Source
Robinson (2017), LIT-1658, §§1, 3–5 of arXiv 1603.01446v3 (pp. 1–2, 7–18, 22–29, 32); the arXiv abstract. Read in NOTE-1221.
What was actually shown
The argument for “canonical” is a definition. Section 3 opens: “we posit a set of six Axioms to specify what this integrated system must be. A sheaf is precisely any mathematical object satisfying all six Axioms. We motivate the Axioms by way of examples” (pp. 7–8). The axioms are the definition of a sheaf of pseudometric spaces on a topology, motivated one at a time by the search-and-rescue and coin-counting examples. The paper also matches Wald’s definition of data fusion and says “This article shows that this is precisely what is provided by sheaf theory” (p. 2).
The word, where it occurs. Sheaves “provide a common, canonical language for heterogeneous datasets” (p. 1); “The foundational and canonical nature of sheaves means that existing approaches … already illuminate some portion of sheaf theory” (p. 2); “Sheaf theory provides canonical computational tools” (p. 4). Its precise uses are standard mathematics: cohomology is “a canonical family of invariants” and the Godement resolution is “the canonical … resolution” (p. 32). No statement, proposition or theorem says sheaves are unique, initial, terminal or universal among sensor-integration models. The arXiv abstract calls the paper “expository”: its purpose is “to explain why sheaves are the canonical data structure”.
The paper’s own concession. “Although there are many co-universal information representations (of which sheaves are one), the stumbling block for integration approaches is the lack of theoretically-motivated computational approaches” (p. 29). The advantage claimed is the tooling, cohomology above all, not uniqueness.
The limits it states. Axiom 3 “is a true limitation” (Remark 7, p. 12). Axiom 6 “can always be satisfied if S(U₁ ∪ U₂) has not been constrained by the sensor models”, by defining the union’s observations as the fibre product, eq. (1) (p. 18). Relational databases “often do not satisfy Axiom 6” (Example 17, p. 22). The claim that Axiom 6 is “a necessary and sufficient constraint” (p. 17) is asserted, not proved.
The tools need less. The consistency radius (Definition 20), the distance bound (Proposition 23) and data fusion (Problem 19) are all stated for a presheaf, Axioms 1–5 (pp. 22–25; THEORY-1077). The sheaf condition and linearity (Axiom 7) enter in Section 5, for cohomology.
What this does not say
- That sheaves are a poor model of sensor systems. The worked examples fit the axioms, and the cohomology examples (30–32) show what the extra structure buys.
- That no precise sense of “canonical” exists. A universal property for some category of sensor models could perhaps be stated and proved; this paper does not do it.
- That later work has not. Only this paper was read.
- That gluing language in the record is unsupported. CLAIM-471 uses gluing as one formal setting among three; it does not lean on the word “canonical”, only on the paper’s being a working example.