Zeros in the critical strip
Zeros on the critical line
The nontrivial zeros of ζ(s) live in the strip 0 < Re(s) < 1. Every one computed so far sits on its centre line, Re(s) = 1/2. This page draws them, counts them, and shows how sign changes of a real function certify that a zero is exactly on the line.
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.
Trivial zeros on the real axis First five zeros, to scale First twenty zeros |ζ(1/2+it)|, t ≤ 50 Where all the colours meet The series at the first zero The average size of ζ on the line Selberg’s bell curve for |ζ| New Zeros of the chopped-off series Hardy Z(t) and Gram points N(T) against its main term S(T) and Selberg’s law Gram intervals at three heights
The critical strip and the line
Settled Settled landscape Open Open claim Separate the two.
The trivial zeros are the zeros at −2, −4, −6, and so on. They are proved. They come from the gamma factor in the functional equation, not from a mystery about the primes. Everything else is a nontrivial zero.
Figure 1 Trivial zeros on the real axis Settled
- Settled Why these points: the functional equation ties ζ(s) to ζ(1 − s) through a factor that includes sin(πs/2). That sine vanishes at every even integer. At s = 0 the pole of ζ(1 − s) cancels it, and at positive even integers the gamma factor cancels it, so only −2, −4, −6, … survive as zeros.
- Settled At the negative odd integers ζ takes rational values tied to Bernoulli numbers: ζ(−n) = (−1)n Bn+1/(n+1). That is where ζ(−1) = −1/12 comes from. It is a value of the continued function, not a sum of 1 + 2 + 3 + … in the ordinary sense.
- Settled The trivial zeros are fully understood, so all the open questions, including the Riemann Hypothesis, concern only the nontrivial zeros in the critical strip drawn below.
Every nontrivial zero lies in the critical strip, the open band where the real part is strictly between 0 and 1. There are no zeros on the line Re(s) = 1. Hadamard and de la Vallée Poussin proved that in 1896, and it is the analytic form of the prime number theorem. The functional equation then excludes zeros on Re(s) = 0. That is a theorem. It is not the Riemann Hypothesis. The Hypothesis is the stronger, still open, statement that the zeros do not merely stay inside the strip: they sit on its center line.
RH: every nontrivial zero ρ has real part 1/2.
Zeros are symmetric in two ways that are proved. They come in conjugate pairs, because the Dirichlet series has real coefficients, so a zero and its complex conjugate vanish together. And if ρ is a zero then so is 1 − ρ, by the functional equation. A zero off the critical line would force a mirror zero the same distance on the other side of the line. The Hypothesis says that departure never happens.
Seeing the mirror: cut straight across the strip
The mirror rule is easier to believe when you can see it. Fix a height t and walk along the horizontal line s = σ + it from left to right, measuring how big ζ is at each step. If the line passes through a zero, the size drops to exactly 0 there.
Figure 2 Cutting across the strip: the mirror rule
- Settled the functional equation Why the two sides are tied: ζ(s) = χ(s) ζ(1 − s), and χ has no zeros inside the strip. So the size of ζ at σ + it is the size at the mirror point 1 − σ + it, times a smooth, never-zero factor |χ|. Panel B shows that factor measured straight from ζ.
- [project] Example on the first cut: |ζ(0.2 + it)| = 0.2694 and |ζ(0.8 + it)| = 0.2113. Their ratio is 1.2753, and |χ(0.2 + it)| = 1.2753. Different sizes, but locked together.
- Settled consequence If ζ were 0 at σ0 + it with σ0 ≠ ½, the factor would force a second zero at 1 − σ0 + it. On the line itself |χ| = 1 exactly, so the gold curve crosses 0 at σ = ½: the line is the mirror.
- Open the Riemann Hypothesis The picture explains why off-line zeros would come in pairs. It does not show they never happen: three cuts at low height are evidence about these cuts only. RH, that every cut through a nontrivial zero meets it at σ = ½, is still unproved.
