Exterior-power truncation for 2D tensor channels
Implementation & benchmark of arXiv:2608.18328 — The Brauer category $\mathcal{B}(2)$ has principal graph $D_\infty$
Source paper: The Brauer category $\mathcal{B}(2)$ has principal graph $D_\infty$ arXiv:2608.18328 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Use the dimension-specific relation A_3=0 to remove all intermediate channels transforming as the third exterior power of the two-dimensional vector representation. In tensor-product attention or equivariant MLPs, this is an exact algebraic pruning rule rather than approximate low-rank compression.
Formulas
Mathematical statement
The paper defines an antisymmetrizer A_3 and proves a leaf relation by setting A_3=0. For V=R^2, the alternating subspace \Lambda^3V is zero, so the antisymmetrizer on three vector slots vanishes. Explicitly, A_3(u\otimes v\otimes w)=\sum_{\pi\in S_3}\operatorname{sgn}(\pi)u_{\pi(1)}\otimes v_{\pi(2)}\otimes w_{\pi(3)}=0. The paper also gives the recursion P_v\otimes\mathrm{id}_1=P_{v+1}+F, where P_v is a recursively defined diagrammatic morphism and F is the complementary diagram; the stated traces are tr(P_v)=2, tr(P_{v+1})=2, and tr(P_v\otimes\mathrm{id}_1)=4. The implementation should first exploit the exact A_3=0 identity: every fully antisymmetric set of three vector indices is discarded before materializing a feature. Higher-order antisymmetric index sets of size at least three are likewise zero for a two-dimensional representation.
Implementation notes
(1) Integrate at feature-basis construction in a tensor-product equivariant layer, polynomial attention kernel, or order-three message-passing block. Do not allocate a channel for the fully antisymmetric part of any triple of 2D vector features. If the implementation initially forms a rank-three tensor T[a,b,c], use the antisymmetrizer only as a debugging check and omit that subspace in production. (2) Pseudocode: form candidate T; compute A3T=(T[a,b,c]+T[b,c,a]+T[c,a,b]-T[b,a,c]-T[a,c,b]-T[c,b,a])/6; assert ||A3T|| is below numerical tolerance; retain T-A3T or, preferably, construct a reduced basis directly from symmetric products, pairwise contractions, and the oriented scalar det(u,v). For order greater than three, canonicalize antisymmetric index subsets and discard every subset of size at least three. (3) The exact mathematical input is dim(V)=2 and A3=0. The empirical quantities are the measured norm of A3T, activation memory, and FLOPs. The P_v recursion can be used as a basis-generation heuristic: expand order v by adding one vector factor and subtract the complementary component F, checking the trace normalization stated above. (4) First experiment: implement order-three polynomial attention on randomly rotated 2D point clouds with hidden width 128. Compare a dense rank-three tensor basis against the truncated Brauer/exterior basis at equal training steps. Record peak activation memory, forward FLOPs, loss variance, equivariance error, and accuracy on held-out rotations/reflections. A successful result is 20–30% lower rank-three intermediate storage with identical accuracy and equivariance, or improved optimization stability from eliminating null channels; failure is a significant accuracy drop indicating that the removed component was not purely antisymmetric.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.