A Very Large Prime

On 12 October 2024, GIMPS recorded Luke Durant’s discovery of the 52nd known Mersenne prime (first flagged as a probable prime the day before), $2^{136{,}279{,}841} - 1$. Writing it out in decimal requires 41,024,320 digits. GpuOwl on a cloud NVIDIA A100 raised the flag and a Lucas-Lehmer test on an H100 confirmed it; both were among thousands of server GPUs Durant ran in 24 data-centre regions [[1]]. The ranking “52nd” is provisional until every smaller exponent has been checked [[2]].

This is the kind of thing that makes headlines, and it deserves them. Mersenne primes are rare and verifying them is genuinely hard. But if I am honest, the more interesting prime story of the last half-century is not about the record-breaking number. It is about a conversation over tea in Princeton in 1972, and the increasingly hard-to-dismiss suspicion that the zeros of the zeta function, which encode the primes, are in a precise statistical sense quantum energy levels, with the primes themselves in the role of the classical orbits behind those levels.

When I say “quantum energy levels”, I mean more than a metaphor and less than an identity. The title is shorthand, and slightly unfair to the zeros: in the analogy they are the levels, and the primes are the orbits. Two things come close to literal. The local statistics of the zeros match those of a random-matrix model of quantum chaos, in a restricted theorem and in extensive numerical evidence, and the formula that builds the zeros out of the primes has the form of the trace formula that builds energy levels out of classical periodic orbits. What is missing is the quantum system itself. Let me explain.

The Riemann Zeta Function Encodes the Primes

Start with the most famous function in number theory. For $\operatorname{Re}(s) > 1$, the Riemann zeta function is defined by the series

$$\zeta(s) = \sum_{n=1}^{\infty} \frac{1}{n^s}.$$

This converges nicely and defines an analytic function. But the real reason to care about it is Euler’s product formula:

$$\zeta(s) = \prod_{p \text{ prime}} \frac{1}{1 - p^{-s}}.$$

This is not obvious — it follows from unique prime factorisation, essentially — but its implications are enormous. The product runs over all primes, and each prime contributes a factor. The primes are encoded in the analytic structure of $\zeta$. If you know $\zeta$, you know the primes; if you understand the zeros of $\zeta$, you understand their distribution.

Riemann’s 1859 paper made this explicit [[3]]. He showed that $\zeta$ extends analytically to the whole complex plane (minus a simple pole at $s = 1$), and he wrote down an explicit formula connecting the prime-counting function

$$\pi(x) = \#\{p \leq x : p \text{ prime}\}$$

to the zeros of $\zeta$. Strictly, his formula is exact for a weighted count that also includes prime powers, $\Pi(x) = \pi(x) + \tfrac{1}{2}\pi(x^{1/2}) + \tfrac{1}{3}\pi(x^{1/3}) + \cdots$:

$$\Pi(x) = \operatorname{Li}(x) - \sum_{\rho} \operatorname{Li}(x^{\rho}) - \ln 2 + \int_x^\infty \frac{dt}{t\,(t^2-1)\ln t},$$

where $\operatorname{Li}(x)$ is the logarithmic integral, taken as a principal value from 0 as Riemann took it (the version $\int_2^x dt/\ln t$ differs by the constant $\operatorname{Li}(2) \approx 1.045$), and the sum runs over the non-trivial zeros $\rho$ of $\zeta$, each paired with its complex conjugate. $\pi(x)$ itself follows by Möbius inversion, $\pi(x) = \Pi(x) - \tfrac{1}{2}\Pi(x^{1/2}) - \tfrac{1}{3}\Pi(x^{1/3}) - \cdots$. The first correction, about $-\tfrac{1}{2}\operatorname{Li}(\sqrt{x})$, is as large as the terms from the zeros.

What are the non-trivial zeros? The zeta function has trivial zeros at the negative even integers $-2, -4, -6, \ldots$ — boring, understood. The non-trivial zeros lie in the critical strip $0 < \operatorname{Re}(s) < 1$, and their imaginary parts are what drive the oscillatory corrections to the prime counts. The cleanest place to see this is the explicit formula for the Chebyshev function $\psi(x) = \sum_{p^k \leq x} \ln p$, which counts primes and prime powers with logarithmic weights:

$$\psi(x) = x - \sum_{\rho} \frac{x^\rho}{\rho} - \ln(2\pi) - \frac{1}{2}\ln(1 - x^{-2}), \tag{1}$$

valid for $x > 1$, with the zeros summed in conjugate pairs in order of height (at a prime power the right-hand side gives the midpoint of the jump). A zero $\rho = \frac{1}{2} + it_n$ and its conjugate contribute $-2x^{1/2}\cos(t_n \ln x - \arg\rho)/|\rho|$, an oscillation in $\ln x$ with frequency $t_n$; the prime staircase is a superposition of these oscillations, one per pair (Fig. 1).

