# Direct numerical check of F₁₀ for the 1951 / 2141 2I fields, including the constant c — 15 Nov 2026, ~8:45–9:30 PM PST (routine sprint, hypercomplex-critical-line-sprint)

Follows `icosa-exponent-20261113.md` (exact identity for F₁₀, the 1/24 theorem), `icosa-oddW-aligned-exponent-20261114.md` (1951/2141: m = (1,1,0), exponent 1/24), and `a5-central-values-20261115b.md` (A₅ central values, φ_A₅). Fields: LMFDB 2.3806401.120.a (p₀ = 1951) and 2.4583881.120.a (p₀ = 2141). Scripts and outputs: `f10direct1115/`. Figure: `figures/f10-direct-check.png` (+ `.svg`).

Labels: **[known]** published, with source; **[project]** computed or proved here; **[speculative]** unproved. All numbers are floating point (PARI at 160 bits with exact ℤ[φ] coefficients; double-precision theta sums; numpy/mpmath Euler products). **None is an interval enclosure.** Every value at or near s = ½ presupposes Artin's conjecture (entire completed L-function with the Artin functional equation) for ρ, ρ′, Sym²ρ, Sym²ρ′, ρ⊗ρ′, Sym³ρ, Sym⁴ρ of these even/odd icosahedral representations, which is **not known** [known: Booker, Experiment. Math. 15 (2006)]. This checks the branch of one product of L-functions at s = ½. It finds, moves or excludes no zero off the line; **no claim about the Riemann Hypothesis**.

## 0. What "direct" means here

F₁₀(s) = ∏_{p ≠ p₀, Frob_p in the 10-fold slice} (1 − ψ_p p^{−s})^{−1} (ψ_p = e^{iθ}, θ ∈ [0, π] read off a_p(ρ) = 2cos θ). Its Euler product converges only for Re s > 1. Near s = ½ the function exists only by continuation, and the only continuation available is the exact identity of `icosa-exponent-20261113.md` §2: F₁₀ = ∏_χ L(χ, s)^{a_χ} L(χ, 2s)^{b_χ} · H(s), H holomorphic and zero-free on Re s > 1/3. So the check has two parts:

1. **Re s > 1 (independent routes) [project].** Compare the truncated Euler product of F₁₀ itself (no L-function code) with the right-hand side built from L-values computed through their functional equations (PARI `lfun`, theta sums) plus H. The routes share only the Frobenius data.
2. **s = ½ + ε, ε = 10⁻¹ … 10⁻⁶ (11 points, half-decade steps) [project].** Evaluate every factor of the right-hand side at s and 2s, and H by its Euler product (absolutely convergent at Re s > 1/3), then fit log|F₁₀| and arg F₁₀ against log ε, and compute c two ways.

Nothing about F₁₀ near ½ is "measured" independently of the identity. What the check tests is that the identity holds globally (part 1), that each factor has the claimed behaviour at ½ (simple zeros of L(ρ), L(ρ′), nonzero L(Sym³ρ, ½), finite nonzero values everywhere else), and that no factor (H, the 2s-pieces, L(6)) does anything unexpected.

## 1. Result

**[project, floats; conditional on Artin holomorphy]** For both fields, F₁₀(½ + ε) ∼ c · ε^{1/24 + iγ}, γ = (s₁+s₂)/10 = 0.15388417686 (s_j = sin(jπ/5)), confirmed numerically:

| | 1951² | 2141² | predicted |
|---|---:|---:|---:|
| modulus slope, last interval (3.2·10⁻⁶ → 10⁻⁶) | 0.0416646753 | 0.0416618771 | 1/24 = 0.0416666667 |
| modulus slope, fit e·log ε + C + dε (6 smallest ε) | **0.0416666645** (max resid 3.3·10⁻⁹) | **0.0416663923** (4.1·10⁻⁷) | 0.0416666667 |
| phase slope, last interval | 0.1538877531 | 0.1538902643 | 0.1538841769 |
| phase slope, same fit | **0.1538841938** (2.6·10⁻⁸) | **0.1538842070** (4.6·10⁻⁸) | 0.1538841769 |
| \|c\| (symmetric branch, see §3) | **2.13255132** | **0.78493939** | — |
| arg c (rad, all branches) | **+0.0092020** | **+0.2233186** | — |
| \|c\| via upper / lower path around s = 1 | 5.6080 / 0.81094 | 2.0642 / 0.29849 | ratio e^{2π(s₁+s₂)/5} = 6.92 |

Local slopes approach 1/24 and γ linearly in ε (the residual drops by ≈ √10 per half-decade, figure panel (c)), as expected from the O(ε) corrections of smooth factors. The 2141 modulus fit is less tight because L(Sym³ρ, s) is small near ½ there (0.0381 at ½, minimum 0.0343 near σ ≈ 0.52) and varies fast, so the O(ε²) term matters more; its last local slope is still within 4.8·10⁻⁶ of 1/24 at ε ≈ 2·10⁻⁶.

