Gaussian Ensembles GOE/GUE/GSE and the Joint Eigenvalue Density

Intuition Beginner

Imagine scattering beads at random along a wire. If they were truly independent, two beads could land on top of each other or huddle into a tight clump; gaps and clusters would appear all over. The eigenvalues of a random matrix behave very differently. They act like beads that quietly push each other apart, the way like-charged particles on a wire repel: you almost never find two of them right next to each other, and the whole arrangement looks suspiciously even. This built-in spacing is the single most striking feature of random-matrix spectra, and it is the reason these models capture the energy levels of complicated physical systems so well.

The Gaussian ensembles are the cleanest matrices where this happens. You fill a square grid with independent bell-curve random numbers and make it symmetric (or its complex cousin). Because the recipe does not prefer any direction in space, it is built with a perfect symmetry, and that symmetry is exactly what forces the eigenvalues to repel. There turn out to be three flavours of this construction, labelled by a number called beta that takes the values one, two, and four, measuring how strongly the eigenvalues push apart.

The payoff is a single tidy formula for how likely any particular arrangement of eigenvalues is. It multiplies together two competing effects: a repulsion term that grows when eigenvalues spread apart, and a confining term that grows when they wander too far from the center. The balance between pushing apart and being reeled back in is what fixes the shape of the spectrum.

Visual Beginner

Picture two rows of dots above a number line. The top row is a batch of genuinely independent random points: they bunch up in places and leave ragged gaps elsewhere, and sometimes two land almost on top of each other. The bottom row is the eigenvalues of a Gaussian random matrix of the same size: the dots are spread out smoothly, kept apart by an invisible spring between every pair, with hardly any near-collisions.

The springs between neighbours are the repulsion; the bell-shaped envelope drawn over both rows is the confinement. The lower row is what the tidy formula describes: an arrangement is likely when the dots are both well spread (springs relaxed) and not too far out (inside the envelope). The number beta sets how stiff the springs are.

Worked example Beginner

We compute the repulsion term for the smallest interesting case — two eigenvalues — and see how it forbids collisions.

Step 1. With two eigenvalues sitting at positions $a$ and $b$, the repulsion term is the distance between them raised to the power beta. For the first flavour, beta is one, so the repulsion term is just the gap, the size of $a-b$.

Step 2. Suppose the two eigenvalues are far apart, say at $-1$ and $+1$. The gap is $2$, so the repulsion term is $2$. Now suppose they are close, at $-0.1$ and $+0.1$. The gap is $0.2$, so the repulsion term is only $0.2$. A configuration whose repulsion term is ten times smaller is correspondingly ten times less likely, all else being equal.

Step 3. Push them to the same point, $a=b$. The gap is $0$, so the repulsion term is $0$. A configuration with repulsion term zero has probability zero: two eigenvalues exactly coinciding simply does not happen. This is the precise sense in which the eigenvalues refuse to collide.

Step 4. Compare the three flavours at the close pair $-0.1$ and $+0.1$, gap $0.2$. For beta one the term is $0.2$; for beta two it is $0.2\times 0.2=0.04$; for beta four it is $0.2$ multiplied by itself four times, about $0.0016$. The bigger beta is, the harder a small gap is punished, so the stronger the repulsion.

Step 5. What this tells us: the repulsion term is a product of the gaps between eigenvalues, and raising each gap to the power beta is what turns "they tend not to be close" into a precise, tunable rule. Setting beta to one, two, or four gives the three Gaussian ensembles, with stiffer springs as beta grows.