Figure 1. The zeros rebuild the primes. Each panel shows the Chebyshev function $\psi(x)$, which jumps by $\ln p$ at every prime power $p^k$ (wide grey staircase), and the right-hand side of Eq. (1) truncated after 0, 5, 20 and 100 conjugate pairs of zeros (thin line). The smooth terms give the trend; each pair of zeros adds one oscillation in $\ln x$, and a hundred pairs place every step between 2 and 30, with the overshoot of a truncated Fourier series at each jump. Zeros computed with mpmath.

How many zeros are there? Counting those with $0 < t_n \leq T$ gives the Riemann–von Mangoldt formula

$$N(T) = \frac{T}{2\pi} \ln \frac{T}{2\pi} - \frac{T}{2\pi} + \frac{7}{8} + S(T) + O\!\left(\frac{1}{T}\right), \tag{2}$$

where $S(T) = \pi^{-1} \arg\zeta(\tfrac{1}{2} + iT)$ is a small, irregular fluctuation that averages to zero [[4]]. Riemann gave the first two terms [[3]]. Call the first three terms the smooth count $\bar{N}(T)$; its derivative gives the mean spacing of the zeros at height $t$, $2\pi/\ln(t/2\pi)$.

The Riemann Hypothesis is the claim that all non-trivial zeros lie on the critical line $\operatorname{Re}(s) = \frac{1}{2}$. Gourdon reported in 2004 that the first $10^{13}$ zeros do [[5]], in computations never published in a refereed journal; in 2021 Platt and Trudgian verified rigorously, with interval arithmetic, that all zeros up to height $3 \cdot 10^{12}$ lie on the line [[6]]. It has not been proved. It remains open more than 160 years after Riemann, and many mathematicians regard it as the most important unsolved problem in mathematics.

Tea with Dyson

In 1972, Hugh Montgomery was visiting the Institute for Advanced Study in Princeton. He was working on a specific question: if you take the imaginary parts of the non-trivial zeros of $\zeta$ and normalise them so that their mean spacing is 1, what is the distribution of spacings between them?

More precisely, he was computing the pair correlation function of the normalised zeros. Normalising means unfolding the heights with the smooth count, $\tilde{t}_n = \bar{N}(t_n)$, so that the mean spacing is 1 at every height. The pair correlation function $R_2(r)$ then measures the density of pairs $(\tilde{t}_m, \tilde{t}_n)$ with $\tilde{t}_n - \tilde{t}_m \approx r$.

Montgomery proved, assuming the Riemann Hypothesis, that the Fourier transform of this pair correlation (the form factor) equals $|\tau|$ for $|\tau| < 1$, which is exactly the random-matrix form factor in that range [[4]] [[7]]. Because the theorem constrains only test functions whose Fourier transforms vanish outside that range, it fixes averages of $R_2$ over windows about one mean spacing wide or wider, not the curve point by point. The full formula is his conjecture:

$$R_2(r) = 1 - \left(\frac{\sin \pi r}{\pi r}\right)^2. \tag{3}$$

He mentioned this to Freeman Dyson over tea. Dyson — who had spent years on quantum mechanics and random matrix theory — recognised the formula immediately. That expression, $1 - (\sin \pi r / \pi r)^2$, is exactly the pair correlation function of eigenvalues of large random matrices drawn from the Gaussian Unitary Ensemble.

Montgomery had not been thinking about quantum mechanics. Dyson had not been thinking about primes. The formula matched.

The Gaussian Unitary Ensemble

Let me say a few words about where that formula comes from in physics, because it is not obvious.

The Gaussian Unitary Ensemble (GUE) is a probability distribution over $N \times N$ Hermitian matrices. Specifically, it is the distribution proportional to $e^{-\operatorname{tr}(H^2)}$ on the space of Hermitian matrices, which is invariant under conjugation $H \mapsto U H U^\dagger$ for any unitary $U$. The entries on the diagonal are real Gaussians; the off-diagonal entries are complex Gaussians with independent real and imaginary parts.

In the limit $N \to \infty$, the eigenvalues of a GUE matrix, rescaled by $\sqrt{N}$, distribute globally according to Wigner’s semicircle law. But the local statistics — the fine-grained distribution of spacings between nearby eigenvalues — follow a universal law. After unfolding to unit mean spacing, the pair correlation function is

$$R_2^{\text{GUE}}(r) = 1 - \left(\frac{\sin \pi r}{\pi r}\right)^2,$$

the same expression as Eq. (3).

