# Field-level check of the 2O and 2I barrier exponents

Sprint run, 12 November 2026, ~9:14–10:30 PM PST. Follows `tetra-central-simplicity-20261112.md` (the 2T central zeros are simple) and the open item in `open-questions.md`: compute actual Frobenius quaternions on real number fields with Galois groups 2O and 2I, build the aligned Model A Euler product as for Q8 and 2T, and measure the boundary exponent β near ½ against the already-certified minima 0.1487 (2O) and 0.0698 (2I), together with the forced-zero branch power.

Labels: [known] established, with source; [project] computed here; [speculative] unproved. **No RH claim.** This is a numerical check of an already-proved group-theoretic exponent against actual Frobenius classes, not a new zero of any L-function, and not a proof of the slope.

## 1. The fields

LMFDB’s HTML search is behind a reCAPTCHA from this host. The Artin API, recovered for a while with a desktop User-Agent, gave 12 quaternionic 24T201 records for 2I (character field ℚ(√5), projective image A5, indicator −1; smallest conductor 1 965 604 = 2²·701²). It did not return a 2O record in the first 200 two-dimensional, indicator −1 representations (those were Q8, 2T, a few dicyclic, and the 12 icosahedral ones); pagination past offset 200 was recaptcha’d again. Coefficient tables were not fetched. The measurement therefore uses PARI’s galpol polynomials, as the task allowed.

**2O is not GL(2,3).** The binary octahedral group is GAP4(48,28) = C₂ . S₄, the minus-type double cover of S₄, which embeds in SU(2) ⊂ H^×. GL(2,3) is GAP4(48,29), the plus-type cover; it does not embed in SU(2). LMFDB’s alias “GL(2,3) = 8T23 / 16T66 / 24T22” is the wrong group for quaternionic Frobenius. The faithful permutation degree of 2O is 16 (cosets of a C₃ that does not contain −1), not 8 (any index-8 subgroup contains the centre, and the action factors through S₄). Among the degree-16 order-48 groups with centre of order 2, 16T65 is the 2O candidate; 16T66 is GL(2,3).

### 2O, from galpol

**[project]** PARI galpol-4 real polynomial of degree 48, identified by `galoisidentify` as [48,28]. The field is totally real (signature [48,0]). The absolute discriminant is 2¹³⁶ · 5³⁶ · 17³², so the only ramified primes are 2, 5 and 17. Two faithful two-dimensional Artin characters (columns 4 and 5 of the PARI character table) have conductor 924 800 = 2⁷·5²·17² and Γ-type Vga = [0,0]:

- character 5 (call it ρ): root number −1, `lfunorderzero` = 1 (a simple forced zero at s = ½);
- character 4 (ρ ⊗ ε): root number +1, L(½) ≈ 4.73498, order 0.

The sign character ε has conductor 5: the field contains the unexpected quadratic subfield ℚ(√5). Script `poly1112/measure_2O.py`.

### 2I, from galpol

**[project]** PARI galpol real polynomial of degree 120 (we work with the `polredbest` model), Galois group SL(2,5) = GAP4(120,5). The field is totally real (signature [120,0]). The absolute discriminant is 2³⁶⁰ · 11¹⁰⁰ · 23⁸⁰, so the only ramified primes are 2, 11 and 23. `galoisinit` fails on this polynomial (the group is not weakly supersolvable; PARI returns 0 / “incorrect type in checkgal”). All 120 automorphisms were obtained from `nfgaloisconj` and cached. An explicit group isomorphism Gal(K/ℚ) → 2I ⊂ H^×, built from a generating pair and the composition table of the automorphisms, separates the two order-5 classes and the two order-10 classes, which share a cycle type in the degree-24 permutation representation and cannot be told apart by factoring a stem polynomial alone. The 12 LMFDB 24T201 records confirm the group, the character field ℚ(√5), and the projective image A5; they are not the field used for the Euler product. Script `poly1112/measure_2I.py`.

## 2. The aligned product and the measurement

The group-theoretic exponent β(I) = Σ_C δ_C sin²θ_C min_{g∈C} sin²∠(axis g, I) is already a theorem (Chebotarev densities, Jacquet–Shalika non-vanishing on Re s = 1, and the local expansion of Result 3d in the 2T note), and the minima

- 2O: 0.148712605374…, smallest root of 192β³ − 288β² + 126β − 13,
- 2I: 0.069765059740…, smallest real root of the recorded degree-6 polynomial,

are certified global minima (`beta1112/certify_global.py`). The field-level check is: take the actual Frobenius class of each unramified prime, choose the representative in that class closest to the certified minimiser I*, form the Model A reduced-norm Euler product, and read the slope of log nrd against log(2σ − 1) as σ → ½⁺.

