Related work
Formalization work is growing quickly, and very little of it concerns numerical methods. Two figures, redrawn from our paper, show this. Lean libraries of numerical analysis have appeared only since the rise of AI agents: NLA-lib, a library of numerical linear algebra meant also for research done by agents, and our own VSCL. The reading list names, by topic, the work that we build on and stand beside.
Formalization work is growing, and little of it is numerical
arXiv submissions per month
29 submissions on numerical topics, against 1,271 on formalization.
- Formalization, 1,271 in all
- of which mention Lean or Mathlib, 806 in all
- Numerical topics, 29 in all
New public GitHub repositories per month
140 repositories on numerical topics with Lean 4, against 12,523 that mention Lean 4 or Mathlib.
- Lean 4 or Mathlib, 12,523 in all
- Autoformalization, 667 in all
- Lean 4 with numerical topics, 140 in all
Method and data of this figure
The first panel counts arXiv submissions per month, February 2024 to September 2026, whose title or abstract mentions autoformalization, or a form of the word formalization together with a proof assistant. The dashed line counts those that mention Lean or Mathlib. The bars count submissions on floating-point arithmetic, rounding error, numerical analysis or numerical linear algebra together with a proof assistant or formal verification. The second panel counts public GitHub repositories created per month, January 2024 to September 2026, whose name, description or README mentions Lean 4 or Mathlib, autoformalization, or Lean 4 together with floating point, rounding error, numerical analysis or numerical linear algebra. Both counts were retrieved through the public arXiv and GitHub search interfaces on 2026-11-11.
arXiv submissions
| Month | Formalization | of which mention Lean or Mathlib | Numerical topics |
|---|---|---|---|
| Feb 2024 | 17 | 6 | 0 |
| Mar 2024 | 16 | 5 | 0 |
| Apr 2024 | 12 | 3 | 0 |
| May 2024 | 11 | 5 | 0 |
| Jun 2024 | 14 | 6 | 1 |
| Jul 2024 | 20 | 9 | 0 |
| Aug 2024 | 11 | 5 | 0 |
| Sep 2024 | 8 | 4 | 0 |
| Oct 2024 | 27 | 11 | 0 |
| Nov 2024 | 7 | 2 | 0 |
| Dec 2024 | 15 | 3 | 1 |
| Jan 2025 | 23 | 8 | 1 |
| Feb 2025 | 26 | 12 | 1 |
| Mar 2025 | 18 | 12 | 2 |
| Apr 2025 | 23 | 8 | 0 |
| May 2025 | 31 | 17 | 1 |
| Jun 2025 | 23 | 11 | 0 |
| Jul 2025 | 25 | 10 | 1 |
| Aug 2025 | 21 | 10 | 0 |
| Sep 2025 | 35 | 17 | 1 |
| Oct 2025 | 45 | 29 | 1 |
| Nov 2025 | 32 | 19 | 0 |
| Dec 2025 | 31 | 18 | 1 |
| Jan 2026 | 30 | 18 | 0 |
| Feb 2026 | 40 | 22 | 2 |
| Mar 2026 | 60 | 40 | 2 |
| Apr 2026 | 90 | 70 | 1 |
| May 2026 | 76 | 47 | 5 |
| Jun 2026 | 108 | 81 | 1 |
| Jul 2026 | 106 | 78 | 2 |
| Aug 2026 | 112 | 90 | 0 |
| Sep 2026 | 158 | 130 | 5 |
New public GitHub repositories
| Month | Lean 4 or Mathlib | Autoformalization | Lean 4 with numerical topics |
|---|---|---|---|
| Jan 2024 | 61 | 3 | 0 |
| Feb 2024 | 48 | 3 | 0 |
| Mar 2024 | 61 | 7 | 0 |
| Apr 2024 | 82 | 8 | 1 |
| May 2024 | 82 | 7 | 0 |
| Jun 2024 | 75 | 9 | 0 |
| Jul 2024 | 65 | 1 | 1 |
| Aug 2024 | 76 | 2 | 1 |
| Sep 2024 | 86 | 15 | 1 |
| Oct 2024 | 74 | 8 | 1 |
| Nov 2024 | 84 | 4 | 1 |
| Dec 2024 | 87 | 6 | 2 |
| Jan 2025 | 97 | 8 | 1 |
| Feb 2025 | 100 | 13 | 3 |
| Mar 2025 | 104 | 13 | 0 |
| Apr 2025 | 105 | 6 | 0 |
| May 2025 | 156 | 17 | 1 |
| Jun 2025 | 131 | 13 | 0 |
| Jul 2025 | 143 | 14 | 2 |
| Aug 2025 | 132 | 13 | 2 |
| Sep 2025 | 172 | 10 | 0 |
| Oct 2025 | 165 | 19 | 3 |
