# Each kind of zero is the Fourier dual of a prime set

10 November 2026, America/Los_Angeles. A proposal, offered in response to a request to think hard and come up with something: each kind of zero of the D=30 quaternion-order zeta is the Fourier dual of a prime set. A local column (zeros spaced π/log p on the half-line) would be, by Poisson summation, exactly dual to the powers of that one prime: a comb ↔ comb pair. The inherited zeros (quarter- and three-quarter-lines) would be dual to all primes, via the explicit formula.

The proposal is essentially right, once three corrections are made, and it fails in a few places where the wording was too loose.

(Pruned 11 Nov 2026: cross-domain comparison passages removed at Grant's request; original in /workspace/math/pruned-20261111/notes-orig/.)

**Labels.** [known] = established mathematics, with a source. [project] = a proof reconstruction, a certified finite computation, or a numerical check here; not a new number-theory theorem. [speculative] = an unproved additional idea.

This note does not claim RH progress. It was not deployed.

## 0. Verdict

1. **[known, constants derived here]** In Voight’s ideal-index coordinate the local factor is \(1-p^{1-2s}\). Its zeros, spaced \(\pi/\log p\) on the half-line, are by Poisson summation dual to a Dirac comb at \(u=\log x=2k\log p\), i.e. at ideal indices \(p^{2k}\). In the reduced-norm coordinate \(w=2s\), the same duality is a comb at \(p^k\). The proposal’s “spikes at \(p^k\)” is the reduced-norm reading; it is off by a factor of two if one plots the ideal index.
2. **[known / project]** The local-zero sum is *not* the remaining ramified von Mangoldt of \(Z_{30}\) itself. After the Euler factor \((1-p^{1-2s})\) cancels the local factor of \(\zeta(2s-1)\), one has \(\Lambda_Z(p^{2k})=2\log p\). The zeros of \(1-p^{1-2s}\) are dual to the *correction* that was removed from the split zeta, of weight \(-2\,p^k\log p\) at the same points \(p^{2k}\). Those two combs live on the same lattice and have different masses.
3. **[known]** The zeros of \(Z_D\) enter the explicit formula for the Chebyshev function \(\psi_Z(x)=\sum_{n\le x}\Lambda_Z(n)\), not for the ideal-counting staircase \(I(x)=\sum_{n\le x}a(n)\). \(I(x)\) is controlled by the pole of \(Z_D\) at \(s=1\), of residue \(\pi^2/45\) for \(D=30\). The proposal’s “count of order ideals vs reconstruction from each zero kind” therefore has to be read as a reconstruction of \(\psi_Z\), or of a Gaussian-smoothed prime-power measure. That reconstruction works, to \(10^{-11}\) in the plot window.
4. **[project]** Inherited zeros reconstruct the *split* contribution \(2\Lambda(n)(1+n)\) at every square \(n^2\), including the ramified primes at their larger split weight. The “non-ramified part” is inherited minus those ramified split terms, or equivalently inherited plus the local correction.
5. **[project]** A single local column is a picket fence. The merged half-line of the three incommensurate combs \(p=2,3,5\) is *not* Poisson: the support of the unfolded gap law is bounded (maximum \( \log 30/\log 5\approx 2.113\)), there is an atom at that maximum, small gaps are common, and the variance is 0.433, not 1.

## 1. The explicit formula, term by term

**[known]** Voight, *Quaternion Algebras*, Theorem 26.4.9 / (26.4.10), in the ideal-index coordinate:

\[
Z_D(s)=\zeta(2s)\,\zeta(2s-1)\prod_{p\mid D}(1-p^{1-2s}).
\]

For \(D=30\) the ramified primes are \(2,3,5\). Write \(Z_{\mathrm{split}}(s)=\zeta(2s)\zeta(2s-1)\), so \(Z_{30}=Z_{\mathrm{split}}\prod_{p\mid 30}(1-p^{1-2s})\). At a ramified prime the extra factor cancels the local factor of \(\zeta(2s-1)\), leaving \((1-p^{-2s})^{-1}\).

Define \(\Lambda_Z\) by \(-Z_D'/Z_D(s)=\sum_n\Lambda_Z(n)\,n^{-s}\). Expanding the Euler products gives the exact weights, supported only on perfect squares (as they must be: the index of a compatible left ideal is a square of a reduced norm).

