Non-Asymptotic Wigner Matrices

• Non-Asymptotic Wigner matrices are finite-dimensional random matrix models controlled via explicit finite-N spectral estimates and local laws.
• They use deterministic approximations and free-probability techniques to derive high-probability error bounds and quantitative convergence rates.
• These methods enable practical insights into eigenvalue rigidity, operator norm convergence, and fluctuation behavior in structured and deformed ensembles.

Non-asymptotic Wigner matrices are Wigner and Wigner-type random matrix models studied through finite-$N$ estimates, quantitative large-$N$ approximations, and spectral control that is meaningful before passage to the limit. In the cited literature, this includes explicit high-probability bounds, local laws down to the smallest spectral resolution scale, deterministic approximations for traces and matrix elements, convergence rates for operator norms, eigenvalue-exclusion statements, and singularity estimates. It also includes a persistent contrast with asymptotic-only universality theory: Tao–Vu’s bulk universality work proves convergence of local $k$-point correlation functions to the Dyson sine process and explicitly states that it “does not claim non-asymptotic universality bounds,” while the critical heavy-tail edge paper proves a new limiting law for the largest eigenvalue but is “essentially asymptotic only” and does not derive explicit finite-$n$ tail inequalities, concentration inequalities, or quantitative convergence rates (Tao et al., 2011, Diaconu, 2022).

1. Models, normalizations, and deterministic reference objects

The standard Wigner model in this literature is an $N\times N$ real symmetric or complex Hermitian matrix with independent upper-triangular entries, centered, and normalized so that off-diagonal variances are of order $N^{-1}$. Several papers extend this to generalized or Wigner-type ensembles with a deterministic variance profile

$s_{ij}:=\mathbb E|h_{ij}|^2,$

or to polynomial and finite-rank deformations involving deterministic matrices (Ajanki et al., 2015, Pizzo et al., 2011).

For general Wigner-type matrices, the semicircle law is replaced by a self-consistent density of states determined by the quadratic vector equation

$-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$

whose solution $\mathbf m(z)=(m_1(z),\dots,m_N(z))$ gives the deterministic diagonal approximation to the resolvent $G(z)=(H-z)^{-1}$ (Ajanki et al., 2015). This shift from a scalar Stieltjes transform to a vector-valued deterministic profile is one of the main structural differences between classical Wigner ensembles and general variance-profile models.

A parallel deterministic reference appears in free-probability formulations. For Hermitian polynomials in independent Wigner and deterministic matrices, the limiting support is described by evaluating the polynomial on a free semicircular system and the deterministic limit variables. In this setting, the support of the free operator’s spectral distribution acts as the deterministic spectral boundary for large finite matrices (Belinschi et al., 2016). For products of functions of self-adjoint polynomials in Wigner and deterministic matrices, deterministic approximations are formulated in the free-product model $N$0, with the free trace $N$1 and the conditional expectation onto $N$2 providing the leading finite-$N$3 terms (Parraud et al., 2022).

This suggests that non-asymptotic Wigner theory is organized around two kinds of deterministic objects: self-consistent analytic objects such as the QVE solution, and operator-valued free-probability models used to predict finite-dimensional spectra and observables.

2. Local laws, rigidity, and the boundary with asymptotic bulk universality

The modern quantitative core of the subject is the local law. For general self-adjoint $N$4 matrices with centered independent entries and arbitrary nonnegative symmetric variance profile $N$5, Ajanki, Erdős, and Krüger prove that the resolvent is close to $N$6 down to the optimal resolution scale

$N$7

and obtain bulk universality for both real symmetric and complex Hermitian symmetry classes (Ajanki et al., 2015). In the bulk, where $N$8, the entrywise local law simplifies to

$N$9

and the averaged law improves to order $k$0 (Ajanki et al., 2015).

These estimates yield large-but-finite spectral control. The same paper derives convergence of the empirical spectrum to the density of states, rigidity of individual eigenvalues relative to their deterministic quantiles, absence of eigenvalues outside $k$1 except near the correct microscopic scale, and complete delocalization

$k$2

for all eigenvectors (Ajanki et al., 2015). The deterministic density may have multiple intervals, internal gaps, square-root edges, cubic-root singularities, and cusp regimes, so the local law is not tied to a single-cut semicircle geometry.

By contrast, Tao–Vu’s bulk universality paper is explicitly asymptotic. It studies the local $k$3-point correlation functions

$k$4

obtained by rescaling around a bulk energy $k$5 at the semicircle spacing scale $k$6, and proves convergence to the determinantal Dyson sine process under weak distributional assumptions. Under a sufficiently high finite-moment assumption, the convergence is vague; under an additional regularity hypothesis on the atom density, it upgrades to local $k$7 convergence (Tao et al., 2011). The paper is careful about the distinction between vague convergence, local $k$8, weak convergence, averaged vague convergence, and local uniform convergence, and it emphasizes that local $k$9 is stronger than vague convergence and only makes sense for continuous ensembles (Tao et al., 2011).

