# Which octonion count hears the zeros?

11 November 2026, PST. Grant pt199u. Mathematical research/figure; no messages.

## Plain explanation

**[project, proved from established factors]** There is an important correction to the suggested picture. Counting *all* octonion right ideals does not create waves from the zeros of its zeta function. Those zeros are not poles of the integral used to recover the coefficients. Here the count is a finite sum of rounded eighth roots, and its absolute error from the pole main term is bounded by192,162,8,1 for G,S1,R,E8 respectively—even as x tends to infinity.

**[known + project adaptation]** Zeros enter after taking a *logarithmic derivative*. This produces a different, signed diagnostic B. Its common part is8 times the ordinary prime-power staircase at x^(1/8); its remaining jumps occur only at powers of2. These are not counts of octonion prime ideals. In fact, all three nonmaximal kinds have negative weights at index16. Their local zero terms can grow much faster than x^(1/16), so “a small log-periodic wobble” would also be misleading.

## 1. Exact ordinary ideal count: no zero residues

Retain the established notion: a right ideal is a finite-index additive lattice L in the order S with L*S contained in L, not an arbitrary principal set a*S. Let a_K(n) count these ideals of additive index n and A_K(x)=sum_{n<=x}a_K(n), right-continuous at jumps.

Write z=2^(-s) and Z_K(s)=zeta(8s)C_K(z). Ascending coefficient lists c_j of the finite correction C are:

- G: [1,1,7,7,15,15,35,15,30,14,28,8,16].
- S1: [1,2,15,35,15,16,15,35,14,14].
- R: [1,0,3,0,1,0,3].
- E8: [1].

They come from the already checked 2-local factors, not a new ideal-lattice enumeration. Expanding the product gives the exact identity for every real x>0:

    A_K(x) = sum_j c_j floor((x/2^j)^(1/8)).

Set mu_K=C_K(2^(-1/8)). All c_j are nonnegative, hence

    -C_K(1) < A_K(x) - mu_K*x^(1/8) <= 0.

The residue of Z at1/8 is mu/8; dividing by s in Perron's integral gives the main term mu*x^(1/8). A zero of Z contributes **no residue** to Z(s)x^s/s. We are not asserting that no artificial representation involving zeros is conceivable; rather, the proposed zero-residue decomposition is for the wrong observable. The bounded floor error is the direct exact answer for this A.

At x=2^128:

| Kind | Exact A | Pole main term | Error | Absolute bound |
|---|---:|---:|---:|---:|
| G | 6,845,563 | 6,845,674.405260117 | -111.405260117 | <192 |
| S1 | 6,870,661 | 6,870,756.812485783 | -95.812485783 | <162 |
| R | 394,101 | 394,107.728655271 | -6.728655271 | <8 |
| E8 | 65,536 | 65,536 | 0 | <1 |

## 2. The distinct signed diagnostic

Define the Dirichlet coefficients lambda_K(n) of -Z'_K/Z_K, initially sufficiently far right for the expansion to converge. This region must be to the right of the rightmost local zero, not merely Re s>1/8. With

    z C'_K(z)/C_K(z) = sum_{k>=1} b_K(k) z^k,

logarithmic differentiation gives exactly

    -Z'_K/Z_K(s) = 8(-zeta'/zeta)(8s) + (log2) z C'_K(z)/C_K(z).

Thus lambda_K(m^8) receives8*Lambda(m), and lambda_K(2^k) additionally receives(log2)*b_K(k); contributions add when the indices coincide. Define B_K,0 with half of the coefficient at equality, as in the classical explicit formula. Then, for x>1,

    B_K,0(x) = 8 psi_0(x^(1/8))
                + (log2) [sum_{1<=k<log2(x)} b_K(k)
                          + (1/2)b_K(log2(x)) if log2(x) is integral].

For right-continuous B replace the bracket by the sum with2^k<=x and use right-continuous psi. This does **not** define a positive prime-ideal counting function. For k=4 the respective b values are -39,-446,-14,0. The corresponding nonmaximal diagnostic has a negative jump at16; no inherited m^8 contribution occurs there.

## 3. Inherited-zero terms, with the correct observable

**[known]** von Mangoldt's explicit formula, with the zero sum taken symmetrically by height and half-weights at prime powers, is

    psi_0(y) = y - sum_rho y^rho/rho - log(2*pi)
                 - (1/2)log(1-y^(-2)), y>1.

Substitute y=x^(1/8) and multiply by8:

    8 psi_0(x^(1/8)) = 8x^(1/8)
       - sum_rho x^(rho/8)/(rho/8)
       - 8log(2*pi) - 4log(1-x^(-1/4)).

A pair rho=1/2 +/- i*gamma contributes

    -16 x^(1/16) Re[exp(i*(gamma/8)*log(x))/rho].

That is the requested x^(1/16) carrier scale and gamma/8 frequency, **but for B, not A**. A hypothetical zero with real part beta off1/2 would instead have x^(beta/8). No statement about all zeros being on the line is needed. The displayed first100 pairs already have project interval certificates on the ordinary line; their midpoints are reused, not newly certified. The sum's convergence is conditional and additional terms do not improve every individual x.

## 4. Local columns, exactly and not just metaphorically

Let C(z)=product_l(1-alpha_l*z), including all roots of the whole correction, not only Q. Here alpha_l is the reciprocal of a root and alpha_l!=1 since C(1)>0. Then

    b_K(k) = -sum_l alpha_l^k.

