Local Eigenvalue Statistics (LES)

• Local Eigenvalue Statistics is the study of microscopic eigenvalue behavior using rescaled coordinates, spacing, and k-point correlations.
• In delocalized ensembles, LES exhibits universal limits via sine and Airy kernels that underpin gap distributions and robust correlation functions.
• In localized operators, LES typically follows Poisson or compound Poisson distributions, highlighting regimes with level clustering and clock processes.

Local eigenvalue statistics describe the fine-scale behavior of eigenvalues after an appropriate local rescaling. For ordered eigenvalues $\lambda_1\le \cdots \le \lambda_N$ and a bulk energy $E$ with limiting density $\rho(E)>0$, the microscopic scale is the mean spacing $\Delta_N(E)=1/(N\rho(E))$, and the basic observables are rescaled coordinates, gaps, $k$-point correlation functions, and local point processes. Across the literature, LES separates sharply according to spectral regime: in delocalized ensembles it is governed by universal Wigner–Dyson limits, while in localized operators it is frequently Poisson, with notable exceptions such as compound Poisson and clock processes (Erdos et al., 2011, Bourgain, 2013, Hislop et al., 28 Aug 2025).

1. Definitions, rescaling, and basic observables

For a random matrix or finite-volume operator with eigenvalues $\lambda_1\le \cdots \le \lambda_N$, LES probe eigenvalues on the microscopic scale where neighboring eigenvalues interact. At a bulk energy $E$ with limiting density $\rho(E)$, one introduces

$$\Delta_N(E):=\frac{1}{N\rho(E)},\qquad x_i:=\frac{\lambda_i-E}{\Delta_N(E)},\qquad s_i:=\frac{\lambda_{i+1}-\lambda_i}{\Delta_N(E)}.$$

The $k$-point correlation functions are the marginals of the joint eigenvalue density, and their local rescaling is the standard object in bulk universality statements. Bulk refers to energies in the interior of the support of $E$0 where $E$1, while the soft edge refers to energies near the spectral boundary where $E$2 vanishes continuously (Erdos et al., 2011).

For random Schrödinger operators and finite-volume restrictions, the same microscopic principle is typically encoded by a rescaled point process. In the one-dimensional Anderson–Bernoulli model, for example,

$E$3

and Poisson LES means that $E$4 converges to a unit-intensity Poisson process, equivalently that the spacing distribution converges to $E$5 (Bourgain, 2013). In fixed-bandwidth random band matrices the analogous rescaling is by $E$6, with unit intensity recovered after multiplying by the limiting density of states $E$7 (Brodie et al., 2020).

2. Universal limits in delocalized ensembles

The central discovery of random matrix theory is that local statistics are universal: they depend only on the symmetry class and the spectral regime, and are independent of microscopic details such as entry distributions or the confining potential. For invariant $E$8-ensembles, bulk universality is expressed by convergence of the rescaled $E$9-point functions to limits independent of $\rho(E)>0$0; for $\rho(E)>0$1 the limiting bulk process is determinantal with the sine kernel

$\rho(E)>0$2

while for $\rho(E)>0$3 the limits are Pfaffian. At the soft edge, after $\rho(E)>0$4 scaling, the universal limit is governed by the Airy kernel and Tracy–Widom laws. Small-gap behavior in the bulk satisfies $\rho(E)>0$5, reflecting level repulsion (Erdos et al., 2011).

