# Zero side of the 2T pure part — 13 Nov 2026 (PST), sprint

Closes the open item "Zero side for the 2T pure part" from `open-questions.md` (binary tetrahedral section). Fields: LMFDB 2.26569.24t7.b.a and .a.a. Code and data: `tetra-zero1113/`. Figure: `figures/tetra-pure-zero-side.png` (+ `.svg`, plain matplotlib, not site-styled). Labels: **[known]** published theorem, **[project]** this project, **[speculative]**. All numerics are floating point (mpmath 30 digits / numpy-scipy doubles), PARI zeros are not interval certificates. Nothing here says anything about the Riemann Hypothesis.

## 1. In plain words

The quaternion Frobenius of a prime has a real part and a "pure" (i,j,k) part. Earlier notes showed the real part is half the trace of ρ, so it is heard by the zeros of L(ρ) alone. The pure part is not: in the aligned convention it is a mixture of **other** L-functions. This sprint rebuilds that mixture from the zeros of those L-functions, and it balances to 10⁻¹⁵–10⁻¹¹ (T ≥ 1.5), inside an honest error budget.

- The **u-direction** (the order-4 axis) is made from ζ, L(ω), L(ω̄) and L(ψ3), with weights ¼, ¼, ¼, −¼.
- The **n-direction** (the order-6 axis) is made *only* from zeros of the complex L-function L(ρ⊗ω) (and its mirror L(ρ⊗ω̄)). It cannot be seen with an ordinary even test function: it is the **difference between the zeros above and below the real axis**, and needs an *odd* test function. This is the new feature compared with Q8.
- The central zero of L(ρ) contributes **nothing** to the pure part (weight a_ρ = 0), checked three ways below.

## 2. What was computed

**Zeros [project, PARI 2.x via cypari2, uncertified].** For each field all seven irreducible Artin L-functions were built with `lfunartin` (functional equation checked, 119–131 digits) and `lfunzeros` run to height 110, for L and for its dual (`zeros-b.a.json`, `zeros-a.a.json`, `zeros_export.py`). Counts at height 110, b.a: ζ 33, L(ω) 122, L(ω̄) 122, L(ψ3) 276, L(ρ) 244 + central zero, L(ρ⊗ω) 155 positive and 154 for its dual (not symmetric: 155 vs 154); a.a: 243/243 (counts not compared with a proved N(T); a missed zero would show up in the residuals below).

**Explicit formula, non-self-dual form [known: Weil 1952; Iwaniec–Kowalski, Thm 5.12].** For a test function h with Fourier transform w(x) = (1/2π)∫h(r)e^{−irx}dr (not necessarily even), ρ = ½ + iγ:

 Σ_ρ h(γ) = [pole terms] + (1/2π)∫h(r)[log N + Σ_j Re ψ((½+κ_j+ir)/2) − d log π]dr − Σ_n [Λ_π(n)w(log n) + conj(Λ_π(n))w(−log n)]/√n.

Consequences used here (derived from the contour argument, then checked numerically):
- **even h**: zeros of π and of its dual π̄ together (positive ordinates of both lists), prime side −2 Σ Re Λ_π(n) w(log n)/√n. It sees only Re Λ.
- **odd h = iH**: the archimedean and pole terms vanish and Σ_ρ H(γ) = Σ(γ>0 for π) H − Σ(γ>0 for π̄) H = −2 Σ Im Λ_π(n) W(log n)/√n. So **Im Λ_π, i.e. the n-direction, is a sum over upper zeros minus lower zeros**. The positive zeros of π̄ are the conjugates of the negative zeros of π, which is why the zeros of ρ⊗ω̄ "pair" with those of ρ⊗ω.
- Combination of coefficients: n-component = −(i/2)·Im-part-formula for ρ⊗ω + (i/2)·… for ρ⊗ω̄ = (H_π − H_π̄)/2 = H_π, a real number (since H_π̄ = −H_π for the odd test function). Coefficients checked by exact inner products with the character table: ρ⊗ω −i/2, ρ⊗ω̄ +i/2, all other n-coefficients 0 (`coeff_check.py`, `coeffs-*.json`); u-coefficients ζ, ω, ω̄ = ¼, ψ3 = −¼, **ρ = 0 exactly** (both fields).

