Diagonalization — conditions, procedure: Visual walkthrough

4.5.31 · D2 · Maths › Linear Algebra (Full) › Diagonalization — conditions, procedure

Koi bhi symbol aane se pehle, aao picture pe agree kar lete hain.

Step 0 — Matrix hai kya, actually?

Arrows ki poori grid lo aur har ek ko $A$ ke through daalo. Zyaadatar arrows alag disha mein bahar aate hain jitni andar gayi thi — woh ghoom gayi. Yahi turning hai jo $A^{100}$ (machine ko 100 baar apply karo) ko itna painful banata hai: har application mein sab kuch phir se ulajh jaata hai.

Figure — Diagonalization — conditions, procedure

Figure mein, grey arrows inputs hain. Orange arrows woh hain jo $A$ ne unke saath kiya. Notice karo woh do khaas arrows jo mote draw hain: woh bahar aaye bilkul usi disha mein jis mein andar gaye the — sirf unki length badi. Unhe yaad rakho. Neeche sab kuch unhi ki talash hai.

Step 1 — Sapna: ek machine jo sirf stretch karti hai

KYUN chahiye. Arrow $(x, y)$ ko $D$ mein daalo: $D (x y) = (λ_{1} x λ_{2} y) .$

$λ_{1} x$ ::: horizontal part $x$ , sirf $λ_{1}$ se scale hua — $y$ ka koi zikr nahi.
$λ_{2} y$ ::: vertical part $y$ , sirf $λ_{2}$ se scale hua — $x$ ka koi zikr nahi.

PICTURE. Axes ke saath bani ek grid, grid hi rehti hai — cells sirf badhe/chhote hote hain, kabhi tilt nahi hote. Ise 100 baar apply karne par sirf har factor 100vi power tak pahunch jaata hai. Yahi poora appeal hai.

Teal (horizontal) aur plum (vertical) directions dekho: har ek sirf scale hota hai. Ek diagonal machine boring hai — aur boring bilkul wahi hai jo hum compute karne ke liye chahte hain.

Step 2 — Zyaadatar matrices boring NAHI hain — toh hum apna graph paper badal dete hain

Yeh Change of Basis ka idea hai: transformation $A$ fix hai, lekin hum jo numbers uske liye likhte hain woh depend karte hain ki humne kaun se axes choose kiye. Axes ko cleverly choose karo aur numbers diagonal ho jaate hain.

Left panel: standard graph paper (grey), jahan $A$ grid ko visibly skew karta dikhta hai. Right panel: do khaas arrows (teal & plum) ke saath drawn graph paper. Right frame mein wahi machine sirf har naye axis ke saath stretch karti hai. Wahi machine, smarter paper.

Step 3 — Pieces ko naam do: $P$ aur $D$ banao

KYUN columns? Ek matrix jiske columns tumhare chosen axes hain, exactly woh machine hai jo "naye paper mein coordinates" ko "standard paper mein coordinates" mein translate karti hai. Use new-paper vector $(1, 0)$ daalo aur bahar aata hai $p_{1}$ — pehla naya axis, purane coordinates mein expressed. Yahi $P$ ka kaam hai.

Figure mein $P$ do arrows ke roop mein columns ke taur par rakhe hain aur $D$ diagonal pe stretch numbers ki table ke roop mein hai. Humne pieces ko naam diya hai — ab hum prove karenge ki naming forced hai.

Step 4 — Eigenvalue equation derive karo (dil ki baat)

Goal se start karo aur dono sides ko right mein $P$ se multiply karo (taaki messy $P^{- 1}$ cancel ho jaaye): $A = P D P^{- 1} ⟹ A P = P D .$

$A P$ ::: machine $A$ ko $P$ ke har column (arrow) par chalao.
$P D$ ::: $P$ ke har column ko $D$ ke matching diagonal number se scale karo.
$P^{- 1} P = I$ ::: inverse cancel ho gaya, yahi reason hai humne $P$ se multiply kiya aur kisi aur se nahi.

Ab ise column by column padho. Left side $A P$ ka $j$ -va column hai $A p_{j}$ . Right side $P D$ ka $j$ -va column hai $λ_{j} p_{j}$ (kyunki $D$ sirf column $j$ ko $λ_{j}$ se scale karta hai). Unhe equal karte hain:

KYUN yeh punchline hai. Humne kabhi decide nahi kiya ki eigenvectors use karenge. Humne sirf poocha "kya $A$ kisi paper mein diagonal dikh sakta hai?" aur algebra ne wapas spat kiya: paper ke axes must woh arrows hone chahiye jo $A$ turn nahi karta. Dekho Eigenvalues and Eigenvectors.

Figure mein ek column arrow $p_{1}$ (teal) dikhata hai: input aur output usi line par hain — output sirf $λ_{1}$ times lamba hai. Ek grey non-eigenvector uske side mein draw hai; uska output naye direction mein swing karta hai. Sirf on-line arrows $P$ ke columns ke roop mein survive karte hain.

