# Quaternion Euler products from actual Q8 Frobenius elements

Sprint run, November 10, 2026, 12:30 PM PT. Open-questions item 5 (a motivated multi-direction coefficient system). No RH claim.

Labels: [known] established mathematics; [project] checked here, not new mathematics; [speculative] unproved.

## The object

[known] For a Galois field K/Q with group Q8 = {±1, ±i, ±j, ±k}, the faithful two-dimensional representation rho is Q8 acting on H = C². At an unramified prime p, Frob_p is an element of Q8 defined only up to conjugacy (the choice of a prime of K above p). The Artin L-function uses det(1 − rho(Frob_p) p^{-s})^{-1}, which only sees the class. Classes: {1} (p splits completely, trace 2), {−1} (residue degree 2, trace −2), {±i}, {±j}, {±k} (residue degree 4, trace 0). The letter i/j/k is fixed by which quadratic subfield p splits in (Frob lies in the index-2 subgroup ⟨i⟩ exactly when p splits in its fixed field).

[project] So the coefficients are literally quaternions, the most natural "multi-direction coefficient system" available: no invented convention. Define the ORDERED product, for real s > 1,

    E_P(s) = ∏_{p ≤ P, p unramified, in a chosen order} (1 − g_p p^{-s})^{-1},   g_p a chosen element of the class of Frob_p.

## Results

1. [project — exact elementary proof] For real s, each factor 1 − g x (|g| = 1, x real) has reduced norm 1 − 2Re(g)x + x² = det(1 − rho(g)x). Reduced norm is multiplicative in H, so
   nrd E_P(s) = |E_P(s)|² = ∏ det(1 − rho(Frob_p)p^{-s})^{-1} = L_P(rho, s),
   the partial Artin Euler product, for EVERY choice of representatives and EVERY order. As P → ∞ (absolute convergence for s > 1), |E(s)|² = L(rho, s). The quaternion's length is arithmetic; nothing else is guaranteed to be.
2. [project — exact rational check] For primes below 200 at s = 2, six variants (increasing order, decreasing order, four random orders with random ± signs inside the f = 4 classes) give six DIFFERENT exact rational quaternions with identical exact reduced norm equal to the rational local product. Both fields: 2.2304.8t5.b.a (polynomial x^8+12x^6+36x^4+36x^2+9) and 2.42025.8t5.a.a.
3. [project — floating point, not intervals] Primes below 20,000 (2,260 unramified primes per field; Frobenius counts 271/279/571/568/571 and 268/281/567/572/572 for 1/−1/i/j/k, close to Chebotarev's 1:1:2:2:2). Every residue-degree assignment matched PARI's a_p for p < 3000; each f = 4 prime split in exactly one quadratic subfield (8,821 checks total).
   - |E_P|² across 202 choices agrees to 3e-14 (float noise). Gap to PARI's full L(rho, s): 4.8e-10 and 6.1e-9 at s = 2; 4.2e-7 and 8.8e-7 at s = 3/2, consistent with the omitted primes.
   - Direction is NOT arithmetic: versus the canonical choice (increasing primes, + signs), random choices move the unit quaternion by up to 6.0° (2304) and 32.7° (42025) at s = 2, and 17.6° and 53.4° at s = 3/2. Merely reversing the order moves it 0.13°/3.9° (s = 2) and 0.89°/10.6° (s = 3/2). Even the real part fraction Re E/|E| moves (0.9986–0.9993 and 0.960–0.968 at s = 2; 0.990–0.998 and 0.900–0.944 at s = 3/2).
4. [project — interpretation] This is the precise "tiny way" noncommutativity enters genuine nonabelian arithmetic: the quaternion product carries extra data, but that data depends on non-canonical choices (primes above p, order of primes). The only choice-free part is the norm, which is exactly the classical Artin L-value. This explains at the level of formulas why Artin L-functions are built from determinants/traces. Compare the toy counterexample in angles.md (a_2 = i, a_3 = j), which had no arithmetic meaning; here the noncommuting coefficients are actual Frobenius elements.
5. [known/project — limits] Only real s > 1 is treated. For complex s, p^{-s} lies in the complex slice of the variable's imaginary unit and does not commute with g_p, so nrd(1 − g p^{-s}) is no longer det(1 − rho(g)p^{-s}); this formula does not extend to the critical strip, and nothing here concerns the critical line, ordinary zeta, or RH. A global conjugation h E h^{-1} preserves everything; individual representative changes are not global conjugations, which is why the direction moves.

## Files

q8-frobenius-quaternions.py, q8-frobenius-quaternions-results.json, q8-frobenius-quaternions-run.log, q8-frobenius-figures.py, figures/q8-frobenius-directions.svg/.png.