| Nov 2025 | 223 | 15 | 1 |
| Dec 2025 | 280 | 21 | 2 |
| Jan 2026 | 327 | 33 | 6 |
| Feb 2026 | 459 | 31 | 6 |
| Mar 2026 | 723 | 41 | 4 |
| Apr 2026 | 699 | 37 | 10 |
| May 2026 | 872 | 39 | 11 |
| Jun 2026 | 909 | 60 | 12 |
| Jul 2026 | 1,287 | 69 | 13 |
| Aug 2026 | 1,754 | 48 | 14 |
| Sep 2026 | 2,804 | 74 | 41 |
The lines count formalization work per month, and the bars count formal work on numerical topics. Part of the 2026 rise in repositories reflects ones that coding agents generated. Redrawn from our paper on linear scales.
Reading list
A listing is not an endorsement in either direction.
- Machine-checked theorems about whole algorithms
- Automatic tools for rounding-error bounds
- Formal floating-point arithmetic
- AI-assisted formalization pipelines
- Checking statements against their sources
- Connecting guarantees to running code
- AI systems in mathematics and its formalization
- Formal libraries of mathematics
- Benchmarks for formal theorem proving
- Numerical software
- Other software by our group
Machine-checked theorems about whole algorithms
Most of these works prove rounding-error results for whole algorithms and programs in a proof assistant. They cover the Cholesky factorization, inner and matrix products, the Jacobi iteration, a wave-equation program, an ODE solver, and fast Fourier transforms, and some connect their theorems to C code. Machine-checked versions of the whole-algorithm theorems in textbooks remain few.
- Accuracy and Stability of Numerical Algorithms, second editionN. J. Higham, SIAM, 2002Textbook of rounding-error theorems for whole algorithms, including the Oettli-Prager backward error theorem.
- Formal Proofs of Rounding Error Bounds - With Application to an Automatic Positive Definiteness CheckP. Roux, J. Autom. Reason., 2016Proves a bound for the floating-point Cholesky factor and uses it to check positive definiteness.
- LAProof: A Library of Formal Proofs of Accuracy and Correctness for Linear Algebra ProgramsA. E. Kellison and others, ARITH 2023Coq library of accuracy proofs for inner, matrix-vector, and matrix products, accounting for underflow.
- VCFloat2: Floating-Point Error Analysis in CoqA. W. Appel and A. Kellison, CPP 2024Coq tool for floating-point reasoning, used to connect LAProof's sparse matrix-vector product to C code.
- Verified Software ToolchainA. W. Appel, NFM 2012Coq toolchain for verifying C programs, used with VCFloat.
- Verified Correctness, Accuracy, and Convergence of a Stationary Iterative Linear Solver: Jacobi MethodM. Tekriwal and others, CICM 2023Verifies the Jacobi iteration in floating point, including its stopping test and an overflow check.
- Wave Equation Numerical Resolution: A Comprehensive Mechanized Proof of a C ProgramS. Boldo and others, J. Autom. Reason., 2013Verifies a finite-difference program, written in C, for the wave equation.
- Numerical Analysis of Ordinary Differential Equations in Isabelle/HOLF. Immler and J. Hölzl, ITP 2012Analysis of numerical methods for ODEs in Isabelle/HOL, with enclosures for discretization and rounding errors.
- A Verified ODE Solver and the Lorenz AttractorF. Immler, J. Autom. Reason., 2018Verified executable ODE solver that chooses its time steps adaptively, applied to the Lorenz attractor.
- A Methodology for the Formal Verification of FFT Algorithms in HOLB. Akbarpour and S. Tahar, FMCAD 2004Proves roundoff recurrences for fast Fourier transforms and that a hardware design implements its fixed-point specification.
