NOTE-1217: Stewart, Golubitsky & Pivato — Symmetry groupoids and patterns of synchrony

Contribution

It answers, in the negative, “whether symmetry is the only mechanism that can create such states in a coupled cell system” (abstract, p. 609), and replaces the symmetry group by a groupoid built from the network’s input sets. Two theorems follow. A pattern of synchrony is forced by the architecture alone, for every admissible dynamics, exactly when its equivalence relation on cells is balanced (Theorem 6.5, p. 630). The dynamics restricted to it are those of a quotient network, a smaller coupled cell network built from the classes (§7, Theorems 8.3, 8.4 and 9.2, pp. 634–644). The paper is the foundation of what later work calls the groupoid formalism for coupled cell networks; its stated scope is “a preliminary step toward a formal understanding of pattern-formation in general, not necessarily symmetric, coupled cell networks” (p. 645).

Key insight

A cell’s equation depends only on its input set, so two cells are forced to evolve alike whenever their inputs look alike, whether or not any permutation of the whole network says so. “This permutation is not a symmetry of the whole network, but it preserves enough structure to create a flow-invariant subspace” (p. 613). Colour the cells by class: the pattern is robust “if and only if every pair of ⋈-equivalent cells is connected by a color preserving groupoid element” (p. 630), that is, if each cell of a colour receives the same number of inputs of each colour through each kind of arrow. Synchrony is then a statement about equal input, not about the synchronized cells being coupled to each other.

Assumptions

• Network (Definition 2.1, p. 616). A finite set of cells C; arrows E ⊆ C × C, so at most one arrow of a given direction between two cells; an equivalence ∼_C on cells (cell type) and ∼_E on arrows (coupling type), with equivalent arrows having equivalent heads and tails. Every cell is “active”, carrying an internal edge (c, c) (eq. 2.1). The authors say finiteness and activity can be relaxed (p. 616); they do not do it.
• Dynamics (Definition 4.1, p. 622). Each cell has a phase space P_c, a nonzero finite-dimensional real vector space, equal for cells of one type (p. 621). A vector field f is admissible if (a) f_c depends only on the cells of the input set I(c) (domain condition, eq. 4.3) and (b) f_β(c)(x) = f_c(β*(x)) for every input isomorphism β (equivariance condition, eq. 4.5). Nothing else is assumed: the coupling need not be additive, “the time-evolution of cell C can in principle be any (smooth) function of the states of C, A, and B” (p. 610).
• “Robust” means invariant for every admissible vector field (Definition 6.2, p. 628): f(Δ_⋈) ⊆ Δ_⋈ for all f ∈ F_G^P. It does not mean stable, and it does not mean invariant for a particular model family inside the admissible class.
• The relation ⋈ refines cell type (p. 628), so that the polydiagonal Δ_⋈ = {x : x_c = x_d whenever c ⋈ d} is defined.
• Existence, not stability. “in this paper we are studying existence, not stability, and we do not wish to rule out unstable synchrony because it still corresponds to an invariant subspace for the dynamics” (p. 626).

Key results

Robust synchrony without symmetry (§1, pp. 610–615)

• Figure 3, eq. (1.5), p. 612. Cells 1 and 2 (same type) are coupled to each other by equivalent arrows and both receive an equivalent arrow from cell 3; cell 3 receives only from cell 1. ẋ₁ = f(x₁, x₂, x₃), ẋ₂ = f(x₂, x₁, x₃), ẋ₃ = g(x₃, x₁). “This network has no symmetry, but the network structure forces the “polydiagonal” subspace Y = {x : x₁ = x₂} to be flow-invariant.” The permutation (1 2) maps the input sets of 1 and 2 onto each other but is not a symmetry, since cell 3’s equation changes under it; it agrees on Y, “and this is enough to make Y flow-invariant” (p. 613).
• Seven-cell chain with a back connection, Figures 5–8, pp. 613–615. A feed-forward chain with cell 1 fed back from cell 3. All input sets are isomorphic, the network has “many groupoid symmetries (42 in all)” (p. 614), and setting x₇ = x₄ = x₁, x₆ = x₃, x₅ = x₂ (eq. 1.8) reduces it to the three-cell directed ring (1.9), whose discrete rotating wave by Hopf bifurcation lifts to a travelling wave of the chain (Figure 8). The authors find it “curious” that the back-coupled chain, “less regular in form”, is the one where travelling waves are typical (p. 614).

