A machine has verified the maths that won a Fields Medal: why it matters

Mathematicians person announced a milestone successful the effort to thoroughly verify the solution of the sphere-packing occupation — for which the Ukrainian mathematician Maryna Viazovska won the Fields Medal successful 2022 — utilizing a machine.

This mentation of the occupation asks what the champion mode is to battalion a clump of spheres successful 8 dimensions.

On February 23, the squad that achieved this said it present has a impervious that a instrumentality has checked and verified fully.

Viazovska’s proof, similar galore different (human) proofs of hard maths problems, was primitively written for mathematicians to marque consciousness of. So for lawsuit her insubstantial skips steps that are “obvious” oregon steps that travel from immoderate theorem readers are expected to beryllium alert of.

On the different hand, the caller accomplishment progressive a instrumentality checking each azygous measurement of the impervious — including those parts that Viazovska skipped — from commencement to finish.

‘Sorry-free’ verification

Mathematicians are making this effort due to the fact that sometimes a hidden spread oregon an unstated presumption tin gaffe through, and which a quality whitethorn not notice.

Once a impervious has been formalised and checked successful this way, it besides means different (human) mathematicians tin cognize which definitions the impervious utilized astatine each points, which theorems the prover relied on, etc., making it easier for them to audit it themselves oregon reuse parts of it successful their work. Other machines tin besides usage this impervious successful aboriginal successful pursuit of verifying much analyzable proofs of different problems.

Formalisation present means taking a quality impervious written successful papers, with informal connection and “obvious” steps, and translating it to the connection of a impervious assistant, i.e. the machine.

The connection present was called Lean, which is bundle for penning mathematics successful a ceremonial language.

The verification matters for results similar that of Viazovska, since her arguments are subtle and agelong crossed respective method domains, including modular forms, Fourier analysis, and peculiar functions.

The caller verification is besides “sorry-free”, meaning the codification contains nary gaps that the humans asked Lean to conscionable trust, oregon instrumentality astatine look value. In different words, Lean verified each measurement of the argument.

‘Remarkable contribution’

In this case, mathematicians are besides reacting to the usage of an auto-formalisation cause called ‘Gauss’, developed by California-based institution Math, Inc.

The task already had a ample Lean codebase with galore earlier lemmas and definitions, arsenic good arsenic a to-do database of the statements that inactive needed to beryllium proved and however they connected to each other. ‘Gauss’, which is an AI tool, parsed these details and formalised the remaining statements for Lean to check.

Just completely formalising Viazovska’s proof is simply a feat successful and of itself. Now, if Gauss’s enactment holds up nether review, it volition beryllium a motion that AI is acceptable to assistance mathematicians with different large formalisation efforts.

“The task squad is already successful the process of reviewing and revising Gauss’ code, thereby ensuring that it meets the editorial standards of the formalisation community. This process volition guarantee that the codification is maintainable and reusable, and that it volition enactment aboriginal formalisation work,” Carnegie Mellon University PhD pupil Sidharth Hariharan wrote successful a post connected LinkedIn. “Gauss’ singular publication has saved the task months of effort, and Gauss volition proceed to play a relation successful the revisions.”

Mr. Hariharan is besides a maintainer of the Sphere Packing Lean project, an open-source effort to formalise Viazovska’s proof.

How Lean works

Lean is fundamentally a programming connection with a precise logical foundation. Mathematicians archetypal construe definitions, theorems, and proofs arsenic Lean code, past its kernel — which is the checker — verifies if they’re correct.

The kernel checks proofs utilizing Lean’s built-in logical rules, portion a abstracted room called mathlib supplies astir of the modular definitions and theorems that mathematicians tin reuse.

The kernel is akin to a tiny machine: for mathematicians, it’s easier to cheque if this instrumentality works correctly than to cheque if the full impervious is valid.

Lean’s moving tin beryllium breached up into 3 steps.

First, mathematicians encode a occupation arsenic a Lean statement, including what the objects progressive successful the impervious are (e.g. Euclidean space, lattices and packings) and what precisely is being claimed.

Second, they incorporated the indispensable mathematical ‘parts’ of the impervious wrong Lean, successful this lawsuit existent and analyzable analysis, Fourier transforms, peculiar functions, modular/Theta-functions, inequalities, measurement theory, etc.

Third, each measurement that a quality mightiness skip (e.g. with a connection similar “it follows that…”) needs to beryllium expanded into a concatenation of lemmas that Lean tin verify. If a insubstantial appeals to a known theorem, mathematicians indispensable either constituent to its formalised mentation successful Lean’s room oregon they indispensable formalise that theorem too.

