The exponential distribution is a continuous probability distribution used to describe the time until an event occurs at a constant rate. It is defined on nonnegative values and has a single positive parameter, usually written as a rate . Its defining feature is memorylessness: conditional on an event not yet having occurred, the distribution of the additional waiting time is the same as the original waiting-time distribution.
Exponential models connect event counts, waiting times, and random changes of state. They describe the intervals between events in a homogeneous Poisson process, the holding times of many continuous-time stochastic models, and idealized component lifetimes when failure risk does not change with age. Their simple mathematical structure supports exact statistical inference and efficient simulation, but the constant-rate assumption also imposes substantial restrictions on their use.
Definition, parameterizations, and basic quantities
A random variable has an exponential distribution with rate , written , if its probability density function is
The exponential function makes the density decrease continuously from its maximum at zero. Although the density is positive at zero, the probability of any individual value, including zero, is zero. The distribution is continuous, not a distribution with a positive probability mass at the origin.
The cumulative distribution function and survival function are
The survival function gives the probability that the waiting time exceeds a specified duration. It is often more convenient than the cumulative distribution function in reliability calculations because “still functioning at time ” directly corresponds to . For ,
An alternative parameterization uses the scale , giving
Here has the units of the measured duration, whereas has inverse-duration units. For example, a rate of events per minute corresponds to a mean waiting time of minutes. Confusing rate and scale reverses this relationship, so a specification such as “exponential with parameter 5” is incomplete unless the convention is stated. Software libraries differ in which parameterization they adopt.
The quantile function follows by solving :
Consequently, the median is , approximately times the mean. The 95th percentile is approximately . About of observations exceed the mean, since . These values illustrate the distribution’s pronounced right asymmetry: most observations lie below the mean, but a relatively long right tail pulls the mean upward.
Multiplying by a positive constant changes its rate according to
Thus a change from minutes to seconds changes the numerical rate but not the underlying probabilistic model. Adding a fixed delay instead produces a shifted exponential distribution with a lower endpoint above zero; that model is distinct from the ordinary one-parameter exponential distribution.
Moments, transforms, and variability
The expected value and variance are
The standard deviation therefore equals the mean. Their ratio, known as the coefficient of variation, is exactly one. This is a strong structural restriction: changing the rate rescales the observations but cannot independently alter their relative variability. Data with a standard deviation much smaller or larger than their mean may require a different model, although sampling variation and incomplete observation also affect empirical comparisons.
More generally, for any real ,
where is the gamma function. For a nonnegative integer , this becomes
The formula follows by substituting into the defining integral for the expectation. All nonnegative integer moments are finite. Negative moments require care: , for example, diverges because the density remains positive arbitrarily close to zero.
The moment-generating function is
Its domain matters: the expectation is infinite when . The logarithm of this expression is the cumulant-generating function,
Expanding it around zero gives the th cumulant
The standardized third central moment, or skewness, is , and the excess kurtosis is . Neither depends on the rate because all members of the family have the same shape after rescaling.
The Laplace transform takes a particularly useful form:
For independent nonnegative random variables, the transform of their sum is the product of their transforms. This observation gives a direct route from exponential waiting times to gamma and related distributions. The characteristic function, which always exists for real , is
The exponential tail decreases faster than any inverse power of , but more slowly than a Gaussian tail of the form . It is therefore not heavy-tailed in the usual sense requiring the moment-generating function to be infinite at every positive argument. Nevertheless, unusually long waits remain appreciably more common than in many tightly concentrated lifetime models. The identity makes the scale of such tail events explicit.
Memorylessness and constant hazard
The exponential distribution satisfies
Using conditional probability, the left side becomes
Thus surviving for time does not change the distribution of the additional waiting time. If a component’s lifetime is exponential with a mean of ten hours, then, conditional on surviving its first seven hours, its expected remaining lifetime is still ten hours. Its expected total lifetime under that condition is seventeen hours.
Under the usual nondegeneracy assumptions, memorylessness characterizes the exponential distribution among nonnegative, proper continuous distributions. If and the memoryless identity holds, then
Setting gives the additive equation . The monotonicity inherited from a survival function rules out pathological additive solutions and yields . Hence , with for a nondegenerate proper distribution.
The same property can be expressed through the hazard function,
For small ,
The hazard is an instantaneous conditional rate, not a probability. It has inverse-time units and may numerically exceed one. A rate of per hour does not mean a probability of ; the corresponding one-hour event probability is .
For a general absolutely continuous nonnegative lifetime,
A constant hazard therefore gives an exponential survival function. The cumulative hazard,
is the straight line . Exponential survival curves become linear when plotted on a logarithmic survival scale, making this representation useful for examining constant-rate assumptions.
Memorylessness should not be confused with statistical independence between different lifetimes. Several variables can each have exponential marginal distributions while being dependent. Nor does memorylessness mean that elapsed time is irrelevant to every question: the probability of surviving another five hours is unchanged by age, but the probability of having survived the first five hours is not one. The property concerns the conditional distribution of residual life, not the removal of information from all aspects of a model.
