divisors compatible with the function.

2 Analytic structure and matrix-valued poles/zeros

2.1 Poles of rational matrix functions

2.1.1 Scalar vs. matrix poles (determinant perspective)

Poles can be described in terms of how \(F(z)\) behaves near a point \(z_0\). A useful determinant perspective is that if \(F(z)=P(z)Q(z)^{-1}\), then singularities often correlate with zeros of \(\det Q(z)\). Still, matrix poles are subtler than scalar poles: even when \(\det Q(z_0)=0\), some directions in the matrix may remain well-behaved because the singularity may act only on particular subspaces.

2.1.2 Order of poles and local behavior

The order of a pole is the degree to which \(F(z)\) diverges as \(z\to z_0\). In matrix terms, one can examine local Laurent expansions or valuation-like behavior via canonical forms. The presence of different invariant factors can yield different pole orders across different invariant subspaces, producing a structured rather than uniform blow-up.

2.2 Zeros and rank-deficiency points

2.2.1 Kernel and image structure near singularities

Zeros of \(F(z)\) are linked to rank loss: if \(F(z_0)\) drops rank, then there exist nonzero vectors \(x\) such that \(F(z_0)x=0\). Locally, how the kernel and image change with \(z\) influences the analytic order of zeros. Because kernel dimension and image dimension can change in structured ways, zero behavior is naturally encoded by invariant factors rather than by determinants alone.

2.2.2 “Zero” notions: determinant zeros and invariant factors

Different notions of “zero” coexist. A determinant zero is a point where \(\det F(z_0)=0\), but this does not fully capture how the matrix behaves on subspaces. Canonical invariant-factor notions refine this by expressing the rational matrix function in a form where zeros correspond to polynomial factors in the invariant divisors; this yields a more detailed pole/zero multiplicity structure.

2.3 Domain of definition

2.3.1 Cuts, excluded points, and meromorphic extension

Rational matrix functions are naturally meromorphic: they are holomorphic on a punctured domain and admit a controlled extension across isolated points. In the simplest single-variable setting, one typically excludes finitely many points where inverses fail in a chosen representation. In a more conceptual view, the function extends meromorphically by allowing poles of various orders. “Cuts” are mainly a concern when working with multi-valued functions; for rational functions, the singular set is discrete and extension is governed by pole behavior rather than branch cuts.

3 Algebraic invariants and canonical forms

3.1 Smith normal form for polynomial matrices

3.1.1 Invariant factors

For a polynomial matrix \(P(z)\), the Smith normal form states that there exist unimodular polynomial matrices \(U(z)\), \(V(z)\) (invertible over the polynomial ring) such that \[ U(z)P(z)V(z)=\operatorname{diag}(d_1(z),\dots,d_r(z),0,\dots,0), \] with \(d_i(z)\) dividing \(d_{i+1}(z)\). The polynomials \(d_i(z)\) are invariant factors, and they classify the module presented by \(P(z)\). These invariants are the polynomial analog of canonical structure that later reappears in rational settings.

3.1.2 Determinantal divisors

Closely related are determinantal divisors, defined via gcd/lcm patterns of minors. For each \(k\), one can consider all \(k\times k\) minors of \(P(z)\) and form their greatest common divisor (up to units). These divisors determine the invariant factors and provide an alternative route to computing canonical data.

3.2 Smith–McMillan form for rational matrix functions

3.2.1 Invariant polynomials in a rational setting

The Smith–McMillan form generalizes Smith normal form to rational matrix functions. For a rational matrix \(F(z)\), one can find invertible polynomial matrices \(U(z)\), \(V(z)\) and diagonal rational factors such that \(F\) becomes diagonal with entries that are quotients of polynomials in a divisibility-compatible manner. The diagonal entries encode invariant polynomials that record pole/zero data in a canonical way.

3.2.2 Minimality and uniqueness up to equivalence

The invariant polynomials in Smith–McMillan form are unique up to multiplication by units and consistent equivalences, making them canonical invariants of the rational matrix function. This “uniqueness up to equivalence” is crucial: while many representations \(P(z)Q(z)^{-1}\) exist, the Smith–McMillan data captures the essential singular structure without dependence on arbitrary factor choices.

3.3 Degree concepts and normalization

3.3.1 Row/column degrees

When expressing rational matrix functions in polynomial-fraction form, degrees can be tracked through row and column degrees of polynomial matrices in a representation. These degree measures help compare representations: a representation with unnecessarily high degrees can obscure minimality. Degree tracking aligns with algorithmic steps in canonical form computations, since many procedures reduce and compare polynomial degrees to detect cancellations and invariant divisors.

