The Exponential Distribution for Waiting Times

Waiting Times and the Exponential Model

Many questions concern not how many events occur in a time period, but how long it takes until one event occurs. For example, a laboratory may study the time until a sensor registers a signal, or a service desk may model the time until the next request. Under suitable conditions, a continuous probability model can describe such a waiting time.

The exponential distribution is a model for a nonnegative waiting duration when the underlying event rate is constant over time. The model is useful only when that assumption is reasonable for the situation being studied; it is not a universal model for every duration.

Definition: A random variable $T$ has an exponential distribution with rate parameter $\lambda$, written $T\sim\operatorname{Exp}(\lambda)$, if $\lambda>0$, $T$ is a continuous, nonnegative waiting time, and its probability density function is $$ f_T(t)= \begin{cases} \lambda e^{-\lambda t}, & t\geq 0,\\ 0, & t<0. \end{cases} $$ The parameter $\lambda$ is the constant rate in the model. Its units are the reciprocal of the time units used for $T$. For example, if time is measured in minutes, $\lambda$ is measured per minute.

The rate parameter is not itself a waiting duration. A larger $\lambda$ makes the density decline more quickly as time increases, so the model assigns less probability to long waits. The numerical value of $\lambda$ must be paired with the time unit: a rate of $2$ per hour is not the same numerical rate as $2$ per minute.

Density, Cumulative Probability, and Survival

Because $T$ is continuous, a probability is an area under its density over an interval. The height $f_T(t)$ at one time is a density, not the probability that the wait equals that exact time. In particular, $P(T=t)=0$ for every individual time $t$.

The probability that the wait is at most $t$, for $t\geq0$, is the cumulative distribution function. Integrating the density gives this probability:

P(T\leq t)=\int_0^t \lambda e^{-\lambda u}\,du = \left[-e^{-\lambda u}\right]_0^t =1-e^{-\lambda t}.

Here $u$ is an integration variable representing possible waiting times. The complementary probability is the chance the wait lasts longer than $t$. This is called the survival function:

Result: For $T\sim\operatorname{Exp}(\lambda)$ and $t\geq0$, $$ F_T(t)=P(T\leq t)=1-e^{-\lambda t}, \qquad S_T(t)=P(T>t)=e^{-\lambda t}. $$ For $t<0$, $P(T\leq t)=0$, since waiting times cannot be negative.

The equality between the tail probability and $e^{-\lambda t}$ follows by taking the complement of $P(T\leq t)$. Since a continuous variable has zero probability at any single point, $P(T>t)=P(T\geq t)$. The same fact means that probabilities for an interval do not depend on whether its endpoints are included.

P(a<T\leq b)=F_T(b)-F_T(a) =e^{-\lambda a}-e^{-\lambda b}, \qquad 0\leq a\leq b.

This interval formula can also be obtained directly by integrating the density from $a$ to $b$. Both approaches give the same area. Before calculating, confirm that the cutoff times and rate use compatible units. If the rate is per hour, convert minutes to hours or convert the rate to per minute.

Using the Exponential Model

The examples below assume the exponential model is appropriate: the waiting time begins at a clearly specified starting point, it is continuous and nonnegative, and the event rate is reasonably constant under the conditions described. Each calculation uses the rate in the same time unit as its cutoff values.

Worked Example: Probability of Waiting Within a Time Interval

Suppose a hypothetical sensor's waiting time $T$ from activation until its next signal is modeled exponentially, with a constant rate of $3$ signals per hour under stable operating conditions. Find the probability that the wait is between 10 and 30 minutes.

The model assumption is that the waiting time is continuous and nonnegative, and that the signal rate remains constant in the operating conditions described. Convert both cutoff times to hours to match the rate: $10$ minutes is $1/6$ hour, and $30$ minutes is $1/2$ hour. Thus $\lambda=3$ per hour, $a=1/6$ hour, and $b=1/2$ hour.

Use the interval formula:

\begin{aligned} P(1/6<T\leq1/2) &=e^{-3(1/6)}-e^{-3(1/2)}\\ &=e^{-0.5}-e^{-1.5}\\ &\approx 0.6065-0.2231\\ &\approx 0.3834. \end{aligned}

Under this model, there is about a $0.3834$, or $38.34\%$, probability that the sensor waits between 10 and 30 minutes for a signal. The endpoints could be written with strict or inclusive inequalities without changing this probability.

Worked Example: Probability of Exceeding a Threshold

In a hypothetical setting, a service desk models the time $T$ until the next request as exponential with rate $4$ requests per hour. Find the probability that the wait exceeds 20 minutes.

