Source
Hansen and Ghrist (2019), LIT-1659, §§2.2.2, 3.1–3.2, 8.1 and 8.6 of arXiv 1808.01513v2 (pp. 7, 9–11, 34–35, 39); read in NOTE-1233.
What was actually shown
Sections are the kernel. A cellular sheaf assigns a vector space to each cell and a linear restriction map to each incidence; a global section is an assignment that agrees under every map (Def. 2.5, p. 6). The coboundary δ^k collects the disagreements, and “H^0(X; F) is naturally isomorphic to Γ(X; F), the space of global sections” (p. 7). Theorem 3.1 (p. 10) is discrete Hodge theory for any complex of finite-dimensional Hilbert spaces: ker Δ^k ≅ H^k, with C^k = ℋ^k ⊕ im δ^{k−1} ⊕ im (δ^k). In degree 0 on a graph, L = δδ, and Lx = 0 exactly when F_{u⊴e}x_u = F_{v⊴e}x_v on every edge: agreement in the edge’s comparison space, not equality of the vertex values.
The flow. Proposition 8.1 (p. 34): “The dynamical system ẋ = −Δ^k_F x has as its space of equilibria ℋ^k(X; F), the space of harmonic k-cochains of F. The trajectory of this dynamical system initialized at x₀ converges exponentially quickly to the orthogonal projection of x₀ onto ℋ^k(X; F).” The proof writes x(t) = Σᵢ e^{−tλᵢ}⟨vᵢ, x₀⟩vᵢ over an orthonormal eigenbasis, so each component decays at its eigenvalue and the slowest decay is at the smallest nonzero one. For k = 0, “a distributed system can reach consensus on the nearest global section to an initial condition” (p. 35). For the discrete scheme x[t + 1] = (I − αL)x[t] the best step is α = 2/(λ_max + λ_min) and the contraction r = (λ_max − λ_min)/(λ_max + λ_min) (p. 39).
The weights matter, and the Laplacian forgets. The stalks’ inner products are data: normalizing is reweighting the sheaf (§3.4, p. 15), and the weights on an approximating sheaf “will be crucial to the spectral behavior of their Laplacians” (p. 39). Two sheaves with different stalk dimensions and maps have the same Laplacian (Fig. 3.1, p. 11): “one cannot hear the shape of a sheaf”.
What could have come out otherwise: the identifications and the flow are checkable line by line, and the Fig. 3.1 example by multiplying out two 2 × 2 products.
What this does not say
- Anything about time scales. The paper never uses “relaxation rate” or “time-scale separation”. That the nonzero spectrum is a hierarchy of rates, that a gap separates time scales, the two-cluster window and the n² chain are readings and computations of the exchange of 2026-10-27 (CLAIM-481), consistent with the proof’s formula but not in the paper.
- Where the restriction maps come from. The paper takes them as given and, in its applications, builds them from measurements (§§8.2, 8.7) or designs them (§8.5). It does not say interactions generate them (CLAIM-471).
- That the sections are nontrivial. The zero cochain is always a section, and a sheaf whose maps are zero has every cochain as a section (p. 11). Questions of interest are about nontrivial sections (§9.2, p. 42), as THEORY-299 found for knowledge sheaves.
- That the spectrum identifies the sheaf, or the correspondences it encodes (Fig. 3.1).
- Anything about nonlinear coupling, time-varying maps or oscillators. The flow is linear with a fixed Laplacian. “Synchronization” in this paper (§8.7) is the estimation problem of finding an approximate global section of measured pairwise relations, not the dynamical synchronization of coupled oscillators.
- Anything about infinite complexes or stalks in general. §3.3 gives conditions for bounded and compact coboundaries; where δ is compact on an infinite-dimensional space the eigenvalues accumulate at zero, so “there will be no smallest nontrivial eigenvalue” (p. 14).