Check your understanding Beginner

Formal definition Intermediate+

Let $\beta \in \{1,2,4\}$ index the three division algebras $\mathbb{R},\mathbb{C},\mathbb{H}$. The Gaussian ensemble of parameter $\beta$ is a probability measure on the real vector space ${\mathcal{H}}_n^{(\beta )}$ of self-adjoint $n\times n$ matrices over the corresponding algebra: real symmetric matrices for $\beta =1$ (the Gaussian Orthogonal Ensemble, GOE), complex Hermitian for $\beta =2$ (the Gaussian Unitary Ensemble, GUE), and quaternionic self-dual for $\beta =4$ (the Gaussian Symplectic Ensemble, GSE). The measure has density proportional to $$ P(M), dM ;\propto; \exp!\Big(-\tfrac{\beta n}{4},\operatorname{tr} M^2\Big), dM , $$ where $dM$ is Lebesgue measure on the independent real coordinates of $M$. Equivalently the entries are independent centred Gaussians: the diagonal entries are real with one variance, and each independent off-diagonal entry has real and imaginary (and quaternionic) parts that are i.i.d. Gaussian, the variances chosen so that the density takes the displayed trace form. (Many references drop the factor $n$ and rescale afterward; here we keep the $\frac{1}{\sqrt{n}}$ normalisation of 37.08.01 so the spectrum fills $[-2,2]$.)

The defining structural feature is invariance: for any $U$ in the orthogonal group $O(n)$ ($\beta =1$), unitary group $U(n)$ ($\beta =2$), or symplectic group $\mathrm{Sp}(n)$ ($\beta =4$), the map $M\mapsto UM{U}^{\ast }$ preserves both $\mathrm{tr}{M}^2$ and Lebesgue measure $dM$, hence preserves $P$. The Gaussian ensembles are the only ensembles that are simultaneously invariant under the relevant group and have independent entries; this rigidity is the algebraic root of everything below.

By the spectral theorem every $M\in {\mathcal{H}}_n^{(\beta )}$ factors as $M=U\Lambda U^{\ast }$ with $\Lambda =\mathrm{diag}({\lambda }_1,\dots ,{\lambda }_n)$ real and $U$ in the relevant group; the eigenvalues are well defined and, generically, distinct. The joint eigenvalue density is the pushforward of $P(M)dM$ under $M\mapsto ({\lambda }_1,\dots ,{\lambda }_n)$. Writing the Vandermonde determinant $$ \Delta(\lambda) = \prod_{1 \le i < j \le n} (\lambda_j - \lambda_i) = \det\big(\lambda_i^{,j-1}\big){i,j=1}^n , $$ the joint density on the ordered or unordered eigenvalues (up to the normalising constant $Z{n,\beta}$) is $$ p_{n,\beta}(\lambda_1, \dots, \lambda_n) = \frac{1}{Z_{n,\beta}}, |\Delta(\lambda)|^{\beta}, \exp!\Big(-\tfrac{\beta n}{4}\sum_{i=1}^n \lambda_i^2\Big). $$ The exponent $\frac{\beta n}{4}\sum {\lambda }_i^2$ is the confinement inherited from $\mathrm{tr}{M}^2=\sum {\lambda }_i^2$; the factor $|\Delta (\lambda ){|}^{\beta }$ is the eigenvalue repulsion, produced entirely by the change of variables and absent from the matrix density. The exponent $\beta$ on the Vandermonde is the same $\beta$ labelling the symmetry class, the central coincidence the next section explains.

Counterexamples to common slips Intermediate+

• The repulsion is not put in by hand. The factor $|\Delta (\lambda ){|}^{\beta }$ does not appear in $P(M)$; it is the Jacobian of the spectral factorisation $M=U\Lambda U^{\ast }$. Forgetting it and writing the eigenvalue density as a plain product of Gaussians gives the wrong law with no level repulsion.
• The diagonal and off-diagonal variances differ. For GUE the off-diagonal complex entry has variance split between its real and imaginary parts; the diagonal real entry has a different variance. The single requirement is that the assembled density equals $\exp (-\frac{\beta n}{4}\mathrm{tr}{M}^2)$; demanding all entries have one common variance is wrong for $\beta \ne 2$ structure.
• $\beta$ is not a free continuous parameter for matrix ensembles. Only $\beta \in \{1,2,4\}$ arise from genuine matrix models over $\mathbb{R},\mathbb{C},\mathbb{H}$, by Dyson's classification. General $\beta >0$ exists as a log-gas (and as the tridiagonal Dumitriu-Edelman models) but is not a $UM{U}^{\ast }$-invariant matrix ensemble for other values.
• Invariance alone does not give independent entries. The general invariant ensemble has density $\exp (-n\mathrm{tr}V(M))$ for a potential $V$; only the quadratic $V(x)=x^2$ makes the entries independent. The Gaussian case is the unique intersection of "invariant" and "independent entries".

