Abstract
This paper presents a unified framework for determining the mathematical structure and basis of solutions to the matrix equation $G^*JG = J$. By decomposing the problem into solution and cosolution spaces related to the centralizer of the matrix's "cosquare," the authors simplify a computationally intensive $O(n^6)$ problem into an efficient process applicable to real, complex, and singular matrices.
Introduction
In linear algebra and physics, isometry groups represent sets of matrices that preserve specific bilinear or sesquilinear forms, such as distance or area. For a given matrix $J$, these groups are defined by the equation $G^*JG = J$. The corresponding Lie algebra, representing the tangent space at the identity, consists of matrices $X$ satisfying $X^*J + JX = 0$.
Directly solving this linear equation is computationally challenging; a standard approach using Singular Value Decomposition on a flattened $n^2 \times n^2$ matrix scales as $O(n^6)$. While solutions for the complex case exist, a general, efficient method for real matrices, quaternions, and singular $J$ matrices has remained elusive until now.
The Cosquare and Centralizer Decomposition
The authors introduce a methodological breakthrough using the cosquare of a matrix. For a nonsingular $J$, the cosquare is defined as $C = J^{-T}J$. The key observation is that the centralizer of the cosquare—the space of all matrices that commute with $C$—is the direct sum of the "solution" space ($X^TJ + JX = 0$) and a "cosolution" space ($X^TJ - JX = 0$):
$$cent(cosq(J)) = sol(J) \oplus cosol(J)$$
This allows the problem to be solved by finding a basis for the centralizer and then applying projection maps to isolate the solution set.
Classification and Geometric Structure
The paper provides a complete classification of the nine types of real $2 \times 2$ isometry groups. These are categorized based on the eigenvalues of the cosquare and the signature of $J$.
As shown in the classification table, these groups correspond to various topological structures, including circles, hyperbolas, and more complex manifolds like $SL(2, \mathbb{R})$.
Visualization and Numerical Implementation
To handle singular matrices where the cosquare is undefined, the authors employ matrix pencils $(J - \lambda J^T)$ and the Kronecker canonical form to describe the generalized eigenstructure. This approach allows for the visualization of high-dimensional manifolds.
Figure 1 demonstrates how different eigenvalue configurations of the cosquare result in distinct 3D projections of 16-dimensional manifolds. To ensure numerical robustness, the authors recommend implementing these calculations using the staircase format (via the GUPTRI algorithm), which avoids the inherent instabilities of Jordan form computations.
Conclusion
By unifying the treatment of isometry groups through the lens of the cosquare and matrix pencils, Edelman and Jeong provide a powerful framework for both theoretical classification and efficient computation. Their work offers explicit dimension formulas and robust algorithms for sampling from these important mathematical groups, bridging the gap between abstract algebra and numerical analysis.
Summary updated