NOTE-1218: Torres-Hugas et al. — Cycle holonomy and remote synchronization

Contribution

It puts a U(1) connection Laplacian (the “twisted Laplacian”) beside the Kuramoto-type phase models used for remote synchronization. It states the spectral criterion for when edge-wise phase relations can be met globally, gives the isolated-cycle spectrum, and builds an associated connection whose holonomies reproduce the cycle mismatches that a symmetric Sakaguchi–Kuramoto lag produces on a cactus graph. On the one motif worked through, a triangle and a pentagon sharing a vertex, it shows that the pentagon’s maximal mismatch falls exactly where the locked branch becomes a two-group state, and close to but after the branch’s transverse instability. The paper’s own framing is that “effective higher-order dynamical constraints emerge naturally on ordinary graphs, provided the interaction carries nontrivial topological structure” (abstract, p. 1).

Key insight

A constraint can belong to a cycle and to none of its edges. Each edge relation zⱼ = e^{iαᵢⱼ}zᵢ can always be met on its own, but going round a cycle composes them, and if the accumulated phase is not 0 mod 2π no assignment of nonzero states meets them all: “global compatibility is obstructed not by local pairwise mismatches, but by intrinsic topological frustration on cycles” (abstract). The second insight is the paper’s own caution against its framing: the geometric point where that frustration is largest is not where the synchronized pattern loses stability. “The twisted Laplacian thus encodes the geometric reorganization of the cycle-compatibility structure, whereas the Jacobian determines the stability of the nonlinear phase-locked state” (p. 7).

Assumptions

• A connected graph with a fixed reference orientation, and a phase αᵢⱼ on each oriented edge with αⱼᵢ = −αᵢⱼ (eq. 1, p. 2). The orientation “is only a bookkeeping convention” (p. 2).
• The relaxed linear setting for the spectral results: complex states zᵢ ∈ ℂ, not unit-modulus phases (p. 2). “ker Lα ≠ {0} if and only if the edge compatibility conditions can be satisfied globally in the relaxed linear setting” (p. 3). Unit amplitude is imposed only afterwards (eq. 11).
• Identical units and identical coupling. Eq. (12) has identical intrinsic frequencies (in a co-rotating frame); eq. (14) has one frequency ω, one coupling κ and one lag α, on an unweighted graph.
• A uniform lag in the gauge model means αᵢⱼ = α “for every edge i → j in the chosen reference orientation” (eq. 13, p. 3), so its cycle holonomies depend on how the edges were oriented.
• Cactus graphs for the path–holonomy correspondence: “every edge belongs to at most one simple cycle” (p. 4). Graphs with overlapping cycles are left as an extension (p. 8).
• One motif for all the dynamics: the triangle–pentagon graph of Nicosia et al. (ref. [11]), with symmetry classes {1}, {2, 3}, {4, 7}, {5, 6} (p. 6), and only the branch “connected continuously to complete synchronization at α = 0” (p. 6; SM p. 9).

Key results

Gauge invariance and holonomy (§II A, p. 2)

• Holonomy, eq. (2): Hol_α(C) = Σ_{(i,j)∈C} αᵢⱼ (mod 2π), invariant under the gauge transformation zᵢ ↦ e^{iθᵢ}zᵢ, αᵢⱼ ↦ αᵢⱼ + θⱼ − θᵢ (eqs. 3, 4).
• “What remains invariant under such transformations is completely determined by the holonomies on a cycle basis. On trees, where no cycles exist, every phase lag configuration can be removed by a gauge transformation” (p. 2).

The twisted Laplacian and its zero mode (§II B, pp. 2–3)

