The central limit theorem says that averages behave far more predictably than the values they average. Draw a random sample of size $n$ from almost any population with mean $\mu$ and standard deviation $\sigma$, and compute the sample mean $\bar{X}$. As $n$ grows, the distribution of $\bar{X}$ gets closer and closer to a normal distribution with mean $\mu$ and standard deviation $\sigma / \sqrt{n}$: $$\bar{X} \approx N\!\left(\mu, \frac{\sigma}{\sqrt{n}}\right).$$ OpenStax Introductory Statistics stresses the remarkable part: the original population can have any shape.

Without the fog

Roll one die and the result is flat: 1 through 6, each equally likely. Roll ten dice and average them, then do that again and again. Averages near 3.5 come up often, because many combinations of dice land there. Averages near 1 or 6 need every die to cooperate, so they are rare. Plot all those averages and a bell appears, even though a single die has no bell at all. OpenStax uses this same picture of repeatedly rolling ten dice.

The two numbers that matter

  • Center. The sample means are centered at the population mean $\mu$.
  • Spread. Their standard deviation, called the standard error of the mean, is $\sigma / \sqrt{n}$. OpenStax describes it as how far, on average, a sample mean lands from the population mean in repeated samples of size $n$.

The square root is the practical lesson. Quadrupling the sample size only halves the standard error.

Worked example 1: ten dice

One fair die has mean $\mu = 3.5$. Its variance is the average of the squared distances from 3.5: $$\frac{2.5^2 + 1.5^2 + 0.5^2 + 0.5^2 + 1.5^2 + 2.5^2}{6} = \frac{17.5}{6} = \frac{35}{12} \approx 2.917,$$ so $\sigma = \sqrt{35/12} \approx 1.708$. For the average of $n = 10$ dice, the standard error is $1.708 / \sqrt{10} \approx 0.540$. So averages of ten dice cluster around 3.5, and most land within about one unit of it, roughly between 2.4 and 4.6 (two standard errors either side).

Worked example 2: coin flips

Score a fair coin flip as 1 for heads and 0 for tails. One flip has mean $0.5$ and standard deviation $0.5$. The average of $n = 100$ flips is the proportion of heads, with standard error $0.5 / \sqrt{100} = 0.05$. By the normal approximation, about 95% of the time the proportion of heads falls within about $1.96 \times 0.05 \approx 0.098$ of one half, so roughly between 0.40 and 0.60. With 10,000 flips the standard error drops to 0.005, and the band tightens to roughly 0.49 to 0.51.

Worked example 3: a probability from OpenStax

OpenStax gives a population with unknown shape, mean 90 and standard deviation 15, and draws samples of size $n = 25$. The standard error is $15 / \sqrt{25} = 3$, so $\bar{X}$ is approximately $N(90, 3)$. To find the probability that the sample mean is between 85 and 92, convert to $z$ scores: $$z = \frac{85 - 90}{3} \approx -1.67, \qquad z = \frac{92 - 90}{3} \approx 0.67.$$ The area between them under the standard normal curve is about 0.70; OpenStax reports 0.6997. The value two standard errors above the mean is $90 + 2 \times 3 = 96$. You can check areas like this on the normal CDF calculator.

What the theorem does not say

  • It does not make individual data normal. A skewed population stays skewed. Only the distribution of means, or sums, becomes approximately normal.
  • $n$ is the sample size, not the number of samples. OpenStax is explicit: $n$ is the number of values averaged together, not the number of times the experiment is repeated.
  • It needs assumptions. The standard version assumes independent draws from the same distribution with a finite standard deviation. Very heavy tailed distributions can break it.
  • "Large enough" depends on the population. Fairly symmetric populations look normal in their means quickly; strongly skewed ones need larger samples. Rules of thumb such as 30 are guides, not guarantees.

Why it matters

Confidence intervals and many hypothesis tests for means rest on this theorem, because it tells you how much a sample mean wobbles from sample to sample. The same idea explains why a binomial distribution with many trials is well approximated by a normal curve: a binomial count is a sum of many independent 0 or 1 outcomes.

Checkpoints

  1. A population has mean 50 and standard deviation 12. What is the standard error for samples of size 36?
  2. How large must a sample be to cut the standard error in Checkpoint 1 in half?
  3. A population is strongly right skewed. Does the central limit theorem say a single observation is approximately normal? What does it say?
  4. Estimate the range that holds about 95% of the proportions of heads in samples of 400 fair coin flips.

Common mistakes

  • Dividing by $n$ instead of $\sqrt{n}$. The standard error uses the square root.
  • Using $\sigma$ when the question is about $\bar{X}$. Probabilities about a sample mean need the standard error, not the population standard deviation.
  • Confusing the law of large numbers with the central limit theorem. The law of large numbers says the sample mean gets close to $\mu$. The central limit theorem describes the shape and spread of its distribution along the way.
  • Applying it to tiny samples from skewed data. The approximation can be poor when $n$ is small and the population is far from symmetric.

Related reading

Related tools

The normal CDF calculator finds areas like the one in worked example 3, and the binomial probability calculator lets you compare an exact binomial answer with the normal approximation. Work out the standard error first.

Sources

Last reviewed: 12 November 2026.

Common questions

What does the central limit theorem say in simple terms?

If you take many random samples of the same size from almost any population and compute each sample's mean, those means pile up in a bell shape. The bigger the sample size, the closer that shape is to a normal distribution, centered at the population mean.

What is the standard error of the mean?

It is the standard deviation of the sample mean, equal to the population standard deviation divided by the square root of the sample size. It shrinks as samples get larger, so bigger samples give means that sit closer to the population mean.

Does the central limit theorem make the data normal?

No. The individual values keep whatever distribution they had. It is the distribution of sample means, or sums, that becomes approximately normal.

What does n mean in the central limit theorem?

n is the number of values averaged together in each sample, not the number of samples or repetitions of the experiment.