3.3.2 Proper vs. strictly proper forms

A matrix rational function is often classified as proper if its “numerator degree” does not exceed its “denominator degree” in a compatible sense, and strictly proper if the numerator degree is strictly less. This classification is representation-dependent unless stated via a canonical or invariant description. In practice, one uses normalized realizations or canonical forms to ensure that the properness notion reflects the underlying function rather than an arbitrary quotient choice.

4 Operations and closure properties

4.1 Addition and multiplication

4.1.1 Common denominator techniques

Given rational matrix functions \(F(z)=P(z)Q(z)^{-1}\) and \(G(z)=R(z)S(z)^{-1}\), one can form sums and products by rewriting them over common denominators. For instance, \[ F(z)+G(z)=P(z)Q(z)^{-1}+R(z)S(z)^{-1} \] can be expressed as a single quotient by combining over \(Q(z)S(z)\) (with attention to noncommutativity only arising if matrices sizes differ; in standard setups the products are well-defined). After combining, simplification may require cancellation of common factors.

4.1.2 Preservation of rationality

Rational matrix functions are closed under addition and multiplication: the result can always be written as a quotient of polynomial matrices. While cancellation can reduce degree or remove apparent singularities, the essential statement is that operations do not produce non-rational behavior.

4.2 Inversion and adjugate constructions

4.2.1 When \(F^{-1}(z)\) remains rational

If \(F(z)\) is invertible as a rational matrix function (meaning it is invertible for generic \(z\), with a rational inverse), then \(F^{-1}(z)\) is again rational. A necessary and sufficient algebraic condition can be expressed via determinant behavior and canonical forms: the inverse exists as a rational function precisely when \(F\) has full rank generically and the invariant data allows inversion without introducing essential singularities beyond poles/zeros already present.

4.2.2 Adjugate and determinant-based formulas

For square matrices over a field of rational functions, one can use adjugate formulas: \[ F(z)^{-1}=\frac{\operatorname{adj}(F(z))}{\det F(z)}, \] where \(\operatorname{adj}(F)\) is the transpose of the cofactor matrix. Since \(\operatorname{adj}(F(z))\) has entries that are polynomials in the entries of \(F(z)\) and \(\det F(z)\) is a rational function, the quotient is rational wherever the determinant does not vanish. This describes inversion through determinant and cofactor structure.

4.3 Transpose, conjugate, and polynomial mapping rules

4.3.1 Compatibility with canonical decompositions

Operations such as transpose (and, in complex settings, conjugation) interact with canonical forms in structured ways. Polynomial matrices can be transformed while preserving invariant-factor relationships up to predictable adjustments (for example, exchanging left/right perspectives corresponds to transposition in certain conventions). Canonical decompositions therefore provide a consistent bookkeeping method: invariants remain stable under these algebraic symmetries.

4.3.2 Variable substitutions and scaling

A rational matrix function can be transformed by changing variables, e.g. replacing \(z\) by a rational function of a new variable or scaling \(z\mapsto az+b\). Such substitutions transform polynomial degrees and move singular points accordingly. When substitutions are Möbius transforms (fractional linear transformations), the pole/zero structure is shifted by the mapping rule on the complex plane while preserving multiplicity data in the invariant-factor sense.

5 Factorizations of rational matrix functions

5.1 Polynomial factorizations

5.1.1 Factoring numerator/denominator matrices

Given \(F(z)=P(z)Q(z)^{-1}\), one can further factor \(P\) and \(Q\) into polynomial matrices (e.g., \(P=P_1P_2\), \(Q=Q_1Q_2\)) when algebraic structure permits. Such factorizations help isolate common factors responsible for cancellations and clarify which pieces contribute to poles versus removable singularities.

5.1.2 Reducing common factors

A common issue is that \(P\) and \(Q\) may share polynomial factors in the matrix sense, leading to cancellations in \(F\). Reducing by such common factors can yield a more economical quotient representation. Canonical minimality notions align with removing these redundancies.

5.2 Stable/causal-style factorizations (algebraic viewpoint)

5.2.1 Lifting factor structures from scalar cases

Factorization ideas often originate in scalar rational functions, where numerator and denominator are decomposed into factors associated with regions of the complex plane. In the matrix case, one seeks analogous decompositions but must handle noncommutativity and matrix rank structure. The algebraic strategy is to mimic scalar factor placement while using polynomial matrix methods to preserve divisibility and invariant structure.

