Zeta Zeros
Display
Theme
Depth
Motion

One conjecture about the zeros

The Riemann Hypothesis

In 1859 Bernhard Riemann suggested that every nontrivial zero of ζ(s) has real part 1/2. It is still open. This page keeps the status straight: what is proven, what is computed here, what is open, and what is the project’s own reading.

Start here

In 1859 Riemann suggested that every nontrivial zero of ζ(s) has real part ½. Open The Hypothesis is unproved. It is a Millennium Prize problem; nothing on this site claims to settle it.

Settled Checking has reached every zero up to height 3·1012, and each lies on the line (Platt and Trudgian, 2021). That is a finite check. Proofs about all the zeros say less: no zeros on the line Re(s) = 1, a thin zero-free region beside it, Hardy’s infinitely many zeros on the critical line, and later theorems that a positive proportion of zeros lie on the line.

Settled The same open question can be restated without naming the zeros of ζ: as a claim about zeros of ζ′, about divisor sums (Robin), about harmonic numbers (Lagarias), about a power series built from ζ(2), ζ(4), ζ(6), … (Riesz), about how large the Mertens function can grow, about Li’s numbers staying positive, about Jensen polynomials having real roots, or about one constant under heat flow (de Bruijn–Newman). Each restatement is a theorem of equivalence; none makes the Hypothesis easier, and none is a proof.

If you only read one thing: Honest status. The first sentence there is the whole point: “The Riemann Hypothesis is unproved.”

Stage 1 Where things stand

At a glance: proven, open, metaphor

Four kinds of sentence on this site. Read the flag before reading the sentence. Added 6 October 2026 PT.
KindExamples on this page
Settled ProvenEuler product; analytic continuation and functional equation; trivial zeros; no zeros on Re(s) = 1 (prime number theorem); explicit formula for whatever zeros exist; Riemann–von Mangoldt count; Hardy’s infinitely many zeros on the line and the positive-proportion theorems; verification of every zero up to height 3·1012 (Platt–Trudgian).
Settled Computed hereOrdinates, |ζ| and Hardy Z(t) on short windows, the N(T) staircase and its remainder S(T) up to the 100 000th zero, π(x) against li(x) and R(x), the explicit-formula truncations, Gram-interval counts, an interactive explicit-formula truncation, and spacing, pair-correlation and number-variance statistics from Odlyzko’s tables, low down and near the 1012-th and 1022-th zero. Finite checks. Each is evidence about the range drawn and nothing beyond it.
OpenThe Riemann Hypothesis itself; Montgomery’s pair-correlation and GUE spacing conjectures; Hilbert–Pólya; Lindelöf.

Honest status

Open Open problem The Riemann Hypothesis is unproved. It is a Millennium Prize problem of the Clay Mathematics Institute. Bombieri’s official problem description is the short orientation written for that list. Surveys that stay close to the proofs: Conrey’s article in the Notices of the American Mathematical Society (2003), and the books of Titchmarsh (Heath-Brown’s revision), Edwards, and Ivić. Wikipedia is a pointer. It is not an authority for this page.

A proposed geometric reading of the same open question — zeros as eigenvalues of a flow on a foliated space, and RH as a positivity or Hodge-type condition on a still unconstructed cohomology — is described, with its sources and its limits, on Foliations and Deninger’s picture. That page does not claim any progress on the Hypothesis.

Stage 2 What checking and proofs each reach

How far the checking has reached

The Hypothesis is a statement about infinitely many zeros, so no computation can prove it. What computation can do is check the zeros one height at a time, and it has been doing that since Riemann’s own notes. The method that makes the checks rigorous is on the critical line page: find sign changes of Hardy’s Z(t), then use Turing’s method to confirm that the count of zeros up to that height matches, so none can be hiding off the line.

Figure 1 How many zeros have been checked Settled

A timeline from 1850 to 2030 with a logarithmic vertical axis from 1 to beyond 100 trillion zeros. A dark green staircase climbs from Riemann's 3 zeros, through Gram's 15 in 1903, Backlund's 79 in 1914, Titchmarsh and Comrie's 1,041 in 1936 (the end of the shaded hand-calculation era), Turing's 1,104 in 1953 (the first electronic computation), Lehmer's 25,000 in 1956, Rosser, Yohe and Schoenfeld's 3.5 million in 1968, and van de Lune, te Riele and Winter's 1.5 billion in 1986, to Gourdon and Demichel's 10 trillion in 2004 and Platt and Trudgian's roughly 12.4 trillion in 2021. Two hollow circles mark van de Lune's unpublished 10 billion in 2001 and the ZetaGrid project's roughly 0.9 trillion in 2004.
Record counts of nontrivial zeros shown to lie on the critical line, 1859 to 2021, from the table of numerical calculations in the English Wikipedia article “Riemann hypothesis” (read 19 October 2026 PT), which cites the original papers; Riemann’s count of 3 is as reported by Siegel (1932), and its date is uncertain. Odlyzko’s computations of zeros near heights 1020 to 1022 are left out because they check zeros in windows high up, not every zero from the bottom. Two counts are checked here with mpmath 1.4.1: nzeros(200) returns 79 and nzeros(300) returns 138, matching Backlund (1914) and Hutchinson (1925). The Riemann–von Mangoldt main term (T/2π) log(T/2πe) + 7/8 at T = 3·1012 gives 1.23624·1013, matching Platt and Trudgian’s count of about 1.2363·1013. The staircase draws itself on load; motion is off if your system asks for reduced motion. Settled Every zero in these ranges is on the line and simple.
  • Settled The jump in 1953 is Turing: the first time a digital computer was used on the zeros, and the first use of his counting trick, which is still the standard way to rule out zeros hiding off the line.
  • Settled The vertical scale is logarithmic, so the steady climb since the 1950s means the count has grown about tenfold every seven years since Turing (1,104 zeros in 1953 to about 1.2·10¹³ in 2021), far faster than the hand-calculation era before it.
  • Open Twelve trillion zeros on the line is strong evidence, but it is not a proof. Number theory has famous patterns that hold far beyond any computer’s reach and then fail, for example π(x) < li(x), which is true for every x ever computed and still false somewhere, as Littlewood proved in 1914.

What is proved, and what is only checked

Checking zeros one by one only ever reaches a finite height. Proofs work differently: they say something about all the zeros at once, but so far they say less. This figure puts the two side by side. The top panel shows where zeros are proved not to be, and where they have been checked. The bottom panel shows what share of all the zeros, out to infinity, is proved to sit on the line.

Figure 2 What is proved, and what is only checked

Two panels. The top panel is the critical strip, with the real part sigma from 0 to 1 running up the page and the height t running across on a logarithmic scale from 2 to 10 to the 16. Thin dark red slivers hug the edges sigma = 1 and sigma = 0: the proved zero-free region, 0.068 wide at the first zero, 0.013 wide at height one million and only 0.0063 wide at 3 times 10 to the 12. The middle line sigma = 1/2 carries the first 100 computed zeros, and the whole strip up to 3 times 10 to the 12 is tinted green because every zero there has been checked to be on the line. Beyond that height the strip is hatched: open. Dashed gold lines at sigma = 7/8 and 1/8 mark a zero-free bound claimed in October 2026 but not yet verified. The bottom panel is a bar chart of the proved share of zeros on the line: Hardy 1914, infinitely many but 0 percent; Selberg 1942, a small positive share; Levinson 1974, more than one third; Conrey 1989, more than two fifths; Pratt, Robles, Zaharescu and Zeindler 2020, more than five twelfths; and a hollow dashed bar for Alpöge and Furman's claimed two thirds in 2026. A line at 100 percent marks what the Hypothesis says.
Top panel: the zero-free region is the simplest of the four explicit regions proved by Mossinghoff, Trudgian and Yang (Research in Number Theory 10, 2024; arXiv 2212.06867): no zero has σ ≥ 1 − 1/(5.558691 log t) once t ≥ 2. The functional equation mirrors it to the edge σ = 0. The widths are computed from that formula, and the first 100 zeros with mpmath 1.4.1 (zetazero(1) = 1/2 + 14.1347…i, zetazero(100) = 1/2 + 236.524…i). The verified height 3.0001753328·1012 is Platt and Trudgian (2021). Bottom panel: proved lower bounds for the share of zeros on the line, as summarised in the English Wikipedia article “Riemann hypothesis” (read 20 October 2026 PT), with the 5/12 result from arXiv 1802.10521. The hollow bar is the August 2026 preprint arXiv 2608.13637 by Alpöge and Furman, which states at least two thirds (0.6725 with a further refinement); the dashed σ = 7/8 lines are an October 2026 claim, released by OpenAI with a Lean formalization, that no zeros have real part above 7/8. Wikipedia describes both as claimed proofs, and says the 7/8 proof had not been independently verified as of October 2026, so both are drawn hollow and dashed. Bars grow and labels fade in on load; motion is off if your system asks for reduced motion.
  • Settled The red slivers are thin, and they get thinner as you go up, so the proved zero-free region shrinks toward nothing. Proving the Hypothesis would mean widening them until they meet in the middle. Even a fixed-width strip, say no zeros with σ > 0.99, would be a historic result.
  • Settled Hardy (1914) showed infinitely many zeros lie on the line, but infinitely many can still be 0% of them. Selberg (1942) was the first to prove a positive share. Levinson (1974), Conrey (1989) and Pratt, Robles, Zaharescu and Zeindler (2020) pushed the share past one third, two fifths and five twelfths.
  • Open Claimed The two 2026 results would be large jumps if they hold up: a share of two thirds instead of five twelfths, and a fixed zero-free strip to the right of 7/8. Neither is the Hypothesis, which needs 100% and nothing off the line at all. This site will move them to settled only when they are independently confirmed.

