From "how long until the next one?" to $\lambda e^{-\lambda x}$
Picture a bus stop where buses don't follow a schedule — they just show up at random. You can't predict the exact moment, but you can ask: how long until the next one? Sometimes it's seconds, sometimes many minutes, and short waits happen far more often than long ones. That lopsided "lots of short waits, a few long ones" shape is exactly what the exponential distribution captures: the waiting time until the next random event.
Think of popcorn in the microwave. Right after one pop, you wait for the next. The kernels don't "remember" how long you've already been waiting — the next pop is just as likely now as it was a moment ago. Most gaps between pops are short, but every so often there's a long quiet stretch. If you wrote down every gap length and stacked them up, you'd get a curve that starts tall on the left and slides down toward zero: many small waits, few big ones.
Now attach a number. The rate $\lambda$ counts how many events happen per unit of time. If your phone gets $\lambda = 2$ texts per hour on average, the typical gap between texts is the mean $1/\lambda = 0.5$ hour, or 30 minutes. The chance of waiting longer than a time $x$ takes a beautifully simple form:
$$P(X>x)=e^{-\lambda x}$$
So the chance of waiting more than a full hour for a text is $P(X>1)=e^{-2}\approx 0.135$ — about a 14% shot. A bigger $\lambda$ means events arrive faster, which makes long waits rarer.
The full distribution has density $f(x)=\lambda e^{-\lambda x}$ and cumulative form $F(x)=P(X\le x)=1-e^{-\lambda x}$ for $x\ge 0$. Its mean is $1/\lambda$ and — unusually — its standard deviation is also $1/\lambda$, so $\text{Var}(X)=1/\lambda^2$. The signature feature is memorylessness: $P(X>s+t\mid X>s)=P(X>t)$. Having already waited $s$ tells you nothing about the wait still to come. In the sim above, the λ rate slider sets how fast events arrive (a steeper curve), and the x query point slider picks the time $x$ at which the page reports $P(X\le x)$ and shades the tail $P(X>x)$.
Slide $\lambda$ from $0.5$ up to $5$ and watch the curve steepen while the mean line $\mu=1/\lambda$ rushes toward zero — faster events, shorter typical waits. Open the Memoryless tab and notice every conditional curve lands exactly on top of the others: a visual proof that the past wait doesn't matter. Finally, open Live Sampling, press Play, and watch the random bars pile up into the smooth $\lambda e^{-\lambda x}$ shape as the sample mean closes in on $1/\lambda$.
Exponential Distribution — memoryless continuous distribution
$$\boxed{f(x;\lambda)=\lambda e^{-\lambda x},\quad F(x)=1-e^{-\lambda x},\quad x\ge0}$$
$E[X]=1/\lambda,\quad\text{Var}(X)=1/\lambda^2,\quad\text{Median}=\ln2/\lambda$
| Symbol | Meaning | Range | Interpretation |
|---|---|---|---|
| $\lambda$ | Rate parameter | $(0,\infty)$ | Events per unit time (from Poisson process) |
| $1/\lambda$ | Scale (mean) | $(0,\infty)$ | Average waiting time between events |
| $x$ | Waiting time | $[0,\infty)$ | Time until next event |
$$P(X>s+t\mid X>s)=P(X>t)$$
The exponential distribution is the ONLY continuous distribution with this property. "If you've waited $s$ minutes, the remaining wait has the same distribution as a fresh wait." This is why it models waiting times for random events (radioactive decay, call arrivals, hardware failures).
Start from $P(X>s+t)=P(X>s)P(X>t)$ for all $s,t\ge0$. Let $G(t)=P(X>t)$. Then $G(s+t)=G(s)G(t)$ — the only continuous solution is $G(t)=e^{-\lambda t}$ for some $\lambda>0$. Therefore $F(x)=1-e^{-\lambda x}$ and $f(x)=\lambda e^{-\lambda x}$.
$$E[X]=\int_0^\infty x\lambda e^{-\lambda x}\,dx=\frac{1}{\lambda}\quad\text{(integration by parts)}$$
Problem: Light bulbs fail at rate λ=0.01/hour (mean life = 100 hours). Find P(bulb lasts more than 150 hours) and P(lasts between 50–150 hours).
$$P(X>150)=e^{-0.01\times150}=e^{-1.5}\approx\mathbf{0.2231}$$
$$P(50
"Memoryless means the past doesn't affect anything — all distributions are memoryless."
Memorylessness is a very special property. For the exponential: given you've waited 10 minutes, the remaining wait has the same Exp(λ) distribution as if you just started. Most distributions aren't memoryless — for a Normal waiting time, knowing you've waited longer shifts the remaining time distribution. The geometric distribution is the only discrete memoryless distribution; exponential is the only continuous one.
📖 Ross — Ch. 5.4
"The rate parameter λ and the mean 1/λ are the same thing."
They are reciprocals. If λ=2 failures per hour, the mean time to failure is 1/2 hour = 30 minutes. Always be clear: λ is the rate (events per time), 1/λ is the mean (time per event). In survival analysis, the "hazard rate" h(t)=λ for exponential — constant over time, which is why it's called "memoryless."
📖 DeGroot & Schervish — Ch. 5.7
Confusing P(X > x) and P(X < x) — wrong tail
✅ F(x)=P(X≤x)=1-e^(-λx) is the CDF. P(X>x)=1-F(x)=e^(-λx) is the survival function. For the exponential, it's common to want the survival probability (how long until failure), so use e^(-λx) directly.
🔍 Students use F(x) when they want P(X>x) — off by sign.
Using exponential model when hazard rate is not constant
✅ Exponential has constant hazard h(t)=λ. Real-world failure rates often increase with age (Weibull distribution) or have a "bathtub curve" (early failures + random failures + wear-out). Always check the hazard rate plot before fitting exponential.
🔍 Fitting exponential to light bulbs that have increasing failure rates with age.
Parametrizing by mean (θ=1/λ) vs rate (λ) — software inconsistency
✅ Some software (R: rexp) uses rate λ. Others (scipy.stats.expon) use scale=1/λ. Always check which parametrization your software uses. rexp(n, rate=2) vs scipy.stats.expon(scale=0.5) are the same distribution.
🔍 Getting backwards exponential by confusing rate and scale parameters.