# How far can the edge sit from ½? A uniform bound for every finite G ⊂ H^×

Sprint, 14 November 2026, ~1:00–2:00 PM PST (Grant, 12:57 PM PST: "That seems like a good direction to me"). Follows `binary-polyhedral-barrier-20261112.md` (β > 0 for every non-abelian G), `tetrahedral-frobenius-20261112.md`, `octahedral-exponent-20261113.md`, `icosa-exponent-20261113.md` and its erratum on holomorphy. Scripts and data: `edge1114/`. Figure: `figures/uniform-edge-site.svg` (+ `-mobile.svg`, PNG previews).

Labels: [known] established, with source; [project, proved] proved here by hand; [project, computer-assisted] checked by a script; [project conjecture] unproved. **No RH claim.** The "edge" below is where an auxiliary correction factor of the Model A aligned product stops converging. It is not a statement about zeros of ζ or of any L-function, and nothing here moves, finds or excludes a zero.

## In plain words

Each aligned product on the page is written as (a few Artin L-functions to fractional powers) × (a leftover correction H). H is harmless near ½, but it only makes sense to the right of some vertical line Re s = σ: that line is its *edge*. Earlier notes found σ = ¼ for the 24-element group and ⅓ for the 48- and 120-element groups. The question was whether the edge can wander arbitrarily far from ½ as the group changes.

Answer: along the axis the page always uses (the longest cyclic direction of the group) the edge is **always ⅓ or ¼**, for every finite group of unit quaternions, the infinite dihedral family included. So the distance from ½ is always between 1/6 and 1/4, and both values occur. If one allows any axis at all, the edge can be pushed as close to 0 as one likes (distance approaching ½), but never to 0. The reason is a counting rule: the edge is 1/k₀, where k₀ is the first power of a prime-Frobenius that the two peeled-off layers fail to account for, and k₀ is decided by small group facts (whether the group has elements of order 3, 4 or 6 off the chosen axis, and whether g and g⁻¹ are conjugate).

## 1. The object and two definitions of the edge

**[project] Set-up** (as in the 2T/2O/2I notes). G ⊂ H^× finite, the Galois image of a quaternionic Artin representation of a number field K. Pick a unit pure quaternion I. The slice is A = G ∩ C_I (a cyclic group, C_I ≅ ℂ). For a class C meeting A, ψ(C) ∈ ℂ is its element in A, written as a complex number, chosen with Im ≥ 0 when C ∩ A = {z, z̄}. The aligned product is

    F(s) = ∏_{p unramified, Frob_p meets A} (1 − ψ(Frob_p) p^{−s})^{−1},   log F_p = Σ_k f_k(Frob_p) x^k/k,  x = p^{−s},

with f_k(g) = ψ(cl g)^k if cl g meets A and f_k(g) = 0 otherwise.

**[project, proved] Ghost layers.** Put Φ_d(g) = (1/d) Σ_{e | d} μ(e) f_{d/e}(g^e). Then f_k(g) = Σ_{d | k} d Φ_d(g^{k/d}) (Möbius inversion), each Φ_d is a class function, and so log F_p = Σ_d Σ_m Φ_d(g^m) x^{dm}/m. Expanding Φ_d = Σ_χ c_{d,χ} χ in irreducible characters gives F = ∏_d ∏_χ L(χ, ds)^{c_{d,χ}} as formal Euler products. The page's identities (2T, 2O, 2I, Q8) peel exactly d = 1 and d = 2 (the layers that matter on Re s ≥ ½; L(χ, 2s) carries the pole of ζ(2s) at ½), and H is the rest:

    log H_p(x) = Σ_{k≥3} r_k(g) x^k/k,   r_k(g) = Σ_{d | k, d ≥ 3} d Φ_d(g^{k/d})
             = f_k(g) − f_1(g^k) − [k even](f_2(g^{k/2}) − f_1(g^k)),

so |r_k| ≤ 4 for all k. Let **k₀ = min{d ≥ 3 : Φ_d ≢ 0}** (k₀ = ∞ if there is none) and **c₁ = ⟨Φ_{k₀}, 1⟩** (the trivial-character weight; equals the class mean of r_{k₀} divided by k₀).

**Two edges.**
- (E1) σ_abs(G, I): the abscissa of absolute convergence of the Dirichlet series log H = Σ_{p,k} r_k(Frob_p) p^{−ks}/k (equivalently of the Euler product of H).
- (E2) σ_hol(G, I): the abscissa of the largest right half-plane to which H continues holomorphically (from Re s large).

