# Why octonion orders are strange at 2

11 November 2026, PST. Grant requested this investigation and a new site figure (pt194u). Math only; no RH or novelty claim. This continues the intermediate-order, odd-prime, global-zeta and mirror-line notes without changing their normalizations.

## 1. Plain-language answer

The real octonions have eight coordinates, but an **order** specifies which coordinate combinations count as integral. Adding certain half-coordinate vectors changes that arithmetic. At an odd prime, a half is harmless: 2 is invertible, so our four orders become the same algebra. At 2, a half cannot be removed that way.

Reducing modulo 2 gives four 256-element multiplication systems, not four copies of the same system. The smallest order G becomes **commutative and associative**. The largest, the Coxeter/E8 order O, remains a genuine **noncommutative, nonassociative split Cayley algebra**. In between, nilpotent ideals shrink while the span of associators grows. “Split” does **not** mean degenerate: O has a nonsingular norm even in characteristic 2.

All four remain alternative and satisfy the Moufang identities. Characteristic 2 does not break those identities. Only G has local numerator zeros on Re s=1/2 in this family; the comparison does not establish a zero-forcing rule from a radical dimension, an index, or an associator rank.

## 2. Objects, ideals and conventions

Use one integral chain

    G ⊂ S1 ⊂ R ⊂ O,
    successive indices 2, 2, 4; [S:G] = 1, 2, 4, 16.

There are seven R copies and seven maximal O copies among the 16 orders; the earlier exact symmetry classification lets one representative of each kind suffice. Our representatives use order numbers 0,1,8,12 and the exact lattice bases recorded in `two1111/results.json`. S1=G+Zω, ω=(1+e1+...+e7)/2. **The integral inclusions do not induce embeddings of their mod-2 reductions.** Do not interpret the diagram as a chain of F2-subalgebras.

A right ideal I is an additive subgroup with I*S contained in I (an F2-subspace in the reduction). It is not just a set aS: nonassociativity matters. Associators are (x,y,z)=(xy)z−x(yz), and A is their **additive span**, not an assumed ideal. Define N explicitly as the largest two-sided ideal nilpotent with respect to **all bracketings**: N^n is spanned by every n-fold product, recursively N^n=sum_{i=1}^{n−1} N^i N^(n−i). This is a radical-like nilpotent ideal; for associative G/2G it is the ordinary Jacobson radical. We do not silently import an associative Jacobson-radical definition for the other three algebras.

## 3. Exact structural table

**[project, computer-assisted]** Exact structure constants, nilpotent powers, polar forms, associator tensors and exhaustive ideal enumeration give:

| S | G | S1 | R | O=E8 |
|---|---:|---:|---:|---:|
| [S:G] | 1 | 2 | 4 | 16 |
| polar norm determinant over Z | 256 | 64 | 16 | 1 |
| polar norm rank over F2 | 0 | 2 | 4 | 8 |
| dim N | 7 | 6 | 4 | 0 |
| dims N,N²,N³,... | 7,4,1,0 | 6,3,0 | 4,0 | 0 |
| quotient by N | F2 | F2×F2 | M2(F2) | split Cayley algebra |
| commutative mod 2? | yes | no | no | no |
| associative mod 2? | yes | no | no | no |
| dim associator span over F2 | 0 | 3 | 4 | 7 |
| dim commutator span over F2 | 0 | 6 | 7 | 7 |
| dim nucleus / centre over F2 | 8 / 8 | 4 / 1 | 1 / 1 | 1 / 1 |
| number of right ideals, including 0 and S | 47 | 91 | 9 | 2 |
| number of two-sided ideals | 47 | 41 | 6 | 2 |
| number of square-zero elements, including 0 | 128 | 64 | 64 | 64 |
| number of norm-one units | 128 | 64 | 96 | 120 |

The row about square-zero elements is not a row about nilpotent ideals: O has 64 square-zero elements but **no nonzero nilpotent ideal**. Its nonzero isotropic elements do not destroy simplicity.

The quotient identifications have short finite checks. S1/N has dimension 2 and all four elements are idempotents, hence is F2×F2. R/N is an associative, noncommutative, simple 4-dimensional F2-algebra; the finite simple associative algebra theorem identifies it with M2(F2) (eight idempotents). For O, nonsingularity and multiplicativity of its norm survive reduction; there are 135 nonzero norm-zero elements. It is therefore the split Cayley algebra by **[known]** Cayley classification over a finite field. This last identification does not rest on similarity of the name E8.

In all four examples, N equals

    {v : b_q(v,w)=0 for every w, and q(v)=0}.

