Three inferences, from three fields that never speak, run one move. To measure a whole you can only partly see, distrust the average and read the edge it throws away.
The instinct is that you cannot know what you have not seen. Each of these three results breaks it, and each breaks it the same way: by refusing the average and interrogating a feature of the sample the average throws away. Not the middle of the data, but its edge, the single largest value, the things seen exactly once, the bare scatter from one trial to the next.
To measure a whole you can only partly see, distrust the mean and read the informative edge. And which edge carries the answer is fixed by the kind of unknown: a COUNT is bounded by the MAXIMUM, a RICHNESS is fed by the SINGLETONS, a RATE is exposed by the VARIANCE.
1940. The Allies need Germany’s monthly tank output. Their spies overguess it four to eight times over. A handful of statisticians read the serial numbers stamped on captured tanks and beat them by a wide margin, later vindicated when the German records were opened.
The maximum you saw is not the ceiling; it sits a predictable step below it. The step is the average gap between your serials, and adding it back lands you within a couple of percent of a total you were never shown. The information about the tanks you never caught is carried by the highest-numbered tank you did.
Shakespeare wrote about 31,534 distinct words. How many did he know and never use? The words he used exactly once, the hapax legomena (about 38% of his vocabulary), answer it, through the same mathematics that estimates butterflies you have not yet netted.
At a doubling (t=1) the estimator returns about 11,435 new words, recovering Efron and Thisted’s published 11,430 to a fraction of a percent. Push t past about 1.3 and it flies apart: the method is honest about its own horizon, which is why “he knew roughly 67,000 words” is a model-dependent lower bound, not a headline fact.
1943. Do bacteria mutate because a threat arrives, or were the survivors already there? One culture cannot tell you: both stories predict the same average. Luria and Delbrück read the answer off how violently parallel cultures disagree.
If mutation is induced, the counts are Poisson and the variance equals the mean. If it is spontaneous, an early mutant founds a clone that doubles into a jackpot, and the variance dwarfs the mean by a factor in the hundreds. The scatter that everyone before them had discarded as noise was the proof. And the mean is worse than weak here: it is literally infinite, growing without limit as you grow the culture, so the rate is read instead from p₀, the empty cultures the jackpots cannot corrupt.
Line the three up and the shared machinery is plain. Each reaches a whole that lies outside the data, and each does it by reading exactly the feature the average discards. The edge is not interchangeable: it is chosen by the kind of unknown.
| The thing | The hidden whole | The edge that holds it | The average, and why it fails | The law |
|---|---|---|---|---|
| German tank output | a COUNT (total N) | the MAXIMUM serial | twice the mean (unbiased but wide (5.667× worse MSE)) | N̂ = m + (m−k)/k = m(k+1)/k − 1 |
| Shakespeare's vocabulary | a RICHNESS (unseen words) | the SINGLETONS (hapax) | mean word-frequency (a flat number blind to frequency-zero) | new-words(t) = n₁t − n₂t² + n₃t³ − … |
| bacterial mutation | a RATE (and its history) | the VARIANCE (and p₀) | the sample mean (literally infinite; never settles) | m = −ln(p₀) |
To measure a whole you can only partly see, distrust the mean and read the informative edge. And which edge carries the answer is fixed by the kind of unknown: a COUNT is bounded by the MAXIMUM, a RICHNESS is fed by the SINGLETONS, a RATE is exposed by the VARIANCE.
Two of these layers already hang in other portals, on other axes, and that is disclosed here in full. The tank problem appears in The Signal Never Sent, arranged by how hard a source tried to keep its secret (serials as inventory); here it runs on the estimation axis, the maximum revealing the count. Luria–Delbrück appears in The Man Lightning Kept Finding, on the rare-event heavy-tail axis; here it runs on the variance-as-evidence axis. Neither prior portal makes the count / richness / rate split this one makes, because that split needs all three members present at once.
These are three different probability models, not one formula in three costumes, and the portal does not pretend otherwise. The tank estimator assumes uniform sampling from a contiguous, gap-free run of serials, a designed sequence; reset the serials per factory, or leave gaps, and it breaks. The Shakespeare estimator assumes a fixed latent vocabulary sampled by a Poisson process, and "what counts as a word" is not invariant (our tokeniser gives 7,643 unseen words where Efron and Thisted got 11,430, from the same method). The fluctuation test assumes a branching process with a 1/(k(k+1)) clone-size law, and its estimator is p₀, not any moment. The unification is the epistemic strategy, not the arithmetic. The novelty here is the reading; each member’s mathematics was verified on its own page, and this portal imports those checks unchanged.
Every number on this page is recomputed live in your browser and re-checked offline, and the three members’ own verifiers are re-run so the portal only stands if they still pass:
node research/what-the-average-throws-away/verify.mjs