Usable Computer-Checked Proofs and Computations for Number Theorists


Proof assistants (also called interactive theorem provers) are increasingly used in academia and industry to verify the correctness of hardware, software, and protocols. However, despite the trustworthiness guarantees they offer, most mathematicians find them too laborious to use.

The goal of the Lean Forward project is to collaborate with number theorists to formally prove theorems about research mathematics and to address the main usability issues hampering the adoption of proof assistants in mathematical circles. The theorems will be selected together with our collaborators to guide the development of formal libraries and verified tools.

Our vehicle will be Lean, a disruptive proof assistant developed at Microsoft Research and Carnegie Mellon University. Lean draws on decades of experience in interactive and automatic theorem provers (e.g., Coq, Isabelle/HOL, and Z3). Its logic is very expressive, and emphasis is placed on strong proof automation. The system is easy to extend via metaprograms that rely on the same logical language that is used to express specifications and proofs.

To support the formalization of theorems, we will develop formal libraries of fundamental number theory and explore how to automate proof search in these. Among many other things, we will create libraries for p-adic numbers and integers and design reasoning procedures for these, exploring how to best exploit Lean's metaprogramming framework. Moreover, we will develop techniques and tools that help mathematicians perform accurate computations and reasoning using proof assistants, integrating procedures from computer algebra systems in a foundational, verified fashion. Finally, we will contribute to Lean's development with ideas and features designed to benefit all users, notably an efficient built-in procedure for first-order reasoning and an integration with automatic theorem provers. The ultimate aim is to develop a proof assistant that actually helps mathematicians, by making them more productive and more confident in their results.

The main strength of proof assistants is their trustworthiness: all definitions and lemmas are checked for well-formedness, and even the most trivial proof steps are verified by an inference kernel with respect to a logical system. Indeed, it took a formal proof to dispel doubts about Hales's proof of the Kepler conjecture.


Proof assistants (also called interactive theorem provers) are interactive tools that make it possible to develop computer-checked, formal proofs of theorems. Two early systems, Automath and Mizar, were designed to mechanize mathematics. Landmark achievements include the formal proofs of the Kepler conjecture by Hales et al. and of the odd-order theorem by Gonthier et al.

The main strength of proof assistants is their trustworthiness: all definitions and lemmas are checked for well-formedness, and even the most trivial proof steps are verified by an inference kernel with respect to a logical system. It took a formal proof to dispel doubts about Hales's proof.

Beyond trustworthiness, formal proofs can also clarify arguments, by exposing and explaining difficult steps. Even without developing any proofs, making theorem statements (including the definitions and hypotheses they rely on) precise can be a substantial gain for communicating results and for making the vast body of mathematics analyzable by computers.

Moreover, by keeping track of changes across large developments, proof assistants facilitate experiments with variants and extensions. When changing a definition, the user is alerted to the inferences that need repair and to redundant definitions, lemmas, and hypotheses.

Finally, modern proof assistants provide automatic proof search that can find long deduction chains quickly, easily beating pen-and-paper methods. Automation is especially useful for highly computational proofs with hundreds or thousands of cases. Despite the success stories, proof assistants can be tedious to use. Most mathematicians find them more of a hindrance than a help. Today, the main users are in hardware verification and programming language research.

