The Uniform and Exponential Distributions

How long will you wait for the next bus if arrivals are "completely random"? How do you model a number chosen with no preference across an interval? These two questions lead to the two most important elementary continuous distributions. The uniform distribution spreads probability evenly across an interval. The exponential distribution models waiting times for events that occur at a constant average rate with no memory of the past. Together they form the bridge between discrete models and continuous machinery. The discrete models are built from equally likely outcomes and geometric waiting times. The continuous machinery includes densities, integrals, and transformations, and it dominates the rest of probability. The uniform distribution is the continuous analogue of "all outcomes equally likely." It is the continuous sibling of the discrete uniform distribution on a finite set. The exponential distribution is the continuous analogue of the The Geometric Distribution. Just as the geometric distribution counts the number of discrete trials until a first success, the exponential distribution measures the continuous time until a first arrival. This arrival occurs in a Poisson-type process. Both distributions are simple enough to compute with exactly. They have closed-form CDFs and closed-form moments. Yet they are rich enough to illustrate every major technique in the continuous toolkit. These techniques include integration against a density, moment generating functions, order statistics, and change of variables. The centerpiece of this article is the memorylessness property of the exponential distribution. Conditional on having survived past time ss, the remaining waiting time is exponential all over again. It is exactly as if the clock had been reset. This is not a curiosity — it is a characterization. Among continuous distributions on [0,∞)[0,\infty), memorylessness holds if and only if the distribution is exponential. This rigidity explains why the exponential distribution is the unique continuous distribution. It appears as the limit of scaled geometric waiting times. It is also why it is the canonical model for radioactive decay, component lifetimes, and interarrival times in Poisson processes. This article gives the precise definitions of both densities. It derives their moments and moment generating functions from scratch. It proves the memorylessness theorem in both directions. It shows, via a genuine counterexample, why memorylessness fails outside the exponential family. It closes by pointing toward the The Normal Distribution and the general machinery of Functions of Random Variables. Both of these build directly on the computations developed here.

Key takeaways
  • The uniform distribution U(a,b)U(a,b) has constant density 1b−a\frac{1}{b-a} on [a,b][a,b]: every subinterval's probability is proportional to its length.
  • The exponential distribution with rate λ>0\lambda>0 has density λe−λx\lambda e^{-\lambda x} on [0,∞)[0,\infty) and models waiting times for rare, constant-rate events.
  • Both distributions have closed-form CDFs, means, variances, and moment generating functions computable by direct integration.
  • The exponential distribution is the continuous limit of a rescaled The Geometric Distribution as the number of trials per unit time grows without bound.
  • Memorylessness: P(X>s+t∣X>s)=P(X>t)P(X > s+t \mid X > s) = P(X>t) for an exponential XX and all s,t≥0s,t \ge 0 — the future is independent of the past, given survival.
  • Memorylessness is a characterization: it holds among continuous positive random variables only for the exponential family, a fact provable via Cauchy's functional equation.
  • Key idea: the exponential distribution is the unique "ageless" continuous waiting-time law, and this rigidity is what makes it computationally and structurally central to all of probability.

Where this lives today

The uniform distribution is the computational bedrock of simulation: essentially every pseudo-random number generator produces (an approximation to) Unif(0,1)\mathrm{Unif}(0,1) samples, and the inverse transform method converts these into samples from any target distribution FF via X=F−1(U)X = F^{-1}(U) — a technique whose correctness for continuous FF is a direct consequence of the density computations developed here, and which is the standard bridge toward simulating the The Normal Distribution and other laws studied under Functions of Random Variables. Uniform priors also appear throughout Bayesian statistics as the "no-information" prior on a bounded parameter range, and uniform distributions on discretized intervals underlie hashing and load-balancing algorithms in computer science. The exponential distribution is the default model for inter-arrival times in Poisson processes (customer arrivals, radioactive decay, photon detection, network packet arrivals), for component lifetimes in reliability engineering when failure is "memoryless" (electronic components rather than mechanically wearing parts), and for holding times in queueing theory (the "M" in Kendall's M/M/1 notation stands for memoryless, i.e. exponential). In cryptography, exponential waiting-time assumptions underlie models of adversarial timing attacks; in physics, the exponential decay law for unstable particles is a direct restatement of the memorylessness theorem proved above (a nucleus's probability of decaying in the next instant does not depend on its age). In finance, exponential distributions model time-to-default and time-to-claim in insurance and credit risk. Both distributions, being the simplest non-trivial continuous laws with closed-form CDFs, also serve as the pedagogical proving ground for every later technique — transformations, order statistics, moment generating functions — that culminates in the The Normal Distribution as a limit of sums, via the central limit theorem.