Stage 3 The same question, restated

The same question, asked of the derivative

The Hypothesis can be restated without mentioning the zeros of ζ at all. In 1934 Andreas Speiser proved that it is equivalent to a statement about the derivative ζ′: the Hypothesis holds if and only if ζ′(s) has no zeros with 0 < Re(s) < 1/2. So you can test it by looking for where the slope of ζ vanishes, and checking that none of those points falls in the left half of the strip.

Figure 3 Where the derivative ζ′ vanishes Settled

A chart with the height t from 0 to 100 running across and the real part sigma from 0 to 3 running up. The 29 zeros of zeta up to height 100 are dark dots on the line sigma = 1/2. The 19 zeros of its derivative zeta-prime in the same range are gold diamonds scattered above that line, with real parts between 0.78 and 2.46; the first is at 2.4632 + 23.2983 i. The band between sigma = 0 and sigma = 1/2 is shaded and empty, labelled: no zeros of zeta-prime, argument-principle count 0. Dots and diamonds appear in order of height on load.
Computed with mpmath 1.4.1 at 20 digits, 21 October 2026 PT. The 19 zeros of ζ′ with 1 ≤ t ≤ 100 were located with Newton’s method (findroot on zeta(s, derivative=1)) from a grid of starting points, and the count was checked independently with the argument principle: the winding of ζ′ around the rectangle 1/2 ≤ σ ≤ 4, 1 ≤ t ≤ 100 is 19, and around 0.02 ≤ σ ≤ 1/2, 1 ≤ t ≤ 100 it is 0. The first one, 2.463162 + 23.298320i, agrees with the published value. The 29 zeros of ζ are from zetazero. Points appear in order of height on load; motion is off if your system asks for reduced motion. Settled Speiser’s equivalence is a theorem; the picture is a check up to height 100 only.
  • Settled Speiser’s argument is geometric: he studied the curves on which ζ(s) takes real values and how they pass through the zeros. Levinson and Montgomery (1974) gave an analytic proof and a counting version: ζ and ζ′ have, up to a small error, the same number of zeros in the left half of the strip.
  • Settled Levinson’s proof the same year that more than one third of the zeros lie on the line, the bar labelled Levinson 1974 in the figure above, also works with ζ and ζ′ together.
  • Settled Levinson and Montgomery also showed that, on average, zeros of ζ′ crowd toward σ = 1/2 as t grows, though slowly; here the closest is 0.78, at height 95.3. Several sit to the right of σ = 1, outside the critical strip, which is allowed; only the band 0 < σ < 1/2 matters for the Hypothesis.

The same question, asked of whole numbers

Here is a version of the Hypothesis that needs nothing beyond whole numbers and division. Take a number n and add up all of its divisors: for 12 that is 1 + 2 + 3 + 4 + 6 + 12 = 28. Call the total σ(n). Numbers with many divisors, like 12, 360 or 5040, have a big σ(n) compared with n; primes have the smallest possible, just n + 1.

How big can σ(n) get? In 1913 Thomas Gronwall found the exact growth rate of the record-holders: σ(n) can reach about eγ · n · ln ln n, where γ ≈ 0.5772 is Euler’s constant, so eγ ≈ 1.7811, and no faster. In 1984 Guy Robin proved something surprising: the Riemann Hypothesis is true if and only if σ(n) < eγ · n · ln ln n for every n bigger than 5040. A question about the zeros of ζ turns into a question about divisor sums, with no complex numbers in sight.

Figure 4 Robin’s inequality: divisor sums

A chart over whole numbers n from 3 to ten million on a log scale. For each n the height is the sum of the divisors of n divided by n times ln ln n. A shaded band shows the spread, from primes along the bottom up to a dark jagged upper edge. A dashed red line marks the ceiling e to the power gamma, about 1.7811. Red dots for 26 small numbers, ending at 5040, sit on or above the ceiling. Past 5040 the band's upper edge stays below the line; gold dots mark the highest point in each power-of-ten range: 10,080 at 1.7558, 110,880 at 1.7348, 1,441,440 at 1.7277. The chart fades in on load.
Computed 29 October 2026 PT with a divisor-sum sieve (numpy 2.5.4) for every whole number n from 3 to 107, about 12 seconds; spot checks σ(12) = 28, σ(28) = 56, σ(5040) = 19,344. The band shows, for each of about 400 slices of the log scale, the lowest, middle and highest value of σ(n) / (n ln ln n). The 26 numbers from 3 to 5040 on or above the ceiling are 3, 4, 5, 6, 8, 9, 10, 12, 16, 18, 20, 24, 30, 36, 48, 60, 72, 84, 120, 180, 240, 360, 720, 840, 2520, 5040. From 5041 to 107 every n is below the ceiling, and the highest value is at n = 10,080 (1.7558). Fade-in is off if your system asks for reduced motion.

Two things in the picture are worth noticing. The red dots stop at 5040, which is why Robin’s statement starts at 5041. And past 5040 the highest points do not rush toward the ceiling; in this range they even sag a little, from 1.7558 at 10,080 to 1.7277 at 1,441,440. Gronwall’s theorem says they must eventually climb back and come as close to the ceiling as you like, but only at astronomically large n. That is why checking up to ten million, or far beyond, cannot settle the question: the danger zone lies out among numbers with thousands of digits and more.

  • Settled Gronwall (1913): the largest values of σ(n) / (n ln ln n) approach eγ as n grows (the limit superior is eγ).
  • Settled Robin (1984): the Riemann Hypothesis holds if and only if σ(n) < eγ n ln ln n for all n > 5040. Robin also proved, with no hypothesis, a slightly weaker bound that holds for every n ≥ 3. Lagarias (2002) turned this into an even plainer equivalent using the harmonic numbers 1 + 1/2 + … + 1/n.
  • Settled Observed here No n from 5041 to 107 breaks the inequality. Far larger checks exist: Briggs (2006) tested the most dangerous candidates (the colossally abundant numbers) much further out. None of this proves the Hypothesis, which remains open.
  • Settled Robin also showed that if the Hypothesis fails, then some colossally abundant number breaks the inequality. These are numbers built like 5040 = 24 · 32 · 5 · 7, packed with small primes, and they are why searches focus on such numbers.

The same question, asked with harmonic numbers

Robin’s inequality uses the growth rate of the divisor sum σ(n). Jeffrey Lagarias found a close cousin that needs only the harmonic numbers Hn = 1 + 1/2 + … + 1/n — the numbers you meet in calculus, with no Euler–Mascheroni constant written out. For every whole number n ≥ 1 define the elementary bound

B(n) = Hn + exp(Hn) · ln(Hn).

Lagarias’s ceiling. At n = 1 one has H1 = 1 and ln(1) = 0, so B(1) = 1.

Lagarias proved in 2002 that the Riemann Hypothesis is true if and only if σ(n) ≤ B(n) for every n ≥ 1, with equality only at n = 1. So the same open question about the zeros of ζ becomes a claim about divisor sums and harmonic numbers alone. It does not prove RH: checking the inequality for many n is a check of an equivalent form, not a proof. Next to Robin’s inequality (equivalent to RH for every n > 5040), Lagarias’s form is the plainer elementary cousin — one bound that covers every n from 1 on.

Figure 5 Lagarias: harmonic-number form of RH

Two panels. Top: for every whole number n from 1 to 10,000,000 on a log scale, the height is σ(n) divided by B(n), where B(n) = H_n + exp(H_n) ln(H_n) and H_n is the n-th harmonic number. A dashed red line marks the bound 1. A red dot at n = 1 sits exactly on the line (equality). Every other point stays below it; the closest approach is n = 12 at 0.9886. Bottom: the twelve highest ratios after n = 1, as bars. The chart fades in on load.
Computed 6 November 2026 PT with a divisor-sum sieve and cumulative harmonics (numpy 2.5.4, Python 3.13.5) for every whole number n from 1 to 107, about 12 seconds. Sanity: at n = 1, σ(1) = 1, H1 = 1, B(1) = 1, so the ratio is exactly 1 (equality, as Lagarias requires). For n ≥ 2 every ratio is strictly below 1; the closest approach in this run is n = 12 at 0.988636. The next closest among highly composite n are 120 at 0.9834, 60 at 0.9826, 2,520 at 0.9783, 5,040 at 0.9752. Fade-in is off if your system asks for reduced motion.

What the picture is saying. The top panel is the Lagarias ratio σ(n)/B(n) against n. The dashed red line is the bound 1. The single red point at n = 1 sits on the line; everything else in the run sits below it, and the gold dots mark the highest ratio in each power-of-ten range. The bottom panel zooms on the twelve closest approaches after n = 1 — small, highly composite numbers such as 12, 60, 120, 2520 and 5040. As n grows the band sags away from the ceiling in this range, the same qualitative lesson as Robin’s picture: the finite check is easy, and the danger (if any) lies far beyond ten million.

  • Settled Lagarias (2002): the inequality σ(n) ≤ Hn + exp(Hn) ln(Hn) for every integer n ≥ 1, with equality only at n = 1, is equivalent to the Riemann Hypothesis.
  • Settled Relation to Robin (1984): Robin’s inequality σ(n) < eγ n ln ln n for all n > 5040 is also equivalent to RH. Lagarias’s form is a close elementary cousin that starts at n = 1 and writes the ceiling with harmonic numbers alone.
  • Settled Observed here Every n from 1 to 107 satisfies the inequality; equality holds only at n = 1; the closest approach afterward is n = 12 at 0.9886. That is a finite check of an equivalent form, not a proof.
  • Open The Riemann Hypothesis itself. Nothing on this page claims it is settled; see Honest status.