This distribution has a crucial qualitative feature called level repulsion: as $r \to 0$, $R_2(r) \to 0$. Eigenvalues of random Hermitian matrices strongly avoid each other. A Poisson distribution — which is what you would get for eigenvalues that were statistically independent — would give $R_2(r) = 1$ everywhere, with no such repulsion. The GUE formula suppresses small gaps quadratically: $R_2(r) \sim \pi^2 r^2 / 3$ for small $r$.

Why do random-matrix statistics arise in physics at all? In 1984 Bohigas, Giannoni and Schmit found that the levels of the quantum Sinai billiard, a chaotic system with time-reversal symmetry, follow the Gaussian Orthogonal Ensemble (GOE), and argued that such fluctuation laws are universal [[8]]. The conjecture that carries their names, which by now has overwhelming numerical support, says that quantum systems whose classical limit is chaotic have the level statistics of the random-matrix ensemble with the same symmetry: GOE when time-reversal symmetry holds (for spinless particles), GUE when it is broken. Nuclear energy levels, quantum billiards with the right shapes, molecular spectra — all of them, when appropriately normalised, show the statistics of the matching ensemble, mostly GOE.

The universality is the point. It does not matter what the specific Hamiltonian is. If the system is sufficiently chaotic, the eigenvalue statistics are universal.

Odlyzko’s Computation

Montgomery’s result was conditional and too coarse to fix the curve point by point. The natural next step was numerical verification: actually compute a large number of Riemann zeros and measure their pair correlation.

Andrew Odlyzko did exactly this, in a series of computations beginning in the 1980s. His 1987 paper, on the first $10^5$ zeros and the $10^5$ after zero number $10^{12}$, found the agreement with the random-matrix predictions “generally good”, improving “at larger heights” [[9]]. For the spacings of the first million zeros he later found “a reasonably good fit, but it is far from perfect”; near zero number $2 \times 10^{20}$, “almost perfect!” [[10]]

Odlyzko then pushed further: ten billion zeros near the $10^{22}$-nd zero, at height about $1.37 \times 10^{21}$ [[11]]. Same statistics. The agreement kept improving with height. This is not a small-sample artefact, and I do not think it is a coincidence.

The plots from Odlyzko’s computations are, in my view, some of the most beautiful images in mathematics. You draw the GUE prediction — a smooth curve, starting at zero, rising to approach 1 — and you overlay the empirical histogram from the Riemann zeros. To the eye, they are the same curve.

To the eye. Near the $10^{16}$-th zero the distribution of nearest-neighbour spacings differs from its limit by about one per cent, with a structure Odlyzko could see in his data. Bogomolny and colleagues argued, with a heuristic formula, that to leading order this is the deviation of a random unitary matrix of finite size $N_{\text{eff}} \approx 0.23 \ln(t/2\pi)$: 7.7 there, 11.3 near the $10^{23}$-rd zero [[12]]. The correction falls off like $1/N_{\text{eff}}^2$, that is like $1/\ln^2(t/2\pi)$. The approach to GUE is real but logarithmically slow, and low down the correction is large: among the first 10,000 zeros $N_{\text{eff}}$ is below 2.

Fig. 2 shows both ends. Among the first 10,000 zeros close pairs are far rarer than GUE predicts: between 0.1 and 0.2 mean spacings the pair correlation is $0.037 \pm 0.006$, half the GUE value of 0.074. In the 10,000 zeros just after the $10^{22}$-nd, from Odlyzko’s published table [[13]], $N_{\text{eff}} \approx 10.8$; the same bin gives $0.072 \pm 0.008$, and a correction of about half a per cent is far below what so few zeros can resolve.

Figure 2. Near zero number $10^{22}$ the pair correlation follows the GUE curve within the counting errors; among the first 10,000 zeros close pairs are much rarer than GUE predicts, the five bins below half a mean spacing lying 2.6 to 6.4 standard errors low. Points: histogram (bins 0.1 wide) of the distances between zeros after unfolding to unit mean spacing, with the smooth count of Eq. (2) for zeros 1 to 10,000 (computed with mpmath) and with the local mean density for zeros $10^{22}+1$ to $10^{22}+10^4$ (Odlyzko’s table [[13]]); bars are counting errors. Curve: Eq. (3). Dashed line: uncorrelated (Poisson) levels. Against the curve, $\chi^2$ is 148 for the low zeros and 32 for the high ones, over 30 bins.

Berry, Keating, and the Missing Hamiltonian

If the zeros of $\zeta$ are energy levels, there should be a Hamiltonian $H$ — a self-adjoint operator — whose spectrum is exactly $\{t_n\}$, the imaginary parts of the non-trivial zeros (assuming the Riemann Hypothesis, so that all zeros are of the form $\frac{1}{2} + it_n$).

