Main content starts here
ScienceInsider
Math
ET
A visualization of shifting gluon fields, which bind quarks together within a proton, as described by the Yang-Mills theory.Derek Leinweber/Wikimedia Commons
1.0x
00:00
1.0x
One by one, artificial intelligence (AI) is felling some of the biggest trees in the mathematical forest. In September, OpenAI announced it had cracked the notoriously difficult Navier-Stokes fluid dynamics problem, one of math’s seven Millennium Prize challenges. A month later, the company released more than 700 math papers, including a major advance on the Riemann hypothesis, another famous unsolved Millennium Prize Problem involving prime numbers. Its latest release, on 28 October, included a refutation of a decades-old conjecture in computer science—as well as a proof that one of physics’ most successful theories is on firm mathematical footing.
Controversially, OpenAI did not release the 1147-page proof of the physics result, known as the 4D Yang-Mills existence problem. Instead, it appointed a committee of physicists and mathematicians headed by Nobel laureate David Gross to “shepherd” the result. On 10 November, the group released a validation of a formalized version of the proof in Lean, a computer language that can check mathematical arguments against a set of axioms. Although it doesn’t mean the still-unreleased proof is correct, it’s a sign of its logical consistency.
For physicists, the result is unlikely to change their day-to-day work: The Yang-Mills theory has already been remarkably successful at explaining the behavior of particles at accelerators such as the Large Hadron Collider. But important mysteries still remain, including how most of the mass of protons and neutrons arises, and why the quarks that make them up can never be found alone. A rigorous proof could offer fresh insights into these phenomena. “Maybe there’s something we’re missing,” says David Tong, a physicist at the University of Cambridge.
For mathematicians, the work is already a landmark achievement, bringing them closer to yet another Millennium Prize Problem. More broadly, it highlights AI’s usefulness in tackling mathematical problems that languished because the calculations needed to prove them were prohibitively laborious.
“If you can find a problem where you think there’s a plausible strategy, but the obstruction has been the amount of work involved, that’s a pretty decent thing to try to send the [AI model] towards,” says Ethan Sussman, a mathematician at Northwestern University who, with Anthropic’s Levent Alpöge, independently disclosed AI-assisted progress on Yang-Mills last month.
Developed by physicists Chen Ning Yang and Robert Mills in 1954, the Yang-Mills theory provides a framework for describing two of nature’s fundamental forces: the weak nuclear force responsible for certain kinds of radioactive decay, and the strong nuclear force, which binds quarks together inside protons and neutrons. It is a central component of the standard model of particle physics.
But putting it on firm mathematical footing has been difficult. As a quantum field theory, it describes particles as excitations of fields that permeate space, where even empty space is alive with quantum fluctuations. At small scales, these fluctuations can involve arbitrarily high energies, causing Yang-Mills calculations to spiral toward infinity—a no-no in math.
Physicists have developed mathematical techniques to tame these infinities and make reliable predictions. However, they haven’t been able to prove these calculations approximate a mathematically consistent theory, says Martin Hairer, a mathematician at the Swiss Federal Institute of Technology Lausanne.
Establishing the existence of such a theory is one part of the Yang-Mills Millennium Prize Problem. The other is to demonstrate that the theory has a property known as a “mass gap.”
The Yang-Mills theory describes force-carrying particles called gluons, which have no intrinsic mass. Yet the strong force can bind them into composite particles called glueballs, which show up heavy in experiments. “They seemingly somehow conjure a mass out of nothing,” Tong says. Physicists have some intuition for how this happens, he adds: Quantum fields interact with themselves, producing excitations that have mass even though the underlying particles are massless. But this explanation for the mass gap still lacks a mathematical foundation.
OpenAI and the Alpöge-Sussman collaboration claim to have tackled the existence problem, but not the mass gap.
Both teams built on the work of mathematician Tadeusz Bałaban, who came close to solving the existence problem in the 1980s. His approach involved dividing space-time into a grid, making the otherwise daunting calculations manageable. The challenge was to show the Yang-Mills theory remained mathematically consistent as the grid became ever finer, eventually approaching the continuous space-time of the real world.
Bałaban rigorously constructed the theory in one and two dimensions, but the 4D case remained unsolved.
“He himself admitted that there were really complicated calculations he still would need to do, and it’s just too painful for a human to carry that out,” says Antti Kupiainen, a mathematician at the University of Helsinki.
The mountain of calculations made the problem a natural candidate for AI, Hairer says. But it wasn’t simply a matter of finishing Bałaban’s arithmetic. Sussman says Anthropic’s AI model, Claude, found new ways to organize the proof and address gaps Bałaban had left unresolved, including how particles and fields influence one another.
Sussman and Alpöge have published their working process on GitHub, and Sussman is now translating the AI-generated math into a manuscript that humans can follow. He expects to have a version ready within weeks.
However, that translation can be tedious: AI-generated proofs can be hundreds of pages long, filled with dense and unfamiliar language. “It uses strange terminology for familiar things, and so you have to do a lot of completely idiotic work before you even go to the mathematics,” Kupiainen says.
Sussman has encountered similar problems, noting the proofs can obscure why particular approaches were chosen or the mechanisms behind them. “The only way to read one of these AI-written manuscripts is with AI,” he says.
That matters because mathematicians at OpenAI or Anthropic will likely need all the insights they can get in tackling the mass gap problem—where mathematicians have no clear path to a proof, says Sourav Chatterjee of Stanford University. But some researchers think it won’t be long until AI finds a path forward. Among the hundreds of papers released by OpenAI in October is work establishing a mass gap in a toy 2D model.
Given the potential significance of the breakthrough, OpenAI’s decision to withhold the full proof has attracted criticism. The Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, has called for its immediate release. Hairer also questions the choice of Gross to lead the committee: Previously, Gross had dismissed efforts to put quantum field theories such as Yang-Mills on rigorous mathematical foundations as “useless” and “ugly.”
For Sussman, the better approach is to make the research open from the outset. “I want to invite everyone to a collaboration,” he says. Rather than presenting hundreds of pages of AI-generated math as a finished product, he wants to work with other researchers to produce a human-written account of the Yang-Mills theory that can be understood and built on.
But there’s no denying AI is changing how mathematicians work, he adds. Historically, researchers often favored problems with elegant solutions, Sussman says, whereas vast technical undertakings like Bałaban’s struggled to attract sustained attention. If AI can shoulder much of that labor, important but neglected avenues of research could become more attractive. “I’m hoping the effect is to recenter our sense of what’s important,” Sussman says, “so that it realigns with what’s actually physically important.”
doi: 10.1126/science.zubcwgg
Math
Artificial intelligence
Can researchers bring order to the confusing world of AI ‘benchmarks’?
By
Ananya
To deter bioterrorism, U.S. imposes new rules on ordering synthetic DNA
By
Phie Jacobs
Researchers sue to block NSF from demolishing social science directorate