Known theorems · the share of zeros on the line
How much of the critical line can we prove?
Hardy showed in 1914 that infinitely many zeros of ζ lie on the critical line. How many, as a share of all of them, can actually be proved? Today’s best classical answer is a little over 41.7%. This page explains how a number like that is proved: Hardy’s real function Z(t), Selberg’s tamed averages, and Levinson’s trick of counting the zeros that are not on the line. It is not progress on the Riemann Hypothesis.
[known] How to read the labels. Every claim carries one tag. [known] is published mathematics, with the source named (full list in the sources). [project] is something we computed or checked ourselves in November 2026, at small heights; a [project] number is evidence, never a theorem. [speculative] is our own reading, not proved. Click any tag to see every claim that carries it.
Nothing on this page is progress on the Riemann Hypothesis. RH says every non-trivial zero is on the line. The theorems here prove that a fixed share of them is: a different, much weaker statement, and no method described here is expected to reach 100%.
Background elsewhere on the site: Hardy’s Z(t) and Gram’s law are worked through on the critical-line page, and what is proved, and what is only checked shows the proved-share chart. This page goes one level deeper: how the shares are proved.
0. The answer in one paragraph
The non-trivial zeros of ζ(s) come in mirror pairs across the vertical line Re s = ½, unless a zero sits exactly on that line [known] the functional equation ξ(s) = ξ(1 − s); Iwaniec (17.1)–(17.3). Up to height T there are about N(T) ≈ (T/2π) log(T/2πe) of them [known] Riemann–von Mangoldt; Iwaniec ch. 15. Nobody can prove that all of them sit on the line (that is RH). What has been proved is that a fixed fraction κ of them does, at every large height; the best fraction proved by the classical method is a little over 41.7% [known] PRZZ 2020. There are two classical ways to prove such a fraction:
- Selberg (1942) looks for zeros on the line directly, as sign changes of a real function, after "taming" the function with a correcting factor called a mollifier. It proves that a positive fraction of the zeros is on the line, but the fraction it gives is tiny [known] Titchmarsh §10.9, Thm 10.22.
- Levinson (1974) does the opposite: he counts the zeros of a related function G that are off the line on the right-hand side, shows there are few of them, and subtracts. With a mollifier again, this gives more than one third [known] Levinson, Adv. Math. 13 (1974); Iwaniec ch. 16–23.
Computers have checked that every zero up to height 3·10¹² is on the line [known] Platt–Trudgian 2021, as cited in Knausgård 2026 §1. That does not prove any fraction "in the long run": the theorems are about all heights at once, and a finite check says nothing about heights beyond it. That is why "about 41% proved" and "100% of the ones we have looked at" are both true at the same time.
1. The ladder of proved fractions
κ = liminf N₀(T)/N(T), where N₀(T) counts the zeros on the line up to height T; κ* is the same for zeros that are simple (not repeated). Values exactly as in the source notes.
| Year | Who | Proved share on the line | Method | Source |
|---|---|---|---|---|
| 1914 | Hardy | infinitely many | sign changes of Ξ, Mellin integral | C. R. 158, 1012–1014 |
| 1921 | Hardy–Littlewood | N₀(T) ≫ T (vanishing share) | sign changes in windows of fixed length | Math. Z. 10 |
| 1942 | Selberg | κ > 0 (tiny, not computed in the paper) | mollified sign changes, windows ≍ 1/log T | Skr. Norske Vid. Akad. 1942 no. 10; Titchmarsh Thm 10.22 |
| 1974 | Levinson | κ ≥ 0.3420 (> 1/3) | count G-zeros right of the line, mollified mean value | Adv. Math. 13 |
| 1975 | Levinson | κ ≥ 0.3474 | "semi-optimal" mollifier ("only 1.4%" better) | PNAS 72, 294–297 |
| 1983 | Conrey | κ ≥ 0.365 | general P, θ = ½ | J. Number Theory 16 |
| 1989 | Conrey | κ ≥ 0.4077 (Thm 1), > 0.4088 (note in proof); κ* ≥ 0.401 | mollifier length θ = 4/7 (Deshouillers–Iwaniec) | Crelle 399 |
| 2011 | Bui–Conrey–Young | κ ≥ 0.4105 | two-piece mollifier | Acta Arith. 150 |
| 2012 | Feng | κ ≥ 0.4128 (gap later found and repaired, Pratt–Robles 2018) | ζ-derivative-type pieces | J. Number Theory 132 |
| 2020 | Pratt–Robles–Zaharescu–Zeindler | κ > 0.417293 (> 5/12); κ* ≥ 0.407511 | richer V and mollifier | Res. Math. Sci. 7 |
PRZZ 2020 is the record for the mollifier method (κ > 0.417293) [known] PRZZ 2020, Res. Math. Sci. 7. It is not the only kind of result any more — see the box.
2026 preprints. [recent preprint, 2026] Unconditional ≥ 67.25% of zeros simple and on the line (Alpöge–Furman, argument found by an AI system and checked by the authors; independent human proof by Lamzouri; small further gain to 67.35% by Knausgård, computer-assisted). Not yet refereed; an expository re-proof by Arias de Reyna says the first write-up needed (easy) corrections.
These use a different method (Montgomery’s pair correlation of zeros plus linear algebra; no mollifier), and they are not progress on RH — the papers say so themselves. [recent preprint] arXiv 2608.13637 §1.4 “What the results are not”; see the source notes, 2026 section
The rest of this primer is about the classical (mollifier) line: Hardy → Hardy–Littlewood → Selberg → Levinson → Conrey → PRZZ.
2. Hardy's Z(t): a zero on the line is a sign change
Intuition first. On the critical line s = ½ + it the value ζ(½ + it) is a complex number: as t increases it moves around the complex plane. It turns out that almost all of that movement is a predictable rotation. If you undo the rotation, what is left can only move back and forth along one line through 0, so it is a real number that can be positive or negative. Real numbers that change sign must pass through 0 — and those passages are zeros of ζ on the critical line.
