# Beyond the first zeros: exact counts, and a different quaternion zeta

## Scope and labels

[known] An intrinsic slice extension, an Artin L-function, an ideal-counting zeta function of a quaternion order, and a lattice Epstein zeta function are different objects. A statement about one is not automatically a statement about the others. This note makes no RH-progress or novelty claim.

[project] The machine-readable certificate is `q8-extended-results.json`; the historical height-20 run is `q8-extended-20-results.json`. The final results table is generated below from the accepted run. `quaternion-orders-results.json` contains the separate ideal-zeta calculations. Numerical roots of the classical Riemann zeta function used in the order illustration are not new interval certificates.

## 1. Certified Q8 counts, not just a search along the line

[known] The three degree-two representations are induced from nontrivial characters of a cyclic subgroup of order four. Their L-functions are entire Hecke L-functions after induction. Thus the entire completed functions used here do not require an assumption of the general Artin holomorphy conjecture. Their Euler factors have unit-modulus roots or missing roots, so their real coefficients obey |a_n| <= d(n) <= n. Sources: Artin induction formalism; the PARI `lfunartin` documentation; the representation/field certificates in our earlier notes.

[project] For each representation we independently enclose a real function on the critical line and the image of an enclosing rectangle. Opposite interval-certified signs locate zeros. The winding of the enclosing rectangle counts *all* zeros inside, with multiplicity, including any off-line zeros that a line search would miss. Matching those totals proves there are no missed zeros in this finite-height region and that each located zero is simple. This performs the completeness task associated with Turing's method, but uses the argument principle, not a new Turing-method implementation. It is not a claim about all heights.

### The completed function and an explicit integral

[known/project checked] Put Q=sqrt(N), nu=0 or 1, c=2 pi/Q, and W=+1 or -1. With our normalization,

    Lambda(s) = Q^s pi^(-s-nu) Gamma((s+nu)/2)^2 L(s),
    Theta(u)  = 4 (u/Q)^nu sum_(n>=1) a_n n^nu K_0(c n u),
    Lambda(s) = integral_1^infinity Theta(u) [u^s+W u^(1-s)] du/u.

The Mellin transform of K_0 gives the first identity; the known functional equation and theta transformation give the last. For s=1/2+it, Lambda is real for W=+1 and purely imaginary for W=-1. We use H(t)=Lambda(s) in the first case and H(t)=Lambda(s)/i in the second. Thus the real kernels are twice cosine and twice sine respectively. Our earlier direct-PARI computations cross-check this normalization.

### Explicit tails

[project] Let M be the coefficient cutoff and U the integral cutoff, and define

    C=8 Q^(-nu) sqrt(pi/(2c)),
    S2(r,m)=r^m [m^2/(1-r)+2 m r/(1-r)^2+r(1+r)/(1-r)^3].

For -1/2 <= Re(s) <= 3/2, uniformly in Im(s), we use

    E_M = C (1/c+1/c^2) S2(exp(-c),M+1),
    E_U = C (U/c+1/c^2) r(1+r)/(1-r)^3, r=exp(-cU).

Proof: K_0(x) <= sqrt(pi/(2x)) exp(-x) for x>0 (its cosh integral and cosh(v)>=1+v^2/2). Bound |a_n| by n, u^nu by u for u>=1, and n^(nu+1/2) by n^2. Integrating u exp(-cnu) gives the displayed bounds after discarding improving factors 1/n. These bounds do not rely on sampled numerical tail decay.

### Rigorous reusable quadrature

[project] Each dyadic interval [a,2a], a=1,2,...,U/2, uses n-point Gauss-Legendre quadrature, with roots and weights enclosed by Arb. The script asserts that the final interval really ends at U. On the complex disk centred at 3a/2 with radius a, Re(u)>=a/2, |u|<=5a/2, and |arg(u)|<=alpha=arcsin(2/3). For |Im(s)|<=T a bound for the full truncated integrand is

    M_a = (16/a) (5a/(2Q))^nu (5a/2)^(3/2) exp(T alpha)
          sum_(n<=M) |a_n| n^nu K_0(cna/2).

Here |K_0(z)|<=K_0(Re z) follows directly from its integral. The slightly loose power bound also covers |u|<1 in the first disk. Taylor expansion in u about 3a/2 has coefficients bounded by M_a/a^k. Gauss is exact through degree 2n-1; the remaining Taylor tail on [a,2a] is <=2 M_a 2^(-2n). Both the integral and the positive-weight quadrature have mass a. Hence the error is at most **4 a M_a 2^(-2n)**. Sum this across all bands.

