DB Category: Denotational Semantics for View-based Database Mappings

We present a categorical denotational semantics for a database mapping, based on views, in the most general framework of a database integration/exchange. Developed database category DB, for databases (objects) and view-based mappings (morphisms) between them, is different from Set category: the morphisms (based on a set of complex query computations) are not functions, while the objects are database instances (sets of relations). The logic based schema mappings between databases, usually written in a highly expressive logical language (ex. LAV, GAV, GLAV mappings, or tuple generating dependency) may be functorially translated into this “computation” category DB. A new approach is adopted, based on the behavioral point of view for databases, and behavioral equivalences for databases and their mappings are established. By introduction of view-based observations for databases, which are computations without side-effects, we define a fundamental (Universal algebra) monad with a power-view endofunctor T. The resulting 2-category DB is symmetric, so that any mapping can be represented as an object (database instance) as well, where a higher-level mapping between mappings is a 2-cell morphism. Database category DB has the following properties: it is equal to its dual, complete and cocomplete. Special attention is devoted to practical examples: a query definition, a query rewriting in GAV Database-integration environment, and the fixpoint solution of a canonical data integration model.

💡 Research Summary

The paper proposes a novel categorical denotational semantics for database mappings that are based on views, addressing the limitations of traditional schema‑level mappings such as LAV, GAV, and GLAV. The authors introduce a new category, denoted DB, whose objects are whole database instances (i.e., sets of relations) and whose morphisms are not ordinary functions but view‑based mappings: complex query trees built from SPJR U (Select‑Project‑Join‑Union) operations. This shift from functions to query‑operad trees allows the representation of highly intricate partial mappings that arise in modern data‑integration, peer‑to‑peer, and data‑warehouse scenarios.

A central technical contribution is the definition of a view‑based observation mechanism. For any database instance (A), the authors collect all possible views obtainable by SPJR U queries into a new instance (T A). The operator (T) is shown to be an endofunctor on DB and, crucially, a monad: it satisfies (A \subseteq T A) and (T(T A) = T A), providing a closure property for views. The construction of (T) relies on R‑operads, where the set (R) of relation symbols serves as types, and the operad’s operations correspond to abstract query constructors. An R‑algebra interprets these abstract operations as concrete query functions over actual relation instances, thereby grounding the operadic syntax in relational algebra.

The authors further lift DB to a 2‑category. Morphisms (view‑based mappings) become objects of a higher level, and 2‑cells represent transformations between mappings. This structure yields a symmetric self‑dual category (DB ≅ DBᵒᵖ) and guarantees both completeness and cocompleteness (existence of all limits and colimits). Such properties are essential for reasoning about composition, pushouts, and pullbacks of complex mappings.

Two notions of observational equivalence are introduced. Strong observation equivalence holds when two databases generate exactly the same set of views; weak observation equivalence holds when they share the same canonical (minimal) view representation. These equivalences align with the “certain answer” semantics used in data integration, providing a formal bridge between logical query answering and categorical abstraction.

The paper demonstrates practical relevance through three examples. First, a concrete query definition illustrates how a conjunctive query translates into a morphism in DB. Second, a GAV‑based query rewriting scenario is modeled as a functorial transformation, showing that traditional rewriting algorithms can be expressed as categorical morphisms. Third, the authors define a fixpoint operator for the canonical data‑integration model, handling potentially infinite chains of view generation and proving that the resulting fixpoint is a well‑defined object in DB.

Overall, the work offers a comprehensive framework that unifies view‑based query computation, operadic algebra, monadic semantics, and 2‑categorical structure. By treating database mappings as complex, non‑functional transformations, the DB category captures the richness of real‑world integration tasks while preserving desirable mathematical properties such as duality, completeness, and the existence of universal constructions. This positions the framework as a solid foundation for future research on formal verification, automated reasoning, and tool support for large‑scale, heterogeneous data‑integration systems.