The same question, asked with ζ(2), ζ(4), ζ(6), …

In 1916 Marcel Riesz built a function out of nothing but the values of ζ at the even numbers 2, 4, 6, … — numbers Euler already knew, such as ζ(2) = π²/6. For a positive number x he set

F(x) = x/ζ(2) − x²/(1! · ζ(4)) + x³/(2! · ζ(6)) − x⁴/(3! · ζ(8)) + … = Σk≥1 (−1)k+1 xk / ((k−1)! · ζ(2k)).

Riesz’s function. The same function can be written as F(x) = x · Σn≥1 μ(n) n−2 e−x/n², where μ is the Möbius function from the Mertens section.

A way to see what F measures. Every ζ(2k) is a little bigger than 1, and it gets closer to 1 as k grows. If every ζ(2k) were exactly 1, the series would add up to x · e−x, which shrinks to almost nothing as x grows: huge positive and negative terms cancel almost perfectly. Because the ζ(2k) are not quite 1, the cancellation is not perfect, and F(x) is what is left over. Riesz proved that the size of that leftover is controlled by the zeros of ζ.

His theorem: the Riemann Hypothesis is true if and only if F(x) grows no faster than x1/4 + ε for every ε > 0 — in symbols, F(x) = O(x1/4+ε). “O(…)” means “at most a fixed multiple of …, once x is large enough”. Without assuming anything, one can prove the weaker ceiling F(x) = O(x1/2+ε) (Titchmarsh, §14.32). So the open question becomes: is the true growth rate of this leftover the exponent ¼, or something larger, up to ½?

Why ¼? Each zero ρ = β + iγ of ζ adds a wave to F(x) of size about xβ/2. The “iγ” part does not change the size: xiγ/2 = cos(½γ ln x) + i sin(½γ ln x) is a point turning around the unit circle, so it only makes the term swing back and forth. If every zero has β = ½, every wave has size about x1/4. A single zero with β bigger than ½ would add a wave that grows faster than x1/4, and F would eventually break the bound. Riesz himself was cautious about how much this helps; he wrote that he could not yet decide whether the condition would make the hypothesis easier to verify.

Figure 6 Riesz: a power series in ζ(2k)

Two panels. Top: on log-log axes, the size of Riesz’s function |F(x)| for 361 values of x from 1 to one million, with two dashed reference lines, x to the one-half (red) and x to the one-quarter (gold). |F| rises to about 0.82 near x = 7, then falls along a dotted curve 16.42/x, and after about x = 20,000 turns into a wave whose height grows like 7.78e-05 times x to the one-quarter, far below both reference lines. Bottom: that wave alone, divided by x to the one-quarter: a steady oscillation of constant height between x = 3,000 and one million. The chart fades in on load.
Computed 9 November 2026 PT with mpmath 1.4.1 (Python 3.13.5) at 361 values of x, 60 per decade from 1 to 106, about 53 seconds. The alternating series needs many digits because its terms swell to about ex before cancelling: at x = 1000 it took 3,727 terms at 514 digits. For the whole range we used the Möbius form (terms with n ≤ 20,100 summed directly, the rest expanded exactly through 1/ζ), at 30 and at 50 digits; the two agree to a relative 1e-26, and at x = 1, 2, 5, …, 1000 they agree with the power series itself (run at two precisions) to better than 10−25. Examples: F(1) = 0.043982, F(10) = -0.780676, F(1000) = -0.016638, F(106) = -1.775e-03. Fade-in is off if your system asks for reduced motion.

What the picture is saying. In the top panel the green curve is |F(x)|, and both axes are logarithmic, so a power of x shows up as a straight line. The red and gold dashed lines are x1/2 and x1/4. Across the whole run the largest value of |F(x)| is 0.820 near x = 7.1; the largest ratio |F(x)| / x1/4 is 0.519, near x = 5.6, and the largest |F(x)| / x1/2 is 0.347, near x = 4.5. After that, F falls along the dotted curve −16.42/x (it is negative there): this part comes from the trivial zeros −2, −4, … of ζ, not from the zeros on the critical line. Near x = 20,000 that falling part has shrunk enough for a wave to show: F starts changing sign (at about 19,953, 23,263, 42,987, 60,720, …), and the wave’s height rides along the dotted gold line 7.775e-05 · x1/4. Between 105 and 106 the largest |F(x)| / x1/4 is 8.101e-05.

The bottom panel takes away the falling part (the terms in 1/x, 1/x², 1/x³) and divides what is left by x1/4. If that leftover grew faster than x1/4, this curve would fan out. It doesn’t: it swings with a constant height of ±7.775 (in units of 10−5). That height, and the rhythm of the swing (one full swing each time x grows by a factor of about e4π/14.13 ≈ 2.4), come from the first zero of ζ, ½ + 14.1347… i. Marek Wolf found the same height, about 7.775 · 10−5, in 2008. A steady wave over x up to a million is exactly what RH predicts, but it is evidence for small x only: a zero far up the line, if one existed off it, would only start to matter at values of x far beyond any computation.

A cousin by Hardy and Littlewood. In their long 1918 memoir, Hardy and Littlewood proved by similar methods that RH is also equivalent to Σk≥1 (−x)k / (k! · ζ(2k+1)) = O(x−1/4+ε) for every ε > 0 — the same idea using ζ at the odd numbers 3, 5, 7, … instead. It is computed in the next section.

  • Settled Riesz (1916): the Riemann Hypothesis is equivalent to F(x) = O(x1/4+ε) for every ε > 0, with F(x) = Σ (−1)k+1 xk / ((k−1)! ζ(2k)). Unconditionally F(x) = O(x1/2+ε).
  • Settled Hardy and Littlewood (Acta Mathematica 41): RH is equivalent to Σ (−x)k / (k! ζ(2k+1)) = O(x−1/4+ε) for every ε > 0.
  • Settled Observed here From x = 1 to 106, |F(x)| / x1/4 never exceeds 0.519; beyond x ≈ 20,000 the oscillation has the constant height 7.775e-05 · x1/4 set by the first zero. The split into a 1/x part and a wave is our own residue bookkeeping, and it matches the computed F to within 2e-11 for x ≥ 3,000. That is a finite check of an equivalent form, not a proof.
  • Open The Riemann Hypothesis itself, the conjecture Riemann made in 1859. Nothing on this page claims it is settled; see Honest status.

The same question, asked with ζ(3), ζ(5), ζ(7), …

Two years after Riesz, G. H. Hardy and J. E. Littlewood found a twin of his test that uses ζ at the odd numbers instead. Unlike ζ(2) = π²/6, nobody has a neat formula for ζ(3), ζ(5), …; they are just numbers that can be computed to as many digits as we like. Hardy and Littlewood set

H(x) = −x/(1! · ζ(3)) + x²/(2! · ζ(5)) − x³/(3! · ζ(7)) + … = Σk≥1 (−x)k / (k! · ζ(2k+1)) = Σn≥1 μ(n)/n · e−x/n².

The Hardy–Littlewood function. The last form uses the Möbius function μ from the Mertens section; it hides a famous fact, that μ(1)/1 + μ(2)/2 + μ(3)/3 + … adds up to exactly 0.

The idea is the same as before. If every ζ(2k+1) were exactly 1, the series would add up to e−x − 1, and the huge positive and negative terms would cancel down to almost −1. Because ζ(3), ζ(5), … are slightly bigger than 1, the cancellation goes further, and H(x) is the small leftover. Their theorem: the Riemann Hypothesis is true if and only if H(x) shrinks at least as fast as x−1/4 + ε for every ε > 0. So here the question is not “how slowly does it grow?” but “how quickly does it die away?”

Where −¼ comes from. Each zero ρ = β + iγ of ζ adds a wave to H of size about x(β−1)/2. With β = ½ that is x−1/4; a zero with β bigger than ½ would add a wave that dies away more slowly, and H would eventually break the bound. The trivial zeros −2, −4, −6, … add a separate, quickly vanishing part, which starts with −14.55 · x−3/2.

Figure 7 Hardy–Littlewood: a series in ζ(3), ζ(5), …

Two panels. Top: on log-log axes, the size of the Hardy–Littlewood function |H(x)| for 361 values of x from 1 to one million, with a dashed gold reference line x to the minus one-quarter, which falls only slowly. |H| starts at 0.48, falls steeply along a dotted line 14.55 times x to the minus three-halves, and after about x = 40,000 turns into a wave whose height shrinks only like 2.93e-05 times x to the minus one-quarter. Bottom: that wave alone, multiplied by x to the one-quarter: a steady oscillation of constant height between x = 3,000 and one million. The chart fades in on load.
Computed 10 November 2026 PT with mpmath 1.4.1 (Python 3.13.5) at 361 values of x, 60 per decade from 1 to 106, about 52 seconds. As with Riesz’s function, the plain series needs many digits because its terms swell before cancelling: at x = 1000 it took 3,720 terms at 514 digits. For the whole range we used the Möbius form (terms with n ≤ 20,100 summed directly, the rest expanded exactly), at 30 and at 50 digits; the two agree to a relative 2e-24, and at x = 1, 2, 5, …, 1000 they agree with the series itself to better than 10−50. Examples: H(1) = -0.480534, H(10) = -0.267646, H(1000) = -4.581688e-04, H(106) = -9.372e-07. Fade-in is off if your system asks for reduced motion.