• Twisted incidence, eq. (5): (δα z)ᵢⱼ = zⱼ − e^{iαᵢⱼ}zᵢ, which for α = 0 is the ordinary incidence operator, so the construction “directly generalizes ordinary diffusive coupling [18]” (p. 2; ref. [18] is Hansen and Ghrist).
• Energy, eq. (6): E(z) = ½ Σ_{(i,j)∈E} |zⱼ − e^{iαᵢⱼ}zᵢ|² = ½‖δα z‖², zero exactly when every edge constraint zⱼ = e^{iαᵢⱼ}zᵢ (eq. 7) holds.
• Lα = δα*δα (eq. 8), E(z) = ½⟨z, Lα z⟩ (eq. 9).
• The zero-mode criterion (p. 3): “When all cycle holonomies vanish, Eq. (10) admits nontrivial solutions. If some cycle has nonzero holonomy, then ker Lα = {0} and the smallest eigenvalue of Lα is strictly positive.” The abstract and discussion state it as “if and only if” (pp. 1, 7). The paper gives no proof beyond the energy argument; the result is the standard one for discrete magnetic and connection Laplacians (refs. [14]–[16]). The step it leaves out is short: on a connected graph a nonzero z meeting every edge constraint has all |zᵢ| equal and nonzero, is fixed along a spanning tree by its value at one node, and closes on each remaining edge exactly when that edge’s fundamental cycle has holonomy 0 mod 2π.
• Dynamics, eqs. (11, 12): with |zᵢ| = 1, E(ϕ) = Σ[1 − cos(ϕⱼ − ϕᵢ − αᵢⱼ)], whose gradient descent is ϕ̇ᵢ = κ Σ_{j∼i} sin(ϕⱼ − ϕᵢ − αᵢⱼ), “a Sakaguchi–Kuramoto-type flow in a co-rotating frame” (p. 3). The uniform antisymmetric connection “defines a reciprocal gradient system, whereas the symmetric Sakaguchi–Kuramoto model uses the same phase lag in both interaction directions” (p. 3).
• What the spectrum says about this nonlinear flow is stated loosely: “The eigenvectors associated with small eigenvalues describe the modes that best accommodate the incompatibility induced by nonzero cycle holonomy” (p. 3). Nothing is computed for eq. (12) itself.

Cycle mismatch in the symmetric Sakaguchi–Kuramoto model (§II C 1, pp. 3–4)

• The model, eq. (14): ϕ̇ᵢ = ω + κ Σⱼ Aᵢⱼ sin(ϕⱼ − ϕᵢ − α), “where the same phase lag α appears in both interaction directions.”
• Path mismatch, eqs. (15)–(18): for an edge (u, v) of a simple ℓ-cycle, the coupling argument along the edge is ϕᵥ − ϕᵤ − α and the sum along the complementary path is ϕᵥ − ϕᵤ − (ℓ − 1)α, so h_α(C) = (ℓ − 2)α (mod 2π), “independent of the phase configuration” and of which edge is chosen.
• The two models are distinct: “The two models have different local coupling laws … The correspondence therefore occurs at the level of path compatibility” (p. 3).

The associated connection on a cactus (§II C 2, p. 4)

• On each cycle C_r, assign +α to every edge of the complementary path from u_r to v_r and +α to the direct edge u_r → v_r, which contributes −α when the cycle is closed. Then Hol_α(C_r) = (ℓ_r − 1)α − α = (ℓ_r − 2)α (eq. 21), so Hol_α(C_r) = h_α(C_r) (eq. 22). Bridges get phase zero by a gauge choice, and the whole assignment is a single connection “uniquely up to a node-wise gauge transformation” (p. 4).
• The correspondence is by construction: the connection’s phases “are chosen so that their cycle holonomies reproduce the path mismatches induced by the symmetric Sakaguchi–Kuramoto lag” (p. 2).

The spectral encoding (§II C 3, pp. 4–5)

• λ₁(α) = min_{‖z‖=1}⟨z, Lα z⟩ (eq. 26), zero iff Hol_α(C_r) ≡ 0 for every cycle (eqs. 27, 28). “A positive value of λ₁(α) measures the minimum normalized residual incompatibility after optimization over all node states” (p. 5).
• Isolated cycle, eqs. (29, 30): the branches are 2 − 2cos((2πm + Hol_α(C_r))/ℓ_r), m = 0, …, ℓ_r − 1. λ₁ vanishes at holonomy 0 and is largest at holonomy π (eqs. 31, 32), where “two adjacent winding sectors become degenerate” (p. 5). I reproduced eq. (30) for ℓ = 3 and 5: at holonomy π, λ₁ = 2 − 2cos(π/ℓ), 1 for the triangle and 0.382 for the pentagon.
• For the whole cactus, λ₁(α) “is not, in general, the sum of the isolated-cycle eigenvalues, because cycles may interact through shared vertices and attached trees” (p. 5; Fig. 1, a 14-node cactus).

