! " ! # $ ! % ! & ' ( ) * + , - . . / 2 ) ; ( ln 1 ) ` var( 0 0 0 x ! " See Appendix 3A Ior Derivation Sometimes it is easier to Iind the CRLB this way. This also gives a new viewpoint oI the CRLB: From Gardner`s Paper (#""" %&'()* +( #(,+ %-.+&/, July 1979) Consider the Normalized version oI this Iorm oI CRLB Posted on BB ! " ! # $ ! % ! & ' ( ) * + , - . . / 2 2 2 ) ; ( ln 1 ) ` var( 0 0 0 0 0 x ! " We`ll 'derive this in a way that will re- interpret the CRLB 2 Consider the 'Incremental Sensitivity oI !(x;0 ) to changes in 0 0 II 0 1 0 120, then it causes !(x;0 ) 1 !(x;0 120 ) How sensitive is !(x;0 ) to that change?? ( ) * + , - ( ) * + , - 2 2 3 3 ( ) * + , - 2 ( ) * + , - 2 3 2 ) ; ( ) ; ( in change ) ; ( in change ) ; ( ) ; ( ) ( ~ 0 0 0 0 0 0 0 0 0 0 0 x x x x x x ! ! ! ! ! 2 ! Now let 20 1 0: 0 0 0 0 0 0 0 0 0 0 . . 3 ( ) * + , - ( ) * + , - . . 3 3 1 2 ) ; ( ln ) ; ( ) ; ( ) ( ~ lim ) ( 0 x x x x x ! ! ! 2 2 ! ! Recall Irom Calculus: 3 3 , 3 , 3 3 , . . 3 . . ) ( ) ( 1 ) ( ln 4 5 " # $ % & ' 3 ! " ! # $ ! % ! & ' ( ) * + , - . . / 2 2 2 2 2 ) ( 1 ) ; ( ln 1 ) ` var( x x ! 2 " ! " 0 0 0 0 0 0 0 Interpretation Norm. CRLB Inverse Mean Square Sensitivity 3 Definition of Fisher Information The denominator in CRLB is called the Fisher InIormation #(0 ) It is a measure oI the 'expected goodness oI the data Ior the purpose oI making an estimate ! " ! # $ ! % ! & ' . . 6 3 2 2 ) ; ( ln ) ( 0 0 0 x ! " # Has the needed properties Ior 'inIo (as does 'Shannon InIo): 1. #(0 ) / 0 (easy to see using the alternate Iorm oI CRLB) 2. #(0 ) is additive Ior independent observations Iollows Irom: 4 5 7 8 3 ( ( ) * + + , - 3 ( ( ( 3 ! ( 3 ! ! ) |; | ( ln ) |; | ( ln ) ; ( ln 0 0 0 x II each # ( (0 ) is the same: #(0 ) 49#(0 ) 4 3.5 CRLB for Signals in AWGN When we have the case that our data is 'signal AWGN then we get a simple Iorm Ior the CRLB: Signal Model: 3|(| )|(;0 | 5|(|, ( 0, 1, 2, . , 4-1 White, Gaussian, Zero Mean Q: What is the CRLB? First write the likelihood Iunction: : ; : ; ! " ! # $ ! % ! & ' 6 6 3 7 6 3 1 0 2 2 2 / 2 | ; | | | 2 1 exp 2 1 ) ; ( 4 ( 4 ( ) ( 3 ! 0 < =< 0 x DiIIerentiate Log LF twice to get: : ; 7 6 3 ! " ! # $ ! % ! & ' ( ) * + , - . . 6 . . 6 3 . . 1 0 2 2 2 2 2 2 | ; | | ; | | ; | | | 1 ) ; ( ln 4 ( ( ) ( ) ( ) ( 3 ! 0 0 0 0 0 < 0 0 x Depends on random 3|(| so must take E} 5 > ? 2 1 0 2 1 0 2 2 2 0 | ; | 2 2 2 | ; | | ; | | ; | | ; | | | 1 ) ; ( ln < 0 0 0 0 0 0 0 < 0 0 0 7 7 6 3 6 3 3 ( ) * + , - . . 6 3 ! ! " ! ! # $ ! ! % ! ! & ' ( ) * + , - . . 6 . . @ @ @ A B C C C D E 6 3 ! " ! # $ ! % ! & ' . . 