The Prime Number Theorem via Complex Analysis

⚡ Key Takeaways

The Prime Number Theorem (PNT) states that the number of primes up to $x$ , denoted $\pi(x)$ , is asymptotic to $x/\log x$ . Equivalently, $\psi(x) \sim x$ , where $\psi$ is the Chebyshev function. The proof via complex analysis hinges on the non-vanishing of the Riemann zeta function $\zeta(s)$ on the line $\operatorname{Re}(s)=1$ and a Tauberian argument. First proved independently by Hadamard and de la Vallée Poussin in 1896, following Riemann's 1859 program. Applies throughout analytic number theory, cryptography heuristics, and the distribution of arithmetic objects.

💡 The Core Insight

Encode the primes in the Euler product for $\zeta(s)$ . Taking a logarithmic derivative converts the product into a Dirichlet series whose coefficients are the von Mangoldt function $\Lambda(n)$ , summing to the Chebyshev $\psi(x)$ . A contour-integral (Perron/Mellin) representation of $\psi(x)$ then expresses it as a sum over the poles of $-\zeta'(s)/\zeta(s)$ . The pole at $s=1$ contributes the main term $x$ ; the absence of zeros on $\operatorname{Re}(s)=1$ guarantees all other contributions are $o(x)$ . Thus $\psi(x)\sim x \iff \zeta(1+it)\neq 0$ .

DefinitionPrime-counting and Chebyshev functions

For real $x \geq 2$ define:

1

\quad \pi(x) = \sum_{p \le x} 1, \qquad \vartheta(x) = \sum_{p\le x} \log p, \qquad \psi(x) = \sum_{p^k \le x} \log p = \sum_{n\le x}\Lambda(n)

Derivation hint: $\pi$ counts primes; $\vartheta$ weights each prime by $\log p$ ; $\psi$ also counts prime powers, since $\Lambda(n)=\log p$ if $n=p^k$ and $0$ otherwise. Here $\Lambda$ is the von Mangoldt function. These three functions are asymptotically equivalent in a precise sense and PNT can be stated through any of them. The logarithmic integral

2

\quad \operatorname{li}(x) = \int_0^x \frac{dt}{\log t} \;(\text{principal value}), \qquad \operatorname{li}(x) \sim \frac{x}{\log x}

Derivation hint: integrate by parts repeatedly to get $\operatorname{li}(x)=\frac{x}{\log x}\sum_{k=0}^{N}\frac{k!}{(\log x)^k}+O\!\big(x/(\log x)^{N+2}\big)$ ; the leading term is $x/\log x$ . is the sharpest elementary main term.

TheoremPrime Number Theorem

As $x\to\infty$ ,

3

\quad \pi(x) \sim \frac{x}{\log x}, \qquad \text{equivalently}\qquad \pi(x)\sim \operatorname{li}(x), \qquad \vartheta(x)\sim x, \qquad \psi(x)\sim x

Derivation hint: the equivalences follow from partial summation (Abel summation) linking $\pi$ to $\vartheta$ and from $\psi(x)-\vartheta(x)=\sum_{k\ge2}\vartheta(x^{1/k})=O(\sqrt{x}\log x)$ . The notation $f\sim g$ means $f(x)/g(x)\to 1$ .

DefinitionThe Riemann zeta function and its Euler product

For $\operatorname{Re}(s)=\sigma>1$ ,

4

\quad \zeta(s) = \sum_{n=1}^{\infty} \frac{1}{n^s} = \prod_{p\text{ prime}} \left(1-p^{-s}\right)^{-1}

Derivation hint: expand each factor as a geometric series $\sum_{k\ge0}p^{-ks}$ and multiply; unique factorization of integers gives every $n^{-s}$ exactly once. $\zeta$ extends to a meromorphic function on $\mathbb{C}$ with a single simple pole at $s=1$ with residue $1$ . The logarithmic derivative generates $\Lambda$ :

5

\quad -\frac{\zeta'(s)}{\zeta(s)} = \sum_{n=1}^{\infty} \frac{\Lambda(n)}{n^s}, \qquad \sigma>1

Derivation hint: $\log\zeta(s)=-\sum_p\log(1-p^{-s})=\sum_p\sum_{k\ge1}\frac{p^{-ks}}{k}$ ; differentiate term-by-term to get $\sum_p\sum_k (\log p)\,p^{-ks}=\sum_n\Lambda(n)n^{-s}$ .

