# Off-axis quaternion-order zeros: a complete finite interval map

[project] Latest extension, November 10, 2026, 01:48 PM PST: complete10000 map and literal Turing proof, counts47188/48478/55806 and certified inherited/local vertical gaps. See [order-height10000.md](order-height10000.md); original30 and100/1000 evidence retained.

## Extension changelog: November10,2026, PST

[project] The complete maps now reach **100 and1000**, using fresh Hardy-Z
sign brackets plus a **literal Turing integral bound**, not a larger version
of the old contour. At100 the positive totals for D2,D3,D30 are180,192,265;
at1000 they are3254,3383,4115. All nonreal roots are simple; centre orders0,0,2.
The split at1000 is1517 inherited on each outer line, plus220/349/1081 local
middle-line roots. Details, proof, checkpoints, timings and the unexecuted10k
plan: `order-height.md`; dataset `order-height/summary.json`, full replay
`order-height/verification.json`. Figures `figures/order-height-growth.svg`
and `figures/order-height-density.svg`, with mobile SVG/PNG variants.

[known] This is exact transport of ordinary-zeta finite verification, not new
RH information. Platt–Trudgian already verified ordinary RH to about3×10^12.
The original height30 record below is retained as a separate historical
argument-principle certificate; its hashes and narrower brackets are unchanged.

---


[project] November 10, 2026, PST. This closes the **finite off-axis part of
open item 6 to height 30**, for maximal-order ideal-index zetas of discriminants
D=2,3,30. It replaces the earlier 48 floating-point samples with a complete
finite classification. It is **not** a new zero theorem or RH progress.

## 1. What was actually certified

Positive count means every nonreal zero with **0 < Im(s) ≤ 30**, with
multiplicity, anywhere in the complex plane. The centre is not included in that
positive count. The quarter and three-quarter locations below are certified in
this finite window, not asserted for all heights.

| D | Re(s)=1/4 | Re(s)=1/2, nonreal | Re(s)=3/4 | Positive total | Centre order | Full symmetric contour |
|---|---:|---:|---:|---:|---:|---:|
| 2 | 13 | 6 | 13 | **32** | 0 | **64** |
| 3 | 13 | 10 | 13 | **36** | 0 | **72** |
| 30 | 13 | 31 | 13 | **57** | 2 | **116** |

[project] Every nonreal zero in the table is simple. For D=30, the extra
real central zero is double. No other nonreal zeros occur through height30.
The local positive counts for p=2,3,5 separately are **6,10,15**.

[project] Figure: `figures/order-off-axis-map.svg` and `.png`.
A separately designed `order-off-axis-map-mobile.svg`/`.png` shows D=30 with
readable direct labels at 360px. Markers are much larger than the zero brackets.

## 2. Identify the object before transporting a zero

[known] Voight, *Quaternion Algebras*, Theorem26.4.9, equation26.4.10:

    Z_D(s) = ζ(2s) ζ(2s−1) ∏_{p|D} (1 − p^(1−2s)).

This is the **ideal-index** zeta of a maximal order in a rational quaternion
algebra, not the Q8 Artin L-function and not an intrinsic slice extension.
The earlier project checked actual maximal-order bases for D=2 and3.
D=30 is a definite rational algebra ramified at2,3,5 and infinity; its
order-zeta formula is independent of a particular maximal-order basis.

[known] In the reduced-norm variable **w=2s**, the same divisor is renamed:

    W_D(w) = ζ(w) ζ(w−1) ∏_{p|D} (1 − p^(1−w)).

The displayed three lines then have real parts1/2,1,3/2, and the height
cutoff becomes60. Neither the change of units nor the word quaternionic
creates a different ordinary-zero statement.

Primary source, retrieved before this run:
https://jvoight.github.io/quat-book.pdf
(read-only local copy in `literature/sources/voight-quaternion-algebras.pdf`;
text at26.4.9 was inspected directly). Completion conventions are discussed
in §26.8; we do **not** use a completed-order functional equation to count here.