The groupoid and the admissible vector fields (§§3–4, pp. 617–625)

• Input isomorphism (Definition 3.2, p. 617). A bijection β : I(c) → I(d) with β(c) = d and (i, c) ∼_E (β(i), d) for every i. The symmetry groupoid B_G is the (disjoint) union of all sets B(c, d) (Definition 3.5, p. 618); B(c, c) is the vertex group. Its connected components are the input-equivalence classes (Lemma 3.8, p. 620), and B(c, c) = S_k₁ × ⋯ × S_kᵣ, permutations within each class of equivalent input arrows (p. 620).
• Admissible = groupoid-equivariant (Definition 4.1). Example 4.4 (pp. 622–623) shows the admissible fields of Figure 9’s network are f₁ = A(x₁, x₃, x₄), f₂ = B(x₂, x₁), f₃ = B(x₃, x₁), f₄ = C(x₄, x₂, x₃) with C symmetric in x₂, x₃, though the network “is not symmetric under the 2-cycle (2 3)”.
• Lemma 4.5 and Proposition 4.6 (pp. 624–625). A B(q, q)-invariant map at one cell of an input class extends uniquely to an admissible field supported on that class, and F_G^P is the direct sum of these pieces over the input classes. These supply the special vector fields the proofs use.

The main theorem on polysynchrony (§6, pp. 627–632)

• Lemma 6.3, p. 629. A robustly polysynchronous relation refines input equivalence: only input-isomorphic cells can be robustly synchronous.
• Definition 6.4, p. 629 (balanced). ⋈ is balanced if for all c ⋈ d, c ≠ d, there is γ ∈ B(c, d) with i ⋈ γ(i) for all i ∈ I(c).
• Theorem 6.5, p. 630. “Let ⋈ be an equivalence relation on a coupled cell network. Then ⋈ is robustly polysynchronous if and only if ⋈ is balanced.” Sufficiency is immediate (f_c = f_d on Δ_⋈). Necessity is proved by building, for each class of input arrows K_s at c, the admissible field h_c(x) = M(Σ_{i∈K_s} x_i) with M a nonzero linear map (eq. 6.4), and evaluating it on a vector equal to v on one ⋈-class U and 0 elsewhere. Invariance forces the counting condition |U ∩ K_s| = |U ∩ L_s| (eq. 6.2) for every class U and arrow type s, which builds the colour-preserving γ (pp. 631–632).
• Example 5.1 and its modification (pp. 626–629). In the ten-cell network of Figure 11 the classes {0, 1}, {2, 3, 6, 8}, {4, 5, 7, 9} are balanced. Recolouring cell 6 breaks it: g₁ = A(u, w, w, v, w) differs from g₀ because of “multiplicity”, three w’s against two (p. 629). Cells 4, 5, 7, 9 are robustly synchronous without any arrow between them: each receives one input of type ○ and one from the class {2, 3, 6, 8} (Figure 12, p. 626). Cells 2, 3, 6, 8, which have no inputs but themselves, “can be considered synchronous in our formalism. Of course such a form of synchrony is dynamically unstable” (p. 626).

Quotient networks (§§7–8, pp. 632–641)