:::quiz[For X∼Exp(λ)X \sim \mathrm{Exp}(\lambda), which quantity always equals the mean E[X]E[X]?]

  • The variance Var(X)\mathrm{Var}(X)
  • The standard deviation Var(X)\sqrt{\mathrm{Var}(X)} ✓
  • The median of XX
  • MX(1)M_X(1)

Ask the tutor: How does the memorylessness of the exponential distribution generalize to the "lack of aftereffect" property defining a continuous-time Markov process?

DefinitionUniform and exponential densities

A continuous random variable XX is uniformly distributed on [a,b][a,b], written X∼Unif(a,b)X \sim \mathrm{Unif}(a,b), if it has probability density function

fX(x)={1b−aa≤x≤b,0otherwise.f_X(x) = \begin{cases} \dfrac{1}{b-a} & a \le x \le b, \\[4pt] 0 & \text{otherwise.} \end{cases}

A continuous random variable XX is exponentially distributed with rate λ>0\lambda > 0, written X∼Exp(λ)X \sim \mathrm{Exp}(\lambda), if it has probability density function

fX(x)={λe−λxx≥0,0x<0.f_X(x) = \begin{cases} \lambda e^{-\lambda x} & x \ge 0, \\[2pt] 0 & x < 0. \end{cases}

Both are genuine densities: fX≥0f_X \ge 0 everywhere, and in each case ∫−∞∞fX(x) dx=1\int_{-\infty}^{\infty} f_X(x)\,dx = 1, the latter following since ∫ab1b−a dx=1\int_a^b \frac{1}{b-a}\,dx = 1 and ∫0∞λe−λx dx=[−e−λx]0∞=1\int_0^\infty \lambda e^{-\lambda x}\,dx = \left[-e^{-\lambda x}\right]_0^\infty = 1.

Intuition

Picture the uniform density as a flat plateau of height 1b−a\frac{1}{b-a} sitting over the interval [a,b][a,b] and zero elsewhere — literally "spread the unit of probability mass evenly." The probability that XX lands in any subinterval [c,d]⊆[a,b][c,d]\subseteq[a,b] is exactly d−cb−a\frac{d-c}{b-a}, proportional to length and nothing else: no point or sub-region is favored. This is the natural continuous generalization of picking a uniformly random element from a finite set, and it is the model of choice whenever a quantity is known only to lie in a range, with no further information distinguishing values within that range. The exponential density, by contrast, is a decaying curve: tallest at x=0x=0 with height λ\lambda, falling off geometrically in the continuous sense. Think of λ\lambda as a rate — the average number of arrivals per unit time in a Poisson-type process — so that 1/λ1/\lambda is the mean waiting time. The defining feature, made precise below, is that the exponential "forgets": having already waited ss minutes for the bus without success tells you nothing about how much longer you'll wait, a fact utterly false for, say, a uniformly distributed waiting time on [0,10][0,10], where waiting 9 minutes already tells you the bus is imminent.

:::svg[Uniform density on [a,b]=[2,7][a,b]=[2,7]: constant height 1/(b−a)=0.21/(b-a)=0.2, area under any subinterval proportional to its width] x f(x) a = 2 b = 7 1/(b−a) = 0.2 P(4≤X≤6)=0.4

Figure
Exponential density with rate λ=0.5, mean 1/λ=2: memorylessness illustrated by the shifted tail matching the original curve
x f(x) s = 3 tail beyond s, rescaled = same shape as f(t) λ = 0.5
TheoremMoments, MGF, and memorylessness

Part A (Uniform moments). Let X∼Unif(a,b)X \sim \mathrm{Unif}(a,b). Then

E[X]=a+b2,Var(X)=(b−a)212.E[X] = \frac{a+b}{2}, \qquad \mathrm{Var}(X) = \frac{(b-a)^2}{12}.

Part B (Exponential moments and MGF). Let X∼Exp(λ)X \sim \mathrm{Exp}(\lambda), λ>0\lambda > 0. Then for t<λt < \lambda,

MX(t)=E[etX]=λλ−t,M_X(t) = E[e^{tX}] = \frac{\lambda}{\lambda - t},

and consequently

E[X]=1λ,Var(X)=1λ2.E[X] = \frac{1}{\lambda}, \qquad \mathrm{Var}(X) = \frac{1}{\lambda^2}.