ProofPNT via the Newman–Tauberian route

We prove $\psi(x)\sim x$ . We follow D. J. Newman's analytic simplification (1980), the cleanest modern complex-analytic proof.

Step 1 — Reduce to a Dirichlet series. From [5], the generating identity holds for $\sigma>1$ . Define

6

\quad \Phi(s) = \sum_{p} \frac{\log p}{p^s} = -\frac{\zeta'(s)}{\zeta(s)} - \sum_p\sum_{k\ge2}\frac{\log p}{p^{ks}}, \qquad \sigma>1

Derivation hint: split [5] into $k=1$ (the $\Phi$ sum) and $k\ge2$ ; the latter converges for $\sigma>\tfrac12$ , hence is holomorphic there. So $\Phi(s)$ and $-\zeta'/\zeta$ differ by a function holomorphic on $\sigma>\tfrac12$ .

Step 2 — Pole and non-vanishing. Since $\zeta$ has a simple pole at $s=1$ , $-\zeta'/\zeta$ has a simple pole at $s=1$ with residue $+1$ . The crucial analytic input is:

7

\quad \zeta(1+it) \neq 0 \quad\text{for all } t\in\mathbb{R},\qquad t\neq 0

Derivation hint: use $3+4\cos\theta+\cos2\theta = 2(1+\cos\theta)^2 \ge 0$ applied to $\log|\zeta^3(\sigma)\zeta^4(\sigma+it)\zeta(\sigma+2it)|$ ; a zero at $1+it$ would force this expression $\to-\infty$ , a contradiction. Consequently $\Phi(s)-\dfrac{1}{s-1}$ extends holomorphically to an open neighborhood of the closed half-plane $\operatorname{Re}(s)\ge1$ .

Step 3 — Integral representation. Using $\psi(x)=\sum_{n\le x}\Lambda(n)$ and partial summation, the key object is

8

\quad \frac{\Phi(s)}{s} = s\int_1^\infty \frac{\vartheta(x)}{x^{s+1}}\,dx \;\bigg/\; s = \int_1^\infty \frac{\vartheta(x)}{x^{s+1}}\,dx, \qquad \sigma>1

Derivation hint: $\Phi(s)=\sum_p (\log p)p^{-s}=s\int_1^\infty \vartheta(x)x^{-s-1}dx$ by Abel summation, since $\vartheta$ is the partial-sum (step) function of the coefficients $\log p$ .

Step 4 — The Tauberian theorem (Newman). Substitute $x=e^t$ , $\vartheta(e^t)=e^t f(t)+\ldots$ ; the relevant convergence theorem is: > Analytic Theorem. Let $g(t)$ ( $t\ge0$ ) be bounded and locally integrable, with Laplace transform $G(z)=\int_0^\infty g(t)e^{-zt}dt$ holomorphic for $\operatorname{Re}(z)>0$ . If $G$ extends holomorphically to $\operatorname{Re}(z)\ge0$ , then $\int_0^\infty g(t)\,dt$ converges and equals $G(0)$ . Apply with $g(t)=\vartheta(e^t)e^{-t}-1$ . One checks $G(z)=\dfrac{\Phi(z+1)}{z+1}-\dfrac1z$ , which by Step 2 is holomorphic on $\operatorname{Re}(z)\ge0$ . Boundedness of $g$ requires Chebyshev's bound $\vartheta(x)=O(x)$ .

