Sum-Product Conjecture Disproved for Real Numbers

Quick answer: The Erdős–Szemerédi sum-product conjecture predicts that any large set of n numbers must have nearly n² distinct pairwise sums or nearly n² distinct pairwise products. In May 2026, Thomas Bloom, Will Sawin, Carl Schildkraut and Dmitrii Zhelezov disproved the real-number version by building sets where both counts stay below n^(2−c). The integer version is still open, and the rational-number version is equivalent to it.

Start with four numbers: \(\{1,2,3,4\}\). Adding every pair gives seven different sums; multiplying every pair gives nine different products. Change the set to \(\{1,2,4,8\}\), and the balance shifts: ten sums, but only seven products. The sum-product conjecture is about what happens as sets get much larger. It predicted that at least one of the two counts must grow nearly quadratically. In 2026 that prediction was shown to fail for some sets of real numbers.

Where Does the Conjecture Stand?

Number systemStatus of the near-quadratic conjectureWhy
Integers \(\mathbb{Z}\)OpenThe real-number counterexample does not apply.
Rational numbers \(\mathbb{Q}\)Open; equivalent to the integer caseMultiplying a set by a common denominator does not change how many distinct sums or products it has.
Real numbers \(\mathbb{R}\)Disproved (2026)Arbitrarily large counterexamples were constructed.
Complex numbers \(\mathbb{C}\)DisprovedEvery real counterexample is also a complex one.

Sums and Products of a Set

For a finite set \(A\), write \(A+A\) for the set of all sums \(a+b\) and \(A\cdot A\) for the set of all products \(ab\), with \(a\) and \(b\) in \(A\). The size of a set, its number of elements, is written \(|A|\). Repeated answers count only once, and using the same element twice is allowed. Two small examples show the tension:

Set \(A\)\(A+A\)\(|A+A|\)\(A\cdot A\)\(|A\cdot A|\)
\(\{1,2,3,4\}\) (arithmetic progression)\(\{2,3,4,5,6,7,8\}\)7\(\{1,2,3,4,6,8,9,12,16\}\)9
\(\{1,2,4,8\}\) (geometric progression)\(\{2,3,4,5,6,8,9,10,12,16\}\)10\(\{1,2,4,8,16,32,64\}\)7

Evenly spaced numbers have few sums but many products. Powers of 2 have few products but many sums. With 4 elements the effect is small, but for large sets it becomes dramatic: \(\{1,2,\dots,n\}\) has only \(2n-1\) sums, yet its number of products grows almost like \(n^2\).

Why don’t these small examples already break the conjecture, since neither reaches \(4^2 = 16\)? Because the conjecture is about growth for very large sets, and because order does not matter (\(2+3 = 3+2\)), a set of \(n\) numbers can have at most \(n(n+1)/2\) distinct sums or products. “Nearly quadratic” is a statement about the exponent, not about hitting exactly \(n^2\).

What Is the Sum-Product Conjecture?

Erdős published versions of the conjecture for integers, real numbers and complex numbers in 1976, with integers his main interest. In 1983, Erdős and Endre Szemerédi proved that one of the two sets must be noticeably bigger than \(A\) itself and stated the conjecture in its famous form:

\[ \max\big(|A+A|,\ |A\cdot A|\big) \ \ge\ |A|^{2-\varepsilon} \quad \text{for every } \varepsilon > 0 \text{ and all large enough } A. \]

In words: since there are about \(|A|^2\) pairs, one of the two operations should produce almost as many different answers as there are pairs.

ResultWhat it says about \(\max(|A+A|,|A\cdot A|)\)Status
Erdős–Szemerédi (1983)at least \(|A|^{1+c}\) for some small \(c > 0\)Proved
Solymosi (2009), real numbersabout \(|A|^{4/3}\), up to a log factorProved
Later refinements (Konyagin–Shkredov, Rudnev–Stevens, Cushman)exponent about \(4/3 + 10/4407 \approx 1.3356\)Proved
The conjectureat least \(|A|^{2-\varepsilon}\)False for real numbers (2026)
Integers \(\mathbb{Z}\)at least \(|A|^{2-\varepsilon}\)?Open

The 2026 Counterexample

On May 27, 2026, Thomas Bloom, Will Sawin, Carl Schildkraut and Dmitrii Zhelezov posted “The sum-product conjecture is false for real numbers.” Their main theorem says there is an absolute constant \(c > 0\) and arbitrarily large finite sets \(A\) of real numbers with

\[ \max\big(|A+A|,\ |A\cdot A|\big) \ \le\ |A|^{2-c}. \]

The constant is tiny. The authors sketch \(c \ge 0.00000087\) and say not to take that number too seriously. Why does such a tiny gap matter? The conjecture has to hold for every \(\varepsilon > 0\). Choose \(\varepsilon = c/2\). Then \(|A|^{2-c} < |A|^{2-\varepsilon}\), so arbitrarily large sets fall below a bound the conjecture says they must eventually exceed.

