{
 "artifacts": null,
 "category": "architecture",
 "description": "Replace a dense token- or channel-mixing matrix with a product of local braid generators acting on adjacent coordinates. Each generator is an exactly invertible 2-by-2 transformation, while the braid and far-commutativity identities give multiple equivalent factorizations of the same global operator. This creates a sparse, reversible mixer with O(kn) cost for a braid word of length k, rather than O(n^2) cost for a dense matrix.",
 "formulas_latex": [
  "$$G=M(\\sigma_i)|_{i,i+1}=\\begin{pmatrix}0\u00261\\\\-1\u00262\\end{pmatrix},\\qquad b_i=a_{i+1},\\quad b_{i+1}=-a_i+2a_{i+1}.$$",
  "$$G^{-1}=M(\\sigma_i^{-1})|_{i,i+1}=\\begin{pmatrix}2\u0026-1\\\\1\u00260\\end{pmatrix},\\qquad b_i=2a_i-a_{i+1},\\quad b_{i+1}=a_i.$$",
  "$$M(\\sigma_i)M(\\sigma_{i+1})M(\\sigma_i)=M(\\sigma_{i+1})M(\\sigma_i)M(\\sigma_{i+1}),\\qquad M(\\sigma_i)M(\\sigma_j)=M(\\sigma_j)M(\\sigma_i)\\ \\text{if }|i-j|\\geq2.$$",
  "$$M(\\beta)=M(\\sigma_{i_k})\\cdots M(\\sigma_{i_1}),\\qquad M(\\beta)\\mathbf{1}=\\mathbf{1}.$$"
 ],
 "id": 2804,
 "implementation": "Integrate the mixer at the token-mixing or channel-mixing line of a small Transformer or MLP, immediately before the feed-forward block or in place of a dense linear projection. Let the input be X in R^(B x n x d). Choose a braid word w = ((i_1,s_1),...,(i_k,s_k)), where i_t is an adjacent pair in {1,...,n-1} and s_t in {+1,-1} selects the generator or inverse. For each batch and feature dimension, apply this pseudocode: `Y = X; for (i,s) in word: if s == +1: (Y[...,i],Y[...,i+1]) = (Y[...,i+1], -Y[...,i] + 2*Y[...,i+1]); else: (Y[...,i],Y[...,i+1]) = (2*Y[...,i] - Y[...,i+1], Y[...,i]); Y = Y / sqrt(2)`. Use the same scale convention in every branch and optionally learn one scalar gate g_t per crossing, initialized at zero, with `Y \u003c- Y + g_t*(cross(Y)-Y)`. For exact reversibility, omit gates and use the explicit inverse word in reverse order. The paper supplies the local update, inverse, braid identities, and invariant constant direction; empirical choices are word length, index pattern, sign pattern, normalization, and gating. Log the ratio ||Y||_2/||X||_2 and compute singular values of the composed operator on a probe matrix. First run on CIFAR-10 with a 6-layer, width-256 MLP-Mixer or a 4-layer, 8-head Transformer. Compare against a fixed permutation, depthwise convolution, butterfly mixing, and a dense learned n-by-n mixer at matched FLOPs. Measure accuracy, loss at equal optimizer steps, wall-clock throughput, activation-norm drift, and parameter count. Success means comparable or better accuracy with at least 2x lower mixer FLOPs and bounded activation norms; failure means norm explosion, severe rank deficiency, or slower loss descent than the permutation baseline.",
 "math_summary": "Let B_n be the braid group generated by sigma_1,...,sigma_{n-1}. Over a field F_p, the paper maps each generator to the identity except on coordinates i and i+1, where M(sigma_i) has block G = [[0,1],[-1,2]]. Thus for a vector a in F_p^n, the local update is b_i = a_{i+1} and b_{i+1} = -a_i + 2a_{i+1}, with all other coordinates unchanged. The inverse generator uses G^{-1} = [[2,-1],[1,0]], so every crossing is reversible. The representation satisfies M(sigma_i)M(sigma_{i+1})M(sigma_i) = M(sigma_{i+1})M(sigma_i)M(sigma_{i+1}) and M(sigma_i)M(sigma_j) = M(sigma_j)M(sigma_i) whenever |i-j| \u003e= 2. For a word beta = sigma_{i_1}...sigma_{i_k} read top-to-bottom, M(beta) = M(sigma_{i_k})...M(sigma_{i_1}). The all-ones vector 1 is fixed by every generator because G(1,1)^T = (1,1)^T. Adapt these matrices over the real numbers, with optional normalization, and apply them independently to token or feature dimensions.",
 "math_tags": [
  "representation-theory",
  "algebra",
  "linear-algebra",
  "dynamical-systems"
 ],
 "ml_areas": [
  "transformer",
  "mlp",
  "initialization",
  "training-dynamics"
 ],
 "paper": {
  "arxiv_id": "2608.29046",
  "arxiv_url": "https://arxiv.org/abs/2608.29046",
  "summary_what_math_gives_to_ml": "The paper constructs an explicit finite-dimensional representation of the braid group using local, invertible 2-by-2 crossing operators, and proves that these operators satisfy braid and far-commutativity relations exactly. The transferable asset is not Fox coloring itself, but a principled library of local reversible channel-mixing transformations whose global behavior is invariant under braid-word rewrites. These operators can replace or augment token-mixing, channel-mixing, or reversible residual blocks, providing structured mixing with linear cost in the number of local crossings rather than quadratic cost for a dense matrix. The most direct experiment is a braid-word mixer inserted into a small Transformer or MLP and compared with permutations, butterfly mixing, and dense linear mixing at equal parameter count and FLOPs.",
  "title": "Fox $p$-Colorings as Fixed Points of Braid Representations",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 4,
  "novelty": 7,
  "usefulness": 5
 },
 "solves": [
  "speedup",
  "stability",
  "scalability"
 ],
 "title": "Braid-word reversible mixer",
 "url": "https://synthcore.org/idea/2804/braid-word-reversible-mixer",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}