# Independent subfield computation and recovery of ordinary-zeta zeros

[project] 10 November 2026, PST. Requested independent check; no deployment,
external message, routine change, or RH-progress claim.

## 1. What was certified

[project — interval certificate] We evaluated the zetas of the three quadratic
fields F5=Q(√5), F41=Q(√41), F205=Q(√205), and their biquadratic compositum
E=Q(√5,√41) **without evaluating ordinary ζ or any Dirichlet/Artin L-factor**.
Each calculation starts from the field's own prime-ideal norms, its discriminant,
and certified class-number/fundamental-unit data. The numerical method is a
Gaussian smoothed approximate functional equation, not the earlier factored
Hurwitz/Bessel evaluation.

The recovered results agree with the known ordinary data:

| target / statistic | ζ_F5 | ζ_F41 | ζ_F205 | ζ_E | recovered ordinary ζ | known ordinary ζ |
|---|---:|---:|---:|---:|---:|---:|
| order at 1/2 | 0 | 0 | 0 | 0 | **0** | 0 |
| order at first zero, height ≈14.13472514173469 | 1 | 1 | 1 | 1 | **1** | 1 |
| order at second zero, height ≈21.02203963877155 | 1 | 1 | 1 | 1 | **1** | 1 |
| order at third zero, height ≈25.01085758014569 | 1 | 1 | 1 | 1 | **1** | 1 |
| positive-height count N(30) | **14** | **24** | **31** | **63** | **3** | 3 |

Here 0 means nonzero, 1 means a simple zero; the last row is a count, not a
pointwise order. Positive-height counts include multiplicities and all real
parts in the critical strip, not merely zeros sampled on the critical line.

[project — exact extraction] The projector has an indispensable factor 1/2:

    2 ord ζ = ord ζ_F5 + ord ζ_F41 + ord ζ_F205 − ord ζ_E,
    2 N_ζ(30) = N_F5(30)+N_F41(30)+N_F205(30)−N_E(30).

Thus the independent whole-window recovery is

    N_ζ(30) = (14+24+31−63)/2 = 3.

The ordinary function is not numerically evaluated to obtain these numbers.
Only after the independent certificate and its verifier succeed is the saved
ordinary reference data opened for comparison. The new data also certify no
real field-zeta zeros in −1/2≤s≤3/2, which is necessary before halving a
conjugate-symmetric contour count to obtain a positive-height count.

**Figure:** `figures/independent-projector-match.png` and `.svg`.

## 2. Exactly what “independent” means

[project — pipeline independence] The three numerical generator stages consume:

1. The four defining polynomials, exact prime decompositions and certified units.
2. Their own finite prime-ideal Euler products on Re(z)=8.
3. Their own functional equations, gamma factors and explicitly enclosed residues.
4. Fixed rational target coordinates supplied by the question.

There are **no calls** to Riemann-zeta evaluation, Dirichlet-L evaluation, Artin-L
evaluation, PARI lfun/zero finding, or an earlier factor certificate. There is no
multiplication of ζ×L(χ) pieces. A static call audit checks the forbidden evaluator
names; the input/hash chain and source are available for inspection. This is
stronger than rerunning the old factors at different precision.

[project — methodological independence] The previous certificates used ordinary
ξ and quadratic Hurwitz sums, together with a K0 theta integral for the Artin
factor. This one uses a Gaussian contour formula, pole residues from units,
prime-ideal Euler products and a different tail argument. It shares the trusted
Arb/PARI libraries and standard field-zeta functional equations with earlier
work; it is not an independent implementation of those libraries.

[known identity; project scope] Mathematical information is **not** independent:
for example a quadratic field's coefficients equal the familiar convolution
1*χ, and the extraction identity is a consequence of the same known Galois
factorizations. An alternate computation does not create new constraints on RH.
The independence is of input provenance and the evaluation/proof pipeline, not
of the underlying functions or of the exact algebraic theorem.

[project — prime cutoff versus series cutoff] For speed we use the finite
Euler product over prime ideals above rational primes p≤30000, retaining all
powers of these primes. This is an approximation to the field's **own**
ideal-count Dirichlet series. It is not a Dirichlet polynomial truncated at
n=30000; its coefficients agree with the true field coefficients for n≤30000,
and omitted coefficients first occur at n>30000. The tail proof below accounts
for every omitted ideal, including products involving a larger prime.