Always σ_hol ≤ σ_abs. "The edge" in the 2T/2O/2I notes ("H holomorphic for Re s > 1/4, > 1/3") is E1, and the 2I erratum's "branch point at 1/3" is the statement E2 = E1 there.

**Validation [project, computer-assisted].** `edge1114/validate.py` reproduces every closed form for H already in the notes: 2T (6-fold axis) only class 4a, r₄ = r₈ = −1, i.e. H = ∏(1 − p^{−4s})^{1/4}, c₁ = −1/16; 2O (8-fold) C4a r₃ = r₇ = −2i, C4b r₄ = −1, C6 (1 + x³)^{1/3}, C3 (1 − x³)^{1/3}, c₁ = −i/12; 2I (10-fold) class mean of r₃ = −0.2351i (c₁ = −0.07837i), as in the erratum; Q8 c₁ = −i/6.

## 2. General theorems

**Theorem 1 [project, proved; uses Chebotarev, known].** If k₀ < ∞ then σ_abs(G, I) = 1/k₀ exactly, independent of the field K. If k₀ = ∞ then H ≡ 1.
*Proof.* For σ > 1/k₀: Σ_p Σ_{k≥k₀} 4 p^{−kσ}/k < ∞. At σ = 1/k₀: pick a class C with r_{k₀}(C) = k₀Φ_{k₀}(C) ≠ 0; the terms k = k₀, Frob_p ∈ C alone give |r_{k₀}(C)|/k₀ · Σ_{p∈C} p^{−1} = ∞ by Chebotarev (positive density) [known: Chebotarev density theorem]. ∎

**Theorem 2 [project, proved; uses known facts on Artin L-functions at Re s = 1].** If k₀ < ∞ and c₁ ∉ ℤ_{≤0}, then σ_hol = σ_abs = 1/k₀, and H has a genuine singularity at s = 1/k₀: near it, H = (s − 1/k₀)^{−c₁} × (holomorphic, nonzero).
*Proof.* log H − Σ_χ c_{k₀,χ} log L(χ, k₀ s) has vanishing p^{k₀}-coefficients and bounded others, so it converges absolutely on Re s > 1/(k₀+1). For χ ≠ 1 irreducible, L(χ, w) is holomorphic and nonzero near w = 1 [known: Brauer induction writes L(χ, w) as a product of integer powers of Hecke L-functions, nonzero on Re w = 1 except the pole of ζ_F; the order at w = 1 is ⟨χ, 1⟩ = 0; Brauer 1947, Ann. of Math. 48]. The finitely many ramified factors are holomorphic and nonzero for Re s > 0. So H = ζ(k₀ s)^{c₁} · (holomorphic nonzero) near 1/k₀, and ζ(k₀ s) ~ (k₀ s − 1)^{−1}. (s − a)^{−c} is single-valued and holomorphic at a only for c ∈ ℤ_{≤0}. ∎
**No Artin conjecture is used** in Theorems 1–4, for any group, 2I included: only Chebotarev and the behaviour of Artin L-functions on Re s = 1.

**Theorem 3 [project, proved; Jordan's lemma, known]. Dichotomy.** k₀ = ∞ (H ≡ 1) iff G ⊂ C_I (G cyclic and I its own axis). Otherwise 3 ≤ k₀ ≤ max(4, P(|G|)), where P(n) is the largest prime factor of n.
*Proof.* If G ⊂ C_I, every class is a single element of A and ψ is a character of G, so f_k = ψ^k = f_1(g^k) and Φ_d = 0 for d ≥ 2. Otherwise some class C misses A: for non-abelian G by Jordan's lemma (a proper subgroup never meets every conjugacy class; Jordan 1872) and for abelian G ⊄ C_I because then A ≠ G and classes are singletons. Take g ∈ C and let k₁ ≥ 2 be the least k with g^k in a class meeting A. Choose a prime q | k₁, odd if possible, and put h = g^{k₁/q}. Then h^e (1 ≤ e < q) is off-slice by minimality of k₁, and h^q is on it. If q is odd, Φ_q(h) = (1/q)[f_q(h) − f_1(h^q)] = −ψ(h^q)/q ≠ 0. If k₁ is a power of 2, take q = 2: then Φ₄(h) = (1/4)[f₄(h) − f₂(h²)] = −ψ(h²)²/4 ≠ 0. So k₀ ≤ max(4, q). Finally q divides the order of h (otherwise h is a power of h^q, hence in a slice class), so q | |G|. ∎

**Corollary (all axes) [project, proved].** For every finite G ⊂ H^× and every unit I with G ⊄ C_I,

    1/max(4, P(|G|)) ≤ σ_abs(G, I) ≤ 1/3,   i.e.   1/6 ≤ ½ − σ_abs ≤ ½ − 1/max(4, P(|G|)) < ½.

