LRA Interpolants from No Man’s Land

Abstract

Interpolation is becoming a standard technique for over-approximating state spaces in software model checking with Satisfiability Modulo Theories (SMT). In particular when modelling programs with linear arithmetics, the standard state-of-the-art technique might provide either interpolants that are too specific or too generic to be useful for a given application. In this work we introduce the SI-LRA interpolation system for linear real arithmetics that allows the tuning of interpolants based on shifting between the primal and dual interpolants. We prove a strength relation between the interpolants constructed by SI-LRA, and integrate SI-LRA into a propositional interpolator in an SMT solver. Our evaluation, performed using a state-of-the-art software model checker, reveals that correct tuning with SI-LRA can reduce the number of needed refinements by up to one third and provide lower runtimes.