The Prime Number Theorem tells us that \pi(x), the count of primes up to x, is asymptotic to x/\ln x. That is true — but it is also a slightly lazy way to state the theorem, because x/\ln x is a surprisingly loose fit. There is a cleaner, sharper "main term" hiding just underneath it: the logarithmic integral \operatorname{Li}(x). It hugs \pi(x) so closely that at a million it is off by barely a hundredth of a percent, and — as we'll see — the exact size of the gap between \pi(x) and \operatorname{Li}(x) is one of the deepest open questions in mathematics.

The idea behind it is almost childishly simple. If a whole number near t is prime with "probability" about 1/\ln t, then to count all the primes up to x we should just add up those probabilities — and adding up a density is exactly what an integral does. That single instinct produces the best elementary estimate for \pi(x) that anyone knows.

The definition

The two differ only by a constant: \operatorname{Li}(x) = \operatorname{li}(x) - \operatorname{li}(2), where \operatorname{li}(2) \approx 1.045. Starting the clock at 2 instead of 0 is convenient for prime counting — the first prime is 2, and it avoids the pole at t = 1 entirely. The integrand 1/\ln t blows up as t \to 1^+, which is exactly why \operatorname{li}(x) needs the principal-value dodge and \operatorname{Li}(x) does not.

There is no elementary closed form for either integral — you cannot write \int dt/\ln t in terms of powers, roots, exponentials and logs. It is a genuinely new "special function", as fundamental to prime counting as the Gamma function is to interpolation.

The heuristic: summing a density

Why should \int_2^x dt/\ln t count primes? Gauss's teenage insight was that the primes thin out at a predictable rate: near t, about one integer in every \ln t is prime, so the local density of primes is 1/\ln t. Near t = 100 that density is 1/\ln 100 \approx 0.217 (roughly one prime in five); near t = 10^9 it has fallen to 1/\ln 10^9 \approx 0.048 (about one in twenty-one).

To count all the primes up to x, add up that density across every value of t. A discrete sum \sum_{t=2}^{x} 1/\ln t becomes, in the continuous limit, the integral

This is more honest than x/\ln x. The crude estimate x/\ln x pretends the density stays fixed at its final value 1/\ln x across the whole range [2, x] — but early on the primes are much denser, so that undercounts. \operatorname{Li}(x) uses the correct, higher density at every point along the way, which is precisely why it comes out larger, and closer to the truth.

The asymptotic expansion — where the extra accuracy comes from

Integrate \operatorname{Li}(x) by parts repeatedly (take u = 1/\ln t, dv = dt) and a beautiful expansion falls out:

Read this carefully. The very first term is x/\ln x — so \operatorname{Li}(x) agrees with the crude estimate to leading order, which is why both satisfy the PNT. But \operatorname{Li}(x) then adds an entire tail of correction terms 1!/\ln x, 2!/\ln^2 x, \dots that x/\ln x simply throws away. That first correction, x/\ln^2 x, is not small — it is smaller than the main term by only a single factor of \ln x — so ignoring it (as x/\ln x does) is what leaves that estimate stuck several percent adrift.

Worked example — watching the corrections land at x = 10^6. Here \ln x = 13.816, so x/\ln x = 72{,}382. Multiply in the successive factors k!/(\ln x)^k and accumulate:

The total is climbing toward \operatorname{Li}(10^6) \approx 78{,}628 — and the true count is \pi(10^6) = 78{,}498. Notice how the crude first line, 72{,}382, is more than six thousand short, while just a few correction terms carry us to within a hundred or so of the answer. That tail of factorials is the whole story of why \operatorname{Li} wins.

Seeing all three at once

Below, the jagged staircase is the true \pi(x), jumping by one at every prime. The two smooth curves are the estimates: \operatorname{Li}(x) rides just above the staircase, tracing its shape almost perfectly, while x/\ln x sags noticeably below. Even at this modest scale the difference is unmistakable — and it only widens in \operatorname{Li}'s favour as you zoom out.

Watch how \operatorname{Li}(x) stays a hair above the staircase across the whole window. That persistent overshoot — \operatorname{Li}(x) > \pi(x) — is not an accident of this range; as we'll see, it holds for every x anyone has ever computed, and yet is provably not the final word.

The numbers, side by side

Nothing sells \operatorname{Li}'s superiority like the raw figures. Here are the two estimates against the truth at three scales, together with their errors (using the offset \operatorname{Li}):

Look at the two error columns. At x = 10^9, the crude estimate is short by over two and a half million primes; \operatorname{Li} is off by just 1{,}701 — a relative error of roughly 0.000003\%. And there is a striking pattern in the \operatorname{Li} column: its error grows only like \sqrt{x} (compare 10, 130, 1701 against the square roots \sqrt{10^3}\approx 32, \sqrt{10^6}=1000, \sqrt{10^9}\approx 31{,}623). That square-root-sized error is not a coincidence — it is the fingerprint of the Riemann zeta zeros.

The connection to the zeta zeros

The Prime Number Theorem is best written not with x/\ln x but with \operatorname{Li} as its true main term:

and the entire mystery of the primes lives in the error E(x). Riemann's explicit formula expresses that error as a sum over the non-trivial zeros \rho of \zeta, schematically E(x) \approx -\sum_{\rho} \operatorname{Li}(x^{\rho}). Each zero \rho = \beta + i\gamma contributes an oscillation of size about x^{\beta}: the further right a zero sits, the larger and more damaging its wobble.

This is why the error column above hovers near \sqrt{x}: if RH holds, the gap \operatorname{Li}(x) - \pi(x) can never grow faster than about \sqrt{x}\,\ln x. A single zero drifting off the critical line to \Re(s) = \beta > \tfrac12 would let the error swell to x^{\beta} and spoil the fit. So the astonishing tightness of \operatorname{Li} is, quite literally, empirical evidence for the Riemann Hypothesis.