Correctness by construction for probabilistic programs
Abstract
The "correct by construction" paradigm is an important component of modern Formal Methods, and here we use the probabilistic Guarded-Command Language pGCL to illustrate its application to probabilistic programming. pGCL extends Dijkstra's guarded-command language GCL with probabilistic choice, and is equipped with a correctness-preserving refinement relation () that enables compact, abstract specifications of probabilistic properties to be transformed gradually to concrete, executable code by applying mathematical insights in a systematic and layered way. Characteristically for "correctness by construction", as far as possible the reasoning in each refinement-step layer does not depend on earlier layers, and does not affect later ones. We demonstrate the technique by deriving a fair-coin implementation of any given discrete probability distribution. In the special case of simulating a fair die, our correct-by-construction algorithm turns out to be "within spitting distance" of Knuth and Yao's optimal solution.