The mechanism. Define the Riemann–Siegel theta function θ(t) = arg Γ(¼ + it/2) − (t/2) log π, and
Then Z(t) is real for real t and |Z(t)| = |ζ(½ + it)| [known] Titchmarsh ch. X; Iwaniec (15.2)–(15.4); it follows from the functional equation. Z is continuous, so if Z(a) > 0 and Z(b) < 0 there is a zero of Z — hence of ζ on the line — between a and b (intermediate value theorem). Counting sign changes of Z certifies zeros on the line without any further complex analysis. Two limitations: a double zero (Z touches 0 without crossing) is invisible, and two zeros very close together can hide between two sample points [known] elementary.
Figure 1
compute/em_numpy.py, compute/zt_gram.py → compute/zt_gram.json). On a 0.01 grid the two Z's differ by at most 2.4·10⁻¹²; zeros agree to 1.1·10⁻¹², Gram points to 1.4·10⁻¹⁴; both see 13 sign changes = mpmath's count of zeros below 60.To watch Z(t) cross zero as t grows, see the sign-change animation on the critical-line page: Watch Z(t) change sign; the full story of the Gram points is in Gram’s law at three heights.
Hardy's 1914 idea. Hardy did not yet have a way to count sign changes. He showed there are infinitely many by contradiction: he wrote an integral of Ξ (essentially Z, with the Gamma factor attached) against an explicit weight using a Mellin (Cahen–Mellin) formula, and showed that if Ξ kept one sign beyond some height the integral would behave in a way it provably does not [known] Hardy, C. R. 158 (1914) 1012–1014, read in the Wikisource transcription; the source notes. This is a "there must be sign changes" argument, not a counting argument.
3. Hardy–Littlewood (1921): sign changes in windows, and why that gives only ≫ T
Intuition. Cut the line into windows [t, t + Δ]. If Z does not change sign in a window, then the integral of Z over the window equals (up to sign) the integral of |Z| — nothing cancels. So if we can show that on average the integral of Z is much smaller than the integral of |Z|, then most windows must contain a sign change.
The mechanism [known] Iwaniec ch. 15, Thm 15.1 and Lemma 15.2; Titchmarsh §§10.5–10.7. Put
No sign change in [t, t + Δ] ⇒ |I(t)| = J(t).
- J is large: the leading term of ζ(½ + it) is 1, so on average |Z| is at least about 1 and J is at least about Δ.
- I is small on average: Z oscillates, its integral mostly cancels, and a mean-square bound gives ∫T2T |I(t)|² dt ≪ ΔT (Iwaniec Lemma 15.2, for Δ ≥ 1 and T ≥ Δ⁶).
Comparing the two, the set of t where |I| = J has measure ≪ T/Δ. Take Δ to be a large constant: most of [T, 2T] is covered by windows that must contain a sign change, and packing disjoint ones gives N₀(T) ≫ T [known] Hardy–Littlewood, Math. Z. 10 (1921); Iwaniec Thm 15.1.
Why this is not a positive fraction. N(T) grows like T log T, but fixed-length windows can certify only about one zero per window, i.e. about T/Δ zeros: a vanishing fraction (about 1/log T) of all of them [known] Iwaniec ch. 15–16. To get a fixed fraction you need windows about as short as the average gap between zeros, Δ ≈ c/log T. In such short windows the comparison above breaks down, because |ζ| on the line is very uneven: usually moderate, occasionally huge, and the averages are dominated by the rare huge values [known] Iwaniec ch. 16; Titchmarsh §10.9.
What short windows actually contain [project] compute/hl_windows.json, Z from our own Riemann–Siegel program, 40/40 windows re-checked with mpmath. Windows [t, t + c/log t] tiling 1000 ≤ t ≤ 2000:
| c | windows | fraction with a sign change | expected zeros per window λ | if zeros were independent random points, 1 − e−λ |
|---|---|---|---|---|
| 1 | 7294 | 0.119 | 0.119 | 0.112 |
| 2 | 3646 | 0.238 | 0.238 | 0.212 |
| 4 | 1823 | 0.471 | 0.476 | 0.379 |
| 8 | 911 | 0.830 | 0.952 | 0.614 |
| 16 | 455 | 0.998 | 1.905 | 0.851 |
For short windows the fraction equals λ almost exactly: a window essentially never holds two zeros, because zeros repel each other (independent random points would give the last column, noticeably lower) [project] observation at small height; it matches the repulsion in Montgomery's 1973 pair-correlation work, which is proved only in part. Selberg's theorem says a fixed positive fraction of such windows contains a sign change at every height; the true fraction here is far larger than anything the proof delivers — that gap between "true at these heights" and "provable at all heights" is the whole subject.
4. The mollifier: what it is and why it is needed
The problem. Averages of |ζ(½ + it)|² are dominated by rare spikes: the mean square up to T grows like log T, while a typical value is much smaller [known] Iwaniec ch. 16; the source notes. Inequalities such as Cauchy–Schwarz then waste almost everything, because the spikes set the size of every average.
The fix. Multiply ζ by a short finite sum M(s) = Σn ≤ y a(n) n−s (a Dirichlet polynomial) chosen to approximate 1/ζ(s). Then ζ·M is close to 1 most of the time, so its averages are no longer dominated by spikes; and because M is a finite sum, the averages of |ζM|² can be actually computed and proved [known] Selberg 1942 via Titchmarsh §10.9 ff.; Iwaniec ch. 16, 20. Multiplying by M never removes a zero of ζ: ζM vanishes wherever ζ does.
The coefficients. 1/ζ(s) = Σ μ(n) n−s (μ = the Möbius function), so the naive choice is a(n) = μ(n) for n ≤ y. A sharp cut-off at y works badly; the key trick is smooth tapering,
so a(1) = 1 and the coefficients fade out as n approaches y [known] Levinson P(x) = x; Conrey 1983/1989 general P; Young eq. for ψ; Iwaniec (23.7). Selberg's version is the square of a smoothed partial sum of ζ−1/2 — he needs a positive weight |φ|², see §5 [known] Iwaniec ch. 16; Titchmarsh §10.9.
The length. y = Tθ. θ is the single most important parameter: a longer mollifier imitates 1/ζ better. But the averages of |ζM|² can only be proved when θ is below a barrier set by how well one can control the "off-diagonal" terms in the expanded sum: θ < ½ by classical means, θ < 4/7 for Conrey's coefficient shape using the Deshouillers–Iwaniec bounds for Kloosterman sums, θ < 17/33 for generic coefficients (Bettin–Chandee–Radziwiłł) [known] Conrey 1989 §1; PRZZ 2020 §1. More on this in §7.
