# Barrier exponent β for every finite non-abelian Frobenius group in H^×

Sprint run, 12 November 2026, ~4:25–5:00 PM PST. Extends `tetrahedral-frobenius-20261112.md` Result 3d (general law β = Σ_C δ_C sin²θ_C sin²α_C for the Model A norm nrd E_A(σ+tI) ~ C(t)(2σ−1)^β near Re s = ½). Closes the open items "2O/2I: what is min_I β?" and "[speculative] β_min > 0 always for non-abelian G", and refutes the [speculative] line "for 2O and 2I the minimum should be smaller still".

Labels: [known] established, with source; [project] proved or computed here; [speculative] unproved. **Nothing here is RH progress.** β only measures how the Model A *norm* (a product over primes whose Frobenius we place in H) degenerates as σ → ½⁺ because quaternion phases that point off the slice C_I get averaged away. No zero of any L-function is found, moved or excluded.

Setting: G ⊂ H^× finite non-abelian (the Galois image of a quaternionic 2-dim Artin representation, Chebotarev density δ_C = |C|/|G| [known]). A convention picks one representative per class; α_C is the angle between its axis and the unit I. "Best representative" = minimise sin²α_C within each class. β(I) := Σ_C δ_C sin²θ_C min_{g∈C} sin²∠(axis g, I).

## Results

1. **[project, proved] c = 3/4 for every such G.** Σ_C δ_C sin²θ_C = 1 − avg_g (Re g)² = 1 − ¼·avg|χ_ρ|² = 3/4, since χ_ρ(g) = 2 Re g and ρ is irreducible (first orthogonality, [known]). So β(I) = 3/4 − Iᵀ M(I) I with M = Σ δ_C sin²θ_C a_C a_Cᵀ over the chosen axes a_C; at an interior minimum β_min = 3/4 − λ_max(M).

2. **[project, proved; uses Jordan's lemma, known] β_min > 0 for every finite non-abelian G ⊂ H^×.** β(I) = 0 would need every non-central class to contain an element whose axis is I, i.e. every class meets A = G ∩ C_I. A is cyclic and proper (G is non-abelian, C_I is commutative). A proper subgroup of a finite group never meets every conjugacy class (Jordan 1872; G ≠ ∪ gAg⁻¹). So β(I) > 0 for each I; the sphere is compact and β is continuous (min of continuous functions), so min β > 0. Consequence: **no choice of slice and representatives removes the barrier at Re s = ½ for a non-abelian quaternion Frobenius group** — the extra directions always cost a positive power of (2σ−1).

3. **[project, proved] Binary dihedral Dic_n (order 4n, includes Q8 = Dic_2):**
   β_min = 1/4 for n odd, β_min = ½ − ¼ cos(π/n) for n even (Q8: ½ ✓).
   Proof: with φ the angle from the cyclic axis, the cyclic elements give ¼ sin²φ (Σ_{m=1}^{2n−1} sin²(πm/n) = n). The 2n elements j·a^k have θ = π/2 and axes at angles kπ/n in the perpendicular plane, split into k even / k odd classes. For n odd the lines of j a^k and j a^{k+n} coincide and lie in opposite classes, so I on one line kills both classes: total ¼. For n even, the b-part is ¼[2 − sin²φ(1 + cos(π/n)cos(π/n−2x))], minimised at sin φ = 1, x = π/(2n). Numerics (Nelder–Mead from 2·10⁵ sphere points) agree for n = 2..8 to 1e-15: 0.5, 0.25, 0.3232233, 0.25, 0.2834936, 0.25, 0.2690301.

4. **[project, proved for 2T; project, computer-assisted, interval arithmetic for 2O and 2I] Binary polyhedral minima.** The values below are exact at a critical point (3/4 − λ_max of the active-axis matrix M). Global optimality is a theorem for 2T (the value (9−√33)/24 already proved in the 2T note) and, for 2O and 2I, an Arb/FLINT interval branch-and-bound on the sphere (§6 below). A dense scan of 2·10⁶ random directions + 40 local refinements per group was the original discovery step and is no longer the proof of globality.

| G | order | min β | exact (smallest root of) | max β (scan) |
|---|---|---|---|---|
| 2T = SL(2,3) | 24 | 0.1356432231 | 12β² − 9β + 1 (= (9−√33)/24 ✓ matches the earlier note) | 1/3 (order-4 axis) |
| 2O | 48 | 0.1487126054 | 192β³ − 288β² + 126β − 13 | 0.2912 |
| 2I = SL(2,5) | 120 | 0.0697650597 | 960β³ − 1440β² + (640 − 24√5)β + 16√5 − 70 over Q(√5); over Q: 46080β⁶ − 138240β⁵ + 165120β⁴ − 98880β³ + 30416β² − 4288β + 181 | 0.1612 |

   At each minimiser the best representative of every class is unique up to g ↔ g⁻¹ (no ties), the minimiser is the top eigenvector of M (distance to the numeric argmin ≤ 7e-9), and β there equals 3/4 − λ_max(M) to 1e-61 (60-digit mpmath). Characteristic data of M: 2O tr 3/4, e₂ = 3/32, det = 1/384; 2I tr 3/4, e₂ = 5/48 − √5/40, det = 1/192 − √5/480. No class lies in the slice at any minimiser, so (as for generic I in 2T) there is no complex Euler product F there.

