The Central Limit Theorem for Means

The central limit theorem is simply the fact that averages of many random values follow a bell curve, even when the individual values don't. It's "central" because it sits at the center of statistics (nearly every test and interval in this course leans on it), and "limit" because the bell shape gets more exact as the sample size grows. The statement $\bar X \approx N\!\left(\mu,\ \frac{\sigma}{\sqrt n}\right)$ looks dense, but read it piece by piece: $\bar X$ (the average of one random sample), $\approx N$ (is roughly normal, a bell curve), $\mu$ (centered at the true population mean) and $\frac{\sigma}{\sqrt n}$ (with a spread that shrinks as the sample size $n$ grows). That's four simple ideas in one line, and it's what lets us make confident statements about a whole population from one sample. The rest of this lesson just unpacks them.

The sampling distribution. Imagine taking a sample, computing $\bar x$, then doing it again with a fresh sample. You'd get a slightly different $\bar x$ each time. The distribution of all those possible values is the sampling distribution of the mean. The CLT says:

1. Its mean is $\mu$: sample means are unbiased.
2. Its standard deviation, the standard error, is $\frac{\sigma}{\sqrt n}$.
3. It's approximately normal for large $n$ (commonly $n \ge 30$), even if the population is skewed. If the population itself is normal, it's exactly normal for every $n$.

Why does averaging help? Think about what happens when you average. One unusually big value gets diluted by all the ordinary ones around it, and big and small values tend to cancel. The more values you average, the harder it is for the result to stray far from $\mu$. That's the standard error shrinking. And the bell shape shows up because the many small, independent ups and downs add together, the same way they do for anything built out of lots of little random pieces.

The $\sqrt n$ law: to cut the standard error in half you need four times as much data. It's $\sqrt n$, not $n$, so precision gets expensive fast.

Using it: standardize with the standard error: $z = \dfrac{\bar x - \mu}{\sigma/\sqrt n}$. It's the same z-score you already know, except the spread is the spread of averages, not of single values.

A common mistake. It's tempting to think the CLT makes the data themselves normal once you have enough of them. It doesn't. A big sample from a skewed population is still skewed. What becomes normal is the distribution of the sample mean $\bar x$, across many imagined samples.

Polls, quality control charts, A/B tests and confidence intervals all rely on the CLT: it's why the normal distribution shows up for averages of almost anything.