Logical Characterizations of Recurrent Graph Neural Networks with Reals and Floats

TL;DR

This paper provides exact logical characterizations of recurrent graph neural networks with reals and floats, establishing their expressive power and linking them to modal logics and distributed automata.

Contribution

It introduces precise logical frameworks matching recurrent GNNs with reals and floats, extending prior work to new settings and connecting to distributed automata.

Findings

Recurrent GNNs with reals are characterized by an infinitary modal logic.

Recurrent GNNs with floats are characterized by a rule-based modal logic with counting.

Recurrent GNNs with reals and floats have equivalent expressive power over MSO-definable properties.

Abstract

In pioneering work from 2019, Barcel\'o and coauthors identified logics that precisely match the expressive power of constant iteration-depth graph neural networks (GNNs) relative to properties definable in first-order logic. In this article, we give exact logical characterizations of recurrent GNNs in two scenarios: (1) in the setting with floating-point numbers and (2) with reals. For floats, the formalism matching recurrent GNNs is a rule-based modal logic with counting, while for reals we use a suitable infinitary modal logic, also with counting. These results give exact matches between logics and GNNs in the recurrent setting without relativising to a background logic in either case, but using some natural assumptions about floating-point arithmetic. Applying our characterizations, we also prove that, relative to graph properties definable in monadic second-order logic (MSO), our infinitary and rule-based logics are equally expressive. This implies that recurrent GNNs with reals and floats have the same expressive power over MSO-definable properties and shows that, for such properties, also recurrent GNNs with reals are characterized by a (finitary!) rule-based modal logic. In the general case, in contrast, the expressive power with floats is weaker than with reals. In addition to logic-oriented results, we also characterize recurrent GNNs, with both reals and floats, via distributed automata, drawing links to distributed computing models.