# Binary tetrahedral Frobenius: 2T = SL(2,3) inside the quaternions

Sprint run, 12 November 2026, ~7:27–8:30 AM PST. Extends the Q8 notes `q8-explicit-formula-20261111.md` (quaternion Frobenius in the explicit formula), `q8-half-hecke-20261111.md` (aligned product as half an Artin L-function) and `frobenius-complex-s-20261111.md` (Model A) from Q8 to the binary tetrahedral group. Closes or narrows the two 2T items in open-questions.md (Q8 explicit-formula section, item "other finite G ⊂ H^×"; half-L section, item "binary tetrahedral/octahedral/icosahedral").

Labels: [known] established mathematics, with source; [project] proved or computed here (elementary, nothing deep); [speculative] unproved. **Nothing here is RH progress.** Every zero involved belongs to an ordinary classical L-function, and no zero is found, moved or excluded.

## 0. The group, the representation, the fields

- [known] 2T = SL(2,3) is the group of the 24 unit Hurwitz quaternions ±1, ±i, ±j, ±k, (±1±i±j±k)/2. It has 7 classes: 1, −1, the 6 elements of order 4 (call the class 4a), two classes of order 6 (6a ∋ (1+i+j+k)/2, 6b ∋ its inverse) and two of order 3 (3a = −6a, 3b = −6b). Its irreducible characters are 1, ω, ω̄ (through 2T → A4 → C3), ψ3 (3-dim, through A4), and three 2-dim ones: ρ, ρ⊗ω, ρ⊗ω̄. (Standard character table; recomputed by PARI `galoischartable`.)
- [project, checked on all 24 elements] **The quaternionic representation has a real character.** Writing H ≅ C² makes 2T act through a 2-dim representation ρ with χ_ρ(q) = 2 Re q (checked: error 0). Its Frobenius–Schur indicator is −1 (computed as (1/24) Σ χ_ρ(g²) = −1), so ρ is quaternionic. The other two 2-dim reps ρ⊗ω, ρ⊗ω̄ have complex characters (indicator 0); the 1- and 3-dim ones: ω, ω̄ indicator 0, ψ3 indicator +1. So the task's object, "the 2-dim rep with real character and indicator −1", exists and is exactly the restriction of H^×. Its determinant is trivial (ρ lands in SL2).
- [known, LMFDB, fetched 12 Nov 2026 ~7:30 AM PST; pages saved in `tetra1112/sources/`] LMFDB files these representations under the label "24t7" (the regular permutation container 24T7). Their degree-8 stem fields have Galois group 8T12 ≅ SL(2,3), which is how the search `dimension=2&group=8T12&frobenius_schur_indicator=-1` (278 matches) finds them. We use the two of smallest conductor:
  - **2.26569.24t7.b.a**: conductor 26569 = 163², root number −1, indicator −1, χ(complex conjugation) = −2 (so ρ(c) = −1 and Γ-factor Γ_R(s+1)²), stem field 8.0.705911761.1, f = x⁸ − x⁷ + x⁶ − 4x⁵ + 5x⁴ − 8x³ + 4x² − 8x + 16. https://www.lmfdb.org/ArtinRepresentation/2.26569.24t7.b.a
  - **2.26569.24t7.a.a**: conductor 163², root number −1, χ(c) = +2 (totally real; Γ_R(s)²), stem field 8.8.18755369578009.1, f = x⁸ − x⁷ − 44x⁶ + 43x⁵ + 442x⁴ − 32x³ − 1311x² − 1156x − 241. https://www.lmfdb.org/ArtinRepresentation/2.26569.24t7.a.a
  Both have projective image A4 with projective stem field 4.4.26569.1 (the same A4 field for both).
- [project, PARI 2.x via cypari2] PARI agrees: `galoisidentify` of the degree-24 Galois closure is [24,3] = SL(2,3); `lfunparams` gives conductor 26569 and Vga [1,1] (b.a) / [0,0] (a.a); `lfunrootres` gives ε = −1 and `lfunorderzero` = 1 for both. The other constituents (PARI): ω, ω̄ conductor 163 (even cubic characters); ψ3 conductor 26569; ρ⊗ω, ρ⊗ω̄ conductor **163** for b.a (Vga [1,1]) and 26569 for a.a, with complex root numbers and L(ρ⊗ω, ½) ≠ 0 (order 0). The conductor-163 twists of the first field (b.a) are the LMFDB Galois orbit 2.163.8t12.a, i.e. [2.163.8t12.a.a](https://www.lmfdb.org/ArtinRepresentation/2.163.8t12.a.a) and [2.163.8t12.a.b](https://www.lmfdb.org/ArtinRepresentation/2.163.8t12.a.b): dimension 2, group SL(2,3), conductor 163, Frobenius–Schur indicator 0, stem field 8.0.705911761.1 (the same polynomial as b.a), determinants 1.163.3t1.a.a / .a.b, projective image A4 with stem 4.4.26569.1. LMFDB records that this pair has the smallest conductor of any 2-dimensional representation with projective image A4. The character table matches χ_ρ · ω and χ_ρ · ω̄ (values 2, −2, −ζ₃ / ζ₃+1, 0, ζ₃ / −ζ₃−1). For the second field (a.a) the twists are the orbit 2.26569.8t12.a (conductor 26569, same determinants, stem polynomial identical to a.a). Fetched 12 November 2026, 4:13 PM PT, from the LMFDB API `artin_reps` (the HTML search page is behind a reCAPTCHA from this host); JSON saved in `tetra1112/sources/`.

