# 2I aligned product: the 1/24 central exponent (Sym³ρ has a forced zero)

Research sprint, 13 November 2026, ~4:25–5:05 PM PST. Follows `octahedral-exponent-20261113.md` (the 2O 5/24 identity), `icosa-central-20261113.md` (2I root numbers and central values of ρ, ρ′) and `polyhedral-fields-20261112.md` (the galpol 2I field). Scripts and outputs: `icosa-exp1113/`.

Labels: [known] established, with source; [project] computed or proved here; [speculative] unproved. **No RH claim.** This writes the aligned Model A product of one 2I field as a product of Artin L-functions times a correction factor, and reads off its branch at s = ½. It finds, moves or excludes no zero off the line, and it is not evidence for or against the Riemann Hypothesis.

**Headline.** The expected answer was “−1/8, from the pole of ζ(2s) alone”, because both faithful 2-dimensional L-functions have root number +1 here. That prediction is **wrong**. The faithful 4-dimensional representation 4′ = Sym³ρ has root number **−1**, so L(4′, s) has a zero of odd order at ½. It appears to the power 1/6, and the exponent becomes 1/6 − 1/8 = **1/24**, provided that zero is simple (L′(4′, ½) ≈ 14.974 in floating point).

## 1. The object and the choice of axis

**[project]** K = PARI galpol real Galois field of degree 120, Gal(K/ℚ) = 2I = SL(2,5), ramified at 2, 11 and 23 only. Frobenius quaternions are known at all 33 857 unramified p < 400 000 (`icosa1113/`, `poly1112/`). An earlier note said 33 860; that is π(400 000), and the three ramified primes are not included. This sprint extended the table to all unramified p ≤ 5 499 979 (380 800 primes, `icosa-exp1113/frob/`, same F_q method plus the p-maximal-order method at 11 index divisors).

Conjugacy classes, labelled by the angle θ with eigenvalues e^{±iθ}:

| class | size | order | θ | χ_ρ = 2cos θ | in the 10-fold slice? |
|---|---:|---:|---|---|---|
| C1 | 1 | 1 | 0 | 2 | yes, ψ = 1 |
| C−1 | 1 | 2 | π | −2 | yes, ψ = −1 |
| C4 | 30 | 4 | π/2 | 0 | no |
| C3 | 20 | 3 | 2π/3 | −1 | no |
| C6 | 20 | 6 | π/3 | 1 | no |
| C10a | 12 | 10 | π/5 | (1+√5)/2 | yes, ψ = e^{iπ/5} |
| C5a | 12 | 5 | 2π/5 | (√5−1)/2 | yes, ψ = e^{2iπ/5} |
| C10b | 12 | 10 | 3π/5 | (1−√5)/2 | yes, ψ = e^{3iπ/5} |
| C5b | 12 | 5 | 4π/5 | −(1+√5)/2 | yes, ψ = e^{4iπ/5} |

Irreducible characters (all real): 1, 3 = Sym²ρ, 3′ = Sym²ρ′, 4 = ρ⊗ρ′, 5 = Sym⁴ρ (these five factor through A₅), and the faithful ρ, ρ′ (Galois conjugate under √5 → −√5), 4′ = Sym³ρ, 6 = Sym⁵ρ. Exact orthogonality is checked in ℚ(ζ₆₀) (`derive_identity.py`).

**Why the 10-fold axis.** The maximal cyclic subgroups of 2I have orders 10, 6 and 4. The aligned products already on the page use the largest one: C₆ ⊂ 2T and C₈ ⊂ 2O. A C₁₀ meets six of the nine classes, 50 of the 120 elements. It contains −1, so the slice is closed under g ↦ −g. That symmetry is what splits the exponents cleanly into real ones on the faithful characters and imaginary ones on the A₅ characters (§2). The 5-fold subgroup C₅ = ⟨e^{2iπ/5}⟩ meets only C1, C5a and C5b (25 elements) and does not contain −1. It would be a strictly coarser, less symmetric slice, so it is not used.

**Definition.** F₁₀(s) = ∏_{p unramified, Frob_p in the slice} (1 − ψ_p p^{−s})^{−1}, where ψ_p is the slice representative with Im ≥ 0 (table). Classes C3, C4, C6 are omitted; they are absorbed into H.

## 2. The exact identity

**[project, proved — exact computer algebra].** For every unramified class, the following holds as an identity of formal Euler factors (all power-sum coefficients, every k ≥ 1). Here s₁ = sin(π/5) and s₂ = sin(2π/5).