• Example 7.1, pp. 632–633. A homogeneous five-cell network (Figure 15) whose admissible fields are (7.1), with f symmetric in its last two arguments. On the balanced polydiagonal {(x, y, x, y, z)} they restrict to ẋ = f(x, y, z), ẏ = f(y, z, x), ż = f(z, x, y) (eq. 7.2), the general field of a three-cell bidirectional ring, which is the general D₃-equivariant field on (Rᵏ)³. “So it is possible for a quotient network to have symmetry even when the original network has none.” Rotating waves, double-frequency states and symmetric chaos are simulated in the five-cell system (Figures 16–18).
• Figure 19, p. 635. Three cells: 1 and 2 coupled both ways, 1 → 3. Cells 2 and 3, each with one input from cell 1, are balanced; the quotient is a two-cell network with Z₂ symmetry. “This is the same as the “master-slave” network of Pecora and Carroll [12]” (p. 633).
• Construction (A)–(F), pp. 634–636. Cells of the quotient are the ⋈-classes; arrows are projections of arrows not joining two distinct equivalent cells; two quotient arrows are equivalent if some input isomorphism maps the cells of one colour onto the cells of the other (eqs. 7.3, 7.4). Lemma 7.3 makes this independent of representatives.
• Definition 8.1, p. 638 (quotient map): surjective on cells, input arrows lift, input isomorphisms lift (eq. 8.1). Lemma 8.2(c): the relation a quotient map induces is balanced. Theorem 8.3, p. 639: the natural quotient by a balanced relation is a quotient map.
• Non-uniqueness and universality. Example 8.5 and Figure 21 (pp. 639–640): the seven-cell chain has two non-isomorphic three-cell ring quotients with the same relation. Theorem 8.4, p. 639: the natural quotient is universal; every other quotient with the same relation factors through it. Corollary 8.8, p. 641: every such quotient is an edge-refinement of the natural one, and conversely.

The second main theorem (§9, pp. 641–644)

• Theorem 9.2, p. 642. “For any f ∈ F_{G₁}^P, the induced vector field f̄ lies in F_{G₂}^P̄,” where f̄(y) = α⁻¹(f(α(y))) (eq. 9.3) is f restricted to Δ_φ and read in the quotient’s coordinates. So “interesting dynamics (rotating waves, symmetric chaos) in this subspace for the cell system G₂ corresponds to the same dynamics in the cell system G₁, in which ⋈-equivalent cells are synchronous” (p. 642).
• Not every quotient field lifts (Remark 7.5, pp. 637–638). In the four-cell network of Figure 20 the restricted fields have linear terms αu₁ + β(u₂ + 2u₃), whereas a general quotient field has αu₁ + βu₂ + γu₃: “vector fields that are admissible with respect to the quotient network G_⋈ do not always lift”. The authors call this “a groupoid analogue of hidden symmetry”. The induced map φ̂ “need not be surjective when φ is natural. It is never surjective when φ is not the natural quotient map” (p. 644).

Final comments (§10, p. 645)

• The formalism is said to carry over to Hamiltonian systems, coupled map lattices, cellular automata, Markov chains (constraints on the stationary density), SDEs and delay equations. None of these is worked out.

Claims

id	claim	strength	support
C1	A synchrony pattern is invariant under every admissible vector field of a coupled cell network iff its equivalence relation is balanced	strong	Theorem 6.5, p. 630, with proof (pp. 630–632)
C2	Only input-isomorphic cells can be robustly synchronous	strong	Lemma 6.3, p. 629, with proof
C3	Robust synchrony can occur in networks with no symmetry	strong	worked examples: Figure 3 / eq. (1.5), p. 612; Example 4.4, p. 623; Figures 15 and 19 (pp. 632–635), the text stating each lacks symmetry
C4	The restriction of an admissible field to a balanced polydiagonal is an admissible field of the quotient network	strong	Theorem 9.2, p. 642, with proof (pp. 642–644)
C5	Not every admissible field of the quotient arises as such a restriction	strong	counterexample, Remark 7.5, pp. 637–638
C6	Quotients with a given relation are unique up to edge-refinement of a universal natural quotient	strong	Theorem 8.4, p. 639; Corollary 8.8, p. 641
C7	A quotient can have symmetry the original network lacks, and so carry symmetric dynamics (rotating waves, multirhythms, symmetric chaos) into synchronous dynamics of the original	moderate	Example 7.1 shown exactly (eq. 7.2); the dynamics shown by simulation (Figures 16–18) and by citing the H/K theorem and symmetric Hopf bifurcation [7, 9]
C8	The groupoid formalism extends to discrete-time, stochastic, delay and Markov-chain settings	weak	asserted, §10, p. 645
C9	Synchrony of input-free identical cells is “dynamically unstable”	weak	asserted in a parenthesis, p. 626; true only for some internal dynamics (see Limitations)