## 1. Part (a): actual quaternion Frobenius  [project]

Method A: degree-24 Galois closure (nfsplitting + polredbest), `idealfrobenius` for the Frobenius of a chosen prime 𝔓 | p, and an explicit isomorphism Gal(K/Q) → 2T found by generator search and checked to be a homomorphism on all 576 pairs. The composition convention was fixed by the equivariance test Frob_{τ𝔓} = τ Frob_𝔓 τ⁻¹ (6/6 τ pass). PARI caveat: `idealfrobenius` applied to the prime ideal returned by `nfgaloisapply` gave a wrong class, so we locate τ𝔓 in `idealprimedec` instead. Method B: PARI `lfunartin` Dirichlet coefficients a_n (n ≤ 20000) of all seven characters, which do not use our isomorphism.

Checks (both fields, all unramified p ≤ 20000, i.e. 2261 primes. For the 429 primes p ≤ 3000, every prime of K above p was used):
- class of the quaternion vs PARI: |a_p(χ) − χ(class of Frob_p)| ≤ 2.5·10⁻¹⁶ for all 7 characters and all p;
- **Re(Frob^k) = ½χ_ρ(Frob^k)**: exact for k = 1..12 (character table), and vs Λ_ρ(p^k)/log p rebuilt from PARI's a_n (2325 prime powers ≤ 20000): max error 8.9·10⁻¹⁶;
- choice of prime above p: the Frobenii of the primes above p run through **every** element of the class (each equally often: e.g. p = 7, order 6, 4 primes → the 4 elements (1±i±j±k)/2 with an even number of minus signs; p = 5, order 4, 6 primes → ±i, ±j, ±k; p = 2, order 3, 8 primes → 4 elements twice each);
- the pure (i,j,k) part flips with the choice: two primes above the same p differ in pure part by at least √2 (any non-central class);
- averaging: for k = 1..6 the average of Frob^k over the primes above p is exactly ½χ_ρ(Frob^k) (pure part 0, all 429 primes), as proved in §2.