A toy mollifier, by numbers [project] compute/toy_mollifier.py → compute/toy_mollifier.json. Take Levinson's taper P(x) = x and mollifiers of length y = 1 (no mollifier), 5, 10, 20, 32 (≈ T1/2) and 60, and sample t from 1000 to 1100 in steps of 0.01 (10,001 points). Two columns of interest: the spread (variance) of log|ζM|, and the concavity gap ½ log(mean |ζM|²) − mean log|ζM| — exactly the amount that Levinson's Step D (§6) throws away. Double-computed: A = mpmath ζ with the mollifier summed in mpmath; C = our own Euler–Maclaurin ζ in numpy (compute/em_numpy.py). |ζM| agrees to ≤ 7·10⁻⁷ at every point, and every number below agrees between A and C to within 5·10⁻⁴ (the largest difference is in the variance on the line, where log|ζM| has spikes to −∞ at the zeros).
| y | on the line σ = ½: var log|ζM| | mean |ζM|² | concavity gap | share of t with ½ ≤ |ζM| ≤ 2 | at σ₀ = ½ − 1.3/log 1050 ≈ 0.3131: var log|ζM| | mean |ζM|² | concavity gap |
|---|---|---|---|---|---|---|---|
| 1 (none) | 1.585 | 6.67 | 0.944 | 0.446 | 0.567 | 21.17 | 0.566 |
| 5 | 1.263 | 3.57 | 0.633 | 0.486 | 0.285 | 11.42 | 0.259 |
| 10 | 1.116 | 2.87 | 0.525 | 0.536 | 0.200 | 9.82 | 0.185 |
| 20 | 1.013 | 2.50 | 0.455 | 0.578 | 0.150 | 9.09 | 0.138 |
| 32 | 0.958 | 2.34 | 0.423 | 0.607 | 0.134 | 8.94 | 0.123 |
| 60 | 0.881 | 2.14 | 0.377 | 0.705 | 0.112 | 8.66 | 0.102 |
Reading it [project]:
- The mean of log|ζM| does not move (≈ 0.002–0.005 on the line for every y; ≈ 0.96 at σ₀). What the mollifier changes is the spread: on the line the variance falls from 1.59 to 0.88 and the mean square from 6.67 to 2.14; the spikes shrink (largest |ζM| 9.85 → 2.93). That is the sense of "flattening".
- The concavity gap — the loss in Step D — shrinks as y grows: 0.94 → 0.38 on the line, 0.57 → 0.10 at σ₀. A longer mollifier loses less, which is the θ-dependence of §7 in miniature.
- At σ₀ (left of the line) the mean log|ζ| is about 0.96, not 0. That is the functional equation: left of the line |ζ| is larger by roughly (t/2π)½−σ₀, and (½ − σ₀)·log(1050/2π) = 0.957. Levinson's actual set-up handles this by working with G and a shifted mollifier (§6, Step E); this toy only shows the flattening of the spread.
- Small-height illustration only: one t-range, simple mollifier, ζ instead of Levinson's G. Not a theorem.
Figure 2 The mollifier flattens |ζ|: the spikes shrink, the zeros stay
compute/fig_toy_mollifier.py. |ζ(½ + it)| (blue) andStep the mollifier longer: the spikes flatten, the zeros stay
Blue is |ζ| on the critical line; red is |ζ| multiplied by a mollifier M(s) = Σ μ(n)·P(log(y/n)/log y)·n−s of length y. The slider moves between six lengths. Nothing in between is drawn: each position is its own computation.
y = 60: largest |ζ·M| on this stretch = 2.89 (largest |ζ| = 9.04)
[project] Curves are on 1000 ≤ t ≤ 1030 at 1,201 points (step 0.025) from our own Euler–Maclaurin ζ in numpy, with Levinson’s taper P(x) = x; |ζ·M| was checked against mpmath at 61 points (largest difference 1.3·10−7). Largest values: 9.04, 4.76, 3.72, 3.18, 3.07, 2.89 for y = 1, 5, 10, 20, 32, 60. A small-height illustration of flattening, not a theorem. Data: anim_mollifier_curves.json.
| ζ·M(½ + it) | (red) for 1000 ≤ t ≤ 1030 with the y = 32 toy mollifier; the dashed line is size 1. Both curves vanish at the |
|---|
same points (the zeros of ζ), but the two big spikes of |ζ| (about 9) become bumps of about 3: the largest value on this stretch drops from 9.04 to 3.07. Curves from our numpy program; |ζM| checked against mpmath at 61 points (largest difference 1.3·10⁻⁷).
5. Selberg (1942): mollified sign changes in short windows
Intuition. Keep Hardy–Littlewood's plan — compare "integral of the function" with "integral of its size" over each window — but (a) use short windows, about one average zero-gap long, and (b) first multiply Z by a positive weight that flattens out the spikes. A positive weight cannot create or destroy sign changes, so every sign change it finds is still a zero of ζ on the line. Because the windows are now as short as the gaps, finding a sign change in a fixed share of them means finding a fixed share of all zeros.
The mechanism, as written out in Titchmarsh §§10.9–10.22 (Theorem 10.22 is Selberg's theorem; Titchmarsh follows Selberg's method). Selberg's own 59-page paper (Skr. Norske Vid. Akad. Oslo 1942 no. 10) was not opened for this page; his announcement (Norske Vid. Selsk. Forh. 15 (1942), Theorem 1: N₀(T) > A T log T) was read [known] the source notes.
- The mollifier. Let α_ν be the coefficients of ζ(s)−1/2 (so α₁ = 1), taper them as β_ν = α_ν (1 − log ν / log X) for ν < X, and put φ(s) = Σ β_ν ν−s. The weight is |φ(½ + it)|², a smoothed version of |ζ|−1; so Z·|φ|² behaves like ζ·ζ−1/2·ζ−1/2 ≈ 1 in size [known] Titchmarsh §10.9 ff.; Iwaniec ch. 16: "Selberg's mollifier is the square of the Dirichlet polynomial which is a smooth truncation of ζ(s)−1/2".
- The real function. Titchmarsh works with a real function F(t), built from Ξ(t) times |φ(½ + it)|² times explicit smooth factors; on the line it has the sign of Z(t) and size about t−1/4 e−δt/2 |ζφ²(½ + it)|, with δ = 1/T. The parameters are X = Tc (c small) and window length h = (ac log T)−1, a constant times the mean gap between zeros [known] Titchmarsh §10.15 and Thm 10.22.