There are infinitely many nontrivial zeros. Up to height T their number is
N(T) = (T/2π) log(T/2π) − T/2π + O(log T).
| n | γn | Gap |
|---|---|---|
| 1 | 14.134725 | — |
| 2 | 21.022040 | 6.89 |
| 3 | 25.010858 | 3.99 |
| 4 | 30.424876 | 5.41 |
| 5 | 32.935062 | 2.51 |
| 6 | 37.586178 | 4.65 |
| 7 | 40.918719 | 3.33 |
| 8 | 43.327073 | 2.41 |
| 9 | 48.005151 | 4.68 |
| 10 | 49.773832 | 1.77 |
| n | γn | Gap |
|---|---|---|
| 11 | 52.970321 | 3.20 |
| 12 | 56.446248 | 3.48 |
| 13 | 59.347044 | 2.90 |
| 14 | 60.831779 | 1.48 |
| 15 | 65.112544 | 4.28 |
| 16 | 67.079811 | 1.97 |
| 17 | 69.546402 | 2.47 |
| 18 | 72.067158 | 2.52 |
| 19 | 75.704691 | 3.64 |
| 20 | 77.144840 | 1.44 |
All twenty were checked: real part 1/2, and |ζ| at each point under 10−18. The six-decimal values for n = 1…8 match the rounding already on this page. Settled These twenty zeros lie on the critical line. Open Being on the line for n = 1…20 does not prove the Riemann Hypothesis. The Riemann–von Mangoldt count says there are infinitely many further zeros.
Figure 3 First five zeros, to scale Settled
Figure 4 First twenty zeros Open
Figure 5 |ζ(1/2+it)|, t ≤ 50 Settled
Follow ζ along the line: it runs through 0
As t grows, the value ζ(½ + it) is a point moving in the complex plane. Every zero on the critical line is a moment when that point passes through the origin.
t = 40.00 · six passes through 0 so far: t = 14.13, 21.02, 25.01, 30.42, 32.94, 37.59
Settled A plot of a known function on 0 ≤ t ≤ 40: mpmath 1.4.1 at 25 digits, step 0.02 (2,001 values), cross-checked against Z(t)e−iθ(t) to 3·10−25. The six passes are the zeros 14.13, 21.02, 25.01, 30.42, 32.94, 37.59 from the tables above. It does not show that later zeros lie on the line.
Seeing the zeros as places where all the colours meet
Every figure above looks along the line itself. This one looks at a whole patch of the plane at once. ζ(s) is a complex number, so it has a size and an angle. Colour each point s by the angle alone. Where ζ is zero the angle is undefined, and every angle occurs in any small circle around that point, so all the colours crowd together there. You can find the zeros by eye.
Figure 6 Where all the colours meet Settled
- Settled A zero and a pole look alike at a glance, but the colours go round in opposite orders. That is the argument principle: the number of times the angle turns around a closed loop equals zeros minus poles inside it.
- Settled To the right of σ = 1 the colours barely change, because ζ stays close to 1 there and has no zeros. Toward the left the bands bunch up as |ζ| grows; the trivial zeros at −2, −4, … are further left than this window.
- Settled Nothing in the picture forces the meeting points onto the dashed line. In this window they happen to be there, and Platt and Trudgian (2021) checked every zero up to height 3·10¹², but a picture of finitely many zeros cannot settle the Riemann Hypothesis.
The series does not converge on the line, yet ζ is still zero there
The familiar definition ζ(s) = 1 + 2−s + 3−s + … only converges when Re(s) > 1. On the critical line every term has size n−1/2, so the sum never settles. People sometimes take this to mean that “ζ = 0 on the line” is a trick. The figure shows what actually happens at the first zero.
Figure 7 The series at the first zero Settled
- Settled Panel A in words: SN ≈ ζ(s) + N1−s/(1 − s). The second piece has length √N / |1 − s| and turns as N grows, so the sums circle the true value ζ(s) on a widening spiral. At a zero the centre of that spiral is exactly 0. Subtracting the drifting piece (the gold track) is one standard way to compute ζ on the line.
- Settled Panel B in words: the alternating series η(s) = 1 − 2−s + 3−s − … does converge for Re(s) > 0, and η(s) = (1 − 21−s) ζ(s). The factor in front is not zero at ρ1, so η vanishes exactly where ζ does. That gives a second, independent way to see the same zero.
- Settled Either route defines the same analytically continued ζ. The zeros on this page are zeros of that function, which is why they can be computed and checked to many digits.
Chop the series off: the zeros leave the line
A natural first guess is to cut the series after N terms, ζN(s) = 1 + 2−s + … + N−s, and look for its zeros instead. That gives a different function, and its zeros behave differently. We found every zero of ζN up to height 60 for N = 10, 30 and 100.