Later 2026 preprints by other authors report sharper numbers, including counterexamples with an exponent around 1.958 instead of just below 2. These have not yet been peer reviewed.

The sets are built from algebraic integers, which are not ordinary integers. For example, \(\sqrt2\) is an algebraic integer because it solves \(x^2 – 2 = 0\), and numbers like \(a + b\sqrt2\) (with integers \(a, b\)) illustrate the type. The actual construction uses number fields of very high degree. Roughly, \(A\) is a product of a box of “units” (numbers whose reciprocals are also algebraic integers, which keeps products under control) and algebraic integers arranged in an additive box (which keeps sums under control). The degree of the field has to grow with the size of \(A\), which is exactly why the construction says nothing about ordinary integers.

Where Did AI Come In?

The paper has a section titled “The role of AI in this proof.” The authors credit OpenAI’s May 2026 counterexample to Erdős’s unit distance conjecture, which used similar number-field ideas, as inspiration. They used GPT-5.5 Pro during early discussions and credit it with one lemma suggestion that simplified part of the argument. They describe the main ideas as almost entirely human-generated and say they wrote the paper themselves.

What Is Still Open?

  • Integers. The authors note that the conjecture may still be true for integers, Erdős’s original setting, and for number fields of bounded degree.
  • The true exponent for reals. The best proved lower bound is about \(1.336\), and counterexamples sit just below 2. Nobody knows where the truth lies in between.
  • Many sums and products. The same paper also disproves a stronger version about sums and products of many elements at once, \(a_1 + \dots + a_k\) and \(a_1 \cdots a_k\), for real numbers.

Why Mathematicians Care

The sum-product problem is a central question in additive combinatorics: can a set be structured for addition and multiplication at the same time? Proved sum-product estimates feed into harmonic analysis, expander graphs and estimates for exponential sums, and those results all still stand. Only the near-quadratic prediction failed.

Common Misconceptions

“The sum-product problem is solved.” Only the real-number version of the conjecture is settled, and it was settled in the negative. The integer version is open.

“The Erdős–Szemerédi theorem was disproved.” No. Their 1983 theorem, that one of the two sets grows faster than \(|A|\), is still true. What failed is the stronger conjecture.

“AI disproved it.” The authors credit an AI result for the inspiration and a model for one simplification, but the proof is theirs.

Watch: The Sum-Product Conjecture in Video

If you learn better by watching, these videos are worth your time.

Math News: The Sum-Product conjecture was disproven!! (Dr. Trefor Bazett). A clear short explanation of the conjecture and the 2026 counterexample.
Sum-product problem and incidence geometry (MIT OpenCourseWare). Yufei Zhao’s lecture on the classical sum-product bounds, for readers who want the background.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. For A = {1, 2, 3, 4}, how many different sums a + b are there?
Found a bug? Let us know!
2. Which kind of set has few products but many sums?
Found a bug? Let us know!
3. What did the 2026 paper prove?
Found a bug? Let us know!
4. Why does the construction not settle the integer case?
Found a bug? Let us know!
5. How did AI contribute to the result?
Found a bug? Let us know!

Frequently Asked Questions

What is the sum-product conjecture?

It is the conjecture of Erdős and Szemerédi that for any large finite set of numbers, either the set of pairwise sums or the set of pairwise products has at least |A| to the power 2 minus epsilon elements, for any epsilon greater than 0.

Has the sum-product conjecture been disproved?

Yes, for real numbers. In May 2026, Thomas Bloom, Will Sawin, Carl Schildkraut and Dmitrii Zhelezov constructed arbitrarily large sets of real numbers whose sum set and product set both have at most |A| to the power 2 minus c elements, for a fixed c greater than 0.

Is the sum-product conjecture still open for integers?

Yes. The 2026 counterexample uses number fields of growing degree, so it does not apply to ordinary integers. The authors note the conjecture may still be true there. The rational-number case is equivalent to the integer case.

Are algebraic integers ordinary integers?

No. An algebraic integer is a root of a polynomial with integer coefficients and leading coefficient 1, such as the square root of 2. Ordinary integers are a special case.

Does this disprove the Erdős–Szemerédi theorem?

No. Their 1983 theorem, that the larger of the two sets grows faster than the set itself, remains true. Only the stronger near-quadratic conjecture failed for real numbers.

What is the best known lower bound?

For real numbers, the larger of the two sets has at least about |A| to the power 4/3 + 10/4407, roughly 1.3356, building on Solymosi’s 2009 bound of 4/3.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Sum-Product Conjecture Disproved for Real Numbers - Effortless Math"?

No one replied yet.

Leave a Reply