Automatic tools for rounding-error bounds
Some tools bound the rounding error of a given program automatically. They handle straight-line code and loops with a fixed bound, and they validate their results in different ways, for example with proofs checked in Coq, HOL Light, or PVS, or with a formally verified certificate checker. Type systems such as Bean track backward error through a program.
- Certification of bounds on expressions involving rounded operatorsM. Daumas and G. Melquiond, ACM Trans. Math. Softw., 2010Bounds rounding errors automatically and produces proofs that the Coq proof assistant checks.
- Rigorous Estimation of Floating-Point Round-Off Errors with Symbolic Taylor ExpansionsA. Solovyev and others, ACM Trans. Program. Lang. Syst., 2019Bounds errors with symbolic Taylor expansions, supports subnormals, uncertain inputs, and mixed precision, and produces proof certificates checked in HOL Light.
- Daisy - Framework for Analysis and Optimization of Numerical Programs (Tool Paper)E. Darulova and others, TACAS 2018Analyzes mixed-precision assignments and searches for equivalent expressions with smaller error bounds.
- Rigorous Floating-Point Round-Off Error Analysis in PRECiSA 4.0L. Titolo and others, FM 2024Produces PVS-checked proofs and bounds errors when rounding changes a branch decision.
- A Verified Certificate Checker for Finite-Precision Error Bounds in Coq and HOL4H. Becker and others, FMCAD 2018Certificate checker, formally verified in Coq and HOL4, that checks Daisy's error-bound certificates.
- Bean: A Language for Backward Error AnalysisA. E. Kellison and others, Proc. ACM Program. Lang., 2025Type system that tracks backward error through a program by composing per-input bounds.
Formal floating-point arithmetic
Rounding-error proofs rest on assumptions about how each arithmetic operation rounds. Formal arithmetic libraries help justify those assumptions for particular formats, and some also prove bounds for sums and dot products. In Lean, three such libraries were published in 2026.
- Flocq: A Unified Library for Proving Floating-Point Algorithms in CoqS. Boldo and G. Melquiond, ARITH 2011Coq rounding theory for any radix and precision, a basis for much later floating-point work in Coq.
- Interim Report on Binary Floating-point Formats for Machine LearningIEEE SA P3109 Working Group, 2026Interim report on the IEEE P3109 family of narrow formats, which FLoPS and FloatLib formalize.
- FLoPS: Semantics, Operations, and Properties of P3109 Floating-Point Representations in LeanT.-C. Chang and others, CoRR, 2026Lean formalization of P3109 formats with saturation and stochastic rounding, proving properties of FastTwoSum and ExtractScalar.
- Semantics, Operations, and Properties of P3109 Floating-Point Representations in Lean: FLoPS artifactT.-C. Chang and others, GitHub repository, 2026FLoPS repository: executable arithmetic proved to refine the bit-level specification, and stochastic-rounding bounds for sums.
- FP: IEEE 754 arithmetic and floating-point numerical analysis in LeanM. A. Tunnell, Zenodo, 2026Lean library of IEEE binary32 and binary64 arithmetic with error bounds for sums and dot products.
- FloatLib: Verified Floating-Point Arithmetic in LeanR. J. George, W. Adkisson and A. Anandkumar, CoRR, 2026Lean library modeling IEEE, posit, P3109, block-scaled, and user-defined formats at the level of bits.
- Berkeley TestFloatJ. R. Hauser, 2018Source of more than a hundred million test cases used to check FloatLib's bit-level model.
AI-assisted formalization pipelines
These projects describe ways to organize AI-assisted formalization and to review its results, for textbooks, course material, and a catalogue of open problems. Some divide the work among agents in separate roles, and others run it as an open protocol or platform to which anyone's agents can contribute.
- Automatic Textbook FormalizationF. Gloeckle and others, CoRR, 2026RepoProver formalized a graduate textbook with agents that sketch statements, fill in proofs, and review the results.
- Formalizing Mathematics at ScaleA. Rammal and others, CoRR, 2026Atlas reports applying RepoProver's approach to further online material from MIT OpenCourseWare.
- Prove2Me: An Open Collaborative Platform for Scaling Math FormalizationS. Chen and others, CoRR, 2026Open platform on which anyone's agents can contribute to formalizing mathematics.
