Wigner’s Semicircle Law

• Wigner’s semicircle law is a universal limit theorem defining the eigenvalue distribution of suitably normalized random matrices with a semicircular density supported on [-2,2].
• The law is derived using both the moment method, which links eigenvalue moments to Catalan numbers, and the resolvent method, establishing precise local spectral control.
• Modern extensions apply the law to matrices with non-identical variances, sparse graphs, and dependent entries, unifying diverse models in random matrix theory and free probability.

Wigner’s semicircle law is the statement that the empirical spectral distribution of a suitably normalized Hermitian or real symmetric random matrix converges to a deterministic probability measure supported on $[-2,2]$ with density

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$

In its classical form, the law governs Wigner matrices with independent upper-triangular entries of mean zero and variance $1/n$; in modern formulations it extends to non-identically distributed variance profiles, local spectral statistics, sparse graphs, dependent-entry ensembles, algebraic constructions from linear codes, and free-probability analogues (Chin, 2019, Benaych-Georges et al., 2016, Chin, 2021).

1. Classical formulation and canonical objects

Let $H=(H_{ij})_{1\le i,j\le N}$ be an $N\times N$ Hermitian random matrix whose upper-triangular entries are independent, centered, and satisfy $\mathrm{Var}(H_{ij})=1/N$ for $i\neq j$, with the diagonal typically of order $O(1/N)$. If $\lambda_1\le\cdots\le \lambda_N$ are its eigenvalues, the empirical spectral distribution is

$\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$

Its Stieltjes transform is

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$0

The global semicircle law states that $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$1 in probability for each fixed $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$2, equivalently $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$3 converges weakly to the semicircle distribution (Benaych-Georges et al., 2016).

The deterministic Stieltjes transform $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$4 is characterized by

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$5

with the branch chosen so that $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$6 for $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$7. The support of the limiting measure is $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$8, and its even moments are Catalan numbers while odd moments vanish: $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$9 where $1/n$0 (Chin, 2019, Benaych-Georges et al., 2016).

A strengthened global formulation no longer requires identical distributions or exact variance $1/n$1. For Hermitian matrices $1/n$2 with independent upper-triangular entries, mean zero, row-variance control $1/n$3, and a Lindeberg condition, the semicircle law still holds when

$1/n$4

as $1/n$5 (Chin, 2019).

2. Moment method, resolvents, and the mechanism of the law

The classical proof expands

$1/n$6

and interprets each summand as a closed walk. Because the entries are independent and centered, any edge traversed exactly once contributes zero. The leading contribution comes from “double-tree” walks: each edge is traversed twice, and the canonical representatives are in bijection with Dyck paths. Their number is the Catalan number, which reproduces the moments of $1/n$7 (Chin, 2019).

The resolvent method encodes the same limit through a self-consistent equation. Writing

$1/n$8

Schur-complement identities lead to

$1/n$9

and passing to the limit yields the quadratic equation for $H=(H_{ij})_{1\le i,j\le N}$0. Stieltjes inversion then recovers the density $H=(H_{ij})_{1\le i,j\le N}$1 (Chin, 2019).

The two methods emphasize different structural features. The moment method exposes Catalan combinatorics and survives in many dependent or algebraic settings through walk-counting. The resolvent method is the basis of local semicircle laws and of quantitative estimates for eigenvalue rigidity and eigenvector delocalization (Benaych-Georges et al., 2016, Götze et al., 2015).

An orthogonal-polynomial manifestation makes the same mechanism explicit. For the $H=(H_{ij})_{1\le i,j\le N}$2-th monic Hermite polynomial $H=(H_{ij})_{1\le i,j\le N}$3, if $H=(H_{ij})_{1\le i,j\le N}$4 are its zeros and $H=(H_{ij})_{1\le i,j\le N}$5, then the empirical measure

$H=(H_{ij})_{1\le i,j\le N}$6

converges weakly to the semicircle distribution on $H=(H_{ij})_{1\le i,j\le N}$7 with density $H=(H_{ij})_{1\le i,j\le N}$8, equivalently to the standard $H=(H_{ij})_{1\le i,j\le N}$9 semicircle after rescaling by $N\times N$0. Kornyik and Michaletzky gave both a generating-function proof via the fixed-point equation $N\times N$1 and a moment-polynomial proof in which the leading coefficient is Catalan (Kornyik et al., 2015).