Part C (Memorylessness). Let X∼Exp(λ)X \sim \mathrm{Exp}(\lambda). Then for all s,t≥0s,t \ge 0,

P(X>s+t∣X>s)=P(X>t).P(X > s+t \mid X > s) = P(X > t).

Conversely, suppose XX is a continuous random variable with P(X>x)>0P(X>x) > 0 for all x≥0x \ge 0, and suppose XX satisfies

P(X>s+t∣X>s)=P(X>t)for all s,t≥0.P(X > s+t \mid X > s) = P(X > t) \quad \text{for all } s,t \ge 0.

Then X∼Exp(λ)X \sim \mathrm{Exp}(\lambda) for some λ>0\lambda > 0.

Proof

Part A. By definition of expectation for a continuous random variable,

E[X]=∫abx⋅1b−a dx=1b−a[x22]ab=b2−a22(b−a)=(b−a)(b+a)2(b−a)=a+b2.E[X] = \int_a^b x \cdot \frac{1}{b-a}\,dx = \frac{1}{b-a}\left[\frac{x^2}{2}\right]_a^b = \frac{b^2 - a^2}{2(b-a)} = \frac{(b-a)(b+a)}{2(b-a)} = \frac{a+b}{2}.

For the second moment,

E[X2]=∫abx2⋅1b−a dx=1b−a[x33]ab=b3−a33(b−a)=a2+ab+b23,E[X^2] = \int_a^b x^2 \cdot \frac{1}{b-a}\,dx = \frac{1}{b-a}\left[\frac{x^3}{3}\right]_a^b = \frac{b^3-a^3}{3(b-a)} = \frac{a^2+ab+b^2}{3},

using the factorization b3−a3=(b−a)(a2+ab+b2)b^3 - a^3 = (b-a)(a^2+ab+b^2). Then

Var(X)=E[X2]−(E[X])2=a2+ab+b23−(a+b)24.\mathrm{Var}(X) = E[X^2] - (E[X])^2 = \frac{a^2+ab+b^2}{3} - \frac{(a+b)^2}{4}.

Putting both terms over denominator 1212:

Var(X)=4(a2+ab+b2)−3(a+b)212=4a2+4ab+4b2−3a2−6ab−3b212=a2−2ab+b212=(b−a)212.\mathrm{Var}(X) = \frac{4(a^2+ab+b^2) - 3(a+b)^2}{12} = \frac{4a^2+4ab+4b^2 - 3a^2 - 6ab - 3b^2}{12} = \frac{a^2 - 2ab + b^2}{12} = \frac{(b-a)^2}{12}.

Part B. For t<λt < \lambda,

MX(t)=∫0∞etxλe−λx dx=λ∫0∞e−(λ−t)x dx=λ⋅1λ−t=λλ−t,M_X(t) = \int_0^\infty e^{tx} \lambda e^{-\lambda x}\,dx = \lambda \int_0^\infty e^{-(\lambda - t)x}\,dx = \lambda \cdot \frac{1}{\lambda - t} = \frac{\lambda}{\lambda-t},

where the integral converges precisely because λ−t>0\lambda - t > 0, and evaluates via ∫0∞e−cx dx=1/c\int_0^\infty e^{-cx}\,dx = 1/c for c>0c>0. (For t≥λt \ge \lambda the integral diverges, so the MGF is defined only on (−∞,λ)(-\infty,\lambda); this is enough to generate all moments by differentiating at t=0t=0.) Differentiating, MX′(t)=λ(λ−t)−2M_X'(t) = \lambda (\lambda - t)^{-2}, so E[X]=MX′(0)=λ/λ2=1/λE[X] = M_X'(0) = \lambda/\lambda^2 = 1/\lambda. Differentiating again, MX′′(t)=2λ(λ−t)−3M_X''(t) = 2\lambda(\lambda-t)^{-3}, so E[X2]=MX′′(0)=2λ/λ3=2/λ2E[X^2] = M_X''(0) = 2\lambda/\lambda^3 = 2/\lambda^2. Hence

Var(X)=E[X2]−(E[X])2=2λ2−1λ2=1λ2.\mathrm{Var}(X) = E[X^2] - (E[X])^2 = \frac{2}{\lambda^2} - \frac{1}{\lambda^2} = \frac{1}{\lambda^2}.

(Alternatively one may compute these moments directly via integration by parts on ∫0∞xnλe−λx dx\int_0^\infty x^n \lambda e^{-\lambda x}\,dx; the MGF route is given here because it will recur for Functions of Random Variables.)

