The music of the primes · part 5

The right result begins with a mistake

7.9 million L-functions, one number each: how loud is the voice at its centre? On the way to the answer this fifth study made one mistake — and caught it before publication, because every result was checked by more than one method.

David Hessabout 15 min readfollows parts 1–4

Part 3 listened to a choir of 795,381 L-functions and asked where each voice falls silent — where its zeros lie. This time the question sounds simpler: how loud is each voice at its very centre?

The fifth study computes the central value L(½, χ_D) of 7.9 million quadratic Dirichlet L-functions — every fundamental discriminant up to 10⁷ and complete stretches at 10⁸, 10⁹ and 10¹⁰ — and, for four million of them, the low zeros as well. What we measured agrees with the conjectures of the last twenty-five years, from the simplest averages to a fine link between values and zeros.

The study also carries a lesson worth more than any single number. While it was being prepared, one of our working formulas gave a result that looked like a small discovery — and passed two tests we had written down in advance. Only a second, independent calculation showed that it was an error of ours. It was corrected before the study was published, and the result is now confirmed in three independent ways. Before you find the right result, you sometimes have to make a mistake — and the only reliable way to notice it is to check everything by more than one method.

Where the error was — and where it was not

The error described in section 07 arose during the work on this fifth study, in the implementation of a formula that appears here for the first time (the density of zeros weighted by the central value). It was found and corrected before publication, so it is in no published text. Studies I–IV — parts 1 to 4 of this series — and the lab note do not use that formula and are not affected. The unweighted prediction used in part 3 and in the addendum of part 4 was recomputed by the independent code on the same occasion, and it agreed.

The study has nine figures; all of them are redrawn here from the published results file, panel by panel, each labelled with its number in the paper.

7.9 Mcentral values: every |D| up to 10⁷, stretches at 10⁸, 10⁹ and 10¹⁰
4.0 Mof these L-functions also with their low zeros
0.65 σlargest deviation of the moments from the full conjecture in 70 tests
13.3 σdifference between two independent calculations that revealed the error before publication
Quick primer · what is a central value

Every fundamental discriminant D (… −8, −7, −4, −3, 5, 8, 12, 13 …) has its own L-function, L(s, χ_D) = Σ χ_D(n)/nˢ, where χ_D(n) is +1, −1 or 0 by a simple rule about D and n. Its functional equation mirrors s into 1 − s, so s = ½ is the centre of symmetry. The number L(½, χ_D) is the central value: one number per L-function, positive in all 7.9 million cases.

D > 0 gives an even character, D < 0 an odd one. The two halves of the family behave slightly differently, so the charts always show them apart: even in blue circles, odd in gold squares. Random matrix theory models these L-functions by symplectic matrices USp(2N); the conjectures tested here say how the primes modify that picture at any finite size of D. Predictions that include the primes are drawn in green, random matrices alone in orange.

01Five moments, eighty tests

The crudest question about six million numbers is their average — and the average of their squares, cubes and so on: the moments. In 2005 Conrey, Farmer, Keating, Rubinstein and Snaith (CFKRS) conjectured the complete answer: the mean of L(½, χ_D)ᵏ is a polynomial of degree k(k+1)/2 in ln|D|. Its leading coefficient comes from random matrices (Keating and Snaith, 2000); all the other coefficients carry arithmetic.

We cut |D| ≤ 10⁷ into seven bands and compared the moments k = 1…5 for both parities: 70 tests, plus 10 more on 1.16 million L-functions at 10⁸.

Fig. 1 · leftEighty tests, none further than one error
Deviation of the measured moment from the CFKRS polynomial in units of its standard error. Each cluster is one band of |D|; inside a cluster k runs from 1 to 5 from left to right.
even (D > 0)odd (D < 0)band of ±1 error
−1.5−1.0−0.50.00.51.01.510²10³10⁴10⁵10⁶10⁷10⁸conductor |D| (centre of the band, log axis)(data − CFKRS) / errorband of ±1 error
The largest deviation is 0.65 errors in the seven bands and 0.57 at 10⁸. The errors are cautious — section 02 explains why — so the points huddle closer to zero than pure chance would put them.