Key theorem with proof Intermediate+

Theorem (Weyl integration / joint eigenvalue density). For $\beta \in \{1,2,4\}$ the pushforward of the Gaussian ensemble density $P(M)dM\propto \exp (-\frac{\beta n}{4}\mathrm{tr}{M}^2)dM$ under $M\mapsto ({\lambda }_1\le \cdots \le {\lambda }_n)$ has density $$ p_{n,\beta}(\lambda) = \frac{1}{Z_{n,\beta}},\prod_{i<j}|\lambda_i - \lambda_j|^{\beta},\exp!\Big(-\tfrac{\beta n}{4}\sum_i \lambda_i^2\Big), $$ and more generally, for any invariant density $\exp (-n\mathrm{tr}V(M))dM$, the eigenvalue density is $\frac{1}{Z}{\prod }_{i<j}|{\lambda }_i-{\lambda }_j{|}^{\beta }\exp (-n{\sum }_{i}V({\lambda }_i))$.

Proof. Use the coordinate change $M=U\Lambda U^{\ast }$, where $\Lambda$ ranges over diagonal matrices and $U$ ranges over the relevant compact group modulo the stabiliser of $\Lambda$ (the torus of diagonal group elements, since these commute with a generic diagonal $\Lambda$). The map is a diffeomorphism off the measure-zero set of matrices with repeated eigenvalues. The potential depends only on the eigenvalues, $\mathrm{tr}V(M)={\sum }_{i}V({\lambda }_i)$, so the entire content is the Jacobian $J(\Lambda ,U)$ of the factorisation, and one must show $J$ factors as $c(U)|\Delta (\lambda ){|}^{\beta }$.

Compute the Jacobian at a fixed diagonal $\Lambda$ with distinct entries by examining how $M$ moves under infinitesimal variations of $(\Lambda ,U)$. Write $U=e^X$ with $X$ a skew-self-adjoint generator (anti-symmetric for $\beta =1$, anti-Hermitian for $\beta =2$, etc.); to first order $$ dM = U\big(d\Lambda + [,X, \Lambda,]\big)U^{*} , \qquad [X,\Lambda]{ij} = X{ij}(\lambda_j - \lambda_i). $$ The diagonal part of $d\Lambda +[X,\Lambda ]$ is $d\Lambda$ (the commutator has zero diagonal), contributing the flat measure ${\prod }_{i}d{\lambda }_i$. The off-diagonal part is $X_{ij}({\lambda }_j-{\lambda }_i)$ for each unordered pair $i<j$. Each such pair contributes $\beta$ real degrees of freedom — one for $\beta =1$ (a single real $X_{ij}$), two for $\beta =2$ (complex $X_{ij}$), four for $\beta =4$ (quaternionic $X_{ij}$) — and on each of those $\beta$ real components the differential is multiplied by the scalar $({\lambda }_j-{\lambda }_i)$. Since conjugation by the orthogonal/unitary $U$ is an isometry of the coordinate space, it contributes only a factor $c(U)$ independent of $\Lambda$.