The finite Euler product is **never substituted at Re(s)=1/2** as an
unjustified approximation. It is evaluated on an absolutely convergent line,
and the smoothed functional equation transfers the result to the target.

## 3. Arithmetic inputs and coefficient cross-checks

[project — exact arithmetic] The fields are defined independently by

| field | polynomial | discriminant | degree | class number h |
|---|---|---:|---:|---:|
| F5 | x²−x−1 | 5 | 2 | 1 |
| F41 | x²−x−10 | 41 | 2 | 1 |
| F205 | x²−x−51 | 205 | 2 | 2 |
| E | x⁴−92x²+1296 | 42025 | 4 | 1 |

For each rational prime p≤30000, `idealprimedec` gives ramification indices e
and residue degrees f. We check Σ e f=[F:Q]. The local ideal-count factor is

    ∏_(prime ideals pfrak over p) (1−p^(−f(pfrak)s))^(−1).

The e's determine the degree check but are **not** exponents in this Euler
factor. Bad primes are included by their actual prime-ideal norms, not removed
or guessed. All 3245 rational primes up to the cutoff are represented. Grouping
identical local norm patterns in the numerical loop is only an optimization;
no group is identified or evaluated as an ordinary or character L-function.

[project — 4000 coefficient checks] A separate integer Dirichlet-convolution
sieve expands these ideal-norm factors for n=1,…,1000 in each field. All 4000
coefficients agree exactly with PARI `dirzetak`. The saved-witness verifier also
recomputes them by multiplicativity and local binomial coefficients, a different
integer algorithm. These checks establish arithmetic consistency, not independent
reimplementation of PARI's underlying number-field routines.

[project — residue certification] Let d=[F:Q], D=disc(F), and

    A_F = √D / π^(d/2),
    Λ_F(s) = A_F^s Γ(s/2)^d ζ_F(s),
    X_F(s) = s(s−1)Λ_F(s).

All four fields are totally real. Standard analytic class-number and functional-
equation formulas give poles of Λ_F at 1 and 0 with residues R_F and −R_F, where

    R_F = 2^(d−1) h_F Reg_F.

`bnfcertify` returns 1 for every field, certifying the class/unit data without
leaving a GRH assumption. Fundamental units are stored as exact rational
polynomials in the defining generator. Arb evaluates their absolute logarithms
at explicit real embeddings and takes a (d−1)×(d−1) determinant to enclose the
regulator. No L-value is used to determine R_F.

For the quadratic fields we use roots (1±√D)/2. For E, the four roots are
±√5±√41. The sum of each unit's embedding logarithms contains zero, and the
regulator determinant is bounded strictly away from zero. The completed
residues are enclosed near

    R_F5   =   0.962423650119206895,
    R_F41  =   8.318254269252360026,
    R_F205 =  15.042635374130202093,
    R_E    = 120.426594906145740137.

The JSON stores the full interval balls and exact unit polynomials, not just
these rounded summaries. Standard field theory identifies E with the compositum;
its defining generator can be chosen as √5+√41.

## 4. The Gaussian formula, including the pole terms

[known contour identity; project specialization] This is the real-self-dual,
pole-containing version of the smoothed approximate functional equation in
Rubinstein [R05, §3.2, Theorem 1], whose original proof was retrieved and read.
For α=1/10000 and c=8, write

    I_F(s) = (1/(2πi)) ∫_(Re z=c)
                 Λ_F(z) exp(α(z−s)²)/(z−s) dz.

Move the contour left, collecting its pole at z=s and the poles of Λ_F at
0,1, then use Λ_F(z)=Λ_F(1−z). The result is

    Λ_F(s) = I_F(s)+I_F(1−s)
             − R_F[exp(α(1−s)²)/(1−s)+exp(αs²)/s].       (1)

The signs matter. Multiplying out the pole-removal factor gives an everywhere
regular formula in the working strip:

    X_F(s) = s(s−1)[I_F(s)+I_F(1−s)]
             + R_F[s exp(α(1−s)²)−(s−1)exp(αs²)].        (2)

In particular X_F(0)=X_F(1)=R_F>0. We approximate only the absolutely convergent
Dirichlet series inside I_F. The pole contribution is supplied independently
by the unit/regulator calculation, not fitted to a known zeta value.