**Test functions.**
- even Gaussian h = e^{−(r/T)²}, T = 1, 1.5, 2, 3, 5, 8, 12, 16, 20 (as for Q8), plus the cos-shifted family h = e^{−(r/T)²}cos(x₀r), T = 8, x₀ = 1…8.5.
- odd, for the n-direction: (G) w(x) = g(x)·erf(8x) ("sgn × Gaussian", equal to the Gaussian weight g(log n) of the earlier tables to better than 10⁻²⁰ for n ≥ 2), T = 1…8 (larger T leaves the prime side below 10⁻¹⁹ and H decays too slowly for the zero list); (S) the sine-shifted family h = i e^{−(r/T)²} sin(x₀r), w = ½[g(x−x₀) − g(x+x₀)], T = 8, x₀ = 1…8.5 and T = 2,3,5 with x₀ = 1,2,3. There is no odd analogue at x₀ = 0 (it is identically 0).

**Prime side.** Method A: actual aligned quaternion g_p (representatives: the six powers of g₆ = (1+i+j+k)/2, and i for the order-4 class), exact prime powers p^k ≤ 20000, unramified p. Method B: Λ from PARI a_n by Dirichlet recursion. The ramified prime 163 is **not** in the quaternion sum; for ρ⊗ω its term in the n-direction is exactly 0 (Λ_{ρ⊗ω}(163^k) is real, the local factor is 1+T), so the n-direction needs no ramified correction. In the u-direction the quaternion sum differs from ¼(Λ_1+Λ_ω+Λ_ω̄−Λ_ψ3) by the **prime-power/ramified correction Δ_u** (the k ≥ 2 terms, the ε_k signs and p = 163), which is computed on the prime side only: |Δ_u| = 4.8·10⁻⁴ (T=1), 4.9·10⁻⁷ (1.5), 2.4·10⁻¹¹ (2) and ≤ 10⁻¹⁸ for T ≥ 3. With the shifted windows near log 163 = 5.09 (x₀ = 5) it is 0.405 (the windows pick out p = 163 and p², k ≥ 2 terms). So the zero side explains the **linear** (k = 1) L-function part of the u-direction; the correction Δ_u is not an L-function sum [project].

**Error budget [heuristic].** zeros above 110 estimated from the zero-density main term (1/2π)(log N + d log(t/2π)) with a count margin E = 10 (same heuristic as Q8, **not** a proved explicit N(T) bound), plus the truncation of the prime sum at n ≤ 20000 via |Λ| ≤ d log n. Zero accuracy of PARI taken as exact at double level (floor 10⁻¹²).

**Double computation.** Code A (mpmath, 30 digits: arch. integrals, zero sums, quaternion prime sums) vs code B (numpy/scipy doubles, Λ from a_n): zero sums agree to ≤ 1.1·10⁻¹⁴, residuals A vs B to ≤ 8·10⁻¹⁴, quaternion prime sums A vs B (u) ≤ 2.7·10⁻¹⁴, (n) ≤ 8·10⁻¹⁵; identity "quaternion n-sum + ramified term = character-side prime sum" to 9.5·10⁻¹⁶. Second check on the zero cutoff (`cutoff_check.py`): truncating the zeros at height 40/60/80/110 changes the T = 16, 20 residuals exactly as predicted (L(ρ) real part: omitted zeros −3.88·10⁻¹ vs density estimate −3.72·10⁻¹ at T=20, H=40; −5.5·10⁻¹³ vs −6.6·10⁻¹³ at H=110). For the u-combination the smooth density **cancels** (1+2−3 = 0 in degree and 2·log 163 − log 26569 = 0 in conductor, because the combination ζ·L(ω)L(ω̄)/L(ψ3) has no net zero density), so only the unsigned estimate is a bound: at T=20, H=40 actual 8.6·10⁻³ ≤ 2.1·10⁻¹, H=60: 2.1·10⁻⁶ ≤ 1.1·10⁻³.

