If bus waiting time is exponential with mean 10 minutes, why does waiting 10 minutes not shorten the expected remaining wait?

## The distribution An exponential distribution models a continuous, non-negative waiting time with a single parameter, the rate `lambda`. Its density, CDF and survival function are: ``` f(t) = lambda * exp(-lambda * t) for t >= 0 F(t) = 1 - exp(-lambda * t) probability the wait is at most t S(t) = exp(-lambda * t) probability the wait exceeds t ``` Its mean is `1 / lambda` and its variance is `1 / lambda^2`. A mean wait of 10 minutes therefore means `lambda = 0.1` arrivals per minute. Note the reciprocal relationship: a bigger rate means a shorter wait. Confusing `lambda` with the mean is the most common arithmetic slip on this topic. ## Memorylessness, stated precisely A non-negative random variable T is memoryless when ``` P(T > s + t | T > s) = P(T > t) for all s, t >= 0 ``` In words: given that the wait has already exceeded s, the probability it lasts another t is exactly what it was at the very start. The proof for the exponential is one line of algebra using the definition of conditional probability, `P(A | B) = P(A and B) / P(B)`. Since `T > s + t` already implies `T > s`, the joint event is just `T > s + t`: ``` P(T > s + t | T > s) = exp(-lambda(s + t)) / exp(-lambda s) = exp(-lambda s - lambda t + lambda s) = exp(-lambda t) = P(T > t) ``` So the conditional distribution of the residual wait, given any amount already elapsed, is the same exponential with the same rate. Its expectation is `1 / lambda = 10` minutes again. The ten minutes you have already stood at the stop buy you nothing. ## The hazard rate view The hazard rate is `h(t) = f(t) / S(t)` — the instantaneous rate of the event happening given it has not happened yet. For the exponential: ``` h(t) = lambda * exp(-lambda t) / exp(-lambda t) = lambda ``` Constant, for all t. Memorylessness and constant hazard are two descriptions of the same fact. This is the most useful reframing in practice, because it says exactly what the model assumes about the world: nothing ages, nothing wears out, nothing becomes overdue. ## Uniqueness The exponential is the **only** continuous distribution that is memoryless. The functional equation `S(s + t) = S(s) * S(t)` with S monotone forces an exponential form, so any continuous memoryless waiting time is exponential for some rate. In the discrete world the geometric distribution plays the same role: the number of trials until the first success has no memory of how many failures have already occurred. Being able to name both, and to say that no other continuous law qualifies, is usually what an interviewer is fishing for. ## Mean, median and spread Setting `F(t) = 0.5` gives the median: ``` median = ln(2) / lambda ≈ 0.693 / lambda ``` For a 10-minute mean that is about 6.93 minutes. The mean exceeds the median because the exponential is right-skewed: most waits are short, and a long tail of rare long waits drags the average up. Also note `SD = 1 / lambda = mean`, so the coefficient of variation is exactly 1. That is a handy diagnostic: if an observed set of waiting times has a standard deviation far from its mean, an exponential model is already in trouble. ## When the assumption is wrong Memorylessness is a strong claim, and plenty of real waits violate it: - **Wear-out.** A mechanical part becomes more likely to fail the longer it has run — an increasing hazard. An exponential cannot represent that, and will overstate survival late in life. - **Infant mortality.** Some systems fail most in their first hours and then settle — a decreasing hazard. Again not exponential. - **Scheduled service.** A bus on a strict timetable is precisely the opposite of memoryless: the longer you have waited, the closer the next scheduled departure. The exponential fits an unscheduled, irregular arrival stream, not a timetable. A gamma distribution with shape greater than 1 gives an increasing hazard, and with shape less than 1 a decreasing one; shape exactly 1 is the exponential itself. Recognising which hazard shape your problem needs is the actual modelling decision hiding behind the maths. ## The interview answer in one breath Name the property, write the conditional probability, cancel the exponentials, say "constant hazard", say "unique among continuous distributions", and finish with a case where the assumption fails. That sequence covers everything the question is testing.