Over all 6,079,285 L-functions up to 10⁷ the first moment differs from the polynomial by a few parts in a million, well inside a measurement uncertainty of 0.007–0.008 %. This is not the first test of CFKRS: Alderson and Rubinstein (2012) went much further in |D| and found agreement to seven or eight digits for k = 1. What is added here are both parities separately, with error bars, up to k = 5.

Fig. 1 · rightThe leading term is not enough
Relative difference from the data over all |D| ≤ 10⁷, for the full polynomial and for its leading term alone. Logarithmic axis: each gridline is a factor of one hundred.
full polynomial, evenfull polynomial, oddleading term only (random matrices × arithmetic factor)error of the measurement
10⁻⁶10⁻⁴10⁻²110²10⁴12345moment k|relative difference from the data|9 % · 12 %×11,000 · ×64,000dotted: error of the measurementfull polynomial: below the error for every k
For k = 1 the leading term alone misses by 9 % (even) and 12 % (odd); for k = 5 by factors of 11,000 and 64,000. The full polynomial stays below the error of the measurement for every k.
Picture this · a GPS without its corrections

A satellite navigation built on the “leading term” alone — on clocks that tick the same everywhere — would be useless by the next morning. Because of relativity, the clocks on GPS satellites run ahead of clocks on the ground by 38 millionths of a second a day; leave that correction out, and your position drifts by about ten kilometres every day.

The CFKRS polynomial tells the same story, only more sharply. With the leading term alone the fifth moment is off by a factor of 64,000 — a navigation that announces a trip around the whole Earth as a walk of some 600 metres. Everything that makes the result usable sits in the corrections: the lower-order terms, where the primes live.

One clean detail: the theory says that odd voices are on average louder than even ones by exactly A₀·π/2 = 1.10654. The data give 1.10649 ± 0.00060.

02Neighbours cancel

How uncertain is an average over 20,000 L-functions? If they were 20,000 random members of the family, the answer would follow from their spread. But ours are consecutive — every discriminant in a stretch — and that changes everything.

Fig. 2 · leftConsecutive blocks are far calmer than random ones
Scatter of the sums of L(½)ᵏ over blocks of w consecutive |D|, divided by the scatter of random samples of the same size.
|D| ≤ 10⁷, even, k = 1|D| ≤ 10⁷, odd, k = 1k = 2 (open symbols)at 10⁸, k = 1 (diamonds)
0.000.250.500.751.0010²10³10⁴10⁵width of the block w (consecutive |D|, log axis)scatter of block sums / random samplesrandom samples = 1w = 10⁵: 0.062
For k = 1 and even characters the ratio falls from 0.48 at w = 100 to 0.27 at 1,000, 0.15 at 10,000 and 0.062 at 100,000 — sixteen times calmer than chance. At 10⁸ (diamonds) the same width cancels somewhat less.

Blocks of 20,000 consecutive discriminants scatter three to nine times less than random samples (k = 1–4). The reason is a hidden periodicity: for a fixed n, the value χ_D(n) repeats in D with period n (or 4n). A block of width w therefore averages every term with n up to about w over complete periods — and a wave averaged over whole periods disappears.

Picture this · measuring the depth of a harbour

Dip a ruler into a harbour at random moments and the waves spoil every reading. Record the level over a whole number of wave periods and the waves cancel by themselves. Consecutive discriminants do this automatically for every “wave” χ_D(n) whose period fits into the block. Part 3 found the same effect in the zeros; here it is in the values.

How does the cancellation depend on the size of D? We wrote down two guesses in advance. The data rejected both.

Fig. 2 · rightTwo scaling laws, both rejected
Effective exponent α for which a band near |D| ≈ 2·10⁶ and a band at 10⁸ cancel equally at the same w/|D|^α.
even, k = 1odd, k = 1k = 2 (open symbols)the two rejected laws
0.00.10.20.30.40.51,0003,00010,00030,000width of the block w at |D| ≈ 10⁸ (log axis)effective exponent α in w/|D|^αscaling with w/√|D| (α = ½): rejectedw alone (α = 0): rejected
If only the length of the functional equation mattered, α would be ½; if only the width mattered, 0. The data sit in between (0.16–0.39) and drift with w. A law for the cancellation remains open.