Poisson processes and continuous-time dynamics
The most direct process interpretation arises from a homogeneous Poisson process with rate . Its count over an interval of length has the Poisson distribution with mean , and counts over disjoint intervals are independent. If is the first event time, then
Thus is exponential. Successive interarrival times are independent exponentials with the same rate, and conversely, accumulating such waiting times constructs a homogeneous Poisson process.
This equivalence distinguishes a count distribution from a waiting-time distribution. A rate of three events per hour gives a Poisson count with mean three over one hour and an exponential interarrival time with mean one-third of an hour. The probability of no event in twenty minutes is , while the probability of exactly one event in that interval is also . These equal values arise from different events and different probability laws.
The relationship changes when the event rate depends on calendar time. For a nonhomogeneous Poisson process with deterministic intensity , the waiting time from time satisfies
This need not be exponential in ordinary time. Measuring intervals by their accumulated intensity can recover standard exponential waiting times, but a fixed-rate exponential model generally cannot represent predictable daily or seasonal rate variation.
Exponential holding times also underlie a continuous-time Markov chain. If the chain is in state , and is the transition rate to state , its total exit rate is
For , the time before leaving state is exponential with rate . The next state is with probability , independently of that holding time. If , the state is absorbing rather than having a proper finite exponential holding time.
The memoryless property explains why the current state can contain sufficient information about future evolution: the model need not also record how long the chain has occupied that state. By contrast, nonexponential holding times generally require an elapsed-time variable or a broader process framework. The Markov property concerns conditional dependence on the present state; exponential holding times are its characteristic realization for time-homogeneous, pure-jump models with discrete states and finite exit rates.
Competing clocks and order statistics
Suppose are independent exponential variables with respective rates . Their minimum
is exponential with rate
Indeed,
This is the basic mathematics of competing exponential clocks: adding possible independent event mechanisms increases the total event rate.
The probability that clock rings first is
More strongly, the identity of the first clock and the time of its occurrence are independent. Their joint density is
which factors into the density of and the probability . Ties have probability zero. This factorization supports event-driven simulation: draw one waiting time using the total rate, then select the event type in proportion to its rate.
In reliability theory, a series system fails when its first component fails. If component lifetimes are independent exponentials, the system lifetime is therefore exponential with rate equal to the sum of component failure rates. A parallel system, which survives until all components have failed, involves the maximum instead. For independent components with a common rate ,
Except when , this is not an exponential distribution.
For an independent sample of size from a common exponential distribution, let
denote the order statistics. Define and spacings . The spacings are independent, with
Initially, clocks compete. After the first event, memorylessness leaves independent residual exponential clocks; the argument then repeats.
Consequently,
The expected minimum is , whereas the expected maximum is , where . The latter grows approximately as , with Euler’s constant. More components therefore shorten the time to the first failure but lengthen the time until every component has failed. The independence assumption is essential: shared environmental shocks or other dependencies can invalidate both the summed-rate rule and the spacing representation.
Relationships to other probability distributions
The exponential distribution is the shape-one case of the gamma distribution. In rate parameterization, a gamma variable with shape has density
Setting gives the exponential density. If are independent exponentials with the same rate, their sum is gamma:
For integer shape this is also called an Erlang distribution and represents the time to the th event of a homogeneous Poisson process.
Equal rates are important. A sum of independent exponentials with different rates generally has a hypoexponential distribution, not a gamma distribution. A different construction selects one exponential rate at random and then draws a single waiting time. This produces a mixture of exponentials rather than a sum. Such mixtures can have decreasing hazards because long-surviving observations increasingly come from their lower-rate components. Neither arbitrary sums nor arbitrary mixtures preserve memorylessness.
The chi-squared distribution is another gamma special case. If , then
For an independent sample,
This exact identity is central to confidence intervals and tests for the exponential rate. It also shows why a chi-squared distribution with two degrees of freedom has a decreasing density, unlike the more centrally peaked shapes obtained at higher degrees of freedom.
The geometric distribution is the discrete analogue of the exponential distribution. Under the convention that it counts trials until the first success, its survival probabilities satisfy
It is memoryless on the nonnegative integer time scale. If trials occur every units of time and , then the scaled waiting times converge in distribution to an exponential variable as tends to zero.
The Weibull distribution generalizes exponential survival to
Its hazard is
The case is exponential; gives increasing hazard, and gives decreasing hazard.
A further relationship comes from conditioning independent exponential variables on their sum. If and all rates are equal, the proportions are independent of and uniformly distributed over the probability simplex. For two variables, is uniform on . These normalized representations are useful in simulation and in constructing more general distributions on proportions.