The direct Euler product at σ = ½ does not converge well enough to be used. As in the 2T work, we form the completed remainder

log nrd E_A(σ+tI) + β Σ_p log(1 − p^{−2σ}) + β log ζ(2σ)

and plot that completed log-norm against log(2σ − 1). At a generic height t the predicted slope is β itself. At t = 0 the quaternionic product tracks L(ρ,σ) (up to ramified factors): if the root number is −1, L(ρ) has a forced central zero and the predicted branch power is the order of that zero (here 1, once it is confirmed to be simple), not β.

## 3. Results

### 2O

**[project, float numerics]** Unramified primes p ≤ 20 000: 2 259 classes, none unrecognised. Chebotarev counts against the eight conjugacy-class densities give χ² = 3.93 on 7 degrees of freedom. The sum of densities times local exponents, using the aligned representative of each class, recovers the predicted β to 1.4·10^{−16}.

PARI’s L-function (functional equation, not a partial Euler product) gives a clean forced zero of ρ: the slope of log |L(ρ, ½+ε)| against log ε is 0.9397, 0.9941, 0.9994, 0.99994 as ε = 10^{−1}, …, 10^{−5}, consistent with a simple zero. The twist ρ ⊗ ε has L(½) ≈ 4.73498 and slopes −0.062, −0.0059, −0.00059, −0.000058, consistent with a nonzero finite value. This is a PARI floating-point check of the already-reported `lfunorderzero`, not an Arb interval certificate.

Completed Model A slopes, last decade (ε = 10^{−3} → 10^{−4}), against the predicted 0.14871:

| height t | last-decade slope | residual vs 0.14871 |
|---|---|---|
| 0.0 | 0.15424 | +0.0055 |
| 5.0 | 0.15844 | +0.0097 |
| 11.3 | 0.15087 | +0.0022 |
| 23.5 | 0.15382 | +0.0051 |

The closest approach is at t = 11.3, residual about 0.002 at this cutoff. The slopes are monotonically decreasing toward β across the three decades in every height we tried. The residual is a truncation effect (primes only to 20 000, ε only to 10^{−4}), not a disagreement with the predicted exponent. The stored log-norms were recomputed into slopes independently of the original script; the two agree exactly, and the Chebotarev χ² and the β-sum were recomputed from the stored class list.

### 2I

**[project, float numerics]** Unramified primes p ≤ 20 000, omitting 2, 11, 23 (ramified) and the 16 other primes up to 20 000 that divide a denominator of an automorphism polynomial in the power basis of the `polredbest` generator (those primes are not q-integral in ℤ[α]): 2 242 classes, none unrecognised. All nine conjugacy classes appear. Chebotarev counts give χ² = 10.86 on 8 degrees of freedom. The two rarest classes (the identity and −1, density 1/120) are a little light (7 and 11 against 18.7); the other seven classes match their expected counts to a few percent. The sum of densities times local exponents, using the aligned representative, recovers the predicted β to 2.5·10^{-16}.

The character field of ρ is ℚ(√5), and `galoisinit` fails on the degree-120 polynomial, so this run does not evaluate L(ρ, ½) in PARI. The forced-zero branch at height 0 is therefore not given a field-level L-value here (unlike 2O). The completed Model A product at height 0 still approaches the group-theoretic β, as it did for 2O: the partial Euler product does not by itself display a central zero.

Completed Model A slopes, last decade (ε = 10^{-3} → 10^{-4}), against the predicted 0.06977:

| height t | last-decade slope | residual vs 0.06977 |
|---|---|---|
| 0.0 | 0.07082 | +0.0011 |
| 5.0 | 0.07823 | +0.0085 |
| 11.3 | 0.07302 | +0.0033 |
| 23.5 | 0.06885 | −0.0009 |

The closest approaches are at t = 23.5 (residual about 0.0009) and at t = 0 (residual about 0.001). At every generic height the slopes decrease toward β across the three decades. The stored log-norms were recomputed into slopes independently of the original script; the two agree exactly.

## 4. What this is not

The minima 0.1487 and 0.0698 are already theorems of the group, independent of any particular field. This run checks that real fields with those Galois groups produce Frobenius classes whose aligned product has the predicted slope, to the accuracy of a cutoff at a few thousand or twenty thousand primes. It is not an interval certificate of the slope, not a determination of new zeros, and not a statement about the Riemann Hypothesis. The 2O forced zero at ½ is a floating PARI confirmation that the root-number −1 zero is simple; it is not an Arb enclosure of L′(ρ, ½).

## 5. Files

`poly1112/measure_2O.py`, `measure-2O-results.json`; `poly1112/measure_2I.py`, `measure-2I-results.json`; `poly1112/groups_action.py`, `cycle-types.json`; `poly1112/sources/` (galpol 2I `polredbest` polynomial, LMFDB 24T201 API records). The 19 MB automorphism list for the degree-120 field is kept with the research files and is not published on the site. Figure script `poly1112/site_figure.py`.
