From Poisson to renewal: dropping the exponential
This whole rung was built on one special clock. The Poisson process counts events whose gaps — the interarrival times — are independent and identically exponentially distributed, and that single choice is what gave us all the magic: independent increments, the memoryless restart, the clean rate lambda. But real arrivals are rarely so obedient. Buses do not arrive memorylessly; a lightbulb that has burned for a while is more likely to fail soon, not equally likely as a fresh one. So we ask the natural next question: what survives if the gaps are i.i.d. but NOT exponential?
That generalization is a renewal process. Pick any distribution F for a positive interarrival time with mean mu = E[X] (we no longer demand X be exponential). Draw gaps X_1, X_2, X_3, ... independently from F. The arrival times are the running sums S_n = X_1 + ... + X_n, and the count N(t) is simply how many arrivals have happened by time t. Each arrival is a "renewal": at that instant the process forgets everything and starts the clock for the next gap from scratch. See the renewal process. The Poisson process is the one special member where F is exponential — and only then does the memoryless property hold.
Counting in the long run: the renewal function and the elementary theorem
The first thing we want to know is: how many renewals should we expect by time t? That expected count is the renewal function, m(t) = E[N(t)]. It is harder than it looks, because N(t) depends on a whole sum of random gaps, but its long-run behaviour is beautifully simple. See the renewal function. The headline result, the elementary renewal theorem, says that the long-run rate of renewals is exactly one over the mean gap: m(t)/t -> 1/mu as t grows. If buses come on average every 10 minutes, then over a long day you see about t/10 of them. No surprise — but it holds for ANY gap distribution with mean mu, exponential or not.
Why is the rate 1/mu, intuitively? By the strong law of large numbers, the n-th arrival time S_n = X_1 + ... + X_n grows like n·mu for large n: the average gap settles to mu. So the n-th renewal happens at roughly time n·mu, which is the same as saying that by time t we have had roughly t/mu renewals. The strong law does the heavy lifting; the renewal theorem is its statement read through the counting process. This is the renewal world's version of 'lambda = 1/mean gap' that you already met for the Poisson process — now freed from the exponential.
Pinned down, the objects are tidy. The gaps X_1, X_2, ... are i.i.d. from F with mean mu = E[X]; the arrival times are S_n = X_1 + ... + X_n; the count by time t is N(t) = max{ n : S_n <= t }; and the renewal function is m(t) = E[N(t)]. The elementary renewal theorem then reads m(t)/t -> 1/mu as t -> infinity. In the Poisson special case, F is Exponential(lambda) so mu = 1/lambda, and the theorem gives 1/mu = lambda — the very rate you have used all rung. A useful relative is the renewal-reward theorem: if each renewal earns a reward, your long-run reward per unit time is E[reward]/mu, the same 1/mu cadence scaled by the average reward.
The bus that is always late: the inspection paradox
Now the famous twist. Buses arrive as a renewal process with mean gap mu = 10 minutes. You show up at a random moment, knowing nothing, and wait for the next bus. Common sense whispers that you land in a 'typical' gap, so your wait should average about 5 minutes — half of 10. Common sense is wrong. The gap you fall into is systematically LONGER than the average gap, and you wait longer than half of mu. This is the inspection paradox. See the inspection paradox.
The picture explains it instantly. Lay all the gaps out on a timeline as intervals. A long gap covers more of the timeline than a short gap — a 20-minute gap occupies twice the road that a 10-minute gap does. So when you drop in at a random instant, you are MORE likely to land inside a long interval simply because long intervals are bigger targets. This is a length-biased sample of the gaps: each gap is effectively weighted by its own length. Short gaps are under-sampled, long gaps over-sampled, and the gap you experience is stretched accordingly. The same effect makes the average class size students experience larger than the average class size the registrar reports.
ordinary gap mean: E[X] = mu length-biased gap mean: E[X^2]/mu = mu + Var(X)/mu >= mu ( equality only when Var(X) = 0, i.e. every gap is exactly mu ) your expected wait for the next event = ( E[X^2] / mu ) / 2 (Poisson case below)
Even the Poisson process feels it
Here is the part that genuinely surprises people. Surely the memoryless Poisson process escapes the paradox? It does not — and the way it 'escapes' is itself the punchline. Suppose buses are a Poisson process with rate lambda = 1 per 10 minutes, so mu = 10 and gaps are exponential. By memorylessness, when you arrive your expected wait for the NEXT bus is the full 10 minutes, not 5: the exponential forgot the bus already gone, so your forward wait looks like a fresh, complete gap. That alone is striking — your wait equals the whole mean gap, not half.
Now look at the gap you actually fell into: it is your forward wait PLUS the time since the previous bus (the 'age' behind you). By memorylessness the backward age is also exponential with mean 10, independent of the forward wait. So the total gap you landed in has expected length 10 + 10 = 20 minutes — twice the ordinary mean of 10! The exponential's own variance (Var(X) = mu^2 here) feeds straight into the length-biased formula E[X^2]/mu = 2·mu. The Poisson process does not dodge the inspection paradox; it exhibits the cleanest possible version of it.
- Decide what you are sampling. The inspection paradox bites when you pick a random TIME point, not a random gap — time-sampling weights each gap by its length.
- Compute the length-biased mean E[X^2]/mu, not the ordinary mean mu, for the gap you will actually experience.
- Split that observed gap into forward wait + backward age; for a Poisson process each piece is exponential with mean mu by memorylessness.
- Read off the moral: your expected wait is (E[X^2]/mu)/2, which equals mu for the Poisson case — never the naive mu/2.
Why this matters, and what not to conclude
Renewal theory is the quiet workhorse behind reliability engineering (when does a repeatedly-replaced component fail?), queueing and the M/M/1 queue, inventory restocking, and any system that resets and runs again. The elementary renewal theorem tells you the long-run rate; refinements like the key renewal theorem describe the steady-state age and residual life. And the inspection paradox is not a curiosity — it is a recurring trap in real measurement: average download speed felt by users, average wait felt by callers, average crowd felt by riders, all skew long because experience is length-biased.
Two honest cautions. First, the paradox is NOT a paradox in the logical sense and nothing is wrong — it simply reflects that two different sampling schemes (pick a gap vs. pick a time) give two different, both-correct averages. The transit company's '10-minute average' is true for gaps; your '20-minute gap' is true for time-points. Neither lies. Second, do not confuse this with the gambler's fallacy or wishful 'evening out' — the long bus you waited through tells you nothing about the next gap, since the gaps are independent. The paradox is about the gap you are IN, decided by where you sampled, not a prediction about gaps to come.