1 Introduction

Fermat’s Last Theorem is the statement that if \(a,b,c,n\) are positive whole numbers with \(n\geq 3\), then \(a^n+b^n\not=c^n\). It is thus the claim that a family of Diophantine equations (\(a^3+b^3=c^3, a^4+b^4=c^4,\ldots \)) has no positive integer solutions. Diophantus was a Greek mathematician who lived around 1800 years ago, and he would have been able to understand the statement of the theorem (he knew about positive integers, addition and multiplication).

Fermat’s Last Theorem was explicitly raised by Fermat in 1637, and was proved by Wiles (with the proof completed in joint work with Taylor) in 1994. There are now several proofs but all of them go broadly in the same direction, using elliptic curves and modular forms.

Lean is an interactive theorem prover; it checks mathematical arguments with super-human accuracy. Explaining a proof of Fermat’s Last Theorem to Lean is in some sense like explaining the proof to Diophantus; for example, the proof starts by observing that before we go any further it’s convenient to first invent/discover zero and negative numbers, and one can point explicitly at places in Lean’s source code here and here where these things happen. However we will adopt a more efficient approach: we will assume all of the theorems both in core Lean and in its mathematics library mathlib, and proceed from there. To give some idea of what this entails: mathlib at the time of writing contains most of an undergraduate mathematics degree and parts of several relevant Masters level courses (for example, the definitions and basic properties of elliptic curves and modular forms are in mathlib). Thus our task can be likened to teaching a graduate level course on Fermat’s Last Theorem to a computer. The computer is quite a challenging audience member – it will insist on being given all technical details of all arguments, and it will not accept proof by intimidation or by appeal to higher authority. Most mathematicians know humans who also behave in this manner. However, it is worse than this; in 2025 at least, the computer will only start filling in details of arguments by itself once the arguments are mathematically utterly obvious. Thus, currently, formalization can be a very time-consuming process.

1.1 Which proof is being formalised?

At the time of writing, these notes do not contain anywhere near a proof of FLT, or even a sketch proof, although we are currently actively working on fixing this.

From 2024 to 2029 we will be beginning to build a proof of FLT, following a strategy constructed by Taylor, taking into account Buzzard’s comments on what would be easy or hard to do in Lean. Our strategy uses refinements of the original Taylor–Wiles method by Diamond/Fujiwara, Khare–Wintenberger, Skinner–Wiles, Kisin, Taylor and others – one could call it a 21st century proof of the theorem. During this initial phase of the project, we shall also be assuming many nontrivial theorems without proof, as long as they were published by 31st December 1989. To get technical for just a second – we shall for example be assuming the existence of Galois representations attached to weight 2 Hilbert modular forms, we will assume Langlands’ cyclic base change theorem for \(\operatorname{GL}_2\), Mazur’s theorem bounding the torsion subgroup of an elliptic curve over the rationals, and several other nontrivial results which were known by the end of the 1980s.

The upshot of this is that, by 2029 at the end of this first phase, the project should contain a complete proof that FLT follows from results that were known to humanity in the 1980s. In particular, one naive way of understanding the goal is that it is a “formalization of the papers of Wiles and Taylor–Wiles, assuming the results in the references of those papers”. However, as noted above, we will actually be taking a slightly different path.

1.2 The structure of this blueprint

This blueprint is a nonlinear document, comprising many chapters. The chapters are not designed to be read in order. Each chapter is self-contained and has a well-defined goal, typically stated at the top of the chapter.

After this chapter, you should next read chapter 2, which explains how to reduce FLT to two highly nontrivial statements about the \(p\)-torsion in a certain elliptic curve (the Frey curve). One of these statements was proved in the 1970s by Mazur, and we shall not be concentrating on it until after the first phase is complete. The other is a theorem of Wiles, and this is what we will be concentrating on in the remainder of the blueprint.

1.3 Remarks

The actual blueprint currently also contains a lot of disorganised ideas. Currently these should be disregarded.

Chapter 4 is an extremely sketchy overview of how the rest of the proof goes. This is currently being expanded and should be ignored right now.

All of the remaining chapters are experiments, and most of them are what I am currently calling “mini-projects”. A mini-project is a bottom-up project, typically at early graduate student level, with a concrete goal. The ultimate goal of many of these projects is to actually get some result into mathlib. We have had one success so far – the Frobenius mini-project is currently being PRed to mathlib by Thomas Browning. Currently most of my efforts are going into running mini-projects, with the two most active ones currently being the adeles mini-project and the quaternion algebra mini-project. These projects do not logically depend on each other for the most part, and one can pick and choose how one reads them.

There is also an appendix, which is again very sketchy, and comprises mostly of a big list of nontrivial theorems many of which we will be assuming without proof in the FLT project.

The next chapter to read, where the proof begins, is chapter 2.