## 5. Explicit error bounds

[project — analytic certificate] Computations use Arb at **384 working bits**.
This is not a claim that every output has 384 accurate bits. Arb propagates
special-function, Gauss-node/weight, input and rounding uncertainty; the bounds
below additionally enclose analytic truncation and quadrature errors.

### 5.1 Omitted prime-ideal coefficients

Let P=30000 and Z_P(z) be the field's finite prime-ideal Euler product. Its
Dirichlet coefficients b(n) satisfy 0≤b(n)≤a_F(n), with equality for n≤P.
Coefficientwise comparison of local Euler factors gives

    0≤a_F(n)≤d_d(n),

where d_d counts ordered d-fold divisor factorizations. Consequently, for b=64,

    |ζ_F(b+iv)−Z_P(b+iv)|
      ≤ Σ_(n>P) d_d(n)n^(−b)
      ≤ P^(2−b) Σ_n d_d(n)n^(−2)
      < 2^d P^(2−b).                                      (3)

The final inequality follows from the elementary p-series bound Σn^−2<2,
not from a numerical Riemann-zeta call.

The difference of the two contour integrals is analytic to the right of 1,
so shift that **error integral** from c=8 to b=64. No functional equation is
assumed for the finite Euler product itself. For
−3/2≤Re(s)≤5/2, the Gaussian integral and |Γ(x+iy)|≤Γ(x) give a uniform error
for the sum of the two I terms of at most

    E_series = 2 A_F^b Γ(b/2)^d 2^d P^(2−b)
                 exp(α(b+3/2)²) sqrt(π/α) / [2π(b−5/2)].  (4)

No discarded coefficient is assumed small because a numerical run stabilized.

### 5.2 Vertical truncation

We integrate only |Im z|≤V=160. Since c=8,

    |Γ(4+iv/2)|²
      = [π(v/2)/sinh(πv/2)] ∏_(j=1)^3 [j²+(v/2)²].

For v≥V this is bounded by

    C_V v^7 exp(−πv/2),
    C_V = (π/64)(1+36/V²)^3/(1−exp(−πV)).

The finite Euler product on Re z=8 is bounded by (8/7)^d, using the elementary
Dirichlet-series bound at 8. Both halves of both I terms are therefore covered by

    E_vertical = (2/π) A_F^8 (8/7)^d exp(α(19/2)²)/(11/2)
                    · C_V^(d/2) ∫_V^∞ v^(7d/2)exp(−πdv/4)dv. (5)

The last integral is evaluated by the exact finite integration-by-parts
recurrence for an integer power times an exponential. This bound drops the
additional Gaussian decay and is conservative.

### 5.3 Gauss quadrature with an analytic-circle bound

The positive half [0,160] is divided into 40 intervals of length 4, using
144-point Gauss–Legendre on each. Extend each integrand analytically to the
circle of radius 4 about the interval midpoint. Its z-real part then lies in
[4,12]. For the same s-domain as above, |z−s|≥3/2. Log-convexity gives
|Γ(z/2)|≤120 there, and the finite Euler product is bounded by (4/3)^d.
A valid supremum for each of the four integrands is

    B_F = max(A_F^4,A_F^12) 120^d (4/3)^d
                                  exp(α(27/2)²)/(3/2).

On a real interval the half-width/radius ratio is 1/2. Cauchy's estimate,
Gauss exactness through degree 287, and positive Gauss weights bound that
interval's quadrature error by 4(length) B_F 2^−288. Summing intervals and the
four integrands, including their 1/(2π) factors, gives

    E_quad = (2/π)·4V B_F 2^−288.                          (6)

### 5.4 Uniform bounds and Taylor enclosures

On |s−(1/2+it)|≤2, |t|≤30, we have |s(s−1)|≤1225. Thus

    ε_F = 1225(E_series+E_vertical+E_quad)

bounds the analytic error of (2). The full interval upper bounds are approximately

| field | ε_F |
|---|---:|
| F5 | 4.477×10^−78 |
| F41 | 8.965×10^−74 |
| F205 | 1.401×10^−69 |
| E | 6.870×10^−54 |

Each rounded number here is larger than the corresponding recorded bound.
Actual interval balls, not this table, are used in the implementation.