The upper bound on the distance is attained (Dic_p along a 4-fold axis, §3), so **sup over all (G, I) of ½ − σ_abs is ½, approached as p → ∞ but never attained**. For a fixed group the distance is bounded by ½ − 1/max(4, P(|G|)): by ¼ whenever |G| has no prime factor above 3 (Q8, 2T, 2O, Dic_n with n = 2^a 3^b), and by 3/10 for 2I (where every axis actually gives ¼ or ⅓, so that bound is not sharp).

**Useful criteria [project, proved, from the definition].** Φ₃(g) ≠ 0 iff either g is off-slice and g³ is on it (e.g. off-slice elements of order 3 or 6), or g is on the slice and ψ(g)³ ≠ ψ(g³) (a *fold*: cl g ∋ g⁻¹, and tripling the angle θ ∈ (0, π) crosses π, i.e. θ ∈ (π/3, 2π/3)). Off-slice elements of order 4 always give Φ₄ = −1/4 (since g² = −1 ∈ A).

## 3. Every family

### Standard axis (the longest cyclic direction; the convention of the page)

**Theorem 4 [project, proved; computer-assisted cross-check `survey.py`, `check_theorems.py`].**

| G | order | slice | k₀ | σ = σ_abs = σ_hol | distance ½ − σ | c₁ | why |
|---|---:|---|---:|---|---|---|---|
| cyclic C_m, own axis | m | C_m | ∞ | none (H ≡ 1) | — | — | Theorem 3 |
| Q8 = Dic₂ | 8 | C₄ | 3 | 1/3 | 1/6 | −i/6 | fold of ±i (θ = π/2) |
| Dic₃ | 12 | C₆ | 4 | **1/4** | **1/4** | −1/8 + i√3/24 | no fold (θ = π/3, 2π/3 triple to ±1), no off-slice order 3/6; off-slice j has order 4 |
| Dic_n, n ≥ 4 | 4n | C_{2n} | 3 | 1/3 | 1/6 | (i/3n) Σ_{n/3<k<2n/3} sin(3πk/n) → −2i/(9π) | fold at θ = πk/n ∈ (π/3, 2π/3) |
| 2T | 24 | C₆ | 4 | **1/4** | **1/4** | −1/16 | classes of the slice are not inverse-closed (no fold); only 4a off-slice |
| 2O | 48 | C₈ | 3 | 1/3 | 1/6 | −i/12 | off-slice C3, C6; fold of C4a |
| 2I | 120 | C₁₀ | 3 | 1/3 | 1/6 | −0.07837… i | off-slice C3, C6 |

*Proof of the Dic_n rows.* Dic_n = ⟨a, j⟩, a = e^{iπ/n}, j² = −1, j a j⁻¹ = a⁻¹. Classes: {1}, {−1}, {a^{±k}} (1 ≤ k ≤ n−1, ψ = e^{iπk/n}), and two classes of j a^k (order 4, j-cubes are off-slice, so they contribute nothing to Φ₃). For θ = πk/n, Φ₃ = (1/3)(e^{3iθ} − e^{i·fold(3θ)}) = (2i/3) sin 3θ when θ ∈ (π/3, 2π/3) and 0 otherwise. An integer k ∈ (n/3, 2n/3) exists iff n = 2 or n ≥ 4. Summing over the 2-element classes gives c₁ = (i/3n) Σ sin(3πk/n), a sum of negative terms, so c₁ ∉ ℝ and Theorem 2 applies. For n = 3, Φ₃ ≡ 0, the j-classes give Φ₄ = −1/4, the class of a² gives r₄ = 2i sin(8π/3) = i√3, so c₁ = (−6 + 2i√3)/48 ∉ ℤ. 2T, 2O, 2I: the class-by-class reasons in the table are finite checks by hand, matched by the script and by the closed forms of earlier notes. ∎

**Uniform bound, standard axis [project, proved].** For every finite G ⊂ H^× that is not cyclic,

    σ_G ∈ {1/4, 1/3},   so   1/6 ≤ ½ − σ_G ≤ 1/4,

with both edges also the true holomorphy edge (σ_hol = σ_abs). **C = 1/4 is sharp and attained** (Dic₃ and 2T); the lower value 1/6 is attained by Q8, every Dic_n with n ≥ 4, 2O and 2I. Along the dihedral family the edge is constant (1/3) from n = 4 on, so nothing is "only approached" here.

