Kronecker-Sum Spectral Structure

• Kronecker-Sum Spectral Structure is defined as operators formed by combining lower-dimensional matrices via the Kronecker sum, yielding additive eigenvalues and eigenvectors.
• It enables efficient computation of matrix functions by decomposing large multidimensional operators into simpler tensor products, reducing computational complexity.
• This framework underpins applications in covariance estimation, PDE discretizations, and statistical modeling by ensuring scalable and precise spectral analysis.

The Kronecker-sum spectral structure refers to the precise algebraic and spectral characteristics of operators, matrices, or tensors constructed via the Kronecker sum. The Kronecker sum is of central importance in numerical linear algebra, operator theory, multivariate statistics, and the spectral analysis of higher-order arrays, enabling the diagonalization and efficient manipulation of multi-way operators. This structure determines the spectrum, eigen-decomposition, and functional calculus of large multidimensional operators from their lower-dimensional components, underlies theoretical results on spectral interlacing and bounds, and drives state-of-the-art scalable algorithms in high-dimensional inference, covariance estimation, and exponential integration.

1. Definition and Algebraic Construction

Let $A\in\mathbb{C}^{n_1\times n_1}$ and $B\in\mathbb{C}^{n_2\times n_2}$ be square matrices. The Kronecker sum is defined as

$$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$

where $\otimes$ is the Kronecker product, and $I_{n_1},I_{n_2}$ are identity matrices. In $d$ dimensions, $\bigoplus_{k=1}^d A^{(k)}$ denotes the sum of $A^{(k)}$ acting in each tensor leg. This definition extends to hypermatrices (“hypermatrix Kronecker sum”) within the Bhattacharya–Mesner (BM) algebra, where side-length-2 cubic hypermatrices admit direct sum, Kronecker product, and Kronecker sum constructions generalizing classical matrix concepts (Filmus et al., 2015).

The Kronecker sum frequently appears as the generator of linear operators for multivariate problems, such as symmetry-adapted Laplacians, covariance models for matrix-variate normal laws, or as the matrix form for discrete differential operators on Cartesian products of graphs (Benzi et al., 2015, Zhou et al., 5 Feb 2025, Yoon et al., 2021).

2. Spectral Decomposition of the Kronecker Sum

A central property is the complete spectral description of $A \oplus B$ in terms of the spectra of $A$ and $B\in\mathbb{C}^{n_2\times n_2}$0:

• If $B\in\mathbb{C}^{n_2\times n_2}$1 has eigenvalues $B\in\mathbb{C}^{n_2\times n_2}$2 with eigenvectors $B\in\mathbb{C}^{n_2\times n_2}$3, $B\in\mathbb{C}^{n_2\times n_2}$4 has eigenvalues $B\in\mathbb{C}^{n_2\times n_2}$5 with eigenvectors $B\in\mathbb{C}^{n_2\times n_2}$6, then

$B\in\mathbb{C}^{n_2\times n_2}$7

for all $B\in\mathbb{C}^{n_2\times n_2}$8. The $B\in\mathbb{C}^{n_2\times n_2}$9 eigenvalues are $$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$0, and the eigenvectors are the Kronecker products $$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$1 (Benzi et al., 2015, Croci et al., 2022, Yoon et al., 2021, Zhou et al., 5 Feb 2025).

This spectral structure generalizes to higher dimensions and to hypermatrices in the BM framework, where spectral parameters similarly “add” in Kronecker-sum constructions (Filmus et al., 2015). For Sylvester operators $$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$2, classical results show that the spectrum consists of all possible sums $$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$3 (Dressler et al., 2022).

The following table summarizes the spectral structure for key Kronecker constructions:

Construction	Spectrum	Eigenvectors
$$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$4	$$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$5	$$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$6
$$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$7	$$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$8	$$A \oplus B = A \otimes I_{n_2} + I_{n_1} \otimes B,$$9
Direct sum $\otimes$0 (block)	$\otimes$1	Blocks from $\otimes$2 or $\otimes$3

The Kronecker-sum structure allows the full eigen-decomposition of large operators to be “lifted” from those of the lower-dimensional factors without explicit formation of the full matrix (Benzi et al., 2015, Croci et al., 2022).

3. Generalizations, Bounds, and Interlacing

Kronecker-sum structures admit refined analysis of spectra, including spectral bounds for sums of Kronecker products and spectral interlacing phenomena:

• Lototsky provides sharp two-sided bounds for the spectral radii and abscissas of discrete and continuous-time stochastic Kronecker sums $\otimes$4 and $\otimes$5 in terms of associated Hermitian matrices $\otimes$6 and $\otimes$7 (Lototsky, 2014).
• For the Jordan–Kronecker product $\otimes$8, the spectrum splits according to the commutation matrix into “even” (symmetric eigenvector) and “odd” (skew-symmetric) parts. For symmetric/skew-symmetric inputs of rank at most two, the odd eigenvalues interlace the evens; for $\otimes$9, full interlacing holds, but this fails generally for $I_{n_1},I_{n_2}$0 when both $I_{n_1},I_{n_2}$1 have rank at least 3 (Kalantarova et al., 2018).
• In BM-algebraic hypermatrix contexts, Kronecker-sum spectra are additive, and explicit characteristic polynomials, determinants, and eigen-hypermatrix decompositions can be constructed for side-length-2 cases, further generalizing classical spectra (Filmus et al., 2015).

