The Beck-Fiala conjecture is a classic problem in combinatorics about how evenly you can "color" or assign weights to elements of a system. Imagine you have a collection of sets, where each element belongs to at most d sets. The conjecture says you can always assign each element a value of +1 or -1 such that every set has a nearly balanced total, with imbalance at most proportional to the square root of d. This kind of balancing problem has broad applications in computer science, statistics, and algorithm design. Despite decades of effort, proving this conjecture has remained elusive except under restrictive conditions.
A recent landmark result by Bansal and Jiang proved the conjecture holds when d is at least the square of the logarithm of the number of elements. The present paper pushes this boundary significantly further, showing the conjecture holds all the way down to d being just slightly larger than a single logarithm of the number of elements. More impressively, this is achieved by an online algorithm, meaning the elements arrive one at a time and the algorithm must make irrevocable coloring decisions on the fly, without knowing what comes next. This is much harder than the offline setting where you can see everything before deciding. The result is essentially the best possible, since it is already known that online algorithms must fail below the logarithmic threshold.
The algorithm works using a technique called a compactly supported Metropolis fixed-point walk, which is a carefully designed random process that keeps the running imbalances under control as elements arrive sequentially. The approach synthesizes ideas from several recent advances on related balancing problems. Notably, the authors also disclose that the proof was developed in collaboration with ChatGPT 5.6 Pro, with the human authors providing high-level direction across several rounds of prompting and then manually verifying and rewriting the output, making this an interesting example of human-AI collaboration in advanced mathematical research.