Bounded Gaps: Zhang's Theorem

The primes thin out as you climb — near x the average gap between consecutive primes grows like \ln x, so gaps march off to infinity on average. Yet twin primes — pairs like (11,13) and (101,103) that differ by only 2 — keep appearing, seemingly forever. The Twin Prime Conjecture says there are infinitely many of them, and it had resisted every attack for over a century. Nobody could even prove the far weaker statement that some fixed gap size recurs infinitely often. For all anyone could prove, the gaps might swell without limit and small gaps might eventually run out.

Then, in April 2013, an unknown mathematician settled that weaker statement in one astonishing stroke. This page tells the story and, more importantly, opens the engine: how an extra sliver of arithmetic information — a level of distribution pushed just past the classical barrier — turns a sieve that "nearly" finds two primes into one that provably does, infinitely often.

The theorem, and the number that stunned everyone

Read that \liminf carefully. It does not say every gap is below 70 million — the gaps still average \ln x and grow arbitrarily large. It says the gap sequence dips back below 70{,}000{,}000 over and over, no matter how far out you go. Before 2013 even that was out of reach; the best unconditional result (Goldston–Pintz–Yıldırım, 2005) showed the gaps are infinitely often much smaller than average — but "much smaller than \ln x" still tends to infinity. Zhang made the bound a genuine constant.

The story: an unknown lecturer and the Annals

Yitang Zhang (born 1955) earned his PhD from Purdue in 1991, then spent years outside academia — reportedly working odd jobs, including at a Subway sandwich shop — before landing a position as a lecturer (not a professor) at the University of New Hampshire. He was almost 58, had published very little, and was completely unknown to the wider number-theory community. In the quiet, he had been chipping away for years at the bounded-gaps problem.

On 17 April 2013 he submitted a paper titled "Bounded gaps between primes" to the Annals of Mathematics — arguably the most prestigious journal in the field, where papers routinely wait years for a verdict. The referees, experts in exactly these techniques, checked the argument and found it correct. It was accepted in about three weeks — practically unheard of. Overnight, a man most of the profession had never heard of had cracked a problem the giants of the subject had circled for decades. It is one of mathematics' great late-blooming, against-the-odds stories.

Speed here is a sign of clarity, not haste. Zhang built entirely on machinery the referees already trusted — the Goldston–Pintz–Yıldırım sieve and the deep exponential-sum estimates of Fouvry–Iwaniec, Friedlander–Iwaniec, and the Weil/Deligne bounds behind them. His contribution was to assemble these into a new distribution estimate and thread it through GPY. Because each ingredient was familiar and the writing was careful, the experts could verify it quickly. A murky proof of an easy claim can take longer to referee than a clean proof of a hard one.

The mechanism, part 1: what GPY needed

The engine underneath is the Goldston–Pintz–Yıldırım (GPY) method. It fixes a set of shifts \mathcal H = \{h_1,\dots,h_k\} and studies the tuple n+h_1,\ n+h_2,\ \dots,\ n+h_k as n runs over a long range. Using a cleverly weighted sieve it tries to prove that, for infinitely many n, at least two of the k entries are prime. If it succeeds, those two primes differ by at most the diameter h_k-h_1 of the tuple — a fixed bound.

Whether the sieve can detect two primes depends on how much you know about primes in arithmetic progressions — captured by the level of distribution \theta. Loosely, \theta is the exponent up to which primes are known to be evenly spread across residue classes to modulus q, on average over q \le x^{\theta}. The Bombieri–Vinogradov theorem gives, unconditionally, any level \theta < \tfrac12 — "as good as the Generalized Riemann Hypothesis on average," but stuck at the barrier \theta = \tfrac12.

So the entire game reduced to one thing: break the \tfrac12 barrier. GPY themselves had flagged this as the missing ingredient back in 2005 and could not supply it. That is exactly the gap Zhang filled.

The mechanism, part 2: Zhang's extra level of distribution

Zhang proved a Bombieri–Vinogradov-type theorem that reaches past \sqrt{x} — but only for a restricted, well-behaved class of moduli. His level is

\theta = \tfrac12 + \tfrac{1}{584}, \qquad\text{so } q \le x^{\,1/2 + 1/584},

which is just a hair above \tfrac12 — and by the GPY dichotomy, a hair is all you need. The catch, and the genius, is the restriction: the estimate holds only for smooth (friable) moduli — integers q all of whose prime factors are small (below x^{\delta} for a tiny \delta). Smoothness lets you factor q flexibly and reduce the problem to bounding exponential sums — and there the heavy artillery of the subject takes over:

Zhang assembled these into a distribution estimate at level \theta = \tfrac12 + \tfrac{1}{584} valid for smooth moduli — and, crucially, the GPY sieve can be arranged to use only smooth moduli. The restricted estimate is enough. That is the whole trick: not an all-purpose extension of Bombieri–Vinogradov (which would be far beyond current technology), but a bespoke one, tailored so the sieve can drink from it.