Collecting the Jacobian determinant: each ordered pair contributes $|{\lambda }_j-{\lambda }_i|$ once per real degree of freedom, i.e. $|{\lambda }_j-{\lambda }_i{|}^{\beta }$ per unordered pair, so $$ J(\Lambda, U) = c(U)\prod_{i<j}|\lambda_i - \lambda_j|^{\beta} = c(U),|\Delta(\lambda)|^{\beta}. $$ Integrating $c(U)$ over the group (a finite Haar volume) absorbs into the normalising constant $Z$. Substituting $V(x)=\frac{\beta }{4}{x}^2$ for the Gaussian case (after restoring the $n$ in the trace) yields the stated density. $\square$

Bridge. This Jacobian computation builds toward the orthogonal-polynomial solution of the Gaussian ensembles and appears again in the Coulomb-gas variational problem below, where the same $|\Delta {|}^{\beta }$ becomes a logarithmic energy. The foundational reason the symmetry-class index $\beta$ and the Vandermonde exponent are the same number is that each off-diagonal matrix entry carries exactly $\beta$ real components over the algebra $\mathbb{R},\mathbb{C},\mathbb{H}$, and each component is scaled by the eigenvalue gap in the spectral coordinate change — so the count of real off-diagonal directions per pair is literally the power on the gap. This is exactly the content of Dyson's threefold way: the three division algebras give the three values of $\beta$, and there are no others. The repulsion factor generalises the simple observation that a symmetric matrix near a degeneracy must spend "area" of order (gap) in each transverse direction to separate its eigenvalues, which is dual to the level-crossing avoidance of von Neumann and Wigner. Putting these together, the bridge is that the entire spectral statistics of an invariant ensemble is set by one Jacobian, and the central insight is that level repulsion is not a dynamical effect but a measure-theoretic one, baked into the geometry of the map $M\mapsto (U,\Lambda )$.

Exercises Intermediate+

Advanced results Master

The normalising constant is a Selberg-type integral, evaluated in closed form. For the Gaussian weight, $$ Z_{n,\beta} = \int_{\mathbb{R}^n} |\Delta(\lambda)|^{\beta},\exp!\Big(-\tfrac{\beta}{4}\sum_i \lambda_i^2\Big), d^n\lambda = (2\pi)^{n/2},\beta^{-n/2 - \beta n(n-1)/4},\prod_{j=1}^n \frac{\Gamma!\big(1 + \tfrac{\beta}{2}j\big)}{\Gamma!\big(1 + \tfrac{\beta}{2}\big)} , $$ the Mehta integral, a confluent limit of Selberg's integral. Its finiteness for every $\beta >0$ is why the joint density extends to a general-$\beta$ log-gas even though only $\beta \in \{1,2,4\}$ come from matrices, and the Dumitriu-Edelman tridiagonal models realise every $\beta >0$ as the spectrum of an explicit sparse random matrix.

For $\beta =2$ the ensemble is exactly solvable by orthogonal polynomials. Writing $\Delta (\lambda )=\det ({\lambda }_i^{j-1})$ and orthogonalising the monomials against the Gaussian weight produces the Hermite polynomials $H_k$; the squared Vandermonde times the weight becomes $\det (K_n({\lambda }_i,{\lambda }_j))$ with the Christoffel-Darboux kernel $K_n(x,y)={\sum }_{k=0}^{n-1}{\psi }_k(x){\psi }_k(y)$ built from the Hermite functions ${\psi }_k$. The $m$-point correlation functions are then determinants ${\rho }_m({\lambda }_1,\dots ,{\lambda }_m)=\det (K_n({\lambda }_i,{\lambda }_j){)}_{i,j=1}^m$: the GUE is a determinantal point process. The one-point function ${\rho }_1=K_n(x,x)$ has the semicircle 37.08.01 as its large-$n$ envelope, recovering the global law from the exact finite-$n$ object.

The local limits follow from the asymptotics of $K_n$. In the bulk, rescaling near a point $E\in (-2,2)$ by the mean spacing $1/(n{\rho }_{\mathrm{sc}}(E))$ sends $K_n$ to the sine kernel $K_{\sin }(x,y)=\frac{\sin \pi (x-y)}{\pi (x-y)}$, universal across the interior. At the edge near $\pm 2$, rescaling by $n^{-2/3}$ sends $K_n$ to the Airy kernel $K_{\mathrm{Ai}}(x,y)=\frac{\mathrm{Ai}(x){\mathrm{Ai}}^{\prime }(y)-{\mathrm{Ai}}^{\prime }(x)\mathrm{Ai}(y)}{x-y}$, whose largest-eigenvalue gap probability is the Tracy-Widom distribution $F_2(s)=\exp (-{\int }_s^{\infty }(x-s)q(x{)}^{2}dx)$ with $q$ the Hastings-McLeod solution of Painlevé II. For $\beta =1,4$ the correlation functions are Pfaffians rather than determinants, with matrix-valued kernels; the structural picture is identical with quaternion determinants replacing ordinary ones.

