Law of Larger Numbers: Why Averages Settle

The law of larger numbers in everyday life

Imagine estimating how often a coin lands heads. After a short run, the fraction you see might feel surprisingly high or low. Keep flipping under the same conditions, and that fraction has a better chance of being close to the coin’s underlying probability. That is the idea people often mean by the law of larger numbers. Its standard name in statistics is the law of large numbers.

The law is about averages across many observations. For a fair coin, the average of heads indicators is also the proportion of heads: record heads as one outcome and tails as the other, then average those records. Under suitable conditions, that average gets closer to its expected value as the number of observations grows. OpenStax discusses this long-run behavior in its treatment of expected value.

This is a statement about what becomes likely, not a promise about the next result. A long series can contain unusual stretches, and adding an observation can move the running average either toward or away from its expected value. The distinction matters whenever someone says an outcome is “due” because recent results have gone the other way.

A synthetic coin-flip example

Here is a synthetic, illustrative sequence, not a report of an experiment. Suppose a fair coin lands heads on 7 of its first 10 synthetic flips. The illustrative heads proportion is 7 ÷ 10 = 0.70. That result is above the fair coin’s assumed heads probability of 0.50, but it does not tell us that tails must dominate the next set of flips.

Now suppose the same synthetic sequence contains 54 heads after 100 illustrative flips. Its running proportion is 54 ÷ 100 = 0.54. After 1,000 illustrative flips, suppose it contains 506 heads, for a proportion of 506 ÷ 1,000 = 0.506. Each checkpoint was chosen to make the calculation easy to inspect; none is empirical evidence or a typical result you should expect to reproduce.

In this invented sequence, the proportion moves closer to 0.50, but the law does not require every larger checkpoint to improve on the previous one. A different synthetic sequence might move away before moving closer again. Nor does the law say the counts of heads and tails must become equal. It concerns the proportion or average, whose fluctuations tend to matter less relative to a growing total.

Why short runs can look dramatic

A small sample leaves considerable room for chance variation. With only a few observations, each new outcome makes a noticeable difference to the average. As the total grows, one additional outcome has a smaller effect. This is why a running average can look restless at first and steadier later, even though the underlying process has not changed.

Consider a café manager tracking the share of customers who choose tea. One morning’s share could differ from the usual share simply because only a handful of customers have ordered. Looking across more customers may give a more stable estimate, provided the customers being counted represent the group the manager actually wants to understand. This is an everyday analogy, not a claim about any café’s sales.

Stability is not the same as certainty. A larger sample can still produce an average that differs from the underlying mean, and the law supplies no guarantee that a particular run will stay within a chosen margin. It describes long-run behavior under its assumptions. For an estimate you plan to act on, you still need to think about the sampling method and the uncertainty around the result.

What more observations cannot repair

The law of large numbers needs a suitable relationship between the observations and the quantity you want to estimate. If you ask only café customers who arrive before work whether they buy tea, adding more early-morning customers may make that group’s average more stable. It does not necessarily make the average representative of all customers throughout the day.

Changing conditions pose another problem. Suppose a café changes its menu midway through a survey. An average combining responses from before and after the change may describe that combined period, but it may not describe the current menu. Similarly, observations that strongly influence one another should not automatically be treated like independent repeats of an unchanged process.

Before collecting more data, name the target: whose average, over what period, under what conditions? Then ask how each observation enters the sample. A precise answer to the wrong question remains the wrong answer. More observations can reduce random fluctuation without removing a systematic mismatch between the sample and the population of interest.

A related idea that answers a different question

The law of large numbers concerns where a sample average tends to go as the sample grows. The central limit theorem addresses a different issue: under its conditions, it describes how the distribution of sample averages behaves across repeated samples. OpenStax treats that distribution in its section on the central limit theorem for sample means.

Picture repeatedly taking a fresh sample from the same process and calculating an average for each sample. You could ask whether those averages cluster around the population mean, or you could ask what shape their distribution takes. The first question points toward the law of large numbers; the second points toward the central limit theorem. They are related, but one is not a substitute for the other.

This distinction is useful when reading a poll or a performance dashboard. Seeing an average settle over time is not, by itself, a complete measure of uncertainty. To judge how far an estimate might be from the target, you need additional information about variation and how the sample was collected. A calm-looking chart can still be based on a biased source.

How to use the idea responsibly

Start with the process, not the chart. Identify what an observation means, whether the process stays reasonably consistent, and whether the observations speak to the population you care about. Then treat a growing average as evidence that random swings may have less influence, not as proof that every concern has disappeared.

When someone argues that a result is bound to reverse after an unusual run, check whether they are confusing a long-run average with a prediction of the next outcome. When someone argues that a huge dataset must be accurate, check who or what was left out. The law of large numbers is most useful when paired with those questions.

Practical checklist:

• Define the population, outcome, and average you want to estimate.
• Check whether observations come from a suitable, reasonably consistent process.
• Expect a larger sample to help with random fluctuation, not to guarantee a smooth running average.
• Look for selection bias or changing conditions before trusting a stable-looking result.