Abbreviated: CRLB or sometimes just CRB CRLB is a lower bound on the variance oI any !"#$%&'( estimator: The CRLB tells us the best we can ever expect to be able to do (w/ an unbiased estimator) then , oI estimator unbiased an is ` II ! ! ) ( ) ( ) ( ) ( ` ` ` 2 ` ! ! " ! ! " ! ! ! ! !"#$ !"#$ # $ # What is the Cramer-Rao Lower Bound 1. Feasibility studies ( e.g. Sensor useIulness, etc.) Can we meet our speciIications? 2. Judgment oI proposed estimators Estimators that don`t achieve CRLB are looked down upon in the technical literature 3. Can sometimes provide Iorm Ior MVU est. 4. Demonstrates importance oI physical and/or signal parameters to the estimation problem e.g. We`ll see that a signal`s BW determines delay est. accuracy $ Radars should use wide BW signals Some Uses of the CRLB Q: What determines how well you can estimate ! ? Recall: Data vector is x 3.3 Est. Accuracy Consideration samples Irom a random process that depends on an ! $ the PDF describes that dependence: %(x;! ) Clearly iI %(x;! ) depends strongly/weakly on ! .we should be able to estimate ! well/poorly. See surIace plots vs. x & ! Ior 2 cases: 1. Strong dependence on ! 2. Weak dependence on ! $ Should look at %(x;! ) as a Iunction oI ! Ior Iixed value oI observed data x Surface Plot Examples of )(x;! ) % & ' ' ( ) * * + , - - . 2 2 2 2 ) | 0 | ( exp 2 1 |; 0 | " /" & ' & ' % '|0| & (|0| Ex. 3.1: PDF Dependence for DC Level in Noise (|0| ~ )(0," 2 ) Then the parameter-dependent PDF oI the data point '|0| is: & '|0| & 3 %('|0|3;! ) Say we observe *0] 3. So ~Slice at *0] 3 The LF the PDF %(x;! ) .but as a Iunction oI parameter ! w/ the data vector x Iixed Define: Likelihood Function (LF) We will also oIten need the Log Likelihood Function (LLF): LLF lnLF} ln %(x;! )} LF Characteristics that Affect Accuracy Intuitively: 'sharpness oI the LF sets accuracy. But How??? Sharpness is measured using curvature: % & value true data given 2 2 ; ln . . 0 0 - ! ! ! x x % Curvature ! $ PDF concentration ! $ Accuracy ! But this is Ior a particular set oI data. we want 'in general: So.Average over random vector to give the average curvature: % & value true 2 2 ; ln . 1 2 1 3 4 1 5 1 6 7 0 0 - ! ! ! x % * 'Expected sharpness oI LF *} is w.r.t %(x;! ) Theorem 3.1 CRLB for Scalar Parameter Assume 'regularity condition is met: ! ! ! 8 . 2 3 4 5 6 7 0 0 0 ) ; ( ln ' % * *} is w.r.t %(x;! ) Then % & value true 2 2 2 ` ; ln 1 . 1 2 1 3 4 1 5 1 6 7 0 0 - # ! ! ! ! " x % * 3.4 Cramer-Rao Lower Bound % & % & +' ' % % % * ) ; ( ; ln ; ln 2 2 2 2 ! ! ! ! ! 9 0 0 . 1 2 1 3 4 1 5 1 6 7 0 0 x x Right-Hand Side is CRLB 1. Write log 1ikelihood Iunction as a Iunction oI !: ln %(x;! ) 2. Fix x and take 2 nd partial oI LLF: 0 2 ln %(x;! )/0! 2 3. II result still depends on x: Fix ! and take expected value w.r.t. x Otherwise skip this step 4. Result may still depend on !: Evaluate at each speciIic value oI ! desired. 5. Negate and Iorm reciprocal Steps to Find the CRLB Need likelihood Iunction: % & : ; % & % & : ; % & ' ' ' ' ( ) * * * * + , - < - . ' ' ( ) * * + , - - . - . - . = 2 2 1 0 2 2 1 0 2 2 2 2 exp 2 1 2 exp 2 1 ; " /" " /" & , ' & , ' & % ) , ) ) , x Example 3.3 CRLB for DC in AWGN '|,| & (|,|, , 0, 1, . , ) 1 (|,| ~ )(0," 2 ) & white Due to whiteness Property oI exp % & : ; % & ! ! ! " ! ! ! # $ ! ! " ! ! # $ ? (~~) 1 0 2 2 0 (~~) 2 2 2 1 2 ln ) ; ( ln . 0 0 - . . 0 0 < - - ' ' ( ) * * + , - . & ) , & ) & , ' & % " /" x Now take ln to get LLF: : ; % & % & & ' ) & , ' & % & ) , - . - . 0 0 < - . 2 1 0 2 1 ) ; ( ln " " x Now take Iirst partial w.r.t. &: sample mean (!) Now take partial again: 2 2 2 ) ; ( ln " ) & % & - . 0 0 x Doesn`t depend on x so we don`t need to do E} Since the result doesn`t depend on x or & all we do is negate and Iorm reciprocal to get CRLB: % & ) % * !"