Here is a problem that feels like it should have an easy answer. You have a big pile of items — each described by many numbers at once: a cost, a weight, a risk score, a dozen more — and you have to split the pile into two groups so that every one of those numbers comes out as balanced as possible between them. Not just the total cost even, but the total weight even, and the total risk even, all at the same time.

Try it and you find the catch: you can balance one attribute perfectly, or a few of them roughly, but forcing every attribute even at once usually can’t be done. Something is always a little lopsided. The interesting question is not whether you can reach perfect balance — you generally can’t — but how small you can force that unavoidable leftover imbalance to be. The leftover has a name, discrepancy, and there is a whole branch of mathematics devoted to it.

A ±1 coloring of the items+1−1+1+1−1−1+1−1each attribute’s signed total lands near zero0attr aattr bdiscrepancy = 2attr cdiscrepancy = the largest imbalance over every attribute
A ±1 coloring of the items forces every attribute's signed total near zero; the largest leftover imbalance is the discrepancy.

A conjecture about imbalance that never grows

In the early 1980s the mathematician János Komlós made a bold guess about that leftover. Line the items up as vectors — one number per attribute — and suppose each vector has length at most one. Komlós conjectured that you can always split the pile, giving each item a +1 or a −1, so that the imbalance in every single attribute stays below one universal constant. Not a constant that grows with the number of items. Not one that grows with the number of attributes. The same fixed ceiling, always.

That is a genuinely startling claim. Pile on a billion items measured across a thousand dimensions, arranged adversarially to be as hard as possible, and the conjecture says a near-perfect balance still exists — the worst attribute off by no more than a bounded amount. For nearly forty years nobody could prove it, and nobody could find the counterexample that would kill it.

The breakthrough

That is what moved in late 2025. The computer scientists Nikhil Bansal and Haotian Jiang made the first major advance on the Komlós conjecture in nearly thirty years.2 They did not pin down the constant — that holy grail is still open — but they came dramatically close: they proved the imbalance can always be held to roughly the fourth root of the logarithm of N, the number of items.1

It is worth sitting with how small that is. The logarithm of a huge number is already small; its fourth root is smaller still. As Quanta Magazine put it in reporting the result, you will not, anywhere in this universe, meet an N for which the fourth root of log N exceeds five.2 For any problem you could physically write down, the bound is nearly the constant Komlós was reaching for.

The best bound anyone had proved — the square root of log N — had stood since the late 1990s. Shrinking a square root to a fourth root sounds like a footnote. It is the difference between a guarantee that slowly drifts and one that, for every problem that will ever exist, holds flat.

The method, as reported, is a careful divide-and-conquer: split the vectors into two halves and track the dependency between attributes — how nudging the balance of one drags the others — so the imbalance is never allowed to pool into any single dimension.2 Controlling that spill between dimensions is the whole game.

number of items N →worst-case imbalance= 5 (never exceeded, anywhere)√(log N) — old bound, drifts up⁴√(log N) — Bansal–Jiang 2025, holds flat
The bound that shrank: the old √(log N) drifts past 5 as N grows, while the 2025 ⁴√(log N) result holds flat below it for every N that will ever exist.

Why an optimization shop reads a pure-math result

Let us be honest about what this is and isn’t. It will not route your fleet next quarter. It is foundational mathematics — a proof that a good balance exists, not a fast algorithm you point at your data on Monday. So why would a firm that ships routing and scheduling systems care?

Because discrepancy is our own kind of question, distilled to its purest form. Strip a load-balancing problem of everything incidental and what remains is exactly this: split things so every dimension stays even, and prove how even you can guarantee it. The word that carries the weight is guarantee — a promise that holds no matter how the inputs conspire against you. That worst-case stance is the spine of the serious end of operations research, and it is the line between optimization as a science and optimization as a model wrapper.

Where it touches what we build

Once you see the shape, it turns up all over the work on this site.

  • Balance under a guarantee. Splitting items so every dimension stays even is the abstract heart of load-balancing — demand spread across vehicles, work across shifts, exposure across a portfolio. We solve the applied versions; discrepancy studies the outer limit of how balanced any split can possibly be.
  • A ±1 split is a coloring. Giving every item a plus or a minus is a two-coloring of a set — the same combinatorial-optimization world as our separation-layout demo, which colors a conflict graph to seat people a safe distance apart. Different problem, identical discipline.
  • Worst-case, not average-case. A discrepancy bound holds against the most hostile arrangement of inputs, not a typical one. That is precisely the posture of robust optimization: don’t plan for the average day and hope — guarantee the plan across the whole range of what could happen.

That last one is close to home. Renan’s doctoral work produced “Hardness,” a measure of how much punishment a solution can absorb before it becomes infeasible — a worst-case guarantee bolted directly onto a decision. The Komlós story and that research rhyme: both refuse to ask “how good on a good day?” and insist on “what can you promise on the worst one?” It is no accident, either, that Bansal is a figure from approximation algorithms — bin packing, scheduling — the same lineage as the minimum-fleet math behind our transit work.

The honest line

We do not do discrepancy theory. We build optimization and AI systems that move a measurable number for a client, and a proof about existence bounds is not that. But we read this field closely, because the frontier of what is provable quietly sets the ceiling on what is buildable. Today’s existence bound is tomorrow’s algorithm and next decade’s solver you stop thinking about. Linear programming was once an academic curiosity; now it prices the seat you fly in.

So the thing we take from a breakthrough like this is not a feature to ship. It is a reminder of what the word optimal is supposed to mean — a promise you can defend against the worst case, not a lucky number that looked good on a calm afternoon. That standard is older and deeper than any of our code, and it is the one we hold the work to.