- Two kinds of mean square — the heart of the proof [known] Titchmarsh Lemmas 10.17–10.21:
- size of the windows (first moment): ∫₀T dt ∫tt+h |F(u)| du > A h T3/4 (Lemma 10.21; it comes from ∫ ζφ²(½ + it) dt ~ T/2 over [T/2, T], i.e. ζφ² has average 1, Lemma 10.20);
- signed window integrals: ∫ |∫tt+h F(u) du|² dt = O(h T1/2 / log X) (Lemma 10.17);
- unsigned window integrals: ∫ (∫tt+h |F(u)| du)² dt = O(h² T1/2 log T / log X) (Lemma 10.19). Dividing, signed/unsigned = (h/log X)/(h² log T/log X) = 1/(h log T) = ac: in mean square the signed window integral is smaller than the unsigned one by the factor ac, which can be made as small as we like by taking a small [project] this "ratio = ac" reading is our arithmetic on Titchmarsh's stated bounds. Plainly: after mollifying, the function cancels inside a window much more than its size would allow if it kept one sign. Without the mollifier these two mean squares cannot be separated at this window length.
- The good set has positive density. Let E be the set of t in (0, T) where ∫tt+h|F| > |∫tt+h F| — on E the function F changes sign inside (t, t + h). Off E the two integrals agree, so all the excess of "unsigned" over "signed" lives on E. Cauchy–Schwarz on E with the three bounds in step 3 gives {m(E)}1/2 > A₁ c1/2 T1/2 − A₂ (ac)1/2 T1/2, so for a small enough, m(E) > A₃ T (m = length/measure) [known] Titchmarsh Thm 10.22, p. 279 [project] the two displayed inequalities were re-derived from Lemmas 10.17/10.19/10.21 and match.
- From windows to zeros. Split (0, T) into intervals (nh, (n+1)h). Since m(E) > A₃T, at least [A₃T/h] of them meet E; each forces a zero of ζ(½ + iu) in (nh, (n+2)h), and each zero is counted at most twice. Hence N₀(T) ≥ ½[A₃T/h] > A T log T — a positive fraction of the N(T) ≈ (T/2π) log T zeros [known] Titchmarsh p. 279.
- Why the fraction is tiny. Every step loses a constant: Cauchy–Schwarz, the comparison of moments, the factor ½ from double counting, and the fact that a window holding two zeros still counts once [known] PRZZ 2020 §1; Iwaniec ch. 16. Titchmarsh §10.9: "The numerical value of the constant A in Selberg's theorem is very small", with a footnote that it "was calculated in an Oxford dissertation by S. H. Min" (we have not seen that dissertation or its number). Explicit versions of a related sign-change method (Atkinson's) give κ > 0.0001049, improved to about 0.0087 by optimising a parameter [known] Baluyot, "On the zeros of Riemann's zeta-function on the critical line", §1 — numbers as quoted there.
Iwaniec calls Selberg's route "safe": it can never produce a negative bound, only a small one [known] Iwaniec ch. 16.
6. Levinson (1974): count the zeros that are not on the line
Intuition. Instead of hunting for zeros on the line, Levinson counts how many zeros a cousin G of ζ has to the right of the line, shows this number is small, and converts "few G-zeros on the right" into "many ζ-zeros on the line". Three facts make it work, and they are the five steps below:
- ζ(½ + it) = 0 exactly when a certain angle, the angle of H(s)G(s), points straight up or down (±90°).
- That angle has to turn about N(T) half-turns in total; every zero of G on the right side can absorb some turning that would otherwise have produced a crossing.
- The number of G-zeros on the right is controlled by the average of log|G·M| along a vertical line slightly left of ½ (Littlewood's lemma), and that average is controlled by one mean square (concavity), which is small only because of the mollifier M.
Levinson's own Adv. Math. paper was not opened for this page; every step below is cited to the later expositions (Iwaniec ch. 17–27; Conrey 1989 §§1–4; Young), which state that they follow Levinson.
Step A — a mirror identity [known] Iwaniec (17.4)–(17.7). Write ξ(s) = H(s)ζ(s) with H(s) = ½ s(1 − s) π−s/2 Γ(s/2), so ξ(s) = ξ(1 − s). Put G = ζ + λζ′ (λ real). Then
[project] identity re-derived by hand — the ξ′ terms cancel because ξ′(1 − s) = −ξ′(s). On the line, 1 − s is the complex conjugate of s and everything has real coefficients, so the right side is 2·Re(H·G). Y has no zeros for large |s| (Iwaniec: Y(s) ∼ −λ log|s|). So ζ(½ + it) = 0 exactly when arg(H·G) ≡ 90° (mod 180°) (up to finitely many exceptions). Levinson's choice is G ≈ ζ + ζ′/log T; as a finite sum G ≈ Σl ≤ T Q(log l / log T) l−s with Q(x) = 1 − x [known] Iwaniec (23.4)–(23.6).
Step B — the angle must turn; zeros of G on the right absorb turning [known] Iwaniec ch. 19, (19.14). Between heights T and T + U the number of crossings is at least (1/π)·(total change of arg H·G) − 1. Split the change into arg H plus arg G. The arg H part is essentially the change of Hardy's θ(t), i.e. about π·N(T, U): this is where the main term comes from. The arg G part is computed by the argument principle on a rectangle to the right of the line: each zero of G there absorbs a full turn, i.e. two potential crossings. Result:
or in Conrey's words: "if the proportion of zeros of [G] to the right of the critical line is ≤ p, then the proportion of zeros of ζ on the critical line is ≥ 1 − 2p" [known] Conrey 1989 §1. The zeros found this way are automatically simple [known] Heath-Brown, Bull. LMS 11 (1979); Selberg independently; cited in Conrey 1989 p. 1 and Iwaniec ch. 23. Why add ζ′/log T at all? Adding a small multiple of the derivative pushes the on-line zeros of ζ sideways to the left, so G has few zeros on the right [project] intuition, consistent with Levinson–Montgomery, Acta Math. 133 (1974) 49–65, and Speiser, Math. Ann. 110 (1935) 514–521: RH ⇔ ζ′ has no zeros with 0 < Re s < ½ — statement as in Farr–Pauli Thm 1.
