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

4.9.9 · D2 · Maths › Probability Theory & Statistics › Chi-squared, t, F distributions — definition, degrees of fre

---

Step 1 — Ek bell curve, aur "squaring" uske saath kya karta hai

KYA. Hum ek single random number $Z$ se shuru karte hain. Symbol $Z$ ka matlab sirf itna hai "standard bell curve se draw kiya hua number": zyaadatar draws $0$ ke paas land karte hain, bade draws (jo $0$ se door hain) rare hote hain. Phir hum $Z^2$ dekhte hain — woh number squared.

KYUN. Squaring sign ko throw away kar deta hai. $-2$ ki miss aur $+2$ ki miss dono equally buri hain agar hum sirf centre se distance ki parwah karte hain. Squaring "kitna door, left ya right?" ko "kitna wobble, bas?" mein badal deta hai. Yahi sab kuch ka seed hai: ${\chi }^2$ kuch nahi hoga sirf jode gaye squared wobble ke alawa.

PICTURE. Figure dekho. Blue bell $Z$ hai. $-1.5$ aur $+1.5$ par do yellow points dono (red arrows) squared axis par usi height $2.25$ par map hote hain — sign erase ho gaya. Dekho small $Z$ barely hilta hai, large $Z$ explode karta hai: squaring rare far draws ko magnify karta hai, isliye total ka ek lamba right tail hoga.

---

Step 2 — $k$ independent wobbles stack karna → ${\chi }_k^2$ shape

KYA. $k$ independent bell-draws $Z_1,\dots ,Z_k$ lo, har ek ko square karo, aur add karo: ${\chi }_k^2=\underset{\text{wobble 1}}{\underbrace{Z_1^2}}+\underset{\text{wobble 2}}{\underbrace{Z_2^2}}+\cdots +\underset{\text{wobble }k}{\underbrace{Z_k^2}}.$

KYUN. Independent matlab har draw doosron ko ignore karta hai — ye $k$ alag-alag directions of variation hain, yaani $k$ degrees of freedom. Squared distances add karna exactly $k$-dimensional space mein ek point ki squared length hai (Pythagoras). Toh ${\chi }_k^2$ = "ek random $k$-dimensional dart origin se kitni door land karta hai, squared".

PICTURE. Figure $k=1,2,5,10$ ke liye curve dikhata hai. $k=1$ ke liye $0$ ke paas pile up hota hai (ek small wobble usually tiny hota hai). Jaise $k$ badhta hai peak right ki taraf $k$ ke aas-paas slide karta hai aur curve fatten hota hai — zyaada wobbles add karne se bada, zyaada spread total milta hai.

---

Step 3 — df constraint: $n-1$ kyun, $n$ nahi

KYA. Real data mein hum true centre $\mu$ nahi jaante; hum sample mean $\bar{x}=\frac{1}{n}\sum x_i$ use karte hain. Deviations $x_i-\bar{x}$ phir ek hidden rule follow karte hain: ${\sum }_{i=1}^n(x_i-\bar{x})=0.$

KYUN. Jab ek baar $n-1$ deviations pata hain, aakhri wala forced hai — woh hona chahiye jo bhi sum ko zero banaye. Freedom ka ek piece $\bar{x}$ ko pin karne mein spend ho gaya. Toh wobble sirf $n-1$ free directions mein rehta hai → ${\chi }_{n-1}^2$.

PICTURE. $n=3$ ke saath teen deviations ka sum $0$ hona chahiye: yeh origin se guzarne wala ek flat plane (green) hai 3D space ke andar. Point us plane par trapped hai — $2=n-1$ directions mein move kar sakta hai, $3$ mein nahi. Woh trapped dimension count hi degrees of freedom hai.

---

Step 4 — Mean ko standardise karo: ek clean $Z$ appear hota hai

KYA. Sample mean $\bar{x}$ ka khud ek bell curve hota hai, data se $\sqrt{n}$ se narrower. Usse standardise karne par exact $Z$ milta hai: $\frac{\bar{x}-\mu }{\sigma /\sqrt{n}}\sim N(0,1)=Z.$

KYUN. $\bar{x}$ $n$ normals ka average hai, toh yeh normal hai centre $\mu$ aur spread $\sigma /\sqrt{n}$ ke saath (averaging wobble cancel karta hai). Uska centre subtract karke aur uske spread se divide karke isse the standard bell par rescale karta hai — numerator jo $t$ ke liye chahiye.

PICTURE. Do bells: wide data bell (width $\sigma$) aur narrow sample-mean bell (width $\sigma /\sqrt{n}$). Arrow dikhata hai $\mu$ subtract karna (centre par shift) phir $\sigma /\sqrt{n}$ se divide karna (width $1$ par rescale) use unit bell par collapse karta hai.

---

Step 5 — Unknown $\sigma$ ko khatam karo: $t$ ratio janam leta hai

KYA. Hum $\sigma$ nahi jaante, toh Step 4 ka $Z$ directly use nahi kar sakte. $\sigma$ ki jagah uska estimate $s$ rakho. Dekho $\sigma$ cancel ho jaata hai: $\frac{\bar{x}-\mu }{s/\sqrt{n}}=\frac{(\bar{x}-\mu )/(\sigma /\sqrt{n})}{s/\sigma }=\frac{\stackrel{\text{Step 4}}{\overbrace{Z}}}{\sqrt{\frac{(n-1)s^2/{\sigma }^2}{n-1}}}=\frac{Z}{\sqrt{V/k}},V=\frac{(n-1)s^2}{{\sigma }^2}\sim {\chi }_k^2,k=n-1.$

KYUN. Top aur bottom dono ko $\sigma$ se divide karne par top exactly Step 4 ka $Z$ ban jaata hai aur bottom exactly Step 3 ka ek ${\chi }_{n-1}^2$, apne df se divided aur square-rooted. Woh bottom $\sigma$ ka ek random estimate hai — yeh wobble karta hai.

