Statistics › Section E · Continuous random variables › Topic E.3
The exponential distribution
The Poisson topic counted events in a window; this topic measures the gaps between them. When events occur randomly at a constant rate, the waiting time for the next one follows the exponential distribution - the continuous twin of the geometric.
Waiting for the next event
How long until the next call, the next radioactive decay, the next customer through the door? When events occur randomly at a constant rate - independently, with no memory of the last one - the waiting time between consecutive events follows the exponential distribution. It is the continuous twin of the geometric: where the geometric counts trials until the first success, the exponential measures the time until the next event. Because time is continuous, its probabilities are areas under a density curve, as in the previous two topics.
The exponential model and its rate
One parameter drives the model: the rate \(\lambda\), the average number of events per unit of time. Its density \(f(x) = \lambda e^{-\lambda x}\) on \(x \ge 0\) starts at height \(\lambda\) and decays toward zero - short waits are common, long waits fade away. The mean and the standard deviation are the same number, \(\frac{1}{\lambda}\): the rate flipped into an average gap. Six calls per hour means \(\lambda = 6\) per hour, so the mean gap is \(\frac{1}{6}\) of an hour - 10 minutes. Some books, including OpenStax, call the rate \(m\), the decay parameter, and write \(X \sim \text{Exp}(m)\); it is the same quantity.
The survival formula
Exponential probabilities come from one formula:
- the probability that the wait exceeds \(t\). The complement gives the other side: \(P(X < t) = 1 - e^{-\lambda t}\). With 6 calls per hour, the chance the next call takes more than 15 minutes is \(e^{-6 \times 0.25} = e^{-1.5} \approx 0.223\), so the chance it arrives within 15 minutes is \(1 - 0.223 = 0.777\). Note the discipline: \(\lambda\) and \(t\) must share time units, so that \(\lambda t\) is a pure number.
Memorylessness
The exponential is the only continuous distribution that is memoryless: having already waited \(s\) minutes changes nothing about the remaining wait - \(P(X > s+t \mid X > s) = P(X > t)\). After any elapsed time, the process restarts as a fresh copy. This fits radioactive decay and ideal arrival streams, where the next event's chance carries no wear. It does not fit devices that age: a lightbulb with 10,000 hours burned is not as good as new, and its failure risk rises rather than resets. Checking the memoryless assumption is checking the model itself.
The Poisson connection
The Poisson and exponential distributions are two views of one process. If the count of events per window is Poisson with mean \(\lambda\) per unit time, then the gaps between consecutive events are exponential with the same rate \(\lambda\) - and vice versa. Six calls per hour: the count in an hour is Poisson with mean 6, and the wait for the next call is exponential with a 10-minute mean. Counts answer "how many"; gaps answer "how long" - one random process, two questions, matched parameters.
Worked example: the help desk's gaps
A help desk receives 6 calls per hour, at random. 1. Average gap: \(\frac{1}{6}\) hour = 10 minutes. 2. Over 20 minutes: \(\lambda t = 6 \times \frac{20}{60} = 2\), so \(P(X > 20) = e^{-2} \approx 0.135\) - long silences are the minority. 3. Under 5 minutes: \(\lambda t = 0.5\), so \(P(X < 5) = 1 - e^{-0.5} \approx 0.393\). 4. Check: the three regions - under 5, between 5 and 20, over 20 - cover every possible wait, so their probabilities must total 1. The middle one is a difference of two survivals:
Worked example: the fresh-copy property
The desk has been silent for 10 minutes. What is the chance of at least 10 more minutes of silence? 1. Intuition says the call is "due" - the gambler's fallacy again. 2. Memorylessness: \(P(X > 20 \mid X > 10) = P(X > 10) = e^{-1} \approx 0.368\). 3. Compare a fresh start: \(P(X > 10) = e^{-1} \approx 0.368\) - identical, as the fresh-copy property promises. 4. Check: \(\frac{P(X>20)}{P(X>10)} = \frac{e^{-2}}{e^{-1}} = e^{-1}\), which is exactly the conditional probability - memorylessness working algebraically.
Percentiles and the median
Exponential percentiles invert the survival formula. Setting \(P(X < k) = p\) gives \(1 - e^{-\lambda k} = p\), so \(k = \frac{-\ln(1-p)}{\lambda}\). The median - the 50th percentile, where half of waits fall below - is \(\frac{\ln 2}{\lambda} \approx \frac{0.693}{\lambda}\). For the 6-per-hour desk: median gap \(= \frac{0.693}{6} \approx 0.116\) hours, about 6.9 minutes. Note the median sits below the 10-minute mean: the right skew drags the average above the typical wait, exactly as the skewness topic's rule predicts.
Common mistakes to avoid
First, mismatching units: \(\lambda\) per hour times \(t\) in minutes is meaningless - convert so that \(\lambda t\) is unitless. Second, reading \(\frac{1}{\lambda}\) as a probability: it is an average waiting time. Third, believing an elapsed wait makes the event due - memorylessness forbids it. Fourth, applying the exponential to devices that wear out - lightbulbs, batteries - whose failure risk rises with age. Fifth, confusing the exponential (a gap) with the Poisson (a count). Sixth, expecting the mean and median to agree: the right skew keeps them apart.
Key takeaways
Exponential models the waiting time between events of a random constant-rate process: density \(\lambda e^{-\lambda x}\), mean and SD both \(\frac{1}{\lambda}\), survival \(P(X > t) = e^{-\lambda t}\), left-area \(1 - e^{-\lambda t}\). It is memoryless - elapsed time restarts the clock - which fits no-wear processes like decay and arrivals, and fails for anything that ages. Counts in a window are Poisson; gaps between events are exponential; the same \(\lambda\) drives both. Percentiles come from \(k = \frac{-\ln(1-p)}{\lambda}\), and the median \(\frac{\ln 2}{\lambda}\) sits below the mean - the signature of the right-skewed shape.
Flashcards
Flashcard
Waiting time between events of a random constant-rate process
Test your understanding
Question 1 of 5
A call centre receives 4 calls per hour, at random. What is the mean waiting time between calls?
Question 2 of 5
At 6 calls per hour, what is the probability the next call takes more than 20 minutes?
Question 3 of 5
The desk has been silent 10 minutes (6 calls per hour). What is the probability of at least 10 more silent minutes?
Question 4 of 5
Which of these fits an exponential model?
Question 5 of 5
At 6 calls per hour, what is the probability the next call arrives within 5 minutes?
