What haven't you seen?

You've sampled a population — bugs from a fuzzer, words in a text, users on a page — and counted the unique bugs that turned up. How many haven't appeared yet, and what fraction of the next draw will be brand new? Two estimators answer that from the same two crumbs of evidence — the singletonsA type that showed up exactly once across your whole sample. The fewer of these you have relative to n, the better-explored the population is. and doubletonsA type that showed up exactly twice. Together with the singletons they're enough to estimate how big the iceberg is under the water. .

Inobservations I've seenunique.
showed up exactly once,showed up exactly twice.
That leaves 67 bugs you've seen three or more times — the well-explored core.
96.5%coverage (Good–Turing)Good–Turing's coverage estimate: the share of the next observation's probability that lands on a type you've already seen. Equivalently, 1 − f1/n.

On the next draw there's a 3.5% chance it's a brand-new bug — your singletons-to-n ratio is the whole answer.

≈ 34unseen bugs (Chao1)Chao1's lower-bound estimate of the number of types still hidden: f1² / (2·f2). When f2 = 0 we switch to the bias-corrected f1·(f1−1)/2.

The estimator points at about 34 more bugs hiding beyond the 120 you've already counted — a total of 154.

120 seen~34 still hidden (22.1% of the iceberg)
f₁, f₂The frequency-of-frequencies counts that drive both estimators. Everything else is built from these and n.35 · 18singletons · doubletons of 1,000 draws
Good–Turing96.5%C = 1 − f₁ ⁄ n = 1 − 351,000
P(new on next draw)3.5%= 1 − C = f₁ ⁄ n. Sometimes called the missing mass.
Chao1 unseen34Û = f₁² ⁄ (2·f₂) = 35² ⁄ (2·18)
Chao1 total Ŝ154Ŝ = S + Û = 120 + 34. A lower bound — the true total can be higher.
  • Both estimators assume your sample is roughly i.i.d. If draws are correlated — bursts of the same bug, sampling-with-replacement from a hot subset — coverage is overstated and Chao1 understates what's hiding.
  • Chao1 is a lower bound, not a point estimate. The actual richness can sit well above Ŝ when the distribution has a long, thin tail.
the unseen-species problem · Good–Turing coverage + Chao1 lower bound, from the same f₁, f₂ pair