The aim of this paper is to provide a general explanation of the “algorithmic content” of proofs, according to a point of view adequate to computer science. Differently from the more usual attitude of program synthesis, where the “algorithmic content” is captured by translating proofs into standard algorithmic languages, here we propose a “direct” interpretation of “proofs as programs”. To do this, a clear explanation is needed of what is to be meant by “proof-execution”, a concept which must generalize the usual “program-execution”. In the first part of the paper we discuss the general conditions to be satisfied by the executions of proofs and consider, as a first example of proof-execution, Prawitz’s normalization. According to our analysis, simple normalization is not fully adequate to the goals of the theory of programs: so, in the second section we present an execution-procedure based on ideas more oriented to computer science than Prawitz’s. We provide a soundness theorem which states that our executions satisfy an appropriate adequacy condition, and discuss the sense according to which our “proof-algorithms” inherently involve parallelism and non determinism. The Properties of our computation model are analyzed and also a completeness theorem involving a notion of “uniform evaluation” of open formulas is stated. Finally, an “algorithmic completeness” theorem is given, which essentially states that every flow-chart program proved to be totally correct can be simulated by an appropriate “purely logical proof”.
Accelerating search-based program synthesis using learned probabilistic models