Step C — Littlewood's lemma on a box with left edge a = ½ − R/log T [known] Iwaniec ch. 20–21, (20.9), (21.1)–(22.1); Conrey 1989 (37). Multiply G by a mollifier M (this only adds zeros, so an upper bound for the zeros of F = G·M on the right is enough). Littlewood's lemma says that, for F analytic on a rectangle,
and the other three edges contribute only O(U/log T). The left edge sits slightly left of ½, at a = ½ − R/log T, as a buffer: every zero on or right of the line is then at distance at least R/log T from the edge, so it is counted with a decent weight. Hence the number of zeros of F with Re ρ ≥ ½ is at most (log T/R)·(1/2π)∫ log|F(a + it)| dt plus small terms [known] Iwaniec (20.9).
Step D — concavity: the average of log is at most the log of the average [known] Iwaniec Thm 22.1; Conrey 1989 (38); Young (1.1). (1/U)∫ log|F| = ½(1/U)∫ log|F|² ≤ ½ log((1/U)∫|F|²). Putting this into Step C gives Levinson's principal inequality
κ ≥ 1 − (1/R) · log( (1/T) ∫₁T |G·M(a + it)|² dt ) + o(1), a = ½ − R/log T.
This is where the mollifier is indispensable: without it the mean square of |G| grows like log T and the bound becomes useless; with M ≈ 1/G it tends to a constant c, and the closer c is to 1 the bigger κ. Concavity loses only the "spread" of log|F| around its mean, which a good mollifier makes small [project] intuition.
Step E — one mean value [known] Young Thm 1 and Lemma 5; Iwaniec ch. 23–27; Conrey 1989 §§3–4. With M(s) = Σh ≤ Tθ μ(h) h−(s + ½ − a) P(log(Tθ/h)/log Tθ), the mean square tends to an explicit constant
so κ ≥ 1 − (log c)/R: everything reduces to a double integral. Levinson's choices (P(x) = x, Q(x) = 1 − x, R = 1.3, θ = ½) give c ≈ 2.35 and κ ≥ 0.34 [known] Young. The hard part is proving the mean value: expanding |G·M|² gives a sum over four integers, the "diagonal" terms (hm = kn) give c, and the "off-diagonal" terms must be shown to cancel. That is what limits θ (§7). Iwaniec calls this route "a gamble": unlike Selberg's, if the mean square comes out a little too big, the bound is negative and you get nothing [known] Iwaniec ch. 16.
Worked illustration — the whole argument at small heights [project] compute/levinson_*.json; all double-computed; none of this is asymptotic or a theorem.
Figure 3
Sweep up the line: each crossing of ±90° is a zero of ζ
Levinson’s first step says ζ(½ + it) = 0 exactly when the angle of H·G points straight up or straight down. Follow the angle as t rises; every time it crosses one of the dashed gold lines, a zero of ζ sits there.
t = 100.00 · angle of H·G = -3.4° · crossings of ±90° so far: 29 of 29
[project] G = ζ + ζ′/L with L = log(100/2π) = 2.767; 4,501 points on 10 ≤ t ≤ 100 (step 0.02), computed twice (mpmath and our own numpy program). The 29 crossings sit on the 29 zeros of ζ there (largest offset 3.5·10−6). Over the range arg H turns by 28.979 half-turns and arg G by only 0.005. Small-height illustration of Step A and Step B; not a proof. Data: levinson_HG_angle.json.
Numbers behind Steps B–D:
| check | result |
|---|---|
| Step A identity Y·ξ = 2 Re(H·G) on the line | agrees to 8 digits at 5 points; Y(½ + it) ≥ 1 on [10, 100] |
| Step B: G-zeros with σ ≥ ½, 10 < t < T (argument principle / Newton) | 0 / 0 for T = 100, and for T = 300 with L = log T and with L = log(T/2π) |
| Step C buffer, T = 300, Levinson's L = log T, R = 1.3, a = 0.2721 | 16 G-zeros in the buffer a ≤ σ < ½ (both methods); 122 further left |
| Littlewood's lemma on that box | 2π·Σ(β − a) = 1.48557 from the zeros vs 1.48766 from the four edge integrals (0.14% apart) |
| Step C conclusion | Σ(β − a)/(R/log T) = 1.04, so "at most 1 zero right of ½" — true (there are 0) |
| Step D without a mollifier, same edge | mean log|G| = 0.00588, but ½ log mean|G|² = 0.699 (mean|G|² = 4.05) |
| Step D bound shape 1 − (1/R) log mean|G|² | −0.076 (useless), versus 0.991 using the exact log-average |
The last two rows are the case for the mollifier in one line: even at height 300 the spikes of |G| inflate the mean square so much that the concavity step throws away everything [project] illustration only: finite T, edge terms ignored. Two cautions from the data: the 16 buffer zeros are the G-partners of the 15 lowest ζ-zeros and of the 18th (t ≈ 72.06), not "the 16 lowest"; and from t ≈ 67 the G-zeros sit within about 10⁻³ of the edge a, so at this height the edge happens to run through where the G-zeros cluster — the classification is still exact (Newton positions are accurate to about 10⁻¹²; fine-step re-runs agree) [project].
7. Longer mollifier ⇒ bigger share: the role of θ, and why 4/7 matters
Intuition. The mollifier M has length y = Tθ. The longer it is, the better it imitates 1/G, the flatter |G·M| is, the smaller the mean square c, and the bigger κ = 1 − (log c)/R. In Young's formula θ enters most visibly through the 1/θ in front of the double integral: a longer mollifier shrinks the "left-over roughness" term [project] reading of Young's formula.
