A hidden variable resultant method for the polynomial multiparameter eigenvalue problem

Emil Graf, Alex Townsend

TL;DR

This paper introduces a general-purpose, global algorithm to solve Polynomial Multiparameter Eigenvalue Problems (PMEPs), eliminating the need for researchers to build tedious, custom solvers for every new problem. By combining techniques from algebraic geometry and linear algebra into a "hidden variable tensor Dixon resultant," the method transforms complex, multi-variable matrix equations into standard eigenvalue problems that can be solved with high accuracy using existing numerical tools.

The Challenge of PMEPs

Polynomial Multiparameter Eigenvalue Problems (PMEPs) appear in a variety of complex physics and engineering applications, such as analyzing aeroelastic flutter (how aircraft wings vibrate in the air) or tracking wave propagation in layered materials. A PMEP requires finding a set of eigenvalues (a coordinate point) and distinct eigenvectors that simultaneously solve multiple non-linear matrix equations.

Historically, there has been no general-purpose solver for these problems. Because PMEPs combine the difficulties of non-linear polynomial rootfinding with the headache of managing distinct eigenvectors for every equation, researchers have had to hand-craft custom "linearizations" tailored to the specific degree and sparsity of their exact problem. This case-by-case mathematical legwork is incredibly time-consuming and discourages scientists from tackling more complicated PMEPs. This paper aims to provide a universal, out-of-the-box solver.

The Proposed Solution: Hidden Variable Tensor Dixon Resultant

The authors build their solver by merging two distinct mathematical concepts: hidden variable resultants (from algebraic geometry) and operator determinants (from linear algebra).

Conceptual mapping of the Multivariate Polynomial Eigenproblem (MPE) solver relative to classical methods.
  1. The Hidden Variable Trick: Given a problem with $d$ variables, the algorithm "hides" one variable ($x_d$), treating the system as a set of polynomials in $d-1$ variables where the coefficients are functions of $x_d$.
  2. The Tensor Dixon Resultant: To find out when these equations share a common solution, the authors generalize the "Dixon resultant"—a classic tool for finding common roots of scalar polynomials—to work with matrices. Crucially, they use Kronecker products to multiply the matrices. This is vital because it properly links the distinct eigenvectors of the different equations together into one unified structure.
  3. Linearization and Solving: This process yields a univariate Polynomial Eigenvalue Problem (PEP) dependent only on $x_d$. The authors linearize this PEP and solve it using the standard, highly robust QZ algorithm to find the $x_d$ coordinates.
  4. Extracting the Rest: Because of the specific mathematical structure of the resultant, the eigenvectors of this new problem have a highly predictable, repeating pattern (a "multivariate block Vandermonde" structure). The authors can extract the remaining $x_1, \dots, x_{d-1}$ coordinates simply by taking ratios of specific entries within these eigenvectors.

Experimental Validation

As a mathematical methods paper, the authors benchmark their algorithm against established physics and engineering problems rather than standard machine learning datasets:

  • Aeroelastic Flutter: A 2-parameter quadratic eigenvalue problem.
  • Leaky Waves in Layered Structures: A 3-parameter eigenvalue problem with non-polynomial dependencies (which are resolved into a polynomial form).

Key Results

  • Theoretical Guarantees: The authors mathematically prove that their resultant matrix is nonsingular (meaning it can be reliably solved) for generic systems up to 3 dimensions, and they computationally verified this property up to 10 dimensions.
  • High Accuracy: On the aeroelastic flutter benchmark, the general method matched the accuracy of a highly specialized, custom-built solver (quad_twopareig) up to $10^{-12}$ for all solutions.
  • Out-of-the-box Versatility: The algorithm successfully solved the complex 3-parameter leaky wave problem directly. It matched the exact solutions of an ad-hoc linearization from previous literature, proving that practitioners can now skip manual mathematical legwork and use this general solver.

Key Insights and Ideas

  • Naive generalizations fail: You cannot simply take the standard scalar Dixon resultant and swap numbers for matrices. That approach fails because it ignores the fact that different matrix polynomials have different eigenvectors.
  • Kronecker products are the glue: By using Kronecker products (inspired by the "operator determinants" method used for linear multiparameter problems), the algorithm successfully captures the distinct eigenvector structures, merging multiple separate problems into one massive, solvable system.
  • Eigenvectors hide the answers: The structure of the eigenvectors generated by the tensor Dixon resultant is a goldmine. Once you solve for the hidden variable ($x_d$), you don't have to start from scratch to find the others. The remaining variables are baked into the ratios of the eigenvector's entries.

Limitations and Future Work

  • Singular Pencils in Real-World Problems: While the algorithm works perfectly for "generic" (randomized) systems, real-world problems often have specific structures that cause the resulting matrix pencil to be singular. The authors mitigate this using a projection method, but it adds computational complexity.
  • Null Space Noise: In singular problems, extracting the remaining coordinates from the eigenvectors can be corrupted by numerical noise in the null space. The authors implemented a filtering technique to only use eigenvector entries that are sufficiently independent of this noise.
  • Theoretical Gap: The formal proof that the resultant is nonsingular for generic systems is only complete for $d \le 3$. Proving this holds true for any number of dimensions ($d$) remains an open mathematical question.

Summary updated

BibTeX
@inproceedings{graf2025hidden,
  author = {Emil Graf and Alex Townsend},
  title = {{A hidden variable resultant method for the polynomial multiparameter eigenvalue problem}},
  booktitle = {Linear},
  year = {2025},
  doi = {10.1016/j.laa.2025.07.031},
}
Figure