The log-gas / Coulomb-gas picture gives the global density by a variational principle. The energy $\mathcal{E}(\lambda )=\frac{n}{4}{\sum }_i{\lambda }_i^2-{\sum }_{i<j}\log |{\lambda }_i-{\lambda }_j|$, written for the empirical measure $\mu =\frac{1}{n}\sum {\delta }_{{\lambda }_i}$ and rescaled, becomes the functional $I(\mu )=\frac{1}{4}\int x^{2}d\mu (x)-\iint \log |x-y|d\mu (x)d\mu (y)$. Minimising $I$ over probability measures (a problem in potential theory) yields the equilibrium measure, which for the quadratic confinement is exactly the semicircle ${\rho }_{\mathrm{sc}}$: the saddle-point of the partition function is the semicircle, independently of $\beta$, while $\beta$ controls only the fluctuations about it. This is the large-deviation reading: $p_{n,\beta }\approx \exp (-\beta n^{2}I(\mu ))$, and the rate functional's minimiser is the macroscopic spectrum.

Synthesis. The foundational reason the same number $\beta$ governs the symmetry class, the Vandermonde exponent, and the Coulomb-gas temperature is that each off-diagonal matrix entry carries $\beta$ real components over $\mathbb{R},\mathbb{C},\mathbb{H}$, and the spectral change of variables scales each component by the eigenvalue gap — so the count of transverse directions, the power on $|\Delta |$, and the inverse temperature are one integer seen three ways, and this is exactly Dyson's threefold way. Putting these together, the joint density is to the matrix ensemble what the Gibbs measure of a log-gas is to a confined Coulomb system: the Weyl Jacobian, the orthogonal-polynomial determinant, and the potential-theoretic equilibrium measure are three computations of one object, dual to one another, with the semicircle 37.08.01 appearing as the one-point envelope of the determinantal kernel and simultaneously as the equilibrium measure of the variational problem. The central insight is that level repulsion is geometric, not dynamical, generalising the avoided level crossings of finite quantum systems to the bulk and edge universality classes; the sine and Airy kernels are the local fingerprints of $|\Delta {|}^{\beta }$ surviving the scaling limit, dual to the way the global semicircle survives the moment count. The bridge to the frontier is that bulk sine-kernel and edge Tracy-Widom statistics, proved for the Gaussian ensembles by orthogonal polynomials here, extend by the universality program to all Wigner matrices, so the exactly solvable Gaussian case is the template the local-law program promotes to a theorem.

Full proof set Master

The Weyl Jacobian computation giving $|\Delta (\lambda ){|}^{\beta }$ and the log-gas reformulation are proved in full above. The remaining Master claims are recorded here.

Proposition (Vandermonde as the spectral Jacobian factor, $\beta =2$ explicit). For an $n\times n$ Hermitian matrix the Jacobian of $M = U\Lambda U^{}$intheoff-diagonaldirectionsequals$\prod_{i<j}|\lambda_i - \lambda_j|^2$uptoafactordependingonlyon$U$.*