## 3. The exact divisor map — and why no cancellation was overlooked

[known] Write an arbitrary ordinary nontrivial zero as ρ=β+iγ, of order m.
The order zeta inherits zeros at

    s = ρ/2                 and                 s = (ρ+1)/2,

both of order m. The first lies in0<Re(s)<1/2 and the second in
1/2<Re(s)<1. At the first, the other zeta factor is ζ(ρ−1), in the
nonreal negative half-plane and nonzero by the functional equation and
Euler-product zero exclusion. At the second it is ζ(ρ+1), in Re>1
and nonzero by the Euler product. The local factors vanish only at Re(s)=1/2,
so cannot cancel or add to either inherited zero. There are no poles at
these nonreal points.

[known] Each ramified prime contributes exactly

    s = 1/2 + iπk/log(p),        k a nonzero integer.

Its derivative is2log(p) at the zero. At these points the two zeta factors
are on Re=1 and Re=0 and are nonzero for nonzero height, by the classical
zero-free boundary line and functional equation. Families for two distinct
primes cannot meet: k/log(p)=l/log(q) would give q^k=p^l, contradicting
unique prime factorization for positive k,l. Thus every such local zero is
simple. The finite run **also** interval-encloses the complementary factors
at every listed point, rather than relying only on those general theorems.

[known] At the centre, put h=s−1/2 and r=# {p:p|D}. Laurent expansion gives

    Z_D(1/2+h) = −2^(r−2) (∏ log(p)) h^(r−1) + O(h^r).

Hence centre order r−1: zero for neither D2 norD3, double forD30.
This is cancellation of a classical pole by local zeros, not odd Artin
root-sign forcing. The only pole at s=1 lies outside our order contour.
All real trivial zeros are outside that contour and were handled in
`order-real-axis.md`.

[known] These observations classify the entire nonreal divisor symbolically
without RH: inherited ordinary zeros plus explicit local roots, and nothing
else. The assertion that **all inherited roots**, at all heights, lie on the
quarter/three-quarter lines is equivalent to ordinary RH. It is not a new
quaternionic RH. There is no cancellation loophole in that equivalence.

## 4. First complete certificate: ordinary ξ through height60

[project] We openly evaluate ordinary ζ in this run. Unlike the independent
subfield projector, no methodological independence from ordinary ζ is claimed.
The entire function is

    ξ(s) = [s(s−1)/2] π^(−s/2) Γ(s/2) ζ(s).

Arb384-bit evaluations enclose the image of **every full segment**, not
merely a mesh of endpoints, on the counterclockwise rectangle

    −1/2 ≤ Re(s) ≤ 3/2,       −60 ≤ Im(s) ≤ 60.

There are4150 accepted image boxes. Each is a convex complex rectangle
excluding0. An endpoint ratio with strictly positive real part isolates
the principal argument increment. Convex zero exclusion puts the entire
segment image in a half-plane through0, so no hidden full turn can occur.
The sum of rigorous increments divided by2π encloses exactly the integer
**26**. The argument principle applies to the entire ξ; apparent formula
singularities at0 and1 are removable and not on the contour.

[project] Thirteen numerical `mpmath.zetazero` outputs choose exact rational
bracket centres. They do **not** certify root existence, simplicity, numbering
or completeness. Fresh Arb values of ξ(1/2+it) have opposite real signs
at endpoints of thirteen disjoint brackets of width **2×10^−20**. The
functional equation and conjugation make that restriction exactly real.
Consequently each bracket contains a zero. Their thirteen conjugates and
the global count26 exhaust the entire window: all are simple, and there
is no room for an extra off-line or repeated zero. This is an argument-
principle certificate, **not literal Turing averaging**.

[project] Dividing the ordinate brackets by2 gives thirteen certified
heights on each inherited line through30, each width **10^−20**. Arb also
excludes0 from the complementary factor products in all **78** cases
(13 roots ×2 inherited families ×3 discriminants).

## 5. Second contour: directly wind the order zetas

[project] The raw product Z_D itself is interval-evaluated simultaneously
on the rectangle

    1/8 ≤ Re(s) ≤ 7/8,       −30 ≤ Im(s) ≤ 30.

The centre has already been proved removable, of order r−1. The pole at1
and all negative real trivial zeros are outside. Thus winding counts zeros,
not zeros minus an unaccounted pole. There are29090 accepted common segments.
The enclosed windings are **64,72,116**. For example,

    116 = 2 × (13 + 31 + 13) + 2.

This agrees exactly with the inherited-plus-local classification. The narrow
order rectangle alone would not exclude inherited zeros close to Re(s)=0
or1; the full ordinary-ξ contour and exact divisor map supply that missing
step. Hence the final nonreal classification really is for the full plane,
not merely the narrower drawn rectangle.

