The music of the primes · a lab note

Where the sum falls silent

A million zeta zeros, one irrational angle and the prime 2. The explicit formula says where their sum must vanish — and it does, at all 33 places.

David Hessabout 9 min reada side note to parts 1–4

Add up a million rotating arrows, one for each zero of the zeta function, and the sum usually points somewhere. But at certain places it falls completely silent — and the primes tell us in advance exactly where.

Think of a concert hall where a million people clap, each at a slightly different pace. Most of the time you hear a steady roar. But now and then the whole hall falls silent at once, as if everyone had agreed on it — and the score that says exactly when is written by the prime numbers. This lab note looks at those silent places. They are called feet here, because they sit at the foot of the peak that the prime 2 raises in the sum. Every number below was computed on one million consecutive zeros near the height t ≈ 4.7·10⁶ and checked against a null model in which the same gaps are shuffled.

33 / 33predicted feet where the sum vanishes
2.0·10⁻⁶median |S| at the feet (null model 8.9·10⁻⁴)
πthe phase jumps by an odd multiple of π at every foot
1.2·10⁻⁷all that is left after the primes up to 10⁵
Quick primer · what is being summed

Take N = 100,000 consecutive zeros, rescale (“unfold”) them so that their mean spacing is 1, and give each zero an arrow of length 1 turned by the angle θ·x, where x is its unfolded position. The average of the arrows is the sum S(θ; N, c) = (1/N) Σ exp(iθx); c is the zero at the centre of the window.

The angle stays fixed at θ₅ = (ln 5)/5 rad, which is ln 5 · 36/π = 18.4427999542…°. It is irrational, so the arrows never repeat exactly.

The explicit formula (Landau, Gonek) says that the sum is not random: each prime p raises a peak at θ = 2π ln p / H, where H = ln(t/2π) slowly grows with the height. As the window moves up through the zeros, the peak of the prime 2 drifts across the fixed angle θ₅. The window itself has its own pattern of zeros (a Dirichlet kernel). Where the two meet, the sum must vanish. That is a foot.

01Anatomy of one foot

Here is the sum around one predicted foot, window by window. With the real zeros, |S| dives by almost three orders of magnitude exactly where the explicit formula puts the foot. With the same gaps in a random order, nothing happens.

A foot: the sum goes silent and its phase turns over
|S| (top, logarithmic axis) and the phase of S with its linear trend removed (bottom), for windows centred up to 1,500 zeros on either side of foot m = 1.
zeta zerosshuffled gaps (null model)phase − trend
10⁻⁶10⁻⁵10⁻⁴10⁻³10⁻²-1,500-1,000-50005001,0001,500window position c − foot (in zeros)|S|foot: |S| ≈ 4.1·10⁻⁶0.00.51.0-1,500-1,000-50005001,0001,500window position c − foot (in zeros)phase / πjump of π
At the foot |S| drops to about 4·10⁻⁶, below one step of the discrete window, and the phase jumps by π — the signature of a phase singularity, the point where a wave has no amplitude and so no defined phase.
Picture this · the eye of a whirlpool

Waves on water have calm points where crests and troughs cancel exactly. Around such a point the wave circulates like water around a drain. Physicists call them wave dislocations or phase singularities (Nye and Berry, 1974). The feet of this sum behave the same way — only the “water” is made of the zeros of the zeta function.

02All 33 feet, not just one

Across the million zeros the peak of the prime 2 meets 33 zeros of the window kernel. They were predicted in advance from the formula alone; the data were then checked at those places and at random control places.

Preregistered · a blind test before the note

The rule for where the feet lie was first put to a blind test. On 10 November 2026 three predictions were written down before the zeros were computed, far beyond this million: foot m = −61 at n = 11,231,255, foot m = −122 at n = 13,447,540, and the top of the peak at 18° at n = 13,487,775, with success defined as a hit within ±1,000 zeros. Then 140,001 new zeros were computed. Both feet landed exactly on the predicted window (scan step 100 zeros) and the peak 100 zeros away: 3 out of 3.

It is the same game as predicting an eclipse: first write down the date, then look at the sky.

