# Counting ideals of the Coxeter integral octonions: what zeta comes out

Night Block 3, Wednesday 11 November 2026, 12:38 AM to 12:53 AM PST (box clock). Notes, scripts and figures only. Nothing deployed, the site and `zeta-zeros-build` untouched, no messages sent.

## Summary

- **[known]** Every one-sided ideal of the Coxeter integral octonions O is two-sided, and the two-sided ideals are exactly mO (m ≥ 1). The source I read directly is H. P. Petersson, "Integral octonions", Malaga workshop notes (2018, dated 1 Jan 2019), Thm. 1(a),(b) and Remark 2. It attributes the statement to Petersson–Racine (monograph in progress, Exc. 87, 107) and says it generalizes Mahler 1942 and van der Blij–Springer 1959, p. 414, "which reaches the conclusion of (b) for the Coxeter octonions". Allcock, J. Algebra 220 (1999) 396–400, is cited as another proof. Copy: `octonion-sources/petersson-ass-rem-int-oct.pdf`.
- **[project, proved, double-computed]** An independent self-contained check: the Z-algebra generated by the left-multiplication matrices L_x (x ∈ O) is **all of M_8(Z)**, with index 1. The same holds for right multiplications, and for both together. So any additive subgroup L ⊂ O with O·L ⊂ L is an M_8(Z)-submodule of Z^8, which forces L = mO. This works over every Z_p at once, because M_8(Z) ⊗ Z_p = M_8(Z_p), and no separate discriminant step is needed.
- **[project, proved]** The "left-ideal counting zeta" is therefore Σ_L [O:L]^{-s} = Σ_m m^{-8s} = **ζ(8s)**. If the ideal norm N(mO) = m² is used (the gcd of the element norms, the analogue of the reduced norm), it is ζ(2s). The same series counts right ideals and two-sided ideals.
- **[project] Critical lines:** none of this gives a new line. ζ(8s) has its nontrivial zeros on Re s = 1/16, but only if RH is true, and they are the ordinary zeta zeros rescaled by 1/8. The element-count (E8 Epstein) zeta 240·2^{-s}ζ(s)ζ(s−3), already in `octonion-analogues.md` and not re-derived here, has zeros on Re s = 1/2 and 7/2, again only under RH. Both are **inherited**. Neither has a local factor, so neither has anything like the "local" ½-line zeros of quaternion orders.
- **[project, proved — computer-assisted, double-computed; Block 4, Nov 11, see `gravesian-periodicity-proof.md`]** The non-maximal *Gravesian* order Z^8 does have a 2-local factor, Q(2^{-s})/(1−2^{-4s}), where Q has degree 8 and satisfies an exact s ↦ 1−s symmetry. Four of its zeros lie on Re s = ½ and four lie off it, at Re s ≈ 1.183 and −0.183. So for this octonion count, the quaternion-Eichler "all local zeros on ½" pattern fails. This is a finite local factor and says nothing about RH.

## Sources: what I could and could not read

| Source | Status | What it gives |
|---|---|---|
| Petersson, *Integral octonions* (Malaga 2018 notes) | **read** (PDF saved) | Thm 1: two-sided ideals ↔ ideals of k; over a Dedekind k every one-sided ideal is two-sided. Proof sketch: over a field, Zorn matrices show one-sided ideals are 0 or C; over a DVR, Nakayama. Says Mahler's description (principal, generated by a multiple of an element of norm 1, 2 or 4) is "explicit but still incomplete": not every such element gives an ideal. |
| Gross, "Embeddings into the integral octonions", Pacific J. Math. 181 (1997) | **read** (PDF saved) | Explicit basis (1.1) of the Coxeter order R for the table e_i e_{i+1} = e_{i+3}, plus "R/pR is an octonion algebra over Z/pZ for all primes p". The basis was used for method A. |
| Mahler, Proc. Roy. Irish Acad. A 48 (1942) 123–133 | **not read** | Content known only through Petersson's summary and a secondary web page (Köplinger). |
| van der Blij–Springer, Indag. Math. 21 (1959) 406–418 | **not read** (ScienceDirect returned only a bot-check page) | Known only through Petersson Remark 2 (p. 414 result). |
| Allcock, J. Algebra 220 (1999) | **not read** | Cited by Petersson as an alternative proof. |
| Conway–Smith, *On Quaternions and Octonions* | **not read this block** | HandWiki's octonion page repeats "all left and right ideals are two-sided; only nO" without a primary citation. |

## Computation (two independent codes)

**Method A** (`octonion_ideal_count.py`, outputs `octonion-ideal-count-A-results.json`, `-A-run.txt`, matrices in `octonion-ideal-count-matrices.json`). Builds the octonions from the Fano table e_i e_{i+1} = e_{i+3} with Gross's basis (1.1). Results:
- closed under multiplication;
- Gram determinant 1 for (x,y) = N(x+y) − N(x) − N(y);
- 240 norm-1 elements, counted two ways: by closure of the unit loop, and by enumerating every vector with entries in {0, ±1/2, ±1} that lies in the lattice;
- exact integer HNF closure of the Z-algebra generated by the L_x has rank 64 and index **1** in M_8(Z), and the same for R_x and for L and R together. Dimension mod p is 64 for p = 2, 3, 5, 7, 11, 13, which follows from index 1 anyway.