What the picture is saying. Up to about x = 10 000, H(x) is almost exactly the steep dotted line 14.55 · x−3/2 from the trivial zeros. That part shrinks faster than the wave from the zeros on the critical line, so near x ≈ 41,368 the wave takes over: H starts changing sign (first at about 41,368, 50,119, 92,612, …) and its height follows the dotted gold line 2.926e-05 · x−1/4. Between 105 and 106 the largest value of |H(x)| · x1/4 is 3.319e-05.

The bottom panel removes the trivial-zero part and multiplies what is left by x1/4. A wave dying away more slowly than x−1/4 would make this curve fan out; instead it swings with constant height ±2.926 (in units of 10−5), set by the first zero ½ + 14.1347… i, which is the same zero that set the height of Riesz’s wave. Our formula for H as “trivial part plus one wave per zero” (100 zeros used) matches the computed H to within 5e-20 · x−1/4 for x ≥ 3000. As with Riesz, this is what RH predicts for x up to a million, and it says nothing about zeros far up the line.

  • Settled Hardy and Littlewood (Acta Mathematica 41, 1918; see Titchmarsh §14.32): RH is equivalent to Σ (−x)k / (k! ζ(2k+1)) = O(x−1/4+ε) for every ε > 0.
  • Settled Observed here From x = 3000 to 106 the wave left after removing the trivial-zero part has constant height 2.926e-05 · x−1/4. The split into parts is our own residue bookkeeping, checked against the computed values. A finite check of an equivalent form, not a proof.
  • Open The Riemann Hypothesis itself. See Honest status.

A coin-flip walk over the whole numbers

Here is another way to hear the Hypothesis using nothing but factoring. Give every whole number a score μ(n), the Möbius function: +1 if n is a product of an even number of different primes (1, 6 = 2·3, 10 = 2·5), −1 for an odd number (2, 3, 30 = 2·3·5), and 0 if some prime divides n twice (4, 12, 18). Then walk along the numbers and keep a running total, M(x) = μ(1) + μ(2) + … + μ(x). This total is the Mertens function. It starts 1, 0, −1, −1, −2, −1, −2, … The function itself is not a hypothesis: it is a fixed count that anyone can compute. The open questions are about how large it can get. Mertens’ own guess about that size is the Mertens conjecture, which has been disproved, and RH is a different, looser bound on the same function, still open.

If the +1s and −1s behaved like fair coin flips, the total would drift about as far as √x, the way a random walk does. In 1912 John Littlewood showed that the Riemann Hypothesis is true if and only if M(x) grows no faster than x1/2 + ε for every ε > 0, so RH says the walk never strays much further than a coin-flip walk would. The figure divides M(x) by √x so that coin-flip-sized wandering stays at a steady height.

Figure 8 The Mertens function to one billion

A chart over x from 10 to one billion on a log scale. The height is the Mertens function M(x) divided by the square root of x. A shaded green band zigzags around zero, mostly between about minus 0.5 and plus 0.5, never settling. Dashed red lines at plus 1 and minus 1 mark Mertens' 1897 guess, which the band never touches in this range. Gold dots mark M at each power of ten: −1 at 10 to the 1, 1 at 10 to the 2, 2 at 10 to the 3, −23 at 10 to the 4, −48 at 10 to the 5, 212 at 10 to the 6, 1,037 at 10 to the 7, 1,928 at 10 to the 8, −222 at 10 to the 9. The chart fades in on load.
Computed 30 October 2026 PT with a segmented Möbius sieve (numpy 2.5.4) for every x from 1 to 109, about 54 seconds. The band shows the lowest and highest M(x)/√x in each of 150 slices of the log scale. Spot checks against published values: M(106) = 212, M(107) = 1,037, M(108) = 1,928, M(109) = −222. Up to 109, M first goes negative at x = 3, changes sign 42,971 times, and equals exactly 0 at 141,121 values of x. Fade-in is off if your system asks for reduced motion.

Franz Mertens computed M(x) by hand up to 10 000 in 1897 and guessed something much stronger than RH: that |M(x)| stays below √x forever, which would put the band inside the red lines for good. The picture makes that guess look safe. After x = 1000 the band never gets further than 0.472 from zero in this range. Yet the guess is false. Andrew Odlyzko and Herman te Riele proved in 1985, using the first 2000 zeros of ζ computed to high precision, that M(x)/√x climbs above 1.06 and below −1.009 somewhere. Nobody has found such an x; it is known only to lie below an enormous bound. It is the same lesson as the prime race and Robin’s inequality: a billion data points can be a very small sample.

  • Settled Littlewood (1912): RH holds if and only if M(x) = O(x1/2 + ε) for every ε > 0.
  • Settled Odlyzko and te Riele (1985) disproved the Mertens conjecture |M(x)| < √x. Pintz (1987) showed a counterexample exists below exp(3.21 × 1064); Kotnik and te Riele (2006) lowered that to below exp(1.59 × 1040). Hurst (2018) computed M(x) to 1016 and pushed the known swings of M(x)/√x past +1.826 and below −1.837.
  • Settled Observed here Up to 109 the band stays well inside ±1, as the published computations to 1016 also find. That supports nothing beyond this range, and the Hypothesis remains open.
  • Speculative Heuristic only The coin-flip picture is a guide, not a theorem: μ(n) is completely determined by factoring, not random. It does predict the right √x scale, and under RH plus further conjectures about the zeros, M(x)/√x is expected to be unbounded but to grow extremely slowly, like (log log log x)5/4 (a conjecture of Gonek, discussed by Ng 2004).

A ladder of numbers that must all stay positive

The last two sections turned the Hypothesis into statements about whole numbers. Xian-Jin Li found a version that is a single list of numbers, λ1, λ2, λ3, …, with one rule: the Riemann Hypothesis is true if and only if none of them is ever negative. They are called the Keiper–Li coefficients, after Jerry Keiper, who introduced them in 1992, and Li, who proved the “if and only if” in 1997.

Each λn is built from the zeros. Take a zero ρ (a complex number, a point in the plane) and form the number 1 − 1/ρ. Its size is the distance from ρ to 1 divided by the distance from ρ to 0. On the critical line ρ sits exactly halfway between 0 and 1 from left to right, so those two distances are equal and 1 − 1/ρ lands exactly on the circle of radius 1. Raising it to the n-th power just turns it further round that circle, and “1 minus a point on the circle” never has a negative real part. So zeros on the line can only push λn = Σρ [1 − (1 − 1/ρ)n] upward. A zero off the line comes with a mirror partner nearer to 0 than to 1; for that partner the point lies outside the circle, its powers spiral outward ever faster, and sooner or later one swing drags some λn below zero. That is the whole idea of the criterion.

You do not need the zeros to compute these numbers. The same λn are the coefficients of a power series for the completed function ξ, so they can be read off from known constants. That is what the figure does.

Figure 9 Li’s criterion: 500 positive numbers

Two stacked charts over n from 1 to 500. Top: one dot for each Keiper–Li number λ_n, rising smoothly from 0.0231 at n = 1 to about 992 at n = 500, lying almost exactly on a dashed gold curve that shows how fast the numbers must grow if the Riemann Hypothesis is true. An inset magnifies n = 1 to 20 as bars, all small but above zero. Bottom: the gap between the dots and the gold curve, a slow red wobble between about −1.6 and 4.7. The chart fades in on load.
Computed 2 November 2026 PT with mpmath 1.4.1: the Taylor series of log ξ(s) at s = 1 from the Stieltjes constants γ0, …, γ499, then the substitution s = 1/(1 − z) that turns it into λ1, …, λ500. The whole run was done twice, at 180 and at 240 significant digits, and the two agree to better than 10−31 for every n. Checks: λ1 = 1 + γ/2 − ½ log 4π = 0.0230957089661…; λ2 = 0.092345735228…; λ1 to λ5, λ10, λ30 and λ100 = 118.6037753768… match published 50-digit tables; and summing 1 − (1 − 1/ρ)n directly over Andrew Odlyzko’s first 100 000 zeros (plus a small estimate for the rest) reproduces every λn here to within 4·10−5; the small gap comes from approximating the zeros beyond the table. Fade-in is off if your system asks for reduced motion.

Two things stand out. First, every value is positive, including the tiny first few (λ1 ≈ 0.023). Second, the dots ride almost exactly on the gold curve (n/2)(log n − 1 − log 2π + γ), where γ ≈ 0.5772 is Euler’s constant. Keiper noticed this growth in 1992, and later work showed it is exactly what RH demands: if RH is true, λn follows the curve with an error far smaller than n, while if RH is false the numbers eventually swing wildly, with swings growing exponentially in n. The red line in the lower panel is that small error for n ≤ 500: a slow wobble between about −1.6 and 4.7, set mostly by the lowest zeros. Under RH that wobble may keep growing, but never faster than about √n · log n, so it stays tiny next to λn itself.

The positive values do not, by themselves, count as evidence for RH. The first λn are dominated by the lowest zeros, which were long ago checked to lie on the line, so their positivity was guaranteed in advance. A zero off the line at height T would barely register until n reached roughly T2 (an observation of Joseph Oesterlé reported by Krzysztof Maślanka), and zeros are already known to be on the line far higher than any computed n could reach. Li’s criterion is a new way to say the Hypothesis, not a step towards proving it.

