Zeta Zeros
Display
Theme
Motion

Statistics of the zeros

Statistics of the zeros

Zeta zeros repel each other. Their spacings, pair correlation and number variance match the eigenvalues of large random Hermitian matrices (GUE) more closely the higher you go. That match is a tested conjecture, not a theorem.

  • Settled standard mathematics, or computed here on a stated range
  • Open a conjecture or an unfinished item

Figures on this page

Every thumbnail is computed from real data or cited tables. Click to jump to the full figure and its provenance caption.

Evidence, not a proof

Three proved or computed layers, and one statistical layer. Each stops short of the Hypothesis. Read the flags. “Verified up to a height” means a finite check. “A positive proportion” means a theorem about a fraction, not about every zero. “GUE” is a conjecture about spacings, tested numerically.

Numerical verification

Settled Proved up to a height Platt and Trudgian, Bulletin of the London Mathematical Society 53 (2021), using interval arithmetic: every nontrivial zero with imaginary part at most 3,000,175,332,800 has real part 1/2, and those zeros are simple. That is the first 12,363,153,437,138 nontrivial zeros. Earlier computations reached comparable heights. This peer-reviewed bound is the one this page relies on. The Riemann–von Mangoldt formula says there are infinitely many further zeros. A check, however high, is evidence. It is not a proof.

Zeros forced onto the line, but not all of them

Settled Proved partial results The published chain, each step unconditional:

  • Hardy, 1914: infinitely many zeros lie on the critical line.
  • Selberg, 1942: a positive proportion of them do.
  • Levinson, 1974: more than one third.
  • Conrey, 1989: more than two fifths.
  • Pratt, Robles, Zaharescu, and Zeindler, Research in the Mathematical Sciences 7 (2020): more than five twelfths.

Two 2026 preprints, Alpöge–Furman (arXiv:2608.13637, August 2026) and Lamzouri (arXiv:2609.02882, September 2026), claim that more than 67.25% of the nontrivial zeros are simple and lie on the critical line. They are weeks old as of this page. This site does not rest the evidence stack on them. Even a proved two thirds would leave the Hypothesis open: a lower density is not “every zero.”

Zero-free regions

Settled Proved, and short of the line There is a classical region, going back to de la Vallée Poussin, in which ζ(σ + it) ≠ 0 for

σ ≥ 1 − c / log(|t| + 2), for some c > 0.

A wedge shrinking toward the line Re(s) = 1. It never reaches Re(s) = 1/2.

Vinogradov and Korobov, independently in 1958, widened that region to the shape

σ ≥ 1 − c / ((log(|t|+2))^(2/3) (log log(|t|+3))^(1/3)).

Still a neighborhood of the line Re(s) = 1, not of the critical line. The constant c is positive and absolute. Record explicit constants are drawn in the map below.

That region is why the best unconditional error in the prime number theorem has the shape x exp(−c (log x)3/5 / (log log x)1/5), for some c > 0. Much larger than the square-root error RH would give. The zero-free region is real progress. It is not the Hypothesis.

How thin is the zero-free region? A map at true scale

The formulas above say “some constant c”. Here are the record explicit constants, drawn to scale, so you can see how much of the strip is actually ruled out. Picture the strip as a road 1 unit wide, with the critical line painted down the middle. A zero-free region is a stretch of road where we have proved no zero can sit.

Figure 1 How thin the proved zero-free region is

Two panels. Top: the critical strip drawn sideways at true scale, heights on a log scale from 2 to 10 to the 30. Up to 3 times 10 to the 12 the strip is shaded except the middle line, because every zero there has been checked to lie on it. Above that, only a hairline along the edge Re s = 1 is proved zero-free: 0.0063 wide at 3 times 10 to the 12, 0.63% of the strip, shown again in a 25 times zoom. Bottom: the width of the proved band against height on doubly logarithmic axes. The classical region, width 1 over 5.558691 log t, is wider until log t is about 8928; after that the Vinogradov–Korobov region is wider. Both stay far below the width 1/2 that would reach the critical line.
Drawn 12 November 2026 PT from published constants; nothing here is a new bound. Classical region: no zeros with σ ≥ 1 − 1/(5.558691·log t) for t ≥ 2; Vinogradov–Korobov region: no zeros with σ ≥ 1 − 1/(55.241·(log t)2/3(log log t)1/3) for t ≥ 3 (both Mossinghoff, Trudgian and Yang, 2024). Checked range: every zero with 0 < t ≤ 3·1012 is on the line (Platt and Trudgian, 2021). We evaluated the formulas with mpmath 1.4.1 at 40 digits and again in ordinary floating point (largest relative difference 4e-16). The crossing point was found two ways (a root-finder, and the fixed-point form log t = (55.241/5.558691)3·log log t): log t = 8928.20 both times, matching the paper’s remark that the classical region is wider for t < e8928. The dashed curve uses 4.896 from a 2026 preprint (Bellotti, Trudgian and Yang). Fade-in is off if your system asks for reduced motion.