This subset is verified to be a linear, two-sided nilpotent ideal. All two-sided ideals were enumerated, so maximality was checked, not guessed from norm degeneracy. Alternatively, each displayed quotient has no nonzero nilpotent ideal, which proves maximality after the quotient identification. For S1, A=N² (dimension 3); for R, A=N (dimension 4). For O, A is the 7-dimensional trace-zero space. For G, A=0 modulo 2.

### Direct G proof [project, proved]

The integral imaginary basis units square to −1 and distinct ones anticommute. Their commutators and associators have even coefficients. Modulo 2 the signs disappear, multiplication becomes the elementary-abelian group table, and

    G/2G ≅ F2[(Z/2)^3] ≅ F2[a,b,c]/(a²,b²,c²).

For example take a=1+e1, b=1+e2, c=1+e3 in our Fano convention. The augmentation ideal (a,b,c) is N; its powers have dimensions 7,4,1,0. Every element of N squares to zero, although products of distinct elements of N need not vanish. Thus “each element squares to zero” is not the same statement as “N²=0”. The original integral G is still nonassociative: its associators first reappear modulo 4.

### Identities at 2 [known / project, proved]

All reductions satisfy x(xy)=(xx)y, (yx)x=y(xx), flexibility, and the Moufang identity (xy)(zx)=(x(yz))x. These are polynomial identities inherited from the real octonions by restriction to an integral order and then quotienting. They do not require dividing by 2. Our integral tensor also checks alternativity exactly: repeated-index associators vanish and the tensor is alternating over Z. In characteristic 2 an alternating trilinear tensor is symmetric under transpositions, but still vanishes when arguments repeat; that does not make it zero. The quadratic identity survives as well:

    x² − t(x)x + q(x)1 = 0,   t(x)=b_q(x,1),
    bar(x)=t(x)1−x.

The familiar orthogonal decomposition using t(x)/2 is what fails at 2, not these division-free identities.

## 4. How much associativity fails modulo 4 and 8

Z/4 and Z/8 are not fields, so there is no honest vector-space dimension to report there. Form the integral 8×512 matrix with columns all basis associators. Trilinearity makes this the full additive associator span. Its seven nonzero Smith invariant factors are:

| S | invariant factors | log2 size mod 2 | log2 size mod 4 | log2 size mod 8 |
|---|---|---:|---:|---:|
| G | 2,2,2,2,2,2,2 | 0 | 7 | 14 |
| S1 | 1,1,1,2,2,2,2 | 3 | 10 | 17 |
| R | 1,1,1,1,2,2,2 | 4 | 11 | 18 |
| O | 1,1,1,1,1,1,1 | 7 | 14 | 21 |

**[project, computer-assisted]** If a nonzero invariant has valuation v, its image modulo 2^k has size 2^max(k−v,0); multiply these sizes. For example, G's mod-4 span is (Z/2)^7 and its mod-8 span is (Z/4)^7. All four integral spans have rational rank 7. The table is an additive measure, not a normalized real-valued “amount of nonassociativity”, nor a prediction about zeros.

## 5. Why this prime is exceptional, and the historical construction

For G, q(x)=sum x_i² and its polar matrix is 2I. Modulo 2, q is the linear function sum x_i on F2-points, while its polar form is zero. This reduction is a **degenerate composition system**, not a Cayley algebra in the usual nonsingular definition. Adding half-coordinate vectors changes the integral Gram matrix, until O has polar determinant 1. O's characteristic-2 norm is nonsingular and isotropic, hence split; its multiplication does not collapse to an associative group algebra.

**[project, proved]** The polar rank sequence also follows conceptually from the indices. Let S be one of these integral overlattices, and S# its polar dual. Since G⊂S, S#⊂G#=(1/2)G, hence 2S#⊂G⊂S. Thus the discriminant group S#/S has exponent 2: all polar Smith invariants are 1 or 2. The determinant is 2^8/[S:G]^2, so

    rank_F2 b_q = 2 log2 [S:G] = 0,2,4,8.

This controls the norm rank, not the zero count.

**[known]** Orthogonal Cayley–Dickson doubling in the standard integral basis produces G and a diagonal sum-of-squares norm. One must not use a characteristic-not-2 polarization formula with 1/2 at 2, or conclude from this degenerate model that all characteristic-2 octonions are degenerate. Petersson's *Integral octonions*, printed pp. 2–3, makes the distinction explicit and gives a nonorthogonal half-coordinate description

    Cox(O)=Ga(H) ⊕ Ga(H)p,  p=(1+i+j+l)/2,

where O=H⊕Hl and l²=−1. Ordinary nonsingular integral quaternion-doubling does not by itself construct this positive-definite unimodular Z-algebra; Petersson explains the obstruction and points to his 2017 nonorthogonal construction.

