PaperlayerView exact revision on arXiv

Source-linked AI summary

Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops

Aakash Sarkar, Ananya Mukherjee, Motoshige Sato

arXiv:2610.37656v1cs.LGcs.AIphysics.comp-ph

TL;DR

The paper asks what relation-operator interactions are permitted by per-entity frame freedom on general relational graphs. It derives a forced Wilson-loop update and tests composition-based holonomy attention, which reaches 0.89 versus 0.63 for a single path on sparse S4 graphs at p=0.2.

  • Problem

    The paper asks what relation-operator interactions are permitted by per-entity frame freedom on general relational graphs.

    Source: §2 · p. 3

  • Method

    It derives cycle-based Wilson-loop scores and conjugation-equivariant edge updates, then evaluates holonomy attention using composition through shared vertices.

    Source: §2 · p. 3, §3 · p. 3

  • Results

    At p=0.2 on sparse S4 graphs, holonomy attention reaches 0.89 versus 0.63 for the single path, with cycle composition 26 to 45 points above it on every row.

    Source: §3 · p. 4

  • Takeaways & Limitations

    Relations compose through shared entities by a forced form; learning the operators themselves remains.

    Source: §4 · p. 5

    Generating operators from context rather than retrieving them from a library remains a next step.

    Source: §4 · p. 4

Abstract

from arXiv · show

Relations compose through the entity they share - Alice is Bob's mother and Bob is Carol's father, so Alice is Carol's grandmother - and the order matters, since a father's mother is not a mother's father; verifying a composed relation means closing the loop. We show that if each entity's coordinate frame may be rotated independently (per-entity frame freedom), three things follow. In the pure-gauge sector, relations that share no cycle cannot interact at all. The only gauge-invariant score of a triangle is a class function of its holonomy $H = T_{ij} T_{jk} T_{ki}$ - a Wilson loop. And the only equivariant update of an edge is $U = \Psi(H)\, T_{ij}$ with $\Psi$ conjugation-equivariant. Edge attention by Wilson loops is therefore not one design among many but the unique form the symmetry permits. Three communities hold pieces of this form: the commuting-operator embeddings of knowledge graphs are its abelian (commuting) sector; AlphaFold's triangle update and the Edge Transformer have its shape - an edge updated through the third vertex - without its constraints; and lattice QCD builds it equivariantly on lattices. We instantiate the equivariant sub-family on general relational graphs as holonomy attention. With no training and composition alone, it infers unobserved $S_4$ relations on sparse graphs at 0.88 with a fifth of the edges corrupted, at or slightly below specialist synchronization algorithms, where every commutative architecture is capped at 0.083; its holonomy score flags corrupted triangles at AUC 0.99 where commuting scores are capped at 0.76 by a second selection rule, and it separates clusters without labels.

1 Frame freedom and the transport sector

Per-entity frame freedom makes relation transports necessary for comparing entities across independently chosen coordinate systems. The paper maps this structure to lattice gauge theory, with entities as sites and relation triangles as Wilson loops.

1 Frame freedom and the transport sector

  • Per-entity frame freedom requires transported scores between entities and plaquette-mediated interaction between relation operators.The paper maps sites to entities, link variables to relation transports, and plaquette holonomies to Wilson loops of relation triangles.Source: §1 · p. 1
  • A closed path accumulates a holonomy whose trace is a Wilson loop independent of the coordinate system at each vertex.The paper motivates frame freedom by noting that entities introduced on the fly have no intrinsic coordinate system.Source: §1 · p. 1
  • Under O(d)^n symmetry, admissible attention scores depend on entities through x_i and T_ij x_j; in the flat limit, standard attention is recovered.When every loop has trivial holonomy, a shared frame exists and standard attention realizes the resulting dot-product form.Source: §1 · p. 2
  • The paper derives this gauge-theoretic structure from per-entity frame freedom on general relational graphs, extending prior lattice formulations.It identifies its contribution as a closed-form O(d) statement, including the role of reflections.Source: §1 · p. 2

2 Forced plaquettes between relations

In the pure-gauge sector, symmetry prevents relations from interacting without a cycle and restricts cycle-mediated interaction to Wilson-loop holonomies. The resulting edge update has a unique conjugation-equivariant form, which connects the paper’s general-graph derivation to existing models and methods.