For Wigner matrices, the Four Moment Theorem shows that local bulk statistics are determined by the first four moments of the off-diagonal entries and the first two moments of the diagonal entries; as a consequence, gap distributions and $\rho(E)>0$6-point correlations agree with the GOE or GUE limits under mild tail assumptions (0906.0510). Universality persists in deformed settings: for $\rho(E)>0$7, where $\rho(E)>0$8 is a deterministic diagonal external source, the global semicircle law fails, but the bulk correlation functions still satisfy sine-kernel universality (O'Rourke et al., 2013). It also persists in constrained and heavy-tailed models. For uniform random $\rho(E)>0$9-regular graphs with $\Delta_N(E)=1/(N\rho(E))$0, the bulk gaps and locally averaged $\Delta_N(E)=1/(N\rho(E))$1-point correlation functions coincide with those of the GOE (Bauerschmidt et al., 2015). For Lévy matrices with i.i.d. $\Delta_N(E)=1/(N\rho(E))$2-stable entries, bulk universality and complete delocalization hold in the regimes established in the paper, and the local statistics converge to those of the GOE after unfolding by the free $\Delta_N(E)=1/(N\rho(E))$3-stable density (Aggarwal et al., 2018).

3. Poisson statistics in localized operators

In localized regimes, LES are expected to display absence of level repulsion. The one-dimensional Anderson–Bernoulli Hamiltonian at small coupling, under an explicit algebraic condition on the disorder and for energies in the interior window $\Delta_N(E)=1/(N\rho(E))$4, exhibits Poisson local eigenvalue statistics: the rescaled point process converges to a unit-intensity Poisson process and the nearest-neighbor spacings converge to the exponential law (Bourgain, 2013).

The same Poisson picture appears in several other localized models. For one-dimensional random band matrices of fixed bandwidth, the local eigenvalue statistics are Poisson with intensity given by the limiting density of states $\Delta_N(E)=1/(N\rho(E))$5; after microscopic rescaling by $\Delta_N(E)=1/(N\rho(E))$6, the limit has unit intensity (Brodie et al., 2020). For one-dimensional random band matrices with bandwidth growing as $\Delta_N(E)=1/(N\rho(E))$7, the paper proves that, in the Gaussian case for $\Delta_N(E)=1/(N\rho(E))$8, and more generally under assumed localization bounds for $\Delta_N(E)=1/(N\rho(E))$9, any nontrivial limit points of the local counting variables are Poisson distributed (Hislop et al., 2021). In the continuum, Schrödinger operators with random point interactions on $k$0, $k$1, have Poisson LES in the localization regime, with intensity measure $k$2 at energies where the density of states is positive (Hislop et al., 2019).

Higher-rank models can also become purely Poisson when multiplicity-producing mechanisms are sufficiently controlled. For the higher-rank lattice Anderson model with uniform rank $k$3, the paper proves that near the two spectral band edges the unfolded point process converges to a Poisson process with intensity $k$4, improving an earlier compound Poisson result (Herschenfeld et al., 2022).

4. Compound Poisson, clock processes, and other exceptional limits

Localized spectra do not always produce simple Poisson limits. For generalized lattice Anderson models with finite-rank perturbations

$k$5

the local counting variables are compound Poisson distributed in the localized regime. In the lattice case, the Lévy measure is supported on $k$6, where $k$7 is the rank of the single-site projection; rank one recovers Poisson, while $k$8 permits persistent local multiplicity and clustering (Hislop et al., 2014). A related phenomenon occurs on the canopy tree: for higher-rank perturbations of rank

$k$9

the eigenvalue-counting point process converges, at large disorder, to a compound Poisson process with cluster sizes supported in $\lambda_1\le \cdots \le \lambda_N$0 (A., 2017).

Random polymer models provide a different exceptional class. For one-dimensional random polymer models, the unfolded LES centered at finitely many critical energies is a uniform clock process, while at noncritical energies it is a Poisson point process under local Hölder assumptions on the IDS and its inverse. At a critical energy $\lambda_1\le \cdots \le \lambda_N$1, the strong clock spacing statement is

$\lambda_1\le \cdots \le \lambda_N$2

and the paper further shows that the transition from clock to Poisson in the unfolded LES is sharp at the critical energies (Hislop et al., 28 Aug 2025).

Outside the self-adjoint setting, the local statistics can be governed by zeros of Gaussian analytic functions. For one-dimensional non-selfadjoint semiclassical pseudodifferential operators with small random perturbations, the rescaled local spectrum converges either to the zero set of a product of independent Gaussian analytic functions in the random-potential case, or to the zero set of the determinant of a matrix of independent Gaussian analytic functions in the random-matrix case. Here the local statistics depend on the perturbation type and the local spectral density, but not on the detailed law of the perturbation under the stated moment assumptions (Nonnenmacher et al., 2017).