  • Settled Li (1997): RH holds if and only if λn ≥ 0 for every n ≥ 1. Bombieri and Lagarias (1999) showed the same positivity test works for any suitable collection of points, not just zeta zeros, and linked it to Weil’s positivity criterion.
  • Settled If RH holds, λn = (n/2)(log n − 1 − log 2π + γ) + O(√n log n) (Lagarias 2007); Voros (2006) showed that this kind of tame growth holds if and only if RH does, and Arias de Reyna (2011) gave a precise form of Keiper’s prediction equivalent to RH.
  • Settled Observed here λ1 to λ500 are all positive. Much longer runs agree: Keiper (1992) went to n = 7000, Maślanka (2004) published about 3300 values, and Fredrik Johansson (2015) computed every λn up to n = 100 000. None of this settles anything about larger n, and the Hypothesis remains open.
  • Settled Reformulation Li’s criterion restates RH; it does not make it easier. A finite list of positive λn, however long, can never prove RH.

Polynomials that copy ξ, and must all have real roots

A polynomial is the simplest kind of function: a finite sum like 3 − 2x + x2. Its roots (where it equals zero) are easy to find, and a natural question is whether they are all ordinary real numbers or whether some are complex. The completed function ξ (see the ξ figure) is not a polynomial, but it has a Taylor series, an endless polynomial, and you can cut that series short. George Pólya showed in 1927 that if you cut it in one particular way, due to Johan Jensen, then the Riemann Hypothesis is true if and only if every one of these cut-off polynomials has only real roots. Mathematicians call a polynomial with only real roots hyperbolic.

Here is why that is natural. Write Ξ(t) = ξ(½ + it), so that the zeros on the critical line become real values of t. RH says every zero of Ξ is real. Jensen’s cut-off of degree d is a polynomial Jd,0 whose coefficients are the first d + 1 Taylor coefficients of Ξ, each multiplied by a binomial number. As d grows, Jd,0, suitably rescaled, turns into Ξ itself, and its roots turn into the zeros of Ξ. If Ξ had a zero off the line, some cut-off would eventually show a non-real pair of roots. One can also start the cut-off further along the series, at a shift n, giving Jd,n; RH is equivalent to all of them, every d and every n, being hyperbolic.

Figure 10 Jensen polynomials: real roots only

Two stacked charts. Top: for each degree d from 1 to 100, a column of dots marking the real roots of the Jensen polynomial of degree d, shown as heights. As d grows the lowest dots rise slowly toward dashed gold lines at the first four zeta zero heights, 14.13, 21.02, 25.01 and 30.42; at d = 100 the lowest is at 13.84. Bottom: a grid of squares for degrees 2 to 8 and shifts 0 to about 98, every square filled to show that all roots came out real. The chart fades in on load.
Computed 3 November 2026 PT with mpmath 1.4.1. The Taylor coefficients of Ξ come from Riemann’s integral Ξ(t) = 8∫0∞ Φ(u) cos(2tu) du with Φ(u) = Σn (2π2n4e9u − 3πn2e5u) e−πn²e4u, using the normalization of Griffin, Ono, Rolen and Zagier (2019). Checks: the series gives Ξ(0) = ξ(½) = 0.49712077818831410991… and Ξ(10) = 0.037967850310…, both matching ξ computed directly from ζ to all 20 digits shown. The whole run was done at 160 and at 220 significant digits, and the roots agree. Top: roots of Jd,0 for d = 1 to 100, each root X shown as the height t = √(−dX). Bottom: for d = 2 to 8 and shifts n up to about 98, every root of Jd,n came out real. Fade-in is off if your system asks for reduced motion.

Two things show up. First, every root found is real, in every polynomial computed. Second, the roots of Jd,0 do march toward the zeta zeros, but slowly: at degree 100 the lowest root is at height 13.838, while the first zero is at 14.135. Doubling the degree roughly halves the gap (it was 0.704 at d = 40 and 0.367 at d = 80), so these polynomials are a faithful but lazy copy of ξ.

  • Settled Pólya (1927), building on Jensen (1913): RH holds if and only if Jd,n is hyperbolic for all d ≥ 1 and n ≥ 0. Because hyperbolicity survives taking derivatives, the shift-zero polynomials Jd,0 alone would be enough.
  • Settled Degrees 2 and 3 are hyperbolic for every shift (Csordas, Norfolk and Varga 1986; Dimitrov and Lucas 2011). Griffin, Ono, Rolen and Zagier (2019) proved every degree d is hyperbolic for all sufficiently large n, and every n when d ≤ 8; for large n the polynomials look like Hermite polynomials, matching a random-matrix prediction. Chasse (2013) confirmed Jd,0 is hyperbolic for every d up to 2·1017.
  • Settled Observed here All roots computed above are real. That repeats known results; it adds no new evidence.
  • Settled Reformulation “For all large n” leaves finitely many exceptions open in each degree, and infinitely many degrees remain. The Hypothesis itself is still open.

Heat flow on ξ, and a constant that pins RH to a single number

Here is one more way to restate the Riemann Hypothesis, and it turns RH into a question about one number. Start from Riemann’s integral for ξ, the same one used in the Jensen section above, and add a dial t to it:

Ht(z) = ∫0∞ et u² Φ(u) cos(zu) du,

The dial t multiplies the integrand by et u². At t = 0 it is 1 and nothing changes.

with Riemann’s fast-shrinking function Φ(u) = Σn (2π2n4e9u − 3πn2e5u) e−πn²e4u. At t = 0 the dial does nothing, and H0(z) = ⅛ Ξ(z/2), where Ξ(T) = ξ(½ + iT). So the zeros of H0 are the zeta zeros, just stretched: a zero of ζ at ½ + iγ shows up as a zero of H0 at z = 2γ. If RH is true, all of them are real numbers z, sitting on the ordinary number line rather than off it in the complex plane.

Turning the dial is like running heat forward or backward in time. Moving t up smooths Ht out: zeros that are real stay real, they push each other apart, and their spacing becomes more regular. Moving t down does the opposite: neighbouring zeros can be pulled together until two of them meet and then leave the number line as a pair of complex zeros. Nicolaas de Bruijn proved in 1950 that once t ≥ ½ every zero of Ht is real. Charles Newman showed in 1976 that there is one exact cut-off number, now called the de Bruijn–Newman constant Λ: all zeros of Ht are real exactly when t ≥ Λ. Since t = 0 is ξ itself, RH is the same statement as Λ ≤ 0.

Figure 11 de Bruijn–Newman: zeros under heat flow

Two stacked charts. Top: twenty dark green curves, one for each of the lowest real zeros of the function H_t, drawn as t runs from −0.5 to 0.5. Every curve drifts slowly downward as t grows; the lowest goes from 28.56 at t = −0.5 to 28.27 at t = 0 and 27.98 at t = 0.5. On the line t = 0 gold dots mark twice the heights of the first twenty zeta zeros, and the curves pass exactly through them. A gold band from t = 0 to t = 0.2 is labelled as the range where the de Bruijn–Newman constant is known to lie; right of it every zero is proved real. Bottom: smallest gap divided by largest gap among the twenty zeros, rising from 0.168 at t = −0.5 to 0.209 at t = 0 and 0.239 at t = 0.5, with a second dashed curve for gaps measured against the average spacing. The chart fades in on load.
Computed 4 November 2026 PT with mpmath 1.4.1. Ht(x) was evaluated by a 2304-point Gauss–Legendre rule on 0 ≤ u ≤ 1.3 at 60 significant digits (Φ is below 10−200 past u = 1.3). Checks: H0(x) matches ξ(½ + ix/2)/8 computed directly from ζ at x = 0, 20, 60 and 150 to at least 40 digits; at t = 0 the twenty zeros found agree with 2γ1, …, 2γ20 (2 × 14.134725… = 28.269450, 2 × 21.022040… = 42.044079, …) to every one of the 15 digits stored; a recheck at 90 digits with a finer rule puts the twenty t = 0 zeros within 10−68 of 2γn, and reproduces the zeros at t = −0.5, 0 and 0.5 to all stored digits. Zeros were found by sign changes on a grid of step 0.25 in x, refined by root-finding, at 51 values of t from −0.5 to 0.5 in steps of 0.02. Fade-in is off if your system asks for reduced motion.

What the picture shows, for these twenty zeros only. Every one stays real across the whole window from t = −0.5 to 0.5; none collide. They all slide gently toward smaller x as t grows (the lowest moves from 28.560 to 27.980), and their spacing gets more even: the smallest gap divided by the largest is 0.168 at t = −0.5, 0.209 at t = 0 and 0.239 at t = 0.5. Part of that unevenness is only because zeros crowd together higher up, so the dashed curve divides each gap by the average spacing at its height (the zero density (1/4π) log(x/4π)); that ratio goes from 0.257 to 0.340 to 0.411. Running the clock further backward, outside the drawn window (t was scanned down to −1.3 in steps of 0.05), the first collisions among these twenty arrive near t ≈ −1.28: zeros 19 and 20 meet at t ≈ −1.2755 (at x ≈ 154.05), and zeros 13 and 14 just after, at t ≈ −1.2853 (x ≈ 120.98). A meeting point was found by solving Ht(x) = 0 and its slope = 0 together. Just past it the two zeros leave the number line as a complex pair, for example 154.073 ± 0.200i at t = −1.296.

Do not read Λ off this chart. Λ is decided by the zeros at every height at once, and the ones that matter are unusually close pairs far up the line (the Lehmer pairs), which collide almost immediately when t goes below 0. Before Rodgers and Tao, such pairs had already shown Λ > −1.1·10−11 (Saouter, Gourdon and Demichel, 2011). The low zeros drawn here survive until about t ≈ −1.28, which says nothing about whether Λ is 0 or a little above it.

