Contribution
It derives, from one symmetry assumption, the form an attention mechanism over relation operators must take. If each entity’s frame can be changed independently and relations transform covariantly, relations interact only around cycles, and they interact through the holonomy. It reads three existing literatures as pieces of that form: commuting knowledge-graph embeddings as its abelian sector, AlphaFold’s triangle update and the Edge Transformer as its unconstrained superset, and lattice gauge smearing as its lattice case. It then measures an untrained instance, “holonomy attention”, on group synchronization with outliers. Its companion paper (LIT-1803, ref. [1]) supplies the node-level half.
Key insight
Fix the gauge along a spanning tree and see what is left. Every tree edge becomes I, so whatever is invariant lives on the edges that close cycles, and on a triangle the only freedom left is conjugating its holonomy: “Setting gᵢ = I, gⱼ = Tᵢⱼ, gₖ = TᵢⱼTⱼₖ sends the triple to (I, I, H), and the only freedom left is the diagonal, H ↦ gHg⁻¹” (p. 2). The rest of the paper follows from that move.
Assumptions
- Per-entity frame freedom. Each edge carries Tᵢⱼ ∈ O(d) with Tⱼᵢ = Tᵢⱼ⁻¹, and a gauge transformation acts by Tᵢⱼ ↦ gᵢTᵢⱼgⱼ⁻¹ with an independent gᵢ ∈ O(d) at every entity (p. 2). The paper says “The symmetry is the task’s, not a network’s: a fresh entity’s frame carries no information, so any predictor must be invariant to it” (p. 2). Its companion paper calls the same premise “an inductive bias, not a claim that fresh names logically imply a physical symmetry” (LIT-1803, p. 2).
- Invertible, orthogonal relations. Every theorem is for O(d). The companion paper’s CLUTRR control, where orthogonality costs 64.6 points out of distribution, is not discussed here.
- The pure-gauge sector. “No cycle, no interaction” holds for functions of the transports alone. With entity vectors, xᵢxⱼᵀ “is an equivariant edge term with no cycle” (p. 2).
- Given transports. In the experiments the relations are elements of S4 in its faithful permutation representation (d = 4), loaded “as a library” (p. 3). Nothing is learned: “No parameters are learned” (p. 3). Operators “generated from context rather than retrieved from a library” are “the next step” (p. 4).
- A smaller symmetry in the implementation. After each round the update is snapped “to the nearest group element in Frobenius norm” (p. 3). With the library fixed, that step is equivariant under the permutation gauge but not under arbitrary O(4) frames. The continuous step is the O(d)-equivariant one.
Key results
-
No cycle, no interaction (p. 2). On an acyclic edge set gauge fixing sends every transport to I, so “Any function of the transports invariant under O(d)^V … is therefore constant”, and an equivariant update of an edge is ±Tᵢⱼ by Schur’s lemma. “Two edges sharing a vertex but no closing edge form a tree, so no ‘global edge attention’ respects the symmetry.”
-
Plaquette invariants (p. 2). The orbits of a triangle are the conjugacy classes of H, and “the continuous invariants are exactly the class functions of H”, generated by tr Hʳ, r ≤ d (Procesi, ref. [8]). “The reflections are load-bearing”: in SO(2), R(±θ) have one characteristic polynomial and are not conjugate. For several cycles the invariants are functions of (H₁, …, H_b) under simultaneous conjugation.
-
Theorem 1 (pp. 3, 7). “Every continuous equivariant update of Tᵢⱼ from a cycle through it, with holonomy H based at i, is U = Ψ(H)Tᵢⱼ with Ψ(gHg⁻¹) = gΨ(H)g⁻¹, and Ψ in the closure of {Σᵣ cᵣ(H)Hʳ}, cᵣ class functions, intersected with O(d).” Ψ = H⁻¹ gives composition through the third vertex, U = TᵢₖTₖⱼ; several third vertices with class-function scores give U = Π(Σₖ αₖTᵢₖTₖⱼ), Π the polar projection, equivariant “wherever the sum is invertible” (p. 3).
-
The abelian sector (p. 3). Commuting operators are simultaneously diagonalizable, so RotatE, ComplEx and DistMult compose coordinate-wise; there “edges couple only linearly, through the synchronization Laplacian”.
