Memorylessness of the Exponential Distribution

What Memorylessness Says

A waiting time is memoryless when learning that it has already lasted a given amount of time does not change the distribution of the additional wait. For a lifetime $X$, this is a statement about conditional probabilities: after surviving to time $s$, the chance of surviving a further $t$ units is the same as the original chance of surviving $t$ units.

The exponential distribution has this property because its survival function decays at a constant multiplicative rate over equal-length time intervals. The conditional-probability calculation is short, but the conditions matter: the conditioning event must have positive probability, and the time origin matters. A shifted exponential distribution, for example, does not have ordinary memorylessness.

Definition: A nonnegative random variable $X$ is memoryless if, for every $s,t\geq 0$ such that $P(X>s)>0$, $$ P(X>s+t\mid X>s)=P(X>t). $$ The left side is the conditional probability of lasting at least $t$ further units, given survival beyond $s$. The right side is the probability of lasting at least $t$ units from the original time origin.

Write $S(x)=P(X>x)$ for the survival function. Since $s+t\geq s$, the event $\{X>s+t\}$ is contained in $\{X>s\}$. Thus, whenever $S(s)>0$, the definition of conditional probability gives

P(X>s+t\mid X>s)=\frac{S(s+t)}{S(s)}.

Consequently, memorylessness is equivalent to the multiplicative survival identity $S(s+t)=S(s)S(t)$. This identity is about the survival function, not the density. It says that survival over two consecutive time intervals factors into the survival probability for each interval.

Why the Exponential Distribution Is Memoryless

Let $X\sim\operatorname{Exponential}(\lambda)$, with rate $\lambda>0$. Its survival function is $S(x)=e^{-\lambda x}$ for $x\geq0$. This is the survival formula for the exponential law; here it is used directly to evaluate the conditional probability.

Theorem: If $X\sim\operatorname{Exponential}(\lambda)$ with $\lambda>0$, then for every $s,t\geq0$, $$ P(X>s+t\mid X>s)=P(X>t)=e^{-\lambda t}. $$

Proof. The exponential survival probability is positive for every finite $s\geq0$, so the conditional probability is defined. Apply the survival-function ratio:

\begin{aligned} P(X>s+t\mid X>s) &=\frac{P(X>s+t)}{P(X>s)}\\ &=\frac{e^{-\lambda(s+t)}}{e^{-\lambda s}}\\ &=e^{-\lambda t}\\ &=P(X>t). \end{aligned}

The elapsed time $s$ cancels from the ratio. This proves the identity for all $s,t\geq0$. $\square$

Worked Example: Conditional Survival After Three Time Units

Suppose $X\sim\operatorname{Exponential}(0.4)$, where time is measured in hours. Find the probability that the wait lasts at least two more hours, given that it has already lasted more than three hours.

The event in the numerator is $X>5$, while the conditioning event is $X>3$. The conditional probability is therefore

\begin{aligned} P(X>5\mid X>3) &=\frac{P(X>5)}{P(X>3)}\\ &=\frac{e^{-0.4(5)}}{e^{-0.4(3)}}\\ &=\frac{e^{-2}}{e^{-1.2}}\\ &=e^{-0.8}\approx 0.4493. \end{aligned}

Without conditioning on the first three hours, the probability of a wait longer than two hours is $P(X>2)=e^{-0.4(2)}=e^{-0.8}\approx0.4493$. The two probabilities agree, as memorylessness requires.

Interpreting the Remaining Waiting Time

The conditional probability statement can also be expressed using the remaining time after $s$. Given $X>s$, define the residual waiting time to be $X-s$. For $t\geq0$, the event $X-s>t$ is exactly the event $X>s+t$. Thus, for an exponential $X$,

P(X-s>t\mid X>s)=P(X>t).

This equality says that the residual wait, conditional on having waited $s$ units, has the same survival probabilities as a fresh exponential waiting time. It does not say that the actual elapsed time disappears, or that the past event $X>s$ becomes impossible. Rather, it says that knowing this event does not alter probabilities about how much longer the wait will last.

Worked Example: A Remaining Wait After Five Hours

A component lifetime $X$ is exponential with rate $\lambda=0.2$ per hour. The component has already lasted more than five hours. Find the conditional probability that it lasts more than four additional hours.

The event “more than four additional hours” is $X>5+4=9$. Directly using the conditional probability formula,

\begin{aligned} P(X>9\mid X>5) &=\frac{P(X>9)}{P(X>5)}\\ &=\frac{e^{-0.2(9)}}{e^{-0.2(5)}}\\ &=\frac{e^{-1.8}}{e^{-1}}\\ &=e^{-0.8}\approx0.4493. \end{aligned}

A new component with the same exponential lifetime has probability $P(X>4)=e^{-0.2(4)}=e^{-0.8}\approx0.4493$ of lasting more than four hours. The conditional and unconditional probabilities match. This concerns the distribution of additional lifetime; it does not imply that the component’s age is five hours or that its total lifetime has been reset.