    F₁₀,p = L_p(ρ,s)^{1/3} L_p(ρ′,s)^{−1/6} L_p(4′,s)^{1/6}
            · ζ_p(s)^{i(s₁+s₂)/5} L_p(3,s)^{i(s₂−s₁)/5} L_p(3′,s)^{2i s₁/5} L_p(4,s)^{−i(s₁+s₂)/5}
            · ζ_p(2s)^{1/8 − i(s₁+s₂)/10} L_p(3,2s)^{−1/8 − i s₁/5} L_p(3′,2s)^{−1/8 − i(s₂−s₁)/10}
              L_p(5,2s)^{1/8} L_p(4,2s)^{i(s₁+s₂)/10}
              L_p(ρ,2s)^{i s₂/5} L_p(ρ′,2s)^{−i(s₁+s₂)/10} L_p(4′,2s)^{i(s₂−s₁)/10} L_p(6,2s)^{−i(s₂−s₁)/10}
            · H_p.

(L(5, s) and L(6, s) have exponent 0 at s.) The k = 1 layer is the expansion of the class function "ψ on the slice, 0 off it" in characters, so it is unique; likewise the k = 2 layer. **Parity:** under g ↦ −g the real part of ψ is odd and the imaginary part is even. So at s, the real exponents sit only on the faithful characters (ρ: 1/3, ρ′: −1/6, 4′: 1/6, 6: 0) and the imaginary ones only on the A₅ characters. The real k = 1 exponents are the same three numbers as for 2O (1/3, −1/6, 1/6), and ζ(2s) again gets real weight 1/8.

**H.** Write log H_p(x) = Σ_{k≥3} r_k x^k/k with x = p^{−s}. On a class of order m, r_k is periodic in k with period 2m. This is checked exactly for k ≤ 120, and it holds for all k because every term in the definition of r_k is periodic. Hence d log H_p/dx is a rational function with denominator dividing 1 − x^{2m}:

- C1, C−1: H = 1
- C4: H = (1 − x⁴)^{1/4}; C3: H = (1 − x³)^{1/3}; C6: H = (1 + x³)^{1/3}
- C5a, C5b: d log H/dx = (polynomial)/(x¹⁰ − 1); C10a, C10b: (polynomial)/(x²⁰ − 1). The explicit forms over ℚ(√(5±√5)) are in `derive-identity-results.json`.

All poles lie on |x| = 1 (roots of unity), and |r_k| is bounded, so each H_p is holomorphic and nonzero on |x| < 1. Globally, Σ_p |log H_p(p^{−s})| ≤ C Σ_p p^{−3σ} < ∞ for σ > 1/3, so **H is holomorphic and zero-free on Re s > 1/3**, which contains ½.

**Correction to the 2O wording.** The 2O note and page say "holomorphic and zero-free for Re s > 0". The argument only gives **Re s > 1/3**, and here the half-plane cannot be extended for free. The class average of r₃ is −0.2351 i ≠ 0, so log H contains (−0.0784 i)·log ζ(3s) plus terms holomorphic near s = 1/3, and H has a logarithmic branch point at s = 1/3. The average of r₄ is −¼ + 0.073 i, which gives another singularity at s = ¼. The analogous statement for 2O (and 2T) should be read as Re s > 1/3. None of this affects s = ½.

**Verified numerically** (`check_product.py`, H evaluated independently from the closed-form rational d log H by Gauss–Legendre quadrature):
- local factors: max |F_p − RHS_p| over the nine classes is 1.4·10⁻¹⁶ at x = 0.1, 3.1·10⁻¹⁶ at x = 0.5, and 6.8·10⁻¹⁵ at x = 0.9;
- global products over the 33 857 unramified p < 400 000: relative error 8.0·10⁻¹⁵ (σ = 2), 6.0·10⁻¹⁴ (σ = 1.5), 3.1·10⁻¹⁵ (σ = 1.2). Because both sides are products of the same local factors, this is a consistency check of the closed forms, not an independent theorem.

## 3. Root numbers: W(4′) = −1

**[project, proof from standard local-constant facts; checked numerically in §4].** The decomposition groups come from PARI `idealprimedec` on a p-maximal order (`decomp.py`):

| p | e | f | D_p ⊂ 2I | inertia |
|---|---:|---:|---|---|
| 2 | 8 | 1 | Q₈ | Q₈ (wild) |
| 11 | 6 | 2 | Dic₃ (order 12) | C₆ (tame) |
| 23 | 3 | 4 | Dic₃ | C₃ (tame) |

(At 11 and 23, the decomposition group cannot be abelian: 6 ∤ 10 and 3 ∤ 22.)

