The music of the primes · part 4 of 4

Three zeros at a time

After pairs come triples. 147 million zeros of the zeta function, one formula from the ratios conjecture — and a comparison from 2008 that finally gets error bars.

David Hessabout 10 min readfollows parts 1–3

Parts 1 and 2 listened to the zeros of the zeta function two at a time: how often does a zero have a neighbour at a given distance? Part 3 turned to a whole choir of L-functions. Now we return to the solo singer — and listen to three zeros at once.

The fourth study measures the three-point correlation of the Riemann zeros on 147 million zeros computed by David Platt and distributed by the LMFDB, from the very first zero up to the height t = 3·10¹⁰. Every number comes with an error bar, and every claim is checked against null models.

147.2 Mzeros of ζ in 14 blocks
3.06·10¹⁰highest height t on the critical line
0.993–1.010amplitude λ₃ in every block (1 = formula)
15.4 σquadratic L-functions at |D| = 10¹² still above the USp limit
Quick primer · what is the three-point correlation

Rescale (“unfold”) the zeros so that their mean spacing is exactly 1. The pair correlation R₂(a) asks how often a zero has another zero at distance a. The three-point correlation R₃(a, b) asks how often a zero has a neighbour at distance a and another one a further b beyond it. A value of 1 means “as often as for independent points”.

Far up the critical line, R₃ approaches the prediction of random matrices (GUE) — for restricted test functions this is a theorem of Hejhal and of Rudnick and Sarnak. At any finite height there are corrections that carry the primes. Conrey and Snaith (2008) derived all of them from the ratios conjecture, down to the constant term. That formula is what this study tests.

01Close triples are still rare

The clearest place to look is where all three zeros sit close together: both gaps of the triple about half a mean spacing. Random matrices predict R₃ ≈ 0.281 in this bin. The zeros fall short — and catch up only very slowly.

Three close zeros at fourteen heights
R₃ in the bin where both gaps lie between 0.5 and 0.6 mean spacings, against the median height of each block.
zeros (Platt, LMFDB)formula (ratios conjecture)GUE
0.150.180.210.240.2781114172023height L = ln(t/2π) (median of the block)R₃ in the bin (0.5–0.6)²GUE (random matrices) 0.2806t = 3.06·10¹⁰: 31 errors below GUE
From t ≈ 4·10⁴ to 3.06·10¹⁰ the value rises from 0.158 to 0.2533 — still about 31 standard errors (10 %) below GUE, and only 1.6 errors from the formula.

For comparison: pairs at the same distance are only 3–5 % below GUE in the lowest blocks and 1.2 % at the top. Close triples are much more sensitive, because near the origin R₃ grows as the product of the squares of the distances — every small correction gets multiplied.

Picture this · three people on one bench

Two strangers on a park bench keep a polite distance. Three strangers on the same bench are even more careful: each of them has to keep away from two neighbours at once. The zeros of ζ behave the same way — and the primes decide exactly how much room they leave each other.

02A comparison from 2008, repeated

Conrey and Snaith already compared their formula with the first 100,000 zeros. They reported differences with a standard deviation of about 0.03 and a maximum of 0.164, while they expected the formula itself to be accurate to about T−1/2 ≈ 0.003. The differences were ten times larger, and they put this down to the small height.

We repeated the comparison on the same zeros — but with an error attached to every bin. Each block is cut into 40 consecutive pieces, and the scatter of the pieces gives the error.

The differences are the size of the noise
Data minus formula on the first 100,000 zeros: standard deviation and maximum, compared with pure noise (the formula plus a random error of each bin, 2,000 simulations).
0.000.050.100.15difference data − formula (first 100,000 zeros)expected error of the formula T⁻¹ᐟ² ≈ 0.003Conrey–Snaith 2008reported · standard deviation≈ 0.03our data − formulastandard deviation0.0271pure noisestandard deviation0.0265 ± 0.0007Conrey–Snaith 2008reported · maximum0.164our data − formulamaximum0.090pure noisemaximum (90 % of simulations)0.088–0.129
Our differences match pure noise. Their maximum of 0.164 lies above our noise range, but their bins are not known to us — finer bins give a larger maximum, so only the size can be compared.

With the error of every bin, the formula agrees: χ² per bin 1.050, where pure noise gives 1.054; GUE gives 6.37. The differences seen in 2008 were the counting noise of the triples — 100,000 zeros simply do not contain enough triples for a sharper test. Our own first guess, that the formula had been evaluated at a single height, turned out to be wrong: their formula is an integral over heights. The rejected guess stays in the paper too.