A Characterization of the Continuous Memoryless Law

The survival identity also explains why the exponential distribution is the continuous memoryless law on nonnegative time. The following statement assumes a proper continuous lifetime: $X$ is finite with probability one, its survival function $S$ is continuous, and $S(t)>0$ for all finite $t\geq0$. Positivity allows us to take logarithms; properness ensures $S(t)$ tends to zero as $t$ tends to infinity.

Theorem: Suppose $X$ is a proper nonnegative random variable with continuous survival function $S$, where $S(t)>0$ for every finite $t\geq0$. If $X$ is memoryless, then there is a rate $\lambda>0$ such that $$ S(t)=e^{-\lambda t}\qquad(t\geq0). $$

Proof. Memorylessness and the survival-function ratio imply $S(s+t)=S(s)S(t)$ for all $s,t\geq0$. Define $g(t)=-\log S(t)$. Continuity and positivity of $S$ imply that $g$ is continuous, and the multiplicative identity becomes

$$ g(s+t)=g(s)+g(t). $$

We now determine $g$. For any positive integer $n$, repeated addition gives $g(n x)=n g(x)$. In particular, for positive integers $m,n$,

g(m)=m g(1),\qquad n g\!\left(\frac{m}{n}\right)=g(m)=m g(1),

so $g(m/n)=(m/n)g(1)$ for every nonnegative rational number $m/n$. For any real $t\geq0$, choose a sequence of nonnegative rational numbers $q_k$ with $q_k\to t$. Continuity then gives

g(t)=\lim_{k\to\infty}g(q_k) =\lim_{k\to\infty}q_k g(1) =tg(1).

Set $\lambda=g(1)$. Since a survival function is nonincreasing, $g$ is nondecreasing and $\lambda\geq0$. Properness gives $S(t)\to0$, so $g(t)=-\log S(t)\to\infty$. Therefore $\lambda\neq0$, and hence $\lambda>0$. It follows that $S(t)=e^{-g(t)}=e^{-\lambda t}$, as claimed. $\square$

This characterization applies to continuous lifetimes under the stated hypotheses. It does not claim that every memoryless random variable must be exponential without regard to its support or whether time is continuous. In particular, geometric waiting times are the corresponding discrete-time example.

Why a Shifted Exponential Is Not Memoryless

A shift changes the time origin that appears in the memoryless identity. Let $Y\sim\operatorname{Exponential}(\lambda)$ and set $X=c+Y$, where $c>0$ is fixed. Then $X\geq c$, and its survival function is

S_X(x)= \begin{cases} 1, & 0\leq x<c,\\ e^{-\lambda(x-c)}, & x\geq c. \end{cases}

The survival ratio $S_X(s+t)/S_X(s)$ does not in general equal $S_X(t)$. In particular, before time $c$, survival is certain, whereas the ordinary memoryless identity compares the conditional residual wait to a new wait measured from zero.

Worked Example: A Shifted Exponential Fails the Identity

Let $X=2+Y$, where $Y\sim\operatorname{Exponential}(0.5)$. Take $s=3$ and $t=1$. Since both $3$ and $4$ exceed the shift $2$, the survival function gives

S_X(3)=e^{-0.5(3-2)}=e^{-0.5}, \qquad S_X(4)=e^{-0.5(4-2)}=e^{-1}.

Consequently,

P(X>4\mid X>3) =\frac{S_X(4)}{S_X(3)} =\frac{e^{-1}}{e^{-0.5}} =e^{-0.5}\approx0.6065.

But $1<2$, so $X>1$ with probability one and $P(X>1)=1$. Thus $P(X>3+1\mid X>3)\neq P(X>1)$, and $X$ is not memoryless in the ordinary sense. Although $X$ is a shifted exponential, the shift prevents its survival function from having the required multiplicative form from time zero.

Common Mistake and Key Takeaway

A common mistake is to replace the conditional probability $P(X>s+t\mid X>s)$ with the unconditional probability $P(X>s+t)$, then compare that quantity with $P(X>t)$. This drops the conditioning event. For an exponential variable, $P(X>s+t)=e^{-\lambda(s+t)}$, which generally does not equal $e^{-\lambda t}$; the equality appears only after dividing by $P(X>s)=e^{-\lambda s}$.

A second caution is that “the wait does not age” is shorthand for a precise equality of conditional distributions, not a claim that all real waiting processes behave this way. The property follows from the exponential survival function, and a shift or a different lifetime model can invalidate it.

Key takeaway: For $X\sim\operatorname{Exponential}(\lambda)$, the conditional survival probability is the survival ratio $S(s+t)/S(s)=e^{-\lambda t}$. The elapsed time cancels, yielding memorylessness. A shifted exponential generally does not satisfy this identity.

Check Your Understanding

Use the survival-function ratio where appropriate.

State the memoryless identity for a nonnegative random variable $X$, including the condition needed for the conditional probability to be defined.
If $X\sim\operatorname{Exponential}(0.3)$, calculate $P(X>7\mid X>2)$ and compare it with $P(X>5)$.
Explain why $P(X>s+t)$ is not the conditional probability $P(X>s+t\mid X>s)$.
For $X=c+Y$, where $c>0$ and $Y$ is exponential, describe how the support shift affects the ordinary memoryless identity.
In the continuous-lifetime characterization, why must the rate $\lambda$ be positive rather than zero?