How the Poisson Distribution Works and When to Use It

You're lying in a field during a meteor shower. A friend who's read up on it says "this shower averages 6 meteors an hour." You settle in and watch for exactly one hour.

How many will you actually see?

Not 6, probably. Maybe 4. Maybe 9. Maybe 2. The average tells you what to expect over many, many hours, not what happens in this one. So what's the actual chance of seeing exactly 4? Or exactly 9? Or none at all?

That's the question the Poisson distribution answers. And once you see where its formula comes from, the famously weird-looking $e^{-\lambda}$ in it stops being weird.

Randomness with a known average

Before the math, let's just watch it happen. Below is a simulated hour of meteor watching. You set the average rate, hit run, and meteors land at random moments across the window.

Run it a bunch of times at the same λ. The count bounces around every single time, even though the average rate never changes. That bounce isn't noise you can average away. It's the actual content of the problem. The Poisson distribution is the exact shape of that bounce.

Notice something else too: where in the hour a meteor lands doesn't affect where the next one lands. Meteor at minute 3 has nothing to do with meteor at minute 40. That independence is the whole reason this problem has a clean answer.

Why not just use the binomial distribution?

Here's an idea. You already know how to handle a fixed number of independent yes/no trials: that's the binomial distribution. So what if we force this problem into that shape?

Chop the hour into $n$ equal slices, say 3,600 one-second slices. In each slice, either a meteor happens or it doesn't. Call the chance of a meteor in any one slice $p$. If meteors land uniformly at random and the average over the whole hour is $\lambda$, then each slice needs:

Now the count of meteors in the hour is just the count of "successes" out of $n$ coin flips, each with success chance $p$. That's exactly the binomial distribution:

But wait: what if two meteors land in the same one-second slice? Our slices are too coarse to count that. So this binomial model is only an approximation, and a shaky one at that.

Slide $n$ up. Watch the green bars, the binomial approximation, sharpen and lock onto the dashed outline. That dashed outline is the Poisson distribution. It's what the binomial model converges to as the slices get so thin that two meteors can never share one, no matter how close together they land.

Taking the limit: where $e^{-\lambda}$ comes from

So the Poisson distribution isn't some separate invention. It's the binomial distribution pushed to its limit: $n \to \infty$ slices, with $p = \lambda/n$ shrinking to keep the average $np = \lambda$ fixed.

Watch what happens to each piece of the binomial formula as $n$ grows huge:

The fraction $\frac{n!}{(n-k)!}$ is a product of $k$ terms all close to $n$, so it cancels almost exactly with the $n^k$ sitting in the denominator of $\left(\frac{\lambda}{n}\right)^k$. What's left over is $\lambda^k$.

The term $\left(1-\frac{\lambda}{n}\right)^{n-k}$ is the classic one. As $n\to\infty$, it slides toward $e^{-\lambda}$, the same limit that defines Euler's number itself.

Put the surviving pieces back together and you get:

That's the Poisson probability mass function. One parameter, $\lambda$, the average rate. Plug in any whole number $k$ and it tells you the exact chance of seeing exactly $k$ events.

One number controls everything: meet $\lambda$

Because $\lambda$ is the only knob, it has to do a lot of work. It sets the average, and it also sets the spread. In fact, for a Poisson distribution:

Same number, both times. That's not true for most distributions you'll meet, and it's a handy way to spot when Poisson is the right tool: if the variance of your counts looks suspiciously close to the mean, you're probably looking at a Poisson process.

Drag $\lambda$ down toward 1 and the distribution piles up near zero and trails off to the right, lopsided. Drag it up past 15 or 20 and it stretches out, centers itself, and starts looking suspiciously like a bell curve. That's not a coincidence, and it's exactly what the next section is about.

So when should you actually reach for a Poisson?

Pull all of this together and the Poisson distribution fits whenever four things are true about whatever you're counting:

• Events happen independently of each other.
• There's a constant average rate over the window you care about (no busy hour vs. quiet hour).
• Two events essentially never happen at the exact same instant.
• You care about the count of events in a fixed window, not their exact timing.

Meteors in an hour, typos on a page, customer arrivals per minute at a quiet counter, radioactive decays per second: all Poisson. Something with clumping, like earthquake aftershocks, or a hard ceiling, like "at most 1 event possible," breaks the assumptions and needs a different model.

There's also a practical shortcut worth knowing. Once $\lambda$ gets large, computing $\lambda^k e^{-\lambda}/k!$ for every $k$ gets tedious, and you might wonder if you can just swap in a normal distribution with the same mean and variance.

Slide $\lambda$ past about 10 and the normal curve tracks the bars closely enough to use in practice. Below that, the Poisson distribution is too lopsided for a symmetric bell to fake convincingly, and you're better off computing the exact formula.

The short version

The Poisson distribution is what you get when you take the binomial distribution and let the number of trials go to infinity while keeping the average fixed. That limit is where the $e^{-\lambda}$ comes from: it's the same limit that defines $e$ itself. The result is a one-parameter distribution, $\lambda$, where the mean and variance are always equal.

Use it whenever you're counting independent events that happen at a steady average rate in a fixed window: meteors per hour, calls per minute, defects per batch. And once $\lambda$ gets past about 10, a plain normal distribution approximates it well enough to save you the arithmetic.

Next time someone tells you an average rate, you'll know exactly how to work out the odds of everything above, below, and exactly at that number.

All visualizations are interactive React components running entirely in your browser. The meteor simulator draws real Poisson-distributed samples, the slicing convergence and normal approximation use exact binomial and normal-CDF calculations. No libraries beyond React.