Kalman gain — minimizes trace of covariance: Question bank

3.5.22 · D5 · Physics › Guidance, Navigation & Control (GNC) › Kalman gain — minimizes trace of covariance

Symbols ke reminders, taaki kuch bhi use hone se pehle samjha ja sake:

$x$ — (unknown) true state. $\overset{x}{^}$ — uska hamara estimate.
$e = x - \overset{x}{^}$ — estimation error, woh vector jo hamare guess se truth ki taraf point karta hai. Iska "measurement se pehle" wala version hai $e^{-} = x - \overset{x}{^}^{-}$ aur "baad" wala version hai $e^{+} = x - \overset{x}{^}^{+}$ .
$\overset{x}{^}^{-}, P^{-}$ — prediction (prior estimate) aur uski uncertainty (covariance) sensor dekhne se pehle. Yahan $P^{-} = E [e^{-} (e^{-})^{⊤}]$ .
$\overset{x}{^}^{+}, P^{+}$ — corrected estimate aur uski uncertainty measurement fold in karne ke baad, $P^{+} = E [e^{+} (e^{+})^{⊤}]$ .
$z = H x + v$ — measurement; $H$ state ko measurement space mein map karta hai; $v$ sensor noise hai jiska covariance $R$ hai.
$K$ — Kalman gain, "trust dial."
$S = H P^{-} H^{⊤} + R$ — innovation covariance (surprise ka spread).
$tr (\cdot)$ — trace, ek matrix ki diagonal entries ka sum.
$J = tr (P^{+})$ — cost function jo hum $K$ ke upar minimize karte hain: total mean-squared error jo hum jitna chhota ho sake utna chahte hain.

Figure — Kalman gain — minimizes trace of covariance

True or false — justify karo

Measurement par zyada trust karna hamesha final uncertainty ko shrink karta hai.

False. Uncertainty

J = tr (P^{+})

K

ka ek U-shaped (parabolic) function hai; optimum ke baad, extra gain prediction error se zyada sensor noise andar le aata hai, toh uncertainty phir badhti hai (figure dekho).

Fused estimate dono prediction aur measurement se individually zyada certain ho sakta hai.

True. Do independent noisy sources combine karne se

P^{+}

mila jo

P^{-}

R

dono se chhota hai — jaise

P^{-} = 4, R = 1

deta hai

P^{+} = 0.8

, dono se neeche. Yahi gain fusion karne ki poori wajah hai.

Simplified update

P^{+} = (I - K H) P^{-}

kisi bhi gain

K

ke liye valid hai jo tum choose karo.

False. Yeh short form sirf optimal $K$ par valid hai, jahan noise term collapse ho jaata hai. Ek arbitrary

K

ke liye tumhe Joseph form

(I - K H) P^{-} (I - K H)^{⊤} + K R K^{⊤}

use karna hoga.

Agar sensor perfect hai (

R \to 0

) aur woh state ko fully observe karta hai (square, invertible

H

), toh filter apni prediction ko bilkul ignore karta hai.

True.

R \to 0

aur

H

invertible hone par,

K \to H^{- 1}

(scalar

H = 1

K \to 1

), toh

\overset{x}{^}^{+} = H^{- 1} z = x

noise tak — estimate measurement ban jaata hai aur prior override ho jaata hai. Agar

H

invertible nahi hai, toh sensor sirf state ke observed part ko pin down karta hai (Edge cases dekho).

Ek useless sensor (

R \to \infty

) estimate ko diverge kara deta hai.

False.

R \to \infty

hone par,

S \to \infty

aur

K \to 0

, toh

\overset{x}{^}^{+} = \overset{x}{^}^{-}

— filter simply apni prediction rakhta hai, calmly, bina kisi nuksan ke.

Kalman gain state ke sabse bade error component ko minimize karta hai.

False. Yeh trace minimize karta hai — sabhi diagonal variances ka sum (total mean-squared error), maximum entry nahi. Yeh poore bundle ko optimize karta hai, sirf worst axis ko nahi.

P

ka trace, estimate aur truth ke beech expected squared distance ke barabar hai.

True (given estimate unbiased hai).

tr (P) = E [e^{⊤} e] = E [∥ x - \overset{x}{^} ∥^{2}]

, kyunki diagonal variances sum karna

E [e_{i}^{2}]

sum karna hai — exactly total mean-squared error jab

E [e] = 0

Bada prior uncertainty

P^{-}

gain

K

ko higher push karta hai.

