# Zero side of the 2I Model A quaternion prime sum — 15 Nov 2026, ~12:45–1:25 PM PST (routine sprint)

Closes the open item "zero side for 2I (numerics)" from `open-questions.md`, the analogue of `tetra-pure-zero-side-20261113.md` (2T) and `octa-pure-zero-side-20261113.md` (2O). Field: PARI galpol real Galois field of degree 120, Gal = SL(2,5) = 2I, ramified at 2, 11, 23 (`polyhedral-fields-20261112.md`, `poly1112/`, Frobenius classes `icosa1113/`, `icosa-exp1113/frob/`). Code and data: `icosa-zero1115/`. Figure: `figures/icosa-pure-zero-side.png` (+ `.svg`; plain matplotlib, same layout family as the tetra and octa figures, 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 **uncertified floats**. Nothing here says anything about the Riemann Hypothesis.

## 1. In plain words

A Frobenius quaternion has a real part and a pure (i, j, k) part. As for 2T and 2O, the **real part** is exactly half the prime side of L(ρ), at every prime power (Re g = ½ tr ρ(g)). In this field L(ρ) has root number +1 and no forced central zero, so there is no central-zero share to talk about (unlike 2T and 2O, where the forced zero feeds the real part).

The **pure part** has three directions here, one per axis type of the icosahedral group: the 5-fold axis v (6 of the 9 classes: 1, −1 and the four of order 5 and 10), the 3-fold axis n (orders 3 and 6) and the 2-fold axis m (order 4). In Model A (k = 1) each is an ordinary, real, signed combination of the **A₅ L-functions** ζ, L(3), L(3′), L(4), L(5) only. The faithful L-functions L(ρ), L(ρ′), L(4′), L(6) get weight **exactly 0** at k = 1 and come back only at the prime-square layer (k = 2). All characters of 2I are real, so every L-function is its own mirror image and there is **no odd (up/down) channel**: the 2T mechanism cannot occur.

What is different from 2O: the 5-fold axis needs **both** Galois-conjugate L-functions L(3) and L(3′) with √5-irrational weights, and, remarkably, the ramified Euler factors (at 2, 11, 23) cancel out of the 5-fold direction completely.

## 2. Exact decomposition [project, proved — exact sympy computation in ℚ(√5); `icosa-zero1115/coeffs.py`, `coeffs.json`]

**Classes and characters.** θ ∈ [0, π] is the half-angle (eigenvalues e^{±iθ}); class sizes: C1 1 (θ = 0), C10a 12 (π/5), C5a 12 (2π/5), C6 20 (π/3), C4 30 (π/2), C10b 12 (3π/5), C3 20 (2π/3), C5b 12 (4π/5), C−1 1 (π). Irreducible characters (all real, field ℚ(√5)): 1, ρ, ρ′ (√5 → −√5), 3 = Sym²ρ, 3′ = Sym²ρ′, 4 = ρ⊗ρ′, 4′ = Sym³ρ, 5 = Sym⁴ρ, 6 = Sym⁵ρ (A₅ ones: 1, 3, 3′, 4, 5). The 9 × 9 orthogonality table was verified exactly in ℚ(√5) (cf. `icosa-exponent-20261113.md` §1).

**Model A representatives.** One fixed unit quaternion per class, g = cos θ + sin θ·a (sin θ ≥ 0), on the axis a of its class: v (5-fold) for C1, C−1, C10a, C5a, C10b, C5b; n (3-fold) for C6, C3; m (2-fold) for C4. The k-th layer of the prime sum is cos kθ (real) and sin kθ (along the axis). Real part and each axis component are class functions of Frob_p for every k. (The three axes are not orthogonal; "direction" means one of the scalar sums S_v, S_n, S_m, as in 2T/2O.) Write s₁ = sin(π/5), s₂ = sin(2π/5).

**k = 1 layer (exact character inner products):**

| | ζ | L(3) | L(3′) | L(4) | L(5) | L(ρ) | faithful ρ′, 4′, 6 |
|---|---|---|---|---|---|---|---|
| real part | 0 | 0 | 0 | 0 | 0 | **½** | 0 |
| v (5-fold) | (s₁+s₂)/5 = 0.3078 | (s₂−s₁)/5 = 0.0727 | 2s₁/5 = 0.2351 | −(s₁+s₂)/5 | 0 | **0** | **0** |
| n (3-fold) | √3/6 | 0 | 0 | √3/6 | −√3/6 | **0** | **0** |
| m (2-fold) | ¼ | −¼ | −¼ | 0 | ¼ | **0** | **0** |

Cross-check: the v row equals the imaginary parts of the k = 1 exponents of the aligned F₁₀ identity in `icosa-exponent-20261113.md` §2 (i(s₁+s₂)/5 on ζ, i(s₂−s₁)/5 on L(3), 2i s₁/5 on L(3′), −i(s₁+s₂)/5 on L(4)), as it must be: the F₁₀ phase *is* the v-component. Note s₁+s₂ = (s₂−s₁)+2s₁, so the four v-weights sum to **0** (used in §5).

