Complex Elliptic Ginibre Ensemble

• Complex elliptic Ginibre ensemble is a family of non-Hermitian random matrices interpolating between the Ginibre and GUE ensembles via a tunable Hermiticity parameter.
• It exhibits an exact determinantal point process structure with elliptical spectral laws and clear regimes of strong and weak non-Hermiticity reflected in kernel limits.
• The model provides practical insights into eigenvector overlaps, spectral diagnostics, and crossover phenomena crucial for understanding universality in random matrix theory.

The complex elliptic Ginibre ensemble is a one-parameter Gaussian family of non-Hermitian random matrices that interpolates between the complex Ginibre ensemble and the Gaussian unitary ensemble (GUE). It is simultaneously a matrix model with correlated entries, a determinantal point process in the complex plane, and a two-dimensional one-component Coulomb gas at inverse temperature $\beta=2$ in a quadrupolar field. Its interpolation parameter controls the passage from a genuinely two-dimensional complex spectrum to an almost Hermitian regime, and the resulting model supports a detailed theory of global elliptic laws, weak- and strong-non-Hermiticity limits, eigenvector non-orthogonality, characteristic polynomial asymptotics, and time-domain spectral diagnostics (Akemann et al., 2022, François et al., 2023).

1. Definition and exact determinantal structure

A standard formulation defines the ensemble on complex $n\times n$ matrices $M_n$ by

$$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$

so that $\tau=0$ gives the complex Ginibre ensemble and $\tau\uparrow1$ approaches the Hermitian regime (Bothner et al., 2022). Equivalent Gaussian realizations write the matrix as a linear combination of two independent GUE matrices; one form is

$$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$

with $X_n$ and $Y_n$ independent GUE matrices, while another uses the Hermitian and anti-Hermitian parts of a complex Ginibre matrix (François et al., 2023, Byun et al., 24 Feb 2026). Different papers use $\tau$, $n\times n$0, or $n\times n$1 for the same Hermiticity parameter and either fix the entry variance to $n\times n$2 or divide the matrix by $n\times n$3; these are normalization conventions rather than distinct models.

For $n\times n$4, the ensemble is a determinantal point process with kernel

$n\times n$5

with elliptic Gaussian weight

$n\times n$6

The correlation functions are determinants of this kernel, and the orthogonal-polynomial input is supplied by planar Hermite polynomials (Akemann et al., 2022). This exact solvability underlies most asymptotic results for the model.

2. Global spectral geometry and the numerical range

At macroscopic scale, the empirical eigenvalue distribution converges to the uniform measure on an ellipse. In one standard normalization the limiting support is

$n\times n$7

with semiaxes $n\times n$8 and $n\times n$9; at $M_n$0 this is the circular law, while $M_n$1 degenerates to the interval $M_n$2 in the Hermitian limit (François et al., 2023). In the kernel formalism, the limiting density is

$M_n$3

which is the elliptic law in projection-kernel form (Akemann et al., 2022).

The eigenvalue ellipse does not exhaust the geometry relevant to non-normality. For the elliptic Ginibre ensemble, the numerical range $M_n$4 has a distinct large-$M_n$5 limit. If $M_n$6 is the elliptic Ginibre matrix with Ginibre entry variance $M_n$7, then

$M_n$8

where

$M_n$9

Thus the limiting numerical range is again an ellipse, but with semiaxes enlarged by a factor $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$0 relative to the spectral support (Byun et al., 24 Feb 2026).

This distinction is structurally important. The numerical range is always convex; for normal matrices it equals the convex hull of the spectrum, whereas for non-normal matrices it contains the spectrum and captures instability, pseudospectral effects, and eigenvector non-orthogonality. In the elliptic Ginibre case, the spectrum lies in the ellipse with axes $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$1 and $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$2, while the numerical range converges to the “fatter” ellipse with axes $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$3 and $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$4. In the Hermitian limit $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$5, the numerical-range ellipse collapses to $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$6, matching the convex hull of the limiting GUE spectrum (Byun et al., 24 Feb 2026).

