Why random matrix theory keeps appearing in Riemann zeros

Montgomery's 1973 observation — that the pair correlation of Riemann zeros matches the GUE ensemble of random matrix theory — is one of those results that shouldn't work, and yet has been confirmed numerically to extraordinary precision.

The standard framing treats this as a deep mystery, or as evidence for some unknown operator whose eigenvalues are the zeros. But our Grade-Shadow formalization suggests the mechanism is more direct.

The Grade-Shadow explanation

The Euler product is a multiplicative object: \(\zeta(s) = \prod_p (1-p^{-s})^{-1}\). When you decompose this in a graded tensor algebra, the dominant structure is grade 2 — pairwise interactions between primes.

The key quantity is the grade ratio \(\delta = O(1/\log\log T) \to 0\). This is exactly the regime where Wigner matrix universality kicks in. In other words:

• The Euler product's multiplicative structure forces pairwise dominance
• Pairwise dominance at small \(\delta\) is precisely the GUE universality condition
• GUE statistics are therefore a consequence of arithmetic, not an assumption

What falls out as a corollary

The Rudnick–Sarnak restricted support result (1996) was proved by analytic methods. In our framework it's a corollary: the Fourier support restriction exists because grade-2 truncation kills exactly those frequencies. Their result isn't a theorem about zeta zeros — it's a shadow of the grade structure.

What's formalized

This route has 328 Grade-Shadow declarations with 87 proved and 100% type-checked. The combinatorial core of the Superquadratic Growth Theorem is Lean 4 verified with 0 axioms and 0 sorry.

The conditional part remains the moment hypothesis: \(m_{2k}(T) \leq C_k (\log T)^{k^2+\varepsilon}\), which is weaker than Lindelöf and widely believed but unproved.

Open question

Is there a way to bootstrap from the unconditional SGT to remove the moment hypothesis entirely? The grade decomposition suggests the moments should satisfy the bound — the question is whether you can prove it from grade-2 dominance alone without invoking the moments as an axiom.

This is where I'm currently stuck, and I'd welcome ideas from anyone working on moment problems for \(L\)-functions.