Ginibre Spectral Correlations

• Ginibre spectral correlations are the joint statistics of complex eigenvalues, singular values, and overlap observables in non-Hermitian random-matrix ensembles, revealing key determinantal and Pfaffian structures.
• The theory demonstrates how global spectral distributions converge to the circular law under quadratic confinement and logarithmic repulsion while finite-scale structures are encoded by specialized kernels.
• Extensions to products, singular values, and parametric perturbations highlight the rich interplay between local universality classes and regime-dependent deviations in non-Hermitian systems.

Ginibre spectral correlations are the joint statistics of complex eigenvalues, singular values, and overlap observables in non-Hermitian random-matrix ensembles of Ginibre type. In the complex Ginibre ensemble, the global eigenvalue distribution converges to the circular law on the unit disk, while finite-scale structure is encoded by determinantal correlation kernels; in adjacent settings one encounters Pfaffian processes, overlap-weighted determinantal structures, Meijer-$G$ weights for products, and new local kernels at edges, singular points, and under parametric perturbations (Liu et al., 2022, Gu, 20 Oct 2025, Fyodorov et al., 18 Jun 2026).

1. Basic ensembles and the meaning of spectral correlations

For the complex Ginibre ensemble, one considers random $N\times N$ matrices $G$ with i.i.d. complex Gaussian entries of mean $0$ and variance $1/N$. The eigenvalues are complex, the empirical spectral measure converges to the uniform measure on the unit disk, and “spectral correlations” refer to the joint statistics of eigenvalues at finite scales: $k$-point correlation functions, spacing statistics, and the structure of the limiting point process in bulk, edge, and outlier regimes. In the complex case the eigenvalues form a determinantal point process; in the real quaternion case they form a Pfaffian process (Liu et al., 2022).

A standard Coulomb-gas representation appears in the complex Ginibre ensemble with macroscopic scaling. If $z_1,\dots,z_N\in\mathbb C$ are the eigenvalues, then the joint density can be written as

$$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$

which makes explicit the competition between quadratic confinement and logarithmic repulsion (Allez et al., 2013). This representation underlies both global large-deviation problems and local determinantal asymptotics.

A second basic distinction is between complex eigenvalues and singular values. For non-Hermitian products $Y_M=X_M\cdots X_1$, singular values are the eigenvalues of the positive Hermitian matrix $Y_M^*Y_M$, or equivalently the squared singular values $N\times N$0. Their correlation theory is often more tractable because determinantal structures, $N\times N$1-function integral formulas, and steepest-descent methods apply directly (Gu, 20 Oct 2025).

2. Exact correlation structures: determinantal, Pfaffian, and overlap-weighted forms

For products of $N\times N$2 independent complex Ginibre matrices, the complex eigenvalues of the product matrix $N\times N$3 again form a determinantal point process. The joint density has the form

$N\times N$4

with radial weight

$N\times N$5

and kernel

$N\times N$6

All $N\times N$7-point functions are determinants of this kernel (Akemann et al., 2012). The Meijer-$N\times N$8 weight is the signature of multiplicative non-Hermiticity in the complex plane.

The same determinantal mechanism persists when eigenvector overlaps are inserted, although the kernel changes. For the induced Ginibre unitary ensemble, the $N\times N$9-th correlation functions weighted by on- and off-diagonal overlaps retain a determinantal structure. In the strongly non-unitary regime, the bulk and edge scaling limits of these overlap-weighted $G$0-point functions are universal, and in the weakly non-unitary regime as well as at the singular origin the overlap kernels acquire new functional forms tied to the corresponding spectral regime (Noda, 2023).

Not every Ginibre-type deformation remains determinantal at finite $G$1. In the fixed trace induced Ginibre ensemble, the eigenvalue density contains a collective factor $G$2, so the usual product-form weight is lost and the ensemble is not determinantal. Nevertheless, all finite-$G$3 $G$4-point correlation functions can still be computed exactly by Laplace-transforming the correlation functions of the canonical induced Ginibre ensemble (Akemann et al., 2017).

Pfaffian structures enter in non-complex symmetry classes. For additive finite-rank deformations of the real quaternion Ginibre ensemble, the critical real edge is Pfaffian, while complex-edge limits remain determinantal and are governed by new kernels expressed through repeated complementary-error-function integrals (Liu et al., 2022).

3. Global laws, constrained spectra, and macroscopic deformations

At the global level, the circular law is the basic reference point, but Ginibre spectral correlations are equally shaped by anisotropy, conditioning, and constraints. In the elliptic Ginibre ensemble, the spectral support is an ellipse rather than a disk. For strong non-Hermiticity, the boundary point is

$G$5

and the limiting support is the corresponding elliptic droplet; in higher dimensions this becomes an ellipsoid in $G$6 (Crumpton et al., 2024, Akemann et al., 2022).

