The Memoryless Property of Exponential Waiting

Why an Event-Free Period Does Not Change an Exponential Wait

Suppose a waiting time has an exponential distribution. If no event has occurred for a while, how does that affect the probability of waiting still longer? The answer is a distinctive feature of the exponential distribution: conditional on having waited through a fixed period, the remaining wait has the same distribution as the original wait. The conditioning matters. The unconditional chance of a long wait and the chance of a long residual wait given survival so far are different probabilities.

Work on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). Let \(T\) be a nonnegative, finite waiting time with exponential rate \(\lambda>0\). Its survival function is \(\mathbb{P}(T>u)=e^{-\lambda u}\) for \(u\ge0\). In particular, for any fixed \(s\ge0\), the event \(T>s\) has positive probability \(e^{-\lambda s}\), so conditioning on that event is meaningful. The distribution has no atoms: its distribution function \(1-e^{-\lambda u}\) is continuous, including at \(u=0\).

The identity concerns a fixed elapsed time \(s\) and a fixed additional waiting time \(t\). The event in the conditioning has positive probability for every finite \(s\). The cases \(s=0\) and \(t=0\) are included: when \(s=0\), the conditioning event has probability \(1\), and when \(t=0\), both sides equal \(1\).

Proof. By the definition of conditional probability and the inclusion \(\{T>s+t\}\subseteq\{T>s\}\),

For \(t\ge0\), the event that the residual wait exceeds \(t\), given \(T>s\), is exactly \(\{T>s+t\}\) under the same conditioning. Thus the conditional survival function of \(T-s\) is \(e^{-\lambda t}\), which is the exponential survival function with rate \(\lambda\). This proves both statements. \(\square\)

The proof uses only a ratio of survival probabilities. The elapsed period \(s\) appears in both numerator and denominator, so it cancels. That cancellation is the mathematical source of the memoryless property; it is not a claim that the waiting time literally restarts or that an event-free period has not occurred.

Applying the Identity to a Poisson Arrival

For a direct arrival-time interpretation, let \((N_t)_{t\ge0}\) be a Poisson process on the probability space above, with continuous time index \(t\ge0\), state space \(S=\{0,1,2,\ldots\}\), and rate \(\lambda>0\). Define its first arrival time by \(T_1=\inf\{t\ge0:N_t\ge1\}\). The earlier result that the first arrival time is exponential gives \(\mathbb{P}(T_1>u)=e^{-\lambda u}\). The event of no arrivals by time \(s\), \(N_s=0\), is the event \(T_1>s\), up to an endpoint that has probability zero. Therefore, the theorem applies to the first arrival after observing an event-free period.

There is also a useful formula for probabilities over an interval of residual waiting times. If \(0\le a<b\), the residual wait is in \((a,b]\) exactly when the first arrival occurs after \(s+a\) and no later than \(s+b\), given survival through \(s\). Since the exponential distribution has no atoms, either choice of strictness at a finite endpoint gives the same probability.

Proof. The event \(a<T-s\le b\), together with \(T>s\), is the event \(s+a<T\le s+b\). Its probability is \(\mathbb{P}(T>s+a)-\mathbb{P}(T>s+b)\). Dividing by \(\mathbb{P}(T>s)=e^{-\lambda s}>0\) gives $$ \frac{e^{-\lambda(s+a)}-e^{-\lambda(s+b)}}{e^{-\lambda s}} =e^{-\lambda a}-e^{-\lambda b}. $$ The right side is nonnegative because \(a<b\) and \(\lambda>0\). This establishes the formula. \(\square\)

Worked Residual-Wait Calculations

Common Mistake: Leaving Out the Conditioning

A frequent error is to calculate the unconditional chance of surviving from time \(0\) all the way to time \(s+t\), then report it as the chance of surviving another \(t\) given survival to \(s\). In the first worked example, the unconditional probability of no arrival by \(\tfrac34\) hour is \(e^{-3/2}\). It is not the requested conditional probability. The conditioning event, no arrival by \(\tfrac12\) hour, has probability \(e^{-1}\), so the correct calculation divides:

The distinction is important in applications: a probability conditional on an event-free observation period describes the residual wait among cases known to have survived that period. It does not describe the original, unconditioned waiting time.

The memoryless identity is specific to exponential waiting times. It does not hold for every waiting-time distribution. For instance, if a wait is uniformly distributed on \([0,2]\), then conditional on surviving one time unit, the chance of surviving one further unit is \(0\), whereas the original chance of surviving one unit is \(\tfrac12\). The conditional and original survival probabilities differ. Thus, an event-free period can change residual waiting probabilities unless the exponential model applies.

For a positive-rate exponential wait, the practical rule is concise: when conditioning on survival through a fixed time \(s\), use the exponential survival formula for the additional time \(t\), not for the total time \(s+t\). For an interval of residual times, subtract the two residual survival probabilities. Always check that the conditioning event has positive probability; here it does because \(e^{-\lambda s}>0\) for finite \(s\).