Direct interval substitution would lose the cancellation in (2). We instead
construct Taylor coefficients at each target. For a node w=z−s0, expand
exp(α(w−δ)²)/(w−δ) by a two-term exponential recurrence and a geometric-series
recurrence. The elementary residue exponentials are expanded separately.
The field functional equation and conjugation symmetry determine exact
real/imaginary parity of coefficients; no RH is used for that symmetry.

A separately computed analytic supremum M_F bounds the exact X_F on the circle
of radius 6 about each center 1/2+it, |t|≤30. That circle stays inside the
I-integral analytic strip −7<Re(s)<8. To obtain M_F, bound the two integrals by
A_F^8(8/7)^d, |Γ(4+iv/2)|≤6 for |v|≤V, the explicit tail (5), and their distance
at least 3/2 from the kernel poles; then include |s(s−1)|≤1600 and the two
residue exponentials. The resulting M_F are below 10^30; exact formulas and
balls are in `independent_kernel.py` and `kernel.json`.

For a model of degree n and |δ|≤r<2, Cauchy's coefficient-error estimate and
Taylor's analytic remainder give the final additional radius

    ε_F/(1−r/2) + M_F(r/6)^(n+1)/(1−r/6).                 (7)

Full complex segment images are enclosed, not just sampled endpoints. Degrees
are 16 for the three small noncentral boxes and 128 at t=0 and t=30, where the
same model also covers a full real/top edge. Saved coefficient balls retain
rounding uncertainty in addition to (7).

## 6. Local counts, actual root existence, and orders

[project — interval argument principle] Each small square has rational
half-width 10^−6 in both coordinates, centered at 1/2 plus one of the stated
rational heights. At every accepted contour segment, the complete Taylor image
for each field is enclosed in a convex complex rectangle excluding zero, and
the endpoint ratio has positive real part. The principal endpoint-argument
increment is therefore the actual continuous increment along that segment.
Summing those intervals isolates an integer winding, hence a zero count.
The four squares required 4,7,8,8 accepted segments respectively.

The decimals specifying the target centers are **not** asserted to be exact
zeros. We freshly certify opposite signs of X_E on its critical line at a
rational height bracket of half-width 10^−12 about each noncentral seed. The
functional equation makes X_E real on that line. Each sign change gives a zero;
the field contour count one makes it unique and simple. For the third target,
for example, the sign enclosures are near ±5.50427×10^−43, with radii less than
6.870×10^−54, so this is not a floating-point “almost zero” decision.

[known bridge; project application] E/F_D is Galois, so Aramata–Brauer makes
ζ_E/ζ_FD entire. Every F_D has one zero in the square; E has just one. They
must therefore have the **same** simple zero, not four unrelated nearby zeros.
Applying the exact order projector at that common point gives ordinary order
(1+1+1−1)/2=1. This supplies an ordinary zero at that point without evaluating
ordinary ζ. At the center all four small squares have count zero, so ordinary
ζ is nonzero there as well. This resolves counts into actual pointwise orders,
not merely a match between rounded root lists.

## 7. Whole-window counts: a complete contour, not a mesh or heuristic

[project — interval argument principle] The rectangle is
−1/2≤Re(s)≤3/2, |Im(s)|≤30. Instead of an expensive vertical mesh we use an
analytic logarithm on Re(s)=3/2, where the field Euler products converge
absolutely. For s_R=3/2+30i, put

    φ_F = Im[log s_R + log(s_R−1) + s_R log A_F
             + d log Γ(s_R/2)
             − Σ_(p≤P) Σ_(pfrak|p) log(1−p^(−f(pfrak)s_R))].

This is an analytically continued argument anchored at the **positive real**
value X_F(3/2). The logs of s_R and s_R−1 and log Γ have their unambiguous
analytic branches in the right half-plane. Every Euler-factor argument lies
in the disk centered at 1 of radius less than 1, so its log has the corresponding
anchored analytic branch. The omitted logarithmic tail is at most

    L_F = 2d/[√P (1−(P+1)^(−3/2))].                        (8)

This interval is added **once at the endpoint**, not repeatedly at every
vertical mesh step. It bounds the total vertical argument uncertainty without
accumulating hundreds of unrelated tail errors.