Every predicted foot is silent
The smallest |S| near each of the 33 predicted feet (hover over the points), for the zeros and for the null model, with the median of random control places.
zeta zerosnull modelcontrol places (median)
10⁻⁶10⁻⁵10⁻⁴10⁻³10⁻²-15-10-5051015foot number m (0 = top of the peak, no foot there)smallest |S|control places (median) 1.1·10⁻³peak p = 2
Median at the feet 2.0·10⁻⁶; random control places 1.1·10⁻³; null model 8.9·10⁻⁴. All 33 feet lie far below every control.
The phase jumps by an odd multiple of π
The jump of the phase across each foot, in units of π.
-3-113-15-10-5051015foot number mphase jump / π
All 33 jumps are odd multiples of π (±1π or ±3π), within 0.0065π. An even jump would mean the sum only came close to zero; an odd one means it went through it.

03The feet are points of a line

So far only the window position c moved. Let the angle move too and the picture becomes a plane (θ, c). In that plane the sum vanishes at isolated points — nodes — around which the phase turns once, either anticlockwise (+1) or clockwise (−1).

Nodes in the (θ, c) plane
Colour: |S| on a logarithmic scale (gold = small, dark = large). Circles: nodes with index +1; crosses: index −1. Switch to the null model with the buttons.
-3-2-10123-800-4000400800window position c − foot (in zeros)(θ − θ₅)·10⁷ (rad)null model: 0 nodes
The nodes lie on one line through the foot (dashed gold), with the slope predicted for the foot as θ changes (−4.264·10⁸ zeros per radian, predicted −4.261·10⁸), and they alternate in sign: 7 + 7 in this window, 11 + 11 in a wider scan, net zero. The shuffled gaps give none.

04When the window itself sings along

Two tuning forks that are slightly out of tune produce a slow wobble — beats. The window of zeros does something similar with the angle: at some sizes it stays quiet, at others it hums along.

The window adds its own term — a lattice term |sin(Nθ/2)| / (N sin(θ/2)). For N = 100,000 it is only 2.5·10⁻⁹, so nothing disturbs the feet. But choose N = 50,000, where Nθ₅/2π = 2,561.5 sits exactly half-way between two whole numbers (an antinode), and the lattice term grows to 1.25·10⁻⁴.

Node and antinode of the window size
|S| around the predicted foot for three window sizes.
N = 100,000N = 200,000N = 50,000 (antinode)
10⁻⁶10⁻⁵10⁻⁴10⁻³10⁻²-1,500-1,000-50005001,0001,500c − predicted foot (in zeros)|S|lattice term for N = 50,000: 1.25·10⁻⁴
At the antinode the single foot is replaced by two minima on both sides, 280–1,160 zeros away, and |S| no longer reaches zero: the lattice term fills the hole.

05The arithmetic of the angle

Why 18.4427999542…°? It is the same problem as the calendar. A year is not a whole number of days (365.2422…), so the calendar needs leap years, and the best leap-year rules come from the continued fraction of that number. Because the angle is irrational, its full turns never close exactly, and the way they almost close is written in its continued fraction: 360/θ₅ = 19.5198… = [19; 1, 1, 12, 8, 2, 5, …].

How close the steps come to whole turns
The best approximations of 360/θ₅: after this many steps the angle is this far from a whole number of turns.
0.0010.010.1110distance from a whole number of turns (degrees, log axis)19 steps1 turn−9.587°20 steps1 turn+8.856°39 steps2 turns−0.731°488 steps25 turns+0.0864°3,943 steps202 turns−0.0398°8,374 steps429 turns+0.0068°45,813 steps2,347 turns−0.0057°
19 steps fall short of a turn by 9.587°, 20 steps overshoot by 8.856°; their average is exactly the half step θ₅/2 = 9.2214°. 39 steps miss by 0.731°, 488 steps by only 0.086°.

The same angle drives a quadratic sum Σ exp(iπτn²) with τ = θ₅/2π = ln 5/(10π). Plotted step by step, such sums draw spirals called curlicues (Berry and Goldberg, 1988). Its continued fraction [0; 19, 1, 1, 12, 8, 2, 5, …] has the same partial quotients.

The curlicue of θ₅
Partial sums Σ exp(iπτn²) for n up to 6,500 (25 cycles of 260 steps); hover over the orange points.
2605201,3003,2506,500startcolour: N from 0 (blue) to 6,500 (gold)
The curve winds into spirals and leaves them along straight runs — like a figure skater who spins on the spot and then glides across the rink. Where it turns is set by the continued fraction of τ.

Counted in cycles of 260 steps, the linear phase moves by 115.127988° per cycle ([0; 3, 7, 1, 7, 8, 2, 2, 1, 8] of a turn). The quadratic phase moves differently.