**[known, historical source access stated]** Coxeter, *Integral Cayley numbers*, Duke Math. J. 13(4) (1946), 561–578, DOI 10.1215/S0012-7094-46-01347-6. Its original full text was not accessible in this run; Petersson pp. 2–3 explicitly attributes the corrected closed order to Coxeter/Bruck and points to Coxeter §4. Kirmse's naive E8-shaped lattice was not closed under octonion multiplication; having an E8 lattice is not, by itself, enough. Baez's book review, printed pp. 10–11, actually read, describes G, the half-sum/Kleinian order, seven double-Hurwitzian orders, and the half-coordinate maximal orders.

Conway–Smith, *On Quaternions and Octonions* (2003), provides the requested reading pointers: ch. 9 *Octavian Integers*, §9.1 p.99, §9.2 p.100, §9.3 p.105; also §6.3 *Doubling Laws* p.69, §6.9 *Other Fields* p.76, §7.4 *Different Forms of the Moufang Laws* p.88, §11.2 *The E8 Lattice Modulo 2* p.135 and §11.4 *The Remaining Subrings Modulo 2* p.140. **Only the publisher preview/contents were accessible, not those chapter bodies.** They are pointers, not falsely claimed quotations. The UCR file named `conway_smith.pdf` is Baez's review, not the book.

Finally, 2 is exceptional **for this family** because all the displayed indices are powers of 2 and the standard polar matrix contains 2. Maximal O has no exceptional ideal-count factor at 2: its factor is 1/(1−2^(−8s)), as at odd primes. Other nonmaximal orders can be made with odd-primary index; this is not a claim that octonion orders can only be bad at 2.

## 6. From the reductions to ideals and the numerators

Let B_S(x) count right ideals in S/2S by codimension. Taking inverse images counts precisely the integral right ideals **containing 2S**, not all ideals of index a power of 2. Both algorithms give coefficients at x^0 through x^8:

    G:  1,1,7,7,15,7,7,1,1
    S1: 1,2,15,35,15,8,7,7,1
    R:  1,0,3,0,1,0,3,0,1
    O:  1,0,0,0,0,0,0,0,1.

These right-subspace counts are independent exact checks of the first visible layer of the old ideal enumeration. R/N=M2(F2) explains its three smallest proper right ideals (codimension 2); its one-sided simple factors have dimension 2. O has no proper nonzero right ideal, in agreement with **[known]** Petersson 2021, Theorem 4.1(i), valid over any commutative base for genuine rank-8 composition algebras. G and S1 have 1-dimensional simple factors; S1's two codimension-one ideals agree with its F2×F2 quotient.

With x=2^(−s), the previously proved local factors remain

    F_G = Q_G/(1−x^4),
    Q_G = 1+x+7x²+7x³+14x⁴+14x⁵+28x⁶+8x⁷+16x⁸;
    F_S1 = Q_S1/(1−x^4),
    Q_S1 = 1+2x+15x²+35x³+14x⁴+14x⁵;
    F_R = (1+3x²)/(1−x^4),
    F_O = 1/(1−x^8).

B_S is **not** Q_S, nor the whole F_S. For example the x^4 coefficient of B_G and B_S1 is 15, while each Q has 14. The first missing full-count lifts occur at index 32 for G and S1: B_G has 7 there but F_G has 15; B_S1 has 8 but F_S1 has 16. The denominator periodicity and the higher lifts must be supplied separately. For R and O the displayed B agrees with the full series through degree 8, but the full series continues indefinitely.

### Same mod-2 algebra, different cyclic lifting data [project, proved; computer-checked]

A useful warning beyond this four-kind family: T_m=Z*1+2^m O, m=1,2, are orders with the **same** reduction F2⊕V, dim V=7, V²=0. Nevertheless they have respectively **0 and 128 right ideals with cyclic additive quotient Z/8**. This is a statement about that quotient type, not an uncomputed assertion about the total index-8 coefficient.

Choose an integral O basis 1,b1,...,b7 and write b_i b_j=c_ij,0+sum_l c_ij,l b_l. Any cyclic quotient character can be normalized by phi(1)=1; the unique mod-2 character forces phi(2^m b_i) even. For m=1, dividing the product congruences by 4 and reducing mod 2 forces alpha_i=phi(2b_i)/2 mod2 to define a unital algebra homomorphism O/2O→F2. There is none: the split Cayley algebra is simple and has dimension 8. For m=2, all basis products vanish under phi modulo 8 and the square equations force phi(4b_i) to be 0 or 4; all 2^7 choices work. Stability of a cyclic right kernel is equivalent to these character equations, since right multiplication on Z/8 acts by scalars. The independent verifier exhausts all 4^7 even-value assignments in both cases. Thus mod-2 multiplication alone does not determine all higher right-ideal lifting data.