#$ 2 value true 2 2 ; ln 1 " ! ! ! . 1 2 1 3 4 1 5 1 6 7 0 0 - . . x ) & 2 } ` var " # Doesn`t depend on + Increases linearly with " 2 Decreases inversely with , CRLB For Iixed ) " 2 & CRLB For Iixed ) & " 2 CRLB Doubling Data Halves CRLB! ) For Iixed " 2 Continuation of Theorem 3.1 on CRLB There exists an unbiased estimator that attains the CRLB iff: : ; ! ! ! ! - . 0 0 ) ( ) ( ) ; ( ln x x . / % Ior some Iunctions /(! ) and .(x) Furthermore, the estimator that achieves the CRLB is then given by: ) ( ` x . . ! (!) Since no unbiased estimator can do better. this is the MVU estimate!! This gives a possible way to Iind the MVU: Compute 0ln %(x;! )/0! (need to anyway) Check to see iI it can be put in Iorm like (!) II so. then .(x) is the MVU esimator with !"#$ / . . ) ( 1 } ` var ! ! Revisit Example 3.3 to Find MVU Estimate % & & ' ) & % & - . 0 0 2 ) ; ( ln " x For DC Level in AWGN we Iound in (!) that: Has Iorm oI /(&)|g(x) &| : ; < - . . . . 1 0 1 ) ( ` ) , , ' ) ' . x ! !"#$ ) & ) & / . . $ . 2 2 } ` var ) ( " " So. for the DC Level in AWGN: the sample mean is the MVUE!! Definition: Efficient Estimator An estimator that is: unbiased and attains the CRLB is said to be an 'EIIicient Estimator Notes: Not all estimators are eIIicient (see next example: Phase Est.) Not even all MVU estimators are eIIicient So. there are times when our ~1 st partial test won`t work!!!! Example 3.4: CRLB for Phase Estimation This is related to the DSB carrier estimation problem we used Ior motivation in the notes Ior Ch. 1 Except here. we have a pure sinusoid and we only wish to estimate only its phase Signal Model: | | ) 2 cos( | | | ; | , ( , 0 & , ' 1 , 2 1 1 > > . ! ! ! " ! ! ! # $ ? ? / AWGN w/ zero mean & " 2 Assumptions: 1. 0 0 1 ( 0 1 is in cycles/sample) 34 & and 0 1 are known (we`ll remove this assumption later) Signal-to-Noise Ratio: Signal Power & 2 /2 Noise Power " 2 2 2 2" & 5)" . Problem: Find the CRLB Ior estimating the phase. We need the PDF: % & % & : ; % & ' ' ' ' ( ) * * * * + , > - < - . - . 2 2 1 0 2 2 2 ) 2 cos( exp 2 1 ; " ? / /" ? , 0 & , ' % 1 ) , ) x Exploit Whiteness and Exp. Form Now taking the log gets rid oI the exponential, then taking partial derivative gives (see book Ior details): % & : ; 2 1 0 2 ) 2 4 sin( 2 ) 2 sin( ; ln @ A B C D E > - > < - . 0 0 - . ? / ? / " ? ? , 0 & , 0 , ' & % 1 1 ) , x Taking partial derivative again: % & : ; % & ) 2 4 cos( ) 2 cos( ; ln 1 0 2 2 2 ? / ? / " ? ? > - > < - . 0 0 - . , 0 & , 0 , ' & % 1 1 ) , x Still depends on random vector x. so need E} Taking the expected value: % & : ; % & : ; F G % & ) 2 4 cos( ) 2 cos( ) 2 4 cos( ) 2 cos( ; ln 1 0 2 1 0 2 2 2 ? / ? / " ? / ? / " ? ? > - > < . 2 3 4 5 6 7 > - > < . 1 2 1 3 4 1 5 1 6 7 0 0 - - . - . , 0 & , 0 , ' * & , 0 & , 0 , ' & * % * 1 1 ) , 1 1 ) , x *'|,|} A cos(2/ 0 1 , 6 ? ) So. plug that in, get a cos 2 term, use trig identity, and get % & 5)" ) )& , 0 & % * ) , 1 ) , H . I ' ' ( ) * * + , > - . 1 2 1 3 4 1 5 1 6 7 0 0 - < < - - - . 2 2 1 0 1 0 2 2 2 2 2 ) 2 4 cos( 1 2 ; ln " ? / " ? ? x ) ) iI 0 1 not near 0 or , )-1 Now. invert to get CRLB: 5)" ) H # 1 } ` var? CRLB Doubling Data Halves CRLB! ) For Iixed 5)" Non-dB CRLB Doubling SNR Halves CRLB! 5)" (non-dB) For Iixed ) Halve CRLB Ior every 3B in SNR Does an eIIicient estimator exist Ior this problem? The CRLB theorem says there is only iI : ; ! ! ! ! - . 0 0 ) ( ) ( ) ; ( ln x x . / % % & : ; 2 1 0 2 ) 2 4 sin( 2 ) 2 sin( ; ln @ A B C D E > - > < - . 0 0 - . ? / ? / " ? ? , 0 & , 0 , ' & % 1 1 ) , x Our earlier result was: EIIicient Estimator does NOT exist!!! We`ll see later though, an estimator Ior which !"#$ J } ` var? as ) J K or as 5)" J K Such an estimator is called an 'asymptotically eIIicient estimator (We`ll see such a phase estimator in Ch. 7 on MLE) CRB ) } ` var?