Part C, forward direction. First compute the survival function (tail probability) of X∼Exp(λ)X\sim\mathrm{Exp}(\lambda):

P(X>x)=∫x∞λe−λu du=e−λx,x≥0.P(X > x) = \int_x^\infty \lambda e^{-\lambda u}\,du = e^{-\lambda x}, \qquad x \ge 0.

Now fix s,t≥0s,t \ge 0. By the definition of conditional probability, since {X>s+t}⊆{X>s}\{X > s+t\} \subseteq \{X>s\} (as t≥0t\ge0), the event {X>s+t}∩{X>s}\{X>s+t\}\cap\{X>s\} equals {X>s+t}\{X>s+t\}, so

P(X>s+t∣X>s)=P(X>s+t)P(X>s)=e−λ(s+t)e−λs=e−λt=P(X>t).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).

This proves the forward direction; note P(X>s)=e−λs>0P(X>s) = e^{-\lambda s} > 0 for all finite ss, so the conditioning is always well defined.

Part C, converse direction. Let Fˉ(x):=P(X>x)\bar F(x) := P(X>x) for x≥0x \ge 0; by hypothesis Fˉ(x)>0\bar F(x) > 0 for all x≥0x \ge 0. The memorylessness hypothesis states

P(X>s+t∣X>s)=P(X>s+t)P(X>s)=P(X>t)for all s,t≥0P(X>s+t \mid X>s) = \frac{P(X>s+t)}{P(X>s)} = P(X>t) \quad \text{for all } s,t\ge 0

(the first equality holding because again {X>s+t}⊆{X>s}\{X>s+t\}\subseteq\{X>s\}), which rearranges to the functional equation

Fˉ(s+t)=Fˉ(s) Fˉ(t)for all s,t≥0.(∗)\bar F(s+t) = \bar F(s)\,\bar F(t) \qquad \text{for all } s,t \ge 0. \tag{$\ast$}

We claim Fˉ\bar F is non-increasing: this holds for any survival function, since {X>s+t}⊆{X>s}\{X>s+t\}\subseteq\{X>s\} implies Fˉ(s+t)≤Fˉ(s)\bar F(s+t)\le \bar F(s) for t≥0t\ge 0. Also Fˉ\bar F is right-continuous (a general property of survival functions 1−FX1-F_X where FXF_X is a CDF, itself right-continuous), and 0<Fˉ(x)≤10 < \bar F(x) \le 1 for all x≥0x\ge 0 by hypothesis together with Fˉ(0)=P(X>0)≤1\bar F(0) = P(X>0) \le 1. Define g(x):=ln⁡Fˉ(x)g(x) := \ln \bar F(x), well-defined since Fˉ(x)>0\bar F(x) > 0. Equation (∗)(\ast) becomes, after taking logarithms,

g(s+t)=g(s)+g(t)for all s,t≥0.(∗∗)g(s+t) = g(s) + g(t) \qquad \text{for all } s,t\ge 0. \tag{$\ast\ast$}

This is Cauchy's functional equation restricted to [0,∞)[0,\infty). We solve it under the regularity condition that gg is monotone (equivalently Fˉ\bar F is monotone, which we established above — Fˉ\bar F non-increasing means gg is non-increasing). Step 1: g(nx)=n g(x)g(nx) = n\,g(x) for all positive integers nn and x≥0x \ge 0. By induction on (∗∗)(\ast\ast): the base case n=1n=1 is trivial, and if g(nx)=ng(x)g(nx) = n g(x) then g((n+1)x)=g(nx+x)=g(nx)+g(x)=ng(x)+g(x)=(n+1)g(x)g((n+1)x) = g(nx + x) = g(nx) + g(x) = ng(x) + g(x) = (n+1)g(x). Step 2: g(x/m)=g(x)/mg(x/m) = g(x)/m for positive integers mm. Apply Step 1 with xx replaced by x/mx/m: g(m⋅(x/m))=m g(x/m)g(m \cdot (x/m)) = m\, g(x/m), i.e. g(x)=m g(x/m)g(x) = m\,g(x/m), so g(x/m)=g(x)/mg(x/m) = g(x)/m. Step 3: g(qx)=q g(x)g(qx) = q\,g(x) for all positive rationals q=n/mq = n/m. Combine Steps 1–2:

g(nmx)=n g(xm)=n⋅g(x)m=nmg(x).g\left(\frac{n}{m}x\right) = n\,g\left(\frac{x}{m}\right) = n \cdot \frac{g(x)}{m} = \frac{n}{m}g(x).