Figure 8 Zeros of the chopped-off series
Step N up: the zeros of the chopped series jump around instead of landing on ½
One frame per N, with no in-between motion: each new term makes a new function with its own zeros. Gold rings are the true zeros on ½; for N = 2…40 no green dot comes closer than 0.062 to one.
N = 40 · 23 zeros in this window · the nearest to ½ + 14.13i is 0.380 away
Settled Zeros of ζN(s) = Σn≤N n−s with −1.5 ≤ Re s ≤ 1.9, 0 < t ≤ 40, for every N = 2…40. Found by Newton’s method in double precision; for each N the count equals an argument-principle count around the window, and each zero was re-found at 30 digits by mpmath (agreement 4·10−14). Frames are discrete: the dots do not travel from one N to the next. Finite N and a finite window only.
What the picture shows. Next to every zero of ζ (red ticks) there is a zero of ζN (filled dot), within 0.4 of it. But it is not on the line: its real part wobbles between about 0.24 and 0.74. The reason is the drifting piece from the spiral above: roughly, ζN(s) is ζ(s) plus a correction of size about N1−σ/t, and that correction nudges each zero sideways instead of removing it. On top of these there are extra zeros (hollow dots): 9, 20 and 31 of them. In total ζN has about (T/2π)·log N zeros up to height T (22.0, 32.5 and 44.0 here), more than ζ’s 13, and the extras form a band that moves toward σ = 1 as N grows.
- Settled Turán (1948) showed that if ζN had no zeros with σ > 1 for all large N, the Riemann Hypothesis would follow. Montgomery (1983) showed that this fails: for large N there are zeros with σ > 1 + c·log log N / log N for any c < 4/π − 1, and Montgomery and Vaughan (2001) showed 4/π − 1 is the right constant. Platt and Trudgian (2016) settled the small cases: there are no zeros with σ > 1 exactly for N ≤ 18 and N = 20, 21, 28. Those zeros are high up; none appear in our window.
- Settled The count of zeros of ζN up to height T is (T/2π)·log N + O(N) (Gonek and Ledoan, 2010).
- Settled Observed here The pairing of filled dots with zeta zeros and the drift of the extra band are what this finite computation shows for three values of N up to height 60. The sentence about the correction term is our reading via the Euler–Maclaurin formula, which is only a rough guide once t is larger than about πN.
- Open None of this says anything about whether the zeros of ζ itself are all on the line.
How big is ζ on the line, on average?
The zeros are where |ζ(1/2 + it)| drops to nothing, but between them it can be large, and the peaks slowly grow. A natural question is how large it is on average. Squaring and averaging gives a clean answer: up to height T, the average of |ζ(1/2 + it)|² is very close to log(T/2π) + 2γ − 1, where γ ≈ 0.5772 is Euler’s constant. Hardy and Littlewood found the leading term log T in 1918, and Ingham found the next one in 1926. None of this depends on the Riemann Hypothesis.
Figure 9 The average size of ζ on the line Settled
- Settled Atkinson (1949) wrote E(T) exactly as a sum of oscillating terms, a cousin of the explicit formula on the zeros and primes page. That is why the bottom curve looks like a wave rather than noise. Good (1977) proved E(T) is sometimes as large as a constant times T1/4, and it is known to be at most about T1/3.
- Open Whether E(T) grows no faster than T1/4 + ε for every ε > 0 is a standing conjecture.
- Settled The fourth power is also known: Ingham (1926) showed its average grows like (log T)4/(2π²). Open For the sixth and higher powers only conjectures exist; the leading constants predicted by Conrey and Ghosh, and in general by Keating and Snaith using random matrices, match numerical checks but are unproved. This is the same random-matrix picture shown on the zero statistics page.
How big is ζ at a random point? Selberg’s bell curve
The average above says how big |ζ(1/2 + it)| is on average. A sharper question is how its size is spread out: pick a height t at random, and how likely is ζ there to be tiny, ordinary, or huge? Atle Selberg answered this in the 1940s. Take the logarithm of the size, log|ζ(1/2 + it)|, and divide by √(½ log log T). As T grows, the values you get line up with the familiar bell curve (the normal distribution with average 0 and spread 1). This is Selberg’s central limit theorem. It is proved without assuming the Riemann Hypothesis.
