Universality, Dyson Brownian Motion, and Log-Gases¶

Job: develop the universality paradigm (bulk / edge / hard edge), Dyson Brownian motion as the eigenvalue SDE / log-gas, the Erdős–Schlein–Yau three-step approach and the Tao–Vu replacement principle, and the rigidity of eigenvalues — with the DBM relaxation to the sine kernel as the worked example.

Core¶

The universality paradigm¶

The local eigenvalue statistics of a Wigner matrix depend only on the symmetry class ($\beta=1,2,4$), not on the entry distribution — the universality theorem (Erdős–Schlein–Yau; Tao–Vu). Three regimes:

Regime	Limiting kernel	Scaling	Universality class
Bulk	sine kernel $K_{\mathrm{sine}}(x,y)=\frac{\sin\pi(x-y)}{\pi(x-y)}$	gaps $\sim 1/N$, unfolded to mean spacing $1$	all Wigner (matching moments)
Soft edge	Airy kernel $K_{\mathrm{Ai}}$	$\lambda_{\max}-2\sim N^{-2/3}$	Tracy–Widom $F_\beta$
Hard edge (Wishart)	Bessel kernel $K_{\mathrm{Bess}}$	near $0$, spacing $\sim N^{-1}$	sample covariance at $0$

Dyson Brownian motion (DBM)¶

The eigenvalue process $\lambda_1(t)\le\cdots\le\lambda_N(t)$ of a Wigner matrix with a GOE/GUE Brownian-motion entry dynamics satisfies the Dyson Brownian motion SDE $$ d\lambda_i=\frac{\beta}{2}\sum_{j\ne i}\frac{dt}{\lambda_i-\lambda_j}+\sqrt{\frac{1}{N}}\,dB_i, $$ the Coulomb repulsion $\frac{\beta}{2}\cdot\frac{1}{\lambda_i-\lambda_j}$ from the Vandermonde Jacobian, plus the independent Brownian noise $\sqrt{1/N}\,dB_i$. The eigenvalues form a log-gas: particles repelling logarithmically (the Coulomb gas) and confined by the Gaussian potential.

DBM data	Description
SDE	$d\lambda_i=\frac{\beta}{2}\sum_{j\ne i}\frac{dt}{\lambda_i-\lambda_j}+\sqrt{1/N}\,dB_i$
Log-gas potential	$V(x)=x^2/2$ (Gaussian confinement); the repulsion $\sum_{i<j}\log
Invariant measure	the GUE / GOE eigenvalue density $\propto\prod_{i<j}
Local equilibrium time	$t\gg N^{-1}$ (the local statistics converge to the DBM equilibrium)

The three-step approach (Erdős–Schlein–Yau)¶

Step	Content	Key estimate
(1) Local semicircle law	the Stieltjes transform $g(z)$ concentrates; the ESD is close to the semicircle down to scale $N^{-1+\varepsilon}$	$\\|g-g_{\mathrm{sc}}\\|=O(N^{-1/2+\varepsilon})$ for $\mathrm{Im}\,z\gg N^{-1+\varepsilon}$
(2) DBM relaxation	the local gap statistics converge to the DBM equilibrium in time $t\gg N^{-1}$	the gap distribution at time $t$ matches the GUE gap for $t\gg N^{-1}$
(3) DBM matching	the DBM equilibrium is the GUE / GOE (the sine kernel); the matching via the Harish-Chandra–Itzykson–Zuber integral / the Brézin–Hikami formula	the gap statistics of $H_t$ (DBM-perturbed) match those of $H_0$

The Tao–Vu approach (the moment method, the replacement principle): replace a Wigner entry by a Gaussian entry without changing the local statistics, via the Lindeberg swapping (the replacement principle — the local statistics are unchanged under the entry-by-entry Gaussian replacement, provided the moments match to order $4$).

Log-gases and the Coulomb gas interpretation¶

The eigenvalue density of the $\beta$-ensemble $$ P(\lambda_1,\dots,\lambda_N)\propto\prod_{i<j}|\lambda_i-\lambda_j|^\beta\,e^{-N\sum_i V(\lambda_i)} $$ (the general potential $V$, the $\beta$ generalization of GOE/GUE/GSE) is the log-gas: the particles $\lambda_i$ repel logarithmically (the Vandermonde) and are confined by $V$. The equilibrium measure $\rho_V$ is the minimizer of the log-gas energy $$ I(\rho)=\iint\log|x-y|^{-1}\rho(x)\rho(y)\,dx\,dy+\int V(x)\rho(x)\,dx. $$

Log-gas data	Description
Energy $I(\rho)$	the log-gas energy (the Coulomb self-energy + the confinement)
Equilibrium measure $\rho_V$	the minimizer of $I$; for $V=x^2/2$, the semicircle
Rigidity	$

The rigidity theorem (Erdős–Schlein–Yau): the eigenvalues are close to their classical locations $\gamma_i$ (the $i$-th quantile of $\rho_V$) — the eigenvalues do not fluctuate much from the equilibrium positions (the bulk fluctuation $O(N^{-1}(\log N)^c)$, the edge $O(N^{-2/3}(\log N)^c)$).