5.2.2 Coprime factorization ideas

Coprime factorization aims to represent \(F\) as a ratio of polynomial matrices with minimal common divisors. In a matrix setting, “coprimeness” is defined through unimodular completion or through Smith normal form constraints. Coprime representations are important because they avoid hidden cancellations and often simplify pole/zero interpretation.

5.3 Linear fractional transformations (LFT) and equivalence

5.3.1 Fractional representation through block matrices

Linear fractional transformations express a map between matrix functions using block matrices. In algebraic form, a transformation can be encoded as \[ \mathcal{T}(F) = (A + BFC)(C_1 + DFC_2)^{-1} \] for suitable blocks, provided inverses exist. Such constructions preserve rationality and allow systematic equivalence transformations.

5.3.2 Fractional representation through block matrices

Block-matrix LFT formulations are also used to compare different representations of the same rational function. By embedding a quotient into a larger algebraic framework, one can track how invariant factors transform under equivalences, making it possible to prove that two rational matrices are related by structured transformations.

6 Realizations and state-space connections (algebraic form)

6.1 Realization via controllable/observable forms

6.1.1 Constructing a realization from a rational function

A realization represents a rational matrix function in terms of matrices that act linearly on a state vector, typically in the form \[ F(z)=C(zI-A)^{-1}B + D, \] or a closely related convention. The realization encodes the rational function’s behavior through finite-dimensional algebra. Constructing such a realization from a given rational matrix function amounts to producing polynomial or matrix data whose elimination reproduces the original quotient form.

6.1.2 Minimal realizations and uniqueness up to similarity

A realization is minimal when it cannot be reduced while preserving the same input-output behavior. Minimality is captured by controllability/observability-like conditions in typical settings, and algebraically it corresponds to a no-redundant-dynamics property. Minimal realizations are often unique up to similarity transformations: different minimal realizations describe the same rational function but use different state coordinates.

6.2 Transfer-function style representation

6.2.1 Block-matrix description

Transfer-function representations package the realization into block matrices (often a system matrix) whose Schur complements yield the transfer matrix \(F(z)\). This block approach makes it straightforward to compute \(F\) from state matrices and to verify equivalence by comparing Schur complements.

6.2.2 Properness and feedthrough terms

In transfer-function form, the constant term \(D\) is the feedthrough and is tied to properness. When \(F\) is strictly proper, one often has \(D=0\) under a realization convention. Properness affects whether the state-space expression includes an explicit direct term versus only a strictly dynamic component.

6.3 Polynomial realization parameters

6.3.1 Dependence on numerator/denominator degrees

The sizes and polynomial degrees involved in a realization reflect the pole structure and the degrees of invariant factors. Roughly, larger pole orders require larger or more complex state-space structures. Degree considerations also influence algorithmic construction: different methods yield realizations whose state dimension and matrix polynomial degrees track canonical data derived from the original rational function.

7 Transformations and equivalence relations

7.1 Similarity and equivalence of rational matrix functions

7.1.1 Left-right multiplication by unimodular matrices

A fundamental equivalence is obtained by multiplying a rational matrix function by unimodular polynomial matrices on the left and right (i.e., invertible over the polynomial ring). Such operations do not change the essential invariant structure of the function; they change representation while preserving canonical data up to equivalence. In terms of quotient forms, this corresponds to altering \(P\) and \(Q\) in ways that keep the rational map intact.

7.1.2 Effects on poles/zeros and canonical forms

Under these equivalences, the set of poles and zeros may be moved only in ways consistent with canonical invariant factors. Since canonical forms summarize the singular structure, equivalence transformations are best understood by how they preserve (or symmetrically transform) the invariant polynomials in Smith–McMillan data.

7.2 Minimality under equivalence

7.2.1 Preserving Smith–McMillan data

Minimality in realization theory aligns with uniqueness of canonical invariant data. If two realizations represent equivalent rational matrix functions (in the sense that their transfer matrices coincide), then their Smith–McMillan forms match. Conversely, distinct invariant data imply that no similarity transformation can relate the functions without changing their pole/zero structure.

7.3 Variable transformations

7.3.1 Möbius transforms and their algebraic effect

A Möbius transform \(z\mapsto \frac{az+b}{cz+d}\) maps points in the complex plane bijectively (away from a possible point at infinity). When applied to a rational matrix function, poles and zeros are carried to the images of their original locations, with multiplicities preserved in the invariant-factor sense after normalization. Algebraically, this comes from substituting into polynomial matrices and then re-clearing denominators to return to canonical form.