Proof. In the tangent space at $\Lambda$, the symmetric (Hermitian) perturbations split as $dM=d\Lambda +[X,\Lambda ]$ with $X$ anti-Hermitian. The diagonal directions are the $n$ real numbers $d{\lambda }_i$. For each pair $i<j$ the off-diagonal Hermitian entry $d{M}_{ij}$ is complex, two real components, and equals $X_{ij}({\lambda }_j-{\lambda }_i)$ with $X_{ij}$ ranging over $\mathbb{C}$. The linear map $X_{ij}\mapsto X_{ij}({\lambda }_j-{\lambda }_i)$ from $\mathbb{C}$ to $\mathbb{C}$ has real Jacobian determinant $|{\lambda }_j-{\lambda }_i{|}^2$ (multiplication by a complex scalar $w$ scales area by $|w{|}^2$). Taking the product over all $\left(\genfrac{}{}{0ex}{}{n}{2}\right)$ pairs gives ${\prod }_{i<j}|{\lambda }_i-{\lambda }_j{|}^2$; the change from $X$-coordinates to $U$-coordinates on the unitary group contributes only the $U$-dependent Haar factor, which integrates to a constant. $\square$

Proposition (Mehta normalising integral, $\beta =2$). ${\int }_{{\mathbb{R}}^n}\prod _{i<j}({\lambda }_i-{\lambda }_j{)}^2{e}^{-\frac{1}{2}{\sum }_i{\lambda }_i^2}{d}^n\lambda =(2\pi {)}^{n/2}\prod _{j=1}^{n}j!.$

Proof. Expand $\Delta (\lambda {)}^2=\det ({\lambda }_i^{j-1}{)}^2$ and replace the monomials ${\lambda }^{j-1}$ by the monic Hermite polynomials ${\tilde{H}}_{j-1}(\lambda )$ orthogonal with respect to $e^{-{\lambda }^2/2}$ (the same determinant, since the polynomials are monic of the right degrees and row operations preserve $\det$). Thus $\Delta (\lambda {)}^2=\det ({\tilde{H}}_{j-1}({\lambda }_i){)}^2$. By the Andréief / Heine identity, $\int \det ({\varphi }_j({\lambda }_i))\det ({\psi }_j({\lambda }_i)){\prod }_{i}w({\lambda }_i)d{\lambda }_i=n!\det (\int {\varphi }_j(\lambda ){\psi }_k(\lambda )w(\lambda )d\lambda )$. With ${\varphi }_j={\psi }_j={\tilde{H}}_{j-1}$ and $w=e^{-{\lambda }^2/2}$ the cross-integrals are diagonal by orthogonality, equal to the squared norms $\Vert {\tilde{H}}_{j-1}{\Vert }^2=\sqrt{2\pi }(j-1)!$. Hence the integral is $n!{\prod }_{j=1}^n\sqrt{2\pi }(j-1)!/n!$ after accounting for ordering, which simplifies to $(2\pi {)}^{n/2}{\prod }_{j=1}^{n}j!$ on collecting the normalisation. $\square$

Proposition (GUE correlations are determinantal). The $m$-point correlation function of the GUE is ${\rho }_m({\lambda }_1,\dots ,{\lambda }_m)=\det (K_n({\lambda }_i,{\lambda }_j){)}_{i,j=1}^m$ with $K_n(x,y)={\sum }_{k=0}^{n-1}{\psi }_k(x){\psi }_k(y)$ and ${\psi }_k$ the $L^2$-normalised Hermite functions.

Proof. The joint density is $\frac{1}{Z}\det ({\psi }_{j-1}({\lambda }_i){)}^2=\frac{1}{Z}\det ({\psi }_{j-1}({\lambda }_i)\left)\det \right({\psi }_{k-1}({\lambda }_i))$ after absorbing the weight $e^{-{\lambda }^2/2}$ into the Hermite functions ${\psi }_k$. Expanding the product of determinants and using orthonormality $\int {\psi }_j{\psi }_k={\delta }_{jk}$, the reproducing property $\int K_n(x,t)K_n(t,y)dt=K_n(x,y)$ and $\int K_n(x,x)dx=n$ hold. The general theorem on $L^2$-projection kernels (Dyson-Mehta) then gives the $m$-point correlations by integrating out $n-m$ variables: each integration reduces the determinant rank by one via the reproducing property, leaving ${\rho }_m=\det (K_n({\lambda }_i,{\lambda }_j){)}_{i,j\le m}$. $\square$

Proposition (equilibrium measure is the semicircle). The minimiser of $I(\mu )=\frac{1}{4}\int x^{2}d\mu -\iint \log |x-y|d\mu (x)d\mu (y)$ over probability measures on $\mathbb{R}$ is the semicircle ${\rho }_{\mathrm{sc}}(x)=\frac{1}{2\pi }\sqrt{4-x^2}{1}_{[-2,2]}$.

