The music of the primes · part 3 of 3

795,381 voices, one choir

After the zeta function comes a whole family of its relatives — the quadratic Dirichlet L-functions. Millions of zeros, one conjecture that governs them all, and a puzzle: why do “neighbours” make less noise than a random sample?

David Hessabout 10 min readfollows parts 1 and 2

In the first two parts we listened to the solo of a single singer — the Riemann zeta function. Now a whole choir stands before us: 795,381 L-functions, each with its own zeros, all singing from the same score.

The third study focuses on the family of quadratic Dirichlet L-functions L(s, χD). The core family was later joined by the extended E100 family with another 121,557 functions. Together that is millions of computed zeros — and tests that decide between two great theories.

795,381L-functions in the main family
+121,557functions in the extended E100 family
10¹⁰largest conductor |D|
0.54–1.19χ² per bin for the ratios conjecture
Quick primer · what is L(s, χD)

Take the zeta function 1 + 1/2s + 1/3s + … and give each term a sign χD(n). For a prime p, χD(p) = +1 if D is a square modulo p and −1 if it is not (and 0 if p divides D). Each number D — the discriminant — thus gives its own function. The size |D| is its conductor, a kind of “complexity”.

Unlike the zeta function, here the central point s = ½ plays a special role: the zeros are arranged symmetrically around it and — as we will see — avoid it.

01The ratios conjecture wins; the USp limit is still out of sight

According to Katz and Sarnak, in the limit of infinitely large conductors the zeros of such a family behave like the eigenvalues of random matrices from the unitary symplectic group USp. Their most famous signature: the density of zeros near the central point drops to zero — the zeros are repelled from it.

Zeros avoid the central point
The Katz–Sarnak theoretical limit WUSp(x) = 1 − sin(2πx)/(2πx) and the mean density in the window [0, ½].
USp limit (Katz–Sarnak)measured mean in the window [0, ½]
0.000.250.500.751.001.2500.511.522.53distance from the central point x (in mean spacings)zero densitywindow [0, ½]dotted: density 1 far from the centre|D| ≈ 2·10³ → 0.772|D| = 10¹⁰ → 0.527limit USp → 0.411
The limit value 0.411 is exactly the mean of WUSp over [0, ½]. The data approach it, but very slowly: even at |D| = 10¹⁰ they are still 12.6 to 39 standard errors above it.

The trouble is that the limit really is a limit. The density near the central point falls from 0.772 (at |D| ≈ 2·10³) to 0.527 (at |D| = 10¹⁰) — and still sits 12.6 to 39 standard errors above the theoretical limit of 0.411.

Picture this · a train bound for USp station

The train is heading the right way, but each order of magnitude in the conductor moves it only a little. The corrections fade only like 1/ln |D| — and the logarithm grows incredibly slowly. From 10³ to 10¹⁰ is seven orders of magnitude, yet ln |D| grows only about threefold.

So what does describe the data? The ratios conjecture in the form of Conrey and Snaith, which also accounts for lower-order terms at finite conductor.

Which theory fits the data?
A χ² per bin of about 1 means the deviations are consistent with pure statistical noise.
0.11101001,000χ² per bin (log axis)χ² ≈ 1: agreement within noiseRatios conjectureConrey–Snaith0.54–1.19Pure USp limitKatz–Sarnakhundreds to thousands
The gap is about three orders of magnitude. The USp limit is right “at infinity”, but at finite conductors the ratios conjecture clearly beats it.

02The puzzle of contiguous families

The study compared two kinds of families: contiguous bands of consecutive discriminants and random samples of the same size. The contiguous families did suspiciously well in the tests — with a lower χ². Why?

Neighbours make less noise
How much low-frequency “noise” contiguous families have compared with random samples of the same size.
0 %25 %50 %75 %100 %low-frequency fluctuation power (random sample = 100%)Random sample of Dreference level100 %Contiguous band of Dconsecutive discriminants23–48 %
That is why contiguous families have a lower χ²: part of the fluctuation the test would otherwise see cancels out by itself.

The mechanism is elegant. The value χD(p) repeats periodically as D grows. In a stretch of consecutive discriminants of length Δ, the averages of χD(p) for primes p < Δ therefore run through almost complete periods — and their contributions cancel.

