Random matrix theory
Random matrix theory is the mathematical study of matrices whose entries, eigenvalues, or invariant subspaces are governed by probability distributions. Rather than treating a matrix as a fixed array of numbers, the theory treats it as a random variable taking values in a space of matrices. Its principal objects include the resulting eigenvalue distributions, the statistical behavior of eigenvectors, and the limiting spectral measures obtained as matrix dimensions increase.
The subject originated in multivariate statistics and later developed into a framework for describing complicated spectra in nuclear physics, number theory, and statistical mechanics. Many of its characteristic results are universal: after an appropriate rescaling, local spectral statistics often depend on symmetry and dimensional regime rather than on the detailed probability law of each matrix entry.
Mathematical formulation
A random matrix is a measurable map
[ H:\Omega\longrightarrow \mathbb F^{N\times N}, ]
where ((\Omega,\mathcal F,\mathbb P)) is a probability space and (\mathbb F) is generally the real or complex number field. Additional conditions may require (H) to be symmetric, Hermitian, positive semidefinite, or invariant under a specified group action.
For a Hermitian random matrix, the eigenvalues are real and may be ordered as
[ \lambda_1\leq \lambda_2\leq\cdots\leq\lambda_N. ]
They then define an empirical spectral measure,
[ \mu_H=\frac{1}{N}\sum_{j=1}^{N}\delta_{\lambda_j}, ]
which assigns equal mass to every eigenvalue. A basic global question concerns the convergence of (\mu_H) as (N) tends to infinity. Local questions instead examine eigenvalue spacings on scales comparable with the mean separation between neighboring levels.
The distribution of a random matrix may be specified through its entries. It may alternatively be imposed directly on matrix space through a density of the form
[ d\mathbb P(H)=Z_N^{-1}\exp!\left[-N\operatorname{Tr}V(H)\right]dH, ]
where (V) is a confining potential and (Z_N) is the normalizing partition function. When the measure is invariant under orthogonal or unitary conjugation, the matrix eigenvectors decouple from the joint eigenvalue density. The change of variables from entries to eigenvalues introduces the Vandermonde determinant,
[ \Delta(\lambda)=\prod_{i<j}(\lambda_j-\lambda_i), ]
whose absolute value appears with a symmetry-dependent exponent.
Classical ensembles
The Gaussian orthogonal ensemble consists of real symmetric matrices with a probability law invariant under orthogonal conjugation. Its eigenvalue density contains (|\Delta(\lambda)|), corresponding to the Dyson index (\beta=1).
The Gaussian unitary ensemble consists of complex Hermitian matrices whose law is invariant under unitary conjugation. Its density contains (|\Delta(\lambda)|^2), so its Dyson index is (\beta=2).
The Gaussian symplectic ensemble is formulated using self-dual quaternionic matrices, or an equivalent complex representation. Its eigenvalue density contains (|\Delta(\lambda)|^4), corresponding to (\beta=4).
These ensembles are unified by the Gaussian (\beta)-ensemble density
[ p(\lambda_1,\ldots,\lambda_N)
\frac{1}{Z_{N,\beta}} \exp!\left( -\frac{\beta}{4}\sum_{j=1}^{N}\lambda_j^2 \right) \prod_{i<j}|\lambda_i-\lambda_j|^\beta. ]
The Vandermonde factor suppresses coincident eigenvalues and produces level repulsion. For small normalized spacing (s), the nearest-neighbor spacing density behaves proportionally to (s^\beta). This behavior distinguishes random-matrix spectra from a Poisson point process, in which nearby points do not repel one another.
Another major family is formed by Wishart matrices. If (X) is a rectangular random data matrix, then (X^\ast X) is positive semidefinite and models a sample covariance matrix. Its large-dimensional spectral density is described by the Marchenko–Pastur distribution.
Historical development
James Wishart introduced the distribution now bearing his name in 1928 while studying sample covariance matrices. This statistical line of development established random matrices before the subject acquired its later connection with quantum spectra.
During the 1950s, Eugene Wigner used large random real symmetric matrices to model the resonance levels of complicated atomic nuclei. The model discarded detailed nuclear interactions while retaining the spectral correlations implied by time-reversal symmetry. Wigner also identified the limiting global density now called the Wigner semicircle distribution.
Freeman Dyson organized the Gaussian ensembles according to their symmetry under time reversal and spin transformations. His 1962 classification associated the three classical invariant ensembles with the indices (\beta=1), (\beta=2), and (\beta=4). Dyson also introduced a matrix-valued Brownian motion whose eigenvalues evolve as mutually repelling stochastic particles.
Exact correlation functions developed concurrently through the work of Madan Lal Mehta and Michel Gaudin. Their analysis connected invariant matrix measures with integral kernels constructed from orthogonal polynomials, allowing multipoint eigenvalue statistics to be expressed by determinants in the unitary case.