## 2. Part (b): choice-free statistics and fixed conventions

**Theorem 1 [project, proved; elementary].** Let G ⊂ H^× be any finite non-abelian group (binary dihedral incl. Q8, 2T, 2O, 2I). If λ: H → R is linear and λ(hgh⁻¹) = λ(g) for all g, h ∈ G, then λ = c·Re. In words: **the reduced trace is the only choice-free linear statistic of quaternion Frobenii, for every non-abelian G, not just Q8.** The same applies to Frob^k for each k, and to vector-valued λ coordinatewise.
*Proof.* The real span of G is a subalgebra of H containing two non-commuting elements; the only such subalgebra is H itself (the others are R and the slices C_u). So λ(hxh⁻¹) = λ(x) for all x ∈ H by linearity. On the pure quaternions, conjugation by G acts by the rotation group G/{±1} ⊂ SO(3). A finite rotation group fixing a non-zero vector consists of rotations about that axis, so it is cyclic. G/{±1} is not cyclic (G is non-abelian), so it fixes no non-zero pure quaternion, and λ restricted to Im H (a fixed vector under the dual action) is 0. ∎ Numerically for 2T, the constraint matrix {hgh⁻¹ − g} has singular values 16.97, 16.97, 16.97, 0 with null vector (1,0,0,0).

**Theorem 2 [project, proved; elementary].** For 2T and any fixed convention (one representative c_C chosen in each class, g_p := c_{class(Frob_p)}), the class function C ↦ Im c_C ∈ R³ always has non-zero weight on the faithful 2-dim reps {ρ, ρ⊗ω, ρ⊗ω̄}. So **it never decomposes into L-functions of proper subquotients (A4, C3) alone.** Its ρ-weight can be zero or non-zero depending on the convention.
*Proof.* The faithful reps are exactly those with −1 ↦ −1. A class function lies in the span of the A4 characters iff f(−g) = f(g). The pure parts of the 6a elements are (±i±j±k)/2 with an even number of minus signs, and those of 3a = −6a have an odd number. These are disjoint sets, so f(3a) ≠ f(6a) for every choice. ∎ Exhaustive enumeration of all 6·4⁴ = 1536 conventions: 168 have zero ρ-weight, **none** has zero faithful weight (minimum norm 1/√6 ≈ 0.408), and the largest ρ-weight norm is √2/3 ≈ 0.471. So the guess "zero ρ-weight and only proper subquotients" is half right at best: ρ can be avoided, but the cubic twists ρ⊗ω, ρ⊗ω̄ cannot.

**The aligned convention and its exact decomposition [project, proved + checked].** Take g6 = (1+i+j+k)/2 = ½ + (√3/2)n with n = (i+j+k)/√3, choose the representatives of the six classes 1, −1, 6a, 6b, 3a, 3b to be the six powers of g6 (they are one per class), and choose u = i for the order-4 class. In field terms: for each p we take the prime of K above p whose Frobenius lies in the cyclic group ⟨g6⟩ (or equals i), and §1 shows such a prime always exists. Define ω by ω(g6) = e^{2πi/3}. Then for every unramified p and every k ≥ 1:

    Im(g_p^k) = Im χ_{ρ⊗ω}(Frob_p^k) · n  +  ε_k · [Frob_p ∈ 4a, k odd] · u,   ε_k = +1 (k ≡ 1 mod 4), −1 (k ≡ 3 mod 4),
    with  1_{4a} = ¼ (1 + ω + ω̄ − ψ3).

So the k = 1 pure part is the class function n·Im χ_{ρ⊗ω} + u·¼(1 + ω + ω̄ − ψ3). The coefficients are a_1 = a_ω = a_ω̄ = u/4, a_ψ3 = −u/4, **a_ρ = 0**, a_{ρ⊗ω} = −(i/2)n, a_{ρ⊗ω̄} = +(i/2)n (printed in `results-ab.json`). (The pure part lies in the plane spanned by n and u. These are coordinates in that non-orthogonal basis.) The n-direction needs no correction at any k: summed over prime powers it is exactly the imaginary part of the prime side of −L′/L(ρ⊗ω, s) (unramified part). Checked: the formula holds on every class for k = 1..12 (error 1.9·10⁻¹⁶).