4 ( 4 ( ( ) ( ) ( ) ( ) ( ) ( 3 " ! " ! ! ! " ! ! ! # $ " # $ x Then using this we get the CRLB for Signal in AWGN: 7 6 3 ( ) * + , - . . / 1 0 2 2 | ; | ) ` var( 4 ( ( ) 0 0 < 0 2 | ; | ( ) * + , - . . 0 0 ( ) Note: tells how sensitive signal is to parameter II signal is very sensitive to parameter change. then CRLB is small . can get very accurate estimate! 6 Ex. 3.5: CRLB of Frequency of Sinusoid Signal Model: 1 , , 2 , 1 , 0 0 | | ) 2 cos( | | 2 1 6 3 F F G G 3 4 ( , ( 5 ( , 6 ( 3 + + % H = 0 0.1 0.2 0.3 0.4 0.5 0 2 4 6 x 10 -4 I o (cycles/sample) C R L B
( c y c l e s / s a m p l e ) 2 0 0.1 0.2 0.3 0.4 0.5 0.01 0.015 0.02 0.025 I o (cycles/sample) C R L B 1 / 2
( c y c l e s / s a m p l e ) Error in Book Bound on Variance Bound on Std. Dev. 4 5 7 6 3 G 9 / 1 0 2 ) 2 sin( 2 1 ) ` var( 4 ( + ( , ( 247 H = = 0 Signal is less sensitive iI , + near 0 or 7 3.6 Transformation of Parameters Say there is a parameter 0 with known CRLB 0 But imagine that we instead are interested in estimating some other parameter I that is a Iunction oI 0 : I 8(0 ) Q: What is CRLB I ? 0 I 0 0 I 97:; 8 97:; 2 ) ( ) var( @ A B C D E . . 3 / Captures the sensitivity oI I to 0 Proved in Appendix 3B Large .g/. 0 1 small error in 0 gives larger error in I 1 increases CRLB (i.e., worsens accuracy) 8 Example: Speed of Vehicle From Elapsed Time Known Distance < start Laser Sensor Sensor Laser stop Measure Elapsed Time % Possible Accuracy Set by CRLB % % % % = 97:; < = 97:; % < 97:; % < % 97:; 9 3 9 @ A B C D E 6 3 9 ( ) * + , - @ A B C D E . . 3 2 4 2 2 But. really want to measure speed = >/% Find the CRLB = : ) / ( 2 ) ? 97:; < = % = / < Accuracy Bound Less accurate at High Speeds (quadratic) More accurate over large distances 2 9 Effect of Transformation on Efficiency Suppose you have an eIIicient estimator oI 0 : 0 ` But. you are really interested in estimating I 8(0 ) Suppose you plan to use ) ` ( ` 0 I 8 3 Q: Is this an eIIicient estimator oI I ??? A: Theorem: II 8(0 ) has Iorm 8(0 ) '0 @, then is eIIicient. ) ` ( ` 0 I 8 3 ~affine transform Proof: First: : ; : ; : ; 0 0 0 I 97:; ' ' @ ' 2 2 ` var ` var ` var 3 3 G 3 ' because 'eIIicient Now, what is CRB I ? Using transIormation result: : ; 0 0 I 0 0 97:; ' 97:; @ ' 97:; ' 2 2 2 3 ( ) * + , - . G . 3 3 ! !" ! !# $ : ; I I 97:; 3 ` var EIIicient! 10 Asymptotic Efficiency Under Transformation II the mapping I 8(0 ) is not aIIine. this result does NOT hold But. iI the number oI data samples used is large, then the estimator is approximately eIIicient ('Asymptotically Efficient) 0 ` ) ` ( ), ` ( ` 0 0 I ! 8 3 0 ` oI pdI Small ! Case PDF is widely spread over nonlinear mapping 0 ` ) ` ( ), ` ( ` 0 0 I ! 8 3 0 ` oI pdI Large ! Case PDF is concentrated onto linearized section