TL;DR: Quaternion, octonion, sedenion, and Clifford-algebra neural networks keep appearing in the literature with claims of parameter efficiency and special expressivity. We spent two months building them, benchmarking them — and honestly trying to make them win. Whenever we added one properly matched real-valued baseline, the advantage evaporated. Every time. We ended up proving the failure is inevitable, not accidental, killed one of our own headline results along the way, and distilled the whole thing into a one-line rule any reviewer can apply in 30 seconds. Preprint DOI: 10.5281/zenodo.21623615 — code and raw per-seed results: github.com/mxkuzn/hypercomplex-placebo-audit

The seduction

If you’ve never met the Cayley–Dickson ladder, it’s genuinely beautiful. Start with real numbers. Double them: complex numbers. Double again: quaternions — multiplication stops commuting, and you get 3D rotations for free. Double again: octonions — multiplication stops associating, which sounds broken but is mathematically profound. Once more: sedenions, where you meet zero divisors — nonzero numbers that multiply to zero.

Each level has properties that feel tailor-made for machine learning. Norm-preserving multiplication? Exactly what you want for stable recurrent networks. Non-commutativity? Order-sensitivity for free. Zero divisors? A built-in “mutual exclusion” primitive — perfect for logical contradictions, surely?

The literature agrees. Quaternion CNNs, octonion knowledge-graph embeddings, parameterized hypercomplex multiplication (PHM) layers, hypercomplex adapters for LLMs — a steady stream of papers reports wins, usually framed as “same accuracy with 1/n the parameters.”

We drank the Kool-Aid too. Our project ran dozens of experiments trying to build reasoning modules on these algebras. Early results looked great. One of them — “octonion state-space models dominate non-Abelian group tracking” — was our headline for a while.

The control the field forgot

Here’s the uncomfortable question almost no hypercomplex paper asks: matched against what?

The typical baseline is a plain dense network or a diagonal recurrence. But a hypercomplex layer isn’t just “a network” — it’s a compact, structured, non-diagonal bilinear operation. If you compare it against something unstructured, you can’t tell whether the win comes from the exotic algebra or from mundane structure.

So we added four boring controls, each parameter-matched: (1) a real rotation built from Householder reflections — isolates “is it non-commutativity, or just a faithful rotation?”; (2) a shuffled structure tensor — take the algebra’s multiplication table and randomly permute it; if shuffled works equally well, the specific algebra is irrelevant; (3) real holographic binding (HRR, 1995 technology) — isolates “is it the algebra, or just binding?”; (4) plain orthogonality — isolates “are zero divisors special, or is dot(x,y)=0 an equally good exclusion primitive?”

Then we re-ran everything: associative memory, group tracking, Bach chorale modeling, liquid neural control, contradiction detection, attention variants, autoencoders. Three to five seeds each, pre-registered kill criteria.

The result across the original eight tasks: adding control #1 alone removed the claimed advantage in 8 out of 8.

Our own headline died the same way. The octonion model that “dominated” non-Abelian group tracking had only ever been compared to a diagonal SSM. Against a param-matched real rotation it lost outright — 0.38 vs 1.00 on the dihedral group D4. The octonion won only on groups that literally embed inside the quaternions. That’s not algebra magic; that’s structure-match.

The steelman that failed

Maybe our tasks just didn’t need non-associativity? Fine. We built the most non-associative target we could find: the associator [a,b,c] = (ab)c − a(bc) from the exceptional Jordan algebra — an object that is identically zero for any associative algebra. If anything needs octonions, it’s this.

A parameter-matched real bilinear network learned it just as well as the octonion network (cosine 0.991 vs 1.000). The lesson generalizes and, we think, is the most transferable idea in the paper: non-associativity in the data generator does not require non-associativity in the learned representation. Any fixed algebra product is just a fixed real bilinear tensor — a real universal approximator absorbs it.