### Other axes [project, proved; checked by `survey.py` for Dic₂…Dic₄₀, 2T, 2O, 2I, several cyclic groups]

| G | axis | k₀ | c₁ |
|---|---|---:|---|
| Dic_n, n even | 4-fold axis of j | 3 | −i/6 (fold of j ~ j⁻¹) |
| Dic_n, n odd | 4-fold axis of j | **smallest prime factor of n** | 0 |
| Dic_n | generic | 3 if 3 ∣ n, else 4 | — |
| 2T | 4-fold (C₄) / generic | 3 / 3 | −i/6 / 0 |
| 2O | 3-fold (C₆) / 4-fold (C₄) / generic | 4 / 3 / 3 | −3/32 + 0.0722i / −i/6 / 0 |
| 2I | 3-fold (C₆) / 4-fold (C₄) / generic | 4 / 3 / 3 | −1/16 + 0.0722i / −i/6 / 0 |
| C_m off its axis | generic | e.g. p for C_p (p odd prime), 4 for C_{2^a} | — |

*Proof of the key row (Dic_n, n odd, j-axis).* The slice is {±1, ±j}; j and j⁻¹ = j aⁿ lie in different classes (n odd), so ψ is a character on the slice and the slice classes give Φ_d = 0 for d ≥ 2. Off the slice are the a^k (k ≠ 0, n), of order m | 2n. Their first slice power is m or m/2, an odd divisor of n greater than 1. With p the smallest prime factor of n, elements of order p and 2p give Φ_p = ∓1/p, and no smaller d ≥ 3 works. Φ_p(g) ≠ 0 exactly when g^p = ±1, i.e. g has order p or 2p, and there are φ(p) = φ(2p) = p − 1 elements of each order with opposite values, so c₁ = 0. ∎ So **σ_abs(Dic_p, j-axis) = 1/p → 0**.

## 4. What is open, and what the two definitions do differently

- [project conjecture / open] When c₁ = 0 (Dic_p on a j-axis, generic axes of 2T, 2O, 2I) Theorem 2 does not apply. H then continues across Re s = 1/k₀ near the real axis, but its half-plane of holomorphy depends on zeros of the L(χ, k₀ s), χ ≠ 1, with Re(k₀ s) < 1, which no known zero-free half-plane controls. So σ_hol is not determined there; only σ_hol ≤ 1/k₀ is proved. Under GRH and Artin for those characters one would expect σ_hol ≤ max(1/(2k₀), the next layer with nonzero trivial weight). Not pursued: it needs zero information, and none is claimed.
- [open] Whether H (or F) has a natural boundary further left (Estermann/Dahlquist-type behaviour for Euler products with non-integer exponents), the "natural boundaries" item of open-questions.
- **Scope.** The uniform bound concerns the correction factor H. The central exponents of F at ½ (3/8, 5/24, 1/24) are a separate matter. They depend on root numbers and central L-values of the specific field, and the 2I value is conditional on Artin's conjecture. This note does not touch them.
- Relation to β: β (`binary-polyhedral-barrier-20261112.md`) measures how the full Model A *norm* degenerates on the approach to Re s = ½; the edge measures where the leftover H stops converging, to the left of ½. Both are group-theoretic and field-independent, but they are different quantities.

## 5. Files

- `edge1114/edge.py`: groups (closure from generators; 2T, 2O, 2I, Dic_n, C_m), classes, slices, ψ, ghost layers Φ_d, k₀, c₁ (float, tolerance 10⁻⁹; nonzero values are all ≥ 0.05 in size).
- `edge1114/validate.py`: reproduces the earlier closed forms of H (2T, 2O, 2I, Q8).
- `edge1114/survey.py` → `survey.json`: every axis type of 2T, 2O, 2I, Dic₂…Dic₄₀ (main axis, both j-axis orbits, generic), C_m.
- `edge1114/check_theorems.py`: checks Theorem 3's bound, the Dic_n main-axis formula for k₀ and c₁ and the j-axis rule on every row (149 rows, 0 violations).
- `edge1114/site_figure.py` → `figures/uniform-edge-site.svg`, `-mobile.svg` and PNG previews.
