NOTE-1222: Hansen & Ghrist — Opinion dynamics on discourse sheaves

Contribution

It recasts linear network opinion dynamics on a cellular sheaf, the “discourse sheaf”. A discourse sheaf gives each agent its own opinion space and each edge its own space of topics under discussion, and it attaches an expression map from each agent to each edge. The Laplacian diffusion, harmonic extension and control results of the classical graph models (Taylor, DeGroot, Friedkin–Johnsen) carry over. Global sections, sheaf cohomology and relative cohomology replace consensus, connectivity and “no consensus vanishes on the input set”. The new elements are these: the limit is agreement in expression, not identity of opinion (§§3–4); the expression maps themselves can be evolved, alone or jointly with the opinions (§§8–9); and bounded-confidence and antagonistic dynamics are written as nonlinear Laplacians (§§10–12). The paper is mathematics with toy examples. It contains no data and no social-scientific test.

Key insight

The paper puts consensus on the edges and leaves the vertices alone. Agents u and v “have expressed consensus when F_{u⊴e}(x_u) = F_{v⊴e}(x_v) … Note that this does not imply that u and v have the same opinions: it means that their expressions of personally held opinions have the appearance of agreement” (§3, p. 7). Diffusion drives the network to the kernel of the coboundary, so everything the theory can say about the limit is about expressions. Opinions held privately can be distinct, opposite or “incomparable” (p. 3). The same split makes the expression maps a second place where disagreement can be removed. If the maps rather than the opinions are evolved, discord falls because what agents say changes, which the paper calls “learning to lie” (§8).

Assumptions

• Linear algebra on a graph. Stalks are real, finite-dimensional inner product spaces with a fixed orthonormal basis, and restriction maps are linear (§2, p. 3). Edges are undirected and pairwise. Multi-agent interactions and probabilistic elements are named as future work (§13, p. 22).
• Topics are basis vectors, and opinions are intensities. “points on each axis correspond to negative, neutral, or positive opinions on the topic” (§3, pp. 6–7). An edge’s discourse topics need not be any agent’s basis topics (p. 7, the pub example on p. 8).
• Continuous time, deterministic. There is no noise anywhere. The paper says discrete-time versions follow by Euler discretization (§1.1, p. 2).
• Expression is fixed in §§4–7. The maps are given, and only the opinions move. In §8 the opinions are fixed and only the maps move. In §9 both move.
• What an agent can see (§8, p. 13). Each agent “is able to modify all its restriction maps F_{v⊴e} (for edges e incident to v) and is able to observe their neighbors’ translated opinions F_{u⊴e}x_u”. Its only objective is that “apparent consensus is reached”.
• The model is not offered as realistic. Evolving opinions “is perhaps not how real people engage in discourse” (p. 8). §8 sets aside “the sociological questions of whether a typical person in the face of opposition actually changes opinions or simply “learns to communicate better"" (p. 13).

Key results

The structure (§§2–3, pp. 3–8)

• Coboundary (δx)e = F{v⊴e}x_v − F_{u⊴e}x_u (eq. 2.3, p. 4); global sections H⁰(G; F) = ker δ (eq. 2.4, p. 5); sheaf Laplacian L_F = δᵀδ (eq. 2.5, p. 5); Hodge theorem H⁰(G; F) = ker L_F (Theorem 2.2, p. 6). Relative cohomology H⁰(G, A; F) is “global sections of F on G that vanish on A” (p. 5).
• The zero sheaf: replacing the constant sheaf’s identity maps with zero maps “gives a very different sheaf in which all communication is devoid of information. Such a sheaf is a jumble” (§2.1, p. 4).
• What a discourse sheaf permits (§3, p. 8): private basis topics, so agents “need not have any basis topics in common”; edge topics unrelated to anyone’s basis opinions; “the formation of policies from principles”; exaggeration and modulation by positive scalars; and lying by negative ones, “selectively. What C says to B need not match what C says to A (even if they are discussing the same topic)”.

Opinion diffusion (§4, pp. 8–9)

• Theorem 4.1 (p. 9). Solutions of dx/dt = −αL_F x (eq. 4.1) converge to the orthogonal projection of x(0) onto H⁰(G; F), exponentially, at a rate set by the Laplacian’s spectrum.
• “If this sheaf has no nontrivial global sections, the only stable opinions will be everywhere zero: an uninteresting solution” (p. 9).
• Example 4.2, “In polite company” (p. 9, Fig. 4 on p. 10). Four agents on a cycle, all stalks one-dimensional. Two agents hold a positive opinion of a politician and two a negative one. The right-hand pair use the map −1 toward the left-hand pair and +1 toward each other: “The two agents on the right tell a polite lie to their neighbors on the left but are frank with each other. A global section of this sheaf maintains the public agreement.” Any initial distribution converges to one of these polarized distributions. “although each agent knows the veracity of their opinion expressions, they do not know the veracity of their neighbors’ expressed opinions.”

