Automated Theorem Proving: After 25 Years

Formal verification involves the use of logical and computational methods to establish claims that are expressed in precise mathematical terms. These can include ordinary mathematical theorems, as well as claims that pieces of hardware or software, network protocols, and mechanical and hybrid systems meet their specifications. In practice, there is not a sharp distinction between verifying a piece of mathematics and verifying the correctness of a system: formal verification requires describing hardware and software systems in mathematical terms, at which point establishing claims as to their correctness becomes a form of theorem proving. Conversely, the proof of a mathematical theorem may require a lengthy computation, in which case verifying the truth of the theorem requires verifying that the computation does what it is supposed to do. The gold standard for supporting a mathematical claim is to provide a proof, and twentieth-century developments in logic show most if not all conventional proof methods can be reduced to a small set of axioms and rules in any of a number of foundational systems. With this reduction, there are two ways that a computer can help establish a claim: it can help find a proof in the first place, and it can help verify that a purported proof is correct. Automated theorem proving focuses on the “finding” aspect. Resolution theorem provers, tableau theorem provers, fast satisfiability solvers, and so on provide means of establishing the validity of formulas in propositional and first-order logic. Other systems provide search procedures and decision procedures for specific languages and domains, such as linear or nonlinear expressions over the integers or the real numbers. Architectures like SMT (“satisfiability modulo theories”) combine domain-general search methods with domain-specific procedures. Computer algebra systems and specialized mathematical software packages provide means of carrying out mathematical computations, establishing mathematical bounds, or finding mathematical objects. A calculation can be viewed as a proof as well, and these systems, too, help establish mathematical claims. Automated reasoning systems strive for power and efficiency, often at the expense of guaranteed soundness. Such systems can have bugs, and it can be difficult to ensure that the results they deliver are correct. In contrast, interactive theorem proving focuses on the “verification” aspect of theorem proving, requiring that every claim is supported by a proof in a suitable axiomatic foundation. This sets a very high standard: every rule of inference and every step of a calculation has to be justified by appealing to prior definitions and theorems, all the way down to basic axioms and rules. In fact, most such systems provide fully elaborated “proof objects” that can be communicated to other systems and checked independently. Constructing such proofs typically requires much more input and interaction from users, but it allows us to obtain deeper and more complex proofs. The Lean Theorem Prover aims to bridge the gap between interactive and automated theorem proving, by situating automated tools and methods in a framework that supports user interaction and the construction of fully specified axiomatic proofs. The goal is to support both mathematical reasoning and reasoning about complex systems, and to verify claims in both domains.

Automated and interactive theorem proving are the two main directions in the field of deduction. Most chapters of this book belong to either the one or the other, whether focusing on theory, on methods or on systems. This reflects the fact that, for a long time, research in computer-aided reasoning was divided into these two directions, driven forward by different communities. Both groups offer powerful tools for different kinds of tasks, with different solutions, leading to different performance and application profiles. Some important examples are: ACL2 (Kaufmann and Moore, 1988), HOL (Gordon, 1988), IMPS (Farmer et al., 1996), Isabelle (Paulson, 1994), KIV (Reif et al., 1997) (see also Chapter II.1.1), NQTHM (Boyer and Moore, 1979), and PVS (Owre et al., 1992) for the interactive (or tactical) theorem proving community; and KoMeT (Bibel et al., 1994), Otter (Wos et al., 1992), Protein (Baumgartner and Furbach, 1994), Setheo (Goller et al., 1994), Spass (Weidenbach et al., 1996), and 3 TAP (Beckert et al., 1996) for the automated theorem proving community.KeywordsTheorem ProverSequent CalculusSoftware VerificationAutomate Theorem ProverProof SearchThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

We introduce a heuristic for automatically checking the validity of first-order formulas of the form forallalpham existbetan middot Psi(alpham,betan) that are encountered in inductive proofs of hardware correctness. The heuristic introduced in this paper is used to automatically check the validity of k-step induction formulas needed to verify hardware designs. The heuristic works on word-level designs that can have data and address buses of arbitrary widths. Our refinement heuristic relies on the idea of predicate instantiation introduced in (H. Abu-Haimed et al., 2003). The heuristic proves quantified formulas by the use of a validity checker, CVC (A. Stump et al., 2002), and a first-order theorem prover, Otter (W.W. McCune, 1994). Our heuristic can be used as a stand-alone technique to verify word-level designs or as a component in an interactive theorem prover. We show the effectiveness of this heuristic for hardware verification by verifying a number of hardware designs completely automatically. The large size of the quantified formulas encountered in these examples shows the effectiveness of our heuristic as a component of a theorem prover