2 Forced plaquettes between relations

  • In the pure-gauge sector, a relation edge cannot be moved by other edges unless they form a cycle.Gauge-fixing a cycle-free edge set sends every transport to I, making invariant functions constant; with entity vectors, an equivariant edge term can exist without a cycle.Source: §2 · p. 2
  • For a triangle, continuous gauge-invariant scores are exactly class functions of its holonomy H, invariant under conjugation.These functions see eigenvalues rather than basis; for several cycles, invariants depend on their holonomies under simultaneous conjugation.Source: §2 · p. 2
  • Every continuous equivariant edge update from a cycle through it has the form U=Ψ(H)T_ij, with Ψ conjugation-equivariant.For several cycles through an edge, the update is a function of their holonomies under simultaneous conjugation.Source: §2 · p. 3
  • Setting Ψ=H^-1 gives composition through a third vertex, while aggregating several vertices with class-function scores uses polar projection back to O(d).The projected aggregate remains in the forced family wherever the sum is invertible.Source: §2 · p. 3
  • Commuting-operator knowledge-graph embeddings occupy the abelian sector, where operators compose coordinate-wise and lack an edge-level update.The paper contrasts this with non-abelian, edge-level composition through shared vertices.Source: §2 · p. 3
  • AlphaFold’s and the Edge Transformer’s triangle updates have the shape of third-vertex edge updates but lack the forced form’s constraints.Their states are not transports, their bilinear maps are unconstrained, and their scores are not invariant.Source: §2 · p. 3
  • A single cycle detects inconsistency but cannot localize a wrong edge; localization requires scoring with loops that avoid the candidate edge.The square loop formed by two triangles can avoid the edge, as in the Edge Transformer’s logit design.Source: §2 · p. 3
  • The update is APE smearing with two-link staples and Wilson-loop weights; the paper’s addition is proving this is the only form on general graphs.It also presents the cross-literature reading and two selection rules predicted by the form.Source: §2 · p. 3

3 The non-abelian sector: holonomy attention

Holonomy attention updates relations by composing them through shared vertices and weighting candidate compositions by consistency around loops. On sparse S4 graphs it improves unobserved-pair inference over single-path composition, while its Wilson-loop score detects corrupted triangles and separates clusters without labels.

3 The non-abelian sector: holonomy attention

  • For an absent edge, candidate relation compositions are scored by the holonomy of a square loop that avoids that edge; scoring with a triangle containing it costs 5–12 points.The evaluation uses hidden S4 frames on fourteen entities, with observed relations corrupted with probability p.Source: §3 · p. 3
  • Holonomy attention is not a better synchronizer: Singer’s eigenvector method ties it or leads by 0–5 points with one reweighting pass, while cycle-edge message passing leads on a matched n=12 pilot.The pilot differences were 3, 3, 0, and 11 points at p=0.1, 0.2, 0.3, and 0.45, respectively, with standard error approximately 0.09.Source: §3 · pp. 3–4
  • At p=0.2 on sparse S4 graphs, holonomy attention reaches 0.89 accuracy on unobserved pairs versus 0.63 for a single path; commuting architectures are capped at 0.083.Composition through cycles is 26–45 points above the single path on every row.Source: §3 · p. 4
  • On dense graphs, holonomy weighting’s gain over uniform weighting opens as clean two-hop paths become a minority, and iterated reweighting performs better still.The uniform mean bends near p≈0.29; paired gains are +0.026±0.004 at p=0.3 and +0.078±0.006 at p=0.45.Source: §3 · p. 4
  • The Wilson-loop score tr H/d reaches AUC 0.99 for corrupted-triangle detection; on a two-cluster graph it separates within- from cross-cluster triangles at AUC 0.98 without group labels.For S4, the commuting score is 0.759 against the holonomy score’s 0.993.Source: §3 · p. 4

4 Discussion

The discussion joins node-level matter propagation with edge-level gauge propagation: vectors use transported messages, while relation operators compose through shared vertices and Wilson-loop scores. It points to context-generated operators and applications where relation order matters.

4 Discussion

  • A full layer alternates node updates using transported messages with edge updates that refine relations, then propagates matter again in the refined field.The post-attention resolvent is exact exactly when the effective connection is flat on the attention support.Source: §4 · p. 4
  • The current next step is to generate operators from context rather than retrieve them from a library; generated transports must come from relation data outside the entity pair and interact through Wilson loops.The forced interaction form is Ψ(H)Tij.Source: §4 · p. 4
  • The proposed non-abelian setting includes multi-plane lensing, cryo-EM orientation recovery, and kinematic chains, where relation order is physically attested.The passage reports that commutative emulators miss image rotation in multi-plane lensing, with 0.51 against 0.92.Source: §4 · p. 5

A Forced Form and Flat Limit (from [1])

With independent frames for each entity, invariant scores can use only frame-independent entity information, while equivariant messages and transports are tightly constrained. In the flat limit, the transports reduce to a shared-frame representation, whereas a standard head composes only through depth.

A Forced Form and Flat Limit (from [1])

B Proof of Theorem 1, the cubic coefficient, and the lensing cell

The lensing cell favors matrix composition over additive and elementwise emulators across three seeds, and reversing plane order reverses the label exactly. The proof derives the conjugation-equivariant update and its cubic coefficient.

the lensing cell

  • The 2×2 matrix-composition model scores 0.923/0.920/0.922 across three seeds, versus 0.507/0.515/0.510 for additive and 0.509/0.515/0.518 for elementwise emulators.Reversing the planes maps M to M^T and ω to −ω exactly.Source: B Proof of Theorem 1, the cubic… · p. 7

Proof of Theorem 1

the cubic coefficient

  • The cubic coefficient is −1/2: the cyclic commutator traces collapse to a single term, and the residual ratio converges numerically to 1.000 as ∥A∥ approaches zero.Source: B Proof of Theorem 1, the cubic… · p. 7
Loading 2610.37656v1…