Piecewise Linear Expectation Analysis via $k$-Induction for Probabilistic Programs

Abstract: Quantitative analysis of probabilistic programs aims at deriving tight numerical bounds for probabilistic properties such as expectation and assertion probability, and plays a crucial role in the verification of probabilistic programs. Along this line of research, most existing works consider numerical bounds over the whole state space monolithically and do not consider piecewise bounds. Clearly, monolithic bounds are either conservative, or not expressive and succinct enough in general. To derive more succinct, expressive and precise numerical bounds for probabilistic properties, we propose a novel approach for synthesizing piecewise linear bounds in this work. To this end, we first show how to extract a piecewise feature w.r.t. a given quantitative property from a probabilistic program using latticed $k$-induction that captures a wide and representative class of piecewise bound functions. Second, we develop an algorithmic approach to synthesize piecewise linear upper and lower bounds from the piecewise feature, for which we show that the synthesis of piecewise linear bounds can be reduced to bilinear programming. Third, we implement our approach with the bilinear programming solver Gurobi. The experimental results indicate that our approach is capable of generating tight or even accurate piecewise linear bounds for an extensive set of benchmarks compared with the state of the art.