And-Inverter Graphs (AIGs) are a popular way to represent Boolean functions (like circuits). AIG simplification algorithms can dramatically reduce an AIG, and play an important role in modern hardware verification tools like equivalence checkers. In practice, these tricky algorithms are implemented with optimized C or C++ routines with no guarantee of correctness. Meanwhile, many interactive theorem provers can now employ SAT or SMT solvers to automatically solve finite goals, but no theorem prover makes use of these advanced, AIG-based approaches. We have developed two ways to represent AIGs within the ACL2 theorem prover. One representation, Hons-AIGs, is especially convenient to use and reason about. The other, Aignet, is the opposite; it is styled after modern AIG packages and allows for efficient algorithms. We have implemented functions for converting between these representations, random vector simulation, conversion to CNF, etc., and developed reasoning strategies for verifying these algorithms. Aside from these contributions towards verifying AIG algorithms, this work has an immediate, practical benefit for ACL2 users who are using GL to bit-blast finite ACL2 theorems: they can now optionally trust an off-the-shelf SAT solver to carry out the proof, instead of using the built-in BDD package. Looking to the future, it is a first step toward implementing verified AIG simplification algorithms that might further improve GL performance.

This volume contains the proceedings of the Workshop on Partiality and Recursion in Interactive Theorem Provers (PAR 2010) which took place on July 15 in Edinburgh, UK. This workshop was held as a satellite workshop of the International Conference on Interactive Theorem Proving (ITP 2010), itself part of the Federated Logic Conference 2010 (FLoC 2010). This workshop is a venue for researchers working on new approaches to cope with partial functions and terminating general (co)recursion in theorem provers.

The formalization of mathematics in practice relies heavily on proof assistants and automatic theorem provers, therefore we studied what are the state of the art proof assitants and their limitations to understand what are the main challenges in making formalized mathematics common practice among mathematicians. We found out that curretly the two major dificulties in formalizing mathematics with proof assistants are due to steep learning curves in how to use these tools and due to a wide gap between the notation employed in these proof assistants and the currently used mathematical notation. We also developed a C++ library to develop proof assistants with great notational flexibility.

Theorem proving procedures for the propositional calculus have traditionally relied on syntactic manipulations of the formula to derive a proof. In particular, clausal theorem provers sometimes lose some of the obvious semantics present in the theorem, in the process of converting the theorem into an unnatural normal form. Most existing propositional theorem provers do not incorporate substitution of equals for equals as an inference rule. In this paper we develop a “natural” proof system for the propositional calculus, with the goal that most succinct mathematical proofs should be encodable as short formal proofs within the proof system. The main distinctive features of NPS are: We show that from a complexity theoretic viewpoint NPS is at least as powerful as the resolution procedure. We further demonstrate formulas on which NPS fares better than resolution. Finally, since proofs in NPS usually resemble manual proofs, we feel that NPS is easily amenable to an interactive theorem prover.

This note considers the design of user interfaces for interactive theorem provers. The basic rules of interface design are reviewed, and their applicability to theorem provers is discussed, leading to considerations about the particular challenges of interface design for theorem provers. A short overview and classification of existing interfaces is given, followed by suggestions of possible future work in the area.

Many modern program verifiers are based on automated theorem provers, which enable full hiding of proof details and allow users to focus all their effort on the program text. This has the advantage that the additional expertise of theorem provers is not required, but has the drawback that when the prover fails to verify a valid program, the user has to annotate the program text with guidance for the verifier. This can be tedious, low-level and repetitive, and may impact on the annotation overhead, readability of the program text and overall development time. Inspired by proof tactics for interactive theorem provers [19], a notion of 'tactics' for the state-of-the-art Dafny program verifier, called Tacny, is developed. With only minor extensions to the Dafny syntax, a user can encode high-level proof patterns as Dafny tactics, liberating herself from low-level and repetitive search tasks, whilst still working with familiar Dafny programming constructs. Manual search and guidance can be replaced with calls to such tactics, which will automate this task. We provide syntax and semantics for Tacny, and show feasibility through a prototype implementation, applied to several examples.