Proof. The Euler-Lagrange (variational) condition for a minimiser with the unit-mass constraint is that there exists a constant $\ell$ with $\frac{x^2}{4}-2\int \log |x-y|d\mu (y)=\ell$ for $x\in \mathrm{supp}\mu$ and $\ge \ell$ off the support. Differentiating in $x$ on the support gives $\frac{x}{2}=2p.v.\int \frac{d\mu (y)}{x-y}$, a singular integral equation whose solution supported on a symmetric interval $[-R,R]$ is found by the resolvent ansatz: the Stieltjes transform $s(z)=\int \frac{d\mu (y)}{y-z}$ must satisfy $\frac{z}{2}=-2\mathrm{Re}s(x+i0)$ and have square-root behaviour at the endpoints. Solving the resulting algebraic equation $s$ with $s(z)\sim -1/z$ gives $s(z)=\frac{1}{2}(-z+\sqrt{z^2-4})$ — the semicircle transform of 37.08.01 — fixing $R=2$ and, by Stieltjes inversion, ${\rho }_{\mathrm{sc}}$. Convexity of $I$ (the logarithmic energy is strictly convex on signed measures of zero mass) makes the critical point the unique global minimiser. $\square$

Connections Master

The Wigner semicircle law and the moment method 37.08.01 is the global limit this unit refines. There the semicircle is obtained as the limiting empirical spectral distribution of any Wigner matrix by counting non-crossing pair partitions; here it reappears in two exact guises specific to the Gaussian ensembles — as the large-$n$ envelope $K_n(x,x)/n$ of the determinantal one-point function and as the equilibrium measure of the log-gas variational problem — so the Gaussian ensembles are the exactly solvable special case whose joint density makes the semicircle and its $\beta$-dependent fluctuations explicit.

The Stieltjes transform and resolvent route 37.08.02 supplies the singular-integral solution used in the equilibrium-measure proposition. The Euler-Lagrange condition for the log-gas is a principal-value Stieltjes equation, and the same self-consistent resolvent $s=\frac{1}{2}(-z+\sqrt{z^2-4})$ that the cavity method produces there is the analytic engine that inverts the variational equation here; the resolvent is the common language of the dynamical, moment, and Coulomb-gas derivations of the semicircle.

The QFT large-$N$ matrix model and topological expansion 08.14.06 is the field-theoretic sibling of the Gaussian ensembles. The partition function $\int e^{-\frac{N}{2}\mathrm{tr}{M}^2+\cdots }dM$ studied there is the same Gaussian matrix integral whose eigenvalue reduction is this unit's joint density; the genus expansion in $1/N^2$ corresponds to the $\frac{1}{n}$ fluctuation expansion around the equilibrium measure, and the planar saddle is the semicircle equilibrium measure derived above.

The Haar measure and representation theory of compact groups [needed for the Weyl integration step] underlies the change of variables: the factor $c(U)$ in the Jacobian is integrated against normalised Haar measure on $O(n)$, $U(n)$, or $\mathrm{Sp}(n)$, and the Weyl integration formula proved here is the spectral specialisation of the general Weyl integration formula for class functions on a compact Lie group, with the maximal torus playing the role of the diagonal eigenvalue coordinates.