Scaling with w/√|D| fails by up to 25 standard errors, a logarithmic variant by up to 16. Both rejected hypotheses stay in the paper. A practical consequence: error bars taken from consecutive blocks are cautious — which is why the eighty points of the first chart sit so close to zero.

03The shape of the crowd

Averages are one thing, the whole distribution another. In the limit of infinitely large |D|, the logarithm of the central value follows a bell curve (Keating and Snaith). At any |D| we can reach, the curve is skewed and shifted. A hybrid model — an arithmetic factor from the primes times the characteristic polynomial of a random symplectic matrix, with the matrix size fixed by the first moment and no free parameter — says by how much.

Fig. 3 · leftThe width fits; the skewness does not quite
Mean, variance and third cumulant of ln L(½) relative to the hybrid model, by bands of |D|. For the mean the plain difference is shown.
variance, evenvariance, oddthird cumulant (open symbols)mean (small open symbols)hybrid model
−0.3−0.2−0.10.010⁴10⁵10⁶10⁷conductor |D| (centre of the band, log axis)data / hybrid − 1hybrid modelvariance: agreement to 0.26 %third cumulant
The variance agrees to 0.26 % for |D| ≥ 10⁵. For |D| ≥ 10⁶ the third cumulant is 4–7 % less negative than the model and the mean lies about 0.02 lower — small, but far outside the errors, and rejected by the rule written beforehand.
Fig. 3 · rightA one-percent deviation on its way to zero?
Fractional moments E[L^½] (full symbols) and E[L^−½] (open) relative to the hybrid model, ±2 errors. The axis is 1/ln|D|, so infinity lies at a finite distance — on the right.
E[L^½], evenE[L^½], oddE[L^−½] (open symbols)hybrid model
−1.5−1.0−0.50.00.51.01.510⁶10⁸10¹²10²⁴∞conductor |D| (axis scaled as 1/ln|D|; infinity on the right)data / hybrid − 1 (%)hybrid modellines: a + b/ln|D|limit ± 1 error →
The deviation of about 1 % shrinks steadily up to 10⁸, and straight lines extrapolate to −0.09 ± 0.23, +0.10 ± 0.38, −0.26 ± 0.21 and −0.03 ± 0.43 % — compatible with zero. But the rule written beforehand asked for a drop of more than three errors between the band 10⁶–10⁷ and 10⁸ and got 0.5–2.4. Verdict: undecided.
Picture this · how far is infinity?

On this axis the step from 10⁶ to 10⁸ is exactly as long as the step from 10⁸ to 10¹², and as the one from 10¹² to 10²⁴. To move our last point one tick to the right, the conductor would have to grow ten-thousandfold. This is why the honest answer here is “undecided”, not “converges”.

04Nobody is silent

Can a voice be completely silent — can L(½, χ_D) be zero? Chowla conjectured that it never happens, and Soundararajan proved in 2000 that it does not happen for at least 87.5 % of one large subfamily. Among all 6,079,285 L-functions up to 10⁷ none vanishes. The quietest is D = −115,147 with L(½) = 0.0000604 (a known example, due to Watkins — the gold dot in the header), and only 66 values lie below 0.001.

Random matrices say exactly how rare quiet voices should be: the number of values below ε falls like ε^(3/2). Halve the threshold and about 35 % of the values remain.

Fig. 4 · leftSmall values follow the 3/2 law
Number of discriminants with L(½) below ε among all |D| ≤ 10⁷. Both axes logarithmic.
even (D > 0)odd (D < 0)slope 3/2 (random matrices)
110¹10²10³10⁴10⁵10⁶10⁻⁴10⁻³10⁻²10⁻¹1threshold ε (log axis)number of D with L(½) < ε (|D| ≤ 10⁷)slope 3/2 (random matrices)quietest value: 6.04·10⁻⁵
The fitted exponent is 1.518 ± 0.015 (even 1.525 ± 0.022, odd 1.511 ± 0.021) against 3/2 from USp.
Fig. 4 · right…but there are fewer of them than the model says
Observed number of values below ε = 0.1 (full symbols) and 0.03 (open) divided by the prediction of the hybrid model, by bands of |D|; ±1 error.
even, ε = 0.1odd, ε = 0.1ε = 0.03 (open symbols)hybrid model
0.60.70.80.91.01.11.210⁵10⁶10⁷conductor |D| (centre of the band, log axis)observed / hybridhybrid model = 1|D| ≥ 10⁶: 11–17 % fewer
For |D| ≥ 10⁶ there are 11–17 % fewer small values than the hybrid model predicts — the same deficit in all nine classes of (χ_D(2), χ_D(3)).