Why the noise cancels: periodicity in D
The value χD(p) repeats with period p as D grows. Each row is one prime, each column one D.
window Δ = 35 consecutive Dwindow meanp = 30.00p = 50.00p = 70.00p = 110.00p = 130.00p = 53+0.29χ = +1χ = −10 (p divides D)columns: D = 778, 779, …, 847
Small primes (p < Δ) run through several complete periods within the window, so their +1s and −1s nearly cancel. The prime 53 > Δ cannot complete a period and its mean stays non-zero. For clarity we show all integers D; the study uses fundamental discriminants.

A contiguous family thus acts as a low-pass filter in prime frequency: it damps all the “tones” up to the cut-off frequency

uc = ln Δ / ln(|D|/π)

Similar effects have been studied in recent years in the literature on so-called murmurations of Dirichlet characters — unexpected waves that emerge when you average over large families.

Test · are residue classes enough?

A simpler explanation suggests itself: perhaps all that matters is how the residue classes D mod M are represented in the family. The study checked this for 19 different moduli from 3 to 840. Result: balancing residue classes does not fully reproduce the effect. Residue classes fix only the primes dividing the modulus — whereas the low frequencies are affected by all primes up into the thousands.

03The first zero: even and odd characters differ

The zero closest to the central point, x₁, is the most sensitive test of the theory. And here the family splits into two temperaments:

Odd characters · D < 0Even characters · D > 0
Modela single effective matrix size Neff from the ratios conjecturepure USp(2Neff) model
Fit✓ excellent
describes the whole distribution of x₁ for every conductor from 10³ to 10¹⁰
≠ offset
the first zero sits systematically higher — by a lower-order term
Trendstable across 7 orders of magnitudethe shape and offset difference shrinks with growing conductor — in the E100 family at |D| = 10¹⁰ even faster than the 1/ln(|D|/π) estimate

04The explicit formula and the dual term

The last test asks: how do the zeros respond to individual primes? The explicit formula lets us measure this dependence and compare it with theory. If the theory is right, the “data vs. theory” relation has a slope of exactly 1.

The dual term completes the match
A slope of 1 means the theory exactly explains how the zeros respond to individual primes.
0.000.250.500.751.00slope of “data vs. theory”perfect agreement = 1Diagonal terms onlyno dual term0.52–0.60+ dual termratios conjecture0.995–0.999
The dual term corresponds to the sine part of WUSp(x) = 1 − sin(2πx)/(2πx) — the part that repels zeros from the central point.

The pure diagonal terms give the wrong slope — roughly half. Only once the dual term of the ratios conjecture is included do theory and data match perfectly: the slope jumps from 0.52–0.60 to 0.995–0.999.

05What follows (and what does not)

✓ What the study showed

The behaviour of families of L-functions on both long and short scales obeys a deep arithmetic order that the ratios conjecture describes precisely. The USp limit is the right destination, but it is not enough to describe real conductors.

✗ What does not follow

The zeros were tested on the critical line. The Generalised Riemann Hypothesis (GRH) is neither confirmed nor refuted by this.

Glossary
Discriminant D and conductor |D|
The number that defines the quadratic character χD. Its size |D| measures the “complexity” of the L-function and sets the density of its zeros.
USp
The unitary symplectic group. Random matrices from it have eigenvalues repelled from the point 1 — the analogue of the central point s = ½.
Katz–Sarnak philosophy
The zero statistics of a family of L-functions match, in the limit, one of the classical groups of random matrices.
Ratios conjecture
A recipe (Conrey, Farmer, Zirnbauer) for averages of ratios of L-functions over a family. It also gives lower-order terms at finite conductor; Conrey and Snaith worked it out for the one-level density.
χ² per bin
A measure of how well a histogram fits the theory. A value around 1 means the differences are consistent with statistical noise.
Neff
Effective matrix size — the number that makes a random matrix the best stand-in for a family with finite conductor.
GRH
The Generalised Riemann Hypothesis: all non-trivial zeros of these L-functions lie on the critical line.
Paper, data and code

Study III · Quadratic Dirichlet L-functions (Katz–Sarnak)

The whole series on Zenodo: Study I · Study II · Study III · 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 *