PICTURE. Split screen. Left: fixed denominator $=1$ crisp normal deta hai. Right: denominator $\sqrt{V/k}$ ek rattling value hai; jab yeh chance se small hota hai, divide karna ratio ko blow up karta hai — red spikes tails mein bahut door. Denominator mein woh randomness hi fat tail hai.

---

Step 6 — Limiting case: $t_k\to Z$ as $k\to \infty$ kyun

KYA. Jaise df $k$ badhta hai, denominator $\sqrt{V/k}$ rattling band kar deta hai aur $1$ par lock ho jaata hai, toh $t_k$ ek ordinary $Z$ ban jaata hai.

KYUN. $V/k$ $k$ squared normals ka average hai, har ek ka mean $1$ hai. Law of large numbers se (kin to the Central Limit Theorem), unhe average karne par $V/k\to 1$ pin ho jaata hai, toh $\sqrt{V/k}\to 1$ aur ratio ka denominator effectively fixed hai — ab fat tails nahi.

PICTURE. Standard normal (dashed) ke upar $k=2,5,30$ ke liye $t$ curves. Small $k$: visibly fat tails aur lower peak. $k=30$: almost perfectly normal par. Tails deflate hote hain jaise $k$ badhta hai.

---

Step 7 — Do wobbles compare karna: $F$ ratio

KYA. Do independent chi-squareds $U\sim {\chi }_{d_1}^2$ aur $V\sim {\chi }_{d_2}^2$ lo. Har ek ko "per degree of freedom" par normalise karo, phir divide karo: $F_{d_1,d_2}=\frac{\stackrel{\text{avg wobble, group 1}}{\overbrace{U/d_1}}}{\underset{\text{avg wobble, group 2}}{\underbrace{V/d_2}}}.$

KYUN. Ek raw ${\chi }_k^2$ ka mean $k$ hota hai — bade groups "bade" lagte hain free mein. Har ek ko uske apne df se divide karna dono ko mean $\approx 1$ par reset karta hai, toh ratio measure karta hai group 1 group 2 se kitni zyaada wobbly hai, fairly. Agar do populations truly ek variance share karti hain, dono parts $1$ ke paas hote hain aur $F\approx 1$.

PICTURE. Do normalised chi-squared bars, dono height $1$ ke aas-paas hover karte hain. Ratio dial $1$ ke paas point karta hai equal variance ke under; jab group 1 genuinely zyaada spread hoti hai, numerator swell karta hai aur dial $1$ se upar swing karta hai. Yeh ANOVA aur variances ke liye Hypothesis Testing ke andar ka engine hai.

---

Ek-picture summary

Square → Sum (ek constraint ke saath) → Scale & Split. Ek bell curve $Z$, squared aur stacked, ${\chi }^2$ ban jaata hai; ek akela $Z$ scaled ${\chi }^2$ ke upar $t$ ban jaata hai; do scaled ${\chi }^2$'s divided $F$ ban jaate hain. Neeche har arrow ek step hai jo tum abhi walk kar ke aaye ho.

Recall Feynman: poora walkthrough simple words mein

Ek shaky number se shuru karo bell curve se. Usse square karo taaki left-aur-right misses same count hon — woh ek wobble score hai. $k$ independent wobbles add karo aur chi-squared milta hai; uska typical size sirf $k$ hai, kyunki har wobble average $1$ hota hai. Agar tumne pehle apne data se average compute kiya tha, ek wobble-direction use ho jaati hai, toh tum $n-1$ free ones count karte ho, $n$ nahi. Ab, jab mean test karna ho aur tumhe apni true steadiness nahi pata, tum usse estimate karte ho — aur kyunki woh estimate khud rattle karta hai, usse divide karne par crazy answers zyaada common ho jaate hain: fat-tailed $t$. Itna data daalo ki estimate rattle karna band kar de, toh $t$ quietly plain bell ban jaata hai. Finally, poochhne ke liye "kya player A player B se wobblier hai?", har player ki average wobble ko $1$ ke aas-paas karo usse apne free-count se divide karke, phir ratio lo: $F$, jo $1$ ke paas baitha rehta hai jab dono equally shaky hain. Square, sum, scale, split — yahi puri family hai.

Recall Quick self-check

Hum $Z$ ko square kyun karte hain $|Z|$ ki jagah? ::: Squaring ek smooth, differentiable "distance²" deta hai jo exactly $k$ dimensions mein ek Pythagorean squared length hai aur clean mean/variance ($1$ aur $2$) rakhta hai. Sample-variance chi-squared mein $n-1$ df kyun hote hain? ::: $\bar{x}$ estimate karna $\sum (x_i-\bar{x})=0$ force karta hai, deviations ko ek $(n-1)$-dimensional plane par trap karta hai — ek free direction lost. $t$ ke tails mote kyun hote hain? ::: Uska denominator $\sqrt{V/k}$ $\sigma$ ka ek random estimate hai; jab woh chance se small hota hai toh ratio blow up karta hai. $t_k^2$ kya hai? ::: $F_{1,k}$.