[project] Write z=s-1/2. The quadrature becomes a finite exponential sum

    G(1/2+z)=sum B_j [exp(z ell_j)+W exp(-z ell_j)],
    ell_j=log u_j, B_j=weight_j Theta_M(u_j)/sqrt(u_j).

Every quantity is an Arb ball. On a contour segment centred at z0 with displacement |delta|<=h, expand in delta to degree d. Since |Re z0|<=1, a valid remainder is

    sum_j 2 |B_j| u_j exp(h ell_j) (h ell_j)^(d+1)/(d+1)!.

Horner interval evaluation encloses the whole segment, not just its endpoints. The full complex error is padded conservatively in both real and imaginary coordinates. Wider intermediate Arb balls are retained, not replaced by the analytic quadrature error.

[project] Development checks caught Bessel K_0 cancellation at 256-bit working precision: a true tiny positive value could have an interval radius around 1e-19. Increasing precision, rather than ignoring that radius, fixed the signs. The final accepted run uses the settings in JSON. A change of U also received an explicit last-band assertion before any deeper result was accepted.

### Winding certificate and interpretation

[project] The contour is the positively oriented boundary of [-1/2,3/2] x [-T,T]. Each segment image is enclosed by a convex rectangle missing zero. Adjacent image rectangles intersect because they both enclose the shared true endpoint. Choose a point in each intersection and join successive points inside the corresponding image rectangle. The actual image curve is homotopic to this polygon in C minus {0}; thus they have the same winding. Arb sums the principal arguments of successive vertex ratios. The resulting interval isolates one integer. Since the completed L-function is entire, the argument principle makes that integer the total zero multiplicity. The absolutely convergent Euler product and functional equation put all its zeros within 0<=Re(s)<=1, so this rectangle encloses the whole relevant strip through height T.

[project] PARI finds candidate heights only; Arb certifies the bracket signs and independent contour count. An incomplete candidate search would fail the final count equality. Each disjoint positive bracket contributes at least one zero, its negative partner contributes another, and an odd functional equation contributes a central zero. Equality with the total count proves completeness and simplicity. Boundary images exclude zero, so there are no ambiguous boundary zeros.

[known] For positive heights the smooth leading term is

    M_N(T)=(T/pi) log(sqrt(N) T/(2 pi e)).

This is an asymptotic main term, not an exact integer prediction. Constants, the gamma phase, and the argument term matter at small height. The JSON additionally records the exact gamma phase and the count-minus-phase residual. The latter is inferred from the certified count; it is *not* an independently certified Turing S(T) integral.


### Accepted result table [project: interval proof]

| Representation | Sign | N+(10) | N+(20) | N+(50) | Total through absolute height 50 | Smooth M(50) |
|---|---:|---:|---:|---:|---:|---:|
| 2.2304.8t5.b.a | -1 | 10 | 25 | 79 | 159 | 78.707641239 |
| 2.2304.8t5.a.a | 1 | 11 | 26 | 79 | 158 | 78.707641239 |
| 2.42025.8t5.a.a | -1 | 15 | 34 | 101 | 203 | 101.813898610 |

[project] All 259 positive-height brackets have width 1e-12. The complete symmetric total across the three examples is 520, including two central zeros. All three contours use 3264 zero-avoiding segment images at T=50. Accepted compute time: 235.006 seconds (under the 1200-second budget); Arb working precision 768 bits. These finite-height Artin facts are not claims of progress on RH.

## 2. Maximal quaternion-order zeta: genuinely different arithmetic

[known] Voight, *Quaternion Algebras*, Theorem 26.4.9, Corollary 26.4.11 and §26.8, give for a maximal order in a definite rational quaternion algebra of finite discriminant D:

    Z_D(s)=zeta(2s) zeta(2s-1) product_(p|D) (1-p^(1-2s)).

This convention sums |O/I|^(-s) over nonzero integral right ideals, where |O/I|=nrd(I)^2. Its initial half-plane is Re(s)>1. A different exponent convention changes displayed zero-line locations. Sources: https://jvoight.github.io/quat-book.pdf and https://doi.org/10.1007/978-3-030-56694-4_26 .