Taking x=1x=1, this gives g(q)=q g(1)g(q) = q\,g(1) for every positive rational qq. Step 4: extend to all real x≥0x \ge 0 using monotonicity. Let c:=g(1)c := g(1). We know c≤0c \le 0: indeed g(1)=ln⁡Fˉ(1)≤0g(1) = \ln \bar F(1) \le 0 since Fˉ(1)≤1\bar F(1) \le 1; and in fact c<0c<0 strictly, for if c=0c=0 then g(q)=0g(q)=0 for all rationals q>0q>0, and by the monotonicity of gg (sandwiching any real x>0x>0 between rationals arbitrarily close to it) this would force g≡0g \equiv 0 on [0,∞)[0,\infty), i.e. Fˉ≡1\bar F \equiv 1, contradicting that Fˉ\bar{F} is a proper survival function of a random variable with P(X>x)→0P(X>x)\to 0 necessarily as x→∞x \to \infty for a genuine (non-degenerate, a.s.-finite) continuous random variable on [0,∞)[0,\infty). [If instead Fˉ≡1\bar F \equiv 1 were allowed, XX would not be a.s. finite, contradicting that XX is a real-valued random variable.] So c<0c < 0; write λ:=−c>0\lambda := -c > 0. Now fix any real x≥0x \ge 0 and let (qn)(q_n), (rn)(r_n) be sequences of positive rationals with qn↑xq_n \uparrow x and rn↓xr_n \downarrow x. Monotonicity of gg (non-increasing) gives

g(rn)≤g(x)≤g(qn).g(r_n) \le g(x) \le g(q_n).

By Step 3, g(qn)=−λqn→−λxg(q_n) = -\lambda q_n \to -\lambda x and g(rn)=−λrn→−λxg(r_n) = -\lambda r_n \to -\lambda x as n→∞n\to\infty. Squeezing, g(x)=−λxg(x) = -\lambda x for every x≥0x \ge 0. Hence Fˉ(x)=eg(x)=e−λx\bar F(x) = e^{g(x)} = e^{-\lambda x} for all x≥0x \ge 0, which is exactly the survival function of Exp(λ)\mathrm{Exp}(\lambda). Since a CDF (equivalently its survival function) determines the distribution of XX uniquely, we conclude X∼Exp(λ)X \sim \mathrm{Exp}(\lambda). ■\blacksquare

∎

Worked examples

Example 1 (Uniform: a randomized quality-control cutoff). A machine cuts steel rods; because of small errors, the length in centimeters of a rod nominally 50 cm50\text{ cm} long is modeled as X∼Unif(48.5,51.5)X \sim \mathrm{Unif}(48.5, 51.5). A rod is acceptable if its length lies in [49,51][49, 51], and is sold at a premium if additionally it lies in [49.8,50.2][49.8, 50.2]. Probability of acceptance. Since XX is uniform on [48.5,51.5][48.5,51.5], an interval of length 33,

P(49≤X≤51)=51−4951.5−48.5=23≈0.667.P(49 \le X \le 51) = \frac{51-49}{51.5-48.5} = \frac{2}{3} \approx 0.667.

Probability of premium quality, given acceptance. We need P(49.8≤X≤50.2∣49≤X≤51)P(49.8 \le X \le 50.2 \mid 49 \le X \le 51). Since [49.8,50.2]⊂[49,51][49.8,50.2] \subset [49,51],

P(49.8≤X≤50.2∣49≤X≤51)=P(49.8≤X≤50.2)P(49≤X≤51)=(50.2−49.8)/32/3=0.4/32/3=0.42=0.2.P(49.8 \le X \le 50.2 \mid 49 \le X \le 51) = \frac{P(49.8\le X\le 50.2)}{P(49\le X\le 51)} = \frac{(50.2-49.8)/3}{2/3} = \frac{0.4/3}{2/3} = \frac{0.4}{2} = 0.2.

Mean and variance of length. By Part A of the theorem with a=48.5a=48.5, b=51.5b=51.5:

E[X]=48.5+51.52=50,Var(X)=(51.5−48.5)212=912=0.75.E[X] = \frac{48.5+51.5}{2} = 50, \qquad \mathrm{Var}(X) = \frac{(51.5-48.5)^2}{12} = \frac{9}{12} = 0.75.

So the average length is exactly the nominal 50 cm50\,\mathrm{cm}, with standard deviation 0.75≈0.866 cm\sqrt{0.75}\approx 0.866\,\mathrm{cm}.

Example 2 (Exponential: server response times and memorylessness in action). A web server's response time TT (in seconds) to a request is modeled as T∼Exp(λ)T \sim \mathrm{Exp}(\lambda) with λ=0.4\lambda = 0.4 (mean response time 1/λ=2.51/\lambda = 2.5 seconds). Probability the response takes more than 4 seconds.

