From Relation Algebra to Semi-join Algebra: An Approach to Graph Query Optimization

Many graph query languages rely on composition to navigate graphs and select nodes of interest, even though evaluating compositions of relations can be costly. Often, this need for composition can be reduced by rewriting toward queries using semi-joins instead, resulting in a significant reduction of the query evaluation cost. We study techniques to recognize and apply such rewritings. Concretely, we study the relationship between the expressive power of the relation algebras, which heavily rely on composition, and the semi-join algebras, which replace composition in favor of semi-joins. Our main result is that each fragment of the relation algebras where intersection and/or difference is only used on edges (and not on complex compositions) is expressively equivalent to a fragment of the semi-join algebras. This expressive equivalence holds for node queries evaluating to sets of nodes. For practical relevance, we exhibit constructive rules for rewriting relation algebra queries to semi-join algebra queries and prove that they lead to only a well-bounded increase in the number of steps needed to evaluate the rewritten queries. In addition, on sibling-ordered trees, we establish new relationships among the expressive power of Regular XPath, Conditional XPath, FO-logic and the semi-join algebra augmented with restricted fixpoint operators.

THE VARIETY OF COSET RELATION ALGEBRAS

Givant [6] generalized the notion of an atomic pair-dense relation algebra from Maddux [13] by defining the notion of a measurable relation algebra, that is to say, a relation algebra in which the identity element is a sum of atoms that can be measured in the sense that the “size” of each such atom can be defined in an intuitive and reasonable way (within the framework of the first-order theory of relation algebras). In Andréka--Givant [2], a large class of examples of such algebras is constructed from systems of groups, coordinated systems of isomorphisms between quotients of the groups, and systems of cosets that are used to “shift” the operation of relative multiplication. In Givant--Andréka [8], it is shown that the class of these full coset relation algebras is adequate to the task of describing all measurable relation algebras in the sense that every atomic and complete measurable relation algebra is isomorphic to a full coset relation algebra.Call an algebra $\mathfrak{A}$ a coset relation algebra if $\mathfrak{A}$ is embeddable into some full coset relation algebra. In the present article, it is shown that the class of coset relation algebras is equationally axiomatizable (that is to say, it is a variety), but that no finite set of sentences suffices to axiomatize the class (that is to say, the class is not finitely axiomatizable).