The classes themselves differ enormously. Discriminants with χ_D(2) = χ_D(3) = −1 produce 4,161 values below 0.03; those with +1, +1 only 21 — just as their arithmetic factors say. Two small primes decide, two hundred to one, who will be quiet.

05The centre knows where the nearest silence is

For four million L-functions we know both the central value and the zeros near the centre. The two are tightly linked: if the first zero sits close to the central point, the function has to come down to meet it, and the central value is small.

Fig. 5 · leftValue and first zero: tied more tightly than in random matrices
Correlation of ln L(½) with the height of the first zero in four families (±2 errors). Orange bars: USp(2N) with N taken from the first moment.
even (D > 0)odd (D < 0)USp(2N)
0.800.820.840.860.880.90B10 · 10⁴–10⁵C100 · 10⁶–2·10⁶R10 · 10⁶–10⁷D100 · 10⁸family and range of |D|correlation of ln L(½) and the first zero
The correlation is 0.83–0.89. In seven of the eight groups it lies above the random-matrix value, by up to 17 errors. Shuffling the values among the discriminants gives 0.000 ± 0.006.
Picture this · a string touched near the middle

Touch a vibrating string lightly at one point and it cannot swing there. A zero is such a touch. The closer the first zero lies to the centre, the less room the function has to rise: the mean central value grows from about 0.7–1.0 in the fifth of the L-functions with the lowest first zero to 7.5–19 in the fifth with the highest.

Perhaps the random matrix simply had the wrong size? The ratios conjecture supplies an effective size N_eff with no free parameter.

Fig. 5 · rightThe right matrix size explains the first zero — but not the coupling
Data divided by USp(2N_eff): the mean first zero (small open symbols) and the shift of the mean first zero when every L-function is weighted by its central value (full symbols); ±2 errors.
even (D > 0)odd (D < 0)mean first zeroUSp(2N_eff)
1.001.051.101.15B10 · 10⁴–10⁵C100 · 10⁶–2·10⁶R10 · 10⁶–10⁷D100 · 10⁸family and range of |D|data / USp(2N_eff)USp(2N_eff) = 1shift of the first zero under weighting: 9–16 % largermean first zero: agrees
With N_eff the mean first zero agrees (within 3.2 errors, except C100 even with 6.9). The shift under weighting stays 9–16 % larger than in random matrices (8–42 errors).

So the extra coupling is not an effect of a finite matrix size. Something arithmetic is left — and the next section shows that the ratios conjecture accounts for it.

06Weighing the zeros by loudness

Take the density of zeros near the central point — the subject of part 3 — and let every L-function vote in proportion to its central value. Loud voices keep their first zero further out, so the weighted density has fewer zeros near the centre and more beyond it. Alessandro Fazzari (2024) derived from the ratios conjecture exactly how much, with all lower-order terms.