True. Prediction mein kam confidence hone par, filter measurement par zyada lean karta hai;

K = P^{-} H^{⊤} S^{- 1}

badhta hai

P^{-}

ke saath (ceiling tak jo

R

set karta hai).

Joseph form accidentally ek aisa covariance produce kar sakta hai jisme negative variances hon.

False. Joseph form

(I - K H) P^{-} (I - K H)^{⊤} + K R K^{⊤}

do positive-semidefinite pieces ka sum hai, toh yeh symmetric aur PSD rehta hai — isliye implementers ise numerical safety ke liye rakhte hain.

Innovation

z - H \overset{x}{^}^{-}

ka zero hona matlab estimate exactly correct hai.

False. Zero innovation sirf yeh batata hai ki sensor prediction se agree karta hai; dono phir bhi saath mein galat ho sakte hain. Yeh "no surprise" signal karta hai, "no error" nahi.

Spot the error

"Kyunki

R

sensor noise hai, gain

K = P^{-} H^{⊤} / R

hai."

Denominator

S = H P^{-} H^{⊤} + R

hona chahiye, sirf

R

nahi. Isme measurement space mein mapped prior uncertainty bhi shamil hai — dono sources ki confidence gain set karti hai.

"Optimum wahan hai jahan

\partial J / \partial K = 0

, toh yeh trace ka maximum bhi ho sakta hai."

Nahi ho sakta. Scalar case mein second derivative literally

2 S

hai jahan

S = H P^{-} H^{⊤} + R > 0

hai, toh yeh ek genuine minimum hai. Matrix case mein

K

matrix-valued hai, toh koi single "

2 S

" number nahi hai — balki Hessian (quadratic form

δ K \mapsto tr (δ K S δ K^{⊤})

) positive-definite hai kyunki

S ≻ 0

, yahi guarantee karta hai ki stationary point ek minimum hai.

P^{+}

derive karte waqt

K R K^{⊤}

term drop kar sakte hain kyunki sensor noise chhota hai."

General

K

ke liye kabhi drop mat karo — yahi woh cheez hai jo trade-off ko parabola banata hai. Sirf optimal

K

par yeh collapse hota hai: wahan

K S = P^{-} H^{⊤}

, toh

K R K^{⊤} = K (S - H P^{-} H^{⊤}) K^{⊤}

extra pieces cancel kar deta hai, tidy

P^{+} = (I - K H) P^{-}

chhod jaata hai.

E [e^{-} v^{⊤}]

jaisi cross terms

P^{+}

mein appear honi chahiye."

Yeh vanish ho jaate hain kyunki prior error

e^{-} = x - \overset{x}{^}^{-}

aur sensor noise

v

uncorrelated assume kiye jaate hain (aur dono zero-mean hain), toh

E [e^{-} v^{⊤}] = 0

. Yahi woh cheez hai jo covariance ko do clean terms mein split hone deti hai.

"Update

\overset{x}{^}^{+} = \overset{x}{^}^{-} + K (z - H \overset{x}{^}^{-})

measurement mein nonlinear hai, toh yeh sirf luck se optimal hai."

Yeh

z

mein linear hai — sabse simple unbiased blend — aur Gaussian noise ke liye yeh linear form provably optimal hai, lucky nahi.

tr (K H P^{-})

aur

tr (P^{-} H^{⊤} K^{⊤})

alag terms hain, toh derivative mein 2 ka factor nahi aata."

Yeh equal hain, kyunki ek matrix ka trace uske transpose ke trace ke barabar hota hai; do equal terms add karne se derivative mein 2 ka factor aata hai.

Why questions

P

ka determinant ki jagah trace minimize kyun karein?

Kyunki

tr (P) = E [∥ x - \overset{x}{^} ∥^{2}]

total mean-squared error hai — natural "main kitna galat hun" number — aur yeh ek clean linear-algebra minimization deta hai jo standard Kalman gain deta hai.

tr (P)

state components ke variances ka sum kyun hai?

Diagonal entry

P_{ii} = E [e_{i}^{2}]

i

-th error component ka variance hai (

E [e_{i}] = 0

use karke); trace diagonal sum karta hai, toh woh variances ko ek total-error scalar mein sum karta hai.

General nonlinear correction ki jagah linear update form kyun choose kiya jaata hai?

Linear sabse simple unbiased blend hai, aur Gaussian noise ke under yeh exactly optimal hai, toh ise complicated banane se koi accuracy gain nahi hoti.

Trace

K

mein parabola (U-shape) ki tarah behave kyun karta hai?

K R K^{⊤}

term

K

mein quadratic aur positive-definite hai, jabki prediction term smoothly shrink hoti hai — unka sum ek positive quadratic hai jisme ek minimum hai, jaise figure ka green curve dikhata hai.

