Chi-squared, t, F distributions — definition, degrees of freedom: Visual walkthrough

Step 1 — One bell curve, and what "squaring" does to it

WHAT. We begin with a single random number $Z$ . The symbol $Z$ just means "a number drawn from the standard bell curve": most draws land near $0$ , big draws (far from $0$ ) are rare. We then look at $Z^{2}$ — the number squared.

WHY. Squaring throws away the sign. A miss of $- 2$ and a miss of $+ 2$ are equally bad if we only care about distance from the centre. Squaring turns "how far off, left or right?" into "how much wobble, period?". That is the seed of everything: $χ^{2}$ will be nothing but added-up squared wobble.

PICTURE. Look at the figure. The blue bell is $Z$ . The two yellow points at $- 1.5$ and $+ 1.5$ both map (red arrows) to the same height $2.25$ on the squared axis — sign erased. Notice small $Z$ barely moves, large $Z$ explodes: squaring magnifies rare far draws, which is why the total will have a long right tail.

Figure — Chi-squared, t, F distributions — definition, degrees of freedom

Step 2 — Stacking $k$ independent wobbles → the $χ_{k}^{2}$ shape

WHAT. Take $k$ independent bell-draws $Z_{1}, \dots, Z_{k}$ , square each, and add them: $χ_{k}^{2} = wobble 1 Z_{1}^{2} + wobble 2 Z_{2}^{2} + \dots + wobble k Z_{k}^{2} .$

WHY. Independent means each draw ignores the others — they are $k$ separate directions of variation, i.e. $k$ degrees of freedom. Adding squared distances is exactly the squared length of a point in $k$ -dimensional space (Pythagoras). So $χ_{k}^{2}$ = "how far a random $k$ -dimensional dart lands from the origin, squared".

PICTURE. The figure shows the curve for $k = 1, 2, 5, 10$ . For $k = 1$ it piles up near $0$ (one small wobble is usually tiny). As $k$ grows the peak slides right to about $k$ and the curve fattens — more wobbles to add means a bigger, more spread total.

Step 3 — The df constraint: why $n - 1$ , not $n$

WHAT. In real data we don't know the true centre $μ$ ; we use the sample mean $\overset{x}{ˉ} = \frac{1}{n} \sum x_{i}$ . The deviations $x_{i} - \overset{x}{ˉ}$ then obey a hidden rule: $\sum_{i = 1}^{n} (x_{i} - \overset{x}{ˉ}) = 0.$

WHY. Once you know $n - 1$ of the deviations, the last one is forced — it must be whatever makes the sum zero. One piece of freedom has been spent to pin down $\overset{x}{ˉ}$ . So the wobble lives in only $n - 1$ free directions → $χ_{n - 1}^{2}$ .

PICTURE. With $n = 3$ the three deviations must sum to $0$ : that is a flat plane through the origin (green) inside 3D space. The point is trapped on that plane — free to move in $2 = n - 1$ directions, not $3$ . That trapped dimension count is the degrees of freedom.

Step 4 — Standardise the mean: a clean $Z$ appears

WHAT. The sample mean $\overset{x}{ˉ}$ itself has a bell curve, narrower than the data by $n$ . Standardising it gives an exact $Z$ : $\frac{x ˉ - μ}{σ / n} \sim N (0, 1) = Z .$

WHY. $\overset{x}{ˉ}$ is an average of $n$ normals, so it's normal with centre $μ$ and spread $σ / n$ (averaging cancels wobble). Subtracting its centre and dividing by its spread rescales it to the standard bell — the numerator we need for $t$ .

PICTURE. Two bells: the wide data bell (width $σ$ ) and the narrow sample-mean bell (width $σ / n$ ). The arrow shows subtracting $μ$ (shift to centre) then dividing by $σ / n$ (rescale to width $1$ ) collapsing it onto the unit bell.

Step 5 — Kill the unknown $σ$ : the $t$ ratio is born

WHAT. We don't know $σ$ , so we can't use the $Z$ of Step 4 directly. Replace $σ$ by its estimate $s$ . Watch $σ$ cancel: $\frac{x ˉ - μ}{s / n} = \frac{( x ˉ - μ ) / ( σ / n )}{s / σ} = \frac{Z Step 4}{\frac{( n - 1 ) s ^{2} / σ ^{2}}{n - 1}} = \frac{Z}{V / k}, V = \frac{( n - 1 ) s ^{2}}{σ ^{2}} \sim χ_{k}^{2}, k = n - 1.$

WHY. Dividing top and bottom by $σ$ makes the top exactly the $Z$ from Step 4 and the bottom exactly a $χ_{n - 1}^{2}$ from Step 3, divided by its df and square-rooted. That bottom is a random estimate of $σ$ — it wobbles.