**Method B** (`octonion_ideal_count_B.py`, `octonion_ideal_count_B2.py`, outputs `octonion-ideal-count-B-results.json`, `-B2-results.json`, run logs). Uses a different construction, Cayley–Dickson doubling of the Hamilton quaternions, and finds the maximal order **without** a literature basis: starting from ±basis, it searches for a half-unit whose Moufang-loop closure has 240 elements (seed (1+e1+e2+e4)/2 in CD coordinates). Results:
- Gram determinant 1 and 240 units;
- **brute force over every nonzero vector of O/pO for p = 2, 3, 5** (projectively, up to 97,656 vectors): the left submodule it generates, and likewise the right submodule, is always 8-dimensional. So O/pO is a simple left module and a simple right module, and there are no nonzero proper left-stable subspaces;
- **brute force over all 65,280 surjective functionals Z^8 → Z/4**: 0 kernels are left-stable, so there are no left ideals with quotient Z/4. This is consistent with the theorem, under which the only left ideal of 2-power index ≤ 2^7 would be O itself;
- the F_p-span of words in the L_x reaches 64 for p = 2 at word length 5 (dims by length 0–5: 8, 28, 56, 63, 64, counting the identity), and for p = 3 at length 3. p = 2 needs longer words; this is not an obstruction.

**Elementary argument, checked** (`octonion_ideal_count_checks.py`, output `octonion-ideal-count-checks.json`):
- [known, alternative-algebra identity] x̄(xy) = N(x)y. It was verified exactly on 2000 random pairs. Consequence [project, proved]: if a left ideal L contains an x with p ∤ N(x), then L ⊇ N(x)O, so L + pO = O.
- This argument alone does not exclude ideals made of norm-≡0 (isotropic) vectors mod p, since the norm form mod p has 4-dimensional totally singular subspaces. That gap is exactly what the brute-force simplicity check and the index-1 computation close.
- **Corollary [project, proved; checked]:** for x of norm 2, xO has index N(x)^4 = 16, which is not an 8th power, so xO is never a one-sided ideal. All 1298 norm-2 elements in the coefficient box [−2,2]^8 were tested (that box does not contain all 2160 in this skew basis): none of xO or Ox is left- or right-stable. This is the precise sense in which Mahler's "generated by an element of norm 2 or 4" can fail, as Petersson notes.

**Proof status.** [project, proved] The argument is complete. Index 1 is an exact integer-HNF computation, independently corroborated by the mod-p simplicity brute force. "M_8(Z)-submodules of Z^8 of finite index are mZ^8" is standard (Morita equivalence, or directly: the submodule is stable under all elementary matrices, so it equals I·Z^8 for the ideal I of first coordinates). It agrees with the [known] Petersson–Racine / van der Blij–Springer theorem. **No novelty claim**: the computation re-verifies a known result.

## Contrast with the quaternion case  [known + project]

| | Hurwitz order H (associative) | Coxeter order O (nonassociative) |
|---|---|---|
| one-sided ideals | many: p+1 right ideals of reduced norm p (p odd), one of norm 2^e | only mO |
| ideal-counting zeta (index) | ζ(2s)ζ(2s−1)(1−2^{1−2s}) [known, Voight §26; project-checked] | ζ(8s) |
| zeros | inherited on Re s = 1/4, 3/4 (if RH) + local zeros of the 2-factor on Re s = 1/2 (exact) | inherited only, Re s = 1/16 (if RH) |
| element count by norm | 24·Σ_{d∣n, d odd} d [known, Hurwitz/Jacobi] | E8: 240σ_3-type, Epstein 240·2^{-s}ζ(s)ζ(s−3) [known] |

The reason [project, proved here for O; known in general]: in H the left multiplications generate only a copy of H inside M_4 (an algebra of rank 4, not 16), leaving room for many submodules. In O, nonassociativity means the left multiplications generate everything (L_x L_y ≠ L_{xy}), and that kills the ideal theory. This is the Azumaya-like behaviour Petersson's Remark 2 points to.

## Control: the Gravesian (integer-coordinate) octonions DO have a local factor at 2  [project, computational, double-computed]

G = Z^8, with every coordinate an integer, is a non-maximal order of index 16 in O. For odd p, G ⊗ Z_p = O ⊗ Z_p, so all new structure sits at p = 2.