- OpenProver: Agentic and Interactive Theorem Proving with Lean 4M. Kripner and M. Straka, CICM 2026Open-source prover for agentic and interactive theorem proving in Lean 4.
- Choir: An Open Protocol for Distributed Multi-Agent AutoformalizationY. Qi and M. Weber, CoRR, 2026Open protocol in which workers' agents claim and prove statements, with an automated gate on contributions.
- Open Problems in Numerical Linear AlgebraA. Townsend, 2026Catalogue of open problems in numerical linear algebra, with Lean statements fixed and reviewed before proving.
- Numerical Methods for Ordinary Differential EquationsJ. C. Butcher, Wiley, 2016Textbook on numerical methods for ODEs, part of which the OpenMath pipeline formalized.
- Formalizing Numerical Analysis: An Agent Pipeline and Quality Audit Beyond Kernel AcceptanceT. Meek and others, CoRR, 2026OpenMath, an agent pipeline that formalized 84 of 175 statements of Butcher's ODE textbook in exact arithmetic and measured statement faithfulness.
Checking statements against their sources
The Lean comparator checks that a proof establishes a statement as written, not that the statement is faithful to its source. Palomar and Tau Ceti address faithfulness directly, through AI review of statements and through review among agents under rubrics written by people. In Choir's demonstration project, a condition omitted when two theorems were transcribed was found by a later review and corrected.
- Comparator: A trustworthy judge for Lean proofsLean FRO, GitHub repository, 2025Checks that a proof establishes the formal statement as written, used by Choir and Townsend's catalogue.
- Formal Conjectures: An Open and Evolving Benchmark for Verified Discovery in MathematicsM. Firsching and others, CoRR, 2026Benchmark of formalized conjectures that warns its statements may not capture the original conjectures.
- Palomar: a registry of Lean verified mathematicsT. Tao, Blog post, What's new, 2026Registry of Lean proofs, opened in August 2026, with mechanical checks and AI review of statements.
- Tau Ceti: a Lean library downstream of Mathlib, written and reviewed with AI agentsTau Ceti Project, 2026Lean library downstream of Mathlib, written by AI agents that review each other's contributions.
- Probabilistic Methods in CombinatoricsY. Zhao, Lecture notes for MIT course 18.226 (Fall 2022)Lecture notes on probabilistic methods, the source of two theorems in Choir's demonstration project.
- Pull request 143, ProbMethodCombinatoricsY. Qi, 2026Pull request, closed unmerged, whose closing comment traces the failure to a condition that the formal statements had omitted.
Connecting guarantees to running code
A theorem about an algorithm does not by itself establish a guarantee for a particular kernel. Evidence about a kernel can be a program proof, a sound output checker, or a test. The first three works give an example of each. The last two are AI-generated solvers that come with Lean files.
- LAProof, version 2.0A. E. Kellison and others, GitHub repository, 2026C code for dense Cholesky and triangular solves, proved to implement functional models.
- Verification methods: Rigorous results using floating-point arithmeticS. M. Rump, Acta Numer., 2010Validated numerics, which computes for one input an interval that provably contains the exact answer.
- How to grade the accuracy of the BLASJ. Demmel and others, CoRR, 2026Proposes grading BLAS kernels by the error bound they satisfy, with randomized tests for each grade.
- AdvectionDiffusionLanyon AI, GitHub repository, 2026AI-generated C solvers for advection-diffusion equations, with Lean files about the same schemes.
- GeneralRelativisticMaxwellLanyon AI, GitHub repository, 2026AI-generated C solvers for general relativistic Maxwell equations, with Lean files about the same schemes.
AI systems in mathematics and its formalization
These are reports on mathematics done with AI systems, most of them from AI laboratories and companies, and several come with Lean formalizations.
- Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical OlympiadT. Luong and E. Lockhart, Google DeepMind blog, 2025Google DeepMind's report on its system at the 2025 International Mathematical Olympiad.
- Advancing Mathematics Research with AI-Driven Formal Proof SearchG. Tsoukalas and others, CoRR, 2026AlphaProof Nexus resolved nine Erdős problems with Lean-checked proofs.
- openai/math: Mathematical manuscripts and supporting proof artifacts produced by an internal OpenAI modelOpenAI, GitHub repository, 2026OpenAI's release of 722 machine-generated manuscripts on 2026-10-06.