Stubbornness, control and reluctance (§§5–7, pp. 9–13)

• Theorem 5.1 (p. 10). With the opinions on U held fixed, the rest converge exponentially to the harmonic extension of u = x₀|_U nearest to x₀. The limit minimizes ‖δx‖² subject to x|_U = u, and “harmonic extensions are not in general global sections” (p. 11). Unique when H⁰(G, U; F) = 0 (p. 10).
• Theorem 6.1 (p. 11). If H⁰(G, U; F) = 0 the control system (6.1), with input on U, is stabilizable. Theorem 6.2 (p. 12), “presented without proof”: if H⁰(G, Y; F) = 0 it is detectable from Y. Example 6.3: on the constant sheaf over a connected graph, influence over one agent suffices to steer global consensus (p. 12).
• §7 (pp. 12–13). Reluctance γ_v is modelled by giving each agent a stubborn “parent” vertex joined by maps √γ_v I, so that dx_v/dt = −α(L_F x)_v + αγ_v((x₀)_v − x_v) (eq. 7.1) and the limit is a harmonic extension on the augmented graph.

Learning to lie (§8, pp. 13–14)

• The dynamics. dF_{v⊴e}/dt = −β(F_{v⊴e}x_v − F_{u⊴e}x_u)x_vᵀ (eq. 8.1), or dδ/dt = −βP_δ(δxxᵀ) with P_δ the projection onto the coboundary sparsity pattern (eq. 8.3). The stated rule: agents “should alter their restriction maps to remove the part of the image that contributes to disagreement with neighbors” (p. 13).
• Theorem 8.1 (p. 13). With x fixed, F(t) converges to “the nearest sheaf such that x is a global section”, in the Frobenius distance ‖δ − δ′‖²_F (eq. 8.4). The proof is edge by edge: each edge’s flow is a projection onto {δ_e : δ_e x_e = 0}. The paper notes “a sort of duality” with the heat equation (p. 14).
• Example 8.2 (p. 14, Fig. 6). The constant sheaf on one edge, with opinions −4 and +1. One map changes sign: the agent “learns to lie about their opinion … changes their expressed opinion’s sign, but also downplays its magnitude.” The figure labels the final maps −0.2 and 0.8. My computation of Theorem 8.1’s limit from these data gives −3/17 ≈ −0.18 and 12/17 ≈ 0.71, so both agents express about 0.71. The ratio agrees with the figure; the figure’s magnitudes do not (see Limitations).

Joint opinion–expression diffusion (§9, pp. 14–17)

• Eq. (9.1). dx/dt = −αδᵀδx and dδ_e/dt = −βδ_e x_e x_eᵀ together.
• Lemma 9.1 (p. 15). Ψ(x, δ) = ½xᵀδᵀδx = ½‖δx‖² is nonincreasing, and dΨ/dt vanishes iff δx = 0. Eq. (9.1) “is actually a gradient descent equation on Ψ” for the inner product (1/α)xᵀx′ + (1/β)Σ tr(δ_eᵀδ′_e), so trajectories converge to pairs of a sheaf and a global section of it.
• “Because the stationary points of (9.1) are not isolated, there is no global asymptotic stability” (p. 15). The equilibria form a singular algebraic variety. Theorem 9.2 (p. 15): its smooth points are Lyapunov stable (center-manifold argument).
• The trivial solution, named and guarded against (p. 16). The highest-dimensional parts of the variety have ker δ = 0, so “This would seem to indicate that we should expect most trajectories to converge to points with x = 0. However, this is not the case.” Theorem 9.3: if a diagonal block of αδ₀ᵀδ₀ − βx₀x₀ᵀ is not semidefinite, then x_∞ ≠ 0. (The statement prints x₀ᵀx₀; the proof uses xxᵀ.) The proof shows that the diagonal blocks of M are constant in time, so an indefinite block stays indefinite and xxᵀ cannot go to zero. Hence “given any initial δ₀ and x₀, there exists some scaling factor κ such that the initial condition (δ₀, κx₀) converges to a sheaf with a nontrivial global section.”
• Theorem 9.4 (p. 16). ‖δ‖²_F, ‖x‖², ‖δx‖² and ‖δx‖²/‖x‖² are nonincreasing. The last, the Rayleigh quotient, means that “x(t) becomes a global section of F(t) at least as quickly as it approaches zero” (p. 17; the proof uses Jensen’s inequality).
• Example 9.5 (p. 17, Fig. 7). The same edge with α = β = 1. One agent again ends up lying, “because the discrepancy between opinions is so great, the sheaf is able to adapt more easily than the opinions themselves.”