Prime sums (Gaussian explicit-formula weights −2Λ(n)g(log n)/√n, g(x) = T/(2√π)·e^{−T²x²/4}, n ≤ 20000, p = 163 omitted). The quaternion side (method A, actual Frobenius) is compared with the character side rebuilt from PARI a_n (method B):

| field | T | n-dir: quaternion | Im prime side L(ρ⊗ω) | u-dir k=1: quaternion | ¼(1+ω+ω̄−ψ3) via a_p | Re part vs ½·ρ side |
|---|---|---|---|---|---|---|
| b.a | 1 | −0.0820867988 | −0.0820867988 | −0.4220345142 | −0.4220345142 | 0.1533284468 (both) |
| b.a | 2 | 0.0502928440 | 0.0502928440 | −0.0623258371 | −0.0623258371 | 0.2939510994 (both) |
| b.a | 3 | 0.1752833166 | 0.1752833166 | −0.0035863864 | −0.0035863864 | 0.1799726800 (both) |

All differences ≤ 1.2·10⁻¹⁶ at T = 1, 1.5, 2, 3 for both fields, and on a 65-point T grid (figure, panel b). The full u-direction with the ε_k corrections also agrees to 1.1·10⁻¹⁶. This agreement is expected, since the identity holds prime by prime. What it confirms is that the class labels from `idealfrobenius` + our isomorphism and PARI's Artin coefficients are consistent, and that the decomposition has no slips. Empirical ρ-correlation of the aligned pure part over p ≤ 20000 is ∓8.4·10⁻⁴ per component (Chebotarev noise; the exact value is 0).

**Interpretation [project].** As for Q8, the explicit formula sees only the reduced trace through L(ρ). Under the aligned convention the pure part is a disguised combination of other L-functions: L(ρ⊗ω), L(ρ⊗ω̄) (the cubic twists, faithful, complex character) and ζ, L(ω), L(ω̄), L(ψ3). These would be heard by those functions' zeros, never by L(ρ)'s. A zero-side check (rebuilding the n-direction from PARI zeros of L(ρ⊗ω)) was **not** done in this sprint. New feature vs Q8: in Q8 the fixed-convention pure part came only from proper quotients (quadratic characters), but for 2T faithful twists are unavoidable (Theorem 2).

## 3. Part (c): the aligned Model A product for 2T

Model A with the complex unit I = n (the axis of ⟨g6⟩), s = σ + t·n, p^{−s} ∈ C_n ≅ C. Primes whose Frobenius has order 1, 2, 3 or 6 have representatives in the slice. They give a complex Euler product F₆(s) = ∏ (1 − ψ_p p^{−s})^{−1} with ψ_p = e^{iθ} the slice value, ψ_p = ½χ_ρ + i·Im χ_{ρ⊗ω} (Frob_p). Order-4 primes (density ¼) cannot be put in the slice: their axes ±i, ±j, ±k make an angle with n whose cosine is c = 1/√3. In the full quaternion product they do not commute with the rest, but the Model A norm is still order-free:

    nrd E_A(σ + t n) = |F₆(s)|² · O₄(σ,t),   O₄ = ∏_{Frob_p ∈ 4a} (1 − 2c p^{−σ} sin(t log p) + p^{−2σ})^{−1}   (rep with u·n = +c).

(The ordered quaternion product over slice primes stays in C_n and equals the complex product to 6·10⁻¹⁵; nrd multiplicativity checked to 1.4·10⁻¹⁴.)

**Result 3a [project, exact local identity + float numerics].**

    F₆(s) = L(ρ,s)^{1/2} · [ L(ρ⊗ω,s) / L(ρ⊗ω̄,s) ]^{1/2} · [ ζ(2s) L(ω,2s) L(ω̄,2s) / L(ψ3,2s) ]^{1/8} · H(s),

all L-functions partial (p = 163 removed; PARI `lfuneuler`). H is an Euler product whose local logs are O(p^{−4s}) (they come from order-4 primes: (1 + p^{−2s})^{1/2} against the ⅛-power). They are exactly 0 for every other class (local check: |log H_p| = 2.5·10⁻⁹ at x = 10⁻² for 4a, scaling like x⁴; ≤ 5·10⁻¹⁸ otherwise). So H is holomorphic and zero-free on Re s > ¼. Compare Q8: L(ρ)^{1/2}·[ζL(χ_u)/(L(χ_v)L(χ_w))]^{i/4}·[ζ(2s)/L(χ_u,2s)]^{1/4}. For 2T the exponents are ½ (on L(ρ)), ½ (on the conjugate pair ratio L(ρ⊗ω)/L(ρ⊗ω̄)) and ⅛ (on the A4 combination at 2s). **All are real.** The i/4 of Q8 is replaced by the ratio of complex-conjugate twists, whose ½-power is a pure phase on the real axis.
Checks, t ∈ [0,40] step 0.1, PARI `lfun` for the right side, direct product over p ≤ X for the left:

| field | σ = 2, X = 5·10⁴ → 2·10⁵ | σ = 3/2, X = 5·10⁴ → 2·10⁵ | naive F₆ ≈ L(ρ)^{1/2} off by |
|---|---|---|---|
| b.a | 2.8·10⁻⁸ → 2.4·10⁻⁹ | 7.9·10⁻⁶ → 1.3·10⁻⁶ | 0.088–0.57 |
| a.a | 2.4·10⁻⁸ → 3.3·10⁻⁹ | 7.1·10⁻⁶ → 1.8·10⁻⁶ | 0.076–0.57 |

**Result 3b [project, consequence of 3a; checked].** Branch structure. A simple zero of L(ρ), L(ρ⊗ω) or L(ρ⊗ω̄) multiplies F₆ by −1. A zero of the A4 combination at 2s (i.e. at s = ρ₀/2) turns the phase by ±π/4. Since every exponent is real, **|F₆| is single-valued** on Re s > ¼ away from zeros, unlike Q8 where |F| jumped by e^{±π/2}. At s = ½ the simple Artin zero (ε = −1) and the ⅛-power of ζ(2s)'s pole combine: **F₆(s) ≈ c·(s − ½)^{3/8}** (½ − ⅛). This assumes L(ρ⊗ω, ½) ≠ 0 (now an interval certificate for both fields, §6: |L(½)| > 1.8283 for the conductor-163 twists and |L(½)| > 0.72674 for the conductor-26569 twists) and L(ω,1), L(ψ3,1) ≠ 0 ([known]: non-vanishing on Re s = 1 for these automorphic L-functions; tetrahedral ρ is automorphic by Langlands 1980, *Base change for GL(2)*, ψ3 = Ad ρ is automorphic on GL(3) by Gelbart–Jacquet 1978 (Ann. Sci. ENS 11), and non-vanishing on Re s = 1 is Jacquet–Shalika 1976, Invent. Math. 38; primary sources not re-read in this sprint). Measured slope of log|F₆| vs log(s − ½) at s − ½ = 10⁻⁵: **0.37495 (b.a), 0.37501 (a.a)**. The phase change once round s = ½ (radius 10⁻³) is 2.35619449 = 3π/4 in both fields, as predicted. Q8 had (s − ½)^{±1/4}.

**Result 3c [project, closed form + float numerics].** The order-4 primes in the norm. Expanding −log(1 − 2c r sin φ + r²) to second order, and using that squares never lie in 4a (so 1_{4a}(F²) = 0), gives

    log O₄ = −(c/2)·arg Λ₄(s) − (c²/4)·log|Λ₄(2s)| − ((1 − c²)/4)·log Λ₄(2σ) + log K,   Λ₄ = ζ L(ω) L(ω̄)/L(ψ3) (partial),

with K's local logs O(p^{−3σ}). Checked against the direct product: 1.1·10⁻⁷ at σ = 2 (from 4.6·10⁻⁷ at X/4) and 8.9·10⁻⁵ at σ = 3/2 (from 1.9·10⁻⁴). O₄ is the **same function for both fields**, since 4a membership and Λ₄ only see the common A4 field 4.4.26569.1.
Consequence: Λ₄(2σ) has ζ's pole at σ = ½, so **O₄ ~ C(t)·(2σ − 1)^{(1−c²)/4} = (2σ − 1)^{1/6}** for t ≠ 0. At t = 0 it is (2σ − 1)^{1/4}, because the |Λ₄(2s)| term joins in. The direct product converges too slowly near ½ to show this, so (as for Q8) the limit rests on the closed form (with K summed to p ≤ 10⁵). Slopes of log O₄ vs log(2σ − 1) at 2σ − 1 = 2·10⁻⁴: 0.2499 (t = 0), 0.1673 (t = 7), 0.1677 (t = 23.5), creeping towards 1/6. Since |F₆|² stays finite at generic points of the line, **the Model A norm still has a barrier at Re s = ½, but it is a sixth root, not a square root.** If O₄ is continued through the closed form, the arg term makes it path-dependent by e^{∓π/√3} ≈ 0.163 / 6.13 per enclosed zero of Λ₄ (not checked numerically).

