Buses come every ten minutes on average, so your wait should average five. It doesn't. Show up at a random moment and you will wait longer than that — sometimes twice as long — and no amount of good scheduling fixes it, because the fault is in the word average.
Here is the complaint, and it is older than buses: you arrive at the stop, you wait, and the wait feels longer than it has any right to. The timetable swears the buses come every ten minutes on average. Ten minutes, average — so a person dropping in at a random time should wait, on average, five. Half the gap. That's the intuition, and it is wrong, and the gap between what it promises and what happens is a real theorem with a real name.
The trap is hiding in plain sight. When you arrive at a random moment, you don't sample a random gap between buses — you sample a random moment, and a moment is more likely to fall inside a long gap than a short one, precisely because a long gap covers more of the timeline. A ten-minute gap is ten times likelier to catch your arrival than a one-minute gap. So the gap you land in is not a typical gap. It is a gap weighted by its own length, and that weighting drags it longer. Watch it happen — drop arrivals onto a real timeline and see which gaps they keep falling into.
The surprise has an exact size, and it is the same clean line that governs a dozen other paradoxes. Write μ for the mean gap and σ² for its variance — how unevenly the gaps are spread. The mean length of the gap a random arrival lands in is
where CV = σ/μ is the coefficient of variation — the spread measured in units of the mean. The extra term σ²/μ is the whole paradox. It is never negative, so the gap you land in is always at least the average gap, and strictly longer the instant the gaps vary at all. Your wait is exactly half of that spread gap, E[X²]/(2·E[X]), by the symmetry that the arrival lands uniformly inside whatever gap it caught.
Everything now lives in one number, CV — the raw variability of the timetable — and the multiplier is 1 + CV². Three worlds sit on that curve:
The memoryless case deserves its own beat, because it is the sharpest. If buses arrive as a Poisson process, then no matter how long you have already been standing there, your expected further wait is still the full mean — ten minutes, always. Waiting doesn't earn you anything. This is the "hitchhiker's paradox," and the verifier checks it directly: filtering to moments already fifteen minutes into a gap with no bus, the expected remaining wait is still ten. The past is spent; the process forgot it.
Now step away from buses, because the identity does not care what is being measured — only that you sample the thing you land in. A university advertises an average class size of 27. It is not lying; average over its classes and you get 27. But ask a random student what size class they are sitting in, and the honest answer is much larger — because a big lecture hall contains many students to ask, and a tiny seminar contains few. You sample classes weighted by their own size, exactly as an arrival samples gaps weighted by their own length. Same identity, same σ²/μ.
This is why the "average class size" a college quotes and the class size a student actually experiences are two different numbers that both deserve to be called an average — and why the brochure figure always flatters. It is the honest arithmetic of who gets counted, not a marketing lie, and knowing the shape of it lets you ask the sharper question: averaged over what?
Once you can see the shape, you cannot stop seeing it. The statistician Allen Downey collected a gallery of it under the title "The inspection paradox is everywhere," and the examples all have the same skeleton — you sample intervals in proportion to their length, so the interval you inspect is inflated by σ²/μ:
The unifying picture is worth holding onto. There is a true distribution out there — of gaps, of class sizes, of family sizes, of friend-counts — and between you and it sits a sampling scheme that weights each thing by its own magnitude. That length-biased lens can only push the mean up, never down, and it pushes it up by exactly the variance over the mean. The whole family of "paradoxes" is one theorem wearing different clothes: the friendship paradox on graphs, the queue that won't move in operations research, and this — the gap you land in — in time.
The core is a theorem and is exact. The mean of the interval containing a random instant is E[X²]/E[X] = μ + σ²/μ; the mean forward wait is half of it; both are standard renewal theory (Cox 1962; Feller, Vol. II), and both are re-derived here two independent ways — a closed form from the raw moments, and a genuine timeline simulation that drops uniformly-random arrivals onto a long strip of renewals and reads off the gap each one lands in, with no length-weighting assumed anywhere. The bias has to emerge on its own, and it does: 31/31 checks in research/inspection-paradox/verify.mjs.
The honest edges, stated rather than hidden. The four timetables — regular, Erlang, Poisson, bursty — are chosen to span the range of CV from 0 upward; a real bus route is somewhere in the middle, and the point is not that buses are exactly Poisson but that any variability at all (σ² > 0) forces the gap open. The full 2× doubling is the Poisson extreme; real routes see less, but never zero. The college is a labelled illustrative model — 100 classes, the sizes named on the panel — not a specific institution's registrar data; the effect is real and well documented (it is why quoted "average class size" understates the lived one), only the exact figures are a model. And the instruments are seeded simulations of the mechanism; the exact numbers they converge to are the theorems above. None of that dents the spine. Your bus really is, on average, later than the timetable's average gap would suggest. It just isn't the driver's fault — it's yours, for arriving at a random time, which is the only kind of time there is.