Reducibility of linear representations, free ideals, and Kippenhahn's conjecture

Abstract: This paper presents a comprehensive study of the characteristic polynomial of matrix tuples and finite dimensional group representations. Among other things, several key concepts are introduced, including the minimal polynomial, spectral index, spectral stability, and characteristic graph. Notably, this framework provides a complete resolution to Kippenhahn's conjecture, settling a long-standing and influential problem in the theory of matrix tuples.

• The paper develops free Cayley–Hamilton ideals and minimal polynomials for matrix tuples, showing that every tuple of k×k matrices satisfies nontrivial annihilating relations of degree at most k.
• The authors prove that irreducibility is equivalent to the existence of a one-step algebraic extension with irreducible characteristic polynomial, while introducing spectral indices to measure factorization and multiplicity.
• For Hermitian tuples, the paper provides a bounded reducibility algorithm terminating in at most ν(A)−κ(A) steps and gives a necessary-and-sufficient criterion that resolves Kippenhahn’s conjecture.

Motivation and historical context

The paper by Stessin and Yang addresses a central question in representation theory: given a finitely generated group or algebra acting linearly on a finite-dimensional space, does the tuple of operators representing the generators admit a nontrivial common invariant subspace? The authors approach this through the lens of the joint characteristic polynomial $Q_A(z_0,z)=\det(z_0I+z_1A_1+\cdots+z_nA_n)$, a notion rooted in Dedekind and Frobenius's theory of group determinants from the 1890s. Frobenius's classical theorem states that for a finite group $G$ with left regular representation $\lambda$, the group determinant factors as $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$, where each $Q_\pi$ is irreducible and determines $\pi$ up to unitary equivalence. The paper revisits this factorization in the modern framework of projective joint spectra, introduced in multivariable operator theory, and develops an algebraic-geometric toolkit—free ideals, minimal polynomials, spectral indices, and characteristic graphs—for studying reducibility.

Cayley–Hamilton ideals and free polynomials

The first structural contribution is a multivariable Cayley–Hamilton theorem. Expanding $Q_A(z_0,z)$ as a polynomial in $z_0$ and substituting $z_0=-A_*(z)$ yields free polynomials $p_\tau$ in the free algebra $G$0 that annihilate the tuple $G$1. These generate the Cayley–Hamilton ideal $G$2, contained in the full annihilating ideal $G$3. A notable consequence is that every tuple of $G$4 matrices satisfies nontrivial annihilating relations of degree at most $G$5—sharper than the Amitsur–Levitzki bound of $G$6, which applies uniformly to all tuples. The Pauli matrix example illustrates the mechanism: the quadratic characteristic polynomial $G$7 forces the Clifford algebra relations $G$8 and anticommutation relations, showing that the Pauli algebra is intrinsic to the Minkowski-type quadratic form rather than to the specific matrices.

For group representations, the authors define free ideals $G$9 and prove that weak containment of unitary representations implies inclusion of the corresponding ideals ($\lambda$0), so weakly equivalent representations share identical ideals. This connects the spectral framework to standard representation-theoretic equivalence notions.

Minimal polynomials and cyclicity

The paper defines a minimal polynomial for a matrix tuple as the lowest-degree homogeneous polynomial $\lambda$1 satisfying $\lambda$2. Three results form the core of this section:

• Uniqueness and divisibility: the minimal polynomial exists, is unique, and divides $\lambda$3; moreover, $\lambda$4, proved via Gauss's Lemma after establishing that relevant rational coefficients are locally bounded near their singular loci (an application of the Riemann extension theorem).
• Linear cyclicity: $\lambda$5 is minimal if and only if some pencil $\lambda$6 is a cyclic matrix.
• Hermitian case: for Hermitian tuples, $\lambda$7 is minimal exactly when it has no repeated irreducible factors—a direct generalization of the single-matrix criterion of simple spectrum.

These results are summarized in a diagram relating five properties: irreducibility of the tuple, irreducibility of $\lambda$8, minimality of $\lambda$9, linear cyclicity, and cyclicity of the generated algebra. Importantly, the implications are strict: irreducibility of the tuple neither implies nor is implied by minimality of $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$0. A dihedral-group example makes this concrete: the direct sum of two inequivalent irreducible representations of $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$1 has a characteristic polynomial with two distinct irreducible quadratic factors—hence minimal—yet the tuple is reducible. This gap between spectral data and reducibility is precisely what motivates Kippenhahn's conjecture and the machinery developed later in the paper.