Linear and quadratic phase after n cycles
Residue (degrees) of the linear part 260n·θ₅ and of the quadratic part θ₅(260n)²/2 after n cycles.
linear 260n·θ₅quadratic θ₅(260n)²/2
-180-900901801102025304050number of cycles n (of 260 steps)residue (degrees)
After 25 cycles the linear part is back to −1.800°, while the quadratic part sits at a quarter turn, −90.968°. After 50 cycles both are near zero: −3.601° and −3.873°.

06What is left after the primes

Imagine a recording of an orchestra in which you switch off the instruments one by one. First the drums of the prime 2, then the strings of the small primes, then whole sections. In the end you hear only the faint hiss of the tape. The question is whether some instrument we did not know about is hiding in that hiss.

If the explicit formula explains the feet, how much of the sum does it explain overall? To find out, we built a model of S from the primes alone — no free parameter, prime powers up to a cut-off X — and measured what remains. The rules and the verdicts were written down before the computation.

The residual shrinks with every block of primes
RMS of |S − model| across the peak, at the 33 feet and elsewhere, against the cut-off X.
across the peak33 feetelsewhere
1·10⁻⁷2·10⁻⁷5·10⁻⁷1·10⁻⁶2·10⁻⁶1101001,00010,000100,000prime powers up to X (log axis)RMS |S − model|1.2·10⁻⁷
From 2.7·10⁻⁶ with the prime 2 alone down to 1.2·10⁻⁷ with prime powers up to 10⁵ — about 0.6× per decade, and the same at the feet as anywhere else. Across the peak the model explains 99.9992 % of S.
The residual along the peak
|S| and the residual for three cut-offs, window by window across the peak of the prime 2 (n ≈ 9.34–9.50 million).
|S|residual, X = 2residual, X = 100residual, X = 100,000
10⁻⁹10⁻⁷10⁻⁵10⁻³10⁻¹9.349.369.389.409.429.449.469.489.50zero index at the window centre (millions)size
The residual does not follow the peak or the feet; it only gets smaller as more primes are added — the hiss of the tape, not a hidden instrument. With shuffled gaps the same model fails completely (relative residual 19.6).

07What follows (and what does not)

✓ What the note showed

At every one of the 33 places predicted by the explicit formula, the sum over a million zeta zeros vanishes, its phase jumps by an odd multiple of π, and the nodes form an alternating chain along one line. Shuffled gaps show none of this. The primes up to 10⁵ explain the sum to 1.2·10⁻⁷.

✗ What does not follow

Nothing here goes beyond the explicit formula: the feet are a precise picture of known theory at one angle, not a new phenomenon. Where the feet lie was preregistered and passed a blind test (3/3) before the note, and the residual test followed rules written before the computation (version 1.1). The finer measurements of the feet on this million — depth, phase jumps, nodes, antinode — were exploratory and cover one height range, so their confirmation at other heights is planned. Nothing here bears on the Riemann hypothesis.

Next time · the music at the centre

Here we listened to the zeros of one function. Next time we move to a whole choir: over three million quadratic L-functions, each with its own zeros, and the question of what they sing at the very centre, at s = ½. The study is being prepared for publication — and it includes a mistake we found and corrected, and then checked on fresh data under a rule written in advance.

Glossary
Unfolding
Rescaling the zeros by their local density so that the mean spacing is exactly 1.
Explicit formula
An exact link between the zeros of ζ and the primes. For sums like S it predicts a peak for every prime power (Landau; Gonek).
Foot
A place where the peak of the prime 2 meets a zero of the window kernel, so that S vanishes.
Phase singularity (node)
A point where a complex wave is exactly zero; the phase turns by ±2π around it. The sign is its index.
Lattice term
The part of S that comes from the mean density of the zeros alone; it depends only on N and θ.
Antinode
A window size N for which Nθ/2π is a half-integer, where the lattice term is largest.
Continued fraction
The expansion x = a₀ + 1/(a₁ + 1/(a₂ + …)); cutting it gives the best rational approximations.
Curlicue
The path of the partial sums of a quadratic exponential sum, studied by Berry and Goldberg (1988).
Null model
The same zeros with their gaps shuffled: same density, no arithmetic.
Note, data and code

Phase singularities of a windowed exponential sum over the Riemann zeros near the Landau–Gonek peak of p = 2 (version 1.2)

This note: version 1.0 (the task and the predicted feet) · version 1.1 (plus the residual test) · version 1.2 (plus the blind test).
The studies on Zenodo: Study I · Study II · Study III · Study IV · licence CC BY 4.0

The charts are interactive — hover over the curves and points; the node map can be switched to the null model.

Napsat komentář

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