Fig. 6 · leftLoud voices keep their zeros away from the centre
Weighted minus unweighted density of the low zeros in the family C100 (10⁶ < |D| ≤ 2·10⁶, 607,941 L-functions) in five bins; the ±2 error bars are smaller than the symbols.
even (D > 0)odd (D < 0)ratios conjecture with weight L(½)USp alone
−0.8−0.6−0.4−0.20.00.20.40.60.00.20.40.60.81.0unfolded height x (centre of a bin of width 0.2; 1 = mean spacing of zeros)weighted − unweighted densityfewer zeros near the centremore further out
The ratios conjecture with weight L(½) (green) passes through the points: χ² = 1.76 (even) and 0.52 (odd) for five bins. Random matrices alone (orange, dashed) fail in every family, with χ² between 82 and 1,811.
Fig. 6 · rightFrom 10⁶ to 10¹⁰: nothing left over
Data minus prediction in the five bins for five conductors and both parities (±2 errors); 3.95 million L-functions in this chart.
even (D > 0)odd (D < 0)prediction
−0.04−0.020.000.020.040.10.30.50.70.9unfolded height x (bin); in each bin from the left |D| ≈ 10⁶, 10⁷, 10⁸, 10⁹, 10¹⁰data − prediction (±2 errors)ratios conjecture weighted by L(½)10⁶10⁷10⁸10⁹10¹⁰
χ² for five bins lies between 0.2 and 4.3 in all twenty groups of the study; the smallest p-value is 0.51. The families at 10⁷ and 10⁹ were computed after the correction described next, under rules written before their data existed.
Picture this · one person, one vote — or one share, one vote

An election counted by heads and the same election counted by shares give different results, and the difference tells you how wealth and opinion go together. Here the shares are central values and the opinions are positions of zeros. The ratios conjecture predicts the difference between the two counts to a few thousandths — at every conductor from ten thousand to ten billion.

07Checked twice: the term that was not there

Now the most instructive part of the study. While it was being prepared — for six laboratory steps — the last chart looked very different. Our first implementation of Fazzari’s formula agreed with the odd characters and disagreed, massively, with the even ones: χ² = 202 and 285 where about 5 is expected. The disagreement had a clean shape, and its size fell with the conductor roughly like |D|^(−1/4). It looked like an unknown lower-order term for even characters.

We did what one is supposed to do. We fitted the law, wrote down in advance what it implied for discriminants not yet computed at 10⁸, computed their zeros — seven and eight hours on nine cores — and the expectation came true. Twice.

Then, still before publishing, came one more step that had also been written down in advance — the one that matters most: an independent implementation of the prediction, sharing no code and no algebraic shortcuts with the first — local factors summed numerically instead of closed forms, a numerical derivative, a numerical limit. It disagreed with the first implementation by up to 13.3 standard errors, for even characters only. The cause was a single factor: in one term of the recipe our working code had X(½ + α + β) where the formula has X(½ + α)·X(½ + β).

Fig. 7 · leftA term that obeyed a law — and vanished
Amplitude of the “extra term” for even characters with the erroneous first implementation (open grey) and with the corrected prediction (blue, ±2 errors); odd characters with the corrected prediction in gold.
even, erroneous first implementationeven, correctedodd, correctedthe “law”no extra term
0.00.51.01.52.02.510⁴10⁵10⁶10⁷10⁸10⁹10¹⁰conductor |D| (log axis)amplitude of the apparent term (C100, erroneous = 1)ratios conjecture: no extra termB10C100R10S710⁸S9LEpower law ≈ |D|^(−1/4)fitted to the erroneous points
With the corrected prediction the amplitude is zero within the errors at every conductor: 0.065 ± 0.059 (C100), −0.038 ± 0.050 (all three families at 10⁸), 0.081 ± 0.185 (E). The families S7 and S9L were computed only after the correction.

Why had two implementations agreed before? Because the second had been derived from the first and inherited the same factor. And why had the expectations come true? Because they were tested against the same wrong formula. The “law” was real — it was the true size of the error, which does shrink with the conductor.

Once the error was found we set ourselves one more test, before any new data: the corrected prediction has to describe a family never analysed before. A desktop computer then spent six hours on the 6.6 million zeros of all 1,519,830 L-functions with 7.5·10⁶ < |D| ≤ 10⁷. Result: χ² = 0.21 and 0.64. The erroneous formula would have given 240.7 — and an amplitude of 0.62, just what its scaling law had “foretold” (0.608 ± 0.045). A second new family at 10⁹ — 486,344 L-functions, almost twenty hours — gave 1.75 and 1.31.

