Matrix Pencil

Treat a pair of same-sized matrices as the affine one-parameter family A−λB, whose finite and infinite generalized eigenvalues and singular structure remain meaningful even when B cannot be inverted.

Version

v3 · 2026-09-06 · History

Domain-specific #

2243

Origin domain

mathematics

Subdomain

linear algebra

Aliases

Linear matrix pencil, Pencil of matrices, Matrix pair, Generalized eigenvalue pencil

Core Idea

A Matrix Pencil is the affine one-parameter matrix family

\[ L(\lambda)=A-\lambda B, \]

where (A) and (B) are matrices of the same size over a field, usually the real or complex numbers. For square pencils, nonzero vectors satisfying \(Ax=\lambda Bx\) define the generalized eigenvalue problem. The ordinary eigenvalue problem is the special case (B=I). The pencil formulation is essential when (B) is singular or ill-conditioned: multiplying by (B^{-1}) may be impossible or numerically destructive, while the pair ((A,B)) still has intrinsic finite, infinite, and possibly singular structure.^[1]

For a square pencil, regularity means that \(det(A-\lambda B)\) is not identically zero. A regular pencil has a generalized spectrum on the projective line, so eigenvalues at infinity are legitimate structural values rather than errors to be discarded. A singular pencil requires more than a list of roots: left and right minimal indices and Kronecker blocks encode directions and constraints not visible in an ordinary determinant.^[2]

The recognition invariant is:

a same-sized matrix pair + affine dependence on one scalar parameter + equivalence applied to the pair/family rather than one evaluated matrix + generalized spectral or singular information that must remain valid without assuming the coefficient of the parameter is invertible.

Polynomial matrix pencils \(P(\lambda)=A_0+\lambda A_1+\cdots+\lambda^dA_d\) are an established extension. In much numerical literature, “pencil” without qualification means the affine degree-one object, and that standard identity is used here.

Structural Signature

The defining roles are:

• Coefficient pair: same-sized matrices (A) and (B), neither silently privileged as invertible.
• Parameter: a scalar \(\lambda\) over the coefficient field, extended projectively when infinity matters.
• Affine family: \(A-\lambda B\), considered as a single algebraic object rather than unrelated evaluated matrices.
• Generalized eigenrelation: \(Ax=\lambda Bx\) for a nonzero right vector, with an analogous left relation.
• Regularity test: whether the determinant polynomial is nonzero for at least one parameter value.
• Finite and infinite spectrum: roots of the homogeneous determinant relation, including the direction associated with a singular leading coefficient.
• Pencil equivalence: transformations \(L(\lambda)\mapsto P L(\lambda) Q\) with nonsingular (P,Q), preserving the appropriate invariant structure.
• Singular structure: minimal indices and Kronecker blocks when the determinant cannot summarize the object.
• Stable numerical reduction: simultaneous transformations such as the QZ algorithm that avoid explicit inversion.^[1]
• Application encoding: a physical, dynamical, or polynomial problem represented as a matrix pair.

What It Is Not

It is not an arbitrary matrix-valued function. That wording lacks the affine algebra and is too broad to inherit generalized eigenvalues, regularity, QZ reduction, and Kronecker structure. Nonlinear matrix-valued functions lead to nonlinear eigenvalue problems; polynomial matrix functions are higher-degree matrix polynomials or polynomial pencils.

It is not one matrix \(A-\lambda_0B\) evaluated at a fixed parameter. The identity lies in the entire family and the coupled transformation of both coefficients. It is not ordinary matrix similarity: generalized pencil equivalence permits simultaneous left and right transformations and retains a different invariant package.

It is not the engineering matrix pencil method for exponential parameter estimation, though that method constructs pencils and exploits generalized eigenvalues. The method is an application of the object, not its definition.

Scope of Application

Generalized eigenvalue problems arise whenever the natural equation is \(Ax=\lambda Bx\): vibration with a stiffness and mass matrix, constrained optimization, stability of descriptor systems, and discretized differential equations. If (B) is positive definite, symmetric-definite algorithms can exploit additional structure. If (B) is singular, infinite eigenvalues can encode algebraic constraints.