- **∞:** K is totally real, so every representation is even; W_∞ = 1.
- **2:** 4′|_{Q₈} = 2·ρ_{Q₈}, so W₂(4′) = W₂(ρ_{Q₈})² = 1, since a self-dual ρ_{Q₈} has W = ±1. Conductor: every nontrivial ramification subgroup of 2I contains −1, which acts as −1 on any faithful representation. So the 2-exponent of a faithful V is dim V × (f₂(ρ)/2) = 4·4 = **16**.
- **11:** ρ|_{D} = ρ_a = Ind_{C₆}^{Dic₃} χ, where χ is a tame character of ℚ_{121}^× of order 6, trivial on 𝔽₁₁^×. det ρ = 1 forces χ|_{ℚ₁₁^×} = the unramified quadratic character, so χ(11) = −1. Tate's formula gives W₁₁(ρ) = χ(11)·g(χ̄)/11. The Gauss sum over 𝔽₁₂₁ is +11 (computed in `gauss_tame.py`; this matches the classical evaluation for characters trivial on the prime field, Berndt–Evans–Williams, *Gauss and Jacobi Sums*, ch. 11). So W₁₁(ρ) = −1. Next, 4′|_D = ρ_a ⊕ λ ⊕ λ^{−1}, with λ(j) = i and λ restricted to inertia the quadratic character. Then W(λ)W(λ^{−1}) = λ(−1) = (−1/11) = −1, so W₁₁(4′) = (−1)(−1) = +1. Conductor exponent 4 (no inertia invariants).
- **23:** in the same way W₂₃(ρ) = χ(23)·g(χ̄)/23 = (−1)(+23)/23 = −1. Here 4′|_D = ρ_a ⊕ λ ⊕ λ^{−1} with λ unramified, so λ(−1) = 1 and W₂₃(4′) = −1. Inertia invariants of 4′ are 2-dimensional and Frobenius acts on them by ±i. So the conductor exponent is 2, and the **Euler factor at 23 is (1 + 23^{−2s})^{−1}**.

So **W(4′) = 1·1·(+1)·(−1) = −1** and N(4′) = 2¹⁶·11⁴·23² = 507 582 152 704. Also W(ρ) = W(ρ′) = W₂(ρ_{Q₈})·(−1)·(−1) = W₂(ρ_{Q₈}). This agrees with the measured +1 for both (it is the same local factor at 2 for ρ and ρ′), but it does not prove it.

**[known] input:** Artin L-functions are meromorphic and satisfy Λ(s, V) = W(V) Λ(1 − s, V^∨) (Brauer induction, Brauer 1947, Ann. of Math. 48; Hecke), and W(V) is the product of the local constants (Langlands–Deligne). Since 4′ is self-dual with W = −1, **ord_{s=½} L(4′, s) is odd**. This holds unconditionally, but it does not exclude ord = −1 (a pole), because holomorphy of icosahedral Artin L-functions is not known in general.

## 4. L(4′) numerically

**[project, float numerics].** PARI `lfun` overflowed a 6.5 GB stack at this conductor. Instead `lambda_sym3.py` sums the theta/Mellin series directly in double precision: Λ(s) = Σ a_n [I_s(n/√N) + W I_{1−s}(n/√N)], with the Γ_ℝ(s)⁴ kernel from PARI `gammamellininv`, using 5.5·10⁶ coefficients (kernel < 10⁻¹⁵ beyond).

- Theta relation θ(1/y) = W y θ(y): ratio −0.99999999719, −0.99999999857, −0.99999999933 at y = 1.05, 1.1, 1.2, so **W = −1**, as in §3.
- L(4′, 2): the formula with W = −1 gives 1.18060502037, against the Dirichlet series 1.18060502036 and the Euler product over p ≤ 5.5·10⁶ 1.18060502039. W = +1 gives 1.18060599, which is off by 10⁻⁶.
- **L′(4′, ½) ≈ 14.9739** (and L(½+ε)/ε → 14.9735 at ε = 10⁻⁵). If Λ(4′) is entire, the zero is **simple** (floating point, not an interval enclosure).

These numbers presuppose that Λ(4′) is entire with the stated functional equation (Artin's conjecture for 4′). The 10⁻⁹ agreement of the theta relation also confirms the conductor 2¹⁶·11⁴·23² and the Euler factor at 23.

