Exponential Memorylessness and Residual Waiting Times

Why the Remaining Wait Can Ignore the Elapsed Time

Suppose a Poisson arrival clock has not rung yet, even though it has already been running for some time. How much longer should we expect to wait? For exponential interarrival times, the answer has a striking feature: conditional on the clock still being active, its remaining wait has the same distribution as a fresh interarrival time. This is the memoryless property. It is a statement about a conditional distribution, not a claim that the elapsed time has disappeared from the history.

Work on a probability space $(\Omega,\mathcal{F},\mathbb{P})$. Let $(N_t)_{t\ge0}$ be a Poisson process with state space $\{0,1,2,\ldots\}$, continuous time $t\ge0$, and rate $\lambda>0$. Write $S_k$ for its $k$th arrival time and $X_k=S_k-S_{k-1}$ for its $k$th interarrival time, with $S_0=0$. The interarrival times are independent and each has exponential distribution with rate $\lambda$. In particular, for a generic interarrival time $X$, $\mathbb{P}(X>x)=e^{-\lambda x}$ for $x\ge0$.

Definition: For a finite age $s\ge0$, the residual waiting time of an interarrival clock, conditional on its having lasted longer than $s$, is represented by $R_s=X-s$ under the conditional law given $X>s$. This conditioning is meaningful because $\mathbb{P}(X>s)=e^{-\lambda s}>0$. For a clock whose current interarrival is $X_k$, the corresponding residual is $X_k-s$ given $X_k>s$.

Thus $R_s$ is not the original interarrival time: it is the time left after age $s$, under a condition that the arrival has not yet occurred. Because an exponential random variable has a continuous distribution, conditioning on $X>s$ or $X\ge s$ gives the same conditional law. The strict inequality is convenient for defining the event of no arrival by age $s$.

The Exponential Survival Identity

Theorem: If $X$ has exponential distribution with rate $\lambda>0$, then for all finite $s,t\ge0$, $$ \mathbb{P}(X>s+t\mid X>s)=\mathbb{P}(X>t). $$ Equivalently, $R_s$ has the same exponential distribution with rate $\lambda$ as $X$.

Proof. The conditional probability is defined because $\mathbb{P}(X>s)>0$. By the definition of conditional probability and the exponential survival function,

\mathbb{P}(X>s+t\mid X>s) =\frac{\mathbb{P}(X>s+t,\ X>s)}{\mathbb{P}(X>s)} =\frac{\mathbb{P}(X>s+t)}{\mathbb{P}(X>s)} =\frac{e^{-\lambda(s+t)}}{e^{-\lambda s}} =e^{-\lambda t} =\mathbb{P}(X>t).

The second equality holds because $s,t\ge0$ implies $X>s+t$ entails $X>s$. For the equivalent statement, for every $t\ge0$, the event $R_s>t$ under the condition $X>s$ is exactly $X>s+t$. The identity therefore gives the survival function of $R_s$, and hence its distribution, as exponential with rate $\lambda$. $\square$

The theorem also yields the full conditional distribution, not just its survival probabilities. For $t\ge0$, the conditional probability that the remaining wait is at most $t$ is one minus its survival probability. Differentiating that distribution function gives its density.

Residual distribution: For $t\ge0$, the conditional distribution function and density of $R_s$ are $$ \mathbb{P}(R_s\le t\mid X>s)=1-e^{-\lambda t}, \qquad f_{R_s\mid X>s}(t)=\lambda e^{-\lambda t}. $$ In particular, the conditional mean residual wait is $1/\lambda$, independently of the age $s$.

To verify the mean, integrate the density. Integration by parts gives $\int_0^\infty t\lambda e^{-\lambda t}\,dt =[-te^{-\lambda t}]_0^\infty+\int_0^\infty e^{-\lambda t}\,dt=1/\lambda$; the boundary term is zero since $te^{-\lambda t}\to0$ as $t\to\infty$. This calculation applies for every $\lambda>0$.

Calculating Residual Probabilities

Worked Example: A Clock Already Running for Half an Hour

Let the interarrival time $X$ have exponential rate $\lambda=2$ per hour. Given that the current gap has lasted longer than $s=\tfrac12$ hour, find the probability that it lasts longer than one additional half-hour, and the probability that the arrival occurs within that additional half-hour.

The first event is $X>s+t$, where $t=\tfrac12$. The survival identity gives

\mathbb{P}(X>1\mid X>\tfrac12) =\mathbb{P}(X>\tfrac12) =e^{-2(1/2)} =e^{-1}.

The probability of arrival within the next half-hour is the complementary conditional probability: $\mathbb{P}(X\le1\mid X>\tfrac12)=1-e^{-1}$. For a direct check, the ratio of survival probabilities is $\mathbb{P}(X>1)/\mathbb{P}(X>\tfrac12)=e^{-2}/e^{-1}=e^{-1}$, as above. The specified age changes the event being conditioned on, but not the distribution of the remaining wait.

Worked Example: Distribution and Mean of the Remaining Wait

Interarrivals have rate $\lambda=3$ per day. The current gap is known to exceed $s=\tfrac13$ day. Find the conditional probability that the residual wait is at most $\tfrac13$ day, the probability that it lies between $\tfrac13$ and $\tfrac23$ day, and its mean.