- Unramified prime power, index \(q^{2k}\): \(\Lambda_Z=2\log q\,(1+q^k)\).
- Ramified prime power, index \(p^{2k}\): \(\Lambda_Z=2\log p\).
- The local factor \(F_p=1-p^{1-2s}\) on its own contributes \(\Lambda_F(p^{2k})=-2\,p^k\log p\), which is exactly the difference \(\Lambda_Z-\Lambda_{\mathrm{split}}\) at those points.

The split log-derivative has residue \(+1\) at \(s=1/2\) (the pole of \(\zeta(2s)\)) and \(+1\) at \(s=1\) (the pole of \(\zeta(2s-1)\)). Those are the main terms \(\sqrt{x}\) and \(x\) in unsmoothed \(\psi\). The same residues, tested against a unit-mass Gaussian \(g\) in \(u=\log x\), become the Mellin values \(G(1/2)\) and \(G(1)\). Trivial zeros of \(\zeta(2s)\) and \(\zeta(2s-1)\) sit at \(s=-m\) and \(s=\tfrac12-m\) and enter with the classical minus sign.

**Local Poisson summation, [known].** The heights \(\gamma_m=m\pi/\log p\), \(m\in\mathbb{Z}\), form a lattice of spacing \(\Delta=\pi/\log p\). For a Schwartz test function, Poisson summation sends this lattice to a lattice of frequencies \(k/\Delta\), which in the variable \(u=\log x\) is the comb \(u=2k\log p\). The identity was checked at seven centres for \(p=2,3,5\), with a maximum absolute error of \(1.1\times 10^{-12}\). The zero sum equals the *two-sided* comb, including a smooth \(k=0\) term and functional-equation spikes at \(p^{-2|k|}\). Subtracting those \(k\le 0\) terms recovers the one-sided Dirichlet series of \(-F_p'/F_p\), of weight \(-2p^k\log p\), to \(4\times 10^{-12}\).

This is the same mechanism as the local terms of the Weil explicit formula, or of a Selberg trace formula: a periodic column of spectral points is dual to a periodic column of geometric lengths. It is not a new duality.

**Inherited zeros, [known].** The zeros of \(\zeta(2s)\) are at \(s=\rho/2\) (the quarter-line, on the ordinary RH, which is used here only as a description of the already certified on-line list). The zeros of \(\zeta(2s-1)\) are at \(s=(\rho+1)/2\) (the three-quarter-line). With the classical sign,

\[
\sum_n\Lambda_{\mathrm{split}}(n)\,g(\log n)
=G(\tfrac12)+G(1)-\sum_{\rho}G(\rho/2)-\sum_{\rho}G((\rho+1)/2)-\sum_{\mathrm{trivial}}G(s).
\]

**[project]** Using the 22,491 certified ordinary zeros to ordinary height 20,000 (inherited height 10,000) and a Gaussian of width 0.08 in \(u\), this identity holds to \(1.0\times 10^{-11}\) on the safe window \(0.8\le u\le 7.36\). The two families are dual to different weights: \(\zeta(2s)\) to \(2\Lambda(n)\) at \(n^2\), \(\zeta(2s-1)\) to \(2n\Lambda(n)\) at \(n^2\). Together they give every prime, not a translate of the ramified comb.

Adding the local reconstruction recovers \(\Lambda_Z\), to \(1.2\times 10^{-11}\) in the same window. The arithmetic identity \(\Lambda_{\mathrm{split}}+\Lambda_F=\Lambda_Z\) is exact (residuals \(4.5\times 10^{-13}\), ordinary floating-point).

## 2. What did not work as proposed

- **Spikes at \(p^k\).** In the project’s standing ideal-index coordinate the comb is at \(p^{2k}\). The reduced-norm coordinate \(w=2s\), already used in the height notes, puts the same spikes at \(p^k\). Both are exact; the figure is labelled in the ideal-index log.
- **“The ramified-prime contribution.”** The local zeros reconstruct the *correction* of weight \(-2p^k\log p\), not the leftover \(\Lambda_Z=2\log p\). The leftover is what remains of \(\zeta(2s)\) at those primes, and is dual to inherited zeros plus the cancelled pole at \(s=1/2\), not to the half-line column.
- **“Count of order ideals.”** \(I(x)=\sum_{n\le x}a(n)\) has a simple main term \((\pi^2/45)x\) from the pole of \(Z_{30}\) at \(s=1\). Zeros of \(Z_D\) do not appear as residues of \(Z_D(s)x^s/s\). The dual of the zeros is \(\psi_Z\). The figure plots the Gaussian-smoothed prime-power measure, which is the right object.
- **“Inherited zeros reproduce the non-ramified part.”** They reproduce the split contribution of *every* prime. To isolate the non-ramified part one must subtract the ramified split weights, or add the local correction.
- **An “exact Gram law” was not part of this brief**; the merged-column spacing law, however, is exact, and it is not Poisson. See §3.