-
A design rule (p. 3). “one cycle through an edge has the same holonomy whichever of its edges is wrong, so one loop detects but cannot localize; two triangles do, via the square tr(HₖHₗ⁻¹) that avoids the edge”. Scoring an edge by a triangle containing it (“literal”) costs 5–12 points (Table 1).
-
Table 1 (p. 4), accuracy on unobserved S4 pairs, 1000 instances per cell:
protocol p single path uniform holonomy literal spectral spectral+1 sparse, deg. 5 0.2 0.625 0.872 0.886 0.753 0.897 0.912 dense, deg. 8 0.2 0.645 0.982 0.988 0.935 0.987 0.990 dense, deg. 8 0.3 0.488 0.912 0.937 0.820 0.933 0.949 dense, deg. 8 0.45 0.309 0.623 0.700 0.559 0.678 0.753 Paired holonomy − uniform: +0.014, +0.006, +0.026, +0.078. Commuting architectures are capped at |S4ᵃᵇ|/|S4| = 2/24 = 0.083.
-
Corruption detection (p. 4). tr H/d separates corrupted from clean triangles at AUC 0.99, and the paper says why: “clean triangles score exactly 1, one corrupted edge never closes a loop, so the only ties are triangles with two or more cancelling corruptions (1.4% in closed form at p = 0.3), and the AUC sits at its ceiling 0.993”. A commuting score sees only G/[G, G], capping its AUC at ½(1 + (|G| − |[G, G]|)/(|G| − 1)): 0.76 for S4 (measured 0.759), chance for A5.
-
Clusters (p. 4). On a two-cluster graph with random cross-cluster relations, the same score separates within- from cross-cluster triangles at AUC 0.98 without labels.
-
Lensing (p. 7). On twelve-plane weak lensing, the sign of the image rotation is predicted at 0.92 by 2×2 matrix composition and at 0.51 by additive or elementwise emulators, which reversal symmetry holds at ½.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Under per-entity O(d) frame freedom, transport-only invariants live on cycles and are class functions of holonomies | strong | §2 gauge-fixing argument; standard (Procesi; Gao et al., ref. [29]; Giles, ref. [31]) |
| C2 | Every continuous equivariant edge update from one cycle is Ψ(H)Tᵢⱼ with Ψ conjugation-equivariant | strong | Theorem 1, proof in App. B; the form of Ψ by Procesi’s polynomial classification and Stone–Weierstrass, sketched |
| C3 | Edge attention by Wilson loops is “the unique form the symmetry permits” | weak as worded | C2 fixes a family; the particular scores, softmax, mean and polar projection are choices within it |
| C4 | Composition through cycles is the capability on S4 synchronization | strong for this benchmark | Table 1: 26–45 points over a single path |
| C5 | Wilson-loop routing improves on uniform averaging where clean paths become a minority | moderate | Table 1 paired gains, +0.6 to +7.8 points; spectral+1 is higher on every row |
| C6 | The holonomy score is the invariant corruption statistic, at AUC 0.99 | strong but near-tautological | the AUC is the ceiling the clean-is-exactly-1 construction sets, as the paper computes |
| C7 | Standard attention is the flat-connection limit, and AlphaFold’s pair stack is learned flattening | weak | a reading of shape (App. A), and a prediction the paper proposes but does not test (p. 4) |
Method
Holonomy attention (p. 3). For an absent edge (i, j), each mediator k gives a candidate Cₖ = TᵢₖTₖⱼ. Score each candidate by its mean agreement with the others, sₖ = mean over l ≠ k of tr(CₖCₗ⁻¹)/d, where CₖCₗ⁻¹ = TᵢₖTₖⱼTⱼₗTₗᵢ is the holonomy of a square that avoids (i, j). Set αₖ ∝ exp(βsₖ), β = 8, aggregate Π(Σₖ αₖCₖ), and after four rounds snap to the nearest library element. Observed edges are held fixed.
Concepts
- per-entity frame freedom — independent O(d) changes of coordinates at each entity, with relation transports covariant.
- holonomy, Wilson loop — the product of transports around a closed path, and its trace.