By the residual distribution, the conditional cumulative distribution function at $t=\tfrac13$ is $1-e^{-3(1/3)}=1-e^{-1}$. For the interval probability, subtract the cumulative probabilities at its two endpoints. Since the residual distribution is continuous, using either strict or non-strict endpoints gives the same probability:

\mathbb{P}\bigl(\tfrac13<R_s\le\tfrac23\mid X>\tfrac13\bigr) =(1-e^{-3(2/3)})-(1-e^{-3(1/3)}) =e^{-1}-e^{-2}.

The density is $3e^{-3t}$ for $t\ge0$, so the mean is $\int_0^\infty 3te^{-3t}\,dt=1/3$ day, using the integration-by-parts calculation above with $\lambda=3$. As a check on the interval probability, integrate the density directly: $\int_{1/3}^{2/3}3e^{-3t}\,dt=[-e^{-3t}]_{1/3}^{2/3}=e^{-1}-e^{-2}$. The same distribution and mean apply for any finite age $s$ for which the clock is conditioned to have lasted longer than $s$.

Reading the Identity as a Poisson-Clock Calculation

In a Poisson process starting at time $0$, the first arrival time is $X_1$. The event $X_1>s$ means exactly that no arrival has occurred by time $s$. Given this event, the time from $s$ to the first arrival is $X_1-s$. Consequently, for finite $s,t\ge0$,

\mathbb{P}(X_1-s>t\mid X_1>s) =\mathbb{P}(X_1>s+t\mid X_1>s) =e^{-\lambda t}.

The same calculation applies within any interarrival gap: if $X_k$ is the gap under consideration, then conditional on $X_k>s$, the remaining portion $X_k-s$ has exponential rate $\lambda$. This is a statement about the residual time on that exponential interarrival clock. It does not assert that elapsed time is unknown, or that every waiting-time model has this property.

Worked Example: A Poisson Clock Has Not Rung by the Observation Time

A Poisson clock has rate $\lambda=4$ per hour and starts at time $0$. Given that it has had no arrival by $s=\tfrac12$ hour, find the probability of at least one arrival in the following $\tfrac14$ hour, and the conditional expected time from the observation time to the first arrival.

Because the first arrival time is $X_1$, the event of no arrival by $\tfrac12$ is $X_1>\tfrac12$. An arrival in the next $\tfrac14$ hour is the event that the residual is at most $\tfrac14$. Thus

\mathbb{P}(X_1\le\tfrac34\mid X_1>\tfrac12) =1-\mathbb{P}(X_1>\tfrac34\mid X_1>\tfrac12) =1-e^{-4(1/4)} =1-e^{-1}.

The conditional expected remaining time is $1/\lambda=1/4$ hour. To check the probability by a ratio, the conditional probability of no arrival during the following interval is $\mathbb{P}(X_1>\tfrac34)/\mathbb{P}(X_1>\tfrac12) =e^{-4(3/4)}/e^{-4(1/2)}=e^{-1}$. Its complement is $1-e^{-1}$, matching the result.

When the Memoryless Calculation Does Not Apply

The formula depends on the exponential survival function. It cannot be applied to an arbitrary waiting time merely because the clock has already run for $s$. For example, let $X$ be uniform on $(0,2)$, a nonexponential waiting time. With $s=\tfrac12$ and $t=\tfrac12$, the conditional probability of another half-unit of waiting is

\mathbb{P}(X>1\mid X>\tfrac12) =\frac{\mathbb{P}(X>1)}{\mathbb{P}(X>\tfrac12)} =\frac{1/2}{3/4} =\frac23.

Here $\mathbb{P}(X>\tfrac12)=\tfrac34$, which is not $\tfrac23$. The error would be to replace the conditional probability by the unconditional survival probability at $t=\tfrac12$. That replacement is justified for an exponential variable by the theorem, not by conditional probability alone.

Common pitfall. Do not confuse “the gap has already lasted longer than $s$” with “the total gap is longer than $s+t$.” The conditional event has probability $\mathbb{P}(X>s+t)/\mathbb{P}(X>s)$, not simply $\mathbb{P}(X>s+t)$. For an exponential gap, the factors involving $s$ cancel, leaving $e^{-\lambda t}$. Also, the rate must be positive: when $\lambda=0$, there are no arrivals and no finite exponential interarrival time to which this residual distribution applies.

This property is useful whenever the state of a Poisson clock is described by its current age but the quantity of interest is the time remaining until its next arrival. The age may be known exactly, yet the conditional residual law stays exponential with the same rate. In calculations, the safe procedure is to identify the event $X>s$, express the requested residual event using $X-s$, and then use the conditional-probability ratio. That procedure keeps the conditioning explicit and makes clear which assumptions support the simplification.

Check Your Understanding

For each question, state the conditioning event and distinguish the residual time from the full interarrival time.

If $X$ is exponential with rate $\lambda>0$, derive $\mathbb{P}(X>s+t\mid X>s)$ for finite $s,t\ge0$.
A rate-$5$ Poisson clock has an ongoing gap known to exceed $s=\tfrac15$ hour. What is the conditional probability that the residual wait exceeds $\tfrac25$ hour?
For exponential rate $\lambda$, write the conditional density of $R_s$ and compute its mean.
Why is conditioning on $X>s$ meaningful for every finite $s$ when $X$ is exponential with positive rate?
For $X$ uniform on $(0,2)$, compute $\mathbb{P}(X>1\mid X>\tfrac12)$ and explain why the exponential identity cannot be used.