How the gap between primes came down to 246 Image: Voice of America (public domain)

How the gap between primes came down to 246

The twin prime conjecture says that there are infinitely many pairs of primes that differ by 2, such as 11 and 13, 101 and 103, or 1,000,000,007 and 1,000,000,009. It is older than anyone can trace and nobody has proved it. Until 2013 nobody had even proved that there is any number N at all such that infinitely many pairs of primes differ by less than N. The average gap between primes near x is ln x, which grows without bound, and for all anyone knew the smallest gaps might grow too.

Zhang

On 17 April 2013 the Annals of Mathematics received a paper from Yitang Zhang, a lecturer at the University of New Hampshire who had published one paper in his life and had spent years after his doctorate without an academic job, at one point working in a sandwich shop. The paper proved that there are infinitely many pairs of primes differing by less than 70,000,000. The referees reported within three weeks, which is unheard of for the Annals, and the paper appeared in 2014. The number 70 million was not optimized; Zhang's point was that it was finite.

The proof built on a method of Daniel Goldston, János Pintz, and Cem Yıldırım from 2005, known as GPY, which came within a hair of bounded gaps and could not close it. GPY needed a certain statement about how evenly primes are distributed among arithmetic progressions to hold a little beyond the level that the Bombieri-Vinogradov theorem provides, and nobody could push it beyond. Zhang found that a weakened form of the stronger statement, restricted to moduli with only small prime factors, was both enough for GPY and provable with the heavy machinery of Deligne's work on exponential sums. Unit 16 of the masterclass sets this out at the level of what each piece does.

Polymath, and Maynard

Within weeks of Zhang's announcement, Terence Tao proposed an online collaboration to reduce the number, and the Polymath8 project ran through the summer of 2013 on a public blog and wiki, with contributors from professors to graduate students. The number fell week by week, through 4,680 in July, by sharpening every part of Zhang's argument, and a record table on the wiki recorded each step with its proof.

Then in November 2013 James Maynard, a postdoctoral researcher at the Université de Montréal who had finished his doctorate in Oxford that year, posted a paper that did something different. Instead of improving Zhang's distribution theorem, he went back to GPY and replaced its single sieve weight by a weight depending on each element of the prime tuple separately, a multidimensional version of Selberg's sieve. With the ordinary Bombieri-Vinogradov theorem, no Deligne, no Zhang, the new weights gave infinitely many pairs of primes differing by at most 600. Terence Tao found the same idea independently at the same time. Maynard's method also proved something Zhang's could not: for every m there are infinitely many intervals of bounded length containing m primes, not just two. The paper, Small gaps between primes, appeared in the Annals in 2015 and is the research spotlight of Unit 16.

Polymath8b, a second phase of the project, combined Maynard's weights with Zhang's distribution results and optimized the whole apparatus by computer. By the spring of 2014 the bound was 246, and there it has stayed. Under the Elliott-Halberstam conjecture, an unproved but widely believed statement about primes in progressions, the same method gives 12, and under a generalized form of it, 6. Not 2.

What was actually proved

The statement "infinitely many pairs of primes differ by at most 246" does not say which difference occurs infinitely often. It says that among the differences 2, 4, 6, ..., 246, at least one occurs infinitely often. A twin prime conjecture for some unknown even number below 246 is what has been proved, and that is a good deal weaker than the conjecture for 2. The method works with admissible tuples, sets of k integers that avoid covering all the residues modulo any prime, and shows that infinitely often at least two members of a shifted tuple are prime. The narrowest admissible tuple with 50 elements has width 246, and the proof shows that among 50 well-chosen numbers in an interval of that width, two are prime infinitely often. To get 2 one would need a tuple of 2 elements, which means showing that both members of {0, 2} are prime infinitely often, which means the method would have to find two primes among two candidates, and it cannot.

Why 6 is a wall

The reason it cannot is called the parity problem, and it was identified by Selberg in the 1940s. A sieve estimates how many members of a set survive the removal of all multiples of small primes. The estimate is built from information about multiples, and that information cannot distinguish a number with an odd number of prime factors from a number with an even number: the sieve weights see 15 = 3 × 5 and 17 in the same way, as "no factor below 3". Any pure sieve argument therefore proves at best that there are infinitely many pairs with both members "prime or a product of two primes", and when the aim is two primes among two candidates, the parity obstruction stops it one step short. Chen's theorem of 1973, that every large even number is a prime plus a number with at most two prime factors, is the same wall seen from the Goldbach side. Unit 15 of the masterclass explains the parity problem with Selberg's own example.

Getting past it needs information that is not of sieve type, which is to say information about primes themselves, such as Friedlander and Iwaniec found for the polynomial a² + b⁴ in 1998 by exploiting the special structure of that polynomial. Nobody has found the analogue for n and n + 2.

Where the bound stands

Since 2014 the number 246 has not moved, and the Polymath project's final paper lists it as the end of the road for the current methods without a new idea on the distribution side. The record for three primes in a bounded interval is 398,130, for four it is 24,797,814, and the lengths grow quickly with the number of primes, which is itself a theorem of Maynard's method. On the other side, in 2014 Ford, Green, Konyagin, and Tao, and independently Maynard, proved that large gaps are larger than the bound Rankin had set in 1938, settling a question for which Erdős had offered 10,000 dollars, his largest prize.

The twin prime conjecture itself is in the position it was in: believed by everyone, supported by the Hardy-Littlewood prediction that the number of twin prime pairs below x is about 1.32 x / (ln x)², which fits the 8,169 pairs below a million and the counts beyond to within a fraction of a percent, and unproved. The difference between 2013 and 2012 is that "infinitely many primes close together" is now a theorem, with a number attached, and the number is 246.

Unit 16 of the masterclass, the last of the program, follows the chain from the average gap to the records on both sides, explains the GPY idea and Maynard's weights at the level of what each step achieves, and ends with a table of the bounds from 70 million to 246 with their dates. It is the unit that ends where the research is.

Study the integers properly

The NT Masterclass: weekly units, quizzes, a final exam, and a certificate.
About the program
Keep reading

Related articles