An interactive essay · Probability
Chaos, added up, always draws the same curve.
Coin flips, dice rolls, measurement errors, the heights of strangers — pile up enough independent randomness and one shape shows up, every single time: the bell curve. Mathematicians call this the Central Limit Theorem. This page is a hands-on tour of why it happens — built for intuition, not prerequisites.
▾ balls are already falling
No ball is predictable. Twelve bounces of pure luck. And still, the pile refuses to surprise us.
That tension — individual chaos, collective certainty — is the whole story. The five acts above are five different answers to the same question: why this shape? Take them in order; each one hands off to the next.
Act ICounting
More ways to be in the middle
Start with one die: six outcomes, all equally likely. The distribution is a flat slab — nothing bell-shaped about it.
Now roll two dice and add them. A total of 7 can happen six ways (1+6, 2+5, 3+4, 4+3, 5+2, 6+1). A total of 12 happens exactly one way. The middle isn't luckier — it's roomier. Extremes are fussy: to roll 12, both dice must cooperate. To land in the middle, they're free to disagree.
Add more dice, and the imbalance compounds — fast:
See the numbers behind this chart
| Sum | Ways to roll it | Probability |
|---|
Two things happen as n grows. The distribution drifts right (the mean grows like n) and spreads out — but only like √n, much slower. If you ride a camera that pans and zooms at exactly that rate, the shape stops moving entirely. That frozen silhouette is the theorem: it's what's left when you subtract the boring growth and look at the pure shape of the accumulated luck.
But wait — maybe dice are special. Six neat symmetric faces. Would something messier still funnel into the bell?
Act IIThe fixed point
Draw anything. It becomes the bell.
The bell isn't a shape some distributions have — it's a place all distributions go.
Below is a blank distribution. Draw on it — spikes, cliffs, two islands, something lopsided and spiteful. Then press Blend: your distribution gets replaced by the distribution of the average of two independent draws from it. Blend again and you're averaging 4 copies, then 8, then 16… Watch what survives.
Averaging is a smoother. It erases your drawing's sharpest details first — the spikes — then its lopsidedness, and in the end remembers exactly two numbers: the mean (where your drawing sat) and the spread (how wide it was). Every other feature of your art gets washed away.
The bell is simply what "nothing left but a mean and a spread" looks like. In mathematical terms it's a fixed point: the one shape that blending maps onto itself. Every other shape is on a conveyor belt heading toward it.
For the curious: why exactly two numbers survive
Blending preserves the mean, and (after the rescaling that keeps the picture in frame) it preserves the spread — but it strictly smooths everything else. Information theorists make this precise: each blend increases entropy, a measure of how spread-out and featureless a distribution is. So the process must end at the maximum-entropy distribution with your mean and spread — and that distribution is the Gaussian. Your drawing's fine structure is literally information, and averaging destroys information. Act IV makes this budget idea something you can hold.
Fine — everything flows to one destination. But that only sharpens the mystery. Why is the destination this exact curve — why e−x² and not some other pleasant hump?
Act IIIThe truce
A truce between two pressures
Here is a swarm of particles obeying two rules.
Jitter. Every instant, each particle takes a tiny random step — left or right, no preference. This spreads the swarm out, for a bluntly democratic reason: there are vastly more ways to be spread out than to be bunched up.
Tether. Each particle is pulled gently back toward the center, with a force that grows the farther it strays. Being far away is expensive.
Alone, each rule is a disaster. Together, they negotiate:
The swarm settles into the shape where, at every single point, the outward push of spreading and the inward pull of the tether cancel exactly. That treaty line is the bell curve.
And this is why the bell is universal: it doesn't care about dice or coins or your drawing. Any process that mixes "spread out randomly" with "but wandering far is costly" strikes the same truce. Adding up random variables is exactly such a process — each new term jitters the sum, while the √n rescaling acts as the tether.
For the curious: the two-line force balance
Write p(x) for the swarm's density. Spreading pushes mass from crowded spots toward empty ones, with strength proportional to the relative slope p′(x)/p(x). The tether pushes mass inward with strength proportional to the distance x. At equilibrium they cancel at every point:
p′(x) / p(x) = −2ax ⟹ p(x) ∝ e−ax²
The square in the exponent is no accident — it's the fingerprint of a restoring force that grows linearly with distance. (Physicists will recognize the swarm as diffusion in a spring potential; the spreading rule alone is the heat equation. The bell is warmth held by a spring.)
That's the physics telling. Information theory tells the same story with a ledger instead of forces — and it explains something the swarm can't: what, exactly, is being optimized.
Act IVInformation
The entropy budget
Entropy is the accountant's measure of spread: how unpredictable a draw is — how many yes/no questions you'd need, on average, to pin one down. A spike has almost none (you already know the answer). A wide, even wash of probability has a lot.
If entropy were all that mattered, the champion would be the flat slab — the uniform distribution. Perfectly even, no favoritism, maximum surprise for its width. Cap the width, and flat is genuinely unbeatable.
But sums of randomness don't come with a width cap. They come with a variance budget: probability placed at distance x from the center costs x² — being far out is quadratically expensive. (That's Act III's tether, wearing a green eyeshade.) So the real question is: with variance pinned at 1, which allocation of probability buys the most entropy?
