The idea in one sentence
If you repeat the same random process many times, the proportion of times each outcome occurs tends to get closer to that outcome’s true probability. That is the law of large numbers, and it is the reason a probability is a meaningful number at all.
A fair coin has a 50% chance of heads. You will not get exactly five heads in ten flips most of the time, and you will not get exactly 500 in 1,000. But the share of heads, as a fraction of all flips, tends to hover closer and closer to one half as the flips pile up.
How quickly it settles
The spread of a proportion shrinks as the number of trials grows. For a fair coin, a typical gap between the observed share of heads and 50% (one standard deviation) is about the square root of 0.25 / n, where n is the number of flips.
| Flips | Typical gap in the share of heads |
|---|---|
| 100 | about 5 percentage points |
| 1,000 | about 1.6 percentage points |
| 10,000 | about 0.5 percentage points |
So after 100 flips, a result anywhere between about 45% and 55% heads is ordinary. After 10,000 flips, a share more than a point or so away from 50% would be surprising. Each tenfold increase in trials narrows the typical gap by a factor of roughly three, since the square root of 10 is a little over 3.
A dice example
Roll a fair die 6,000 times. You would expect about 1,000 sixes. The standard deviation of the count is the square root of 6,000 × (1/6) × (5/6), which is about 28.9. A result within roughly two standard deviations, so about 942 to 1,058, would be unremarkable. As a share of rolls, that is about 15.7% to 17.6%, compared with the true 16.67%.
Notice that the count can easily miss 1,000 by dozens, yet the proportion is already close to one in six.
What the law does not say
This is where most confusion arises.
It does not say luck evens out. After ten heads in a row, the next flip is still 50% heads. The coin has no memory and no obligation to catch up. Believing otherwise is the gamblers fallacy. What actually happens is dilution: the early streak is swamped by the thousands of flips that follow, so its share of the total fades, even though it is never cancelled.
It does not say raw differences shrink. The gap between the number of heads and the expected half typically grows with more flips, roughly in proportion to the square root of the flips. After 100 flips, the typical difference from an even split is about 5 heads. After 10,000 flips, it is about 50. The share gets closer to 50%, and the count gap gets larger. Both statements are true at once.
It does not guarantee anything about a short run. Whether your next hour of play matches the probabilities is a matter of chance, and it often will not. The behaviour of small samples is a separate topic with its own cautions.
It does not apply to processes with changing odds. If probabilities shift, for example when cards leave a deck, the frequencies settle toward the true, changing probabilities, not a single fixed one.
Why it matters for games
The law of large numbers is why expected value is useful: over many repeats, the average result settles near the expected value. It is also why a casino with many bets per day sees results close to its mathematical margin, while an individual player over a single evening can see something very different. The house edge works through large numbers; it makes no claim about any one session.
For designers, the law explains why games with many small random events feel steadier than games with a few large ones. For players tracking their own results, it explains why a win rate from five matches says little, and one from five hundred says considerably more. Tools such as confidence intervals put numbers on how much confidence a given sample deserves.
A practical way to see it
Simulation makes the law tangible. A short program or spreadsheet that rolls a die thousands of times and plots the running share of sixes will show a wobbly line early on that gradually flattens toward 1/6. Running the experiment several times shows the early wobble differs each time while the later behaviour looks much the same. That convergence, not any sort of correction mechanism, is the whole content of the law.
Remember its scope. It describes what happens to proportions over very many independent repeats. It offers no way to predict a single outcome and no way to turn a game with a negative expectation into a profitable one.