**Counts.** a_k is the number of left-stable sublattices of index 2^k; right-stable gives identical numbers.

| k | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | … 18 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| a_k | 1 | 1 | 7 | 7 | 15 | 15 | 35 | 15 | 31 | 15 | 35 | 15 | 31 | then period 4: 15, 35, 15, 31 |

Three codes:
1. `octonion_gravesian_count.py`: brute force over dual functionals mod 2 and mod 4, k ≤ 2. Gives 1 and 7, with 0 cyclic Z/4 quotients. The same code returns 0, 0 on the Coxeter order O as a control.
2. `octonion_gravesian_levels.py`: pure-Python HNF breadth-first search through index-2 steps, k ≤ 18, left and right. Outputs `octonion-gravesian-levels-{left,right}.json`.
3. `octonion_gravesian_levels_pari.py`: PARI mathnf, with a *direct* stability test H⁻¹L_iH ∈ M_8(Z) on every index-2 sublattice, k ≤ 12. Output `octonion-gravesian-levels-pari.json`. It agrees with code 2 at every k.

The breadth-first search is complete because G/2G is local. In characteristic 2 the signs vanish, so G/2G ≅ F_2[(Z/2)^3] ≅ F_2[x,y,z]/(x²,y²,z²), using the [known] twisted-group-algebra description of O (Albuquerque–Majid; not re-read tonight). Its only simple module is F_2, so every stable lattice has a chain of stable index-2 steps. Structure cross-check: that local ring has 1 ideal of codimension 1 and 7 of codimension 2, matching a_1 and a_2.

**Local factor.** With x = 2^{-s}, the 2-part of Σ_L [G:L]^{-s} is P(x) = Σ a_k x^k. Within the computed range (k ≤ 18),

P(x) = Q(x)/(1 − x⁴),  Q(x) = 1 + x + 7x² + 7x³ + 14x⁴ + 14x⁵ + 28x⁶ + 8x⁷ + 16x⁸.

Update, Block 4 (Nov 11, 1:40 AM PST): this now holds for all k, **[project, proved — computer-assisted]**. The multiplication algebra contains 8·M_8(Z), so primitive stable lattices have index ≤ 2^24; three runs to index 2^26 then pin the numerator as Q(x)(1+x⁴) over 1−x⁸. Proof and records: `gravesian-periodicity-proof.md`, `block4/`. (Earlier wording: a [project conjecture] supported by 18 + 12 levels.)

**Functional equation [project, exact on Q].** Q(x) = (2x²)⁴ Q(1/(2x)), equivalently c_{8−k} = 2^{4−k} c_k. This is the s ↦ 1 − s symmetry of the factor (x ↦ 1/(2x)), the same palindromic shape as the quaternion Eichler/Solomon polynomials.

**Zeros of the factor [project, double-computed].** Script `octonion_gravesian_factor.py`, output `octonion-gravesian-factor-results.json`. numpy roots, mpmath polyroots at 50 digits, and an exact PARI Sturm count agree. The exact count works on P(y)P(−y) in Y = y², y = √2x, then W = Y + 1/Y, and finds exactly 2 of the 4 W-roots in [−2,2]. Result:
- **4 of the 8 zeros lie on Re s = 1/2**, at Im s ≈ ±1.52014 and ±3.56122 (mod 2π/log 2 ≈ 9.0647);
- **4 zeros lie OFF the half-line**, at Re s ≈ 1.18337 and −0.18337, Im s ≈ ±2.30895. They form a functional-equation mirror pair, since |x₁|·|x₂| = 0.44032 · 1.13554 = 1/2.

The off-line zeros are zeros of one finite local factor: they say nothing about RH or about ζ. This is not a "new line" and not a counterexample to anything proved. It does show that the quaternion Eichler "circle theorem" pattern (every local zero on Re s = 1/2) **does not carry over** to this non-maximal octonion order. [analogy] In Grant's language: the Gravesian factor has "local" zeros on the half-line *and* a mirror pair that has left it, like the leave-the-line event in zero-movement.md §5, but occurring in an actual arithmetic count, not a model family.

## Plain statement for Grant

When you count ideals, octonions give nothing new. The integer octonions have almost no one-sided ideals, only the obvious ones made by multiplying everything by a whole number, so their ideal count is just the ordinary zeta function with s stretched by 8. Its zeros are the usual zeta zeros moved onto the line Re s = 1/16, if the Riemann hypothesis is true. That is an inherited line, not a new one. Quaternions are different: they have many genuine one-sided ideals, and that is where the extra "local" zeros on the half-line come from. The cruder integer-coordinate octonions, which are not the best ring, do have extra ideals at the prime 2. Their extra zeros are half on the half-line and half off it, in a mirror pair.

## Figures

- `figures/octonion-ideal-count.{svg,png}` (desktop) and `figures/octonion-ideal-count-mobile.{svg,png}` (360 px), made by `octonion_ideal_count_figures.py`: "what counting gives which zeta" for H vs O.
