The Prime Number Theorem is, at heart, a statement about a single vertical line. Everything — the fact that \pi(x) \sim x/\ln x, the size of the error, the rhythm of the primes — turns on one deceptively small question: can \zeta(s) vanish on the line \Re(s) = 1?

The answer, proved independently by Hadamard and de la Vallée Poussin in 1896, is no: \zeta(1 + it) \neq 0 for every real t \neq 0. That one non-vanishing is exactly what lets the primes' density settle down. In this lesson we prove it with a startlingly cheap trick — a single trigonometric inequality — and then push a little to the left of the line to carve out a thin zero-free region whose width directly controls how sharp the Prime Number Theorem is.

Why the line \Re(s)=1 is the whole ballgame

The bridge from primes to zeros runs through the logarithmic derivative of zeta. For \Re(s) > 1 the Euler product gives the Dirichlet series

where \Lambda(n) is the von Mangoldt function (\log p on prime powers p^k, else 0). Perron's formula turns this into a contour integral for the prime-counting sum \psi(x) = \sum_{n\le x}\Lambda(n), and the explicit formula expresses \psi(x) as a main term x minus a sum over the zeros \rho of \zeta:

A zero \rho = \beta + i\gamma contributes a term of size x^{\beta}/|\rho|. If some zero had \beta = 1, that term would be of order x itself — the same size as the main term — and the primes would not thin out at the predicted rate. So the PNT, in its rawest form, is precisely the claim that no zero reaches the line \beta = 1.

This is not a lucky sufficient condition that happens to work — it is a genuine equivalence (via the Wiener–Ikehara Tauberian theorem). Getting even a sliver to the left of the line will upgrade "\sim" into a quantitative error bound.

The 3–4–1 trick

Here is the whole engine, and it fits on a postcard. Start from the elementary identity — just expand (1+\cos\theta)^2 and use \cos 2\theta = 2\cos^2\theta - 1:

A perfect square, so it is never negative. Now take real parts of the Dirichlet series above at three heights. Because \Re\big(n^{-(\sigma+it)}\big) = n^{-\sigma}\cos(t\ln n),

Form the combination with weights 3, 4, 1 at the heights 0, t, 2t, and set \theta = t\ln n inside the sum:

Every \Lambda(n) \ge 0, every n^{-\sigma} > 0, and the bracket is \ge 0 term by term — so the whole sum is \ge 0. That single inequality is enough to forbid a zero on the line. The magic is that the weights 3,4,1 are chosen so the positive pull of the pole at s=1 (weight 3) can never quite be overwhelmed by the negative pull a would-be zero at 1+it exerts (weight 4).

Worked example: deriving the contradiction

Suppose, for contradiction, that \zeta(1 + it_0) = 0 for some fixed t_0 \neq 0. We feed three standard estimates (each proved from the pole at s=1 and the Hadamard partial-fraction bound for \zeta'/\zeta) into the 3–4–1 inequality. For 1 < \sigma \le 2 and t = t_0 there is a constant A > 0 with:

Substitute these into 3(\cdot) + 4(\cdot) + (\cdot) \ge 0 and write \delta = \sigma - 1 > 0. With the assumed zero \beta = 1 we have \sigma - \beta = \delta, so:

Rearranged, this says

where C(t_0) is a fixed finite number once t_0 is chosen. But we are free to let \sigma \to 1^{+}, i.e. \delta \to 0^{+}, which sends the left side 1/\delta \to +\infty. A quantity racing to +\infty cannot stay below a fixed constant — contradiction. Therefore no zero sits at 1 + it_0, and since t_0 was arbitrary, \zeta(1+it)\neq 0 everywhere on the line. \blacksquare

From a point to a region: de la Vallée Poussin

The same three bounds, without pretending \beta = 1, prove something stronger. Keep \beta general in the middle bound. The 3–4–1 combination gives

Now play the two sides against each other by choosing \sigma - 1 = \delta proportional to 1/\ln|t|. Optimising the constant, the inequality can only hold if \beta stays a definite distance to the left of 1:

So there is not just a zero-free line but a whole zero-free strip hugging it — pinched thinner and thinner the higher you climb.

Picturing the sliver

Below, the horizontal axis is \Re(s) and the vertical axis is \Im(s). Step through: the critical strip 0 < \Re(s) < 1; the wall \Re(s)=1 where zeta cannot vanish; and finally the curved boundary \sigma = 1 - c/\ln(|t|+2). The zeros of \zeta are banished from the thin region between that curve and the wall — a region that grows narrower as you go up.

The shaded sliver is the prize. It is not a fat region — for large |t| its width c/\ln|t| is minuscule — but it is a non-empty margin, and in analytic number theory a non-empty margin is worth a great deal.

Width controls the error term

Why chase a wider region if we already know the primes thin out? Because the shape of the boundary is precisely what the error term in the PNT reads off. Pushing the contour of the explicit formula just inside a zero-free region of the form \sigma > 1 - \eta(t) and bounding x^{\rho} there produces a bound of the shape \psi(x) - x \ll x\,e^{-\,\text{(gain from the width)}}. Wider region ⇒ larger exponent gain ⇒ smaller error. Concretely:

The classical region buys the famous e^{-c'\sqrt{\ln x}} savings — enough for the PNT with a real error bound, but far weaker than the square-root error the Riemann Hypothesis would give. The Vinogradov–Korobov region, gained by a delicate estimate of exponential sums, is the widest anyone has proved unconditionally, and it is still the current record — no essentially wider region is known. Every improvement to the error term since 1958 has come from pushing the boundary curve, not from any new idea about the primes themselves.