In 1964, You Watanabe derived a finite-dimensional Pfaffian representation for the two-level correlation function of the Gaussian orthogonal ensemble. The calculation translated the quaternion-determinant formulation into a kernel built from skew-orthogonal polynomial pairs and made the finite-(N) correction at the spectral boundary explicit. This representation became part of the standard treatment of orthogonal-ensemble correlation functions during the subsequent development of Pfaffian point processes.
Later work by Oleksandr Marchenko and Leonid Pastur established the limiting law for high-dimensional sample covariance spectra. Craig Tracy and Harold Widom subsequently identified the limiting distributions of extreme eigenvalues at the soft spectral edge. These distributions are related to solutions of the Painlevé equations.
Global spectral laws
For a Wigner matrix with centered independent entries of appropriately scaled variance, the empirical spectral measure converges to the semicircle law. In one common normalization, its density is
[ \rho_{\mathrm{sc}}(x)
\frac{1}{2\pi}\sqrt{4-x^2}, \mathbf 1_{{|x|\leq 2}}. ]
The law describes eigenvalues on the scale of the entire spectrum. It remains valid for broad classes of entry distributions, provided that the normalization and moment assumptions prevent a small number of entries from dominating the matrix.
For a sample covariance matrix (X^\ast X), with the matrix aspect ratio approaching a positive constant (c), the limiting density is the Marchenko–Pastur law. Its continuous component is supported on
[ \lambda_{\pm}=(1\pm\sqrt c)^2, ]
and has density
[ \rho_{\mathrm{MP}}(x)
\frac{\sqrt{(\lambda_+-x)(x-\lambda_-)}}{2\pi c x} \mathbf 1_{[\lambda_-,\lambda_+]}(x), ]
together with an atom at zero when required by the limiting rank deficiency.
Both laws concern macroscopic spectral behavior. They do not by themselves determine the fluctuations of individual eigenvalues or the statistics of spacings after local rescaling.
Local correlations and universality
In the interior of a unitary-ensemble spectrum, the rescaled correlation kernel converges to the sine kernel,
[ K_{\mathrm{sine}}(x,y)
\frac{\sin\pi(x-y)}{\pi(x-y)}. ]
The associated determinantal point process describes bulk correlations after the local mean spacing has been normalized to one. Orthogonal and symplectic ensembles possess related matrix-valued kernels and Pfaffian correlation functions.
Near a regular upper spectral boundary, the bulk scaling ceases to apply. The rescaled kernel instead approaches the Airy kernel, and the largest eigenvalue converges to a symmetry-dependent Tracy–Widom distribution. At a hard boundary such as the origin of certain covariance ensembles, the corresponding limit is governed by a Bessel kernel.
Universality states that these limiting forms persist beyond the Gaussian ensembles. Entry distributions can differ substantially while producing the same local limits, provided that the matrices remain in the same symmetry class and satisfy the relevant regularity conditions. The symmetry index therefore controls many local statistics more directly than the precise density of the entries.
Coulomb-gas interpretation
The joint eigenvalue density can be rewritten as a Gibbs distribution for particles on a line. Taking its logarithm gives an energy of the form
[ \mathcal H(\lambda_1,\ldots,\lambda_N)
N\sum_{j=1}^{N}V(\lambda_j)
\sum_{i<j}\log|\lambda_i-\lambda_j|. ]
The potential (V) confines the particles, whereas the logarithmic interaction separates them. The parameter (\beta) acts as an inverse temperature. This interpretation connects random matrix theory with logarithmic potential theory, large-deviation principles, and equilibrium measures.
The equilibrium measure minimizes a continuum energy functional obtained from (\mathcal H). Its density determines the leading global eigenvalue distribution, while fluctuations around equilibrium generate finer spectral statistics. The Coulomb-gas formulation also explains why the Vandermonde determinant produces repulsion without requiring a direct interaction among the original matrix entries.
Spectral applications
In quantum systems whose classical dynamics is chaotic, unfolded energy levels frequently exhibit the local correlations of an invariant random-matrix ensemble with matching symmetries. Integrable systems instead commonly produce Poissonian spacing statistics. This distinction forms the basis of the Bohigas–Giannoni–Schmit conjecture.
Random matrix statistics also occur in the zeros of the Riemann zeta function. High zeros on the critical line have pair correlations matching those of the Gaussian unitary ensemble after normalization by the local zero density. The correspondence concerns statistical limits and does not identify zeta zeros as eigenvalues of a known finite random matrix.
In high-dimensional statistics, sample covariance matrices provide null models for principal-component spectra. Finite-rank perturbations can create eigenvalue outliers once the perturbation passes a critical threshold, a transition described by the Baik–Ben Arous–Péché transition. Below that threshold, the leading sample eigenvalue remains attached to the edge of the continuous spectrum.