The triangle–pentagon motif (§II C 4–6, pp. 5–7; SM pp. 9–10)

• h_α(C₃) = α, h_α(C₅) = 3α (eqs. 33, 34). The pentagon reaches holonomy π first, at α = π/3, when the triangle’s is π/3 (eqs. 36, 37; Figs. 2, 3).
• Complete synchrony is excluded by the degrees, not the holonomy: equal phases would need Ω = ω − κkᵢ sin α for every i, “which cannot hold when the degrees kᵢ differ and sin α ≠ 0” (p. 6).
• The locked branch, with ϕ₂ = ϕ₃ = Ωt + x, ϕ₄ = ϕ₇ = Ωt + y, ϕ₅ = ϕ₆ = Ωt + z and z = y + x, satisfies eqs. (38, 39): 0 = sin(y + α) − 2 sin x cos α − sin α and 0 = 2 sin(x − α) + 2 sin(y − α) + sin(x + α) + sin α. Setting x = 0, y = z = α leaves sin α(2cos α − 1) = 0, so α = π/3, that is 3α = π and h_α(C₅) = π (eq. 40). There the seven nodes form two groups, ϕ₁ = ϕ₂ = ϕ₃ and ϕ₄ = … = ϕ₇ = ϕ₁ + π/3 (p. 7). I checked numerically that the branch from α = 0 passes through x = 0, y = π/3 at α = π/3.
• Stability, eq. (41) and SM: the transverse perturbations split by symmetry into a triangular mode η on {2, 3}, with rate λ = −[cos(x + α) + 2cos α], and a two-dimensional pentagonal block M⊥ on the pairs {4, 7} and {5, 6} (SM p. 9). The first zero of det M⊥ along the branch is at αc ≃ 1.0164, with xc ≃ 0.0631, yc ≃ 0.9657, zc ≃ 1.0288, σ(M⊥) ≃ {0, −1.7038} and η-rate ≃ −1.5247 (SM p. 10). “Hence this transverse crossing is also the first instability of the full linearized dynamics” (SM p. 10). My continuation gave the crossing at α ≈ 1.017 with the same branch values to two decimals.
• The result: αc ≃ 1.0164 < απ = π/3 ≃ 1.0472 (eq. 42). “The maximally frustrated pentagonal holonomy is therefore not a stability criterion … this configuration lies on the transversely unstable continuation of the locked branch” (p. 7).
• The physical phase differences still close around the cycle: h_α(C₅) = 3α “is not the sum of the physical phase differences. It is the mismatch between the coupling argument along the direct edge and the accumulated coupling arguments along the complementary path” (p. 7).

Discussion (pp. 7–8)

• “no monotone relation between an individual cycle holonomy and the global eigenvalue λ₁(α) should be expected in general” (p. 8). For the motif I computed λ₁ of the associated connection: 0.275 at α = π/3, with its local maximum near α ≈ 1.09, not at π/3, which fits the paper’s caution. The figure’s vertical line marks π/3 as the isolated-pentagon value, not as the motif’s maximum (Fig. 3 caption).
• Extensions named: heterogeneous lags, weighted graphs, overlapping cycles (p. 8). Code: github.com/llui2/twisted-laplacian; “No data were used” (p. 8).

Claims

id	claim	strength	support
C1	On a connected graph with an antisymmetric edge connection, Lα has a nonzero kernel iff every cycle holonomy is 0 mod 2π; on a tree every connection is gauge-trivial	strong	stated with an energy argument (pp. 2–3); standard result (refs. [14]–[16]); proof not written out; checked numerically on cycles
C2	For an isolated ℓ-cycle the twisted spectrum is 2 − 2cos((2πm + Hol)/ℓ), and λ₁ is largest at Hol = π, where two winding sectors are degenerate	strong	closed form, eqs. (29)–(32); reproduced for ℓ = 3, 5
C3	The symmetric Sakaguchi–Kuramoto lag gives each simple ℓ-cycle a phase-independent path mismatch (ℓ − 2)α	strong	three-line derivation, eqs. (15)–(18)
C4	On a cactus graph these mismatches are the holonomies of one associated connection, unique up to gauge	strong, but by construction	eqs. (19)–(24); the connection is built to match, so the correspondence is a representation, not a derivation of SK dynamics
C5	In the triangle–pentagon motif, the locked branch is an exact two-group state at α = π/3, where the pentagon’s mismatch is π	strong	exact algebra, eqs. (38)–(40); checked numerically
C6	That branch loses transverse stability in the pentagonal sector at αc ≃ 1.0164 < π/3, so the maximal-mismatch point is not a stability boundary	strong	numerical continuation and det M⊥ = 0, SM pp. 9–10; reproduced
C7	Remote synchronization “can be understood as a manifestation of effective higher-order constraints inherent in the network’s topology”	weak	proposed (p. 1); shown only as a geometric marker on one motif that coincides with a two-group state and is not the stability threshold
C8	λ₁(α) of the whole cactus measures how the cycles’ combined incompatibility is distributed	moderate	definitional (eq. 26) and Fig. 1; no relation to dynamics is shown, and none monotone is expected (p. 8)

