Central limit theorem: Question bank — Hinglish deep dive

1.3.15 · D5 · AI-ML › Probability & Statistics › Central limit theorem

Quick symbol refresher taaki koi bhi cheez yahan use hone se pehle uska matlab samajh aaye:

$X_{1}, \dots, X_{n}$ — raw samples, i.i.d. (independent, same distribution).
$μ$ — true population mean; $σ^{2}$ — true population variance.
$\overset{ˉ}{X}_{n} = \frac{1}{n} \sum_{i} X_{i}$ — sample mean (size $n$ ke ek sample ka ek number).
$σ / n$ — standard error (SE): $\overset{ˉ}{X}_{n}$ kitna $μ$ ke aas-paas wobble karta hai.
$N (0, 1)$ — standard normal (bell curve, mean 0, spread 1).

True or false — justify

CLT se raw data

X_{i}

normally distributed ho jaata hai jaise

n

badhta hai.

False. Individual

X_{i}

kabhi shape nahi badlte — ek die hamesha uniform rahega. Sirf average

\overset{ˉ}{X}_{n}

ki distribution normal ke kareeb aati hai.

Bada sample size sample mean ki distribution ko wider aur zyada spread out banata hai.

False. SE

= σ / n

jaise

n

badhta hai shrink karta hai, isliye

\overset{ˉ}{X}_{n}

ki distribution narrower hoti hai aur

μ

par tighter concentrate hoti hai.

CLT ke liye underlying distribution symmetric honi chahiye.

False. Ye skewed distributions (exponential, Bernoulli with

p = 0.05

) ke liye bhi kaam karta hai; skew ka matlab bas itna hai ki bell shape convincing hone se pehle tumhe bada

n

chahiye.

Agar

n = 30

hai, toh sample mean exactly normal hai.

False. CLT ek limit hai (

n \to \infty

); kisi bhi finite

n

par ye ek approximation hai. "

n \geq 30

" ek rough rule of thumb hai, guarantee nahi.

CLT sum

\sum_{i} X_{i}

par bhi laagu hota hai, sirf mean

\overset{ˉ}{X}_{n}

par nahi.

True. Sum bas

n \overset{ˉ}{X}_{n}

hai; ye

N (n μ, n σ^{2})

ki taraf tend karta hai. Mean aur sum usi ek normal fact ki do scalings hain.

CLT hold hone ke liye variance

σ^{2}

finite hona chahiye.

True. Proof ko

E [Y_{i}^{2}] = 1

chahiye (finite second moment). Infinite variance wali distributions (jaise Cauchy) theorem ko break kar deti hain — unke sample means Cauchy hi rehte hain.

CLT aur Law of Large Numbers ek hi cheez kehte hain.

False. LLN kehta hai

\overset{ˉ}{X}_{n} \to μ

(ek point). CLT us point ke aas-paas wobble ki shape aur spread describe karta hai — ye finer, second-order statement hai.

\overset{ˉ}{X}_{n}

σ

se nahi balki

σ / n

se divide karke standardize karna galti hai.

False. Ye bilkul sahi hai:

\overset{ˉ}{X}_{n}

ka SD

σ / n

hai, isliye

N (0, 1)

tak pahunchne ke liye tumhe raw

σ

se nahi, balki uske SD se divide karna chahiye.

Spot the error

\overset{ˉ}{X}_{n} \sim N (μ, σ^{2})

large

n

ke liye."

Variance galat hai. Ye

σ^{2} / n

hona chahiye —

n

numbers ka mean ek single number se

n

times kam variable hota hai.

"Standard error

= σ / n

, isliye jab main sample double karta hoon toh ye half ho jaata hai."

= σ / n

hai. SE ko halve karne ke liye tumhe

n

ko quadruple karna hoga (square root ki wajah se), double nahi.

"Mera data heavily skewed hai, toh main bas

n = 10

par CLT apply kar lunga jaise usual."

Strong skew ke liye,

n = 10

bahut chhota hai; approximation abhi bhi clearly lopsided hai. Skew convergence slow karta hai — tumhe shayad saikdon samples chahiye honge.

Z_{n} = \frac{X ˉ _{n} - μ}{σ}

N (0, 1)

ki taraf converge karta hai."

n

missing hai. Sirf

σ

se divide karne par ek aisa variable milta hai jiska variance

\to 0

(ye 0 par ek spike mein collapse ho jaata hai), koi fixed bell nahi.

"Samples time ke saath correlated hain, lekin CLT phir bhi jaisi bataya gaya hai waise apply hota hai."

Classic CLT ko independence chahiye; variance-of-a-sum step mein

Var (X + Y) = Var (X) + Var (Y)

use hua tha, jo correlation ke under fail karta hai. Tumhe dependent-data version chahiye hoga.

P (a \leq \overset{ˉ}{X}_{n} \leq b) = Φ (b) - Φ (a)

Tum endpoints standardize karna bhool gaye. Ye

Φ (\frac{b - μ}{σ / n}) - Φ (\frac{a - μ}{σ / n})

hona chahiye.

