What do Shannon-type Inequalities, Submodular Width, and Disjunctive Datalog have to do with one another?

TL;DR

This paper establishes a deep connection between information theory, database query bounds, and algorithm design by extending bounds to disjunctive datalog and introducing the PANDA algorithm, which matches submodular width runtimes.

Contribution

It introduces Shannon flow inequalities and proof sequences to derive tight bounds and algorithms for disjunctive datalog, generalizing previous results to broader query classes.

Findings

PANDA algorithm matches submodular width runtimes.

Bounds apply to disjunctive datalog with functional dependencies.

Improves runtime bounds over previous algorithms.

Abstract

Recent works on bounding the output size of a conjunctive query with functional dependencies and degree constraints have shown a deep connection between fundamental questions in information theory and database theory. We prove analogous output bounds for disjunctive datalog rules, and answer several open questions regarding the tightness and looseness of these bounds along the way. Our bounds are intimately related to Shannon-type information inequalities. We devise the notion of a "proof sequence" of a specific class of Shannon-type information inequalities called "Shannon flow inequalities". We then show how such a proof sequence can be interpreted as symbolic instructions guiding an algorithm called "PANDA", which answers disjunctive datalog rules within the time that the size bound predicted. We show that PANDA can be used as a black-box to devise algorithms matching precisely the fractional hypertree width and the submodular width runtimes for aggregate and conjunctive queries with functional dependencies and degree constraints. Our results improve upon known results in three ways. First, our bounds and algorithms are for the much more general class of disjunctive datalog rules, of which conjunctive queries are a special case. Second, the runtime of PANDA matches precisely the submodular width bound, while the previous algorithm by Marx has a runtime that is polynomial in this bound. Third, our bounds and algorithms work for queries with input cardinality bounds, functional dependencies, and degree constraints. Overall, our results show a deep connection between three seemingly unrelated lines of research; and, our results on proof sequences for Shannon flow inequalities might be of independent interest.