**k = 2 layer:** v: (2s₂/5) ρ − ((s₁+s₂)/5) ρ′ + ((s₂−s₁)/5)(4′ − 6); n: (√3/6)(ρ + ρ′ − 4′); m: 0 (sin 2·π/2 = 0). Real part: −½ζ + ½·L(3). **k = 3:** v: ((s₂−s₁)/5)ζ + (2s₂/5)·3 − ((s₁+s₂)/5)·3′ − ((s₂−s₁)/5)·4; real −½ρ + ½·4′ (the faithful ones return in the real part at odd k ≥ 3).

**Why [project, proved].** Central parity: g ↦ −g sends the class with representative e^{θa} to the one with e^{(π−θ)a} on the same axis, so sin kθ ↦ (−1)^{k+1} sin kθ and cos kθ ↦ (−1)^k cos kθ. A class function even under g ↦ −g has zero inner product with every faithful character (χ(−g) = −χ(g)); an odd one with every A₅ character. So the pure part is A₅-only for odd k and faithful-only for even k; the real part is the reverse. Reality: all weights are real, so every combined Λ is real and no odd test function sees anything.

**The "tiny way" the extra directions change things.**
- vs **2T**: 2T has non-real characters (ω, ω̄), so its n-direction is an up/down zero asymmetry of L(ρ⊗ω) visible only to odd test functions. 2I has no such channel; its three directions are all even zero sums, as in 2O.
- vs **2O**: the 2O u and n weights live in ℚ(√2, √3) over five S₄ L-functions. Here the 3-fold and 2-fold weights are the same kind of thing (√3/6 and ±¼ on ζ, L(4), L(5) resp. ζ, L(3), L(3′), L(5)), but the **5-fold direction requires the Galois-conjugate pair L(3), L(3′)** and L(4) = L(ρ⊗ρ′) with √5-weights; L(3′) is the √5-conjugate of L(3), a different L-function.
- The 2-fold direction has no k = 2 term in any group; the 3-fold direction has a k = 2 term only through the faithful ρ, ρ′, 4′, as in 2O.

## 3. New L-function data [project, float; PARI `lfuncheckfeq`, uncertified]

Total reality of the field gives Γ-factor Γ_ℝ(s)^d for every piece. Frobenius classes of all unramified p ≤ 5.4·10⁶ (`icosa-exp1113/frob/`) give a_n; ramified Euler factors from V^I and Frobenius (explicit SU(2) model of D ≅ Dic₃ at 11 and 23: g = diag(ζ₆ or ζ₃), j acts by e₁→−e₂, e₂→e₁). At 11 and 23: L(3), L(3′): (1+p^{−s})^{−1}, exponent 2; L(4): (1−p^{−2s})^{−1}, exponent 2; L(5): (1−p^{−s})^{−1}, exponent 4. Conductor exponent at 2 and the Euler factor at 2 found by scanning lfuncheckfeq (`scan_cond.py`, `scan5.py`); the 2-part is consistent with the Artin conductor formula for inertia at 2 mapping onto V₄ ⊂ A₅ (Q₈ ⊂ 2I): with a ramification steps whose group is Q₈ and b whose group is C₄ (higher steps at most {±1}, which acts trivially on A₅ pieces), the Artin conductor formula gives f(3) = f(4) = f(5) = 3a + b (codimensions of the Q₈- and C₄-invariants are 3, 3, 3 and 2, 2, 2), and f(ρ) = 8 forces a + b/2 + c/4 = 4; the observed exponent 8 for 3 and 4 then predicts 8 for 5, which the scan confirmed. [project, argument; the image V₄ is itself inferred from dim V^I = 0, 1 for 3, 4.]

| L | degree | conductor | Euler factor at 2 | W | checkfeq |
|---|---|---|---|---|---|
| ρ | 2 | 2⁸·11²·23² = 16 386 304 | 1 | +1 | 2⁻⁵⁵ (`icosa-central-20261113.md`) |
| 3 = Sym²ρ | 3 | 2⁸·11²·23² | 1 | +1 | 2⁻⁵⁴ |
| 3′ | 3 | 2⁸·11²·23² | 1 | +1 | 2⁻⁵⁵ |
| 4 = ρ⊗ρ′ | 4 | 2⁸·11²·23² | (1−2^{−s})^{−1} | +1 | 2⁻⁵⁴ |
| 5 = Sym⁴ρ | 5 | 2⁸·11⁴·23⁴ = 1.05·10¹² | (1−2^{−s})^{−2} | +1 | 2⁻⁵⁰ |