[project] We check two explicit maximal orders using exact rational arithmetic:

* D=2, B=(-1,-1): basis 1,i,j,(1+i+j+k)/2 (the Hurwitz order).
* D=3, B=(-1,-3): basis 1,i,(1+j)/2,(i+k)/2.

Here i^2=-1, j^2=-D (for these two algebras), and k=ij=-ji. All 16 basis products have integer coordinates. The determinants of the reduced-trace Gram matrices are -4 and -9, giving reduced discriminants 2 and 3. The Hilbert symbols show ramification at {2,infinity} and {3,infinity}, respectively. For (-1,-3), (-1/3)=-1 gives ramification at 3; at 2 the odd-unit formula gives +1; at infinity both defining parameters are negative. Equality of order and algebra discriminants proves maximality. This verifies the orders, not merely a formal Euler product.

### Three different families, with an important missing centre

[known consequence/project checked] A nontrivial Riemann zeta zero rho supplies zeros at rho/2 and (rho+1)/2. Our illustrated zeros therefore sit at real parts 1/4 and 3/4 because the selected classical zeros have real part 1/2. **The claim that every inherited nontrivial zero lies on those two lines is conditional on RH.** No such assumption is needed for the selected numerical examples or for the local-factor family below. Other real zeros/cancellations are not claimed to be exhaustively plotted.

[known consequence] For D=p, the local factor supplies exact zeros

    s=1/2 + i pi k/log(p), k an integer, k != 0.

They are genuinely on the half line. But at k=0 its zero cancels the pole of zeta(2s). Thus

    Z_p(1/2) = -log(p)/2,

not zero. For p=2 this is the Hurwitz value -0.34657359027997265..., and for p=3 it is -0.54930614433405485.... This is unrelated to the root-number -1 Q8 Artin centre. Higher local-factor zeros here cannot be cancelled by a pole: neither shifted zeta argument is its pole at those nonzero heights.

[project] `quaternion-orders.py` checks 48 selected nonreal zeros, central limits from both sides, 5000 exact ideal-count coefficients per order, and the completed functional equation at 80-digit precision. The checks of roots and values use mpmath, **not interval arithmetic**. The exact local-factor zero locations and the analytic cancellation argument do not depend on those numerical checks.

[known/project checked] In the present normalization

    Xi_D(s)=D^s 2^(2-2s) pi^(1-2s) Gamma(2s) Z_D(s)
    Xi_D(1-s)=Xi_D(s).

Thus the completion has centre 1/2 despite the inherited quarter/three-quarter families. A functional-equation centre is not a universal zero line.

### An extra control: a different way to get a central zero

[known consequence/project checked] D=30 is also a valid definite discriminant (three finite ramified primes plus infinity). Three local-factor zeros cancel one pole, leaving a *double* central zero:

    Z_30(1/2+h)=-2 log(2)log(3)log(5) h^2 + O(h^3).

The leading coefficient is approximately -2.45117397417. More generally r finite ramified primes give central order r-1, since each local factor has a simple zero and zeta(2s) one pole; a definite rational algebra requires r odd. This is a transparent local-factor mechanism, not the odd-sign Artin mechanism and not a new theorem. It usefully shows why the blanket phrase “quaternion zeta has/has not a central zero” is underspecified.

## 3. The small differences, separated cleanly

[known/project checked] Non-real coefficients can single out one direction instead of a whole zero sphere. A quaternionic Galois representation can have root number -1 and a forced central Artin zero. Octonionic star-polynomial roots can depend on brackets, whereas finite pointwise products cannot. Quaternion-order ideal counting introduces shifted and local-factor zero families. The intrinsic xi zero spheres leave essential loops at infinity, without proving a horned-sphere embedding. Lattice dimension changes an Epstein functional-equation centre under a fixed norm exponent. These are six distinct mechanisms; none is evidence that an intrinsic change from i to j changes the ordinary zeta zeros.

## Reproduce

    /workspace/tools/pari-venv/bin/python q8-coefficients.py
    /workspace/tools/pari-venv/bin/python q8-extended.py
    python3 quaternion-orders.py
    python3 extended-tests.py

The interval computation has a hard 1200-second budget. If a sign or count cannot be certified it raises an error rather than accepting an approximate answer. Previously completed field receipts remain historical evidence, but a final success requires all three fields at the declared target.
