The Generalised Riemann Hypothesis

The Riemann Hypothesis is a claim about a single function, \zeta(s), and it controls a single headline count: how many primes there are up to x. But primes come in flavours. Split them by their remainder on division by q — the primes that are 1 \bmod 10 (11, 31, 41, 61, \dots), the ones that are 3 \bmod 10 (3, 13, 23, 43, \dots), and so on — and you get a whole family of prime-counting problems, one for each residue class.

Each of those finer problems is governed not by \zeta but by a Dirichlet L-function L(s,\chi). The Generalised Riemann Hypothesis (GRH) is the breathtaking claim that every last one of these functions obeys the same law that RH imposes on \zeta: all their non-trivial zeros line up on \Re(s) = \tfrac12. It is not one hypothesis but infinitely many, bundled into a single sentence — and its payoff is control over the primes not just in bulk, but in every arithmetic progression at once.

From one L-function to infinitely many

A Dirichlet character \chi modulo q is a way of colouring the residue classes so the colours multiply nicely (\chi(mn) = \chi(m)\chi(n)). To each one Dirichlet attached a series that, like \zeta, has an Euler product over the primes:

L(s,\chi) = \sum_{n=1}^{\infty} \frac{\chi(n)}{n^{s}} = \prod_{p} \left(1 - \frac{\chi(p)}{p^{s}}\right)^{-1}.

The trivial (principal) character \chi_0 just gives back \zeta(s) up to a few Euler factors — so \zeta is the simplest member of the family, and RH is the special case of GRH for that one member. Every other character \chi carries genuinely new arithmetic: its zeros encode how the primes distribute across the residue classes mod q. There is one L(s,\chi) for each of the infinitely many characters, of each of the infinitely many moduli.

The hypothesis

For every Dirichlet character \chi, all non-trivial zeros of L(s,\chi) lie on the critical line \Re(s) = \tfrac12.
The Extended Riemann Hypothesis says the same for the \zeta_K(s) — the Dedekind zeta function of any number field K.
The Grand Riemann Hypothesis pushes it to all automorphic L-functions: every "principal" L-function in the Langlands sense has its non-trivial zeros on \Re(s) = \tfrac12.

RH is the bottom rung: it is exactly GRH restricted to the single function \zeta = L(s,\chi_0). So GRH is strictly stronger — \text{GRH} \Rightarrow \text{RH}, but not the other way round. Proving RH alone would leave the entire infinite family of Dirichlet L-functions untouched.

Picturing "all of them at once"

RH pins the zeros of one function to one line. GRH stacks that same picture up, once per L-function, and insists every stack behaves identically. Below, each panel is the critical strip of a different L(s,\chi). Step through: the shared strip, the shared critical line, and then the zeros — every one landing on \Re(s) = \tfrac12, in every panel.

The heights of the zeros differ from one \chi to the next — that is where the distinct arithmetic lives. What GRH asserts is that the horizontal coordinate is always the same: dead-centre, for the whole family, forever.

Why GRH is much more than RH

RH controls the overall prime count \pi(x). GRH controls the count in each arithmetic progression,

\pi(x;q,a) = \#\{\, p \le x : p \equiv a \pmod q \,\},

where \gcd(a,q)=1. The primes ought to split evenly among the \varphi(q) admissible residue classes, so the smooth prediction is \tfrac{1}{\varphi(q)}\operatorname{Li}(x). The whole question is how big the error can be — and, crucially, how that error behaves as q grows with x. GRH gives the sharpest possible answer, uniformly in q:

Uniform prime counting in progressions. Assuming GRH, \pi(x;q,a) = \dfrac{1}{\varphi(q)}\operatorname{Li}(x) + O\!\big(\sqrt{x}\,\log x\big), with the constant uniform in q — no Siegel–Walfisz restriction, and no exceptional Siegel zeros.
Least prime in a progression (explicit Linnik). The smallest prime p \equiv a \pmod q satisfies p \ll \varphi(q)^{2}(\log q)^{2} — an explicit power of q, in place of Linnik's ineffective q^{L}.
Miller's primality test is polynomial-time. Assuming GRH, a composite n always has a witness below 2(\log n)^2, so Miller's test is deterministic and runs in polynomial time.
Few Miller–Rabin witnesses suffice. Testing the small prime bases up to 2(\log n)^2 is enough to prove primality outright.
Class numbers and Euclidean rings. GRH yields effective bounds relating class numbers to L(1,\chi), and settles which rings of integers are norm-Euclidean in borderline cases.