A prototypical macroscopic constraint is the Ginibre index problem. If

$G$7

then for both complex and real Ginibre ensembles

$G$8

with $G$9 for the complex ensemble and $0$0 for the real ensemble, and with the same explicit rate function $0$1. The minimizer is $0$2, and $0$3 has a third-order phase transition at $0$4. As a consequence, central fluctuations of $0$5 occur on the $0$6 scale rather than the Gaussian $0$7 scale familiar from one-dimensional log-gases (Allez et al., 2013).

Another global conditioning problem arises in the real Ginibre ensemble when one imposes $0$8 real eigenvalues with $0$9. In that regime the empirical measure satisfies a constrained large deviations principle, and the equilibrium measure decomposes as

$1/N$0

The real component is supported on a compact interval, whereas the complex component has constant density on a bounded domain $1/N$1 that is separated from the real axis. The resulting two-phase log-gas interpolates between the circular law at $1/N$2 and the semicircle law at $1/N$3 (Molino et al., 2015).

These macroscopic problems show that Ginibre correlations are not exhausted by the unconditioned disk law. Conditioning can create annuli, singular circle measures, phase-separated supports, and third-order transitions without leaving the Ginibre variational framework. This suggests that global Ginibre spectral correlations are best understood as Coulomb-gas equilibria under geometry- and constraint-dependent admissible measures.

4. Local universality, edge kernels, and non-Hermiticity crossovers

The local theory is dominated by universality statements, but the relevant universal object depends strongly on regime. In the elliptic Ginibre ensembles, the strong-non-Hermiticity edge density is the same for the real and complex ensembles: $1/N$4 with no angular dependence in the leading density. At weak non-Hermiticity, however, the two ensembles diverge: in the complex case the edge density factorizes into a Gaussian in the imaginary direction and an Airy-type integral in the real direction, whereas in the real case the imaginary dependence becomes $1/N$5 and the Airy factor is replaced by $1/N$6. The same regime dependence appears in eigenvector self-overlaps (Crumpton et al., 2024).

A higher-dimensional synthesis is available for the elliptic Ginibre ensemble. A single contour-integral representation of the kernel gives the global elliptic law, boundary-dominated two-point correlations, bulk Ginibre-type kernels in $1/N$7, and weakly non-Hermitian crossovers to $1/N$8-dimensional sine- and Airy-type kernels in $1/N$9. In one complex dimension this recovers the usual interpolation between complex Ginibre and GUE-type local statistics; in higher dimensions it yields product Ginibre bulk kernels and harmonic-oscillator fermion kernels as two limits of the same steepest-descent problem (Akemann et al., 2022).

Finite-rank deformations introduce a different local transition. When the deformation has eigenvalues on the unit circle but no supercritical spikes, the edge statistics are no longer the standard Ginibre edge. Instead, the critical edge kernel becomes

$k$0

where $k$1 is the number of Jordan blocks associated with the spike eigenvalue, and the kernel is expressed in terms of repeated complementary-error-function integrals $k$2. In the complex case this yields a new determinantal edge universality class; in the real quaternion case the real edge is Pfaffian with kernel $k$3 and complex edges remain determinantal. Remarkably, the critical edge depends on geometric multiplicity but not on Jordan block sizes, whereas supercritical outlier fluctuations do depend on the full Jordan structure (Liu et al., 2022).

Taken together, these results isolate a recurring pattern: bulk correlations are often robust, but edge correlations are sensitive to deformation type, symmetry class, and the strong-versus-weak non-Hermiticity scaling. This suggests that “Ginibre universality” is local and regime-specific rather than monolithic.

5. Products, singular values, rectangularity, and multiplicative transitions

Products of Ginibre matrices produce several distinct correlation theories. For complex eigenvalues of the product $k$4, the macroscopic density is the $k$5-th power of the circular law after rescaling, and the finite-$k$6 kernel is explicitly determinantal. In the large-$k$7 limit with fixed $k$8, the microscopic bulk and edge correlations, after unfolding, are identical to those of the single Ginibre ensemble. By contrast, the origin is non-universal: for $k$9 the kernel becomes a Meijer-$z_1,\dots,z_N\in\mathbb C$0 weight times a hypergeometric

$z_1,\dots,z_N\in\mathbb C$1

which generalizes the $z_1,\dots,z_N\in\mathbb C$2-Bessel law in the complex plane at $z_1,\dots,z_N\in\mathbb C$3 (Akemann et al., 2012).