9

\quad \int_1^\infty \frac{\vartheta(x)-x}{x^2}\,dx \quad\text{converges}

Derivation hint: this is the change of variables $t=\log x$ applied to $\int_0^\infty g(t)\,dt = G(0)$ from the analytic theorem.

Step 5 — From convergence to asymptotics. Suppose $\vartheta(x)\ge \lambda x$ for some $\lambda>1$ on a sequence $x\to\infty$ . Since $\vartheta$ is nondecreasing,

10

\quad \int_x^{\lambda x}\frac{\vartheta(t)-t}{t^2}\,dt \ge \int_x^{\lambda x}\frac{\lambda x - t}{t^2}\,dt = \int_1^{\lambda}\frac{\lambda-u}{u^2}\,du > 0

Derivation hint: bound $\vartheta(t)\ge\vartheta(x)\ge\lambda x$ on $[x,\lambda x]$ , substitute $t=ux$ ; the resulting positive constant is independent of $x$ , contradicting Cauchy convergence of [9]. The same with $\lambda<1$ (using $\vartheta(t)\le\vartheta(\lambda x)\le\lambda x$ , wait—) gives the lower contradiction; both force $\vartheta(x)/x\to 1$ . Hence $\vartheta(x)\sim x$ , and by [3] $\psi(x)\sim x$ and $\pi(x)\sim\operatorname{li}(x)$ . $\blacksquare$

∎

⚠️ The Chebyshev bound is not optional

Step 4 silently uses $\vartheta(x)=O(x)$ to guarantee $g$ is bounded. This is Chebyshev's elementary estimate (1852), proved via the central binomial coefficient $\binom{2n}{n}$ . Without it the Laplace transform may not even define a holomorphic function in the required half-plane. PNT is not purely complex-analytic in a vacuum — it rests on a prior elementary upper bound.

BR

Bernhard Riemann

1826–1866

Riemann's eight-page 1859 memoir "Über die Anzahl der Primzahlen unter einer gegebenen Größe" introduced the modern viewpoint: study $\pi(x)$ via the analytic continuation of $\zeta(s)$ to all of $\mathbb{C}$ , its functional equation relating $s\leftrightarrow1-s$ , and its nontrivial zeros. He wrote down an exact "explicit formula" for $\pi(x)$ as a sum over zeros and conjectured the Riemann Hypothesis (all nontrivial zeros on $\operatorname{Re}(s)=\tfrac12$ ). He did not prove PNT — several of his assertions were established only decades later by von Mangoldt, Hadamard, and de la Vallée Poussin. The complex-analytic proof of PNT is the realization, in its minimal form, of Riemann's program.

🧠 Intuition

Think of $\psi(x)=\sum_{n\le x}\Lambda(n)$ as a signal. The Mellin/Perron formula

11

\quad \psi(x) = x - \sum_{\rho}\frac{x^\rho}{\rho} - \log(2\pi) - \tfrac12\log(1-x^{-2})

Derivation hint: shift the contour in $\psi(x)=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\big(-\frac{\zeta'}{\zeta}(s)\big)\frac{x^s}{s}ds$ across all poles; $s=1$ gives $x$ , each nontrivial zero $\rho$ gives $-x^\rho/\rho$ , the trivial zeros give the log terms. decomposes $\psi$ into a main term $x$ plus oscillations indexed by the zeros $\rho=\beta+i\gamma$ . Each zero contributes a wave $x^\beta$ in amplitude and $\cos(\gamma\log x)$ in phase. PNT, $\psi(x)\sim x$ , holds precisely when every such wave has amplitude $o(x)$ , i.e. $\beta<1$ for all $\rho$ . The Riemann Hypothesis would put every $\beta=\tfrac12$ , making the error $O(\sqrt{x}\,\log^2 x)$ — the smallest conceivable. So: primes are as regularly distributed as the zeros of $\zeta$ are far left of the line $\operatorname{Re}(s)=1$ .