Fig. 7 · rightSame data, two formulas
χ² for five bins in all twenty groups: erroneous first implementation (open) and corrected prediction (full). Logarithmic axis.
even, erroneouseven, correctedodd, erroneousodd, correctedp = 0.001
0.1110100B10C100R10S7D100D100+bD100c10⁸S9LEfamily (ordered by conductor)χ² for five bins (log axis)threshold p = 0.001202285241
The correction changes almost nothing for odd characters and everything for even ones. Above the dashed line a model is rejected at p = 0.001.
What we took from it

On the way to a right result a mistake can happen to anyone; what matters is having a way to find it. A prediction written down beforehand protects against a flexible analysis. It does not protect against an error shared by the prediction and the analysis. Only a second method that shares nothing with the first one does that.

Recomputing the data analysis with independent code was already routine in this series. Study V extended the routine to the theoretical prediction itself — and that is where it paid off: the error never reached a published page.

Picture this · two clocks set by the same watch

Two clocks set by the same wrong watch agree perfectly — and tell you nothing. In 2011 the OPERA experiment measured neutrinos that seemed to outrun light by six standard errors. The statistics were sound; the cause turned out to be a loose fibre-optic connector. Agreement between results that share a flaw is not confirmation.

08Two ledgers that must balance

Two implementations and two new families are already two ways of checking. A third one, of a completely different kind, does not use the ratios conjecture at all: the explicit formula, an exact identity that ties the zeros of every L-function to the primes. Sum a test function over the zeros — the zero side. Sum the matching expression over the primes, weighted by χ_D(p) — the prime side. The two must be equal, L-function by L-function.

We evaluated both sides of the weighted-minus-unweighted averages on two families: C100 (607,941 L-functions) and S7 (1,519,830). The prime side uses only the characters and the measured central values — no zeros and no conjecture.

Fig. 8 · left and middleWhere the effect comes from: mostly small primes
The prime side split into the prime 2, odd primes up to √(|D|/π) (“small”), odd primes above (“large”) and prime powers; white diamonds: the zero side. Pale bars: C100, full bars: S7.
evenoddzero side
−0.5−0.4−0.3−0.2−0.10.0p = 2small odd plarge odd pprime powerssum = zero side?parts of the prime side (in each group from the left: C100 even, S7 even, C100 odd, S7 odd)weighted − unweighted (contribution)
The sums are −0.3685 (even) and −0.4549 (odd) on C100, −0.3868 and −0.4655 on S7 — and the zero side lands on them. About three quarters of the effect come from the small odd primes, most of the rest from the prime 2.
Fig. 8 · rightThe books balance — if the last cent is counted properly
Zero side minus prime side in units of the paired error. Open: the part above the computed zeros taken from the smooth density alone; full: from the densities of the ratios conjecture, as the rule written beforehand prescribed.
tail from the smooth density alonetail from the ratios-conjecture densitiesband of ±3 errors
−60−40−200C100 · evenC100 · oddS7 · evenS7 · oddfamily and parity(zero side − prime side) / error−5.1+2.2−9.0+1.6−53.6−2.0−23.2+0.1band of ±3 errors
With the proper tail the two sides agree to 2.2 and 1.6 errors on C100 and to 2.0 and 0.1 errors on S7. The paired error is tiny — 4.4·10⁻⁷ for S7 even — so even a tail of 2·10⁻⁵ decides.

Our zeros stop at an unfolded height of 4.3, and the test function still has a sliver of weight above it. At this precision the sliver matters: on S7 the explicit formula demands 2.35·10⁻⁵ and 3.13·10⁻⁵ for it, and the ratios conjecture supplies 2.26·10⁻⁵ and 3.14·10⁻⁵.

Picture this · double-entry bookkeeping

An accountant records every transaction twice, once on each side of the ledger. If the two columns do not add up to the same total, something is wrong — however plausible each column looks on its own. The explicit formula is the double entry of number theory: zeros in one column, primes in the other. After section 07, this was the audit — a third, independent road to the same answer.

09A tuning fork for the method

One last figure, from the appendix. Before trusting a statistical method on something unknown, try it on something known. Our tuning fork is a result of Ford and Zaharescu (2005) about the zeros of the Riemann zeta function itself.