PICTURE. Split screen. Left: fixed denominator $= 1$ gives the crisp normal. Right: the denominator $V / k$ is a rattling value; when it happens to be small, dividing blows the ratio up — the red spikes far out in the tails. That randomness in the denominator is the fat tail.

Step 6 — Limiting case: why $t_{k} \to Z$ as $k \to \infty$

WHAT. As the df $k$ grows, the denominator $V / k$ stops rattling and locks onto $1$ , so $t_{k}$ becomes an ordinary $Z$ .

WHY. $V / k$ is an average of $k$ squared normals, each with mean $1$ . By the law of large numbers (kin to the Central Limit Theorem), averaging many of them pins $V / k \to 1$ , so $V / k \to 1$ and the ratio's denominator is effectively fixed — no more fat tails.

PICTURE. The $t$ curves for $k = 2, 5, 30$ over the standard normal (dashed). Small $k$ : visibly fat tails and a lower peak. $k = 30$ : almost perfectly on the normal. The tails deflate as $k$ climbs.

Step 7 — Two wobbles compared: the $F$ ratio

WHAT. Take two independent chi-squareds $U \sim χ_{d_{1}}^{2}$ and $V \sim χ_{d_{2}}^{2}$ . Normalise each to "per degree of freedom", then divide: $F_{d_{1}, d_{2}} = \frac{U / d _{1} avg wobble, group 1}{avg wobble, group 2 V / d _{2}} .$

WHY. A raw $χ_{k}^{2}$ has mean $k$ — bigger groups look "bigger" for free. Dividing each by its own df resets both to mean $\approx 1$ , so the ratio measures how many times more wobbly group 1 is than group 2, fairly. If the two populations truly share a variance, both parts sit near $1$ and $F \approx 1$ .

PICTURE. Two normalised chi-squared bars, each hovering near height $1$ . The ratio dial points near $1$ under equal variance; when group 1 is genuinely more spread, the numerator swells and the dial swings above $1$ . This is the engine inside ANOVA and Hypothesis Testing for variances.

The one-picture summary

Square → Sum (with a constraint) → Scale & Split. One bell curve $Z$ , squared and stacked, becomes $χ^{2}$ ; a lone $Z$ over a scaled $χ^{2}$ becomes $t$ ; two scaled $χ^{2}$ 's divided become $F$ . Every arrow below is one step you just walked.

Recall Feynman: the whole walkthrough in plain words

Start with one shaky number from the bell curve. Square it so left-and-right misses count the same — that's a wobble score. Add up $k$ independent wobbles and you get chi-squared; its typical size is just $k$ , because each wobble averages $1$ . If you first had to compute an average from your data, one wobble-direction gets used up, so you count $n - 1$ free ones, not $n$ . Now, to test a mean when you don't know your true steadiness, you estimate it — and because that estimate itself rattles, dividing by it makes crazy answers more common: fat-tailed $t$ . Throw enough data at it and the estimate stops rattling, so $t$ quietly becomes the plain bell again. Finally, to ask "is player A wobblier than player B?", make each player's average wobble about $1$ by dividing by its own free-count, then take the ratio: $F$ , which sits near $1$ when they're equally shaky. Square, sum, scale, split — that's the entire family.

Recall Quick self-check

Why do we square $Z$ instead of using $∣ Z ∣$ ? ::: Squaring gives a smooth, differentiable "distance²" that is exactly a Pythagorean squared length in $k$ dimensions and has clean mean/variance ( $1$ and $2$ ). Why does the sample-variance chi-squared have $n - 1$ df? ::: Estimating $\overset{x}{ˉ}$ forces $\sum (x_{i} - \overset{x}{ˉ}) = 0$ , trapping the deviations on an $(n - 1)$ -dimensional plane — one free direction lost. What makes $t$ 's tails fat? ::: Its denominator $V / k$ is a random estimate of $σ$ ; when it's small by chance the ratio blows up. What is $t_{k}^{2}$ ? ::: $F_{1, k}$ .