  • Settled de Bruijn (1950): all zeros of Ht are real for t ≥ ½. Newman (1976): there is a constant Λ ≤ ½ with all zeros real exactly when t ≥ Λ; RH is equivalent to Λ ≤ 0. Ki, Kim and Lee (2009) sharpened de Bruijn’s bound to Λ < ½.
  • Settled Λ ≥ 0 (Rodgers and Tao, 2020), proving a conjecture Newman made in 1976 as “a quantitative version of the dictum that the Riemann hypothesis, if true, is only barely so.” Together with RH ⇔ Λ ≤ 0, this means RH is now the same as the single equation Λ = 0.
  • Settled Λ ≤ 0.22 (the Polymath 15 project, 2019), improved to Λ ≤ 0.2 (Platt and Trudgian, 2021) using their check that RH holds up to height 3·1012. So today 0 ≤ Λ ≤ 0.2, the gold band in the figure.
  • Settled Observed here The twenty lowest zeros of Ht stay real for −0.5 ≤ t ≤ 0.5 and grow more evenly spaced as t rises. That illustrates the heat-flow picture; it adds no evidence about Λ.
  • Open Whether Λ = 0, which is exactly the Riemann Hypothesis. It is still open.

Stage 4 The same question for other L-functions

The same question for other L-functions

ζ is the first of a large family. Replace the 1 in front of every term of 1 + 1/2s + 1/3s + ⋯ with a repeating pattern χ(n) that respects multiplication, and you get a Dirichlet L-function. Dirichlet built them in 1837 to prove that every progression like 1, 5, 9, 13, … (numbers that leave remainder 1 after dividing by 4) contains infinitely many primes. These functions have zeros too, and the natural guess, the Generalized Riemann Hypothesis, is that their nontrivial zeros also all sit on the line with real part 1/2.

Figure 12 Zeros of other L-functions Open

Three stacked charts over heights t from 0 to 40, each showing the size of a function on the line with real part one half. Top, teal: the Riemann zeta function, touching zero 6 times, first at 14.1347. Middle, gold: the L-function of the character mod 4, touching zero 14 times, first at 6.0209. Bottom, red: the L-function of the character mod 3, touching zero 13 times, first at 8.0397. Each zero is a dot on the axis; the curves draw left to right and the dots appear in order of height on load.
Computed with mpmath 1.4.1 at 25 digits, 23 October 2026 PT, using dirichlet(s, χ) for the two characters and zeta(s) for ζ. Each function’s size on the line s = 1/2 + it was sampled every 0.02 from t = 0 to 40, each dip was refined with Newton’s method (findroot) in the complex plane, and every zero landed back on real part 1/2 to working precision. The counts are 6, 14 and 13; the first zeros, 1/2 + 6.020949i for mod 4 and 1/2 + 8.039737i for mod 3, agree with the values listed in the LMFDB. The L-functions have more zeros than ζ up to the same height because their zero count grows like (T/2π) log(qT/2πe) for modulus q, about 14.2 and 12.4 here. Motion is off if your system asks for reduced motion. Open The Generalized Riemann Hypothesis is unproved; this picture checks three functions up to height 40 only.
  • Settled Large computer checks exist here too. Rumely (1993) and later Platt (2016) verified that the zeros of all Dirichlet L-functions with modulus up to large bounds lie on the line up to substantial heights; no counterexample has ever been found.
  • Settled Many theorems are proved conditionally, assuming the Generalized Hypothesis: for example, sharp bounds on the least prime in a progression and fast deterministic primality tests. If the hypothesis failed, these arguments would need new proofs.
  • Open The zeros of the mod 4 function also explain Chebyshev’s observation that primes of the form 4k + 3 usually outnumber primes of the form 4k + 1. Rubinstein and Sarnak (1994) measured this bias exactly, but only by assuming the Generalized Hypothesis and that the zero heights have no hidden linear relations. See the race below.

Chebyshev’s bias: a prime race steered by the mod 4 zeros

Every odd prime leaves remainder 1 or 3 when you divide it by 4: 5, 13, 17, 29 are of the form 4n + 1; 3, 7, 11, 19 are of the form 4n + 3. In the long run the two teams are the same size (Dirichlet’s theorem, made precise by the prime number theorem for progressions). But in 1853 Chebyshev wrote to Fuss that the 4n + 3 team “contains a lot more” at every size he checked. Picture a race: walk up the whole numbers, step up one whenever you pass a prime 4n + 3, step down one for a prime 4n + 1. The height you are at is the lead of the 4n + 3 team.

Here that walk was carried out for every whole number up to ten billion. Two things show at once. The 4n + 3 team really is almost always ahead. And it is not always ahead: the 4n + 1 team grabs the lead for the first time at x = 26,861, exactly where John Leech found it in 1957.

Figure 13 Chebyshev’s prime race mod 4

Two charts of the race between primes of the form 4n + 3 and 4n + 1. Top: on a log scale from 1 000 to 10 billion, the lead of the 4n + 3 primes, rescaled by ln x over the square root of x, wiggles in a dark band that mostly stays between 0 and 2.5 and hovers around a dashed line at 1. It dips below zero, into a pink zone where the 4n + 1 primes lead, only in 5 short stretches, starting at 26,861, 616,841, 12,306,137, 951,784,481, 6,309,280,709. Bottom: a zoom on the raw lead between 25 500 and 28 500, a staircase that falls to minus 1 at x = 26 861 and climbs back. Footer: up to 10 billion the 4n + 3 primes are strictly ahead for 99.28% of whole numbers x. The charts fade in on load.
Computed 28 October 2026 PT with a segmented sieve of Eratosthenes (numpy 2.5.4) up to N = 1010, about 123 seconds. The lead is checked at every whole number x ≤ N, not only at round values. Totals: 227,529,235 primes 4n + 3 and 227,523,275 primes 4n + 1, so the lead at 1010 is 5,960; at 104, 105, 106, 107 it is 10, 25, 147, 218. The 4n + 1 team leads in 5 separate stretches, starting at 26,861, 616,841, 12,306,137, 951,784,481, 6,309,280,709; its biggest lead in each is 1, 8, 24, 48, 1374. The first four starts agree with Leech (1957) and with Bays and Hudson as reported by Granville and Martin (2006); in the fifth stretch the race first ties at 6,309,280,697, the number they quote, and the 4n + 1 team goes strictly ahead at the next prime 4n + 1, 6,309,280,709 (checked with an independent second sieve). The biggest lead near 623,681 is 8, as Leech reported. Fade-in is off if your system asks for reduced motion.

Why should one team be favoured, and why not always? The answer runs through zeros. The steady push comes from squares: every odd prime squared, 9, 25, 49, …, has the form 4n + 1. Counting prime powers fairly makes the race even on average, so counting only the primes leaves the 4n + 3 team ahead by about half the number of prime squares, roughly √x / ln x. That is why the top chart divides by √x / ln x: on that scale the steady push is the dashed level 1.

The wiggles around that level come from the zeros of L(s, χ), the mod 4 function in the figure above. In the explicit formula each zero 1/2 + iγ adds a wave in ln x whose speed is γ and whose size shrinks like 1/γ. The lowest zero, at γ ≈ 6.0209, gives the slowest and largest wave. The 4n + 1 team can only lead when enough of these waves happen to point downward at once, strongly enough to beat the steady +1. Leech noticed exactly this: at x = 620,000, 16 of the first 20 zero pairs pull in the negative direction.

  • Settled The lead changes sign infinitely often: the 4n + 1 team takes the lead for arbitrarily large x (Littlewood, 1914). This is proved without assuming any Riemann Hypothesis.
  • Settled Observed here Up to 1010, the 4n + 3 team is strictly ahead at 99.28% of whole numbers x, tied at 0.0015%, and behind at 0.72%. Most of that 0.72% comes from the single long fifth stretch.
  • Open Conditional Counting each x with weight 1/x instead of 1 (the “logarithmic” share), Rubinstein and Sarnak (1994) showed that in the limit the 4n + 3 team leads 99.59% of the time (density 0.9959…). This number rests on two unproved assumptions: the Generalized Riemann Hypothesis for L(s, χ), and that the zero heights γ have no hidden whole-number relations among them (linear independence). With the same 1/x weighting, the 4n + 1 team’s share in the data here is only 0.081%, against the limit’s 0.41%; a finite range need not match a limit, and the long stretches where 4n + 1 leads are rare and far apart.
  • Open Conditional Assuming the Generalized Riemann Hypothesis, Kaczorowski (1995) showed that with the ordinary weight 1 the share of x where 4n + 3 leads has no limit at all; it keeps drifting up and down forever (Sarnak found the same independently; see Granville and Martin, 2006). That is why the 1/x weighting is used, and why the 99.28% above is a fact about this range, not a law.

Where the same question is already a theorem: curves over finite fields

For the ordinary zeta function ζ(s), the Riemann Hypothesis is still open. There is a parallel question for curves over finite fields where the answer is known. Take an elliptic curve E given by y² = x³ + A x + B, and reduce it modulo a prime p that does not divide the discriminant. Count its points over the field with p elements, including the point at infinity: write N = #E(Fp) and ap = p + 1 − N. Hasse proved in 1933 that

|ap| ≤ 2√p,

Equivalently, the two roots of 1 − ap T + p T² = 0 both have absolute value p−1/2. That is the Riemann Hypothesis for the zeta function of E over Fp.

So here the zeros really do lie on the critical circle |T| = p−1/2, and the proof is finished. Weil extended the same statement to curves of every genus in 1948. None of this settles RH for ζ; it is a theorem about a different zeta function, built from counting points on a curve over a finite field rather than from the primes of Z.