In 1999, Michael Berry and Jon Keating proposed a candidate [[4]]. Their suggestion was the classical Hamiltonian

$$H_{\text{cl}} = xp,$$

where $x$ is position and $p$ is momentum, quantised in the simplest Hermitian ordering:

$$\hat{H} = \frac{1}{2}(\hat{x}\hat{p} + \hat{p}\hat{x}).$$

Classically, $H = xp$ describes phase-space trajectories $xp = E = \text{const}$, with $x(t) = x_0 e^t$ and $p(t) = p_0 e^{-t}$ — exponential expansion in position and contraction in momentum. It is a hyperbolic dilation, not by itself a conventional bounded chaotic system. It does have one property the zeros demand: because $\dot{x} = x$ ties the velocity to the position, the motion cannot be run backwards, so there is no time-reversal symmetry [[4]].

The semiclassical estimate of the number of states below energy $E$ is the phase-space area under the hyperbola $xp = E$, in units of the Planck cell $2\pi\hbar$ (Berry and Keating set $\hbar = 1$). The area is infinite, so they cut it off by hand at $x \geq \ell_x$ and $p \geq \ell_p$, with $\ell_x \ell_p = 2\pi$, one Planck cell [[4]]. Then

$$\frac{A(E)}{2\pi} = \frac{1}{2\pi}\int_{\ell_x}^{E/\ell_p} \left(\frac{E}{x} - \ell_p\right) dx = \frac{E}{2\pi} \ln \frac{E}{2\pi} - \frac{E}{2\pi} + 1.$$

Berry and Keating add a Maslov correction $\alpha/4\pi$ for the angle $\alpha$ through which the orbit turns in phase space, and state that the result has the constant $7/8$. I get that from the quarter turn, $\alpha = \pi/2$, if the correction subtracts one eighth:

$$N(E) \approx \frac{E}{2\pi} \ln \frac{E}{2\pi} - \frac{E}{2\pi} + \frac{7}{8},$$

which is the smooth part of Eq. (2) with $T$ replaced by $E$. Berry and Keating call closing the phase space to make the motion bounded “a central unsolved problem”.

The problem is that $\hat{H} = \frac{1}{2}(\hat{x}\hat{p} + \hat{p}\hat{x})$ as an operator on $L^2(\mathbb{R})$ is not bounded below and has a continuous spectrum, not a discrete one. Turning it into an operator with a discrete spectrum matching the Riemann zeros requires boundary conditions that have not been found. This is the crux: Berry and Keating have a classical system with the right smooth counting and the right lack of time-reversal symmetry, but its trajectories are unbounded and it has no real periodic orbits, and the quantum boundary conditions that would supply both are missing. A recent proposal by Bender, Brody and Müller does not close the gap: their operator, with classical limit $2xp$, has the zeros as eigenvalues under a suitable boundary condition, is not Hermitian in the conventional sense, and would imply the Riemann Hypothesis if shown rigorously to be self-adjoint [[14]].

What would be profound about finding $\hat{H}$? If $\hat{H}$ is self-adjoint, its eigenvalues are all real numbers. If those eigenvalues are the imaginary parts of the zeros, then all zeros have real part exactly $\frac{1}{2}$ — which is the Riemann Hypothesis. A proof of the existence of such a Hamiltonian would, in one stroke, resolve the most important open problem in mathematics.

Primes as Periodic Orbits: The Gutzwiller Analogy

The quantum chaos connection goes deeper than pair correlations. In semiclassical quantum mechanics, the Gutzwiller trace formula relates the density of quantum energy levels to a sum over classical periodic orbits:

$$d(E) = \bar{d}(E) + \frac{1}{\pi\hbar}\sum_{\gamma} A_\gamma \cos\!\left(\frac{S_\gamma(E)}{\hbar} - \phi_\gamma\right), \tag{4}$$

where the sum runs over all classical periodic orbits $\gamma$ (each primitive orbit and its repetitions), $S_\gamma$ is the classical action of the orbit, $\phi_\gamma$ is a phase (Maslov index correction), and the amplitude $A_\gamma$ is the period of the primitive orbit divided by the square root of a stability determinant. For an unstable orbit traversed $m$ times it falls off like $e^{-m\lambda T/2}$, with $T$ the period of the primitive orbit and $\lambda$ its Lyapunov exponent [[4]]. The smooth part $\bar{d}(E)$ comes from the Thomas-Fermi approximation; the oscillatory part encodes quantum interference between orbits.