Differential-algebraic equations \(E\dot{x}=Ax\) produce the pencil \(A-\lambda E\). Regularity distinguishes systems with a well-posed finite-dimensional generalized spectral structure; the infinite part reflects constraints and differentiation index. Control theory uses related system pencils to analyze zeros, controllability, and realization properties.

Polynomial eigenvalue problems such as \((\lambda^2 M+\lambda C+K)x=0\) are often linearized into a larger affine pencil. A good linearization preserves finite and infinite eigenstructure, not merely the determinant’s finite roots.^[3]

Clarity

Sign convention does not define a new pencil identity: \(A-\lambda B\), \(A+\lambda B\), and \(\lambda B-A\) can represent equivalent conventions when coefficients are adjusted consistently. Matrix dimensions do matter. Square regular pencils support a determinant-based spectrum; rectangular or singular pencils demand Kronecker-style structure.

An “infinite eigenvalue” should not be paraphrased as an infinitely large numerical output. Homogenize the pencil as \(\alpha A-\beta B\). A generalized eigenvalue is the ratio \(\lambda=\alpha/\beta\); \(\beta=0\) gives the projective point at infinity. This coordinate treatment keeps degree loss caused by singular (B) inside the spectrum.

Regular and singular are properties of the entire pencil, not synonyms for invertible and singular coefficient matrices. Both (A) and (B) may be singular while \(A-\lambda B\) is regular.

Manages Complexity

The pencil packages two coupled operators without forcing one into the role of denominator. This prevents premature inversion, preserves constraint structure, and allows common transformations to expose triangular, block, or canonical form. The QZ algorithm generalizes QR by reducing the pair while never requiring (B^{-1}); it remains defined when (B) is singular and is usually preferable when (B) is ill-conditioned.^[1]

Canonical and condensed forms separate the finite spectrum, infinite spectrum, and singular chains. Applications can then reason blockwise: stable versus unstable eigenvalues, dynamic versus algebraic modes, controllable versus hidden directions, or finite versus infinite elementary divisors. The abstraction thus turns one entangled equation into an invariant inventory.

Abstract Reasoning

To analyze a matrix pencil:

1. State the field, matrix dimensions, and coefficient pair.
2. Preserve the pair (A,B); do not form (B^{-1}A) before checking invertibility and conditioning.
3. Test whether the pencil is regular.
4. If regular, compute finite and infinite generalized eigenvalues and relevant deflating subspaces.
5. If singular, determine normal rank, left/right minimal indices, and appropriate Kronecker information.
6. Exploit symmetry, Hermitian structure, definiteness, sparsity, palindromicity, or other declared structure.
7. Use simultaneous stable transformations and report backward error for the pair.
8. Translate the invariant blocks back into the originating application’s modes or constraints.

The central reasoning move is to ask what remains invariant under transformations of the family, rather than what is true of one coefficient or one sampled value.

Knowledge Transfer

The affine pair and its invariants transfer literally across numerical linear algebra, descriptor systems, vibration analysis, control, and polynomial eigenvalue linearization. Each domain changes the interpretation of (A), (B), and \(\lambda\), but regularity, projective eigenvalues, deflating subspaces, and pairwise reduction remain the same mathematics.

The broader catalog parent is Matrix: each coefficient and every evaluation is a matrix, and standard matrix operations provide the substrate. The pencil adds parameter coupling, pair equivalence, and a generalized invariant structure; these are not captured by the parent’s single-array identity.

Examples

Ordinary eigenproblem. \(L(\lambda)=A-\lambda I\). Its generalized eigenvalues are the ordinary eigenvalues of (A).

Mechanical vibration. \(Kx=\lambda Mx\), where (K) is stiffness and (M) is mass. The pair preserves physical structure that forming (M^{-1}K) can obscure.

Descriptor system. \(E\dot{x}=Ax\) yields \(A-\lambda E\). Singular (E) can produce infinite eigenvalues representing algebraic constraints.

Polynomial linearization. A quadratic polynomial \(\lambda^2M+\lambda C+K\) is represented by a larger companion pencil whose generalized spectrum reproduces the polynomial eigenvalues.

Non-example. A neural-network layer \(W(\lambda)\) with arbitrary nonlinear parameter dependence is a matrix-valued function, not automatically a matrix pencil.

