Observability matrix — rank test: Question bank — Hinglish deep dive

3.5.33 · D5 · Physics › Guidance, Navigation & Control (GNC) › Observability matrix — rank test

Yaad dilane ke liye — inme jo objects hain (parent note mein poori tarah build kiye gaye): system hai $\overset{x}{˙} = A x + B u, y = C x + D u$ jisme state $x \in R^{n}$ aur output $y \in R^{p}$ hai.

Figure — Observability matrix — rank test

Figure dikhata hai ki "ek unobservable direction" kaisi lagti hai: sensor row $C$ (cyan) ek axis span karti hai; amber state vector jo untouched axis ke along hai wo uspar zero shadow daalta hai — chahe dynamics $A$ use jitna bhi push kare, meter kuch nahi padhta.

True or false — justify

TF: Zyada outputs wala system (bada

p

) hamesha kam outputs wale se zyada observable hota hai.

False.

p

sirf

O

ko lamba banata hai (

p n

rows), lekin observability ke liye full column rank

n

chahiye; badly-aligned

C

kaafi saari rows ke saath bhi ek state direction miss kar sakta hai, jabki ek achhi tarah choose ki gayi single row sab capture kar sakti hai.

TF: Agar

rank (O) = n

hai toh initial state

x (0)

output se uniquely recover ki ja sakti hai.

True. Full column rank ka matlab hai

ker (O) = {0}

, toh map

x (0) \mapsto (y (0), \overset{y}{˙} (0), \dots)

injective hai aur unique

x (0)

ke liye invert ki ja sakti hai.

TF: Observability guarantee karta hai ki system stable hai.

False. Observability purely

C

aur

A

ke beech ek geometric relationship hai; eigenvalue locations ke baare mein kuch nahi kehta. Ek unstable system perfectly observable ho sakta hai, aur ek stable wala ek poora state hide kar sakta hai.

TF: Ek stable system automatically observable hota hai kyunki uske states predictably settle ho jaate hain.

False. Settling behaviour (eigenvalues) aur sensor ko visibility (

O

ka kernel) independent hain. Upar wala build structurally sahi hai phir bhi ek poora invisible state hai.

TF: Agar

C = 0

hai (koi sensor kuch nahi padhta), toh

n \geq 1

ke liye system kabhi observable nahi ho sakta.

True. Har block

C A^{k} = 0

hai, toh

O = 0

ka rank

0 < n

hai; koi output nahi toh free response identically zero hai aur

x (0)

completely hidden hai.

TF: Observability matrix square honi chahiye.

False. Ye

p n \times n

hai; square sirf tab hoti hai jab

p = 1

ho.

p > 1

ke liye ye tall hai, aur rank ab bhi

n

se cap ki gayi hai.

TF:

(A, C)

observable hona

(A^{⊤}, C^{⊤})

ke controllable hone ke equivalent hai.

True — ye duality statement hai. Transpose karna "stack

C

vertically with

C A^{k}

" ko "stack

C^{⊤}

horizontally with

A^{⊤ k} C^{⊤}

" mein badal deta hai.

TF:

O

mein row block

C A^{n}

add karne se uska rank badh sakta hai.

False. Cayley–Hamilton theorem ke according — jo kehta hai ki ek matrix apni characteristic equation satisfy karta hai, toh

A^{n}

I, \dots, A^{n - 1}

ka combination hota hai — block

C A^{n}

existing rows ka linear combination hai aur koi naya rank contribute nahi karta.

TF: Agar koi bhi single scalar output

y_{i}

apne aap par unobservable hai, toh poora system unobservable hai.

False. Observability poore

O

ki property hai; alag-alag output channels alag-alag state directions cover kar sakte hain, toh individually-weak sensors ka set jointly observable ho sakta hai.

Spot the error

Error: "Maine

O

C A^{n - 1}

tak compute kiya, rank

n - 1

mila, phir

C A^{n}

aur

C A^{n + 1}

add karke rank

n

pahunch gaya — toh system observable hai."

Added rows Cayley–Hamilton-redundant hain aur rank raise nahi kar sakti; actual rank

n - 1

thi, toh system genuinely not observable hai. Reasoning ne dependent rows se rank fabricate ki.

Error: "

O

p n = 6

rows hain

n = 2

ke liye, toh uska rank 6 hai aur system observable hai."

Rank independent columns count karta hai, jo ki at most

n = 2

ho sakta hai; rows ki sankhya irrelevant hai. Observable hone ke liye rank

= n = 2

chahiye,

= 6

nahi.

Error: "Do identical decoupled modes (repeated eigenvalue

λ

do eigenvectors ke saath) ko ek single scalar output se alag kiya ja sakta hai agar main kaafi derivatives lun."

Kitni bhi derivatives kaam nahi aati: har

C A^{k}

us eigenspace par

C λ^{k}

ki tarah act karta hai, toh saari rows wahan proportional hain — do modes indistinguishable hain. Tumhare paas

p \geq 2

independent output rows chahiye.

Error: "

O

ke kernel mein ek nonzero

x_{0}

hai, lekin system phir bhi observable hai kyunki

y (t)

eventually nonzero ho jaata hai."

Agar

x_{0} \in ker O

hai toh

C e^{A t} x_{0} \equiv 0

sab

t

ke liye — output identically zero hai, kabhi "eventually" nonzero nahi hota. Nonzero kernel exactly unobservability ki definition hai.

Error: "Observability

B

aur

D

par depend karta hai, toh maine

O

banate waqt unhe include kiya."

u

B

D

y

par contribution known hai aur subtract kiya ja sakta hai;

x (0)

reconstruct karne ke liye jo bacha rehta hai wo free response hai, jo sirf

C

aur

A

se govern hota hai.