3. Strong and weak non-Hermiticity regimes

At fixed $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$7, the ensemble is in the strong non-Hermiticity regime. The global support remains elliptic, and local bulk correlations are Ginibre. For both the standard elliptic ensemble and certain non-Gaussian fixed-trace or trace-squared extensions, the microscopic bulk kernel converges to

$$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$8

so the strong-non-Hermitian bulk universality class is the same as Ginibre (Akemann et al., 2016). A complementary contour-integral analysis shows that, in the bulk, the residue at $$\mathbb{P}_{n,\tau}(M_n)=Z_{n,\tau}^{-1}\exp\left\{-\frac{n}{1-\tau^2}\Big(M_n^\dagger M_n-\frac{\tau}{2}\big(M_n^2+(M_n^\dagger)^2\big)\Big)\right\}\,dM_n,\qquad \tau\in[0,1),$$9 dominates and produces the Ginibre kernel, whereas boundary saddle points control macroscopic edge contributions (Akemann et al., 2022).

Weak non-Hermiticity is obtained by letting $\tau=0$0 approach $\tau=0$1 with $\tau=0$2. In the bulk, one standard scaling is

$\tau=0$3

or equivalently $\tau=0$4 at a bulk point $\tau=0$5. In this regime the global support collapses to the real axis with semicircular density, but the local point process remains genuinely complex and converges to the Fyodorov–Khoruzhenko–Sommers kernel

$\tau=0$6

As $\tau=0$7, this collapses to the Hermitian sine-kernel limit after integrating out the imaginary directions; as $\tau=0$8, it recovers Ginibre statistics after the appropriate rescaling (Akemann et al., 2016).

Two observable-specific weak-non-Hermiticity crossover laws have been derived in detail. For bulk spacings of ordered real parts, the spacing distribution is a generalized Gaudin–Mehta law $\tau=0$9, defined through a Fredholm determinant of a two-dimensional weakly non-Hermitian sine kernel. It interpolates between the GUE Gaudin–Mehta distribution $\tau\uparrow1$0 as $\tau\uparrow1$1 and the Poisson law $\tau\uparrow1$2 as $\tau\uparrow1$3 (Bothner et al., 2022). At the right spectral edge, with

$\tau\uparrow1$4

the limiting distribution $\tau\uparrow1$5 of the rightmost eigenvalue is represented by a Fredholm determinant and by an integro-differential Painlevé-II-type equation. It interpolates between the Tracy–Widom GUE law as $\tau\uparrow1$6 and a Gumbel limit after the appropriate centering and scaling as $\tau\uparrow1$7 (Bothner et al., 2022).

4. Eigenvectors, overlaps, and non-normal dynamics

Unlike the GUE, the complex elliptic Ginibre ensemble is intrinsically non-normal, so left and right eigenvectors are distinct and their overlaps are nontrivial. In the stochastic-process formulation, one considers a matrix-valued diffusion $\tau\uparrow1$8 with Hermiticity parameter $\tau\uparrow1$9. Its complex eigenvalues satisfy Dyson-like stochastic differential equations

$$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$0

while the quadratic covariations are governed exactly by eigenvector overlaps: $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$1 Thus the overlap processes are not auxiliary observables; they are the local covariance structure of eigenvalue motion. For $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$2 one recovers Dyson Brownian motion, for $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$3 the drift vanishes and the eigenvalues are complex martingales, and for all $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$4 eigenvalue collisions are absent almost surely (Yabuoku, 2020).

At finite $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$5, the mean diagonal overlap in the complex elliptic ensemble admits an exact formula in terms of the finite-$$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$6 density $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$7 and an auxiliary Hermite-polynomial sum $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$8: $$A_{n,t}=\sqrt{\frac{1+t}{2}}\,X_n+i\sqrt{\frac{1-t}{2}}\,Y_n,$$9 In the strong-non-Hermitian bulk, after scaling $X_n$0 with $X_n$1,

$X_n$2

This is the elliptically deformed analogue of the Ginibre bulk overlap profile (Crumpton et al., 2024).