Why θ is capped. In the expanded mean square, an off-diagonal term oscillates like (hm/kn)it. With h, k ≤ Tθ and m, n ≲ T, two different integers hm ≠ kn are far enough apart that integrating over a t-range of length about T kills the term — but only when θ < ½ [known] Young Lemma 5 (requires θ < ½) and §6; Iwaniec ch. 27 [project] paraphrase of the exponent bookkeeping. At θ = ½ the off-diagonal terms stop being noise that integrates away and become structured arithmetic sums of Kloosterman type. Conrey 1989 shows they still cancel, on average over the moduli, up to θ = 4/7: "mollifier of length y = Tθ with θ = 4/7 − ε — whereas in Levinson's theorem θ = ½ − ε. The work of Deshouillers and Iwaniec on averages of Kloosterman sums is what allows us to use a longer mollifier" [known] Conrey 1989 §1, quoted. The proof combines Balasubramanian–Conrey–Heath-Brown 1985 (reducing the error to incomplete Kloosterman sums), Conrey 1983 (the main term) and Deshouillers–Iwaniec 1982 (the cancellation, via the spectral theory of automorphic forms) [known] Conrey 1989 §1. That jump from ½ to 4/7 is what lifted the proved share from 0.365 (Conrey 1983, θ = ½) to 0.4077/0.4088 (Conrey 1989).
Figure 4
compute/levinson_kappa.json and levinson_extrap.json; two independent integrators (Gauss–Legendre in numpy, mpmath.quad) agree to about 10⁻¹⁴:| set-up | θ | best R | κ bound | compare with the literature |
|---|---|---|---|---|
| Young's fixed choice P = x, Q = 1 − x, R = 1.3 | ½ | 1.3 | 0.34274 (c = 2.350068) | Young: "c = 2.35…, κ ≥ 0.34…"; Levinson's paper 0.3420 |
| P = x, linear Q, R optimised | ½ | 1.246 | 0.34744 | Levinson 1975: 0.3474 |
| P = x, linear Q, R optimised | 4/7 | 1.142 | 0.39113 | — |
| P cubic, linear Q | ½ | 1.242 | 0.35627 | Conrey–Iwaniec–Soundararajan remark: 0.356 at θ = ½ |
| P cubic, linear Q | 4/7 | 1.132 | 0.40219 | Conrey 1989: κ* ≥ 0.401 |
| P = x, linear Q | 0.1 / 0.2 / 0.3 / 0.4 | −0.129 / 0.056 / 0.180 / 0.274 | too short a mollifier gives a useless (even negative) bound |
The ratio 0.34744/0.34274 ≈ 1.014 matches the "only 1.4%" improvement in Levinson's 1975 PNAS abstract [project] consistent with sources. Conrey's 0.4077/0.4088 at θ = 4/7 is higher than our 0.40219 because he optimises more freely [known] Conrey 1989 Thm 1.
Past 4/7 — formula only, no theorem [project]. Evaluating the same formula at θ = 0.6, 0.7, 0.8, 0.9, 1.0 gives 0.419, 0.471, 0.516, 0.554, 0.5866. These are not theorems: for ζ itself no mean value with θ > 4/7 (for this shape) is proved. The θ = 1 value reproduces the 0.5865 that Conrey–Iwaniec–Soundararajan quote for the optimal mollifier at θ = 1 in the family of Dirichlet L-functions, where θ = 1 is provable on average [known] CIS, arXiv 1105.1177. Even a hypothetical θ = 1 mollifier of these shapes gives only about 59%: the method has a ceiling well below 100% unless θ grows without bound [project] PRZZ §1 cites Farmer's "θ = ∞" heuristic and Bettin–Gonek (θ = ∞ ⇒ RH). The later records (0.4105, 0.4128, 0.417293) came from richer G (more derivatives) and multi-piece mollifiers rather than longer θ [known] BCY 2011, Feng 2012, PRZZ 2020.
8. What these methods do not give
- A lower bound for a proportion in the long run (a liminf as T → ∞). Nothing about any particular finite stretch; and conversely finite computations (all zeros to height 3·10¹² on the line) say nothing about the liminf [known].
- Constant losses at every step — Cauchy–Schwarz and double counting for Selberg; the factor 2 in Step B, the buffer (the 1/R), and concavity for Levinson. The records come from better G, better M and longer M [known] PRZZ 2020 §1; Conrey 1989 §1.
- No route to 100%. No version of Levinson's method is expected to reach 1 [known] PRZZ 2020 intro, and none of this is progress on RH.
- Selberg's and Levinson's methods are "diametrically opposed" (Conrey's comparison, quoted in PRZZ §1): Selberg counts on-line zeros directly and can never give a negative bound; Levinson counts off-line zeros of a cousin and subtracts, and gets 1/3+ from a single mean value but can fail outright if the mean value is too big [known] PRZZ 2020 §1; Iwaniec ch. 16.
9. Worked example: Z(t) sign changes versus N(T) up to height 2000
[project] double-computed: A = mpmath (siegelz on a 0.02 grid, grampoint, nzeros), B = our own Riemann–Siegel program in numpy (main sum + first correction term, own θ, own Gram points, 0.005 grid). Pointwise the two Z's agree to ≤ 1.1·10⁻² (t < 100), ≤ 3.9·10⁻³ (100–500), ≤ 1.1·10⁻³ (500–2000); A and B agree on every entry.
| T | N(T) (mpmath.nzeros) | sign changes on the grid (A = B) | Gram points gn ≤ T (from n = −1) | sign changes in the sequence Z(gn) | "good" Gram points |
|---|---|---|---|---|---|
| 100 | 29 | 29 | 30 | 29 | 30 |
| 200 | 79 | 79 | 80 | 79 | 80 |
| 500 | 269 | 269 | 270 | 257 | 264 |
| 1000 | 649 | 649 | 649 | 604 | 627 |
| 1500 | 1069 | 1069 | 1070 | 975 | 1023 |
| 2000 | 1517 | 1517 | 1517 | 1358 | 1438 |
- Sign changes on a fine grid = N(T) at every checkpoint: every zero up to 2000 shows up as a sign change of Z, i.e. is on the line and simple. This is long known — Gram 1903 (15 zeros), Backlund 1914 (79, to t = 200), Hutchinson 1925 (138, to t = 300), Titchmarsh 1935–36 with computations planned by Comrie (1041 zeros, to t = 1468) [known] the source notes, "Early numerical verification". We only re-saw it. (Caveat: the N(T) column is mpmath's own count; the genuinely double-computed part is the agreement of the two independent sign-change counts.)
- Sampling Z only at Gram points catches fewer: 1358/1517 ≈ 89.5% up to 2000, and the miss rate grows with T (Gram's law fails more and more often). How you look decides how many zeros you can certify — the story of §3 → §5 → §6 in miniature: Hardy–Littlewood look in fixed windows, Selberg mollifies and looks in short windows, Levinson stops looking for sign changes and counts the off-line zeros of a cousin instead [project].