4. Functional Calculus and Applications

The spectral diagonalization of Kronecker sums underpins exact and highly efficient computation of matrix functions and exponentials:

• If $I_{n_1},I_{n_2}$2 is analytic, $I_{n_1},I_{n_2}$3, where $I_{n_1},I_{n_2}$4 acts entrywise on the diagonal $I_{n_1},I_{n_2}$5. Thus, $I_{n_1},I_{n_2}$6 (Benzi et al., 2015).
• This enables efficient Krylov, quadrature, or exponential integrator routines, where the large problem is decomposed into tensor products of much smaller problems. For example, in exponential time integration, $I_{n_1},I_{n_2}$7; no large matrix exponentials are required (Croci et al., 2022, Benzi et al., 2015).

Exploiting the Kronecker-sum structure in numerical algorithms results in orders-of-magnitude speed-ups and reduced memory requirements. For a dimension $I_{n_1},I_{n_2}$8, matrix-vector computations can be reduced from $I_{n_1},I_{n_2}$9 to $d$0 for moderate $d$1 (Croci et al., 2022).

5. Statistical Models and Covariance Estimation

The Kronecker sum appears as a natural covariance structure in matrix-variate subgaussian models, spatio-temporal inference, and graphical model estimation:

• Observed data $d$2 is modeled as $d$3, where $d$4 governs the covariance. The spectrum of $d$5 controls identifiability, convergence rates, and regularization in high-dimensional settings (Zhou et al., 5 Feb 2025, Yoon et al., 2021).
• In sparse Kronecker-sum inverse covariance estimation, the explicit spectral decomposition allows Newton-type (EiGLasso) or composite-gradient methods to scale to $d$6 in the thousands or more. The entire negative log-likelihood, gradient, and Hessian diagonalize in the Kronecker basis, leading to quadratic or linear rates of convergence depending on the accuracy of Hessian approximation. Trace-non-identifiability is addressed by post hoc redistribution (Yoon et al., 2021).
• Computational complexity is reduced from $d$7 to $d$8 per iteration, with empirical gains of 2–4 orders of magnitude in real data (Yoon et al., 2021).

6. Low-Rank and Spectral-Norm Approximation

Operators acting on matrix spaces can often be approximated as sums of few Kronecker products—Kronecker-sum decompositions—leading to efficient low-rank spectral approximations and model compression:

• Spectral-norm-optimal approximations are achieved via alternating semidefinite programming, as opposed to the classical SVD (optimal for Frobenius norm). This iterative block-alternation yields partial optima and converges monotonically, with feasibility for large problems when $d$9 (Kronecker rank) is small (Dressler et al., 2022).
• Special cases including Sylvester operators and Lyapunov equations exhibit spectra and inverse spectra captured exponentially rapidly by low-rank Kronecker-sum expansions (Dressler et al., 2022).
• For each decomposition $\bigoplus_{k=1}^d A^{(k)}$0, the spectral norm is bounded by $\bigoplus_{k=1}^d A^{(k)}$1. Alternating SDP reveals the dominant spectral structure numerically.

7. Extensions, Limitations, and Unification

The Kronecker-sum spectral structure generalizes seamlessly from matrices to higher-order arrays and hypermatrices under the BM algebra, using the same block-diagonalization and product-of-eigenvector principles (Filmus et al., 2015). Spectral results are universally valid for block-diagonalizable operators and for stochastic Kronecker sum forms, but may not generalize to arbitrary sums $\bigoplus_{k=1}^d A^{(k)}$2 without further symmetry or covariance structure (Lototsky, 2014, Kalantarova et al., 2018).

The “uncorrelated tuple” property ensures spectral decomposability under Kronecker-sum and product operations. This algebraic property underpins all fast algorithms exploiting Kronecker-sum spectral structure, including those used in statistical inference, partial differential equations, and high-dimensional operator theory. Explicit determinant and characteristic polynomial identities arise for small sizes, extending classical results to tensor-valued contexts (Filmus et al., 2015).

The Kronecker-sum spectral structure thus forms the mathematical infrastructure for scalable, theoretically-underpinned computation in multidimensional problems, allowing spectral characterization, efficient matrix-function evaluation, and principled inverse-problem regularization across a diverse range of applications.