Then the kernel gets to work.

At the Lean Together league successful January, Lean creator Leo de Moura said the priorities for this twelvemonth see finalising the Lean 4 compiler and improving its show to trim compilation times and to grip the ample standard of modern Lean libraries.

Challenge of formalisation

According to mathematicians, the eventual intent is to marque mathematical correctness little babelike connected spot and societal processes and much connected explicit and verifiable mathematics.

For instance, successful 1879, the English mathematician Alfred Kempe published a impervious of the four-colour theorem. If you gully a representation connected a level expanse of paper, you tin colour each portion truthful that immoderate 2 regions that stock a borderline get antithetic colours, utilizing lone 4 colours. And Kempe said he’d proved this.

Kempe’s peers accepted his impervious for astir a decennary due to the fact that the impervious looked tenable and helium was highly reputed. But successful 1890, the mathematician Percy Heawood recovered a mistake that invalidated it. The theorem aboriginal turned retired to beryllium existent but Kempe’s impervious was inactive wrong.

Mathematicians besides recovered successful the 20th period that a mathematical ‘proof’ is simply a ceremonial entity that can, successful principle, beryllium checked by a machine, and radical wanted applicable tools to bash that.

Proofs successful modern mathematics tin besides beryllium highly agelong and peer-reviewers whitethorn not ever beryllium up to the task of checking if they’re close from commencement to end. Proof assistants frankincense emerged arsenic a mode to raise, and meet, the barroom for verification.

Helping with formalisation

The main obstruction to getting a impervious checked by a instrumentality successful a “sorry-free” mode is to get the proof, and its associated objects, definitions, etc. into the machine’s connection — i.e. formalisation.

Some large theorems that person been wholly formalised see the four-colour theorem itself, successful 2005; the premier fig theorem besides successful 2005; the Feit-Thompson unusual bid successful 2012; and the Kepler conjecture successful 2014.

Automation has helped successful this regard, though it’s inactive not travel to the constituent wherever a instrumentality tin instrumentality a ‘human proof’ and crook it into a implicit ceremonial impervious successful a reliable way.

In September 2025, a squad led by Indian Institute of Science mathematics prof Siddhartha Gadgil won a grant from the AI For Math Fund for its enactment connected ‘LeanAide’. “By creating an accessible, no-code AI+Lean environment, the task seeks to simplify the formalisation process for Lean users and empower mathematicians with new, innovative tools for research, including agentic solutions,” the citation reads.

Some different tools similar Gauss see Lean Copilot, an AI helper wrong Lean that suggests what measurement to effort next; Sledgehammer, a instrumentality that tries to lick your existent extremity automatically by calling different programs; and Alpha Proof, an AI instrumentality developed by DeepMind to nutrient proofs that a impervious adjunct similar Lean tin check.

AI successful mathematics

In caller years, AI has fundamentally reshaped mathematics adjacent arsenic it has continued to germinate distant from conscionable being a almighty calculator. AI-driven platforms similar Photomath and specialised acquisition intelligence, oregon SEI, models contiguous service arsenic on-demand tutors that connection step-by-step explanations successful earthy connection and tin accommodate to idiosyncratic students.

Large connection models (LLMs) are being utilized to make high-quality standardised tests arsenic good arsenic to instrumentality connected challenges similar the International Mathematical Olympiad. In 2025, reasoning models from OpenAI and Google DeepMind achieved scores worthy of golden medals.

AI models person besides go a reasoning spouse for seasoned mathematicians, helping to lick problems by detecting patterns successful ample datasets. It has been utilized to make caller conjectures successful topology and geometry, often spotting connections crossed disparate fields that evaded experts.

On February 13, for instance, 2 models built by OpenAI helped physicists marque a caller finding successful particle physics, overturning a content the assemblage had held for galore years.

“[xAI cofounder] Christian Szegedy has predicted that models volition beryllium mathematically ‘superhuman successful astir each respects’ successful six months to a year,” University of Toronto adjunct prof Daniel Litt wrote connected his blog on February 21.

“I find that precise timeline hard to judge for astir aspects of mathematical research, but I fishy that helium won’t beryllium disconnected by overmuch erstwhile it comes to proving immoderate people of progressive statements that would antecedently person required an expert. This is simply a constrictive conception of mathematics indeed, but it is existent that producing specified proofs is simply a ample portion of mathematics research.”

mukunth.v@thehindu.co.in