Root number +1 and no forced central zero for any of them (lfunorderzero = 0 where computed). Wrong Euler-factor options fail with checkfeq 2⁻³ … 2⁻⁶ (L(5) options (1+x)^{−2}, (1−x²)^{−1}, (1+x²)^{−1}). **Artin's conjecture:** for these (even, icosahedral) representations it is **not known** in general [known: only meromorphy, Brauer 1947]; every check below presupposes holomorphy and is consistent with it (a pole in the range would show as an unbalanced residual).

**Zeros.** `zeros_export.py`: exact coefficients in ℤ[φ] (`an_exact.py`: integer pairs, the e_j of each class found by rounding at 60 digits and asserted to 10⁻³⁰) converted to PARI at 288 bits (`set_real_precision_bits`), then `lfunzeros` to height 70 (ζ: 110). **Lesson [project]:** with double-precision a_n, `lfunzeros` silently drops zeros above height ≈ 20 (88 instead of 183 zeros for ρ at height 60). Exact coefficients at ≥ 192 bits fix it. Counts of 0 < γ ≤ 70: L(ρ) 216, L(3) 232, L(3′) 232, L(4) 248, ζ 33 (to 110); main-term (H/2π)(log N + d·log(H/2πe)): 216.5, 232.2, 232.2, 247.9, 32.6 — within 1. Not a proved N(T) bound.

**L(5): skipped.** conductor 1.05·10¹², degree 5. lfunzeros needs about 3.1·10⁷ coefficients (PARI warning, any height and precision tried: 14–22, 72–128 bits); only 5.4·10⁶ unramified primes are classified (the 2I Frobenius table is the expensive part, see `icosa-exponent-20261113.md`). Not attempted: extending the table to 3·10⁷.

## 4. What was verified

**Explicit formula, self-dual form [known: Weil 1952; Iwaniec–Kowalski, *Analytic Number Theory*, Thm 5.12]**, exactly as in the 2T/2O notes: m₀h(0) + 2Σ_{γ>0}h(γ) = pole (ζ only) + (1/π)∫₀^∞ h(r)[log N + d·(Re ψ(¼+ir/2) − log π)]dr − 2ΣΛ_χ(n)w(log n)/√n, h even. Test functions: Gaussian h = e^{−(r/T)²}, T = 1, 1.5, 2, 3, 5, 8, 12; cos-shifted T = 8, x₀ = 1, 1.5, 2, 3, …, 8. Prime sums: unramified p^k ≤ 20 000 (ramified Euler factors included in the L-function side). Code A: mpmath 30 digits, closed-form cos kθ / sin kθ; code B: numpy/scipy doubles, explicit quaternion multiplication. Per-L residual (zero side − arch − pole − prime side):

| | per-L (ζ, ρ, 3, 3′, 4), max, code A | real part | v (5-fold) | budget |
|---|---|---|---|---|
| T = 1 | 3.9·10⁻¹⁰ (ζ; others ≤ 2.5·10⁻¹¹) | 5.5·10⁻¹² | 1.1·10⁻¹⁰ | 10⁻⁸ (prime truncation at 20 000) |
| Gaussian T = 1.5 … 8 | ≤ 6.9·10⁻¹⁶ | ≤ 1.4·10⁻¹⁶ | ≤ 1.3·10⁻¹⁶ | 10⁻¹² floor |
| Gaussian T = 12 | 1.3·10⁻¹⁴ | 6.2·10⁻¹⁵ | 8.5·10⁻¹⁷ | 3·10⁻¹⁴ (zero tail above 70) |
| shifted T = 8, x₀ = 1 … 8 | ≤ 1.1·10⁻¹⁴ | ≤ 4.7·10⁻¹⁵ | ≤ 1.2·10⁻¹⁵ | 10⁻¹² floor |

Code B per-L residuals ≤ 2.1·10⁻¹⁴ (T = 1: 3.9·10⁻¹⁰). Zero sums A vs B agree to ≤ 1.2·10⁻¹⁴; quaternion axis sums A vs B (closed form vs explicit quaternion products) ≤ 3.4·10⁻¹⁴ over all rows and all four directions. All within budget. For T ≥ 3 these are rounding level, not a sharper claim.

