The Riemann zeta function \zeta(s)=\sum_{n\ge1} n^{-s} is a single instrument that hears all the primes at once. But arithmetic often asks a sharper question: how are the primes distributed among the residue classes — how many primes are \equiv 1 \pmod 4 versus \equiv 3 \pmod 4? To listen to one residue class at a time, we need a whole family of zeta-like functions, one tuned to each class. That family is the Dirichlet L-functions.

The trick is to weight each term n^{-s} by a Dirichlet character \chi — a periodic, completely multiplicative "colouring" of the integers mod q. Dirichlet invented them in 1837 for exactly one purpose: to prove that every arithmetic progression a, a+q, a+2q, \dots with \gcd(a,q)=1 contains infinitely many primes. That is Dirichlet's theorem, and L-functions are the machine that proves it.

The definition

Set \chi(n)=1 for all n and you recover \zeta(s) exactly — so \zeta is just the simplest member of the family. Everything you know about the Euler product for \zeta transfers, with 1 replaced by the character value \chi(p) at each prime.

Where the Euler product comes from

The engine is complete multiplicativity: \chi(mn)=\chi(m)\chi(n) for all m,n (not merely coprime ones). This means \chi respects prime factorisation perfectly: if n = p_1^{a_1}\cdots p_k^{a_k} then \chi(n) = \chi(p_1)^{a_1}\cdots\chi(p_k)^{a_k}. Start from the geometric series attached to a single prime p:

Multiply one such series over every prime and expand. Each term picks a prime power from each factor and multiplies them; by unique factorisation the product of the chosen bases is a distinct integer n, and by complete multiplicativity the product of the chosen numerators is exactly \chi(n). Collecting terms rebuilds \sum_n \chi(n)n^{-s}. The character slots into the zeta proof without changing a single step.

Three convergence regimes

Exactly where the series and product make sense depends on which character you took:

The extra convergence for \chi\neq\chi_0 comes from cancellation: over any full period the character values sum to zero (\sum_{n \bmod q}\chi(n)=0, a character orthogonality relation), so the partial sums of \sum \chi(n) stay bounded. Abel/Dirichlet summation then squeezes convergence down to \Re(s)>0. The principal character has no such cancellation — its terms are all +1 or 0 — so it keeps \zeta's pole.

Worked example: the characters mod 4

There are exactly two characters modulo 4, and they are the cleanest illustration of the whole story. Both vanish on even numbers (which share the factor 2 with the modulus):

The non-principal character. Its Euler product runs over the odd primes, with sign + on primes \equiv1 and - on primes \equiv3 \pmod 4:

This is the Dirichlet beta function \beta(s). It is entire, and its value at s=1 is the Leibniz series \beta(1) = 1-\tfrac13+\tfrac15-\tfrac17+\cdots = \tfrac{\pi}{4} \neq 0 — that non-vanishing is the seed of Dirichlet's theorem for the classes 1,3 \bmod 4.

The principal character. Its Euler product runs over the odd primes with all signs +, so it is just \zeta with the p=2 factor deleted:

Since \zeta has a simple pole at s=1 with residue 1, and the factor (1-2^{-s}) equals \tfrac12 there, L(s,\chi_0) has a single simple pole at s=1 of residue \tfrac12. In general the residue of L(s,\chi_0) at s=1 is \prod_{p\mid q}(1-p^{-1}) = \varphi(q)/q.

Seeing it: an entire L-function on the real axis

Here is the non-principal real character mod 4, plotted on the positive real axis as \beta(s) = 1-3^{-s}+5^{-s}-7^{-s}+\cdots. Unlike \zeta (and unlike L(s,\chi_0)), it has no pole — it glides smoothly through s=1, where it equals the Leibniz value \pi/4 \approx 0.785, and tends to 1 as s\to\infty (only the leading 1 survives). That it never dips to zero on the real axis — in particular at s=1 — is precisely the non-vanishing Dirichlet needed.

A smooth, pole-free curve is the visual signature of a non-principal L-function. Swap in the principal character and this graph would rocket to a vertical asymptote at s=1 instead.

Analytic continuation and the functional equation

Just as with \zeta, every L(s,\chi) continues to the whole plane and satisfies a reflection s \mapsto 1-s — but now the reflection swaps \chi for its conjugate \bar\chi, and the symmetry factor carries a Gauss sum \tau(\chi) = \sum_{n \bmod q}\chi(n)\,e^{2\pi i n/q} (which has magnitude |\tau(\chi)| = \sqrt{q} for a primitive character). The precise shape depends on the parity of \chi.

The Gamma factor is \Gamma(s/2) for even characters and \Gamma((s+1)/2) for odd ones — the same shift by the parity a that appears in the completed zeta function (\zeta is the even character a=0). Its poles at the negative integers manufacture the trivial zeros of L(s,\chi): at the negative even integers when \chi is even, at the negative odd integers when it is odd.

The logarithmic derivative: primes, filtered by χ

The reason L-functions count primes is the same as for \zeta: take a logarithm of the Euler product and differentiate. For \Re(s) > 1,

where \Lambda(n) is the von Mangoldt function (\log p when n is a power of the prime p, and 0 otherwise). This is \zeta's prime-counting sum with each term tinted by the character \chi(n). Averaging these tinted sums over all characters mod q — using orthogonality to select a single residue class a — is exactly how one counts the primes \equiv a \pmod q. Poles and zeros of L(s,\chi) become the main and oscillating terms in that count.

Zeros in the critical strip

Because L(s,\chi) \neq 0 for \Re(s)>1 (the Euler product is a product of non-zero factors) and the functional equation reflects that to \Re(s)<0, all the interesting zeros of a non-principal L(s,\chi) lie either at the trivial spots forced by the Gamma factor, or inside the critical strip 0 \le \Re(s) \le 1. The Generalised Riemann Hypothesis predicts that every non-trivial zero sits exactly on the critical line \Re(s) = \tfrac12 — the same conjecture as for \zeta, now for the whole family at once. The single most consequential fact anyone has actually proved, though, is much more modest and lives at the edge of the strip: L(1,\chi) \neq 0 for every non-principal \chi.