Figure 14 Hasse: RH for curves over finite fields

Two panels. Top: for five elliptic curves over the finite field with p elements, the normalized count a_p/(2√p) is plotted against the prime p from 5 to 10,000 on a log scale. All 6,132 points lie between −1 and 1, the strip Hasse proved in 1933. The closest approach to the boundary in this run is 0.9998 at p = 8,839 on x³+1. Bottom left: a histogram of the same normalized traces, filling the strip. Bottom right: for the curve y² = x³ − x + 1, each prime places a pair of points on the unit circle at angles ±θ with a_p = 2√p cos θ. The chart fades in on load.
Computed 5 November 2026 PT in Python 3.13.5 (numpy 2.5.4). For each of the five curves y² = x³ − x, y² = x³ + 1, y² = x³ − x + 1, y² = x³ + x + 1, y² = x³ − 2x + 1, and for every prime p from 5 to 10,000 of good reduction (1227 primes; 3 primes of bad reduction skipped), #E(Fp) was counted by testing the Legendre symbol of x³ + A x + B at every x in Fp and adding the point at infinity — 6,132 pairs (curve, p) in 11.7 seconds. Check: for y² = x³ − x over F5, N = 8 and a5 = -2 (the seven affine points (0,0), (1,0), (2,±1), (3,±2), (4,0) plus infinity). Every normalized trace satisfies |ap|/(2√p) ≤ 1; the largest value in the run is 0.9998 at p = 8,839 on x³+1. Fade-in is off if your system asks for reduced motion.

What the picture is saying. The top panel is the Hasse strip: height ap/(2√p) against p. Every computed point sits inside [−1, 1], which is exactly the theorem. The bottom-left histogram shows those normalized traces filling the strip; the bottom-right panel rewrites ap = 2√p cos θ as a pair of points e±iθ on the unit circle for the single curve y² = x³ − x + 1 — the eigenvalues of Frobenius, after dividing out √p. This is RH as a theorem for these zeta functions. It is evidence about the shape of the analogy, not a proof of RH for ζ.

  • Settled Hasse (1933, full proof 1936): for every elliptic curve over a finite field Fq, |#E(Fq) − (q + 1)| ≤ 2√q. Equivalently, the zeros of its zeta function lie on |T| = q−1/2.
  • Settled Weil (1948): the same Riemann Hypothesis holds for the zeta function of every nonsingular projective curve over a finite field (any genus). The later Weil conjectures for higher-dimensional varieties were proved by Deligne (1974).
  • Settled Observed here 6,132 normalized Frobenius traces for five curves and every good prime up to 10,000 all satisfy |ap|/(2√p) ≤ 1; the closest to the bound is 0.9998. That is a check of the computation against a theorem already proved, not new evidence.
  • Open The Riemann Hypothesis for the ordinary zeta function ζ(s) of the integers. It is still open; nothing on this page claims otherwise.

Lifting a dimension: when it forces zeros onto the line, and when it can’t

A flat problem sometimes becomes easy when you see it as the shadow of a solid one. This section follows that one move through four places. In Monge’s circle theorem it is a neat trick. For curves over finite fields it is Weil’s proof that every zero lies on the critical line. For the ordinary ζ(s) nobody knows the space you would lift into. And over the octonions the extra dimension is simply not available.

1. The model move: Monge’s three circles

[known] Take three circles in a plane, of three different sizes, none inside another. Each pair has two outer tangent lines, which touch both circles without passing between them. Those two lines meet at a point. It is the point from which one circle looks like a scaled copy of the other. Monge’s theorem: the three meeting points lie on one straight line.

[known] The 3D proof. Make each circle the equator of a ball, so the three centres sit in the original plane. Lay one flat sheet on top of the three balls, touching all of them, and another underneath. Each pair of balls fits snugly inside a cone, and the tip of that cone is exactly the meeting point of the two outer tangents. The cone touches both sheets, so its tip lies in both sheets. Two sheets cross in a line, so all three tips lie on that line. A plane statement was proved by stepping into space and coming back down. (This quick version needs the two sheets to exist, which fails for example when the smallest circle sits between the other two. Proofs by similar cones or by scaling maps cover the remaining cases.)

[known] Equal sizes. If two circles have the same radius, their outer tangents are parallel and never meet in the ordinary plane. Projective geometry adds one “point at infinity” for each direction. The theorem then says the other two meeting points lie on a line parallel to those two tangents. If all three radii are equal, all three points are at infinity, on the “line at infinity”. A one-line algebra proof covers every case with different radii: each meeting point is pij = (ricj − rjci)/(ri − rj). The combination r1(r2 − r3)p23 + r2(r3 − r1)p13 + r3(r1 − r2)p12 is 0, and its three coefficients also add to 0. That is exactly the condition for three points to be collinear.

Left: three circles with the outer tangent lines of each pair meeting at three points on one dashed line; below, the same circles as equators of balls between a top and a bottom sheet that cross along that line. Right: a schematic square for the surface C × C of the curve y² = x³ + x + 1 over F₅, with the diagonal and the graph of Frobenius meeting at nine marked points, and the inequality |Δ·Γ − q − 1| ≤ 2g√q.

[known] Left: Monge’s configuration at the last frame of the motion below. The 3D view is computed: the two tangent sheets were solved for exactly, and the three cone tips lie on their crossing line to within 1.4·10−14. Right: [project] a schematic, not a picture of the real surface. That surface lives over the algebraic closure of F5 and cannot be drawn. Only the count is literal: the nine meeting points are the nine points of the curve over F5, counted below.

Move the circles; the three points stay on a line

The circles drift and change size. Each frame recomputes, from scratch, where each pair’s outer tangents meet. Nothing in the drawing pushes those points onto a line; they land on one because of the theorem.

Monge’s three circles, movingThree circles of different sizes drift and breathe. For each pair, the two outer tangent lines are drawn in one colour and meet at a red point. In every frame the three red points are recomputed from the circles and lie on the dashed gold line.gold: the line through the three meeting points

Final frame: the triangle spanned by the three meeting points has area ÷ (longest side)² = 3.3e-17, rounding error only.

[known] Monge’s theorem. [project, checked computation] The motion is a closed loop of radii between 18 and 90 px, kept distinct so every meeting point stays on screen; over 2,001 test frames the largest defect is 2·10−16. Each meeting point is p = (ricj − rjci)/(ri − rj) for centres c and radii r; the motion never lets two radii be equal, the case where that point runs off to infinity (see the text).

2. Where the lift proves the Riemann Hypothesis: Weil’s curves

[known] Fix a finite field Fq and a smooth curve C over it, of genus g. Count its points with coordinates in Fq, then in the bigger fields Fq², Fq³, and so on: N1, N2, N3, …. Pack the counts into one function Z(C, T), defined by d log Z/dT = N1 + N2T + N3T² + ⋯. By the work of Artin, F. K. Schmidt and Hasse, this is a fraction P(T)/((1 − T)(1 − qT)) with P a polynomial of degree 2g. Put T = q−s and you have a zeta function of s, with a functional equation linking s and 1 − s, just as for ζ.

[known] Write P(T) = (1 − a1T)⋯(1 − a2gT). Its zeros lie on Re s = ½ exactly when every |ai| = √q. That holds exactly when the counts never stray far from qn + 1:

|Nn − qn − 1| ≤ 2g qn/2 for every n.

[known] The lift. Weil went from the curve C (one dimension) to the surface C × C (two dimensions). On that surface sit two special curves. One is the diagonal Δ, the points (x, x). The other is the graph Γ of the Frobenius map, the points (x, xq). They meet exactly where xq = x, and those are the points with coordinates in Fq. So the count N1 becomes an intersection number Δ·Γ, the number of crossings of two curves on a surface. On a surface, such numbers obey an inequality, the Hodge index theorem. On C × C it takes the form of the Castelnuovo–Severi inequality D² ≤ 2d1d2. Apply it to combinations of Δ and Γ and out comes |N1 − q − 1| ≤ 2g√q. Do the same over Fqn and the Riemann Hypothesis for C follows. The extra dimension supplies a positivity (a square that cannot be negative), and the positivity pins the zeros to the line.

[known] Credits. Hasse proved the genus-1 case in the 1930s (see the section above). Weil announced the general case in 1940 and the C × C proof in 1941, and gave full details in his 1948 book. The modern short form, which derives the inequality from the Hodge index theorem, is due to Mattuck–Tate and Grothendieck (1958); Milne’s survey sets it out step by step. A completely different, elementary route that needs no surface was found by Stepanov (1969) and simplified by Bombieri (1973). So the lift is one proof, not the only one. For varieties of every dimension, Deligne proved the corresponding statement in 1974.

A worked example: y² = x³ + x + 1 over F5

[project, checked computation] We counted the points of this elliptic curve (genus 1) over the fields with 5, 25, …, 78,125 elements. Each count was done twice by brute force, once with Euler’s test for squares and once with a full table of squares; both agree. Over F5 itself there are 9 points: (0, ±1), (2, ±1), (3, ±1), (4, ±2) and the point at infinity. So a = 5 + 1 − 9 = −3 and

Z(E, T) = (1 + 3T + 5T²) / ((1 − T)(1 − 5T)).

