Learning Summaries of Recursive Functions

We describe a learning-based approach for verifying recursive functions. The Boolean formula learning algorithm CDNF is used to automatically infer function summaries for recursive functions. In contrast to traditional iterative fix point computation-based approaches, ours can quickly guess summaries and verify purported summaries. When purported summaries are incorrect, the learning algorithm refines them by posing queries. We solve examples that are unattainable by a mature model checker for recursive programs.