The slab loses for a reason you can feel: its cliff edges. A hard edge is an absurd promise — "nothing, ever, beyond this line" — and softening it into a gentle tail buys a big chunk of entropy for a tiny sliver of variance. Heavy tails lose the other way: rare far-out draws are so expensive that the middle must bunch into a near-spike to stay on budget, and spikes are an entropy catastrophe. The bell is the allocation with no waste left: every cliff already softened, every extravagance already trimmed.
Don't take the dial's word for it, though — it only searched one family of shapes. Sculpt your own and try to beat the gold line:
Nobody beats the gold line, and Shannon proved nobody can: among all distributions with variance 1, the bell has the maximum entropy. The spill you just watched is the proof made visible — as long as any cliff or spike remains, some local rearrangement still gains entropy at no budget cost, so the process can only come to rest at the bell.
And notice: Act II was this act in disguise. Every "blend" increases entropy, while the rescaling preserves the budget — so blending has no choice but to march uphill to the entropy summit. That is the information-theory proof of the Central Limit Theorem, in one breath.
For the curious: the audit in three lines
Maximize H = −∫ p ln p subject to ∫ p = 1 and ∫ x²p = σ². Lagrange multipliers turn the constrained problem into an unconstrained one, and setting the derivative to zero gives:
ln p(x) = −1 − λ₁ − λ₂x² ⟹ p(x) = A·e−λ₂x²
The x² in the exponent is the variance constraint's signature — the price tag, imported directly into the shape. (Fine print: the meter measures differential entropy, which can go negative for very spiky shapes. Think of a negative score as information-debt.)
One picture remains — the strangest, and once you see it, the most convincing. It's the reason the formula has a square in it, told with pure geometry.
Act VGeometry
The cloud that looks the same from every angle
Take two independent copies of a random quantity — call them X and Y — and plot each pair as a point (X, Y). Thousands of samples form a cloud.
Here's the trick that makes this picture powerful: the sum X+Y lives along the 45° diagonal. Adding random variables is casting the cloud's shadow onto a tilted line. Rotating the cloud sweeps through every possible weighted sum of X and Y.
So: drag the cloud, and watch its shadow.
The bell's cloud is round. Rotate it however you like — its shadow is identical. It is the only distribution whose cloud manages that while X and Y stay independent. And that double demand — independence and rotation-proofness — is so restrictive it forces the exact formula.
For the curious: how roundness forces e−x²
Independence means the cloud's density factors: f(x)·f(y). Roundness means it can only depend on the distance from center: g(x² + y²). Set them equal, take logs, and let h = log f:
h(x) + h(y) = G(x² + y²) ⟹ h′(x)/x = h′(y)/y for all x, y
A function of x alone equalling a function of y alone must be a constant, say −2a. Then h′(x) = −2ax, so h(x) = −ax² + c, and
f(x) ∝ e−ax²
Two innocent-looking symmetry demands, one inevitable curve.
And now the theorem in one sentence: adding many independent copies makes the cloud round. With n copies the cloud lives in n dimensions, and the same lurching-shadow effect you saw with coins gets averaged over so many directions that every shadow smooths toward the same silhouette — the bell. You already watched this happen in Act II: each blend was one more dimension of rounding.
FinaleThe summit
Many paths, one mountain
Five arguments, from five different corners of mathematics, and they all land on the same curve:
Act I · CountingThe middle is roomier
Extreme totals require every die to cooperate; middling totals can happen countless ways. Sheer headcount builds the hump.
Act II · Fixed pointAll roads lead to it
Averaging smooths every distribution toward the one shape that averaging leaves unchanged. The bell is the destination, not a coincidence.
Act III · BalanceSpreading vs. the tether
Entropy pushes mass outward, cost pulls it home. The bell is the shape where the two pressures cancel at every point.
Act IV · BudgetingThe perfect spend
Entropy wants width; variance taxes distance squared. The bell is the only allocation with no waste left in the ledger — the maximum-entropy shape for its budget.
Act V · GeometryThe only round cloud
Independence plus same-from-every-angle is a demand only e−x² can meet — and adding copies makes every cloud round.
When five unrelated roads all reach the same summit, you're not looking at a trick. You're looking at real terrain. That's what makes the Central Limit Theorem a theorem about the world, not just about dice: measurement errors, molecular collisions, genetics, polling — anywhere many small independent influences add up, the bell is already waiting.
Where the bell cracks — the fine print
- Independence. The pieces must not conspire. Stock-market losses look bell-ish on calm days — then a panic correlates everyone, and the "impossible" tail arrives. This footnote has bankrupted several hedge funds.
- Finite spread. If single draws can be so wild that the variance is infinite (heavy tails), averaging never tames them. Average a million draws from the Cauchy distribution and you've learned nothing — the average is exactly as wild as one draw. Such sums funnel to other shapes (the stable distributions — the bell's feral cousins).
- Enough addends. The bell is a limit. With three dice it's an approximation; with thirty it's excellent; the far tails take longest to fall in line.
Going further:
- 3Blue1Brown, But what is the Central Limit Theorem? — beautiful animated companion.
- Seeing Theory (Brown University) — a visual course on all of basic probability.
- Francis Galton's original 1889 bean machine — the hero of our opening scene, built in brass.
- For the entropy road: E. T. Jaynes, Probability Theory: The Logic of Science, ch. 7 — and the entropy-power inequality, if you want the full rigor of Act IV.