See it¶

DBM relaxation to the sine kernel in the bulk. Start with a Wigner matrix $H_0$ (with a general entry distribution, not necessarily Gaussian), and add a small GOE Brownian-motion perturbation $H_t=H_0+\sqrt{t}\,G$ (the DBM time $t$, $G$ a GOE matrix). The eigenvalues evolve by the DBM SDE. The three-step: 1. The local semicircle law holds for $H_t$ for all $t$ (the ESD is close to the semicircle down to scale $N^{-1+\varepsilon}$, the Stieltjes transform concentrates). 2. The DBM relaxation: the local gap statistics at time $t$ are close to the DBM equilibrium (the GUE gap statistics) for $t\gg N^{-1}$ (the local equilibration, the relaxation to the invariant measure in the local statistics). 3. The matching: the gap statistics of $H_t$ are the same as $H_0$ (the DBM perturbation preserves the gap statistics — the DBM matching via the Brézin–Hikami formula / the Harish-Chandra–Itzykson–Zuber integral, the gap distribution is invariant under the DBM flow up to the local equilibration).

Hence the gap statistics of $H_0$ (the original Wigner matrix) are the GUE sine-kernel statistics — the universality. The key point: the Dyson Brownian motion is the mechanism — the eigenvalue dynamics relax to the universal equilibrium in a short time ($t\sim N^{-1}$), and the matching preserves the gap statistics, giving the universality for arbitrary Wigner matrices.

Check¶

Why is the DBM relaxation time $N^{-1}$ (not $O(1)$), and what does this say about the local equilibration?

The DBM SDE has the Coulomb drift $\frac{\beta}{2}\sum_{j\ne i}\frac{1}{\lambda_i-\lambda_j}$ (the repulsion) and the noise $\sqrt{1/N}\,dB_i$. The local relaxation: the gap $\lambda_{i+1}-\lambda_i\sim 1/N$ in the bulk, so the drift $\sim\beta N/2$ (the nearest-neighbor repulsion), and the noise variance $\sim 1/N$. The relaxation time for a gap to equilibrate is the time for the drift to overcome the noise: $t_{\mathrm{relax}}\sim(\text{gap})^2/(\text{diffusion})\sim(1/N)^2/(1/N)=N^{-1}$. So the local equilibration happens in time $N^{-1}$ (the local, not the global, equilibrium — the global / the ESD equilibration is $O(1)$, but the local gap statistics equilibrate in $N^{-1}$).

Show that the rigidity (the eigenvalue closeness to the classical location) implies the local semicircle law down to scale $N^{-1+\varepsilon}$.

The rigidity says $|\lambda_i-\gamma_i|=O(N^{-2/3}(\log N)^c/\hat i^{1/3})$. In the bulk ($i\sim N/2$), this is $O(N^{-1}(\log N)^c)$ — the eigenvalues are within $O(N^{-1}(\log N)^c)$ of the classical locations. The Stieltjes transform $g(z)=\frac1N\sum\frac{1}{\lambda_i-z}$ for $\mathrm{Im}\,z\gg N^{-1+\varepsilon}$ (i.e. above the rigidity scale) is then close to $g_{\mathrm{sc}}(z)$ (the eigenvalues are close to the classical locations, the sum is close to the integral against $\rho_{\mathrm{sc}}$), with the error controlled by the rigidity and the imaginary part. So the local semicircle law (the concentration of $g$) follows from the rigidity for $\mathrm{Im}\,z\gg N^{-1+\varepsilon}$.

Pitfalls¶

The universality (Erdős–Schlein–Yau, Tao–Vu) is for Wigner matrices with finite variance / moment hypotheses (the matching moments to order $4$, the finite variance) — heavy tails give Lévy limits (not the semicircle), sparse matrices give different limits (the tree spectral measures, the Biane–Bordenave tree-like RMT).
The three-step approach is the bulk universality — the edge needs the separate Airy / Tracy–Widom analysis (the edge scaling, the Airy kernel, the Deift–Zhou steepest descent). Do not conflate the bulk and the edge universality.
The DBM relaxation time is $N^{-1}$ (the local equilibrium, not a global one) — the global / the ESD equilibration is $O(1)$, the local gap statistics equilibrate in $N^{-1}$.
The rigidity is the eigenvalue closeness (the $N^{-2/3}$ edge, the $N^{-1}$ bulk) — the rigidity is the input to the local semicircle law and the DBM relaxation, not a consequence.
The Tao–Vu replacement principle requires the moment matching (the moments match to order $4$ for the four-moment theorem; the Lindeberg swapping preserves the local statistics under the moment condition), not the distributional matching.
The log-gas / the $\beta$-ensemble is the general-$\beta$ generalization; the DBM is the $\beta=1,2$ (the GOE / GUE) specific dynamics, the general $\beta$ requires the stochastic Airy operator (the $\beta$-Tracy–Widom, see determinantal & $\beta$-ensembles).

Refs¶

Erdős–Schlein–Yau, "Universality of random matrices and log-gases" (CMP 2010); "Rigidity of eigenvalues" (2012)
Tao–Vu, "Random matrices: universality of local eigenvalue statistics" (2010); "Random matrices: the four-moment theorem"
Dyson, "A Brownian-motion model for the eigenvalues of a random matrix" (1962)
Anderson–Guionnet–Zeitouni, An Introduction to Random Matrices; Erdős–Yau, A Dynamical Approach to Random Matrix Theory (AMS 2017)

Next¶

RMT index · Foundations · Determinantal & β-Ensembles · Spiked Models & Statistics · Non-Hermitian & Circular · Free probability · FP / free stochastic — the DBM as the large-$N$ limit / the free Brownian motion · Probability