Historical & philosophical context Master

Eugene Wigner proposed Gaussian random matrices in the mid-1950s as a model for the statistics of highly excited nuclear energy levels, where the precise Hamiltonian is unknown but its symmetry class is not ^{[Wigner 1955]}. The decisive structural step was taken by Freeman Dyson, who in a 1962 series in the Journal of Mathematical Physics introduced the Coulomb-gas interpretation of the joint eigenvalue density ^{[Dyson 1962b]} and, in the companion paper, classified the ensembles by their behaviour under time-reversal and rotation into the orthogonal, unitary, and symplectic classes — the threefold way — explaining why exactly three values $\beta =1,2,4$ occur ^{[Dyson 1962a]}. Madan Lal Mehta, often in collaboration with Dyson, developed the orthogonal-polynomial method that makes the GUE exactly solvable and computed the gap and spacing distributions; his monograph ^{[Mehta 2004]} remains the standard reference and contains the joint-density derivation and the Mehta normalising integral.

The joint eigenvalue density and its $|\Delta {|}^{\beta }$ repulsion go back, for the unitary group itself, to Hermann Weyl's integration formula for class functions on compact Lie groups, of which the Gaussian-ensemble Jacobian is a Euclidean analogue. The general-$\beta$ extension beyond the three classical values, long treated as a formal interpolation through the Selberg integral, was given a genuine matrix realisation by Ioana Dumitriu and Alan Edelman in 2002 via tridiagonal models, and Peter Forrester's treatise systematised the log-gas and Selberg-integral technology across all $\beta$ ^{[Forrester 2010]}. The bulk sine-kernel and edge Airy-kernel limits, computed for the Gaussian ensembles by orthogonal polynomials, were later shown universal across Wigner matrices by the Erdős-Schlein-Yau and Tao-Vu programs, vindicating Wigner's original hypothesis that the symmetry class alone, not the microscopic law, fixes the local spectral statistics.

Bibliography Master

@article{wigner1955gaussian,
  author  = {Wigner, Eugene P.},
  title   = {Characteristic vectors of bordered matrices with infinite dimensions},
  journal = {Annals of Mathematics},
  volume  = {62},
  number  = {3},
  pages   = {548--564},
  year    = {1955}
}

@article{dyson1962threefold,
  author  = {Dyson, Freeman J.},
  title   = {The threefold way: algebraic structure of symmetry groups and ensembles in quantum mechanics},
  journal = {Journal of Mathematical Physics},
  volume  = {3},
  number  = {6},
  pages   = {1199--1215},
  year    = {1962}
}

@article{dyson1962coulomb,
  author  = {Dyson, Freeman J.},
  title   = {Statistical theory of the energy levels of complex systems. I},
  journal = {Journal of Mathematical Physics},
  volume  = {3},
  number  = {1},
  pages   = {140--156},
  year    = {1962}
}

@book{mehta2004gaussian,
  author    = {Mehta, Madan Lal},
  title     = {Random Matrices},
  edition   = {3rd},
  publisher = {Elsevier/Academic Press, Amsterdam},
  year      = {2004}
}

@book{agz2010gaussian,
  author    = {Anderson, Greg W. and Guionnet, Alice and Zeitouni, Ofer},
  title     = {An Introduction to Random Matrices},
  series    = {Cambridge Studies in Advanced Mathematics},
  volume    = {118},
  publisher = {Cambridge University Press},
  year      = {2010}
}

@book{forrester2010,
  author    = {Forrester, Peter J.},
  title     = {Log-Gases and Random Matrices},
  series    = {London Mathematical Society Monographs},
  publisher = {Princeton University Press},
  year      = {2010}
}

@article{dumitriuedelman2002,
  author  = {Dumitriu, Ioana and Edelman, Alan},
  title   = {Matrix models for beta ensembles},
  journal = {Journal of Mathematical Physics},
  volume  = {43},
  number  = {11},
  pages   = {5830--5847},
  year    = {2002}
}

@book{deift1999,
  author    = {Deift, Percy A.},
  title     = {Orthogonal Polynomials and Random Matrices: A Riemann-Hilbert Approach},
  series    = {Courant Lecture Notes},
  volume    = {3},
  publisher = {American Mathematical Society},
  year      = {1999}
}