# Q8 Frobenius-quaternion Euler products at complex s

Sprint run, 11 November 2026, ~12:30 PM PST. Open-questions item 5 ("any extension to complex s"). Follows `frobenius-quaternions.md` (real s > 1 only). Everything here is in the region of absolute convergence (Re s = 2 and 3/2). Nothing here concerns the critical line, zeros, or RH.

Labels: [known] established mathematics; [project] checked or proved here, not new mathematics in any deep sense; [speculative] unproved.

## Setup

[known] K/Q Galois with group Q8; rho the faithful 2-dimensional representation; Frob_p is a Q8 element up to conjugacy, classes {1}, {−1}, {±i}, {±j}, {±k}. Artin: L(rho,s) = ∏ det(1 − rho(Frob_p) p^{−s})^{−1}. For a class {±u} (u in {i,j,k}) the prime p has residue degree 4 and splits in exactly one quadratic subfield; identifying Gal(K/Q) with the unit group {±1,±i,±j,±k} ⊂ H labels those three subfields by i, j, k (a labeling choice: automorphisms of Q8 permute i, j, k).

To write p^{−s} for complex s inside H we must say what "the complex unit" of s is. Two honest options:

- **Model A (unit inside H).** s = σ + tI with I a unit imaginary quaternion, so p^{−s} = p^{−σ}(cos(t log p) − I sin(t log p)) lies in the slice C_I ⊂ H. E_P(s) = ordered ∏ (1 − g_p p^{−s})^{−1}, g_p a chosen element of the Frobenius class.
- **Model B (central unit).** Work in H ⊗_R C ≅ M2(C), where the complex unit commutes with everything; g_p acts by its 2×2 matrix rho(g_p).

## Results

1. [project — exact elementary proof; checked to 30 digits at 200 random points, max error < 1e-25] For g = cos θ + u sin θ (u unit imaginary) and x = r(cos φ − I sin φ),

       nrd(1 − g x) = 1 − 2r( cos θ cos φ + sin θ sin φ (u·I) ) + r².

   Only the component u·I of the Frobenius axis along the chosen complex unit enters. Compare det(1 − rho(g)z) with z = r e^{−iφ}: its modulus is |1 − e^{iθ}z| · |1 − e^{−iθ}z|. So:
   - classes ±1 (θ = 0, π): nrd = |1 ∓ z|², matching |det| exactly;
   - the class aligned with I (u = ±I, θ = π/2): nrd = |1 ∓ iz|², i.e. ONE of the two eigenvalue factors of |1 + z²| = |1 − iz||1 + iz|, squared, and which one depends on the sign choice ±u;
   - the classes orthogonal to I (u·I = 0): nrd = 1 + r², with **no t-dependence at all**: those primes stop oscillating.
   At t = 0 (real s) all three cases reduce to det(1 − rho(g)x), recovering the earlier real-s theorem.

2. [project — consequence] In Model A, nrd E_P(s) is still independent of the ORDER of the factors (nrd is multiplicative; checked to 1e-11 for three orders at several t), but it is no longer independent of the representative CHOICES once t ≠ 0, and it is no longer |L_P(rho,s)|. The choice of complex unit I = i, j or k singles out one quadratic subfield; a generic I (we used (i+j+k)/√3) makes every degree-4 prime choice-sensitive. Choosing the sign of g_p in the aligned class amounts to choosing one of the two primes of the aligned quadratic field above p (on which the order-4 character takes values ±i); no such choice is canonical. Neither averaging over signs nor over I ∈ {i,j,k} restores |1 + z²| for every class (the orthogonal classes always give 1 + r²).

3. [project — floating point, not intervals] Primes < 20,000 (2,260 unramified per field), σ = 2 and 3/2, t from 0 to 40 in steps of 1/4, six sign-choice sets (all +, all −, four random):
   - Model A deviates from log|L_P(rho,s)| by up to 0.033–0.087 (2304 field) and 0.17–0.54 (42025 field) at σ = 2, and up to 0.12–0.24 and 0.38–0.85 at σ = 3/2, depending on I. The spread across sign choices reaches 0.06–0.18 and 0.05–1.03 at σ = 2, and 0.23–0.51 and 0.20–1.54 at σ = 3/2 (natural-log units; 1.54 is a factor of about 4.7 in the norm). At t = 0 every variant agrees with log L_P to ≤ 5e-15.
   - In the 42025 field, the k-label is the subfield Q(√41), where 2 splits, so with I = k the prime 2 is choice-sensitive and dominates the curve with period close to 2π/log 2 ≈ 9.06 (visible in the figure). With I = j (Q(√5), where 2 is inert) the choice spread is small (0.049 at σ = 2).
   - Model B: log det E_P agrees with log L_P(rho,s) to ≤ 1.7e-14 and across all sign/order choices to ≤ 7e-16. Against PARI's full lfun at t = 0, 10, 25, 40, the gaps (truncation at 20,000) are ≤ 1.2e-7 at σ = 2 and ≤ 2e-5 at σ = 3/2. The TRACE of E_P is not choice-free (spreads up to 0.0012/0.014 at σ = 2 and 0.018/0.085 at σ = 3/2). So among the conjugation invariants of a 2×2 product, only the determinant survives: this is [known] Artin's construction, recovered here as the unique choice-free invariant in these experiments.

4. [project — interpretation] This closes the "complex s" half of item 5 in the negative for Model A and in the classical direction for Model B. If the variable's imaginary unit lives inside the same quaternions as the Frobenius elements, the arithmetic is broken in a specific, small way: the norm sees only the projection of each Frobenius axis onto that unit, so one quadratic subfield is privileged, the other two lose their oscillation, and the result depends on non-canonical choices of primes. Making the complex unit central (complexifying H to 2×2 complex matrices) removes the problem and gives back exactly the Artin L-function, and nothing more that is choice-free. [speculative] This is a concrete reason why a "quaternion-valued s" in a nonabelian Euler product does not carry new arithmetic information; we do not claim that no other construction could.

## Not claimed

No zeros, no continuation, nothing in the critical strip, no RH information. The Artin L-functions themselves are unchanged. Numerics are double-precision floats, not interval certificates; the closed form in result 1 is an exact identity.

## Files

`complexs1111/complex_s_frobenius.py` (run with /workspace/tools/pari-venv/bin/python3; 1 min 47 s; 4,378 checks passed), `complexs1111/complex-s-results.json`, `complexs1111/run.log`, `complexs1111/fig.py`, figure `figures/q8-frobenius-complex-s.svg/.png`.

## Page status

Not folded into /critical-line-hypercomplex in this run: a live session was editing build.py, the gallery and the zeros-and-primes page at 12:06–12:25 PM PST, so this run did research and notes only. Ready to fold in as a short section (one figure, the closed form, and result 4) on the next run when the build is idle.