**Axis identity "axis sum = k=1 zero side + Δ".** The real part and the **v (5-fold) direction** are complete from ζ, L(3), L(3′), L(4) and L(ρ), so both close as above. **Δ_v = 6.8·10⁻⁷ at T = 1**, 1.0·10⁻¹² at T = 1.5, 7·10⁻¹⁸ at T = 3 (`delta_split.py`): Δ = Q₂ + Q₃₊ − R − PP with R (ramified primes 2, 11, 23 inside the L-functions) **identically 0** along v — the four weights sum to 0 and L(ζ), L(3), L(3′), L(4) have Λ(p) = (1, −1, −1, 0)·log p at p = 11, 23 and (1, 0, 0, 1)·log 2 at p = 2, so the k = 1 ramified terms cancel — and Q₂ + Q₃₊ = PP to 7·10⁻⁷ (−0.0700 each at T = 1). So **the 5-fold direction is, to this accuracy, just the k = 1 zero side**. [project; the near-equality Q₂ + Q₃₊ ≈ PP is observed, not proved.]

**n (3-fold) and m (2-fold)** need L(5), so their zero side is **not** checked (what is available: the partial shares, T = 1: n: ζ −0.458, L(4) −0.052, missing L(5); m: ζ −0.396, L(3) +0.126, L(3′) −0.037, missing L(5)). Their Δ (prime side only, exact L(5) coefficients, `delta_n3m2.py`): Δ_n = 0.0241 at T = 1 (R −2.3·10⁻⁴, PP −0.0119), 1.4·10⁻³ at T = 1.5, 7·10⁻¹¹ at T = 3; **Δ_m = 0.452 at T = 1, 0.354 at 1.5, 0.217 at T = 3, 0.100 (T = 8, x₀ = 4)**, all from R at p = 2 (m has no k ≥ 2 terms; at p = 2 ζ, L(3), L(3′), L(5) have Λ(2) = (1, 0, 0, 2)·log 2, so the weighted ramified term does not cancel). So m is dominated by the ramified prime 2, like 2O's m.

Values (Model A, T = 1): quaternion sums real −0.28140, v −0.43485, n −0.30473, m −0.01103. Zero-side shares in v (weight × (zeros − arch − pole)): ζ −0.488, L(3) −0.037, L(3′) +0.035, L(4) +0.055 (total −0.435, plus Δ_v 7·10⁻⁷). ζ with its pole is again the largest single share.

## 5. Observations

1. **[project, proved + numerics] No asymmetry, no odd channel**, in any convention; unlike 2T. All three directions are even zero sums of A₅ L-functions; faithful L(ρ), L(ρ′) enter only through prime squares (k = 2).
2. **[project, numerics] Real part:** half of L(ρ)'s zero side, with no forced central zero (W = +1), identity exact to 5.5·10⁻¹² (T = 1) and ≤ 6·10⁻¹⁵ (T ≥ 1.5).
3. **[project] The 5-fold direction cancels all ramified Euler factors** (weights sum to 0 over ζ, L(3), L(3′), L(4)); so ramification at 2, 11, 23 does not enter the v-direction at k = 1. In contrast the 2-fold direction is dominated by p = 2.
4. **[project]** The A₅ pieces L(3), L(3′), L(4) all have conductor 2⁸·11²·23², the same as L(ρ); L(5) has 2⁸·11⁴·23⁴ (the 11 and 23 exponents double). L(3) and L(3′) have the same conductor, as √5-conjugates must.

## 6. What this does and does not show

- It **verifies, for one 2I field, a known identity** (the explicit formula) in a new combination, proves exactly which L-functions carry each quaternion direction, and records new conductor / Euler-factor / root-number data (floats). No new theorem about zeros. **No RH progress is claimed or implied.**
- Zeros are PARI floats, uncertified; completeness is checked via counts vs the main term and indirectly by the residuals; tail margin E = 10 is heuristic. Artin holomorphy is presupposed (open for even icosahedral ρ).
- **Skipped:** L(5) zeros (hence the zero side of the 3-fold and 2-fold directions), heights above 70 (ζ: 110), the k = 2 layer as a zero sum, an independent proof (not just checkfeq) of the 2-part conductor and Euler factors of 3, 3′, 4, 5.

## 7. Files

`icosa-zero1115/`: `coeffs.py` → `coeffs.json` (exact weights k = 1, 2, 3); `an_build.py` (float a_n from classes), `an_exact.py` (exact ℤ[φ] a_n); `scan_cond.py`, `scan5.py` (+ `scan-*.log`: conductor / Euler-factor scans); `zeros_export.py` → `zeros-{zeta,rho,three,threep,four}.json` (logs `zx70-*.log`); `run.py` → `results-A.json`, `run-A.log` (3 min on 7 cores); `delta_split.py` → `delta-split.json`; `delta_n3m2.py` → `delta-n3m2.json`; `fig.py` → `figures/icosa-pure-zero-side.png/.svg`.
