We introduce a new way of composing proofs in rule-based proof systems that generalizes tree-like and dag-like proofs. In the new definition, proofs are directed graphs of derived formulas, in which cycles are allowed as long as every formula is derived at least as many times as it is required as a premise. We call such proofs circular. We show that, for all sets of standard inference rules, circular proofs are sound. For Frege we show that circular proofs can be converted into tree-like ones with at most polynomial overhead. For Resolution the translation can no longer be a Resolution proof because, as we show, the pigeonhole principle has circular Resolution proofs of polynomial size. Surprisingly, as proof systems for deriving clauses from clauses, Circular Resolution turns out to be equivalent to Sherali-Adams, a proof system for reasoning through polynomial inequalities that has linear programming at its base. As corollaries we get: 1) polynomial-time (LP-based) algorithms that find circular Resolution proofs of constant width, 2) examples that separate circular from dag-like Resolution, such as the pigeonhole principle and its variants, and 3) exponentially hard cases for circular Resolution. the hypotheses, and the formula that needs to be proved displays strictly positive balance. With this interpretation of flows, circular proofs have the appealing flavour of a network in which demands are fulfilled by the hypotheses, and flow towards the conclusions, which produce surplus. Accordingly, and in analogy with the theory of classical network flows, it makes no difference whether the flows are required to be integers or real numbers, and valid flow assignments can be found efficiently, when they exist, by linear programming techniques.While proof-graphs with unrestricted cycles are, in general, unsound, we show that circular proofs are sound. We offer two very different proofs of this fact. The first one is combinatorial in nature and is phrased in the style of traditional soundness proofs in standard proof systems. Concretely, given a truth assignment that falsifies the conclusion, the soundness proof constructs a path of falsified formulas until it reaches a hypothesis, and does so by induction on the total flow-sum of the flow assignment that satisfies the flow-balance condition. The second proof is (semi-)algebraic and is phrased in the style of the duality theorem for linear programming. Concretely, we phrase the existence of a flow assignment that satisfies the flow-balance condition as the feasibility of a linear program, and observe that the infeasibility of its dual witnesses the soundness of the proof.
Proof complexity of circular proofsWith all the definitions in place, we proceed to studying the power of circular proofs from the perspective of propositional proof complexity. For Frege systems, which operate with arbitrary propositional formulas through the standard textbook inference rules, we show that circularity adds no power: the circular, dag-like and tree-like variants of Frege polynomially ...