Functions, limits, and continuity — Hinglish notes

1.2.1 · AI-ML › Calculus & Optimization Basics

1. Functions — the input→output machine

"Exactly one" rule kyun? Agar ek input do outputs de sake, toh machine ambiguous ho jaati hai — aap uske behaviour ko predict nahi kar sakte, aur calculus (jo the output track karta hai) toot jaata.

ML mein kaise dikhta hai: Ek neural network ek giant function $f (x; θ) = \overset{y}{^}$ hai. Input = pixels, output = class score. Poora game yahi hai ki $θ$ aisa choose karo ki machine achha behave kare.

2. Limits — where is the output heading?

$ε$ – $δ$ ko kaise padhe (Feynman version): Aap (challenger) kehte ho "Main bet lagata hoon ki tum output ko $ε$ ke andar $L$ ke nahi rakh sakte." Main (defender) kehta hoon "Mujhe koi bhi $ε$ do, aur main tumhe ek $δ$ dunga: $c$ ke $δ$ ke andar raho aur main guarantee karta hoon." Agar main hamesha jeet sakta hoon, limit $L$ hai. Note karo $0 < ∣ x - c ∣$ $x = c$ khud ko exclude karta hai — ye "point ki parwah mat karo" clause hai.

Deriving a limit from scratch (no plugging in)

Claim: $x \to 3 lim (2 x + 1) = 7$ . Isko $ε$ – $δ$ se prove karte hain.

Hum chahte hain $∣ f (x) - 7∣ < ε$ , i.e. $∣2 x + 1 - 7∣ = ∣2 x - 6∣ = 2∣ x - 3∣ < ε$ . Ye step kyun? Humne output-gap ko input-gap $∣ x - 3∣$ ke terms mein rewrite kiya, kyunki $δ$ exactly usi ko control karta hai.

Toh $∣ x - 3∣ < ε /2$ . Choose karo $δ = ε /2$ . Tab $∣ x - 3∣ < δ \Rightarrow 2∣ x - 3∣ < 2 δ = ε$ . ✅ Done — limit genuinely $7$ hai, definition se banaya gaya, guess nahi kiya.

One-sided limits & the removable hole

Figure — Functions, limits, and continuity

3. Continuity — heading matches arrival

ML kyun care karta hai: Gradient descent ko $f$ (loss) continuous aur differentiable chahiye taaki usable slopes milein. ReLU $max (0, x)$ har jagah continuous hai (koi jump nahi) lekin $0$ par ek kink hai — continuity guaranteed hai, differentiability nahi. Isliye ReLU theek se train hota hai lekin $0$ par subgradient ki zaroorat padti hai.

4. Common Mistakes (steel-manned)

5. Flashcards

Ek function ki do defining properties kya hain?

Domain ka har input exactly one output par map karta hai; har domain element ko ek output assign ki jaati hai.

lim_{x \to c} f (x) = L

ε

–

δ

definition batao.

\forall ε > 0 \exists δ > 0

s.t.

0 < ∣ x - c ∣ < δ \Rightarrow ∣ f (x) - L ∣ < ε

Limit definition

0 < ∣ x - c ∣

(strict) kyun use karta hai?

Point

x = c

khud ko exclude karne ke liye — limit nearby behaviour track karta hai,

c

par value nahi.

c

par continuity ki teen conditions kya hain?

f (c)

exist karta hai;

lim_{x \to c} f (x)

exist karta hai; dono equal hain.

0/0

tumhe kya karne ko kehta hai?

Ye indeterminate hai — simplify karo (factor/rationalize/L'Hôpital); limit phir bhi exist ho sakti hai.

Kya ReLU

max (0, x)

0

par continuous hai?

0

par differentiable hai?

Continuous haan (koi jump nahi); differentiable nahi (corner hai).

Teen discontinuity types ke naam batao.

Removable (hole), jump, infinite.

lim_{x \to 1} \frac{x ^{2} - 1}{x - 1}

compute karo.

Factor karke

x + 1

milta hai, answer

2

hai.

Vertical line test kya check karta hai?

Kya ek curve ek function hai (koi vertical line use do baar nahi kaatti).

Recall Feynman: 12-year-old ko samjhao

Socho tum ek doorway ki taraf chal rahe ho. Limit woh jagah hai jis taraf tum clearly ja rahe ho, chahe door ke right par floor mein ek tiny hole ho — tum phir bhi bata sakte ho kidhar ja rahe ho. Ek function continuous hai agar jab tum actually door par pahunchte ho, tum exactly usi jagah land karte ho jahan ja rahe the, koi jump nahi, koi hole nahi, koi cliff se girna nahi. AI mein, computer ek curve par ball ko downhill roll karke seekhta hai; agar curve mein jumps ya cliffs hain, toh ball smoothly roll nahi kar sakti — isliye hume smooth, continuous curves chahiye.

Connections

Derivatives and gradients — derivative ek difference quotient ki limit hai; pehle continuity chahiye.
Gradient Descent — continuous, (mostly) differentiable loss surfaces chahiye.
Activation functions (ReLU, sigmoid) — kinks par continuity vs differentiability.
L'Hôpital's Rule — $0/0$ aur $\infty/\infty$ indeterminate limits resolve karna.
Continuity of composite functions — continuous functions ki chain continuous hoti hai (backprop ki backbone).