the second hundred · metaphor 174
A lucky break can feel like a gift with no invoice — it simply happened, against all odds. But the odds are not vague. For the right kind of luck, improbability has a precise, brutal arithmetic: it falls off exponentially, and there is a number that names exactly how fast.
Average many small independent things — coin flips, deals, days, opinions, mutations — and the average almost always lands near its expected value. That's the law of large numbers, and it's comforting. Large-deviation theory asks the darker question: when the average lands somewhere it shouldn't — a run of luck, a freak collapse — how rare is that, exactly? And the answer is not "very." The answer is a specific exponential, steeper the further from ordinary you go.
The crucial word is exponentially. Doubling the number of things you average doesn't halve the odds of a given miracle — it squares how unlikely it was. A break that was one-in-a-hundred over fifty draws is one-in-ten-thousand over a hundred. There's a function, the rate function, that fixes the exponent to the decimal. This page lets you set the size of the miracle and watch its price computed, live, from real random draws.
The rate function
Take n independent draws, average them, and ask for the chance the average lands at least a above where it belongs. For small departures the central limit theorem gives a bell and calls it a day. But out in the tail — a large deviation, a fixed distance a from the mean no matter how big n grows — something sharper is true. The probability shrinks like e^(−n·I(a)): exponential in the number of draws, with an exponent set by a single function of how far out you're reaching.
That function I(a) is the rate function, and it is the whole story. It is zero at the mean (the ordinary outcome costs nothing) and rises, faster than linearly, as a moves outward. A break twice as far from ordinary doesn't cost twice as much improbability — it costs more than twice, because I curves upward. The rate function is the exact price list of the atypical: hand it a deviation and it returns the exponent, to the decimal.
Cramér's theorem makes the promise precise: −(1/n)·ln P(average ≥ a) → I(a) as n grows. Read it backward and it says the log-probability of a miracle is, for large samples, just −n·I(a) — a straight line down, its slope the rate. The instrument measures that slope from real draws and watches it settle onto I(a).
What to try
Everything is measured from live coin draws — the shaded odds, the exponent, all of it. Start by dragging n up: the sampling distribution on the left narrows around zero (the average sharpens), so the shaded tail beyond your bar shrinks fast — and the right panel shows the measured exponent −ln P̂ / n descending toward the amber I(a) line. That descent is Cramér's theorem happening in front of you: the crude odds carry a polynomial fudge factor, but the exponent per draw locks onto the rate function.
Now drag a outward — make the miracle bigger. The odds don't ease off; they fall off a cliff, and the "odds against" chip races into the thousands and then past what any number of draws can resolve. When it does, the panel says so honestly: Monte Carlo simply stops seeing events that rare, and the disclaimer reports the floor. That failure is the point — a true miracle is so exponentially unlikely that even a machine flipping coins millions of times will never witness it, yet the rate function still prices it exactly.
The mapping
Here is the lens. Many outcomes we call luck — a career that broke exactly right, a portfolio that compounded through a decade of good calls, an epidemic that fizzled, a scandal that never landed — are aggregates: a long sum of small, roughly independent contributions. For such things, an extraordinary result isn't a coin toss that came up magic. It is many ordinary tosses that all happened to lean the same way, and large-deviation theory says exactly how much improbability that leaning costs: e^(−n·I), priced by the rate function, steep and specific.
This reframes both awe and dread. The lucky break was not free; it was purchased at a precise exponential rate, which is why it is rare and why it doesn't repeat — the price is paid again in full each time. And the catastrophe you fear has an exponent too: the further the departure that would ruin you, the more disproportionately unlikely it is that enough small things align to produce it. Miracles and disasters live in the same tail, governed by the same curve. The theory doesn't dissolve the wonder. It tells you the wonder had a number.
Read as life lessons
Because I(a) curves upward, reaching twice as far from ordinary is far more than twice as rare. The extraordinary gets exponentially lonelier the further out you go — luck does not scale linearly.
The exponent carries a factor of n. The more independent things a great outcome required to all align, the more astronomically improbable it was — and the less it says about the next attempt.
An exponential is small, not zero. Across enough lives, even e^(−n·I) events are guaranteed to happen to someone — you just can't count on being them, and you'll only ever hear the survivors' side.
In the wild
Large deviations are the modern language of entropy and thermodynamics: the overwhelming improbability of a gas spontaneously huddling in one corner is a rate function, and equilibrium is simply where I is minimized.
Ruin theory and risk management price the exponentially rare — the chance reserves are exhausted, the tail loss that breaks a firm. Getting the exponent right is the difference between solvency and collapse.
Error exponents in communication, and the rate of rare events in queues and networks, are large-deviation statements — the vanishing-but-precise odds that a well-designed system fails.
The mapping, exactly
| Mathematics | Life |
|---|---|
| average of n draws | An outcome built from many small, roughly independent contributions — talent, effort, timing, a thousand ordinary days. |
| the mean | The unremarkable result all those pieces usually add up to — the life that "makes sense." |
| a large deviation (avg ≥ a) | The lucky break or the catastrophe — an aggregate that lands wildly far from what was likely. |
| e^(−n·I(a)) | The improbability of that miracle, falling off exponentially in how many pieces had to lean the same way. |
| the rate function I(a) | The exact price of the departure — the exponent per contribution, the miracle itemized to the decimal. |
| I convex, I(mean)=0 | Ordinary is free; the further from ordinary you reach, the more disproportionately it costs. Luck is not linear. |
The honest model
Each draw is a fair coin worth ±1, so the sample average sits at 0 on average. For your chosen n the panel draws thousands of independent samples of that average, bins them into the pink distribution, and counts the fraction that clear the bar at a — that measured fraction is P̂, the odds of the miracle, computed and nothing else. The "odds against" is simply 1/P̂, and the measured exponent is −ln P̂ / n.
The amber line is the true rate function for a fair coin, I(a) = ½[(1+a)·ln(1+a) + (1−a)·ln(1−a)] — the same binary-entropy curve that governs coin-flip surprises. It is drawn from the formula, not fit to the data; the measured exponents fall toward it on their own as n grows, exactly as Cramér's theorem requires. When a setting is rarer than the trials can resolve — zero events out of everything sampled — the panel refuses to invent a number and reports the Monte-Carlo floor instead.
Where the metaphor tears
The clean exponential rate needs independent draws with light tails — no single contribution can dominate. Much of real luck is heavy-tailed: one viral post, one acquisition, one inheritance. There the miracle happens by a single giant jump, not a conspiracy of many small ones, the decay is merely polynomial, and I(a) simply doesn't apply. Ask whether your break was a thousand small leanings or one big roll.
The rate function is a fact about the ensemble — about repetition, replicas, the many samples the panel draws. Your single lucky life offers exactly one outcome and no way to estimate the exponent that governed it. The precise odds are a property of the model, not something legible in the one thing that actually happened.
I(a) tells you how rare a deviation of a given magnitude is — never which deviation, or what it means. And rare is not never: over billions of independent lives, exponentially unlikely breaks are certain to strike someone. Survivorship does the rest — you meet the winners, and mistake a paid-for tail event for a law.