**Result 3d [project, elementary; general barrier law].** Averaging the r² term over the phase gives the exponent for any finite G ⊂ H^×, any unit I and any representatives:

    nrd E_A(σ + tI) ~ C(t)·(2σ − 1)^β,   β = Σ_C δ_C · sin²θ_C · sin²α_C,

where δ_C is the Chebotarev density, e^{±iθ_C} the eigenvalues, and α_C the angle between the representative's axis and I. β does not depend on the sign choices, only on the angles. Q8 aligned: β = 2·¼·1·1 = ½ ✓ (the known square root). 2T: I along an order-6 axis gives β = ¼·(2/3) = **1/6**; along an order-4 axis, β = (2/3)·(3/4)·(2/3) = 1/3. The minimum over all I (best representative in each class) is **(9 − √33)/24 ≈ 0.1356**, reached in the plane of an order-4 and an order-6 axis, 14.75° from the order-6 axis (numeric scan of 2·10⁵ directions + closed-form 2×2 eigenvalue). For such a generic I no class lies in the slice, so there is no complex F. [speculative] For 2O and 2I the minimum should be smaller still, but never 0: a non-abelian G cannot put all its classes on one axis.

## 4. Overnight cross-checks from 11–12 Nov (oddcond1111/), read-only

- **P6 at p = 11, brute force** (`genp-P6-11.txt`, `results-genp-P6-p11-brute.json`, finished 10:40 PM PST Nov 11): P_O = 1 + x² + 12x⁴ + x⁶ + 11x⁸ and P_Λ = 1 + 12x² + x⁴ + 11x⁶, with 26 = 2p + 4 primitive lattices. This is exactly the proof's prediction (P_O = 1 + x² + (p+1)x⁴ + x⁶ + p x⁸, P_Λ = 1 + (p+1)x² + x⁴ + p x⁶), which by Lemma 1 gives (1 + 12x²)/(1 − x⁴) = (1 + (p+1)x²)/(1 − x⁴) at p = 11. **Agrees.** [project]
- **Q at p = 7, method B** (`Bp-Q-7.txt`): **did not finish.** The last write was 10:59 PM PST Nov 11, no process is running now, and there is no `results-Bp-Q-p7.json`. Its three completed levels (16, 66, 16 new lattices; running totals 17, 83, 99) match method A's first three levels exactly. Method A then found 112 more (total 211, the proof's prediction). So this is a partial agreement (99 of 211), not an independent full check. [project]
- Machine-readable: `tetra1112/crosscheck-overnight.json` (from `crosscheck_overnight.py`; it writes nothing into oddcond1111/).

## Not claimed

No information about zeros of any L-function, no RH or GRH statement. PARI zeros were not used (the zero side for L(ρ⊗ω) is untested). Floats, not intervals. The near-½ limits rest on closed forms plus PARI values. LMFDB data are as fetched and the ρ⊗ω conductor is PARI's. Theorem 1 and 2 proofs are elementary group theory.

## Files

`tetra1112/`: `certify_exponents.py` (this catch-up; → `certify-exponents-results.json`), `export_pari.py` (PARI export → `pari-data.json`, 34 s), `analyze.py` (parts a, b → `results-ab.json`, `run-ab.log`), `halfprod.py` (part c → `results-c.json`, `run-c.log`, ~6 min), `crosscheck_overnight.py` (→ `crosscheck-overnight.json`), `fig.py` (→ `figures/tetrahedral-frobenius.png`, `fig-data.json`), `site_figure.py` (site SVGs), `sources/` (LMFDB pages). Run PARI scripts with `/workspace/tools/pari-venv/bin/python3`.

