Eigenvalue collisions for periodic matrix families associated with Ginibre matrices
Carlos Vargas · 2025
Read Eigenvalue collisions for periodic matrix families associated with Ginibre matrices by Carlos Vargas (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-17178.txt.
Introduction
Carlos Vargas studies the rare points where two eigenvalues of a matrix family collide, events called exceptional points. His families blend a fixed random matrix from the Ginibre ensemble with a rotating diagonal matrix whose entries travel along circles or other curves. By tracking the eigenvalues as the parameters loop, he can spot when they swap places, which betrays a collision inside the loop. Software to compute and visualize them accompanies the paper.
The counts show a regularity: for circular paths the number of collisions came out as the matrix dimension times itself minus one, under fairly general conditions. Elliptical and other curves produced even more. The pattern links random matrix theory to the topology of eigenvalue motion. Exceptional points matter in physics, where they mark sharp changes in a system's behavior. Vargas offers both a counting phenomenon to explain and tools to explore it.
Catalog record
| CREATOR | Carlos Vargas |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-17178
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Eigenvalue collisions for periodic matrix families associated with Ginibre matrices Carlos Vargas Abstract We study, count and locate the exceptional points where eigenvalues collide for certain families of matrices R ( s , t ) = cos ( s π / 2 ) C + sin ( s π / 2 ) U ( t ) , s , t ∈ [ 0 , 1 ] R(s,t)=\cos(s\pi/2)C+\sin(s\pi/2)U(t),\quad s,t\in[0,1] where C C is a realization of a Ginibre random matrix, or a closely related matrix, and U ( t ) U(t) is a t t -periodic diagonal matrix whose eigenvalues move along the unit circle or other parameterized curves (simple or not). To do this, we track eigenvalues continuously along simple curves, searching for eigenvalue discrepancies after looping, indicating such collision-events inside said curves. Although stochastic in nature, the collision count for the circle cases repeatedly yielded N ( N − 1 ) N(N-1) under fairly general circumstances, where N N is the matrix dimension. More collisions were observed for the elliptic generalization and for other curves. We include a package to compute and visualize these results. 1 Introduction We consider the two-parameter family of matrices R(s,t)=\cos(s\pi/2)C+\sin(s\pi/2)U(t),\quad s,t\in[0,1] where C\in M_{N}(\mathbb{C}) is a realization of a Ginibre matrix, and U(t) is a t -periodic diagonal matrix with eigenvalues flowing along the unit circle, or more generally, along a curve. In the circular case, we set U(t)=\mathrm{diag}\left(\omega^{tN},\omega^{tN+1},\omega^{tN+2},\dots,\omega^{tN+N-1}\right), where \omega is the first non-trivial counter-clockwise N -th root of unity. In this case the model may be re-written as R(s,t)=\cos(s\pi/2)C+\sin(s\pi/2)\exp(2i\pi t)U(0). The random matrix R(s,0) has been studied in the large- N limit. In particular, Zhong [ 10 ] computed its asymptotic spectral distribution, using ideas from Haagerup and Larsen [ 5 ] on the Brown measure of R -diagonal elements. In this work, we focus on how the eigenvalues evolve as s and t vary, with particular attention to locating the exceptional points (s,t) where eigenvalues collide. Since the dependence on (s,t) is analytic (see Section 3.5 ), eigenvalues can be continuously tracked. We are especially interested in the monodromy of the eigenvalues arising from the evolution in t : For fixed s , as t runs from 0 to 1 , the eigenvalues of R(s,t) rotate counter-clockwise at speeds inversely correlated to their moduli. After one full period in t , the eigenvalues must return to the original configuration, but they don’t generally end up at their same starting positions, thus defining a permutation \sigma(s) , for all s\in[0,1] . At s=0 and s=1 , this permutation is trivial, but for intermediate values of s , \sigma(s) is a piece-wise constant process, with different values explained by composing consecutive transpositions associated to eigenvalue collisions. By tracking the evolution of \sigma(s) finely across s\in[0,1] , we locate the points where \sigma(s) jumps, corresponding to collisions. We repeatedly observed a count of N(N-1) of such collision-events for the circular case, an increased (stochastic) number of collisions for other curves, and also some variations in the elliptic case. The count for the circular case, although surprisingly constant, is indeed stochastic: the commutative version of