Concepts

• Connection (here) — the collection α = {αᵢⱼ} of edge phase shifts with αⱼᵢ = −αᵢⱼ, acting as “a discrete parallel transport” (p. 2).
• Holonomy — the accumulated phase Σ αᵢⱼ around an oriented cycle, mod 2π (eq. 2); gauge-invariant; its sign flips with the cycle’s orientation.
• Gauge transformation — a node-wise phase redefinition (eqs. 3, 4). Connections related by one are gauge-equivalent.
• Twisted incidence operator and twisted Laplacian — δα and Lα = δα*δα (eqs. 5, 8); elsewhere the magnetic or connection Laplacian, and the U(1) case of a sheaf Laplacian.
• Topological frustration — a connection with some nonzero holonomy, so that the edge relations cannot be met together though each can be met alone (p. 2).
• Path mismatch h_α(C) — the difference between the coupling argument on one edge and the sum along the rest of its cycle, (ℓ − 2)α in the symmetric SK model (eq. 17).
• Cactus graph — a graph whose simple cycles share at most vertices (p. 4).
• Remote synchronization — “nodes synchronize despite being separated by intermediate, incoherent nodes” (p. 1); in the motif, the nonadjacent pair {4, 7} (p. 6).

Connections

• Pecora et al. 2014 (ref. [12]; LIT-960, read in NOTE-775) and Sorrentino et al. 2016 (ref. [13]): the symmetry clusters and the symmetry-separated transverse blocks the motif analysis uses (p. 6; SM p. 9).
• Hansen and Ghrist 2019 (ref. [18]; LIT-1659) and 2021 (ref. [19]; LIT-1646), both registered unread: the twisted incidence operator is said to generalize diffusive coupling “[18]” (p. 2).
• Shubin 1994 (ref. [14], discrete magnetic Laplacian), Singer and Wu 2012 (ref. [15]) and Bandeira, Singer and Spielman 2013 (ref. [16]): the mathematical lineage of the zero-mode criterion. None is held; the record’s Singer (LIT-1655, angular synchronization, 2011) is a different paper.
• Nicosia et al. 2013 (ref. [11]): the motif and the remote synchronization transition the paper says π/3 is “close to” (p. 2). Not held, and the comparison was not checked.
• Manik, Timme and Witthaut 2017 (ref. [21]): cycle flows and multistability in oscillator networks, cited for “the broader role of cycle constraints” (p. 1). Not held.