## 5. Interval certificates for L(ρ, ½), L(ρ′, ½)

**[project, computer-assisted, interval arithmetic — conditional on Artin's conjecture for ρ, ρ′]** `certify_icosa.py` adapts `octa1113/certify_octa.py`: exact termwise Mellin integrals with `acb.integral`, 160 bits, c·n ≤ 120, M = 77 312 terms, coefficients exact in ℤ[(1+√5)/2], 122 s.

- **L(ρ, ½) ∈ 5.851618798 ± 1.2·10⁻¹⁰** (PARI float 5.85161879797644)
- **L(ρ′, ½) ∈ 0.5132584851 ± 8.8·10⁻¹¹** (PARI float 0.513258485079238)
- θ(½)/(2θ(2)) = 1 ± 5.4·10⁻¹¹ (ρ) and 1 ± 5.8·10⁻¹⁰ (ρ′), consistent with W = +1 and excluding −1. This checks the theta relation at one point. It does not prove it.

**Hypothesis:** these enclosures use Λ(ρ) = 4Σ a_n G₀(c n), which holds if Λ(ρ, s) is entire with Λ(s) = Λ(1 − s). For even icosahedral ρ this is Artin's conjecture, which is **not known in general** [known: Booker, "Artin's conjecture, Turing's method and the Riemann hypothesis", Experiment. Math. 15 (2006), arXiv:math/0507502, which treats even icosahedral holomorphy as open]. For 2O, by contrast, Langlands–Tunnell applies. If ρ is automorphic, Kim–Shahidi (Ann. of Math. 155, 2002) make Sym³ρ automorphic on GL(4), and then L(4′) is entire too.

## 6. The central exponent

Along the real axis as s → ½⁺, the modulus of each factor behaves as follows:

- ζ(2s)^{1/8 − i(...)}: ζ(2σ) > 0 with a simple pole, contributing (s − ½)^{−1/8} [known].
- L(3, 2s), L(3′, 2s), L(5, 2s) with real exponents ∓1/8: at s = ½ these are L(χ, 1) for nontrivial irreducible χ. These values are finite and nonzero for every Artin L-function [known: Brauer induction writes L(χ, s) as a product of integer powers of Hecke L-functions, which are holomorphic and nonzero on Re s = 1 apart from the pole of ζ_F; the poles cancel because ⟨χ, 1⟩ = 0]. No automorphy is needed for this.
- All other 2s-factors and all A₅ factors at s have purely imaginary exponents. Every character of 2I is real, so L(χ, σ) is real for real σ > ½ and |L^{ic}| = e^{−c·arg L} is locally constant. These factors do not change the size. (If one of L(3, ½), L(3′, ½), L(4, ½) vanished, the complex exponent would gain an imaginary part. Those orthogonal L-functions have W = +1 [known: Fröhlich–Queyrut 1973], hence even order. Their values were not computed.)
- H(½) ∈ ℂ^× (§2).

**Theorem [project; exact identity + the inputs listed].** Let m_ρ, m_ρ′, m_{4′} be the orders of L(·, s) at s = ½ (integers, since Artin L-functions are meromorphic). Then, along the real axis,

    |F₁₀(σ)| ≍ (σ − ½)^{e},   e = −1/8 + m_ρ/3 − m_ρ′/6 + m_{4′}/6,

and m_{4′} is **odd** (W(4′) = −1, §3). If moreover Λ(ρ), Λ(ρ′) and Λ(4′) are entire (Artin's conjecture for these three; automorphy of ρ would give 4′ = Sym³ρ by Kim–Shahidi, while ρ′ is assumed separately), then m_ρ = m_ρ′ = 0 by the interval certificates of §5. If in addition L′(4′, ½) ≠ 0 (float value 14.974), then

    **e = 1/6 − 1/8 = 1/24**,

and F₁₀(s) ∼ c (s − ½)^{1/24 + i(s₁+s₂)/10} with c ≠ 0, ∞, provided the A₅ values L(3, ½), L(3′, ½), L(4, ½) are nonzero (not computed). The imaginary part (s₁ + s₂)/10 ≈ 0.1539 comes from ζ(2s). Without simplicity, e ∈ {1/24, 3/8, 17/24, …} (m_{4′} = 1, 3, 5, …).

Comparison: 2T 3/8 = ½ − ⅛ (simple zero of L(ρ) at weight ½); 2O 5/24 = ⅓ − ⅛ (simple zero of L(ρ) at weight ⅓); 2I 1/24 = ⅙ − ⅛ (here L(ρ) does not vanish, and the forced zero sits on the 4-dimensional Sym³ρ instead).