Figure

Critical strip, the line Re(s)=1, and the contour of the Perron integral

```svg

```

ExampleNumerical check of the main term and error

Given: $x = 10^9$ . Compare $\pi(x)$ with $x/\log x$ and $\operatorname{li}(x)$ .

Steps:

$\log(10^9) = 9\log 10 \approx 20.7233$ .
$x/\log x \approx 10^9 / 20.7233 \approx 4.8255\times10^7$ .
$\operatorname{li}(10^9)\approx 5.0849234\times10^7$ (numerically).
True value: $\pi(10^9) = 50\,847\,534$ .

Comparison: Relative error of $x/\log x$ : $(50847534 - 48254942)/50847534 \approx 5.1\%$ — slowly improving as predicted. Relative error of $\operatorname{li}(x)$ : $(50847534 - 50849234)/50847534 \approx -3.3\times10^{-5}$ .

\boxed{\pi(10^9)=50\,847\,534,\quad \operatorname{li}(10^9)\approx 50\,849\,234,\quad \text{error }\approx 1700 \approx 3.4\sqrt{x}/\log x}

The vastly superior accuracy of $\operatorname{li}$ illustrates why it, not $x/\log x$ , is the genuine main term.

ExampleFrom

\psi

to

\pi

via partial summation

Given: Assume $\psi(x) = x + O(x e^{-c\sqrt{\log x}})$ (de la Vallée Poussin error). Derive the corresponding error for $\pi(x)$ .

Steps:

First $\vartheta(x)=\psi(x)-O(\sqrt x\log x)=x+O(xe^{-c\sqrt{\log x}})$ .
Abel summation: $\pi(x)=\dfrac{\vartheta(x)}{\log x}+\displaystyle\int_2^x \dfrac{\vartheta(t)}{t\log^2 t}\,dt$ .
Substitute $\vartheta(t)=t+E(t)$ , $E(t)=O(te^{-c\sqrt{\log t}})$ :

\pi(x)=\frac{x}{\log x}+\int_2^x\frac{dt}{\log^2 t}+\frac{E(x)}{\log x}+\int_2^x\frac{E(t)}{t\log^2 t}dt.

The first two terms recombine into $\operatorname{li}(x)$ ; the error integrals are dominated by $O(xe^{-c'\sqrt{\log x}})$ .

\boxed{\pi(x)=\operatorname{li}(x)+O\!\big(x\,e^{-c\sqrt{\log x}}\big)}

This is the classical PNT with error term, equivalent to a zero-free region $\sigma > 1 - c/\log(|t|+2)$ .

⚠️ PNT does not require the Riemann Hypothesis

A frequent error is to believe RH is needed for PNT. It is not. PNT follows from the strictly weaker fact $\zeta(1+it)\neq0$ . RH only sharpens the error term to $O(\sqrt{x}\log^2 x)$ . Conversely, knowing PNT gives essentially no zero-free region beyond the line $\sigma=1$ itself.

⚠️

x/\log x

vs.

\operatorname{li}(x)

— both are correct, but unequal in quality

Both $x/\log x$ and $\operatorname{li}(x)$ are valid statements of " $\pi(x)\sim\,\cdot$ ", yet $\pi(x)-x/\log x \sim x/\log^2 x \to \infty$ in absolute terms while staying $o(x/\log x)$ . Asymptotic equivalence ( $\sim$ ) is about ratios, not differences. $\operatorname{li}(x)$ captures the next-order corrections and is dramatically more accurate numerically.