## Page status

Deployed ~8:20 AM PST, 12 Nov 2026, as /critical-line-hypercomplex#binary-tetrahedral-frobenius (version 93c203c6-bdd5-4c08-a3a4-54e7671c5879). The page text is plain-language, with labels [known]/[project] and the no-RH sentence. Module: /workspace/math/zeta-zeros-build/tetra_frobenius_page.py. Site figure: `figures/tetrahedral-frobenius-site{,-mobile}.svg` (published as img/hyper-binary-tetrahedral-frobenius{,-mobile}.svg). Checks: see DEPLOY-LOG.md.


## 5. Certification of the exponents — 12 Nov 2026, ~4:30 PM PT (catch-up sprint)

Labels as in §0. Still **no RH progress**.

**LMFDB citation [known].** See the updated paragraph in §0. API records: `tetra1112/sources/lmfdb-artin-2.163.8t12.a.json`, `lmfdb-artin-cond-26569-dim2.json`, `lmfdb-twist-citation-extract.json`. The NFGal polynomials match the two stem fields exactly.

**Exact local identity for F₆ and H [project, proved — computer-assisted].** For each of the seven classes, the 8th-power identity

    P_ρ⁴ P_{ρ⊗ω}⁴ P_ζ(x²) P_ω(x²) P_{ω̄}(x²)  =  P_F⁸ P_{ρ⊗ω̄}⁴ P_{ψ3}(x²) H⁸

holds as an equality of polynomials in Q(ζ₁₂)[x], where P_F = 1−ψx (or 1 on 4a) and H⁸ = (1−x⁴)² on 4a (and 1 otherwise). Thus, with p=163 removed,

    F₆(s) = L(ρ,s)^{1/2} [L(ρ⊗ω,s)/L(ρ⊗ω̄,s)]^{1/2} [ζ(2s) L(ω,2s) L(ω̄,2s)/L(ψ3,2s)]^{1/8} H(s),

with **H(s) = ∏_{Frob_p ∈ 4a} (1−p^{−4s})^{1/4}**. This product converges absolutely for Re s>1/4, is holomorphic and zero-free for Re s>0, and H(½)=∏(1−p^{−2})^{1/4}∈(0,1). Script: `certify_exponents.py` (all seven classes `ok: true`).

**Central exponent 3/8 [project, proved, with listed hypotheses].** A simple zero of L(ρ) at s=½ contributes (s−½)^{1/2}; the simple pole of ζ(2s) contributes (s−½)^{−1/8}. If L(ρ⊗ω,½)≠0 (hence also the conjugate) and L(ω,1), L(ω̄,1), L(ψ3,1)≠0, then F₆(s)∼c(s−½)^{3/8} with c≠0,∞, and the monodromy is 3π/4. The non-vanishing on Re s=1 is [known] (Jacquet–Shalika 1976, Invent. Math. 38; ω is Dirichlet; ψ3=Ad ρ is automorphic by Gelbart–Jacquet 1978; tetrahedral ρ is automorphic by Langlands 1980). L(ρ⊗ω,½)≠0 is now an interval certificate for both fields (§6: |L(½)| > 1.8283 for the conductor-163 twists 2.163.8t12.a.a and .a.b, and |L(½)| > 0.72674 for the conductor-26569 twists 2.26569.8t12.a.a and .a.b). L(ρ) has ε=−1 and PARI order 1 at ½.

**Barrier exponents 1/6 and 1/4, and the general law [project, proved].** The r²-average of −log(1−2a r+r²) is exactly −sin²θ sin²α (symbolic integral). Chebotarev plus the pole of ζ at 1 then give

    nrd E_A(σ+tI) ∼ C(t)·(2σ−1)^β,    β = Σ_C δ_C sin²θ_C sin²α_C,

using [known] non-vanishing of the relevant L-functions on Re s=1. Special values: Q8 aligned β=1/2; 2T along an order-6 axis β=1/6; along an order-4 axis β=1/3. The remainder K after the closed form of log O₄ has vanishing r¹ and r² terms (trig identities). For 0≤r≤1/2, an explicit majorant from the degree-≥3 remainders of −log(1−u) and the artanh series for log Λ₄ gives |log K(r,φ)| ≤ C r³ with C=9.304885…. The tail over p≥5 at σ=1/2 is at most 8.830…. The Euler product for K therefore converges uniformly on Re s≥1/2. This proves O₄ ∼ (2σ−1)^{1/6} for t≠0 and (2σ−1)^{1/4} at t=0, with no further numerical slope.