There are many obstacles preventing the wide adoption of proof assistants in mathematics.
A zoo of systems
There are many actively developed proof assistants, with incompatible logics and libraries. For example, Coq is based on an expressive logic with dependent types (which facilitate the encoding of abstract algebraic structures), and extensive algebra libraries have been developed for it, but it offers relatively weak proof automation; Isabelle/HOL and HOL Light feature stronger automation and comprehensive analysis libraries but a weaker logic without dependent types; the set-theoretic Mizar system is oriented towards mathematics, but it offers little automation or programmability.
Incomplete mathematical libraries
Formal libraries—consisting of definitions, lemmas, and proofs—are a prerequisite for most formal developments. Before users can apply proof assistants to their own research, they need to formalize a lot of undergraduate and graduate-level mathematics. Even the largest formal libraries cover only a tiny fraction of mathematics.
Weak automation of mathematics
Proof assistants typically combine general-purpose logical automation and procedures for arithmetic. But without domain-specific automation, a single sentence in an informal proof can correspond to dozens or hundreds of lines of formal proof.
Poor interoperability with computer algebra systems
Notwithstanding the existence of a few prototypes, interoperability between computer algebra systems and proof assistants is an open problem. A trustworthy integration of algebra systems in proof assistants requires the validation of certificates produced by the algebra systems. However, most procedures do not generate certificates; for some algorithms, it is not even clear what certificates would look like. Conversely, users of algebra systems could exploit the proof assistants' support for expressing and finding proofs.
Difficult extensibility
Extending proof assistants with procedures for domain-specific reasoning requires a high level of expertise. The programmatic interfaces of the main proof assistants have evolved over several decades and are difficult to learn. For example, to extend Coq, one must typically master both the Ltac tactic language and the low-level OCaml interfaces, in addition to the Gallina specification language.

Despite the many obstacles, there is a strong feeling in various parts of the research community that mathematics deserves to be formalized. Fields medalist Vladimir Voevodsky has been one of the strongest advocates of this view. His research area was plagued by flawed theorems, to the point where he stopped believing pen-and-paper arguments.

With the emergence of huge computational proofs, the issue of peer reviewing is becoming more pressing. There are many disputed theorems, including Mochizuki's claimed proof of the abc conjecture. It sometimes happens that articles pass peer review, and get published in prestigious journals such as the Annals of Mathematics, before it emerges that a lemma is flawed and the entire proof collapses. More applied branches of research, such as cryptography, are also subject to controversies. By relying to a greater extent on computers to develop and check proofs, researchers can raise the reviewing standards, even for computational proofs, thereby increasing confidence in the results' validity.

Proof assistants are powerful, trustworthy tools. If we computer scientists care to listen to interested mathematicians and cooperate with them, we can learn how to direct our tools towards their goals.

& Objectives

Interactive theorem proving is steadily gaining ground. In some areas of computer science, it is common for research papers to be accompanied by formal proofs. Proof assistants are even deployed in the classroom, replacing pen and paper. These circumstances point to a future where these tools will be routinely used, resulting in more reliable science. But to have a substantial impact on mathematical practice, we must narrow the usability gap.

The difficulties are as much social as technical. Proof assistants are developed primarily by computer scientists. Much of the formalized mathematics is motivated by hardware or software verification. Mathematicians largely dismiss proof assistants as impractical, so the technology improves only slowly. Sadly, even routine operations such as factorization of polynomials can become small challenges when moving to a formal-logic environment. We want to break the vicious circle by working closely with mathematicians. We aim to bring a proof assistant, its automation, and its libraries further, guided by the needs of mathematicians who understand the value of proof assistants.

Specifically, we will collaborate with Sander Dahmen and his team at the VU Amsterdam to formalize parts of the team's research in number theory and recent results in related areas, addressing usability issues as they arise. Cooperation with research mathematicians will benefit Lean Forward substantially: not only will they guide us through their field and help us carry out the formalization, they will provide frank feedback on our technology and act as advocates for formal verification in their community.

Our vehicle will be the Lean proof assistant. Lean is a new open source system developed by Leonardo de Moura (Microsoft Research), Jeremy Avigad (Carnegie Mellon), and their colleagues. The system's design and engineering is unusually clean and efficient. Lean attempts to combine the best from two leading proof assistants:

Lean's logical foundation is a variant of Coq's calculus of inductive constructions, a dependent type theory. Lean distinguishes itself with its small inference kernel and strong automation. Independent proof checkers provide additional guarantees. Lean's support for dependent types is smoother than Coq's, thanks to flexible pattern matching and a generalized congruence closure algorithm. A mechanism for introducing quotient types and a transfer tool facilitate reasoning up to isomorphism without resorting to setoids or homotopy type theory.

For the design of basic algebraic libraries, Lean's developers turned to Isabelle/HOL for inspiration. The libraries rely on type classes, a mechanism to categorize types and their operations (e.g., "(Z, +) forms a group"). Type classes interact well with Lean's dependent types. By contrast, in Isabelle/HOL, there is no way to use type classes to reason about the integers modulo n as a ring.

