Law of Large Numbers vs. the Gambler's Fallacy

A cat has discovered a suspended toy that swings freely. She bats it left or right with equal probability on each swipe, and she swipes it a lot. After a thousand swipes, the count is close to 500 left and 500 right. After ten thousand swipes, closer still to the 50/50 mark.

That is the law of large numbers. Now suppose she has just batted the toy left five times in a row. Someone watching concludes that a rightward swipe is now overdue, that the universe is correcting for the imbalance.

That is the gambler's fallacy. The toy has no memory.

The Law of Large Numbers

The weak law of large numbers states that for a sequence of independent, identically distributed random variables $X_1, X_2, \ldots$ with finite mean $\mu$, the sample mean converges in probability to $\mu$:

$$\bar{X}_n = \frac{1}{n}\sum_{i=1}^n X_i \xrightarrow{p} \mu$$

Convergence in probability means that for any $\varepsilon > 0$, the probability that $\bar{X}_n$ deviates from $\mu$ by more than $\varepsilon$ goes to zero as $n \to \infty$:

$$P\!\left(|\bar{X}_n - \mu| > \varepsilon\right) \to 0$$

The strong law sharpens this: convergence holds almost surely, meaning the sample mean converges to $\mu$ with probability one, not just for a specific $\varepsilon$. The strong law is a harder result to prove and a stronger statement to make, but the practical message is the same.

For the cat with the toy, each swipe is a Bernoulli trial with $p = 0.5$. The expected proportion of leftward swipes is 0.5. The law guarantees that over enough trials, the observed proportion will be arbitrarily close to that value.

What Independence Actually Means

The key condition in both laws is independence. Trial $i$ is independent of trial $j$ if knowing the outcome of $j$ tells you nothing about the probability of any outcome of $i$.

This is not a statement about the sequence of outcomes. It is a statement about the mechanism that generates them. The toy does not remember its previous position. The cat's paw does not carry a debt from past results. Each swipe is generated by the same process with the same probability, regardless of history.

Independence means that $P(X_{n+1} = k \mid X_1, X_2, \ldots, X_n) = P(X_{n+1} = k)$. The conditional probability equals the unconditional probability. Past outcomes contain no information about future ones.

The Fallacy

The gambler's fallacy is treating a theorem about long-run averages as if it generates corrective pressure on individual trials.

After five leftward swipes, the proportion of lefts is 5/5 = 1.0, well above the expected 0.5. The law of large numbers says this proportion will approach 0.5 over many trials. The fallacy is concluding that the next trial must therefore be right-biased to correct the imbalance.

That is not how the convergence works. What actually happens is dilution: the five extra lefts get divided by an increasingly large denominator. After 1,000 more swipes split evenly, the proportion is 505/1005 ≈ 0.502. The early excess is swamped by subsequent data, not reversed by it. No individual trial is correcting anything.

The fallacy mistakes a statistical property of aggregates for a causal constraint on individual outcomes. These are different things.

The Hot-Hand Fallacy

The opposite error also exists. In basketball, observers have historically believed that a player who has made several consecutive shots is "hot" and more likely to make the next one. This is called the hot-hand fallacy: treating independent trials as positively correlated when they are not.

For a long time, the research consensus was that the hot hand was entirely illusory. More recent work suggests the picture is complicated and that genuine skill streaks may exist in some contexts. The lesson is not that streaks never carry information. It is that claiming they do requires evidence of dependence, not just the existence of the streak.

If trials are independent, past outcomes are irrelevant in both directions. The streak proves nothing about what comes next.

Where the Confusion Comes From

The law of large numbers is a statement about a ratio approaching a limit. The gambler's fallacy is a claim about the next term in a sequence. Confusing them means confusing the behavior of an aggregate with the behavior of a single draw.

The cat's lifetime ratio of left-to-right swipes will approach 1:1. That convergence is guaranteed. The next swipe is still a fair coin flip. Both things are true, and they do not conflict.

The theorem describes what happens over many trials. It does not describe what must happen next.