The contrast is conceptually important. Local laws and rigidity supply finite-$n$0 control and stability estimates; bulk universality in the sense of $n$1-point correlations supplies limiting local statistics but, in the cited asymptotic formulation, not non-asymptotic error bounds.

3. Spectral norm, regular edges, and outliers

A central non-asymptotic problem is control of the spectral edge. For Wigner-type ensembles with variance matrix $n$2, the norm is asymptotically given by the maximum of the support of the self-consistent density of states. Alt, Erdős, and Krüger improve the earlier bound $n$3 by proving

$n$4

where $n$5 is the smallest positive root of

$n$6

This yields a high-probability finite-$n$7 bound on $n$8 for sequences satisfying the stated moment assumptions (Erdős et al., 2018). The proof uses a tree/Dyck-path expansion of the Dyson equation, a chopping procedure converting long branches into factors $n$9, and an effective Markov chain approximation for weighted Dyck paths (Erdős et al., 2018).

For Hermitian quadratic polynomials in several independent Wigner matrices,

$N\times N$0

Ajanki, Cipolloni, and Erdős prove that, except for specific reducible cases, the limiting spectral density has regular square-root edges and the operator norm converges to the deterministic edge at the optimal rate $N\times N$1 (Fronk et al., 2023). The paper explicitly states that the exponent is optimal, and relates this to the $N\times N$2 edge spacing implied by square-root behavior.

Finite-rank deformations give a second major edge regime. For

$N\times N$3

with $N\times N$4 a deterministic Hermitian finite-rank perturbation, Capitaine, Knowles, and others show that spikes satisfying $N\times N$5 produce outliers converging to

$N\times N$6

under the finite-moment assumptions of uniformly bounded fifth moments off-diagonal and uniformly bounded third moments on the diagonal (Pizzo et al., 2011). The corresponding eigenvalues fluctuate on the $N\times N$7-scale, with non-universal limits in the localized-eigenvector case and universal GOE/GUE limits in the delocalized-eigenvector case (Pizzo et al., 2011).

A more recent generalized Wigner result studies

$N\times N$8

where the variance profile is $N\times N$9 and the spike scale satisfies

$N^{-1}$0

In this regime the top $N^{-1}$1 eigenvalues become outliers with first-order behavior

$N^{-1}$2

their centered fluctuations are jointly Gaussian, the eigenvectors align with the perturbation directions, and the full eigenvector process converges in $N^{-1}$3 for every $N^{-1}$4 (Bhattacharya et al., 15 Jan 2026).

By contrast, the borderline heavy-tail edge paper identifies the asymptotic transition law

$N^{-1}$5

but explicitly “does not contain” finite-$N^{-1}$6 tail inequalities, non-asymptotic operator norm bounds with constants, or quantitative convergence rates (Diaconu, 2022).

4. Finite-$N^{-1}$7 observables: functions, traces, and quantitative freeness

Non-asymptotic control is not limited to eigenvalues. For an $N^{-1}$8 Wigner matrix $N^{-1}$9, the matrix elements of $s_{ij}:=\mathbb E|h_{ij}|^2,$0 fluctuate on scale $s_{ij}:=\mathbb E|h_{ij}|^2,$1 for any $s_{ij}:=\mathbb E|h_{ij}|^2,$2. Khorunzhy and collaborators show that $s_{ij}:=\mathbb E|h_{ij}|^2,$3 and $s_{ij}:=\mathbb E|h_{ij}|^2,$4 have explicit Gaussian limits whose variances depend on $s_{ij}:=\mathbb E|h_{ij}|^2,$5, $s_{ij}:=\mathbb E|h_{ij}|^2,$6, and the distribution of the corresponding diagonal entry, with error $s_{ij}:=\mathbb E|h_{ij}|^2,$7 if $s_{ij}:=\mathbb E|h_{ij}|^2,$8 and $s_{ij}:=\mathbb E|h_{ij}|^2,$9 otherwise (Erdős et al., 2016). The proof replaces Helffer–Sjöstrand by Pleijel’s inversion formula and combines it with local semicircle-law input, thereby relaxing earlier regularity assumptions to bounded variation (Erdős et al., 2016).

Quantitative free-probability approximations at finite $-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$0 are developed for products

$-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$1

where $-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$2 is a tuple of independent Wigner matrices and $-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$3 a tuple of deterministic matrices of uniformly bounded operator norm. The trace observable is approximated by its free-product analogue with high-probability error

$-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$4

while vector or matrix-entry observables admit the isotropic approximation with error

$-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$5

For GOE/GUE, the function norm improves from $-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$6 to $-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$7 in the Fourier-space argument (Parraud et al., 2022). These estimates are explicitly finite-$-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$8, high-probability statements rather than pure limit theorems.