Nonlinear Laplacians (§§10–12, pp. 17–21)

• L^Φ_F = δᵀ ∘ Φ ∘ δ with Φ applied edge by edge (§10). With Φ = ∇U from edge potentials, the heat equation is gradient descent on Ψ(x) = Σ U_e(δ_e x). Proposition 10.1 (p. 18): if each U_e is radially unbounded with a unique local minimum at 0, the limit is the same projection onto H⁰ as in the linear case.
• Bounded confidence (§11). U_e(y) = ψ_e(‖y‖²) with ψ′e > 0 below a threshold D_e and 0 above it. Theorem 11.1 (p. 19): x is an equilibrium iff every edge whose expressions disagree disagrees by at least D_e. Equilibria are the global sections of the sheaf F_x on the subgraph of edges still in confidence. Theorem 11.2 (p. 20): near an interior equilibrium, trajectories converge to the nearest global section of F{x₀}.
• Antagonism (§12). U_e = −‖y‖² on negative edges gives the signed Laplacian L^S_F = δᵀSδ, which is linear and possibly indefinite. Proposition 12.1 (p. 20): if the negative edges form a cutset and some local section on one side does not extend by zero, L^S_F is indefinite. For the constant sheaf this always holds, “and hence unstable opinion dynamics” (p. 21). Flipping a restriction map’s sign instead models an “agreement to disagree”, the approach of Altafini [2, 3] (pp. 20–21).

What the authors leave open (§13, pp. 21–22)

“Deeper analysis of the nonlinear dynamical systems (9.1) and (11.4) is needed”, along with probabilistic elements and simplicial complexes for multi-agent interaction. “sheaf dynamics are applicable to more than simply opinion dynamics”, for example neuroscience and game theory. Much sheaf-theoretic language was “excised” for accessibility.

Claims

id	claim	strength	support
C1	Sheaf-Laplacian opinion diffusion converges to the orthogonal projection of the initial opinions onto H⁰	strong	Theorem 4.1, proof by spectral decomposition, p. 9
C2	Its limits are agreement in expression, not in held opinion; stable polarized distributions exist	strong	definitional (p. 7) plus Example 4.2, p. 9
C3	With some agents stubborn, the rest converge to the nearest harmonic extension, unique when H⁰(G, U; F) = 0	strong	Theorem 5.1, pp. 10–11 (one slip in the proof’s condition, see Limitations)
C4	H⁰(G, U; F) = 0 makes the system stabilizable from U, and H⁰(G, Y; F) = 0 makes it detectable from Y	strong / weak	Theorem 6.1 proved (p. 11); Theorem 6.2 stated without proof (p. 12)
C5	Expression-map diffusion with fixed opinions converges to the Frobenius-nearest sheaf of which the opinions are a section	strong	Theorem 8.1, edge-wise projection, pp. 13–14
C6	Joint diffusion is gradient descent on ½‖δx‖² and converges to a sheaf with a section; smooth equilibria are Lyapunov stable	strong	Lemma 9.1, Theorem 9.2, pp. 15–16
C7	Joint diffusion does not collapse the opinions to zero when a diagonal block of αδ₀ᵀδ₀ − βx₀x₀ᵀ is indefinite	strong	Theorem 9.3, conservation of the diagonal blocks, p. 16; checked numerically here, with the α, β weights
C8	‖δ‖_F, ‖x‖, ‖δx‖ and the Rayleigh quotient are nonincreasing under joint diffusion	strong	Theorem 9.4, pp. 16–17
C9	Bounded-confidence equilibria are global sections of the in-confidence subgraph’s sheaf, and locally attracting	strong	Theorems 11.1, 11.2, pp. 19–20
C10	A cutset of negative edges makes the signed Laplacian indefinite and the dynamics unstable	strong	Proposition 12.1, p. 20
C11	Discourse sheaves model selective expression, lying, policies from principles, and learning to communicate	moderate	constructions and toy examples (§3, Examples 4.2, 8.2, 9.5); no data, no test against observed discourse
C12	Sheaf methods add expressiveness with “no increase in difficulty of computation or analysis”	weak	asserted in §13, p. 21; no complexity comparison given

