How the Law of Large Numbers Works

Quick Answer

Theoretical probability counts outcomes: a fair coin gives heads with P = 1/2. Empirical probability measures experiments: the fraction of heads you actually saw. The law of large numbers connects them: as independent trials pile up, the empirical frequency converges on the theoretical value. For example, 10 flips can easily show 70 percent heads, but 10,000 flips will almost surely sit near 50 percent. Open the lab, run trials at up to 60 per second, and watch the convergence happen on the graph.

How the Law of Large Numbers Works

Every trial adds one data point, and the empirical frequency is the running average of them all. Early on, each new flip moves that average a lot: one streak of heads can push 10 flips to 0.8. But by trial 1,000 a single flip shifts the average by at most 0.001, so old streaks matter less and less. That dilution is the whole mechanism. Therefore the law promises convergence of the AVERAGE over the long run; it never promises the next flip, and it never balances the books after a streak.

• Theoretical: P(face) = 1/n from counting equally likely outcomes
• Empirical: successes divided by trials, measured by experiment
• Convergence comes from dilution, not correction

Why Small Samples Swing Wildly

The noise in an empirical frequency shrinks roughly with the square root of the sample size, so precision is expensive: four times the trials buys only half the wobble. For example, at 100 flips a range like 0.4 to 0.6 is routine, while at 10,000 flips the frequency rarely strays past 0.51. This square-root law is why polls quote margins of error, why quality control samples in bulk, and why a probability estimated from a handful of trials deserves no trust. The gap readout in the lab makes the slow grind visible.

Streaks and the Gambler's Fallacy

After five heads in a row, tails feels "due", but each flip is independent and the coin has no memory: P(tails) is still exactly 1/2. The law of large numbers gets misread as a promise of correction; really it is a promise of dilution. The five extra heads never get cancelled, they just become invisible inside ten thousand later flips. For example, run the lab past a streak and watch the graph drift back to the line without any answering streak in the other direction.

Using Empirical Probability in Practice

When outcomes are not equally likely (a bent coin, a real-world process), theory cannot hand you P, so measurement is all there is: run trials, count successes, divide. That is exactly what insurers, pollsters, and scientists do, with the square-root law dictating their sample sizes. In a classroom, project the lab as a coin flip simulator or a dice roll simulator and run a live probability experiment: every coin flip and dice roll lands in the tally with nobody flipping by hand. However, the lab uses fair objects on purpose: because you KNOW the true value is 1/n, you can watch how fast, and how reluctantly, an honest experiment reveals it.

Common Mistakes