Dirichlet Series

The zeta function looks like a one-off — a special sum someone happened to notice. It isn't. It is just the simplest member of a huge family: pick any sequence of numbers a(1), a(2), a(3), \ldots that answers some question about integers, and you can build an analytic function out of it the same way. Different choices of a(n) encode wildly different number-theoretic information — and once the information is wearing the clothes of a function, the whole machinery of calculus and complex analysis becomes available to attack it.

The general form

Given an arithmetic function a(n), its Dirichlet series is

D(s) = \sum_{n=1}^{\infty} \frac{a(n)}{n^{s}}.

Every choice of a(n) is a different lens on the integers, turned into a function of a single variable s.

Worked example 1: the plainest choice

Take the simplest possible sequence — every term equal to 1, so a(n) = 1 for every n. Its Dirichlet series is

D(s) = \sum_{n=1}^{\infty}\frac{1}{n^{s}} = 1 + \frac{1}{2^s} + \frac{1}{3^s} + \frac{1}{4^s} + \cdots = \zeta(s).

The Riemann zeta function isn't a special new object at all — it's just the Dirichlet series of the constant sequence. That single observation is what makes zeta the natural gateway into this whole family.

Worked example 2: the reciprocal twin

Now try a much stranger-looking sequence: the Möbius function \mu(n), which is +1, -1, or 0 depending on the prime factors of n. Its Dirichlet series turns out to be the exact reciprocal of zeta:

\sum_{n=1}^{\infty} \frac{\mu(n)}{n^{s}} = \frac{1}{\zeta(s)}.

That is a remarkable thing for two totally different-looking sums to satisfy. Let's watch it happen numerically at s=2, where \zeta(2) = \pi^2/6 \approx 1.6449, so 1/\zeta(2) \approx 0.6079. Multiplying a truncated \zeta(2) by a truncated Möbius series should creep toward 1:

function mobius(n: number): number { if (n === 1) return 1; let x = n, primeFactors = 0; for (let p = 2; p * p <= x; p++) { if (x % p === 0) { x /= p; if (x % p === 0) return 0; // squared prime factor primeFactors++; } } if (x > 1) primeFactors++; return primeFactors % 2 === 0 ? 1 : -1; } const s = 2; let zetaPart = 0; let mobiusPart = 0; for (let n = 1; n <= 10; n++) { zetaPart += 1 / Math.pow(n, s); mobiusPart += mobius(n) / Math.pow(n, s); } console.log("Partial zeta(2), n up to 10: ", zetaPart.toFixed(4)); console.log("Partial mobius series, n<=10: ", mobiusPart.toFixed(4)); console.log("Their product (should -> 1): ", (zetaPart * mobiusPart).toFixed(4));

With only ten terms each side, the product already lands near 0.955 — heading straight for the exact answer, 1. This is the analytic shadow of Möbius inversion: \mu is built precisely to "undo" the constant sequence, and multiplying their Dirichlet series makes that cancellation visible as \zeta(s)\cdot(1/\zeta(s)) = 1.

Here is the same race plotted as the cutoff grows: the partial zeta sum climbs toward \pi^2/6, the partial Möbius sum sinks toward 1/\zeta(2), and their product — the one we actually care about — flattens out at 1 almost immediately:

Multiplication mirrors convolution

The product of two Dirichlet series re-sums by divisors. If

A(s) = \sum \frac{a(n)}{n^s}, \qquad B(s) = \sum \frac{b(n)}{n^s},

then

A(s)\,B(s) = \sum_{n=1}^{\infty} \frac{(a * b)(n)}{n^{s}}, \qquad (a*b)(n) = \sum_{d\mid n} a(d)\,b(n/d).

This combination, a * b, is called Dirichlet convolution — it's the arithmetic operation lurking underneath ordinary multiplication of series. The sequence that behaves like "1" for this multiplication is \varepsilon(n), equal to 1 when n=1 and 0 otherwise — and indeed \mu * 1 = \varepsilon is exactly the identity at the heart of Möbius inversion, which is precisely why worked example 2 above landed on 1.

A function being multiplicative is exactly what gives its Dirichlet series an Euler product of its own, broken into one independent factor per prime. Analysis and arithmetic translate cleanly into one another.

A tiny concrete check of the convolution formula: let a(n) = 1 for every n (so A(s) = \zeta(s)) and also b(n) = 1 for every n (so B(s) = \zeta(s) too). At n = 4, the divisors are 1, 2, 4, so

(a*b)(4) = \sum_{d \mid 4} a(d)\,b(4/d) = a(1)b(4) + a(2)b(2) + a(4)b(1) = 1+1+1 = 3.

Three is exactly the number of divisors of 4 — which makes sense, since convolving the all-ones sequence with itself just counts pairs (d, n/d), one for every divisor. So \zeta(s)^{2} is the Dirichlet series of the divisor-counting function d(n) — another small arithmetic fact, caught instantly by squaring a series. The same trick generalises: raising \zeta(s) to any whole power k produces the Dirichlet series that counts, for each n, the number of ways to write n as an ordered product of k positive integers.

A Dirichlet series does not converge for every value of s just because you can write the sum down. Each series has its own region of convergence — typically a half-plane \Re(s) > \sigma for some threshold \sigma that depends on how fast a(n) grows.

\zeta(s) = \sum 1/n^s needs \Re(s) > 1. The Möbius series, whose terms are always at most 1 in size, still needs the same threshold for the sum to converge absolutely as written above. But the threshold genuinely depends on the coefficients: take a(n) = n instead, and \sum n/n^{s} = \sum 1/n^{s-1} only converges once \Re(s) - 1 > 1, i.e. \Re(s) > 2 — a whole unit further right than \zeta itself.

Treating a Dirichlet series formula as valid "everywhere" — plugging in whatever s you like without checking convergence first — is a very common overreach, and it's exactly the kind of shortcut that made the early history of this subject so treacherous.

Dirichlet built the very first non-trivial example of this idea to attack a problem Euler's methods couldn't touch: are there infinitely many primes among numbers of the form a, a+d, a+2d, a+3d, \ldots (an arithmetic progression)? His L-functions attach a Dirichlet series to each "character" of a residue class, and unlock exactly this question — see Dirichlet's theorem on arithmetic progressions.

For a taste: to tell primes of the form 4k+1 (like 5, 13, 17) apart from primes of the form 4k+3 (like 3, 7, 11), Dirichlet used a coefficient sequence \chi(n) that is +1 on numbers \equiv 1 \pmod 4, -1 on numbers \equiv 3 \pmod 4, and 0 on even numbers. The resulting Dirichlet series L(s,\chi) = \sum \chi(n)/n^{s} behaves well enough at s=1 to prove both families of primes are infinite — a question the plain zeta function has no way of even asking.

That was just the opening move. L-functions today are attached to elliptic curves, modular forms, and Galois representations, and they sit at the centre of some of the biggest open problems in mathematics — including the Birch and Swinnerton-Dyer conjecture, one of the seven Millennium Prize Problems. A humble reciprocal power series turned out to be a doorway into the frontier of modern research.

Why build them

Dirichlet series convert questions about sequences into questions about functions — where poles, zeros and growth rates can be brought to bear. Dirichlet's own creation, the L-functions built from characters, attach a series to each residue class and unlock primes in arithmetic progressions. They are the prototype of the vast theory of L-functions at the frontier of modern number theory.