WHY the "exactly one" rule? If one input could give two outputs, the machine is ambiguous — you couldn't predict its behaviour, and calculus (which tracks the output) would break.
WHAT it looks like in ML: A neural network is a giant function f(x;θ)=y^. Input = pixels, output = class score. The whole game is choosing θ so the machine behaves well.
HOW to read ε–δ (Feynman version): You (the challenger) say "I bet you can't keep the output within ε of L." I (defender) reply "Give me any ε, and I'll hand you a δ: stay within δ of c and I guarantee it." If I can always win, the limit is L. Note 0<∣x−c∣excludesx=c itself — that's the "don't care about the point" clause.
We want ∣f(x)−7∣<ε, i.e. ∣2x+1−7∣=∣2x−6∣=2∣x−3∣<ε.
Why this step? We rewrote the output-gap in terms of the input-gap ∣x−3∣, because δ controls exactly that.
So ∣x−3∣<ε/2. Choose δ=ε/2. Then ∣x−3∣<δ⇒2∣x−3∣<2δ=ε. ✅ Done — the limit is genuinely 7, built from the definition, not guessed.
WHY ML cares: Gradient descent needs f (loss) continuous and differentiable to have usable slopes. The ReLUmax(0,x) is continuous everywhere (no jump) but has a kink at 0 — continuity is guaranteed, differentiability is not. That's why ReLU trains fine but needs a subgradient at 0.
What are the two defining properties of a function?
Each input in the domain maps to exactly one output; every domain element is assigned an output.
State the ε–δ definition of limx→cf(x)=L.
∀ε>0∃δ>0 s.t. 0<∣x−c∣<δ⇒∣f(x)−L∣<ε.
Why does the limit definition use 0<∣x−c∣ (strict)?
To exclude the point x=c itself — the limit tracks nearby behaviour, not the value at c.
What are the three conditions for continuity at c?
f(c) exists; limx→cf(x) exists; they are equal.
What does 0/0 tell you to do?
It's indeterminate — simplify (factor/rationalize/L'Hôpital); the limit may still exist.
Is ReLU max(0,x) continuous at 0? Differentiable at 0?
Continuous yes (no jump); differentiable no (corner).
Name the three discontinuity types.
Removable (hole), jump, infinite.
Compute limx→1x−1x2−1.
Factor to x+1, gives 2.
Vertical line test checks what?
Whether a curve is a function (no vertical line hits it twice).
Recall Feynman: explain to a 12-year-old
Imagine walking toward a doorway. The limit is which spot you're clearly heading toward, even if there's a tiny hole in the floor right at the door — you can still tell where you're going. A function is continuous if when you actually reach the door, you land exactly on the spot you were heading to, with no jump, no hole, no falling off a cliff. In AI, the computer learns by rolling a ball downhill on a curve; if the curve has jumps or cliffs, the ball can't roll smoothly — so we want smooth, continuous curves.
Socho ek function ek machine hai: ek number daalo, ek hi number bahar aata hai. Yahi "exactly one output" wala rule important hai — warna prediction possible hi nahi hoga. ML mein poora neural network bhi ek bada function hi hai: input pixels daalo, output class score milta hai.
Limit ka matlab hai — "input jab c ke paas ja raha hai, to output kahan ki taraf ja raha hai?" Point c pe kya ho raha hai, usse limit ko koi matlab nahi. Jaise x−1x2−1 pe x=1 daalo to 0/0 aata hai (undefined), lekin factor karke (x+1) ban jaata hai, aur limit 2 nikal aati hai. Yaad rakho: 0/0 ka matlab "limit exist nahi karti" nahi hai — yeh sirf ishaara hai ki simplify karo.
Continuity ka matlab hai pen uthaye bina graph banana — no hole, no jump, no cliff. Teen conditions: f(c) exist kare, limit exist kare, aur dono barabar ho (mnemonic: HEA — Here, Expected, Agree). ML mein yeh isliye zaroori hai kyunki gradient descent ek gend ko slope pe neeche rol karta hai; agar curve mein jump ya cliff ho to gend theek se roll nahi karegi.
Ek important baat: continuous hona differentiable hone se kamzor condition hai. ReLU max(0,x) har jagah continuous hai lekin 0 pe ek "kink" (corner) hai, isliye wahan slope unique nahi. Isliye ReLU train to hota hai, par 0 pe ek subgradient use karna padta hai. Yeh distinction interview aur real training dono mein kaam aata hai.