**The minimum of β [project, proved].** In the plane of an order-4 axis and an order-6 axis, the critical-point equation reduces to 33w²−33w+2=0 for w=cos²φ. At the larger root w=[33+5√33]/66 one has β=(9−√33)/24, which is smaller than the axis values 1/6 and 1/3. The squaring identity 25w(1−w)=2(2w−1)² holds exactly. A previous 2·10⁵-direction scan found no smaller value; by symmetry of the 4+6 axes the global min is in such a plane.

**Files.** `tetra1112/certify_exponents.py`, `tetra1112/certify-exponents-results.json`. Page: /critical-line-hypercomplex#binary-tetrahedral-exponents (anchor added 12 Nov, 4:30 PM PT).

## 6. Interval certificate for L(ρ⊗ω, ½)

**[project, computer-assisted, interval arithmetic].** The 3/8 central exponent for the aligned product uses the hypothesis L(ρ⊗ω, ½) ≠ 0 (and hence L(ρ⊗ω̄, ½) ≠ 0 by conjugation). For the conductor-163 twists of the first field (LMFDB 2.163.8t12.a.a and 2.163.8t12.a.b) this is now certified. The completed L-function for two Γ_R(s+1) factors has a rapidly convergent expansion in the Bessel function K_0, split at t=1 and using the functional equation. The Dirichlet coefficients a_n lie in ℚ(ζ_3) and are read from the exact conjugacy class of each Frobenius (already computed as a unit quaternion) together with the exact character table; the Euler factor at the ramified prime 163 is 1+T. The root number is the cubic Gauss sum of the determinant character, ε = G(χ_det)/√163, evaluated as a finite Arb sum (162 terms). Integrals and K_0 values are Arb/FLINT interval arithmetic (`certify_lhalf.py`). The resulting enclosures are

- 2.163.8t12.a.a: L(½) ∈ 1.82725 + 0.06243 i  with radius < 1.5·10^{-6}, so |L(½)| > 1.8283;
- 2.163.8t12.a.b: the complex conjugate, with the same lower bound.

A floating-point two-split check of the same theta integrals recovers ε to 10^{-14}, in agreement with the Gauss-sum formula and with PARI’s `lfunrootres`.

For the conductor-26569 twists of the second field (LMFDB 2.26569.8t12.a.a and 2.26569.8t12.a.b; Vga = [0, 0], so two Γ_R(s) factors) the same exact coefficients in ℚ(ζ_3) are used, with ramified Euler factor 1 at 163. Direct integration of the K_0 theta over a long range is slow at this conductor, so the Mellin integral is evaluated termwise: writing α = 2π/√N = 2π/163,

A(X) = 4 ∑_{n} a_n (α n)^{-1/2} F(α n X),    F(z) = ∫_z^∞ u^{-1/2} K_0(u) du,

with F computed from the classical complete integral 2^{-3/2} Γ(1/4)^2 minus a cumulative Arb integral, and a rigorous near-zero remainder together with the bound 0 < F(z) < z^{-1/2} K_0(z). The root number has no equally simple closed form as at conductor 163; it is recovered from the two-X identity ε = (A(1) − A(2))/(A^*(1/2) − A^*(1)), itself a consequence of the functional equation, and then checked a posteriori by the residual Λ − ε conj(Λ) at s = ½ (absolute value less than 6.2·10^{-8}). The enclosures, at 40 decimal digits with 1500 coefficients, are

- 2.26569.8t12.a.a: L(½) ∈ 0.603107737 + 0.405483136 i  with radius < 1.5·10^{-10}, so |L(½)| > 0.72674;
- 2.26569.8t12.a.b: the complex conjugate, with the same lower bound.

These agree with PARI’s floating values. No RH statement is made.