Bearing on the record

• CLAIM-471: the holonomy bullet is right in content, slightly overstated in warrant, and the paper is a weak case for “generate”. The claim’s line that “g₁₂g₂₃g₃₁ ≠ 1 forbids a consistent nonzero assignment though every edge is satisfiable” is what the paper says (p. 2: “the local phase relations are globally incompatible … that cannot be resolved edge by edge”), with the qualification “mod 2π” and in the relaxed linear setting (p. 3). “Prove” is too strong for this paper: it states the criterion (pp. 1, 3, 7) with an energy argument and leaves the cycle-basis step unwritten; the result is a standard one it cites (refs. [14]–[16]), so it is safe, but the paper is not its proof (THEORY-1082). For the claim’s thesis, note that the connection here is never produced by an interaction: the gauge model takes α as given (eq. 1), and the SK correspondence builds an associated connection to reproduce mismatches (p. 2, eq. 21). In the claim’s own table this is a correspondence rule; the gradient flow (12) is its enforcement; nothing is evidence. So it illustrates the “composed correspondences determine what is admissible” half, and is neutral on the “interaction fixes the correspondence” half, which is defeater (a)‘s ground. The A21 §5 caution the claim carries is consistent with the text: nonzero holonomy is “a genuine global obstruction” to a compatible state (p. 2), not a failure of the graph.
• CLAIM-481: an extension and a limit. The claim’s Laplacian-flow reading (kernel = exact compatibility, nonzero spectrum = relaxation rates) carries over to Lα, with one change: under nonzero holonomy the kernel is empty (p. 3), so the linear flow ż = −κLα z has no compatible state to settle on, and λ₁ > 0 is a floor of incompatibility rather than the first rate above a collective mode. The paper computes no relaxation rates. The limit is its §II C 6 (p. 7; SM pp. 9–10): for the nonlinear SK dynamics, the twisted spectrum does not set stability; the Jacobian does, and its threshold αc ≃ 1.0164 differs from the spectral marker π/3. That fits the claim’s “Anything about nonlinear coupling” exclusion and gives it a worked instance.
• QUESTION-281: the frustration route, checked, is narrower than the exchange made it. The paper supports frustration as a structural fact on cycles (“intrinsic topological frustration”, abstract, p. 2) and λ₁ as the residual left after optimizing over states (p. 5). It does not show that frustration stabilizes or organizes a form. In the motif, the nonuniform locked state is forced by unequal degrees, not by holonomy (p. 6); the remote-synchronization pattern is continued smoothly from complete synchrony at α = 0 (p. 6), so no transition creates it; and it ends, as a stable state, at αc ≃ 1.0164 when a transverse eigenvalue in the pentagonal sector crosses zero (SM p. 10), before the frustration peak. For the question that is a stated system with the transition that unmakes a form named, separately from the geometry that is said to shape it; the bifurcation’s type and what lies past it are not given (THEORY-1087).
• THEORY-276: a further instance, consistent and in its terms. The motif analysis uses the reflection-symmetry classes {1}, {2, 3}, {4, 7}, {5, 6} as the admissible clusters, citing Pecora et al. and Sorrentino et al. (p. 6), and its transverse perturbations “separate according to the graph symmetries” (p. 7; SM p. 9) into a triangular η-mode and one 2 × 2 block coupling the pairs {4, 7} and {5, 6}. Read with THEORY-276, the automorphism group is two commuting reflections, one moving {2, 3} and one moving {4, 7} and {5, 6} together; the latter two are intertwined, and the pentagonal block loses stability at αc while the triangular mode stays stable (SM p. 10). That is the pattern THEORY-276 predicts for isolated desynchronization, here for phase oscillators with a lag, though the paper does not use that language or follow the dynamics past αc. Two caveats: SK coupling depends on ϕⱼ − ϕᵢ, not on xⱼ alone as in THEORY-276’s eq. (1), though the equivariance argument of its 2026-10-19 note holds for any identical pairwise coupling; and this is a computation for one model, not the second experiment QUESTION-147 asks for. It also bears out THEORY-276’s last point: {4, 7} are in step with no edge between them (p. 6).
• QUESTION-130: a connection that is chosen, not found. The question asks whether a transport between contexts can be “fixed by independently established relations rather than chosen”. This paper’s associated connection is chosen, by the modeller, to reproduce a mismatch the dynamics defines (eq. 21), and only its gauge class is fixed (p. 4). The holonomy (ℓ − 2)α is real in the sense that it is phase-independent and read off the model (eq. 17); the connection carrying it is a representation. A worked example of the distinction the question draws, outside evaluation.
• No instruction for machine-learning practice.

Limitations

• The zero-mode criterion is stated, not proved, in this paper (p. 3); it is standard, and I supplied the missing step above.
• The SK correspondence is a representation by construction, for cactus graphs only, and gives “a static spectral encoding” (p. 4); nothing in the paper derives the SK dynamics from Lα.
• One motif. The only dynamical result is for a seven-node graph, on one branch. Fig. 1’s 14-node cactus is spectral only.
• “Higher-order” is the authors’ gloss. A cycle constraint on a graph with a U(1) connection is classical (magnetic Laplacian, ref. [14]); the paper’s addition is the remote-synchronization reading.
• “Uniform lag” in the gauge model depends on edge orientations (eq. 13), so its holonomies are those of the orientation chosen; the SK side has no such dependence.
• The comparison with Nicosia et al.’s transition value (p. 2) is not quantified here and was not checked.
• The figures’ data did not survive extraction; only captions and axis ticks were read. My numerical checks used my own script.
• v2 only. Differences from v1 and from the journal version are unknown.

Open questions

• Does the pentagonal-sector instability at αc lead to the “isolated desynchronization” of the pentagon pairs with {2, 3} still in step, as the symmetry structure suggests (THEORY-276)? The paper stops at the threshold.
• On graphs with overlapping cycles, where one edge would need two phases, is there any associated connection, or a sheaf of higher rank (p. 8)?
• Is the closeness of αc ≃ 1.0164 to π/3 ≃ 1.0472 structural, or a coincidence of this motif? A family of motifs with ℓ-cycles of other lengths would say.