Concepts

• Cell — “a system of ODEs” whose output is a canonical observable (p. 609).
• Input set I(c) — c and every cell with an arrow to c (Definition 3.1, p. 617).
• Input isomorphism — a base-point-preserving bijection of input sets that preserves arrow types (Definition 3.2).
• Symmetry groupoid B_G — all input isomorphisms, multiplied by composition where domain and range match (Definition 3.5, p. 618).
• Admissible vector field — one satisfying the domain and equivariance conditions (Definition 4.1, p. 622); “the most general vector fields on P that are consistent with the coupled cell network” (p. 622).
• Polydiagonal Δ_⋈ — states equal on each ⋈-class (p. 628); ⋈-polysynchronous — a trajectory lying in it (Definition 6.1).
• Robustly polysynchronous — Δ_⋈ invariant under every admissible field (Definition 6.2). Not a stability notion.
• Balanced — equivalent cells are joined by an input isomorphism that keeps every input in its class (Definition 6.4); equivalently, each cell of a colour receives the same number of inputs of each colour along each arrow type (eq. 6.2).
• Quotient network G_⋈ — cells the classes, arrows and their types induced as in (A)–(F), pp. 634–636; quotient map (Definition 8.1).
• Model-independent vs model-dependent synchrony — the former “valid for any vector field consistent with the given cell architecture” (p. 625); the paper treats only it.

Connections

• Golubitsky and Stewart’s symmetry methods ([7], [8], [9]): fixed-point subspaces of a symmetry group are flow-invariant (p. 611), and the H/K theorem and symmetric Hopf bifurcation give the spatio-temporal patterns (pp. 611, 614, 633). This paper generalizes the first from groups to groupoids. It does not state, as a theorem, that the orbit partition of a symmetry group is balanced; that follows from the definitions (a network symmetry restricted to I(c) is an input isomorphism to I(σ(c)) that maps each cell to one in the same orbit).
• Pecora and Carroll ([12]): Figure 19 is their master–slave network (p. 633). Boccaletti, Pecora and Pelaez ([2]) and Wang ([14]) are cited as surveys of synchrony.
• Groupoids: Brandt [1], Brown [3], Higgins [10], MacLane [11], Weinstein [15].
• Dias and Stewart ([5], “submitted”) are cited for the proofs about the quotient’s groupoid (Remark 7.4) and the full characterization of when quotient fields lift (Remark 7.5). Golubitsky, Nicol and Stewart ([6], “submitted”) for applications of Theorem 9.2. Neither is held by the record.