Error: "Maine

O = [C A C A^{2} C]

build kiya jisme

A

left par hai."

Ye controllability pattern hai (

A^{k} B

A

left par, horizontally stacked). Observability vertically stack karta hai aur

A

right par hota hai:

C, C A, C A^{2}

Why questions

Why: Hum

y = C e^{A t} x (0)

t = 0

par differentiate kyun karte hain sirf

y (0)

use karne ki jagah?

Ek akele instant se sirf

y (0) = C x (0)

milta hai, jo

x (0)

C

ki rows par projection hai. Derivatives

C A^{k} x (0)

nayi directions reveal karti hain, aur unhe stack karne se wo saari directions span hoti hain jinhe sensor possibly reach kar sakta hai.

Why:

n - 1

exactly stopping power kyun hai —

n

2 n

kyun nahi?

Cayley–Hamilton theorem (ek matrix apni characteristic polynomial satisfy karta hai)

A^{n}

aur usse aage sab kuch

I, \dots, A^{n - 1}

ka combination bana deta hai, toh

k \geq n

ke liye rows

C A^{k}

dependent hain. Powers

0

n - 1

already poora reachable row space span kar dete hain.

Why: Ek unobservable state exactly zero output produce karta hai na ki sirf ek chota wala?

Kyunki

x_{0} \in ker O

ka matlab hai

C A^{k} x_{0} = 0

har

k

ke liye, aur

e^{A t}

A

mein ek power series hai; toh

C e^{A t} x_{0}

zero terms ka sum hai — identically zero, sirf small nahi.

Why: PBH test kabhi kabhi rank test se preferred kyun hota hai?

PBH test (Popov–Belevitch–Hautus) ek baar mein ek eigenvalue

λ

check karta hai yeh poochh ke ki kya

[A - λ I C]

ka full column rank hai; ek rank drop kaunsa specific mode unobservable hai pinpoint karta hai — wo information jo

O

ka single rank number hide kar deta hai.

Why: Sirf

A

aur

C

kyun matter karte hain,

B

aur

D

nahi?

Kyunki

u (t)

ek signal hai jo hum choose karte hain aur jaante hain,

y

par iska poora effect (

B

aur feed-through

D

ke through) compute karke remove kiya ja sakta hai;

x (0)

reconstruct karne ke liye jo bacha rehta hai wo free response hai, jo sirf

C

aur

A

se govern hota hai.

Why: Observability Kalman Filter ke liye kyun matter karta hai?

Kalman Filter noisy outputs se state reconstruct karta hai; agar ek state

ker O

mein rehta hai toh wo koi signal emit nahi karta, toh uska estimate converge nahi ho sakta. Observability ek prerequisite hai estimation error ke zero tak drive hone ke liye.

Why: Kalman decomposition system ko observable aur unobservable parts mein kyun split karta hai?

Kyunki

ker O

A

-invariant subspace hai: iske andar ke states hamesha hidden rehte hain, toh Kalman decomposition system ko ek visible block (reconstructable) aur ek invisible block mein separate karta hai jise koi bhi sensor record kabhi expose nahi kar sakta.

Edge cases

Edge: Jab

n = 1

ho (scalar state) toh

O

kya hai aur uska rank kya hoga?

O = C

hai, ek single

p \times 1

block. Observable iff

C \neq = 0

hai; ek one-dimensional state ki koi bhi nonzero sensor reading already sab kuch dekh leti hai.

Edge: Agar

A = 0

ho (saare states frozen), toh

(0, C)

kab observable hai?

O = [C; 0; \dots; 0]

hai (semicolons

C

ke neeche zero blocks stack karte hain), toh rank

= rank (C)

hai. Observable iff

C

akele rank

n

hai, kyunki frozen dynamics ke saath derivatives kuch nahi add karti.

Edge: Agar

A = λ I

ho (ek single repeated eigenvalue poore

R^{n}

mein), toh ek scalar output kitna observable ho sakta hai?

Har

C A^{k} = λ^{k} C

hai jo

C

ka scalar multiple hai, toh

O

ka rank

1

hai. Ek scalar sensor at most ek direction resolve kar sakta hai; aisi fully-degenerate mode observe karne ke liye tumhare paas

p \geq n

chahiye.

Edge: Kya ek system simultaneously observable aur completely uncontrollable ho sakta hai?

Haan — observability (

A, C

ki property) aur controllability (

A, B

ki property) independent hain. Tum ek aisi state perfectly reconstruct kar sakte ho jise tum bilkul steer nahi kar sakte.

Edge: Agar

O

ka rank exactly

n - 1

ho, toh hidden subspace kitna bada hoga?

Unobservable subspace

ker O

ki dimension

n - rank (O) = 1

hai; exactly ek independent state direction invisible hai, jabki baaki

n - 1

reconstruct ki ja sakti hain. Rank and null space of a matrix dekho.

Edge:

C

ko ek nonzero constant se scale karne se observability badlti hai?

Nahi.

C

c \neq = 0

se multiply karne par

O

ki har row

c

se scale hoti hai, jo column rank unchanged rakhta hai — observability row space ki property hai, row magnitudes ki nahi.

Recall One-line self-test

Agar koi tumhe ek tall $O$ de aur kahe "dekho kitni saari rows hain — zaroor observable hoga," toh tum pehle kya check karte ho? Kya uska column rank $n$ ke barabar hai — rows aur unki sankhya ek distraction hai; sirf independent columns observability decide karte hain.

Connections

Observability matrix — rank test (parent)
Controllability matrix — rank test · Cayley–Hamilton theorem · PBH test
Kalman Filter · Kalman decomposition · State-space representation · Rank and null space of a matrix