5. **[project] Tiny-way observations.**
   - **Bigger group, bigger barrier is false both ways.** 2O ⊃ 2T, yet min β(2O) = 0.1487 > 0.1356 = min β(2T): the extra order-8 elements (θ = π/4, π·3/4) and the 12 extra order-4 elements add off-slice mass faster than the finer axis set lets I align. 2I is the smallest.
   - **Universal floor [project, proved, using the 2O/2I interval certificate]:** by the classification of finite subgroups of H^× (cyclic, Dic_n, 2T, 2O, 2I; [known], e.g. Conway–Smith, *On Quaternions and Octonions*, 2003, ch. 3), every non-abelian quaternion Frobenius group has β ≥ 0.06976…, attained only by the icosahedral group. Dic_n never goes below ¼.
   - [speculative] The 2I case should be realisable by icosahedral Artin representations of LMFDB type 2.?.120t? with indicator −1 (e.g. stem degree 24, "SL(2,5)"); not looked up, and no field-level check was done in this run.

## 6. Interval certificate that the 2O and 2I critical values are global minima

**[project, computer-assisted, interval arithmetic].** β is even, so it is enough to cover a fundamental domain for I ∼ −I. We take the three positive faces of the cube, parametrised as (1,x,y), (x,1,y) and (x,y,1) with (x,y) ∈ [−1,1]². On this domain we run an adaptive branch-and-bound in Arb/FLINT (`beta1112/certify_global.py`), with two discard tests.

**Lipschitz test.** For unit a, I and a unit tangent v, |∂_v (a·I)²| = 2|(a·I)(a·v)| ≤ 2|u|√(1−u²) ≤ 1, so each (a·I)² — and the pointwise max over a finite set of axes — is 1-Lipschitz for geodesic distance on the sphere. Hence |β(I)−β(J)| ≤ c · d_geo(I,J) = (3/4) arccos(I·J). The spherical image of a cube-face rectangle is a spherically convex quadrilateral (the sides are great circles). The geodesic radius from the image of the rectangle centre is therefore attained at a vertex, and can be enclosed in Arb. If β(centre) − (3/4) r is certainly at least the certified floor, the box is discarded.

**Quadratic test.** On any region where a unique choice of best axis in each class is certified (by interval comparison of the (a·I)²), one has β(I) = 3/4 − Iᵀ M I with M fixed, so β(I) ≥ 3/4 − λ_max(M) for every unit I. The seed critical value β_* is 3/4 − λ_max(M_*) at the unique-best-axis cell of a previously computed approximate argmin; the top eigenvector of M_* is verified to lie in that open cell, and β_* is verified to be a root of the known minimal polynomial (to residual < 10^{-20}). If a box has a certified constant active set and 3/4 − λ_max(M) is not certified strictly below β_*, the box is discarded. (When the two enclosures overlap they represent the same algebraic number; the more conservative endpoint is kept as the certified floor.)

The algorithm terminates with no leftover boxes: 2T (a check against the known theorem) after 359 splits, 2O after 1187, 2I after 6679, all in a few seconds at 64 decimal digits. The resulting enclosure of the minimum has width less than 10^{-63} and contains the unique relevant real root of the minimal polynomial already recorded in `minpoly.json`. The same certificate was rerun with only certainly-true interval comparisons in the discard tests.

No smaller critical value occurs: a cell whose unconstrained floor 3/4 − λ_max(M) was certified strictly below β_* would have been reported, and none was.

## Not claimed
No RH statement; no field-level numerics for 2O/2I (only the group-theoretic exponent, which by Result 3d of the 2T note depends only on Chebotarev densities and angles). Global minimality for 2O and 2I is now the interval certificate of §6, not a numerical scan.

## Files
`beta1112/groups.py`, `scan.py` → `scan.json`; `refine.py` → `refine.json`; `exact.py`, `exact2.py` → `exact.json` (60-digit); `minpoly.py` → `minpoly.json` (sympy from /workspace/work/venv-voice); `fig.py` → `figures/binary-polyhedral-beta.png` (β on the sphere, Mollweide, red star = minimiser); `certify_global.py` → `certify-global-results.json` (interval certificate of global minimality).

## Page status
Folded into /critical-line-hypercomplex#binary-polyhedral-beta. Global-minimality labels for 2O and 2I upgraded after the §6 certificate, 12 November 2026, ~5:35 PM PST.