The geometric content is that overlap is largest in the interior of the droplet and vanishes at the boundary at leading order. In the weak-non-Hermiticity regime, the same paper derives an explicit integral formula for the weak-bulk mean self-overlap, showing how non-orthogonality persists in the thin complex cloud around the real axis (Crumpton et al., 2024). The $X_n$3 dynamic calculation gives a particularly transparent picture: diagonal overlap is negatively correlated with the squared eigenvalue gap, so closer eigenvalues carry larger overlap and faster motion (Yabuoku, 2020).

5. Characteristic polynomials and dissipative spectral form factors

Outside the limiting ellipse, the normalized characteristic polynomial has a nontrivial random analytic limit. With

$X_n$4

which conformally maps $X_n$5 onto $X_n$6, define

$X_n$7

Then

$X_n$8

where

$X_n$9

The limiting object is therefore an explicit deterministic analytic factor times the exponential of a Gaussian analytic function with elliptically deformed covariance (François et al., 2023).

A distinct line of work studies the dissipative spectral form factor (DSFF), defined for complex time $Y_n$0 through the Fourier modes of the complex eigenvalues. In the large-$Y_n$1 regime with

$Y_n$2

the DSFF decomposes into disconnected and connected parts and exhibits a dip–ramp–plateau structure throughout the Ginibre–GUE crossover. The rigorous asymptotic phase diagram identifies three ramp types: a quadratic Ginibre-like ramp, a linear GUE-like ramp, and a mixed regime in which both appear. The mesoscopic window

$Y_n$3

is the interpolation regime between two-dimensional non-Hermitian and one-dimensional Hermitian behavior. The associated time scales satisfy

$Y_n$4

so strong non-Hermiticity gives $Y_n$5 and $Y_n$6, whereas weak non-Hermiticity gives $Y_n$7 and $Y_n$8 (Akemann et al., 27 May 2026). Finite-$Y_n$9 exact Laguerre-polynomial formulas display the same dip–ramp–plateau pattern and provide explicit crossover estimates for the Thouless and Heisenberg times (Sen et al., 2024).

6. Extensions, adjacent structures, and mathematical context

The standard Gaussian ensemble sits inside a broader elliptic universality class. Non-Gaussian fixed-trace and trace-squared deformations, although non-determinantal at the eigenvalue level, can be linearized into averages of determinantal elliptic ensembles. In the strong-non-Hermitian regime they retain Ginibre bulk universality, and in the weak regime they converge to the same Fyodorov–Khoruzhenko–Sommers kernel. This provides a rigorous universality theorem beyond the integrable Gaussian model (Akemann et al., 2016).

The kernel technology also extends to higher dimensions. In $\tau$0, the droplet becomes

$\tau$1

the strong-non-Hermitian bulk limit is a product of $\tau$2 Ginibre kernels, and weak non-Hermiticity interpolates between $\tau$3-dimensional real and $\tau$4-dimensional complex universality classes (Akemann et al., 2022). This suggests that the one-dimensional complex elliptic ensemble is the $\tau$5 member of a larger family rather than an isolated model.

Several neighboring theories illuminate the same structure from different angles. The singular values of a spiked elliptic Ginibre extension form a Pfaffian point process, with contour-integral kernels, bulk sine-kernel universality, a BBP transition at the soft edge for fixed $\tau$6, and additional crossover phenomena as $\tau$7 (Liu et al., 2018). On the orthogonal-polynomial side, the elliptic Ginibre weight

$\tau$8

supports a family of polyanalytic Hermite polynomials obtained from Bogoliubov-squeezed creation and annihilation operators; these furnish orthogonal bases for Landau levels and connect the ensemble to two-photon coherent states and the metaplectic representation of $\tau$9 (Demni et al., 28 Jan 2025).

By contrast, some exact structures of the circular Ginibre ensemble do not survive elliptic deformation. The power-map decomposition of Ginibre eigenvalues into independent determinantal blocks relies on full radial symmetry, and this does not directly extend to the elliptic setting, which breaks rotational invariance (Dubach, 2017). This contrast clarifies a basic structural fact: the complex elliptic Ginibre ensemble preserves enough integrability to admit exact kernels, contour formulas, and solvable crossover limits, but not the full radial algebra of the circular model.