For u=log2(x), m=floor(u), u not integral, the exact local term is

    L_K(u) = -(log2) sum_l alpha_l*(1-alpha_l^m)/(1-alpha_l).

Each alpha gives zeros at

    s_l,n = (Log(alpha_l)+2*pi*i*n)/log2, n in Z.

Each fixed-root column has imaginary period2*pi/log2; the union can have a smaller period. The precise column-residue identity is

    L_K(u) = (log2)*C'(1)/C(1)
             - sum_l lim_{M->infinity} sum_{n=-M}^M x^(s_l,n)/s_l,n.

At integral u use the midpoint value. Proof: for a=Log(alpha), Fourier coefficients of alpha*exp(-a*t)/(alpha-1), 0<t<1, are1/(a+2*pi*i*n). Dirichlet convergence gives

    sum_n exp((a+2*pi*i*n)*u)/(a+2*pi*i*n)
        = alpha^(m+1)/(alpha-1),

with the symmetric-index limit and the midpoint at a jump. Multiplication by log2 gives the displayed identity. Finite sums over the finitely many roots can be interchanged with these limits.

This is an honest exact local explicit formula. It does not make A zero-sensitive. Nor does it imply that the local diagnostic is a bounded or purely log-periodic wobble. A single column has an x^(s_l,0) carrier times a unit-periodic function of u. The rightmost numerical local real parts are G1.1833734993, S1 1.9430847813, R0.7924812504; E8 has no correction. Growth and generally nonrational phases matter.

## 5. Computation and independent checks

Run with `/workspace/tools/pari-venv/bin/python explicit1111/compute.py`; `test_math.py` contains small exact regression tests; `figure.py` produces desktop/mobile SVG.

**[project, computed]**

- Exact C-floor counts versus an independently coded odd-m/F2 formula at1,240 arguments up to2^128, including random arguments and exact jump boundaries. Eighth roots use repeated integer square roots versus binary search.
- Independently enumerate394,809 odd-m/index-exponent events at2^128, using the original F2 numerators/denominators, not C. All four large-bound totals agree.
- Enumerate every coefficient and every prefix through2^18 via its odd part and original F2 coefficient; all1,048,576 kind/prefix comparisons match the C-floor formula. These are arithmetic counts derived from the established ideal factors, not new HNF searches.
- Every signed b_k throughk128 agrees between polynomial-division recurrence and traces of powers of an integer companion matrix. Root/geometric sums independently agree at selected m through128 to relative error below5.5e-110 (110-digit numerical arithmetic, not interval enclosures).
- The513-point plot uses integer x rounded from2^u, u in[16.37,64.37]. Counts are exact; plotted curves join these samples and do not resolve every jump. Smooth main terms and signed reconstructions are numerical.
- Shared zero sums with N=5,20,100 are evaluated independently with numpy complex arithmetic and110-digit mpmath arithmetic: maximum difference below1.7e-12. Exact prime-power weights use a sieve versus trial-division prime generation, agreement below4.4e-14.
- For W=(8psi(y)-8y-h(y))/(8sqrt(y)), h(y)=-8log(2*pi)-4log(1-y^-2), normalized RMS errors are0.149343,0.108999,0.066196 for N5,20,100. Maximum normalized errors0.418822,0.368808,0.348314. These are measured grid errors, not rigorous tail bounds or monotonic convergence claims.
- Separate local-column cuts M64,256,1024 at u0.37,4.37,12.37 have maximum scaled errors0.012887,0.004076,0.001134. Scaling is by1 plus the sum of absolute geometric-root contributions. Column-index truncation is not a common-height truncation and these figures are numerical checks, not certified bounds. The figure uses exact b_k, not these approximations.

All coefficients, grid samples, metrics and input hashes are in `explicit1111/results.json`. Known-factor inputs and source hashes are inherited from `global1111/results.json`. No new RH information.

## 6. Sources actually checked and research-program consequence

Opened Kiran S. Kedlaya, *Notes on analytic number theory*, chapter9, §§9.1–9.5, especially Theorem9.9, https://kskedlaya.org/ant/chap-von-mangoldt.html . It gives half-weight convention, Perron integral, zero residues and a truncated explicit formula with remainder. Its rendered Lemma9.2 has an apparent typo in the constant denominator; the correctly stated Theorem9.9 and standard constant log(2*pi) are used here. Davenport, *Multiplicative Number Theory*, §17, is a standard textbook reference; its section heading/formula were verified in indexed text, not by claiming a full book read. Existing Petersson/2-local sources and caveats are retained in the preceding octonion notes.

**Program consequence:** every odd-prime contribution in this sixteen-order family is ordinary zeta after rescaling. Subtracting the exact prime2 term from B leaves8psi(x^(1/8)); counting all ideals leaves bounded rounding error. Neither computation introduces a new independent odd-prime signal or a route to RH.

Genuine next questions: (1) give quantitative smoothed/local-column truncation bounds with jump conventions fixed, rather than treating numerical truncations as certificates; (2) seek a naturally defined arithmetic operator explaining the local polynomial/traces, not merely the generic companion-matrix repackaging used for checking; (3) investigate orders with odd conductor or another precisely defined object if genuinely new odd-prime factors are wanted. Existing conceptual G-reciprocity and outside-referee questions are not closed by this calculation.