For singular values of products of square Ginibre matrices when both the number of factors $z_1,\dots,z_N\in\mathbb C$4 and the matrix size $z_1,\dots,z_N\in\mathbb C$5 grow, the local statistics are again determinantal but now interpolate between two Hermitian universality classes. In the bulk, the limiting kernel interpolates between the GUE sine kernel and a rigid “picket-fence” kernel; at the hard edge the statistics are picket-fence; at the soft edge one obtains either the Airy kernel, a truncated picket-fence kernel, or a transitional kernel depending on the scaling of $z_1,\dots,z_N\in\mathbb C$6. A crucial point is that the soft edge requires a mesoscopic unfolding scale, and the same local kernels appear in Dyson Brownian motion with equidistant initial positions, although that correspondence is only local (Akemann et al., 2020).

Rectangular products shift the picture again. For

$z_1,\dots,z_N\in\mathbb C$7

with $z_1,\dots,z_N\in\mathbb C$8, rectangularity parameters

$z_1,\dots,z_N\in\mathbb C$9

and $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$0, the limiting Stieltjes transform of the squared singular values satisfies

$$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$1

In the homogeneous case $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$2, this admits an explicit parametric solution for the density. More importantly, for the logarithms of the squared singular values, the bulk $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$3-point correlations converge to the sine-kernel determinantal process after centering and scaling by the local density. Rectangularity therefore changes the global law and the microscopic scale factor $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$4, but not the universal bulk kernel (Gu, 20 Oct 2025).

A different multiplicative phenomenon appears at the level of the spectral radius of products of complex Ginibre matrices. If $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$5 are the complex eigenvalues of $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$6 and $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$7, then a suitably rescaled version of $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$8 converges to a non-trivial law $$P(z_1,\cdots,z_N) = \frac{1}{Z_N} \exp\left(- N \sum_{k=1}^N |z_k|^2 + 2 \sum_{i< j} \ln |z_i-z_j|\right),$$9 for $Y_M=X_M\cdots X_1$0, to the Gumbel distribution for $Y_M=X_M\cdots X_1$1, and to the standard normal distribution for $Y_M=X_M\cdots X_1$2, with explicit convergence rates in Kolmogorov and Wasserstein distance. This gives a Gaussian–$Y_M=X_M\cdots X_1$3–Gumbel phase transition in an extreme-value observable of product Ginibre spectra (Ma et al., 9 Oct 2025).

6. Parametric correlations, overlaps, and the limits of universality

Parameter dependence reveals another layer of Ginibre spectral correlations. For the parameter-dependent complex Ginibre family

$Y_M=X_M\cdots X_1$4

the connected two-point density correlator in the bulk satisfies

$Y_M=X_M\cdots X_1$5

with

$Y_M=X_M\cdots X_1$6

The associated parametric number covariance in a microscopic disk has a closed form in terms of $Y_M=X_M\cdots X_1$7, the mean number $Y_M=X_M\cdots X_1$8, and modified Bessel functions. In the small-$Y_M=X_M\cdots X_1$9 limit the same analysis yields the overlap distribution

$Y_M^*Y_M$0

linking parametric spectral motion directly to eigenvector non-orthogonality. Numerical evidence indicates that the same scaling function describes real Ginibre, non-Hermitian Bernoulli Wigner, and bi-unitarily invariant ensembles under suitable global diffusive dynamics (Fyodorov et al., 18 Jun 2026).

Universality, however, has clear boundaries. In open Floquet chaotic systems with localized leaks, the long-lived resonances follow truncated circular orthogonal ensemble statistics rather than Ginibre statistics: the density of states can move toward the Ginibre limit as the leak size increases, yet nearest-neighbor spacing distributions and ratio statistics remain distinct. Short-lived resonances show no clear correspondence with either Ginibre or truncated circular ensembles (Signor et al., 9 Dec 2025).

An even sharper limitation appears in dissipative quantum chaos. For the dissipative kicked top, a systematic exploration of parameter space in a genuine semiclassical limit shows that Ginibre spectral correlations are neither a robust nor a universal diagnostic of dissipative quantum chaos. The correspondence principle formulated as “chaotic attractor $Y_M^*Y_M$1 Ginibre spectrum” fails even in the original kicked model: Ginibre-like complex spacing statistics can appear in regimes with simple classical attractors, and chaotic attractors do not force Ginibre behavior (Villaseñor et al., 24 Jul 2025).

This suggests a precise but qualified conclusion. Ginibre spectral correlations are highly robust in many random-matrix settings—especially for determinantal bulk limits, overlap-weighted kernels, and multiplicative models with appropriate unfolding—but their universality is contingent on the observable, the symmetry class, the geometry of non-Hermiticity, and the scaling regime.