Counterexample

\pi(x)<\operatorname{li}(x)

is NOT always true — Skewes' number

All computed data suggest $\pi(x)<\operatorname{li}(x)$ . Yet Littlewood (1914) proved $\pi(x)-\operatorname{li}(x)$ changes sign infinitely often. The first sign change (Skewes' number) is astronomically large — below $1.4\times10^{316}$ . This warns against extrapolating numerical "patterns": the oscillation terms $\sum_\rho x^\rho/\rho$ eventually overwhelm any finite empirical trend.

connection

PNT sits at the hub of analytic number theory. Its proof is the prototype for the Wiener–Ikehara Tauberian Theorem, which abstracts Step 4. The non-vanishing $\zeta(1+it)\neq0$ generalizes to Dirichlet L-functions, yielding the Prime Number Theorem in Arithmetic Progressions (Dirichlet's theorem with density $1/\varphi(q)$ ). The error-term refinements connect directly to the Riemann Hypothesis and zero-free regions. The same machinery — Euler product, logarithmic derivative, Tauberian inversion — underlies the Chebotarev Density Theorem and the Sato–Tate Conjecture. Finally, the Selberg–Erdős Elementary Proof (1948) shows PNT can be obtained without complex analysis, a celebrated alternative route.

🔬 Where this appears

⏳ History

- 1737 — Euler. Proves

\sum_p 1/p

diverges and establishes the Euler product for

\zeta(s)

, the first analytic link between primes and

\zeta

.

- 1798 / 1808 — Legendre. Conjectures

\pi(x)\approx x/(\log x - B)

with

B\approx1.08366

from numerical data.

- 1792–1849 — Gauss. Conjectures (in private notebooks, age 15)

\pi(x)\sim\operatorname{li}(x)

, the form that proved most accurate.

- 1852 — Chebyshev. Proves

0.921\,x/\log x < \pi(x) < 1.106\,x/\log x

for large

x

and that if

\pi(x)/(x/\log x)

has a limit, it is

1

. Introduces

\vartheta,\psi

.

- 1859 — Riemann. Memoir on

\pi(x)

: analytic continuation, functional equation, explicit formula, and the Riemann Hypothesis. Sets the program but proves no asymptotic.

- 1895 — von Mangoldt. Proves Riemann's explicit formula for

\psi(x)

rigorously, reducing PNT to the non-vanishing of

\zeta

on

\operatorname{Re}(s)=1

.

- 1896 — Hadamard & de la Vallée Poussin (independently). Prove

\zeta(1+it)\neq0

and deduce PNT. De la Vallée Poussin adds the zero-free region and error

O(xe^{-c\sqrt{\log x}})

.

- 1948 — Selberg & Erdős. "Elementary" (no complex analysis) proof of PNT via Selberg's symmetry formula, sparking a famous priority dispute.

- 1980 — D. J. Newman. Drastically simplifies the analytic proof using a short contour-integral Tauberian theorem — now the standard graduate presentation.

open-problem

The qualitative PNT is settled; the frontier is the size of the error term, governed entirely by the location of zeta zeros.

Riemann Hypothesis. $\pi(x)=\operatorname{li}(x)+O(\sqrt{x}\log x)$ is equivalent to RH. Unproven; one of the Clay Millennium Problems.
Best unconditional zero-free region. The Vinogradov–Korobov region $\sigma>1-c/(\log|t|)^{2/3}(\log\log|t|)^{1/3}$ gives the current best error $O\big(x\exp(-c(\log x)^{3/5}(\log\log x)^{-1/5})\big)$ — still far from RH.
Sign changes (Littlewood/Skewes). Pinning the first crossing of $\pi(x)=\operatorname{li}(x)$ and understanding the oscillation $\pi(x)-\operatorname{li}(x)=\Omega_\pm(\sqrt{x}\log\log\log x/\log x)$ remains delicate.
Prime gaps. PNT gives average gap $\log n$ but not maximal/minimal gaps; the twin prime and bounded-gap problems (Zhang, Maynard) lie beyond it.
Effective constants. Making $c$ explicit and uniform in companion theorems (e.g. for $L$ -functions, Siegel zeros) is an active obstacle, e.g. the ineffectivity of Siegel's bound.

The Prime Number Theorem via Complex Analysis

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also