Otomobiltamir
Earth & Environment Science

New Sidon Set Bound Tightens the Leading Correction to √N

Published Oct 29, 2026 Reads 379 By Terence Tao

A new preprint lowers the known upper-bound coefficient for Sidon sets to 0.94304 and extends its method to γ-Golomb rulers.

New Sidon Set Bound Tightens the Leading Correction to √N
I’ll preserve the article’s claims and HTML elements, then expand the explanations and implications using only context supported by the supplied text.

The latest improvement to a long-running Sidon set bound comes from refining the weights used in its proof. A preprint by Daniel Carter, Kevin O’Bryant and Terence Tao establishes an upper bound of \displaystyle k \leq N^{1/2} + 0.94304 N^{1/4} + O(1) for a Sidon set of size k in {1,…,N}. The leading term, √N, sets the basic scale: as the interval grows, the largest possible set grows roughly like its square root. The new result sharpens the next term, reducing the coefficient attached to N1/4. That correction is smaller than the leading term, but it’s where this style of upper bound can distinguish between methods that otherwise agree on the main scale. The authors say their current method appears close to its limit.

The work grew out of a blog discussion last year about using AlphaEvolve to improve the then-best bound. That starting point matters because the result is not presented as a simple application of AI: the authors developed further refinements to the proof strategy, and the paper centers on those mathematical choices. After progress stalled for several months, they posted the results as a preprint, On the diameter and distribution of Sidon sets. They didn’t try the latest 2026 AI tools on their own optimization, so the paper doesn’t claim to settle what automated search might yet find. A different method has since produced a marginally smaller coefficient, 0.942809522; the paper’s focus is the technique, rather than holding the numerical record. That’s a useful distinction in a field where a small improvement can attract attention while the underlying method carries the more lasting value.

Turning unique differences into an inequality

A Sidon set has distinct pairwise sums, or, equivalently, each nonzero difference between two of its elements occurs only once. This condition rules out collisions: two different pairs can’t produce the same sum, and two different pairs can’t account for the same nonzero difference. For a set inside {1,…,N}, the central question is how large it can be before the available sums or differences force a repeat. It’s a clean combinatorial constraint, but converting it into a sharp limit on size is the hard part.

Singer, Bose and Ruzsa gave algebraic constructions with sizes near √N for particular choices of N. Those constructions show that the square-root scale is attainable in important cases, while an upper bound asks how far any set can exceed that scale. Erdős and Turán conjectured an upper bound of √N plus a term of order No(1); Erdős also suggested that a constant-size correction might be too optimistic. The gap between constructions and upper bounds is therefore not just about the leading growth rate. It’s also about the size and form of the excess above √N.

The proof strategy starts from the scarcity of repeated differences. In convolution form, that property gives a pointwise upper bound on the overlap between a set and its reflection. Pairing this bound with a nonnegative autocorrelation function yields an inequality involving two weights, w and t. The weights act as a way to translate local information about overlaps into a global restriction on the set’s size. If their convolution stays above a chosen threshold across {1,…,N}, the inequality constrains how large k can be.

The choice of weights is doing real mathematical work here. The uniqueness condition supplies the raw restriction; the weights determine how effectively the proof collects that restriction across the whole interval. In similar optimization arguments, a convenient choice may make the inequality easy to evaluate but leave some strength unused. A more carefully shaped choice can capture more of the available information, though it may make the analysis harder to control.

Simple cutoff weights already recover the familiar bound k ≤ √N + N1/4 + O(1). Taking w as an indicator of [1,k3/2] and t as an indicator of [−k3/2,N] makes the convolution threshold roughly k3/2. Substituting these choices into the inequality produces the earlier estimate after lower-order terms are set aside. The cutoff construction is valuable because it makes the mechanism visible: a relatively direct pair of weights already yields a meaningful bound, and the new work asks how much more can be extracted by relaxing that rigid shape.

Why the weights matter

The sharper coefficient comes from replacing those abrupt cutoffs with more flexible functions subject to biconvex constraints. Rather than switching abruptly between zero and one, the functions can vary in ways that better fit the inequality. The constraint is that this flexibility must still preserve the properties the proof needs; an arbitrary smoother weight would not automatically improve the result.

The authors go further by choosing the weights randomly: some required properties, including nonnegativity of the autocorrelation, hold in expectation rather than for every individual choice. That shifts part of the argument from checking one fixed construction to controlling the average behavior of a family of choices. It can open room for a stronger estimate, but it also creates more work: the calculations are lengthy and supported by substantial computation. (And this is the part most people overlook.) The numerical improvement depends on a chain of technical conditions, not simply on selecting a more attractive weight profile.

This is more significant than it looks. The coefficient changes only in the correction term, and a different method has already edged it lower, but the result demonstrates how a familiar uniqueness condition can be pushed further by redesigning the proof’s weights. The authors’ view that their method is near its limit also sets a clear boundary on the claim: this is progress within a particular framework, not evidence that the broader problem has been settled. If you’re working in this space, the useful question is whether the expected-property approach can be adapted or whether a genuinely different proof strategy is needed. A narrower margin. A more demanding optimization.

Where the framework may travel

The same framework also applies to γ-Golomb rulers, where each nonzero difference may occur as many as γ times. This extends the setting from strict uniqueness to a controlled allowance for repeated differences. The underlying challenge remains recognizable: use restrictions on how often differences occur to limit the size of a set. That makes the weight design relevant beyond the original Sidon set bound, although the article doesn’t claim that the same numerical improvement transfers unchanged.

The broader takeaway is methodological: carefully designed, sometimes randomized weights can extract stronger bounds from the same underlying uniqueness condition, even as improvements from this approach begin to narrow. That said, the record-setting coefficient isn’t the only measure of progress. The proof offers a structured way to study how weights, convolution thresholds and autocorrelation constraints interact. For readers following additive combinatorics, that’s the durable point: a modest numerical advance can still clarify which parts of a proof are limiting the result, and which might yield to a new choice of weights.

Source: Terence Tao · terrytao.wordpress.com

Discussion

Sign in to join the discussion.