5. Mechanisms of proof

In delocalized random matrix models, the modern analytic mechanism is built from local laws, rigidity, and stochastic evolution. The local semicircle law gives high-probability control of the Stieltjes transform and resolvent entries down to microscopic scales, rigidity controls $\lambda_1\le \cdots \le \lambda_N$3, and eigenvector delocalization rules out concentration of mass on a few coordinates. Dyson Brownian motion then supplies the intrinsic mechanism behind universality: local gap distributions thermalize to the Gaussian equilibrium on time scales $\lambda_1\le \cdots \le \lambda_N$4, and Green function comparison transfers the result from Gaussian-divisible ensembles to general Wigner matrices (Erdos et al., 2011). In the Four Moment approach, resolvent control, lower-tail gap estimates, and a Lindeberg replacement scheme show that local bulk statistics depend only on finitely many moments (0906.0510). In random regular graphs, the corresponding machinery becomes constrained Dyson Brownian motion together with switching dynamics (Bauerschmidt et al., 2015).

In localized operators, the decisive inputs are different. Wegner estimates give linear control of expected eigenvalue counts in small windows, while Minami-type estimates give quadratic control on the probability of multiple eigenvalues in such windows. Localization—via exponential decay of eigenfunctions, fractional-moment bounds, or multiscale analysis—then produces spatial decoupling and approximate independence across subboxes (Brodie et al., 2020, Hislop et al., 2019). For singular Bernoulli disorder, where classical Minami’s inequality does not apply, the Anderson–Bernoulli analysis replaces single-site regularity by IDS and Furstenberg-measure regularity, together with a geometric decoupling of near-resonant quasimodes (Bourgain, 2013). Near band edges in higher-rank models, the Dietlein–Elgart method uses eigenvalue level spacing, Dirichlet–Neumann bracketing, and a Cartan-type lemma to derive a weak Minami estimate sufficient for Poisson LES (Herschenfeld et al., 2022). In random polymer models, transfer matrices, modified Prüfer variables, and bootstrap multiscale analysis identify the clock regime at critical energies and the Poisson regime away from them (Hislop et al., 28 Aug 2025).

6. Related usage: linear eigenvalue statistics

In several papers, the acronym LES denotes Linear Eigenvalue Statistics rather than local eigenvalue statistics. In that usage, the basic object is a centered sum such as

$\lambda_1\le \cdots \le \lambda_N$5

and the main results are central limit theorems rather than point-process limits. For mesoscopic Wigner-type matrices in the bulk, the paper proves a universal CLT

$\lambda_1\le \cdots \le \lambda_N$6

for $\lambda_1\le \cdots \le \lambda_N$7 in the regime $\lambda_1\le \cdots \le \lambda_N$8 (Riabov, 2023). For real-valued random centrosymmetric matrices, the centered LES converges to a Gaussian law with an explicit contour-integral variance, and for a polynomial $\lambda_1\le \cdots \le \lambda_N$9 the variance simplifies to

$E$0

(Jana et al., 30 Sep 2025). For two correlated non-Hermitian random matrices, the centered linear statistics of the combined spectrum converge jointly to a bivariate Gaussian distribution, with variance depending on the correlation structure of the entries and the fourth mixed cumulants (Jana et al., 28 Mar 2025). For Jacobi $E$1-ensembles, the moment-generating function of linear statistics shows that the mean and variance are related to the sine kernel in the bulk and to the Bessel kernel at the hard edge (Min et al., 2021).

This terminological overlap is significant: in current usage, “LES” may refer either to microscopic local point processes and spacing statistics, or to centered linear statistics and their CLTs. The two notions are mathematically distinct, even though both are organized by local spectral scaling and by the same universal kernels in the regimes where universality holds.