Bearing on the record

• THEORY-276 (Pecora et al., symmetry orbits as clusters): generalized, and its “symmetry is sufficient, not necessary” given its theorem. The orbit partition of a network symmetry group is balanced, so THEORY-276’s invariance result is a special case of Theorem 6.5 (p. 630), in a much wider class: arbitrary nonlinear, non-additive coupling (p. 610), several cell and arrow types, directed arrows. The clusters Pecora et al. noted “where input and output couplings match in a cluster, but there is no symmetry” are the balanced relations that are not orbit partitions, of which Figure 3 (p. 612) and Figure 19 (p. 635) are the smallest. And the paper’s necessity half says that, over all admissible dynamics, nothing else is robust. Two limits. Necessity is relative to the whole admissible class: a narrower model family, such as Pecora’s eq. (1) with one fixed H, may have extra invariant subspaces. And the paper says nothing about stability, which is THEORY-276’s other half (the transverse blocks) and QUESTION-147’s.
• CLAIM-471: the attribution holds in substance, but half of it is the paper’s scope statement, not its result. The claim says this paper separates “an invariant synchrony pattern, which the architecture allows, from an attracting one, which the dynamics selects”. The paper draws the first line sharply: it separates model-independent from model-dependent synchrony “because this separates the effect of the general architecture of the system from that of the specific model equations employed” (p. 625), and it studies “existence, not stability” (p. 626). It does not treat attracting synchrony at all, and “the dynamics selects” is nowhere in it. Its one example of the gap is its remark that the synchrony of input-free cells is invariant though “dynamically unstable” (p. 626). A citation for invariance without attractivity is sound; one for how the dynamics selects needs another source.
• QUESTION-063 (common drive or coupling): the theorem puts the two on one footing. A drive is just another cell, and the balance condition counts inputs, not who sends them. In Figure 19 (p. 635) cells 2 and 3 are robustly synchronous though 3 is only driven by cell 1 and 2 also drives cell 1 back; in Example 5.1 (p. 626) cells 4, 5, 7, 9 are robustly synchronous with no arrow among them, each receiving the same coloured inputs. So an architecture-forced synchrony pattern says “equal input”, and by itself says neither “coupled to each other” nor “driven from outside”. This bears on THEORY-276’s use on QUESTION-063: a pattern of clusters is a coupling signature only relative to a known graph that includes every source of input, the shared drive among them. The paper does not discuss drive or inference, so this is the record’s reading.
• CLAIM-481: it supports only the one line that cites A22 §7. “A synchrony subspace can be invariant without being attracting” is the paper’s stated scope (p. 626). Nothing in it bears on time scales, dissipation, Laplacian spectra or strong coupling destabilizing synchrony; the claim’s other content gets no support or correction from it.
• THEORY-755 and CLAIM-157 (exact aggregation without a block). The balance condition is a counting condition on inputs, the coupled-cell analogue of an equitable partition. For a linear network ẋ = Ax, the polydiagonal is invariant when each row of each block of A has the same sum (inputs counted), while THEORY-755’s exact aggregation of block sums needs equal column sums (outputs counted). That transposition is the record’s observation, not the paper’s, which never mentions equitable partitions. It adds a nonlinear case to CLAIM-157’s “groupings that close without being blocks”. The quotient network is an exact reduced description of the synchronous states, and synchronized cells need not interact at all (Example 5.1).
• No instruction for machine-learning practice.

Limitations

• Existence only. No stability, genericity of attraction, or basin result anywhere (p. 626). Which balanced patterns are seen is left to the model.
• The instability remark is loose. “Of course such a form of synchrony is dynamically unstable” (p. 626), said of identical cells with no inputs, holds for some internal dynamics (chaotic or neutral) and not for others: if B has a globally attracting equilibrium, the cells converge and their synchrony is attracting. The point the authors need, that invariance does not imply attraction, stands.
• “Robust” is relative to the admissible class. The necessity half of Theorem 6.5 uses linear admissible fields of the form M(Σ x_i) (eq. 6.4); for a model family that excludes such fields, or fixes coupling functions, more subspaces can be invariant. The theorem does not say what is invariant for one model.
• The 2003 network definition allows at most one arrow per ordered pair of cells (E ⊆ C × C, Definition 2.1), so weighted or multiple arrows are outside it. It is standardly reported that later papers by Golubitsky, Stewart and Török allow multiple arrows; that is not in this text and was not checked.
• The quotient field class is larger than what restriction gives (Remark 7.5), so a quotient’s generic behaviour need not be realizable in the original network; the full lifting criterion is deferred to Dias and Stewart [5], unread.
• Several claims are by simulation or citation: the rotating waves, multirhythms and symmetric chaos of Figures 8 and 16–18, and the “42 in all” groupoid symmetries of Figure 6 (p. 614), are not derived in the text.
• Extraction and figures. The extracted text loses arrow types; the descriptions of Figures 3, 11, 12 and 19 above were checked on the rendered pages. Figures 9, 14, 15, 20 and 21 were read from the text only.
• The published SIAM text was not compared with the accepted version.

Open questions

• Which balanced patterns are attracting, and how does that depend on the cell dynamics? The paper leaves it entirely to the model (p. 626).
• When does every quotient-admissible field lift (Remark 7.5)? Deferred to Dias and Stewart [5].
• Do the §10 extensions (Markov chains, cellular automata, SDEs) give theorems of the same form? Asserted, not shown (p. 645). For Markov chains, the record’s THEORY-822 (strong lumpability) is a candidate answer, joined to this paper only by the record’s reading.