## 2. Ingredients

- **Exact identity [project, proved 13 Nov, exact computer algebra in ℚ(ζ₆₀)].** Weights a_χ, b_χ read from `icosa-exp1113/derive-identity-results.json`. Re-checked here prime-free: log H_p computed as log F₁₀,p − Σ a_χ log L_p(χ, s) − Σ b_χ log L_p(χ, 2s) on each of the 9 classes is O(x³) (|log H_p|/x³ ≤ 0.39 at x = 10⁻², 10⁻³), and the character traces of my eigenvalue lists match the stored table.
- **ρ, ρ′ (W = −1, Γ_ℝ(s+δ)², δ = 0 / 1) and Sym²ρ, Sym²ρ′ (W = +1, Γ_ℝ(s)³, Euler factor (1 − p₀^{−s})^{−1})**: PARI `lfun` at 160 bits, exact ℤ[φ] coefficients from the oriented traces `cert1113/ap-oriented-*.json` (all p < 1.2·10⁵), 119 990 coefficients (PARI cost 3.6–4.1·10⁴ and 8.5–9.4·10⁴). `lfuncheckfeq` 2⁻¹⁷³, 2⁻¹⁷², 2⁻¹⁷⁵, 2⁻¹⁷⁵ (1951); 2⁻¹⁶⁶, 2⁻¹⁶¹, 2⁻¹⁷³, 2⁻¹⁷⁶ (2141). L′(ρ, ½) = 14.6404198416 / 8.3882156982 and L′(ρ′, ½) = 3.7050867341 / 6.1887816359 (agree with `icosa-cert1114/pari-cross-results.json` to all printed digits); L(3, ½), L(3′, ½) reproduce `a5-central-values-20261115b.md` to 20 digits.
- **Sym³ρ, ρ⊗ρ′, Sym⁴ρ (conductor p₀⁴)**: the double-precision theta/Mellin sums of `sym3-1114/` (3.5·10⁷ coefficients from the Frobenius orders; these three characters are rational, so no orientation is needed), now evaluated at all s = ½ + ε, 2s = 1 + 2ε and check points. Reproduce the 14 Nov central values to ≥ 11 digits; Dirichlet series vs theta formula at s = 2, 2.4, 3, 4 agree to ≤ 4·10⁻¹² for s ≤ 3 and ≤ 1.1·10⁻¹⁰ at s = 4; halving the cutoff changes L(½) by ≤ 1·10⁻¹⁰ (2141 Sym³ρ) or less.
- **L(6 = Sym⁵ρ, 2s)** (conductor p₀⁶ for 1951; too large for a functional-equation sum): enters F₁₀ only with the imaginary weight −i(s₂−s₁)/10, so it affects only the phase. Truncated Euler product over p ≤ 3.5·10⁷ (χ₆ is rational: 6, −6, 0, 0, 0, 1, 1, −1, −1 on 1, −1, C4, C3, C6, C5a, C5b, C10a, C10b). Euler factor at p₀ [project, from `decomp_p0.py`]: 1951 has D = I = C₁₀ (e = 10, f = 1 in the degree-24 stem field), so Sym⁵ρ is totally ramified, factor 1; 2141 has D = I = C₅, so ψ^{±5} are unramified and trivial, factor (1 − p₀^{−s})^{−2}. log L(6, 1) = −0.90732 (1951), +0.44286 (2141); partial products at 10⁶, 10⁷, 3.5·10⁷ move by ≤ 1.7·10⁻⁴, i.e. ≤ 6·10⁻⁶ rad in arg c.
- **ζ(s), ζ(2s)**: mpmath, 40 digits.
- **H(s)**: product over primes. p < 1.2·10⁵: exact oriented classes. 1.2·10⁵ < p ≤ 3.5·10⁷: Frobenius order only, so for orders 5 and 10 the two orientations (C5a/C5b, C10a/C10b) are averaged (they have equal Chebotarev density); the RMS of the half-differences is 2.1·10⁻⁷ in log H(½). Beyond 3.5·10⁷: Chebotarev mean m₃x³ + m₄x⁴ (m₃ = −0.0783 i, m₄ = −0.0692 − 0.0058 i), integrated with E₁; size 1.4·10⁻⁶ at s = ½. log H(½) = 0.0751047 − 0.0309936 i (1951), −0.0206940 − 0.1765590 i (2141). At p₀: H_{p₀} = 1/∏ L_{p₀}(χ)^{weights}, with the local factors above (ζ, 3, 3′, 5: (1 − p₀^{−s})^{−1}; ρ, ρ′, 4, 4′: 1).
- **Positivity on the real segment [project, floats]**: L(ρ, σ), L(ρ′, σ) (PARI, σ = 0.51 … 2.00, step 0.01) and L(Sym³ρ, σ) (theta sum, σ = 0.50 … 2.00) have no sign change; with `a5central1115/real-segment.json` (L(3), L(3′), L(4) on [½, 2]) every real-weight and imaginary-weight factor is real and positive on the path, so all logarithms are real there, **except ζ(s)** (§3).