As a reinterpretation of Frobenius's theorem, the authors show that $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$2 is the minimal polynomial of the generating tuple $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$3 when these generate all of $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$4. They note the caveat that when the listed elements are a proper subset of $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$5, irreducibility of individual $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$6 can fail (as known for $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$7), and they pose as an open question whether minimality holds for finite solvable groups.

Spectral indices and algebraic extensions

Two similarity-invariant quantities are introduced: the spectral index $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$8, the number of distinct irreducible factors of $Q_\lambda=\prod_{[\pi]\in\hat G} Q_\pi^{d_\pi}$9, and the spectral multiplicity $Q_\pi$0, the sum of factor multiplicities. For commuting tuples $Q_\pi$1; for the regular representation of a finite group, $Q_\pi$2.

Because $Q_\pi$3 alone cannot decide reducibility, the paper studies algebraic extensions $Q_\pi$4 with $Q_\pi$5. Using Hilbert's irreducibility theorem, the authors show that the lower spectral index $Q_\pi$6 is attained by a single well-chosen extension matrix $Q_\pi$7 (Hermitian when $Q_\pi$8 is), yielding the main structural result: a tuple is irreducible if and only if some one-step algebraic extension has an irreducible characteristic polynomial. This is a clean necessary-and-sufficient condition, though the authors concede that constructing such a $Q_\pi$9 explicitly "can be challenging" in practice, since it amounts to building a basis for $\pi$0. This practical difficulty motivates the algorithmic approach taken for Hermitian tuples.

Local spectral analysis

The technical backbone for the Hermitian case is local analysis of the proper projective joint spectrum in divisor form, $\pi$1. Near regular points of the spectral components, contour-integral projections (Riesz projections) attached to eigenvalues of $\pi$2 allow the authors to extract algebraic identities among words built from the projections $\pi$3, the operators $\pi$4, and derivative coefficients $\pi$5, $\pi$6 of the implicit functions parametrizing the components. These identities, established previously by Stessin, force vanishing of certain aggregated word sums when the tuple is Hermitian—an obstruction that ultimately produces invariant subspaces.

Two prior results are recalled and used: a theorem stating that if the spectrum has a single component of multiplicity $\pi$7 and certain associated pairs $\pi$8 retain single-component spectra of degree $\pi$9, then $Q_A(z_0,z)$0 decomposes as a direct sum of $Q_A(z_0,z)$1 identical copies; and a projection-based criterion under which a union of spectral components corresponds to a common invariant subspace.

Spectrally stable extensions and a bounded reducibility test

The key new concept is spectral stability: an extension $Q_A(z_0,z)$2 is spectrally stable if every eigenvalue-reciprocal point $Q_A(z_0,z)$3 lies on a single component of $Q_A(z_0,z)$4 with nonvanishing derivative—equivalently, each component of $Q_A(z_0,z)$5 sits inside a single component of $Q_A(z_0,z)$6 with unchanged multiplicity. The central theorem asserts:

If $Q_A(z_0,z)$7 is an admissible Hermitian tuple, the canonical extension $Q_A(z_0,z)$8 (built from all words $Q_A(z_0,z)$9 of length at most the number of distinct eigenvalues) is spectrally stable, and not all component multiplicities $z_0$0 equal 1, then $z_0$1 is reducible.

The proof splits into two cases. If multiplicities differ, a rank argument shows that block matrices $z_0$2 must vanish whenever $z_0$3, directly producing a reducing subspace spanned by eigenspaces of equal multiplicity. If all multiplicities equal $z_0$4, the earlier decomposition theorem applies. In both cases the reducing subspace is constructed explicitly, not merely shown to exist.

This feeds into the spectral test of reducibility, an iterative algorithm alternating between graph checks and spectrally stable extensions (with a symmetrization step restoring Hermiticity at each stage). Because each unstable step strictly decreases some component multiplicity, which cannot fall below 1, the test terminates in at most $z_0$5 steps. For example, Waterhouse-type counterexamples to Kippenhahn's conjecture with $z_0$6, $z_0$7 are decided in at most $z_0$8 steps—one step when $z_0$9.

