ADC - DAC — resolution, sampling rate, Nyquist: Visual walkthrough

5.5.4 · D2 · Coding › Embedded Systems & Real-Time Software › ADC - DAC — resolution, sampling rate, Nyquist

Shuru karne se pehle, teen plain-word promises un symbols ke baare mein jo aap milenge:

Neeche sab kuch inhi teen se bana hai.

Step 1 — Ek continuous wave actually hoti kya hai

KYA HAI. Ek pure tone draw karo. Sabse clean wave jo hum choose kar sakte hain woh hai cosine: yeh apne top se shuru hoti hai, bottom tak jaati hai, aur wapis chadh aati hai — hamesha ke liye, smoothly.

COSINE KYUN, KUCH LUMPY KYUN NAHI? Kyunki koi bhi signal pure cosines ke sum mein toda ja sakta hai (yahi toh Fourier Transform & Frequency Domain ka poora point hai). Agar hum samajh lein ki sampling frequency $f$ ki ek cosine ke saath kya karti hai, toh hum samajh jaayenge ki yeh sab ke saath kya karti hai. Toh hum sabse simple brick padhte hain aur wall par trust karte hain.

PICTURE. Smooth violet curve asli signal hai. Abhi tak kuch bhi digital nahi hai — CPU ne isko chua nahi.

Figure — ADC - DAC — resolution, sampling rate, Nyquist

Step 2 — Sampling = wave ko fixed instants par chhurra marna

KYA HAI. Ek converter poori smooth curve nahi dekh sakta. Yeh sirf fixed, evenly spaced moments par dekhta hai aur woh height likh leta hai jo use milti hai. Hum un looks ko $T_{s}$ seconds ke gap par rakhte hain.

EVENLY SPACED KYUN? Hardware ek clock par tick karta hai. Clock by nature regular hoti hai, toh looks $t = 0, T_{s}, 2 T_{s}, 3 T_{s}, \dots$ par padti hain — yaani $t = n T_{s}$ par poore numbers $n = 0, 1, 2, \dots$ ke liye.

PICTURE. Violet curve abhi bhi hai, lekin ab sirf orange dots computer mein survive karte hain. Dots ke beech ki har cheez guess ki jaati hai, measure nahi.

Sampling instants $t = n / f_{s}$ ko apni wave mein plug karne par actually stored numbers milte hain: $s [n] = cos (2 π f \cdot \frac{n}{f _{s}}) = cos (2 π \frac{f}{f _{s}} n)$

$s [n]$ — $n$ -vaan stored number (square brackets = "discrete index", continuous time nahi).
$\frac{f}{f _{s}}$ — show ka star: wave ki frequency sampling rate ke units mein measure ki gayi. Is ratio ko dhyaan se dekho.

Step 3 — Woh disaster: do alag waves, identical dots

KYA HAI. Hamari real wave of frequency $f$ lo. Ab frequency $f + f_{s}$ ki ek faster wave lo. Dono ko same instants $t = n / f_{s}$ par sample karo. Stored numbers compare karo.

YEH PAIR KYUN? Kyunki $f_{s}$ "ek full extra spin per sample" frequency hai. Ise add karne se har dot ke pair ke beech ek invisible full rotation chhup jaani chahiye — invisible precisely isliye kyunki hum kabhi mid-spin nahi dekhte.

Aao math term-by-term check karte hain: $cos (2 π \frac{( f + f _{s} )}{f _{s}} n) = cos (2 π \frac{f}{f _{s}} n + 2 π n)$

Extra piece hai $2 π \frac{f _{s}}{f _{s}} n = 2 π n$ .
$2 π n$ full circles ki poori number hai. Cosine har full circle ke baad exact same value par wapas aata hai.

Toh: $cos (2 π \frac{( f + f _{s} )}{f _{s}} n) = cos (2 π \frac{f}{f _{s}} n) = s [n]$

Dono waves byte-for-byte identical dots produce karti hain.

PICTURE. Magenta fast wave aur violet slow wave dono same orange dots se thread hoti hain. Jab aapke paas sirf dots hain, aap nahi bata sakte kaun si curve ne unhe banaya.

Step 4 — Fast wave kis slow wave ki nakal karti hai?

KYA HAI. Step 3 ne dikhaya ki ek $f_{s}$ add karne se kuch nahi badalta. Toh do, teen, ya $k$ add karne se bhi nahi badalta. Iska matlab hai $f$ , $f + f_{s}$ , $f + 2 f_{s}$ , … sab same dots par collapse ho jaate hain. Apparent frequency find karne ke liye hum jitne poore $f_{s}$ chunks peel kar sakte hain karte hain aur bacha hua rakh lete hain.

FLOOR NAHI, ROUND KYUN? Hum nearest multiple of $f_{s}$ chahte hain, kyunki $f_{s} - f$ par ek wave aur uska mirror bhi dots share karta hai (cosine symmetric hai). Nearest multiple par round karne se hume sabse chhota possible leftover milta hai — woh frequency jo ear/scope actually report karta hai.

PICTURE. Frequencies ki ek number line. Halfway mark $f_{s} /2$ se upar ki har cheez neeche ek low frequency par fold back hoti hai jaise ek hinge ho. Woh folding line ek saath villain aur hero hai.