## 3. Residuals per T (zero side − (arch + poles + prime side), code A)

| T | u-direction (b.a = a.a) | n-direction b.a / a.a (odd, sgn·Gaussian) | control: real part, L(ρ) b.a / a.a | budget |
|---|---|---|---|---|
| 1 | 1.0·10⁻¹⁰ | −3.3·10⁻¹² / 9.9·10⁻¹² | −2.8·10⁻¹² / 3.5·10⁻¹² | 6·10⁻⁹ (prime truncation) |
| 1.5 | 3.1·10⁻¹⁸ | −2.5·10⁻¹⁷ / 5.6·10⁻¹⁷ | −1.3·10⁻¹⁶ / 4.2·10⁻¹⁸ | ≤ 10⁻¹² floor |
| 2 | −1.5·10⁻¹⁷ | −9.0·10⁻¹⁷ / 3.4·10⁻¹⁷ | −1.2·10⁻¹⁶ / −5.3·10⁻¹⁷ | ″ |
| 3 | −5.1·10⁻¹⁷ | −1.5·10⁻¹⁶ / 8.5·10⁻¹⁷ | −2.1·10⁻¹⁶ / −5.0·10⁻¹⁷ | ″ |
| 5 | 1.4·10⁻¹⁷ | −1.4·10⁻¹⁶ / 2.7·10⁻¹⁶ | −2.5·10⁻¹⁶ / 2.2·10⁻¹⁷ | ″ |
| 8 | 9.4·10⁻¹⁷ | −2.3·10⁻¹⁶ / 1.9·10⁻¹⁶ | −3.4·10⁻¹⁶ / 3.8·10⁻¹⁷ | ″ |
| 12 | 8.5·10⁻¹⁷ | (prime side < 10⁻¹⁹; see S family) | −5.6·10⁻¹⁶ / −9.8·10⁻¹⁷ | ″ |
| 16 | 1.1·10⁻¹⁶ | — | −5.3·10⁻¹⁶ / −3.6·10⁻¹⁶ | ≈ 8·10⁻²⁰–10⁻¹² |
| 20 | −6.8·10⁻¹⁵ | — | −5.8·10⁻¹³ / −6.1·10⁻¹³ | 2·10⁻¹² |

Every residual is ≤ its budget (the budget at T ≤ 5 is the 10⁻¹² floor, so those rows are at rounding level, not a sharper claim). The values of the prime sides themselves are O(10⁻¹)–O(1) at T ≤ 3 and tiny for T ≥ 5 (the u-direction is −0.4218 at T=1, −0.1665 at 1.5, −0.0623 at 2, −3.6·10⁻³ at 3; the n-direction −0.08209 / +0.05043 / +0.17528 at T = 1/2/3 for b.a, identical to the earlier table). So at large T the identity is a statement about the zero side being balanced by the archimedean term alone. The shifted families, where the prime side is O(1–10): u-direction residual ≤ 4.6·10⁻¹⁴ (cos-shifted, T = 8, x₀ = 1…8.5, both fields), n-direction (sine-shifted) residual ≤ 6.3·10⁻¹⁵ (b.a) and ≤ 9.9·10⁻¹⁵ (a.a) while the n-sums are of size 0.24–4.0 (e.g. a.a x₀ = 5: −3.993172); L(ρ) control ≤ 1.8·10⁻¹⁴.