- 1 + θ(T)/π (the smooth part of N(T)) is within 0.6 of N(T) at every checkpoint, e.g. 1516.99 vs 1517 [project].
10. Does any of this transfer to the project's hypercomplex objects?
A short version of this section also sits on the quaternion and octonion page.
This section is about the project's own objects (see /workspace/math/lit/quaternionic-zeta/hyper-critical/, especially q8-half-hecke-20261111.md). Labels matter more than usual here. No [project] claim below finds, moves or excludes any zero, and none of it bears on RH.
10.1 The classical Q8 Artin L-function L(ρ, s). Here K/ℚ is Galois with group Q8 and ρ is its 2-dimensional faithful representation; L(ρ, s) equals a Hecke L-function L(F, ψ, s) of a quadratic field F with an order-4 character ψ [known] induction of characters, as set up in the q8-half-hecke note. It is an ordinary L-function: it has a functional equation Λ(s) = W·Λ(1 − s), and since the character of ρ is real (traces 2, −2, 0), its coefficients are real and W = ±1. So Hardy's Z-function idea transfers directly: W−1/2Λ(½ + it) is real, and its sign changes are zeros on the line [project] elementary; this is what PARI's lfunhardy computes.
- Check [project]
compute/q8_hardy_z.py→q8_hardy_z.json; PARI only, so a consistency check rather than an independent double computation: for 2.2304.8t5.b.a (W = −1) and 2.2304.8t5.a.a (W = +1), on 0 < t < 30 (step 0.02) both lfunhardy and our own normalisation of lfunlambda give 42 sign changes with identical signs at every grid point, and PARI's lfunzeros finds 43 and 42 zeros on [0, 30]: the extra one for W = −1 is the forced central zero at t = 0 (order 1), which a sign change on t > 0 cannot see. So at these heights every zero of these two L(ρ) is on the line and simple. - Selberg's method has been carried over to degree-2 L-functions: Hafner proved N₀(T) > A T log T for L-functions of Hecke eigenforms of even weight on the full modular group, using a mollifier that approximates L−1/2, Good's approximate functional equation and spectral theory [known] Hafner, Math. Ann. 264 (1983) 21–37; announcement read in Bull. AMS (N.S.) 8 (1983), no. 2, from p. 340, and later for Maass wave forms of level 1 [known] Hafner, J. reine angew. Math. 377 (1987) 127–158, as cited by Andersen–Thorner; not opened.
- Levinson's method has been carried over too, with an explicit share. Andersen and Thorner prove that for any holomorphic or Hecke–Maass newform f on Γ₀(N) with trivial nebentypus, more than 6.96% of the zeros of L(s, f) are on the critical line (8.96% under the generalised Ramanujan conjecture). They follow Bernard's version of Levinson's method; the mollifier length allowed is ν < ¼ − θ/2 (θ = progress toward Ramanujan, ≤ 7/64), against ½ for ζ, and the extra length over Bernard comes from the Blomer–Harcos spectral decomposition of shifted convolution sums [known] N. Andersen, J. Thorner, "Zeros of GL₂ L-functions on the critical line", arXiv 2004.03581, Thm 1.1 and Thm 2.1 — read on the arXiv page; journal status not checked. Earlier: Bernard, Acta Arith. 167 (2015)
201–237 (κ_f > 0.0297 for holomorphic f of squarefree level; as cited there, not opened).
- Does that cover our L(ρ, s)? Yes. Update 12 November 2026: the standard step is now cited and checked (Q₈ Artin L-function as a Maass newform). The original reasoning follows. Here ρ is even (complex conjugation acts as −I or I), det ρ = 1 (so the nebentypus is trivial), the conductor is 2304, PARI gives the gamma factors Γ_ℝ(s+1)² (2.2304.8t5.b.a) and Γ_ℝ(s)² (2.2304.8t5.a.a), and all three quadratic subfields are real (discriminants 8, 12, 24) [project] PARI lfunparams/nfsubfields. So ρ is induced from a character of a real quadratic field, and by the classical dihedral (Maass) construction L(ρ, s) should be L(s, f) for a Hecke–Maass newform f of level 2304, trivial nebentypus, Laplace eigenvalue ¼ [known] in general; this identification was not re-checked here. Granting it, Andersen–Thorner give κ > 0.0696 for these L(ρ, s). Compare ζ: 0.417293 by the same family of methods. The mollifier is the bottleneck in both cases — for degree 2 the provable length is much shorter (ν < ¼ − θ/2, against θ < ½, or 4/7, for ζ). What remains to check is the Maass identification (a citation), the paper’s journal status, and whether the 8.96% figure applies to dihedral forms.
- Levinson's Steps A–D are pure algebra and complex analysis for any L-function with this kind of functional equation [project] elementary: the Step A algebra used only ξ(s) = ξ(1 − s) and real coefficients; Andersen–Thorner's (2.11) is exactly the degree-2 form of Step D, with log c/(2R) [known] arXiv 2004.03581 (2.11). The hard part is again Step E, the mollified mean square.
10.2 The aligned quaternion "half-L-function" F. The project's Model A product over split primes is (q8-half-hecke note, result 1, [project] exact)
with G zero-free on Re s > 1/3. Measured against the three steps above:
- No Z-function, no mirror identity [project] elementary. The representation above gives F no reflection law about ½: the factor [ζ(2s)/L(χu, 2s)]1/4 reflects about s = ¼ (2s ↔ 1 − 2s), and G is only defined for Re s > 1/3. So neither Hardy's real-valued Z nor Levinson's identity "ξ = 2·Re(H·G)" is available for F. [speculative] F probably has no functional equation of any kind; not proved.
- No single-valued function to count zeros of. F is multivalued: going round a zero of ζ or L(χu) multiplies |F| by e−π/2, round a zero of L(χv) or L(χw) by e+π/2; and the full Model A norm |F|²·O(σ) tends to 0 like (2σ − 1)1/2 as σ → ½⁺ at every height, because half of the primes (the inert ones) lose their phases [project] q8-half-hecke note, results 3 and 5. On the critical line the Model A norm has no sign changes to count.