Why the expected χ² is 1.054, not 1

An error estimated from 40 pieces is itself uncertain. A deviation divided by such an error follows Student’s t distribution with 39 degrees of freedom, whose mean square is (40 − 1)/(40 − 3) = 1.054. Testing against 1 would quietly favour a model that fits slightly worse.

03Only the full formula fits

In all 14 blocks up to 3.06·10¹⁰ the formula agrees bin by bin: χ² per bin 0.905–1.145, mean 1.038, and no block further than three standard deviations from the expectation. Even the largest single deviations (3.0–4.9 errors among 784 bins of a block) are exactly what the t distribution predicts.

Which model fits the triples?
χ² per bin over the 14 blocks, from the best to the worst block; about 1 means agreement within noise.
0.51101001,000χ² per bin over the 14 blocks (log axis)expected 1.054: agreement within noiseRatios conjectureConrey–Snaith 20080.905–1.145Pairs onlyproduct of R₂ corrections2.7–37Without the primesformula, arithmetic factors removed5.0–90GUErandom matrices, infinite size6.4–117Finite matricesCUE of effective size, first order27–446
Removing the primes, or building the triples from the pair corrections, fails. Even a random matrix of a cleverly chosen finite size does worse than GUE.

The lesson: the primes enter the three-point function in a way that cannot be assembled from the two-point corrections, and at these heights a finite random matrix is no substitute for arithmetic.

Test · does the pipeline invent structure?

The same code run on 10 million eigenvalues of random 10×10 unitary matrices gives χ² 1.04 against their exact correlations and 7.27 against GUE: it finds a known structure and tells it apart from a wrong one. Shuffling the spacings of a block destroys the genuine three-point part, and the formula then fails (χ² 1,026). An independent recomputation that shares no code with the main one reproduces every number.

04One number for the whole deviation

Bin-by-bin tests produce 784 numbers per block. To follow the deviation from GUE with height, we summarise it by one amplitude:

R₃(data) − R₃(GUE) = λ₃ · [R₃(formula) − R₃(GUE)]

λ₃ = 0 means pure GUE, λ₃ = 1 means exactly the formula.

The amplitude stays at 1
λ₃ in the 14 blocks with one standard error; 1 = ratios conjecture, 0 = GUE (far below the window).
0.980.991.001.011.0281114172023height L = ln(t/2π) (median of the block)amplitude λ₃λ = 1: ratios conjectureGUE (λ = 0) lies 120–436 errors below this window ↓
λ₃ lies between 0.993 and 1.010 with errors of 0.2–0.8 %. GUE is excluded by 120 to 436 standard errors in every block, and there is no trend with height.

A straight line in the height gives a slope of (5.4 ± 3.4)·10⁻⁴ — no trend within the errors. The pair correlation tells the same story (λ₂ = 0.990–1.018).

Inside the lowest blocks, the amplitude of consecutive pieces falls — by 0.39 ± 0.05 across the first 100,000 zeros. It looked like a crack in the formula, but it is the formula’s own prediction: the height changes a lot inside a low block, and the deviation from GUE is larger at the bottom. Averaged over the heights of each piece, the formula predicts the fall: −0.391 ± 0.053 observed, −0.401 predicted.

Picture this · one dial instead of 784 pixels

Instead of comparing 784 pixels of a photograph, we ask a single question: how far is the “arithmetic filter” turned up? From t ≈ 4·10⁴ to 3·10¹⁰ the dial sits at 100 % — within a few tenths of a percent.

05Smaller sets, much further up

Steven J. Miller suggested testing “much smaller sets of zeros much further up”, because convergence in such problems can be very slow. For the zeta function, the only public zeros far above 3·10¹⁰ are Andrew Odlyzko’s sets of 10,000 zeros near the zeros number 10¹², 10²¹ and 10²².

Height tZerosAmplitude
2.7·10¹¹9,9830.92 ± 0.46
1.4·10²⁰9,983−0.82 ± 1.07
1.4·10²¹9,9830.50 ± 0.86
together0.62 ± 0.38

Consistent with the formula — and also with GUE. Ten thousand zeros do not decide. Near 10²¹, an error of 0.2 would need about 186,000 zeros, an error of 0.1 about 745,000. Odlyzko’s computation of more than a billion zeros near the zero number 10²³ would pin λ₃ down to about 0.003 — but it is not public.

Test · are the error bars honest?

200 replicas of random 100×100 matrices with 10,000 values each give λ₃/σ with a standard deviation of 0.98, and the highest block of Platt’s zeros cut into 2,236 pieces of 10,000 zeros scatters by 0.30 against an estimated error of 0.35. The errors are not underestimated — if anything, slightly conservative.