Concepts

• Discourse sheaf (§3, p. 6): a cellular sheaf on a social graph. Its vertex stalks are opinion spaces, its edge stalks are discourse spaces of topics under discussion, and its restriction maps are expressions of opinion (Fig. 3, p. 7).
• Expression (restriction) map F_{v⊴e}: how agent v states its private opinions on edge e’s topics. It can be selective, combining, scaling (“exaggerating or modulating”) or sign-flipping (“lie”), and it can differ from one edge to another (p. 8).
• Global section / harmonic opinion distribution: x ∈ ker δ = ker L_F, in which every pair’s expressions agree (pp. 5, 8).
• Discord: the value of L_F x at a vertex, “the difference between x_v … and the opinions which would bring v in harmony with all its neighbors” (p. 8).
• Harmonic extension: the minimum-disagreement completion of opinions fixed on U (§5).
• Relative cohomology H⁰(G, A; F): sections vanishing on A, read as “a measure of independence of the agents in A from the rest of the social network” (p. 8).
• Learning to lie (§8): diffusion of the expression maps toward a sheaf of which the current opinions are a section.
• Nonlinear sheaf Laplacian (§10): δᵀ ∘ Φ ∘ δ with an edgewise Φ; the bounded-confidence and signed Laplacians are instances.

Connections

• Hansen and Ghrist, “Toward a spectral theory of cellular sheaves” (ref. [23], LIT-1659), the source of the sheaf Laplacian, harmonic extension and the convergence-rate claims this paper uses.
• Taylor 1968 (ref. [36]): §§5–7 are “extensions of ideas originally introduced by Taylor” (p. 10).
• Hegselmann and Krause (ref. [25]) for bounded confidence; Altafini (refs. [2–4]) for antagonistic interactions and structural balance, which the paper says studied structures like Example 4.2 “implicitly” (p. 9).
• Curry’s thesis (ref. [13]) for cellular sheaves.
• THEORY-299 (Gebhart, Hansen and Schrater; knowledge sheaves) is the same machinery used for embeddings, with the same zero section; see Bearing on the record.

Bearing on the record

• CLAIM-471: the paper is a worked formal case of the claim’s “agreement on what is shared and not identity”, and of a correspondence changed by interaction. Each edge carries its own comparison space, and compatibility is F_{u⊴e}x_u = F_{v⊴e}x_v, with private opinions “distinct (or even incomparable)” (pp. 3, 7). That is the claim’s form exactly. Two refinements follow from the text. First, in §§3–7 the correspondences are fixed background and only the states move. In §8 they move too, and they are driven by the interaction: each agent observes its neighbours’ expressions and revises its own map (p. 13). So the paper has both of the claim’s regimes, and the second is a formal instance against its defeater (a). It is a constructed model, though, not a finding that real correspondences are made this way. Second, the paper shows that compatibility on what is shared can be reached by changing what is shared rather than what is held. A global section then says nothing about the private states. Its own example calls this lying (Example 4.2, p. 9; §8). Of the claim’s three operations, opinion diffusion is enforcement (§4), and §8’s learning is enforcement acting on the correspondence. The paper has no separate evidence operation.
• CLAIM-342: the paper is an instance of the self-confirming loop, and it guards against degeneracy, not against self-confirmation. The expression-learning rule uses only the edge disagreement δx (eqs. 8.1, 9.1). The evidence each agent revises against is its neighbours’ expressions, which are produced by the maps being revised, and nothing in the model scores an expression against the opinion held or against anything outside the network. The paper names the outcome as a loss of fidelity (“learning to lie”) and never as an error. On the risk THEORY-299 recorded, the paper has three guards and one gap. (i) The paper says that a sheaf without nonzero sections sends opinions to zero, “an uninteresting solution” (p. 9). (ii) Theorem 8.1 moves the maps only to the nearest sheaf in Frobenius distance. That is a rank-one change along x_v at each vertex, so the maps are not sent to zero unless zero happens to be nearest. (iii) For joint diffusion it names the risk itself (“we should expect most trajectories to converge to points with x = 0”) and rules out x → 0 when a diagonal block of αδ₀ᵀδ₀ − βx₀x₀ᵀ is indefinite (Theorem 9.3, p. 16). (iv) The gap: it does not address the maps collapsing, and ‖δ‖F is nonincreasing (Theorem 9.4). A consequence of the same conserved quantity, which I derived and the paper does not state, is that a vertex’s maps cannot all go to zero if its block α Σ_e F{v⊴e}ᵀF_{v⊴e} − βx_v x_vᵀ starts with a positive eigenvalue. So the trivial solution is avoided by conservation and minimal change, which are internal to the loop, not by evidence from outside it. The result is coherence, and A22 §10’s distinction between “becoming more coherent” and “becoming more knowledgeable” applies: this paper’s learning is the first by construction.
• QUESTION-285, third answer criterion: a concrete G for an apparatus that revises its correspondences, with what it cannot reach stated. In §8, G is the gradient flow (8.3). It cannot change the graph, the stalks, the sparsity pattern or (in §8) the held opinions, and it is bound to the nearest admissible sheaf. These are structural limits on G of the kind the question asks for. They secure that the revision stays small and keeps the data nonzero. They do not secure answerability to anything the system did not produce, because the model has no outside. The paper is a clean negative case for the question’s last clause: a revising apparatus with well-defined limits on G that converges to agreement, and that the authors themselves describe as having learned to misreport.
• The exchange’s description (A22 §11, “Richer observation and communication maps, including selective expression”) is right about communication and selective expression. The abstract says “selective opinion modulation and lying”. Example 4.2 says the sheaf “permits selective expression of opinion” (p. 9), and p. 8 says deception “can be done selectively”. “Observation maps” overstates it. The sheaf’s maps are expression maps, and what an agent observes is its neighbours’ expressions (p. 13). The only observation operator is the control-theory output C of §6, which is the identity on the observed vertices.
• CLAIM-481. The convergence rate is “related to the spectral properties of the sheaf Laplacian” (p. 9), which is consistent with the claim’s Laplacian-spectrum rates. The paper only cites this to [23] and does no more with it.
• No instruction for machine-learning practice.