And there is no cleverer algebra waiting to be discovered: Hurwitz’s theorem and Frobenius’s theorem prove that normed division algebras exist only in dimensions 1, 2, 4, 8, and every rung of the ladder trades away a property (order, commutativity, associativity, division). “Best of all worlds” is forbidden by classification theorems from the 19th century.

The one exception — and what it really was

One result genuinely surprised us. On a state-tracking task, a quaternion recurrence trained on sequences of length 12 extrapolated to length 1024 at 0.992 accuracy — while tight dimension-matched real controls (learned SO(4), Householder products) failed even on the training distribution.

Representational magic? No. SO(4) provably contains the quaternion solution, and some real-control seeds occasionally found it. SGD just reaches it reliably only inside the quaternion parameterization. The advantage is an optimization basin — “hard to find, easy to install” — not a capability moat. And when we benchmarked against the field’s actual SOTA for long sequences (the LRU, a diagonal complex recurrence), the quaternion lost at half the parameters.

Ironically, the best-performing algebra across our whole study is the most boring non-trivial one: complex numbers. Dimension two. Everything above it pays for properties that turn out to be either reproducible in the reals or actively harmful.

From “it loses” to “it cannot win”

The last third of the paper stops asking whether the algebras win and asks whether they can. Some highlights, each with exact constants rather than vibes. Reflections are unreachable: in every Cayley–Dickson algebra, multiplication by a unit element has determinant greater than zero — if your domain’s symmetry group contains a reflection (music’s transposition/inversion group does), it provably does not live in the algebra’s multiplication. Zero divisors can’t do 3-way exclusion: three mutually annihilating unit sedenions do not exist — the best achievable “annihilation” for a triple is exactly 0.500, for a quadruple exactly 1/sqrt(3); meanwhile plain orthogonality gives you 16-way exclusion in the same 16 dimensions, at zero parameters. Non-associativity forfeits the parallel scan: modern long-sequence models get log-depth training from prefix scans, which require associativity — for octonions/sedenions the scan doesn’t approximate the answer, it computes a different function (130–143% measured disagreement). And you can’t group your way out: tensor products of these algebras collapse into ordinary matrix algebras (we verify HxH is isomorphic to M4(R) with error exactly 0.0) — which is what your GPU already multiplies.

What actually survived

Two things, and neither is an algebra.

First, the methodology. If you take one thing from this article: any structured-layer paper that doesn’t report a param-matched real structured control and a shuffled-structure control hasn’t shown anything yet. Our checklist is five lines long and would have prevented every false positive we chased — including our own.

Second, exactness. The only component that ever hit 1.000 accuracy with zero hallucination at every sequence length was a compiled exact table — not learned, not approximated. On real bAbI data, an exact symbolic front-end beat a trained end-to-end monolith by +0.5 accuracy while every VSA (hypercomplex or real) plateaued around 0.51. Whatever the future of neuro-symbolic systems looks like, our data says the moat is exact computation, not exotic arithmetic.

Why publish a negative result?

Because we couldn’t find this paper when we needed it. Two months of GPU time, ~35 controlled experiments, several “wins” that dissolved under one honest baseline — all of that would have been saved by someone else’s version of this write-up. (In fairness, the field is converging: a controlled study of Clifford-algebra layers published this July independently reached the same conclusion by the same method — scalarization baselines match or beat geometric algebra except under deep group composition.)

We also kept score on ourselves: the paper has a section listing every self-refutation our own controls caught, including one found during the final verification pass of the manuscript — a sedenion score that had been quietly dropped from a table because it broke a “clean” ordering. It’s back in the table, and the claim is weaker and truer.

If you work on structured architectures and think we got something wrong, the entire audit is reproducible: every number in every table is generated by a committed script with committed per-seed JSON results. Open an issue — that’s exactly the kind of comment we published this for.

Preprint: doi.org/10.5281/zenodo.21623615 — Code and raw results: github.com/mxkuzn/hypercomplex-placebo-audit