## 3. The branch at s = 1 (new observation; affects |c|, not the exponent)

**[project, elementary]** The k = 1 layer contains ζ(s)^{i(s₁+s₂)/5}: the Chebotarev mean of the slice class function ψ is i(s₁+s₂)/5, so F₁₀(s) ≈ (s − 1)^{−i(s₁+s₂)/5} × (holomorphic, nonzero) near s = 1. F₁₀ therefore has a **logarithmic branch point at s = 1**, and its value on (½, 1) depends on how the continuation passes s = 1. Since ζ(σ) < 0 on (0, 1), log ζ(σ) = log|ζ(σ)| ∓ iπ for the path above / below s = 1, and ζ(σ)^{iw} = e^{iw log|ζ(σ)|} · e^{±πw}, w = (s₁+s₂)/5 = 0.30777. So:

- the exponent 1/24 + iγ is the same on every branch (the branch only multiplies by a constant);
- **arg c is the same on every branch**;
- **|c| is defined only up to the factor e^{±π(s₁+s₂)/5} = 2.630^{±1}**. The table reports the upper (+), lower (−) and symmetric (|ζ|, factor 1) values.

The 13 Nov note §6 already allowed for this ("|L^{ic}| = e^{−c·arg L} is locally constant"); it was not spelled out for c.

## 4. Re s > 1: identity vs Euler product

**[project, floats]** |log F₁₀(Euler) − log F₁₀(identity)|:

| σ | 1951 | 2141 | orientation-averaging RMS (expected size of the difference) |
|---|---:|---:|---:|
| 1.2 | 1.1·10⁻⁵ | 2.2·10⁻⁵ | 2.5·10⁻⁵ |
| 1.5 | 2.7·10⁻⁷ | 7.4·10⁻⁷ | 6.2·10⁻⁷ |
| 2.0 | 6.0·10⁻¹⁰ | 2.2·10⁻⁹ | 1.5·10⁻⁹ |

(F₁₀(1.2) = exp(0.27787 + 0.40472 i) for 1951.) The Euler product side uses oriented classes to 1.2·10⁵, orientation-averaged factors to 3.5·10⁷, and a Chebotarev-mean tail (size 2.2·10⁻³ at σ = 1.2, 5·10⁻¹⁰ at σ = 2). The agreement is at the level of the orientation noise in every case, for both fields and for both orientations.

## 5. The constant c

**Two routes [project, floats]:** (a) the closed-form limit

    c = L′(ρ,½)^{1/3} L′(ρ′,½)^{−1/6} L(Sym³ρ,½)^{1/6} · 2^{−1/8 + iγ} · |ζ(½)|^{iw} · ∏_{χ=3,3′,4} L(χ,½)^{a_χ} · ∏_χ L(χ,1)^{b_χ} · L(6,1)^{b₆} · H(½)

(the 2^{−1/8+iγ} comes from ζ(2s) = 1/(2ε) + O(1)); (b) c(ε) = F₁₀(½+ε) ε^{−1/24−iγ} on the grid, Richardson-extrapolated from ε = 3.2·10⁻⁶ and 10⁻⁶. They agree to 2.3·10⁻¹¹ (1951) and 1.7·10⁻¹⁰ (2141) relative. (They share the same L-functions at different points, so this is a consistency check, not an independent proof.)

**Phase budget of arg c (rad)**, orientation of the oriented-trace files (ρ has trace 2cos 72° on the "A" order-5 class), and the swapped orientation (ρ ↔ ρ′; F₁₀ itself changes):

| contribution | 1951 | 2141 | 1951 swapped | 2141 swapped |
|---|---:|---:|---:|---:|
| ζ(2s): 2^{iγ} | +0.10666 | +0.10666 | +0.10666 | +0.10666 |
| ζ(s): w·log\|ζ(½)\| | +0.11655 | +0.11655 | +0.11655 | +0.11655 |
| L(3), L(3′), L(4) at ½ (= φ_A₅ of 15 Nov) | **−0.40627** | **+0.26485** | −0.25110 | +0.02623 |
| L(3, 1), L(3′, 1) (imaginary parts of b) | −0.04461 | +0.00832 | +0.01567 | −0.09281 |
| L(4, 1), L(ρ, 1), L(ρ′, 1), L(Sym³ρ, 1), L(6, 1) | +0.26786 | −0.09650 | −0.07100 | −0.25302 |
| arg H(½) | −0.03099 | −0.17656 | −0.06702 | −0.12371 |
| **arg c** | **+0.00920** | **+0.22332** | **−0.15023** | **−0.22010** |
| \|c\| (symmetric) | 2.13255 | 0.78494 | 1.07281 | 0.67422 |