Step 5 — $P$ kyun invertible hona chahiye → " $n$ independent arrows" condition

KYUN. Agar do khaas arrows ek hi line par hote (ya ek, doosron ka blend hota), $P$ space ko flat crush kar deta — woh kuch directions ko zero par crush kar deta, aur tum kabhi un-crush nahi kar sakte. Un-crushing nahi = $P^{- 1}$ nahi = diagonalization nahi.

Toh poori condition yeh hai:

Left: do independent arrows plane ko span karte hain — $P$ honest hai, invertible hai, hum jeette hain. Right: do arrows ek hi line par — jo "grid" woh banate hain woh ek degenerate sliver hai, $det P = 0$ , koi inverse nahi, hum haarte hain.

Step 6 — Degenerate case: woh shear jiske paas arrows khatam ho jaate hain

Shear $A = (2012)$ lo. Iska ek hi eigenvalue hai $λ = 2$ ( $(2 - λ)^{2} = 0$ se). $A x = 2 x$ solve karne par doosra coordinate $0$ hona forced hai, toh har eigenvector horizontal line par hai — sirf ek independent khaas direction hai, do nahi.

Algebraic multiplicity $a (2) = 2$ ::: eigenvalue $2$ characteristic equation ka double root hai (dekho Characteristic Polynomial).
Geometric multiplicity $g (2) = 1$ ::: lekin actually sirf ek independent eigenvector exist karta hai.
$g < a$ ::: kaafi arrows nahi → plane fill nahi ho sakta → not diagonalizable (yeh ek defective matrix hai; iska sahi sabse simple form Jordan Normal Form mein rehta hai).

Shear square ke upar ko right push karta hai neeche ko pin karte hue. Horizontal ke alawa har arrow tilt ho jaata hai. Doosre un-turned direction ka koi existence nahi hai jo doosre axis ke roop mein serve kare — recipe "collect $n$ arrows" par ruk jaati hai.

Step 7 — Ek map par saare cases

Ek chauthi, sneaky case: ek real matrix jiske real eigenvectors hi nahi hote, jaise $9 0^{\circ}$ rotation $(01 - 1 0)$ — yeh har real arrow ko turn karta hai, toh koi real khaas direction exist nahi karti. Yeh hai diagonalizable, lekin sirf tabhi jab hum complex numbers allow karein ( $λ = \pm i$ ). Real matrix, complex answer.

Parent se ordering warning exactly arrow-and-stretch pairing hai: $P$ ka column $j$ woh arrow hona chahiye jiska stretch $D_{j j}$ par baitha ho. Unhe scramble karo aur $A p_{j} = λ_{j} p_{j}$ silently fail ho jaata hai. Dekho bhi Similar Matrices (kyun $P D P^{- 1}$ eigenvalues ko rakhta hai) aur Matrix Powers and Exponentials (payoff: $A^{100} = P D^{100} P^{- 1}$ ). Symmetric matrices ke liye arrows guaranteed perpendicular bhi hote hain — Spectral Theorem.

Ek-picture summary

Ek frame poori derivation compress karti hai: twisting machine $A$ (grey skewed grid) pure-stretch machine $D$ (teal/plum axis grid) ban jaati hai exactly isliye kyunki humne paper ko un arrows ke saath redraw kiya jo $A$ kabhi nahi ghoomata. $P$ tumhe do papers ke beech le jaata hai; $D$ wahi hai jo machine karti dikhti hai jab tum eigenvector axes par khade hote ho.

Recall Feynman: poora walk simple words mein

Hum ek aise machine chahte the jo itni simple ho ki use sau baar repeat karna aasaan ho — ek machine jo sirf har axis ko stretch kare aur kabhi twist na kare. Real machines twist karti hain, toh hum machine nahi badal sakte. Iske bajaaye hum apna graph paper badle: hum un thode khaas arrows ko dhoondhte hain jo machine kabhi nahi ghooma ti, sirf lambai badhaati hai. Hum apne naye axes unhi arrows ke saath line up karte hain. Us paper par machine finally pure stretching jaisi lagti hai — ek diagonal table $D$ . "Kaun sa arrow kaun sa axis hai" ki book matrix $P$ hai; uska inverse $P^{- 1}$ wapas translate karta hai. Equation $A = P D P^{- 1}$ ko squeeze karne ne prove kiya ki arrows eigenvectors hone hi chahiye — hamare paas koi doosra choice nahi tha. Catch yeh hai: kabhi kabhi poora space fill karne ke liye kaafi un-turned arrows nahi hote (ek shear ke paas sirf ek hota hai). Tab naya paper build nahi ho sakta, aur hum machine ko "not diagonalizable" kehte hain.