06Addendum to part 3: the train has still not arrived

Part 3 ended with a train bound for USp station: quadratic L-functions approach the Katz–Sarnak limit, but very slowly. Following the same suggestion of Miller, we computed three new bands far beyond study III — all fundamental discriminants in three windows of width 15,000 at |D| = 10¹²: 27,355 L-functions, about 12 hours of computing per band.

The ratios conjecture describes all three bands (χ² per bin 0.87–1.21); the pure USp limit does not (about 4). Near the central point the density is 0.5040 ± 0.0061 — the ratios conjecture says 0.5054, USp says 0.4105. That is still 15.4 standard errors above the limit.

How slow is slow?
Excess of the zero density in [0, ½] over the USp limit against the size of the conductor; points: families of study III and the three bands at 10¹² (combined), with one standard error. Circles: where the ratios conjecture brings the excess below 0.05 and 0.01.
ratios conjecture (no free parameter)fit c/lfit: power of |D|
0.00.10.20.30.435102050100size of the conductor log₁₀|D| (log axis)excess density in [0, ½] over USpdata|D| ≈ 3.6·10²¹|D| ≈ 2.5·10¹⁰¹
Within the data all three descriptions look alike; beyond them they part by orders of magnitude. At |D| = 10¹²⁰ the fit c/l still leaves 0.0094, the power law 1.6·10⁻⁸. The ratios conjecture, with no free parameter, brings the excess below 0.05 only near |D| ≈ 3.6·10²¹ and below 0.01 near 2.5·10¹⁰¹.

The excess multiplied by l = ln(|D|/π) stays at 2.5–2.6 from |D| ≈ 5·10⁴ to 10¹⁰ and is 2.48 ± 0.16 at 10¹²: it falls only as 1/l, just as the ratios conjecture predicts. The arrival at USp is therefore an inference through the conjecture — not an observation.

A cautionary tale · cubic twists

In 1987 Zagier and Kramarz found that 23.3 % of a family of cubic twists have a vanishing central value and suggested that the share stays positive. Mark Watkins later pushed the data to ten million: 22.9 %, 20.5 %, 17.7 % — the share kept sinking and most likely tends to zero. The data were right; the extrapolation from them most likely was not. Miller pointed us to this story, and it is why we do not extrapolate either.

07What follows (and what does not)

✓ What the study showed

The three-point correlation of the zeros at finite height is described bin by bin by the formula of Conrey and Snaith from the ratios conjecture — on 147 million zeros at 14 heights up to 3·10¹⁰. The deviation from random matrices is exactly as large as the formula says, within 0.2–0.8 %. The differences reported in 2008 were counting noise.

✗ What does not follow

Agreement supports a conjecture; it does not prove it. The zeros were computed on the critical line, so nothing here bears on the Riemann hypothesis. Statements at t = 3·10¹⁰ or |D| = 10¹² are statements at a finite height, not about the limit.

Glossary
Three-point correlation R₃(a, b)
The density of triples of zeros with gaps a and b (in mean spacings), normalised so that independent points give 1.
Unfolding
Rescaling the zeros by their local density ln(t/2π)/2π so that the mean spacing is exactly 1 at every height.
GUE
The Gaussian unitary ensemble — random Hermitian matrices whose eigenvalue statistics match the zeros far up the critical line.
Ratios conjecture
A recipe (Conrey, Farmer, Zirnbauer) for averages of ratios of L-functions. It yields all lower-order terms at finite height; Conrey and Snaith worked it out for the three-point correlation.
Amplitude λ₃
One number for the size of the deviation from GUE in the shape of the formula: 0 = GUE, 1 = formula.
χ² per bin
A measure of how well a histogram fits a model. With errors from 40 pieces its expected value is 1.054.
Platt’s zeros
The zeros of ζ up to t ≈ 3·10¹⁰ computed by David Platt with rigorous interval arithmetic and distributed by the LMFDB.
USp limit
The Katz–Sarnak limit for the zeros of quadratic L-functions near the central point (part 3).
Paper, data and code

Study IV · Three-point correlation of the Riemann zeros (with an addendum to study III)

The whole series on Zenodo: Study I · Study II · Study III · Study IV · licence CC BY 4.0 (text, data) and MIT (code)

The charts are interactive — hover over the curves and points.

2 komentáře

Napsat komentář

Vaše e-mailová adresa nebude zveřejněna. Vyžadované informace jsou označeny *