## 7. Numerical slope

**[project, float numerics].** Using the closed form with PARI `lfun` for L(ρ), L(ρ′) (400 000 coefficients, W = +1), `lambda_sym3.py` for L(4′) (W = −1), and the exact ζ(2s), the slopes of log |F₁₀| against log ε at s = ½ + ε are below. The A₅ factors at 2s, the phases and H are finite and nonzero at ½ and are dropped; they change the slope only by O(ε).

| interval | slope | residual vs 1/24 |
|---|---:|---:|
| 10⁻¹ → 5·10⁻² | −0.033735 | −0.0754 |
| 10⁻² → 5·10⁻³ | 0.035683 | −0.0060 |
| 10⁻³ → 5·10⁻⁴ | 0.041087 | −0.00058 |
| 5·10⁻⁴ → 10⁻⁴ | 0.041468 | −0.00020 |
| 10⁻⁴ → 5·10⁻⁵ | 0.041609 | −0.000058 |
| 5·10⁻⁵ → 10⁻⁵ | **0.041647** | −0.000020 |

Without L(4′), the same partial product has last-interval slope −0.1250086, against −1/8 (`slope-partial-results.json`). This is what one would have predicted from root numbers +1 alone, and it misses the forced zero of Sym³ρ.

## 8. Status summary

- [project, proved — exact computer algebra] Local identity for F₁₀; H holomorphic and zero-free on Re s > 1/3 (not Re s > 0: branch point at 1/3).
- [project, proof from standard local-constant facts, confirmed numerically to 10⁻⁹] W(Sym³ρ) = −1, conductor 2¹⁶·11⁴·23², Euler factor (1 + 23^{−2s})^{−1}. Hence L(Sym³ρ, s) has odd order at ½ (unconditional, given Brauer's functional equation).
- [project, interval arithmetic, conditional on Artin's conjecture for ρ, ρ′] L(ρ, ½) ≈ 5.8516 and L(ρ′, ½) ≈ 0.51326 are nonzero.
- [project, float numerics, assumes Λ(Sym³ρ) entire] L′(Sym³ρ, ½) ≈ 14.974, so the zero is simple.
- [project] Exponent **1/24** under those hypotheses. Measured slope 0.041647.
- [known] L(χ, 1) ≠ 0, ∞ for nontrivial irreducible Artin χ (Brauer); Artin conjecture open for even icosahedral (Booker 2006); Fröhlich–Queyrut for orthogonal root numbers.
- Not claimed: anything about zeros off s = ½, any RH/GRH statement, or holomorphy of any icosahedral Artin L-function.

## 9. Open

- Interval certificate for L′(Sym³ρ, ½) ≠ 0 (conductor 5.1·10¹¹, about 5·10⁶ terms; the Γ_ℝ⁴ kernel needs a rigorous Meijer-G enclosure).
- An unconditional route to ord_{½} L(ρ) = 0 without Artin's conjecture (for example a Booker-style finite-height holomorphy check near ½).
- The A₅ values L(3, ½), L(3′, ½), L(4, ½) (which fix the imaginary part of the exponent), for example via ζ of the quintic and sextic subfields.
- W₂(ρ_{Q₈}) from 2-adic local constants (it equals the measured W(ρ) = +1).
- 2O/2T: replace "Re s > 0" by "Re s > 1/3" in the H statement (page and notes) and decide whether H continues past 1/3 there.

## 10. Files (`icosa-exp1113/`)

- `derive_identity.py`, `derive-identity-results.json`: exact ℚ(ζ₆₀) derivation, exponents, r_k, d log H, root moduli
- `check_product.py`, `check-product-results.json`: local and global checks
- `decomp.py`, `decomp-results.json`, `decomp.log`: e, f at 2, 11, 23
- `gauss_tame.py`, `gauss-tame-results.json`: tame Gauss sums / W₁₁(ρ), W₂₃(ρ)
- `frob_extend.py`, `frob_nonsqfree.py`, `frob/`: Frobenius classes 400 000 < p ≤ 5 499 979. A run to 11·10⁶ was also started and was unfinished or unaudited at writing (its non-squarefree primes were not processed); it is not used.
- `lambda_sym3.py`, `lambda-sym3-results.json`: L(Sym³ρ) by direct theta sums; `lfun_sym3.py` (PARI attempt, stack overflow, kept for the record)
- `certify_icosa.py`, `certify-icosa-results-cn120-p160.json`, `certify-cn120.log`: Arb enclosures
- `slope_closed.py`, `slope-partial-results.json`, `slope_full.py`, `slope-results.json`: slopes
- `site_figure.py` → `figures/icosahedral-exponent-site.svg`, `-mobile.svg`
