The Exponential Distribution
Stand at a bus stop. Sit by a phone in a call centre. Point a Geiger counter at a lump of uranium.
In every case the same question hangs in the air: how long until the next thing
happens? The next bus, the next call, the next click of a decaying atom. The
exponential distribution is the answer — it is the distribution of the
waiting time until the next event.
You've already met its partner. The
Poisson distribution
counts how many events land in a fixed window — 3 calls this minute, 7 cars this hour. The
exponential asks the mirror-image question about the same stream of events: not "how many?"
but "how long between them?". Poisson is the count; exponential is the
wait. One is discrete (whole events), the other continuous (any positive length of
time), and they are two faces of the very same random process — a
Poisson process, in which events arrive independently at a steady average
rate \lambda.
The formula, and what its pieces mean
If events arrive at rate \lambda (events per unit time), the waiting time
X until the next one has probability density
f(x) = \lambda e^{-\lambda x}, \qquad x \ge 0.
Read it off the graph in your head: at x = 0 the density starts high, at
\lambda, and then decays smoothly toward zero as the wait
gets longer. Short waits are the most likely; long waits get rarer and rarer, but never quite
impossible. That relentless downhill slide is the "exponential" in the name.
Two more expressions do most of the day-to-day work. The chance the wait is longer than some
time x is the area in the tail,
P(X > x) = e^{-\lambda x},
and its complement, the cumulative distribution (the chance the event has already
happened by time x), is
F(x) = P(X \le x) = 1 - e^{-\lambda x}.
The tail formula P(X > x) = e^{-\lambda x} is the one to keep in your
pocket — it's the shortcut to almost every exponential question.
Mean and spread: the average wait is 1/\lambda
Here is the single most useful fact, and the one people most often get backwards. If events come at
rate \lambda, the average time between them is
\mathbb{E}[X] = \frac{1}{\lambda}.
It's pure common sense once you say it aloud: if a shop serves \lambda = 2
customers per minute, the typical gap between customers is 1/2 a minute.
Faster rate, shorter wait. The mean is the reciprocal of the rate — it is
1/\lambda, not \lambda.
The variance is the square of that mean,
\operatorname{Var}(X) = \frac{1}{\lambda^2}, \qquad \text{so the standard deviation is also } \frac{1}{\lambda}.
A standard deviation equal to the mean tells you the exponential is a very spread-out,
lopsided distribution: a big pile of short waits near zero, dragged out by a long thin tail of
occasional very long ones. Its long right tail makes the mean sit well to the right of the typical
(most likely) value — nothing like the tidy symmetry of the
normal curve.
Turn the rate up and down
Slide the rate \lambda and watch the whole story move. A
bigger \lambda starts the curve higher and drops it away
faster — events come thick and fast, so long waits become vanishingly unlikely. A
smaller \lambda gives a low, slowly-fading curve — a
sluggish stream where long waits are commonplace. The shaded region is the tail probability
P(X > 1) = e^{-\lambda}: the chance you wait more than one time-unit for
the next event. See how quickly it shrinks as \lambda grows.
- Models the waiting time until the next event of a Poisson process arriving at rate \lambda.
- Density f(x) = \lambda e^{-\lambda x} for x \ge 0; tail P(X > x) = e^{-\lambda x}; CDF F(x) = 1 - e^{-\lambda x}.
- Mean = 1/\lambda (the average wait); variance = 1/\lambda^2.
- It is memoryless — and is the only continuous distribution that is.
Worked example 1 — waiting for a phone call
Calls reach a small help-desk as a Poisson process at \lambda = 2 calls
per minute. You've just answered one. Two questions:
-
How long, on average, until the next call? The mean wait is
\mathbb{E}[X] = \frac{1}{\lambda} = \frac{1}{2}\ \text{minute} = 30\ \text{seconds}.
Two calls a minute means one every half-minute on average — exactly as expected.
-
What's the chance you get a whole quiet minute? That's
P(X > 1), straight from the tail formula:
P(X > 1) = e^{-\lambda \cdot 1} = e^{-2} \approx 0.135.
So a full call-free minute happens only about 13.5% of the time — roughly one
quiet minute in every seven. And the chance a call arrives within the first minute is the
complement, 1 - e^{-2} \approx 0.865.
Worked example 2 — the wait that resets itself
Same help-desk, \lambda = 2 per minute — but now here's the twist that
makes the exponential special. Suppose you've already been waiting
3 minutes and no call has come. What's the chance you wait at least
one more minute?
Intuition screams "surely a call is overdue by now!" — but the maths says otherwise. We want the
conditional probability
P(X > 3 + 1 \mid X > 3) = \frac{P(X > 4)}{P(X > 3)} = \frac{e^{-2\cdot 4}}{e^{-2\cdot 3}} = e^{-2(4-3)} = e^{-2} \approx 0.135.
That is exactly the same e^{-2} we'd get for a fresh, brand-new
wait — the three minutes you already sat through changed nothing at all. The distribution has
no memory of how long you've been waiting. Whatever the clock says, the next minute
looks like the first minute. This is the memoryless property, and it's worth its own
card.
The memoryless property
In symbols, for any already-elapsed wait s and any extra time
t,
P(X > s + t \mid X > s) = P(X > t).
In words: given that you've waited s already, the remaining wait
has the very same distribution as a fresh one. The process doesn't age. A component that
hasn't failed yet is, statistically, as good as new — a brand-new atom and one
that's sat undecayed for a billion years have identical chances of decaying in the next second.
Why does the exponential do this and nothing else? Because its tail is a pure power of
e, and powers add when you multiply:
e^{-\lambda(s+t)} = e^{-\lambda s}\, e^{-\lambda t}. Divide by
e^{-\lambda s} and the elapsed time cancels clean away. In fact the
exponential is the only continuous distribution with this property — it is the
continuous cousin of the (also memoryless) geometric distribution.
It's the neatest link in all of elementary probability. Say events arrive at rate
\lambda. The number in a window of length x is
Poisson with mean \lambda x, so the chance of seeing zero events
in that window is the Poisson formula with k = 0:
e^{-\lambda x}.
But "no event before time x" is exactly the same statement as "the waiting
time is longer than x." So
P(X > x) = e^{-\lambda x} — and there is your exponential tail, popped
straight out of the Poisson count. The gaps between Poisson events are exponentially
distributed, with the identical \lambda. Count and wait, two views of one
clock.
Three traps snare almost everyone meeting the exponential for the first time:
-
Memorylessness is not "I'm due." After a long dry spell people feel the next event
must be coming soon. For an exponential wait that's simply false: the past tells you
nothing about the future wait. The clock resets every single instant. Waiting longer does
not improve your odds one bit.
-
It only fits a constant rate. The exponential assumes a flat
hazard — an equal chance of "the event" in every instant. That's perfect for
radioactive decay or random arrivals, but wrong for things that wear out:
old tyres, ageing engines, human lifespans. Those have a rising hazard (more likely to fail
the older they get), so the exponential underestimates late failures. Reach for a Weibull or another
model there.
-
The mean is 1/\lambda, not \lambda.
A rate of \lambda = 5 per hour means an average wait of
1/5 of an hour (12 minutes), not 5. Rate and mean-wait are reciprocals —
mixing them up is the classic slip.