3. Local semicircle laws and fine spectral control

The global law controls spectral mass on macroscopic $N\times N$2 scales. A local semicircle law controls the resolvent down to spectral windows containing slightly more than one eigenvalue. For Wigner matrices, with

$N\times N$3

the local law on the spectral domain

$N\times N$4

takes the form

$N\times N$5

and, entrywise,

$N\times N$6

where

$N\times N$7

In the bulk $N\times N$8, so $N\times N$9; near the edge $\mathrm{Var}(H_{ij})=1/N$0, the same framework gives optimal control down to $\mathrm{Var}(H_{ij})=1/N$1 (Benaych-Georges et al., 2016).

Under only a finite $\mathrm{Var}(H_{ij})=1/N$2-th moment assumption,

$\mathrm{Var}(H_{ij})=1/N$3

one still obtains quantitative convergence. Götze, Naumov, and Tikhomirov showed that the typical Kolmogorov distance between the empirical spectral distribution $\mathrm{Var}(H_{ij})=1/N$4 of $\mathrm{Var}(H_{ij})=1/N$5 and $\mathrm{Var}(H_{ij})=1/N$6 is of order $\mathrm{Var}(H_{ij})=1/N$7 up to logarithmic correction, more precisely

$\mathrm{Var}(H_{ij})=1/N$8

with $\mathrm{Var}(H_{ij})=1/N$9 and $i\neq j$0. They also proved a short-scale law,

$i\neq j$1

with overwhelming probability for $i\neq j$2, along with eigenvalue rigidity and eigenvector delocalization (Götze et al., 2015).

A sharper local result under the same finite $i\neq j$3-moment regime gives

$i\neq j$4

for $i\neq j$5, leading to

$i\neq j$6

with overwhelming probability, $i\neq j$7 for the Kolmogorov distance, the optimal expectation bound $i\neq j$8, rigidity at scale

$i\neq j$9

and delocalization bounds of GOE/GUE type (Götze et al., 2016).

The same local philosophy extends beyond classical Wigner matrices. For the Hermite $O(1/N)$0 ensemble, Bao and Su proved that if $O(1/N)$1 and $O(1/N)$2 with $O(1/N)$3, then the average number of states in

$O(1/N)$4

converges in probability to $O(1/N)$5, and the index $O(1/N)$6 satisfies

$O(1/N)$7

(Bao et al., 2011).

4. General conditions, dependence, and heavy tails

A precise finite-variance characterization is available. For Hermitian matrices $O(1/N)$8 with independent upper-triangular entries, mean zero, variances $O(1/N)$9, Lindeberg condition

$\lambda_1\le\cdots\le \lambda_N$0

and margin-smallness

$\lambda_1\le\cdots\le \lambda_N$1

the empirical spectral distribution converges weakly almost surely to $\lambda_1\le\cdots\le \lambda_N$2 if and only if

$\lambda_1\le\cdots\le \lambda_N$3

This identifies semicircle convergence with the condition that most rows have total variance near $\lambda_1\le\cdots\le \lambda_N$4 (Chin, 2021).

The finite-variance hypothesis is not definitive. Zhou showed that if the entries are in the normal-domain-of-attraction and the truncated second moment

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

is slowly varying, then Wigner’s semicircle law for the raw empirical spectral distribution $\lambda_1\le\cdots\le \lambda_N$6 still holds almost surely, even when $\lambda_1\le\cdots\le \lambda_N$7 (Zhou, 2011). The necessary-and-sufficient framework of (Chin, 2021) likewise extends to infinite second moments through truncated second moments $\lambda_1\le\cdots\le \lambda_N$8 and weak Lindeberg-type assumptions.

Dependence does not automatically destroy the law, but it changes the admissible hypotheses. Friesen and Löwe introduced an “approximately uncorrelated” criterion for correlated symmetric ensembles $\lambda_1\le\cdots\le \lambda_N$9. Under this criterion the empirical measure converges weakly in probability to the semicircle law. For Curie–Weiss ensembles, the criterion holds for $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$0, while at $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$1 there is a transition in the top eigenvalue: below criticality $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$2 in probability, whereas above criticality the top eigenvalue becomes macroscopic (Hochstättler et al., 2014).

Other dependence patterns are also compatible with semicircle behavior. Alt proved a local semicircle law for matrices with the fourfold symmetry

$\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$3

including the Fourier transform of a GOE and the flip matrix model, with optimal-scale control down to $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$4 up to logarithmic corrections (Alt, 2015). Krajewski, Tanasa, and Vu established a dependent-entry semicircle law through cumulant scaling bounds and a renormalisation-group argument based on replicas: if the non-Gaussian cumulants satisfy the prescribed $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$5-power bounds, then only the Gaussian two-point contractions survive in the large-$\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$6 limit, yielding the semicircle law (Krajewski et al., 2016, Krajewski, 2017).