Limitations

• No empirical content. The social reading (lying, polite company, learning to communicate) is interpretation of the mathematics. The authors themselves disclaim realism (pp. 8, 13).
• A slip in Theorem 5.1’s proof (p. 10). It says L_F[Y, Y] is positive definite “if H⁰(G, Y; F) = 0”. By the paper’s own definition, a section vanishing on A, the condition is H⁰(G, U; F) = 0, which is what the text before the theorem says for uniqueness.
• Theorem 9.3’s weights. The statement uses αδ₀ᵀδ₀ − βx₀ᵀx₀ (read x₀x₀ᵀ). The proof defines M = δᵀδ − xxᵀ but differentiates αδᵀδ − βxxᵀ, and later writes tr(M) = ‖δ‖²_F − ‖x‖². A numerical check here, on a three-vertex path with two-dimensional vertex stalks, found the diagonal blocks of αδᵀδ − βxxᵀ conserved for α = 2, β = 0.5, and those of δᵀδ − xxᵀ not conserved. So the theorem holds with the weights, and the proof’s M and trace are correct only when α = β.
• Figure 6’s labels. The figure’s final maps are −0.2 and 0.8. Theorem 8.1’s limit from the stated data (−4, +1, the constant sheaf) is −3/17 and 12/17. The ratio matches but the magnitudes do not, and the labels lie off the straight path the flow follows. They may be rounded or schematic. The paper does not say.
• Theorem 6.1 proves stabilizability. The following text says “we can ensure that the dynamics converge to any global section” (p. 12). That is argued from uniqueness of harmonic extension, not from a controllability theorem. Theorem 6.2 has no proof.
• The nonlinear and joint systems are only partly analysed. There is “no global asymptotic stability” for (9.1), and stability is shown only at smooth equilibria. The authors say deeper analysis is needed (p. 21).
• Typos. “which is in U by condition (1)” in the proof of Theorem 11.2 (p. 20) should be W. The proof of Theorem 5.1 labels the kernel part y⊥ and the image part y∥.
• The published SIAM text was not compared with the preprint.

Open questions

• What would make expression learning answerable to something it does not control? For example, a penalty on ‖F_{v⊴e} − F⁰_{v⊴e}‖ held by the agent (fidelity to its own truthful map), or an edge observable scored against an outside signal. Neither is in the paper.
• Under joint diffusion, for which initial conditions do some of a vertex’s expression maps go to zero, so that it stops communicating on some edges? The paper bounds x away from zero, not δ.
• Which global sections does joint diffusion select, given that its equilibria form a singular variety (p. 15)?
• Does anything like the “polite lie” sheaf of Example 4.2 describe observed discourse, and how would its restriction maps be measured?