The Spectral Ledger: Why We Keep Looking for the Operator

What a spectrum would actually buy

The reason anyone wants an operator is not aesthetic. Self-adjoint operators on a Hilbert space have real eigenvalues — that is a theorem, not a hope. So if someone exhibits a self-adjoint operator whose eigenvalues are the numbers t in the non-trivial zeros written as one half plus i t, the Riemann Hypothesis follows immediately, because real t is precisely the statement that the zeros lie on the critical line. The hypothesis stops being a question about where zeros happen to sit and becomes a consequence of an operator being what it is.

That is the Hilbert-Polya idea in one sentence, and it explains the persistence of the search. Almost every other route to the hypothesis fights the zeros one region at a time. The spectral route would not fight them at all. It would make their position structural.

It also explains why the standard is so high. A near-miss operator is worth very little. It must be self-adjoint, its spectrum must be exactly the zeros with correct multiplicity, and it must reproduce the counting function — the number of zeros up to height T, which grows like T over two pi times the logarithm of T over two pi e. Anything that gets the statistics right and the counting wrong is a model of the zeros, not their source.

The ledger, entry by entry

Pair correlation. Montgomery computed the two-point statistics of the zeros in 1973, under a restriction on test functions, and Dyson recognised the answer over tea: it is the pair-correlation function of eigenvalues from the Gaussian unitary ensemble, the symmetry class of large random Hermitian matrices. Not a resemblance — the same function. This is the founding entry and it remains the strongest.

Numerics. Odlyzko computed zeros at height ten to the twentieth and beyond and compared the nearest-neighbour spacing distribution against GUE. The agreement is not qualitative. It is agreement to several decimal places over millions of zeros, including the characteristic level repulsion at small spacings — the zeros avoid each other exactly as eigenvalues of a Hermitian matrix do, and nothing like independent random points.

Higher correlations. Rudnick and Sarnak extended the result to n-level correlations for all n, again under restricted test functions, and for a wide class of L-functions rather than zeta alone. That matters: it means the phenomenon is a property of the family, not a coincidence in one function.

The proved analogue. Over function fields the corresponding hypothesis is a theorem — Weil proved it, and Deligne proved the general case — and there the zeros genuinely are eigenvalues, of Frobenius acting on cohomology. Katz and Sarnak then showed that the symmetry statistics predicted for families of L-functions do appear in the function-field setting. So a spectral interpretation is not speculative in principle. In one arithmetic universe it is established fact.

The trace-formula shape. Weil's explicit formula relates a sum over zeros to a sum over prime powers. Selberg's trace formula relates a sum over eigenvalues of the Laplacian on a hyperbolic surface to a sum over closed geodesics. The two have the same architecture, term for term, with primes standing where periodic orbits stand. This is why the analogy has structural weight rather than decorative weight: a formula of that type is what a spectrum with a dynamical system underneath it looks like.

The semiclassical candidate. Berry and Keating argued that if a classical Hamiltonian underlies the zeros it should be the dilation generator, position times momentum, with the correct counting emerging from a Planck-cell truncation of phase space. Connes built a framework on adeles that produces the counting with the zeros appearing as an absorption spectrum. Bender, Brody and Muller later wrote down a differential operator with the right formal eigenvalue condition. None of these is finished: xp has no discrete spectrum without a truncation nobody has justified from first principles, and the BBM operator's self-adjointness depends on a similarity transformation and boundary conditions whose validity is disputed.

The laboratory echo. Level repulsion of GUE and GOE type shows up in neutron-resonance spectra of heavy nuclei, in microwave billiards, in disordered conductors, and — as Chapter 32 recorded — in the Fisher zeros of quenched spin systems crossing into a dynamical phase transition. The recurrence is real and it is instructive. It also cuts the other way, which the next section is about.

What none of it gives

Every entry above is evidence that the zeros behave like a spectrum in the unitary symmetry class. Not one of them is evidence about which operator. Random-matrix universality is the reason for the gap: GUE statistics follow from symmetry class alone, which is why they appear in nuclei, billiards, and conduction fluctuations that share no substrate whatever. Statistics identify the class. They do not identify the member.

The function-field proof does not transfer. There is no known cohomology over the integers playing the role Frobenius plays there, and constructing one is not a technical gap but the central open problem of the programme. Citing the analogue as support is legitimate; citing it as near-completion is not.

The restricted test functions matter too. Montgomery's and Rudnick-Sarnak's results hold in limited ranges, and extending them is itself hard. The numerics are overwhelming but numerics are not proof, and there are known slow-onset effects in zeta statistics that only appear at heights far beyond computation.

And the laboratory echoes are the weakest entry, not the most exciting one. A microwave cavity showing level repulsion tells you the cavity is chaotic and time-reversal broken. It tells you nothing about arithmetic. The volume has to keep saying this because the temptation runs the other way every time an experiment makes the news.

What would count as payment

It is worth writing down the discharge conditions so the debt cannot quietly be renegotiated. A candidate operator has to be self-adjoint on a genuine Hilbert space, with no similarity transformation smuggling the reality of the spectrum in by hand. Its spectrum must be the imaginary parts of the non-trivial zeros, with the right multiplicities. It must reproduce the counting function asymptotically and in its oscillating corrections. And a trace formula for it should return the explicit formula, primes and all, rather than merely resembling it.

Partial credit is possible and would still be enormous. An operator with the correct spectrum but unclear self-adjointness would be a research programme rather than a proof. A rigorous xp truncation would be a major advance even without the full spectrum. What would not count is another statistical match, however precise: that ledger is already full and it has not moved the problem in fifty years.

The negative outcome deserves the same clarity. It is entirely possible that the zeros are not a spectrum and that the GUE agreement reflects a shared symmetry constraint with no operator behind it at all. That would leave the statistics intact, the analogy intact, and the Hilbert-Polya route dead. Nothing currently known excludes it.

Why the accumulation matters anyway

So: talk about spectral. Six independent lines — correlation theory, numerics at astronomical height, higher correlations across a family, a proved analogue in another arithmetic, a trace formula with primes where orbits go, and a semiclassical Hamiltonian that almost works — all point the same direction, and the thing they point at has never been seen. That is an unusual epistemic situation. It is not the situation of a hunch, and it is not the situation of a result.

The right posture is the one this volume keeps arguing for elsewhere: hold the convergence as convergence and the absence as absence, without letting either dissolve the other. The accumulated reasons make the search rational. They do not make the object real. Chapter 41 made a joke of the missing operator because the debt is old; this chapter itemises why the debt is worth carrying rather than writing off.

There is one more reason, less respectable and worth admitting. If the operator is found, a great deal of what this volume has borrowed stops being analogy and becomes inheritance. That prospect is a motive, and a motive is exactly the thing to name out loud and then discount, because wanting a structure to exist is the standard mechanism by which people come to believe it does.