the model (included in the non-commutative model with overwhelmingly small probability) yields N(N-1)/2 collisions (see Section 3.4 ). Eigenvalue collisions, more precisely, correspond to points (s,t,\lambda) where R(s,t) has a repeated eigenvalue \lambda , of order 2 or greater. We detect these using the monodromy action around small square loops in the (s,t) -space: if the eigenvalues do not return to their original ordering after traversing the square loop, then a collision must exist inside the square. The dynamics of eigenvalues as s and t vary are visually striking: swirling trajectories, local repulsion, and a discrete set of cycle-merging/separating-events. When U(t) flows along the circle, the eigenvalues initially revolve around their own centers, before teaming-up to rotate around the origin, until finally each eigenvalue loops by itself, as the value of s increases. We elaborate on the model and its variants in Section 3 . We use permutations to detect and count eigenvalue collisions (Section 4 ). In Section 5 we discuss statistical properties of these collisions and examine the influence of alternative curves, initial eigenvalue distributions and eccentricity on the collision counts. Section 6 briefly describes implementation details. Appendix A provides high-resolution illustrations of the eigenvalue-permutation evolution for N=10 under four different scenarios and three different curves. Appendix B presents analogous results for elliptic deformations. Appendix C includes eigenvalue tracks for traceless Bernoulli matrices, highlighting specific cases with repeated eigenvalues. Appendix D displays eigenvalue tracks for the circular case with N=100 . We provide a software package for computing and visualizing these phenomena https://github.com/carlos-vargas-math/eigenvalue-collisions . To give the reader a general feeling for these eigenvalue dynamics, we prepared some videos showing the evolution and first collisions (w.r.t. the parameter s ) for a case with N=20 (available at https://carlos-vargas-math.github.io/eigenvalue-collisions/animation ). 2 Acknowledgements The author is grateful to O. Arizmendi, A. Beshenov, D. López and J. Santos, for constant feedback and moral support to conclude this project. The author acknowledges P. Zhong, for sharing formula ( [ 10 ] , Th. 8.8) which was used in our model, C. Díaz-Aguilera and N. Sakuma, for suggesting interesting examples, and R. Speicher, for general remarks that improved this paper. The author declares that there is no conflict of interest regarding the publication of this work. 3 Model Description 3.1 Ginibre Matrix Figure 1: Eigenvalues of complex Gaussian Ginibre matrix (left), symmetric Bernoulli Ginibre matrix (center) and 200 independent points in the unit disc (right) A Ginibre matrix is an N\times N random matrix, consisting of independent, identically distributed entries with zero-mean and the same finite variance \sigma^{2} . The well-known circular law [ 2 , 3 , 1 , 4 , 9 ] states that if \sigma^{2}=1/N , the empirical distribution of the eigenvalues converges almost surely, as N\to\infty , to the uniform distribution on the centered unit disk. Similar to the central limit theorem, the circular law is universal: it holds for matrices with independent, centered, finite-variance entries, regardless of the underlying distribution (e.g., complex Gaussian, real Gaussian, symmetric Bernoulli; see Figure 1 ). In Figure 1 , we observe that the real-valued matrix has conjugate eigenvalues (as expected), but otherwise, the two eigenvalue distributions (left and center) appear fairly similar and spread uniformly over the unit disk. For the complex Gaussian case, Ginibre computed the explicit joint density of eigenvalues in 1965 [ 2 ] : \rho(\lambda_{1},\dots,\lambda_{N})=\frac{1}{\pi^{n}\prod_{k=1}^{N}k!}\exp(-\sum_{k=1}^{N}|\lambda_{k}|^{2})\prod_{1\leq j<k\leq N}|\lambda_{k}-\lambda_{j}|^{2} The eigenvalues of the conjugate transpose C^{*} of a Ginibre matrix are simply the conjugates of the eigenvalues of C . More generally, the eigenvalues of the matrix E_{x}=\cos(x)C+\sin(x)C^{*} follow the elliptic law (see [ 7 , 8 ] ): asymptotically, they are uniformly distributed on an ellipse with semi-axes 1+\rho and 1-\rho , where \rho=\sin(2x) . Figure 2 illustrates this behavior for several values of x . Figure 2: Eigenvalues of E_{x} for some values of x\in[0,\pi/4] (respectively \rho=0,0.312,0.436,0.6,0.8,0.917 ) The eigenvalues of Ginibre matrices exhibit repulsion (as seen from the Vandermonde determinant factor), leading to highly uniform distributions and faster, stronger convergence to the asymptotic limit, especially when compared