Multiply the heights γ of the zeros by a number α and keep only the fractional parts. For every α they spread evenly between 0 and 1 — to first order. In the second order the primes leave a fingerprint: if α = ln p / 2π for a prime p, the density carries a ripple whose shape follows from Landau’s formula of 1911, with no free parameter.

Fig. 9The fingerprint of a prime in the fractional parts
Density of the fractional parts of α·γₙ on 2,001,052 zeros of ζ (Odlyzko’s table) in 50 bins, ±2 errors; green: Landau’s formula.
α = ln 2/2πα = ln 7/2πα = √2/2π (control)Landau’s formula
0.70.80.91.01.11.20.000.250.500.751.00u = fractional part of α·γₙdensity (uniform = 1)fewer fractional parts at the edges, more in the middle
For α = ln 2/2π and ln 7/2π the ripple is there, as the formula says; for the control α = √2/2π the density is flat. In all 20 combinations of block and prime (2, 3, 5, 7) the smallest p-value is 0.66, while a flat density is excluded for p = 2 in every block (χ² between 1,383 and 4,940 for 49 degrees of freedom).

This is known theory, and the test confirms only the method. But the ripple is too pretty to leave alone — see the end of this page.

10What follows (and what does not)

✓ What the study showed

On 7.9 million central values the moments follow the full CFKRS polynomials for both parities up to k = 5. Consecutive families cancel far more than random ones. Small values follow the 3/2 law and none vanishes. And the weighted density of low zeros follows the ratios conjecture from 10⁴ to 10¹⁰ — confirmed in three independent ways: by two implementations that share no code, by two families computed afterwards under rules written before their data, and by the explicit formula.

✗ What does not follow

Agreement supports a conjecture; it does not prove it, and nothing here bears on the generalised Riemann hypothesis. Several of our working hypotheses were rejected by the data and stay in the paper (the scaling of the cancellation, the third cumulant, the coupling to the first zero in USp); two are undecided. The “new term” for even characters does not exist: it was an error in our working implementation during this study, corrected before publication. Studies I–IV are not affected.

The paper also tests the splitting conjecture of Gonek, Hughes, Keating and Bui and looks for the X^(3/4) term in the third moment; those parts have no figures and are left out here.

Glossary
Central value
The value L(½, χ_D) of an L-function at the centre of symmetry of its functional equation.
Fundamental discriminant
An integer D that is the discriminant of a quadratic number field; each one gives exactly one primitive real character χ_D. |D| is the conductor.
Even and odd characters
χ_D is even for D > 0 and odd for D < 0. The two parities have different Γ-factors and are analysed separately.
Moment
The mean of L(½, χ_D)ᵏ over a set of discriminants.
CFKRS polynomial
The conjecture of Conrey, Farmer, Keating, Rubinstein and Snaith for the moments: a polynomial of degree k(k+1)/2 in ln|D| with all lower-order terms.
USp(2N)
Random unitary symplectic matrices — the random-matrix model of the quadratic family (Katz and Sarnak).
Hybrid model
ln L = an arithmetic part from the primes + ln of the characteristic polynomial of USp(2N), with N fixed by the first moment.
Ratios conjecture
A recipe (Conrey, Farmer, Zirnbauer) for averages of ratios of L-functions; it yields the density of low zeros with all lower-order terms. Fazzari worked out the version weighted by L(½).
Weighted density of low zeros
The density of zeros near the central point in which every L-function counts in proportion to its central value.
Explicit formula
An exact identity between a sum over the zeros of an L-function and a sum over the primes.
χ² for five bins
A measure of how well five measured values fit a prediction; about 5 is expected when the errors are exact, less when they are cautious.
Rule written beforehand
The criterion for the verdict (supported, rejected, undecided), recorded with its date before the numbers it judges were computed.
Next time · part 6

The ripple of a single prime is only the beginning. What happens when two primes act at once? What does the next order look like? And do the zeros of other L-functions carry the same fingerprint, twisted by their characters? The sixth study follows the ripple — it is on its way.

Paper, data and code

Study V · Central values of quadratic Dirichlet L-functions: moments, consecutive families, the distribution of values, the first zero, and the weighted density of low zeros

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

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

Napsat komentář

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