Universality in Large Random Matrices — Epoche C1
Fill a large symmetric matrix with independent random numbers — draw them from a Gaussian, or let each be $+1$ or $-1$ by a coin toss — divide every entry by $\sqrt{N}$, and plot a histogram of the eigenvalues. The two histograms are the same, and as $N$ grows both settle onto a semicircle. That the answer does not depend on which distribution was used is the phenomenon called universality, and this note is about what it really requires. The received account attributes it to the assumption that the entries are independent and identically distributed. That assumption is sufficient and it is not necessary, and the earlier version of this note was right to say so. But it was too generous about how far the conditions can be relaxed, and the corrections below matter, because the relaxations that work and the ones that fail differ in an instructive way. The objects, and why the scaling is $1/\sqrt{N}$ Fix the terms first. A Wigner matrix $H$ of size $N \times N$ is a Hermitian random matrix — meaning $H_{ji} = \overline{H_{ij}}$, so that its eigenvalues are real — whose entries on and above the diagonal are independent, centred, and identically distributed above the diagonal. The empirical spectral distribution , or ESD, is the probability measure that places mass $1/N$ at each of the $N$ eigenvalues; it is the histogram just described, normalised to have total mass one. Universality is a statement about the limit of this measure and, later, about finer statistics within it. The normalisation is not a convention but a forced choice, and the reason is a one-line calculation. The trace of $H^2$ equals the sum of the squares of the eigenvalues, and it also equals the sum of the squares of the entries, so if each entry has variance $s$ then $$\mathbb{E}\!\left[\sum_{k=1}^{N} \lambda_k^2\right] = \mathbb{E}\big[\operatorname{Tr}(H^2)\big] = \sum_{i,j=1}^{N} \mathbb{E}\big|H_{ij}\big|^2 = N^2 s.$$ For the average of $\lambda_k^2$ to stay bounded as $N$ grows, we need $N^2 s / N = Ns$ to be of order one, so $s$ must be of order $1/N$ and the entries of order $N^{-1/2}$. The exponent is one half because the trace couples $N^2$ entries into $N$ eigenvalues, a ratio of $N$. With $s = 1/N$ the limiting law is the semicircle $$f(x) = \frac{1}{2\pi R^{2}}\sqrt{4R^{2} - x^{2}}, \qquad |x| \le 2R,$$ where $R^2$ is the common variance after rescaling. The constant in front is forced too: $\int_{-2R}^{2R}\sqrt{4R^{2}-x^{2}}\,dx$ is the area under a semicircle of radius $2R$, namely $\tfrac{1}{2}\pi(2R)^{2} = 2\pi R^{2}$, so dividing by $2\pi R^{2}$ makes the total mass one. Why the entry distribution washes out: the moment count Eugene Wigner's own route to this law (1958) is a counting argument, and it is worth following because it shows exactly which features of the entries survive the limit — which is the whole question this essay is about. The idea is to compute the moments of the ESD and recognise them. The $2k$-th moment of the ESD is $\frac{1}{N}\operatorname{Tr}(H^{2k})$, and expanding the matrix product turns this into a sum over closed walks of length $2k$ on the vertex set $\{1,\dots,N\}$: $$\frac{1}{N}\,\mathbb{E}\big[\operatorname{Tr}(H^{2k})\big] = \frac{1}{N}\sum_{i_1,\dots,i_{2k}} \mathbb{E}\big[H_{i_1 i_2} H_{i_2 i_3}\cdots H_{i_{2k} i_1}\big].$$ Now count. Because the entries are independent and centred, a term vanishes unless every edge of the walk is traversed at least twice — an edge used once contributes an independent factor of mean zero. A closed walk of $2k$ steps in which every edge is used at least twice therefore visits at most $k$ distinct edges, and a connected graph with $k$ edges has at most $k+1$ vertices. So the number of index tuples contributing is of order $N^{k+1}$. Each surviving term is a product of $k$ variances, of size $(1/N)^{k}$. With the prefactor $1/N$ the total is of order $$N^{\,k+1}\cdot N^{-k}\cdot N^{-1} = 1,$$ so these terms survive. A walk that uses some edge three or more times spans fewer distinct edges and hence fewer vertices, losing at least one factor of $N$, and vanishes in the limit. The surviving walks are exactly those that traverse each of $k$ edges precisely twice on a tree, and these are in bijection with rooted plane trees with $k$ edges, counted by the Catalan number $C_k = \frac{1}{k+1}\binom{2k}{k}$. Odd moments vanish, since a closed walk of odd length cannot pair up its edges. The Catalan numbers are the even moments of the semicircle law with $R=1$, which identifies the limit. This computation is the mechanism of universality, and it should be read carefully. Only pairs of entries appear in the surviving terms, so only second moments — variances — enter the limit. Third, fourth and higher moments of the entry distribution are attached to walks that reuse an edge, and those are suppressed by powers of $1/N$. Universality at the level of the global density is therefore not a mysterious robustness; it is the statement that the leading-order combinatorics sees nothing but variances. The same fact seen through the resolvent The other standard route makes the same point analytically and generalises better. For a probability measure $\mu$ on the real line, its Stieltjes transform is $m(z) = \int \frac{\mu(dx)}{x - z}$ for $z$ off the real axis; it determines $\mu$ uniquely, and pointwise convergence of transforms implies weak convergence of measures, which is why one works with it. For a matrix, the corresponding object is the normalised trace of the resolvent $(H - zI)^{-1}$, the quantity the earlier version of this note wrote down: $$\frac{1}{N}\operatorname{Tr}\big((H_N - zI)^{-1}\big) \longrightarrow m(z) \quad \text{as } N \to \infty.$$ What the earlier version did not say is what $m$ is, and that is where the content lies. Applying the Schur complement formula to a single diagonal entry of the resolvent expresses it in terms of the resolvent of the matrix with one row and column deleted, and to leading order in $N$ it yields the self-cons