Optimal gain par

K R K^{⊤}

term exactly kyun collapse hota hai?

Optimum par

K S = P^{-} H^{⊤}

, toh back substitute karne par noise term extra prediction pieces cancel kar deta hai, tidy

P^{+} = (I - K H) P^{-}

chhod jaata hai — yeh coincidence sirf minimizing

K

par hota hai.

Real code mein Joseph form kyun rakhein agar

P^{+} = (I - K H) P^{-}

simpler hai?

Joseph form rounding error aur slightly non-optimal gains ke saath bhi symmetric aur positive-semidefinite rehta hai, numerically "impossible" covariances se bachata hai.

Innovation covariance

S

dono

H P^{-} H^{⊤}

aur

R

combine kyun karta hai?

Surprise

z - H \overset{x}{^}^{-}

prediction ki uncertainty (

H

se mapped) aur sensor ki uncertainty dono carry karta hai; tumhe innovation ko us cheez ke total spread se weight karna hoga jo use cause kar sakti hai.

R

badhana (noisier sensor) gain kyun ghatata hai?

Noisier sensor kam trustworthy hai, toh

S

badhta hai aur

K = P^{-} H^{⊤} S^{- 1}

shrink hota hai, estimate ko prediction ki taraf wapas pull karta hai.

Trace ka matlab "total error" hone ke liye estimate unbiased kyun honi chahiye?

Agar

E [e] \neq = 0

, toh covariance

P

sirf (galat) average ke around spread measure karta hai, average miss khud nahi;

tr (P)

minimize karna phir ek systematic bias untouched chhod deta. Zero-mean error

tr (P) = E [∥ e ∥^{2}]

exactly banata hai.

Edge cases

K

aur

\overset{x}{^}^{+}

kya hote hain jab prediction perfectly certain ho,

P^{-} \to 0

K \to 0

, toh

\overset{x}{^}^{+} = \overset{x}{^}^{-}

— ek flawless prior sensor ko bilkul override karta hai, perfect-sensor case ka mirror-image.

Agar prediction aur measurement equally uncertain hain (scalar

P^{-} = R

H = 1

), toh gain kahan land karta hai?

K = P^{-} / (P^{-} + R) = 1/2

— filter exactly halfway difference split karta hai, jaise symmetry demand karta hai.

Agar

H

non-square hai (states se kam measurements, toh ise invert nahi kiya ja sakta)?

Gain phir bhi

K = P^{-} H^{⊤} S^{- 1}

ke zariye exist karta hai, kyunki sirf

S = H P^{-} H^{⊤} + R

(ek chhota matrix) invert hota hai, kabhi

H

nahi. Sensor sirf state ke observed directions ko sharpen karta hai; unobserved directions apni prior uncertainty rakhte hain.

Formula

H^{- 1}

ki jagah

S^{- 1}

kyun use karta hai, given

H

invertible nahi bhi ho sakta?

Kyunki humhe kabhi

H

invert nahi karna — derivation

K

ke upar minimize karta hai aur sirf hamesha-square, positive-definite

S

invert karta hai. "

K \to H^{- 1}

" statement ek special case hai jisme square invertible

H

chahiye; general formula ise completely sidestep karta hai.

Agar

K

deliberately zero set kar diya jaaye toh

P^{+}

ka kya hoga?

Joseph form deta hai

P^{+} = P^{-}

: koi measurement use nahi hoti, toh uncertainty unchanged rehti hai — filter sirf prediction par coast karta hai.

Kya optimal gain ke liye

P^{+}

kabhi

P^{-}

se exceed kar sakta hai?

Nahi — optimum par

P^{+} = (I - K H) P^{-}

sirf

P^{-}

ke barabar ya usse chhota ho sakta hai; jab sahi se weight kiya jaaye toh measurement information kabhi uncertainty nahi badhati.

Zero measurement noise aur zero prediction uncertainty (

R \to 0, P^{-} \to 0

) kya imply karta hai?

Dono sources perfect certainty claim karte hain;

S \to 0

K

ko ill-defined bana deta hai (

0/0

), ek degenerate, over-confident setup signal karta hai jise practice mein regularize karna hoga.

Connections

Kalman gain — minimizes trace of covariance (parent)
Kalman Filter — Update Step
Covariance Matrices and Uncertainty
Innovation and Residuals
Joseph Form Covariance Update
Sensor Fusion in Navigation
Least Squares Estimation
Extended Kalman Filter (EKF) in GNC