[project] The inequalities

    m_p < 30 log(p)/π < m_p+1

are verified as strict Arb inequalities, yielding m_2=6,m_3=10,m_5=15.
All31 recorded local-height intervals are pairwise disjoint. At each height
all applicable complementary products exclude0 (**47** D-specific checks).
The exact local formulas establish existence; the numerical residual alone
is not used to declare a zero.

## 6. Reproducibility, verifier and a caught interval pitfall

[project] Runtime: python-flint0.9.0, generator precision384 bits, verifier
precision512 bits. Seed-only mpmath1.3.0 was added to the existing PARI venv.
The trusted analytic backend is Arb's zeta, gamma, exponential and ball
arithmetic. Their rigorous internal tails are delegated to that library;
we did not independently implement or formally verify those algorithms.
A primary description is Fredrik Johansson, *Arb: efficient arbitrary-precision
midpoint-radius interval arithmetic*, https://arxiv.org/abs/1611.02831.

[project] The first replay correctly stopped: at segment16 a higher-precision
Arb interval extension was **wider**, not narrower. Increasing precision can
change an analytic expansion and its wrapping over a non-point input. We did
not waive zero exclusion. A regression test reproduces this, and the verifier
now subdivides an uncertain original segment and encloses every subpiece.
It replays all **33240** archived segments as **40702** validated512-bit
leaves, with7462 additional subdivisions. Every endpoint argument and every
claimed winding is recomputed. It also checks exact rectangular coverage,
orientation, closure, sign brackets, all complement checks, and strict local
cutoffs. Deliberately corrupted geometry is rejected.

[project] Controls were run red-before-green: polynomial windings+1,+2,
rational winding−1, boundary-zero refusal, expected finite counts, and the
precision-change replay regression. The verifier is an independently invoked
replay, **not** a separately implemented special-functions backend.

Artifacts:
- `order-off-axis.py`: generator;
- `order_contour.py`: full-image enclosure and winding engine;
- `order_zeta.py`: known factorization/evaluators, shared explicitly;
- `order-off-axis-results.json`: complete47,818,588-byte certificate;
- `order-off-axis-results.json.gz`: byte-identical compressed publication copy,
  15,379,253 bytes, below the site's per-asset limit;
- `order-off-axis-summary.json`: counts, local data and full-certificate hash;
- `verify-order-off-axis.py`, `order-off-axis-verification.json`:512-bit replay;
- `test_order_contour.py`, `test_order_off_axis.py`, `test_order_replay.py`;
- generation, replay and red-test logs retained beside these files;
- `order-off-axis-figures.py`, four SVG/PNG outputs.

The uncompressed certificate SHA256 is
`96a6456f91953d3fec1121384fcfff9bd0fc390d8caabae837de90d70c3163ac`.
The site publishes the gzip and summary, not the oversized uncompressed asset.
The verifier reads the gzip if the uncompressed file is absent.

From this directory:

```bash
/workspace/tools/pari-venv/bin/python test_order_contour.py
/workspace/tools/pari-venv/bin/python order-off-axis.py
/workspace/tools/pari-venv/bin/python test_order_off_axis.py
/workspace/tools/pari-venv/bin/python test_order_replay.py
/workspace/tools/pari-venv/bin/python verify-order-off-axis.py
python3 order-off-axis-figures.py
```

The public verifier needs the certificate gzip plus `order_contour.py`,
`order_zeta.py`, and the generator file (its source hash is checked), but it
never calls mpmath or any zero oracle. Re-generation needs mpmath for seeds.

## 7. Verdict and stopping boundary

[project] We now have complete interval-certified maps, not suggestive samples,
for the requested finite off-axis window. The full count checks a real potential
failure mode: missed zeros or wrong centre multiplicities. It matches exactly.

[known] It does not establish a new quaternionic mechanism constraining ordinary
ζ. The order zeta's inherited component is explicitly made from ζ. The middle
line is a different, elementary local-factor family; its abundance is not
evidence for ordinary RH. Certifying more heights would extend the same finite
check, not upgrade this dependence into a downward proof.

[project] Highest-value next arithmetic check remains an independent Artin
coefficient audit past the previously audited2000/5000 terms. Literal
Turing averaging to100 is a distinct unimplemented item, not completed here.