Artificial Intelligence for Theorem Proving (AITP) has given rise to a plethora of benchmarks and methodologies, particularly in Interactive Theorem Proving (ITP). Research in the area is fragmented, with a diverse set of approaches being spread across several ITP systems. This presents a significant challenge to the comparison of methods, which are often complex and difficult to replicate. Addressing this, we present BAIT, a framework for the fair and streamlined comparison of learning approaches in ITP. We demonstrate BAIT’s capabilities with an in-depth comparison, across several ITP benchmarks, of state-of-the-art architectures applied to the problem of formula embedding. We find that Structure Aware Transformers perform particularly well, improving on techniques associated with the original problem sets. BAIT also allows us to assess the end-to-end proving performance of systems built on interactive environments. This unified perspective reveals a novel end-to-end system that improves on prior work. We also provide a qualitative analysis, illustrating that improved performance is associated with more semantically-aware embeddings. By streamlining the implementation and comparison of Machine Learning algorithms in the ITP context, we anticipate BAIT will be a springboard for future research.

Watson is a general-purpose system for formal reasoning. It is an interactive equational higher-order theorem prover. The higher-order logic supported by the prover is distinctive in being type free (it is a safe variant of Quine's i>NF). Watson allows the development of automated proof strategies, which are represented and stored by the prover in the same way as theorems. The mathematical foundations of the prover and the way these are presented to a user are discussed. The paper also contains discussions of experiences with the prover and relations of the prover to other systems.

Agda is a dependently typed functional programming language and a proof assistant in which developing programs and proving their correctness is one activity. We show how this process can be enhanced by integrating external automated theorem provers, provide a prototypical integration of the equational theorem prover Waldmeister, and give examples of how this proof automation works in practice.KeywordsProof ObligationProof AssistantAutomate Theorem ProveEquational LogicFunctional Programming LanguageThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Most computer checked proofs are tied to the particular technology of a prover’s software. While sharing results between proof assistants is a recognized and desirable goal, the current organization of theorem proving tools makes such sharing an exception instead of the rule. In this talk, I argue that we need to turn the current architecture of proof assistants and formal proofs inside-out. That is, instead of having a few mature theorem provers include within them their formally checked theorems and proofs, I propose that proof assistants should sit on the edge of a web of formal proofs and that proof assistant should be exporting their proofs so that they can exist independently of any theorem prover. While it is necessary to maintain the dependencies between definitions, theories, and theorems, no explicit library structure should be imposed on this web of formal proofs. Thus a theorem and its proofs should not necessarily be located at a particular URL or within a particular prover’s library. While the world of symbolic logic and proof theory certainly allows for proofs to be seen as global and permanent objects, there is a lot of research and engineering work that is needed to make this possible. I describe some of the required research and development that must be done to achieve this goal.

Algorithmic methods can successfully automate the proof, and even the discovery, of a large class of identities involving sums of hypergeometric terms. In particular, the Wilf-Zeilberger (WZ) algorithm is a uniform framework for a substantial class of hypergeometric summation problems. This algorithm can produce a rational function certificate that can, on the face of it, be used to verify the result by routine algebraic manipulations, independently of the working of the algorithm that discovered it. It is therefore very natural to consider using this certificate to produce, by automated means, a rigorous deductive proof in an interactive theorem prover. However, naive presentations of the WZ method tend to gloss over trivial-looking but rather knotty questions about zero denominators, which makes their rigorous formalization tricky and their ultimate logical justification somewhat obscure. We describe how we have handled these difficulties to produce rigorous WZ proofs inside the HOL Light theorem prover.

Interactive theorem provers (ITPs) exploit the collaboration between humans and computers, enabling proof of complex theorems. Further, ITPs allow extraction of provably correct implementations from proofs. However, often, the extracted code interface with external libraries containing real-life complexities—proprietary library calls, remote/cloud APIs, complex models like ML models, inline assembly, highly non-linear arithmetic, vector instructions etc. We refer to such functions/operations as closed-box components. For such components, the user has to provide appropriate assumed lemmas to model the behavior of these functions. However, we found instances where these assumed lemmas are inconsistent with the actual semantics of these closed-box components. Hence, even correct-by-construction code extracted from an ITP may still behave incorrectly when interfaced with such closed-box components.To this end, we propose StarFuzz, that allows the $$\text {F}^\star $$ F ⋆ interactive theorem prover to provide better end-to-end assurance on the application— even when interfaced with the closed-box components. Under the hood, StarFuzz rides on Sādhak, an SMT solver that combines fuzz testing to allow satisfiability checking over closed-box components. On the $$\text {F}^\star $$ F ⋆ library that includes external implementations in OCaml, StarFuzz discovered four bugs—one bug that revealed an error on the assumed lemmas for a closed-box function, and three bugs in the external implementations of these components.