Change what you put in the matrix and the local statistics do not notice
Fill a symmetric matrix with Gaussian numbers and the gaps between neighbouring eigenvalues follow a particular law. Fill it with coin flips instead — every entry either +1 or −1, no bell curve anywhere — and the gaps follow the same law. Fill it with numbers drawn flat from an interval: same law again.
This is the deep claim of the subject, and it is much stronger than it first sounds. It is not that the three answers are similar. It is that they converge to the identical limiting distribution, so that a measurement of eigenvalue spacings carries no information whatsoever about what the entries were.
The statement is known as the Wigner–Dyson–Gaudin–Mehta conjecture. It was proved for wide classes of matrices in roughly 2009 to 2012, through work of the Erdős–Yau school and of Tao and Vu, using different methods that arrived at overlapping results — which is why it is fairer to describe it as a body of work over several years than to attach it to a single paper or a single date.
Four entry distributions, one matrix size, one histogram. Three of the four have variance 1 and differ in every other respect. The fourth has no variance at all, and is here because the theorem needs finite variance and it is worth seeing what that hypothesis is doing.
Dense 140 × 140 symmetric matrices, built entry by entry and diagonalised by Jacobi rotations — not the tridiagonal shortcut, because the whole claim here is about what the entries are. Gaps are divided by the mean gap among their immediate neighbours, so no global shape is assumed anywhere.
Try it: Collect a few thousand spacings on Gaussian, note the distance to the Wigner curve, then switch to ±1 coin flips and collect the same amount. The number barely moves. Measured offline over 24 dense 150 × 150 matrices each, the three finite-variance laws gave 0.072, 0.098 and 0.100 — a spread of under 0.03, against a distance to Poisson of about 0.53 for all of them.
A prediction that failed, and what actually happened: the natural expectation is that Cauchy entries — infinite variance, outside the theorem’s hypotheses — will break the spacing statistics. They do not, at any size tested here. Measured with the density estimated locally rather than assumed, the distance from the Cauchy spacings to the Wigner curve was 0.53 at n = 80, 0.14 at n = 160 and 0.09 at n = 320: moving toward the Gaussian answer of about 0.08, while the distance to Poisson stayed above 0.45 throughout. What Cauchy destroys is the global law — the spectrum spans about 1,900 at n = 80, 7,000 at n = 160 and 17,000 at n = 320, against a steady 3.9 for Gaussian entries, and that failure grows with n rather than healing. The honest summary is that infinite variance wrecks the semicircle and leaves the local repulsion standing, which is a more interesting fact than the one we went looking for.
The Cauchy result only makes sense once the two questions are properly separated. Where are the eigenvalues? is a global question, and its answer is a curve: a semicircle for Wigner matrices, something else for other ensembles. How far apart are neighbours? is a local question, and its answer survives changes that rewrite the global curve completely.
The slider is exact rather than statistical: multiplying a matrix by a constant multiplies every eigenvalue by that constant, and the local unfolding divides gaps by neighbouring gaps, so the right-hand histogram cannot move even in principle. The ensemble switch is the real test — sample covariance matrices (120 variables, 240 observations) have a global density with a hard edge away from zero and a long tail, nothing like a semicircle, and their spacings land on the same Wigner curve anyway.
Try it: Drag the scale slider and watch the left chart stretch while the right one sits still — that much is exact arithmetic rather than a statistical claim, since multiplying a matrix by a constant multiplies every eigenvalue by it. Then switch to sample covariance matrices. The global density changes shape entirely, with an edge away from zero and a long right tail, and the spacings land on the same Wigner curve regardless.
Universality is not the claim that everything is a random matrix. The clean way to see the boundary is to build something with the right global density and the wrong local structure — and the easiest such thing is a diagonal matrix, whose eigenvalues are simply its entries.
Draw those entries from the semicircle distribution itself. Now the global spectral density is not approximately a semicircle, it is exactly one. And the spacings collapse to Poisson, because independent points do not repel.
The right-hand column is a diagonal matrix, which is about as structured as a matrix gets: its eigenvalues are simply its entries, and they know nothing about one another. The most likely gap between neighbours is zero, which is exactly what an eigenvalue of a random matrix will not do. Measured over 40 samples of size 300, the independent points sat 0.058 from the Poisson curve and 0.546 from Wigner, while the matrix spectrum sat 0.047 from Wigner and 0.541 from Poisson — the two verdicts are not close, and they point in opposite directions.
Try it: Sample both columns. The top charts converge to the same curve; the bottom ones converge to different ones. The gap at zero on the left is level repulsion, and its absence on the right is what independence looks like.
So the hypothesis that carries the weight is independence of the entries, not the shape of their distribution. Correlate them strongly enough, or impose enough structure, and the eigenvalues stop pushing on one another. That is the direction from which counterexamples come, not from replacing one bell curve with another.
The reason this matters outside mathematics is a licensing argument. Nobody knows the Hamiltonian of a uranium nucleus. Universality says that you do not need to: if you are willing to grant that the operator is complicated and Hermitian, the local statistics of its levels are already determined, and they are the statistics of a matrix full of noise.
The Hamiltonian. Hundreds of nucleons interacting through a force nobody can write down exactly, let alone diagonalise.
That it is a Hermitian operator, and — if the system is invariant under time reversal — that it can be written with real entries.
The gaps between neighbouring energy levels, once rescaled to unit mean, follow the β = 1 spacing law. Level repulsion, with the probability of a near-collision vanishing linearly in the gap.
It predicts the statistics of the levels, not the levels. No random matrix will tell you where any particular resonance sits.
The trade is always the same shape. You give up every specific fact about the system and you get back a distribution — which is worth having only because the distribution is testable, and because there was never any prospect of getting the specific facts.
Try it: Step through the three cases and notice what is constant: in each one, the thing predicted is a distribution, and the thing surrendered is every individual fact. That trade is the whole method.