What the picture shows, in numbers:

  • Settled published theorem, evaluated here Near the first zero, t ≈ 14.1347, the classical region only reaches down to σ ≈ 0.932. The zero itself is at σ = ½, far outside it. At t = 1000 the band is 0.0260 wide; at t = 106, 0.0130.
  • Settled published theorem, evaluated here Just above the checked range, at t = 3·1012, the band is 0.00626 wide. Counting its mirror image near σ = 0 (the functional equation reflects zeros across the line), 1.25% of the strip is ruled out there. The other 98.75% is open.
  • Settled published theorems, crossing computed here The band keeps shrinking, like 1/log t. The Vinogradov–Korobov shape shrinks more slowly, but its constant is ten times bigger, so it only wins beyond log t ≈ 8928, a height with about 3877 digits. With Ford’s older constant 57.54 the crossing is at log t ≈ 10242.
  • Open the Riemann Hypothesis RH says the zero-free region is the whole strip except the line: width ½ on each side, the red dashed level in panel B. No proved region reaches any fixed distance from σ = 1. Not even “no zeros with σ > 0.999” is known.

GUE, Montgomery, Odlyzko

Open Conjectural statistics, with numerical support Montgomery, in 1973, conjectured that the pair correlation of the zeros matches the Gaussian Unitary Ensemble. Dyson recognized the match. For the normalized spacing variable u the conjectured pair correlation is

1 − (sin(πu) / (πu))².

Montgomery’s pair correlation. It vanishes at u = 0, so very close pairs are rare. That vanishing is the precise content of “level repulsion.” The usual statement assumes the zeros are on the critical line.

Nearest-neighbor spacings in the GUE are modeled by the Wigner surmise

p(s) = (32/π²) s² exp(−4s²/π).

GUE Wigner surmise for the gap distribution, after the gaps are scaled to mean 1. A model of local spacing, not a theorem about ζ.

The mean gap between ordinates near height T is about 2π / log T. Comparisons have to be made in that normalization. Odlyzko, Mathematics of Computation 48 (1987), computed the first 105 zeros and 105 zeros near the 1012-th, to about 10−8, and found that the spacing statistics approach the GUE prediction as the height grows. Later computations of his, around height 1020 and above, sharpened the same picture. They are evidence for a conjecture about how zeros repel, among zeros already known to lie on the line at those heights. They are not a proof that every zero has real part 1/2.

The histogram below is drawn from Odlyzko’s published table, not from a simulation. Unfolding: each ordinate γ is mapped to (γ/2π) log(γ/2πe) + 7/8, the smooth count above, so consecutive differences have mean 1 at every height.

Figure 2 Spacing histogram Settled

Histogram of normalized spacings between consecutive zeta zeros, from 0 to 3. Pale teal bars rise from almost nothing near zero to a peak near 1 and fall off by about 2.5. A dark teal curve, the GUE Wigner surmise, follows the bars closely. A dashed gray curve, the Poisson law for uncorrelated points, starts at 1 at spacing zero and decays, matching badly. A gold outline for only the first 1000 zeros is jagged but has the same shape.
Data: the first 100 000 zeros from Odlyzko’s tables (file zeros1, heights 14.13 to 74 920.83, stated accuracy 3·10−9), downloaded 6 October 2026 PT. Its first 1000 ordinates agree with an independent mpmath.zetazero computation (see checkpoints). 99 999 spacings, bin width 0.1. Mean 1.000; variance 0.161, against 0.178 for the Wigner surmise and 1 for Poisson. Spacings below 0.25: 1.20% observed, 1.61% for the surmise, 22.1% for Poisson. The largest gap between the empirical and surmise distribution functions is 0.019; against Poisson it is 0.298. The smallest unfolded spacing is 0.022. The first 1000 zeros alone (gold) are noisier and a little narrower. Settled Level repulsion is plainly visible. Honest caveats: the Wigner surmise is a 2×2-matrix approximation to the exact GUE law; low zeros approach the GUE law only slowly as the height grows, with finite-height corrections analysed by Bogomolny, Bohigas, Leboeuf and Monastra (J. Phys. A 39, 2006); and spacings are not independent, so no significance test is claimed. Open GUE for ζ is conjectural, and this histogram is about spacing among zeros already known to be on the line. It is not a proof of the Hypothesis.