The top edge from 3/2+30i to −1/2+30i is independently enclosed by the Gaussian
AFE model, with 18 accepted segments for all four functions simultaneously.
Let D_F be its certified continuous argument increment. Functional equation
and conjugation supply the other edges: each vertical edge contributes 2φ_F,
and the bottom contributes D_F. Thus the full winding is

    (4φ_F+2D_F)/(2π).

A separate real-axis certificate proves X_F(s)>0 throughout [−1/2,3/2]
using four positive interval segments; hence no real zeros contribute. The
positive-height count is consequently the unique integer in

    N_F(30) ∈ (2φ_F+D_F)/(2π).                              (9)

The radii of these last enclosures are below 0.00736 for the quadratic fields
and 0.0148 for E. They isolate 14,24,31,63 respectively by a wide margin.
All critical-strip zeros through height 30 are inside this rectangle; neither
line sampling nor an assumption of finite-height RH is used. This is an
argument-principle certificate, **not** a claim to implement Turing's particular
averaged-argument inequalities.

## 8. The projector, and why it legitimately recovers ordinary data

[known algebraic identity; project application] Artin formalism for the
biquadratic extension gives the identity of completed functions, up to a
nonzero constant,

    X_F5 X_F41 X_F205 = constant · X_E ξ_ζ².

For example, the gamma degrees are 6−4=2, the discriminant factors leave the
ordinary squared gamma normalization, and the three pole-removal factors minus
one leave exactly two. Taking signed orders, or boundary argument integrals,
gives the stated projector with coefficient 1/2. These are known symbolic
identities; using them is not a numerical call to an ordinary-zeta oracle.

[project — direct normalization audit] We also check the own prime-ideal patterns
at every one of the 3245 audited primes: the local exponents for the three
quadratic fields minus E are exactly two at residue degree one and zero at all
other residue degrees, including ramified primes. Gamma-degree difference is
two, discriminant ratio is one, and pole-removal exponent difference is two.
`scripts/independent_projector_audit.py` records these exact checks in
`results/independent/projector-audit.json`. The all-prime identity is a known
theorem, not an extrapolation from this finite test.

[project] The three local recovered ordinary zeros lie on the critical line
and contribute three positive zeros; the independently recovered full-strip
count is also three. Thus this pipeline verifies that these are all the
positive ordinary zeros through 30, with no omitted or multiple ordinary zeros
in that finite window. This is a re-verification of a long-known finite result,
not a new RH theorem. It makes the **data-flow direction** genuinely downward,
unlike the earlier field product evaluated from its ordinary factor.

## 9. Verification, artifacts and reproducibility

[project] Scripts:

* `scripts/independent_field_arithmetic.py`: prime ideals, class/unit certificates,
  regulator/residue balls and 1000 own ideal coefficients per field.
* `scripts/independent_kernel.py`: all four prime-ideal products, Gaussian contour
  kernel, and explicit error/supremum bounds.
* `scripts/independent_certificate.py`: Taylor models, local contours/signs,
  real-axis exclusion, top contour, Euler-log phases and signed recovery.
* `scripts/verify_independent.py`: source/input hash chain, evaluator-call audit,
  a second integer coefficient algorithm, independent recomputation of all three
  analytic tail/error budgets and three representative kernel nodes, re-enclosure
  of **every** saved segment
  and endpoint increment, real-axis proof and independent Euler-log phase check.
  Only after all those pass does it read the old ordinary reference files.
* `scripts/independent_projector_audit.py`: exact local-prime and completed
  normalization checks for the extraction identity.
* `scripts/plot_independent.py`: matching table from certified counts and the
  post-proof reference comparison.

All JSONs and logs are under `results/independent/`. In particular:
`arithmetic.json`, `kernel.json`, `certificate.json`, `verification.json`, and
`comparison.json`. Inputs and source scripts have a checked SHA-256 chain.
The numerical generators do not import the old ordinary/factor certificates;
that comparison is confined to the final section of the separate verifier.