Estimation and exact frequentist inference
For independent observations from an exponential distribution, the likelihood function is
Its logarithm is . Differentiation gives the maximum likelihood estimator
The maximum likelihood estimator of the scale is simply the sample mean . The likelihood depends on the sample only through , making a sufficient statistic for the rate.
The rate estimator is not unbiased. Since ,
An unbiased estimator is
For ,
The restrictions on sample size reflect divergent inverse moments of the gamma distribution. In particular, the maximum likelihood rate estimator has infinite expectation when , despite being finite with probability one. This distinguishes the behavior of an estimator on typical samples from the existence of its theoretical moments.
The Fisher information in complete observations is . Under large-sample approximation,
An approximate standard error is therefore . Intervals based directly on this normal approximation can extend below zero in small samples, whereas exact intervals respect the positive parameter space.
Let denote the quantile of a chi-squared distribution with degrees of freedom. The pivotal quantity gives an exact confidence interval:
Taking reciprocals and reversing the endpoints gives the corresponding interval for the mean lifetime.
For example, suppose ten complete waiting times total fifty hours. The estimated rate is per hour and the estimated mean is five hours. Using the approximate 2.5th and 97.5th percentiles and of , the exact 95% rate interval is approximately per hour. The corresponding mean interval is approximately hours. These intervals quantify uncertainty under the exponential sampling model; they do not establish that its constant-hazard assumption is correct.
Censoring and Bayesian analysis
Lifetime observations are often incomplete. Under right censoring, observation ends before some events occur. For subject or component , let
where is a censoring time, and let if the event is observed and otherwise. Under independent, noninformative censoring, the contribution involving is
Thus the combined likelihood has the form
The maximum likelihood rate estimate is when : observed events divided by total time at risk. Each censored observation contributes exposure even though it contributes no event. Ignoring censored observations discards this information and generally distorts inference. If no events are observed, the likelihood decreases as increases; its supremum occurs as approaches the boundary value zero rather than at a positive interior estimate.
The likelihood’s simple form does not make all censoring schemes equivalent. The sampling distribution of and depends on how observation stops. Under fixed-time censoring, the number of observed failures is random; under a design that stops at a specified failure count, that count is fixed. Exact procedures must account for these distinctions rather than automatically reusing complete-sample formulas. Left truncation, where subjects enter observation only after surviving to a given age, also requires conditioning on entry.
In Bayesian inference, the gamma distribution is a conjugate prior for the exponential rate. With prior shape and rate ,
The censored-data likelihood gives the posterior distribution
Complete observations are the special case . The posterior mean rate is , while posterior quantiles provide a credible interval for the rate.
Integrating over the posterior produces a different distribution for a future waiting time . If and , then its posterior predictive survival function is
This is a Lomax, or Pareto type II, survival function, not an exponential one. Its hazard is , which decreases with elapsed time. Uncertainty about a fixed underlying rate therefore produces predictive behavior different from that obtained by substituting a single estimated rate. The predictive mean is when , and is infinite otherwise.
Simulation, information, and model assessment
The simplest simulation method uses inverse transformation. If is uniform on , then
has the exponential distribution. Equivalently, . Random-number implementations must handle endpoints carefully: is undefined, and a generator that can return zero requires a compatible convention. When evaluating for small , a specialized log1p operation improves numerical stability. Similarly, can be computed accurately near zero using expm1.
For independent exponential event clocks, simulation can use the summed-rate construction rather than generating every clock separately. With event rates , draw a waiting time with rate , then select event with probability . Rates are updated after the event. This is the basis of the direct Gillespie algorithm for continuous-time reaction models. Its exactness is relative to a jump-process model whose propensities remain constant between jumps; explicitly time-dependent propensities require modified treatment.
The distribution also has a fundamental role in information theory. Among absolutely continuous distributions on with a specified finite mean , the exponential distribution maximizes differential entropy. With natural logarithms,
For any competing density with the same mean and suitable integrability, nonnegativity of the Kullback–Leibler divergence from to the exponential density gives
Differential entropy depends on the measurement scale, so the numerical value presumes a fixed unit. This maximum-entropy result identifies a distribution under stated constraints; it does not demonstrate that an observed event mechanism has constant hazard.
Model assessment focuses on those structural assumptions. An empirical survival curve should be approximately linear after taking its negative logarithm if a constant hazard is plausible. Quantile comparisons and goodness-of-fit tests can detect departures, but test calibration must account for estimating the rate from the same data. A coefficient of variation near one is compatible with an exponential model but does not uniquely identify it. Independence between successive waiting times must also be examined separately from their marginal distribution.
Applications include idealized radioactive-decay waiting times, queueing models with exponential service, and reliability models during approximately constant-hazard operating periods. Aging, early failures, minimum service durations, clustering, and persistent variation between individuals can each undermine the model. An M/M/1 queue, for example, assumes Poisson arrivals and independent exponential service times; its notation encodes substantive assumptions rather than a consequence of knowing average arrival and service rates. A fitted mean alone cannot establish either exponential waiting times or a Poisson event process.