Characterizing Exponential Waiting Times by Memorylessness

Memorylessness as a Condition on Survival

A waiting time is memoryless when knowing that it has already lasted \(s\) units does not change the distribution of the additional wait. The condition is strong: it relates survival probabilities at every pair of times, not just at one selected time. We will show that, for a proper nonnegative waiting time with a continuous survival function, this condition determines the exponential distribution.

Let \(T\) be a nonnegative, finite almost surely random variable on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). Its survival function is \(S(t)=\mathbb{P}(T>t)\), for \(t\ge0\). In this tutorial, continuity of the survival function means continuity on \([0,\infty)\), including at zero. A conditional probability given \(T>s\) is defined only if \(S(s)>0\).

At \(s=0\), the conditioning event is \(\{T>0\}\), so it must have positive probability for the definition to apply. The characterization below assumes \(\mathbb{P}(T>0)>0\). This excludes \(T=0\) almost surely, for which the conditional statement at time zero is not defined. Under the stated assumptions, memorylessness will in fact force \(\mathbb{P}(T>0)=1\).

The Characterization

Proof. We first prove that memorylessness implies the exponential form. Write \(S_0=S(0)>0\). Continuity at zero and \(S_0>0\) imply that \(S(t)>0\) for all sufficiently small \(t\ge0\). Choose one such \(t>0\). The memoryless identity with \(s=0\) gives \(S(t)=S(0)S(t)\). Since \(S(t)>0\), division by \(S(t)\) gives \(S_0=1\).

Consequently, the event \(T>0\) has probability one. Also, \(S(t)\) tends to zero as \(t\) tends to infinity: the events \(\{T>n\}\) decrease to \(\{T=\infty\}\), which has probability zero because \(T\) is finite almost surely. For every \(s\ge0\) with \(S(s)>0\), the conditional identity now yields \(S(s+t)=S(s)S(t)\).

We next show that \(S\) cannot vanish at a finite time. Suppose \(S(a)=0\) for some \(a>0\). Since \(S(0)=1\) and \(S\) is continuous at zero, there is a \(\delta>0\) such that \(S(x)>0\) whenever \(0\le x<\delta\). Choose an integer \(n\ge2\) large enough that \(x=a/n<\delta\). Then \(S(x)>0\). Repeatedly applying the memoryless identity gives \(S(2x)=S(x)^2>0\), and whenever \(S(kx)>0\), it gives \(S((k+1)x)=S(kx)S(x)>0\). Thus \(S(nx)>0\), contradicting \(S(a)=0\). Hence \(S(t)>0\) for every finite \(t\ge0\).

Define \(f(t)=-\log S(t)\). This is well-defined and continuous because \(S(t)>0\) and \(S\) is continuous. The survival identity gives \(f(s+t)=f(s)+f(t)\) for all \(s,t\ge0\). Set \(c=f(1)\). Additivity gives \(f(1/n)=c/n\) for every positive integer \(n\), since \(n f(1/n)=f(1)\). It then gives \(f(m/n)=mc/n\) for every nonnegative integer \(m\). For any \(t\ge0\), choose nonnegative rational numbers \(q_k\) tending to \(t\). Continuity implies \(f(t)=\lim_k f(q_k)=\lim_k c q_k=ct\). Therefore \(S(t)=e^{-ct}\).

Since \(S(t)\le1\), \(c\ge0\). If \(c=0\), then \(S(t)=1\) for every \(t\), contradicting \(S(t)\to0\) for a finite random variable. Thus \(c>0\); set \(\lambda=c\).

For the converse, suppose \(S(t)=e^{-\lambda t}\) with \(\lambda>0\). For every \(s\ge0\), \(S(s)=e^{-\lambda s}>0\), so the conditional probability is defined. Directly, \(\mathbb{P}(T>s+t\mid T>s)=S(s+t)/S(s)=e^{-\lambda(s+t)}/e^{-\lambda s}=e^{-\lambda t}=S(t)\). This proves memorylessness and completes both directions. \(\square\)

The proof also identifies why the assumptions matter. The condition at time zero rules out a positive mass at zero, and finiteness almost surely rules out the constant survival function \(S(t)=1\), which corresponds to an infinite waiting time rather than a proper exponential distribution. Continuity lets us take logarithms and determine the additive function from its values on rational times.

Identifying the Rate

Proof. Since \(t_0>0\) and \(\lambda>0\), \(r=e^{-\lambda t_0}\) satisfies \(0<r<1\). Taking logarithms gives \(\log(r)=-\lambda t_0\), and division by \(t_0\) yields the stated formula. If two positive rates give the same \(r\) at \(t_0\), this same equation gives the same rate for both, proving uniqueness. \(\square\)

The strict bounds on \(r\) are useful when checking data or a proposed model. A survival probability of zero or one at a strictly positive time cannot be produced by an exponential distribution with positive rate. The proposition identifies a rate from a single survival value; it does not by itself verify that an arbitrary waiting-time distribution is exponential. For that, the full survival curve must have the required form.

Direct Verification and a Discrete-Time Contrast

A Common Error and Why the Characterization Matters

Common mistake. Memorylessness does not say \(S(s+t)=S(t)\). That incorrect equation says the elapsed time \(s\) has no effect on the unconditional chance of surviving to \(s+t\), and it omits the probability of having survived the first \(s\) units. For an exponential variable with rate \(2\), take \(s=1\) and \(t=1\). The correct calculation is \(S(2)=e^{-4}=S(1)S(1)=e^{-2}e^{-2}\). The incorrect calculation would give \(S(2)=S(1)=e^{-2}\), which is false. In conditional form, the correct identity is \(S(s+t)/S(s)=S(t)\), and it requires \(S(s)>0\).

The characterization is useful because it turns a qualitative assumption about residual waiting into a precise distributional conclusion. If a model treats the residual wait as having the same law as a fresh wait at every elapsed time, then, under the theorem’s hypotheses, its survival probabilities must decay exponentially at a constant rate. Conversely, the direct ratio calculation shows exactly how an exponential model implements that assumption. A fitted exponential model should therefore be assessed not only by one survival probability but by whether the assumed memoryless description is appropriate over the time range of interest.

Keep the event convention in view: survival means \(T>t\), and conditioning on \(T>s\) is valid only when \(S(s)>0\). The exponential survival function is positive at every finite time, so its conditional probabilities are all well-defined. In the characterization proof, positivity at every finite time was established rather than assumed; this prevents taking logarithms of zero and is essential to the argument.