## 4. Observations

1. **[project, numerics] The n-direction is an asymmetry.** For b.a with sgn·Gaussian weight: T=1: upper zeros +0.9397, lower zeros −1.0218, sum −0.0821 (the n-direction); T=3: +3.0050 and −2.8297, sum 0.1753; T=8: +5.8395 and −5.8387, sum 0.00088. It is a small difference of two large sums, fed only by the **non-symmetry** of the zeros of L(ρ⊗ω) in t ↔ −t (155 vs 154 zeros at height 110 for b.a). A self-dual L-function (such as L(ρ), the reduced trace) has symmetric zeros, so the same odd test function gives exactly 0 for it. The even Gaussian h, as used for the real part, sees only Re Λ_{ρ⊗ω} and therefore **cannot** see the n-direction at all.
2. **[project, numerics] The central zero of L(ρ) contributes nothing to the pure part.** (i) Coefficient a_ρ = 0 exactly (character inner products, both fields). (ii) The pure part was rebuilt from zeros of ρ⊗ω, ρ⊗ω̄, ζ, ω, ω̄, ψ3 only, with no ρ zeros, and balances to ≤ 10⁻¹¹. (iii) The central zero is essential for the real part: its share of the L(ρ) zero side is 0.488 (T=1) → 0.0167 (T=20) for b.a, 0.950 → 0.0170 for a.a; dropping it breaks the real-part identity by exactly 1.00 at every T.
3. **[project, numerics] u-direction is shared out among four L-functions.** For b.a, T=1, ¼(zeros − arch − pole) = ζ −0.3963, L(ω)+L(ω̄) +0.0274, −¼L(ψ3) −0.0534; ζ dominates, with a signed sum −0.4223 (+0.0005 from Δ_u). The pole of ζ is needed (terms h(±i/2) = 2e^{1/(4T²)}). The u-direction is identical for both fields (same A4 data: ζ, L(ω), L(ψ3) agree since they belong to the shared quartic subfield). The n-direction differs between the fields because the twists ρ⊗ω have different conductors (163 vs 26569).
4. **[project] Reduced trace vs pure part.** The reduced trace sees L(ρ) only, a self-dual function with symmetric zeros and a forced central zero. The extra quaternion direction is carried by a complex-character L-function whose zeros are not symmetric, and by ζ and cubic/quadratic-type L-functions of proper subquotients; this is direction-dependent (the choice of g₆, u = i fixes it) and flips with the convention [project; Theorem 2 of `tetrahedral-frobenius-20261112.md`].

## 5. What this does and does not show

- It **verifies, for these two fields, a known identity** (the explicit formula) in a new combination; it introduces no new theorem about zeros. The new items are the odd-test-function reading of the n-direction, and the numerical bookkeeping (coefficients ∓i/2 combine to a real number H_π).
- Zeros are PARI floats, uncertified; the tail margin E = 10 is heuristic; a completeness check of the zero list is only indirect (a missing zero would break the residual at some T). Float numerics are **not** interval certificates.
- Δ_u (k ≥ 2 and ramified terms) is computed from the primes and is not covered by a zero-side statement.
- Untested: 2O, 2I (the same method applies; character tables with √2 / √5), higher T for the odd family, heights > 110.

## 6. Files

`tetra-zero1113/zeros_export.py` (PARI zeros; run with `/workspace/tools/pari-venv/bin/python3 zeros_export.py b.a 110`), `run.py` (A/B explicit formula, `python3 run.py b.a`, ~13 min each), `cutoff_check.py`, `coeff_check.py`, `fig.py`; data `zeros-*.json`, `results-*.json`, `coeffs-*.json`, `cutoff-check-*.json`, logs `run-*.log`; figure `figures/tetra-pure-zero-side.png`/`.svg`.

## 7. Page suggestion

Solid enough for a short subsection ("Where the extra quaternion direction lives"): coefficient table, the plain-words statement of 1., panels (b) and (c) of the figure, with the caveat "PARI zeros, not certified". Folded onto the live page 13 Nov 2026 as #binary-tetrahedral-pure-zero-side (site figure `figures/tetra-pure-zero-side-site{,-mobile}.svg` from `tetra-zero1113/site_figure.py`).