Step 5 — Dekhte hain yeh hota kaise hai: 1800 Hz sampled at 1000 Hz

KYA HAI. Ek real 1800 Hz tone, 1000 baar per second dotted.

YEH EXAMPLE KYUN? $1800$ sampling rate ka lagbhag do guna hai — ek textbook violation. Aao dekhen yeh kaun sa fake produce karta hai.

Har piece compute karo:

$\frac{f}{f _{s}} = \frac{1800}{1000} = 1.8$
$round (1.8) = 2$
$f_{s} \cdot 2 = 2000$
$f_{a l ia s} = ∣1800 - 2000∣ = 200 Hz$

PICTURE. Magenta 1800 Hz truth aur violet 200 Hz impostor har orange dot share karte hain. Aapki recording ek 200 Hz hum play back karti hai jo kabhi room mein thi hi nahi.

Step 6 — Woh rule jo masquerade forbid karta hai

KYA HAI. Ab hum demand karte hain: hamare signal mein koi bhi wave kisi slow wave se collide na kare. Step 4 se, folding $f_{s} /2$ se upar ki kisi bhi frequency ke liye hoti hai. Toh hum simply signal ko itna high kuch contain karne se forbid kar dete hain.

STRICT "GREATER THAN" KYUN, "EQUAL TO" KYUN NAHI? Boundary case $f = f_{s} /2$ dekho: dots har baar exactly zero-crossings par land kar sakte hain, $0, 0, 0, \dots$ read karte hue — wave wahan hai lekin dots "silence" kehte hain. Toh safe condition hai strictly top frequency ke do guna se zyada fast.

PICTURE. Safe zone (fold line $f_{s} /2$ ke neeche) glow karta hai; forbidden zone (uske upar) wahan hai jahan aliases paida hote hain. Apni saari signal energy glow mein rakho.

$f_{ma x} < f_{N y q u i s t}$ guarantee karne ka practical tarika hai ki ADC se pehle ek anti-aliasing low-pass filter se physically fold line ke upar ki sab kuch delete kar do.

Step 7 — Do degenerate cases jinpar aapko kabhi nahi tripna chahiye

KYA HAI. Do boundary scenarios jahan naive intuition fail karti hai.

Case A — exactly Nyquist par, $f = f_{N y q u i s t} = f_{s} /2$ . $cos (2 π \cdot \frac{1}{2} n) = cos (π n) = (- 1)^{n} = + 1, - 1, + 1, - 1, \dots$ — yeh sirf tab kaam karta hai jab hum peaks sample karein. Clock ko ek quarter cycle shift karo aur har dot ek zero-crossing par land karta hai: $0, 0, 0, \dots$ . Wave gayab ho jaati hai. Lesson: equality unsafe hai; yahi wajah hai ki parent note strict inequality aur ek comfort margin par insist karta hai (CD audio $20$ kHz limit ke liye $44.1$ kHz use karta hai, bare $40$ kHz nahi).

Case B — ek constant (DC), $f = 0$ . $cos (0) = 1$ har sample ke liye. Har dot same height par hai, ek flat line. Koi wiggle nahi, koi ambiguity nahi, koi aliasing nahi — DC hamesha safe hai, aur yeh kisi bhi $f_{N y q u i s t}$ se comfortably neeche baithta hai.

PICTURE. Left panel: peak-sampled vs zero-crossing-sampled Nyquist wave — same wave, wildly different dots. Right panel: placid DC line, sab dots equal.

Ek-picture summary

Is page ki har idea ek single frame mein collapse ho jaati hai: ek true wave, uske dots, woh impostor jo un dots mein bhi fit hota hai, aur fold line jo safe ko stolen se alag karti hai.

Recall Feynman retelling — poora walkthrough kitchen words mein

Socho tum ek wiggly wave draw kar rahe ho apni pencil ko steady beat par tap karke, dot, dot, dot. Do taps ke beech tum kabhi nahi dekhte — toh agar wave ek poora extra upar-neeche chhupa leta hai jab tum blink karte ho, toh tumhe kabhi pata nahi chalega. Iska matlab hai ek bahut fast wave aur ek lazy slow wave exactly same trail of dots chhod sakte hain. Jab baad mein tum dots connect karte ho, tum lazy wala draw karte ho — fast tone ko disguise mil gaya. Disguise find karne ke liye, jitne "one-full-extra-wiggle-per-tap" chunks fit hon subtract karo; jo bhi wiggle bach jaaye woh fake pitch hai jo tum sunoge (1800 Hz ko 1000 baar per second tap karne par 200 Hz fake aata hai). Ilaaj ek speed rule hai: jis fastest wiggle ki tumhe parwah hai uske liye do se zyada baar tap karo, aur pencil se pehle ek gate lagao jo kisi bhi faster cheez ko kabhi aane se rok de. Exactly do guna kaafi nahi hai, kyunki tumhare taps un jagahon par land kar sakte hain jahan wave har baar zero cross kar rahi ho — aur phir tumhari drawing ek dead-flat line hai jo pretend karti hai kuch wahan tha hi nahi.

Connections

Parent topic
Anti-Aliasing Filters
Fourier Transform & Frequency Domain
Quantization Noise & SNR
Zero-Order Hold & Reconstruction