Pair correlation is the statistic Montgomery actually conjectured about. Take every pair of unfolded zeros, not just neighbours, and histogram their separations. Montgomery proved, assuming RH, that it matches 1 − (sin πu/πu)² for test functions whose Fourier transform is supported in (−1, 1). Beyond that range it is conjecture.

Figure 3 Pair correlation Open

Graph of pair correlation against normalized separation u from 0 to 3. A teal step line for the first 100,000 zeros rises from almost zero at u = 0, passes 0.5 near u = 0.45, overshoots slightly near u = 1 and then hovers near 1. A smooth black curve, Montgomery's formula, follows it closely. Gold and red dots for zeros near the 10 to the 12th and 10 to the 22nd zero scatter around the same curve.
Computed 6 October 2026 PT from Odlyzko’s tables (index). Teal: all 250 810 pairs with unfolded separation below 3 among the first 100 000 zeros, in bins of width 0.05. In the first bin, [0, 0.05], the value is 0.0022, against 0.0028 for Montgomery’s curve averaged over the bin. Dots: 10 000 zeros from the 1012+1st (gold) and from the 1022+1st (red), in bins of 0.1. Each window is unfolded by the local density log(t/2π)/2π. Root-mean-square distance from the curve: 0.028 low down, 0.016 near 1012, and 0.030 near 1022. The larger 1022 figure is comparable to counting noise, roughly ±0.03 per bin for 10 000 zeros. The low zeros show a small systematic wiggle: about 0.05 above the curve near u ≈ 0.9 and about 0.04 below near u ≈ 1.35. Bogomolny and Keating (Phys. Rev. Lett. 77, 1996) gave a heuristic, prime-based correction that accounts for finite-height wiggles of this kind. Open Montgomery’s full conjecture is open. This is evidence about how zeros already on the line repel each other. It does not place any zero on the line.

The same nearest-neighbour test can be run far higher up. Odlyzko published 10 000 consecutive zeros starting at the 1012+1st (zeros3, guaranteed to 10−8) and at the 1022+1st (zeros5, which the file describes as “not guaranteed, and probably accurate to within 10−6”). There the mean raw gap is 0.257 and 0.134. Each gap is multiplied by the local density log(t/2π)/2π, so the unfolded mean is 1.000.

Figure 4 Spacings at 1012, 1022 Open

Three side-by-side histograms of normalized spacings between consecutive zeta zeros: the first 100,000 zeros, 10,000 zeros from the 10 to the 12th, and 10,000 zeros from the 10 to the 22nd. Each has a teal curve for the exact GUE prediction and a dashed curve for Poisson. The low-zero histogram is slightly taller and narrower than the GUE curve. The two high-zero histograms follow the GUE curve more closely.
Computed 6 October 2026 PT. The teal curve here is the exact GUE spacing density, computed from the Fredholm determinant of the sine kernel by Bornemann’s quadrature method (Math. Comp. 79, 2010; 40 nodes). Its variance comes out as 0.180, and it differs from the Wigner surmise by at most 0.0016 in distribution function. Variance of the unfolded spacings: 0.161 for zeros 1…100 000, 0.177 near 1012, 0.176 near 1022, against 0.180 for GUE and 1 for Poisson. Largest gap between empirical and GUE distribution functions: 0.021, 0.0064, 0.0084; against Poisson it is about 0.28–0.30 in every window. Spacings below 0.25: 1.20%, 1.58%, 1.58%, against 1.63% for GUE. Smallest spacing: 0.022, 0.022, 0.066. The high windows sit much closer to GUE than the low zeros do. That is the trend Odlyzko reported, and it points the same way as the finite-height corrections of Bogomolny et al. cited above. With 10 000 spacings, distribution gaps below about 0.01 are within ordinary sampling scatter. Spacings are not independent, so no formal test is claimed. Open GUE for ζ is a conjecture. These windows are finite. The Hypothesis is untouched.

Spacings and pair correlation look at short distances. Number variance looks at long ones. Drop a window holding L zeros on average at a random place, count the zeros inside, and take the variance of that count. For uncorrelated points it equals L. For GUE eigenvalues it grows only like (1/π²) log L, because repulsion keeps the count unusually steady. Berry (Nonlinearity 1, 1988), treating the zeros as if they were quantum energy levels, predicted that zeta zeros follow GUE for windows much shorter than Lmax = log(E/2π)/(2π log 2) at height E, and stop growing beyond it, oscillating about (1/π²)(log log(E/2π) + 1.4009). The primes set that ceiling. That is a heuristic prediction, not a theorem.

