Canonical Collections in Category Theory

• Canonical collections are concrete representatives of colimits, chosen with systematic naming to be usable in specification languages.
• They are crucial in applications like CASL and DOL, where a fixed signature with explicit symbol names is essential for parameter instantiation and network combinations.
• The approach resolves the gap between abstract colimit existence and practical implementation, ensuring that combined specifications maintain consistent and predictable symbol identities.

Canonical collections, in the sense developed in "Canonical Selection of Colimits" (Mossakowski et al., 2017), are systematic choices of concrete representatives for colimits in category-theoretic semantics. The central issue is that colimits ordinarily exist only up to isomorphism, whereas specification and ontology languages require an actual chosen object with actual symbol names. In this setting, a canonical collection is not merely a combination determined abstractly up to isomorphism, but a specifically selected combined object whose symbols are chosen in a predictable way and can therefore be used reliably by users, refinements, imports, and tools (Mossakowski et al., 2017).

1. The basic problem

Category theory treats a colimit of a diagram $D:\mathcal I\to\mathcal C$ as unique only up to unique isomorphism. For many mathematical purposes, that suffices. The paper isolates a different requirement: in specification and ontology languages, semantics is presented over concrete signatures with concrete symbol names, so an isomorphism class is not operationally adequate (Mossakowski et al., 2017).

This matters because later specifications may import, refine, or translate the resulting specification; users need to know which symbols are available after combination; and tools must compute an explicit result rather than an isomorphism class (Mossakowski et al., 2017). If two representatives differ only by renaming symbols, category theory regards them as equivalent, but a refinement written against one naming choice may be ill-formed against another isomorphic choice. The problem is therefore not the existence of colimits, but the absence of a principled selection of representatives.

The notion of canonical collections addresses exactly this gap. The aim is to define and study selections of colimits, that is, canonical choices of representatives, so that a colimit can serve as the canonical combination of a family of signatures or specifications in a way that remains usable at the syntactic and tool level (Mossakowski et al., 2017).

2. Role in specification languages

The motivation comes from two uses of colimits in institution-independent semantics: instantiations of parameterised specifications in CASL, and combinations of networks of specifications in DOL (Mossakowski et al., 2017).

In CASL, a generic specification $B[P]$ consists of a parameter $P$ included in a body $B$, written $P\hookrightarrow B$. Instantiation with an actual parameter $A$ and a morphism $\sigma:P\to A$ is given by a pushout: $\xymatrix{ P\, \ar@{^{(}->}[r] \ar[d]^{\sigma} & B \ar[d]\ A\, \ar@{^{(}->}[r] & B[A fit \sigma] }$ The resulting specification should preserve the structure of $B$ over $A$ in the same way that $B[P]$0 extends $B[P]$1. In particular, symbols introduced in the body should be preserved when possible (Mossakowski et al., 2017).

In DOL, a network is a graph of specifications and morphisms, and the operator combine computes the colimit of that graph. Depending on the shape of the graph, the colimit may be a pushout or a disjoint union. Whether symbols are identified or kept separate is therefore determined by the diagram shape itself (Mossakowski et al., 2017).

In both settings, colimits implement combination. The practical difficulty is that the combination must be returned as a definite signature or specification, not as a merely abstract object. Canonical collections are thus a semantic and implementation requirement rather than a purely formal embellishment.

3. Why ordinary colimits are insufficient

The paper emphasizes that ordinary colimit theory abstracts away from precisely the information that specification languages need. A colimit representative carries concrete names, and those names affect subsequent well-formedness conditions. This is why the problem cannot be dismissed as a harmless choice of isomorphic presentation (Mossakowski et al., 2017).

The issue is especially visible when specifications are chained. A result of one combination may later be reused as the source or target of imports, refinements, or translations. If a tool silently selects a different but isomorphic representative, the next specification step may fail syntactically even though the underlying categorical semantics is unchanged. This suggests that, in these applications, symbol identity is not an accidental presentation issue but part of the operational semantics.

A canonical collection therefore has to provide more than existence. It must provide a systematic naming policy, so that the combined object is predictable and user-friendly. The paper frames this as a very practical issue that is often glossed over in category-theoretic semantics, even though it is unavoidable in actual specification languages (Mossakowski et al., 2017).

4. Categorical setting

The paper works in inclusive categories with symbols (Mossakowski et al., 2017).

An inclusive category is a category $B[P]$2 together with a broad subcategory that is a partial order. The morphisms in that subcategory are called inclusions and are written

$B[P]$3

when there is an inclusion $B[P]$4. This formalizes the common situation in which some signature morphisms are literally extensions by adding symbols (Mossakowski et al., 2017).

The paper also introduces categories with symbols. The supplied material states that a category with symbols is an inclusive category $B[P]$5 equipped with additional symbol structure, and that this structure is used to track concrete symbol names in the chosen colimit representatives (Mossakowski et al., 2017). Even at this level of abstraction, the purpose is clear: the categorical framework must expose enough naming information for a colimit selection to be meaningful as a concrete specification-language construct.

This setup reflects the intended applications. In signature categories, inclusions model extension, and symbol-level information makes it possible to ask not only whether a colimit exists, but which symbols survive, which are identified, and under what names they appear in the selected representative.

5. Canonical selection of colimits

The core proposal is to replace the informal practice of speaking of “the colimit” with an explicit notion of selection of colimits (Mossakowski et al., 2017). A canonical collection is such a selection applied to the task of combining a family of objects.

What is selected is not an isomorphism class but a representative. The representative must be suitable for later use in a specification language: it must expose a concrete signature, concrete symbol names, and a stable interface for further imports and refinements. The paper describes this as the need for a systematically selected combined object whose elements or symbols are chosen in a predictable, user-friendly way (Mossakowski et al., 2017).

This gives canonical collections a distinctly semantic character. They are not quotient constructions in the usual abstract sense, nor merely amalgams up to isomorphism. They are selected colimit objects intended to function as the canonical result of “combining” or “collecting together” specifications. A plausible implication is that the quality of a colimit semantics in these languages depends not only on categorical existence theorems but also on the discipline with which representatives are chosen.

6. Significance and scope

The paper’s contribution is to make explicit a requirement that is easy to suppress in abstract category theory but unavoidable in formal specification practice. Once colimits are used as semantic operators in languages such as CASL and DOL, the choice of representative becomes part of the semantics as experienced by users and tools (Mossakowski et al., 2017).

The significance is therefore methodological as well as technical. Canonical collections connect institution-independent semantics with implementable language design. They ensure that a combination operation returns something that can actually be named, reused, and checked syntactically. This is especially important in settings where symbol preservation matters, such as parameter instantiation and network combination.

The paper does not argue that ordinary colimit theory is mathematically deficient. Rather, it identifies a mismatch between the abstraction level of standard uniqueness-up-to-isomorphism and the needs of concrete specification environments. Canonical collections resolve that mismatch by insisting that colimit semantics in such environments must include a canonical selection policy, not merely an existence result (Mossakowski et al., 2017).