In an earlier version of this post I called Eq. (1) the trace-formula analogue; it is the dual. Eq. (1) sums over zeros to produce the primes; the analogue sums over primes to produce the zeros. Write $d(t)$ for the density of zeros, a sum of delta functions at the heights $t_n$. Differentiating Eq. (2), and expanding $S(t) = \pi^{-1}\operatorname{Im}\ln\zeta(\tfrac{1}{2} + it)$ with the Euler product, gives [[4]]

$$d(t) = \frac{1}{2\pi}\ln\frac{t}{2\pi} + O\!\left(\frac{1}{t^2}\right) - \frac{1}{\pi}\sum_{p}\sum_{m=1}^{\infty} \frac{\ln p}{p^{m/2}} \cos(t\, m \ln p). \tag{5}$$

The Euler product does not converge on the critical line, so the sum diverges as written (“divergent but formally exact”, in Berry and Keating’s words). Averaged over any smooth window in $t$, it converges and is exact, and that is how Fig. 3 is computed.

Comparing Eqs. (4) and (5) term by term: the heights $t_n$ play the role of the quantum energy levels $E_n$; the primes $p$ — and their prime powers $p^m$ — play the role of the classical periodic orbits $\gamma$. With $\hbar = 1$, the “action” of the orbit corresponding to $p^m$ is $t \cdot m\ln p$, so its period is $m \ln p$. The primes are the primitive periodic orbits; $p^m$ is the $m$-th traversal of that orbit. The amplitude $\ln p / p^{m/2}$ is the primitive period times $e^{-m\ln p/2}$, which corresponds to a Lyapunov exponent of exactly 1 for every orbit [[4]].

The prefactor is a success of the analogy. It is $1/\pi$, not $2/\pi$; with time-reversal symmetry each orbit would interfere coherently with its time-reversed partner and double the sum, so Eq. (5), like the GUE statistics, points to a system without time-reversal symmetry [[4]]. Berry and Keating name two features that do not fit. The exponential decay of long orbits is exact for the zeros but only an approximation to the stability determinant in Eq. (4). And the overall minus sign requires a Maslov phase of $\pi$ for every orbit, which is hard to understand, since an orbit traversed twice should then pick up $2\pi$ [[4]].

Fig. 3 runs Eq. (5) in the direction of the analogy, from primes to zeros.

Figure 3. The primes rebuild the zeros. Each panel shows the density of zeros, in zeros per unit $t$, averaged over a Gaussian window of standard deviation $\sigma = 0.4$ in $t$ and computed from Eq. (5) with only the prime powers up to the stated bound; dotted lines mark the zeros $t_n$. Seven prime powers put broad maxima near most of the zeros, 35 resolve all thirteen below $t = 60$, and with all 78,734 prime powers below $10^6$ the curve matches the same average taken directly over the zeros to within $10^{-5}$. The negative dips in the upper panels are truncation artefacts.

Odlyzko saw Eq. (5) at work in his data: a Fourier sum over 40,000 zeros near the $10^{20}$-th zero, which would be small for random numbers, spikes at $\ln p^m$ [[10]].

This is a precise structural analogy, not an identification of the Riemann zeros with a known physical spectrum. The Selberg trace formula makes an analogous orbit–spectrum correspondence rigorous: for waves on a compact surface of constant negative curvature the periodic-orbit sum is exact [[4]]. It does not make the Riemann zeta function a limit of those systems.

I find it remarkable that the logarithms of primes — the most elementary sequence in arithmetic — appear as the periods of orbits in what would be a quantum chaotic system. Each prime contributes an oscillation to the density of zeros with “frequency” $\ln p$ (Fig. 3). You are, in a sense, hearing the primes as quantum interference.

This connects to a theme that comes up elsewhere on this blog. The falling cat problem involves Berry phase and geometric holonomy — again a situation where deep structure emerges from symmetry and topology. The Schrödinger cat in quantum computing involves the spectacular fragility of quantum coherence. The Riemann zeros would be the spectrum of a quantum system only if a suitable Hilbert–Pólya operator were found; that remains conjectural.

A Brief Detour: Maynard and Primes Without Digits

While we are talking about primes, I cannot resist a detour through two results of James Maynard, who received the Fields Medal in 2022.

The first concerns bounded gaps. Euclid proved that there are infinitely many primes. The Twin Prime Conjecture says there are infinitely many pairs of primes $(p, p+2)$. This remains open. But in 2013, Yitang Zhang proved something extraordinary: there are infinitely many pairs of primes differing by at most 70,000,000 [[15]]. The bound is large, but the qualitative statement — that gaps between primes are bounded infinitely often — was completely new. Shortly thereafter, Maynard proved a much stronger result with a refinement of the Goldston–Pintz–Yıldırım sieve: infinitely many prime pairs with gap at most 600 [[16]]. Terence Tao proved it independently at much the same time, hence the name Maynard-Tao sieve. A crowdsourced effort (Polymath8b) brought the bound down to 246 [[17]]. The Twin Prime Conjecture remains open, but 246 is a long way from 70,000,000.