- pure-gauge sector — functions of the transports alone, with no entity vectors.
- abelian sector — commuting transports, where composition is coordinate-wise and edges couple only linearly.
- literal scoring — scoring a candidate edge by a triangle that contains it; self-confirming in the paper’s sense.
Connections
The construction is standard in two literatures, as the paper says. Lattice gauge theory has had plaquette-mediated, conjugation-equivariant updates since APE smearing (1987), with recent gauge-equivariant networks. Group synchronization uses cycle inconsistency as “the standard diagnostic” (refs. [11], [30]), and Gao, Brodzki and Mukherjee (ref. [29]) classify edge potentials by holonomy. Singer’s eigenvector method (LIT-1655, read in NOTE-1224) is the paper’s own baseline and ties or beats it. The U(1) case is the twisted Laplacian of THEORY-1082. The paper cites Bodnar et al.’s neural sheaf diffusion (ref. [21]) as gauge structure “with no interaction between edge maps”.
Bearing on the record
- QUESTION-130. The paper is QUESTION-130’s construction (local frames, T′ = g T g⁻¹, holonomy around a loop, conjugacy class as the invariant) with the forced form of an equivariant update added (THEORY-1140). It does not answer the question. QUESTION-130’s first objection is that arbitrary transports make curvature vacuous, and here the transports are supplied. It is supporting literature for the question’s “gauge- covariant comparison”, which A19 already rated “mathematically constructible”.
- QUESTION-282. Another formalism whose correspondence structure is an input. The paper names learning the operators as its next step.
- THEORY-1082 and CLAIM-471. The non-abelian generalization of the twisted Laplacian. For d > 1 it brings something U(1) lacks: a non-flat connection can still keep a shared subspace on which every loop is trivial (CLAIM-495).
- CLAIM-342 and THEORY-1088. The design rule (score a candidate only by loops that avoid it, since a loop through it is self-confirming) is an instance of CLAIM-342’s demand for evidence that does not pass through what it tests. It is a weak instance: several squares can share a mistaken edge. Learning transports by minimizing holonomy alone is exposed to THEORY-1088’s way of reaching consensus by changing the maps.
- CASE-179. Holonomy attention is a candidate revision apparatus for CASE-179’s “revision by consistency alone” arm, and the experiment is an instance of that arm’s warning: the data are built flat, so consistency and correctness coincide by construction.
- ML practice. The paper proposes an architecture and a design rule
for it. That is the anthology’s business, and the LIT carries
anthology-candidate.
Limitations
- The experiments test the theory’s sector, not attention. The task is synchronization of given group elements, the representation is supplied, and nothing is learned. Its own sentence: “Holonomy attention is not a better synchronizer” (p. 3).
- The AUC of 0.99 and the cluster AUC of 0.98 measure what the data were built to contain. Clean triangles are exactly flat and cross-cluster relations are random.
- “Unique” holds at the level of Theorem 1’s family. The weights, the softmax, the mean and the projection are choices within it.
- Standard attention as “the flat-connection limit” holds under a representational reading. A standard head’s W_QᵀW_K is not orthogonal and is “covariant only under the global O(d)” (App. A).
- The premise is asserted to be “the task’s” here and called an inductive bias in the companion paper. Neither paper tests it on a task where it could fail, apart from the companion’s CLUTRR control, where it does fail.
Open questions
- Does AlphaFold’s pair representation lose triangle inconsistency with depth and recycling? That is the paper’s own prediction (p. 4), and it is testable on released models.
- Do operators generated from context, not drawn from a library, keep the routing gain on a task with no group given in advance?
- What replaces Wilson loops when the transports are not invertible, as the companion paper’s CLUTRR resets are not?
Note of 2026-10-30: plain spectral synchronization trails holonomy at p = 0.45
The Connections section says Singer’s eigenvector method “ties or beats” holonomy attention. By the table above, that holds for spectral synchronization with one reweighting pass, which is ahead on every row. Plain spectral synchronization trails holonomy weighting on three of four settings: 0.987 against 0.988, 0.933 against 0.937, and at p = 0.45 0.678 against 0.700. It leads only on the sparse graph, 0.897 against 0.886. Found by the status review of this date in narrowing THEORY-1158.