### Do the invariants explain half-line zeros?

The comparison, per one prime-2 period, is:

| S | log2[S:G] | dim N | dim A mod2 | numerator roots on 1/2 / total |
|---|---:|---:|---:|---:|
| G | 0 | 7 | 0 | 4 / 8 |
| S1 | 1 | 6 | 3 | 0 / 5 |
| R | 2 | 4 | 4 | 0 / 2 |
| O | 4 | 0 | 7 | 0 / 0 |

**[project, computer-assisted comparison; earlier root count proved]** In this small family the only order with an associative mod-2 reduction is the only one with half-line zeros. More associators do not automatically create those zeros. The four data points do not establish that these invariants determine the root count, or a monotonic theorem beyond the displayed examples. The actual reason exactly four G roots lie on the circle remains the proved quartic/Sturm argument in `octonion-mirror-lines-20261111.md`; the palindrome pairs roots but does not force all eight onto the circle. We have not deduced that particular polynomial identity from the nilpotent radical. That conceptual bridge remains open, not a new conjectural RH statement.

## 7. Reproduction and verification

`two1111/compute.py` uses exact Fraction lattice products, integral structure constants, all 512 associators, integer Smith form, and a complete cyclic-closure/join enumeration of invariant subspaces. It checks alternativity over Z and enumerates all two-sided ideals, including strong nilpotent powers. Norm radicals are checked against N.

`two1111/verify_independent.py` reconstructs the Fano tensor separately, changes basis using an exact rational matrix inverse, first takes the integral column HNF before its Smith form, and **sieves all 417,199 subspaces of F2^8 for each kind**, in canonical RREF. This does not reuse the cyclic-submodule generation or multiplication table implementation. All right-ideal codimension counts, two-sided totals and associator invariants agree. Shared inputs are the lattice bases and the stated Fano orientation, not an independently sourced order classification. It also verifies the cyclic lifting counterexample.

Run in an environment with numpy and sympy:

    python two1111/compute.py
    python two1111/verify_independent.py
    python two1111/figure.py
    python two1111/test_acceptance.py

Exact outputs: `two1111/results.json`, `two1111/independent-verification.json`. Desktop/mobile SVGs: `figures/octonion-strangeness-at-2(-mobile).svg`. Figure labels distinguish known structure theory from finite project computations; the dots are counts, not zero coordinates. PNG still: `/workspace/math/octonion-strangeness-at-2-20261111.png`. Build/deploy receipt will be recorded in `two1111/live-receipt.json` and `/workspace/math/DEPLOY-LOG.md` after live verification.

## 8. Sources actually checked, with access limits

- H. P. Petersson, *Integral octonions*, Malaga workshop 2018, notes dated 1 January 2019, **printed pp. 1–3** (history, norm nonsingularity, nonorthogonal halves); Theorem 1 and Remark 2 for ideals. Public PDF: https://www.fernuni-hagen.de/mi/fakultaet/emeriti/docs/petersson/ass.-rem.-int.-oct.pdf . Read from cached primary PDF/text.
- H. P. Petersson, *Norm equivalences and ideals of composition algebras*, Münster J. Math. **14** (2021), 283–293, **Theorem 4.1(i), pp. 289–292**. https://doi.org/10.17879/06089651315 ; author PDF https://www.fernuni-hagen.de/mi/fakultaet/emeriti/docs/petersson/norm-equiv.-ideals-comp.-alg.pdf . Primary text read; this applies to O but not the degenerate G/S1/R reductions.
- A. Castillo-Ramirez and A. Elduque, *Some special features of Cayley algebras, and G2, in low characteristics*, **§1, equation (1), Table 1; §4 for characteristic 2**. https://arxiv.org/html/1503.05529 . Primary text read; arbitrary-field definition, quadratic identity, split basis, and the warning about dividing by 2.
- H. S. M. Coxeter, *Integral Cayley numbers*, Duke Math. J. **13(4)** (1946), **561–578**, https://doi.org/10.1215/S0012-7094-46-01347-6 . Metadata verified. Original body not accessible; the §4 pointer and historical account come explicitly from Petersson pp. 2–3, not an invented page quotation. The failed PDF fetch is retained locally as `coxeter-1946-fetch-block.html`, not misrepresented as a PDF.
- J. H. Conway and D. A. Smith, *On Quaternions and Octonions* (2003), **ch. 9, §§9.1–9.3, pp.99–105; §11.2 p.135; §11.4 p.140**. Section/page pointers verified in the publisher preview contents; chapter bodies not read. Publisher preview: https://api.taylorfrancis.com/content/books/mono/download?identifierName=doi&identifierValue=10.1201%2F9781439864180&type=googlepdf . We do not publish the downloaded preview.
- J. Baez, review of Conway–Smith, 12 August 2004, **printed pp.10–11**, actually read. https://math.ucr.edu/home/baez/octonions/conway_smith/conway_smith.pdf . This is a review, not the complete book.