The second result is stranger. In a preprint of 2016, published in 2019, Maynard proved that for any decimal digit $d \in \{0, 1, \ldots, 9\}$, there are infinitely many primes whose decimal representation contains no instance of $d$ [[18]]. There are infinitely many primes with no $7$ in their decimal expansion. There are infinitely many primes with no $3$. The proof combines the Hardy–Littlewood circle method with Harman’s sieve, and the method extends to other sufficiently large bases.

This is one of those results that sounds impossible on first hearing. Surely removing an entire digit should make most large numbers unavailable, so the primes run out? Not so. Below $X$ there are only about $X^{\ln 9/\ln 10} \approx X^{0.954}$ numbers without a given digit, a vanishing fraction of all numbers, but still far too many for their sparseness alone to exclude the primes.

The 52nd Mersenne Prime and What We Do Not Know

Return to $M_{136{,}279{,}841} = 2^{136{,}279{,}841} - 1$. Mersenne primes have the form $2^p - 1$ where $p$ is a prime (though not all such numbers are prime — $2^{11} - 1 = 2047 = 23 \times 89$). They are proved prime by the Lucas-Lehmer test: define the sequence

$$s_0 = 4, \qquad s_{n+1} = s_n^2 - 2.$$

Then, for an odd prime $p$, $M_p = 2^p - 1$ is prime if and only if $s_{p-2} \equiv 0 \pmod{M_p}$.

The test requires $p - 2$ squarings modulo $M_p$. Each squaring involves numbers of $p$ binary digits — 136 million bits, or 41 million decimal digits, for the new prime — and reduction modulo $M_p = 2^p - 1$ is cheap because $2^p \equiv 1$: the high half of a product folds back onto the low half with a shift and an addition. The squarings must run one after another, but each one is done by fast Fourier transform multiplication [[19]], an algorithm that parallelises well; that is why GPUs help so much. Durant’s cloud network was, in effect, a massively distributed modular arithmetic engine. GIMPS clients now search with a Fermat probable-prime test and keep the Lucas-Lehmer test for confirmation [[1]].

We do not know if there are infinitely many Mersenne primes. The heuristic Lenstra-Pomerance-Wagstaff conjecture says yes: the expected number of Mersenne primes $2^p - 1$ with $p \leq x$ is approximately

$$e^\gamma \ln x / \ln 2 \approx 1.78 \cdot \log_2 x,$$

where $\gamma \approx 0.5772$ is the Euler-Mascheroni constant [[20]]. This predicts roughly logarithmic growth in the count — for $p \leq 136{,}279{,}841$ it gives about 48, against the 52 known examples — but is nowhere near proved.

The known Mersenne primes do not form a sequence with obviously regular gaps. The exponents $p$ are: 2, 3, 5, 7, 13, 17, 19, 31, 61, 89, 107, 127, … and then larger, less predictable values. Whether their distribution has GUE-like statistics is not a standard research question (the sample is too small), but the question of whether the primes $p$ for which $2^p - 1$ is prime have any special structure is an active one. For now, the answer is: we do not know.

Why This Matters, and Why It Does Not Prove Anything

Let me be precise about what has and has not been established.

What has been established:

  • Under the Riemann Hypothesis, Montgomery proved a restricted pair-correlation result; the full GUE sine-kernel formula remains conjectural.
  • Odlyzko’s computations show the statistics approaching GUE with height, with deviations that shrink like $1/\ln^2 t$ [[11]] [[12]].
  • The Gutzwiller/Selberg analogy between periodic orbits and primes is mathematically precise in related settings (hyperbolic surfaces, and function fields over finite fields, where the zeros are eigenvalues of a Frobenius operator [[21]]).
  • Rudnick and Sarnak extended Montgomery’s theorem to all $n$-level correlations, for $\zeta$ and other principal $L$-functions, again under restricted Fourier support; it is not a proof of full GUE statistics for the Riemann zeros [[22]].

What has not been established:

  • There is no known Hamiltonian $\hat{H}$ whose spectrum is the set of Riemann zeros.
  • The Riemann Hypothesis remains open.
  • There is no proof that a self-adjoint operator has the Riemann zeros as spectrum, or that full GUE statistics hold for them.