The logarithm turns the product over primes into a sum over primes, and the theorem says that sum behaves like many small independent random pushes, which is exactly what produces a bell curve. The zeros show up too: right at a zero, log|ζ| plunges to minus infinity, so heights close to a zero make a long tail on the left.
Figure 10 Selberg’s bell curve for |ζ|
- Settled Selberg’s central limit theorem, for both log|ζ| and the angle of ζ (Selberg 1946; full proofs in Tsang’s 1984 thesis; a short modern proof by Radziwiłł and Soundararajan, 2017).
- Settled Observed here The data have the bell’s centre and rough shape, the leftward lean eases from -1.08 to -0.77 as the height grows, but the spread stays near 1.21–1.26. That is not a contradiction: the theorem is a statement about the limit, and the normaliser grows with log log T, which is only 2.0, 2.5, 2.8 here. A correction of fixed size, which the theorem allows, is still a large share at such heights. Reaching log log T = 5 would need T near 1064.
- Open How large ζ can get at its very biggest along the line, and the exact shape of the rare extreme values, are still being worked out. Conjectures from random matrices (Keating and Snaith, 2000) and from Fyodorov, Hiary and Keating (2012) predict them; parts have since been proved, but not all.
Hardy’s Z(t): why sign changes certify zeros on the line
Settled |ζ| touches zero at a zero, but a modulus cannot change sign, so a plot of |ζ| alone cannot tell a true zero from a near miss. The functional equation fixes that. Define the Riemann–Siegel theta function θ(t) and set
Z(t) = e^(iθ(t)) ζ(1/2 + it), with θ(t) = arg Γ(1/4 + it/2) − (t/2) log π.
Because Z is real and continuous, every sign change of Z between two values of t traps a zero of ζ exactly on the critical line, by the intermediate value theorem. No rounding can move such a zero off the line, because a zero off the line would not show up as a sign change of Z at all. This is the starting point of large-scale verifications. Checked here: at t = 5, 17.5, 33.3, and 55.5 the imaginary part of eiθζ(1/2+it), computed independently, was below 10−24.
Figure 11 Hardy Z(t) and Gram points Settled
mpmath.siegelz (mpmath 1.4.1, 25 digits) on a grid of spacing 0.05 for 0 ≤ t ≤ 60. Thirteen sign changes; refining each with a root finder reproduced mpmath.zetazero for n = 1…13, from 14.134725 to 59.347044. Z(0) ≈ −1.460, and on this grid the largest |Z| is about 3.665, at t = 45.65 (the same peak as the |ζ| figure above, now with its sign: Z is negative there). Gold: the Gram points g−1 ≈ 9.667 through g11 ≈ 57.545, where θ(gn) = nπ. In this window (−1)nZ(gn) > 0 at every one, which is “Gram’s law.” It is a heuristic, not a theorem: scanning n = 0…400, the first failure is at n = 126 (g126 ≈ 282.455, Z ≈ −0.028), the failure Hutchinson found in 1925, and twelve fail in that range. Settled Thirteen zeros on the line. Open Not a statement about the rest.Watch Z(t) change sign
Z(t) is real, so a zero on the line shows up as a sign change you can see. Each gold tick is one certified zero.
t = 60.00 · Z(t) = 0.587 · sign changes so far: 13
Settled mpmath.siegelz at 25 digits on a grid of step 0.05 (1,201 values). Each sign change was refined with a root finder and agrees with mpmath.zetazero for n = 1…13 to double precision. A sign change proves a zero on the line; it cannot rule out zeros off it.
Counting zeros: why sign changes are not enough by themselves
Settled Settled method Sign changes of Z find zeros on the line. They cannot see a zero that has left it. So verification compares two counts up to a height T: the number of sign changes of Z on [0, T], and the total number N(T) of zeros in the whole critical strip up to T, which comes from the argument principle and is pinned down rigorously by Turing’s method. If the two counts agree, every zero up to T is on the line and simple. Platt and Trudgian’s bound, quoted below, is this comparison carried out with interval arithmetic.