Figure 5 Number variance Settled

Line graph of number variance against window length L from 0 to 10. A dashed Poisson line climbs steeply and leaves the chart at L = 0.8. A black GUE curve rises quickly and then slowly, reaching 0.579 at L = 10. A red curve for zeros near the 10 to the 22nd zero tracks the GUE curve closely and ends at 0.541. A gold curve for zeros near the 10 to the 12th ends at 0.506. A teal curve for the first 100,000 zeros bends away early, levels off near 0.40 around L = 4 to 5, and dips to 0.339 at L = 10.
Computed 6 October 2026 PT from Odlyzko’s tables, unfolded with the smooth zero count (θ(t)/π for the high windows, mpmath at 40 digits). Each point averages 200 000 random window positions. GUE curve: L − 2∫0L(L − r)(sin πr/πr)² dr, computed by quadrature; at L = 10 it is 0.5793, matching its large-L form (1/π²)(log 2πL + γ + 1) = 0.5793. Values at L = 1, 2, 5, 10: first 100 000 zeros 0.321, 0.373, 0.406, 0.339; near 1012 0.340, 0.402, 0.470, 0.506; near 1022 0.342, 0.412, 0.488, 0.541; GUE 0.344, 0.416, 0.509, 0.579; Poisson 1, 2, 5, 10. Berry’s scale Lmax works out to about 2.0 for the low zeros (at the median height, about 40 434), 5.6 near 1012 and 10.8 near 1022. That is the order in which the curves peel away from GUE. His predicted plateau for the low zeros is about 0.36–0.37; the teal curve peaks at 0.407 (L = 4.3) and averages 0.383 over 3 ≤ L ≤ 10. No fit is claimed. Settled Finite windows, standard statistics. Open GUE and Berry’s formula are conjectural for ζ. None of this places a zero on the line.

Partial theorems point the same way and still fall short. Rudnick and Sarnak (1996) match the GUE correlations for test functions whose Fourier transforms are supported in a bounded range, not across the full range Montgomery conjectured. Hilbert–Pólya, the idea that the ordinates are eigenvalues of a self-adjoint operator, would imply the Hypothesis if it were proved. It has not been proved. On this site it stays Open. The zeros are points. They are not themselves fields.

How to read the stack Verification to a great height, a proved positive proportion on the line, a zero-free region near Re(s) = 1, and GUE statistics: four different kinds of support. None of them is the missing step from “supported” to “proved.”

Closest calls: Lehmer pairs

On the critical line, Hardy’s Z(t) is a real function, and every zero shows up as a place where it changes sign. Between two neighbouring zeros the curve has to swing up (or down) and come back. Usually the swing is large. Sometimes two zeros sit so close together that the curve only just makes it across. D. H. Lehmer found the first famous example near t = 7005 in 1956, and such near-misses are now called Lehmer pairs.

Figure 6 Lehmer pairs Settled

Three rows of graphs of Hardy's Z function. Each row shows the curve crossing zero at a pair of neighbouring zeros, with a magnified panel of the stretch between them. A typical pair near t = 7026 rises to about 2.3 between its zeros. Lehmer's pair near t = 7005 rises only to 0.0040. The closest pair among the first 100 000 zeros, near t = 71733, dips only to minus 0.00055 before crossing back.
Hardy Z(t) around three pairs of neighbouring zeros. Ordinates from Odlyzko’s table of the first 100 000 zeros; Z(t) computed with mpmath 1.4.1 siegelz at 30 digits, 7 October 2026 PT. A typical pair (zeros 6732 and 6733, gap 1.000 mean spacings) peaks at −2.32 between its zeros. Lehmer’s pair (zeros 6709 and 6710, gap 0.042 mean spacings) peaks at 0.0040, while its neighbours reach 7.6. The closest pair in the table (zeros 95 248 and 95 249, gap 0.022 mean spacings) peaks at −0.00055. The magnified panels redraw on load; motion is off if your system asks for reduced motion. Settled These zeros are on the line; a finite window, not a proof.
  • Settled If one of these small swings had failed to cross zero, Z(t) would show two fewer sign changes than the zero count N(T) requires. That would put two zeros off the critical line, a counterexample to the Riemann Hypothesis. Lehmer pairs are where that nearly happens, so verification programs check them with extra care.
  • Settled Computed here Among the 99 999 gaps in the first 100 000 zeros, 79 are smaller than 0.1 mean spacings. Independent random points (Poisson) would give about 9 500 such gaps; the GUE Wigner surmise gives about 107. Repulsion makes tiny gaps rare, but they still occur.
  • Settled Lehmer pairs are tied to the de Bruijn–Newman constant Λ, which measures how close ζ is to having a zero off the line under a heat-flow deformation. The Riemann Hypothesis is equivalent to Λ ≤ 0, and Rodgers and Tao (2020) proved Λ ≥ 0. So if the Hypothesis is true, it holds with no room to spare. Open Whether Λ = 0 is exactly the Riemann Hypothesis.