The broader context is the Langlands programme — a still-hypothetical grand unification of number theory, algebraic geometry, and representation theory, sometimes described as a “grand unified theory of mathematics.” The Langlands correspondence predicts deep connections between $L$-functions (generalisations of $\zeta$) and representations of algebraic groups. The spectral interpretation of Riemann zeros — if it could be made precise — would fit naturally into this framework. Some researchers believe that a proof of the Riemann Hypothesis will come from the Langlands side, not from analytic number theory or quantum mechanics. Others think the quantum chaos connection is the right road. Nobody knows.

What would it mean if the Hilbert–Pólya programme succeeded? It would give a spectral route to the Riemann Hypothesis. It would not make prime numbers themselves the eigenvalues: the proposed spectrum is the set of imaginary parts of non-trivial zeta zeros. The primes would be the periodic orbits.

Or it could be a very surprising coincidence. Mathematics has a long history of producing such coincidences — the same numbers appearing in unrelated contexts for reasons that, when understood, turned out not to be coincidences at all. I suspect this is not a coincidence. But suspicion is not proof.

A Closing Reflection

I started this post with the 52nd Mersenne prime because it is the news item that prompted me to write. GPU clusters finding 41-million-digit primes are genuinely impressive technology. But I keep returning to the image of Montgomery and Dyson at tea in 1972, and the formula $1 - (\sin \pi r / \pi r)^2$ connecting two conversations that had nothing to do with each other.

I have spent some time with random matrix theory, and separately with the zeta function, and the thing that still strikes me is how clean the connection is. This is not a numerical coincidence of the form “these two quantities agree to 3 decimal places.” Odlyzko’s plots show agreement across many orders of magnitude of height, for zeros computed more than $10^{22}$ entries into the sequence. The GUE curve and the empirical histogram are, visually, the same curve.

As someone trained as a physicist, I find this both encouraging and slightly unsettling. The zeros have a structure that numerically resembles eigenvalue repulsion in quantum-chaotic systems, and that structure might point towards a proof of the Riemann Hypothesis. But a numerical resemblance is not yet a Hilbert space, a Hamiltonian or a physical system.

If you find a flaw in this picture, or know of a result I have missed, I am genuinely interested. Peer review is welcome — open an issue on GitHub.

Literature and calculations checked through 2026-10-24.


References

[1] Great Internet Mersenne Prime Search (2024, October 21). GIMPS discovers largest known prime number: 2^136,279,841-1. Retrieved 24 October 2026 from https://www.mersenne.org/primes/?press=M136279841

[2] Great Internet Mersenne Prime Search. List of known Mersenne prime numbers. Retrieved 24 October 2026 from https://www.mersenne.org/primes/

[3] Riemann, B. (1859). Ueber die Anzahl der Primzahlen unter einer gegebenen Grösse. Monatsberichte der Berliner Akademie, November 1859. Transcription and English translation by D. R. Wilkins: https://www.maths.tcd.ie/pub/HistMath/People/Riemann/Zeta/

[4] Berry, M. V., & Keating, J. P. (1999). The Riemann zeros and eigenvalue asymptotics. SIAM Review, 41(2), 236–266. DOI: 10.1137/S0036144598347497

[5] Gourdon, X. (2004). The $10^{13}$ first zeros of the Riemann zeta function, and zeros computation at very large height. Unpublished manuscript. http://numbers.computation.free.fr/Constants/Miscellaneous/zetazeros1e13-1e24.pdf

[6] Platt, D., & Trudgian, T. (2021). The Riemann hypothesis is true up to $3 \cdot 10^{12}$. Bulletin of the London Mathematical Society, 53(3), 792–797. DOI: 10.1112/blms.12460

[7] Montgomery, H. L. (1973). The pair correlation of zeros of the zeta function. Analytic Number Theory, Proc. Symp. Pure Math., 24, 181–193. DOI: 10.1090/pspum/024/9944

[8] Bohigas, O., Giannoni, M. J., & Schmit, C. (1984). Characterization of chaotic quantum spectra and universality of level fluctuation laws. Physical Review Letters, 52(1), 1–4. DOI: 10.1103/PhysRevLett.52.1

[9] Odlyzko, A. M. (1987). On the distribution of spacings between zeros of the zeta function. Mathematics of Computation, 48(177), 273–308. DOI: 10.2307/2007890

[10] Odlyzko, A. M. (1990). Primes, quantum chaos, and computers. In Number Theory (pp. 35–46). National Research Council. Author’s copy

[11] Odlyzko, A. M. (2001). The $10^{22}$-nd zero of the Riemann zeta function. In M. van Frankenhuysen & M. L. Lapidus (Eds.), Dynamical, Spectral, and Arithmetic Zeta Functions (Contemporary Mathematics 290, pp. 139–144). American Mathematical Society. DOI: 10.1090/conm/290/04578