Our overall aim will be met by pursuing four scientific objectives, presented below. Our starting point is that there is tremendous value in simultaneously using and developing a proof assistant. Lean Forward combines these two activities.

Our overall aim is to make proof assistants usable by mathematicians, by initially focusing on number theory and related areas, developing the proof automation, tool integrations, and formal libraries guided by actual users' needs.

The Lean Forward Team

Funded by the project


VU Amsterdam
principal investigator


VU Amsterdam
expected to start in October 2019


VU Amsterdam
PhD student
expected to start in September 2019


VU Amsterdam
PhD student
expected to start in January 2020


VU Amsterdam
PhD student


VU Amsterdam
student assistant
April–September 2019

Associated members


VU Amsterdam

Robert Y.

VU Amsterdam


Inria Rennes – Bretagne Atlantique &
VU Amsterdam


RU Nijmegen

Main collaborators


Carnegie Mellon

de Moura

Microsoft Research

van Raamsdonk

VU Amsterdam

Additional collaborators


VU Amsterdam

van Langen

VU Amsterdam


VU Amsterdam


Former members

Le Hénaff

École Polytechnique Paris


U. Amsterdam


VU Amsterdam



Formalizing the solution to the cap set problem

Sander R. Dahmen, Johannes Hölzl, and Robert Y. Lewis. Accepted at 10th International Conference on Interactive Theorem Proving (ITP 2019).

Preprint (PDF)Proof sketch (PDF)

Superposition with lambdas

Alexander Bentkamp, Jasmin Blanchette, Sophie Tourret, Petar Vukmirović, and Uwe Waldmann. Accepted at 27th International Conference on Automated Deduction (CADE-27).

Preprint (PDF)Report (PDF)

Extending a brainiac prover to lambda-free higher-order logic

Petar Vukmirović, Jasmin Christian Blanchette, Simon Cruanes, and Stephan Schulz. In Vojnar, T., Zhang, L. (eds.) 25th International Conference on Tools and Algorithms for the Construction and Analysis of Systems (TACAS 2019), LNCS, Springer, 2019.

Preprint (PDF)Report (PDF)

A formal proof of Hensel's lemma over the p-adic integers

Robert Y. Lewis. In Mahboubi, A., Myreen, M. O. (eds.) 8th ACM SIGPLAN International Conference on Certified Programs and Proofs (CPP 2019), pp. 15–26, ACM, 2019.

Postprint (PDF)


Meta-programming with the Lean Proof Assistant

Pablo Le Hénaff. MSc internship report, École Polytechnique Paris, 2018.

Thesis (PDF)

A Formally Verified Proof of the Mason–Stothers Theorem in Lean

Jens Wagemaker. BSc thesis, Vrije Universiteit Amsterdam, 2018.

Thesis (PDF)


Logical Verification 2019–2020

Period 2 of 2019–2020, Vrije Universiteit Amsterdam, Amsterdam, the Netherlands. Lecturers: Alexander Bentkamp and Jasmin Blanchette.

Logic and Modeling 2018–2019

Period 5 of 2018–2019, Vrije Universiteit Amsterdam, Amsterdam, the Netherlands. Lecturer: Robert Y. Lewis.

Lean Together 2019

7–11 January 2019, Amsterdam, the Netherlands.

Logical Verification 2018–2019

Period 2 of 2018–2019, Vrije Universiteit Amsterdam, Amsterdam, the Netherlands. Lecturers: Jasmin Blanchette and Johannes Hölzl.


Please write to Jasmin Blanchette for queries related to the project.

The Zulip chat room Lean and the #Lean Forward stream of the Zulip chat room Sneeuwbal are used by members of the project.


The project is thankful for the support and advice of the following friends and colleagues: Alessio d'Arielli, Peter Boven, Wan Fokkink, Cristiano Giuffrida, Mojca Lovrenčak, Roy Overbeek, Anja Palatzke, Lawrence Paulson, Thomas Sturm, Rikkert Stuwe, Mark Summerfield, Susanti Tang-Budiwandojo, and Caroline Waij.