Unverified 2026

Alexander-polynomial routing regularizer

Implementation & benchmark of arXiv:2608.12484 — Monodromy of plane curve singularities and quiver mutation

Usefulness5/10

Difficulty5/10

Novelty7/10

Regularization Stability Generalization Attention Moe routing Transformer Combinatorics Linear algebra Algebra

Source paper: Monodromy of plane curve singularities and quiver mutation arXiv:2608.12484 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Use the normalized determinant of a routing or attention interaction matrix as a global spectral signature. Penalize abrupt changes in this Laurent-polynomial signature when the model learns or dynamically rewires its interaction graph, preserving global connectivity patterns while still allowing local edge adaptation.

Formulas

$$D_m(t)=(1-t)D_{m-1}(t)+tD_{m-2}(t),\qquad D_0(t)=1,\qquad D_1(t)=1-t$$

$$D_m(t)=\sum_{j=0}^{m}(-t)^j$$

$$\det\mathcal{E}_{A_{2g}}(t)=1-t+t^{2}-\cdots+t^{2g}=\frac{1+t^{2g+1}}{1+t}$$

$$\mathcal{L}_{\mathrm{sig}}=\sum_{r=1}^{R}\left\|c_r(A_{\mathrm{new}})-c_r(A_{\mathrm{old}})\right\|_2^2,\qquad \det(I-\rho A)=\sum_{r=0}^{n}c_r(A)\rho^r$$

Mathematical statement

For the tridiagonal $A_{2g}$ Euler matrix, the paper defines $D_m(t)$ as the determinant of its $m\times m$ truncation and derives the exact recursion $D_m(t)=(1-t)D_{m-1}(t)+tD_{m-2}(t)$, with $D_0(t)=1$ and $D_1(t)=1-t$. It then obtains $D_m(t)=\sum_{j=0}^{m}(-t)^j$, hence $\det \mathcal{E}_{A_{2g}}(t)=1-t+t^2-\cdots+t^{2g}=(1+t^{2g+1})/(1+t)$. More generally, the paper computes determinants by matchings: for the $E_6$ example, matching counts $(m_0,m_1,m_2,m_3)=(1,5,5,1)$ give $\det E=(1-t)^6+5t(1-t)^4+5t^2(1-t)^2+t^3$. The determinant and, more robustly, the cokernel module are invariant under invertible row and column changes up to Laurent units. The adaptation treats the coefficients or evaluations of $\det(I-\rho A)$ as a differentiable global routing signature.

Implementation notes

Integrate this at the router or attention-logit level. Let $A\in\mathbb{R}^{n\times n}$ be the row-stochastic routing matrix after softmax, with $A_{ij}$ the probability that token or node $i$ communicates with $j$. Form a damped Euler matrix $E(\rho)=I-\rho A$, where $\rho\in(0,1/\|A\|_2)$ is fixed, and compute a short determinant signature without differentiating a full symbolic polynomial. Use $R=4$ probe values $\rho_r$, evaluate $s_r=\log|\det(E(\rho_r)+\varepsilon I)|$ using a differentiable LU decomposition, and penalize $\mathcal{L}_{\mathrm{sig}}=\sum_r(s_r^{\mathrm{new}}-s_r^{\mathrm{old}})^2$ across adjacent training checkpoints or before and after a routing mutation. For sparse large graphs, estimate log determinants with Hutchinson probes: $\log\det E\approx \frac{1}{K}\sum_k z_k^{\top}\log(E)z_k$, using a truncated Chebyshev approximation to $\log$. The exact mathematical transfer is the determinant recursion and its interpretation as a global invariant; the estimator and damping are engineering approximations. Pseudocode: compute router matrix $A$; for each $\rho_r$, compute $s_r=\operatorname{slogdet}(I-\rho_rA)$; store the previous signature; add $\lambda\|s-s_{\mathrm{prev}}\|_2^2$ to the task loss; update the stored signature with an exponential moving average. First run on a 4-layer, 8-head small transformer on WikiText-2, comparing standard attention routing, entropy regularization, and this signature penalty. Track perplexity, gradient norm, attention-collapse rate, and routing churn. Success is lower perplexity or equal perplexity with substantially fewer abrupt attention-pattern changes and fewer unstable training runs.

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.