Matrix representation of linear transformations: Visual walkthrough

4.5.28 · D2 · Maths › Linear Algebra (Full) › Matrix representation of linear transformations

Step 0 — Teen words jo hum hamesha use karenge

Kisi bhi symbol se pehle, teen plain-language ideas. Har ek ek picture hai, formula nahi.

Figure — Matrix representation of linear transformations

Picture mein: do ingredient arrows $b_{1}$ (blue) aur $b_{2}$ (orange) plane pe hain. Ek teesra arrow $v$ (green) ek stretched-aur-added copy ke roop mein drawn hai: $v = 3 b_{1} + 2 b_{2}$ . Numbers $3$ aur $2$ iske coordinates hain. Inhe grid se padho — coordinates hamesha yehi hote hain.

Yahaan se kyun shuru karein? Neeche har symbol ya toh ek arrow hai, ek ingredient arrow hai, ya ek coordinate hai. Agar yeh teen solid hain, toh baad mein kuch bhi surprise nahi kar sakta. Zyada gehri story ke liye dekho Coordinate vectors and bases — kyun numbers $(3, 2)$ unique hote hain.

Step 1 — Machine ka ek rule, drawn

KYA. Hum machine ki ek promise ko ek picture mein state karte hain: pehle mix karo ya baad mein, answer same aata hai.

KYUN. Yahi ek property hai jo hum use karenge. Sab kuch — poori matrix — isi se nikalti hai. Toh hume ise dekhna chahiye, sirf padhna nahi.

PICTURE. Do routes, same destination:

Top route: $b_{1}$ aur $b_{2}$ lo, pehle unhe $v = 3 b_{1} + 2 b_{2}$ mein mix karo, phir $v$ ko $T$ se guzaaro.
Bottom route: $b_{1}$ aur $b_{2}$ ko pehle $T$ se guzaaro taaki $T (b_{1})$ aur $T (b_{2})$ mile, phir unhe same numbers $3$ aur $2$ se mix karo.

Dono ek identical arrow $T (v)$ (red) pe land karte hain. Symbols mein, har piece padhte hue:

= \underbrace{3\,T(b_1)}_{\text{same stretch, moved output}} + \underbrace{2\,T(b_2)}_{\text{same stretch, moved output}}.$$ Numbers $3$ aur $2$ machine se **seedha guzar jaate hain** bina kisi badlaav ke. Yahi poora secret hai: *kisi bhi* arrow pe $T$ ko jaanne ke liye, tumhe sirf **ingredient arrows** pe jaanna hoga. --- ## Step 2 — Machine ko sirf ingredients pe test karo **KYA.** Hum machine mein sirf $b_1$ daalaate hain, phir sirf $b_2$, aur record karte hain ki har ek kahaan jaata hai. **KYUN.** Step 1 ne prove kiya ki yeh do outputs mein *saari* information hai. Baaki sab kuch mixing hai — free, koi naya machine-work nahi chahiye. ![[deepdives/dd-maths-4.5.28-d2-s03.png]] **PICTURE.** Left panel: $b_1$ andar jaata hai, arrow $T(b_1)$ bahar aata hai (blue → red). Right panel: $b_2$ andar jaata hai, $T(b_2)$ bahar aata hai (orange → red). Humne machine ko exactly **do baar** run kiya hai. $n$ ingredient arrows waali machine ke liye, tum exactly $n$ baar run karte ho — isse zyada kabhi nahi. $$T(b_1) = \text{koi output arrow}, \qquad T(b_2) = \text{koi output arrow}.$$ Lekin plane mein floating ek arrow abhi *numbers* nahi hai. Use store karne ke liye, hume ise ingredient arrows ke khilaaf bhi measure karna hoga — yahi agla step hai. --- ## Step 3 — Har output ko output ingredients ke khilaaf measure karo **KYA.** Hum har output arrow $T(b_j)$ ko **codomain** ingredient arrows $c_1, c_2$ (output world $W$ ke ingredients) ke stretch-and-add ke roop mein express karte hain. **KYUN.** Ek arrow ki picture computer mein type nahi ho sakti; numbers ki pair ho sakti hai. Numbers sirf ek *chosen set of ingredients* ke relative exist karte hain. Toh hum output space ke liye ingredients $c_1, c_2$ chunte hain aur unse coordinates padh lete hain. ![[deepdives/dd-maths-4.5.28-d2-s04.png]] **PICTURE.** Red output $T(b_1)$ ko $c_1$ (blue) aur $c_2$ (orange) grid pe decompose kiya jaata hai: dashed guide-lines dikhate hain ki yeh $a_{11}c_1 + a_{21}c_2$ ke barabar hai. Guide-lines padhke: $$T(b_1) = \underbrace{a_{11}}_{\text{kitna }c_1}c_1 + \underbrace{a_{21}}_{\text{kitna }c_2}c_2, \qquad T(b_2) = \underbrace{a_{12}}_{\text{kitna }c_1}c_1 + \underbrace{a_{22}}_{\text{kitna }c_2}c_2.$$ Two-index name $a_{ij}$ padhna: **pehla** number $i$ bolta hai *kaunsa output ingredient* $c_i$; **doosra** number $j$ bolta hai *kaunsa input ingredient* $b_j$ humne daala tha. Toh $a_{21}$ = "$b_1$ daalne par output mein $c_2$ ki maatra." > [!mistake] Yeh measurement skip karna jab $c_1,c_2$ standard ingredients hain > Jab $c_1=(1,0)$, $c_2=(0,1)$, toh output *khud hi apne coordinates jaisa lagta hai*, isliye > log Step 3 skip kar dete hain. Jis moment output ingredients tilted ho jaate hain (jaise > parent ke Example 3 mein), arrow aur uske coordinates alag ho jaate hain — tumhe **zaroor** > $a_{ij}$ solve karne chahiye. Apne dimaag mein Steps 2 aur 3 kabhi merge mat karo. --- ## Step 4 — Numbers ko ek table mein stack karo (matrix ka janam) **KYA.** Hum Step 3 ke coordinate pairs lete hain aur unhe **columns** ke roop mein khada karte hain. **KYUN.** Hum ek aisa layout chahte hain jisse "$v$ daalo, $T(v)$ padho" pure arithmetic ban jaaye. Column layout wahi hai jo yeh kaam karta hai (Step 5 mein prove hoga). ![[deepdives/dd-maths-4.5.28-d2-s05.png]] **PICTURE.** Do coordinate pairs slide karke ek grid ke do vertical columns ban jaate hain: $$[T] = \left[\ \underbrace{\begin{array}{c} a_{11} \\ a_{21} \end{array}}_{\text{jahaan }b_1\text{ gaya, }c\text{'s mein}} \ \middle|\ \underbrace{\begin{array}{c} a_{12} \\ a_{22} \end{array}}_{\text{jahaan }b_2\text{ gaya, }c\text{'s mein}}\ \right].$$ Har column ek machine-test hai. Column $j$ = "ingredient $b_j$ kahaan land kiya, output ingredients mein likha." Parent ka mnemonic — **"Columns Catch Outputs in $C$"** — bilkul yehi picture hai. > [!mistake] Outputs ko rows ki tarah neeche leth dena > Yeh natural lagta hai (hum left-to-right padhte hain), lekin rows se tumhe ==transpose== > milta hai, ek *alag* machine. Agla step dikhata hai kyun columns, aur sirf columns, $T$ ko > reproduce karte hain. --- ## Step 5 — Table-times-list answer kyun deta hai **KYA.** Hum ek general arrow $v = x_1 b_1 + x_2 b_2$ (coordinates $x_1,x_2$) daalaate hain aur dikhate hain ki table list $(x_1,x_2)$ pe multiply hokar output coordinates deta hai. **KYUN.** Yahi payoff hai — yeh "machine chalao" ko "table ke saath arithmetic karo" mein convert karta hai. ![[deepdives/dd-maths-4.5.28-d2-s06.png]] **PICTURE.** Arrows follow karo: $x_1$ column 1 pick karta hai, $x_2$ column 2 pick karta hai, aur woh add ho jaate hain. Term by term: $$T(v) = x_1\,T(b_1) + x_2\,T(b_2) = x_1(a_{11}c_1 + a_{21}c_2) + x_2(a_{12}c_1 + a_{22}c_2).$$ Ab **output ingredient ke hisaab se gather karo** — saare $c_1$ pieces collect karo, phir saare $c_2$ pieces: $$T(v) = \underbrace{(a_{11}x_1 + a_{12}x_2)}_{c_1\text{ ka coordinate}}\,c_1 + \underbrace{(a_{21}x_1 + a_{22}x_2)}_{c_2\text{ ka coordinate}}\,c_2.