The Riemann Hypothesis would tell us that every non-trivial zero of \zeta(s) sits on the critical line \Re(s) = \tfrac12 — none stray to the right. That is a breathtaking demand, and after 165 years nobody can prove it. But here is a liberating thought that drives a whole industry of analytic number theory: for very many applications you do not need every zero on the line — you only need few enough zeros off it.

A zero-density theorem makes that precise. It doesn't locate any single zero; instead it puts a ceiling on how many zeros can possibly lie a given distance to the right of the critical line. If the exceptional zeros — the ones that would violate RH — are guaranteed to be sparse, then the primes still behave on average exactly as RH predicts. This is the art of getting "RH-quality" conclusions out of a strictly weaker, and provable, input.

The counting function N(\sigma, T)

Write a non-trivial zero as \rho = \beta + i\gamma, where \beta = \Re(\rho) is its distance across the strip and \gamma = \Im(\rho) is its height. Fix a threshold \sigma \ge \tfrac12 and a height T, and count the zeros that lie at least as far right as \sigma, up to height T:

With this one function the two competing worldviews become a single equation. The Riemann Hypothesis is the statement

i.e. the box to the right of the line is empty. A zero-density theorem is the humbler, unconditional statement that N(\sigma, T) is merely small when \sigma > \tfrac12 — small enough that the rare off-line zeros can't spoil the average behaviour of the primes.

The density hypothesis

How small is "small enough"? The gold standard short of RH itself is the density hypothesis:

Read the exponent 2(1 - \sigma) as a dial. At the line, \sigma = \tfrac12 gives exponent 1 — so N(\tfrac12, T) \ll T^{1 + \varepsilon}, matching the true count \sim T \log T. As \sigma climbs toward 1 the exponent slides to 0: near the edge of the strip the number of zeros grows barely at all with T. The zeros, in other words, are conjectured to crowd toward the critical line, thinning out dramatically the further right you look.

The density hypothesis is weaker than RH — RH says the box is empty, the density hypothesis merely says it's sparse — yet it is strong enough to reproduce many of RH's arithmetic consequences on average. It is also still open. What we can prove unconditionally is a family of classical bounds that approach it.

The classical unconditional bounds

The first genuinely useful estimate is Ingham's (1940), built on bounds for the fourth moment of \zeta on the critical line:

Compare the two exponents. Both agree at the endpoints — they equal 1 at \sigma = \tfrac12 and 0 at \sigma = 1 — but in between Ingham's \tfrac{3(1-\sigma)}{2-\sigma} sits above the density-hypothesis target 2(1-\sigma), so it permits more zeros. The whole history of the subject is a decades-long campaign to press the provable curve down onto (and past, in ranges) the conjectured one — Huxley, Montgomery, Jutila, Heath-Brown, Bourgain, and most recently the Guth–Maynard breakthrough (2024) that finally proved the density hypothesis's key case near \sigma = \tfrac34.

The chart plots the exponent e(\sigma) in N(\sigma,T) \ll T^{\,e(\sigma)} against the threshold \sigma. Higher curve = more zeros allowed that far right. Everything to the right of \sigma = \tfrac12 and below a curve is the permitted zero density for that bound; RH would flatten the whole picture to the axis (e \equiv 0 for \sigma > \tfrac12). Both known curves squeeze to zero as \sigma \to 1: near the right edge, off-line zeros are provably almost absent.

Worked example — how few off-line zeros tame \psi(x)

The reason N(\sigma, T) matters is the explicit formula, which writes the Chebyshev prime-counting function \psi(x) = \sum_{n \le x}\Lambda(n) as a main term minus a sum over the zeros:

Each zero \rho = \beta + i\gamma contributes a wave of size \left|\dfrac{x^{\rho}}{\rho}\right| = \dfrac{x^{\beta}}{|\rho|}. The real part \beta is what hurts: a zero on the line (\beta = \tfrac12) contributes only x^{1/2}/|\rho|, the smallest possible; a zero pushed right to \beta = 1 - \delta contributes x^{1-\delta}/|\rho| — dangerously close to the main term x. So the error in \psi(x) is dominated by how many zeros lie how far right.

Step 1 — group the zeros by their real part. Sum the contributions up to height T as a Riemann–Stieltjes integral against the counting function:

Step 2 — feed in a density bound. Suppose the density hypothesis N(\sigma, T) \ll T^{2(1-\sigma)} holds. Then the mass at threshold \sigma is controlled by x^{\sigma} T^{2(1-\sigma)}. Rewrite that product to see where it peaks:

Step 3 — read the trade-off. If T^{2} \le x (that is, T \le \sqrt{x}), the bracket is \le 1 and its (1-\sigma) power is largest at \sigma = \tfrac12 — the far-right zeros contribute less, not more. The many possible off-line zeros are held so sparse that their combined effect never overtakes the honest x^{1/2}-sized error you'd get from RH. That is the whole trick: the density bound does the job of RH on average, because a zero at \beta close to 1 is individually loud but provably rare enough that the chorus stays quiet.

The payoff: RH-quality results, unconditionally

Turning that estimate into arithmetic gives two of the most useful theorems in the subject — both proved with no assumption of RH.

The unifying slogan is worth memorising: you don't need every zero on the line, just few enough off it. Bombieri–Vinogradov is precisely why sieve methods (Chen's theorem on p + 2 = \text{prime or semiprime}, and the machinery behind bounded gaps between primes) can proceed without waiting for RH.