P(T>4)=e−λ⋅4=e−1.6≈0.2019.P(T > 4) = e^{-\lambda \cdot 4} = e^{-1.6} \approx 0.2019.

Probability it takes more than 7 seconds total, given it has already taken more than 3 seconds. By memorylessness (Part C), with s=3s=3, t=4t=4:

P(T>7∣T>3)=P(T>4)=e−1.6≈0.2019,P(T > 7 \mid T > 3) = P(T > 4) = e^{-1.6} \approx 0.2019,

exactly the same number as the unconditional P(T>4)P(T>4) computed above — the three seconds already elapsed are entirely irrelevant to the additional wait. Contrast this with the direct (non-memoryless-shortcut) computation as a check:

P(T>7∣T>3)=P(T>7)P(T>3)=e−0.4⋅7e−0.4⋅3=e−2.8e−1.2=e−1.6,P(T>7 \mid T>3) = \frac{P(T>7)}{P(T>3)} = \frac{e^{-0.4\cdot 7}}{e^{-0.4\cdot 3}} = \frac{e^{-2.8}}{e^{-1.2}} = e^{-1.6},

confirming the two computations agree. Variance and a probability sandwich. By Part B, Var(T)=1/λ2=1/0.16=6.25\mathrm{Var}(T) = 1/\lambda^2 = 1/0.16 = 6.25, so the standard deviation is 2.52.5 seconds — as large as the mean itself, a signature feature of the exponential distribution (its coefficient of variation is always exactly 11). Using the MGF, we can also get the third moment via MT′′′(0)M_T'''(0): since MT(t)=λ/(λ−t)M_T(t) = \lambda/(\lambda - t), direct differentiation gives MT′′′(t)=6λ(λ−t)−4M_T'''(t) = 6\lambda(\lambda-t)^{-4}, so E[T3]=6λ/λ4=6/λ3=6/0.064=93.75E[T^3] = 6\lambda/\lambda^4 = 6/\lambda^3 = 6/0.064 = 93.75, a number that would be considerably more painful to extract from repeated integration by parts.

CounterexampleWhy "continuous and positive support" alone is not enough — memorylessness needs the distributional hypothesis, not just support

The converse direction of Part C assumed XX is a continuous random variable on [0,∞)[0,\infty) with a genuine functional equation Fˉ(s+t)=Fˉ(s)Fˉ(t)\bar F(s+t) = \bar F(s)\bar F(t) holding for all s,t≥0s,t\ge 0 — not merely at a single pair (s,t)(s,t), and not merely approximately. Drop the "for all s,ts, t" universality (i.e. only require memorylessness at one particular pair, or only for tt in some restricted set) and the conclusion collapses. Concretely, let X∼Unif(0,1)X \sim \mathrm{Unif}(0, 1), whose survival function is

Fˉ(x)=P(X>x)=1−x,0≤x≤1,\bar F(x) = P(X > x) = 1-x, \qquad 0 \le x \le 1,

(and Fˉ(x)=0\bar F(x) = 0 for x>1x > 1). Check whether the memoryless identity can hold even for a single nontrivial pair s,t>0s, t > 0 with s+t≤1s+t \le 1:

Fˉ(s+t)=1−(s+t),Fˉ(s)Fˉ(t)=(1−s)(1−t)=1−s−t+st.\bar F(s+t) = 1-(s+t), \qquad \bar F(s)\bar F(t) = (1-s)(1-t) = 1 - s - t + st.

These agree only when st=0st = 0, i.e. only in the degenerate case s=0s=0 or t=0t=0. For example with s=0.2s=0.2, t=0.3t=0.3:

Fˉ(0.5)=0.5,Fˉ(0.2)Fˉ(0.3)=(0.8)(0.7)=0.56,\bar F(0.5) = 0.5, \qquad \bar F(0.2)\bar F(0.3) = (0.8)(0.7) = 0.56,

and 0.5≠0.560.5 \ne 0.56. So already at this one pair the uniform distribution fails memorylessness — confirming that memorylessness genuinely distinguishes the exponential family from other continuous laws on [0,∞)[0,\infty), including the uniform distribution studied in the same article. More strikingly, this also shows why the hypothesis "for all s,t≥0s,t\ge0" cannot be relaxed to "for some s,ts,t" in the converse: the forward implication (exponential ⇒\Rightarrow memoryless) is airtight, but one cannot hope to recover exponentiality from a single instance of the identity, since many non-exponential Fˉ\bar F (like the uniform one above restricted appropriately, or any Fˉ\bar F tampered with off a sparse set) can coincidentally satisfy (∗)(\ast) at isolated points while failing it generically. It is precisely the simultaneous validity for every pair — turning (∗)(\ast) into Cauchy's equation on a continuum — that pins down the exponential form uniquely. Intuitively: with the uniform distribution, having survived 0.20.2 units already makes total survival to 0.50.5 more informative (less likely relative to fresh odds) than starting fresh at 0.30.3, because uniform mass is running out near x=1x=1 — the exact opposite of "forgetting."