Characteristic graphs

The second geometric tool is combinatorial. The characteristic graph $z_0=-A_*(z)$0 has vertices $z_0=-A_*(z)$1 and an edge $z_0=-A_*(z)$2 whenever some $z_0=-A_*(z)$3 has a nonzero $z_0=-A_*(z)$4 entry. Two facts anchor the theory:

• A tuple is reducible if and only if some similar tuple has a non-strongly-connected characteristic graph (proved via Burnside's theorem: absence of paths across a strongly connected component prevents elementary matrices from lying in the generated algebra).
• If $z_0=-A_*(z)$5 has no repeated factors, then diagonalizing $z_0=-A_*(z)$6 and checking strong connectedness of the resulting graph decides irreducibility—and strong connectedness is independent of the diagonalizing basis.

Consequently, for Hermitian tuples with minimal $z_0=-A_*(z)$7, the strongly connected components of the graph partition $z_0=-A_*(z)$8 into reducing subspaces on each of which the restriction is irreducible, giving an effective decomposition procedure. The authors note the natural limitation that the graph itself is not similarity-invariant in general (they exhibit similar tuples with connected and disconnected graphs), which is why admissible transformations and diagonalization play an essential role.

Resolution of Kippenhahn's conjecture

Kippenhahn conjectured in 1951 that a Hermitian pair whose joint characteristic polynomial has a repeated factor must be reducible. History ran against the naive statement: Shapiro verified it for sizes up to 5, but Laffey produced a counterexample at size 8 in 1983, followed by Waterhouse's family at sizes $z_0=-A_*(z)$9 for all $p_\tau$0 and further examples by Li–Spitkovsky–Shuka at size 6. The present paper reframes the problem: since repeated factors mean $p_\tau$1 is not minimal, the conjecture asks whether non-minimality forces reducibility for Hermitian tuples.

The final theorem gives a complete answer as a necessary and sufficient condition: a Hermitian tuple with repeated factors in $p_\tau$2 is reducible if and only if either its characteristic graph is not strongly connected, or the graph is strongly connected and the characteristic polynomial of the fully iterated extension $p_\tau$3 still has repeated factors. Since the test terminates in bounded time, this settles the validity question for tuples of arbitrary length. An illustrative analysis of pairs with $p_\tau$4 identifies three scenarios—irreducibility, decomposition into two components detected by the graph, or forced reducibility with an explicit reducing subspace constructed through unitary block structure.

It should be noted that the resolution is a criterion, not a blanket affirmation or refutation: the conjecture holds precisely for those tuples satisfying the stated condition, and the known counterexamples correspond to cases where the iterated extension loses repeated factors while the original graph remains connected.

Limitations and open questions

Several caveats are explicit in the paper. The construction of the extension matrix $p_\tau$5 achieving $p_\tau$6 is acknowledged to be difficult in general, limiting the practical applicability of the abstract irreducibility criterion outside the Hermitian setting. The bounded termination guarantee applies only to Hermitian tuples; extending the graph-based method to tuples whose characteristic polynomials have no repeated factors but which are not Hermitian remains open. The minimality question for $p_\tau$7 when generators form a proper subset of $p_\tau$8 is unresolved even for solvable groups. Finally, the authors pose two concrete problems: whether a single additional group element suffices to make the extended characteristic polynomial irreducible for any irreducible representation, and what the minimal word length of such an element is—questions that would refine the 130-year-old Frobenius factorization theorem.

Conclusion

The paper builds a coherent bridge between the classical theory of group determinants and modern multivariable spectral theory. Its principal contributions are the several-variable Cayley–Hamilton ideal, a satisfactory theory of minimal polynomials for matrix tuples with exact characterizations via linear cyclicity and (for Hermitian tuples) absence of repeated factors, similarity-invariant spectral indices, and two complementary decidability tools—algebraic extensions governed by Hilbert irreducibility, and characteristic graphs coupled with spectrally stable extensions. Together these yield a bounded algorithm deciding reducibility of Hermitian tuples and a necessary-and-sufficient condition that fully resolves Kippenhahn's 1951 conjecture, with all reducing subspaces constructible rather than merely existential.