# Multiplicity-free Schur channel tying

- ID: 1735
- Canonical URL: https://synthcore.org/idea/1735/multiplicity-free-schur-channel-tying
- API JSON: https://synthcore.org/api/idea/1735.json
- API Markdown: https://synthcore.org/api/idea/1735.md
- Verification status: unverified
- Source: [arXiv:2608.02881](https://arxiv.org/abs/2608.02881)
- Category: architecture
- Solves: memory, scalability, generalization
- ML areas: mlp, embedding, attention
- Math tags: representation-theory, combinatorics, algebra
- Ratings: usefulness 4/10; difficulty 5/10; novelty 7/10

## Idea description

Tie neural parameters across feature channels according to the Schur multiplicity pattern of the fermionic coinvariant representation. In the two-fermion case, use one learned parameter block for each Schur degree instead of independently parameterizing every ordered pair of fermionic channels, eliminating redundant copies while preserving the relevant GL2 channel symmetry.

## Mathematical statement

For two auxiliary grading variables u,v, the paper defines [r]_{u,v}=u^{r-1}+u^{r-2}v+...+v^{r-1} and uses s_(a,b)(u,v)=(uv)^b[a-b+1]_{u,v}. Its type-B two-fermion Hilbert series is a sum of single Schur polynomials: Hilb(R_Bn^(0,2);u,v)=sum over the stated bipartitions of binom(n,i)K_{lambda,(1^i)}K_{mu,(1^(n-i))}s_(n-ell-lambda_1,ell+lambda_2)(u,v). Here lambda and mu are partitions forming a bipartition of n, i=|lambda|, lambda=(lambda_1,lambda_2), mu=(2^ell,1^(n-lambda_1-lambda_2-2ell)), and K are Kostka numbers. Each Schur term is a single GL2 representation rather than an arbitrary multiplicity. The engineering interpretation is to parameterize channel mixing by Schur blocks indexed by (a,b), with basis dimension a-b+1, instead of using an independent matrix for every channel ordering.

## Key formulas

- $$[r]_{u,v}=u^{r-1}+u^{r-2}v+\cdots+uv^{r-2}+v^{r-1},\qquad s_{(a,b)}(u,v)=(uv)^b[a-b+1]_{u,v}.$$
- $$\operatorname{Hilb}(R_{\mathfrak{B}_{n}}^{(0,2)};u,v)=\sum_{\substack{(\lambda,\mu)\vdash n,\\ \lambda=(\lambda_1,\lambda_2),\ i=|\lambda|,\\ \mu=(2^{\ell},1^{n-\lambda_1-\lambda_2-2\ell})}}\binom{n}{i}K_{\lambda,(1^{i})}K_{\mu,(1^{n-i})}s_{(n-\ell-\lambda_1,\ell+\lambda_2)}(u,v).$$
- $$W_{a,b}=\operatorname{span}\{u^{a-t}v^{b+t}:0\le t\le a-b\},\qquad \dim W_{a,b}=a-b+1\quad(a\ge b).$$

## Implementation notes

Apply this to a two-stream feature mixer, such as two views, two augmentation streams, two heads, or two fermionic feature channels. Let h^(1) and h^(2) have identical token dimensions. Instead of concatenating them and using a dense unconstrained projection, construct homogeneous monomials u^(a-t)v^(b+t) for each selected pair a>=b; represent the corresponding channel subspace W_(a,b) with a shared parameter matrix B_(a,b) applied identically to all t. Concretely, create q_(a,b)[t]=Mix(h^(1),h^(2)) projected to degree (a-t,b+t), stack q over t, and compute y_(a,b)[t]=sum_s B_(a,b)[t,s]q_(a,b)[s], with B shared across tokens and optionally constrained to commute with the GL2 action. Generate the polynomial basis recursively from the displayed [r]_(u,v) formula. For a small model, use degrees (1,0), (1,1), (2,0), and (2,1), retaining only blocks whose activation variance exceeds a threshold. For small n, compute Hilbert coefficients from the Kostka formula; for larger n, estimate graded feature ranks from activation covariance. Compare against an untied MLP mixer with equal FLOPs on two-view classification or signed-set regression. Success is a 20–40% reduction in mixer parameters and activation memory at equal accuracy, or higher accuracy at equal parameter count. Also test stream swapping: the tied model should degrade less than the untied baseline when the two inputs are exchanged.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.