Worked example: the chain from "extra distribution" to a bounded gap

Let's walk the logic end to end, the way Zhang's proof does. It is a chain of three links.

Link 1 — extra distribution level. Zhang supplies \theta = \tfrac12 + \tfrac{1}{584} > \tfrac12 for smooth moduli. This is the new arithmetic input; everything downstream is bookkeeping.

Link 2 — GPY converts the level into a value of k. The sieve computation turns a level \theta into the smallest tuple size k it can handle. With Zhang's small excess over \tfrac12, the machine needs a large but finite k: about

k \approx 3.5\times10^{6} \quad(\text{Zhang used } k_0 = 3{,}500{,}000).

The conclusion at this stage: any admissible k_0-tuple contains two primes for infinitely many n.

Link 3 — pick an admissible tuple and read off its diameter. A tuple \mathcal H is admissible if, for every prime p, its members do not cover all residue classes mod p (otherwise p would divide one entry every time and block primality — see the Hardy–Littlewood k-tuple conjectures). You now want the narrowest admissible tuple with k_0 entries — take the first k_0 primes past k_0, or optimise harder — and its diameter H_1 = h_{k_0}-h_1 is your bound. For k_0 = 3.5\times10^{6} a valid admissible tuple fits inside

H_1 \approx 70{,}000{,}000.

Chain complete: extra distribution level \Rightarrow GPY detects 2 primes in an admissible k_0-tuple \Rightarrow bounded gap = diameter of the tuple. The two detected primes sit inside a window of width H_1, so they differ by at most H_1 — infinitely often. That is exactly \liminf(p_{n+1}-p_n)\le 70{,}000{,}000.

A larger \theta means you know primes are well-distributed to larger moduli, which sharpens the sieve's ability to distinguish "genuinely prime" from "has only large factors." The sharper the sieve, the fewer candidate slots (k) you need to be sure two of them land on primes. As \theta \to 1 the required k shrinks dramatically; at the GRH-flavoured extreme you would need far fewer entries and get a far smaller bound. Zhang's \theta is only barely above \tfrac12, so his k is huge and his bound is huge — but finite, which is the whole ballgame.

What happened next: Polymath8 and the great collapse

Zhang's 70 million was never meant to be optimal — he chose parameters for a clean proof, not a small number. Within days the world's number theorists pounced. Terence Tao launched Polymath8, a massively collaborative open online project, to optimise every step: better admissible tuples (link 3), a larger level of distribution (link 1), and a more efficient sieve. The bound fell almost weekly.

Then in late 2013 James Maynard (and independently Tao) introduced a multidimensional GPY sieve that was far more efficient and, remarkably, needed only the classical level \theta < \tfrac12 — no exotic distribution estimate at all — to get bounded gaps. Folded into Polymath8, it drove the bound down to 246, where the best unconditional result still stands. Under stronger (conjectural) distribution assumptions the same machinery reaches a gap of 12 or even 6.

Notice the logarithmic vertical axis: the drop from 7\times10^{7} to 246 is more than five orders of magnitude, most of it inside a single year of frenzied collaboration — a vivid picture of how fast an open, well-posed problem can move once the first crack appears.

The single most common misreading is "Zhang proved twin primes." He did not. His bound is 70{,}000{,}000 — and even after Polymath8 and Maynard it is 246, still a long way from the 2 the Twin Prime Conjecture demands. Bounded gaps say some gap size \le H_1 recurs infinitely often; they do not tell you which gap it is, and in particular say nothing about the specific gap 2.

There is also a hard theoretical wall. The GPY/Maynard sieve method provably cannot, on its own, push the bound below 6 — this is the "parity problem" that has haunted sieve theory since Selberg. Getting from 6 (or 246) down to 2 appears to need a genuinely new idea. So the Twin Prime Conjecture remains open. Zhang's achievement is monumental and completely different: the first ever finite bound, not the conjectured value 2.

Why this is a landmark

For a century, "small gaps between primes" was a place where everyone believed a lot and could prove almost nothing. Zhang moved the problem from "we conjecture bounded gaps exist" to "bounded gaps provably exist" — a qualitative jump, from 0 to 1, that no amount of shrinking 70 million to 246 can match in significance. And he did it by combining existing tools with one new, precisely aimed estimate, showing that the barrier at \theta = \tfrac12 was not absolute after all — at least for the smooth moduli a sieve can be coaxed into using. It is a model of how deep algebraic geometry (Deligne's bounds over finite fields) feeds back into the most classical analytic number theory.