[12] Bogomolny, E., Bohigas, O., Leboeuf, P., & Monastra, A. G. (2006). On the spacing distribution of the Riemann zeros: corrections to the asymptotic result. Journal of Physics A: Mathematical and General, 39(34), 10743–10754. DOI: 10.1088/0305-4470/39/34/010

[13] Odlyzko, A. M. Tables of zeros of the Riemann zeta function (zeros number $10^{22} + 1$ through $10^{22} + 10^4$). Retrieved 24 October 2026 from https://www-users.cse.umn.edu/~odlyzko/zeta_tables/index.html

[14] Bender, C. M., Brody, D. C., & Müller, M. P. (2017). Hamiltonian for the zeros of the Riemann zeta function. Physical Review Letters, 118(13), 130201. DOI: 10.1103/PhysRevLett.118.130201

[15] Zhang, Y. (2014). Bounded gaps between primes. Annals of Mathematics, 179(3), 1121–1174. DOI: 10.4007/annals.2014.179.3.7

[16] Maynard, J. (2015). Small gaps between primes. Annals of Mathematics, 181(1), 383–413. DOI: 10.4007/annals.2015.181.1.7

[17] Polymath, D. H. J. (2014). Variants of the Selberg sieve, and bounded intervals containing many primes. Research in the Mathematical Sciences, 1, 12. DOI: 10.1186/s40687-014-0012-7

[18] Maynard, J. (2019). Primes with restricted digits. Inventiones Mathematicae, 217(1), 127–218. DOI: 10.1007/s00222-019-00865-6

[19] Great Internet Mersenne Prime Search. The math. Retrieved 24 October 2026 from https://www.mersenne.org/various/math.php

[20] PrimePages. Heuristics: Deriving the Wagstaff Mersenne conjecture. Retrieved 24 October 2026 from https://t5k.org/mersenne/heuristic.html

[21] Katz, N. M., & Sarnak, P. (1999). Zeroes of zeta functions and symmetry. Bulletin of the American Mathematical Society, 36(1), 1–26. DOI: 10.1090/S0273-0979-99-00766-1

[22] Rudnick, Z., & Sarnak, P. (1996). Zeros of principal L-functions and random matrix theory. Duke Mathematical Journal, 81(2), 269–322. DOI: 10.1215/S0012-7094-96-08115-6


Changelog

  • 2026-02-17: Corrected the date of the Montgomery-Dyson meeting from 1973 to 1972 (the paper was published in the 1973 proceedings volume, but the meeting at the IAS took place in April 1972).
  • 2026-07-11: Distinguished Montgomery’s restricted theorem from the full pair-correlation conjecture, corrected the Riemann–von Mangoldt counting formula, and narrowed the Berry–Keating and Selberg claims to their established scope.
  • 2026-10-03: Corrected the Hilbert-Pólya argument (a self-adjoint operator needs only real eigenvalues, not a non-negative spectrum, to force the zeros onto the critical line); fixed the description of the explicit formula for $\psi(x)$, which is a formula for the Chebyshev function, not the zero-counting function; limited Maynard’s restricted-digit result to base 10 and other sufficiently large bases (it cannot hold in every base, e.g. base 2); dated the 52nd Mersenne prime as GIMPS does (12 October 2024, Lucas-Lehmer confirmation, after the 11 October probable-prime flag).
  • 2026-10-24: Corrected the trace-formula analogy: its number-theoretic counterpart is the density of zeros as a sum over prime powers (new Eq. (5), Fig. 3), the dual of the explicit formula for $\psi(x)$, not that formula itself. Replaced the approximate formula for $\pi(x)$ by Riemann’s exact formula for $\Pi(x)$, with $\operatorname{Li}$ taken from 0. Corrected Odlyzko’s 1987 data ($10^5$ zeros at each of two heights; agreement “generally good”) and separated zeros checked from the start (Gourdon’s unrefereed $10^{13}$; Platt and Trudgian, rigorously, to height $3 \cdot 10^{12}$) from windows computed far up; “billions of entries” became more than $10^{22}$. Corrected the Bohigas-Giannoni-Schmit result (GOE, with time-reversal symmetry), the restriction in Montgomery’s theorem (on the Fourier variable, not on $r$), Rudnick and Sarnak’s scope (individual principal $L$-functions), $p$ digits to $p$ bits, $xp$ as “the right classical system”, and the GIMPS hardware. Added the finite-height deviations from GUE (Fig. 2), the Berry-Keating derivation of the $7/8$, the reconstruction of $\psi(x)$ from zeros (Fig. 1) and the Bender-Brody-Müller proposal.