The φ_A₅ row reproduces `a5central1115/phase.json` to 10⁻¹² (an independent re-computation from the 160-bit values here). **[project]** φ_A₅ is one of six contributions and is not arg c: the trivial character ζ(s) (also an A₅ piece) adds +0.117, ζ(2s) +0.107, the 2s-pieces and H add the rest. In the file orientation arg c is close to 0 for 1951 (the large −0.406 is almost cancelled) and +0.223 for 2141. Swapping the orientation changes both |c| (L′(ρ) and L′(ρ′) trade weights 1/3 ↔ −1/6) and arg c, but not the exponent (fitted slopes for the swapped runs: 0.0416666584 / 0.1538841964 and 0.0416663956 / 0.1538842034).

**Error budget [project, estimates, not rigorous]:** orientation averaging of H (RMS 2.1·10⁻⁷), Chebotarev tail of H (1.4·10⁻⁶, included; its own error is smaller), L(6, 1) truncation (≤ 6·10⁻⁶ rad), theta-sum values (~10⁻¹⁰ relative). So |c| is good to about 2·10⁻⁶ relative and arg c to about 10⁻⁵ rad, at float level.

## 6. Corrections and clarifications

1. **[clarification of `a5-central-values-20261115b.md` §1/§3 and the page sentence "The A₅ pieces only turn the constant c by a fixed angle"]** True for L(3), L(3′), L(4), which is what that sentence was about. But φ_A₅ is not the phase of c (§5), and the A₅ piece ζ(s) does more than rotate: through the branch point at s = 1 it fixes |c| only up to e^{±π(s₁+s₂)/5}. The statement "F₁₀ ∼ c(s−½)^{1/24 + iγ}, c ≠ 0, ∞" stands on every branch.
2. **[update of `icosa-oddW-aligned-exponent-20261114.md` §7]** "a numerical check of the product F₁₀ itself for these fields (needs oriented Frobenius classes)": done here with oriented classes to 1.2·10⁵ and orientation averaging beyond (the averaging error is measured, §4–5).
3. No earlier number changed. L′(ρ, ½), L′(ρ′, ½), L(Sym³ρ, ½), L(4, ½), L(5, ½), L(3, ½), L(3′, ½) all reproduce.

## 7. What was and was not verified

- **Verified [project, floats]:** the exact identity against the direct Euler product at Re s = 1.2, 1.5, 2 (both fields, both orientations); F₁₀(½ + ε) on 11 points down to ε = 10⁻⁶ with slopes 1/24 (to 2·10⁻⁹ / 3·10⁻⁷ by fit) and γ (to 2·10⁻⁸ / 4·10⁻⁸); |c| and arg c by two routes; positivity of every real factor on the real segment.
- **Not verified:** anything by interval arithmetic; Artin holomorphy (assumed); L(6, 1) by a functional equation (only a truncated Euler product, phase-only effect ≤ 6·10⁻⁶ rad); the H tail beyond 3.5·10⁷ (Chebotarev-mean estimate); the orientation-averaged primes individually. F₁₀ near ½ is never computed without the identity, because no such route exists (its Euler product diverges there).
- **Not claimed:** anything about zeros off s = ½, GRH, or RH.

## 8. Files (`f10direct1115/`)

`f10_pari.py` → `pari-<label>-b160.json` (ρ, ρ′, 3, 3′ at all points), `f10_theta.py LABEL {4p,4,5}` → `theta-<rep>-<label>.json`, `decomp_p0.py` → `decomp-p0.json`, `f10_assemble.py LABEL [swap]` → `assemble-<label>[-swap].json` (Re s > 1 check, rows, fits, c, phase budget), `seg_pari.py`, `seg_theta.py` → `seg-rho-*.json`, `seg-4p-*.json`, `summary.json`, `fig.py` → `../figures/f10-direct-check.png/.svg`. Logs `*.log`. Environment `/workspace/work/venv-pari` (cypari2, mpmath, numpy); figure with `/usr/bin/python3`. Run time: theta sums 80 s each, PARI ≈ 2 min per field, assembly ≈ 6 min per field.