Scope, edge cases & common traps

SituationCorrect handling
X∼Unif(a,b)X\sim \mathrm{Unif}(a,b), evaluate P(X=c)P(X=c) for a point ccEquals 00 always — continuous distributions assign zero mass to single points; only intervals carry probability.
Rate vs. mean parametrization of exponentialSome texts use rate λ\lambda (density λe−λx\lambda e^{-\lambda x}), others use mean θ=1/λ\theta=1/\lambda (density 1θe−x/θ\tfrac1\theta e^{-x/\theta}) — always confirm which convention is in force before reading off E[X]E[X].
Applying memorylessness at negative tt or ssThe identity P(X>s+t∣X>s)=P(X>t)P(X>s+t\mid X>s)=P(X>t) is stated, and only proved, for s,t≥0s,t\ge 0; for t<0t<0 the left side isn't even the right conditional event framing.
Believing memorylessness is just "independence of the past" for any random processIt characterizes only the exponential continuous law on [0,∞)[0,\infty) (and, discretely, the geometric law) — no other continuous distribution on [0,∞)[0,\infty) has it, by the theorem's converse.
Confusing "uniform on an interval" with "discrete uniform on integers in that interval"These are different objects with different variances; Var\mathrm{Var} of continuous uniform on [a,b][a,b] is (b−a)2/12(b-a)^2/12, not the discrete uniform formula.
Minimum of independent exponentialsmin⁡(X1,…,Xn)\min(X_1,\dots,X_n) with Xi∼Exp(λi)X_i\sim\mathrm{Exp}(\lambda_i) independent is Exp(λ1+⋯+λn)\mathrm{Exp}(\lambda_1+\cdots+\lambda_n) — a clean consequence of memorylessness/tail multiplicativity, easily misremembered as the maximum instead.

The mathematician behind it

Origins · 1 minSiméon Denis Poisson and how this was discovered
SD
Siméon Denis Poisson
1781–1840 · French mathematician, engineer, and physicist working in probability, mechanics, and mathematical physics

Poisson's 1837 treatise Recherches sur la probabilité des jugements introduced the limiting law now bearing his name, arising as the limit of the binomial distribution under rare-event scaling; the exponential distribution emerges as the natural continuous-time waiting law between events of a Poisson process, tying together discrete counting and continuous waiting in a single framework that dominates queueing theory, reliability engineering, and physics of decay processes to this day.

History
- 1837 — Poisson derives the limiting law of rare events (Poisson distribution), setting the stage for continuous-time exponential waiting times between such events.
- 1900s — Erlang, studying telephone exchange traffic, formalizes exponential interarrival and holding-time models, founding queueing theory.
- 1930s — Kolmogorov's axiomatization of probability places continuous densities like the uniform and exponential on rigorous measure-theoretic footing.
- 1950s — The memorylessness characterization of the exponential distribution is isolated and proved via Cauchy's functional equation, becoming a standard textbook theorem.
- Present — Uniform and exponential distributions remain the two default building blocks for simulation (inverse-transform sampling) and for reliability, survival, and queueing models.
Why is the exponential distribution called "memoryless," and is that the same as independence?

Memorylessness means that conditional on having survived past time ss, the distribution of the remaining lifetime X−sX-s is identical to the distribution of a fresh XX: formally P(X>s+t∣X>s)=P(X>t)P(X>s+t\mid X>s)=P(X>t) for all s,t≥0s,t\ge0. This is a statement about a single random variable's conditional law, not about two separate random variables being statistically independent in the usual sense. It is best read as "the amount of elapsed time carries zero information about the amount of remaining time" — the survival probability resets exactly as if a fresh exponential clock had started at time ss.

Can a uniform distribution ever be memoryless?