Run from this directory:

```sh
/workspace/tools/pari-venv/bin/python scripts/independent_field_arithmetic.py
/workspace/tools/pari-venv/bin/python scripts/independent_kernel.py
/workspace/tools/pari-venv/bin/python scripts/independent_certificate.py
/workspace/tools/pari-venv/bin/python scripts/verify_independent.py
/workspace/tools/pari-venv/bin/python scripts/independent_projector_audit.py
python3 scripts/plot_independent.py
```

The arithmetic stage took about 0.18 seconds, the shared contour kernel about
68 seconds, and the five Taylor/count stages about 15 seconds on this box.
These are observed runtimes, not certified complexity bounds. Libraries are
python-flint 0.9.0 and cypari2 2.2.4. The saved-witness verifier checks 4000
coefficients, 45 contour segments and four real-axis segments; its scope is
explicitly narrower than a formal verification of Arb/PARI internals.

[project — trust scope] The certificate trusts PARI's exact number-field
arithmetic and `bnfcertify`, Arb's interval elementary/gamma routines and
Gauss-root enclosures, and the standard field functional equations/class-number
formula. Analytic truncation, quadrature and Taylor error bounds are derived
above and implemented explicitly. The separate checker re-encloses cached
models; reproducing all special-function kernel balls requires rerunning the
kernel generator. No empirical mesh convergence is accepted as a proof.

## 10. What was not done, and what this changes

[project — optional K not certified] We did **not** independently evaluate the
degree-eight Q8 field zeta. Its completed factor has
A_K=205³/π⁴ and eight gamma factors. Reusing the degree-four cutoff would not
provide a useful error bound. A numerical cost estimate for this same coarse
shifted-series bound, optimized over even shift lines, already needs a rational
prime cutoff around 1.8×10^10 merely to force its absolute X-error below 10^−50.
This is an illustration of this method's cost, not an impossibility theorem or
a precision prescription for degree eight. We neither manufacture a K certificate
nor relabel the earlier factored K result as independent. **K is unnecessary for
the ordinary-zeta projector**, so all requested ordinary recoveries succeed
without it. A more specialized direct degree-eight kernel is future work.

[project — honest verdict] The proposed independent check is now accomplished
for all four required proper subfields, all four local targets, and the complete
height-30 window. It confirms the projector has the right normalization, shows
that direct field arithmetic can carry a real finite payload down to ordinary
ζ, and independently audits the earlier factorwise result with a different
analytic formula. It does not supply the missing uniform analytic inequality
identified in `think-hard-heilbronn.md`. The compatible hypothetical-zero model
Θ_G=m·1_G still survives the known general bookkeeping at unchecked heights.
Here the independent upper count does bite finitely: it leaves room only for
the three isolated ordinary zeros below 30. No analogous all-height bound has
been established. None of these finite
checks advances an RH proof, establishes universal simplicity/noncoincidence,
or creates a new quaternionic/ordinary-zeta field of study.

## References and source status

* [R05] M. Rubinstein, *Computational methods and experiments in analytic number
  theory* (2005), §3.2 Theorem 1 and proof, equations (23)–(28).
  <https://arxiv.org/abs/math/0412181>;
  <https://doi.org/10.1017/CBO9780511550492.015>.
  Original retrieved and relevant theorem/proof read; PDF/text archived at
  `sources/independent/rubinstein-2005.pdf/.txt`. Our Gaussian specialization and
  the finite-prime-product error treatment are derived here, not claimed as
  verbatim the author's implementation.
* Aramata–Brauer and the biquadratic/Heilbronn identities: primary-source audit
  in `think-hard-transfer.md` §§2–4 and `think-hard-heilbronn.md` §§1–2,9.
  No unproved Artin holomorphy is needed for these Galois Dedekind quotients.
* PARI/GP reference documentation, general number fields, `idealprimedec`,
  `dirzetak`, `bnfcertify`, and `bnfinit`; the last two provide the exact
  class/unit certificate trusted here. Coefficient cross-checks and exact
  unit polynomials, rather than numerical PARI L-values, are recorded in the
  arithmetic artifact.

[project — follow-up status] The earlier proposal “independent subfield input”
is closed for this finite set of targets. The independent degree-eight evaluator
and any globally useful family bound remain open. No live page was changed.