Structural Tensions

• Pair preservation versus inversion convenience: (B^{-1}A) looks familiar but may destroy existence, stability, and infinity information.
• Finite roots versus projective completeness: determinant roots in the field omit eigenvalues at infinity.
• Regular spectrum versus singular structure: eigenvalues suffice for regular pencils but not for singular ones.
• Canonical clarity versus numerical sensitivity: exact canonical forms can be discontinuous under perturbation; stable condensed forms are often preferable computationally.
• Linearization convenience versus structure preservation: a larger affine pencil simplifies algorithms but can introduce spurious conditioning or obscure polynomial symmetries.
• General definition versus usable identity: calling every matrix-valued function a pencil erases the special machinery that makes the term valuable.

Structural–Framed Character

The affine family, regularity, equivalence, eigenstructure, and minimal indices are structural. Choice of sign, scaling, linearization, normalization, and numerical tolerance is framed by convention and application.

Structural Core vs. Domain Accent

The portable core is an object represented by a parameterized pair whose invariant properties must be studied jointly rather than by dividing one component into the other. The domain accent—matrices, determinants, fields, eigenvalues, Kronecker blocks, and QZ transformations—is constitutive. The candidate is therefore domain-specific.

Instantiates / Related Primes

Matrix is the proposed immediate parent in the existing catalog: a pencil is a structured parameterized family built from two same-sized matrices. Matrix Similarity and Minimal Polynomial are neighboring single-operator invariant concepts. Polynomial Ring is related through polynomial matrix entries and determinant polynomials but is not the structural genus.

The prospective queue contains one strict edge to domain_specific:matrix. No live DAG mutation is authorized.

Relationships to Other Abstractions

Current abstraction Matrix Pencil Domain-specific

Parents (1) — more general patterns this builds on

• Matrix Pencil is a kind of Matrix Domain-specific

  Matrix is the proposed immediate parent in the existing catalog: a pencil is a structured parameterized family built from two same-sized matrices.

Hierarchy paths (5) — routes to 5 parentless roots

• Matrix Pencil → Matrix → Tensor → Transformation → Function (Mapping)

Show alternative paths (4)

Neighborhood in Abstraction Space

Matrix Pencil sits in a sparse region of the domain-specific corpus (79^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

• Quadratic Equation — 0.83
• Linear fractional transformation — 0.82
• Minimum Relevant Variables in a Linear System — 0.82
• Quadratic Space — 0.82
• Schneider–Lang Theorem — 0.82

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

• Matrix: one rectangular array, rather than an affine family built from a pair.
• Generalized eigenvalue problem: the problem posed on a pencil, not the pencil object itself.
• Matrix similarity: (P^{-1}AP) for one square matrix; pencil equivalence acts on the family.
• Matrix polynomial: a possibly higher-degree polynomial in the parameter with matrix coefficients.
• Nonlinear eigenvalue problem: arbitrary analytic or nonlinear parameter dependence.
• Matrix pencil method: a signal-estimation procedure that uses generalized eigenvalues of constructed pencils.
• Ordinary singular matrix: singularity of one coefficient does not determine regularity of the pencil.

References

[1] Cleve B. Moler and G. W. Stewart, “An Algorithm for Generalized Matrix Eigenvalue Problems,” SIAM Journal on Numerical Analysis 10(2), 1973, 241–256. DOI 10.1137/0710024. registry ↩a ↩b ↩c

[2] Israel Gohberg, Peter Lancaster, and Leiba Rodman, Matrix Polynomials, SIAM Classics in Applied Mathematics 58, 2009. Originally published by Academic Press, 1982. registry ↩

[3] D. Steven Mackey, Niloufer Mackey, Christian Mehl, and Volker Mehrmann, “Vector Spaces of Linearizations for Matrix Polynomials,” SIAM Journal on Matrix Analysis and Applications 28(4), 2006, 971–1004. DOI 10.1137/050628350. registry ↩

[4] Peter Lancaster and Qiang Ye, “Variational and Numerical Methods for Symmetric Matrix Pencils,” Bulletin of the Australian Mathematical Society 43(1), 1991, 1–17. DOI 10.1017/S0004972700028806. registry ↩