Contribution
A textbook, not a research contribution: the results in the parts read are standard order and category theory. What it adds is an organisation of applied category theory around compositionality, with each chapter pairing a modelling situation with one categorical idea, and an introduction to monoidal categories written for readers outside mathematics. Its first chapter makes “generative effect”, from Elie Adam’s thesis, the motivating notion: what an observation of a whole shows that the combined observations of its parts do not.
Key insight
An observation of systems is a structure-preserving map, and the question “what category are you working in?” is the question of which structure it preserves. Observations that preserve order but not joins have generative effects. Being a left adjoint is exactly the property that rules them out, when the source has all joins. So “lossy observation” and “failure of compositionality” become a precise statement about adjoints.
Assumptions
- Preorders (reflexive, transitive) rather than partial orders; meets and joins defined up to equivalence.
- Theorem 1.115 needs the source of a join-preserving map to have all joins (dually, the source of a meet-preserving map all meets).
- In Chapter 3, schemas are finitely presented categories (graphs with path equations) and instances are Set-valued functors; Σ and Π are asserted to exist for every schema map, with details deferred to Spivak’s other work.
Key results
Chapter 1:
- Definition 1.93 and Exercise 1.94. For monotone f with the relevant joins, f(a) ∨ f(b) ≤ f(a ∨ b) always; f has a generative effect if the two sides differ for some a, b. Worked example: Φ on the five partitions of {•, ◦, ∗}, “is • connected to ∗?”, is monotone, and joining two partitions in neither of which • and ∗ are connected can connect them.
- Proposition 1.78. Monotone maps P → Bool correspond to upper sets of P. Exercise 1.66 is the Yoneda lemma for preorders: p ≤ p′ iff ↑p′ ⊆ ↑p.
- Proposition 1.107. f ⊣ g iff p ≤ g(f(p)) and f(g(q)) ≤ q for all p, q.
- Proposition 1.111. Right adjoints preserve meets; left adjoints preserve joins. Example 1.113: a right adjoint need not preserve joins.
- Theorem 1.115 (adjoint functor theorem for preorders). If Q has all meets, g : Q → P preserves meets iff it is a right adjoint, with left adjoint f(p) = ⋀{q | p ≤ g(q)}; dually for joins.
- Section 1.4.2. Any g : S → T induces g! ⊣ g* between partition preorders (push forward and take transitive closure; pull back).
- Example 1.117. For f : A → B, preimage f* has left adjoint the image f! and right adjoint f∗ (buckets all of whose apples are in the subset).
- Section 1.4.4. g∘f is a closure operator for any adjunction; every closure operator j arises from j ⊣ inclusion of its fixed points.
- Section 1.4.5. Reflexive–transitive closure is left adjoint to the inclusion of preorder relations into all relations.
Chapter 3 (sections read):
- Definition 3.68. Pulling back an instance I : D → Set along F : C → D is F ; I, giving Δ_F : D-Inst → C-Inst.
- Definition 3.70 and Example 3.71. Adjunctions as natural hom-set bijections; Galois connections are adjunctions between preorders.
- Section 3.4.3–3.4.4. Σ_F ⊣ Δ_F ⊣ Π_F; for ! : C → 1, Σ_! gives connected components and Π_! a selection (self-sent emails in the example). Σ-operations are built from colimits, Π-operations from limits (stated).
Concepts
- generative effect: failure of a monotone map to preserve joins; “we see something when we observe the combined system that we could not expect by merely combining our observations of the pieces”.
- Galois connection / adjunction: monotone f, g with f(p) ≤ q iff p ≤ g(q).
- closure operator: monotone j with p ≤ j(p) and j(j(p)) ≅ j(p).
- data migration functors Δ, Σ, Π: pullback along a schema map and its left and right adjoints.
- coherence conditions: the axioms that make structures “work well together”; the authors treat their choice as empirical.
Connections
Chapter 1 follows Adam’s thesis on generative effects (which goes on to abelian categories and cohomology) and points to Galois connections in program analysis. Chapter 3 descends from the categorical database work of the mid-1990s and Rosebrugh’s sketches, and from Spivak’s functorial data migration. The later chapters, seen only in summary, draw on Censi’s co-design, graphical linear algebra and Willems’ behavioural approach, Carboni–Walters and Fong’s decorated cospans, and the topos-theory literature.
Bearing on the record
- CLAIM-119 (communicative categories form a concept lattice). The book does not mention formal concept analysis. It supplies the order theory FCA rests on: the derivation operators of a formal context form an antitone Galois connection between subsets of objects and of attributes, and the formal concepts are the fixed points of the induced closure operators (Section 1.4.4’s construction). That identification is my connection, not the book’s; its Galois connections are covariant, and the antitone case is a short step. Bearing: it supports the claim’s formal vocabulary, not the empirical claim about communicative categories.
- Generative effects. Definition 1.93 states precisely when an observation of a composite cannot be computed from observations of the parts. Whether a translation or reading of a combined text has such an effect is a question the definition makes precise; the book does not raise it, and no claim is made here.
- LIT-827 (Backprop as Functor, same first two authors) is the research instance of the book’s “functorial semantics” theme.
- No THEORY filed: the results read are textbook theorems the record has no use stating separately. No instruction for machine-learning practice.
Limitations
- Skimmed: Chapters 2 and 4–7, including the book’s own reference treatment of monoidal categories, were not read beyond openings and summaries; no claim here rests on them.
- The book proves only some of what it states; several equivalences (for instance g! ⊣ g* on partitions, and the existence of Σ and Π in general) are left to exercises or other sources.
- As an introduction it chooses examples “to our personal taste”, by its own account, and does not survey applied category theory.
Open questions
- Whether a full reading of Chapters 2 and 4 (monoidal preorders, enrichment, profunctors, monoidal categories) would support a THEORY on resource theories, which the record does not yet hold.