Figure 12 N(T) against its main term Settled
mpmath.nzeros: N(100) = 29 and N(200) = 79 both ways. The smooth term (T/2π) log(T/2πe) + 7/8 gives 29.002 at T = 100 and 79.193 at T = 200. On 8 ≤ T ≤ 200 the difference never exceeds 0.998 in absolute value, and its average at mid-step is about 0.004. That difference is essentially S(T); it is known to be unbounded, but grows so slowly that it stays tiny at any height a plot can show. Settled A count on a finite range.Count the zeros as the height grows
Sweep the height T upward. The count N(T) jumps by one at each zero, and the smooth formula tracks it; the gap between them is essentially S(T), which stays small on this range.
T = 100.00 · N(T) = 29 · smooth term = 29.002 · difference = −0.002
Settled Zero ordinates from mpmath.zetazero; the staircase agrees with mpmath.nzeros at T = 20, 40, 60, 80, 100, and with Odlyzko’s table used in the figure above. The smooth curve is drawn from T = 8, where the formula is meaningful. A count on a finite range.
S(T) over a longer range
Settled Settled theory, computed range The remainder S(T) = N(T) − θ(T)/π − 1 is exactly (1/π) arg ζ(1/2 + iT). It jumps up by 1 at each zero and slides down between zeros. Two classical facts frame the next figure. Its average is zero: the integral of S from 0 to T is O(log T). And Selberg (Arch. Math. Naturvid. 48, 1946) proved, unconditionally, that its mean square grows like a double logarithm:
(1/T) ∫₀^T S(t)² dt = (1/2π²) log log T + o(1).
Figure 13 S(T) and Selberg’s law Settled
mpmath.siegeltheta to 10−9). Range: −1.573 to +1.622, the maximum near T ≈ 71 129; |S| exceeds 1 on 1.24% of the range. S first reaches −1 at t ≈ 282.455, just before the 127th zero at 282.465. That is the same spot where Gram’s law first fails, at g126 ≈ 282.455 (see the Z(t) figure). The average of S over the whole range is 2·10−6. Bottom: the time-average of S² in six windows of T (14–50, 50–200, 200–1000, 1000–5000, 5000–20 000, 20 000–74 921) is 0.107, 0.129, 0.149, 0.164, 0.174, 0.182. Near the 1012-th zero it is 0.233, and near the 1022-th it is 0.266. At those heights Selberg’s leading term gives 0.166 and 0.197. Those two values use exact zero indices from Odlyzko’s tables, with θ evaluated by mpmath at 30 digits. The data sit above the leading term by an amount that creeps from about 0.05 to about 0.07. That is what a lower-order constant would look like. It is not evidence for or against RH. Settled Finite ranges, standard theory.Once a zero shows up as a sign change, how many zeros can you certify at every height at once? That is the question behind How much of the critical line can we prove? (Hardy–Littlewood windows, Selberg’s mollifier, Levinson’s count), which builds on this section.
Gram’s law at three heights
Settled Computed counts, proved theorems about failure The gold Gram points on the Z(t) figure chop the line into Gram intervals, one zero’s worth of room each on average, because θ(t) advances by π across each interval. Gram’s law is the tendency for each interval to hold exactly one zero. It is a bookkeeping heuristic. Trudgian (Acta Arith. 148, 2011) proved that it fails for a positive proportion of Gram intervals. So the right question is how often it holds, and how that changes with height.
Figure 14 Gram intervals at three heights Settled
mpmath.grampoint: g126 = 282.45472 both ways), giving 99 998 intervals below the 100 000th zero. High windows: θ evaluated by mpmath at 40 digits, 9 998 intervals each. Exactly one zero: 79.4%, 70.6%, 69.1%. Empty: 10.3%, 14.8%, 15.6%. Two zeros: 10.2%, 14.3%, 15.1%. Three zeros: 86, 22 and 23 intervals. Low down the books balance exactly, as they must when there is one zero per interval on average: 10 329 empty intervals against 10 157 doubles plus twice 86 triples. Gram’s law holds less often as the height grows, on this evidence. The trend is consistent with Trudgian’s theorem, but three windows do not measure its rate. Settled Counts on finite ranges. Open Says nothing about zeros off the line: a Gram interval with no zero only means its zero landed in a neighbour.The hand sketch on the Riemann Hypothesis page is a different object. It is not to scale, and it is not these tables.
Why these zeros matter for the primes is the next page: zeros and primes. How they are spaced is on statistics of the zeros.
[project] Beyond the plane: the critical line in quaternions and octonions—an illustrated, numerically checked exploration of zero spheres and what they do not prove.
Reading path · step 3 of 13