Worked example — the uniform range GRH buys you

The reason mathematicians ache for GRH is a single change in how far a theorem reaches. Unconditionally, the best general tool is the Siegel–Walfisz theorem: the primes equidistribute in progressions, but only when the modulus is tiny relative to x. It restricts you to

q \le (\log x)^{A} \qquad (\text{Siegel–Walfisz, any fixed } A).

GRH, by contrast, keeps the same equidistribution valid all the way out to

q \le \sqrt{x}\,/\,(\log x)^{2} \qquad (\text{GRH, essentially } q \le \sqrt{x}).

Put a number on it. Take x = 10^{12}, so \log x \approx 27.6. With A = 2, Siegel–Walfisz only controls the primes for moduli up to q \le (\log x)^2 \approx 763 — after that it says nothing. GRH controls them up to q \le \sqrt{10^{12}} = 10^{6} — a million. That is not a small improvement in a constant; it is the difference between "hundreds" and "a million" residue classes over which you can trust the primes to spread evenly. Whole swathes of analytic number theory that must currently sum over q up to \sqrt{x} become unconditional-looking the moment you are allowed to assume GRH.

Why \sqrt{x}? Because a zero at \Re(s) = \tfrac12 + \delta would contribute an error of size x^{1/2+\delta}; GRH's \delta = 0 pins that to x^{1/2}, the smallest a genuinely oscillating error term can be. The square-root barrier is exactly the critical line, translated from the world of zeros into the world of primes.

Every result in this lesson is conditional. "Assuming GRH, Miller's test is polynomial-time" is a genuine, published theorem — but the "assuming GRH" cannot be dropped, because nobody has proved GRH (nor RH, its baby brother). A paper that says its algorithm is fast "under GRH" is telling you the fast-running is conjectural, however overwhelming the numerical evidence.

And do not confuse the two hypotheses. \text{RH} is only the \chi_0 case — the single function \zeta. \text{GRH} demands the same of the entire infinite family of Dirichlet L-functions, so \text{GRH} \Rightarrow \text{RH} but not conversely. Proving RH tomorrow would win the million-dollar prize and still leave the uniform prime-in-progression estimate — and Miller's deterministic test — completely open. The extra strength lives entirely in the \chi \ne \chi_0 members, and in ruling out the dreaded Siegel zero: a hypothetical real zero of an L(s,\chi) sitting maddeningly close to s = 1, which GRH forbids outright but which no unconditional proof can yet exclude.

The primality-testing payoff, in one line

The cleanest place GRH earns its keep in computer science is Gary Miller's primality test. A composite n is exposed by some witness base a; the only question is how far up you must search to be guaranteed one. GRH answers it:

\text{a composite } n \text{ has a witness } a \le 2(\log n)^2.

So you only ever test the bases a = 2, 3, \dots, \lfloor 2(\log n)^2\rfloor: if none of them is a witness, n is prime — with certainty, deterministically, in polynomial time. This is why for decades the fastest proven-fast primality test was "conditional on GRH." (The unconditional AKS test of 2002 finally removed the assumption, at the cost of a much larger exponent.) The same 2(\log n)^2 bound is what lets a fixed, tiny set of Miller–Rabin bases certify primality outright rather than merely "probably."

It feels like a category error: what does the location of a complex zero have to do with a computer finding a witness quickly? The bridge is that "a small witness exists" is really the statement "the non-witnesses do not form too large a subgroup" — and the size of that subgroup is controlled by how quickly a certain character sum \sum_{a}\chi(a) starts to cancel. That cancellation rate is exactly what the zeros of L(s,\chi) measure. Push all the zeros onto \Re(s) = \tfrac12 and the character sums cancel as fast as they possibly can, which forces a witness to appear by 2(\log n)^2. The critical line is a statement about primes, and — through characters — a statement about how soon structure has to reveal itself. That is why one conjecture in complex analysis quietly underwrites results in cryptography and computation.

The tower of generalisations

It is worth keeping the ladder straight, because the names get thrown around loosely:

Name	Applies to	Zeros on \Re(s)=\tfrac12?
RH	\zeta(s) only	conjectured
GRH	all Dirichlet L(s,\chi)	conjectured
ERH (Extended)	Dedekind \zeta_K(s) of any number field	conjectured
GRH (Grand)	all automorphic L-functions	conjectured

Each rung contains the one below it: the Dedekind zeta of \mathbb{Q} is just \zeta, and factors into Dirichlet L-functions over abelian fields, so ERH sits above GRH sits above RH. The Grand RH — phrased in the language of the Langlands programme — is the single conjecture that would, in one stroke, place the zeros of essentially every L-function that arithmetic has ever produced onto the same critical line.