No. The counterexample above shows explicitly that Unif(0,1)\mathrm{Unif}(0,1) fails the memoryless identity even at a single pair (s,t)=(0.2,0.3)(s,t)=(0.2,0.3), giving Fˉ(0.5)=0.5\bar F(0.5)=0.5 against Fˉ(0.2)Fˉ(0.3)=0.56\bar F(0.2)\bar F(0.3)=0.56. More conceptually, any uniform distribution on a bounded interval has survival probability hitting exactly 00 at the right endpoint, so information about elapsed time is highly informative about remaining time near that boundary — the opposite of forgetting. Memorylessness is provably exclusive to the exponential family among continuous laws on [0,∞)[0,\infty).

How do you simulate an exponential random variable from a uniform one?

Use the inverse transform method. If U∼Unif(0,1)U \sim \mathrm{Unif}(0,1), then X:=−1λln⁡(1−U)X := -\frac{1}{\lambda}\ln(1-U) is distributed Exp(λ)\mathrm{Exp}(\lambda). This follows because the CDF of the exponential is F(x)=1−e−λxF(x) = 1-e^{-\lambda x}, which is a strictly increasing bijection from [0,∞)[0,\infty) to [0,1)[0,1); setting u=F(x)u = F(x) and solving for xx gives x=−ln⁡(1−u)/λ=F−1(u)x = -\ln(1-u)/\lambda = F^{-1}(u), and the general inverse-transform theorem guarantees F−1(U)∼FF^{-1}(U) \sim F whenever U∼Unif(0,1)U\sim\mathrm{Unif}(0,1). Since 1−U1-U is itself Unif(0,1)\mathrm{Unif}(0,1) when UU is, one often simplifies to X=−ln⁡(U)/λX = -\ln(U)/\lambda.

Why does the exponential distribution have variance equal to the square of its mean?

This is a direct computation from the moment generating function MX(t)=λ/(λ−t)M_X(t)=\lambda/(\lambda-t): differentiating gives E[X]=1/λE[X]=1/\lambda and E[X2]=2/λ2E[X^2]=2/\lambda^2, so Var(X)=2/λ2−1/λ2=1/λ2=(E[X])2\mathrm{Var}(X)=2/\lambda^2 - 1/\lambda^2 = 1/\lambda^2 = (E[X])^2. Equivalently, the exponential distribution's coefficient of variation (standard deviation divided by mean) is always exactly 11, a signature numeric fact that instantly distinguishes exponential data from, say, nearly-deterministic lifetimes (coefficient of variation near 00) or highly erratic heavy-tailed lifetimes (coefficient of variation above 11).

What is the relationship between the exponential distribution and the geometric distribution?

The exponential distribution is the continuous-time limit of a rescaled The Geometric Distribution. If you divide time into nn tiny intervals of length 1/n1/n and attempt a "success" (arrival) independently in each with probability λ/n\lambda/n, the number of failed intervals before the first success is geometric, and the time until first success, Tn=(number of trials)/nT_n = (\text{number of trials})/n, converges in distribution to Exp(λ)\mathrm{Exp}(\lambda) as n→∞n\to\infty. This is why the exponential distribution inherits memorylessness directly from the memorylessness of the geometric distribution (a fair coin doesn't "remember" past tails), and it is the standard route by which discrete waiting-time intuition becomes continuous waiting-time theory.

Does every "rate" parametrized model automatically have the exponential distribution?

No — having a constant hazard rate is precisely what forces exponentiality, but many lifetime models have hazard rates that change over time (increasing hazard for wear-out failure, decreasing hazard for "infant mortality" in components), and these are modeled by distributions such as the Weibull or gamma, which strictly contain the exponential as a special case. The exponential distribution is exactly the constant-hazard-rate member of these families, which is equivalent, by the theorem proved here, to memorylessness; any observed increasing or decreasing failure rate is direct empirical evidence against an exponential model.

Why does the proof of the memorylessness converse need Cauchy's functional equation rather than a shortcut?

Because the hypothesis only gives a multiplicative relation Fˉ(s+t)=Fˉ(s)Fˉ(t)\bar F(s+t) = \bar F(s)\bar F(t) for all nonnegative s,ts,t, with no a priori assumption of differentiability or continuity beyond XX being a continuous random variable. Cauchy's functional equation (in additive form, after taking logarithms) is famous for having wild, non-linear, even non-measurable solutions if no regularity is assumed — solutions built via a Hamel basis of R\mathbb{R} over Q\mathbb{Q}. The monotonicity of the survival function Fˉ\bar F (a genuine, provable property of every survival function) is exactly the regularity condition that rules out these pathological solutions and forces the clean linear form g(x)=−λxg(x) = -\lambda x, hence the exponential form for Fˉ\bar F.

Before this

Where this leads