Assume requests follow a stable rate during the period of interest and that $T$ is measured from the start of the waiting period. The rate is per hour, so convert $20$ minutes to $1/3$ hour. Apply the survival function:

\begin{aligned} P(T>1/3) &=e^{-4(1/3)}\\ &=e^{-4/3}\\ &\approx 0.2636. \end{aligned}

The model gives about a $26.36\%$ probability that the wait is longer than 20 minutes. This is a tail probability: it includes all possible waiting times beyond the cutoff, not just a narrow interval near 20 minutes.

Worked Example: Probability of a Wait Between Two Thresholds

Suppose the time $T$ to complete a hypothetical calibration task is modeled exponentially at a rate of $0.25$ completions per day. Under stable conditions, find the probability that completion takes more than 1 day but no more than 5 days.

The model treats the completion time as continuous and nonnegative, with a constant rate during the task conditions. The time cutoffs are already in days, matching the units of the rate. Substitute $a=1$, $b=5$, and $\lambda=0.25$ into the interval formula:

\begin{aligned} P(1<T\leq5) &=e^{-0.25(1)}-e^{-0.25(5)}\\ &=e^{-0.25}-e^{-1.25}\\ &\approx 0.7788-0.2865\\ &\approx 0.4923. \end{aligned}

The probability of completing the calibration task after day 1 and by day 5 is about $0.4923$, or $49.23\%$. The calculation subtracts the probability of waiting beyond 5 days from the probability of waiting beyond 1 day.

Why the Formula Works

The main probability formulas follow directly from the density, rather than from a table of separate cases. For any $t\geq0$, integrate the density from zero up to $t$. The antiderivative of $\lambda e^{-\lambda u}$ is $-e^{-\lambda u}$, so the area is $1-e^{-\lambda t}$. The full area under the density is therefore

\int_0^\infty \lambda e^{-\lambda u}\,du =\lim_{c\to\infty}\left(1-e^{-\lambda c}\right) =1,

because $\lambda>0$ implies $e^{-\lambda c}$ approaches zero as $c$ increases without bound. Thus the density assigns total probability one to nonnegative waiting times. Taking the complement of the cumulative probability gives $P(T>t)=e^{-\lambda t}$; subtracting cumulative probabilities at two cutoffs gives the interval formula.

This calculation also shows how the rate controls the model. For a fixed positive time $t$, increasing $\lambda$ increases $\lambda t$, so $e^{-\lambda t}$ decreases. Consequently, the modeled probability of still waiting beyond $t$ becomes smaller.

Common Mistakes and Model Checks

Treating the density as a point probability. The claim “$f_T(20)=0.26$, so the probability of waiting exactly 20 minutes is $0.26$” is incorrect. A density value is not a probability, and a continuous waiting time has $P(T=20)=0$. Use an interval area or a cumulative or survival probability.
Mixing time units. Substituting a cutoff of 20 into a rate of 4 per hour treats the time as 20 hours, not 20 minutes. Convert the cutoff to hours first, or convert the rate to per minute.
Using the tail formula for an interval. $e^{-\lambda a}$ is the probability of waiting longer than $a$, not the probability of waiting between $a$ and $b$. For an interval, subtract the tail beyond $b$: $e^{-\lambda a}-e^{-\lambda b}$.
Applying the model without checking its conditions. A changing rate, such as a process that becomes much busier at certain times, may make one constant-rate exponential model unsuitable. The formulas are valid for the stated model; their usefulness for a particular setting depends on whether its assumptions are plausible.

Key takeaway: For an exponential waiting time with rate $\lambda$, use $P(T>t)=e^{-\lambda t}$ for a threshold exceedance and $P(a<T\leq b)=e^{-\lambda a}-e^{-\lambda b}$ for an interval. First check that the model is appropriate and that rate and time units agree.

Check Your Understanding

Answer each question using the exponential model formulas and include units in your reasoning.

If a waiting time has rate $2$ per minute, what is the model's survival probability at 30 seconds? What time unit must you use in the calculation?
For $T\sim\operatorname{Exp}(0.5)$, with time measured in hours, write an expression for the probability that $T$ is between 1 and 4 hours.
Explain why the value of the density at a particular time is not the probability that the waiting time equals that value.
A model gives $P(T>3)=e^{-1.2}$. What is the rate parameter, and what time units are implied if the cutoff is 3 hours?
Name one feature of a real waiting-time process that could make the constant-rate exponential model questionable.