Sparse graph models occupy an intermediate regime between independent-entry Wigner matrices and strongly structured combinatorial objects. For centered, rescaled Erdős–Rényi adjacency matrices, Erdős, Knowles, Yau, and Yin proved that as long as $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$7 with at least logarithmic speed, the density of eigenvalues is given by the semicircle law for spectral windows of length larger than $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$8 up to logarithmic corrections, and the eigenvectors are completely delocalized in the sense that their $\mu_N=\frac1N\sum_{k=1}^N\delta_{\lambda_k}.$9-norms are at most of order $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$00 with very high probability (Erdős et al., 2011).

5. Orthogonal polynomials, $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$01-ensembles, and free probability

The semicircle law is not confined to matrix entries. For monic Hermite polynomials $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$02, the appropriately normalized roots converge weakly to the same distribution. The generating function of the moments of the roots satisfies a fixed-point equation whose unique power-series solution is the Catalan generating function $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$03, and the direct moment expansion shows that for even $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$04,

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$05

with $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$06. Kornyik and Michaletzky also recalled the Forrester–Gamburd identity

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$07

for a Wigner matrix $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$08, which makes the relation between Hermite zeros and expected characteristic polynomials exact (Kornyik et al., 2015).

For the Hermite $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$09 ensemble, the Dumitriu–Edelman tridiagonal model gives a concrete matrix realization whose eigenvalues have the log-gas density

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$10

Its normalized empirical measure

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$11

converges to the semicircle law for every $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$12. Bao and Su extended this to a local law in the bulk and obtained a Gaussian fluctuation result for the index with variance $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$13 (Bao et al., 2011).

A free-probability analogue replaces matrices by free self-adjoint noncommutative random variables. For self-normalized sums

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$14

Neufeld proved weak convergence $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$15, where $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$16 is the standard semicircle distribution. In the bounded identically distributed case this gives

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$17

while in the unbounded i.i.d. case with finite fourth moment it yields

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$18

for the Kolmogorov distance (Neufeld, 2024).

6. Structured ensembles, estimators, and contemporary variants

Linear-code ensembles provide a non-classical source of semicircle behavior. If $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$19 is an $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$20 linear code with dual distance $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$21, $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$22 is formed by choosing $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$23 distinct codewords uniformly at random and mapping them via the canonical additive character, and

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$24

then under $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$25 and $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$26 the empirical spectral distribution of $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$27 converges in probability to the standard semicircle law (Chan et al., 2018). Chan and Xiong strengthened this by proving that the dual distance condition $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$28 is sufficient for convergence via the Stieltjes-transform method and that

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$29

uniformly over intervals $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$30, for some $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$31; if $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$32 with $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$33, then

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$34

The same work states that codes of dual distance $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$35 fail to converge to the semicircle law in general, so the condition $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$36 is not merely technical (Chan et al., 2019).

Weighted-network models yield a shifted semicircle law for normalized Laplacians. For random weighted graphs with normalized Laplacian

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$37

the empirical distribution of the nontrivial eigenvalues converges to a density centered at $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$38,

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$39

under the stated degree and link-weight conditions. Equivalently, the shifted eigenvalues $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$40 follow a semicircle law of radius $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$41 (Sakumoto et al., 2020).

The semicircle law can also be estimated nonparametrically. Zhou introduced the kernel density estimator

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$42

and the associated smoothed distribution $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$43. If $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$44, $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$45, and $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$46, then

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$47

Moreover, if $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$48, $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$49, $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$50, and $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$51, $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$52, then

$$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$53

For the Cauchy kernel $$\rho_{sc}(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,\mathbf 1_{[-2,2]}(x).$$54, the estimator is the imaginary-part-of-Stieltjes-transform estimate (Zhou, 2011).

Taken together, these formulations show that Wigner’s semicircle law is both a specific limit theorem for normalized random matrices and a unifying principle across combinatorial, algebraic, analytic, and noncommutative settings. The global law identifies the macroscopic density, the local law resolves fluctuations down to near-microscopic scales, and modern extensions specify with increasing precision which variance structures, dependence patterns, and normalization schemes still lead to the same universal semicircle (Benaych-Georges et al., 2016, Chin, 2021).