Interactive visualization of eigenvalue spacing distributions — Wigner-Dyson vs Poisson, GOE/GUE/GSE ensembles, and universality in complex systems
In 1951, Eugene Wigner proposed that energy level spacings of heavy atomic nuclei follow a universal distribution — not the Poisson distribution one might expect from independent random levels. For the Gaussian Orthogonal Ensemble (GOE, β=1), the spacing distribution is approximated by P(s) = (π/2)·s·exp(-πs²/4), known as the Wigner Surmise. This simple 2×2 formula is remarkably close to the exact result for large matrices. The key feature is "level repulsion": P(s) ~ s^β as s→0, meaning eigenvalues avoid each other — completely unlike Poisson statistics where P(0) = 1 is maximal.
Level repulsion is the tendency of eigenvalues to avoid clustering: for spacing s near zero, P(s) ∝ s^β, where β is the Dyson index. GOE (β=1): linear repulsion; GUE (β=2): quadratic repulsion; GSE (β=4): quartic repulsion. This reflects the underlying symmetry — real symmetric (GOE), complex Hermitian (GUE), quaternion self-dual (GSE). Remarkably, these distributions are universal: they depend only on the symmetry class, not on the detailed distribution of matrix elements. The same statistics appear in nuclear spectra, quantum dots, zeros of the Riemann zeta function, chaotic billiards, and many other systems.
The eigenvalue density of a large N×N random matrix from any Gaussian ensemble converges to ρ(λ) = (2/(πR²))·√(R²-λ²), where R=2√N is the spectral radius. This "semicircle law" is universal — it holds regardless of the distribution of individual matrix elements (under mild conditions). The top canvas shows this density: as you increase N, the histogram of eigenvalues converges to the semicircle. Edge effects cause Tracy-Widom fluctuations at the spectral edges, which themselves follow a universal distribution used in systems from random growth models to financial correlations.
Real symmetric matrices H where H_ij ~ N(0,1) for i
Complex Hermitian matrices with real diagonal H_ii ~ N(0,1) and complex off-diagonal Re(H_ij), Im(H_ij) ~ N(0,½). Invariant under unitary transformations: H → UHU†. The surmise gives P(s) = (32/π²)·s²·exp(-4s²/π). Stronger level repulsion (quadratic near s=0). Appears in systems without time-reversal symmetry: electrons in magnetic fields, microwave cavities, scattering in quantum chromodynamics. The Riemann zeta zeros on the critical line are believed to follow GUE statistics (Montgomery's pair correlation conjecture, supported by Odlyzko's numerical verification).
Self-dual quaternion matrices with invariance under symplectic transformations. The surmise gives P(s) = (2¹⁸/(3⁶π³))·s⁴·exp(-(64/9π)s²) — a more complex expression with even stronger (quartic) repulsion. Applies to time-reversal invariant systems with half-integer spin (Kramers degeneracy). This is the rarest ensemble in physics but appears in certain mesoscopic systems and quantum dots with strong spin-orbit coupling.
For integrable (non-chaotic) quantum systems, Berry and Tabor (1977) showed that energy levels generally follow Poisson statistics: P(s) = exp(-s). There is no level repulsion — P(0)=1 is maximal, and spacings are completely uncorrelated. The number variance is Σ²(L)=L (linear growth). This serves as the baseline contrast to random matrix predictions. The transition from Poisson to Wigner-Dyson statistics as a system parameter changes (e.g., shape deformation of a billiard) signals the onset of quantum chaos.
Wigner's original motivation: energy levels of heavy nuclei follow GOE statistics. Despite the enormous complexity of nuclear interactions, the spectral statistics of highly excited nuclear states are indistinguishable from random matrix predictions. This has been verified experimentally through neutron resonance spectroscopy — the spacing distributions of neutron capture resonances match the Wigner surmise with remarkable precision. The universality suggests that individual nuclear structure details are irrelevant for spectral statistics.
Perhaps the most striking application: the imaginary parts of the zeros of the Riemann zeta function on the critical line Re(s)=½ appear to follow GUE statistics. Hugh Montgomery (1972) conjectured this based on pair correlation analysis, and Freeman Dyson immediately recognized the connection to random matrix theory. Andrew Odlyzko's extensive numerical computations (millions of zeros centered around the 10²⁰-th zero) confirm GUE statistics to high precision. This suggests a deep, still-unproven connection between prime numbers and random matrices.
The Bohigas-Giannoni-Schmit conjecture (1984) states that quantum systems whose classical counterparts are chaotic exhibit random matrix spectral statistics. A stadium billiard (chaotic) shows Wigner-Dyson statistics; a rectangular billiard (integrable) shows Poisson statistics. This has been verified in microwave cavity experiments, quantum dots, and atomic spectra. The transition from integrable to chaotic can be tracked through the spectral statistics, making RMT a diagnostic tool for quantum chaos.
Laloux, Cizeau, Bouchaud, and Potters (1999, 2000) applied RMT to financial correlation matrices. Eigenvalues outside the RMT-predicted range contain genuine market information; those within the RMT band are noise. This "denoising" technique improves portfolio optimization (Markowitz theory). RMT also appears in: correlation analysis of neural activity, machine learning (Hessian eigenvalue spectra of loss landscapes), wireless communications (MIMO channel capacity), and complex networks (spectral analysis of graph adjacency matrices).