"Main

μ

aur

σ

apne sample se exactly jaanta hoon, isliye mera CLT interval exact hai."

Practice mein

σ

data se estimate kiya jaata hai, jo uncertainty add karta hai — isliye chhote samples mein

z

ki jagah

t

-distribution use hoti hai. Confidence Intervals aur Hypothesis Testing dekho.

Why questions

Standard error mein

n

ki jagah

n

kyun aata hai?

Sum ka variance

n

ki tarah badhta hai, aur

\overset{ˉ}{X}_{n}

n

se divide karta hai, jisse variance

σ^{2} / n

milta hai; SD variance ka square root hota hai, isliye

σ / n

CLT Confidence Intervals aur Hypothesis Testing ka engine kyun hai?

Dono ko

\overset{ˉ}{X}_{n}

ki sampling distribution chahiye; CLT hume ek jaani-pahchaani normal shape deta hai, isliye hum probabilities ko

z

-scores aur thresholds mein convert kar sakte hain.

Stochastic Gradient Descent mein ek mini-batch gradient true gradient ka noisy version kyun behave karta hai?

Har batch gradient batch elements par ek sample mean hai, isliye CLT ke anusaar ye

\approx N (true gradient, σ_{g}^{2} / n)

hai — unbiased with noise jo batch size badhne par shrink hota hai.

Basic CLT mein terms identically distributed (sirf independent nahi) kyun honi chahiye?

Proof ek MGF ko

n

-th power tak raise karta hai,

[M_{Y_{1}} (t / n)]^{n}

; wo ek shared MGF tabhi exist karta hai jab har

Y_{i}

ki same distribution ho.

Hum limit lene se pehle standardize kyun karte hain?

Un-standardized,

\overset{ˉ}{X}_{n}

μ

par collapse ho jaata hai (LLN) — dekhne ke liye kuch nahi bachta.

σ / n

se rescaling "zoom in" karta hai taaki residual wobble ek fixed bell mein stabilize ho jaaye.

Bootstrap Methods aur Monte Carlo Methods many resamples ko average karne par kyun rely kar sakte hain?

Unke estimates averages hain, isliye unki sampling error SE

\propto 1/ n

ke saath normal hai — CLT humein batata hai estimate kitni tez settle hota hai.

Maximum Likelihood Estimation problem mein infinite-variance model ke saath CLT kyun kaam nahi karta?

Koi finite

σ^{2}

nahi matlab standardizing SE undefined hai aur Taylor step

E [Y^{2}] = 1

collapse ho jaata hai; sampling distribution simply normal mein converge nahi hoti.

Edge cases

Agar har

X_{i}

ek constant

c

hai (zero variance) toh

\overset{ˉ}{X}_{n}

ka kya hota hai?

\overset{ˉ}{X}_{n} = c

hamesha; SE

= 0

, isliye "distribution" ek single spike hai. CLT vacuously true hai lekin degenerate — standardize karne ke liye kuch nahi.

Agar

n = 1

ho toh?

Toh

\overset{ˉ}{X}_{1} = X_{1}

, isliye sample-mean distribution sirf raw distribution hai — koi normalizing nahi hua hai. CLT sirf badhte

n

ke liye bite karta hai.

Data pehle se hi

N (μ, σ^{2})

hai — kya CLT kuch add karta hai?

Mean exactly

N (μ, σ^{2} / n)

hai har

n

ke liye, koi approximation ki zaroorat nahi. CLT ka promise

n = 1

par hi puri ho jaati hai kyunki normals ke sums normal rehte hain.

Bernoulli with

p = 0.001

(extremely rare event) — kya

n = 100

normality ke liye kaafi hai?

Nahi. Near-zero

p

ke saath zyattar samples 0 hain, isliye sum Poisson-like aur skewed hai; bell trustworthy hone se pehle tumhe

n p

aur

n (1 - p)

dono comfortably large chahiye (common check

\geq 10

Do bahut heavy-tailed lekin finite-variance sources: kya CLT phir bhi normality guarantee karta hai?

Haan eventually, lekin heavy tails ise dramatically slow karta hai — Bias-Variance Tradeoff aur Sample Size Calculation dono isko feel karte hain, kyunki required

n

bahut bada ho sakta hai.

Agar main dependent samples average karun jo positively correlated hain, toh SE ke saath kya galat hota hai?

True SE

σ / n

se bada hota hai (positive covariance variance-of-sum mein add hota hai), isliye ek naive CLT interval over-confident hota hai — bahut narrow.

Recall Ek-line self-test

Mean kisi bhi $n$ ke liye exactly normal hota hai jab data khud ::: normal ho (normals ke sums normal hote hain). $n$ ke $\times 4$ hone par SE kis factor se shrink hota hai? ::: 2 ke factor se, kyunki $4 = 2$ . Woh ek moment condition jo CLT ke bina nahi chal sakta ::: finite variance ( $σ^{2} < \infty$ ).