# How much of the critical line can we prove? Selberg's and Levinson's methods

*Primer for Grant — written in block 4 of the overnight session (Thu Nov 12, 2026, 3:30–5:15 AM PT) from the block 1–3
notes (mechanism-notes.md, sources.md, compute/). Status: draft for the 5:20 AM webdev block.*

**How to read the labels.** Every claim carries one of three tags.
- **[known: source]** — published mathematics; the source is named (full details in §12 and sources.md).
- **[project]** — something we computed or checked ourselves this week (small heights, double-computed unless it says
  otherwise). A [project] number is evidence, never a theorem.
- **[speculative]** — our own reading or guess, not proved.

Short names used below: *Titchmarsh* = E. C. Titchmarsh, *The Theory of the Riemann Zeta-Function*, 2nd ed., revised by
D. R. Heath-Brown (Oxford, 1986); *Iwaniec* = H. Iwaniec, *Lectures on the Riemann Zeta Function* (AMS University
Lecture Series 62, 2014); *Young* = M. P. Young, "A short proof of Levinson's theorem", Arch. Math. 95 (2010) 539–548;
*Conrey 1989* = J. B. Conrey, "More than two fifths of the zeros of the Riemann zeta function are on the critical
line", J. reine angew. Math. 399 (1989) 1–26.

**Nothing in this primer is progress on the Riemann Hypothesis.** RH says *every* non-trivial zero is on the line.
The theorems below prove that *a fixed share* of them is. That is a different, much weaker statement, and no method
described here is expected to reach 100%.

---

## 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 sources.md.

| 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"; sources.md §2026]

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

  Z(t) = e^{iθ(t)} ζ(½ + it).

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].

![Z(t) for 0 < t < 60, with its 13 sign changes and the Gram points](figures/zt-zeros-gram.svg)

*Figure 1 — figures/zt-zeros-gram.svg.* Z(t) for 0 < t < 60 [project]. It changes sign 13 times, at the 13 zeros of ζ
below height 60 (14.1347, 21.0220, 25.0109, 30.4249, 32.9351, 37.5862, 40.9187, 43.3271, 48.0052, 49.7738, 52.9703,
56.4462, 59.3470). The dashed verticals are the **Gram points** g₋₁ … g₁₁, defined by θ(g_n) = nπ; at a Gram point the
rotation e^{iθ} is ±1, so Z(g_n) = ±ζ(½ + ig_n). Below height 60 every Gram point obeys "Gram's law"
(−1)ⁿ Z(g_n) > 0, so consecutive Gram points have opposite signs and each Gram interval holds at least one zero [project]. Higher
up the law fails more and more often (§9 table) [project].
*Double computation [project]:* program A = mpmath (siegelz, zetazero, grampoint); program C = our own Euler–Maclaurin
ζ in numpy with θ from scipy's complex log-gamma, our own bisection for zeros and our own Newton solve for Gram points
(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.

**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;
sources.md]. 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

  I(t) = ∫_t^{t+Δ} Z(u) du,  J(t) = ∫_t^{t+Δ} |Z(u)| du.

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
  ∫_T^{2T} |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; mechanism-notes §3]. 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) = Σ_{n ≤ 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,

  a(n) = μ(n) · P(log(y/n)/log y), with P(0) = 0, P(1) = 1,

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 = T^θ. 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 (≈ T^{1/2}) 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.

![|ζ| and |ζ·M| on the line for 1000 ≤ t ≤ 1030](figures/mollifier-flattening.svg)

*Figure 2 — figures/mollifier-flattening.svg* [project, compute/fig_toy_mollifier.py]. |ζ(½ + it)| (blue) and
|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) could not be opened this week;
his announcement (Norske Vid. Selsk. Forh. 15 (1942), Theorem 1: N₀(T) > A T log T) was read [known: sources.md].

1. **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}"].
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 = T^c (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].
3. **Two kinds of mean square** — the heart of the proof [known: Titchmarsh Lemmas 10.17–10.21]:
   - *size of the windows* (first moment): ∫₀^T dt ∫_t^{t+h} |F(u)| du > A h T^{3/4} (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: ∫ |∫_t^{t+h} F(u) du|² dt = O(h T^{1/2} / log X) (Lemma 10.17);
   - *unsigned* window integrals: ∫ (∫_t^{t+h} |F(u)| du)² dt = O(h² T^{1/2} 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.
4. **The good set has positive density.** Let E be the set of t in (0, T) where ∫_t^{t+h}|F| > |∫_t^{t+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 c^{1/2} T^{1/2} − A₂ (ac)^{1/2} T^{1/2},
   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].
5. **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].
6. **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:
1. ζ(½ + it) = 0 exactly when a certain angle, the angle of H(s)G(s), points straight up or down (±90°).
2. 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.
3. 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 could not be opened this week; 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

  Y(s) ξ(s) = H(s)G(s) + H(1 − s)G(1 − s),  Y(s) = 2 − λ(H′/H(s) + H′/H(1 − s)).