- An upper bound of 9/4 for the matrix multiplication exponentOpenAI, OpenAI Math Release preprint OAI:Matrix-Multiplication-Nine-Fourths-October-2-2026Reported upper bound of 9/4 on the exponent of matrix multiplication, with an accompanying Lean formalization.
- The strong Prime Number Theorem (formalization generated by Gauss)Math Inc., 2025Math Inc.'s Gauss completed the formalization of the strong prime number theorem.
- Formalizing Fermat's Last TheoremAnthropic, Anthropic research blog, 2026Reports a Lean proof of Fermat's Last Theorem produced largely autonomously in eleven days.
- Formalizing the Enormous TheoremAxiom Math, 2026Announced Lean formalization of the classification of finite simple groups, about 43.1 million lines of Lean.
- Formal Reasoning Meets LLMs: Toward AI for Mathematics and VerificationK. Yang and others, Commun. ACM, 2026Article on formal reasoning with large language models, toward AI for mathematics and verification.
Formal libraries of mathematics
Machine-checked formalizations rely on formal libraries, which are collections of definitions, theorems and lemmas that a proof assistant checks. Most of the AI systems above build on Mathlib, the main library of Lean.
- The Lean mathematical libraryThe mathlib Community, CPP 2020Mathlib, a formal library of mathematics for the Lean proof assistant.
- mathlib4Lean Community, 2026The Mathlib repository.
- NLA-libT. Chen and others, Website, 2026A Lean library of numerical linear algebra.
- Archive of Formal ProofsArchive of Formal Proofs, 2026The Archive of Formal Proofs, a formal library for Isabelle.
- Mathematical ComponentsA. Mahboubi and E. Tassi, Zenodo, 2022The Mathematical Components library for the Coq proof assistant.
- The Role of the Mizar Mathematical Library for Interactive Proof Development in MizarG. Bancerek and others, J. Autom. Reason., 2018Describes the Mizar Mathematical Library, the formal library of the Mizar system.
Benchmarks for formal theorem proving
Benchmarks hold formal problems for AI systems to prove. Across four widely used Lean benchmarks, one of the 7,091 problems concerns a numerical method.
- miniF2F: a cross-system benchmark for formal Olympiad-level mathematicsK. Zheng, J. M. Han and S. Polu, ICLR 2022Cross-system benchmark of formal Olympiad-level problems, one of the four Lean benchmarks counted in the figure above.
- PutnamBench: Evaluating Neural Theorem-Provers on the Putnam Mathematical CompetitionG. Tsoukalas and others, NeurIPS 2024Benchmark of Putnam competition problems for neural theorem provers, one of the four Lean benchmarks counted in the figure above.
- ProofNet: Autoformalizing and Formally Proving Undergraduate-Level MathematicsZ. Azerbayev and others, CoRR, 2023Benchmark of undergraduate-level mathematics for autoformalization and proving, one of the four Lean benchmarks counted in the figure above.
- FormalMATH: Benchmarking Formal Mathematical Reasoning of Large Language ModelsZ. Yu and others, CoRR, 2025Lean benchmark whose one numerical-method problem runs three steps of gradient descent in exact arithmetic.
Numerical software
Netlib is a collection of mathematical software, papers and databases. LAPACK, the reference library of dense linear algebra, is distributed there, and its test suite has checked the library's numerical behavior for three decades. The LAPACK Working Notes are a series of reports on its development.
- NetlibWebsiteA collection of mathematical software, papers and databases.
- LAPACK: Linear Algebra PACKage, reference implementation and test suiteReference LAPACK, GitHub repository, 2026The reference implementation and its test suite, which generates matrices with prescribed conditioning and structure and checks scaled residuals against a threshold.
- LAPACK Working NotesWebsiteA series of reports on the development of LAPACK.
Other software by our group
RandBLAS and RandLAPACK are C++ libraries for randomized numerical linear algebra (RandNLA). Members of our group are among their developers.
- RandBLASGitHub repositoryExtends the BLAS tradition to randomized operations, such as sketching dense and sparse data.
- RandLAPACKGitHub repositoryDriver algorithms and building blocks for RandNLA, including randomized SVD and QR with column pivoting.