## 3. Spacing statistics

**[known, Montgomery–Odlyzko, reproduced].** Unfolded nearest-neighbour spacings of the 22,481 certified ordinary zeros above height 50, which are exactly the inherited spacings at half height, follow the GUE Wigner surmise in shape (quadratic repulsion at the origin, a peak near 1, a thin tail). The sample mean is 1.000; the sample variance is 0.157, against 0.178 for the \(2\times 2\) Wigner surmise. The mismatch is not hidden: the surmise is an approximation, and height 20,000 is not an Odlyzko-scale computation. The plot is a reproduction, not a new statistical theorem.

**[known]** Each local column is a picket fence: raw spacing \(\pi/\log p\), and every unfolded gap is exactly 1.

**[project]** The merged half-line is the union of three incommensurate combs, with total density \((\log 30)/\pi\). From 2,165,215 exact gaps up to height two million:

- the unfolded law has mean 1.000 and variance 0.433, not the Poisson value 1;
- the support is bounded, with maximum \(\log 30/\log 5\approx 2.113\), saturated in the sample;
- there is an atom at that maximum, from two consecutive \(p=5\) zeros with no 2 or 3 in the interval;
- small gaps are common (about 6% below 0.1), unlike GUE.

Poisson would have an exponential tail and no upper bound. The merged law is a quasiperiodic gap law, not a random one. A single column is perfectly periodic; the merged half-line is a finite union of periodic columns, already irregular to the eye but still far from Poisson or GUE; the inherited family follows GUE statistics.

## 4. Framing

**[known]** Poisson summation; the Weil / Riemann explicit formula; the local terms of a Selberg trace formula; Montgomery–Odlyzko GUE statistics for ordinary zeros; Voight’s factorisation of the ideal-index zeta.

**[project]** The constants above; the certified-zero reconstruction of the split and order measures; the exact merged-column spacing law; the figures.

**[known]** The explicit formula is an identity between two sides of one function, not a process that converts primes into zeros at a locus.

## 5. Figures

One relation per figure. Desktop and 360px mobile, SVG and PNG.

- `figures/zero-duality-recon.svg` (and `.png`, plus `-mobile`): three-row explicit-formula reconstruction. Local zeros rebuild the ramified correction at \(p^{2k}\); inherited zeros rebuild every split prime power; together they rebuild \(\Lambda_Z\). Residual, in the plot window, about \(10^{-11}\).
- `figures/zero-duality-spacings.svg` (and `.png`, plus `-mobile`): unfolded nearest-neighbour laws for inherited zeros (versus the GUE surmise), one local column (a Dirac at 1), and the merged 2, 3, 5 half-line (versus Poisson, with the atom at the maximum marked).

## 6. What is exact, certified, or numerical

**Exact.** The Euler factorisation and the weights of \(\Lambda_Z\), \(\Lambda_{\mathrm{split}}\) and \(\Lambda_F\); the Poisson duality of a local column to the comb at \(p^{2k}\) (ideal index) or \(p^k\) (reduced norm), including the \(k\le 0\) terms; the residue of \(Z_{30}\) at \(s=1\), equal to \(\pi^2/45\); the identification of inherited heights with ordinary ordinates at half height; the picket-fence gaps; the theoretical maximum merged gap \(\log 30/\log 5\).

**Certified [project].** The 22,491 ordinary zeros to height 20,000, from the existing height certificate (`source_id` `eb130d8f…`, every chunk digest checked); the GUE spacing plot from those midpoints.

**Numerical [project].** The Poisson identity to \(1.1\times 10^{-12}\); the inherited and order reconstructions to \(10^{-11}\) in the safe window, with a Gaussian of width 0.08 (for this width only the first few hundred ordinary zeros contribute appreciably; the full list was used nonetheless); the merged-column histogram from 2.17 million exact heights.

## 7. Reproducibility

From this directory:

```bash
python3 zero_duality.py
python3 zero_duality_figures.py
python3 test_zero_duality.py
```

The computation reads, and does not write, `order-height/`. It does not deploy, and it does not touch `zeta-zeros-build/`.

Artifacts: `zero-duality.md`; `zero_duality.py`, `zero_duality_figures.py`, `test_zero_duality.py`; `zero-duality-results.json` and `zero-duality/results.json`; the eight figure files in `figures/`.