8 Examples and worked calculations

8.1 Simple 2×2 examples

8.1.1 Diagonal rational matrix functions

Consider a diagonal rational matrix \[ F(z)=\begin{pmatrix} \frac{z-1}{z} & 0\\ 0 & \frac{z+2}{z-3} \end{pmatrix}. \] Its poles occur where either denominator vanishes (at \(z=0\) and \(z=3\)), while zeros occur at the corresponding numerator zeros (at \(z=1\) and \(z=-2\)). In this setting, matrix poles and zeros reduce largely to scalar behavior, but it already illustrates how pole locations depend on where the quotient denominators vanish.

8.1.2 Non-diagonal rational functions with shared denominators

Let \[ F(z)=\frac{1}{z(z-1)}\begin{pmatrix} z & 1\\ 0 & z-1 \end{pmatrix}. \] Here the common denominator suggests possible poles at \(z=0\) and \(z=1\). Yet the matrix entries show different cancellation patterns: the \((2,2)\) entry involves \((z-1)/(z(z-1))=1/z\), revealing a pole only at \(z=0\), while the \((1,1)\) entry equals \(z/(z(z-1))=1/(z-1)\), showing a pole only at \(z=1\). This demonstrates how rank and entrywise cancellations influence the effective singular behavior.

8.2 Smith–McMillan computation walkthrough

8.2.1 Finding invariant polynomials

For a rational matrix function \(F(z)=P(z)Q(z)^{-1}\), one can compute invariant factors by converting to Smith normal form on appropriate polynomial matrices built from \(P\) and \(Q\). Practically, algorithms compute gcd-like structures of minors and then produce diagonal invariant quotients. The diagonal entries in Smith–McMillan form record how many times each irreducible factor appears in poles and zeros.

8.2.2 Interpreting the resulting pole/zero structure

Once the Smith–McMillan diagonal form is obtained, the invariant factors can be read as a structured pole/zero list: each diagonal term corresponds to a rational factor whose numerator encodes zeros and whose denominator encodes poles. The multiplicity and ordering follow from divisibility relations among invariant polynomials, which are more informative than using only \(\det F\) or \(\det Q\).

8.3 Realization from a rational matrix function example

8.3.1 Constructing a minimal realization

For a strictly proper scalar rational function \(f(z)=\frac{1}{z-a}\), a minimal realization is \[ f(z)=C(zI-A)^{-1}B \] with \(A=[a]\), \(B=[1]\), \(C=[1]\). For matrix-valued rational functions, one uses block constructions or companion forms that reproduce the same invariant poles and residues. Minimality corresponds to avoiding uncontrollable or unobservable state directions.

8.3.2 Verifying equivalence to the original function

Equivalence is verified by computing the transfer expression \(C(zI-A)^{-1}B + D\) and simplifying to match the original rational matrix function. In a canonical approach, one compares Smith–McMillan data of the realization’s transfer matrix to that of the given function, ensuring that poles, zeros, and their multiplicities coincide.

9 Common pitfalls and practical checks

9.1 Misinterpreting poles and cancellations

9.1.1 Apparent vs. true singularities

A frequent error is to treat every denominator singularity in a particular quotient representation as a genuine pole of the rational matrix function. Cancellations can occur when factors in the numerator annihilate problematic behavior from the denominator. Canonical forms help separate true poles (encoded in invariant denominators) from removable singularities.

9.2 Denominator singularity and evaluation issues

9.2.1 Checking where \(Q(z)^{-1}\) exists

When using \(F(z)=P(z)Q(z)^{-1}\), it is essential to verify invertibility of \(Q(z)\) at the evaluation point. Even if the rational function extends meromorphically, the specific quotient formula may not be valid at excluded points. In computations, one should distinguish between evaluating the quotient form and evaluating the function via an alternative representation that cancels factors.

9.3 Numerical vs. exact algebra considerations

9.3.1 Sensitivity to factor cancellations

Numerical methods may fail to detect exact cancellations, leading to spurious large values near near-singularities. Exact algebra tools such as Smith or Smith–McMillan computations provide reliable singular-structure information, whereas floating-point computations can blur the difference between removable singularities and genuine poles.

9.3.2 Using canonical forms to avoid ambiguity

Canonical forms remove ambiguity arising from arbitrary representation choices. By computing invariant polynomials and comparing them across candidate representations, one can confirm whether two rational matrix functions are equivalent and whether singularities match in a representation-independent manner.