Real Mathematicians Prove Talagrand's Convexity Conjecture

There are quite a few famous unsolved problems in mathematics. Some are so famous they have million dollar bounties on them that will be paid out to the first mathematicians who definitively either prove or disprove them.

And then there are other important conjectures out there that, in addition to academic recognition, might pay smaller rewards to the people who crack them. But which if they are cracked, can be even more valuable if they can successfully be put to use.

One of those smaller conjectures is the Talagrand Convexity Conjecture. Proposed by French mathematician Michael Talagrand in 1995, Talagrand famously offered a $2,000 prize to anyone who solved it. Here is Jules Aknin’s simplified description of the conjecture:

Talagrand’s convexity conjecture is a statement about high‑dimensional geometry: even in enormous, messy clouds of points, simple convex shapes are guaranteed to appear. In other words, no matter how chaotic the configuration looks, there is unavoidable structure hiding inside it.

If such structures are truly unavoidable, they can open the door to practical applications, including in financial markets. Aknin describes how they might contribute to finding order within chaos:

At a deeper level, results like this sit exactly at the intersection of geometry and probability. They don’t just create “pretty shapes” in higher dimensions, they show how structure emerges inside high‑dimensional random systems, which is the same conceptual problem we face when building models for complex datasets. In fact, as the original article notes, this kind of unification between geometric and probabilistic thinking could eventually influence how machines process high‑dimensional data and how we design algorithms that operate in those spaces.

[...]

On the surface, a high‑dimensional cloud of points in geometry and the Australian equity market do not look related. But in practice, we deal with a very similar object every day: hundreds of stocks, multiple time horizons, momentum profiles, risk factors, macro shocks and behavioural flows all interacting at once.

From a distance, that system may appear as pure noise. Talagrand’s result reinforces an idea that systematic managers have believed for a long time: structure is not optional, it is inevitable; the real question is whether you have the tools and discipline to find it.

Even though it doesn’t carry the million-dollar prize of mathematics most famous unsolved problems, the stakes for those who can successfully build on the proven conjecture are still very high. That work is quite possibly worth a lot more than Talagrand’s two-thousand dollar prize for the proof itself.

A team of three mathematicians stand to collect Talagrand’s $2,000. Merrick Hua, Antoine Song, and Stefan Tudose posted a preprint paper of their proof on 11 May 2026, in which they converted the problem from one involving geometry to one involving probability and combinatorics. Like many math stories this year, AI is involved, but unlike those other stories, it’s only a bit player.

The new proof was worked out by Dongming Hua and Antoine Song from the California Institute of Technology, and Stefan Tudose from Princeton University, who joined the other authors after hearing about their work. Together, the mathematicians reformulated Talagrand’s geometric conjecture to a problem of probability theory and random vectors. In their paper published on the arXiv preprint server, they proved an equivalent conjecture for probability, showing that any 1-subgaussian random vector in n dimensions can be expressed as the sum of three standard Gaussian random vectors.

This result solves Talagrand’s convexity problem, proving that for any large enough set in Gaussian space, a convex set of significant measures can be found inside a triple sum of the original set. The solution also confirms a combinatorial analog of the problem, which is important for discrete mathematics.

Initially, Song and Hua say they attempted to work out a solution with the help of ChatGPT. However, while the LLM helped to answer some of their questions and move them closer to a solution, it was Tudose who provided the final proof. Ultimately, the team did not use the work done with ChatGPT. In their paper, the team writes that Tudose’s proof was “more general and conceptual.”

It’s a remarkable achievement. It is also a harbinger of the kind of role AI will find as just another tool used by people.

Image credit: Texture of an iced leaf image by Daniela deGol on Wikimedia Commons. Creative Commons CC By 4.0 Attribution 4.0 International Deed.