- Littlewood's lemma does transfer — with weights equal to the real parts of the exponents [project] elementary, numerically checked. Continue log F horizontally from the right (the usual convention behind S(t)). Then Littlewood's lemma applies factor by factor: ∫ over the boundary of log f ds = −2πi·a·Σ(distances) for f = La, and the usual log|f| version of the lemma is minus the imaginary part of that, i.e. 2π·Re(a)·Σ(distances). So a Levinson-type count run through |F| sees the zeros of L(ρ) with weight ½, zeros of G with weight 1, zeros of ζ(2s) with weight ¼ and of L(χu, 2s) with weight −¼ — and the zeros of ζ, L(χu), L(χv), L(χw) (exponents ±i/4) with weight 0: they only feed the boundary terms. Check on f = ζa, box 0.3 ≤ σ ≤ 3, 10 ≤ t ≤ 40 (six zeros, Σ distances = 1.2) [
compute/littlewood_complex_exponent.py→ .json; numpy Euler–Maclaurin vs mpmath on a coarser grid]:
| a | Littlewood expression (numpy) | (mpmath) | predicted 2π·Re(a)·1.2 |
|---|---|---|---|
| 1 | 7.53982 | 7.53969 | 7.53982 |
| ½ | 3.76991 | 3.76984 | 3.76991 |
| ½ + i/4 | 3.77189 | 3.77972 | 3.76991 |
| i/4 | 0.00198 | 0.00988 | 0 |
| ¼ | 1.88495 | 1.88492 | 1.88496 |
(The residue for imaginary exponents shrinks with the grid step — 0.0099 at t-step 0.04, 0.0020 at 0.005 — as expected from the jumps of arg ζ along the left edge where the horizontal paths pass a zero.) Consequence [project]: any Levinson-type count through |F| is half the count for L(ρ) plus terms from Dirichlet L-functions at 2s; the quaternion directions add no new zero information. Whether the boundary terms coming from the i/4 pieces are small enough not to spoil the count is open.
10.3 Plain summary. Sign changes (Hardy, Selberg) and Levinson's identity need a real-valued function on the line coming from a functional equation. The classical L(ρ) has one; the aligned quaternion product F does not. The one tool that does carry over to F — Littlewood's lemma — shows that F's zero bookkeeping is just "half of L(ρ)", with the imaginary-exponent pieces invisible. [project] A resemblance worth recording without claiming anything: Selberg's mollifier is itself built from a half power (a smoothed partial sum of ζ−1/2), and F's main factor is L(ρ)1/2 [speculative] no mathematical use found for this; it is noted only so it is not rediscovered as if new.
11. Figures and computations used here
All four figures above are drawn from the computations below, each run twice by independent programs unless stated. The scripts and their outputs are in the research folder; start with em_numpy.py, our own Euler–Maclaurin ζ in numpy.
- Figure 1: zt_gram.py → zt_gram.json (Z(t), 13 zeros, Gram points).
- Figure 2 and the toy-mollifier table: toy_mollifier.py, fig_toy_mollifier.py.
- Figure 3: fig_levinson_angle.py with the angle data in levinson_HG_angle.json.
- Figure 4: fig_kappa_theta_svg.py, from levinson_kappa.json and levinson_extrap.json.
- Short-window table: hl_windows.py. Littlewood-lemma check: littlewood_complex_exponent.py. Q8 Hardy Z: q8_hardy_z.py.
The whole write-up, as plain text: primer.md. All four figures are palette SVGs and follow the site’s light and dark themes.
12. Sources
- G. H. Hardy, "Sur les zéros de la fonction ζ(s) de Riemann", C. R. Acad. Sci. Paris 158 (1914) 1012–1014.
- G. H. Hardy, J. E. Littlewood, "The zeros of Riemann's zeta-function on the critical line", Math. Z. 10 (1921) 283–317.
- A. Selberg, "On the zeros of Riemann's zeta-function", Skr. Norske Vid. Akad. Oslo I 1942 no. 10 (not opened; the proof was read in Titchmarsh §§10.9–10.22, and the announcement in Norske Vid. Selsk. Forh. 15 (1942)).
- N. Levinson, "More than one third of zeros of Riemann's zeta-function are on σ = ½", Adv. Math. 13 (1974) 383–436 (not opened); PNAS 71 (1974) 1013–1015 and 3984–3987; PNAS 72 (1975) 294–297 (abstracts read).
- J. B. Conrey, J. Number Theory 16 (1983) 49–74; J. reine angew. Math. 399 (1989) 1–26.
- H. M. Bui, J. B. Conrey, M. P. Young, Acta Arith. 150 (2011) 35–64. S. Feng, J. Number Theory 132 (2012) 511–542.
- K. Pratt, N. Robles, A. Zaharescu, D. Zeindler, "More than five-twelfths of the zeros of ζ are on the critical line", Res. Math. Sci. 7 (2020), Paper 2.
- E. C. Titchmarsh, The Theory of the Riemann Zeta-Function, 2nd ed., rev. D. R. Heath-Brown, Oxford 1986, ch. X.
- H. Iwaniec, Lectures on the Riemann Zeta Function, AMS ULECT 62 (2014). M. P. Young, Arch. Math. 95 (2010) 539–548.
- J. B. Conrey, H. Iwaniec, K. Soundararajan, "Critical zeros of Dirichlet L-functions", arXiv 1105.1177.
- S. Baluyot, "On the zeros of Riemann's zeta-function on the critical line" (expository; Atkinson-type constants).
- J. L. Hafner, "Zeros on the critical line for Dirichlet series attached to certain cusp forms", Math. Ann. 264 (1983)
21–37 (announcement: Bull. AMS (N.S.) 8 (1983), no. 2, from p. 340); Maass forms: J. reine angew. Math. 377 (1987) 127–158 (not opened).
- N. Andersen, J. Thorner, "Zeros of GL₂ L-functions on the critical line", arXiv 2004.03581 (read on the arXiv page; journal not checked). D. Bernard, "Modular case of Levinson's theorem", Acta Arith. 167 (2015) 201–237 (as cited; not opened).
- Early numerics: Gram 1903, Backlund 1914, Hutchinson 1925, Titchmarsh 1935/1936 — see the source notes.
- 2026 preprints: arXiv 2608.13637 (Alpöge–Furman), 2609.02882 (Lamzouri), 2610.08965 (Knausgård); comment arXiv
2610.24612 (Arias de Reyna) — see the source notes, 2026 section.
Reading path · step 4 of 13