[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 ≈ Σ_{l ≤ 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:

  N₀₁(T, U) ≥ N(T, U) − 2 N_G(right) + O(U),

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,

  2π · Σ_{zeros ρ in the box} (distance of ρ from the left edge) = ∫ log|F(a + it)| dt + (three other edge terms),

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) ∫₁^T |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) = Σ_{h ≤ T^θ} μ(h) h^{−(s + ½ − a)} P(log(T^θ/h)/log T^θ), the mean square tends to an explicit constant

  c(P, Q, R, θ) = 1 + (1/θ) ∫₀¹∫₀¹ e^{2Rv} ( d/dx [e^{Rθx} P(x + u) Q(v + θx)] at x = 0 )² du dv,

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].

![Angle of H·G along the line, and zeros of G versus zeros of ζ](figures/levinson-angle-zeros.svg)

*Figure 3 — figures/levinson-angle-zeros.svg.* Top: the angle of H·G along s = ½ + it for 10 < t < 100, with
G = ζ + ζ′/L, L = log(100/2π). It crosses ±90° exactly **29** times, and the 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 — G contributes
essentially no turning, which is what Step B predicts when G has no zeros on the right. Bottom: zeros of G (squares)
all lie left of the line, below the zeros of ζ (dots), with Levinson's box edge a = ½ − R/log T drawn for R = 1.3, T = 100.

Numbers behind Steps B–D (mechanism-notes §9.7):

| 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 = T^θ. 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 ≤ T^θ 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 = T^θ 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).

![Levinson's bound κ as a function of the mollifier length θ](figures/kappa-vs-theta.svg)

*Figure 4 — figures/kappa-vs-theta.svg.* Our evaluation of Young's formula [project, 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; mechanism-notes §8; 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 g_n ≤ T (from n = −1) | sign changes in the sequence Z(g_n) | "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: sources.md,
  "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?

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^{−1/2}Λ(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 in block 4; 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)?** Probably yes, with one standard step we did not re-check. 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, block 4]. 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). Open
  question 1 in open-questions.md keeps the remaining checks.
- 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])

  F(s) = L(ρ, s)^{1/2} · [ζ(s)L(χ_u, s)/(L(χ_v, s)L(χ_w, s))]^{i/4} · [ζ(2s)/L(χ_u, 2s)]^{1/4} · G(s),

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(χ_u, 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(χ_u) multiplies |F| by
  e^{−π/2}, round a zero of L(χ_v) or L(χ_w) 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, new this block;
  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 = L^a, 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(χ_u, 2s) with weight −¼ — and the zeros of ζ, L(χ_u), L(χ_v), L(χ_w) (exponents ±i/4) with weight **0**: they only
  feed the boundary terms. Check on f = ζ^a, 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 question 2.

**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

| figure | file | made by | what it shows |
|---|---|---|---|
| 1 | figures/zt-zeros-gram.svg | compute/zt_gram.py (block 4) | Z(t), 0 < t < 60, 13 zeros, Gram points g₋₁…g₁₁ |
| 2 | figures/mollifier-flattening.svg | compute/fig_toy_mollifier.py (block 4) | \|ζ\| vs \|ζ·M\| on the line, 1000 ≤ t ≤ 1030 |
| 3 | figures/levinson-angle-zeros.svg | compute/fig_levinson_angle.py (block 3) | angle of H·G vs ±90°; G-zeros vs ζ-zeros, t < 100 |
| 4 | figures/kappa-vs-theta.svg | compute/fig_kappa_theta_svg.py (block 3) | Levinson's bound κ against mollifier length θ |

All four are palette SVGs (literal hex colours, opaque #fffdf8 background, no transparency, text kept as text).
Computations: compute/em_numpy.py (independent ζ by Euler–Maclaurin), zt_gram.py/.json, toy_mollifier.py/.json,
fig_toy_mollifier.py/.json,
littlewood_complex_exponent.py/.json, q8_hardy_z.py/.json (block 4); levinson_*.py/.json, hl_windows.py/.json,
progA/progB (blocks 1–3).

## 12. Sources (short list; details and how each was checked: sources.md)

- 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 — read in block 4); 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 block 4; 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 sources.md.
- 2026 preprints: arXiv 2608.13637 (Alpöge–Furman), 2609.02882 (Lamzouri), 2610.08965 (Knausgård); comment arXiv
  2610.24612 (Arias de Reyna) — see sources.md §2026.