$$ Dono brackets *hi* output coordinates hain — aur har bracket exactly **table ki row dotted with list** hai. Yahi rule hai "row dot column." Toh: > [!formula] Fundamental relationship, ab earned > $$\boxed{\,[T(v)]_C = [T]_C^B\,[v]_B\,}$$ > Ise padhna: **left** = output arrow ke coordinates $C$-ingredients mein; **right** = table > times input ke coordinate-list $B$-ingredients mein. Yahaan kuch bhi assume nahi kiya gaya > — har symbol pehle draw kiya gaya tha. Kyunki humne column layout *choose* kiya tha ise kaam karne ke liye, [[Composition of linear maps|composition]] $T\circ S$ ke liye matrix multiplication matrix product ke barabar hona chahiye — same consistency requirement, ek level upar. --- ## Step 6 — Edge case: zero machine aur "do-nothing" machine **KYA.** Hum do degenerate machines check karte hain taaki koi scenario unseen na rahe. **KYUN.** Ek rule jo simplest inputs pe nahi chal sakta, woh rule tum samjhe nahi ho. ![[deepdives/dd-maths-4.5.28-d2-s07.png]] **PICTURE.** Left: **zero machine** har arrow ko origin pe crush kar deta hai. Ingredients test karne par, $T(b_1)=0$ aur $T(b_2)=0$, toh *har* column $\binom{0}{0}$ hai: $$[T]_{\text{zero}} = \begin{pmatrix} 0 & 0 \\ 0 & 0 \end{pmatrix}.$$ Iska [[Kernel and image|image]] ek single point hai aur iska [[Rank and nullity|rank]] $0$ hai — columns mein koi directions nahi hain. Right: **identity machine** har arrow ko untouched chhod deta hai. Tab $T(b_1)=b_1$ aur $T(b_2)=b_2$, aur (same ingredients andar aur bahar ke saath) har output *ek* ingredient hai, toh uske coordinates $(1,0)$ aur $(0,1)$ hain: $$[T]_{\text{id}} = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}.$$ Same recipe, koi exception nahi: **ingredients pe run karo, outputs ko output ingredients mein likho, columns ke roop mein stack karo.** > [!mistake] Yeh sochna ki identity hamesha identity matrix hoti hai > Sirf tabhi jab **input aur output ingredients match karein**. Identity map feed karo lekin > uske outputs ek *alag* set of ingredients mein padho toh tumhe ek [[Change of basis|change-of-basis]] > matrix milega — phir bhi isi exact recipe se bana. --- ## Ek-picture summary ![[deepdives/dd-maths-4.5.28-d2-s08.png]] Ek diagram, poora derivation: ingredient arrows $b_1,b_2$ machine $T$ mein enter karte hain → har output $c_1,c_2$ ke khilaaf measure hota hai → coordinate pairs columns ke roop mein khade ho jaate hain → taiyaar table kisi bhi input list ko multiply karke output list deta hai. Woh closed loop *hi* theorem hai. > [!recall]- Feynman retelling — kisi dost ko batao > Tumhare paas ek magic machine hai jo arrows ko stretch aur spin karti hai, lekin tum har > arrow pe test nahi kar sakte — infinitely many hain. Trick: har arrow secretly kuch "ingredient > arrows" ki recipe hoti hai, jaise *2 parts yeh plus 3 parts woh*. Aur yeh machine fair hai — > yeh recipes respect karti hai: pehle mix karo ya pehle run karo, same result. Toh tum > machine sirf ingredients pe chalate ho, ek ek baar, aur note karte ho ki har ek kahaan > land karta hai. Lekin "kahaan land karta hai" numbers hone chahiye, toh tum har landing spot > ko *output* ingredients ke khilaaf bhi measure karte ho. Woh number-pairs khade karke, side > by side — woh column-table matrix hai. Ab *kisi bhi* arrow ke liye, tum machine ko dobara > nahi chhute: uski recipe dekhte ho, table ko us recipe-list se multiply karte ho, aur bahar > aata hai ki woh kahaan jaata hai. Zero machine ek all-zeros table deti hai; do-nothing machine > $1$s-on-the-diagonal table deti hai. Har baar same recipe. Yahi poora idea hai ek > transformation ko matrix ke roop mein represent karna. > [!recall] Do quick self-checks > Ek plane map ke liye exactly do machine-runs kyun? ::: Do ingredient arrows hain; har column ek run hai, aur map apne ingredients pe action se fix ho jaata hai. > Agar output ingredients tilted hain, toh Step 3 mein kya extra kaam aata hai? ::: Tumhe $T(b_j)=a_{1j}c_1+a_{2j}c_2$ ke liye coordinates solve karne padte hain; arrow ab apne coordinates se equal nahi hota. > Kaunsa layout table-times-list ko $T$ reproduce karata hai, aur kyun? ::: Columns — kyunki output ko ingredient ke hisaab se gather karne par har output coordinate, table ki row aur input list ka dot product ban jaata hai. --- ## Connections - [[Matrix representation of linear transformations]] — parent result jo is page pe visually derive kiya gaya hai. - [[Coordinate vectors and bases]] — Step 0 ke numbers ka matlab. - [[Change of basis]] — jab Step 6 mein in/out ingredients alag hon toh table ka kya hota hai. - [[Composition of linear maps]] — column layout matrix multiplication ko force karta hai. - [[Rank and nullity]] — taiyaar table ke columns se padho (Step 6). - [[Kernel and image]] — zero machine ka collapsed image. - [[Eigenvalues and diagonalization]] — aise ingredients chunna jo table ko simplest banaaein.