Take a plain standard normal — the reference bell centred at 0, one unit wide. Square it. Grab another independent standard normal, square that too. Keep going for a handful of them, then add all the squares together. The pile of numbers you get has a distribution of its very own — and it turns out to be one of the most useful curves in all of statistics. It is the chi-square distribution, written \chi^2.

Formally: if Z_1, Z_2, \dots, Z_k are independent standard normals, then

The single number k — how many squared normals we added up — is called the degrees of freedom. It is the only knob this distribution has, and it controls everything about the shape. That deceptively simple "square-and-add" recipe is the engine behind almost every test that asks "does my data actually fit what I expected?"

Three facts that fall straight out of the recipe

You can read most of the chi-square's personality directly off the definition, before touching any formula.

• It can never be negative. Every term is a square, and squares are \ge 0. So the whole sum lives on [0, \infty) — there is no left tail at all. A chi-square value of -3 is impossible.
• Its mean is k. A single squared standard normal Z^2 averages out to exactly 1 (that's just the variance of a standard normal, \operatorname{Var}(Z)=1). Add k of them and the averages add: \mathbb{E}[X] = k. The mean of a chi-square is its degrees of freedom.
• Its variance is 2k. Each Z^2 contributes a variance of 2, and for independent terms variances also add, giving \operatorname{Var}(X) = 2k. So the spread grows with k, but the standard deviation \sqrt{2k} grows only like \sqrt{k} — slower than the mean.

Because the mean climbs like k while the spread climbs only like \sqrt{2k}, a large-k chi-square is a hump sitting a long way from zero, comfortably clear of the wall at 0 — which is exactly why it stops looking lopsided and starts looking like a bell (more on that below).

Watch the shape change with k

Drag the degrees of freedom slider and watch the curve transform. At k = 1 and k = 2 the density is jammed up against zero — it's spiked near the origin and slides down a long right tail. As you raise k, the hump peels away from the wall, drifts to the right (remember, the mean sits at k), and grows steadily more symmetric: by k = 8 or so it already looks a lot like a gently lopsided bell.

The shaded band on the right is the upper tail beyond a cutoff c. That region is the whole point of a chi-square test: a big chi-square value lands out there, and "how much area is past c" is the probability of seeing a mismatch at least this large by chance. Notice how, for tiny k, almost nothing reaches past c, while for large k the whole hump has marched out to meet it.

Worked example 1 — one squared normal is \chi^2_1

Start at the very smallest case, k = 1. Here X = Z^2: take a single standard normal and square it. What should this distribution look like, and what is its mean?

Reason it through. A standard normal Z is usually close to 0 (values near the peak), so Z^2 is usually a very small positive number — that is the tall spike at the origin. Occasionally Z is out near \pm 2 or \pm 3, and squaring sends those to 4 or 9 — that's the long thin right tail. And because both +2 and -2 square to the same 4, the sign information is thrown away entirely.

For the mean, use the identity \mathbb{E}[Z^2] = \operatorname{Var}(Z) + (\mathbb{E}[Z])^2 = 1 + 0^2 = 1. So the average of a \chi^2_1 is exactly 1 — matching our rule "mean = k." If you simulated a million squared standard normals and averaged them, you'd get a number hugging 1, even though the typical (most likely) value is much nearer 0. That gap between the crowded-near-zero mode and the mean of 1 is the right-skew in a nutshell.

Worked example 2 — adding three of them, and reading the tail

Now let k = 3: three independent standard normals, each squared, all added up,

By our rules the mean is k = 3 and the variance is 2k = 6 (so standard deviation \sqrt{6}\approx 2.45). The curve still starts at the origin and is still right-skewed, but the hump has lifted off the wall and now peaks around x = 1, trailing off toward the right.

Here's how a test uses it. Suppose a calculation on some real data gives a chi-square statistic of X = 9.0. Is that "surprisingly large" for \chi^2_3? We look up the upper-tail critical value: for 3 degrees of freedom, only 5\% of the distribution lies beyond c = 7.81. Our 9.0 > 7.81 sits out in that 5\% tail — a result this extreme happens under "everything's fine" less than 5\% of the time — so at the 5\% level we'd flag the data as a poor fit. Every chi-square test is exactly this move: compute a statistic, compare it to an upper-tail cutoff.

What is it actually used for?

The chi-square distribution is the reference distribution — the yardstick you compare your statistic against — for a whole family of everyday tests:

• Goodness-of-fit. Does a die (or a genetics ratio, or a survey) match the counts you expected? You form \sum \frac{(\text{observed} - \text{expected})^2}{\text{expected}} — a sum of squared, scaled deviations — and it follows a chi-square. See the chi-square goodness-of-fit test.
• Independence / association. In a two-way table of counts (e.g. treatment vs. outcome), are the two variables related, or independent? The same observed-minus-expected machinery gives a chi-square statistic.
• Inference about a variance. Because sample variances are built from squared deviations from a mean, the sampling distribution of a (scaled) sample variance is chi-square — so it's the tool for testing or building intervals for a population variance \sigma^2.

In every case the logic is the same: squared deviations, summed, judged against the upper tail of \chi^2.