At the level of global fluctuations, centered unnormalized traces

$-\frac{1}{m_i(z)} = z + \sum_j s_{ij} m_j(z),$9

satisfy a joint Gaussian limit under mild assumptions, but the limiting covariance is not determined by the limiting $\mathbf m(z)=(m_1(z),\dots,m_N(z))$0-distribution of the deterministic matrices alone. It also depends on Hadamard-product limits $\mathbf m(z)=(m_1(z),\dots,m_N(z))$1, transpose limits $\mathbf m(z)=(m_1(z),\dots,m_N(z))$2 in the general pseudo-variance case, and the Wigner parameters $\mathbf m(z)=(m_1(z),\dots,m_N(z))$3 (Male et al., 2020). The paper emphasizes that general Wigner matrices together with deterministic matrices are not asymptotically free of second order in general (Male et al., 2020).

5. Spectral exclusion, norm convergence, and non-singularity

A distinct non-asymptotic theme is eventual spectral exclusion. For a self-adjoint polynomial $\mathbf m(z)=(m_1(z),\dots,m_N(z))$4 in independent Wigner matrices and deterministic matrices with uniformly bounded operator norm, Belinschi, Capitaine, and Février prove that if an interval $\mathbf m(z)=(m_1(z),\dots,m_N(z))$5 stays a positive distance away from the support of the limiting free-probability spectral measure, then almost surely, for all large $\mathbf m(z)=(m_1(z),\dots,m_N(z))$6, the finite-dimensional polynomial matrix has no eigenvalues in $\mathbf m(z)=(m_1(z),\dots,m_N(z))$7 (Belinschi et al., 2016). This is a strong “no outliers away from support” statement, and the same work establishes strong asymptotic freeness, meaning both normalized traces and operator norms of polynomial expressions converge to their free limits (Belinschi et al., 2016).

The finite-$\mathbf m(z)=(m_1(z),\dots,m_N(z))$8 quantitative version of this perspective appears in later work on asymptotic freeness through Heisenberg evolution. There the approximation is controlled directly at finite $\mathbf m(z)=(m_1(z),\dots,m_N(z))$9, and large time gaps generate asymptotic freeness for deterministic observables conjugated by $G(z)=(H-z)^{-1}$0 (Parraud et al., 2022). A plausible implication is that non-asymptotic freeness results supply an operational bridge between local law technology and free-probability support predictions.

Non-singularity provides another canonical finite-$G(z)=(H-z)^{-1}$1 problem. For a Wigner matrix $G(z)=(H-z)^{-1}$2 with independent upper-triangular entries, possibly non-identically distributed and allowed to depend on $G(z)=(H-z)^{-1}$3, define the largest jump parameter

$G(z)=(H-z)^{-1}$4

Under the sole assumption $G(z)=(H-z)^{-1}$5, and with no moment assumptions, the singularity probability satisfies

$G(z)=(H-z)^{-1}$6

for every $G(z)=(H-z)^{-1}$7 (Manrique et al., 2014). The proof uses the Lévy concentration function, the Kolmogorov–Rogozin inequality, Kesten’s refinement, linear and quadratic concentration bounds, and a rank-growth scheme adapted from Costello–Tao–Vu (Manrique et al., 2014). The decisive feature is universality with respect to the maximal atom size rather than moments or support geometry.

6. Structured and pseudo-random variants

The non-asymptotic viewpoint also extends beyond genuinely random i.i.d. models. Pseudo-Wigner ensembles are symmetric sign matrices whose upper-triangular entries are only $G(z)=(H-z)^{-1}$8-independent rather than fully independent. For appropriate $G(z)=(H-z)^{-1}$9-independent ensembles, the empirical spectral distribution satisfies the Kolmogorov-distance bound

$N$00

with explicit polynomially small failure probabilities in $N$01 (Soloveychik et al., 2017). The same paper gives an explicit construction using dual BCH codes and proves

$N$02

showing that semicircular spectral behavior can be realized with description length on the order of $N$03 bits for fixed precision (Soloveychik et al., 2017).

Other structured variants are more asymptotic in formulation but clarify the combinatorial scope of Wigner methods. For symmetric matrices with independent entries whose moments vary with $N$04, position, sparsity, or tail behavior, the limiting spectral distribution can be governed by special symmetric partitions rather than only non-crossing pair-partitions (Bose et al., 2021). In the standard i.i.d. finite-variance case, this collapses back to the Catalan/non-crossing picture and recovers the semicircle law (Bose et al., 2021).

Wigner-type ensembles with built-in symmetries from Cartan classes DIII and CI also admit explicit fluctuation theory. Using the Schenker–Schulz-Baldes framework plus a calculus of patterns, one obtains a central limit theorem for centered traces of rescaled Chebyshev polynomials, with diagonal asymptotic covariance and limiting variances

$N$05

while $N$06 odd modes have vanishing asymptotic variance (Stolz, 2017). This shows that additional deterministic dependencies can preserve Wigner-type Gaussian fluctuation behavior while altering the underlying combinatorics.

Taken together, these developments define non-asymptotic Wigner matrix theory as a quantitative theory of finite-dimensional spectra and observables: local laws replace pure convergence, self-consistent analytic equations replace heuristic edges, free-probability supports become finite-$N$07 exclusion zones, and even structured or pseudo-random ensembles can be analyzed through explicit spectral error bounds.