[project, checked computation] The roots of 1 + 3T + 5T² are T = (−3 ± i√11)/10. They are complex conjugates and their product is 1/5, so |T|² = 1/5 exactly. With T = 5−s that says Re s = ½ for every zero, with nothing left to chance. The zeros are s = ½ + i(±2.3061 + 2πk)/log 5, repeating every 2π/log 5 ≈ 3.904. Every brute-force count matches 5n + 1 − (αn + ᾱn) with α = (−3 + i√11)/2, and every count sits inside Weil’s bound. At n = 4 it comes within 1 of the edge: 49 against 50.

nfield size 5npoints NnNn − 5n − 1bound 2·5n/2
15934.47
22527110.00
3125108−1822.36
46256754950.00
53,1253,069−57111.80
615,62515,552−74250.00
778,12578,633507559.02
Left: bars for n = 1 to 7 of the normalized error (N − 5ⁿ − 1)/(2·5^(n/2)), all between −1 and 1. Right: the zeros of the curve’s zeta function in the s-plane, all on the line Re s = ½, with poles on Re s = 0 and 1.

[project, checked computation] Counts from lifting-primer-20261111/compute.py (pure Python, 11 November 2026). Field moduli were checked irreducible, and the zeta coefficients re-derived from the closed form, with SymPy 1.14. This checks a theorem; it is not evidence about ζ.

3. Why this does not carry over to ζ

[known, open] ζ(s) is the zeta function of the integers. Its “points” are the primes, much as a curve’s points are counted above. Copying Weil would need three things. First, a surface “Spec Z × Spec Z” that really is two-dimensional. Second, a Frobenius whose graph meets the diagonal in the primes. Third, an intersection theory on that surface with a Hodge-type positivity. None of these is known. The obvious product of the integers with themselves gives back the integers: there is no smaller “field of constants” underneath Z to multiply over, the role Fq plays for a curve. This missing base is what people mean by the hoped-for “field with one element”.

[known, open] The best-known current attempt is the programme of Connes and Consani. Their arithmetic site (2014) is a space built from the positive integers and “tropical” arithmetic, where addition is replaced by taking a minimum. Its points over a suitable base form a quotient of the adèle class space of Q. Its point-counting zeta function is the complete Riemann zeta function, and they realise Frobenius correspondences inside the square of the site. Their scaling site (2015) extends this: for each prime p it carries a curve Cp on which a Riemann–Roch formula holds, with real-valued dimensions. They describe the framework as one in which Weil’s proof might be transposed, and say this would require, in particular, an adequate Riemann–Roch theorem in “characteristic 1”. That step, and the positivity, remain open. Nothing here, and nothing in that programme so far, proves the Riemann Hypothesis. A different way to phrase the missing positivity, Weil’s explicit-formula criterion, is on From higher to lower.

4. Octonions: a plane you cannot lift

[known] Desargues’ theorem. Take two triangles. If the three lines joining matching corners meet at one point, then the three points where matching sides cross lie on one line. In three dimensions the proof is the same move as Monge’s: if the triangles lie in different planes, all three crossing points lie in both planes, and two planes meet in a line. A flat picture is handled by lifting one triangle out of the plane and projecting back.

[known] How Monge and Desargues are related. Monge’s theorem can be derived from Desargues’ theorem. The outer tangents form a triangle A′B′C′ around the circles. The lines from each circle’s centre to the matching corner bisect the corner angles, so they meet at that triangle’s incentre. The triangle of centres and A′B′C′ are therefore in perspective from a point. Desargues then puts the crossings of matching sides, which are the three Monge points, on one line. We do not claim the converse. Monge is a statement about circles in the Euclidean plane, while Desargues is a pure statement about points and lines that makes sense in far more general planes. What they share is the 3D proof.

[known] When can a plane be lifted? Hilbert (Grundlagen der Geometrie, 1899) showed that a plane geometry can sit inside a space of more than two dimensions exactly when Desargues’ theorem holds in it. A projective plane satisfies Desargues exactly when it can be given coordinates from a division ring, a number system with associative multiplication and division, such as the reals, complexes or quaternions. Every projective space of dimension 3 or more is automatically Desarguesian, and is the projective space over some division ring.

[known] The octonions. In 1933 Ruth Moufang built a projective plane from the octonions, now called OP² or the Cayley plane. Octonion multiplication is not associative; it satisfies only the weaker “alternative” laws. In OP² only the restricted “little Desargues” theorem holds, which makes it a Moufang plane, and the full Desargues theorem fails. Two consequences follow. OP² cannot be placed inside any projective 3-space. And there is no octonionic projective 3-space at all, because any projective space of dimension 3 or more would come from an associative division ring. The octonions are exactly where the “step up one dimension” move of Monge and Desargues is not available.

[project] Our own climb into octonions. Our hypercomplex threads went “up” in a different sense: into bigger number systems, quaternion and octonion orders, rather than into a product surface. That climb showed how zeros are built. For the odd-conductor octonion orders we proved the local factors of the shapes P6 and Q for every odd prime, and proved that neither ever has a zero on Re s = ½ (general-p proofs; the p = 3 picture). But it never forced any zero onto ½. None of the odd-conductor local factors we computed has a zero on the line at all, and nothing there touches the zeros of ζ itself. Weil’s lift works because the new dimension brings a positivity that holds the zeros in place. Our climb brought no such inequality.

[speculative] It is tempting to connect the two failures: no octonionic 3-space, so no Desargues, and no forcing in our octonion results. We have no argument linking non-associativity to the absence of a Weil-type positivity. The two facts share a vocabulary (“lifting”, “dimension”), not a proof, and we record the thought only as a question we cannot yet make precise.

Sources for this section

Why the prime product matters: a look-alike that breaks the rule

If the Riemann Hypothesis is true, what makes it true? A tempting answer is the mirror symmetry: ζ, and every Dirichlet L-function, satisfies a functional equation that pairs the value at s with the value at 1 − s. That symmetry is why zeros off the line would have to come in mirror pairs. But symmetry alone is not enough, and there is a clean example that shows it.

In 1936 Davenport and Heilbronn took the two complex L-functions mod 5, L(s, χ) and its mirror twin L(s, χ̄), and mixed them with one carefully chosen complex weight κ = (√(10 − 2√5) − 2)/(√5 − 1) ≈ 0.284079. The mix f(s) = ½(1 − iκ) L(s, χ) + ½(1 + iκ) L(s, χ̄) obeys exactly the same kind of functional equation as an L-function. What it loses is the Euler product: it can no longer be written as a product over primes. And its zeros drift off the line.

Figure 15 A look-alike with zeros off the line

Two sideways pictures of the critical strip, heights 0 to 200. Top: a Dirichlet L-function mod 5 has 130 zeros, all on the dashed middle line. Bottom: the Davenport–Heilbronn function, built from the same pieces with the same symmetry but no Euler product, has 130 zeros; 122 on the line and 8 off it in red mirror pairs, first at 0.809 + 85.70i and 0.191 + 85.70i. On load the dots appear in order of height and the red ones slide off the line.
Computed with mpmath 1.4.1 at 20 digits, 26 October 2026 PT, using dirichlet(s, χ) for the character mod 5 with χ(2) = i. Both functions were sampled every 0.02 along the line and on a grid across the right half of the strip up to t = 200; every dip was refined with Newton’s method (findroot), and the mirror image of each off-line zero was checked to be a zero too. As an independent count, the argument principle around the box 0.55 ≤ σ ≤ 1.6, 1 ≤ t ≤ 200 gives 4 zeros for f and 0 for L(s, χ). Both functions have 130 zeros up to height 200 (the L-function formula (T/2π) log(5T/2πe) gives about 129.5). The four off-line zeros right of the line, 0.808517 + 85.699348i, 0.650830 + 114.163343i, 0.574356 + 166.479306i, 0.724258 + 176.702461i, agree with Spira (1994) to all six published digits, as does the first on-line zero of f at 1/2 + 5.094160i. Motion is off if your system asks for reduced motion.
  • Settled The Davenport–Heilbronn function has infinitely many zeros with real part greater than 1, where no L-function can vanish (Davenport and Heilbronn, 1936), and also zeros inside the strip but off the line (shown by Voronin, 1980). None with real part above 1 occur below height 200; they appear only much higher up.
  • Settled Even so, it has infinitely many zeros exactly on the line (Voronin, 1980; Karatsuba later sharpened the count), and 122 of the 130 zeros found here are on it. A function can look “mostly right” and still fail.
  • Open Conditional Bombieri and Hejhal (1995) showed that, assuming the Generalized Riemann Hypothesis and a spacing hypothesis for the underlying L-functions, almost all (100% in proportion) of the zeros of such mixtures lie on the line. The exceptions are real but rare.
  • Open The lesson most experts draw is that any proof of the Riemann Hypothesis must use the Euler product, the link to the primes, and not only the functional equation. This is a guide to strategy, not a theorem; it does not say anything about whether the Hypothesis is true.

How are these shares proved? Hardy’s sign changes, Selberg’s tamed averages and Levinson’s counting trick, with the ladder from Hardy to 41.7% and the 2026 preprints: How much of the critical line can we prove?

Stage 5 A schematic of the plane

A schematic of the plane

Hand-drawn, and not evidence. The computed ordinates are the table in the critical strip and the line. A short source note lives at sources/zeta-primes-sources.md.

Figure 16

Schematic of the critical line and trivial zeros A complex-plane sketch, not to scale. A shaded critical strip, a vertical critical line with eight illustrative points symmetric across the real axis, and three marks on the negative real axis for trivial zeros. Heights are hand-placed and are not computed imaginary parts. Re(s) = 1/2 Re Im trivial zeros −2 −4 −6 critical strip
Schematic, and not to scale. Not the computed ordinates in the figure above. The shaded band is the critical strip. Axes are drawn over it. Points on the critical line are placed by hand, in mirror pairs, to suggest s ↦ 1 − s. They are not computed zeros. Marks at −2, −4, and −6 stand for trivial zeros read here as controlled defects. The picture is not evidence.

Reading path · step 9 of 13